blob: 3595403921bd5be10c3e5e591bf04916e654423d [file] [log] [blame]
Linus Torvalds1da177e2005-04-16 15:20:36 -07001/*
Peter Zijlstra391e43d2011-11-15 17:14:39 +01002 * kernel/sched/core.c
Linus Torvalds1da177e2005-04-16 15:20:36 -07003 *
4 * Kernel scheduler and related syscalls
5 *
6 * Copyright (C) 1991-2002 Linus Torvalds
7 *
8 * 1996-12-23 Modified by Dave Grothe to fix bugs in semaphores and
9 * make semaphores SMP safe
10 * 1998-11-19 Implemented schedule_timeout() and related stuff
11 * by Andrea Arcangeli
12 * 2002-01-04 New ultra-scalable O(1) scheduler by Ingo Molnar:
13 * hybrid priority-list and round-robin design with
14 * an array-switch method of distributing timeslices
15 * and per-CPU runqueues. Cleanups and useful suggestions
16 * by Davide Libenzi, preemptible kernel bits by Robert Love.
17 * 2003-09-03 Interactivity tuning by Con Kolivas.
18 * 2004-04-02 Scheduler domains code by Nick Piggin
Ingo Molnarc31f2e82007-07-09 18:52:01 +020019 * 2007-04-15 Work begun on replacing all interactivity tuning with a
20 * fair scheduling design by Con Kolivas.
21 * 2007-05-05 Load balancing (smp-nice) and other improvements
22 * by Peter Williams
23 * 2007-05-06 Interactivity improvements to CFS by Mike Galbraith
24 * 2007-07-01 Group scheduling enhancements by Srivatsa Vaddagiri
Ingo Molnarb9131762008-01-25 21:08:19 +010025 * 2007-11-29 RT balancing improvements by Steven Rostedt, Gregory Haskins,
26 * Thomas Gleixner, Mike Kravetz
Linus Torvalds1da177e2005-04-16 15:20:36 -070027 */
28
29#include <linux/mm.h>
30#include <linux/module.h>
31#include <linux/nmi.h>
32#include <linux/init.h>
Ingo Molnardff06c12007-07-09 18:52:00 +020033#include <linux/uaccess.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070034#include <linux/highmem.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070035#include <asm/mmu_context.h>
36#include <linux/interrupt.h>
Randy.Dunlapc59ede72006-01-11 12:17:46 -080037#include <linux/capability.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070038#include <linux/completion.h>
39#include <linux/kernel_stat.h>
Ingo Molnar9a11b49a2006-07-03 00:24:33 -070040#include <linux/debug_locks.h>
Ingo Molnarcdd6c482009-09-21 12:02:48 +020041#include <linux/perf_event.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070042#include <linux/security.h>
43#include <linux/notifier.h>
44#include <linux/profile.h>
Nigel Cunningham7dfb7102006-12-06 20:34:23 -080045#include <linux/freezer.h>
akpm@osdl.org198e2f12006-01-12 01:05:30 -080046#include <linux/vmalloc.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070047#include <linux/blkdev.h>
48#include <linux/delay.h>
Pavel Emelyanovb4888932007-10-18 23:40:14 -070049#include <linux/pid_namespace.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070050#include <linux/smp.h>
51#include <linux/threads.h>
52#include <linux/timer.h>
53#include <linux/rcupdate.h>
54#include <linux/cpu.h>
55#include <linux/cpuset.h>
56#include <linux/percpu.h>
Alexey Dobriyanb5aadf72008-10-06 13:23:43 +040057#include <linux/proc_fs.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070058#include <linux/seq_file.h>
Nick Piggine692ab52007-07-26 13:40:43 +020059#include <linux/sysctl.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070060#include <linux/syscalls.h>
61#include <linux/times.h>
Jay Lan8f0ab512006-09-30 23:28:59 -070062#include <linux/tsacct_kern.h>
bibo maoc6fd91f2006-03-26 01:38:20 -080063#include <linux/kprobes.h>
Shailabh Nagar0ff92242006-07-14 00:24:37 -070064#include <linux/delayacct.h>
Ingo Molnardff06c12007-07-09 18:52:00 +020065#include <linux/unistd.h>
Jens Axboef5ff8422007-09-21 09:19:54 +020066#include <linux/pagemap.h>
Peter Zijlstra8f4d37e2008-01-25 21:08:29 +010067#include <linux/hrtimer.h>
Reynes Philippe30914a52008-03-17 16:19:05 -070068#include <linux/tick.h>
Peter Zijlstraf00b45c2008-04-19 19:45:00 +020069#include <linux/debugfs.h>
70#include <linux/ctype.h>
Steven Rostedt6cd8a4b2008-05-12 21:20:42 +020071#include <linux/ftrace.h>
Tejun Heo5a0e3ad2010-03-24 17:04:11 +090072#include <linux/slab.h>
Carsten Emdef1c6f1a2011-10-26 23:14:16 +020073#include <linux/init_task.h>
Al Viro40401532012-02-13 03:58:52 +000074#include <linux/binfmts.h>
Frederic Weisbecker91d1aa432012-11-27 19:33:25 +010075#include <linux/context_tracking.h>
Gideon Israel Dsouza52f5684c2014-04-07 15:39:20 -070076#include <linux/compiler.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070077
David Howells96f951e2012-03-28 18:30:03 +010078#include <asm/switch_to.h>
Eric Dumazet5517d862007-05-08 00:32:57 -070079#include <asm/tlb.h>
Satyam Sharma838225b2007-10-24 18:23:50 +020080#include <asm/irq_regs.h>
Christian Borntraegerdb7e5272012-01-11 08:58:16 +010081#include <asm/mutex.h>
Glauber Costae6e66852011-07-11 15:28:17 -040082#ifdef CONFIG_PARAVIRT
83#include <asm/paravirt.h>
84#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -070085
Peter Zijlstra029632f2011-10-25 10:00:11 +020086#include "sched.h"
Tejun Heoea138442013-01-18 14:05:55 -080087#include "../workqueue_internal.h"
Thomas Gleixner29d5e042012-04-20 13:05:45 +000088#include "../smpboot.h"
Gregory Haskins6e0534f2008-05-12 21:21:01 +020089
Steven Rostedta8d154b2009-04-10 09:36:00 -040090#define CREATE_TRACE_POINTS
Steven Rostedtad8d75f2009-04-14 19:39:12 -040091#include <trace/events/sched.h>
Steven Rostedta8d154b2009-04-10 09:36:00 -040092
Peter Zijlstra029632f2011-10-25 10:00:11 +020093DEFINE_MUTEX(sched_domains_mutex);
94DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues);
Peter Zijlstradc61b1d2010-06-08 11:40:42 +020095
Peter Zijlstrafe44d622010-12-09 14:15:34 +010096static void update_rq_clock_task(struct rq *rq, s64 delta);
Venkatesh Pallipadi305e6832010-10-04 17:03:21 -070097
Peter Zijlstra029632f2011-10-25 10:00:11 +020098void update_rq_clock(struct rq *rq)
Peter Zijlstra3e51f332008-05-03 18:29:28 +020099{
Peter Zijlstrafe44d622010-12-09 14:15:34 +0100100 s64 delta;
Venkatesh Pallipadi305e6832010-10-04 17:03:21 -0700101
Peter Zijlstra9edfbfe2015-01-05 11:18:11 +0100102 lockdep_assert_held(&rq->lock);
103
104 if (rq->clock_skip_update & RQCF_ACT_SKIP)
Mike Galbraithf26f9af2010-12-08 11:05:42 +0100105 return;
Venkatesh Pallipadiaa483802010-10-04 17:03:22 -0700106
Peter Zijlstrafe44d622010-12-09 14:15:34 +0100107 delta = sched_clock_cpu(cpu_of(rq)) - rq->clock;
Mike Galbraith4036ac12014-06-24 07:49:40 +0200108 if (delta < 0)
109 return;
Peter Zijlstrafe44d622010-12-09 14:15:34 +0100110 rq->clock += delta;
111 update_rq_clock_task(rq, delta);
Peter Zijlstra3e51f332008-05-03 18:29:28 +0200112}
113
Ingo Molnare436d802007-07-19 21:28:35 +0200114/*
Ingo Molnarbf5c91b2007-10-15 17:00:04 +0200115 * Debugging: various feature bits
116 */
Peter Zijlstraf00b45c2008-04-19 19:45:00 +0200117
118#define SCHED_FEAT(name, enabled) \
Peter Zijlstraf00b45c2008-04-19 19:45:00 +0200119 (1UL << __SCHED_FEAT_##name) * enabled |
120
121const_debug unsigned int sysctl_sched_features =
Peter Zijlstra391e43d2011-11-15 17:14:39 +0100122#include "features.h"
Peter Zijlstraf00b45c2008-04-19 19:45:00 +0200123 0;
124
125#undef SCHED_FEAT
126
127#ifdef CONFIG_SCHED_DEBUG
128#define SCHED_FEAT(name, enabled) \
129 #name ,
130
Hiroshi Shimamoto12925312012-05-25 15:41:54 +0900131static const char * const sched_feat_names[] = {
Peter Zijlstra391e43d2011-11-15 17:14:39 +0100132#include "features.h"
Peter Zijlstraf00b45c2008-04-19 19:45:00 +0200133};
134
135#undef SCHED_FEAT
136
Li Zefan34f3a812008-10-30 15:23:32 +0800137static int sched_feat_show(struct seq_file *m, void *v)
Peter Zijlstraf00b45c2008-04-19 19:45:00 +0200138{
Peter Zijlstraf00b45c2008-04-19 19:45:00 +0200139 int i;
140
Peter Zijlstraf8b6d1c2011-07-06 14:20:14 +0200141 for (i = 0; i < __SCHED_FEAT_NR; i++) {
Li Zefan34f3a812008-10-30 15:23:32 +0800142 if (!(sysctl_sched_features & (1UL << i)))
143 seq_puts(m, "NO_");
144 seq_printf(m, "%s ", sched_feat_names[i]);
Peter Zijlstraf00b45c2008-04-19 19:45:00 +0200145 }
Li Zefan34f3a812008-10-30 15:23:32 +0800146 seq_puts(m, "\n");
Peter Zijlstraf00b45c2008-04-19 19:45:00 +0200147
Li Zefan34f3a812008-10-30 15:23:32 +0800148 return 0;
Peter Zijlstraf00b45c2008-04-19 19:45:00 +0200149}
150
Peter Zijlstraf8b6d1c2011-07-06 14:20:14 +0200151#ifdef HAVE_JUMP_LABEL
152
Ingo Molnarc5905af2012-02-24 08:31:31 +0100153#define jump_label_key__true STATIC_KEY_INIT_TRUE
154#define jump_label_key__false STATIC_KEY_INIT_FALSE
Peter Zijlstraf8b6d1c2011-07-06 14:20:14 +0200155
156#define SCHED_FEAT(name, enabled) \
157 jump_label_key__##enabled ,
158
Ingo Molnarc5905af2012-02-24 08:31:31 +0100159struct static_key sched_feat_keys[__SCHED_FEAT_NR] = {
Peter Zijlstraf8b6d1c2011-07-06 14:20:14 +0200160#include "features.h"
161};
162
163#undef SCHED_FEAT
164
165static void sched_feat_disable(int i)
166{
Peter Zijlstrae33886b2015-07-24 15:03:40 +0200167 static_key_disable(&sched_feat_keys[i]);
Peter Zijlstraf8b6d1c2011-07-06 14:20:14 +0200168}
169
170static void sched_feat_enable(int i)
171{
Peter Zijlstrae33886b2015-07-24 15:03:40 +0200172 static_key_enable(&sched_feat_keys[i]);
Peter Zijlstraf8b6d1c2011-07-06 14:20:14 +0200173}
174#else
175static void sched_feat_disable(int i) { };
176static void sched_feat_enable(int i) { };
177#endif /* HAVE_JUMP_LABEL */
178
Mel Gorman1a687c22012-11-22 11:16:36 +0000179static int sched_feat_set(char *cmp)
Peter Zijlstraf00b45c2008-04-19 19:45:00 +0200180{
Peter Zijlstraf00b45c2008-04-19 19:45:00 +0200181 int i;
Mel Gorman1a687c22012-11-22 11:16:36 +0000182 int neg = 0;
Peter Zijlstraf00b45c2008-04-19 19:45:00 +0200183
Hillf Danton524429c2011-01-06 20:58:12 +0800184 if (strncmp(cmp, "NO_", 3) == 0) {
Peter Zijlstraf00b45c2008-04-19 19:45:00 +0200185 neg = 1;
186 cmp += 3;
187 }
188
Peter Zijlstraf8b6d1c2011-07-06 14:20:14 +0200189 for (i = 0; i < __SCHED_FEAT_NR; i++) {
Mathieu Desnoyers77401912010-09-13 17:47:00 -0400190 if (strcmp(cmp, sched_feat_names[i]) == 0) {
Peter Zijlstraf8b6d1c2011-07-06 14:20:14 +0200191 if (neg) {
Peter Zijlstraf00b45c2008-04-19 19:45:00 +0200192 sysctl_sched_features &= ~(1UL << i);
Peter Zijlstraf8b6d1c2011-07-06 14:20:14 +0200193 sched_feat_disable(i);
194 } else {
Peter Zijlstraf00b45c2008-04-19 19:45:00 +0200195 sysctl_sched_features |= (1UL << i);
Peter Zijlstraf8b6d1c2011-07-06 14:20:14 +0200196 sched_feat_enable(i);
197 }
Peter Zijlstraf00b45c2008-04-19 19:45:00 +0200198 break;
199 }
200 }
201
Mel Gorman1a687c22012-11-22 11:16:36 +0000202 return i;
203}
204
205static ssize_t
206sched_feat_write(struct file *filp, const char __user *ubuf,
207 size_t cnt, loff_t *ppos)
208{
209 char buf[64];
210 char *cmp;
211 int i;
Jason Baron5cd08fb2014-07-02 15:52:44 +0000212 struct inode *inode;
Mel Gorman1a687c22012-11-22 11:16:36 +0000213
214 if (cnt > 63)
215 cnt = 63;
216
217 if (copy_from_user(&buf, ubuf, cnt))
218 return -EFAULT;
219
220 buf[cnt] = 0;
221 cmp = strstrip(buf);
222
Jason Baron5cd08fb2014-07-02 15:52:44 +0000223 /* Ensure the static_key remains in a consistent state */
224 inode = file_inode(filp);
225 mutex_lock(&inode->i_mutex);
Mel Gorman1a687c22012-11-22 11:16:36 +0000226 i = sched_feat_set(cmp);
Jason Baron5cd08fb2014-07-02 15:52:44 +0000227 mutex_unlock(&inode->i_mutex);
Peter Zijlstraf8b6d1c2011-07-06 14:20:14 +0200228 if (i == __SCHED_FEAT_NR)
Peter Zijlstraf00b45c2008-04-19 19:45:00 +0200229 return -EINVAL;
230
Jan Blunck42994722009-11-20 17:40:37 +0100231 *ppos += cnt;
Peter Zijlstraf00b45c2008-04-19 19:45:00 +0200232
233 return cnt;
234}
235
Li Zefan34f3a812008-10-30 15:23:32 +0800236static int sched_feat_open(struct inode *inode, struct file *filp)
237{
238 return single_open(filp, sched_feat_show, NULL);
239}
240
Alexey Dobriyan828c0952009-10-01 15:43:56 -0700241static const struct file_operations sched_feat_fops = {
Li Zefan34f3a812008-10-30 15:23:32 +0800242 .open = sched_feat_open,
243 .write = sched_feat_write,
244 .read = seq_read,
245 .llseek = seq_lseek,
246 .release = single_release,
Peter Zijlstraf00b45c2008-04-19 19:45:00 +0200247};
248
249static __init int sched_init_debug(void)
250{
Peter Zijlstraf00b45c2008-04-19 19:45:00 +0200251 debugfs_create_file("sched_features", 0644, NULL, NULL,
252 &sched_feat_fops);
253
254 return 0;
255}
256late_initcall(sched_init_debug);
Peter Zijlstraf8b6d1c2011-07-06 14:20:14 +0200257#endif /* CONFIG_SCHED_DEBUG */
Ingo Molnarbf5c91b2007-10-15 17:00:04 +0200258
259/*
Peter Zijlstrab82d9fd2007-11-09 22:39:39 +0100260 * Number of tasks to iterate in a single balance run.
261 * Limited because this is done with IRQs disabled.
262 */
263const_debug unsigned int sysctl_sched_nr_migrate = 32;
264
265/*
Peter Zijlstrae9e92502009-09-01 10:34:37 +0200266 * period over which we average the RT time consumption, measured
267 * in ms.
268 *
269 * default: 1s
270 */
271const_debug unsigned int sysctl_sched_time_avg = MSEC_PER_SEC;
272
273/*
Peter Zijlstra9f0c1e52008-02-13 15:45:39 +0100274 * period over which we measure -rt task cpu usage in us.
Peter Zijlstrafa85ae22008-01-25 21:08:29 +0100275 * default: 1s
276 */
Peter Zijlstra9f0c1e52008-02-13 15:45:39 +0100277unsigned int sysctl_sched_rt_period = 1000000;
Peter Zijlstrafa85ae22008-01-25 21:08:29 +0100278
Peter Zijlstra029632f2011-10-25 10:00:11 +0200279__read_mostly int scheduler_running;
Ingo Molnar6892b752008-02-13 14:02:36 +0100280
Peter Zijlstrafa85ae22008-01-25 21:08:29 +0100281/*
Peter Zijlstra9f0c1e52008-02-13 15:45:39 +0100282 * part of the period that we allow rt tasks to run in us.
283 * default: 0.95s
Peter Zijlstrafa85ae22008-01-25 21:08:29 +0100284 */
Peter Zijlstra9f0c1e52008-02-13 15:45:39 +0100285int sysctl_sched_rt_runtime = 950000;
286
Rik van Riel3fa08182015-03-09 12:12:07 -0400287/* cpus with isolated domains */
288cpumask_var_t cpu_isolated_map;
289
Dario Faggioli332ac172013-11-07 14:43:45 +0100290/*
Robert P. J. Daycc2a73b2006-12-10 02:20:00 -0800291 * this_rq_lock - lock this runqueue and disable interrupts.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700292 */
Alexey Dobriyana9957442007-10-15 17:00:13 +0200293static struct rq *this_rq_lock(void)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700294 __acquires(rq->lock)
295{
Ingo Molnar70b97a72006-07-03 00:25:42 -0700296 struct rq *rq;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700297
298 local_irq_disable();
299 rq = this_rq();
Thomas Gleixner05fa7852009-11-17 14:28:38 +0100300 raw_spin_lock(&rq->lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700301
302 return rq;
303}
304
Peter Zijlstra8f4d37e2008-01-25 21:08:29 +0100305#ifdef CONFIG_SCHED_HRTICK
306/*
307 * Use HR-timers to deliver accurate preemption points.
Peter Zijlstra8f4d37e2008-01-25 21:08:29 +0100308 */
Peter Zijlstra8f4d37e2008-01-25 21:08:29 +0100309
Peter Zijlstra8f4d37e2008-01-25 21:08:29 +0100310static void hrtick_clear(struct rq *rq)
311{
312 if (hrtimer_active(&rq->hrtick_timer))
313 hrtimer_cancel(&rq->hrtick_timer);
314}
315
316/*
Peter Zijlstra8f4d37e2008-01-25 21:08:29 +0100317 * High-resolution timer tick.
318 * Runs from hardirq context with interrupts disabled.
319 */
320static enum hrtimer_restart hrtick(struct hrtimer *timer)
321{
322 struct rq *rq = container_of(timer, struct rq, hrtick_timer);
323
324 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
325
Thomas Gleixner05fa7852009-11-17 14:28:38 +0100326 raw_spin_lock(&rq->lock);
Peter Zijlstra3e51f332008-05-03 18:29:28 +0200327 update_rq_clock(rq);
Peter Zijlstra8f4d37e2008-01-25 21:08:29 +0100328 rq->curr->sched_class->task_tick(rq, rq->curr, 1);
Thomas Gleixner05fa7852009-11-17 14:28:38 +0100329 raw_spin_unlock(&rq->lock);
Peter Zijlstra8f4d37e2008-01-25 21:08:29 +0100330
331 return HRTIMER_NORESTART;
332}
333
Rabin Vincent95e904c2008-05-11 05:55:33 +0530334#ifdef CONFIG_SMP
Peter Zijlstra971ee282013-06-28 11:18:53 +0200335
Thomas Gleixner4961b6e2015-04-14 21:09:05 +0000336static void __hrtick_restart(struct rq *rq)
Peter Zijlstra971ee282013-06-28 11:18:53 +0200337{
338 struct hrtimer *timer = &rq->hrtick_timer;
Peter Zijlstra971ee282013-06-28 11:18:53 +0200339
Thomas Gleixner4961b6e2015-04-14 21:09:05 +0000340 hrtimer_start_expires(timer, HRTIMER_MODE_ABS_PINNED);
Peter Zijlstra971ee282013-06-28 11:18:53 +0200341}
342
Peter Zijlstra31656512008-07-18 18:01:23 +0200343/*
344 * called from hardirq (IPI) context
345 */
346static void __hrtick_start(void *arg)
Peter Zijlstrab328ca12008-04-29 10:02:46 +0200347{
Peter Zijlstra31656512008-07-18 18:01:23 +0200348 struct rq *rq = arg;
Peter Zijlstrab328ca12008-04-29 10:02:46 +0200349
Thomas Gleixner05fa7852009-11-17 14:28:38 +0100350 raw_spin_lock(&rq->lock);
Peter Zijlstra971ee282013-06-28 11:18:53 +0200351 __hrtick_restart(rq);
Peter Zijlstra31656512008-07-18 18:01:23 +0200352 rq->hrtick_csd_pending = 0;
Thomas Gleixner05fa7852009-11-17 14:28:38 +0100353 raw_spin_unlock(&rq->lock);
Peter Zijlstrab328ca12008-04-29 10:02:46 +0200354}
355
Peter Zijlstra31656512008-07-18 18:01:23 +0200356/*
357 * Called to set the hrtick timer state.
358 *
359 * called with rq->lock held and irqs disabled
360 */
Peter Zijlstra029632f2011-10-25 10:00:11 +0200361void hrtick_start(struct rq *rq, u64 delay)
Peter Zijlstrab328ca12008-04-29 10:02:46 +0200362{
Peter Zijlstra31656512008-07-18 18:01:23 +0200363 struct hrtimer *timer = &rq->hrtick_timer;
xiaofeng.yan177ef2a2014-08-26 03:15:41 +0000364 ktime_t time;
365 s64 delta;
366
367 /*
368 * Don't schedule slices shorter than 10000ns, that just
369 * doesn't make sense and can cause timer DoS.
370 */
371 delta = max_t(s64, delay, 10000LL);
372 time = ktime_add_ns(timer->base->get_time(), delta);
Peter Zijlstrab328ca12008-04-29 10:02:46 +0200373
Arjan van de Vencc584b22008-09-01 15:02:30 -0700374 hrtimer_set_expires(timer, time);
Peter Zijlstra31656512008-07-18 18:01:23 +0200375
376 if (rq == this_rq()) {
Peter Zijlstra971ee282013-06-28 11:18:53 +0200377 __hrtick_restart(rq);
Peter Zijlstra31656512008-07-18 18:01:23 +0200378 } else if (!rq->hrtick_csd_pending) {
Frederic Weisbeckerc46fff22014-02-24 16:40:02 +0100379 smp_call_function_single_async(cpu_of(rq), &rq->hrtick_csd);
Peter Zijlstra31656512008-07-18 18:01:23 +0200380 rq->hrtick_csd_pending = 1;
381 }
Peter Zijlstrab328ca12008-04-29 10:02:46 +0200382}
383
384static int
385hotplug_hrtick(struct notifier_block *nfb, unsigned long action, void *hcpu)
386{
387 int cpu = (int)(long)hcpu;
388
389 switch (action) {
390 case CPU_UP_CANCELED:
391 case CPU_UP_CANCELED_FROZEN:
392 case CPU_DOWN_PREPARE:
393 case CPU_DOWN_PREPARE_FROZEN:
394 case CPU_DEAD:
395 case CPU_DEAD_FROZEN:
Peter Zijlstra31656512008-07-18 18:01:23 +0200396 hrtick_clear(cpu_rq(cpu));
Peter Zijlstrab328ca12008-04-29 10:02:46 +0200397 return NOTIFY_OK;
398 }
399
400 return NOTIFY_DONE;
401}
402
Rakib Mullickfa748202008-09-22 14:55:45 -0700403static __init void init_hrtick(void)
Peter Zijlstrab328ca12008-04-29 10:02:46 +0200404{
405 hotcpu_notifier(hotplug_hrtick, 0);
406}
Peter Zijlstra31656512008-07-18 18:01:23 +0200407#else
408/*
409 * Called to set the hrtick timer state.
410 *
411 * called with rq->lock held and irqs disabled
412 */
Peter Zijlstra029632f2011-10-25 10:00:11 +0200413void hrtick_start(struct rq *rq, u64 delay)
Peter Zijlstra31656512008-07-18 18:01:23 +0200414{
Wanpeng Li86893332014-11-26 08:44:06 +0800415 /*
416 * Don't schedule slices shorter than 10000ns, that just
417 * doesn't make sense. Rely on vruntime for fairness.
418 */
419 delay = max_t(u64, delay, 10000LL);
Thomas Gleixner4961b6e2015-04-14 21:09:05 +0000420 hrtimer_start(&rq->hrtick_timer, ns_to_ktime(delay),
421 HRTIMER_MODE_REL_PINNED);
Peter Zijlstra31656512008-07-18 18:01:23 +0200422}
423
Andrew Morton006c75f2008-09-22 14:55:46 -0700424static inline void init_hrtick(void)
Peter Zijlstra31656512008-07-18 18:01:23 +0200425{
426}
Rabin Vincent95e904c2008-05-11 05:55:33 +0530427#endif /* CONFIG_SMP */
Peter Zijlstrab328ca12008-04-29 10:02:46 +0200428
429static void init_rq_hrtick(struct rq *rq)
Peter Zijlstra8f4d37e2008-01-25 21:08:29 +0100430{
Peter Zijlstra31656512008-07-18 18:01:23 +0200431#ifdef CONFIG_SMP
432 rq->hrtick_csd_pending = 0;
433
434 rq->hrtick_csd.flags = 0;
435 rq->hrtick_csd.func = __hrtick_start;
436 rq->hrtick_csd.info = rq;
437#endif
438
Peter Zijlstra8f4d37e2008-01-25 21:08:29 +0100439 hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
440 rq->hrtick_timer.function = hrtick;
Peter Zijlstra8f4d37e2008-01-25 21:08:29 +0100441}
Andrew Morton006c75f2008-09-22 14:55:46 -0700442#else /* CONFIG_SCHED_HRTICK */
Peter Zijlstra8f4d37e2008-01-25 21:08:29 +0100443static inline void hrtick_clear(struct rq *rq)
444{
445}
446
Peter Zijlstra8f4d37e2008-01-25 21:08:29 +0100447static inline void init_rq_hrtick(struct rq *rq)
448{
449}
450
Peter Zijlstrab328ca12008-04-29 10:02:46 +0200451static inline void init_hrtick(void)
452{
453}
Andrew Morton006c75f2008-09-22 14:55:46 -0700454#endif /* CONFIG_SCHED_HRTICK */
Peter Zijlstra8f4d37e2008-01-25 21:08:29 +0100455
Ingo Molnar1b9f19c2007-07-09 18:51:59 +0200456/*
Peter Zijlstrafd99f912014-04-09 15:35:08 +0200457 * cmpxchg based fetch_or, macro so it works for different integer types
458 */
459#define fetch_or(ptr, val) \
460({ typeof(*(ptr)) __old, __val = *(ptr); \
461 for (;;) { \
462 __old = cmpxchg((ptr), __val, __val | (val)); \
463 if (__old == __val) \
464 break; \
465 __val = __old; \
466 } \
467 __old; \
468})
469
Peter Zijlstrae3baac42014-06-04 10:31:18 -0700470#if defined(CONFIG_SMP) && defined(TIF_POLLING_NRFLAG)
Peter Zijlstrafd99f912014-04-09 15:35:08 +0200471/*
472 * Atomically set TIF_NEED_RESCHED and test for TIF_POLLING_NRFLAG,
473 * this avoids any races wrt polling state changes and thereby avoids
474 * spurious IPIs.
475 */
476static bool set_nr_and_not_polling(struct task_struct *p)
477{
478 struct thread_info *ti = task_thread_info(p);
479 return !(fetch_or(&ti->flags, _TIF_NEED_RESCHED) & _TIF_POLLING_NRFLAG);
480}
Peter Zijlstrae3baac42014-06-04 10:31:18 -0700481
482/*
483 * Atomically set TIF_NEED_RESCHED if TIF_POLLING_NRFLAG is set.
484 *
485 * If this returns true, then the idle task promises to call
486 * sched_ttwu_pending() and reschedule soon.
487 */
488static bool set_nr_if_polling(struct task_struct *p)
489{
490 struct thread_info *ti = task_thread_info(p);
Jason Low316c1608d2015-04-28 13:00:20 -0700491 typeof(ti->flags) old, val = READ_ONCE(ti->flags);
Peter Zijlstrae3baac42014-06-04 10:31:18 -0700492
493 for (;;) {
494 if (!(val & _TIF_POLLING_NRFLAG))
495 return false;
496 if (val & _TIF_NEED_RESCHED)
497 return true;
498 old = cmpxchg(&ti->flags, val, val | _TIF_NEED_RESCHED);
499 if (old == val)
500 break;
501 val = old;
502 }
503 return true;
504}
505
Peter Zijlstrafd99f912014-04-09 15:35:08 +0200506#else
507static bool set_nr_and_not_polling(struct task_struct *p)
508{
509 set_tsk_need_resched(p);
510 return true;
511}
Peter Zijlstrae3baac42014-06-04 10:31:18 -0700512
513#ifdef CONFIG_SMP
514static bool set_nr_if_polling(struct task_struct *p)
515{
516 return false;
517}
518#endif
Peter Zijlstrafd99f912014-04-09 15:35:08 +0200519#endif
520
Peter Zijlstra76751042015-05-01 08:27:50 -0700521void wake_q_add(struct wake_q_head *head, struct task_struct *task)
522{
523 struct wake_q_node *node = &task->wake_q;
524
525 /*
526 * Atomically grab the task, if ->wake_q is !nil already it means
527 * its already queued (either by us or someone else) and will get the
528 * wakeup due to that.
529 *
530 * This cmpxchg() implies a full barrier, which pairs with the write
531 * barrier implied by the wakeup in wake_up_list().
532 */
533 if (cmpxchg(&node->next, NULL, WAKE_Q_TAIL))
534 return;
535
536 get_task_struct(task);
537
538 /*
539 * The head is context local, there can be no concurrency.
540 */
541 *head->lastp = node;
542 head->lastp = &node->next;
543}
544
545void wake_up_q(struct wake_q_head *head)
546{
547 struct wake_q_node *node = head->first;
548
549 while (node != WAKE_Q_TAIL) {
550 struct task_struct *task;
551
552 task = container_of(node, struct task_struct, wake_q);
553 BUG_ON(!task);
554 /* task can safely be re-inserted now */
555 node = node->next;
556 task->wake_q.next = NULL;
557
558 /*
559 * wake_up_process() implies a wmb() to pair with the queueing
560 * in wake_q_add() so as not to miss wakeups.
561 */
562 wake_up_process(task);
563 put_task_struct(task);
564 }
565}
566
Peter Zijlstrafd99f912014-04-09 15:35:08 +0200567/*
Kirill Tkhai88751252014-06-29 00:03:57 +0400568 * resched_curr - mark rq's current task 'to be rescheduled now'.
Ingo Molnarc24d20d2007-07-09 18:51:59 +0200569 *
570 * On UP this means the setting of the need_resched flag, on SMP it
571 * might also involve a cross-CPU call to trigger the scheduler on
572 * the target CPU.
573 */
Kirill Tkhai88751252014-06-29 00:03:57 +0400574void resched_curr(struct rq *rq)
Ingo Molnarc24d20d2007-07-09 18:51:59 +0200575{
Kirill Tkhai88751252014-06-29 00:03:57 +0400576 struct task_struct *curr = rq->curr;
Ingo Molnarc24d20d2007-07-09 18:51:59 +0200577 int cpu;
578
Kirill Tkhai88751252014-06-29 00:03:57 +0400579 lockdep_assert_held(&rq->lock);
Ingo Molnarc24d20d2007-07-09 18:51:59 +0200580
Kirill Tkhai88751252014-06-29 00:03:57 +0400581 if (test_tsk_need_resched(curr))
Ingo Molnarc24d20d2007-07-09 18:51:59 +0200582 return;
583
Kirill Tkhai88751252014-06-29 00:03:57 +0400584 cpu = cpu_of(rq);
Peter Zijlstrafd99f912014-04-09 15:35:08 +0200585
Peter Zijlstraf27dde82013-08-14 14:55:31 +0200586 if (cpu == smp_processor_id()) {
Kirill Tkhai88751252014-06-29 00:03:57 +0400587 set_tsk_need_resched(curr);
Peter Zijlstraf27dde82013-08-14 14:55:31 +0200588 set_preempt_need_resched();
Ingo Molnarc24d20d2007-07-09 18:51:59 +0200589 return;
Peter Zijlstraf27dde82013-08-14 14:55:31 +0200590 }
Ingo Molnarc24d20d2007-07-09 18:51:59 +0200591
Kirill Tkhai88751252014-06-29 00:03:57 +0400592 if (set_nr_and_not_polling(curr))
Ingo Molnarc24d20d2007-07-09 18:51:59 +0200593 smp_send_reschedule(cpu);
Andy Lutomirskidfc68f22014-06-04 10:31:15 -0700594 else
595 trace_sched_wake_idle_without_ipi(cpu);
Ingo Molnarc24d20d2007-07-09 18:51:59 +0200596}
597
Peter Zijlstra029632f2011-10-25 10:00:11 +0200598void resched_cpu(int cpu)
Ingo Molnarc24d20d2007-07-09 18:51:59 +0200599{
600 struct rq *rq = cpu_rq(cpu);
601 unsigned long flags;
602
Thomas Gleixner05fa7852009-11-17 14:28:38 +0100603 if (!raw_spin_trylock_irqsave(&rq->lock, flags))
Ingo Molnarc24d20d2007-07-09 18:51:59 +0200604 return;
Kirill Tkhai88751252014-06-29 00:03:57 +0400605 resched_curr(rq);
Thomas Gleixner05fa7852009-11-17 14:28:38 +0100606 raw_spin_unlock_irqrestore(&rq->lock, flags);
Ingo Molnarc24d20d2007-07-09 18:51:59 +0200607}
Thomas Gleixner06d83082008-03-22 09:20:24 +0100608
Peter Zijlstrab021fe32013-09-17 09:30:55 +0200609#ifdef CONFIG_SMP
Frederic Weisbecker3451d022011-08-10 23:21:01 +0200610#ifdef CONFIG_NO_HZ_COMMON
Thomas Gleixner06d83082008-03-22 09:20:24 +0100611/*
Venkatesh Pallipadi83cd4fe2010-05-21 17:09:41 -0700612 * In the semi idle case, use the nearest busy cpu for migrating timers
613 * from an idle cpu. This is good for power-savings.
614 *
615 * We don't do similar optimization for completely idle system, as
616 * selecting an idle cpu will add more delays to the timers than intended
617 * (as that cpu's timer base may not be uptodate wrt jiffies etc).
618 */
Thomas Gleixnerbc7a34b2015-05-26 22:50:33 +0000619int get_nohz_timer_target(void)
Venkatesh Pallipadi83cd4fe2010-05-21 17:09:41 -0700620{
Thomas Gleixnerbc7a34b2015-05-26 22:50:33 +0000621 int i, cpu = smp_processor_id();
Venkatesh Pallipadi83cd4fe2010-05-21 17:09:41 -0700622 struct sched_domain *sd;
623
Thomas Gleixnerbc7a34b2015-05-26 22:50:33 +0000624 if (!idle_cpu(cpu))
Viresh Kumar6201b4d2014-03-18 16:26:07 +0530625 return cpu;
626
Peter Zijlstra057f3fa2011-04-18 11:24:34 +0200627 rcu_read_lock();
Venkatesh Pallipadi83cd4fe2010-05-21 17:09:41 -0700628 for_each_domain(cpu, sd) {
Peter Zijlstra057f3fa2011-04-18 11:24:34 +0200629 for_each_cpu(i, sched_domain_span(sd)) {
630 if (!idle_cpu(i)) {
631 cpu = i;
632 goto unlock;
633 }
634 }
Venkatesh Pallipadi83cd4fe2010-05-21 17:09:41 -0700635 }
Peter Zijlstra057f3fa2011-04-18 11:24:34 +0200636unlock:
637 rcu_read_unlock();
Venkatesh Pallipadi83cd4fe2010-05-21 17:09:41 -0700638 return cpu;
639}
640/*
Thomas Gleixner06d83082008-03-22 09:20:24 +0100641 * When add_timer_on() enqueues a timer into the timer wheel of an
642 * idle CPU then this timer might expire before the next timer event
643 * which is scheduled to wake up that CPU. In case of a completely
644 * idle system the next event might even be infinite time into the
645 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
646 * leaves the inner idle loop so the newly added timer is taken into
647 * account when the CPU goes back to idle and evaluates the timer
648 * wheel for the next timer event.
649 */
Frederic Weisbecker1c200912011-08-10 23:21:01 +0200650static void wake_up_idle_cpu(int cpu)
Thomas Gleixner06d83082008-03-22 09:20:24 +0100651{
652 struct rq *rq = cpu_rq(cpu);
653
654 if (cpu == smp_processor_id())
655 return;
656
Andy Lutomirski67b9ca72014-06-04 10:31:17 -0700657 if (set_nr_and_not_polling(rq->idle))
Thomas Gleixner06d83082008-03-22 09:20:24 +0100658 smp_send_reschedule(cpu);
Andy Lutomirskidfc68f22014-06-04 10:31:15 -0700659 else
660 trace_sched_wake_idle_without_ipi(cpu);
Thomas Gleixner06d83082008-03-22 09:20:24 +0100661}
Mike Galbraith39c0cbe2010-03-11 17:17:13 +0100662
Frederic Weisbeckerc5bfece2013-04-12 16:45:34 +0200663static bool wake_up_full_nohz_cpu(int cpu)
Frederic Weisbecker1c200912011-08-10 23:21:01 +0200664{
Frederic Weisbecker53c5fa12014-06-04 16:20:21 +0200665 /*
666 * We just need the target to call irq_exit() and re-evaluate
667 * the next tick. The nohz full kick at least implies that.
668 * If needed we can still optimize that later with an
669 * empty IRQ.
670 */
Frederic Weisbeckerc5bfece2013-04-12 16:45:34 +0200671 if (tick_nohz_full_cpu(cpu)) {
Frederic Weisbecker1c200912011-08-10 23:21:01 +0200672 if (cpu != smp_processor_id() ||
673 tick_nohz_tick_stopped())
Frederic Weisbecker53c5fa12014-06-04 16:20:21 +0200674 tick_nohz_full_kick_cpu(cpu);
Frederic Weisbecker1c200912011-08-10 23:21:01 +0200675 return true;
676 }
677
678 return false;
679}
680
681void wake_up_nohz_cpu(int cpu)
682{
Frederic Weisbeckerc5bfece2013-04-12 16:45:34 +0200683 if (!wake_up_full_nohz_cpu(cpu))
Frederic Weisbecker1c200912011-08-10 23:21:01 +0200684 wake_up_idle_cpu(cpu);
685}
686
Suresh Siddhaca380622011-10-03 15:09:00 -0700687static inline bool got_nohz_idle_kick(void)
688{
Suresh Siddha1c792db2011-12-01 17:07:32 -0800689 int cpu = smp_processor_id();
Vincent Guittot873b4c62013-06-05 10:13:11 +0200690
691 if (!test_bit(NOHZ_BALANCE_KICK, nohz_flags(cpu)))
692 return false;
693
694 if (idle_cpu(cpu) && !need_resched())
695 return true;
696
697 /*
698 * We can't run Idle Load Balance on this CPU for this time so we
699 * cancel it and clear NOHZ_BALANCE_KICK
700 */
701 clear_bit(NOHZ_BALANCE_KICK, nohz_flags(cpu));
702 return false;
Suresh Siddhaca380622011-10-03 15:09:00 -0700703}
704
Frederic Weisbecker3451d022011-08-10 23:21:01 +0200705#else /* CONFIG_NO_HZ_COMMON */
Suresh Siddhaca380622011-10-03 15:09:00 -0700706
707static inline bool got_nohz_idle_kick(void)
708{
709 return false;
710}
711
Frederic Weisbecker3451d022011-08-10 23:21:01 +0200712#endif /* CONFIG_NO_HZ_COMMON */
Thomas Gleixner06d83082008-03-22 09:20:24 +0100713
Frederic Weisbeckerce831b32013-04-20 15:15:35 +0200714#ifdef CONFIG_NO_HZ_FULL
715bool sched_can_stop_tick(void)
716{
Frederic Weisbecker3882ec62014-03-18 22:54:04 +0100717 /*
Rik van Riel1e78cdb2015-02-16 15:23:49 -0500718 * FIFO realtime policy runs the highest priority task. Other runnable
719 * tasks are of a lower priority. The scheduler tick does nothing.
720 */
721 if (current->policy == SCHED_FIFO)
722 return true;
723
724 /*
725 * Round-robin realtime tasks time slice with other tasks at the same
726 * realtime priority. Is this task the only one at this priority?
727 */
728 if (current->policy == SCHED_RR) {
729 struct sched_rt_entity *rt_se = &current->rt;
730
731 return rt_se->run_list.prev == rt_se->run_list.next;
732 }
733
734 /*
Frederic Weisbecker3882ec62014-03-18 22:54:04 +0100735 * More than one running task need preemption.
736 * nr_running update is assumed to be visible
737 * after IPI is sent from wakers.
738 */
Viresh Kumar541b8262014-06-24 14:04:12 +0530739 if (this_rq()->nr_running > 1)
740 return false;
Frederic Weisbeckerce831b32013-04-20 15:15:35 +0200741
Viresh Kumar541b8262014-06-24 14:04:12 +0530742 return true;
Frederic Weisbeckerce831b32013-04-20 15:15:35 +0200743}
744#endif /* CONFIG_NO_HZ_FULL */
Ingo Molnar45bf76d2007-07-09 18:51:59 +0200745
Peter Zijlstra029632f2011-10-25 10:00:11 +0200746void sched_avg_update(struct rq *rq)
Peter Zijlstrae9e92502009-09-01 10:34:37 +0200747{
748 s64 period = sched_avg_period();
749
Frederic Weisbecker78becc22013-04-12 01:51:02 +0200750 while ((s64)(rq_clock(rq) - rq->age_stamp) > period) {
Will Deacon0d98bb22010-05-24 12:11:43 -0700751 /*
752 * Inline assembly required to prevent the compiler
753 * optimising this loop into a divmod call.
754 * See __iter_div_u64_rem() for another example of this.
755 */
756 asm("" : "+rm" (rq->age_stamp));
Peter Zijlstrae9e92502009-09-01 10:34:37 +0200757 rq->age_stamp += period;
758 rq->rt_avg /= 2;
759 }
760}
761
Dhaval Giani6d6bc0a2008-05-30 14:23:45 +0200762#endif /* CONFIG_SMP */
Ingo Molnar45bf76d2007-07-09 18:51:59 +0200763
Paul Turnera790de92011-07-21 09:43:29 -0700764#if defined(CONFIG_RT_GROUP_SCHED) || (defined(CONFIG_FAIR_GROUP_SCHED) && \
765 (defined(CONFIG_SMP) || defined(CONFIG_CFS_BANDWIDTH)))
Peter Zijlstraeb755802008-08-19 12:33:05 +0200766/*
Paul Turner82774342011-07-21 09:43:35 -0700767 * Iterate task_group tree rooted at *from, calling @down when first entering a
768 * node and @up when leaving it for the final time.
769 *
770 * Caller must hold rcu_lock or sufficient equivalent.
Peter Zijlstraeb755802008-08-19 12:33:05 +0200771 */
Peter Zijlstra029632f2011-10-25 10:00:11 +0200772int walk_tg_tree_from(struct task_group *from,
Paul Turner82774342011-07-21 09:43:35 -0700773 tg_visitor down, tg_visitor up, void *data)
Peter Zijlstraeb755802008-08-19 12:33:05 +0200774{
775 struct task_group *parent, *child;
776 int ret;
777
Paul Turner82774342011-07-21 09:43:35 -0700778 parent = from;
779
Peter Zijlstraeb755802008-08-19 12:33:05 +0200780down:
781 ret = (*down)(parent, data);
782 if (ret)
Paul Turner82774342011-07-21 09:43:35 -0700783 goto out;
Peter Zijlstraeb755802008-08-19 12:33:05 +0200784 list_for_each_entry_rcu(child, &parent->children, siblings) {
785 parent = child;
786 goto down;
787
788up:
789 continue;
790 }
791 ret = (*up)(parent, data);
Paul Turner82774342011-07-21 09:43:35 -0700792 if (ret || parent == from)
793 goto out;
Peter Zijlstraeb755802008-08-19 12:33:05 +0200794
795 child = parent;
796 parent = parent->parent;
797 if (parent)
798 goto up;
Paul Turner82774342011-07-21 09:43:35 -0700799out:
Peter Zijlstraeb755802008-08-19 12:33:05 +0200800 return ret;
801}
802
Peter Zijlstra029632f2011-10-25 10:00:11 +0200803int tg_nop(struct task_group *tg, void *data)
Peter Zijlstraeb755802008-08-19 12:33:05 +0200804{
805 return 0;
806}
807#endif
808
Ingo Molnar45bf76d2007-07-09 18:51:59 +0200809static void set_load_weight(struct task_struct *p)
810{
Nikhil Raof05998d2011-05-18 10:09:38 -0700811 int prio = p->static_prio - MAX_RT_PRIO;
812 struct load_weight *load = &p->se.load;
813
Ingo Molnardd41f592007-07-09 18:51:59 +0200814 /*
815 * SCHED_IDLE tasks get minimal weight:
816 */
817 if (p->policy == SCHED_IDLE) {
Nikhil Raoc8b28112011-05-18 14:37:48 -0700818 load->weight = scale_load(WEIGHT_IDLEPRIO);
Nikhil Raof05998d2011-05-18 10:09:38 -0700819 load->inv_weight = WMULT_IDLEPRIO;
Ingo Molnardd41f592007-07-09 18:51:59 +0200820 return;
821 }
822
Nikhil Raoc8b28112011-05-18 14:37:48 -0700823 load->weight = scale_load(prio_to_weight[prio]);
Nikhil Raof05998d2011-05-18 10:09:38 -0700824 load->inv_weight = prio_to_wmult[prio];
Ingo Molnar45bf76d2007-07-09 18:51:59 +0200825}
826
Peter Zijlstra371fd7e2010-03-24 16:38:48 +0100827static void enqueue_task(struct rq *rq, struct task_struct *p, int flags)
Gregory Haskins2087a1a2008-06-27 14:30:00 -0600828{
Mike Galbraitha64692a2010-03-11 17:16:20 +0100829 update_rq_clock(rq);
Michael S. Tsirkin43148952013-09-22 17:20:54 +0300830 sched_info_queued(rq, p);
Peter Zijlstra371fd7e2010-03-24 16:38:48 +0100831 p->sched_class->enqueue_task(rq, p, flags);
Ingo Molnardd41f592007-07-09 18:51:59 +0200832}
833
Peter Zijlstra371fd7e2010-03-24 16:38:48 +0100834static void dequeue_task(struct rq *rq, struct task_struct *p, int flags)
Ingo Molnardd41f592007-07-09 18:51:59 +0200835{
Mike Galbraitha64692a2010-03-11 17:16:20 +0100836 update_rq_clock(rq);
Michael S. Tsirkin43148952013-09-22 17:20:54 +0300837 sched_info_dequeued(rq, p);
Peter Zijlstra371fd7e2010-03-24 16:38:48 +0100838 p->sched_class->dequeue_task(rq, p, flags);
Ingo Molnar71f8bd42007-07-09 18:51:59 +0200839}
840
Peter Zijlstra029632f2011-10-25 10:00:11 +0200841void activate_task(struct rq *rq, struct task_struct *p, int flags)
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +0100842{
843 if (task_contributes_to_load(p))
844 rq->nr_uninterruptible--;
845
Peter Zijlstra371fd7e2010-03-24 16:38:48 +0100846 enqueue_task(rq, p, flags);
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +0100847}
848
Peter Zijlstra029632f2011-10-25 10:00:11 +0200849void deactivate_task(struct rq *rq, struct task_struct *p, int flags)
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +0100850{
851 if (task_contributes_to_load(p))
852 rq->nr_uninterruptible++;
853
Peter Zijlstra371fd7e2010-03-24 16:38:48 +0100854 dequeue_task(rq, p, flags);
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +0100855}
856
Peter Zijlstrafe44d622010-12-09 14:15:34 +0100857static void update_rq_clock_task(struct rq *rq, s64 delta)
Venkatesh Pallipadiaa483802010-10-04 17:03:22 -0700858{
Glauber Costa095c0aa2011-07-11 15:28:18 -0400859/*
860 * In theory, the compile should just see 0 here, and optimize out the call
861 * to sched_rt_avg_update. But I don't trust it...
862 */
863#if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
864 s64 steal = 0, irq_delta = 0;
865#endif
866#ifdef CONFIG_IRQ_TIME_ACCOUNTING
Peter Zijlstra8e92c202010-12-09 14:15:34 +0100867 irq_delta = irq_time_read(cpu_of(rq)) - rq->prev_irq_time;
Peter Zijlstrafe44d622010-12-09 14:15:34 +0100868
869 /*
870 * Since irq_time is only updated on {soft,}irq_exit, we might run into
871 * this case when a previous update_rq_clock() happened inside a
872 * {soft,}irq region.
873 *
874 * When this happens, we stop ->clock_task and only update the
875 * prev_irq_time stamp to account for the part that fit, so that a next
876 * update will consume the rest. This ensures ->clock_task is
877 * monotonic.
878 *
879 * It does however cause some slight miss-attribution of {soft,}irq
880 * time, a more accurate solution would be to update the irq_time using
881 * the current rq->clock timestamp, except that would require using
882 * atomic ops.
883 */
884 if (irq_delta > delta)
885 irq_delta = delta;
886
887 rq->prev_irq_time += irq_delta;
888 delta -= irq_delta;
Glauber Costa095c0aa2011-07-11 15:28:18 -0400889#endif
890#ifdef CONFIG_PARAVIRT_TIME_ACCOUNTING
Ingo Molnarc5905af2012-02-24 08:31:31 +0100891 if (static_key_false((&paravirt_steal_rq_enabled))) {
Glauber Costa095c0aa2011-07-11 15:28:18 -0400892 steal = paravirt_steal_clock(cpu_of(rq));
893 steal -= rq->prev_steal_time_rq;
894
895 if (unlikely(steal > delta))
896 steal = delta;
897
Glauber Costa095c0aa2011-07-11 15:28:18 -0400898 rq->prev_steal_time_rq += steal;
Glauber Costa095c0aa2011-07-11 15:28:18 -0400899 delta -= steal;
900 }
901#endif
902
Peter Zijlstrafe44d622010-12-09 14:15:34 +0100903 rq->clock_task += delta;
904
Glauber Costa095c0aa2011-07-11 15:28:18 -0400905#if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
Nicolas Pitre5d4dfdd2014-05-27 13:50:41 -0400906 if ((irq_delta + steal) && sched_feat(NONTASK_CAPACITY))
Glauber Costa095c0aa2011-07-11 15:28:18 -0400907 sched_rt_avg_update(rq, irq_delta + steal);
908#endif
Venkatesh Pallipadiaa483802010-10-04 17:03:22 -0700909}
910
Peter Zijlstra34f971f2010-09-22 13:53:15 +0200911void sched_set_stop_task(int cpu, struct task_struct *stop)
912{
913 struct sched_param param = { .sched_priority = MAX_RT_PRIO - 1 };
914 struct task_struct *old_stop = cpu_rq(cpu)->stop;
915
916 if (stop) {
917 /*
918 * Make it appear like a SCHED_FIFO task, its something
919 * userspace knows about and won't get confused about.
920 *
921 * Also, it will make PI more or less work without too
922 * much confusion -- but then, stop work should not
923 * rely on PI working anyway.
924 */
925 sched_setscheduler_nocheck(stop, SCHED_FIFO, &param);
926
927 stop->sched_class = &stop_sched_class;
928 }
929
930 cpu_rq(cpu)->stop = stop;
931
932 if (old_stop) {
933 /*
934 * Reset it back to a normal scheduling class so that
935 * it can die in pieces.
936 */
937 old_stop->sched_class = &rt_sched_class;
938 }
939}
940
Peter Zijlstra1e3c88b2009-12-17 17:00:43 +0100941/*
Ingo Molnardd41f592007-07-09 18:51:59 +0200942 * __normal_prio - return the priority that is based on the static prio
Ingo Molnar71f8bd42007-07-09 18:51:59 +0200943 */
Ingo Molnar14531182007-07-09 18:51:59 +0200944static inline int __normal_prio(struct task_struct *p)
945{
Ingo Molnardd41f592007-07-09 18:51:59 +0200946 return p->static_prio;
Ingo Molnar14531182007-07-09 18:51:59 +0200947}
948
949/*
Ingo Molnarb29739f2006-06-27 02:54:51 -0700950 * Calculate the expected normal priority: i.e. priority
951 * without taking RT-inheritance into account. Might be
952 * boosted by interactivity modifiers. Changes upon fork,
953 * setprio syscalls, and whenever the interactivity
954 * estimator recalculates.
955 */
Ingo Molnar36c8b582006-07-03 00:25:41 -0700956static inline int normal_prio(struct task_struct *p)
Ingo Molnarb29739f2006-06-27 02:54:51 -0700957{
958 int prio;
959
Dario Faggioliaab03e02013-11-28 11:14:43 +0100960 if (task_has_dl_policy(p))
961 prio = MAX_DL_PRIO-1;
962 else if (task_has_rt_policy(p))
Ingo Molnarb29739f2006-06-27 02:54:51 -0700963 prio = MAX_RT_PRIO-1 - p->rt_priority;
964 else
965 prio = __normal_prio(p);
966 return prio;
967}
968
969/*
970 * Calculate the current priority, i.e. the priority
971 * taken into account by the scheduler. This value might
972 * be boosted by RT tasks, or might be boosted by
973 * interactivity modifiers. Will be RT if the task got
974 * RT-boosted. If not then it returns p->normal_prio.
975 */
Ingo Molnar36c8b582006-07-03 00:25:41 -0700976static int effective_prio(struct task_struct *p)
Ingo Molnarb29739f2006-06-27 02:54:51 -0700977{
978 p->normal_prio = normal_prio(p);
979 /*
980 * If we are RT tasks or we were boosted to RT priority,
981 * keep the priority unchanged. Otherwise, update priority
982 * to the normal priority:
983 */
984 if (!rt_prio(p->prio))
985 return p->normal_prio;
986 return p->prio;
987}
988
Linus Torvalds1da177e2005-04-16 15:20:36 -0700989/**
990 * task_curr - is this task currently executing on a CPU?
991 * @p: the task in question.
Yacine Belkadie69f6182013-07-12 20:45:47 +0200992 *
993 * Return: 1 if the task is currently executing. 0 otherwise.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700994 */
Ingo Molnar36c8b582006-07-03 00:25:41 -0700995inline int task_curr(const struct task_struct *p)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700996{
997 return cpu_curr(task_cpu(p)) == p;
998}
999
Kirill Tkhai67dfa1b2014-10-27 17:40:52 +03001000/*
Peter Zijlstra4c9a4bc2015-06-11 14:46:39 +02001001 * switched_from, switched_to and prio_changed must _NOT_ drop rq->lock,
1002 * use the balance_callback list if you want balancing.
1003 *
1004 * this means any call to check_class_changed() must be followed by a call to
1005 * balance_callback().
Kirill Tkhai67dfa1b2014-10-27 17:40:52 +03001006 */
Steven Rostedtcb469842008-01-25 21:08:22 +01001007static inline void check_class_changed(struct rq *rq, struct task_struct *p,
1008 const struct sched_class *prev_class,
Peter Zijlstrada7a7352011-01-17 17:03:27 +01001009 int oldprio)
Steven Rostedtcb469842008-01-25 21:08:22 +01001010{
1011 if (prev_class != p->sched_class) {
1012 if (prev_class->switched_from)
Peter Zijlstrada7a7352011-01-17 17:03:27 +01001013 prev_class->switched_from(rq, p);
Peter Zijlstra4c9a4bc2015-06-11 14:46:39 +02001014
Peter Zijlstrada7a7352011-01-17 17:03:27 +01001015 p->sched_class->switched_to(rq, p);
Dario Faggioli2d3d8912013-11-07 14:43:44 +01001016 } else if (oldprio != p->prio || dl_task(p))
Peter Zijlstrada7a7352011-01-17 17:03:27 +01001017 p->sched_class->prio_changed(rq, p, oldprio);
Steven Rostedtcb469842008-01-25 21:08:22 +01001018}
1019
Peter Zijlstra029632f2011-10-25 10:00:11 +02001020void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags)
Peter Zijlstra1e5a7402010-10-31 12:37:04 +01001021{
1022 const struct sched_class *class;
1023
1024 if (p->sched_class == rq->curr->sched_class) {
1025 rq->curr->sched_class->check_preempt_curr(rq, p, flags);
1026 } else {
1027 for_each_class(class) {
1028 if (class == rq->curr->sched_class)
1029 break;
1030 if (class == p->sched_class) {
Kirill Tkhai88751252014-06-29 00:03:57 +04001031 resched_curr(rq);
Peter Zijlstra1e5a7402010-10-31 12:37:04 +01001032 break;
1033 }
1034 }
1035 }
1036
1037 /*
1038 * A queue event has occurred, and we're going to schedule. In
1039 * this case, we can save a useless back to back clock update.
1040 */
Kirill Tkhaida0c1e62014-08-20 13:47:32 +04001041 if (task_on_rq_queued(rq->curr) && test_tsk_need_resched(rq->curr))
Peter Zijlstra9edfbfe2015-01-05 11:18:11 +01001042 rq_clock_skip_update(rq, true);
Peter Zijlstra1e5a7402010-10-31 12:37:04 +01001043}
1044
Linus Torvalds1da177e2005-04-16 15:20:36 -07001045#ifdef CONFIG_SMP
Peter Zijlstra5cc389b2015-06-11 14:46:50 +02001046/*
1047 * This is how migration works:
1048 *
1049 * 1) we invoke migration_cpu_stop() on the target CPU using
1050 * stop_one_cpu().
1051 * 2) stopper starts to run (implicitly forcing the migrated thread
1052 * off the CPU)
1053 * 3) it checks whether the migrated task is still in the wrong runqueue.
1054 * 4) if it's in the wrong runqueue then the migration thread removes
1055 * it and puts it into the right queue.
1056 * 5) stopper completes and stop_one_cpu() returns and the migration
1057 * is done.
1058 */
1059
1060/*
1061 * move_queued_task - move a queued task to new rq.
1062 *
1063 * Returns (locked) new rq. Old rq's lock is released.
1064 */
Peter Zijlstra5e16bbc2015-06-11 14:46:51 +02001065static struct rq *move_queued_task(struct rq *rq, struct task_struct *p, int new_cpu)
Peter Zijlstra5cc389b2015-06-11 14:46:50 +02001066{
Peter Zijlstra5cc389b2015-06-11 14:46:50 +02001067 lockdep_assert_held(&rq->lock);
1068
1069 dequeue_task(rq, p, 0);
1070 p->on_rq = TASK_ON_RQ_MIGRATING;
1071 set_task_cpu(p, new_cpu);
1072 raw_spin_unlock(&rq->lock);
1073
1074 rq = cpu_rq(new_cpu);
1075
1076 raw_spin_lock(&rq->lock);
1077 BUG_ON(task_cpu(p) != new_cpu);
1078 p->on_rq = TASK_ON_RQ_QUEUED;
1079 enqueue_task(rq, p, 0);
1080 check_preempt_curr(rq, p, 0);
1081
1082 return rq;
1083}
1084
1085struct migration_arg {
1086 struct task_struct *task;
1087 int dest_cpu;
1088};
1089
1090/*
1091 * Move (not current) task off this cpu, onto dest cpu. We're doing
1092 * this because either it can't run here any more (set_cpus_allowed()
1093 * away from this CPU, or CPU going down), or because we're
1094 * attempting to rebalance this task on exec (sched_exec).
1095 *
1096 * So we race with normal scheduler movements, but that's OK, as long
1097 * as the task is no longer on this CPU.
Peter Zijlstra5cc389b2015-06-11 14:46:50 +02001098 */
Peter Zijlstra5e16bbc2015-06-11 14:46:51 +02001099static struct rq *__migrate_task(struct rq *rq, struct task_struct *p, int dest_cpu)
Peter Zijlstra5cc389b2015-06-11 14:46:50 +02001100{
Peter Zijlstra5cc389b2015-06-11 14:46:50 +02001101 if (unlikely(!cpu_active(dest_cpu)))
Peter Zijlstra5e16bbc2015-06-11 14:46:51 +02001102 return rq;
Peter Zijlstra5cc389b2015-06-11 14:46:50 +02001103
1104 /* Affinity changed (again). */
1105 if (!cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p)))
Peter Zijlstra5e16bbc2015-06-11 14:46:51 +02001106 return rq;
Peter Zijlstra5cc389b2015-06-11 14:46:50 +02001107
Peter Zijlstra5e16bbc2015-06-11 14:46:51 +02001108 rq = move_queued_task(rq, p, dest_cpu);
1109
1110 return rq;
Peter Zijlstra5cc389b2015-06-11 14:46:50 +02001111}
1112
1113/*
1114 * migration_cpu_stop - this will be executed by a highprio stopper thread
1115 * and performs thread migration by bumping thread off CPU then
1116 * 'pushing' onto another runqueue.
1117 */
1118static int migration_cpu_stop(void *data)
1119{
1120 struct migration_arg *arg = data;
Peter Zijlstra5e16bbc2015-06-11 14:46:51 +02001121 struct task_struct *p = arg->task;
1122 struct rq *rq = this_rq();
Peter Zijlstra5cc389b2015-06-11 14:46:50 +02001123
1124 /*
1125 * The original target cpu might have gone down and we might
1126 * be on another cpu but it doesn't matter.
1127 */
1128 local_irq_disable();
1129 /*
1130 * We need to explicitly wake pending tasks before running
1131 * __migrate_task() such that we will not miss enforcing cpus_allowed
1132 * during wakeups, see set_cpus_allowed_ptr()'s TASK_WAKING test.
1133 */
1134 sched_ttwu_pending();
Peter Zijlstra5e16bbc2015-06-11 14:46:51 +02001135
1136 raw_spin_lock(&p->pi_lock);
1137 raw_spin_lock(&rq->lock);
1138 /*
1139 * If task_rq(p) != rq, it cannot be migrated here, because we're
1140 * holding rq->lock, if p->on_rq == 0 it cannot get enqueued because
1141 * we're holding p->pi_lock.
1142 */
1143 if (task_rq(p) == rq && task_on_rq_queued(p))
1144 rq = __migrate_task(rq, p, arg->dest_cpu);
1145 raw_spin_unlock(&rq->lock);
1146 raw_spin_unlock(&p->pi_lock);
1147
Peter Zijlstra5cc389b2015-06-11 14:46:50 +02001148 local_irq_enable();
1149 return 0;
1150}
1151
Peter Zijlstrac5b28032015-05-15 17:43:35 +02001152/*
1153 * sched_class::set_cpus_allowed must do the below, but is not required to
1154 * actually call this function.
1155 */
1156void set_cpus_allowed_common(struct task_struct *p, const struct cpumask *new_mask)
Peter Zijlstra5cc389b2015-06-11 14:46:50 +02001157{
Peter Zijlstra5cc389b2015-06-11 14:46:50 +02001158 cpumask_copy(&p->cpus_allowed, new_mask);
1159 p->nr_cpus_allowed = cpumask_weight(new_mask);
1160}
1161
Peter Zijlstra5cc389b2015-06-11 14:46:50 +02001162void do_set_cpus_allowed(struct task_struct *p, const struct cpumask *new_mask)
1163{
Peter Zijlstra6c370672015-05-15 17:43:36 +02001164 struct rq *rq = task_rq(p);
1165 bool queued, running;
1166
Peter Zijlstra25834c72015-05-15 17:43:34 +02001167 lockdep_assert_held(&p->pi_lock);
Peter Zijlstra6c370672015-05-15 17:43:36 +02001168
1169 queued = task_on_rq_queued(p);
1170 running = task_current(rq, p);
1171
1172 if (queued) {
1173 /*
1174 * Because __kthread_bind() calls this on blocked tasks without
1175 * holding rq->lock.
1176 */
1177 lockdep_assert_held(&rq->lock);
1178 dequeue_task(rq, p, 0);
1179 }
1180 if (running)
1181 put_prev_task(rq, p);
1182
Peter Zijlstrac5b28032015-05-15 17:43:35 +02001183 p->sched_class->set_cpus_allowed(p, new_mask);
Peter Zijlstra6c370672015-05-15 17:43:36 +02001184
1185 if (running)
1186 p->sched_class->set_curr_task(rq);
1187 if (queued)
1188 enqueue_task(rq, p, 0);
Peter Zijlstra5cc389b2015-06-11 14:46:50 +02001189}
1190
1191/*
1192 * Change a given task's CPU affinity. Migrate the thread to a
1193 * proper CPU and schedule it away if the CPU it's executing on
1194 * is removed from the allowed bitmask.
1195 *
1196 * NOTE: the caller must have a valid reference to the task, the
1197 * task must not exit() & deallocate itself prematurely. The
1198 * call is not atomic; no spinlocks may be held.
1199 */
Peter Zijlstra25834c72015-05-15 17:43:34 +02001200static int __set_cpus_allowed_ptr(struct task_struct *p,
1201 const struct cpumask *new_mask, bool check)
Peter Zijlstra5cc389b2015-06-11 14:46:50 +02001202{
1203 unsigned long flags;
1204 struct rq *rq;
1205 unsigned int dest_cpu;
1206 int ret = 0;
1207
1208 rq = task_rq_lock(p, &flags);
1209
Peter Zijlstra25834c72015-05-15 17:43:34 +02001210 /*
1211 * Must re-check here, to close a race against __kthread_bind(),
1212 * sched_setaffinity() is not guaranteed to observe the flag.
1213 */
1214 if (check && (p->flags & PF_NO_SETAFFINITY)) {
1215 ret = -EINVAL;
1216 goto out;
1217 }
1218
Peter Zijlstra5cc389b2015-06-11 14:46:50 +02001219 if (cpumask_equal(&p->cpus_allowed, new_mask))
1220 goto out;
1221
1222 if (!cpumask_intersects(new_mask, cpu_active_mask)) {
1223 ret = -EINVAL;
1224 goto out;
1225 }
1226
1227 do_set_cpus_allowed(p, new_mask);
1228
1229 /* Can the task run on the task's current CPU? If so, we're done */
1230 if (cpumask_test_cpu(task_cpu(p), new_mask))
1231 goto out;
1232
1233 dest_cpu = cpumask_any_and(cpu_active_mask, new_mask);
1234 if (task_running(rq, p) || p->state == TASK_WAKING) {
1235 struct migration_arg arg = { p, dest_cpu };
1236 /* Need help from migration thread: drop lock and wait. */
1237 task_rq_unlock(rq, p, &flags);
1238 stop_one_cpu(cpu_of(rq), migration_cpu_stop, &arg);
1239 tlb_migrate_finish(p->mm);
1240 return 0;
Peter Zijlstracbce1a62015-06-11 14:46:54 +02001241 } else if (task_on_rq_queued(p)) {
1242 /*
1243 * OK, since we're going to drop the lock immediately
1244 * afterwards anyway.
1245 */
1246 lockdep_unpin_lock(&rq->lock);
Peter Zijlstra5e16bbc2015-06-11 14:46:51 +02001247 rq = move_queued_task(rq, p, dest_cpu);
Peter Zijlstracbce1a62015-06-11 14:46:54 +02001248 lockdep_pin_lock(&rq->lock);
1249 }
Peter Zijlstra5cc389b2015-06-11 14:46:50 +02001250out:
1251 task_rq_unlock(rq, p, &flags);
1252
1253 return ret;
1254}
Peter Zijlstra25834c72015-05-15 17:43:34 +02001255
1256int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask)
1257{
1258 return __set_cpus_allowed_ptr(p, new_mask, false);
1259}
Peter Zijlstra5cc389b2015-06-11 14:46:50 +02001260EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
1261
Ingo Molnardd41f592007-07-09 18:51:59 +02001262void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
Ingo Molnarc65cc872007-07-09 18:51:58 +02001263{
Peter Zijlstrae2912002009-12-16 18:04:36 +01001264#ifdef CONFIG_SCHED_DEBUG
1265 /*
1266 * We should never call set_task_cpu() on a blocked task,
1267 * ttwu() will sort out the placement.
1268 */
Peter Zijlstra077614e2009-12-17 13:16:31 +01001269 WARN_ON_ONCE(p->state != TASK_RUNNING && p->state != TASK_WAKING &&
Oleg Nesterove2336f62014-10-08 20:33:48 +02001270 !p->on_rq);
Peter Zijlstra0122ec52011-04-05 17:23:51 +02001271
1272#ifdef CONFIG_LOCKDEP
Peter Zijlstra6c6c54e2011-06-03 17:37:07 +02001273 /*
1274 * The caller should hold either p->pi_lock or rq->lock, when changing
1275 * a task's CPU. ->pi_lock for waking tasks, rq->lock for runnable tasks.
1276 *
1277 * sched_move_task() holds both and thus holding either pins the cgroup,
Peter Zijlstra8323f262012-06-22 13:36:05 +02001278 * see task_group().
Peter Zijlstra6c6c54e2011-06-03 17:37:07 +02001279 *
1280 * Furthermore, all task_rq users should acquire both locks, see
1281 * task_rq_lock().
1282 */
Peter Zijlstra0122ec52011-04-05 17:23:51 +02001283 WARN_ON_ONCE(debug_locks && !(lockdep_is_held(&p->pi_lock) ||
1284 lockdep_is_held(&task_rq(p)->lock)));
1285#endif
Peter Zijlstrae2912002009-12-16 18:04:36 +01001286#endif
1287
Mathieu Desnoyersde1d7282009-05-05 16:49:59 +08001288 trace_sched_migrate_task(p, new_cpu);
Peter Zijlstracbc34ed2008-12-10 08:08:22 +01001289
Peter Zijlstra0c697742009-12-22 15:43:19 +01001290 if (task_cpu(p) != new_cpu) {
Paul Turner0a74bef2012-10-04 13:18:30 +02001291 if (p->sched_class->migrate_task_rq)
1292 p->sched_class->migrate_task_rq(p, new_cpu);
Peter Zijlstra0c697742009-12-22 15:43:19 +01001293 p->se.nr_migrations++;
Peter Zijlstraff303e62015-04-17 20:05:30 +02001294 perf_event_task_migrate(p);
Peter Zijlstra0c697742009-12-22 15:43:19 +01001295 }
Ingo Molnardd41f592007-07-09 18:51:59 +02001296
1297 __set_task_cpu(p, new_cpu);
Ingo Molnarc65cc872007-07-09 18:51:58 +02001298}
1299
Peter Zijlstraac66f542013-10-07 11:29:16 +01001300static void __migrate_swap_task(struct task_struct *p, int cpu)
1301{
Kirill Tkhaida0c1e62014-08-20 13:47:32 +04001302 if (task_on_rq_queued(p)) {
Peter Zijlstraac66f542013-10-07 11:29:16 +01001303 struct rq *src_rq, *dst_rq;
1304
1305 src_rq = task_rq(p);
1306 dst_rq = cpu_rq(cpu);
1307
1308 deactivate_task(src_rq, p, 0);
1309 set_task_cpu(p, cpu);
1310 activate_task(dst_rq, p, 0);
1311 check_preempt_curr(dst_rq, p, 0);
1312 } else {
1313 /*
1314 * Task isn't running anymore; make it appear like we migrated
1315 * it before it went to sleep. This means on wakeup we make the
1316 * previous cpu our targer instead of where it really is.
1317 */
1318 p->wake_cpu = cpu;
1319 }
1320}
1321
1322struct migration_swap_arg {
1323 struct task_struct *src_task, *dst_task;
1324 int src_cpu, dst_cpu;
1325};
1326
1327static int migrate_swap_stop(void *data)
1328{
1329 struct migration_swap_arg *arg = data;
1330 struct rq *src_rq, *dst_rq;
1331 int ret = -EAGAIN;
1332
1333 src_rq = cpu_rq(arg->src_cpu);
1334 dst_rq = cpu_rq(arg->dst_cpu);
1335
Peter Zijlstra74602312013-10-10 20:17:22 +02001336 double_raw_lock(&arg->src_task->pi_lock,
1337 &arg->dst_task->pi_lock);
Peter Zijlstraac66f542013-10-07 11:29:16 +01001338 double_rq_lock(src_rq, dst_rq);
1339 if (task_cpu(arg->dst_task) != arg->dst_cpu)
1340 goto unlock;
1341
1342 if (task_cpu(arg->src_task) != arg->src_cpu)
1343 goto unlock;
1344
1345 if (!cpumask_test_cpu(arg->dst_cpu, tsk_cpus_allowed(arg->src_task)))
1346 goto unlock;
1347
1348 if (!cpumask_test_cpu(arg->src_cpu, tsk_cpus_allowed(arg->dst_task)))
1349 goto unlock;
1350
1351 __migrate_swap_task(arg->src_task, arg->dst_cpu);
1352 __migrate_swap_task(arg->dst_task, arg->src_cpu);
1353
1354 ret = 0;
1355
1356unlock:
1357 double_rq_unlock(src_rq, dst_rq);
Peter Zijlstra74602312013-10-10 20:17:22 +02001358 raw_spin_unlock(&arg->dst_task->pi_lock);
1359 raw_spin_unlock(&arg->src_task->pi_lock);
Peter Zijlstraac66f542013-10-07 11:29:16 +01001360
1361 return ret;
1362}
1363
1364/*
1365 * Cross migrate two tasks
1366 */
1367int migrate_swap(struct task_struct *cur, struct task_struct *p)
1368{
1369 struct migration_swap_arg arg;
1370 int ret = -EINVAL;
1371
Peter Zijlstraac66f542013-10-07 11:29:16 +01001372 arg = (struct migration_swap_arg){
1373 .src_task = cur,
1374 .src_cpu = task_cpu(cur),
1375 .dst_task = p,
1376 .dst_cpu = task_cpu(p),
1377 };
1378
1379 if (arg.src_cpu == arg.dst_cpu)
1380 goto out;
1381
Peter Zijlstra6acce3e2013-10-11 14:38:20 +02001382 /*
1383 * These three tests are all lockless; this is OK since all of them
1384 * will be re-checked with proper locks held further down the line.
1385 */
Peter Zijlstraac66f542013-10-07 11:29:16 +01001386 if (!cpu_active(arg.src_cpu) || !cpu_active(arg.dst_cpu))
1387 goto out;
1388
1389 if (!cpumask_test_cpu(arg.dst_cpu, tsk_cpus_allowed(arg.src_task)))
1390 goto out;
1391
1392 if (!cpumask_test_cpu(arg.src_cpu, tsk_cpus_allowed(arg.dst_task)))
1393 goto out;
1394
Mel Gorman286549d2014-01-21 15:51:03 -08001395 trace_sched_swap_numa(cur, arg.src_cpu, p, arg.dst_cpu);
Peter Zijlstraac66f542013-10-07 11:29:16 +01001396 ret = stop_two_cpus(arg.dst_cpu, arg.src_cpu, migrate_swap_stop, &arg);
1397
1398out:
Peter Zijlstraac66f542013-10-07 11:29:16 +01001399 return ret;
1400}
1401
Linus Torvalds1da177e2005-04-16 15:20:36 -07001402/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07001403 * wait_task_inactive - wait for a thread to unschedule.
1404 *
Roland McGrath85ba2d82008-07-25 19:45:58 -07001405 * If @match_state is nonzero, it's the @p->state value just checked and
1406 * not expected to change. If it changes, i.e. @p might have woken up,
1407 * then return zero. When we succeed in waiting for @p to be off its CPU,
1408 * we return a positive number (its total switch count). If a second call
1409 * a short while later returns the same number, the caller can be sure that
1410 * @p has remained unscheduled the whole time.
1411 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07001412 * The caller must ensure that the task *will* unschedule sometime soon,
1413 * else this function might spin for a *long* time. This function can't
1414 * be called with interrupts off, or it may introduce deadlock with
1415 * smp_call_function() if an IPI is sent by the same process we are
1416 * waiting to become inactive.
1417 */
Roland McGrath85ba2d82008-07-25 19:45:58 -07001418unsigned long wait_task_inactive(struct task_struct *p, long match_state)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001419{
1420 unsigned long flags;
Kirill Tkhaida0c1e62014-08-20 13:47:32 +04001421 int running, queued;
Roland McGrath85ba2d82008-07-25 19:45:58 -07001422 unsigned long ncsw;
Ingo Molnar70b97a72006-07-03 00:25:42 -07001423 struct rq *rq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001424
Andi Kleen3a5c3592007-10-15 17:00:14 +02001425 for (;;) {
1426 /*
1427 * We do the initial early heuristics without holding
1428 * any task-queue locks at all. We'll only try to get
1429 * the runqueue lock when things look like they will
1430 * work out!
1431 */
1432 rq = task_rq(p);
Linus Torvaldsfa490cf2007-06-18 09:34:40 -07001433
Andi Kleen3a5c3592007-10-15 17:00:14 +02001434 /*
1435 * If the task is actively running on another CPU
1436 * still, just relax and busy-wait without holding
1437 * any locks.
1438 *
1439 * NOTE! Since we don't hold any locks, it's not
1440 * even sure that "rq" stays as the right runqueue!
1441 * But we don't care, since "task_running()" will
1442 * return false if the runqueue has changed and p
1443 * is actually now running somewhere else!
1444 */
Roland McGrath85ba2d82008-07-25 19:45:58 -07001445 while (task_running(rq, p)) {
1446 if (match_state && unlikely(p->state != match_state))
1447 return 0;
Andi Kleen3a5c3592007-10-15 17:00:14 +02001448 cpu_relax();
Roland McGrath85ba2d82008-07-25 19:45:58 -07001449 }
Linus Torvaldsfa490cf2007-06-18 09:34:40 -07001450
Andi Kleen3a5c3592007-10-15 17:00:14 +02001451 /*
1452 * Ok, time to look more closely! We need the rq
1453 * lock now, to be *sure*. If we're wrong, we'll
1454 * just go back and repeat.
1455 */
1456 rq = task_rq_lock(p, &flags);
Peter Zijlstra27a9da62010-05-04 20:36:56 +02001457 trace_sched_wait_task(p);
Andi Kleen3a5c3592007-10-15 17:00:14 +02001458 running = task_running(rq, p);
Kirill Tkhaida0c1e62014-08-20 13:47:32 +04001459 queued = task_on_rq_queued(p);
Roland McGrath85ba2d82008-07-25 19:45:58 -07001460 ncsw = 0;
Oleg Nesterovf31e11d2008-08-20 16:54:44 -07001461 if (!match_state || p->state == match_state)
Oleg Nesterov93dcf552008-08-20 16:54:44 -07001462 ncsw = p->nvcsw | LONG_MIN; /* sets MSB */
Peter Zijlstra0122ec52011-04-05 17:23:51 +02001463 task_rq_unlock(rq, p, &flags);
Linus Torvaldsfa490cf2007-06-18 09:34:40 -07001464
Andi Kleen3a5c3592007-10-15 17:00:14 +02001465 /*
Roland McGrath85ba2d82008-07-25 19:45:58 -07001466 * If it changed from the expected state, bail out now.
1467 */
1468 if (unlikely(!ncsw))
1469 break;
1470
1471 /*
Andi Kleen3a5c3592007-10-15 17:00:14 +02001472 * Was it really running after all now that we
1473 * checked with the proper locks actually held?
1474 *
1475 * Oops. Go back and try again..
1476 */
1477 if (unlikely(running)) {
1478 cpu_relax();
1479 continue;
1480 }
1481
1482 /*
1483 * It's not enough that it's not actively running,
1484 * it must be off the runqueue _entirely_, and not
1485 * preempted!
1486 *
Luis Henriques80dd99b2009-03-16 19:58:09 +00001487 * So if it was still runnable (but just not actively
Andi Kleen3a5c3592007-10-15 17:00:14 +02001488 * running right now), it's preempted, and we should
1489 * yield - it could be a while.
1490 */
Kirill Tkhaida0c1e62014-08-20 13:47:32 +04001491 if (unlikely(queued)) {
Thomas Gleixner8eb90c32011-02-23 23:52:21 +00001492 ktime_t to = ktime_set(0, NSEC_PER_SEC/HZ);
1493
1494 set_current_state(TASK_UNINTERRUPTIBLE);
1495 schedule_hrtimeout(&to, HRTIMER_MODE_REL);
Andi Kleen3a5c3592007-10-15 17:00:14 +02001496 continue;
1497 }
1498
1499 /*
1500 * Ahh, all good. It wasn't running, and it wasn't
1501 * runnable, which means that it will never become
1502 * running in the future either. We're all done!
1503 */
1504 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001505 }
Roland McGrath85ba2d82008-07-25 19:45:58 -07001506
1507 return ncsw;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001508}
1509
1510/***
1511 * kick_process - kick a running thread to enter/exit the kernel
1512 * @p: the to-be-kicked thread
1513 *
1514 * Cause a process which is running on another CPU to enter
1515 * kernel-mode, without any delay. (to get signals handled.)
1516 *
Lucas De Marchi25985ed2011-03-30 22:57:33 -03001517 * NOTE: this function doesn't have to take the runqueue lock,
Linus Torvalds1da177e2005-04-16 15:20:36 -07001518 * because all it wants to ensure is that the remote task enters
1519 * the kernel. If the IPI races and the task has been migrated
1520 * to another CPU then no harm is done and the purpose has been
1521 * achieved as well.
1522 */
Ingo Molnar36c8b582006-07-03 00:25:41 -07001523void kick_process(struct task_struct *p)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001524{
1525 int cpu;
1526
1527 preempt_disable();
1528 cpu = task_cpu(p);
1529 if ((cpu != smp_processor_id()) && task_curr(p))
1530 smp_send_reschedule(cpu);
1531 preempt_enable();
1532}
Rusty Russellb43e3522009-06-12 22:27:00 -06001533EXPORT_SYMBOL_GPL(kick_process);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001534
Oleg Nesterov30da6882010-03-15 10:10:19 +01001535/*
Peter Zijlstra013fdb82011-04-05 17:23:45 +02001536 * ->cpus_allowed is protected by both rq->lock and p->pi_lock
Oleg Nesterov30da6882010-03-15 10:10:19 +01001537 */
Peter Zijlstra5da9a0f2009-12-16 18:04:38 +01001538static int select_fallback_rq(int cpu, struct task_struct *p)
1539{
Tang Chenaa00d892013-02-22 16:33:33 -08001540 int nid = cpu_to_node(cpu);
1541 const struct cpumask *nodemask = NULL;
Peter Zijlstra2baab4e2012-03-20 15:57:01 +01001542 enum { cpuset, possible, fail } state = cpuset;
1543 int dest_cpu;
Peter Zijlstra5da9a0f2009-12-16 18:04:38 +01001544
Tang Chenaa00d892013-02-22 16:33:33 -08001545 /*
1546 * If the node that the cpu is on has been offlined, cpu_to_node()
1547 * will return -1. There is no cpu on the node, and we should
1548 * select the cpu on the other node.
1549 */
1550 if (nid != -1) {
1551 nodemask = cpumask_of_node(nid);
1552
1553 /* Look for allowed, online CPU in same node. */
1554 for_each_cpu(dest_cpu, nodemask) {
1555 if (!cpu_online(dest_cpu))
1556 continue;
1557 if (!cpu_active(dest_cpu))
1558 continue;
1559 if (cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p)))
1560 return dest_cpu;
1561 }
Peter Zijlstra2baab4e2012-03-20 15:57:01 +01001562 }
Peter Zijlstra5da9a0f2009-12-16 18:04:38 +01001563
Peter Zijlstra2baab4e2012-03-20 15:57:01 +01001564 for (;;) {
1565 /* Any allowed, online CPU? */
Srivatsa S. Bhate3831ed2012-03-30 19:40:28 +05301566 for_each_cpu(dest_cpu, tsk_cpus_allowed(p)) {
Peter Zijlstra2baab4e2012-03-20 15:57:01 +01001567 if (!cpu_online(dest_cpu))
1568 continue;
1569 if (!cpu_active(dest_cpu))
1570 continue;
1571 goto out;
1572 }
Peter Zijlstra5da9a0f2009-12-16 18:04:38 +01001573
Peter Zijlstra2baab4e2012-03-20 15:57:01 +01001574 switch (state) {
1575 case cpuset:
1576 /* No more Mr. Nice Guy. */
1577 cpuset_cpus_allowed_fallback(p);
1578 state = possible;
1579 break;
1580
1581 case possible:
1582 do_set_cpus_allowed(p, cpu_possible_mask);
1583 state = fail;
1584 break;
1585
1586 case fail:
1587 BUG();
1588 break;
1589 }
1590 }
1591
1592out:
1593 if (state != cpuset) {
1594 /*
1595 * Don't tell them about moving exiting tasks or
1596 * kernel threads (both mm NULL), since they never
1597 * leave kernel.
1598 */
1599 if (p->mm && printk_ratelimit()) {
John Stultzaac74dc2014-06-04 16:11:40 -07001600 printk_deferred("process %d (%s) no longer affine to cpu%d\n",
Peter Zijlstra2baab4e2012-03-20 15:57:01 +01001601 task_pid_nr(p), p->comm, cpu);
1602 }
Peter Zijlstra5da9a0f2009-12-16 18:04:38 +01001603 }
1604
1605 return dest_cpu;
1606}
1607
Peter Zijlstrae2912002009-12-16 18:04:36 +01001608/*
Peter Zijlstra013fdb82011-04-05 17:23:45 +02001609 * The caller (fork, wakeup) owns p->pi_lock, ->cpus_allowed is stable.
Peter Zijlstrae2912002009-12-16 18:04:36 +01001610 */
Peter Zijlstra970b13b2009-11-25 13:31:39 +01001611static inline
Peter Zijlstraac66f542013-10-07 11:29:16 +01001612int select_task_rq(struct task_struct *p, int cpu, int sd_flags, int wake_flags)
Peter Zijlstra970b13b2009-11-25 13:31:39 +01001613{
Peter Zijlstracbce1a62015-06-11 14:46:54 +02001614 lockdep_assert_held(&p->pi_lock);
1615
Wanpeng Li6c1d9412014-11-05 09:14:37 +08001616 if (p->nr_cpus_allowed > 1)
1617 cpu = p->sched_class->select_task_rq(p, cpu, sd_flags, wake_flags);
Peter Zijlstrae2912002009-12-16 18:04:36 +01001618
1619 /*
1620 * In order not to call set_task_cpu() on a blocking task we need
1621 * to rely on ttwu() to place the task on a valid ->cpus_allowed
1622 * cpu.
1623 *
1624 * Since this is common to all placement strategies, this lives here.
1625 *
1626 * [ this allows ->select_task() to simply return task_cpu(p) and
1627 * not worry about this generic constraint ]
1628 */
Peter Zijlstrafa17b502011-06-16 12:23:22 +02001629 if (unlikely(!cpumask_test_cpu(cpu, tsk_cpus_allowed(p)) ||
Peter Zijlstra70f11202009-12-20 17:36:27 +01001630 !cpu_online(cpu)))
Peter Zijlstra5da9a0f2009-12-16 18:04:38 +01001631 cpu = select_fallback_rq(task_cpu(p), p);
Peter Zijlstrae2912002009-12-16 18:04:36 +01001632
1633 return cpu;
Peter Zijlstra970b13b2009-11-25 13:31:39 +01001634}
Mike Galbraith09a40af2010-04-15 07:29:59 +02001635
1636static void update_avg(u64 *avg, u64 sample)
1637{
1638 s64 diff = sample - *avg;
1639 *avg += diff >> 3;
1640}
Peter Zijlstra25834c72015-05-15 17:43:34 +02001641
1642#else
1643
1644static inline int __set_cpus_allowed_ptr(struct task_struct *p,
1645 const struct cpumask *new_mask, bool check)
1646{
1647 return set_cpus_allowed_ptr(p, new_mask);
1648}
1649
Peter Zijlstra5cc389b2015-06-11 14:46:50 +02001650#endif /* CONFIG_SMP */
Peter Zijlstra970b13b2009-11-25 13:31:39 +01001651
Peter Zijlstrad7c01d22011-04-05 17:23:43 +02001652static void
Peter Zijlstrab84cb5d2011-04-05 17:23:55 +02001653ttwu_stat(struct task_struct *p, int cpu, int wake_flags)
Tejun Heo9ed38112009-12-03 15:08:03 +09001654{
Peter Zijlstrad7c01d22011-04-05 17:23:43 +02001655#ifdef CONFIG_SCHEDSTATS
Peter Zijlstrab84cb5d2011-04-05 17:23:55 +02001656 struct rq *rq = this_rq();
Tejun Heo9ed38112009-12-03 15:08:03 +09001657
Peter Zijlstrad7c01d22011-04-05 17:23:43 +02001658#ifdef CONFIG_SMP
1659 int this_cpu = smp_processor_id();
Tejun Heo9ed38112009-12-03 15:08:03 +09001660
Peter Zijlstrad7c01d22011-04-05 17:23:43 +02001661 if (cpu == this_cpu) {
1662 schedstat_inc(rq, ttwu_local);
1663 schedstat_inc(p, se.statistics.nr_wakeups_local);
1664 } else {
1665 struct sched_domain *sd;
1666
1667 schedstat_inc(p, se.statistics.nr_wakeups_remote);
Peter Zijlstra057f3fa2011-04-18 11:24:34 +02001668 rcu_read_lock();
Peter Zijlstrad7c01d22011-04-05 17:23:43 +02001669 for_each_domain(this_cpu, sd) {
1670 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) {
1671 schedstat_inc(sd, ttwu_wake_remote);
1672 break;
1673 }
1674 }
Peter Zijlstra057f3fa2011-04-18 11:24:34 +02001675 rcu_read_unlock();
Peter Zijlstrad7c01d22011-04-05 17:23:43 +02001676 }
Peter Zijlstraf339b9d2011-05-31 10:49:20 +02001677
1678 if (wake_flags & WF_MIGRATED)
1679 schedstat_inc(p, se.statistics.nr_wakeups_migrate);
1680
Peter Zijlstrad7c01d22011-04-05 17:23:43 +02001681#endif /* CONFIG_SMP */
1682
1683 schedstat_inc(rq, ttwu_count);
1684 schedstat_inc(p, se.statistics.nr_wakeups);
1685
1686 if (wake_flags & WF_SYNC)
1687 schedstat_inc(p, se.statistics.nr_wakeups_sync);
1688
Peter Zijlstrad7c01d22011-04-05 17:23:43 +02001689#endif /* CONFIG_SCHEDSTATS */
Tejun Heo9ed38112009-12-03 15:08:03 +09001690}
1691
Peter Zijlstrad7c01d22011-04-05 17:23:43 +02001692static void ttwu_activate(struct rq *rq, struct task_struct *p, int en_flags)
Tejun Heo9ed38112009-12-03 15:08:03 +09001693{
Tejun Heo9ed38112009-12-03 15:08:03 +09001694 activate_task(rq, p, en_flags);
Kirill Tkhaida0c1e62014-08-20 13:47:32 +04001695 p->on_rq = TASK_ON_RQ_QUEUED;
Peter Zijlstrac2f71152011-04-13 13:28:56 +02001696
1697 /* if a worker is waking up, notify workqueue */
1698 if (p->flags & PF_WQ_WORKER)
1699 wq_worker_waking_up(p, cpu_of(rq));
Tejun Heo9ed38112009-12-03 15:08:03 +09001700}
1701
Peter Zijlstra23f41ee2011-04-05 17:23:56 +02001702/*
1703 * Mark the task runnable and perform wakeup-preemption.
1704 */
Peter Zijlstra89363382011-04-05 17:23:42 +02001705static void
Peter Zijlstra23f41ee2011-04-05 17:23:56 +02001706ttwu_do_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
Tejun Heo9ed38112009-12-03 15:08:03 +09001707{
Tejun Heo9ed38112009-12-03 15:08:03 +09001708 check_preempt_curr(rq, p, wake_flags);
Tejun Heo9ed38112009-12-03 15:08:03 +09001709 p->state = TASK_RUNNING;
Peter Zijlstrafbd705a2015-06-09 11:13:36 +02001710 trace_sched_wakeup(p);
1711
Tejun Heo9ed38112009-12-03 15:08:03 +09001712#ifdef CONFIG_SMP
Peter Zijlstra4c9a4bc2015-06-11 14:46:39 +02001713 if (p->sched_class->task_woken) {
1714 /*
Peter Zijlstracbce1a62015-06-11 14:46:54 +02001715 * Our task @p is fully woken up and running; so its safe to
1716 * drop the rq->lock, hereafter rq is only used for statistics.
Peter Zijlstra4c9a4bc2015-06-11 14:46:39 +02001717 */
Peter Zijlstracbce1a62015-06-11 14:46:54 +02001718 lockdep_unpin_lock(&rq->lock);
Tejun Heo9ed38112009-12-03 15:08:03 +09001719 p->sched_class->task_woken(rq, p);
Peter Zijlstracbce1a62015-06-11 14:46:54 +02001720 lockdep_pin_lock(&rq->lock);
Peter Zijlstra4c9a4bc2015-06-11 14:46:39 +02001721 }
Tejun Heo9ed38112009-12-03 15:08:03 +09001722
Steven Rostedte69c6342010-12-06 17:10:31 -05001723 if (rq->idle_stamp) {
Frederic Weisbecker78becc22013-04-12 01:51:02 +02001724 u64 delta = rq_clock(rq) - rq->idle_stamp;
Jason Low9bd721c2013-09-13 11:26:52 -07001725 u64 max = 2*rq->max_idle_balance_cost;
Tejun Heo9ed38112009-12-03 15:08:03 +09001726
Jason Lowabfafa52013-09-13 11:26:51 -07001727 update_avg(&rq->avg_idle, delta);
1728
1729 if (rq->avg_idle > max)
Tejun Heo9ed38112009-12-03 15:08:03 +09001730 rq->avg_idle = max;
Jason Lowabfafa52013-09-13 11:26:51 -07001731
Tejun Heo9ed38112009-12-03 15:08:03 +09001732 rq->idle_stamp = 0;
1733 }
1734#endif
1735}
1736
Peter Zijlstrac05fbaf2011-04-05 17:23:57 +02001737static void
1738ttwu_do_activate(struct rq *rq, struct task_struct *p, int wake_flags)
1739{
Peter Zijlstracbce1a62015-06-11 14:46:54 +02001740 lockdep_assert_held(&rq->lock);
1741
Peter Zijlstrac05fbaf2011-04-05 17:23:57 +02001742#ifdef CONFIG_SMP
1743 if (p->sched_contributes_to_load)
1744 rq->nr_uninterruptible--;
1745#endif
1746
1747 ttwu_activate(rq, p, ENQUEUE_WAKEUP | ENQUEUE_WAKING);
1748 ttwu_do_wakeup(rq, p, wake_flags);
1749}
1750
1751/*
1752 * Called in case the task @p isn't fully descheduled from its runqueue,
1753 * in this case we must do a remote wakeup. Its a 'light' wakeup though,
1754 * since all we need to do is flip p->state to TASK_RUNNING, since
1755 * the task is still ->on_rq.
1756 */
1757static int ttwu_remote(struct task_struct *p, int wake_flags)
1758{
1759 struct rq *rq;
1760 int ret = 0;
1761
1762 rq = __task_rq_lock(p);
Kirill Tkhaida0c1e62014-08-20 13:47:32 +04001763 if (task_on_rq_queued(p)) {
Frederic Weisbecker1ad4ec02013-04-12 01:51:00 +02001764 /* check_preempt_curr() may use rq clock */
1765 update_rq_clock(rq);
Peter Zijlstrac05fbaf2011-04-05 17:23:57 +02001766 ttwu_do_wakeup(rq, p, wake_flags);
1767 ret = 1;
1768 }
1769 __task_rq_unlock(rq);
1770
1771 return ret;
1772}
1773
Peter Zijlstra317f3942011-04-05 17:23:58 +02001774#ifdef CONFIG_SMP
Peter Zijlstrae3baac42014-06-04 10:31:18 -07001775void sched_ttwu_pending(void)
Peter Zijlstra317f3942011-04-05 17:23:58 +02001776{
1777 struct rq *rq = this_rq();
Peter Zijlstrafa14ff42011-09-12 13:06:17 +02001778 struct llist_node *llist = llist_del_all(&rq->wake_list);
1779 struct task_struct *p;
Peter Zijlstrae3baac42014-06-04 10:31:18 -07001780 unsigned long flags;
Peter Zijlstra317f3942011-04-05 17:23:58 +02001781
Peter Zijlstrae3baac42014-06-04 10:31:18 -07001782 if (!llist)
1783 return;
1784
1785 raw_spin_lock_irqsave(&rq->lock, flags);
Peter Zijlstracbce1a62015-06-11 14:46:54 +02001786 lockdep_pin_lock(&rq->lock);
Peter Zijlstra317f3942011-04-05 17:23:58 +02001787
Peter Zijlstrafa14ff42011-09-12 13:06:17 +02001788 while (llist) {
1789 p = llist_entry(llist, struct task_struct, wake_entry);
1790 llist = llist_next(llist);
Peter Zijlstra317f3942011-04-05 17:23:58 +02001791 ttwu_do_activate(rq, p, 0);
1792 }
1793
Peter Zijlstracbce1a62015-06-11 14:46:54 +02001794 lockdep_unpin_lock(&rq->lock);
Peter Zijlstrae3baac42014-06-04 10:31:18 -07001795 raw_spin_unlock_irqrestore(&rq->lock, flags);
Peter Zijlstra317f3942011-04-05 17:23:58 +02001796}
1797
1798void scheduler_ipi(void)
1799{
Peter Zijlstraf27dde82013-08-14 14:55:31 +02001800 /*
1801 * Fold TIF_NEED_RESCHED into the preempt_count; anybody setting
1802 * TIF_NEED_RESCHED remotely (for the first time) will also send
1803 * this IPI.
1804 */
Peter Zijlstra8cb75e02013-11-20 12:22:37 +01001805 preempt_fold_need_resched();
Peter Zijlstraf27dde82013-08-14 14:55:31 +02001806
Frederic Weisbeckerfd2ac4f2014-03-18 21:12:53 +01001807 if (llist_empty(&this_rq()->wake_list) && !got_nohz_idle_kick())
Peter Zijlstrac5d753a2011-07-19 15:07:25 -07001808 return;
1809
1810 /*
1811 * Not all reschedule IPI handlers call irq_enter/irq_exit, since
1812 * traditionally all their work was done from the interrupt return
1813 * path. Now that we actually do some work, we need to make sure
1814 * we do call them.
1815 *
1816 * Some archs already do call them, luckily irq_enter/exit nest
1817 * properly.
1818 *
1819 * Arguably we should visit all archs and update all handlers,
1820 * however a fair share of IPIs are still resched only so this would
1821 * somewhat pessimize the simple resched case.
1822 */
1823 irq_enter();
Peter Zijlstrafa14ff42011-09-12 13:06:17 +02001824 sched_ttwu_pending();
Suresh Siddhaca380622011-10-03 15:09:00 -07001825
1826 /*
1827 * Check if someone kicked us for doing the nohz idle load balance.
1828 */
Vincent Guittot873b4c62013-06-05 10:13:11 +02001829 if (unlikely(got_nohz_idle_kick())) {
Suresh Siddha6eb57e02011-10-03 15:09:01 -07001830 this_rq()->idle_balance = 1;
Suresh Siddhaca380622011-10-03 15:09:00 -07001831 raise_softirq_irqoff(SCHED_SOFTIRQ);
Suresh Siddha6eb57e02011-10-03 15:09:01 -07001832 }
Peter Zijlstrac5d753a2011-07-19 15:07:25 -07001833 irq_exit();
Peter Zijlstra317f3942011-04-05 17:23:58 +02001834}
1835
1836static void ttwu_queue_remote(struct task_struct *p, int cpu)
1837{
Peter Zijlstrae3baac42014-06-04 10:31:18 -07001838 struct rq *rq = cpu_rq(cpu);
1839
1840 if (llist_add(&p->wake_entry, &cpu_rq(cpu)->wake_list)) {
1841 if (!set_nr_if_polling(rq->idle))
1842 smp_send_reschedule(cpu);
1843 else
1844 trace_sched_wake_idle_without_ipi(cpu);
1845 }
Peter Zijlstra317f3942011-04-05 17:23:58 +02001846}
Peter Zijlstrad6aa8f82011-05-26 14:21:33 +02001847
Chuansheng Liuf6be8af2014-09-04 15:17:53 +08001848void wake_up_if_idle(int cpu)
1849{
1850 struct rq *rq = cpu_rq(cpu);
1851 unsigned long flags;
1852
Andy Lutomirskifd7de1e2014-11-29 08:13:51 -08001853 rcu_read_lock();
1854
1855 if (!is_idle_task(rcu_dereference(rq->curr)))
1856 goto out;
Chuansheng Liuf6be8af2014-09-04 15:17:53 +08001857
1858 if (set_nr_if_polling(rq->idle)) {
1859 trace_sched_wake_idle_without_ipi(cpu);
1860 } else {
1861 raw_spin_lock_irqsave(&rq->lock, flags);
1862 if (is_idle_task(rq->curr))
1863 smp_send_reschedule(cpu);
1864 /* Else cpu is not in idle, do nothing here */
1865 raw_spin_unlock_irqrestore(&rq->lock, flags);
1866 }
Andy Lutomirskifd7de1e2014-11-29 08:13:51 -08001867
1868out:
1869 rcu_read_unlock();
Chuansheng Liuf6be8af2014-09-04 15:17:53 +08001870}
1871
Peter Zijlstra39be3502012-01-26 12:44:34 +01001872bool cpus_share_cache(int this_cpu, int that_cpu)
Peter Zijlstra518cd622011-12-07 15:07:31 +01001873{
1874 return per_cpu(sd_llc_id, this_cpu) == per_cpu(sd_llc_id, that_cpu);
1875}
Peter Zijlstrad6aa8f82011-05-26 14:21:33 +02001876#endif /* CONFIG_SMP */
Peter Zijlstra317f3942011-04-05 17:23:58 +02001877
Peter Zijlstrac05fbaf2011-04-05 17:23:57 +02001878static void ttwu_queue(struct task_struct *p, int cpu)
1879{
1880 struct rq *rq = cpu_rq(cpu);
1881
Daniel Hellstrom17d9f312011-05-20 04:01:10 +00001882#if defined(CONFIG_SMP)
Peter Zijlstra39be3502012-01-26 12:44:34 +01001883 if (sched_feat(TTWU_QUEUE) && !cpus_share_cache(smp_processor_id(), cpu)) {
Peter Zijlstraf01114c2011-05-31 12:26:55 +02001884 sched_clock_cpu(cpu); /* sync clocks x-cpu */
Peter Zijlstra317f3942011-04-05 17:23:58 +02001885 ttwu_queue_remote(p, cpu);
1886 return;
1887 }
1888#endif
1889
Peter Zijlstrac05fbaf2011-04-05 17:23:57 +02001890 raw_spin_lock(&rq->lock);
Peter Zijlstracbce1a62015-06-11 14:46:54 +02001891 lockdep_pin_lock(&rq->lock);
Peter Zijlstrac05fbaf2011-04-05 17:23:57 +02001892 ttwu_do_activate(rq, p, 0);
Peter Zijlstracbce1a62015-06-11 14:46:54 +02001893 lockdep_unpin_lock(&rq->lock);
Peter Zijlstrac05fbaf2011-04-05 17:23:57 +02001894 raw_spin_unlock(&rq->lock);
Tejun Heo9ed38112009-12-03 15:08:03 +09001895}
1896
1897/**
Linus Torvalds1da177e2005-04-16 15:20:36 -07001898 * try_to_wake_up - wake up a thread
Tejun Heo9ed38112009-12-03 15:08:03 +09001899 * @p: the thread to be awakened
Linus Torvalds1da177e2005-04-16 15:20:36 -07001900 * @state: the mask of task states that can be woken
Tejun Heo9ed38112009-12-03 15:08:03 +09001901 * @wake_flags: wake modifier flags (WF_*)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001902 *
1903 * Put it on the run-queue if it's not already there. The "current"
1904 * thread is always on the run-queue (except when the actual
1905 * re-schedule is in progress), and as such you're allowed to do
1906 * the simpler "current->state = TASK_RUNNING" to mark yourself
1907 * runnable without the overhead of this.
1908 *
Yacine Belkadie69f6182013-07-12 20:45:47 +02001909 * Return: %true if @p was woken up, %false if it was already running.
Tejun Heo9ed38112009-12-03 15:08:03 +09001910 * or @state didn't match @p's state.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001911 */
Peter Zijlstrae4a52bc2011-04-05 17:23:54 +02001912static int
1913try_to_wake_up(struct task_struct *p, unsigned int state, int wake_flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001914{
Linus Torvalds1da177e2005-04-16 15:20:36 -07001915 unsigned long flags;
Peter Zijlstrac05fbaf2011-04-05 17:23:57 +02001916 int cpu, success = 0;
Peter Zijlstra2398f2c2008-06-27 13:41:35 +02001917
Oleg Nesterove0acd0a2013-08-12 18:14:00 +02001918 /*
1919 * If we are going to wake up a thread waiting for CONDITION we
1920 * need to ensure that CONDITION=1 done by the caller can not be
1921 * reordered with p->state check below. This pairs with mb() in
1922 * set_current_state() the waiting thread does.
1923 */
1924 smp_mb__before_spinlock();
Peter Zijlstra013fdb82011-04-05 17:23:45 +02001925 raw_spin_lock_irqsave(&p->pi_lock, flags);
Peter Zijlstrae9c84312009-09-15 14:43:03 +02001926 if (!(p->state & state))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001927 goto out;
1928
Peter Zijlstrafbd705a2015-06-09 11:13:36 +02001929 trace_sched_waking(p);
1930
Peter Zijlstrac05fbaf2011-04-05 17:23:57 +02001931 success = 1; /* we're going to change ->state */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001932 cpu = task_cpu(p);
Peter Zijlstrad7c01d22011-04-05 17:23:43 +02001933
Peter Zijlstrac05fbaf2011-04-05 17:23:57 +02001934 if (p->on_rq && ttwu_remote(p, wake_flags))
1935 goto stat;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001936
1937#ifdef CONFIG_SMP
Peter Zijlstrae9c84312009-09-15 14:43:03 +02001938 /*
Peter Zijlstrac05fbaf2011-04-05 17:23:57 +02001939 * If the owning (remote) cpu is still in the middle of schedule() with
1940 * this task as prev, wait until its done referencing the task.
Peter Zijlstrae9c84312009-09-15 14:43:03 +02001941 */
Peter Zijlstraf3e94782012-09-12 11:22:00 +02001942 while (p->on_cpu)
Peter Zijlstrae4a52bc2011-04-05 17:23:54 +02001943 cpu_relax();
Peter Zijlstrae4a52bc2011-04-05 17:23:54 +02001944 /*
1945 * Pairs with the smp_wmb() in finish_lock_switch().
1946 */
1947 smp_rmb();
Linus Torvalds1da177e2005-04-16 15:20:36 -07001948
Peter Zijlstraa8e4f2e2011-04-05 17:23:49 +02001949 p->sched_contributes_to_load = !!task_contributes_to_load(p);
Peter Zijlstrae9c84312009-09-15 14:43:03 +02001950 p->state = TASK_WAKING;
Peter Zijlstraefbbd052009-12-16 18:04:40 +01001951
Peter Zijlstrae4a52bc2011-04-05 17:23:54 +02001952 if (p->sched_class->task_waking)
Peter Zijlstra74f8e4b2011-04-05 17:23:47 +02001953 p->sched_class->task_waking(p);
Peter Zijlstraab19cb22009-11-27 15:44:43 +01001954
Peter Zijlstraac66f542013-10-07 11:29:16 +01001955 cpu = select_task_rq(p, p->wake_cpu, SD_BALANCE_WAKE, wake_flags);
Peter Zijlstraf339b9d2011-05-31 10:49:20 +02001956 if (task_cpu(p) != cpu) {
1957 wake_flags |= WF_MIGRATED;
Mike Galbraithf5dc3752009-10-09 08:35:03 +02001958 set_task_cpu(p, cpu);
Peter Zijlstraf339b9d2011-05-31 10:49:20 +02001959 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001960#endif /* CONFIG_SMP */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001961
Peter Zijlstrac05fbaf2011-04-05 17:23:57 +02001962 ttwu_queue(p, cpu);
1963stat:
Peter Zijlstrab84cb5d2011-04-05 17:23:55 +02001964 ttwu_stat(p, cpu, wake_flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001965out:
Peter Zijlstra013fdb82011-04-05 17:23:45 +02001966 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001967
1968 return success;
1969}
1970
David Howells50fa6102009-04-28 15:01:38 +01001971/**
Tejun Heo21aa9af2010-06-08 21:40:37 +02001972 * try_to_wake_up_local - try to wake up a local task with rq lock held
1973 * @p: the thread to be awakened
1974 *
Peter Zijlstra2acca552011-04-05 17:23:50 +02001975 * Put @p on the run-queue if it's not already there. The caller must
Tejun Heo21aa9af2010-06-08 21:40:37 +02001976 * ensure that this_rq() is locked, @p is bound to this_rq() and not
Peter Zijlstra2acca552011-04-05 17:23:50 +02001977 * the current task.
Tejun Heo21aa9af2010-06-08 21:40:37 +02001978 */
1979static void try_to_wake_up_local(struct task_struct *p)
1980{
1981 struct rq *rq = task_rq(p);
Tejun Heo21aa9af2010-06-08 21:40:37 +02001982
Tejun Heo383efcd2013-03-18 12:22:34 -07001983 if (WARN_ON_ONCE(rq != this_rq()) ||
1984 WARN_ON_ONCE(p == current))
1985 return;
1986
Tejun Heo21aa9af2010-06-08 21:40:37 +02001987 lockdep_assert_held(&rq->lock);
1988
Peter Zijlstra2acca552011-04-05 17:23:50 +02001989 if (!raw_spin_trylock(&p->pi_lock)) {
Peter Zijlstracbce1a62015-06-11 14:46:54 +02001990 /*
1991 * This is OK, because current is on_cpu, which avoids it being
1992 * picked for load-balance and preemption/IRQs are still
1993 * disabled avoiding further scheduler activity on it and we've
1994 * not yet picked a replacement task.
1995 */
1996 lockdep_unpin_lock(&rq->lock);
Peter Zijlstra2acca552011-04-05 17:23:50 +02001997 raw_spin_unlock(&rq->lock);
1998 raw_spin_lock(&p->pi_lock);
1999 raw_spin_lock(&rq->lock);
Peter Zijlstracbce1a62015-06-11 14:46:54 +02002000 lockdep_pin_lock(&rq->lock);
Tejun Heo21aa9af2010-06-08 21:40:37 +02002001 }
Peter Zijlstra2acca552011-04-05 17:23:50 +02002002
Tejun Heo21aa9af2010-06-08 21:40:37 +02002003 if (!(p->state & TASK_NORMAL))
Peter Zijlstra2acca552011-04-05 17:23:50 +02002004 goto out;
Tejun Heo21aa9af2010-06-08 21:40:37 +02002005
Peter Zijlstrafbd705a2015-06-09 11:13:36 +02002006 trace_sched_waking(p);
2007
Kirill Tkhaida0c1e62014-08-20 13:47:32 +04002008 if (!task_on_rq_queued(p))
Peter Zijlstrad7c01d22011-04-05 17:23:43 +02002009 ttwu_activate(rq, p, ENQUEUE_WAKEUP);
2010
Peter Zijlstra23f41ee2011-04-05 17:23:56 +02002011 ttwu_do_wakeup(rq, p, 0);
Peter Zijlstrab84cb5d2011-04-05 17:23:55 +02002012 ttwu_stat(p, smp_processor_id(), 0);
Peter Zijlstra2acca552011-04-05 17:23:50 +02002013out:
2014 raw_spin_unlock(&p->pi_lock);
Tejun Heo21aa9af2010-06-08 21:40:37 +02002015}
2016
2017/**
David Howells50fa6102009-04-28 15:01:38 +01002018 * wake_up_process - Wake up a specific process
2019 * @p: The process to be woken up.
2020 *
2021 * Attempt to wake up the nominated process and move it to the set of runnable
Yacine Belkadie69f6182013-07-12 20:45:47 +02002022 * processes.
2023 *
2024 * Return: 1 if the process was woken up, 0 if it was already running.
David Howells50fa6102009-04-28 15:01:38 +01002025 *
2026 * It may be assumed that this function implies a write memory barrier before
2027 * changing the task state if and only if any tasks are woken up.
2028 */
Harvey Harrison7ad5b3a2008-02-08 04:19:53 -08002029int wake_up_process(struct task_struct *p)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002030{
Oleg Nesterov9067ac82013-01-21 20:48:17 +01002031 WARN_ON(task_is_stopped_or_traced(p));
2032 return try_to_wake_up(p, TASK_NORMAL, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002033}
Linus Torvalds1da177e2005-04-16 15:20:36 -07002034EXPORT_SYMBOL(wake_up_process);
2035
Harvey Harrison7ad5b3a2008-02-08 04:19:53 -08002036int wake_up_state(struct task_struct *p, unsigned int state)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002037{
2038 return try_to_wake_up(p, state, 0);
2039}
2040
Linus Torvalds1da177e2005-04-16 15:20:36 -07002041/*
Juri Lellia5e7be32014-09-19 10:22:39 +01002042 * This function clears the sched_dl_entity static params.
2043 */
2044void __dl_clear_params(struct task_struct *p)
2045{
2046 struct sched_dl_entity *dl_se = &p->dl;
2047
2048 dl_se->dl_runtime = 0;
2049 dl_se->dl_deadline = 0;
2050 dl_se->dl_period = 0;
2051 dl_se->flags = 0;
2052 dl_se->dl_bw = 0;
Peter Zijlstra40767b02015-01-28 15:08:03 +01002053
2054 dl_se->dl_throttled = 0;
2055 dl_se->dl_new = 1;
2056 dl_se->dl_yielded = 0;
Juri Lellia5e7be32014-09-19 10:22:39 +01002057}
2058
2059/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07002060 * Perform scheduler related setup for a newly forked process p.
2061 * p is forked by current.
Ingo Molnardd41f592007-07-09 18:51:59 +02002062 *
2063 * __sched_fork() is basic setup used by init_idle() too:
Linus Torvalds1da177e2005-04-16 15:20:36 -07002064 */
Rik van Riel5e1576e2013-10-07 11:29:26 +01002065static void __sched_fork(unsigned long clone_flags, struct task_struct *p)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002066{
Peter Zijlstrafd2f4412011-04-05 17:23:44 +02002067 p->on_rq = 0;
2068
2069 p->se.on_rq = 0;
Ingo Molnardd41f592007-07-09 18:51:59 +02002070 p->se.exec_start = 0;
2071 p->se.sum_exec_runtime = 0;
Ingo Molnarf6cf8912007-08-28 12:53:24 +02002072 p->se.prev_sum_exec_runtime = 0;
Ingo Molnar6c594c22008-12-14 12:34:15 +01002073 p->se.nr_migrations = 0;
Peter Zijlstrada7a7352011-01-17 17:03:27 +01002074 p->se.vruntime = 0;
Peter Zijlstrafd2f4412011-04-05 17:23:44 +02002075 INIT_LIST_HEAD(&p->se.group_node);
Ingo Molnar6cfb0d52007-08-02 17:41:40 +02002076
2077#ifdef CONFIG_SCHEDSTATS
Lucas De Marchi41acab82010-03-10 23:37:45 -03002078 memset(&p->se.statistics, 0, sizeof(p->se.statistics));
Ingo Molnar6cfb0d52007-08-02 17:41:40 +02002079#endif
Nick Piggin476d1392005-06-25 14:57:29 -07002080
Dario Faggioliaab03e02013-11-28 11:14:43 +01002081 RB_CLEAR_NODE(&p->dl.rb_node);
Peter Zijlstra40767b02015-01-28 15:08:03 +01002082 init_dl_task_timer(&p->dl);
Juri Lellia5e7be32014-09-19 10:22:39 +01002083 __dl_clear_params(p);
Dario Faggioliaab03e02013-11-28 11:14:43 +01002084
Peter Zijlstrafa717062008-01-25 21:08:27 +01002085 INIT_LIST_HEAD(&p->rt.run_list);
Nick Piggin476d1392005-06-25 14:57:29 -07002086
Avi Kivitye107be32007-07-26 13:40:43 +02002087#ifdef CONFIG_PREEMPT_NOTIFIERS
2088 INIT_HLIST_HEAD(&p->preempt_notifiers);
2089#endif
Peter Zijlstracbee9f82012-10-25 14:16:43 +02002090
2091#ifdef CONFIG_NUMA_BALANCING
2092 if (p->mm && atomic_read(&p->mm->mm_users) == 1) {
Mel Gorman7e8d16b2013-10-07 11:28:54 +01002093 p->mm->numa_next_scan = jiffies + msecs_to_jiffies(sysctl_numa_balancing_scan_delay);
Peter Zijlstracbee9f82012-10-25 14:16:43 +02002094 p->mm->numa_scan_seq = 0;
2095 }
2096
Rik van Riel5e1576e2013-10-07 11:29:26 +01002097 if (clone_flags & CLONE_VM)
2098 p->numa_preferred_nid = current->numa_preferred_nid;
2099 else
2100 p->numa_preferred_nid = -1;
2101
Peter Zijlstracbee9f82012-10-25 14:16:43 +02002102 p->node_stamp = 0ULL;
2103 p->numa_scan_seq = p->mm ? p->mm->numa_scan_seq : 0;
Peter Zijlstra4b96a29b2012-10-25 14:16:47 +02002104 p->numa_scan_period = sysctl_numa_balancing_scan_delay;
Peter Zijlstracbee9f82012-10-25 14:16:43 +02002105 p->numa_work.next = &p->numa_work;
Iulia Manda44dba3d2014-10-31 02:13:31 +02002106 p->numa_faults = NULL;
Rik van Riel7e2703e2014-01-27 17:03:45 -05002107 p->last_task_numa_placement = 0;
2108 p->last_sum_exec_runtime = 0;
Peter Zijlstra8c8a7432013-10-07 11:29:21 +01002109
Peter Zijlstra8c8a7432013-10-07 11:29:21 +01002110 p->numa_group = NULL;
Peter Zijlstracbee9f82012-10-25 14:16:43 +02002111#endif /* CONFIG_NUMA_BALANCING */
Ingo Molnardd41f592007-07-09 18:51:59 +02002112}
2113
Mel Gorman1a687c22012-11-22 11:16:36 +00002114#ifdef CONFIG_NUMA_BALANCING
Mel Gorman3105b862012-11-23 11:23:49 +00002115#ifdef CONFIG_SCHED_DEBUG
Mel Gorman1a687c22012-11-22 11:16:36 +00002116void set_numabalancing_state(bool enabled)
2117{
2118 if (enabled)
2119 sched_feat_set("NUMA");
2120 else
2121 sched_feat_set("NO_NUMA");
2122}
Mel Gorman3105b862012-11-23 11:23:49 +00002123#else
2124__read_mostly bool numabalancing_enabled;
2125
2126void set_numabalancing_state(bool enabled)
2127{
2128 numabalancing_enabled = enabled;
2129}
2130#endif /* CONFIG_SCHED_DEBUG */
Andi Kleen54a43d52014-01-23 15:53:13 -08002131
2132#ifdef CONFIG_PROC_SYSCTL
2133int sysctl_numa_balancing(struct ctl_table *table, int write,
2134 void __user *buffer, size_t *lenp, loff_t *ppos)
2135{
2136 struct ctl_table t;
2137 int err;
2138 int state = numabalancing_enabled;
2139
2140 if (write && !capable(CAP_SYS_ADMIN))
2141 return -EPERM;
2142
2143 t = *table;
2144 t.data = &state;
2145 err = proc_dointvec_minmax(&t, write, buffer, lenp, ppos);
2146 if (err < 0)
2147 return err;
2148 if (write)
2149 set_numabalancing_state(state);
2150 return err;
2151}
2152#endif
2153#endif
Mel Gorman1a687c22012-11-22 11:16:36 +00002154
Ingo Molnardd41f592007-07-09 18:51:59 +02002155/*
2156 * fork()/clone()-time setup:
2157 */
Dario Faggioliaab03e02013-11-28 11:14:43 +01002158int sched_fork(unsigned long clone_flags, struct task_struct *p)
Ingo Molnardd41f592007-07-09 18:51:59 +02002159{
Peter Zijlstra0122ec52011-04-05 17:23:51 +02002160 unsigned long flags;
Ingo Molnardd41f592007-07-09 18:51:59 +02002161 int cpu = get_cpu();
2162
Rik van Riel5e1576e2013-10-07 11:29:26 +01002163 __sched_fork(clone_flags, p);
Peter Zijlstra06b83b52009-12-16 18:04:35 +01002164 /*
Peter Zijlstra0017d732010-03-24 18:34:10 +01002165 * We mark the process as running here. This guarantees that
Peter Zijlstra06b83b52009-12-16 18:04:35 +01002166 * nobody will actually run it, and a signal or other external
2167 * event cannot wake it up and insert it on the runqueue either.
2168 */
Peter Zijlstra0017d732010-03-24 18:34:10 +01002169 p->state = TASK_RUNNING;
Ingo Molnardd41f592007-07-09 18:51:59 +02002170
Ingo Molnarb29739f2006-06-27 02:54:51 -07002171 /*
Mike Galbraithc350a042011-07-27 17:14:55 +02002172 * Make sure we do not leak PI boosting priority to the child.
2173 */
2174 p->prio = current->normal_prio;
2175
2176 /*
Mike Galbraithb9dc29e2009-06-17 10:46:01 +02002177 * Revert to default priority/policy on fork if requested.
2178 */
2179 if (unlikely(p->sched_reset_on_fork)) {
Dario Faggioliaab03e02013-11-28 11:14:43 +01002180 if (task_has_dl_policy(p) || task_has_rt_policy(p)) {
Mike Galbraithb9dc29e2009-06-17 10:46:01 +02002181 p->policy = SCHED_NORMAL;
Mike Galbraith6c697bd2009-06-17 10:48:02 +02002182 p->static_prio = NICE_TO_PRIO(0);
Mike Galbraithc350a042011-07-27 17:14:55 +02002183 p->rt_priority = 0;
2184 } else if (PRIO_TO_NICE(p->static_prio) < 0)
2185 p->static_prio = NICE_TO_PRIO(0);
2186
2187 p->prio = p->normal_prio = __normal_prio(p);
2188 set_load_weight(p);
Mike Galbraith6c697bd2009-06-17 10:48:02 +02002189
Mike Galbraithb9dc29e2009-06-17 10:46:01 +02002190 /*
2191 * We don't need the reset flag anymore after the fork. It has
2192 * fulfilled its duty:
2193 */
2194 p->sched_reset_on_fork = 0;
2195 }
Lennart Poetteringca94c442009-06-15 17:17:47 +02002196
Dario Faggioliaab03e02013-11-28 11:14:43 +01002197 if (dl_prio(p->prio)) {
2198 put_cpu();
2199 return -EAGAIN;
2200 } else if (rt_prio(p->prio)) {
2201 p->sched_class = &rt_sched_class;
2202 } else {
Hiroshi Shimamoto2ddbf952007-10-15 17:00:11 +02002203 p->sched_class = &fair_sched_class;
Dario Faggioliaab03e02013-11-28 11:14:43 +01002204 }
Ingo Molnarb29739f2006-06-27 02:54:51 -07002205
Peter Zijlstracd29fe62009-11-27 17:32:46 +01002206 if (p->sched_class->task_fork)
2207 p->sched_class->task_fork(p);
2208
Peter Zijlstra86951592010-06-22 11:44:53 +02002209 /*
2210 * The child is not yet in the pid-hash so no cgroup attach races,
2211 * and the cgroup is pinned to this child due to cgroup_fork()
2212 * is ran before sched_fork().
2213 *
2214 * Silence PROVE_RCU.
2215 */
Peter Zijlstra0122ec52011-04-05 17:23:51 +02002216 raw_spin_lock_irqsave(&p->pi_lock, flags);
Peter Zijlstra5f3edc12009-09-10 13:42:00 +02002217 set_task_cpu(p, cpu);
Peter Zijlstra0122ec52011-04-05 17:23:51 +02002218 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
Peter Zijlstra5f3edc12009-09-10 13:42:00 +02002219
Naveen N. Raof6db8342015-06-25 23:53:37 +05302220#ifdef CONFIG_SCHED_INFO
Ingo Molnardd41f592007-07-09 18:51:59 +02002221 if (likely(sched_info_on()))
Chandra Seetharaman52f17b62006-07-14 00:24:38 -07002222 memset(&p->sched_info, 0, sizeof(p->sched_info));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002223#endif
Peter Zijlstra3ca7a442011-04-05 17:23:40 +02002224#if defined(CONFIG_SMP)
2225 p->on_cpu = 0;
Nick Piggin4866cde2005-06-25 14:57:23 -07002226#endif
Peter Zijlstra01028742013-08-14 14:55:46 +02002227 init_task_preempt_count(p);
Dario Faggioli806c09a2010-11-30 19:51:33 +01002228#ifdef CONFIG_SMP
Gregory Haskins917b6272008-12-29 09:39:53 -05002229 plist_node_init(&p->pushable_tasks, MAX_PRIO);
Juri Lelli1baca4c2013-11-07 14:43:38 +01002230 RB_CLEAR_NODE(&p->pushable_dl_tasks);
Dario Faggioli806c09a2010-11-30 19:51:33 +01002231#endif
Gregory Haskins917b6272008-12-29 09:39:53 -05002232
Nick Piggin476d1392005-06-25 14:57:29 -07002233 put_cpu();
Dario Faggioliaab03e02013-11-28 11:14:43 +01002234 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002235}
2236
Dario Faggioli332ac172013-11-07 14:43:45 +01002237unsigned long to_ratio(u64 period, u64 runtime)
2238{
2239 if (runtime == RUNTIME_INF)
2240 return 1ULL << 20;
2241
2242 /*
2243 * Doing this here saves a lot of checks in all
2244 * the calling paths, and returning zero seems
2245 * safe for them anyway.
2246 */
2247 if (period == 0)
2248 return 0;
2249
2250 return div64_u64(runtime << 20, period);
2251}
2252
2253#ifdef CONFIG_SMP
2254inline struct dl_bw *dl_bw_of(int i)
2255{
Paul E. McKenneyf78f5b92015-06-18 15:50:02 -07002256 RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held(),
2257 "sched RCU must be held");
Dario Faggioli332ac172013-11-07 14:43:45 +01002258 return &cpu_rq(i)->rd->dl_bw;
2259}
2260
Peter Zijlstrade212f12013-12-19 11:54:45 +01002261static inline int dl_bw_cpus(int i)
Dario Faggioli332ac172013-11-07 14:43:45 +01002262{
Peter Zijlstrade212f12013-12-19 11:54:45 +01002263 struct root_domain *rd = cpu_rq(i)->rd;
2264 int cpus = 0;
2265
Paul E. McKenneyf78f5b92015-06-18 15:50:02 -07002266 RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held(),
2267 "sched RCU must be held");
Peter Zijlstrade212f12013-12-19 11:54:45 +01002268 for_each_cpu_and(i, rd->span, cpu_active_mask)
2269 cpus++;
2270
2271 return cpus;
Dario Faggioli332ac172013-11-07 14:43:45 +01002272}
2273#else
2274inline struct dl_bw *dl_bw_of(int i)
2275{
2276 return &cpu_rq(i)->dl.dl_bw;
2277}
2278
Peter Zijlstrade212f12013-12-19 11:54:45 +01002279static inline int dl_bw_cpus(int i)
Dario Faggioli332ac172013-11-07 14:43:45 +01002280{
2281 return 1;
2282}
2283#endif
2284
Dario Faggioli332ac172013-11-07 14:43:45 +01002285/*
2286 * We must be sure that accepting a new task (or allowing changing the
2287 * parameters of an existing one) is consistent with the bandwidth
2288 * constraints. If yes, this function also accordingly updates the currently
2289 * allocated bandwidth to reflect the new situation.
2290 *
2291 * This function is called while holding p's rq->lock.
Peter Zijlstra40767b02015-01-28 15:08:03 +01002292 *
2293 * XXX we should delay bw change until the task's 0-lag point, see
2294 * __setparam_dl().
Dario Faggioli332ac172013-11-07 14:43:45 +01002295 */
2296static int dl_overflow(struct task_struct *p, int policy,
2297 const struct sched_attr *attr)
2298{
2299
2300 struct dl_bw *dl_b = dl_bw_of(task_cpu(p));
Steven Rostedt4df16382014-02-19 13:53:35 -05002301 u64 period = attr->sched_period ?: attr->sched_deadline;
Dario Faggioli332ac172013-11-07 14:43:45 +01002302 u64 runtime = attr->sched_runtime;
2303 u64 new_bw = dl_policy(policy) ? to_ratio(period, runtime) : 0;
Peter Zijlstrade212f12013-12-19 11:54:45 +01002304 int cpus, err = -1;
Dario Faggioli332ac172013-11-07 14:43:45 +01002305
2306 if (new_bw == p->dl.dl_bw)
2307 return 0;
2308
2309 /*
2310 * Either if a task, enters, leave, or stays -deadline but changes
2311 * its parameters, we may need to update accordingly the total
2312 * allocated bandwidth of the container.
2313 */
2314 raw_spin_lock(&dl_b->lock);
Peter Zijlstrade212f12013-12-19 11:54:45 +01002315 cpus = dl_bw_cpus(task_cpu(p));
Dario Faggioli332ac172013-11-07 14:43:45 +01002316 if (dl_policy(policy) && !task_has_dl_policy(p) &&
2317 !__dl_overflow(dl_b, cpus, 0, new_bw)) {
2318 __dl_add(dl_b, new_bw);
2319 err = 0;
2320 } else if (dl_policy(policy) && task_has_dl_policy(p) &&
2321 !__dl_overflow(dl_b, cpus, p->dl.dl_bw, new_bw)) {
2322 __dl_clear(dl_b, p->dl.dl_bw);
2323 __dl_add(dl_b, new_bw);
2324 err = 0;
2325 } else if (!dl_policy(policy) && task_has_dl_policy(p)) {
2326 __dl_clear(dl_b, p->dl.dl_bw);
2327 err = 0;
2328 }
2329 raw_spin_unlock(&dl_b->lock);
2330
2331 return err;
2332}
2333
2334extern void init_dl_bw(struct dl_bw *dl_b);
2335
Linus Torvalds1da177e2005-04-16 15:20:36 -07002336/*
2337 * wake_up_new_task - wake up a newly created task for the first time.
2338 *
2339 * This function will do some initial scheduler statistics housekeeping
2340 * that must be done for every newly created context, then puts the task
2341 * on the runqueue and wakes it.
2342 */
Samir Bellabes3e51e3e2011-05-11 18:18:05 +02002343void wake_up_new_task(struct task_struct *p)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002344{
2345 unsigned long flags;
Ingo Molnardd41f592007-07-09 18:51:59 +02002346 struct rq *rq;
Peter Zijlstrafabf3182010-01-21 21:04:57 +01002347
Peter Zijlstraab2515c2011-04-05 17:23:52 +02002348 raw_spin_lock_irqsave(&p->pi_lock, flags);
Peter Zijlstrafabf3182010-01-21 21:04:57 +01002349#ifdef CONFIG_SMP
2350 /*
2351 * Fork balancing, do it here and not earlier because:
2352 * - cpus_allowed can change in the fork path
2353 * - any previously selected cpu might disappear through hotplug
Peter Zijlstrafabf3182010-01-21 21:04:57 +01002354 */
Peter Zijlstraac66f542013-10-07 11:29:16 +01002355 set_task_cpu(p, select_task_rq(p, task_cpu(p), SD_BALANCE_FORK, 0));
Peter Zijlstrafabf3182010-01-21 21:04:57 +01002356#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07002357
Alex Shia75cdaa2013-06-20 10:18:47 +08002358 /* Initialize new task's runnable average */
Yuyang Du540247f2015-07-15 08:04:39 +08002359 init_entity_runnable_average(&p->se);
Peter Zijlstraab2515c2011-04-05 17:23:52 +02002360 rq = __task_rq_lock(p);
Peter Zijlstracd29fe62009-11-27 17:32:46 +01002361 activate_task(rq, p, 0);
Kirill Tkhaida0c1e62014-08-20 13:47:32 +04002362 p->on_rq = TASK_ON_RQ_QUEUED;
Peter Zijlstrafbd705a2015-06-09 11:13:36 +02002363 trace_sched_wakeup_new(p);
Peter Zijlstraa7558e02009-09-14 20:02:34 +02002364 check_preempt_curr(rq, p, WF_FORK);
Steven Rostedt9a897c52008-01-25 21:08:22 +01002365#ifdef CONFIG_SMP
Peter Zijlstraefbbd052009-12-16 18:04:40 +01002366 if (p->sched_class->task_woken)
2367 p->sched_class->task_woken(rq, p);
Steven Rostedt9a897c52008-01-25 21:08:22 +01002368#endif
Peter Zijlstra0122ec52011-04-05 17:23:51 +02002369 task_rq_unlock(rq, p, &flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002370}
2371
Avi Kivitye107be32007-07-26 13:40:43 +02002372#ifdef CONFIG_PREEMPT_NOTIFIERS
2373
Peter Zijlstra1cde2932015-06-08 16:00:30 +02002374static struct static_key preempt_notifier_key = STATIC_KEY_INIT_FALSE;
2375
Peter Zijlstra2ecd9d22015-07-03 18:53:58 +02002376void preempt_notifier_inc(void)
2377{
2378 static_key_slow_inc(&preempt_notifier_key);
2379}
2380EXPORT_SYMBOL_GPL(preempt_notifier_inc);
2381
2382void preempt_notifier_dec(void)
2383{
2384 static_key_slow_dec(&preempt_notifier_key);
2385}
2386EXPORT_SYMBOL_GPL(preempt_notifier_dec);
2387
Avi Kivitye107be32007-07-26 13:40:43 +02002388/**
Luis Henriques80dd99b2009-03-16 19:58:09 +00002389 * preempt_notifier_register - tell me when current is being preempted & rescheduled
Randy Dunlap421cee22007-07-31 00:37:50 -07002390 * @notifier: notifier struct to register
Avi Kivitye107be32007-07-26 13:40:43 +02002391 */
2392void preempt_notifier_register(struct preempt_notifier *notifier)
2393{
Peter Zijlstra2ecd9d22015-07-03 18:53:58 +02002394 if (!static_key_false(&preempt_notifier_key))
2395 WARN(1, "registering preempt_notifier while notifiers disabled\n");
2396
Avi Kivitye107be32007-07-26 13:40:43 +02002397 hlist_add_head(&notifier->link, &current->preempt_notifiers);
2398}
2399EXPORT_SYMBOL_GPL(preempt_notifier_register);
2400
2401/**
2402 * preempt_notifier_unregister - no longer interested in preemption notifications
Randy Dunlap421cee22007-07-31 00:37:50 -07002403 * @notifier: notifier struct to unregister
Avi Kivitye107be32007-07-26 13:40:43 +02002404 *
Mathieu Desnoyersd84525a2015-05-17 12:53:10 -04002405 * This is *not* safe to call from within a preemption notifier.
Avi Kivitye107be32007-07-26 13:40:43 +02002406 */
2407void preempt_notifier_unregister(struct preempt_notifier *notifier)
2408{
2409 hlist_del(&notifier->link);
2410}
2411EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
2412
Peter Zijlstra1cde2932015-06-08 16:00:30 +02002413static void __fire_sched_in_preempt_notifiers(struct task_struct *curr)
Avi Kivitye107be32007-07-26 13:40:43 +02002414{
2415 struct preempt_notifier *notifier;
Avi Kivitye107be32007-07-26 13:40:43 +02002416
Sasha Levinb67bfe02013-02-27 17:06:00 -08002417 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link)
Avi Kivitye107be32007-07-26 13:40:43 +02002418 notifier->ops->sched_in(notifier, raw_smp_processor_id());
2419}
2420
Peter Zijlstra1cde2932015-06-08 16:00:30 +02002421static __always_inline void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2422{
2423 if (static_key_false(&preempt_notifier_key))
2424 __fire_sched_in_preempt_notifiers(curr);
2425}
2426
Avi Kivitye107be32007-07-26 13:40:43 +02002427static void
Peter Zijlstra1cde2932015-06-08 16:00:30 +02002428__fire_sched_out_preempt_notifiers(struct task_struct *curr,
2429 struct task_struct *next)
Avi Kivitye107be32007-07-26 13:40:43 +02002430{
2431 struct preempt_notifier *notifier;
Avi Kivitye107be32007-07-26 13:40:43 +02002432
Sasha Levinb67bfe02013-02-27 17:06:00 -08002433 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link)
Avi Kivitye107be32007-07-26 13:40:43 +02002434 notifier->ops->sched_out(notifier, next);
2435}
2436
Peter Zijlstra1cde2932015-06-08 16:00:30 +02002437static __always_inline void
2438fire_sched_out_preempt_notifiers(struct task_struct *curr,
2439 struct task_struct *next)
2440{
2441 if (static_key_false(&preempt_notifier_key))
2442 __fire_sched_out_preempt_notifiers(curr, next);
2443}
2444
Dhaval Giani6d6bc0a2008-05-30 14:23:45 +02002445#else /* !CONFIG_PREEMPT_NOTIFIERS */
Avi Kivitye107be32007-07-26 13:40:43 +02002446
Peter Zijlstra1cde2932015-06-08 16:00:30 +02002447static inline void fire_sched_in_preempt_notifiers(struct task_struct *curr)
Avi Kivitye107be32007-07-26 13:40:43 +02002448{
2449}
2450
Peter Zijlstra1cde2932015-06-08 16:00:30 +02002451static inline void
Avi Kivitye107be32007-07-26 13:40:43 +02002452fire_sched_out_preempt_notifiers(struct task_struct *curr,
2453 struct task_struct *next)
2454{
2455}
2456
Dhaval Giani6d6bc0a2008-05-30 14:23:45 +02002457#endif /* CONFIG_PREEMPT_NOTIFIERS */
Avi Kivitye107be32007-07-26 13:40:43 +02002458
Linus Torvalds1da177e2005-04-16 15:20:36 -07002459/**
Nick Piggin4866cde2005-06-25 14:57:23 -07002460 * prepare_task_switch - prepare to switch tasks
2461 * @rq: the runqueue preparing to switch
Randy Dunlap421cee22007-07-31 00:37:50 -07002462 * @prev: the current task that is being switched out
Nick Piggin4866cde2005-06-25 14:57:23 -07002463 * @next: the task we are going to switch to.
2464 *
2465 * This is called with the rq lock held and interrupts off. It must
2466 * be paired with a subsequent finish_task_switch after the context
2467 * switch.
2468 *
2469 * prepare_task_switch sets up locking and calls architecture specific
2470 * hooks.
2471 */
Avi Kivitye107be32007-07-26 13:40:43 +02002472static inline void
2473prepare_task_switch(struct rq *rq, struct task_struct *prev,
2474 struct task_struct *next)
Nick Piggin4866cde2005-06-25 14:57:23 -07002475{
Andrew Vagin895dd922012-07-12 14:14:29 +04002476 trace_sched_switch(prev, next);
Michael S. Tsirkin43148952013-09-22 17:20:54 +03002477 sched_info_switch(rq, prev, next);
Peter Zijlstrafe4b04f2011-02-02 13:19:09 +01002478 perf_event_task_sched_out(prev, next);
Avi Kivitye107be32007-07-26 13:40:43 +02002479 fire_sched_out_preempt_notifiers(prev, next);
Nick Piggin4866cde2005-06-25 14:57:23 -07002480 prepare_lock_switch(rq, next);
2481 prepare_arch_switch(next);
2482}
2483
2484/**
Linus Torvalds1da177e2005-04-16 15:20:36 -07002485 * finish_task_switch - clean up after a task-switch
2486 * @prev: the thread we just switched away from.
2487 *
Nick Piggin4866cde2005-06-25 14:57:23 -07002488 * finish_task_switch must be called after the context switch, paired
2489 * with a prepare_task_switch call before the context switch.
2490 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2491 * and do any other architecture-specific cleanup actions.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002492 *
2493 * Note that we may have delayed dropping an mm in context_switch(). If
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01002494 * so, we finish that here outside of the runqueue lock. (Doing it
Linus Torvalds1da177e2005-04-16 15:20:36 -07002495 * with the lock held can cause deadlocks; see schedule() for
2496 * details.)
Oleg Nesterovdfa50b62014-10-09 21:32:32 +02002497 *
2498 * The context switch have flipped the stack from under us and restored the
2499 * local variables which were saved when this task called schedule() in the
2500 * past. prev == current is still correct but we need to recalculate this_rq
2501 * because prev may have moved to another CPU.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002502 */
Oleg Nesterovdfa50b62014-10-09 21:32:32 +02002503static struct rq *finish_task_switch(struct task_struct *prev)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002504 __releases(rq->lock)
2505{
Oleg Nesterovdfa50b62014-10-09 21:32:32 +02002506 struct rq *rq = this_rq();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002507 struct mm_struct *mm = rq->prev_mm;
Oleg Nesterov55a101f2006-09-29 02:01:10 -07002508 long prev_state;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002509
2510 rq->prev_mm = NULL;
2511
2512 /*
2513 * A task struct has one reference for the use as "current".
Oleg Nesterovc394cc92006-09-29 02:01:11 -07002514 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
Oleg Nesterov55a101f2006-09-29 02:01:10 -07002515 * schedule one last time. The schedule call will never return, and
2516 * the scheduled task must drop that reference.
Oleg Nesterovc394cc92006-09-29 02:01:11 -07002517 * The test for TASK_DEAD must occur while the runqueue locks are
Linus Torvalds1da177e2005-04-16 15:20:36 -07002518 * still held, otherwise prev could be scheduled on another cpu, die
2519 * there before we look at prev->state, and then the reference would
2520 * be dropped twice.
2521 * Manfred Spraul <manfred@colorfullife.com>
2522 */
Oleg Nesterov55a101f2006-09-29 02:01:10 -07002523 prev_state = prev->state;
Frederic Weisbeckerbf9fae92012-09-08 15:23:11 +02002524 vtime_task_switch(prev);
Stephane Eraniana8d757e2011-08-25 15:58:03 +02002525 perf_event_task_sched_in(prev, current);
Nick Piggin4866cde2005-06-25 14:57:23 -07002526 finish_lock_switch(rq, prev);
Catalin Marinas01f23e12011-11-27 21:43:10 +00002527 finish_arch_post_lock_switch();
Steven Rostedte8fa1362008-01-25 21:08:05 +01002528
Avi Kivitye107be32007-07-26 13:40:43 +02002529 fire_sched_in_preempt_notifiers(current);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002530 if (mm)
2531 mmdrop(mm);
Oleg Nesterovc394cc92006-09-29 02:01:11 -07002532 if (unlikely(prev_state == TASK_DEAD)) {
Dario Faggiolie6c390f2013-11-07 14:43:35 +01002533 if (prev->sched_class->task_dead)
2534 prev->sched_class->task_dead(prev);
2535
bibo maoc6fd91f2006-03-26 01:38:20 -08002536 /*
2537 * Remove function-return probe instances associated with this
2538 * task and put them back on the free list.
Ingo Molnar9761eea2007-07-09 18:52:00 +02002539 */
bibo maoc6fd91f2006-03-26 01:38:20 -08002540 kprobe_flush_task(prev);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002541 put_task_struct(prev);
bibo maoc6fd91f2006-03-26 01:38:20 -08002542 }
Frederic Weisbecker99e5ada2013-04-20 17:11:50 +02002543
Frederic Weisbeckerde734f82015-06-11 18:07:12 +02002544 tick_nohz_task_switch();
Oleg Nesterovdfa50b62014-10-09 21:32:32 +02002545 return rq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002546}
2547
Gregory Haskins3f029d32009-07-29 11:08:47 -04002548#ifdef CONFIG_SMP
2549
Gregory Haskins3f029d32009-07-29 11:08:47 -04002550/* rq->lock is NOT held, but preemption is disabled */
Peter Zijlstrae3fca9e2015-06-11 14:46:37 +02002551static void __balance_callback(struct rq *rq)
Gregory Haskins3f029d32009-07-29 11:08:47 -04002552{
Peter Zijlstrae3fca9e2015-06-11 14:46:37 +02002553 struct callback_head *head, *next;
2554 void (*func)(struct rq *rq);
2555 unsigned long flags;
Gregory Haskins3f029d32009-07-29 11:08:47 -04002556
Peter Zijlstrae3fca9e2015-06-11 14:46:37 +02002557 raw_spin_lock_irqsave(&rq->lock, flags);
2558 head = rq->balance_callback;
2559 rq->balance_callback = NULL;
2560 while (head) {
2561 func = (void (*)(struct rq *))head->func;
2562 next = head->next;
2563 head->next = NULL;
2564 head = next;
Gregory Haskins3f029d32009-07-29 11:08:47 -04002565
Peter Zijlstrae3fca9e2015-06-11 14:46:37 +02002566 func(rq);
Gregory Haskins3f029d32009-07-29 11:08:47 -04002567 }
Peter Zijlstrae3fca9e2015-06-11 14:46:37 +02002568 raw_spin_unlock_irqrestore(&rq->lock, flags);
2569}
2570
2571static inline void balance_callback(struct rq *rq)
2572{
2573 if (unlikely(rq->balance_callback))
2574 __balance_callback(rq);
Gregory Haskins3f029d32009-07-29 11:08:47 -04002575}
2576
2577#else
2578
Peter Zijlstrae3fca9e2015-06-11 14:46:37 +02002579static inline void balance_callback(struct rq *rq)
Gregory Haskins3f029d32009-07-29 11:08:47 -04002580{
2581}
2582
2583#endif
2584
Linus Torvalds1da177e2005-04-16 15:20:36 -07002585/**
2586 * schedule_tail - first thing a freshly forked thread must call.
2587 * @prev: the thread we just switched away from.
2588 */
Andi Kleen722a9f92014-05-02 00:44:38 +02002589asmlinkage __visible void schedule_tail(struct task_struct *prev)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002590 __releases(rq->lock)
2591{
Oleg Nesterov1a43a142014-10-08 21:36:44 +02002592 struct rq *rq;
Ingo Molnar70b97a72006-07-03 00:25:42 -07002593
Oleg Nesterov1a43a142014-10-08 21:36:44 +02002594 /* finish_task_switch() drops rq->lock and enables preemtion */
2595 preempt_disable();
Oleg Nesterovdfa50b62014-10-09 21:32:32 +02002596 rq = finish_task_switch(prev);
Peter Zijlstrae3fca9e2015-06-11 14:46:37 +02002597 balance_callback(rq);
Oleg Nesterov1a43a142014-10-08 21:36:44 +02002598 preempt_enable();
Steven Rostedtda19ab52009-07-29 00:21:22 -04002599
Linus Torvalds1da177e2005-04-16 15:20:36 -07002600 if (current->set_child_tid)
Pavel Emelyanovb4888932007-10-18 23:40:14 -07002601 put_user(task_pid_vnr(current), current->set_child_tid);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002602}
2603
2604/*
Oleg Nesterovdfa50b62014-10-09 21:32:32 +02002605 * context_switch - switch to the new MM and the new thread's register state.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002606 */
Oleg Nesterovdfa50b62014-10-09 21:32:32 +02002607static inline struct rq *
Ingo Molnar70b97a72006-07-03 00:25:42 -07002608context_switch(struct rq *rq, struct task_struct *prev,
Ingo Molnar36c8b582006-07-03 00:25:41 -07002609 struct task_struct *next)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002610{
Ingo Molnardd41f592007-07-09 18:51:59 +02002611 struct mm_struct *mm, *oldmm;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002612
Avi Kivitye107be32007-07-26 13:40:43 +02002613 prepare_task_switch(rq, prev, next);
Peter Zijlstrafe4b04f2011-02-02 13:19:09 +01002614
Ingo Molnardd41f592007-07-09 18:51:59 +02002615 mm = next->mm;
2616 oldmm = prev->active_mm;
Zachary Amsden9226d122007-02-13 13:26:21 +01002617 /*
2618 * For paravirt, this is coupled with an exit in switch_to to
2619 * combine the page table reload and the switch backend into
2620 * one hypercall.
2621 */
Jeremy Fitzhardinge224101e2009-02-18 11:18:57 -08002622 arch_start_context_switch(prev);
Zachary Amsden9226d122007-02-13 13:26:21 +01002623
Heiko Carstens31915ab2010-09-16 14:42:25 +02002624 if (!mm) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002625 next->active_mm = oldmm;
2626 atomic_inc(&oldmm->mm_count);
2627 enter_lazy_tlb(oldmm, next);
2628 } else
2629 switch_mm(oldmm, mm, next);
2630
Heiko Carstens31915ab2010-09-16 14:42:25 +02002631 if (!prev->mm) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002632 prev->active_mm = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002633 rq->prev_mm = oldmm;
2634 }
Ingo Molnar3a5f5e42006-07-14 00:24:27 -07002635 /*
2636 * Since the runqueue lock will be released by the next
2637 * task (which is an invalid locking op but in the case
2638 * of the scheduler it's an obvious special-case), so we
2639 * do an early lockdep release here:
2640 */
Peter Zijlstracbce1a62015-06-11 14:46:54 +02002641 lockdep_unpin_lock(&rq->lock);
Ingo Molnar8a25d5d2006-07-03 00:24:54 -07002642 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002643
2644 /* Here we just switch the register state and the stack. */
2645 switch_to(prev, next, prev);
Ingo Molnardd41f592007-07-09 18:51:59 +02002646 barrier();
Oleg Nesterovdfa50b62014-10-09 21:32:32 +02002647
2648 return finish_task_switch(prev);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002649}
2650
2651/*
Sha Zhengju1c3e8262013-02-20 17:14:38 +08002652 * nr_running and nr_context_switches:
Linus Torvalds1da177e2005-04-16 15:20:36 -07002653 *
2654 * externally visible scheduler statistics: current number of runnable
Sha Zhengju1c3e8262013-02-20 17:14:38 +08002655 * threads, total number of context switches performed since bootup.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002656 */
2657unsigned long nr_running(void)
2658{
2659 unsigned long i, sum = 0;
2660
2661 for_each_online_cpu(i)
2662 sum += cpu_rq(i)->nr_running;
2663
2664 return sum;
2665}
2666
Tim Chen2ee507c2014-07-31 10:29:48 -07002667/*
2668 * Check if only the current task is running on the cpu.
2669 */
2670bool single_task_running(void)
2671{
2672 if (cpu_rq(smp_processor_id())->nr_running == 1)
2673 return true;
2674 else
2675 return false;
2676}
2677EXPORT_SYMBOL(single_task_running);
2678
Linus Torvalds1da177e2005-04-16 15:20:36 -07002679unsigned long long nr_context_switches(void)
2680{
Steven Rostedtcc94abf2006-06-27 02:54:31 -07002681 int i;
2682 unsigned long long sum = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002683
KAMEZAWA Hiroyuki0a945022006-03-28 01:56:37 -08002684 for_each_possible_cpu(i)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002685 sum += cpu_rq(i)->nr_switches;
2686
2687 return sum;
2688}
2689
2690unsigned long nr_iowait(void)
2691{
2692 unsigned long i, sum = 0;
2693
KAMEZAWA Hiroyuki0a945022006-03-28 01:56:37 -08002694 for_each_possible_cpu(i)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002695 sum += atomic_read(&cpu_rq(i)->nr_iowait);
2696
2697 return sum;
2698}
2699
Peter Zijlstra8c215bd2010-07-01 09:07:17 +02002700unsigned long nr_iowait_cpu(int cpu)
Arjan van de Ven69d25872009-09-21 17:04:08 -07002701{
Peter Zijlstra8c215bd2010-07-01 09:07:17 +02002702 struct rq *this = cpu_rq(cpu);
Arjan van de Ven69d25872009-09-21 17:04:08 -07002703 return atomic_read(&this->nr_iowait);
2704}
2705
Mel Gorman372ba8c2014-08-06 14:19:21 +01002706void get_iowait_load(unsigned long *nr_waiters, unsigned long *load)
2707{
Peter Zijlstra3289bdb2015-04-14 13:19:42 +02002708 struct rq *rq = this_rq();
2709 *nr_waiters = atomic_read(&rq->nr_iowait);
2710 *load = rq->load.weight;
Mel Gorman372ba8c2014-08-06 14:19:21 +01002711}
2712
Ingo Molnardd41f592007-07-09 18:51:59 +02002713#ifdef CONFIG_SMP
2714
Ingo Molnar48f24c42006-07-03 00:25:40 -07002715/*
Peter Zijlstra38022902009-12-16 18:04:37 +01002716 * sched_exec - execve() is a valuable balancing opportunity, because at
2717 * this point the task has the smallest effective memory and cache footprint.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002718 */
Peter Zijlstra38022902009-12-16 18:04:37 +01002719void sched_exec(void)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002720{
Peter Zijlstra38022902009-12-16 18:04:37 +01002721 struct task_struct *p = current;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002722 unsigned long flags;
Peter Zijlstra0017d732010-03-24 18:34:10 +01002723 int dest_cpu;
Peter Zijlstra38022902009-12-16 18:04:37 +01002724
Peter Zijlstra8f42ced2011-04-05 17:23:53 +02002725 raw_spin_lock_irqsave(&p->pi_lock, flags);
Peter Zijlstraac66f542013-10-07 11:29:16 +01002726 dest_cpu = p->sched_class->select_task_rq(p, task_cpu(p), SD_BALANCE_EXEC, 0);
Peter Zijlstra0017d732010-03-24 18:34:10 +01002727 if (dest_cpu == smp_processor_id())
2728 goto unlock;
Peter Zijlstra38022902009-12-16 18:04:37 +01002729
Peter Zijlstra8f42ced2011-04-05 17:23:53 +02002730 if (likely(cpu_active(dest_cpu))) {
Tejun Heo969c7922010-05-06 18:49:21 +02002731 struct migration_arg arg = { p, dest_cpu };
Ingo Molnar36c8b582006-07-03 00:25:41 -07002732
Peter Zijlstra8f42ced2011-04-05 17:23:53 +02002733 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2734 stop_one_cpu(task_cpu(p), migration_cpu_stop, &arg);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002735 return;
2736 }
Peter Zijlstra0017d732010-03-24 18:34:10 +01002737unlock:
Peter Zijlstra8f42ced2011-04-05 17:23:53 +02002738 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002739}
2740
Linus Torvalds1da177e2005-04-16 15:20:36 -07002741#endif
2742
Linus Torvalds1da177e2005-04-16 15:20:36 -07002743DEFINE_PER_CPU(struct kernel_stat, kstat);
Glauber Costa3292beb2011-11-28 14:45:17 -02002744DEFINE_PER_CPU(struct kernel_cpustat, kernel_cpustat);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002745
2746EXPORT_PER_CPU_SYMBOL(kstat);
Glauber Costa3292beb2011-11-28 14:45:17 -02002747EXPORT_PER_CPU_SYMBOL(kernel_cpustat);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002748
2749/*
Hidetoshi Setoc5f8d992009-03-31 16:56:03 +09002750 * Return accounted runtime for the task.
2751 * In case the task is currently running, return the runtime plus current's
2752 * pending runtime that have not been accounted yet.
2753 */
2754unsigned long long task_sched_runtime(struct task_struct *p)
2755{
2756 unsigned long flags;
2757 struct rq *rq;
Stanislaw Gruszka6e998912014-11-12 16:58:44 +01002758 u64 ns;
Ingo Molnar48f24c42006-07-03 00:25:40 -07002759
Peter Zijlstra911b2892013-11-11 18:21:56 +01002760#if defined(CONFIG_64BIT) && defined(CONFIG_SMP)
2761 /*
2762 * 64-bit doesn't need locks to atomically read a 64bit value.
2763 * So we have a optimization chance when the task's delta_exec is 0.
2764 * Reading ->on_cpu is racy, but this is ok.
2765 *
2766 * If we race with it leaving cpu, we'll take a lock. So we're correct.
2767 * If we race with it entering cpu, unaccounted time is 0. This is
2768 * indistinguishable from the read occurring a few cycles earlier.
Mike Galbraith4036ac12014-06-24 07:49:40 +02002769 * If we see ->on_cpu without ->on_rq, the task is leaving, and has
2770 * been accounted, so we're correct here as well.
Peter Zijlstra911b2892013-11-11 18:21:56 +01002771 */
Kirill Tkhaida0c1e62014-08-20 13:47:32 +04002772 if (!p->on_cpu || !task_on_rq_queued(p))
Peter Zijlstra911b2892013-11-11 18:21:56 +01002773 return p->se.sum_exec_runtime;
2774#endif
2775
Hidetoshi Setoc5f8d992009-03-31 16:56:03 +09002776 rq = task_rq_lock(p, &flags);
Stanislaw Gruszka6e998912014-11-12 16:58:44 +01002777 /*
2778 * Must be ->curr _and_ ->on_rq. If dequeued, we would
2779 * project cycles that may never be accounted to this
2780 * thread, breaking clock_gettime().
2781 */
2782 if (task_current(rq, p) && task_on_rq_queued(p)) {
2783 update_rq_clock(rq);
2784 p->sched_class->update_curr(rq);
2785 }
2786 ns = p->se.sum_exec_runtime;
Peter Zijlstra0122ec52011-04-05 17:23:51 +02002787 task_rq_unlock(rq, p, &flags);
Hidetoshi Setoc5f8d992009-03-31 16:56:03 +09002788
2789 return ns;
2790}
2791
Balbir Singh49048622008-09-05 18:12:23 +02002792/*
Christoph Lameter7835b982006-12-10 02:20:22 -08002793 * This function gets called by the timer code, with HZ frequency.
2794 * We call it with interrupts disabled.
Christoph Lameter7835b982006-12-10 02:20:22 -08002795 */
2796void scheduler_tick(void)
2797{
Christoph Lameter7835b982006-12-10 02:20:22 -08002798 int cpu = smp_processor_id();
2799 struct rq *rq = cpu_rq(cpu);
Ingo Molnardd41f592007-07-09 18:51:59 +02002800 struct task_struct *curr = rq->curr;
Peter Zijlstra3e51f332008-05-03 18:29:28 +02002801
2802 sched_clock_tick();
Christoph Lameter7835b982006-12-10 02:20:22 -08002803
Thomas Gleixner05fa7852009-11-17 14:28:38 +01002804 raw_spin_lock(&rq->lock);
Peter Zijlstra3e51f332008-05-03 18:29:28 +02002805 update_rq_clock(rq);
Peter Zijlstrafa85ae22008-01-25 21:08:29 +01002806 curr->sched_class->task_tick(rq, curr, 0);
Alex Shi83dfd522013-06-20 10:18:49 +08002807 update_cpu_load_active(rq);
Peter Zijlstra3289bdb2015-04-14 13:19:42 +02002808 calc_global_load_tick(rq);
Thomas Gleixner05fa7852009-11-17 14:28:38 +01002809 raw_spin_unlock(&rq->lock);
Ingo Molnardd41f592007-07-09 18:51:59 +02002810
Peter Zijlstrae9d2b062010-09-17 11:28:50 +02002811 perf_event_task_tick();
Peter Zijlstrae220d2d2009-05-23 18:28:55 +02002812
Christoph Lametere418e1c2006-12-10 02:20:23 -08002813#ifdef CONFIG_SMP
Suresh Siddha6eb57e02011-10-03 15:09:01 -07002814 rq->idle_balance = idle_cpu(cpu);
Daniel Lezcano7caff662014-01-06 12:34:38 +01002815 trigger_load_balance(rq);
Christoph Lametere418e1c2006-12-10 02:20:23 -08002816#endif
Frederic Weisbecker265f22a2013-05-03 03:39:05 +02002817 rq_last_tick_reset(rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002818}
2819
Frederic Weisbecker265f22a2013-05-03 03:39:05 +02002820#ifdef CONFIG_NO_HZ_FULL
2821/**
2822 * scheduler_tick_max_deferment
2823 *
2824 * Keep at least one tick per second when a single
2825 * active task is running because the scheduler doesn't
2826 * yet completely support full dynticks environment.
2827 *
2828 * This makes sure that uptime, CFS vruntime, load
2829 * balancing, etc... continue to move forward, even
2830 * with a very low granularity.
Yacine Belkadie69f6182013-07-12 20:45:47 +02002831 *
2832 * Return: Maximum deferment in nanoseconds.
Frederic Weisbecker265f22a2013-05-03 03:39:05 +02002833 */
2834u64 scheduler_tick_max_deferment(void)
2835{
2836 struct rq *rq = this_rq();
Jason Low316c1608d2015-04-28 13:00:20 -07002837 unsigned long next, now = READ_ONCE(jiffies);
Frederic Weisbecker265f22a2013-05-03 03:39:05 +02002838
2839 next = rq->last_sched_tick + HZ;
2840
2841 if (time_before_eq(next, now))
2842 return 0;
2843
Kevin Hilman8fe8ff02014-01-15 14:51:38 +01002844 return jiffies_to_nsecs(next - now);
Frederic Weisbecker265f22a2013-05-03 03:39:05 +02002845}
2846#endif
2847
Lai Jiangshan132380a2009-04-02 14:18:25 +08002848notrace unsigned long get_parent_ip(unsigned long addr)
Steven Rostedt6cd8a4b2008-05-12 21:20:42 +02002849{
2850 if (in_lock_functions(addr)) {
2851 addr = CALLER_ADDR2;
2852 if (in_lock_functions(addr))
2853 addr = CALLER_ADDR3;
2854 }
2855 return addr;
2856}
Linus Torvalds1da177e2005-04-16 15:20:36 -07002857
Steven Rostedt7e49fcc2009-01-22 19:01:40 -05002858#if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
2859 defined(CONFIG_PREEMPT_TRACER))
2860
Masami Hiramatsuedafe3a2014-04-17 17:18:42 +09002861void preempt_count_add(int val)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002862{
Steven Rostedt6cd8a4b2008-05-12 21:20:42 +02002863#ifdef CONFIG_DEBUG_PREEMPT
Linus Torvalds1da177e2005-04-16 15:20:36 -07002864 /*
2865 * Underflow?
2866 */
Ingo Molnar9a11b49a2006-07-03 00:24:33 -07002867 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
2868 return;
Steven Rostedt6cd8a4b2008-05-12 21:20:42 +02002869#endif
Peter Zijlstrabdb43802013-09-10 12:15:23 +02002870 __preempt_count_add(val);
Steven Rostedt6cd8a4b2008-05-12 21:20:42 +02002871#ifdef CONFIG_DEBUG_PREEMPT
Linus Torvalds1da177e2005-04-16 15:20:36 -07002872 /*
2873 * Spinlock count overflowing soon?
2874 */
Miguel Ojeda Sandonis33859f72006-12-10 02:20:38 -08002875 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
2876 PREEMPT_MASK - 10);
Steven Rostedt6cd8a4b2008-05-12 21:20:42 +02002877#endif
Thomas Gleixner8f47b182014-02-07 20:58:39 +01002878 if (preempt_count() == val) {
2879 unsigned long ip = get_parent_ip(CALLER_ADDR1);
2880#ifdef CONFIG_DEBUG_PREEMPT
2881 current->preempt_disable_ip = ip;
2882#endif
2883 trace_preempt_off(CALLER_ADDR0, ip);
2884 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002885}
Peter Zijlstrabdb43802013-09-10 12:15:23 +02002886EXPORT_SYMBOL(preempt_count_add);
Masami Hiramatsuedafe3a2014-04-17 17:18:42 +09002887NOKPROBE_SYMBOL(preempt_count_add);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002888
Masami Hiramatsuedafe3a2014-04-17 17:18:42 +09002889void preempt_count_sub(int val)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002890{
Steven Rostedt6cd8a4b2008-05-12 21:20:42 +02002891#ifdef CONFIG_DEBUG_PREEMPT
Linus Torvalds1da177e2005-04-16 15:20:36 -07002892 /*
2893 * Underflow?
2894 */
Ingo Molnar01e3eb82009-01-12 13:00:50 +01002895 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
Ingo Molnar9a11b49a2006-07-03 00:24:33 -07002896 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002897 /*
2898 * Is the spinlock portion underflowing?
2899 */
Ingo Molnar9a11b49a2006-07-03 00:24:33 -07002900 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
2901 !(preempt_count() & PREEMPT_MASK)))
2902 return;
Steven Rostedt6cd8a4b2008-05-12 21:20:42 +02002903#endif
Ingo Molnar9a11b49a2006-07-03 00:24:33 -07002904
Steven Rostedt6cd8a4b2008-05-12 21:20:42 +02002905 if (preempt_count() == val)
2906 trace_preempt_on(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
Peter Zijlstrabdb43802013-09-10 12:15:23 +02002907 __preempt_count_sub(val);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002908}
Peter Zijlstrabdb43802013-09-10 12:15:23 +02002909EXPORT_SYMBOL(preempt_count_sub);
Masami Hiramatsuedafe3a2014-04-17 17:18:42 +09002910NOKPROBE_SYMBOL(preempt_count_sub);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002911
2912#endif
2913
2914/*
Ingo Molnardd41f592007-07-09 18:51:59 +02002915 * Print scheduling while atomic bug:
Linus Torvalds1da177e2005-04-16 15:20:36 -07002916 */
Ingo Molnardd41f592007-07-09 18:51:59 +02002917static noinline void __schedule_bug(struct task_struct *prev)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002918{
Dave Jones664dfa62011-12-22 16:39:30 -05002919 if (oops_in_progress)
2920 return;
2921
Peter Zijlstra3df0fc52009-12-20 14:23:57 +01002922 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
2923 prev->comm, prev->pid, preempt_count());
Satyam Sharma838225b2007-10-24 18:23:50 +02002924
Ingo Molnardd41f592007-07-09 18:51:59 +02002925 debug_show_held_locks(prev);
Arjan van de Vene21f5b12008-05-23 09:05:58 -07002926 print_modules();
Ingo Molnardd41f592007-07-09 18:51:59 +02002927 if (irqs_disabled())
2928 print_irqtrace_events(prev);
Thomas Gleixner8f47b182014-02-07 20:58:39 +01002929#ifdef CONFIG_DEBUG_PREEMPT
2930 if (in_atomic_preempt_off()) {
2931 pr_err("Preemption disabled at:");
2932 print_ip_sym(current->preempt_disable_ip);
2933 pr_cont("\n");
2934 }
2935#endif
Stephen Boyd6135fc12012-03-28 17:10:47 -07002936 dump_stack();
Rusty Russell373d4d02013-01-21 17:17:39 +10302937 add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
Ingo Molnardd41f592007-07-09 18:51:59 +02002938}
Linus Torvalds1da177e2005-04-16 15:20:36 -07002939
Ingo Molnardd41f592007-07-09 18:51:59 +02002940/*
2941 * Various schedule()-time debugging checks and statistics:
2942 */
2943static inline void schedule_debug(struct task_struct *prev)
2944{
Aaron Tomlin0d9e2632014-09-12 14:16:19 +01002945#ifdef CONFIG_SCHED_STACK_END_CHECK
2946 BUG_ON(unlikely(task_stack_end_corrupted(prev)));
2947#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07002948 /*
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01002949 * Test if we are atomic. Since do_exit() needs to call into
Oleg Nesterov192301e2013-11-13 16:45:38 +01002950 * schedule() atomically, we ignore that path. Otherwise whine
2951 * if we are scheduling when we should not.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002952 */
Oleg Nesterov192301e2013-11-13 16:45:38 +01002953 if (unlikely(in_atomic_preempt_off() && prev->state != TASK_DEAD))
Ingo Molnardd41f592007-07-09 18:51:59 +02002954 __schedule_bug(prev);
Paul E. McKenneyb3fbab02011-05-24 08:31:09 -07002955 rcu_sleep_check();
Ingo Molnardd41f592007-07-09 18:51:59 +02002956
Linus Torvalds1da177e2005-04-16 15:20:36 -07002957 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
2958
Ingo Molnar2d723762007-10-15 17:00:12 +02002959 schedstat_inc(this_rq(), sched_count);
Ingo Molnardd41f592007-07-09 18:51:59 +02002960}
2961
2962/*
2963 * Pick up the highest-prio task:
2964 */
2965static inline struct task_struct *
Peter Zijlstra606dba22012-02-11 06:05:00 +01002966pick_next_task(struct rq *rq, struct task_struct *prev)
Ingo Molnardd41f592007-07-09 18:51:59 +02002967{
Peter Zijlstra37e117c2014-02-14 12:25:08 +01002968 const struct sched_class *class = &fair_sched_class;
Ingo Molnardd41f592007-07-09 18:51:59 +02002969 struct task_struct *p;
2970
2971 /*
2972 * Optimization: we know that if all tasks are in
2973 * the fair class we can call that function directly:
2974 */
Peter Zijlstra37e117c2014-02-14 12:25:08 +01002975 if (likely(prev->sched_class == class &&
Peter Zijlstra38033c32014-01-23 20:32:21 +01002976 rq->nr_running == rq->cfs.h_nr_running)) {
Peter Zijlstra606dba22012-02-11 06:05:00 +01002977 p = fair_sched_class.pick_next_task(rq, prev);
Peter Zijlstra6ccdc842014-04-24 12:00:47 +02002978 if (unlikely(p == RETRY_TASK))
2979 goto again;
2980
2981 /* assumes fair_sched_class->next == idle_sched_class */
2982 if (unlikely(!p))
2983 p = idle_sched_class.pick_next_task(rq, prev);
2984
2985 return p;
Ingo Molnardd41f592007-07-09 18:51:59 +02002986 }
2987
Peter Zijlstra37e117c2014-02-14 12:25:08 +01002988again:
Peter Zijlstra34f971f2010-09-22 13:53:15 +02002989 for_each_class(class) {
Peter Zijlstra606dba22012-02-11 06:05:00 +01002990 p = class->pick_next_task(rq, prev);
Peter Zijlstra37e117c2014-02-14 12:25:08 +01002991 if (p) {
2992 if (unlikely(p == RETRY_TASK))
2993 goto again;
Ingo Molnardd41f592007-07-09 18:51:59 +02002994 return p;
Peter Zijlstra37e117c2014-02-14 12:25:08 +01002995 }
Ingo Molnardd41f592007-07-09 18:51:59 +02002996 }
Peter Zijlstra34f971f2010-09-22 13:53:15 +02002997
2998 BUG(); /* the idle class will always have a runnable task */
Ingo Molnardd41f592007-07-09 18:51:59 +02002999}
3000
3001/*
Thomas Gleixnerc259e012011-06-22 19:47:00 +02003002 * __schedule() is the main scheduler function.
Pekka Enbergedde96e2012-08-04 11:49:47 +03003003 *
3004 * The main means of driving the scheduler and thus entering this function are:
3005 *
3006 * 1. Explicit blocking: mutex, semaphore, waitqueue, etc.
3007 *
3008 * 2. TIF_NEED_RESCHED flag is checked on interrupt and userspace return
3009 * paths. For example, see arch/x86/entry_64.S.
3010 *
3011 * To drive preemption between tasks, the scheduler sets the flag in timer
3012 * interrupt handler scheduler_tick().
3013 *
3014 * 3. Wakeups don't really cause entry into schedule(). They add a
3015 * task to the run-queue and that's it.
3016 *
3017 * Now, if the new task added to the run-queue preempts the current
3018 * task, then the wakeup sets TIF_NEED_RESCHED and schedule() gets
3019 * called on the nearest possible occasion:
3020 *
3021 * - If the kernel is preemptible (CONFIG_PREEMPT=y):
3022 *
3023 * - in syscall or exception context, at the next outmost
3024 * preempt_enable(). (this might be as soon as the wake_up()'s
3025 * spin_unlock()!)
3026 *
3027 * - in IRQ context, return from interrupt-handler to
3028 * preemptible context
3029 *
3030 * - If the kernel is not preemptible (CONFIG_PREEMPT is not set)
3031 * then at the next:
3032 *
3033 * - cond_resched() call
3034 * - explicit schedule() call
3035 * - return from syscall or exception to user-space
3036 * - return from interrupt-handler to user-space
Frederic Weisbeckerbfd9b2b2015-01-28 01:24:09 +01003037 *
Frederic Weisbeckerb30f0e32015-05-12 16:41:49 +02003038 * WARNING: must be called with preemption disabled!
Ingo Molnardd41f592007-07-09 18:51:59 +02003039 */
Thomas Gleixnerc259e012011-06-22 19:47:00 +02003040static void __sched __schedule(void)
Ingo Molnardd41f592007-07-09 18:51:59 +02003041{
3042 struct task_struct *prev, *next;
Harvey Harrison67ca7bd2008-02-15 09:56:36 -08003043 unsigned long *switch_count;
Ingo Molnardd41f592007-07-09 18:51:59 +02003044 struct rq *rq;
Peter Zijlstra31656512008-07-18 18:01:23 +02003045 int cpu;
Ingo Molnardd41f592007-07-09 18:51:59 +02003046
Ingo Molnardd41f592007-07-09 18:51:59 +02003047 cpu = smp_processor_id();
3048 rq = cpu_rq(cpu);
Paul E. McKenney38200cf2014-10-21 12:50:04 -07003049 rcu_note_context_switch();
Ingo Molnardd41f592007-07-09 18:51:59 +02003050 prev = rq->curr;
Ingo Molnardd41f592007-07-09 18:51:59 +02003051
Ingo Molnardd41f592007-07-09 18:51:59 +02003052 schedule_debug(prev);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003053
Peter Zijlstra31656512008-07-18 18:01:23 +02003054 if (sched_feat(HRTICK))
Mike Galbraithf333fdc2008-05-12 21:20:55 +02003055 hrtick_clear(rq);
Peter Zijlstra8f4d37e2008-01-25 21:08:29 +01003056
Oleg Nesterove0acd0a2013-08-12 18:14:00 +02003057 /*
3058 * Make sure that signal_pending_state()->signal_pending() below
3059 * can't be reordered with __set_current_state(TASK_INTERRUPTIBLE)
3060 * done by the caller to avoid the race with signal_wake_up().
3061 */
3062 smp_mb__before_spinlock();
Thomas Gleixner05fa7852009-11-17 14:28:38 +01003063 raw_spin_lock_irq(&rq->lock);
Peter Zijlstracbce1a62015-06-11 14:46:54 +02003064 lockdep_pin_lock(&rq->lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003065
Peter Zijlstra9edfbfe2015-01-05 11:18:11 +01003066 rq->clock_skip_update <<= 1; /* promote REQ to ACT */
3067
Oleg Nesterov246d86b2010-05-19 14:57:11 +02003068 switch_count = &prev->nivcsw;
Ingo Molnardd41f592007-07-09 18:51:59 +02003069 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
Tejun Heo21aa9af2010-06-08 21:40:37 +02003070 if (unlikely(signal_pending_state(prev->state, prev))) {
Ingo Molnardd41f592007-07-09 18:51:59 +02003071 prev->state = TASK_RUNNING;
Tejun Heo21aa9af2010-06-08 21:40:37 +02003072 } else {
Peter Zijlstra2acca552011-04-05 17:23:50 +02003073 deactivate_task(rq, prev, DEQUEUE_SLEEP);
3074 prev->on_rq = 0;
3075
Tejun Heo21aa9af2010-06-08 21:40:37 +02003076 /*
Peter Zijlstra2acca552011-04-05 17:23:50 +02003077 * If a worker went to sleep, notify and ask workqueue
3078 * whether it wants to wake up a task to maintain
3079 * concurrency.
Tejun Heo21aa9af2010-06-08 21:40:37 +02003080 */
3081 if (prev->flags & PF_WQ_WORKER) {
3082 struct task_struct *to_wakeup;
3083
3084 to_wakeup = wq_worker_sleeping(prev, cpu);
3085 if (to_wakeup)
3086 try_to_wake_up_local(to_wakeup);
3087 }
Tejun Heo21aa9af2010-06-08 21:40:37 +02003088 }
Ingo Molnardd41f592007-07-09 18:51:59 +02003089 switch_count = &prev->nvcsw;
3090 }
3091
Peter Zijlstra9edfbfe2015-01-05 11:18:11 +01003092 if (task_on_rq_queued(prev))
Peter Zijlstra606dba22012-02-11 06:05:00 +01003093 update_rq_clock(rq);
3094
3095 next = pick_next_task(rq, prev);
Mike Galbraithf26f9af2010-12-08 11:05:42 +01003096 clear_tsk_need_resched(prev);
Peter Zijlstraf27dde82013-08-14 14:55:31 +02003097 clear_preempt_need_resched();
Peter Zijlstra9edfbfe2015-01-05 11:18:11 +01003098 rq->clock_skip_update = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003099
Linus Torvalds1da177e2005-04-16 15:20:36 -07003100 if (likely(prev != next)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003101 rq->nr_switches++;
3102 rq->curr = next;
3103 ++*switch_count;
3104
Oleg Nesterovdfa50b62014-10-09 21:32:32 +02003105 rq = context_switch(rq, prev, next); /* unlocks the rq */
3106 cpu = cpu_of(rq);
Peter Zijlstracbce1a62015-06-11 14:46:54 +02003107 } else {
3108 lockdep_unpin_lock(&rq->lock);
Thomas Gleixner05fa7852009-11-17 14:28:38 +01003109 raw_spin_unlock_irq(&rq->lock);
Peter Zijlstracbce1a62015-06-11 14:46:54 +02003110 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07003111
Peter Zijlstrae3fca9e2015-06-11 14:46:37 +02003112 balance_callback(rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003113}
Thomas Gleixnerc259e012011-06-22 19:47:00 +02003114
Thomas Gleixner9c40cef2011-06-22 19:47:01 +02003115static inline void sched_submit_work(struct task_struct *tsk)
3116{
Thomas Gleixner3c7d5182011-07-17 20:46:52 +02003117 if (!tsk->state || tsk_is_pi_blocked(tsk))
Thomas Gleixner9c40cef2011-06-22 19:47:01 +02003118 return;
3119 /*
3120 * If we are going to sleep and we have plugged IO queued,
3121 * make sure to submit it to avoid deadlocks.
3122 */
3123 if (blk_needs_flush_plug(tsk))
3124 blk_schedule_flush_plug(tsk);
3125}
3126
Andi Kleen722a9f92014-05-02 00:44:38 +02003127asmlinkage __visible void __sched schedule(void)
Thomas Gleixnerc259e012011-06-22 19:47:00 +02003128{
Thomas Gleixner9c40cef2011-06-22 19:47:01 +02003129 struct task_struct *tsk = current;
3130
3131 sched_submit_work(tsk);
Frederic Weisbeckerbfd9b2b2015-01-28 01:24:09 +01003132 do {
Frederic Weisbeckerb30f0e32015-05-12 16:41:49 +02003133 preempt_disable();
Frederic Weisbeckerbfd9b2b2015-01-28 01:24:09 +01003134 __schedule();
Frederic Weisbeckerb30f0e32015-05-12 16:41:49 +02003135 sched_preempt_enable_no_resched();
Frederic Weisbeckerbfd9b2b2015-01-28 01:24:09 +01003136 } while (need_resched());
Thomas Gleixnerc259e012011-06-22 19:47:00 +02003137}
Linus Torvalds1da177e2005-04-16 15:20:36 -07003138EXPORT_SYMBOL(schedule);
3139
Frederic Weisbecker91d1aa432012-11-27 19:33:25 +01003140#ifdef CONFIG_CONTEXT_TRACKING
Andi Kleen722a9f92014-05-02 00:44:38 +02003141asmlinkage __visible void __sched schedule_user(void)
Frederic Weisbecker20ab65e32012-07-11 20:26:37 +02003142{
3143 /*
3144 * If we come here after a random call to set_need_resched(),
3145 * or we have been woken up remotely but the IPI has not yet arrived,
3146 * we haven't yet exited the RCU idle mode. Do it here manually until
3147 * we find a better solution.
Andy Lutomirski7cc78f82014-12-03 15:37:08 -08003148 *
3149 * NB: There are buggy callers of this function. Ideally we
Frederic Weisbeckerc467ea72015-03-04 18:06:33 +01003150 * should warn if prev_state != CONTEXT_USER, but that will trigger
Andy Lutomirski7cc78f82014-12-03 15:37:08 -08003151 * too frequently to make sense yet.
Frederic Weisbecker20ab65e32012-07-11 20:26:37 +02003152 */
Andy Lutomirski7cc78f82014-12-03 15:37:08 -08003153 enum ctx_state prev_state = exception_enter();
Frederic Weisbecker20ab65e32012-07-11 20:26:37 +02003154 schedule();
Andy Lutomirski7cc78f82014-12-03 15:37:08 -08003155 exception_exit(prev_state);
Frederic Weisbecker20ab65e32012-07-11 20:26:37 +02003156}
3157#endif
3158
Thomas Gleixnerc5491ea2011-03-21 12:09:35 +01003159/**
3160 * schedule_preempt_disabled - called with preemption disabled
3161 *
3162 * Returns with preemption disabled. Note: preempt_count must be 1
3163 */
3164void __sched schedule_preempt_disabled(void)
3165{
Thomas Gleixnerba74c142011-03-21 13:32:17 +01003166 sched_preempt_enable_no_resched();
Thomas Gleixnerc5491ea2011-03-21 12:09:35 +01003167 schedule();
3168 preempt_disable();
3169}
3170
Frederic Weisbecker06b1f802015-02-16 19:20:07 +01003171static void __sched notrace preempt_schedule_common(void)
Frederic Weisbeckera18b5d02015-01-22 18:08:04 +01003172{
3173 do {
Frederic Weisbeckerb30f0e32015-05-12 16:41:49 +02003174 preempt_active_enter();
Frederic Weisbeckera18b5d02015-01-22 18:08:04 +01003175 __schedule();
Frederic Weisbeckerb30f0e32015-05-12 16:41:49 +02003176 preempt_active_exit();
Frederic Weisbeckera18b5d02015-01-22 18:08:04 +01003177
3178 /*
3179 * Check again in case we missed a preemption opportunity
3180 * between schedule and now.
3181 */
Frederic Weisbeckera18b5d02015-01-22 18:08:04 +01003182 } while (need_resched());
3183}
3184
Linus Torvalds1da177e2005-04-16 15:20:36 -07003185#ifdef CONFIG_PREEMPT
3186/*
Andreas Mohr2ed6e342006-07-10 04:43:52 -07003187 * this is the entry point to schedule() from in-kernel preemption
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01003188 * off of preempt_enable. Kernel preemptions off return from interrupt
Linus Torvalds1da177e2005-04-16 15:20:36 -07003189 * occur there and call schedule directly.
3190 */
Andi Kleen722a9f92014-05-02 00:44:38 +02003191asmlinkage __visible void __sched notrace preempt_schedule(void)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003192{
Linus Torvalds1da177e2005-04-16 15:20:36 -07003193 /*
3194 * If there is a non-zero preempt_count or interrupts are disabled,
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01003195 * we do not want to preempt the current task. Just return..
Linus Torvalds1da177e2005-04-16 15:20:36 -07003196 */
Frederic Weisbeckerfbb00b52013-06-19 23:56:22 +02003197 if (likely(!preemptible()))
Linus Torvalds1da177e2005-04-16 15:20:36 -07003198 return;
3199
Frederic Weisbeckera18b5d02015-01-22 18:08:04 +01003200 preempt_schedule_common();
Linus Torvalds1da177e2005-04-16 15:20:36 -07003201}
Masami Hiramatsu376e2422014-04-17 17:17:05 +09003202NOKPROBE_SYMBOL(preempt_schedule);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003203EXPORT_SYMBOL(preempt_schedule);
Oleg Nesterov009f60e2014-10-05 22:23:22 +02003204
Oleg Nesterov009f60e2014-10-05 22:23:22 +02003205/**
Frederic Weisbecker4eaca0a2015-06-04 17:39:08 +02003206 * preempt_schedule_notrace - preempt_schedule called by tracing
Oleg Nesterov009f60e2014-10-05 22:23:22 +02003207 *
3208 * The tracing infrastructure uses preempt_enable_notrace to prevent
3209 * recursion and tracing preempt enabling caused by the tracing
3210 * infrastructure itself. But as tracing can happen in areas coming
3211 * from userspace or just about to enter userspace, a preempt enable
3212 * can occur before user_exit() is called. This will cause the scheduler
3213 * to be called when the system is still in usermode.
3214 *
3215 * To prevent this, the preempt_enable_notrace will use this function
3216 * instead of preempt_schedule() to exit user context if needed before
3217 * calling the scheduler.
3218 */
Frederic Weisbecker4eaca0a2015-06-04 17:39:08 +02003219asmlinkage __visible void __sched notrace preempt_schedule_notrace(void)
Oleg Nesterov009f60e2014-10-05 22:23:22 +02003220{
3221 enum ctx_state prev_ctx;
3222
3223 if (likely(!preemptible()))
3224 return;
3225
3226 do {
Frederic Weisbeckerbe690032015-06-04 17:39:07 +02003227 /*
3228 * Use raw __prempt_count() ops that don't call function.
3229 * We can't call functions before disabling preemption which
3230 * disarm preemption tracing recursions.
3231 */
3232 __preempt_count_add(PREEMPT_ACTIVE + PREEMPT_DISABLE_OFFSET);
3233 barrier();
Oleg Nesterov009f60e2014-10-05 22:23:22 +02003234 /*
3235 * Needs preempt disabled in case user_exit() is traced
3236 * and the tracer calls preempt_enable_notrace() causing
3237 * an infinite recursion.
3238 */
3239 prev_ctx = exception_enter();
3240 __schedule();
3241 exception_exit(prev_ctx);
3242
Oleg Nesterov009f60e2014-10-05 22:23:22 +02003243 barrier();
Frederic Weisbeckerbe690032015-06-04 17:39:07 +02003244 __preempt_count_sub(PREEMPT_ACTIVE + PREEMPT_DISABLE_OFFSET);
Oleg Nesterov009f60e2014-10-05 22:23:22 +02003245 } while (need_resched());
3246}
Frederic Weisbecker4eaca0a2015-06-04 17:39:08 +02003247EXPORT_SYMBOL_GPL(preempt_schedule_notrace);
Oleg Nesterov009f60e2014-10-05 22:23:22 +02003248
Thomas Gleixner32e475d2013-11-21 12:41:44 +01003249#endif /* CONFIG_PREEMPT */
Linus Torvalds1da177e2005-04-16 15:20:36 -07003250
3251/*
Andreas Mohr2ed6e342006-07-10 04:43:52 -07003252 * this is the entry point to schedule() from kernel preemption
Linus Torvalds1da177e2005-04-16 15:20:36 -07003253 * off of irq context.
3254 * Note, that this is called and return with irqs disabled. This will
3255 * protect us against recursive calling from irq.
3256 */
Andi Kleen722a9f92014-05-02 00:44:38 +02003257asmlinkage __visible void __sched preempt_schedule_irq(void)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003258{
Frederic Weisbeckerb22366c2013-02-24 12:59:30 +01003259 enum ctx_state prev_state;
Ingo Molnar6478d882008-01-25 21:08:33 +01003260
Andreas Mohr2ed6e342006-07-10 04:43:52 -07003261 /* Catch callers which need to be fixed */
Peter Zijlstraf27dde82013-08-14 14:55:31 +02003262 BUG_ON(preempt_count() || !irqs_disabled());
Linus Torvalds1da177e2005-04-16 15:20:36 -07003263
Frederic Weisbeckerb22366c2013-02-24 12:59:30 +01003264 prev_state = exception_enter();
3265
Andi Kleen3a5c3592007-10-15 17:00:14 +02003266 do {
Frederic Weisbeckerb30f0e32015-05-12 16:41:49 +02003267 preempt_active_enter();
Andi Kleen3a5c3592007-10-15 17:00:14 +02003268 local_irq_enable();
Thomas Gleixnerc259e012011-06-22 19:47:00 +02003269 __schedule();
Andi Kleen3a5c3592007-10-15 17:00:14 +02003270 local_irq_disable();
Frederic Weisbeckerb30f0e32015-05-12 16:41:49 +02003271 preempt_active_exit();
Lai Jiangshan5ed0cec2009-03-06 19:40:20 +08003272 } while (need_resched());
Frederic Weisbeckerb22366c2013-02-24 12:59:30 +01003273
3274 exception_exit(prev_state);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003275}
3276
Peter Zijlstra63859d42009-09-15 19:14:42 +02003277int default_wake_function(wait_queue_t *curr, unsigned mode, int wake_flags,
Ingo Molnar95cdf3b2005-09-10 00:26:11 -07003278 void *key)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003279{
Peter Zijlstra63859d42009-09-15 19:14:42 +02003280 return try_to_wake_up(curr->private, mode, wake_flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003281}
Linus Torvalds1da177e2005-04-16 15:20:36 -07003282EXPORT_SYMBOL(default_wake_function);
3283
Ingo Molnarb29739f2006-06-27 02:54:51 -07003284#ifdef CONFIG_RT_MUTEXES
3285
3286/*
3287 * rt_mutex_setprio - set the current priority of a task
3288 * @p: task
3289 * @prio: prio value (kernel-internal form)
3290 *
3291 * This function changes the 'effective' priority of a task. It does
3292 * not touch ->normal_prio like __setscheduler().
3293 *
Thomas Gleixnerc365c292014-02-07 20:58:42 +01003294 * Used by the rt_mutex code to implement priority inheritance
3295 * logic. Call site only calls if the priority of the task changed.
Ingo Molnarb29739f2006-06-27 02:54:51 -07003296 */
Ingo Molnar36c8b582006-07-03 00:25:41 -07003297void rt_mutex_setprio(struct task_struct *p, int prio)
Ingo Molnarb29739f2006-06-27 02:54:51 -07003298{
Kirill Tkhaida0c1e62014-08-20 13:47:32 +04003299 int oldprio, queued, running, enqueue_flag = 0;
Ingo Molnar70b97a72006-07-03 00:25:42 -07003300 struct rq *rq;
Thomas Gleixner83ab0aa2010-02-17 09:05:48 +01003301 const struct sched_class *prev_class;
Ingo Molnarb29739f2006-06-27 02:54:51 -07003302
Dario Faggioliaab03e02013-11-28 11:14:43 +01003303 BUG_ON(prio > MAX_PRIO);
Ingo Molnarb29739f2006-06-27 02:54:51 -07003304
Peter Zijlstra0122ec52011-04-05 17:23:51 +02003305 rq = __task_rq_lock(p);
Ingo Molnarb29739f2006-06-27 02:54:51 -07003306
Thomas Gleixner1c4dd992011-06-06 20:07:38 +02003307 /*
3308 * Idle task boosting is a nono in general. There is one
3309 * exception, when PREEMPT_RT and NOHZ is active:
3310 *
3311 * The idle task calls get_next_timer_interrupt() and holds
3312 * the timer wheel base->lock on the CPU and another CPU wants
3313 * to access the timer (probably to cancel it). We can safely
3314 * ignore the boosting request, as the idle CPU runs this code
3315 * with interrupts disabled and will complete the lock
3316 * protected section without being interrupted. So there is no
3317 * real need to boost.
3318 */
3319 if (unlikely(p == rq->idle)) {
3320 WARN_ON(p != rq->curr);
3321 WARN_ON(p->pi_blocked_on);
3322 goto out_unlock;
3323 }
3324
Steven Rostedta8027072010-09-20 15:13:34 -04003325 trace_sched_pi_setprio(p, prio);
Andrew Mortond5f9f942007-05-08 20:27:06 -07003326 oldprio = p->prio;
Thomas Gleixner83ab0aa2010-02-17 09:05:48 +01003327 prev_class = p->sched_class;
Kirill Tkhaida0c1e62014-08-20 13:47:32 +04003328 queued = task_on_rq_queued(p);
Dmitry Adamushko051a1d12007-12-18 15:21:13 +01003329 running = task_current(rq, p);
Kirill Tkhaida0c1e62014-08-20 13:47:32 +04003330 if (queued)
Ingo Molnar69be72c2007-08-09 11:16:49 +02003331 dequeue_task(rq, p, 0);
Hiroshi Shimamoto0e1f3482008-03-10 11:01:20 -07003332 if (running)
Kirill Tkhaif3cd1c42014-09-12 17:41:40 +04003333 put_prev_task(rq, p);
Ingo Molnardd41f592007-07-09 18:51:59 +02003334
Dario Faggioli2d3d8912013-11-07 14:43:44 +01003335 /*
3336 * Boosting condition are:
3337 * 1. -rt task is running and holds mutex A
3338 * --> -dl task blocks on mutex A
3339 *
3340 * 2. -dl task is running and holds mutex A
3341 * --> -dl task blocks on mutex A and could preempt the
3342 * running task
3343 */
3344 if (dl_prio(prio)) {
Oleg Nesterov466af292014-06-06 18:52:06 +02003345 struct task_struct *pi_task = rt_mutex_get_top_task(p);
3346 if (!dl_prio(p->normal_prio) ||
3347 (pi_task && dl_entity_preempt(&pi_task->dl, &p->dl))) {
Dario Faggioli2d3d8912013-11-07 14:43:44 +01003348 p->dl.dl_boosted = 1;
Dario Faggioli2d3d8912013-11-07 14:43:44 +01003349 enqueue_flag = ENQUEUE_REPLENISH;
3350 } else
3351 p->dl.dl_boosted = 0;
Dario Faggioliaab03e02013-11-28 11:14:43 +01003352 p->sched_class = &dl_sched_class;
Dario Faggioli2d3d8912013-11-07 14:43:44 +01003353 } else if (rt_prio(prio)) {
3354 if (dl_prio(oldprio))
3355 p->dl.dl_boosted = 0;
3356 if (oldprio < prio)
3357 enqueue_flag = ENQUEUE_HEAD;
Ingo Molnardd41f592007-07-09 18:51:59 +02003358 p->sched_class = &rt_sched_class;
Dario Faggioli2d3d8912013-11-07 14:43:44 +01003359 } else {
3360 if (dl_prio(oldprio))
3361 p->dl.dl_boosted = 0;
Brian Silverman746db942015-02-18 16:23:56 -08003362 if (rt_prio(oldprio))
3363 p->rt.timeout = 0;
Ingo Molnardd41f592007-07-09 18:51:59 +02003364 p->sched_class = &fair_sched_class;
Dario Faggioli2d3d8912013-11-07 14:43:44 +01003365 }
Ingo Molnardd41f592007-07-09 18:51:59 +02003366
Ingo Molnarb29739f2006-06-27 02:54:51 -07003367 p->prio = prio;
3368
Hiroshi Shimamoto0e1f3482008-03-10 11:01:20 -07003369 if (running)
3370 p->sched_class->set_curr_task(rq);
Kirill Tkhaida0c1e62014-08-20 13:47:32 +04003371 if (queued)
Dario Faggioli2d3d8912013-11-07 14:43:44 +01003372 enqueue_task(rq, p, enqueue_flag);
Steven Rostedtcb469842008-01-25 21:08:22 +01003373
Peter Zijlstrada7a7352011-01-17 17:03:27 +01003374 check_class_changed(rq, p, prev_class, oldprio);
Thomas Gleixner1c4dd992011-06-06 20:07:38 +02003375out_unlock:
Peter Zijlstra4c9a4bc2015-06-11 14:46:39 +02003376 preempt_disable(); /* avoid rq from going away on us */
Peter Zijlstra0122ec52011-04-05 17:23:51 +02003377 __task_rq_unlock(rq);
Peter Zijlstra4c9a4bc2015-06-11 14:46:39 +02003378
3379 balance_callback(rq);
3380 preempt_enable();
Ingo Molnarb29739f2006-06-27 02:54:51 -07003381}
Ingo Molnarb29739f2006-06-27 02:54:51 -07003382#endif
Dario Faggiolid50dde52013-11-07 14:43:36 +01003383
Ingo Molnar36c8b582006-07-03 00:25:41 -07003384void set_user_nice(struct task_struct *p, long nice)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003385{
Kirill Tkhaida0c1e62014-08-20 13:47:32 +04003386 int old_prio, delta, queued;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003387 unsigned long flags;
Ingo Molnar70b97a72006-07-03 00:25:42 -07003388 struct rq *rq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003389
Dongsheng Yang75e45d52014-02-11 15:34:50 +08003390 if (task_nice(p) == nice || nice < MIN_NICE || nice > MAX_NICE)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003391 return;
3392 /*
3393 * We have to be careful, if called from sys_setpriority(),
3394 * the task might be in the middle of scheduling on another CPU.
3395 */
3396 rq = task_rq_lock(p, &flags);
3397 /*
3398 * The RT priorities are set via sched_setscheduler(), but we still
3399 * allow the 'normal' nice value to be set - but as expected
3400 * it wont have any effect on scheduling until the task is
Dario Faggioliaab03e02013-11-28 11:14:43 +01003401 * SCHED_DEADLINE, SCHED_FIFO or SCHED_RR:
Linus Torvalds1da177e2005-04-16 15:20:36 -07003402 */
Dario Faggioliaab03e02013-11-28 11:14:43 +01003403 if (task_has_dl_policy(p) || task_has_rt_policy(p)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003404 p->static_prio = NICE_TO_PRIO(nice);
3405 goto out_unlock;
3406 }
Kirill Tkhaida0c1e62014-08-20 13:47:32 +04003407 queued = task_on_rq_queued(p);
3408 if (queued)
Ingo Molnar69be72c2007-08-09 11:16:49 +02003409 dequeue_task(rq, p, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003410
Linus Torvalds1da177e2005-04-16 15:20:36 -07003411 p->static_prio = NICE_TO_PRIO(nice);
Peter Williams2dd73a42006-06-27 02:54:34 -07003412 set_load_weight(p);
Ingo Molnarb29739f2006-06-27 02:54:51 -07003413 old_prio = p->prio;
3414 p->prio = effective_prio(p);
3415 delta = p->prio - old_prio;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003416
Kirill Tkhaida0c1e62014-08-20 13:47:32 +04003417 if (queued) {
Peter Zijlstra371fd7e2010-03-24 16:38:48 +01003418 enqueue_task(rq, p, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003419 /*
Andrew Mortond5f9f942007-05-08 20:27:06 -07003420 * If the task increased its priority or is running and
3421 * lowered its priority, then reschedule its CPU:
Linus Torvalds1da177e2005-04-16 15:20:36 -07003422 */
Andrew Mortond5f9f942007-05-08 20:27:06 -07003423 if (delta < 0 || (delta > 0 && task_running(rq, p)))
Kirill Tkhai88751252014-06-29 00:03:57 +04003424 resched_curr(rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003425 }
3426out_unlock:
Peter Zijlstra0122ec52011-04-05 17:23:51 +02003427 task_rq_unlock(rq, p, &flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003428}
Linus Torvalds1da177e2005-04-16 15:20:36 -07003429EXPORT_SYMBOL(set_user_nice);
3430
Matt Mackalle43379f2005-05-01 08:59:00 -07003431/*
3432 * can_nice - check if a task can reduce its nice value
3433 * @p: task
3434 * @nice: nice value
3435 */
Ingo Molnar36c8b582006-07-03 00:25:41 -07003436int can_nice(const struct task_struct *p, const int nice)
Matt Mackalle43379f2005-05-01 08:59:00 -07003437{
Matt Mackall024f4742005-08-18 11:24:19 -07003438 /* convert nice value [19,-20] to rlimit style value [1,40] */
Dongsheng Yang7aa2c012014-05-08 18:33:49 +09003439 int nice_rlim = nice_to_rlimit(nice);
Ingo Molnar48f24c42006-07-03 00:25:40 -07003440
Jiri Slaby78d7d402010-03-05 13:42:54 -08003441 return (nice_rlim <= task_rlimit(p, RLIMIT_NICE) ||
Matt Mackalle43379f2005-05-01 08:59:00 -07003442 capable(CAP_SYS_NICE));
3443}
3444
Linus Torvalds1da177e2005-04-16 15:20:36 -07003445#ifdef __ARCH_WANT_SYS_NICE
3446
3447/*
3448 * sys_nice - change the priority of the current process.
3449 * @increment: priority increment
3450 *
3451 * sys_setpriority is a more generic, but much slower function that
3452 * does similar things.
3453 */
Heiko Carstens5add95d2009-01-14 14:14:08 +01003454SYSCALL_DEFINE1(nice, int, increment)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003455{
Ingo Molnar48f24c42006-07-03 00:25:40 -07003456 long nice, retval;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003457
3458 /*
3459 * Setpriority might change our priority at the same moment.
3460 * We don't have to worry. Conceptually one call occurs first
3461 * and we have a single winner.
3462 */
Dongsheng Yanga9467fa2014-05-08 18:35:15 +09003463 increment = clamp(increment, -NICE_WIDTH, NICE_WIDTH);
Dongsheng Yangd0ea0262014-01-27 22:00:45 -05003464 nice = task_nice(current) + increment;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003465
Dongsheng Yanga9467fa2014-05-08 18:35:15 +09003466 nice = clamp_val(nice, MIN_NICE, MAX_NICE);
Matt Mackalle43379f2005-05-01 08:59:00 -07003467 if (increment < 0 && !can_nice(current, nice))
3468 return -EPERM;
3469
Linus Torvalds1da177e2005-04-16 15:20:36 -07003470 retval = security_task_setnice(current, nice);
3471 if (retval)
3472 return retval;
3473
3474 set_user_nice(current, nice);
3475 return 0;
3476}
3477
3478#endif
3479
3480/**
3481 * task_prio - return the priority value of a given task.
3482 * @p: the task in question.
3483 *
Yacine Belkadie69f6182013-07-12 20:45:47 +02003484 * Return: The priority value as seen by users in /proc.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003485 * RT tasks are offset by -200. Normal tasks are centered
3486 * around 0, value goes from -16 to +15.
3487 */
Ingo Molnar36c8b582006-07-03 00:25:41 -07003488int task_prio(const struct task_struct *p)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003489{
3490 return p->prio - MAX_RT_PRIO;
3491}
3492
3493/**
Linus Torvalds1da177e2005-04-16 15:20:36 -07003494 * idle_cpu - is a given cpu idle currently?
3495 * @cpu: the processor in question.
Yacine Belkadie69f6182013-07-12 20:45:47 +02003496 *
3497 * Return: 1 if the CPU is currently idle. 0 otherwise.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003498 */
3499int idle_cpu(int cpu)
3500{
Thomas Gleixner908a3282011-09-15 15:32:06 +02003501 struct rq *rq = cpu_rq(cpu);
3502
3503 if (rq->curr != rq->idle)
3504 return 0;
3505
3506 if (rq->nr_running)
3507 return 0;
3508
3509#ifdef CONFIG_SMP
3510 if (!llist_empty(&rq->wake_list))
3511 return 0;
3512#endif
3513
3514 return 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003515}
3516
Linus Torvalds1da177e2005-04-16 15:20:36 -07003517/**
3518 * idle_task - return the idle task for a given cpu.
3519 * @cpu: the processor in question.
Yacine Belkadie69f6182013-07-12 20:45:47 +02003520 *
3521 * Return: The idle task for the cpu @cpu.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003522 */
Ingo Molnar36c8b582006-07-03 00:25:41 -07003523struct task_struct *idle_task(int cpu)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003524{
3525 return cpu_rq(cpu)->idle;
3526}
3527
3528/**
3529 * find_process_by_pid - find a process with a matching PID value.
3530 * @pid: the pid in question.
Yacine Belkadie69f6182013-07-12 20:45:47 +02003531 *
3532 * The task of @pid, if found. %NULL otherwise.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003533 */
Alexey Dobriyana9957442007-10-15 17:00:13 +02003534static struct task_struct *find_process_by_pid(pid_t pid)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003535{
Pavel Emelyanov228ebcb2007-10-18 23:40:16 -07003536 return pid ? find_task_by_vpid(pid) : current;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003537}
3538
Dario Faggioliaab03e02013-11-28 11:14:43 +01003539/*
3540 * This function initializes the sched_dl_entity of a newly becoming
3541 * SCHED_DEADLINE task.
3542 *
3543 * Only the static values are considered here, the actual runtime and the
3544 * absolute deadline will be properly calculated when the task is enqueued
3545 * for the first time with its new policy.
3546 */
3547static void
3548__setparam_dl(struct task_struct *p, const struct sched_attr *attr)
3549{
3550 struct sched_dl_entity *dl_se = &p->dl;
3551
Dario Faggioliaab03e02013-11-28 11:14:43 +01003552 dl_se->dl_runtime = attr->sched_runtime;
3553 dl_se->dl_deadline = attr->sched_deadline;
Harald Gustafsson755378a2013-11-07 14:43:40 +01003554 dl_se->dl_period = attr->sched_period ?: dl_se->dl_deadline;
Dario Faggioliaab03e02013-11-28 11:14:43 +01003555 dl_se->flags = attr->sched_flags;
Dario Faggioli332ac172013-11-07 14:43:45 +01003556 dl_se->dl_bw = to_ratio(dl_se->dl_period, dl_se->dl_runtime);
Peter Zijlstra40767b02015-01-28 15:08:03 +01003557
3558 /*
3559 * Changing the parameters of a task is 'tricky' and we're not doing
3560 * the correct thing -- also see task_dead_dl() and switched_from_dl().
3561 *
3562 * What we SHOULD do is delay the bandwidth release until the 0-lag
3563 * point. This would include retaining the task_struct until that time
3564 * and change dl_overflow() to not immediately decrement the current
3565 * amount.
3566 *
3567 * Instead we retain the current runtime/deadline and let the new
3568 * parameters take effect after the current reservation period lapses.
3569 * This is safe (albeit pessimistic) because the 0-lag point is always
3570 * before the current scheduling deadline.
3571 *
3572 * We can still have temporary overloads because we do not delay the
3573 * change in bandwidth until that time; so admission control is
3574 * not on the safe side. It does however guarantee tasks will never
3575 * consume more than promised.
3576 */
Dario Faggioliaab03e02013-11-28 11:14:43 +01003577}
3578
Steven Rostedtc13db6b2014-07-23 11:28:26 -04003579/*
3580 * sched_setparam() passes in -1 for its policy, to let the functions
3581 * it calls know not to change it.
3582 */
3583#define SETPARAM_POLICY -1
3584
Thomas Gleixnerc365c292014-02-07 20:58:42 +01003585static void __setscheduler_params(struct task_struct *p,
3586 const struct sched_attr *attr)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003587{
Dario Faggiolid50dde52013-11-07 14:43:36 +01003588 int policy = attr->sched_policy;
3589
Steven Rostedtc13db6b2014-07-23 11:28:26 -04003590 if (policy == SETPARAM_POLICY)
Peter Zijlstra39fd8fd2014-01-15 16:33:20 +01003591 policy = p->policy;
3592
Linus Torvalds1da177e2005-04-16 15:20:36 -07003593 p->policy = policy;
Dario Faggiolid50dde52013-11-07 14:43:36 +01003594
Dario Faggioliaab03e02013-11-28 11:14:43 +01003595 if (dl_policy(policy))
3596 __setparam_dl(p, attr);
Peter Zijlstra39fd8fd2014-01-15 16:33:20 +01003597 else if (fair_policy(policy))
Dario Faggiolid50dde52013-11-07 14:43:36 +01003598 p->static_prio = NICE_TO_PRIO(attr->sched_nice);
3599
Peter Zijlstra39fd8fd2014-01-15 16:33:20 +01003600 /*
3601 * __sched_setscheduler() ensures attr->sched_priority == 0 when
3602 * !rt_policy. Always setting this ensures that things like
3603 * getparam()/getattr() don't report silly values for !rt tasks.
3604 */
3605 p->rt_priority = attr->sched_priority;
Steven Rostedt383afd02014-03-11 19:24:20 -04003606 p->normal_prio = normal_prio(p);
Thomas Gleixnerc365c292014-02-07 20:58:42 +01003607 set_load_weight(p);
3608}
Peter Zijlstra39fd8fd2014-01-15 16:33:20 +01003609
Thomas Gleixnerc365c292014-02-07 20:58:42 +01003610/* Actually do priority change: must hold pi & rq lock. */
3611static void __setscheduler(struct rq *rq, struct task_struct *p,
Thomas Gleixner0782e632015-05-05 19:49:49 +02003612 const struct sched_attr *attr, bool keep_boost)
Thomas Gleixnerc365c292014-02-07 20:58:42 +01003613{
3614 __setscheduler_params(p, attr);
Dario Faggiolid50dde52013-11-07 14:43:36 +01003615
Steven Rostedt383afd02014-03-11 19:24:20 -04003616 /*
Thomas Gleixner0782e632015-05-05 19:49:49 +02003617 * Keep a potential priority boosting if called from
3618 * sched_setscheduler().
Steven Rostedt383afd02014-03-11 19:24:20 -04003619 */
Thomas Gleixner0782e632015-05-05 19:49:49 +02003620 if (keep_boost)
3621 p->prio = rt_mutex_get_effective_prio(p, normal_prio(p));
3622 else
3623 p->prio = normal_prio(p);
Steven Rostedt383afd02014-03-11 19:24:20 -04003624
Dario Faggioliaab03e02013-11-28 11:14:43 +01003625 if (dl_prio(p->prio))
3626 p->sched_class = &dl_sched_class;
3627 else if (rt_prio(p->prio))
Peter Zijlstraffd44db2009-11-10 20:12:01 +01003628 p->sched_class = &rt_sched_class;
3629 else
3630 p->sched_class = &fair_sched_class;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003631}
Dario Faggioliaab03e02013-11-28 11:14:43 +01003632
3633static void
3634__getparam_dl(struct task_struct *p, struct sched_attr *attr)
3635{
3636 struct sched_dl_entity *dl_se = &p->dl;
3637
3638 attr->sched_priority = p->rt_priority;
3639 attr->sched_runtime = dl_se->dl_runtime;
3640 attr->sched_deadline = dl_se->dl_deadline;
Harald Gustafsson755378a2013-11-07 14:43:40 +01003641 attr->sched_period = dl_se->dl_period;
Dario Faggioliaab03e02013-11-28 11:14:43 +01003642 attr->sched_flags = dl_se->flags;
3643}
3644
3645/*
3646 * This function validates the new parameters of a -deadline task.
3647 * We ask for the deadline not being zero, and greater or equal
Harald Gustafsson755378a2013-11-07 14:43:40 +01003648 * than the runtime, as well as the period of being zero or
Dario Faggioli332ac172013-11-07 14:43:45 +01003649 * greater than deadline. Furthermore, we have to be sure that
Juri Lellib08278192014-05-13 14:11:31 +02003650 * user parameters are above the internal resolution of 1us (we
3651 * check sched_runtime only since it is always the smaller one) and
3652 * below 2^63 ns (we have to check both sched_deadline and
3653 * sched_period, as the latter can be zero).
Dario Faggioliaab03e02013-11-28 11:14:43 +01003654 */
3655static bool
3656__checkparam_dl(const struct sched_attr *attr)
3657{
Juri Lellib08278192014-05-13 14:11:31 +02003658 /* deadline != 0 */
3659 if (attr->sched_deadline == 0)
3660 return false;
3661
3662 /*
3663 * Since we truncate DL_SCALE bits, make sure we're at least
3664 * that big.
3665 */
3666 if (attr->sched_runtime < (1ULL << DL_SCALE))
3667 return false;
3668
3669 /*
3670 * Since we use the MSB for wrap-around and sign issues, make
3671 * sure it's not set (mind that period can be equal to zero).
3672 */
3673 if (attr->sched_deadline & (1ULL << 63) ||
3674 attr->sched_period & (1ULL << 63))
3675 return false;
3676
3677 /* runtime <= deadline <= period (if period != 0) */
3678 if ((attr->sched_period != 0 &&
3679 attr->sched_period < attr->sched_deadline) ||
3680 attr->sched_deadline < attr->sched_runtime)
3681 return false;
3682
3683 return true;
Dario Faggioliaab03e02013-11-28 11:14:43 +01003684}
3685
David Howellsc69e8d92008-11-14 10:39:19 +11003686/*
3687 * check the target process has a UID that matches the current process's
3688 */
3689static bool check_same_owner(struct task_struct *p)
3690{
3691 const struct cred *cred = current_cred(), *pcred;
3692 bool match;
3693
3694 rcu_read_lock();
3695 pcred = __task_cred(p);
Eric W. Biederman9c806aa2012-02-02 18:54:02 -08003696 match = (uid_eq(cred->euid, pcred->euid) ||
3697 uid_eq(cred->euid, pcred->uid));
David Howellsc69e8d92008-11-14 10:39:19 +11003698 rcu_read_unlock();
3699 return match;
3700}
3701
Wanpeng Li75381602014-11-26 08:44:04 +08003702static bool dl_param_changed(struct task_struct *p,
3703 const struct sched_attr *attr)
3704{
3705 struct sched_dl_entity *dl_se = &p->dl;
3706
3707 if (dl_se->dl_runtime != attr->sched_runtime ||
3708 dl_se->dl_deadline != attr->sched_deadline ||
3709 dl_se->dl_period != attr->sched_period ||
3710 dl_se->flags != attr->sched_flags)
3711 return true;
3712
3713 return false;
3714}
3715
Dario Faggiolid50dde52013-11-07 14:43:36 +01003716static int __sched_setscheduler(struct task_struct *p,
3717 const struct sched_attr *attr,
Peter Zijlstradbc7f062015-06-11 14:46:38 +02003718 bool user, bool pi)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003719{
Steven Rostedt383afd02014-03-11 19:24:20 -04003720 int newprio = dl_policy(attr->sched_policy) ? MAX_DL_PRIO - 1 :
3721 MAX_RT_PRIO - 1 - attr->sched_priority;
Kirill Tkhaida0c1e62014-08-20 13:47:32 +04003722 int retval, oldprio, oldpolicy = -1, queued, running;
Thomas Gleixner0782e632015-05-05 19:49:49 +02003723 int new_effective_prio, policy = attr->sched_policy;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003724 unsigned long flags;
Thomas Gleixner83ab0aa2010-02-17 09:05:48 +01003725 const struct sched_class *prev_class;
Ingo Molnar70b97a72006-07-03 00:25:42 -07003726 struct rq *rq;
Lennart Poetteringca94c442009-06-15 17:17:47 +02003727 int reset_on_fork;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003728
Steven Rostedt66e53932006-06-27 02:54:44 -07003729 /* may grab non-irq protected spin_locks */
3730 BUG_ON(in_interrupt());
Linus Torvalds1da177e2005-04-16 15:20:36 -07003731recheck:
3732 /* double check policy once rq lock held */
Lennart Poetteringca94c442009-06-15 17:17:47 +02003733 if (policy < 0) {
3734 reset_on_fork = p->sched_reset_on_fork;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003735 policy = oldpolicy = p->policy;
Lennart Poetteringca94c442009-06-15 17:17:47 +02003736 } else {
Peter Zijlstra7479f3c2014-01-15 17:05:04 +01003737 reset_on_fork = !!(attr->sched_flags & SCHED_FLAG_RESET_ON_FORK);
Lennart Poetteringca94c442009-06-15 17:17:47 +02003738
Dario Faggioliaab03e02013-11-28 11:14:43 +01003739 if (policy != SCHED_DEADLINE &&
3740 policy != SCHED_FIFO && policy != SCHED_RR &&
Lennart Poetteringca94c442009-06-15 17:17:47 +02003741 policy != SCHED_NORMAL && policy != SCHED_BATCH &&
3742 policy != SCHED_IDLE)
3743 return -EINVAL;
3744 }
3745
Peter Zijlstra7479f3c2014-01-15 17:05:04 +01003746 if (attr->sched_flags & ~(SCHED_FLAG_RESET_ON_FORK))
3747 return -EINVAL;
3748
Linus Torvalds1da177e2005-04-16 15:20:36 -07003749 /*
3750 * Valid priorities for SCHED_FIFO and SCHED_RR are
Ingo Molnardd41f592007-07-09 18:51:59 +02003751 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
3752 * SCHED_BATCH and SCHED_IDLE is 0.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003753 */
Peter Zijlstra0bb040a2014-01-15 17:15:13 +01003754 if ((p->mm && attr->sched_priority > MAX_USER_RT_PRIO-1) ||
Dario Faggiolid50dde52013-11-07 14:43:36 +01003755 (!p->mm && attr->sched_priority > MAX_RT_PRIO-1))
Linus Torvalds1da177e2005-04-16 15:20:36 -07003756 return -EINVAL;
Dario Faggioliaab03e02013-11-28 11:14:43 +01003757 if ((dl_policy(policy) && !__checkparam_dl(attr)) ||
3758 (rt_policy(policy) != (attr->sched_priority != 0)))
Linus Torvalds1da177e2005-04-16 15:20:36 -07003759 return -EINVAL;
3760
Olivier Croquette37e4ab32005-06-25 14:57:32 -07003761 /*
3762 * Allow unprivileged RT tasks to decrease priority:
3763 */
Rusty Russell961ccdd2008-06-23 13:55:38 +10003764 if (user && !capable(CAP_SYS_NICE)) {
Dario Faggiolid50dde52013-11-07 14:43:36 +01003765 if (fair_policy(policy)) {
Dongsheng Yangd0ea0262014-01-27 22:00:45 -05003766 if (attr->sched_nice < task_nice(p) &&
Peter Zijlstraeaad4512014-01-16 17:54:25 +01003767 !can_nice(p, attr->sched_nice))
Dario Faggiolid50dde52013-11-07 14:43:36 +01003768 return -EPERM;
3769 }
3770
Ingo Molnare05606d2007-07-09 18:51:59 +02003771 if (rt_policy(policy)) {
Oleg Nesterova44702e2010-06-11 01:09:44 +02003772 unsigned long rlim_rtprio =
3773 task_rlimit(p, RLIMIT_RTPRIO);
Oleg Nesterov5fe1d752006-09-29 02:00:48 -07003774
Oleg Nesterov8dc3e902006-09-29 02:00:50 -07003775 /* can't set/change the rt policy */
3776 if (policy != p->policy && !rlim_rtprio)
3777 return -EPERM;
3778
3779 /* can't increase priority */
Dario Faggiolid50dde52013-11-07 14:43:36 +01003780 if (attr->sched_priority > p->rt_priority &&
3781 attr->sched_priority > rlim_rtprio)
Oleg Nesterov8dc3e902006-09-29 02:00:50 -07003782 return -EPERM;
3783 }
Darren Hartc02aa732011-02-17 15:37:07 -08003784
Juri Lellid44753b2014-03-03 12:09:21 +01003785 /*
3786 * Can't set/change SCHED_DEADLINE policy at all for now
3787 * (safest behavior); in the future we would like to allow
3788 * unprivileged DL tasks to increase their relative deadline
3789 * or reduce their runtime (both ways reducing utilization)
3790 */
3791 if (dl_policy(policy))
3792 return -EPERM;
3793
Ingo Molnardd41f592007-07-09 18:51:59 +02003794 /*
Darren Hartc02aa732011-02-17 15:37:07 -08003795 * Treat SCHED_IDLE as nice 20. Only allow a switch to
3796 * SCHED_NORMAL if the RLIMIT_NICE would normally permit it.
Ingo Molnardd41f592007-07-09 18:51:59 +02003797 */
Darren Hartc02aa732011-02-17 15:37:07 -08003798 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE) {
Dongsheng Yangd0ea0262014-01-27 22:00:45 -05003799 if (!can_nice(p, task_nice(p)))
Darren Hartc02aa732011-02-17 15:37:07 -08003800 return -EPERM;
3801 }
Oleg Nesterov8dc3e902006-09-29 02:00:50 -07003802
Olivier Croquette37e4ab32005-06-25 14:57:32 -07003803 /* can't change other user's priorities */
David Howellsc69e8d92008-11-14 10:39:19 +11003804 if (!check_same_owner(p))
Olivier Croquette37e4ab32005-06-25 14:57:32 -07003805 return -EPERM;
Lennart Poetteringca94c442009-06-15 17:17:47 +02003806
3807 /* Normal users shall not reset the sched_reset_on_fork flag */
3808 if (p->sched_reset_on_fork && !reset_on_fork)
3809 return -EPERM;
Olivier Croquette37e4ab32005-06-25 14:57:32 -07003810 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07003811
Jeremy Fitzhardinge725aad22008-08-03 09:33:03 -07003812 if (user) {
KOSAKI Motohirob0ae1982010-10-15 04:21:18 +09003813 retval = security_task_setscheduler(p);
Jeremy Fitzhardinge725aad22008-08-03 09:33:03 -07003814 if (retval)
3815 return retval;
3816 }
3817
Linus Torvalds1da177e2005-04-16 15:20:36 -07003818 /*
Ingo Molnarb29739f2006-06-27 02:54:51 -07003819 * make sure no PI-waiters arrive (or leave) while we are
3820 * changing the priority of the task:
Peter Zijlstra0122ec52011-04-05 17:23:51 +02003821 *
Lucas De Marchi25985ed2011-03-30 22:57:33 -03003822 * To be able to change p->policy safely, the appropriate
Linus Torvalds1da177e2005-04-16 15:20:36 -07003823 * runqueue lock must be held.
3824 */
Peter Zijlstra0122ec52011-04-05 17:23:51 +02003825 rq = task_rq_lock(p, &flags);
Peter Zijlstradc61b1d2010-06-08 11:40:42 +02003826
Peter Zijlstra34f971f2010-09-22 13:53:15 +02003827 /*
3828 * Changing the policy of the stop threads its a very bad idea
3829 */
3830 if (p == rq->stop) {
Peter Zijlstra0122ec52011-04-05 17:23:51 +02003831 task_rq_unlock(rq, p, &flags);
Peter Zijlstra34f971f2010-09-22 13:53:15 +02003832 return -EINVAL;
3833 }
3834
Dario Faggiolia51e9192011-03-24 14:00:18 +01003835 /*
Thomas Gleixnerd6b1e912014-02-07 20:58:40 +01003836 * If not changing anything there's no need to proceed further,
3837 * but store a possible modification of reset_on_fork.
Dario Faggiolia51e9192011-03-24 14:00:18 +01003838 */
Dario Faggiolid50dde52013-11-07 14:43:36 +01003839 if (unlikely(policy == p->policy)) {
Dongsheng Yangd0ea0262014-01-27 22:00:45 -05003840 if (fair_policy(policy) && attr->sched_nice != task_nice(p))
Dario Faggiolid50dde52013-11-07 14:43:36 +01003841 goto change;
3842 if (rt_policy(policy) && attr->sched_priority != p->rt_priority)
3843 goto change;
Wanpeng Li75381602014-11-26 08:44:04 +08003844 if (dl_policy(policy) && dl_param_changed(p, attr))
Dario Faggioliaab03e02013-11-28 11:14:43 +01003845 goto change;
Dario Faggiolid50dde52013-11-07 14:43:36 +01003846
Thomas Gleixnerd6b1e912014-02-07 20:58:40 +01003847 p->sched_reset_on_fork = reset_on_fork;
Namhyung Kim45afb172012-07-07 16:49:02 +09003848 task_rq_unlock(rq, p, &flags);
Dario Faggiolia51e9192011-03-24 14:00:18 +01003849 return 0;
3850 }
Dario Faggiolid50dde52013-11-07 14:43:36 +01003851change:
Dario Faggiolia51e9192011-03-24 14:00:18 +01003852
Peter Zijlstradc61b1d2010-06-08 11:40:42 +02003853 if (user) {
Dario Faggioli332ac172013-11-07 14:43:45 +01003854#ifdef CONFIG_RT_GROUP_SCHED
Peter Zijlstradc61b1d2010-06-08 11:40:42 +02003855 /*
3856 * Do not allow realtime tasks into groups that have no runtime
3857 * assigned.
3858 */
3859 if (rt_bandwidth_enabled() && rt_policy(policy) &&
Mike Galbraithf4493772011-01-13 04:54:50 +01003860 task_group(p)->rt_bandwidth.rt_runtime == 0 &&
3861 !task_group_is_autogroup(task_group(p))) {
Peter Zijlstra0122ec52011-04-05 17:23:51 +02003862 task_rq_unlock(rq, p, &flags);
Peter Zijlstradc61b1d2010-06-08 11:40:42 +02003863 return -EPERM;
3864 }
Peter Zijlstradc61b1d2010-06-08 11:40:42 +02003865#endif
Dario Faggioli332ac172013-11-07 14:43:45 +01003866#ifdef CONFIG_SMP
3867 if (dl_bandwidth_enabled() && dl_policy(policy)) {
3868 cpumask_t *span = rq->rd->span;
Dario Faggioli332ac172013-11-07 14:43:45 +01003869
3870 /*
3871 * Don't allow tasks with an affinity mask smaller than
3872 * the entire root_domain to become SCHED_DEADLINE. We
3873 * will also fail if there's no bandwidth available.
3874 */
Peter Zijlstrae4099a52013-12-17 10:03:34 +01003875 if (!cpumask_subset(span, &p->cpus_allowed) ||
3876 rq->rd->dl_bw.bw == 0) {
Dario Faggioli332ac172013-11-07 14:43:45 +01003877 task_rq_unlock(rq, p, &flags);
3878 return -EPERM;
3879 }
3880 }
3881#endif
3882 }
Peter Zijlstradc61b1d2010-06-08 11:40:42 +02003883
Linus Torvalds1da177e2005-04-16 15:20:36 -07003884 /* recheck policy now with rq lock held */
3885 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
3886 policy = oldpolicy = -1;
Peter Zijlstra0122ec52011-04-05 17:23:51 +02003887 task_rq_unlock(rq, p, &flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003888 goto recheck;
3889 }
Dario Faggioli332ac172013-11-07 14:43:45 +01003890
3891 /*
3892 * If setscheduling to SCHED_DEADLINE (or changing the parameters
3893 * of a SCHED_DEADLINE task) we need to check if enough bandwidth
3894 * is available.
3895 */
Peter Zijlstrae4099a52013-12-17 10:03:34 +01003896 if ((dl_policy(policy) || dl_task(p)) && dl_overflow(p, policy, attr)) {
Dario Faggioli332ac172013-11-07 14:43:45 +01003897 task_rq_unlock(rq, p, &flags);
3898 return -EBUSY;
3899 }
3900
Thomas Gleixnerc365c292014-02-07 20:58:42 +01003901 p->sched_reset_on_fork = reset_on_fork;
3902 oldprio = p->prio;
3903
Peter Zijlstradbc7f062015-06-11 14:46:38 +02003904 if (pi) {
3905 /*
3906 * Take priority boosted tasks into account. If the new
3907 * effective priority is unchanged, we just store the new
3908 * normal parameters and do not touch the scheduler class and
3909 * the runqueue. This will be done when the task deboost
3910 * itself.
3911 */
3912 new_effective_prio = rt_mutex_get_effective_prio(p, newprio);
3913 if (new_effective_prio == oldprio) {
3914 __setscheduler_params(p, attr);
3915 task_rq_unlock(rq, p, &flags);
3916 return 0;
3917 }
Thomas Gleixnerc365c292014-02-07 20:58:42 +01003918 }
3919
Kirill Tkhaida0c1e62014-08-20 13:47:32 +04003920 queued = task_on_rq_queued(p);
Dmitry Adamushko051a1d12007-12-18 15:21:13 +01003921 running = task_current(rq, p);
Kirill Tkhaida0c1e62014-08-20 13:47:32 +04003922 if (queued)
Peter Zijlstra4ca9b722012-01-25 11:50:51 +01003923 dequeue_task(rq, p, 0);
Hiroshi Shimamoto0e1f3482008-03-10 11:01:20 -07003924 if (running)
Kirill Tkhaif3cd1c42014-09-12 17:41:40 +04003925 put_prev_task(rq, p);
Dmitry Adamushkof6b53202007-10-15 17:00:08 +02003926
Thomas Gleixner83ab0aa2010-02-17 09:05:48 +01003927 prev_class = p->sched_class;
Peter Zijlstradbc7f062015-06-11 14:46:38 +02003928 __setscheduler(rq, p, attr, pi);
Dmitry Adamushkof6b53202007-10-15 17:00:08 +02003929
Hiroshi Shimamoto0e1f3482008-03-10 11:01:20 -07003930 if (running)
3931 p->sched_class->set_curr_task(rq);
Kirill Tkhaida0c1e62014-08-20 13:47:32 +04003932 if (queued) {
Thomas Gleixner81a44c52014-02-07 20:58:41 +01003933 /*
3934 * We enqueue to tail when the priority of a task is
3935 * increased (user space view).
3936 */
3937 enqueue_task(rq, p, oldprio <= p->prio ? ENQUEUE_HEAD : 0);
3938 }
Steven Rostedtcb469842008-01-25 21:08:22 +01003939
Peter Zijlstrada7a7352011-01-17 17:03:27 +01003940 check_class_changed(rq, p, prev_class, oldprio);
Peter Zijlstra4c9a4bc2015-06-11 14:46:39 +02003941 preempt_disable(); /* avoid rq from going away on us */
Peter Zijlstra0122ec52011-04-05 17:23:51 +02003942 task_rq_unlock(rq, p, &flags);
Ingo Molnarb29739f2006-06-27 02:54:51 -07003943
Peter Zijlstradbc7f062015-06-11 14:46:38 +02003944 if (pi)
3945 rt_mutex_adjust_pi(p);
Thomas Gleixner95e02ca2006-06-27 02:55:02 -07003946
Peter Zijlstra4c9a4bc2015-06-11 14:46:39 +02003947 /*
3948 * Run balance callbacks after we've adjusted the PI chain.
3949 */
3950 balance_callback(rq);
3951 preempt_enable();
Linus Torvalds1da177e2005-04-16 15:20:36 -07003952
3953 return 0;
3954}
Rusty Russell961ccdd2008-06-23 13:55:38 +10003955
Peter Zijlstra7479f3c2014-01-15 17:05:04 +01003956static int _sched_setscheduler(struct task_struct *p, int policy,
3957 const struct sched_param *param, bool check)
3958{
3959 struct sched_attr attr = {
3960 .sched_policy = policy,
3961 .sched_priority = param->sched_priority,
3962 .sched_nice = PRIO_TO_NICE(p->static_prio),
3963 };
3964
Steven Rostedtc13db6b2014-07-23 11:28:26 -04003965 /* Fixup the legacy SCHED_RESET_ON_FORK hack. */
3966 if ((policy != SETPARAM_POLICY) && (policy & SCHED_RESET_ON_FORK)) {
Peter Zijlstra7479f3c2014-01-15 17:05:04 +01003967 attr.sched_flags |= SCHED_FLAG_RESET_ON_FORK;
3968 policy &= ~SCHED_RESET_ON_FORK;
3969 attr.sched_policy = policy;
3970 }
3971
Peter Zijlstradbc7f062015-06-11 14:46:38 +02003972 return __sched_setscheduler(p, &attr, check, true);
Peter Zijlstra7479f3c2014-01-15 17:05:04 +01003973}
Rusty Russell961ccdd2008-06-23 13:55:38 +10003974/**
3975 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
3976 * @p: the task in question.
3977 * @policy: new policy.
3978 * @param: structure containing the new RT priority.
3979 *
Yacine Belkadie69f6182013-07-12 20:45:47 +02003980 * Return: 0 on success. An error code otherwise.
3981 *
Rusty Russell961ccdd2008-06-23 13:55:38 +10003982 * NOTE that the task may be already dead.
3983 */
3984int sched_setscheduler(struct task_struct *p, int policy,
KOSAKI Motohirofe7de492010-10-20 16:01:12 -07003985 const struct sched_param *param)
Rusty Russell961ccdd2008-06-23 13:55:38 +10003986{
Peter Zijlstra7479f3c2014-01-15 17:05:04 +01003987 return _sched_setscheduler(p, policy, param, true);
Rusty Russell961ccdd2008-06-23 13:55:38 +10003988}
Linus Torvalds1da177e2005-04-16 15:20:36 -07003989EXPORT_SYMBOL_GPL(sched_setscheduler);
3990
Dario Faggiolid50dde52013-11-07 14:43:36 +01003991int sched_setattr(struct task_struct *p, const struct sched_attr *attr)
3992{
Peter Zijlstradbc7f062015-06-11 14:46:38 +02003993 return __sched_setscheduler(p, attr, true, true);
Dario Faggiolid50dde52013-11-07 14:43:36 +01003994}
3995EXPORT_SYMBOL_GPL(sched_setattr);
3996
Rusty Russell961ccdd2008-06-23 13:55:38 +10003997/**
3998 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
3999 * @p: the task in question.
4000 * @policy: new policy.
4001 * @param: structure containing the new RT priority.
4002 *
4003 * Just like sched_setscheduler, only don't bother checking if the
4004 * current context has permission. For example, this is needed in
4005 * stop_machine(): we create temporary high priority worker threads,
4006 * but our caller might not have that capability.
Yacine Belkadie69f6182013-07-12 20:45:47 +02004007 *
4008 * Return: 0 on success. An error code otherwise.
Rusty Russell961ccdd2008-06-23 13:55:38 +10004009 */
4010int sched_setscheduler_nocheck(struct task_struct *p, int policy,
KOSAKI Motohirofe7de492010-10-20 16:01:12 -07004011 const struct sched_param *param)
Rusty Russell961ccdd2008-06-23 13:55:38 +10004012{
Peter Zijlstra7479f3c2014-01-15 17:05:04 +01004013 return _sched_setscheduler(p, policy, param, false);
Rusty Russell961ccdd2008-06-23 13:55:38 +10004014}
4015
Ingo Molnar95cdf3b2005-09-10 00:26:11 -07004016static int
4017do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004018{
Linus Torvalds1da177e2005-04-16 15:20:36 -07004019 struct sched_param lparam;
4020 struct task_struct *p;
Ingo Molnar36c8b582006-07-03 00:25:41 -07004021 int retval;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004022
4023 if (!param || pid < 0)
4024 return -EINVAL;
4025 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
4026 return -EFAULT;
Oleg Nesterov5fe1d752006-09-29 02:00:48 -07004027
4028 rcu_read_lock();
4029 retval = -ESRCH;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004030 p = find_process_by_pid(pid);
Oleg Nesterov5fe1d752006-09-29 02:00:48 -07004031 if (p != NULL)
4032 retval = sched_setscheduler(p, policy, &lparam);
4033 rcu_read_unlock();
Ingo Molnar36c8b582006-07-03 00:25:41 -07004034
Linus Torvalds1da177e2005-04-16 15:20:36 -07004035 return retval;
4036}
4037
Dario Faggiolid50dde52013-11-07 14:43:36 +01004038/*
4039 * Mimics kernel/events/core.c perf_copy_attr().
4040 */
4041static int sched_copy_attr(struct sched_attr __user *uattr,
4042 struct sched_attr *attr)
4043{
4044 u32 size;
4045 int ret;
4046
4047 if (!access_ok(VERIFY_WRITE, uattr, SCHED_ATTR_SIZE_VER0))
4048 return -EFAULT;
4049
4050 /*
4051 * zero the full structure, so that a short copy will be nice.
4052 */
4053 memset(attr, 0, sizeof(*attr));
4054
4055 ret = get_user(size, &uattr->size);
4056 if (ret)
4057 return ret;
4058
4059 if (size > PAGE_SIZE) /* silly large */
4060 goto err_size;
4061
4062 if (!size) /* abi compat */
4063 size = SCHED_ATTR_SIZE_VER0;
4064
4065 if (size < SCHED_ATTR_SIZE_VER0)
4066 goto err_size;
4067
4068 /*
4069 * If we're handed a bigger struct than we know of,
4070 * ensure all the unknown bits are 0 - i.e. new
4071 * user-space does not rely on any kernel feature
4072 * extensions we dont know about yet.
4073 */
4074 if (size > sizeof(*attr)) {
4075 unsigned char __user *addr;
4076 unsigned char __user *end;
4077 unsigned char val;
4078
4079 addr = (void __user *)uattr + sizeof(*attr);
4080 end = (void __user *)uattr + size;
4081
4082 for (; addr < end; addr++) {
4083 ret = get_user(val, addr);
4084 if (ret)
4085 return ret;
4086 if (val)
4087 goto err_size;
4088 }
4089 size = sizeof(*attr);
4090 }
4091
4092 ret = copy_from_user(attr, uattr, size);
4093 if (ret)
4094 return -EFAULT;
4095
4096 /*
4097 * XXX: do we want to be lenient like existing syscalls; or do we want
4098 * to be strict and return an error on out-of-bounds values?
4099 */
Dongsheng Yang75e45d52014-02-11 15:34:50 +08004100 attr->sched_nice = clamp(attr->sched_nice, MIN_NICE, MAX_NICE);
Dario Faggiolid50dde52013-11-07 14:43:36 +01004101
Michael Kerriske78c7bc2014-05-09 16:54:28 +02004102 return 0;
Dario Faggiolid50dde52013-11-07 14:43:36 +01004103
4104err_size:
4105 put_user(sizeof(*attr), &uattr->size);
Michael Kerriske78c7bc2014-05-09 16:54:28 +02004106 return -E2BIG;
Dario Faggiolid50dde52013-11-07 14:43:36 +01004107}
4108
Linus Torvalds1da177e2005-04-16 15:20:36 -07004109/**
4110 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
4111 * @pid: the pid in question.
4112 * @policy: new policy.
4113 * @param: structure containing the new RT priority.
Yacine Belkadie69f6182013-07-12 20:45:47 +02004114 *
4115 * Return: 0 on success. An error code otherwise.
Linus Torvalds1da177e2005-04-16 15:20:36 -07004116 */
Heiko Carstens5add95d2009-01-14 14:14:08 +01004117SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy,
4118 struct sched_param __user *, param)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004119{
Jason Baronc21761f2006-01-18 17:43:03 -08004120 /* negative values for policy are not valid */
4121 if (policy < 0)
4122 return -EINVAL;
4123
Linus Torvalds1da177e2005-04-16 15:20:36 -07004124 return do_sched_setscheduler(pid, policy, param);
4125}
4126
4127/**
4128 * sys_sched_setparam - set/change the RT priority of a thread
4129 * @pid: the pid in question.
4130 * @param: structure containing the new RT priority.
Yacine Belkadie69f6182013-07-12 20:45:47 +02004131 *
4132 * Return: 0 on success. An error code otherwise.
Linus Torvalds1da177e2005-04-16 15:20:36 -07004133 */
Heiko Carstens5add95d2009-01-14 14:14:08 +01004134SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004135{
Steven Rostedtc13db6b2014-07-23 11:28:26 -04004136 return do_sched_setscheduler(pid, SETPARAM_POLICY, param);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004137}
4138
4139/**
Dario Faggiolid50dde52013-11-07 14:43:36 +01004140 * sys_sched_setattr - same as above, but with extended sched_attr
4141 * @pid: the pid in question.
Juri Lelli5778fcc2014-01-14 16:10:39 +01004142 * @uattr: structure containing the extended parameters.
Masanari Iidadb66d752014-04-18 01:59:15 +09004143 * @flags: for future extension.
Dario Faggiolid50dde52013-11-07 14:43:36 +01004144 */
Peter Zijlstra6d35ab42014-02-14 17:19:29 +01004145SYSCALL_DEFINE3(sched_setattr, pid_t, pid, struct sched_attr __user *, uattr,
4146 unsigned int, flags)
Dario Faggiolid50dde52013-11-07 14:43:36 +01004147{
4148 struct sched_attr attr;
4149 struct task_struct *p;
4150 int retval;
4151
Peter Zijlstra6d35ab42014-02-14 17:19:29 +01004152 if (!uattr || pid < 0 || flags)
Dario Faggiolid50dde52013-11-07 14:43:36 +01004153 return -EINVAL;
4154
Michael Kerrisk143cf232014-05-09 16:54:15 +02004155 retval = sched_copy_attr(uattr, &attr);
4156 if (retval)
4157 return retval;
Dario Faggiolid50dde52013-11-07 14:43:36 +01004158
Richard Weinbergerb14ed2c2014-06-02 22:38:34 +02004159 if ((int)attr.sched_policy < 0)
Peter Zijlstradbdb2272014-05-09 10:49:03 +02004160 return -EINVAL;
Dario Faggiolid50dde52013-11-07 14:43:36 +01004161
4162 rcu_read_lock();
4163 retval = -ESRCH;
4164 p = find_process_by_pid(pid);
4165 if (p != NULL)
4166 retval = sched_setattr(p, &attr);
4167 rcu_read_unlock();
4168
4169 return retval;
4170}
4171
4172/**
Linus Torvalds1da177e2005-04-16 15:20:36 -07004173 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
4174 * @pid: the pid in question.
Yacine Belkadie69f6182013-07-12 20:45:47 +02004175 *
4176 * Return: On success, the policy of the thread. Otherwise, a negative error
4177 * code.
Linus Torvalds1da177e2005-04-16 15:20:36 -07004178 */
Heiko Carstens5add95d2009-01-14 14:14:08 +01004179SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004180{
Ingo Molnar36c8b582006-07-03 00:25:41 -07004181 struct task_struct *p;
Andi Kleen3a5c3592007-10-15 17:00:14 +02004182 int retval;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004183
4184 if (pid < 0)
Andi Kleen3a5c3592007-10-15 17:00:14 +02004185 return -EINVAL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004186
4187 retval = -ESRCH;
Thomas Gleixner5fe85be2009-12-09 10:14:58 +00004188 rcu_read_lock();
Linus Torvalds1da177e2005-04-16 15:20:36 -07004189 p = find_process_by_pid(pid);
4190 if (p) {
4191 retval = security_task_getscheduler(p);
4192 if (!retval)
Lennart Poetteringca94c442009-06-15 17:17:47 +02004193 retval = p->policy
4194 | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004195 }
Thomas Gleixner5fe85be2009-12-09 10:14:58 +00004196 rcu_read_unlock();
Linus Torvalds1da177e2005-04-16 15:20:36 -07004197 return retval;
4198}
4199
4200/**
Lennart Poetteringca94c442009-06-15 17:17:47 +02004201 * sys_sched_getparam - get the RT priority of a thread
Linus Torvalds1da177e2005-04-16 15:20:36 -07004202 * @pid: the pid in question.
4203 * @param: structure containing the RT priority.
Yacine Belkadie69f6182013-07-12 20:45:47 +02004204 *
4205 * Return: On success, 0 and the RT priority is in @param. Otherwise, an error
4206 * code.
Linus Torvalds1da177e2005-04-16 15:20:36 -07004207 */
Heiko Carstens5add95d2009-01-14 14:14:08 +01004208SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004209{
Peter Zijlstrace5f7f82014-05-12 22:50:34 +02004210 struct sched_param lp = { .sched_priority = 0 };
Ingo Molnar36c8b582006-07-03 00:25:41 -07004211 struct task_struct *p;
Andi Kleen3a5c3592007-10-15 17:00:14 +02004212 int retval;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004213
4214 if (!param || pid < 0)
Andi Kleen3a5c3592007-10-15 17:00:14 +02004215 return -EINVAL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004216
Thomas Gleixner5fe85be2009-12-09 10:14:58 +00004217 rcu_read_lock();
Linus Torvalds1da177e2005-04-16 15:20:36 -07004218 p = find_process_by_pid(pid);
4219 retval = -ESRCH;
4220 if (!p)
4221 goto out_unlock;
4222
4223 retval = security_task_getscheduler(p);
4224 if (retval)
4225 goto out_unlock;
4226
Peter Zijlstrace5f7f82014-05-12 22:50:34 +02004227 if (task_has_rt_policy(p))
4228 lp.sched_priority = p->rt_priority;
Thomas Gleixner5fe85be2009-12-09 10:14:58 +00004229 rcu_read_unlock();
Linus Torvalds1da177e2005-04-16 15:20:36 -07004230
4231 /*
4232 * This one might sleep, we cannot do it with a spinlock held ...
4233 */
4234 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
4235
Linus Torvalds1da177e2005-04-16 15:20:36 -07004236 return retval;
4237
4238out_unlock:
Thomas Gleixner5fe85be2009-12-09 10:14:58 +00004239 rcu_read_unlock();
Linus Torvalds1da177e2005-04-16 15:20:36 -07004240 return retval;
4241}
4242
Dario Faggiolid50dde52013-11-07 14:43:36 +01004243static int sched_read_attr(struct sched_attr __user *uattr,
4244 struct sched_attr *attr,
4245 unsigned int usize)
4246{
4247 int ret;
4248
4249 if (!access_ok(VERIFY_WRITE, uattr, usize))
4250 return -EFAULT;
4251
4252 /*
4253 * If we're handed a smaller struct than we know of,
4254 * ensure all the unknown bits are 0 - i.e. old
4255 * user-space does not get uncomplete information.
4256 */
4257 if (usize < sizeof(*attr)) {
4258 unsigned char *addr;
4259 unsigned char *end;
4260
4261 addr = (void *)attr + usize;
4262 end = (void *)attr + sizeof(*attr);
4263
4264 for (; addr < end; addr++) {
4265 if (*addr)
Michael Kerrisk22400672014-05-09 16:54:33 +02004266 return -EFBIG;
Dario Faggiolid50dde52013-11-07 14:43:36 +01004267 }
4268
4269 attr->size = usize;
4270 }
4271
Vegard Nossum4efbc452014-02-16 22:24:17 +01004272 ret = copy_to_user(uattr, attr, attr->size);
Dario Faggiolid50dde52013-11-07 14:43:36 +01004273 if (ret)
4274 return -EFAULT;
4275
Michael Kerrisk22400672014-05-09 16:54:33 +02004276 return 0;
Dario Faggiolid50dde52013-11-07 14:43:36 +01004277}
4278
4279/**
Dario Faggioliaab03e02013-11-28 11:14:43 +01004280 * sys_sched_getattr - similar to sched_getparam, but with sched_attr
Dario Faggiolid50dde52013-11-07 14:43:36 +01004281 * @pid: the pid in question.
Juri Lelli5778fcc2014-01-14 16:10:39 +01004282 * @uattr: structure containing the extended parameters.
Dario Faggiolid50dde52013-11-07 14:43:36 +01004283 * @size: sizeof(attr) for fwd/bwd comp.
Masanari Iidadb66d752014-04-18 01:59:15 +09004284 * @flags: for future extension.
Dario Faggiolid50dde52013-11-07 14:43:36 +01004285 */
Peter Zijlstra6d35ab42014-02-14 17:19:29 +01004286SYSCALL_DEFINE4(sched_getattr, pid_t, pid, struct sched_attr __user *, uattr,
4287 unsigned int, size, unsigned int, flags)
Dario Faggiolid50dde52013-11-07 14:43:36 +01004288{
4289 struct sched_attr attr = {
4290 .size = sizeof(struct sched_attr),
4291 };
4292 struct task_struct *p;
4293 int retval;
4294
4295 if (!uattr || pid < 0 || size > PAGE_SIZE ||
Peter Zijlstra6d35ab42014-02-14 17:19:29 +01004296 size < SCHED_ATTR_SIZE_VER0 || flags)
Dario Faggiolid50dde52013-11-07 14:43:36 +01004297 return -EINVAL;
4298
4299 rcu_read_lock();
4300 p = find_process_by_pid(pid);
4301 retval = -ESRCH;
4302 if (!p)
4303 goto out_unlock;
4304
4305 retval = security_task_getscheduler(p);
4306 if (retval)
4307 goto out_unlock;
4308
4309 attr.sched_policy = p->policy;
Peter Zijlstra7479f3c2014-01-15 17:05:04 +01004310 if (p->sched_reset_on_fork)
4311 attr.sched_flags |= SCHED_FLAG_RESET_ON_FORK;
Dario Faggioliaab03e02013-11-28 11:14:43 +01004312 if (task_has_dl_policy(p))
4313 __getparam_dl(p, &attr);
4314 else if (task_has_rt_policy(p))
Dario Faggiolid50dde52013-11-07 14:43:36 +01004315 attr.sched_priority = p->rt_priority;
4316 else
Dongsheng Yangd0ea0262014-01-27 22:00:45 -05004317 attr.sched_nice = task_nice(p);
Dario Faggiolid50dde52013-11-07 14:43:36 +01004318
4319 rcu_read_unlock();
4320
4321 retval = sched_read_attr(uattr, &attr, size);
4322 return retval;
4323
4324out_unlock:
4325 rcu_read_unlock();
4326 return retval;
4327}
4328
Rusty Russell96f874e2008-11-25 02:35:14 +10304329long sched_setaffinity(pid_t pid, const struct cpumask *in_mask)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004330{
Rusty Russell5a16f3d2008-11-25 02:35:11 +10304331 cpumask_var_t cpus_allowed, new_mask;
Ingo Molnar36c8b582006-07-03 00:25:41 -07004332 struct task_struct *p;
4333 int retval;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004334
Thomas Gleixner23f5d142009-12-09 10:15:01 +00004335 rcu_read_lock();
Linus Torvalds1da177e2005-04-16 15:20:36 -07004336
4337 p = find_process_by_pid(pid);
4338 if (!p) {
Thomas Gleixner23f5d142009-12-09 10:15:01 +00004339 rcu_read_unlock();
Linus Torvalds1da177e2005-04-16 15:20:36 -07004340 return -ESRCH;
4341 }
4342
Thomas Gleixner23f5d142009-12-09 10:15:01 +00004343 /* Prevent p going away */
Linus Torvalds1da177e2005-04-16 15:20:36 -07004344 get_task_struct(p);
Thomas Gleixner23f5d142009-12-09 10:15:01 +00004345 rcu_read_unlock();
Linus Torvalds1da177e2005-04-16 15:20:36 -07004346
Tejun Heo14a40ff2013-03-19 13:45:20 -07004347 if (p->flags & PF_NO_SETAFFINITY) {
4348 retval = -EINVAL;
4349 goto out_put_task;
4350 }
Rusty Russell5a16f3d2008-11-25 02:35:11 +10304351 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) {
4352 retval = -ENOMEM;
4353 goto out_put_task;
4354 }
4355 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) {
4356 retval = -ENOMEM;
4357 goto out_free_cpus_allowed;
4358 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07004359 retval = -EPERM;
Eric W. Biederman4c44aaa2012-07-26 05:05:21 -07004360 if (!check_same_owner(p)) {
4361 rcu_read_lock();
4362 if (!ns_capable(__task_cred(p)->user_ns, CAP_SYS_NICE)) {
4363 rcu_read_unlock();
Kirill Tkhai16303ab2014-09-22 22:36:30 +04004364 goto out_free_new_mask;
Eric W. Biederman4c44aaa2012-07-26 05:05:21 -07004365 }
4366 rcu_read_unlock();
4367 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07004368
KOSAKI Motohirob0ae1982010-10-15 04:21:18 +09004369 retval = security_task_setscheduler(p);
David Quigleye7834f82006-06-23 02:03:59 -07004370 if (retval)
Kirill Tkhai16303ab2014-09-22 22:36:30 +04004371 goto out_free_new_mask;
David Quigleye7834f82006-06-23 02:03:59 -07004372
Peter Zijlstrae4099a52013-12-17 10:03:34 +01004373
4374 cpuset_cpus_allowed(p, cpus_allowed);
4375 cpumask_and(new_mask, in_mask, cpus_allowed);
4376
Dario Faggioli332ac172013-11-07 14:43:45 +01004377 /*
4378 * Since bandwidth control happens on root_domain basis,
4379 * if admission test is enabled, we only admit -deadline
4380 * tasks allowed to run on all the CPUs in the task's
4381 * root_domain.
4382 */
4383#ifdef CONFIG_SMP
Kirill Tkhaif1e3a092014-09-22 22:36:36 +04004384 if (task_has_dl_policy(p) && dl_bandwidth_enabled()) {
4385 rcu_read_lock();
4386 if (!cpumask_subset(task_rq(p)->rd->span, new_mask)) {
Dario Faggioli332ac172013-11-07 14:43:45 +01004387 retval = -EBUSY;
Kirill Tkhaif1e3a092014-09-22 22:36:36 +04004388 rcu_read_unlock();
Kirill Tkhai16303ab2014-09-22 22:36:30 +04004389 goto out_free_new_mask;
Dario Faggioli332ac172013-11-07 14:43:45 +01004390 }
Kirill Tkhaif1e3a092014-09-22 22:36:36 +04004391 rcu_read_unlock();
Dario Faggioli332ac172013-11-07 14:43:45 +01004392 }
4393#endif
Peter Zijlstra49246272010-10-17 21:46:10 +02004394again:
Peter Zijlstra25834c72015-05-15 17:43:34 +02004395 retval = __set_cpus_allowed_ptr(p, new_mask, true);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004396
Paul Menage8707d8b2007-10-18 23:40:22 -07004397 if (!retval) {
Rusty Russell5a16f3d2008-11-25 02:35:11 +10304398 cpuset_cpus_allowed(p, cpus_allowed);
4399 if (!cpumask_subset(new_mask, cpus_allowed)) {
Paul Menage8707d8b2007-10-18 23:40:22 -07004400 /*
4401 * We must have raced with a concurrent cpuset
4402 * update. Just reset the cpus_allowed to the
4403 * cpuset's cpus_allowed
4404 */
Rusty Russell5a16f3d2008-11-25 02:35:11 +10304405 cpumask_copy(new_mask, cpus_allowed);
Paul Menage8707d8b2007-10-18 23:40:22 -07004406 goto again;
4407 }
4408 }
Kirill Tkhai16303ab2014-09-22 22:36:30 +04004409out_free_new_mask:
Rusty Russell5a16f3d2008-11-25 02:35:11 +10304410 free_cpumask_var(new_mask);
4411out_free_cpus_allowed:
4412 free_cpumask_var(cpus_allowed);
4413out_put_task:
Linus Torvalds1da177e2005-04-16 15:20:36 -07004414 put_task_struct(p);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004415 return retval;
4416}
4417
4418static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
Rusty Russell96f874e2008-11-25 02:35:14 +10304419 struct cpumask *new_mask)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004420{
Rusty Russell96f874e2008-11-25 02:35:14 +10304421 if (len < cpumask_size())
4422 cpumask_clear(new_mask);
4423 else if (len > cpumask_size())
4424 len = cpumask_size();
4425
Linus Torvalds1da177e2005-04-16 15:20:36 -07004426 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
4427}
4428
4429/**
4430 * sys_sched_setaffinity - set the cpu affinity of a process
4431 * @pid: pid of the process
4432 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4433 * @user_mask_ptr: user-space pointer to the new cpu mask
Yacine Belkadie69f6182013-07-12 20:45:47 +02004434 *
4435 * Return: 0 on success. An error code otherwise.
Linus Torvalds1da177e2005-04-16 15:20:36 -07004436 */
Heiko Carstens5add95d2009-01-14 14:14:08 +01004437SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len,
4438 unsigned long __user *, user_mask_ptr)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004439{
Rusty Russell5a16f3d2008-11-25 02:35:11 +10304440 cpumask_var_t new_mask;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004441 int retval;
4442
Rusty Russell5a16f3d2008-11-25 02:35:11 +10304443 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
4444 return -ENOMEM;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004445
Rusty Russell5a16f3d2008-11-25 02:35:11 +10304446 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask);
4447 if (retval == 0)
4448 retval = sched_setaffinity(pid, new_mask);
4449 free_cpumask_var(new_mask);
4450 return retval;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004451}
4452
Rusty Russell96f874e2008-11-25 02:35:14 +10304453long sched_getaffinity(pid_t pid, struct cpumask *mask)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004454{
Ingo Molnar36c8b582006-07-03 00:25:41 -07004455 struct task_struct *p;
Thomas Gleixner31605682009-12-08 20:24:16 +00004456 unsigned long flags;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004457 int retval;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004458
Thomas Gleixner23f5d142009-12-09 10:15:01 +00004459 rcu_read_lock();
Linus Torvalds1da177e2005-04-16 15:20:36 -07004460
4461 retval = -ESRCH;
4462 p = find_process_by_pid(pid);
4463 if (!p)
4464 goto out_unlock;
4465
David Quigleye7834f82006-06-23 02:03:59 -07004466 retval = security_task_getscheduler(p);
4467 if (retval)
4468 goto out_unlock;
4469
Peter Zijlstra013fdb82011-04-05 17:23:45 +02004470 raw_spin_lock_irqsave(&p->pi_lock, flags);
Peter Zijlstra6acce3e2013-10-11 14:38:20 +02004471 cpumask_and(mask, &p->cpus_allowed, cpu_active_mask);
Peter Zijlstra013fdb82011-04-05 17:23:45 +02004472 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004473
4474out_unlock:
Thomas Gleixner23f5d142009-12-09 10:15:01 +00004475 rcu_read_unlock();
Linus Torvalds1da177e2005-04-16 15:20:36 -07004476
Ulrich Drepper9531b622007-08-09 11:16:46 +02004477 return retval;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004478}
4479
4480/**
4481 * sys_sched_getaffinity - get the cpu affinity of a process
4482 * @pid: pid of the process
4483 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4484 * @user_mask_ptr: user-space pointer to hold the current cpu mask
Yacine Belkadie69f6182013-07-12 20:45:47 +02004485 *
4486 * Return: 0 on success. An error code otherwise.
Linus Torvalds1da177e2005-04-16 15:20:36 -07004487 */
Heiko Carstens5add95d2009-01-14 14:14:08 +01004488SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len,
4489 unsigned long __user *, user_mask_ptr)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004490{
4491 int ret;
Rusty Russellf17c8602008-11-25 02:35:11 +10304492 cpumask_var_t mask;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004493
Anton Blanchard84fba5e2010-04-06 17:02:19 +10004494 if ((len * BITS_PER_BYTE) < nr_cpu_ids)
KOSAKI Motohirocd3d8032010-03-12 16:15:36 +09004495 return -EINVAL;
4496 if (len & (sizeof(unsigned long)-1))
Linus Torvalds1da177e2005-04-16 15:20:36 -07004497 return -EINVAL;
4498
Rusty Russellf17c8602008-11-25 02:35:11 +10304499 if (!alloc_cpumask_var(&mask, GFP_KERNEL))
4500 return -ENOMEM;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004501
Rusty Russellf17c8602008-11-25 02:35:11 +10304502 ret = sched_getaffinity(pid, mask);
4503 if (ret == 0) {
KOSAKI Motohiro8bc037f2010-03-17 09:36:58 +09004504 size_t retlen = min_t(size_t, len, cpumask_size());
KOSAKI Motohirocd3d8032010-03-12 16:15:36 +09004505
4506 if (copy_to_user(user_mask_ptr, mask, retlen))
Rusty Russellf17c8602008-11-25 02:35:11 +10304507 ret = -EFAULT;
4508 else
KOSAKI Motohirocd3d8032010-03-12 16:15:36 +09004509 ret = retlen;
Rusty Russellf17c8602008-11-25 02:35:11 +10304510 }
4511 free_cpumask_var(mask);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004512
Rusty Russellf17c8602008-11-25 02:35:11 +10304513 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004514}
4515
4516/**
4517 * sys_sched_yield - yield the current processor to other threads.
4518 *
Ingo Molnardd41f592007-07-09 18:51:59 +02004519 * This function yields the current CPU to other tasks. If there are no
4520 * other threads running on this CPU then this function will return.
Yacine Belkadie69f6182013-07-12 20:45:47 +02004521 *
4522 * Return: 0.
Linus Torvalds1da177e2005-04-16 15:20:36 -07004523 */
Heiko Carstens5add95d2009-01-14 14:14:08 +01004524SYSCALL_DEFINE0(sched_yield)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004525{
Ingo Molnar70b97a72006-07-03 00:25:42 -07004526 struct rq *rq = this_rq_lock();
Linus Torvalds1da177e2005-04-16 15:20:36 -07004527
Ingo Molnar2d723762007-10-15 17:00:12 +02004528 schedstat_inc(rq, yld_count);
Dmitry Adamushko4530d7a2007-10-15 17:00:08 +02004529 current->sched_class->yield_task(rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004530
4531 /*
4532 * Since we are going to call schedule() anyway, there's
4533 * no need to preempt or enable interrupts:
4534 */
4535 __release(rq->lock);
Ingo Molnar8a25d5d2006-07-03 00:24:54 -07004536 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
Thomas Gleixner9828ea92009-12-03 20:55:53 +01004537 do_raw_spin_unlock(&rq->lock);
Thomas Gleixnerba74c142011-03-21 13:32:17 +01004538 sched_preempt_enable_no_resched();
Linus Torvalds1da177e2005-04-16 15:20:36 -07004539
4540 schedule();
4541
4542 return 0;
4543}
4544
Herbert Xu02b67cc32008-01-25 21:08:28 +01004545int __sched _cond_resched(void)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004546{
Konstantin Khlebnikovfe32d3c2015-07-15 12:52:04 +03004547 if (should_resched(0)) {
Frederic Weisbeckera18b5d02015-01-22 18:08:04 +01004548 preempt_schedule_common();
Linus Torvalds1da177e2005-04-16 15:20:36 -07004549 return 1;
4550 }
4551 return 0;
4552}
Herbert Xu02b67cc32008-01-25 21:08:28 +01004553EXPORT_SYMBOL(_cond_resched);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004554
4555/*
Frederic Weisbecker613afbf2009-07-16 15:44:29 +02004556 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
Linus Torvalds1da177e2005-04-16 15:20:36 -07004557 * call schedule, and on return reacquire the lock.
4558 *
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01004559 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
Linus Torvalds1da177e2005-04-16 15:20:36 -07004560 * operations here to prevent schedule() from being called twice (once via
4561 * spin_unlock(), once by hand).
4562 */
Frederic Weisbecker613afbf2009-07-16 15:44:29 +02004563int __cond_resched_lock(spinlock_t *lock)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004564{
Konstantin Khlebnikovfe32d3c2015-07-15 12:52:04 +03004565 int resched = should_resched(PREEMPT_LOCK_OFFSET);
Jan Kara6df3cec2005-06-13 15:52:32 -07004566 int ret = 0;
4567
Peter Zijlstraf607c662009-07-20 19:16:29 +02004568 lockdep_assert_held(lock);
4569
Paul E. McKenney4a81e832014-06-20 16:49:01 -07004570 if (spin_needbreak(lock) || resched) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07004571 spin_unlock(lock);
Peter Zijlstrad86ee482009-07-10 14:57:57 +02004572 if (resched)
Frederic Weisbeckera18b5d02015-01-22 18:08:04 +01004573 preempt_schedule_common();
Nick Piggin95c354f2008-01-30 13:31:20 +01004574 else
4575 cpu_relax();
Jan Kara6df3cec2005-06-13 15:52:32 -07004576 ret = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004577 spin_lock(lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004578 }
Jan Kara6df3cec2005-06-13 15:52:32 -07004579 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004580}
Frederic Weisbecker613afbf2009-07-16 15:44:29 +02004581EXPORT_SYMBOL(__cond_resched_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004582
Frederic Weisbecker613afbf2009-07-16 15:44:29 +02004583int __sched __cond_resched_softirq(void)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004584{
4585 BUG_ON(!in_softirq());
4586
Konstantin Khlebnikovfe32d3c2015-07-15 12:52:04 +03004587 if (should_resched(SOFTIRQ_DISABLE_OFFSET)) {
Thomas Gleixner98d825672007-05-23 13:58:18 -07004588 local_bh_enable();
Frederic Weisbeckera18b5d02015-01-22 18:08:04 +01004589 preempt_schedule_common();
Linus Torvalds1da177e2005-04-16 15:20:36 -07004590 local_bh_disable();
4591 return 1;
4592 }
4593 return 0;
4594}
Frederic Weisbecker613afbf2009-07-16 15:44:29 +02004595EXPORT_SYMBOL(__cond_resched_softirq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004596
Linus Torvalds1da177e2005-04-16 15:20:36 -07004597/**
4598 * yield - yield the current processor to other threads.
4599 *
Peter Zijlstra8e3fabf2012-03-06 18:54:26 +01004600 * Do not ever use this function, there's a 99% chance you're doing it wrong.
4601 *
4602 * The scheduler is at all times free to pick the calling task as the most
4603 * eligible task to run, if removing the yield() call from your code breaks
4604 * it, its already broken.
4605 *
4606 * Typical broken usage is:
4607 *
4608 * while (!event)
4609 * yield();
4610 *
4611 * where one assumes that yield() will let 'the other' process run that will
4612 * make event true. If the current task is a SCHED_FIFO task that will never
4613 * happen. Never use yield() as a progress guarantee!!
4614 *
4615 * If you want to use yield() to wait for something, use wait_event().
4616 * If you want to use yield() to be 'nice' for others, use cond_resched().
4617 * If you still want to use yield(), do not!
Linus Torvalds1da177e2005-04-16 15:20:36 -07004618 */
4619void __sched yield(void)
4620{
4621 set_current_state(TASK_RUNNING);
4622 sys_sched_yield();
4623}
Linus Torvalds1da177e2005-04-16 15:20:36 -07004624EXPORT_SYMBOL(yield);
4625
Mike Galbraithd95f4122011-02-01 09:50:51 -05004626/**
4627 * yield_to - yield the current processor to another thread in
4628 * your thread group, or accelerate that thread toward the
4629 * processor it's on.
Randy Dunlap16addf92011-03-18 09:34:53 -07004630 * @p: target task
4631 * @preempt: whether task preemption is allowed or not
Mike Galbraithd95f4122011-02-01 09:50:51 -05004632 *
4633 * It's the caller's job to ensure that the target task struct
4634 * can't go away on us before we can do any checks.
4635 *
Yacine Belkadie69f6182013-07-12 20:45:47 +02004636 * Return:
Peter Zijlstra7b270f62013-01-22 13:09:13 +05304637 * true (>0) if we indeed boosted the target task.
4638 * false (0) if we failed to boost the target.
4639 * -ESRCH if there's no task to yield to.
Mike Galbraithd95f4122011-02-01 09:50:51 -05004640 */
Dan Carpenterfa933842014-05-23 13:20:42 +03004641int __sched yield_to(struct task_struct *p, bool preempt)
Mike Galbraithd95f4122011-02-01 09:50:51 -05004642{
4643 struct task_struct *curr = current;
4644 struct rq *rq, *p_rq;
4645 unsigned long flags;
Dan Carpenterc3c18642013-02-05 14:37:51 +03004646 int yielded = 0;
Mike Galbraithd95f4122011-02-01 09:50:51 -05004647
4648 local_irq_save(flags);
4649 rq = this_rq();
4650
4651again:
4652 p_rq = task_rq(p);
Peter Zijlstra7b270f62013-01-22 13:09:13 +05304653 /*
4654 * If we're the only runnable task on the rq and target rq also
4655 * has only one task, there's absolutely no point in yielding.
4656 */
4657 if (rq->nr_running == 1 && p_rq->nr_running == 1) {
4658 yielded = -ESRCH;
4659 goto out_irq;
4660 }
4661
Mike Galbraithd95f4122011-02-01 09:50:51 -05004662 double_rq_lock(rq, p_rq);
Shigeru Yoshida39e24d8f2013-11-23 18:38:01 +09004663 if (task_rq(p) != p_rq) {
Mike Galbraithd95f4122011-02-01 09:50:51 -05004664 double_rq_unlock(rq, p_rq);
4665 goto again;
4666 }
4667
4668 if (!curr->sched_class->yield_to_task)
Peter Zijlstra7b270f62013-01-22 13:09:13 +05304669 goto out_unlock;
Mike Galbraithd95f4122011-02-01 09:50:51 -05004670
4671 if (curr->sched_class != p->sched_class)
Peter Zijlstra7b270f62013-01-22 13:09:13 +05304672 goto out_unlock;
Mike Galbraithd95f4122011-02-01 09:50:51 -05004673
4674 if (task_running(p_rq, p) || p->state)
Peter Zijlstra7b270f62013-01-22 13:09:13 +05304675 goto out_unlock;
Mike Galbraithd95f4122011-02-01 09:50:51 -05004676
4677 yielded = curr->sched_class->yield_to_task(rq, p, preempt);
Venkatesh Pallipadi6d1cafd2011-03-01 16:28:21 -08004678 if (yielded) {
Mike Galbraithd95f4122011-02-01 09:50:51 -05004679 schedstat_inc(rq, yld_count);
Venkatesh Pallipadi6d1cafd2011-03-01 16:28:21 -08004680 /*
4681 * Make p's CPU reschedule; pick_next_entity takes care of
4682 * fairness.
4683 */
4684 if (preempt && rq != p_rq)
Kirill Tkhai88751252014-06-29 00:03:57 +04004685 resched_curr(p_rq);
Venkatesh Pallipadi6d1cafd2011-03-01 16:28:21 -08004686 }
Mike Galbraithd95f4122011-02-01 09:50:51 -05004687
Peter Zijlstra7b270f62013-01-22 13:09:13 +05304688out_unlock:
Mike Galbraithd95f4122011-02-01 09:50:51 -05004689 double_rq_unlock(rq, p_rq);
Peter Zijlstra7b270f62013-01-22 13:09:13 +05304690out_irq:
Mike Galbraithd95f4122011-02-01 09:50:51 -05004691 local_irq_restore(flags);
4692
Peter Zijlstra7b270f62013-01-22 13:09:13 +05304693 if (yielded > 0)
Mike Galbraithd95f4122011-02-01 09:50:51 -05004694 schedule();
4695
4696 return yielded;
4697}
4698EXPORT_SYMBOL_GPL(yield_to);
4699
Linus Torvalds1da177e2005-04-16 15:20:36 -07004700/*
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01004701 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
Linus Torvalds1da177e2005-04-16 15:20:36 -07004702 * that process accounting knows that this is a task in IO wait state.
Linus Torvalds1da177e2005-04-16 15:20:36 -07004703 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07004704long __sched io_schedule_timeout(long timeout)
4705{
NeilBrown9cff8ad2015-02-13 15:49:17 +11004706 int old_iowait = current->in_iowait;
4707 struct rq *rq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004708 long ret;
4709
Arjan van de Ven8f0dfc32009-07-20 11:26:58 -07004710 current->in_iowait = 1;
Shaohua Li10d784e2015-05-08 10:51:29 -07004711 blk_schedule_flush_plug(current);
NeilBrown9cff8ad2015-02-13 15:49:17 +11004712
4713 delayacct_blkio_start();
4714 rq = raw_rq();
4715 atomic_inc(&rq->nr_iowait);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004716 ret = schedule_timeout(timeout);
NeilBrown9cff8ad2015-02-13 15:49:17 +11004717 current->in_iowait = old_iowait;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004718 atomic_dec(&rq->nr_iowait);
Shailabh Nagar0ff92242006-07-14 00:24:37 -07004719 delayacct_blkio_end();
NeilBrown9cff8ad2015-02-13 15:49:17 +11004720
Linus Torvalds1da177e2005-04-16 15:20:36 -07004721 return ret;
4722}
NeilBrown9cff8ad2015-02-13 15:49:17 +11004723EXPORT_SYMBOL(io_schedule_timeout);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004724
4725/**
4726 * sys_sched_get_priority_max - return maximum RT priority.
4727 * @policy: scheduling class.
4728 *
Yacine Belkadie69f6182013-07-12 20:45:47 +02004729 * Return: On success, this syscall returns the maximum
4730 * rt_priority that can be used by a given scheduling class.
4731 * On failure, a negative error code is returned.
Linus Torvalds1da177e2005-04-16 15:20:36 -07004732 */
Heiko Carstens5add95d2009-01-14 14:14:08 +01004733SYSCALL_DEFINE1(sched_get_priority_max, int, policy)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004734{
4735 int ret = -EINVAL;
4736
4737 switch (policy) {
4738 case SCHED_FIFO:
4739 case SCHED_RR:
4740 ret = MAX_USER_RT_PRIO-1;
4741 break;
Dario Faggioliaab03e02013-11-28 11:14:43 +01004742 case SCHED_DEADLINE:
Linus Torvalds1da177e2005-04-16 15:20:36 -07004743 case SCHED_NORMAL:
Ingo Molnarb0a94992006-01-14 13:20:41 -08004744 case SCHED_BATCH:
Ingo Molnardd41f592007-07-09 18:51:59 +02004745 case SCHED_IDLE:
Linus Torvalds1da177e2005-04-16 15:20:36 -07004746 ret = 0;
4747 break;
4748 }
4749 return ret;
4750}
4751
4752/**
4753 * sys_sched_get_priority_min - return minimum RT priority.
4754 * @policy: scheduling class.
4755 *
Yacine Belkadie69f6182013-07-12 20:45:47 +02004756 * Return: On success, this syscall returns the minimum
4757 * rt_priority that can be used by a given scheduling class.
4758 * On failure, a negative error code is returned.
Linus Torvalds1da177e2005-04-16 15:20:36 -07004759 */
Heiko Carstens5add95d2009-01-14 14:14:08 +01004760SYSCALL_DEFINE1(sched_get_priority_min, int, policy)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004761{
4762 int ret = -EINVAL;
4763
4764 switch (policy) {
4765 case SCHED_FIFO:
4766 case SCHED_RR:
4767 ret = 1;
4768 break;
Dario Faggioliaab03e02013-11-28 11:14:43 +01004769 case SCHED_DEADLINE:
Linus Torvalds1da177e2005-04-16 15:20:36 -07004770 case SCHED_NORMAL:
Ingo Molnarb0a94992006-01-14 13:20:41 -08004771 case SCHED_BATCH:
Ingo Molnardd41f592007-07-09 18:51:59 +02004772 case SCHED_IDLE:
Linus Torvalds1da177e2005-04-16 15:20:36 -07004773 ret = 0;
4774 }
4775 return ret;
4776}
4777
4778/**
4779 * sys_sched_rr_get_interval - return the default timeslice of a process.
4780 * @pid: pid of the process.
4781 * @interval: userspace pointer to the timeslice value.
4782 *
4783 * this syscall writes the default timeslice value of a given process
4784 * into the user-space timespec buffer. A value of '0' means infinity.
Yacine Belkadie69f6182013-07-12 20:45:47 +02004785 *
4786 * Return: On success, 0 and the timeslice is in @interval. Otherwise,
4787 * an error code.
Linus Torvalds1da177e2005-04-16 15:20:36 -07004788 */
Heiko Carstens17da2bd2009-01-14 14:14:10 +01004789SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid,
Heiko Carstens754fe8d2009-01-14 14:14:09 +01004790 struct timespec __user *, interval)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004791{
Ingo Molnar36c8b582006-07-03 00:25:41 -07004792 struct task_struct *p;
Dmitry Adamushkoa4ec24b2007-10-15 17:00:13 +02004793 unsigned int time_slice;
Thomas Gleixnerdba091b2009-12-09 09:32:03 +01004794 unsigned long flags;
4795 struct rq *rq;
Andi Kleen3a5c3592007-10-15 17:00:14 +02004796 int retval;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004797 struct timespec t;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004798
4799 if (pid < 0)
Andi Kleen3a5c3592007-10-15 17:00:14 +02004800 return -EINVAL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004801
4802 retval = -ESRCH;
Thomas Gleixner1a551ae2009-12-09 10:15:11 +00004803 rcu_read_lock();
Linus Torvalds1da177e2005-04-16 15:20:36 -07004804 p = find_process_by_pid(pid);
4805 if (!p)
4806 goto out_unlock;
4807
4808 retval = security_task_getscheduler(p);
4809 if (retval)
4810 goto out_unlock;
4811
Thomas Gleixnerdba091b2009-12-09 09:32:03 +01004812 rq = task_rq_lock(p, &flags);
Peter Zijlstraa57beec2014-01-27 11:54:13 +01004813 time_slice = 0;
4814 if (p->sched_class->get_rr_interval)
4815 time_slice = p->sched_class->get_rr_interval(rq, p);
Peter Zijlstra0122ec52011-04-05 17:23:51 +02004816 task_rq_unlock(rq, p, &flags);
Dmitry Adamushkoa4ec24b2007-10-15 17:00:13 +02004817
Thomas Gleixner1a551ae2009-12-09 10:15:11 +00004818 rcu_read_unlock();
Dmitry Adamushkoa4ec24b2007-10-15 17:00:13 +02004819 jiffies_to_timespec(time_slice, &t);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004820 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004821 return retval;
Andi Kleen3a5c3592007-10-15 17:00:14 +02004822
Linus Torvalds1da177e2005-04-16 15:20:36 -07004823out_unlock:
Thomas Gleixner1a551ae2009-12-09 10:15:11 +00004824 rcu_read_unlock();
Linus Torvalds1da177e2005-04-16 15:20:36 -07004825 return retval;
4826}
4827
Steven Rostedt7c731e02008-05-12 21:20:41 +02004828static const char stat_nam[] = TASK_STATE_TO_CHAR_STR;
Ingo Molnar36c8b582006-07-03 00:25:41 -07004829
Ingo Molnar82a1fcb2008-01-25 21:08:02 +01004830void sched_show_task(struct task_struct *p)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004831{
Linus Torvalds1da177e2005-04-16 15:20:36 -07004832 unsigned long free = 0;
Paul E. McKenney4e797522012-11-07 13:35:32 -08004833 int ppid;
Tetsuo Handa1f8a7632014-12-05 21:22:22 +09004834 unsigned long state = p->state;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004835
Tetsuo Handa1f8a7632014-12-05 21:22:22 +09004836 if (state)
4837 state = __ffs(state) + 1;
Erik Gilling28d06862010-11-19 18:08:51 -08004838 printk(KERN_INFO "%-15.15s %c", p->comm,
Andreas Mohr2ed6e342006-07-10 04:43:52 -07004839 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
Ingo Molnar4bd77322007-07-11 21:21:47 +02004840#if BITS_PER_LONG == 32
Linus Torvalds1da177e2005-04-16 15:20:36 -07004841 if (state == TASK_RUNNING)
Peter Zijlstra3df0fc52009-12-20 14:23:57 +01004842 printk(KERN_CONT " running ");
Linus Torvalds1da177e2005-04-16 15:20:36 -07004843 else
Peter Zijlstra3df0fc52009-12-20 14:23:57 +01004844 printk(KERN_CONT " %08lx ", thread_saved_pc(p));
Linus Torvalds1da177e2005-04-16 15:20:36 -07004845#else
4846 if (state == TASK_RUNNING)
Peter Zijlstra3df0fc52009-12-20 14:23:57 +01004847 printk(KERN_CONT " running task ");
Linus Torvalds1da177e2005-04-16 15:20:36 -07004848 else
Peter Zijlstra3df0fc52009-12-20 14:23:57 +01004849 printk(KERN_CONT " %016lx ", thread_saved_pc(p));
Linus Torvalds1da177e2005-04-16 15:20:36 -07004850#endif
4851#ifdef CONFIG_DEBUG_STACK_USAGE
Eric Sandeen7c9f8862008-04-22 16:38:23 -05004852 free = stack_not_used(p);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004853#endif
Oleg Nesterova90e9842014-12-10 15:45:21 -08004854 ppid = 0;
Paul E. McKenney4e797522012-11-07 13:35:32 -08004855 rcu_read_lock();
Oleg Nesterova90e9842014-12-10 15:45:21 -08004856 if (pid_alive(p))
4857 ppid = task_pid_nr(rcu_dereference(p->real_parent));
Paul E. McKenney4e797522012-11-07 13:35:32 -08004858 rcu_read_unlock();
Peter Zijlstra3df0fc52009-12-20 14:23:57 +01004859 printk(KERN_CONT "%5lu %5d %6d 0x%08lx\n", free,
Paul E. McKenney4e797522012-11-07 13:35:32 -08004860 task_pid_nr(p), ppid,
David Rientjesaa47b7e2009-05-04 01:38:05 -07004861 (unsigned long)task_thread_info(p)->flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004862
Tejun Heo3d1cb202013-04-30 15:27:22 -07004863 print_worker_info(KERN_INFO, p);
Nick Piggin5fb5e6d2008-01-25 21:08:34 +01004864 show_stack(p, NULL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004865}
4866
Ingo Molnare59e2ae2006-12-06 20:35:59 -08004867void show_state_filter(unsigned long state_filter)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004868{
Ingo Molnar36c8b582006-07-03 00:25:41 -07004869 struct task_struct *g, *p;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004870
Ingo Molnar4bd77322007-07-11 21:21:47 +02004871#if BITS_PER_LONG == 32
Peter Zijlstra3df0fc52009-12-20 14:23:57 +01004872 printk(KERN_INFO
4873 " task PC stack pid father\n");
Linus Torvalds1da177e2005-04-16 15:20:36 -07004874#else
Peter Zijlstra3df0fc52009-12-20 14:23:57 +01004875 printk(KERN_INFO
4876 " task PC stack pid father\n");
Linus Torvalds1da177e2005-04-16 15:20:36 -07004877#endif
Thomas Gleixner510f5ac2011-07-17 20:47:54 +02004878 rcu_read_lock();
Oleg Nesterov5d07f422014-08-13 21:19:53 +02004879 for_each_process_thread(g, p) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07004880 /*
4881 * reset the NMI-timeout, listing all files on a slow
Lucas De Marchi25985ed2011-03-30 22:57:33 -03004882 * console might take a lot of time:
Linus Torvalds1da177e2005-04-16 15:20:36 -07004883 */
4884 touch_nmi_watchdog();
Ingo Molnar39bc89f2007-04-25 20:50:03 -07004885 if (!state_filter || (p->state & state_filter))
Ingo Molnar82a1fcb2008-01-25 21:08:02 +01004886 sched_show_task(p);
Oleg Nesterov5d07f422014-08-13 21:19:53 +02004887 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07004888
Jeremy Fitzhardinge04c91672007-05-08 00:28:05 -07004889 touch_all_softlockup_watchdogs();
4890
Ingo Molnardd41f592007-07-09 18:51:59 +02004891#ifdef CONFIG_SCHED_DEBUG
4892 sysrq_sched_debug_show();
4893#endif
Thomas Gleixner510f5ac2011-07-17 20:47:54 +02004894 rcu_read_unlock();
Ingo Molnare59e2ae2006-12-06 20:35:59 -08004895 /*
4896 * Only show locks if all tasks are dumped:
4897 */
Shmulik Ladkani93335a22009-11-25 15:23:41 +02004898 if (!state_filter)
Ingo Molnare59e2ae2006-12-06 20:35:59 -08004899 debug_show_all_locks();
Linus Torvalds1da177e2005-04-16 15:20:36 -07004900}
4901
Paul Gortmaker0db06282013-06-19 14:53:51 -04004902void init_idle_bootup_task(struct task_struct *idle)
Ingo Molnar1df21052007-07-09 18:51:58 +02004903{
Ingo Molnardd41f592007-07-09 18:51:59 +02004904 idle->sched_class = &idle_sched_class;
Ingo Molnar1df21052007-07-09 18:51:58 +02004905}
4906
Ingo Molnarf340c0d2005-06-28 16:40:42 +02004907/**
4908 * init_idle - set up an idle thread for a given CPU
4909 * @idle: task in question
4910 * @cpu: cpu the idle task belongs to
4911 *
4912 * NOTE: this function does not set the idle thread's NEED_RESCHED
4913 * flag, to make booting more robust.
4914 */
Paul Gortmaker0db06282013-06-19 14:53:51 -04004915void init_idle(struct task_struct *idle, int cpu)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004916{
Ingo Molnar70b97a72006-07-03 00:25:42 -07004917 struct rq *rq = cpu_rq(cpu);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004918 unsigned long flags;
4919
Peter Zijlstra25834c72015-05-15 17:43:34 +02004920 raw_spin_lock_irqsave(&idle->pi_lock, flags);
4921 raw_spin_lock(&rq->lock);
Ingo Molnar5cbd54e2008-11-12 20:05:50 +01004922
Rik van Riel5e1576e2013-10-07 11:29:26 +01004923 __sched_fork(0, idle);
Peter Zijlstra06b83b52009-12-16 18:04:35 +01004924 idle->state = TASK_RUNNING;
Ingo Molnardd41f592007-07-09 18:51:59 +02004925 idle->se.exec_start = sched_clock();
4926
KOSAKI Motohiro1e1b6c52011-05-19 15:08:58 +09004927 do_set_cpus_allowed(idle, cpumask_of(cpu));
Peter Zijlstra6506cf6c2010-09-16 17:50:31 +02004928 /*
4929 * We're having a chicken and egg problem, even though we are
4930 * holding rq->lock, the cpu isn't yet set to this cpu so the
4931 * lockdep check in task_group() will fail.
4932 *
4933 * Similar case to sched_fork(). / Alternatively we could
4934 * use task_rq_lock() here and obtain the other rq->lock.
4935 *
4936 * Silence PROVE_RCU
4937 */
4938 rcu_read_lock();
Ingo Molnardd41f592007-07-09 18:51:59 +02004939 __set_task_cpu(idle, cpu);
Peter Zijlstra6506cf6c2010-09-16 17:50:31 +02004940 rcu_read_unlock();
Linus Torvalds1da177e2005-04-16 15:20:36 -07004941
Linus Torvalds1da177e2005-04-16 15:20:36 -07004942 rq->curr = rq->idle = idle;
Kirill Tkhaida0c1e62014-08-20 13:47:32 +04004943 idle->on_rq = TASK_ON_RQ_QUEUED;
Peter Zijlstra3ca7a442011-04-05 17:23:40 +02004944#if defined(CONFIG_SMP)
4945 idle->on_cpu = 1;
Nick Piggin4866cde2005-06-25 14:57:23 -07004946#endif
Peter Zijlstra25834c72015-05-15 17:43:34 +02004947 raw_spin_unlock(&rq->lock);
4948 raw_spin_unlock_irqrestore(&idle->pi_lock, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004949
4950 /* Set the preempt count _outside_ the spinlocks! */
Peter Zijlstra01028742013-08-14 14:55:46 +02004951 init_idle_preempt_count(idle, cpu);
Jonathan Corbet625f2a32011-04-22 11:19:10 -06004952
Ingo Molnardd41f592007-07-09 18:51:59 +02004953 /*
4954 * The idle tasks have their own, simple scheduling class:
4955 */
4956 idle->sched_class = &idle_sched_class;
Steven Rostedt868baf02011-02-10 21:26:13 -05004957 ftrace_graph_init_idle_task(idle, cpu);
Frederic Weisbecker45eacc62013-05-15 22:16:32 +02004958 vtime_init_idle(idle, cpu);
Carsten Emdef1c6f1a2011-10-26 23:14:16 +02004959#if defined(CONFIG_SMP)
4960 sprintf(idle->comm, "%s/%d", INIT_TASK_COMM, cpu);
4961#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07004962}
4963
Juri Lellif82f8042014-10-07 09:52:11 +01004964int cpuset_cpumask_can_shrink(const struct cpumask *cur,
4965 const struct cpumask *trial)
4966{
4967 int ret = 1, trial_cpus;
4968 struct dl_bw *cur_dl_b;
4969 unsigned long flags;
4970
Mike Galbraithbb2bc552015-01-28 04:53:55 +01004971 if (!cpumask_weight(cur))
4972 return ret;
4973
Juri Lelli75e23e42014-10-28 11:54:46 +00004974 rcu_read_lock_sched();
Juri Lellif82f8042014-10-07 09:52:11 +01004975 cur_dl_b = dl_bw_of(cpumask_any(cur));
4976 trial_cpus = cpumask_weight(trial);
4977
4978 raw_spin_lock_irqsave(&cur_dl_b->lock, flags);
4979 if (cur_dl_b->bw != -1 &&
4980 cur_dl_b->bw * trial_cpus < cur_dl_b->total_bw)
4981 ret = 0;
4982 raw_spin_unlock_irqrestore(&cur_dl_b->lock, flags);
Juri Lelli75e23e42014-10-28 11:54:46 +00004983 rcu_read_unlock_sched();
Juri Lellif82f8042014-10-07 09:52:11 +01004984
4985 return ret;
4986}
4987
Juri Lelli7f514122014-09-19 10:22:40 +01004988int task_can_attach(struct task_struct *p,
4989 const struct cpumask *cs_cpus_allowed)
4990{
4991 int ret = 0;
4992
4993 /*
4994 * Kthreads which disallow setaffinity shouldn't be moved
4995 * to a new cpuset; we don't want to change their cpu
4996 * affinity and isolating such threads by their set of
4997 * allowed nodes is unnecessary. Thus, cpusets are not
4998 * applicable for such threads. This prevents checking for
4999 * success of set_cpus_allowed_ptr() on all attached tasks
5000 * before cpus_allowed may be changed.
5001 */
5002 if (p->flags & PF_NO_SETAFFINITY) {
5003 ret = -EINVAL;
5004 goto out;
5005 }
5006
5007#ifdef CONFIG_SMP
5008 if (dl_task(p) && !cpumask_intersects(task_rq(p)->rd->span,
5009 cs_cpus_allowed)) {
5010 unsigned int dest_cpu = cpumask_any_and(cpu_active_mask,
5011 cs_cpus_allowed);
Juri Lelli75e23e42014-10-28 11:54:46 +00005012 struct dl_bw *dl_b;
Juri Lelli7f514122014-09-19 10:22:40 +01005013 bool overflow;
5014 int cpus;
5015 unsigned long flags;
5016
Juri Lelli75e23e42014-10-28 11:54:46 +00005017 rcu_read_lock_sched();
5018 dl_b = dl_bw_of(dest_cpu);
Juri Lelli7f514122014-09-19 10:22:40 +01005019 raw_spin_lock_irqsave(&dl_b->lock, flags);
5020 cpus = dl_bw_cpus(dest_cpu);
5021 overflow = __dl_overflow(dl_b, cpus, 0, p->dl.dl_bw);
5022 if (overflow)
5023 ret = -EBUSY;
5024 else {
5025 /*
5026 * We reserve space for this task in the destination
5027 * root_domain, as we can't fail after this point.
5028 * We will free resources in the source root_domain
5029 * later on (see set_cpus_allowed_dl()).
5030 */
5031 __dl_add(dl_b, p->dl.dl_bw);
5032 }
5033 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
Juri Lelli75e23e42014-10-28 11:54:46 +00005034 rcu_read_unlock_sched();
Juri Lelli7f514122014-09-19 10:22:40 +01005035
5036 }
5037#endif
5038out:
5039 return ret;
5040}
5041
Linus Torvalds1da177e2005-04-16 15:20:36 -07005042#ifdef CONFIG_SMP
Linus Torvalds1da177e2005-04-16 15:20:36 -07005043
Mel Gormane6628d52013-10-07 11:29:02 +01005044#ifdef CONFIG_NUMA_BALANCING
5045/* Migrate current task p to target_cpu */
5046int migrate_task_to(struct task_struct *p, int target_cpu)
5047{
5048 struct migration_arg arg = { p, target_cpu };
5049 int curr_cpu = task_cpu(p);
5050
5051 if (curr_cpu == target_cpu)
5052 return 0;
5053
5054 if (!cpumask_test_cpu(target_cpu, tsk_cpus_allowed(p)))
5055 return -EINVAL;
5056
5057 /* TODO: This is not properly updating schedstats */
5058
Mel Gorman286549d2014-01-21 15:51:03 -08005059 trace_sched_move_numa(p, curr_cpu, target_cpu);
Mel Gormane6628d52013-10-07 11:29:02 +01005060 return stop_one_cpu(curr_cpu, migration_cpu_stop, &arg);
5061}
Peter Zijlstra0ec8aa02013-10-07 11:29:33 +01005062
5063/*
5064 * Requeue a task on a given node and accurately track the number of NUMA
5065 * tasks on the runqueues
5066 */
5067void sched_setnuma(struct task_struct *p, int nid)
5068{
5069 struct rq *rq;
5070 unsigned long flags;
Kirill Tkhaida0c1e62014-08-20 13:47:32 +04005071 bool queued, running;
Peter Zijlstra0ec8aa02013-10-07 11:29:33 +01005072
5073 rq = task_rq_lock(p, &flags);
Kirill Tkhaida0c1e62014-08-20 13:47:32 +04005074 queued = task_on_rq_queued(p);
Peter Zijlstra0ec8aa02013-10-07 11:29:33 +01005075 running = task_current(rq, p);
5076
Kirill Tkhaida0c1e62014-08-20 13:47:32 +04005077 if (queued)
Peter Zijlstra0ec8aa02013-10-07 11:29:33 +01005078 dequeue_task(rq, p, 0);
5079 if (running)
Kirill Tkhaif3cd1c42014-09-12 17:41:40 +04005080 put_prev_task(rq, p);
Peter Zijlstra0ec8aa02013-10-07 11:29:33 +01005081
5082 p->numa_preferred_nid = nid;
Peter Zijlstra0ec8aa02013-10-07 11:29:33 +01005083
5084 if (running)
5085 p->sched_class->set_curr_task(rq);
Kirill Tkhaida0c1e62014-08-20 13:47:32 +04005086 if (queued)
Peter Zijlstra0ec8aa02013-10-07 11:29:33 +01005087 enqueue_task(rq, p, 0);
5088 task_rq_unlock(rq, p, &flags);
5089}
Peter Zijlstra5cc389b2015-06-11 14:46:50 +02005090#endif /* CONFIG_NUMA_BALANCING */
Linus Torvalds1da177e2005-04-16 15:20:36 -07005091
5092#ifdef CONFIG_HOTPLUG_CPU
Ingo Molnar48f24c42006-07-03 00:25:40 -07005093/*
5094 * Ensures that the idle task is using init_mm right before its cpu goes
Linus Torvalds1da177e2005-04-16 15:20:36 -07005095 * offline.
5096 */
5097void idle_task_exit(void)
5098{
5099 struct mm_struct *mm = current->active_mm;
5100
5101 BUG_ON(cpu_online(smp_processor_id()));
5102
Martin Schwidefskya53efe52012-10-26 17:17:44 +02005103 if (mm != &init_mm) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07005104 switch_mm(mm, &init_mm, current);
Martin Schwidefskya53efe52012-10-26 17:17:44 +02005105 finish_arch_post_lock_switch();
5106 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07005107 mmdrop(mm);
5108}
5109
Peter Zijlstra48c5cca2010-11-13 19:32:29 +01005110/*
Peter Zijlstra5d180232012-08-20 11:26:57 +02005111 * Since this CPU is going 'away' for a while, fold any nr_active delta
5112 * we might have. Assumes we're called after migrate_tasks() so that the
5113 * nr_active count is stable.
5114 *
5115 * Also see the comment "Global load-average calculations".
Peter Zijlstra48c5cca2010-11-13 19:32:29 +01005116 */
Peter Zijlstra5d180232012-08-20 11:26:57 +02005117static void calc_load_migrate(struct rq *rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005118{
Peter Zijlstra5d180232012-08-20 11:26:57 +02005119 long delta = calc_load_fold_active(rq);
5120 if (delta)
5121 atomic_long_add(delta, &calc_load_tasks);
Thomas Gleixnerdce48a82009-04-11 10:43:41 +02005122}
Peter Zijlstra48c5cca2010-11-13 19:32:29 +01005123
Peter Zijlstra3f1d2a32014-02-12 10:49:30 +01005124static void put_prev_task_fake(struct rq *rq, struct task_struct *prev)
5125{
5126}
5127
5128static const struct sched_class fake_sched_class = {
5129 .put_prev_task = put_prev_task_fake,
5130};
5131
5132static struct task_struct fake_task = {
5133 /*
5134 * Avoid pull_{rt,dl}_task()
5135 */
5136 .prio = MAX_PRIO + 1,
5137 .sched_class = &fake_sched_class,
5138};
5139
Peter Zijlstra48c5cca2010-11-13 19:32:29 +01005140/*
5141 * Migrate all tasks from the rq, sleeping tasks will be migrated by
5142 * try_to_wake_up()->select_task_rq().
5143 *
5144 * Called with rq->lock held even though we'er in stop_machine() and
5145 * there's no concurrency possible, we hold the required locks anyway
5146 * because of lock validation efforts.
5147 */
Peter Zijlstra5e16bbc2015-06-11 14:46:51 +02005148static void migrate_tasks(struct rq *dead_rq)
Peter Zijlstra48c5cca2010-11-13 19:32:29 +01005149{
Peter Zijlstra5e16bbc2015-06-11 14:46:51 +02005150 struct rq *rq = dead_rq;
Peter Zijlstra48c5cca2010-11-13 19:32:29 +01005151 struct task_struct *next, *stop = rq->stop;
5152 int dest_cpu;
5153
5154 /*
5155 * Fudge the rq selection such that the below task selection loop
5156 * doesn't get stuck on the currently eligible stop task.
5157 *
5158 * We're currently inside stop_machine() and the rq is either stuck
5159 * in the stop_machine_cpu_stop() loop, or we're executing this code,
5160 * either way we should never end up calling schedule() until we're
5161 * done here.
5162 */
5163 rq->stop = NULL;
5164
Frederic Weisbecker77bd3972013-04-12 01:50:58 +02005165 /*
5166 * put_prev_task() and pick_next_task() sched
5167 * class method both need to have an up-to-date
5168 * value of rq->clock[_task]
5169 */
5170 update_rq_clock(rq);
5171
Peter Zijlstra5e16bbc2015-06-11 14:46:51 +02005172 for (;;) {
Peter Zijlstra48c5cca2010-11-13 19:32:29 +01005173 /*
5174 * There's this thread running, bail when that's the only
5175 * remaining thread.
5176 */
5177 if (rq->nr_running == 1)
5178 break;
5179
Peter Zijlstracbce1a62015-06-11 14:46:54 +02005180 /*
5181 * Ensure rq->lock covers the entire task selection
5182 * until the migration.
5183 */
5184 lockdep_pin_lock(&rq->lock);
Peter Zijlstra3f1d2a32014-02-12 10:49:30 +01005185 next = pick_next_task(rq, &fake_task);
Peter Zijlstra48c5cca2010-11-13 19:32:29 +01005186 BUG_ON(!next);
5187 next->sched_class->put_prev_task(rq, next);
5188
5189 /* Find suitable destination for @next, with force if needed. */
Peter Zijlstra5e16bbc2015-06-11 14:46:51 +02005190 dest_cpu = select_fallback_rq(dead_rq->cpu, next);
Peter Zijlstra48c5cca2010-11-13 19:32:29 +01005191
Peter Zijlstracbce1a62015-06-11 14:46:54 +02005192 lockdep_unpin_lock(&rq->lock);
Peter Zijlstra5e16bbc2015-06-11 14:46:51 +02005193 rq = __migrate_task(rq, next, dest_cpu);
5194 if (rq != dead_rq) {
5195 raw_spin_unlock(&rq->lock);
5196 rq = dead_rq;
5197 raw_spin_lock(&rq->lock);
5198 }
Peter Zijlstra48c5cca2010-11-13 19:32:29 +01005199 }
5200
5201 rq->stop = stop;
5202}
Linus Torvalds1da177e2005-04-16 15:20:36 -07005203#endif /* CONFIG_HOTPLUG_CPU */
5204
Nick Piggine692ab52007-07-26 13:40:43 +02005205#if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
5206
5207static struct ctl_table sd_ctl_dir[] = {
Alexey Dobriyane0361852007-08-09 11:16:46 +02005208 {
5209 .procname = "sched_domain",
Eric W. Biedermanc57baf12007-08-23 15:18:02 +02005210 .mode = 0555,
Alexey Dobriyane0361852007-08-09 11:16:46 +02005211 },
Eric W. Biederman56992302009-11-05 15:38:40 -08005212 {}
Nick Piggine692ab52007-07-26 13:40:43 +02005213};
5214
5215static struct ctl_table sd_ctl_root[] = {
Alexey Dobriyane0361852007-08-09 11:16:46 +02005216 {
5217 .procname = "kernel",
Eric W. Biedermanc57baf12007-08-23 15:18:02 +02005218 .mode = 0555,
Alexey Dobriyane0361852007-08-09 11:16:46 +02005219 .child = sd_ctl_dir,
5220 },
Eric W. Biederman56992302009-11-05 15:38:40 -08005221 {}
Nick Piggine692ab52007-07-26 13:40:43 +02005222};
5223
5224static struct ctl_table *sd_alloc_ctl_entry(int n)
5225{
5226 struct ctl_table *entry =
Milton Miller5cf9f062007-10-15 17:00:19 +02005227 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL);
Nick Piggine692ab52007-07-26 13:40:43 +02005228
Nick Piggine692ab52007-07-26 13:40:43 +02005229 return entry;
5230}
5231
Milton Miller6382bc92007-10-15 17:00:19 +02005232static void sd_free_ctl_entry(struct ctl_table **tablep)
5233{
Milton Millercd7900762007-10-17 16:55:11 +02005234 struct ctl_table *entry;
Milton Miller6382bc92007-10-15 17:00:19 +02005235
Milton Millercd7900762007-10-17 16:55:11 +02005236 /*
5237 * In the intermediate directories, both the child directory and
5238 * procname are dynamically allocated and could fail but the mode
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01005239 * will always be set. In the lowest directory the names are
Milton Millercd7900762007-10-17 16:55:11 +02005240 * static strings and all have proc handlers.
5241 */
5242 for (entry = *tablep; entry->mode; entry++) {
Milton Miller6382bc92007-10-15 17:00:19 +02005243 if (entry->child)
5244 sd_free_ctl_entry(&entry->child);
Milton Millercd7900762007-10-17 16:55:11 +02005245 if (entry->proc_handler == NULL)
5246 kfree(entry->procname);
5247 }
Milton Miller6382bc92007-10-15 17:00:19 +02005248
5249 kfree(*tablep);
5250 *tablep = NULL;
5251}
5252
Namhyung Kim201c3732012-08-16 17:03:24 +09005253static int min_load_idx = 0;
libinfd9b86d2013-04-08 14:39:12 +08005254static int max_load_idx = CPU_LOAD_IDX_MAX-1;
Namhyung Kim201c3732012-08-16 17:03:24 +09005255
Nick Piggine692ab52007-07-26 13:40:43 +02005256static void
Alexey Dobriyane0361852007-08-09 11:16:46 +02005257set_table_entry(struct ctl_table *entry,
Nick Piggine692ab52007-07-26 13:40:43 +02005258 const char *procname, void *data, int maxlen,
Namhyung Kim201c3732012-08-16 17:03:24 +09005259 umode_t mode, proc_handler *proc_handler,
5260 bool load_idx)
Nick Piggine692ab52007-07-26 13:40:43 +02005261{
Nick Piggine692ab52007-07-26 13:40:43 +02005262 entry->procname = procname;
5263 entry->data = data;
5264 entry->maxlen = maxlen;
5265 entry->mode = mode;
5266 entry->proc_handler = proc_handler;
Namhyung Kim201c3732012-08-16 17:03:24 +09005267
5268 if (load_idx) {
5269 entry->extra1 = &min_load_idx;
5270 entry->extra2 = &max_load_idx;
5271 }
Nick Piggine692ab52007-07-26 13:40:43 +02005272}
5273
5274static struct ctl_table *
5275sd_alloc_ctl_domain_table(struct sched_domain *sd)
5276{
Alex Shi37e6bae2014-01-23 18:39:54 +08005277 struct ctl_table *table = sd_alloc_ctl_entry(14);
Nick Piggine692ab52007-07-26 13:40:43 +02005278
Milton Millerad1cdc12007-10-15 17:00:19 +02005279 if (table == NULL)
5280 return NULL;
5281
Alexey Dobriyane0361852007-08-09 11:16:46 +02005282 set_table_entry(&table[0], "min_interval", &sd->min_interval,
Namhyung Kim201c3732012-08-16 17:03:24 +09005283 sizeof(long), 0644, proc_doulongvec_minmax, false);
Alexey Dobriyane0361852007-08-09 11:16:46 +02005284 set_table_entry(&table[1], "max_interval", &sd->max_interval,
Namhyung Kim201c3732012-08-16 17:03:24 +09005285 sizeof(long), 0644, proc_doulongvec_minmax, false);
Alexey Dobriyane0361852007-08-09 11:16:46 +02005286 set_table_entry(&table[2], "busy_idx", &sd->busy_idx,
Namhyung Kim201c3732012-08-16 17:03:24 +09005287 sizeof(int), 0644, proc_dointvec_minmax, true);
Alexey Dobriyane0361852007-08-09 11:16:46 +02005288 set_table_entry(&table[3], "idle_idx", &sd->idle_idx,
Namhyung Kim201c3732012-08-16 17:03:24 +09005289 sizeof(int), 0644, proc_dointvec_minmax, true);
Alexey Dobriyane0361852007-08-09 11:16:46 +02005290 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx,
Namhyung Kim201c3732012-08-16 17:03:24 +09005291 sizeof(int), 0644, proc_dointvec_minmax, true);
Alexey Dobriyane0361852007-08-09 11:16:46 +02005292 set_table_entry(&table[5], "wake_idx", &sd->wake_idx,
Namhyung Kim201c3732012-08-16 17:03:24 +09005293 sizeof(int), 0644, proc_dointvec_minmax, true);
Alexey Dobriyane0361852007-08-09 11:16:46 +02005294 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx,
Namhyung Kim201c3732012-08-16 17:03:24 +09005295 sizeof(int), 0644, proc_dointvec_minmax, true);
Alexey Dobriyane0361852007-08-09 11:16:46 +02005296 set_table_entry(&table[7], "busy_factor", &sd->busy_factor,
Namhyung Kim201c3732012-08-16 17:03:24 +09005297 sizeof(int), 0644, proc_dointvec_minmax, false);
Alexey Dobriyane0361852007-08-09 11:16:46 +02005298 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct,
Namhyung Kim201c3732012-08-16 17:03:24 +09005299 sizeof(int), 0644, proc_dointvec_minmax, false);
Zou Nan haiace8b3d2007-10-15 17:00:14 +02005300 set_table_entry(&table[9], "cache_nice_tries",
Nick Piggine692ab52007-07-26 13:40:43 +02005301 &sd->cache_nice_tries,
Namhyung Kim201c3732012-08-16 17:03:24 +09005302 sizeof(int), 0644, proc_dointvec_minmax, false);
Zou Nan haiace8b3d2007-10-15 17:00:14 +02005303 set_table_entry(&table[10], "flags", &sd->flags,
Namhyung Kim201c3732012-08-16 17:03:24 +09005304 sizeof(int), 0644, proc_dointvec_minmax, false);
Alex Shi37e6bae2014-01-23 18:39:54 +08005305 set_table_entry(&table[11], "max_newidle_lb_cost",
5306 &sd->max_newidle_lb_cost,
5307 sizeof(long), 0644, proc_doulongvec_minmax, false);
5308 set_table_entry(&table[12], "name", sd->name,
Namhyung Kim201c3732012-08-16 17:03:24 +09005309 CORENAME_MAX_SIZE, 0444, proc_dostring, false);
Alex Shi37e6bae2014-01-23 18:39:54 +08005310 /* &table[13] is terminator */
Nick Piggine692ab52007-07-26 13:40:43 +02005311
5312 return table;
5313}
5314
Joe Perchesbe7002e2013-06-12 11:55:36 -07005315static struct ctl_table *sd_alloc_ctl_cpu_table(int cpu)
Nick Piggine692ab52007-07-26 13:40:43 +02005316{
5317 struct ctl_table *entry, *table;
5318 struct sched_domain *sd;
5319 int domain_num = 0, i;
5320 char buf[32];
5321
5322 for_each_domain(cpu, sd)
5323 domain_num++;
5324 entry = table = sd_alloc_ctl_entry(domain_num + 1);
Milton Millerad1cdc12007-10-15 17:00:19 +02005325 if (table == NULL)
5326 return NULL;
Nick Piggine692ab52007-07-26 13:40:43 +02005327
5328 i = 0;
5329 for_each_domain(cpu, sd) {
5330 snprintf(buf, 32, "domain%d", i);
Nick Piggine692ab52007-07-26 13:40:43 +02005331 entry->procname = kstrdup(buf, GFP_KERNEL);
Eric W. Biedermanc57baf12007-08-23 15:18:02 +02005332 entry->mode = 0555;
Nick Piggine692ab52007-07-26 13:40:43 +02005333 entry->child = sd_alloc_ctl_domain_table(sd);
5334 entry++;
5335 i++;
5336 }
5337 return table;
5338}
5339
5340static struct ctl_table_header *sd_sysctl_header;
Milton Miller6382bc92007-10-15 17:00:19 +02005341static void register_sched_domain_sysctl(void)
Nick Piggine692ab52007-07-26 13:40:43 +02005342{
Peter Zijlstra6ad4c182009-11-25 13:31:39 +01005343 int i, cpu_num = num_possible_cpus();
Nick Piggine692ab52007-07-26 13:40:43 +02005344 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
5345 char buf[32];
5346
Milton Miller73785472007-10-24 18:23:48 +02005347 WARN_ON(sd_ctl_dir[0].child);
5348 sd_ctl_dir[0].child = entry;
5349
Milton Millerad1cdc12007-10-15 17:00:19 +02005350 if (entry == NULL)
5351 return;
5352
Peter Zijlstra6ad4c182009-11-25 13:31:39 +01005353 for_each_possible_cpu(i) {
Nick Piggine692ab52007-07-26 13:40:43 +02005354 snprintf(buf, 32, "cpu%d", i);
Nick Piggine692ab52007-07-26 13:40:43 +02005355 entry->procname = kstrdup(buf, GFP_KERNEL);
Eric W. Biedermanc57baf12007-08-23 15:18:02 +02005356 entry->mode = 0555;
Nick Piggine692ab52007-07-26 13:40:43 +02005357 entry->child = sd_alloc_ctl_cpu_table(i);
Milton Miller97b6ea72007-10-15 17:00:19 +02005358 entry++;
Nick Piggine692ab52007-07-26 13:40:43 +02005359 }
Milton Miller73785472007-10-24 18:23:48 +02005360
5361 WARN_ON(sd_sysctl_header);
Nick Piggine692ab52007-07-26 13:40:43 +02005362 sd_sysctl_header = register_sysctl_table(sd_ctl_root);
5363}
Milton Miller6382bc92007-10-15 17:00:19 +02005364
Milton Miller73785472007-10-24 18:23:48 +02005365/* may be called multiple times per register */
Milton Miller6382bc92007-10-15 17:00:19 +02005366static void unregister_sched_domain_sysctl(void)
5367{
Markus Elfring781b0202015-07-04 09:06:32 +02005368 unregister_sysctl_table(sd_sysctl_header);
Milton Miller6382bc92007-10-15 17:00:19 +02005369 sd_sysctl_header = NULL;
Milton Miller73785472007-10-24 18:23:48 +02005370 if (sd_ctl_dir[0].child)
5371 sd_free_ctl_entry(&sd_ctl_dir[0].child);
Milton Miller6382bc92007-10-15 17:00:19 +02005372}
Nick Piggine692ab52007-07-26 13:40:43 +02005373#else
Milton Miller6382bc92007-10-15 17:00:19 +02005374static void register_sched_domain_sysctl(void)
5375{
5376}
5377static void unregister_sched_domain_sysctl(void)
Nick Piggine692ab52007-07-26 13:40:43 +02005378{
5379}
Peter Zijlstra5cc389b2015-06-11 14:46:50 +02005380#endif /* CONFIG_SCHED_DEBUG && CONFIG_SYSCTL */
Nick Piggine692ab52007-07-26 13:40:43 +02005381
Gregory Haskins1f11eb62008-06-04 15:04:05 -04005382static void set_rq_online(struct rq *rq)
5383{
5384 if (!rq->online) {
5385 const struct sched_class *class;
5386
Rusty Russellc6c49272008-11-25 02:35:05 +10305387 cpumask_set_cpu(rq->cpu, rq->rd->online);
Gregory Haskins1f11eb62008-06-04 15:04:05 -04005388 rq->online = 1;
5389
5390 for_each_class(class) {
5391 if (class->rq_online)
5392 class->rq_online(rq);
5393 }
5394 }
5395}
5396
5397static void set_rq_offline(struct rq *rq)
5398{
5399 if (rq->online) {
5400 const struct sched_class *class;
5401
5402 for_each_class(class) {
5403 if (class->rq_offline)
5404 class->rq_offline(rq);
5405 }
5406
Rusty Russellc6c49272008-11-25 02:35:05 +10305407 cpumask_clear_cpu(rq->cpu, rq->rd->online);
Gregory Haskins1f11eb62008-06-04 15:04:05 -04005408 rq->online = 0;
5409 }
5410}
5411
Linus Torvalds1da177e2005-04-16 15:20:36 -07005412/*
5413 * migration_call - callback that gets triggered when a CPU is added.
5414 * Here we can start up the necessary migration thread for the new CPU.
5415 */
Paul Gortmaker0db06282013-06-19 14:53:51 -04005416static int
Ingo Molnar48f24c42006-07-03 00:25:40 -07005417migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005418{
Ingo Molnar48f24c42006-07-03 00:25:40 -07005419 int cpu = (long)hcpu;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005420 unsigned long flags;
Tejun Heo969c7922010-05-06 18:49:21 +02005421 struct rq *rq = cpu_rq(cpu);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005422
Peter Zijlstra48c5cca2010-11-13 19:32:29 +01005423 switch (action & ~CPU_TASKS_FROZEN) {
Gautham R Shenoy5be93612007-05-09 02:34:04 -07005424
Linus Torvalds1da177e2005-04-16 15:20:36 -07005425 case CPU_UP_PREPARE:
Thomas Gleixnera468d382009-07-17 14:15:46 +02005426 rq->calc_load_update = calc_load_update;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005427 break;
Ingo Molnar48f24c42006-07-03 00:25:40 -07005428
Linus Torvalds1da177e2005-04-16 15:20:36 -07005429 case CPU_ONLINE:
Gregory Haskins1f94ef52008-03-10 16:52:41 -04005430 /* Update our root-domain */
Thomas Gleixner05fa7852009-11-17 14:28:38 +01005431 raw_spin_lock_irqsave(&rq->lock, flags);
Gregory Haskins1f94ef52008-03-10 16:52:41 -04005432 if (rq->rd) {
Rusty Russellc6c49272008-11-25 02:35:05 +10305433 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
Gregory Haskins1f11eb62008-06-04 15:04:05 -04005434
5435 set_rq_online(rq);
Gregory Haskins1f94ef52008-03-10 16:52:41 -04005436 }
Thomas Gleixner05fa7852009-11-17 14:28:38 +01005437 raw_spin_unlock_irqrestore(&rq->lock, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005438 break;
Ingo Molnar48f24c42006-07-03 00:25:40 -07005439
Linus Torvalds1da177e2005-04-16 15:20:36 -07005440#ifdef CONFIG_HOTPLUG_CPU
Gregory Haskins08f503b2008-03-10 17:59:11 -04005441 case CPU_DYING:
Peter Zijlstra317f3942011-04-05 17:23:58 +02005442 sched_ttwu_pending();
Gregory Haskins57d885f2008-01-25 21:08:18 +01005443 /* Update our root-domain */
Thomas Gleixner05fa7852009-11-17 14:28:38 +01005444 raw_spin_lock_irqsave(&rq->lock, flags);
Gregory Haskins57d885f2008-01-25 21:08:18 +01005445 if (rq->rd) {
Rusty Russellc6c49272008-11-25 02:35:05 +10305446 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
Gregory Haskins1f11eb62008-06-04 15:04:05 -04005447 set_rq_offline(rq);
Gregory Haskins57d885f2008-01-25 21:08:18 +01005448 }
Peter Zijlstra5e16bbc2015-06-11 14:46:51 +02005449 migrate_tasks(rq);
Peter Zijlstra48c5cca2010-11-13 19:32:29 +01005450 BUG_ON(rq->nr_running != 1); /* the migration thread */
Thomas Gleixner05fa7852009-11-17 14:28:38 +01005451 raw_spin_unlock_irqrestore(&rq->lock, flags);
Peter Zijlstra5d180232012-08-20 11:26:57 +02005452 break;
Peter Zijlstra48c5cca2010-11-13 19:32:29 +01005453
Peter Zijlstra5d180232012-08-20 11:26:57 +02005454 case CPU_DEAD:
Peter Zijlstraf319da02012-08-20 11:26:57 +02005455 calc_load_migrate(rq);
Gregory Haskins57d885f2008-01-25 21:08:18 +01005456 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005457#endif
5458 }
Peter Zijlstra49c022e2011-04-05 10:14:25 +02005459
5460 update_max_interval();
5461
Linus Torvalds1da177e2005-04-16 15:20:36 -07005462 return NOTIFY_OK;
5463}
5464
Paul Mackerrasf38b0822009-06-02 21:05:16 +10005465/*
5466 * Register at high priority so that task migration (migrate_all_tasks)
5467 * happens before everything else. This has to be lower priority than
Ingo Molnarcdd6c482009-09-21 12:02:48 +02005468 * the notifier in the perf_event subsystem, though.
Linus Torvalds1da177e2005-04-16 15:20:36 -07005469 */
Paul Gortmaker0db06282013-06-19 14:53:51 -04005470static struct notifier_block migration_notifier = {
Linus Torvalds1da177e2005-04-16 15:20:36 -07005471 .notifier_call = migration_call,
Tejun Heo50a323b2010-06-08 21:40:36 +02005472 .priority = CPU_PRI_MIGRATION,
Linus Torvalds1da177e2005-04-16 15:20:36 -07005473};
5474
Paul Gortmaker6a82b602015-04-27 18:47:50 -04005475static void set_cpu_rq_start_time(void)
Corey Minyarda803f022014-05-08 13:47:39 -05005476{
5477 int cpu = smp_processor_id();
5478 struct rq *rq = cpu_rq(cpu);
5479 rq->age_stamp = sched_clock_cpu(cpu);
5480}
5481
Paul Gortmaker0db06282013-06-19 14:53:51 -04005482static int sched_cpu_active(struct notifier_block *nfb,
Tejun Heo3a101d02010-06-08 21:40:36 +02005483 unsigned long action, void *hcpu)
5484{
5485 switch (action & ~CPU_TASKS_FROZEN) {
Corey Minyarda803f022014-05-08 13:47:39 -05005486 case CPU_STARTING:
5487 set_cpu_rq_start_time();
5488 return NOTIFY_OK;
Jan H. Schönherrdd9d3842015-08-12 21:35:56 +02005489 case CPU_ONLINE:
5490 /*
5491 * At this point a starting CPU has marked itself as online via
5492 * set_cpu_online(). But it might not yet have marked itself
5493 * as active, which is essential from here on.
5494 *
5495 * Thus, fall-through and help the starting CPU along.
5496 */
Tejun Heo3a101d02010-06-08 21:40:36 +02005497 case CPU_DOWN_FAILED:
5498 set_cpu_active((long)hcpu, true);
5499 return NOTIFY_OK;
5500 default:
5501 return NOTIFY_DONE;
5502 }
5503}
5504
Paul Gortmaker0db06282013-06-19 14:53:51 -04005505static int sched_cpu_inactive(struct notifier_block *nfb,
Tejun Heo3a101d02010-06-08 21:40:36 +02005506 unsigned long action, void *hcpu)
5507{
5508 switch (action & ~CPU_TASKS_FROZEN) {
5509 case CPU_DOWN_PREPARE:
Juri Lelli3c18d442015-03-31 09:53:37 +01005510 set_cpu_active((long)hcpu, false);
Tejun Heo3a101d02010-06-08 21:40:36 +02005511 return NOTIFY_OK;
Juri Lelli3c18d442015-03-31 09:53:37 +01005512 default:
5513 return NOTIFY_DONE;
Tejun Heo3a101d02010-06-08 21:40:36 +02005514 }
5515}
5516
Eduard - Gabriel Munteanu7babe8d2008-07-25 19:45:11 -07005517static int __init migration_init(void)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005518{
5519 void *cpu = (void *)(long)smp_processor_id();
Akinobu Mita07dccf32006-09-29 02:00:22 -07005520 int err;
Ingo Molnar48f24c42006-07-03 00:25:40 -07005521
Tejun Heo3a101d02010-06-08 21:40:36 +02005522 /* Initialize migration for the boot CPU */
Akinobu Mita07dccf32006-09-29 02:00:22 -07005523 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
5524 BUG_ON(err == NOTIFY_BAD);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005525 migration_call(&migration_notifier, CPU_ONLINE, cpu);
5526 register_cpu_notifier(&migration_notifier);
Eduard - Gabriel Munteanu7babe8d2008-07-25 19:45:11 -07005527
Tejun Heo3a101d02010-06-08 21:40:36 +02005528 /* Register cpu active notifiers */
5529 cpu_notifier(sched_cpu_active, CPU_PRI_SCHED_ACTIVE);
5530 cpu_notifier(sched_cpu_inactive, CPU_PRI_SCHED_INACTIVE);
5531
Thomas Gleixnera004cd42009-07-21 09:54:05 +02005532 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005533}
Eduard - Gabriel Munteanu7babe8d2008-07-25 19:45:11 -07005534early_initcall(migration_init);
Christoph Lameter476f3532007-05-06 14:48:58 -07005535
Peter Zijlstra4cb98832011-04-07 14:09:58 +02005536static cpumask_var_t sched_domains_tmpmask; /* sched_domains_mutex */
5537
Ingo Molnar3e9830d2007-10-15 17:00:13 +02005538#ifdef CONFIG_SCHED_DEBUG
Ingo Molnar4dcf6af2007-10-24 18:23:48 +02005539
Peter Zijlstrad039ac62012-05-31 21:20:16 +02005540static __read_mostly int sched_debug_enabled;
Mike Travisf6630112009-11-17 18:22:15 -06005541
Peter Zijlstrad039ac62012-05-31 21:20:16 +02005542static int __init sched_debug_setup(char *str)
Mike Travisf6630112009-11-17 18:22:15 -06005543{
Peter Zijlstrad039ac62012-05-31 21:20:16 +02005544 sched_debug_enabled = 1;
Mike Travisf6630112009-11-17 18:22:15 -06005545
5546 return 0;
5547}
Peter Zijlstrad039ac62012-05-31 21:20:16 +02005548early_param("sched_debug", sched_debug_setup);
5549
5550static inline bool sched_debug(void)
5551{
5552 return sched_debug_enabled;
5553}
Mike Travisf6630112009-11-17 18:22:15 -06005554
Mike Travis7c16ec52008-04-04 18:11:11 -07005555static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level,
Rusty Russell96f874e2008-11-25 02:35:14 +10305556 struct cpumask *groupmask)
Ingo Molnar4dcf6af2007-10-24 18:23:48 +02005557{
5558 struct sched_group *group = sd->groups;
Ingo Molnar4dcf6af2007-10-24 18:23:48 +02005559
Rusty Russell96f874e2008-11-25 02:35:14 +10305560 cpumask_clear(groupmask);
Ingo Molnar4dcf6af2007-10-24 18:23:48 +02005561
5562 printk(KERN_DEBUG "%*s domain %d: ", level, "", level);
5563
5564 if (!(sd->flags & SD_LOAD_BALANCE)) {
Peter Zijlstra3df0fc52009-12-20 14:23:57 +01005565 printk("does not load-balance\n");
Ingo Molnar4dcf6af2007-10-24 18:23:48 +02005566 if (sd->parent)
Peter Zijlstra3df0fc52009-12-20 14:23:57 +01005567 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
5568 " has parent");
Ingo Molnar4dcf6af2007-10-24 18:23:48 +02005569 return -1;
5570 }
5571
Tejun Heo333470e2015-02-13 14:37:28 -08005572 printk(KERN_CONT "span %*pbl level %s\n",
5573 cpumask_pr_args(sched_domain_span(sd)), sd->name);
Ingo Molnar4dcf6af2007-10-24 18:23:48 +02005574
Rusty Russell758b2cd2008-11-25 02:35:04 +10305575 if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) {
Peter Zijlstra3df0fc52009-12-20 14:23:57 +01005576 printk(KERN_ERR "ERROR: domain->span does not contain "
5577 "CPU%d\n", cpu);
Ingo Molnar4dcf6af2007-10-24 18:23:48 +02005578 }
Rusty Russell758b2cd2008-11-25 02:35:04 +10305579 if (!cpumask_test_cpu(cpu, sched_group_cpus(group))) {
Peter Zijlstra3df0fc52009-12-20 14:23:57 +01005580 printk(KERN_ERR "ERROR: domain->groups does not contain"
5581 " CPU%d\n", cpu);
Ingo Molnar4dcf6af2007-10-24 18:23:48 +02005582 }
5583
5584 printk(KERN_DEBUG "%*s groups:", level + 1, "");
5585 do {
5586 if (!group) {
Peter Zijlstra3df0fc52009-12-20 14:23:57 +01005587 printk("\n");
5588 printk(KERN_ERR "ERROR: group is NULL\n");
Ingo Molnar4dcf6af2007-10-24 18:23:48 +02005589 break;
5590 }
5591
Rusty Russell758b2cd2008-11-25 02:35:04 +10305592 if (!cpumask_weight(sched_group_cpus(group))) {
Peter Zijlstra3df0fc52009-12-20 14:23:57 +01005593 printk(KERN_CONT "\n");
5594 printk(KERN_ERR "ERROR: empty group\n");
Ingo Molnar4dcf6af2007-10-24 18:23:48 +02005595 break;
5596 }
5597
Peter Zijlstracb83b622012-04-17 15:49:36 +02005598 if (!(sd->flags & SD_OVERLAP) &&
5599 cpumask_intersects(groupmask, sched_group_cpus(group))) {
Peter Zijlstra3df0fc52009-12-20 14:23:57 +01005600 printk(KERN_CONT "\n");
5601 printk(KERN_ERR "ERROR: repeated CPUs\n");
Ingo Molnar4dcf6af2007-10-24 18:23:48 +02005602 break;
5603 }
5604
Rusty Russell758b2cd2008-11-25 02:35:04 +10305605 cpumask_or(groupmask, groupmask, sched_group_cpus(group));
Ingo Molnar4dcf6af2007-10-24 18:23:48 +02005606
Tejun Heo333470e2015-02-13 14:37:28 -08005607 printk(KERN_CONT " %*pbl",
5608 cpumask_pr_args(sched_group_cpus(group)));
Nicolas Pitreca8ce3d2014-05-26 18:19:39 -04005609 if (group->sgc->capacity != SCHED_CAPACITY_SCALE) {
Nicolas Pitre63b2ca32014-05-26 18:19:37 -04005610 printk(KERN_CONT " (cpu_capacity = %d)",
5611 group->sgc->capacity);
Gautham R Shenoy381512c2009-04-14 09:09:36 +05305612 }
Ingo Molnar4dcf6af2007-10-24 18:23:48 +02005613
5614 group = group->next;
5615 } while (group != sd->groups);
Peter Zijlstra3df0fc52009-12-20 14:23:57 +01005616 printk(KERN_CONT "\n");
Ingo Molnar4dcf6af2007-10-24 18:23:48 +02005617
Rusty Russell758b2cd2008-11-25 02:35:04 +10305618 if (!cpumask_equal(sched_domain_span(sd), groupmask))
Peter Zijlstra3df0fc52009-12-20 14:23:57 +01005619 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
Ingo Molnar4dcf6af2007-10-24 18:23:48 +02005620
Rusty Russell758b2cd2008-11-25 02:35:04 +10305621 if (sd->parent &&
5622 !cpumask_subset(groupmask, sched_domain_span(sd->parent)))
Peter Zijlstra3df0fc52009-12-20 14:23:57 +01005623 printk(KERN_ERR "ERROR: parent span is not a superset "
5624 "of domain->span\n");
Ingo Molnar4dcf6af2007-10-24 18:23:48 +02005625 return 0;
5626}
5627
Linus Torvalds1da177e2005-04-16 15:20:36 -07005628static void sched_domain_debug(struct sched_domain *sd, int cpu)
5629{
5630 int level = 0;
5631
Peter Zijlstrad039ac62012-05-31 21:20:16 +02005632 if (!sched_debug_enabled)
Mike Travisf6630112009-11-17 18:22:15 -06005633 return;
5634
Nick Piggin41c7ce92005-06-25 14:57:24 -07005635 if (!sd) {
5636 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
5637 return;
5638 }
5639
Linus Torvalds1da177e2005-04-16 15:20:36 -07005640 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
5641
Ingo Molnar4dcf6af2007-10-24 18:23:48 +02005642 for (;;) {
Peter Zijlstra4cb98832011-04-07 14:09:58 +02005643 if (sched_domain_debug_one(sd, cpu, level, sched_domains_tmpmask))
Linus Torvalds1da177e2005-04-16 15:20:36 -07005644 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005645 level++;
5646 sd = sd->parent;
Miguel Ojeda Sandonis33859f72006-12-10 02:20:38 -08005647 if (!sd)
Ingo Molnar4dcf6af2007-10-24 18:23:48 +02005648 break;
5649 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07005650}
Dhaval Giani6d6bc0a2008-05-30 14:23:45 +02005651#else /* !CONFIG_SCHED_DEBUG */
Ingo Molnar48f24c42006-07-03 00:25:40 -07005652# define sched_domain_debug(sd, cpu) do { } while (0)
Peter Zijlstrad039ac62012-05-31 21:20:16 +02005653static inline bool sched_debug(void)
5654{
5655 return false;
5656}
Dhaval Giani6d6bc0a2008-05-30 14:23:45 +02005657#endif /* CONFIG_SCHED_DEBUG */
Linus Torvalds1da177e2005-04-16 15:20:36 -07005658
Dinakar Guniguntala1a20ff22005-06-25 14:57:33 -07005659static int sd_degenerate(struct sched_domain *sd)
Suresh Siddha245af2c2005-06-25 14:57:25 -07005660{
Rusty Russell758b2cd2008-11-25 02:35:04 +10305661 if (cpumask_weight(sched_domain_span(sd)) == 1)
Suresh Siddha245af2c2005-06-25 14:57:25 -07005662 return 1;
5663
5664 /* Following flags need at least 2 groups */
5665 if (sd->flags & (SD_LOAD_BALANCE |
5666 SD_BALANCE_NEWIDLE |
5667 SD_BALANCE_FORK |
Siddha, Suresh B89c47102006-10-03 01:14:09 -07005668 SD_BALANCE_EXEC |
Nicolas Pitre5d4dfdd2014-05-27 13:50:41 -04005669 SD_SHARE_CPUCAPACITY |
Vincent Guittotd77b3ed2014-04-11 11:44:40 +02005670 SD_SHARE_PKG_RESOURCES |
5671 SD_SHARE_POWERDOMAIN)) {
Suresh Siddha245af2c2005-06-25 14:57:25 -07005672 if (sd->groups != sd->groups->next)
5673 return 0;
5674 }
5675
5676 /* Following flags don't use groups */
Peter Zijlstrac88d5912009-09-10 13:50:02 +02005677 if (sd->flags & (SD_WAKE_AFFINE))
Suresh Siddha245af2c2005-06-25 14:57:25 -07005678 return 0;
5679
5680 return 1;
5681}
5682
Ingo Molnar48f24c42006-07-03 00:25:40 -07005683static int
5684sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
Suresh Siddha245af2c2005-06-25 14:57:25 -07005685{
5686 unsigned long cflags = sd->flags, pflags = parent->flags;
5687
5688 if (sd_degenerate(parent))
5689 return 1;
5690
Rusty Russell758b2cd2008-11-25 02:35:04 +10305691 if (!cpumask_equal(sched_domain_span(sd), sched_domain_span(parent)))
Suresh Siddha245af2c2005-06-25 14:57:25 -07005692 return 0;
5693
Suresh Siddha245af2c2005-06-25 14:57:25 -07005694 /* Flags needing groups don't count if only 1 group in parent */
5695 if (parent->groups == parent->groups->next) {
5696 pflags &= ~(SD_LOAD_BALANCE |
5697 SD_BALANCE_NEWIDLE |
5698 SD_BALANCE_FORK |
Siddha, Suresh B89c47102006-10-03 01:14:09 -07005699 SD_BALANCE_EXEC |
Nicolas Pitre5d4dfdd2014-05-27 13:50:41 -04005700 SD_SHARE_CPUCAPACITY |
Peter Zijlstra10866e622013-08-19 16:57:04 +02005701 SD_SHARE_PKG_RESOURCES |
Vincent Guittotd77b3ed2014-04-11 11:44:40 +02005702 SD_PREFER_SIBLING |
5703 SD_SHARE_POWERDOMAIN);
Ken Chen54364992008-12-07 18:47:37 -08005704 if (nr_node_ids == 1)
5705 pflags &= ~SD_SERIALIZE;
Suresh Siddha245af2c2005-06-25 14:57:25 -07005706 }
5707 if (~cflags & pflags)
5708 return 0;
5709
5710 return 1;
5711}
5712
Peter Zijlstradce840a2011-04-07 14:09:50 +02005713static void free_rootdomain(struct rcu_head *rcu)
Rusty Russellc6c49272008-11-25 02:35:05 +10305714{
Peter Zijlstradce840a2011-04-07 14:09:50 +02005715 struct root_domain *rd = container_of(rcu, struct root_domain, rcu);
Peter Zijlstra047106a2009-11-16 10:28:09 +01005716
Rusty Russell68e74562008-11-25 02:35:13 +10305717 cpupri_cleanup(&rd->cpupri);
Juri Lelli6bfd6d72013-11-07 14:43:47 +01005718 cpudl_cleanup(&rd->cpudl);
Juri Lelli1baca4c2013-11-07 14:43:38 +01005719 free_cpumask_var(rd->dlo_mask);
Rusty Russellc6c49272008-11-25 02:35:05 +10305720 free_cpumask_var(rd->rto_mask);
5721 free_cpumask_var(rd->online);
5722 free_cpumask_var(rd->span);
5723 kfree(rd);
5724}
5725
Gregory Haskins57d885f2008-01-25 21:08:18 +01005726static void rq_attach_root(struct rq *rq, struct root_domain *rd)
5727{
Ingo Molnara0490fa2009-02-12 11:35:40 +01005728 struct root_domain *old_rd = NULL;
Gregory Haskins57d885f2008-01-25 21:08:18 +01005729 unsigned long flags;
Gregory Haskins57d885f2008-01-25 21:08:18 +01005730
Thomas Gleixner05fa7852009-11-17 14:28:38 +01005731 raw_spin_lock_irqsave(&rq->lock, flags);
Gregory Haskins57d885f2008-01-25 21:08:18 +01005732
5733 if (rq->rd) {
Ingo Molnara0490fa2009-02-12 11:35:40 +01005734 old_rd = rq->rd;
Gregory Haskins57d885f2008-01-25 21:08:18 +01005735
Rusty Russellc6c49272008-11-25 02:35:05 +10305736 if (cpumask_test_cpu(rq->cpu, old_rd->online))
Gregory Haskins1f11eb62008-06-04 15:04:05 -04005737 set_rq_offline(rq);
Gregory Haskins57d885f2008-01-25 21:08:18 +01005738
Rusty Russellc6c49272008-11-25 02:35:05 +10305739 cpumask_clear_cpu(rq->cpu, old_rd->span);
Gregory Haskinsdc938522008-01-25 21:08:26 +01005740
Ingo Molnara0490fa2009-02-12 11:35:40 +01005741 /*
Shigeru Yoshida05159732013-11-17 12:12:36 +09005742 * If we dont want to free the old_rd yet then
Ingo Molnara0490fa2009-02-12 11:35:40 +01005743 * set old_rd to NULL to skip the freeing later
5744 * in this function:
5745 */
5746 if (!atomic_dec_and_test(&old_rd->refcount))
5747 old_rd = NULL;
Gregory Haskins57d885f2008-01-25 21:08:18 +01005748 }
5749
5750 atomic_inc(&rd->refcount);
5751 rq->rd = rd;
5752
Rusty Russellc6c49272008-11-25 02:35:05 +10305753 cpumask_set_cpu(rq->cpu, rd->span);
Gregory Haskins00aec932009-07-30 10:57:23 -04005754 if (cpumask_test_cpu(rq->cpu, cpu_active_mask))
Gregory Haskins1f11eb62008-06-04 15:04:05 -04005755 set_rq_online(rq);
Gregory Haskins57d885f2008-01-25 21:08:18 +01005756
Thomas Gleixner05fa7852009-11-17 14:28:38 +01005757 raw_spin_unlock_irqrestore(&rq->lock, flags);
Ingo Molnara0490fa2009-02-12 11:35:40 +01005758
5759 if (old_rd)
Peter Zijlstradce840a2011-04-07 14:09:50 +02005760 call_rcu_sched(&old_rd->rcu, free_rootdomain);
Gregory Haskins57d885f2008-01-25 21:08:18 +01005761}
5762
Pekka Enberg68c38fc2010-07-15 23:18:22 +03005763static int init_rootdomain(struct root_domain *rd)
Gregory Haskins57d885f2008-01-25 21:08:18 +01005764{
5765 memset(rd, 0, sizeof(*rd));
5766
Pekka Enberg68c38fc2010-07-15 23:18:22 +03005767 if (!alloc_cpumask_var(&rd->span, GFP_KERNEL))
Li Zefan0c910d22009-01-06 17:39:06 +08005768 goto out;
Pekka Enberg68c38fc2010-07-15 23:18:22 +03005769 if (!alloc_cpumask_var(&rd->online, GFP_KERNEL))
Rusty Russellc6c49272008-11-25 02:35:05 +10305770 goto free_span;
Juri Lelli1baca4c2013-11-07 14:43:38 +01005771 if (!alloc_cpumask_var(&rd->dlo_mask, GFP_KERNEL))
Rusty Russellc6c49272008-11-25 02:35:05 +10305772 goto free_online;
Juri Lelli1baca4c2013-11-07 14:43:38 +01005773 if (!alloc_cpumask_var(&rd->rto_mask, GFP_KERNEL))
5774 goto free_dlo_mask;
Gregory Haskins6e0534f2008-05-12 21:21:01 +02005775
Dario Faggioli332ac172013-11-07 14:43:45 +01005776 init_dl_bw(&rd->dl_bw);
Juri Lelli6bfd6d72013-11-07 14:43:47 +01005777 if (cpudl_init(&rd->cpudl) != 0)
5778 goto free_dlo_mask;
Dario Faggioli332ac172013-11-07 14:43:45 +01005779
Pekka Enberg68c38fc2010-07-15 23:18:22 +03005780 if (cpupri_init(&rd->cpupri) != 0)
Rusty Russell68e74562008-11-25 02:35:13 +10305781 goto free_rto_mask;
Rusty Russellc6c49272008-11-25 02:35:05 +10305782 return 0;
5783
Rusty Russell68e74562008-11-25 02:35:13 +10305784free_rto_mask:
5785 free_cpumask_var(rd->rto_mask);
Juri Lelli1baca4c2013-11-07 14:43:38 +01005786free_dlo_mask:
5787 free_cpumask_var(rd->dlo_mask);
Rusty Russellc6c49272008-11-25 02:35:05 +10305788free_online:
5789 free_cpumask_var(rd->online);
5790free_span:
5791 free_cpumask_var(rd->span);
Li Zefan0c910d22009-01-06 17:39:06 +08005792out:
Rusty Russellc6c49272008-11-25 02:35:05 +10305793 return -ENOMEM;
Gregory Haskins57d885f2008-01-25 21:08:18 +01005794}
5795
Peter Zijlstra029632f2011-10-25 10:00:11 +02005796/*
5797 * By default the system creates a single root-domain with all cpus as
5798 * members (mimicking the global state we have today).
5799 */
5800struct root_domain def_root_domain;
5801
Gregory Haskins57d885f2008-01-25 21:08:18 +01005802static void init_defrootdomain(void)
5803{
Pekka Enberg68c38fc2010-07-15 23:18:22 +03005804 init_rootdomain(&def_root_domain);
Rusty Russellc6c49272008-11-25 02:35:05 +10305805
Gregory Haskins57d885f2008-01-25 21:08:18 +01005806 atomic_set(&def_root_domain.refcount, 1);
5807}
5808
Gregory Haskinsdc938522008-01-25 21:08:26 +01005809static struct root_domain *alloc_rootdomain(void)
Gregory Haskins57d885f2008-01-25 21:08:18 +01005810{
5811 struct root_domain *rd;
5812
5813 rd = kmalloc(sizeof(*rd), GFP_KERNEL);
5814 if (!rd)
5815 return NULL;
5816
Pekka Enberg68c38fc2010-07-15 23:18:22 +03005817 if (init_rootdomain(rd) != 0) {
Rusty Russellc6c49272008-11-25 02:35:05 +10305818 kfree(rd);
5819 return NULL;
5820 }
Gregory Haskins57d885f2008-01-25 21:08:18 +01005821
5822 return rd;
5823}
5824
Nicolas Pitre63b2ca32014-05-26 18:19:37 -04005825static void free_sched_groups(struct sched_group *sg, int free_sgc)
Peter Zijlstrae3589f62011-07-15 10:35:52 +02005826{
5827 struct sched_group *tmp, *first;
5828
5829 if (!sg)
5830 return;
5831
5832 first = sg;
5833 do {
5834 tmp = sg->next;
5835
Nicolas Pitre63b2ca32014-05-26 18:19:37 -04005836 if (free_sgc && atomic_dec_and_test(&sg->sgc->ref))
5837 kfree(sg->sgc);
Peter Zijlstrae3589f62011-07-15 10:35:52 +02005838
5839 kfree(sg);
5840 sg = tmp;
5841 } while (sg != first);
5842}
5843
Peter Zijlstradce840a2011-04-07 14:09:50 +02005844static void free_sched_domain(struct rcu_head *rcu)
5845{
5846 struct sched_domain *sd = container_of(rcu, struct sched_domain, rcu);
Peter Zijlstrae3589f62011-07-15 10:35:52 +02005847
5848 /*
5849 * If its an overlapping domain it has private groups, iterate and
5850 * nuke them all.
5851 */
5852 if (sd->flags & SD_OVERLAP) {
5853 free_sched_groups(sd->groups, 1);
5854 } else if (atomic_dec_and_test(&sd->groups->ref)) {
Nicolas Pitre63b2ca32014-05-26 18:19:37 -04005855 kfree(sd->groups->sgc);
Peter Zijlstradce840a2011-04-07 14:09:50 +02005856 kfree(sd->groups);
Peter Zijlstra9c3f75c2011-07-14 13:00:06 +02005857 }
Peter Zijlstradce840a2011-04-07 14:09:50 +02005858 kfree(sd);
5859}
5860
5861static void destroy_sched_domain(struct sched_domain *sd, int cpu)
5862{
5863 call_rcu(&sd->rcu, free_sched_domain);
5864}
5865
5866static void destroy_sched_domains(struct sched_domain *sd, int cpu)
5867{
5868 for (; sd; sd = sd->parent)
5869 destroy_sched_domain(sd, cpu);
5870}
5871
Linus Torvalds1da177e2005-04-16 15:20:36 -07005872/*
Peter Zijlstra518cd622011-12-07 15:07:31 +01005873 * Keep a special pointer to the highest sched_domain that has
5874 * SD_SHARE_PKG_RESOURCE set (Last Level Cache Domain) for this
5875 * allows us to avoid some pointer chasing select_idle_sibling().
5876 *
5877 * Also keep a unique ID per domain (we use the first cpu number in
5878 * the cpumask of the domain), this allows us to quickly tell if
Peter Zijlstra39be3502012-01-26 12:44:34 +01005879 * two cpus are in the same cache domain, see cpus_share_cache().
Peter Zijlstra518cd622011-12-07 15:07:31 +01005880 */
5881DEFINE_PER_CPU(struct sched_domain *, sd_llc);
Peter Zijlstra7d9ffa82013-07-04 12:56:46 +08005882DEFINE_PER_CPU(int, sd_llc_size);
Peter Zijlstra518cd622011-12-07 15:07:31 +01005883DEFINE_PER_CPU(int, sd_llc_id);
Mel Gormanfb13c7e2013-10-07 11:29:17 +01005884DEFINE_PER_CPU(struct sched_domain *, sd_numa);
Preeti U Murthy37dc6b52013-10-30 08:42:52 +05305885DEFINE_PER_CPU(struct sched_domain *, sd_busy);
5886DEFINE_PER_CPU(struct sched_domain *, sd_asym);
Peter Zijlstra518cd622011-12-07 15:07:31 +01005887
5888static void update_top_cache_domain(int cpu)
5889{
5890 struct sched_domain *sd;
Mel Gorman5d4cf992013-12-17 09:21:25 +00005891 struct sched_domain *busy_sd = NULL;
Peter Zijlstra518cd622011-12-07 15:07:31 +01005892 int id = cpu;
Peter Zijlstra7d9ffa82013-07-04 12:56:46 +08005893 int size = 1;
Peter Zijlstra518cd622011-12-07 15:07:31 +01005894
5895 sd = highest_flag_domain(cpu, SD_SHARE_PKG_RESOURCES);
Peter Zijlstra7d9ffa82013-07-04 12:56:46 +08005896 if (sd) {
Peter Zijlstra518cd622011-12-07 15:07:31 +01005897 id = cpumask_first(sched_domain_span(sd));
Peter Zijlstra7d9ffa82013-07-04 12:56:46 +08005898 size = cpumask_weight(sched_domain_span(sd));
Mel Gorman5d4cf992013-12-17 09:21:25 +00005899 busy_sd = sd->parent; /* sd_busy */
Peter Zijlstra7d9ffa82013-07-04 12:56:46 +08005900 }
Mel Gorman5d4cf992013-12-17 09:21:25 +00005901 rcu_assign_pointer(per_cpu(sd_busy, cpu), busy_sd);
Peter Zijlstra518cd622011-12-07 15:07:31 +01005902
5903 rcu_assign_pointer(per_cpu(sd_llc, cpu), sd);
Peter Zijlstra7d9ffa82013-07-04 12:56:46 +08005904 per_cpu(sd_llc_size, cpu) = size;
Peter Zijlstra518cd622011-12-07 15:07:31 +01005905 per_cpu(sd_llc_id, cpu) = id;
Mel Gormanfb13c7e2013-10-07 11:29:17 +01005906
5907 sd = lowest_flag_domain(cpu, SD_NUMA);
5908 rcu_assign_pointer(per_cpu(sd_numa, cpu), sd);
Preeti U Murthy37dc6b52013-10-30 08:42:52 +05305909
5910 sd = highest_flag_domain(cpu, SD_ASYM_PACKING);
5911 rcu_assign_pointer(per_cpu(sd_asym, cpu), sd);
Peter Zijlstra518cd622011-12-07 15:07:31 +01005912}
5913
5914/*
Ingo Molnar0eab9142008-01-25 21:08:19 +01005915 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
Linus Torvalds1da177e2005-04-16 15:20:36 -07005916 * hold the hotplug lock.
5917 */
Ingo Molnar0eab9142008-01-25 21:08:19 +01005918static void
5919cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005920{
Ingo Molnar70b97a72006-07-03 00:25:42 -07005921 struct rq *rq = cpu_rq(cpu);
Suresh Siddha245af2c2005-06-25 14:57:25 -07005922 struct sched_domain *tmp;
5923
5924 /* Remove the sched domains which do not contribute to scheduling. */
Li Zefanf29c9b12008-11-06 09:45:16 +08005925 for (tmp = sd; tmp; ) {
Suresh Siddha245af2c2005-06-25 14:57:25 -07005926 struct sched_domain *parent = tmp->parent;
5927 if (!parent)
5928 break;
Li Zefanf29c9b12008-11-06 09:45:16 +08005929
Siddha, Suresh B1a848872006-10-03 01:14:08 -07005930 if (sd_parent_degenerate(tmp, parent)) {
Suresh Siddha245af2c2005-06-25 14:57:25 -07005931 tmp->parent = parent->parent;
Siddha, Suresh B1a848872006-10-03 01:14:08 -07005932 if (parent->parent)
5933 parent->parent->child = tmp;
Peter Zijlstra10866e622013-08-19 16:57:04 +02005934 /*
5935 * Transfer SD_PREFER_SIBLING down in case of a
5936 * degenerate parent; the spans match for this
5937 * so the property transfers.
5938 */
5939 if (parent->flags & SD_PREFER_SIBLING)
5940 tmp->flags |= SD_PREFER_SIBLING;
Peter Zijlstradce840a2011-04-07 14:09:50 +02005941 destroy_sched_domain(parent, cpu);
Li Zefanf29c9b12008-11-06 09:45:16 +08005942 } else
5943 tmp = tmp->parent;
Suresh Siddha245af2c2005-06-25 14:57:25 -07005944 }
5945
Siddha, Suresh B1a848872006-10-03 01:14:08 -07005946 if (sd && sd_degenerate(sd)) {
Peter Zijlstradce840a2011-04-07 14:09:50 +02005947 tmp = sd;
Suresh Siddha245af2c2005-06-25 14:57:25 -07005948 sd = sd->parent;
Peter Zijlstradce840a2011-04-07 14:09:50 +02005949 destroy_sched_domain(tmp, cpu);
Siddha, Suresh B1a848872006-10-03 01:14:08 -07005950 if (sd)
5951 sd->child = NULL;
5952 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07005953
Peter Zijlstra4cb98832011-04-07 14:09:58 +02005954 sched_domain_debug(sd, cpu);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005955
Gregory Haskins57d885f2008-01-25 21:08:18 +01005956 rq_attach_root(rq, rd);
Peter Zijlstradce840a2011-04-07 14:09:50 +02005957 tmp = rq->sd;
Nick Piggin674311d2005-06-25 14:57:27 -07005958 rcu_assign_pointer(rq->sd, sd);
Peter Zijlstradce840a2011-04-07 14:09:50 +02005959 destroy_sched_domains(tmp, cpu);
Peter Zijlstra518cd622011-12-07 15:07:31 +01005960
5961 update_top_cache_domain(cpu);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005962}
5963
Linus Torvalds1da177e2005-04-16 15:20:36 -07005964/* Setup the mask of cpus configured for isolated domains */
5965static int __init isolated_cpu_setup(char *str)
5966{
Rusty Russellbdddd292009-12-02 14:09:16 +10305967 alloc_bootmem_cpumask_var(&cpu_isolated_map);
Rusty Russell968ea6d2008-12-13 21:55:51 +10305968 cpulist_parse(str, cpu_isolated_map);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005969 return 1;
5970}
5971
Ingo Molnar8927f492007-10-15 17:00:13 +02005972__setup("isolcpus=", isolated_cpu_setup);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005973
Andreas Herrmann49a02c52009-08-18 12:51:52 +02005974struct s_data {
Peter Zijlstra21d42cc2011-04-07 14:09:48 +02005975 struct sched_domain ** __percpu sd;
Andreas Herrmann49a02c52009-08-18 12:51:52 +02005976 struct root_domain *rd;
5977};
5978
Andreas Herrmann2109b992009-08-18 12:53:00 +02005979enum s_alloc {
Andreas Herrmann2109b992009-08-18 12:53:00 +02005980 sa_rootdomain,
Peter Zijlstra21d42cc2011-04-07 14:09:48 +02005981 sa_sd,
Peter Zijlstradce840a2011-04-07 14:09:50 +02005982 sa_sd_storage,
Andreas Herrmann2109b992009-08-18 12:53:00 +02005983 sa_none,
5984};
5985
Peter Zijlstrac1174872012-05-31 14:47:33 +02005986/*
5987 * Build an iteration mask that can exclude certain CPUs from the upwards
5988 * domain traversal.
5989 *
5990 * Asymmetric node setups can result in situations where the domain tree is of
5991 * unequal depth, make sure to skip domains that already cover the entire
5992 * range.
5993 *
5994 * In that case build_sched_domains() will have terminated the iteration early
5995 * and our sibling sd spans will be empty. Domains should always include the
5996 * cpu they're built on, so check that.
5997 *
5998 */
5999static void build_group_mask(struct sched_domain *sd, struct sched_group *sg)
6000{
6001 const struct cpumask *span = sched_domain_span(sd);
6002 struct sd_data *sdd = sd->private;
6003 struct sched_domain *sibling;
6004 int i;
6005
6006 for_each_cpu(i, span) {
6007 sibling = *per_cpu_ptr(sdd->sd, i);
6008 if (!cpumask_test_cpu(i, sched_domain_span(sibling)))
6009 continue;
6010
6011 cpumask_set_cpu(i, sched_group_mask(sg));
6012 }
6013}
6014
6015/*
6016 * Return the canonical balance cpu for this group, this is the first cpu
6017 * of this group that's also in the iteration mask.
6018 */
6019int group_balance_cpu(struct sched_group *sg)
6020{
6021 return cpumask_first_and(sched_group_cpus(sg), sched_group_mask(sg));
6022}
6023
Peter Zijlstrae3589f62011-07-15 10:35:52 +02006024static int
6025build_overlap_sched_groups(struct sched_domain *sd, int cpu)
6026{
6027 struct sched_group *first = NULL, *last = NULL, *groups = NULL, *sg;
6028 const struct cpumask *span = sched_domain_span(sd);
6029 struct cpumask *covered = sched_domains_tmpmask;
6030 struct sd_data *sdd = sd->private;
Zhihui Zhangaaecac42014-08-01 21:18:03 -04006031 struct sched_domain *sibling;
Peter Zijlstrae3589f62011-07-15 10:35:52 +02006032 int i;
6033
6034 cpumask_clear(covered);
6035
6036 for_each_cpu(i, span) {
6037 struct cpumask *sg_span;
6038
6039 if (cpumask_test_cpu(i, covered))
6040 continue;
6041
Zhihui Zhangaaecac42014-08-01 21:18:03 -04006042 sibling = *per_cpu_ptr(sdd->sd, i);
Peter Zijlstrac1174872012-05-31 14:47:33 +02006043
6044 /* See the comment near build_group_mask(). */
Zhihui Zhangaaecac42014-08-01 21:18:03 -04006045 if (!cpumask_test_cpu(i, sched_domain_span(sibling)))
Peter Zijlstrac1174872012-05-31 14:47:33 +02006046 continue;
6047
Peter Zijlstrae3589f62011-07-15 10:35:52 +02006048 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
Suresh Siddha4d78a222011-11-18 15:03:29 -08006049 GFP_KERNEL, cpu_to_node(cpu));
Peter Zijlstrae3589f62011-07-15 10:35:52 +02006050
6051 if (!sg)
6052 goto fail;
6053
6054 sg_span = sched_group_cpus(sg);
Zhihui Zhangaaecac42014-08-01 21:18:03 -04006055 if (sibling->child)
6056 cpumask_copy(sg_span, sched_domain_span(sibling->child));
6057 else
Peter Zijlstrae3589f62011-07-15 10:35:52 +02006058 cpumask_set_cpu(i, sg_span);
6059
6060 cpumask_or(covered, covered, sg_span);
6061
Nicolas Pitre63b2ca32014-05-26 18:19:37 -04006062 sg->sgc = *per_cpu_ptr(sdd->sgc, i);
6063 if (atomic_inc_return(&sg->sgc->ref) == 1)
Peter Zijlstrac1174872012-05-31 14:47:33 +02006064 build_group_mask(sd, sg);
Peter Zijlstrae3589f62011-07-15 10:35:52 +02006065
Peter Zijlstrac3decf02012-05-31 12:05:32 +02006066 /*
Nicolas Pitre63b2ca32014-05-26 18:19:37 -04006067 * Initialize sgc->capacity such that even if we mess up the
Peter Zijlstrac3decf02012-05-31 12:05:32 +02006068 * domains and no possible iteration will get us here, we won't
6069 * die on a /0 trap.
6070 */
Nicolas Pitreca8ce3d2014-05-26 18:19:39 -04006071 sg->sgc->capacity = SCHED_CAPACITY_SCALE * cpumask_weight(sg_span);
Peter Zijlstrac1174872012-05-31 14:47:33 +02006072
6073 /*
6074 * Make sure the first group of this domain contains the
6075 * canonical balance cpu. Otherwise the sched_domain iteration
6076 * breaks. See update_sg_lb_stats().
6077 */
Peter Zijlstra74a5ce22012-05-23 18:00:43 +02006078 if ((!groups && cpumask_test_cpu(cpu, sg_span)) ||
Peter Zijlstrac1174872012-05-31 14:47:33 +02006079 group_balance_cpu(sg) == cpu)
Peter Zijlstrae3589f62011-07-15 10:35:52 +02006080 groups = sg;
6081
6082 if (!first)
6083 first = sg;
6084 if (last)
6085 last->next = sg;
6086 last = sg;
6087 last->next = first;
6088 }
6089 sd->groups = groups;
6090
6091 return 0;
6092
6093fail:
6094 free_sched_groups(first, 0);
6095
6096 return -ENOMEM;
6097}
6098
Peter Zijlstradce840a2011-04-07 14:09:50 +02006099static int get_group(int cpu, struct sd_data *sdd, struct sched_group **sg)
Linus Torvalds1da177e2005-04-16 15:20:36 -07006100{
Peter Zijlstradce840a2011-04-07 14:09:50 +02006101 struct sched_domain *sd = *per_cpu_ptr(sdd->sd, cpu);
6102 struct sched_domain *child = sd->child;
6103
6104 if (child)
6105 cpu = cpumask_first(sched_domain_span(child));
6106
Peter Zijlstra9c3f75c2011-07-14 13:00:06 +02006107 if (sg) {
Peter Zijlstradce840a2011-04-07 14:09:50 +02006108 *sg = *per_cpu_ptr(sdd->sg, cpu);
Nicolas Pitre63b2ca32014-05-26 18:19:37 -04006109 (*sg)->sgc = *per_cpu_ptr(sdd->sgc, cpu);
6110 atomic_set(&(*sg)->sgc->ref, 1); /* for claim_allocations */
Peter Zijlstra9c3f75c2011-07-14 13:00:06 +02006111 }
Peter Zijlstradce840a2011-04-07 14:09:50 +02006112
Linus Torvalds1da177e2005-04-16 15:20:36 -07006113 return cpu;
6114}
Linus Torvalds1da177e2005-04-16 15:20:36 -07006115
Ingo Molnar48f24c42006-07-03 00:25:40 -07006116/*
Peter Zijlstradce840a2011-04-07 14:09:50 +02006117 * build_sched_groups will build a circular linked list of the groups
6118 * covered by the given span, and will set each group's ->cpumask correctly,
Nicolas Pitreced549f2014-05-26 18:19:38 -04006119 * and ->cpu_capacity to 0.
Peter Zijlstrae3589f62011-07-15 10:35:52 +02006120 *
6121 * Assumes the sched_domain tree is fully constructed
Ingo Molnar48f24c42006-07-03 00:25:40 -07006122 */
Peter Zijlstrae3589f62011-07-15 10:35:52 +02006123static int
6124build_sched_groups(struct sched_domain *sd, int cpu)
Siddha, Suresh B1e9f28f2006-03-27 01:15:22 -08006125{
Peter Zijlstradce840a2011-04-07 14:09:50 +02006126 struct sched_group *first = NULL, *last = NULL;
6127 struct sd_data *sdd = sd->private;
6128 const struct cpumask *span = sched_domain_span(sd);
Peter Zijlstraf96225f2011-04-07 14:09:57 +02006129 struct cpumask *covered;
Peter Zijlstradce840a2011-04-07 14:09:50 +02006130 int i;
6131
Peter Zijlstrae3589f62011-07-15 10:35:52 +02006132 get_group(cpu, sdd, &sd->groups);
6133 atomic_inc(&sd->groups->ref);
6134
Viresh Kumar09366292013-06-11 16:32:43 +05306135 if (cpu != cpumask_first(span))
Peter Zijlstrae3589f62011-07-15 10:35:52 +02006136 return 0;
6137
Peter Zijlstraf96225f2011-04-07 14:09:57 +02006138 lockdep_assert_held(&sched_domains_mutex);
6139 covered = sched_domains_tmpmask;
6140
Peter Zijlstradce840a2011-04-07 14:09:50 +02006141 cpumask_clear(covered);
6142
6143 for_each_cpu(i, span) {
6144 struct sched_group *sg;
Viresh Kumarcd08e922013-06-11 16:32:44 +05306145 int group, j;
Peter Zijlstradce840a2011-04-07 14:09:50 +02006146
6147 if (cpumask_test_cpu(i, covered))
6148 continue;
6149
Viresh Kumarcd08e922013-06-11 16:32:44 +05306150 group = get_group(i, sdd, &sg);
Peter Zijlstrac1174872012-05-31 14:47:33 +02006151 cpumask_setall(sched_group_mask(sg));
Peter Zijlstradce840a2011-04-07 14:09:50 +02006152
6153 for_each_cpu(j, span) {
6154 if (get_group(j, sdd, NULL) != group)
6155 continue;
6156
6157 cpumask_set_cpu(j, covered);
6158 cpumask_set_cpu(j, sched_group_cpus(sg));
6159 }
6160
6161 if (!first)
6162 first = sg;
6163 if (last)
6164 last->next = sg;
6165 last = sg;
6166 }
6167 last->next = first;
Peter Zijlstrae3589f62011-07-15 10:35:52 +02006168
6169 return 0;
Siddha, Suresh B1e9f28f2006-03-27 01:15:22 -08006170}
Srivatsa Vaddagiri51888ca2006-06-27 02:54:38 -07006171
Linus Torvalds1da177e2005-04-16 15:20:36 -07006172/*
Nicolas Pitre63b2ca32014-05-26 18:19:37 -04006173 * Initialize sched groups cpu_capacity.
Siddha, Suresh B89c47102006-10-03 01:14:09 -07006174 *
Nicolas Pitre63b2ca32014-05-26 18:19:37 -04006175 * cpu_capacity indicates the capacity of sched group, which is used while
Siddha, Suresh B89c47102006-10-03 01:14:09 -07006176 * distributing the load between different sched groups in a sched domain.
Nicolas Pitre63b2ca32014-05-26 18:19:37 -04006177 * Typically cpu_capacity for all the groups in a sched domain will be same
6178 * unless there are asymmetries in the topology. If there are asymmetries,
6179 * group having more cpu_capacity will pickup more load compared to the
6180 * group having less cpu_capacity.
Siddha, Suresh B89c47102006-10-03 01:14:09 -07006181 */
Nicolas Pitre63b2ca32014-05-26 18:19:37 -04006182static void init_sched_groups_capacity(int cpu, struct sched_domain *sd)
Siddha, Suresh B89c47102006-10-03 01:14:09 -07006183{
Peter Zijlstrae3589f62011-07-15 10:35:52 +02006184 struct sched_group *sg = sd->groups;
Siddha, Suresh B89c47102006-10-03 01:14:09 -07006185
Viresh Kumar94c95ba2013-06-11 16:32:45 +05306186 WARN_ON(!sg);
Peter Zijlstrae3589f62011-07-15 10:35:52 +02006187
6188 do {
6189 sg->group_weight = cpumask_weight(sched_group_cpus(sg));
6190 sg = sg->next;
6191 } while (sg != sd->groups);
6192
Peter Zijlstrac1174872012-05-31 14:47:33 +02006193 if (cpu != group_balance_cpu(sg))
Siddha, Suresh B89c47102006-10-03 01:14:09 -07006194 return;
6195
Nicolas Pitre63b2ca32014-05-26 18:19:37 -04006196 update_group_capacity(sd, cpu);
6197 atomic_set(&sg->sgc->nr_busy_cpus, sg->group_weight);
Siddha, Suresh B89c47102006-10-03 01:14:09 -07006198}
6199
6200/*
Mike Travis7c16ec52008-04-04 18:11:11 -07006201 * Initializers for schedule domains
6202 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
6203 */
6204
Hidetoshi Seto1d3504f2008-04-15 14:04:23 +09006205static int default_relax_domain_level = -1;
Peter Zijlstra60495e72011-04-07 14:10:04 +02006206int sched_domain_level_max;
Hidetoshi Seto1d3504f2008-04-15 14:04:23 +09006207
6208static int __init setup_relax_domain_level(char *str)
6209{
Dimitri Sivanicha841f8c2012-06-05 13:44:36 -05006210 if (kstrtoint(str, 0, &default_relax_domain_level))
6211 pr_warn("Unable to set relax_domain_level\n");
Li Zefan30e0e172008-05-13 10:27:17 +08006212
Hidetoshi Seto1d3504f2008-04-15 14:04:23 +09006213 return 1;
6214}
6215__setup("relax_domain_level=", setup_relax_domain_level);
6216
6217static void set_domain_attribute(struct sched_domain *sd,
6218 struct sched_domain_attr *attr)
6219{
6220 int request;
6221
6222 if (!attr || attr->relax_domain_level < 0) {
6223 if (default_relax_domain_level < 0)
6224 return;
6225 else
6226 request = default_relax_domain_level;
6227 } else
6228 request = attr->relax_domain_level;
6229 if (request < sd->level) {
6230 /* turn off idle balance on this domain */
Peter Zijlstrac88d5912009-09-10 13:50:02 +02006231 sd->flags &= ~(SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
Hidetoshi Seto1d3504f2008-04-15 14:04:23 +09006232 } else {
6233 /* turn on idle balance on this domain */
Peter Zijlstrac88d5912009-09-10 13:50:02 +02006234 sd->flags |= (SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
Hidetoshi Seto1d3504f2008-04-15 14:04:23 +09006235 }
6236}
6237
Peter Zijlstra54ab4ff2011-04-07 14:10:03 +02006238static void __sdt_free(const struct cpumask *cpu_map);
6239static int __sdt_alloc(const struct cpumask *cpu_map);
6240
Andreas Herrmann2109b992009-08-18 12:53:00 +02006241static void __free_domain_allocs(struct s_data *d, enum s_alloc what,
6242 const struct cpumask *cpu_map)
6243{
6244 switch (what) {
Andreas Herrmann2109b992009-08-18 12:53:00 +02006245 case sa_rootdomain:
Peter Zijlstra822ff792011-04-07 14:09:51 +02006246 if (!atomic_read(&d->rd->refcount))
6247 free_rootdomain(&d->rd->rcu); /* fall through */
Peter Zijlstra21d42cc2011-04-07 14:09:48 +02006248 case sa_sd:
6249 free_percpu(d->sd); /* fall through */
Peter Zijlstradce840a2011-04-07 14:09:50 +02006250 case sa_sd_storage:
Peter Zijlstra54ab4ff2011-04-07 14:10:03 +02006251 __sdt_free(cpu_map); /* fall through */
Andreas Herrmann2109b992009-08-18 12:53:00 +02006252 case sa_none:
6253 break;
6254 }
6255}
6256
6257static enum s_alloc __visit_domain_allocation_hell(struct s_data *d,
6258 const struct cpumask *cpu_map)
6259{
Peter Zijlstradce840a2011-04-07 14:09:50 +02006260 memset(d, 0, sizeof(*d));
6261
Peter Zijlstra54ab4ff2011-04-07 14:10:03 +02006262 if (__sdt_alloc(cpu_map))
6263 return sa_sd_storage;
Peter Zijlstra21d42cc2011-04-07 14:09:48 +02006264 d->sd = alloc_percpu(struct sched_domain *);
Peter Zijlstradce840a2011-04-07 14:09:50 +02006265 if (!d->sd)
6266 return sa_sd_storage;
Andreas Herrmann2109b992009-08-18 12:53:00 +02006267 d->rd = alloc_rootdomain();
Peter Zijlstradce840a2011-04-07 14:09:50 +02006268 if (!d->rd)
Peter Zijlstra21d42cc2011-04-07 14:09:48 +02006269 return sa_sd;
Andreas Herrmann2109b992009-08-18 12:53:00 +02006270 return sa_rootdomain;
6271}
6272
Peter Zijlstradce840a2011-04-07 14:09:50 +02006273/*
6274 * NULL the sd_data elements we've used to build the sched_domain and
6275 * sched_group structure so that the subsequent __free_domain_allocs()
6276 * will not free the data we're using.
6277 */
6278static void claim_allocations(int cpu, struct sched_domain *sd)
6279{
6280 struct sd_data *sdd = sd->private;
Peter Zijlstradce840a2011-04-07 14:09:50 +02006281
6282 WARN_ON_ONCE(*per_cpu_ptr(sdd->sd, cpu) != sd);
6283 *per_cpu_ptr(sdd->sd, cpu) = NULL;
6284
Peter Zijlstrae3589f62011-07-15 10:35:52 +02006285 if (atomic_read(&(*per_cpu_ptr(sdd->sg, cpu))->ref))
Peter Zijlstradce840a2011-04-07 14:09:50 +02006286 *per_cpu_ptr(sdd->sg, cpu) = NULL;
Peter Zijlstrae3589f62011-07-15 10:35:52 +02006287
Nicolas Pitre63b2ca32014-05-26 18:19:37 -04006288 if (atomic_read(&(*per_cpu_ptr(sdd->sgc, cpu))->ref))
6289 *per_cpu_ptr(sdd->sgc, cpu) = NULL;
Peter Zijlstradce840a2011-04-07 14:09:50 +02006290}
6291
Peter Zijlstracb83b622012-04-17 15:49:36 +02006292#ifdef CONFIG_NUMA
Peter Zijlstracb83b622012-04-17 15:49:36 +02006293static int sched_domains_numa_levels;
Rik van Riele3fe70b2014-10-17 03:29:50 -04006294enum numa_topology_type sched_numa_topology_type;
Peter Zijlstracb83b622012-04-17 15:49:36 +02006295static int *sched_domains_numa_distance;
Rik van Riel9942f792014-10-17 03:29:49 -04006296int sched_max_numa_distance;
Peter Zijlstracb83b622012-04-17 15:49:36 +02006297static struct cpumask ***sched_domains_numa_masks;
6298static int sched_domains_curr_level;
Vincent Guittot143e1e22014-04-11 11:44:37 +02006299#endif
Peter Zijlstracb83b622012-04-17 15:49:36 +02006300
Vincent Guittot143e1e22014-04-11 11:44:37 +02006301/*
6302 * SD_flags allowed in topology descriptions.
6303 *
Nicolas Pitre5d4dfdd2014-05-27 13:50:41 -04006304 * SD_SHARE_CPUCAPACITY - describes SMT topologies
Vincent Guittot143e1e22014-04-11 11:44:37 +02006305 * SD_SHARE_PKG_RESOURCES - describes shared caches
6306 * SD_NUMA - describes NUMA topologies
Vincent Guittotd77b3ed2014-04-11 11:44:40 +02006307 * SD_SHARE_POWERDOMAIN - describes shared power domain
Vincent Guittot143e1e22014-04-11 11:44:37 +02006308 *
6309 * Odd one out:
6310 * SD_ASYM_PACKING - describes SMT quirks
6311 */
6312#define TOPOLOGY_SD_FLAGS \
Nicolas Pitre5d4dfdd2014-05-27 13:50:41 -04006313 (SD_SHARE_CPUCAPACITY | \
Vincent Guittot143e1e22014-04-11 11:44:37 +02006314 SD_SHARE_PKG_RESOURCES | \
6315 SD_NUMA | \
Vincent Guittotd77b3ed2014-04-11 11:44:40 +02006316 SD_ASYM_PACKING | \
6317 SD_SHARE_POWERDOMAIN)
Peter Zijlstracb83b622012-04-17 15:49:36 +02006318
6319static struct sched_domain *
Vincent Guittot143e1e22014-04-11 11:44:37 +02006320sd_init(struct sched_domain_topology_level *tl, int cpu)
Peter Zijlstracb83b622012-04-17 15:49:36 +02006321{
6322 struct sched_domain *sd = *per_cpu_ptr(tl->data.sd, cpu);
Vincent Guittot143e1e22014-04-11 11:44:37 +02006323 int sd_weight, sd_flags = 0;
6324
6325#ifdef CONFIG_NUMA
6326 /*
6327 * Ugly hack to pass state to sd_numa_mask()...
6328 */
6329 sched_domains_curr_level = tl->numa_level;
6330#endif
6331
6332 sd_weight = cpumask_weight(tl->mask(cpu));
6333
6334 if (tl->sd_flags)
6335 sd_flags = (*tl->sd_flags)();
6336 if (WARN_ONCE(sd_flags & ~TOPOLOGY_SD_FLAGS,
6337 "wrong sd_flags in topology description\n"))
6338 sd_flags &= ~TOPOLOGY_SD_FLAGS;
Peter Zijlstracb83b622012-04-17 15:49:36 +02006339
6340 *sd = (struct sched_domain){
6341 .min_interval = sd_weight,
6342 .max_interval = 2*sd_weight,
6343 .busy_factor = 32,
Peter Zijlstra870a0bb2012-05-11 00:26:27 +02006344 .imbalance_pct = 125,
Vincent Guittot143e1e22014-04-11 11:44:37 +02006345
6346 .cache_nice_tries = 0,
6347 .busy_idx = 0,
6348 .idle_idx = 0,
Peter Zijlstracb83b622012-04-17 15:49:36 +02006349 .newidle_idx = 0,
6350 .wake_idx = 0,
6351 .forkexec_idx = 0,
6352
6353 .flags = 1*SD_LOAD_BALANCE
6354 | 1*SD_BALANCE_NEWIDLE
Vincent Guittot143e1e22014-04-11 11:44:37 +02006355 | 1*SD_BALANCE_EXEC
6356 | 1*SD_BALANCE_FORK
Peter Zijlstracb83b622012-04-17 15:49:36 +02006357 | 0*SD_BALANCE_WAKE
Vincent Guittot143e1e22014-04-11 11:44:37 +02006358 | 1*SD_WAKE_AFFINE
Nicolas Pitre5d4dfdd2014-05-27 13:50:41 -04006359 | 0*SD_SHARE_CPUCAPACITY
Peter Zijlstracb83b622012-04-17 15:49:36 +02006360 | 0*SD_SHARE_PKG_RESOURCES
Vincent Guittot143e1e22014-04-11 11:44:37 +02006361 | 0*SD_SERIALIZE
Peter Zijlstracb83b622012-04-17 15:49:36 +02006362 | 0*SD_PREFER_SIBLING
Vincent Guittot143e1e22014-04-11 11:44:37 +02006363 | 0*SD_NUMA
6364 | sd_flags
Peter Zijlstracb83b622012-04-17 15:49:36 +02006365 ,
Vincent Guittot143e1e22014-04-11 11:44:37 +02006366
Peter Zijlstracb83b622012-04-17 15:49:36 +02006367 .last_balance = jiffies,
6368 .balance_interval = sd_weight,
Vincent Guittot143e1e22014-04-11 11:44:37 +02006369 .smt_gain = 0,
Jason Low2b4cfe62014-04-23 18:30:34 -07006370 .max_newidle_lb_cost = 0,
6371 .next_decay_max_lb_cost = jiffies,
Vincent Guittot143e1e22014-04-11 11:44:37 +02006372#ifdef CONFIG_SCHED_DEBUG
6373 .name = tl->name,
6374#endif
Peter Zijlstracb83b622012-04-17 15:49:36 +02006375 };
Peter Zijlstracb83b622012-04-17 15:49:36 +02006376
6377 /*
Vincent Guittot143e1e22014-04-11 11:44:37 +02006378 * Convert topological properties into behaviour.
Peter Zijlstracb83b622012-04-17 15:49:36 +02006379 */
Vincent Guittot143e1e22014-04-11 11:44:37 +02006380
Nicolas Pitre5d4dfdd2014-05-27 13:50:41 -04006381 if (sd->flags & SD_SHARE_CPUCAPACITY) {
Vincent Guittotcaff37e2015-02-27 16:54:13 +01006382 sd->flags |= SD_PREFER_SIBLING;
Vincent Guittot143e1e22014-04-11 11:44:37 +02006383 sd->imbalance_pct = 110;
6384 sd->smt_gain = 1178; /* ~15% */
Vincent Guittot143e1e22014-04-11 11:44:37 +02006385
6386 } else if (sd->flags & SD_SHARE_PKG_RESOURCES) {
6387 sd->imbalance_pct = 117;
6388 sd->cache_nice_tries = 1;
6389 sd->busy_idx = 2;
6390
6391#ifdef CONFIG_NUMA
6392 } else if (sd->flags & SD_NUMA) {
6393 sd->cache_nice_tries = 2;
6394 sd->busy_idx = 3;
6395 sd->idle_idx = 2;
6396
6397 sd->flags |= SD_SERIALIZE;
6398 if (sched_domains_numa_distance[tl->numa_level] > RECLAIM_DISTANCE) {
6399 sd->flags &= ~(SD_BALANCE_EXEC |
6400 SD_BALANCE_FORK |
6401 SD_WAKE_AFFINE);
6402 }
6403
6404#endif
6405 } else {
6406 sd->flags |= SD_PREFER_SIBLING;
6407 sd->cache_nice_tries = 1;
6408 sd->busy_idx = 2;
6409 sd->idle_idx = 1;
6410 }
6411
6412 sd->private = &tl->data;
Peter Zijlstracb83b622012-04-17 15:49:36 +02006413
6414 return sd;
6415}
6416
Vincent Guittot143e1e22014-04-11 11:44:37 +02006417/*
6418 * Topology list, bottom-up.
6419 */
6420static struct sched_domain_topology_level default_topology[] = {
6421#ifdef CONFIG_SCHED_SMT
6422 { cpu_smt_mask, cpu_smt_flags, SD_INIT_NAME(SMT) },
6423#endif
6424#ifdef CONFIG_SCHED_MC
6425 { cpu_coregroup_mask, cpu_core_flags, SD_INIT_NAME(MC) },
6426#endif
Vincent Guittot143e1e22014-04-11 11:44:37 +02006427 { cpu_cpu_mask, SD_INIT_NAME(DIE) },
6428 { NULL, },
6429};
6430
6431struct sched_domain_topology_level *sched_domain_topology = default_topology;
6432
6433#define for_each_sd_topology(tl) \
6434 for (tl = sched_domain_topology; tl->mask; tl++)
6435
6436void set_sched_topology(struct sched_domain_topology_level *tl)
6437{
6438 sched_domain_topology = tl;
6439}
6440
6441#ifdef CONFIG_NUMA
6442
Peter Zijlstracb83b622012-04-17 15:49:36 +02006443static const struct cpumask *sd_numa_mask(int cpu)
6444{
6445 return sched_domains_numa_masks[sched_domains_curr_level][cpu_to_node(cpu)];
6446}
6447
Peter Zijlstrad039ac62012-05-31 21:20:16 +02006448static void sched_numa_warn(const char *str)
6449{
6450 static int done = false;
6451 int i,j;
6452
6453 if (done)
6454 return;
6455
6456 done = true;
6457
6458 printk(KERN_WARNING "ERROR: %s\n\n", str);
6459
6460 for (i = 0; i < nr_node_ids; i++) {
6461 printk(KERN_WARNING " ");
6462 for (j = 0; j < nr_node_ids; j++)
6463 printk(KERN_CONT "%02d ", node_distance(i,j));
6464 printk(KERN_CONT "\n");
6465 }
6466 printk(KERN_WARNING "\n");
6467}
6468
Rik van Riel9942f792014-10-17 03:29:49 -04006469bool find_numa_distance(int distance)
Peter Zijlstrad039ac62012-05-31 21:20:16 +02006470{
6471 int i;
6472
6473 if (distance == node_distance(0, 0))
6474 return true;
6475
6476 for (i = 0; i < sched_domains_numa_levels; i++) {
6477 if (sched_domains_numa_distance[i] == distance)
6478 return true;
6479 }
6480
6481 return false;
6482}
6483
Rik van Riele3fe70b2014-10-17 03:29:50 -04006484/*
6485 * A system can have three types of NUMA topology:
6486 * NUMA_DIRECT: all nodes are directly connected, or not a NUMA system
6487 * NUMA_GLUELESS_MESH: some nodes reachable through intermediary nodes
6488 * NUMA_BACKPLANE: nodes can reach other nodes through a backplane
6489 *
6490 * The difference between a glueless mesh topology and a backplane
6491 * topology lies in whether communication between not directly
6492 * connected nodes goes through intermediary nodes (where programs
6493 * could run), or through backplane controllers. This affects
6494 * placement of programs.
6495 *
6496 * The type of topology can be discerned with the following tests:
6497 * - If the maximum distance between any nodes is 1 hop, the system
6498 * is directly connected.
6499 * - If for two nodes A and B, located N > 1 hops away from each other,
6500 * there is an intermediary node C, which is < N hops away from both
6501 * nodes A and B, the system is a glueless mesh.
6502 */
6503static void init_numa_topology_type(void)
6504{
6505 int a, b, c, n;
6506
6507 n = sched_max_numa_distance;
6508
Aravind Gopalakrishnane2378822015-08-10 20:20:48 -05006509 if (sched_domains_numa_levels <= 1) {
Rik van Riele3fe70b2014-10-17 03:29:50 -04006510 sched_numa_topology_type = NUMA_DIRECT;
Aravind Gopalakrishnane2378822015-08-10 20:20:48 -05006511 return;
6512 }
Rik van Riele3fe70b2014-10-17 03:29:50 -04006513
6514 for_each_online_node(a) {
6515 for_each_online_node(b) {
6516 /* Find two nodes furthest removed from each other. */
6517 if (node_distance(a, b) < n)
6518 continue;
6519
6520 /* Is there an intermediary node between a and b? */
6521 for_each_online_node(c) {
6522 if (node_distance(a, c) < n &&
6523 node_distance(b, c) < n) {
6524 sched_numa_topology_type =
6525 NUMA_GLUELESS_MESH;
6526 return;
6527 }
6528 }
6529
6530 sched_numa_topology_type = NUMA_BACKPLANE;
6531 return;
6532 }
6533 }
6534}
6535
Peter Zijlstracb83b622012-04-17 15:49:36 +02006536static void sched_init_numa(void)
6537{
6538 int next_distance, curr_distance = node_distance(0, 0);
6539 struct sched_domain_topology_level *tl;
6540 int level = 0;
6541 int i, j, k;
6542
Peter Zijlstracb83b622012-04-17 15:49:36 +02006543 sched_domains_numa_distance = kzalloc(sizeof(int) * nr_node_ids, GFP_KERNEL);
6544 if (!sched_domains_numa_distance)
6545 return;
6546
6547 /*
6548 * O(nr_nodes^2) deduplicating selection sort -- in order to find the
6549 * unique distances in the node_distance() table.
6550 *
6551 * Assumes node_distance(0,j) includes all distances in
6552 * node_distance(i,j) in order to avoid cubic time.
Peter Zijlstracb83b622012-04-17 15:49:36 +02006553 */
6554 next_distance = curr_distance;
6555 for (i = 0; i < nr_node_ids; i++) {
6556 for (j = 0; j < nr_node_ids; j++) {
Peter Zijlstrad039ac62012-05-31 21:20:16 +02006557 for (k = 0; k < nr_node_ids; k++) {
6558 int distance = node_distance(i, k);
6559
6560 if (distance > curr_distance &&
6561 (distance < next_distance ||
6562 next_distance == curr_distance))
6563 next_distance = distance;
6564
6565 /*
6566 * While not a strong assumption it would be nice to know
6567 * about cases where if node A is connected to B, B is not
6568 * equally connected to A.
6569 */
6570 if (sched_debug() && node_distance(k, i) != distance)
6571 sched_numa_warn("Node-distance not symmetric");
6572
6573 if (sched_debug() && i && !find_numa_distance(distance))
6574 sched_numa_warn("Node-0 not representative");
6575 }
6576 if (next_distance != curr_distance) {
6577 sched_domains_numa_distance[level++] = next_distance;
6578 sched_domains_numa_levels = level;
6579 curr_distance = next_distance;
6580 } else break;
Peter Zijlstracb83b622012-04-17 15:49:36 +02006581 }
Peter Zijlstrad039ac62012-05-31 21:20:16 +02006582
6583 /*
6584 * In case of sched_debug() we verify the above assumption.
6585 */
6586 if (!sched_debug())
6587 break;
Peter Zijlstracb83b622012-04-17 15:49:36 +02006588 }
Andrey Ryabininc1235882014-11-07 17:53:40 +03006589
6590 if (!level)
6591 return;
6592
Peter Zijlstracb83b622012-04-17 15:49:36 +02006593 /*
6594 * 'level' contains the number of unique distances, excluding the
6595 * identity distance node_distance(i,i).
6596 *
Viresh Kumar28b4a522013-04-05 16:26:46 +05306597 * The sched_domains_numa_distance[] array includes the actual distance
Peter Zijlstracb83b622012-04-17 15:49:36 +02006598 * numbers.
6599 */
6600
Tang Chen5f7865f2012-09-25 21:12:30 +08006601 /*
6602 * Here, we should temporarily reset sched_domains_numa_levels to 0.
6603 * If it fails to allocate memory for array sched_domains_numa_masks[][],
6604 * the array will contain less then 'level' members. This could be
6605 * dangerous when we use it to iterate array sched_domains_numa_masks[][]
6606 * in other functions.
6607 *
6608 * We reset it to 'level' at the end of this function.
6609 */
6610 sched_domains_numa_levels = 0;
6611
Peter Zijlstracb83b622012-04-17 15:49:36 +02006612 sched_domains_numa_masks = kzalloc(sizeof(void *) * level, GFP_KERNEL);
6613 if (!sched_domains_numa_masks)
6614 return;
6615
6616 /*
6617 * Now for each level, construct a mask per node which contains all
6618 * cpus of nodes that are that many hops away from us.
6619 */
6620 for (i = 0; i < level; i++) {
6621 sched_domains_numa_masks[i] =
6622 kzalloc(nr_node_ids * sizeof(void *), GFP_KERNEL);
6623 if (!sched_domains_numa_masks[i])
6624 return;
6625
6626 for (j = 0; j < nr_node_ids; j++) {
Peter Zijlstra2ea45802012-05-25 09:26:43 +02006627 struct cpumask *mask = kzalloc(cpumask_size(), GFP_KERNEL);
Peter Zijlstracb83b622012-04-17 15:49:36 +02006628 if (!mask)
6629 return;
6630
6631 sched_domains_numa_masks[i][j] = mask;
6632
6633 for (k = 0; k < nr_node_ids; k++) {
Peter Zijlstradd7d8632012-05-11 00:56:20 +02006634 if (node_distance(j, k) > sched_domains_numa_distance[i])
Peter Zijlstracb83b622012-04-17 15:49:36 +02006635 continue;
6636
6637 cpumask_or(mask, mask, cpumask_of_node(k));
6638 }
6639 }
6640 }
6641
Vincent Guittot143e1e22014-04-11 11:44:37 +02006642 /* Compute default topology size */
6643 for (i = 0; sched_domain_topology[i].mask; i++);
6644
Vincent Guittotc515db82014-05-13 11:11:01 +02006645 tl = kzalloc((i + level + 1) *
Peter Zijlstracb83b622012-04-17 15:49:36 +02006646 sizeof(struct sched_domain_topology_level), GFP_KERNEL);
6647 if (!tl)
6648 return;
6649
6650 /*
6651 * Copy the default topology bits..
6652 */
Vincent Guittot143e1e22014-04-11 11:44:37 +02006653 for (i = 0; sched_domain_topology[i].mask; i++)
6654 tl[i] = sched_domain_topology[i];
Peter Zijlstracb83b622012-04-17 15:49:36 +02006655
6656 /*
6657 * .. and append 'j' levels of NUMA goodness.
6658 */
6659 for (j = 0; j < level; i++, j++) {
6660 tl[i] = (struct sched_domain_topology_level){
Peter Zijlstracb83b622012-04-17 15:49:36 +02006661 .mask = sd_numa_mask,
Vincent Guittot143e1e22014-04-11 11:44:37 +02006662 .sd_flags = cpu_numa_flags,
Peter Zijlstracb83b622012-04-17 15:49:36 +02006663 .flags = SDTL_OVERLAP,
6664 .numa_level = j,
Vincent Guittot143e1e22014-04-11 11:44:37 +02006665 SD_INIT_NAME(NUMA)
Peter Zijlstracb83b622012-04-17 15:49:36 +02006666 };
6667 }
6668
6669 sched_domain_topology = tl;
Tang Chen5f7865f2012-09-25 21:12:30 +08006670
6671 sched_domains_numa_levels = level;
Rik van Riel9942f792014-10-17 03:29:49 -04006672 sched_max_numa_distance = sched_domains_numa_distance[level - 1];
Rik van Riele3fe70b2014-10-17 03:29:50 -04006673
6674 init_numa_topology_type();
Peter Zijlstracb83b622012-04-17 15:49:36 +02006675}
Tang Chen301a5cb2012-09-25 21:12:31 +08006676
6677static void sched_domains_numa_masks_set(int cpu)
6678{
6679 int i, j;
6680 int node = cpu_to_node(cpu);
6681
6682 for (i = 0; i < sched_domains_numa_levels; i++) {
6683 for (j = 0; j < nr_node_ids; j++) {
6684 if (node_distance(j, node) <= sched_domains_numa_distance[i])
6685 cpumask_set_cpu(cpu, sched_domains_numa_masks[i][j]);
6686 }
6687 }
6688}
6689
6690static void sched_domains_numa_masks_clear(int cpu)
6691{
6692 int i, j;
6693 for (i = 0; i < sched_domains_numa_levels; i++) {
6694 for (j = 0; j < nr_node_ids; j++)
6695 cpumask_clear_cpu(cpu, sched_domains_numa_masks[i][j]);
6696 }
6697}
6698
6699/*
6700 * Update sched_domains_numa_masks[level][node] array when new cpus
6701 * are onlined.
6702 */
6703static int sched_domains_numa_masks_update(struct notifier_block *nfb,
6704 unsigned long action,
6705 void *hcpu)
6706{
6707 int cpu = (long)hcpu;
6708
6709 switch (action & ~CPU_TASKS_FROZEN) {
6710 case CPU_ONLINE:
6711 sched_domains_numa_masks_set(cpu);
6712 break;
6713
6714 case CPU_DEAD:
6715 sched_domains_numa_masks_clear(cpu);
6716 break;
6717
6718 default:
6719 return NOTIFY_DONE;
6720 }
6721
6722 return NOTIFY_OK;
Peter Zijlstracb83b622012-04-17 15:49:36 +02006723}
6724#else
6725static inline void sched_init_numa(void)
6726{
6727}
Tang Chen301a5cb2012-09-25 21:12:31 +08006728
6729static int sched_domains_numa_masks_update(struct notifier_block *nfb,
6730 unsigned long action,
6731 void *hcpu)
6732{
6733 return 0;
6734}
Peter Zijlstracb83b622012-04-17 15:49:36 +02006735#endif /* CONFIG_NUMA */
6736
Peter Zijlstra54ab4ff2011-04-07 14:10:03 +02006737static int __sdt_alloc(const struct cpumask *cpu_map)
6738{
6739 struct sched_domain_topology_level *tl;
6740 int j;
6741
Viresh Kumar27723a62013-06-10 16:27:20 +05306742 for_each_sd_topology(tl) {
Peter Zijlstra54ab4ff2011-04-07 14:10:03 +02006743 struct sd_data *sdd = &tl->data;
6744
6745 sdd->sd = alloc_percpu(struct sched_domain *);
6746 if (!sdd->sd)
6747 return -ENOMEM;
6748
6749 sdd->sg = alloc_percpu(struct sched_group *);
6750 if (!sdd->sg)
6751 return -ENOMEM;
6752
Nicolas Pitre63b2ca32014-05-26 18:19:37 -04006753 sdd->sgc = alloc_percpu(struct sched_group_capacity *);
6754 if (!sdd->sgc)
Peter Zijlstra9c3f75c2011-07-14 13:00:06 +02006755 return -ENOMEM;
6756
Peter Zijlstra54ab4ff2011-04-07 14:10:03 +02006757 for_each_cpu(j, cpu_map) {
6758 struct sched_domain *sd;
6759 struct sched_group *sg;
Nicolas Pitre63b2ca32014-05-26 18:19:37 -04006760 struct sched_group_capacity *sgc;
Peter Zijlstra54ab4ff2011-04-07 14:10:03 +02006761
Peter Zijlstra5cc389b2015-06-11 14:46:50 +02006762 sd = kzalloc_node(sizeof(struct sched_domain) + cpumask_size(),
Peter Zijlstra54ab4ff2011-04-07 14:10:03 +02006763 GFP_KERNEL, cpu_to_node(j));
6764 if (!sd)
6765 return -ENOMEM;
6766
6767 *per_cpu_ptr(sdd->sd, j) = sd;
6768
6769 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
6770 GFP_KERNEL, cpu_to_node(j));
6771 if (!sg)
6772 return -ENOMEM;
6773
Igor Mammedov30b4e9e2012-05-09 12:38:28 +02006774 sg->next = sg;
6775
Peter Zijlstra54ab4ff2011-04-07 14:10:03 +02006776 *per_cpu_ptr(sdd->sg, j) = sg;
Peter Zijlstra9c3f75c2011-07-14 13:00:06 +02006777
Nicolas Pitre63b2ca32014-05-26 18:19:37 -04006778 sgc = kzalloc_node(sizeof(struct sched_group_capacity) + cpumask_size(),
Peter Zijlstra9c3f75c2011-07-14 13:00:06 +02006779 GFP_KERNEL, cpu_to_node(j));
Nicolas Pitre63b2ca32014-05-26 18:19:37 -04006780 if (!sgc)
Peter Zijlstra9c3f75c2011-07-14 13:00:06 +02006781 return -ENOMEM;
6782
Nicolas Pitre63b2ca32014-05-26 18:19:37 -04006783 *per_cpu_ptr(sdd->sgc, j) = sgc;
Peter Zijlstra54ab4ff2011-04-07 14:10:03 +02006784 }
6785 }
6786
6787 return 0;
6788}
6789
6790static void __sdt_free(const struct cpumask *cpu_map)
6791{
6792 struct sched_domain_topology_level *tl;
6793 int j;
6794
Viresh Kumar27723a62013-06-10 16:27:20 +05306795 for_each_sd_topology(tl) {
Peter Zijlstra54ab4ff2011-04-07 14:10:03 +02006796 struct sd_data *sdd = &tl->data;
6797
6798 for_each_cpu(j, cpu_map) {
he, bofb2cf2c2012-04-25 19:59:21 +08006799 struct sched_domain *sd;
6800
6801 if (sdd->sd) {
6802 sd = *per_cpu_ptr(sdd->sd, j);
6803 if (sd && (sd->flags & SD_OVERLAP))
6804 free_sched_groups(sd->groups, 0);
6805 kfree(*per_cpu_ptr(sdd->sd, j));
6806 }
6807
6808 if (sdd->sg)
6809 kfree(*per_cpu_ptr(sdd->sg, j));
Nicolas Pitre63b2ca32014-05-26 18:19:37 -04006810 if (sdd->sgc)
6811 kfree(*per_cpu_ptr(sdd->sgc, j));
Peter Zijlstra54ab4ff2011-04-07 14:10:03 +02006812 }
6813 free_percpu(sdd->sd);
he, bofb2cf2c2012-04-25 19:59:21 +08006814 sdd->sd = NULL;
Peter Zijlstra54ab4ff2011-04-07 14:10:03 +02006815 free_percpu(sdd->sg);
he, bofb2cf2c2012-04-25 19:59:21 +08006816 sdd->sg = NULL;
Nicolas Pitre63b2ca32014-05-26 18:19:37 -04006817 free_percpu(sdd->sgc);
6818 sdd->sgc = NULL;
Peter Zijlstra54ab4ff2011-04-07 14:10:03 +02006819 }
6820}
6821
Peter Zijlstra2c402dc2011-04-07 14:10:01 +02006822struct sched_domain *build_sched_domain(struct sched_domain_topology_level *tl,
Viresh Kumar4a850cb2013-06-04 16:12:43 +05306823 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
6824 struct sched_domain *child, int cpu)
Peter Zijlstra2c402dc2011-04-07 14:10:01 +02006825{
Vincent Guittot143e1e22014-04-11 11:44:37 +02006826 struct sched_domain *sd = sd_init(tl, cpu);
Peter Zijlstra2c402dc2011-04-07 14:10:01 +02006827 if (!sd)
Peter Zijlstrad069b912011-04-07 14:10:02 +02006828 return child;
Peter Zijlstra2c402dc2011-04-07 14:10:01 +02006829
Peter Zijlstra2c402dc2011-04-07 14:10:01 +02006830 cpumask_and(sched_domain_span(sd), cpu_map, tl->mask(cpu));
Peter Zijlstra60495e72011-04-07 14:10:04 +02006831 if (child) {
6832 sd->level = child->level + 1;
6833 sched_domain_level_max = max(sched_domain_level_max, sd->level);
Peter Zijlstrad069b912011-04-07 14:10:02 +02006834 child->parent = sd;
Viresh Kumarc75e0122013-06-10 16:27:19 +05306835 sd->child = child;
Peter Zijlstra6ae72df2014-07-22 11:47:40 +02006836
6837 if (!cpumask_subset(sched_domain_span(child),
6838 sched_domain_span(sd))) {
6839 pr_err("BUG: arch topology borken\n");
6840#ifdef CONFIG_SCHED_DEBUG
6841 pr_err(" the %s domain not a subset of the %s domain\n",
6842 child->name, sd->name);
6843#endif
6844 /* Fixup, ensure @sd has at least @child cpus. */
6845 cpumask_or(sched_domain_span(sd),
6846 sched_domain_span(sd),
6847 sched_domain_span(child));
6848 }
6849
Peter Zijlstra60495e72011-04-07 14:10:04 +02006850 }
Dimitri Sivanicha841f8c2012-06-05 13:44:36 -05006851 set_domain_attribute(sd, attr);
Peter Zijlstra2c402dc2011-04-07 14:10:01 +02006852
6853 return sd;
6854}
6855
Mike Travis7c16ec52008-04-04 18:11:11 -07006856/*
Dinakar Guniguntala1a20ff22005-06-25 14:57:33 -07006857 * Build sched domains for a given set of cpus and attach the sched domains
6858 * to the individual cpus
Linus Torvalds1da177e2005-04-16 15:20:36 -07006859 */
Peter Zijlstradce840a2011-04-07 14:09:50 +02006860static int build_sched_domains(const struct cpumask *cpu_map,
6861 struct sched_domain_attr *attr)
Linus Torvalds1da177e2005-04-16 15:20:36 -07006862{
Viresh Kumar1c632162013-06-10 16:27:18 +05306863 enum s_alloc alloc_state;
Peter Zijlstradce840a2011-04-07 14:09:50 +02006864 struct sched_domain *sd;
Andreas Herrmann49a02c52009-08-18 12:51:52 +02006865 struct s_data d;
Peter Zijlstra822ff792011-04-07 14:09:51 +02006866 int i, ret = -ENOMEM;
Rusty Russell3404c8d2008-11-25 02:35:03 +10306867
Andreas Herrmann2109b992009-08-18 12:53:00 +02006868 alloc_state = __visit_domain_allocation_hell(&d, cpu_map);
6869 if (alloc_state != sa_rootdomain)
6870 goto error;
Mike Travis7c16ec52008-04-04 18:11:11 -07006871
Peter Zijlstradce840a2011-04-07 14:09:50 +02006872 /* Set up domains for cpus specified by the cpu_map. */
Rusty Russellabcd0832008-11-25 02:35:02 +10306873 for_each_cpu(i, cpu_map) {
Peter Zijlstraeb7a74e62011-04-07 14:10:00 +02006874 struct sched_domain_topology_level *tl;
6875
Peter Zijlstra3bd65a82011-04-07 14:09:54 +02006876 sd = NULL;
Viresh Kumar27723a62013-06-10 16:27:20 +05306877 for_each_sd_topology(tl) {
Viresh Kumar4a850cb2013-06-04 16:12:43 +05306878 sd = build_sched_domain(tl, cpu_map, attr, sd, i);
Viresh Kumar22da9562013-06-04 15:41:15 +05306879 if (tl == sched_domain_topology)
6880 *per_cpu_ptr(d.sd, i) = sd;
Peter Zijlstrae3589f62011-07-15 10:35:52 +02006881 if (tl->flags & SDTL_OVERLAP || sched_feat(FORCE_SD_OVERLAP))
6882 sd->flags |= SD_OVERLAP;
Peter Zijlstrad1102352011-07-20 18:42:57 +02006883 if (cpumask_equal(cpu_map, sched_domain_span(sd)))
6884 break;
Peter Zijlstrae3589f62011-07-15 10:35:52 +02006885 }
Peter Zijlstradce840a2011-04-07 14:09:50 +02006886 }
Peter Zijlstra21d42cc2011-04-07 14:09:48 +02006887
Peter Zijlstradce840a2011-04-07 14:09:50 +02006888 /* Build the groups for the domains */
6889 for_each_cpu(i, cpu_map) {
6890 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
6891 sd->span_weight = cpumask_weight(sched_domain_span(sd));
Peter Zijlstrae3589f62011-07-15 10:35:52 +02006892 if (sd->flags & SD_OVERLAP) {
6893 if (build_overlap_sched_groups(sd, i))
6894 goto error;
6895 } else {
6896 if (build_sched_groups(sd, i))
6897 goto error;
6898 }
Peter Zijlstra1cf519022011-04-07 14:09:47 +02006899 }
Peter Zijlstraa06dadb2011-04-07 14:09:44 +02006900 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07006901
Nicolas Pitreced549f2014-05-26 18:19:38 -04006902 /* Calculate CPU capacity for physical packages and nodes */
Peter Zijlstraa9c9a9b2011-04-07 14:09:49 +02006903 for (i = nr_cpumask_bits-1; i >= 0; i--) {
6904 if (!cpumask_test_cpu(i, cpu_map))
6905 continue;
Linus Torvalds1da177e2005-04-16 15:20:36 -07006906
Peter Zijlstradce840a2011-04-07 14:09:50 +02006907 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
6908 claim_allocations(i, sd);
Nicolas Pitre63b2ca32014-05-26 18:19:37 -04006909 init_sched_groups_capacity(i, sd);
Peter Zijlstradce840a2011-04-07 14:09:50 +02006910 }
Siddha, Suresh Bf712c0c2006-07-30 03:02:59 -07006911 }
John Hawkes9c1cfda2005-09-06 15:18:14 -07006912
Linus Torvalds1da177e2005-04-16 15:20:36 -07006913 /* Attach the domains */
Peter Zijlstradce840a2011-04-07 14:09:50 +02006914 rcu_read_lock();
Rusty Russellabcd0832008-11-25 02:35:02 +10306915 for_each_cpu(i, cpu_map) {
Peter Zijlstra21d42cc2011-04-07 14:09:48 +02006916 sd = *per_cpu_ptr(d.sd, i);
Andreas Herrmann49a02c52009-08-18 12:51:52 +02006917 cpu_attach_domain(sd, d.rd, i);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006918 }
Peter Zijlstradce840a2011-04-07 14:09:50 +02006919 rcu_read_unlock();
Srivatsa Vaddagiri51888ca2006-06-27 02:54:38 -07006920
Peter Zijlstra822ff792011-04-07 14:09:51 +02006921 ret = 0;
Srivatsa Vaddagiri51888ca2006-06-27 02:54:38 -07006922error:
Andreas Herrmann2109b992009-08-18 12:53:00 +02006923 __free_domain_allocs(&d, alloc_state, cpu_map);
Peter Zijlstra822ff792011-04-07 14:09:51 +02006924 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07006925}
Paul Jackson029190c2007-10-18 23:40:20 -07006926
Rusty Russellacc3f5d2009-11-03 14:53:40 +10306927static cpumask_var_t *doms_cur; /* current sched domains */
Paul Jackson029190c2007-10-18 23:40:20 -07006928static int ndoms_cur; /* number of sched domains in 'doms_cur' */
Ingo Molnar4285f5942008-05-16 17:47:14 +02006929static struct sched_domain_attr *dattr_cur;
6930 /* attribues of custom domains in 'doms_cur' */
Paul Jackson029190c2007-10-18 23:40:20 -07006931
6932/*
6933 * Special case: If a kmalloc of a doms_cur partition (array of
Rusty Russell42128232008-11-25 02:35:12 +10306934 * cpumask) fails, then fallback to a single sched domain,
6935 * as determined by the single cpumask fallback_doms.
Paul Jackson029190c2007-10-18 23:40:20 -07006936 */
Rusty Russell42128232008-11-25 02:35:12 +10306937static cpumask_var_t fallback_doms;
Paul Jackson029190c2007-10-18 23:40:20 -07006938
Heiko Carstensee79d1b2008-12-09 18:49:50 +01006939/*
6940 * arch_update_cpu_topology lets virtualized architectures update the
6941 * cpu core maps. It is supposed to return 1 if the topology changed
6942 * or 0 if it stayed the same.
6943 */
Gideon Israel Dsouza52f5684c2014-04-07 15:39:20 -07006944int __weak arch_update_cpu_topology(void)
Heiko Carstens22e52b02008-03-12 18:31:59 +01006945{
Heiko Carstensee79d1b2008-12-09 18:49:50 +01006946 return 0;
Heiko Carstens22e52b02008-03-12 18:31:59 +01006947}
6948
Rusty Russellacc3f5d2009-11-03 14:53:40 +10306949cpumask_var_t *alloc_sched_domains(unsigned int ndoms)
6950{
6951 int i;
6952 cpumask_var_t *doms;
6953
6954 doms = kmalloc(sizeof(*doms) * ndoms, GFP_KERNEL);
6955 if (!doms)
6956 return NULL;
6957 for (i = 0; i < ndoms; i++) {
6958 if (!alloc_cpumask_var(&doms[i], GFP_KERNEL)) {
6959 free_sched_domains(doms, i);
6960 return NULL;
6961 }
6962 }
6963 return doms;
6964}
6965
6966void free_sched_domains(cpumask_var_t doms[], unsigned int ndoms)
6967{
6968 unsigned int i;
6969 for (i = 0; i < ndoms; i++)
6970 free_cpumask_var(doms[i]);
6971 kfree(doms);
6972}
6973
Dinakar Guniguntala1a20ff22005-06-25 14:57:33 -07006974/*
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01006975 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
Paul Jackson029190c2007-10-18 23:40:20 -07006976 * For now this just excludes isolated cpus, but could be used to
6977 * exclude other special cases in the future.
Dinakar Guniguntala1a20ff22005-06-25 14:57:33 -07006978 */
Peter Zijlstrac4a88492011-04-07 14:09:42 +02006979static int init_sched_domains(const struct cpumask *cpu_map)
Dinakar Guniguntala1a20ff22005-06-25 14:57:33 -07006980{
Milton Miller73785472007-10-24 18:23:48 +02006981 int err;
6982
Heiko Carstens22e52b02008-03-12 18:31:59 +01006983 arch_update_cpu_topology();
Paul Jackson029190c2007-10-18 23:40:20 -07006984 ndoms_cur = 1;
Rusty Russellacc3f5d2009-11-03 14:53:40 +10306985 doms_cur = alloc_sched_domains(ndoms_cur);
Paul Jackson029190c2007-10-18 23:40:20 -07006986 if (!doms_cur)
Rusty Russellacc3f5d2009-11-03 14:53:40 +10306987 doms_cur = &fallback_doms;
6988 cpumask_andnot(doms_cur[0], cpu_map, cpu_isolated_map);
Peter Zijlstradce840a2011-04-07 14:09:50 +02006989 err = build_sched_domains(doms_cur[0], NULL);
Milton Miller6382bc92007-10-15 17:00:19 +02006990 register_sched_domain_sysctl();
Milton Miller73785472007-10-24 18:23:48 +02006991
6992 return err;
Dinakar Guniguntala1a20ff22005-06-25 14:57:33 -07006993}
6994
Dinakar Guniguntala1a20ff22005-06-25 14:57:33 -07006995/*
6996 * Detach sched domains from a group of cpus specified in cpu_map
6997 * These cpus will now be attached to the NULL domain
6998 */
Rusty Russell96f874e2008-11-25 02:35:14 +10306999static void detach_destroy_domains(const struct cpumask *cpu_map)
Dinakar Guniguntala1a20ff22005-06-25 14:57:33 -07007000{
7001 int i;
7002
Peter Zijlstradce840a2011-04-07 14:09:50 +02007003 rcu_read_lock();
Rusty Russellabcd0832008-11-25 02:35:02 +10307004 for_each_cpu(i, cpu_map)
Gregory Haskins57d885f2008-01-25 21:08:18 +01007005 cpu_attach_domain(NULL, &def_root_domain, i);
Peter Zijlstradce840a2011-04-07 14:09:50 +02007006 rcu_read_unlock();
Dinakar Guniguntala1a20ff22005-06-25 14:57:33 -07007007}
7008
Hidetoshi Seto1d3504f2008-04-15 14:04:23 +09007009/* handle null as "default" */
7010static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur,
7011 struct sched_domain_attr *new, int idx_new)
7012{
7013 struct sched_domain_attr tmp;
7014
7015 /* fast path */
7016 if (!new && !cur)
7017 return 1;
7018
7019 tmp = SD_ATTR_INIT;
7020 return !memcmp(cur ? (cur + idx_cur) : &tmp,
7021 new ? (new + idx_new) : &tmp,
7022 sizeof(struct sched_domain_attr));
7023}
7024
Paul Jackson029190c2007-10-18 23:40:20 -07007025/*
7026 * Partition sched domains as specified by the 'ndoms_new'
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01007027 * cpumasks in the array doms_new[] of cpumasks. This compares
Paul Jackson029190c2007-10-18 23:40:20 -07007028 * doms_new[] to the current sched domain partitioning, doms_cur[].
7029 * It destroys each deleted domain and builds each new domain.
7030 *
Rusty Russellacc3f5d2009-11-03 14:53:40 +10307031 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'.
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01007032 * The masks don't intersect (don't overlap.) We should setup one
7033 * sched domain for each mask. CPUs not in any of the cpumasks will
7034 * not be load balanced. If the same cpumask appears both in the
Paul Jackson029190c2007-10-18 23:40:20 -07007035 * current 'doms_cur' domains and in the new 'doms_new', we can leave
7036 * it as it is.
7037 *
Rusty Russellacc3f5d2009-11-03 14:53:40 +10307038 * The passed in 'doms_new' should be allocated using
7039 * alloc_sched_domains. This routine takes ownership of it and will
7040 * free_sched_domains it when done with it. If the caller failed the
7041 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1,
7042 * and partition_sched_domains() will fallback to the single partition
7043 * 'fallback_doms', it also forces the domains to be rebuilt.
Paul Jackson029190c2007-10-18 23:40:20 -07007044 *
Rusty Russell96f874e2008-11-25 02:35:14 +10307045 * If doms_new == NULL it will be replaced with cpu_online_mask.
Li Zefan700018e2008-11-18 14:02:03 +08007046 * ndoms_new == 0 is a special case for destroying existing domains,
7047 * and it will not create the default domain.
Max Krasnyanskydfb512e2008-08-29 13:11:41 -07007048 *
Paul Jackson029190c2007-10-18 23:40:20 -07007049 * Call with hotplug lock held
7050 */
Rusty Russellacc3f5d2009-11-03 14:53:40 +10307051void partition_sched_domains(int ndoms_new, cpumask_var_t doms_new[],
Hidetoshi Seto1d3504f2008-04-15 14:04:23 +09007052 struct sched_domain_attr *dattr_new)
Paul Jackson029190c2007-10-18 23:40:20 -07007053{
Max Krasnyanskydfb512e2008-08-29 13:11:41 -07007054 int i, j, n;
Heiko Carstensd65bd5e2008-12-09 18:49:51 +01007055 int new_topology;
Paul Jackson029190c2007-10-18 23:40:20 -07007056
Heiko Carstens712555e2008-04-28 11:33:07 +02007057 mutex_lock(&sched_domains_mutex);
Srivatsa Vaddagiria1835612008-01-25 21:08:00 +01007058
Milton Miller73785472007-10-24 18:23:48 +02007059 /* always unregister in case we don't destroy any domains */
7060 unregister_sched_domain_sysctl();
7061
Heiko Carstensd65bd5e2008-12-09 18:49:51 +01007062 /* Let architecture update cpu core mappings. */
7063 new_topology = arch_update_cpu_topology();
7064
Max Krasnyanskydfb512e2008-08-29 13:11:41 -07007065 n = doms_new ? ndoms_new : 0;
Paul Jackson029190c2007-10-18 23:40:20 -07007066
7067 /* Destroy deleted domains */
7068 for (i = 0; i < ndoms_cur; i++) {
Heiko Carstensd65bd5e2008-12-09 18:49:51 +01007069 for (j = 0; j < n && !new_topology; j++) {
Rusty Russellacc3f5d2009-11-03 14:53:40 +10307070 if (cpumask_equal(doms_cur[i], doms_new[j])
Hidetoshi Seto1d3504f2008-04-15 14:04:23 +09007071 && dattrs_equal(dattr_cur, i, dattr_new, j))
Paul Jackson029190c2007-10-18 23:40:20 -07007072 goto match1;
7073 }
7074 /* no match - a current sched domain not in new doms_new[] */
Rusty Russellacc3f5d2009-11-03 14:53:40 +10307075 detach_destroy_domains(doms_cur[i]);
Paul Jackson029190c2007-10-18 23:40:20 -07007076match1:
7077 ;
7078 }
7079
Xiaotian Fengc8d2d472013-08-06 20:06:42 +08007080 n = ndoms_cur;
Max Krasnyanskye761b772008-07-15 04:43:49 -07007081 if (doms_new == NULL) {
Xiaotian Fengc8d2d472013-08-06 20:06:42 +08007082 n = 0;
Rusty Russellacc3f5d2009-11-03 14:53:40 +10307083 doms_new = &fallback_doms;
Peter Zijlstra6ad4c182009-11-25 13:31:39 +01007084 cpumask_andnot(doms_new[0], cpu_active_mask, cpu_isolated_map);
Li Zefanfaa2f982008-11-04 16:20:23 +08007085 WARN_ON_ONCE(dattr_new);
Max Krasnyanskye761b772008-07-15 04:43:49 -07007086 }
7087
Paul Jackson029190c2007-10-18 23:40:20 -07007088 /* Build new domains */
7089 for (i = 0; i < ndoms_new; i++) {
Xiaotian Fengc8d2d472013-08-06 20:06:42 +08007090 for (j = 0; j < n && !new_topology; j++) {
Rusty Russellacc3f5d2009-11-03 14:53:40 +10307091 if (cpumask_equal(doms_new[i], doms_cur[j])
Hidetoshi Seto1d3504f2008-04-15 14:04:23 +09007092 && dattrs_equal(dattr_new, i, dattr_cur, j))
Paul Jackson029190c2007-10-18 23:40:20 -07007093 goto match2;
7094 }
7095 /* no match - add a new doms_new */
Peter Zijlstradce840a2011-04-07 14:09:50 +02007096 build_sched_domains(doms_new[i], dattr_new ? dattr_new + i : NULL);
Paul Jackson029190c2007-10-18 23:40:20 -07007097match2:
7098 ;
7099 }
7100
7101 /* Remember the new sched domains */
Rusty Russellacc3f5d2009-11-03 14:53:40 +10307102 if (doms_cur != &fallback_doms)
7103 free_sched_domains(doms_cur, ndoms_cur);
Hidetoshi Seto1d3504f2008-04-15 14:04:23 +09007104 kfree(dattr_cur); /* kfree(NULL) is safe */
Paul Jackson029190c2007-10-18 23:40:20 -07007105 doms_cur = doms_new;
Hidetoshi Seto1d3504f2008-04-15 14:04:23 +09007106 dattr_cur = dattr_new;
Paul Jackson029190c2007-10-18 23:40:20 -07007107 ndoms_cur = ndoms_new;
Milton Miller73785472007-10-24 18:23:48 +02007108
7109 register_sched_domain_sysctl();
Srivatsa Vaddagiria1835612008-01-25 21:08:00 +01007110
Heiko Carstens712555e2008-04-28 11:33:07 +02007111 mutex_unlock(&sched_domains_mutex);
Paul Jackson029190c2007-10-18 23:40:20 -07007112}
7113
Srivatsa S. Bhatd35be8b2012-05-24 19:46:26 +05307114static int num_cpus_frozen; /* used to mark begin/end of suspend/resume */
7115
Linus Torvalds1da177e2005-04-16 15:20:36 -07007116/*
Tejun Heo3a101d02010-06-08 21:40:36 +02007117 * Update cpusets according to cpu_active mask. If cpusets are
7118 * disabled, cpuset_update_active_cpus() becomes a simple wrapper
7119 * around partition_sched_domains().
Srivatsa S. Bhatd35be8b2012-05-24 19:46:26 +05307120 *
7121 * If we come here as part of a suspend/resume, don't touch cpusets because we
7122 * want to restore it back to its original state upon resume anyway.
Linus Torvalds1da177e2005-04-16 15:20:36 -07007123 */
Tejun Heo0b2e9182010-06-21 23:53:31 +02007124static int cpuset_cpu_active(struct notifier_block *nfb, unsigned long action,
7125 void *hcpu)
Linus Torvalds1da177e2005-04-16 15:20:36 -07007126{
Srivatsa S. Bhatd35be8b2012-05-24 19:46:26 +05307127 switch (action) {
7128 case CPU_ONLINE_FROZEN:
7129 case CPU_DOWN_FAILED_FROZEN:
7130
7131 /*
7132 * num_cpus_frozen tracks how many CPUs are involved in suspend
7133 * resume sequence. As long as this is not the last online
7134 * operation in the resume sequence, just build a single sched
7135 * domain, ignoring cpusets.
7136 */
7137 num_cpus_frozen--;
7138 if (likely(num_cpus_frozen)) {
7139 partition_sched_domains(1, NULL, NULL);
7140 break;
7141 }
7142
7143 /*
7144 * This is the last CPU online operation. So fall through and
7145 * restore the original sched domains by considering the
7146 * cpuset configurations.
7147 */
7148
Max Krasnyanskye761b772008-07-15 04:43:49 -07007149 case CPU_ONLINE:
Srivatsa S. Bhat7ddf96b2012-05-24 19:46:55 +05307150 cpuset_update_active_cpus(true);
Srivatsa S. Bhatd35be8b2012-05-24 19:46:26 +05307151 break;
Max Krasnyanskye761b772008-07-15 04:43:49 -07007152 default:
7153 return NOTIFY_DONE;
7154 }
Srivatsa S. Bhatd35be8b2012-05-24 19:46:26 +05307155 return NOTIFY_OK;
Max Krasnyanskye761b772008-07-15 04:43:49 -07007156}
Tejun Heo3a101d02010-06-08 21:40:36 +02007157
Tejun Heo0b2e9182010-06-21 23:53:31 +02007158static int cpuset_cpu_inactive(struct notifier_block *nfb, unsigned long action,
7159 void *hcpu)
Tejun Heo3a101d02010-06-08 21:40:36 +02007160{
Juri Lelli3c18d442015-03-31 09:53:37 +01007161 unsigned long flags;
7162 long cpu = (long)hcpu;
7163 struct dl_bw *dl_b;
Omar Sandoval533445c2015-05-04 03:09:36 -07007164 bool overflow;
7165 int cpus;
Juri Lelli3c18d442015-03-31 09:53:37 +01007166
Omar Sandoval533445c2015-05-04 03:09:36 -07007167 switch (action) {
Tejun Heo3a101d02010-06-08 21:40:36 +02007168 case CPU_DOWN_PREPARE:
Omar Sandoval533445c2015-05-04 03:09:36 -07007169 rcu_read_lock_sched();
7170 dl_b = dl_bw_of(cpu);
Juri Lelli3c18d442015-03-31 09:53:37 +01007171
Omar Sandoval533445c2015-05-04 03:09:36 -07007172 raw_spin_lock_irqsave(&dl_b->lock, flags);
7173 cpus = dl_bw_cpus(cpu);
7174 overflow = __dl_overflow(dl_b, cpus, 0, 0);
7175 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
Juri Lelli3c18d442015-03-31 09:53:37 +01007176
Omar Sandoval533445c2015-05-04 03:09:36 -07007177 rcu_read_unlock_sched();
Juri Lelli3c18d442015-03-31 09:53:37 +01007178
Omar Sandoval533445c2015-05-04 03:09:36 -07007179 if (overflow)
7180 return notifier_from_errno(-EBUSY);
Srivatsa S. Bhat7ddf96b2012-05-24 19:46:55 +05307181 cpuset_update_active_cpus(false);
Srivatsa S. Bhatd35be8b2012-05-24 19:46:26 +05307182 break;
7183 case CPU_DOWN_PREPARE_FROZEN:
7184 num_cpus_frozen++;
7185 partition_sched_domains(1, NULL, NULL);
7186 break;
Tejun Heo3a101d02010-06-08 21:40:36 +02007187 default:
7188 return NOTIFY_DONE;
7189 }
Srivatsa S. Bhatd35be8b2012-05-24 19:46:26 +05307190 return NOTIFY_OK;
Tejun Heo3a101d02010-06-08 21:40:36 +02007191}
Max Krasnyanskye761b772008-07-15 04:43:49 -07007192
Linus Torvalds1da177e2005-04-16 15:20:36 -07007193void __init sched_init_smp(void)
7194{
Rusty Russelldcc30a32008-11-25 02:35:12 +10307195 cpumask_var_t non_isolated_cpus;
7196
7197 alloc_cpumask_var(&non_isolated_cpus, GFP_KERNEL);
Yong Zhangcb5fd132009-09-14 20:20:16 +08007198 alloc_cpumask_var(&fallback_doms, GFP_KERNEL);
Nick Piggin5c1e1762006-10-03 01:14:04 -07007199
Chris Metcalf8cb97642015-05-06 18:04:26 +02007200 /* nohz_full won't take effect without isolating the cpus. */
7201 tick_nohz_full_add_cpus_to(cpu_isolated_map);
7202
Peter Zijlstracb83b622012-04-17 15:49:36 +02007203 sched_init_numa();
7204
Peter Zijlstra6acce3e2013-10-11 14:38:20 +02007205 /*
7206 * There's no userspace yet to cause hotplug operations; hence all the
7207 * cpu masks are stable and all blatant races in the below code cannot
7208 * happen.
7209 */
Heiko Carstens712555e2008-04-28 11:33:07 +02007210 mutex_lock(&sched_domains_mutex);
Peter Zijlstrac4a88492011-04-07 14:09:42 +02007211 init_sched_domains(cpu_active_mask);
Rusty Russelldcc30a32008-11-25 02:35:12 +10307212 cpumask_andnot(non_isolated_cpus, cpu_possible_mask, cpu_isolated_map);
7213 if (cpumask_empty(non_isolated_cpus))
7214 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus);
Heiko Carstens712555e2008-04-28 11:33:07 +02007215 mutex_unlock(&sched_domains_mutex);
Max Krasnyanskye761b772008-07-15 04:43:49 -07007216
Tang Chen301a5cb2012-09-25 21:12:31 +08007217 hotcpu_notifier(sched_domains_numa_masks_update, CPU_PRI_SCHED_ACTIVE);
Tejun Heo3a101d02010-06-08 21:40:36 +02007218 hotcpu_notifier(cpuset_cpu_active, CPU_PRI_CPUSET_ACTIVE);
7219 hotcpu_notifier(cpuset_cpu_inactive, CPU_PRI_CPUSET_INACTIVE);
Max Krasnyanskye761b772008-07-15 04:43:49 -07007220
Peter Zijlstrab328ca12008-04-29 10:02:46 +02007221 init_hrtick();
Nick Piggin5c1e1762006-10-03 01:14:04 -07007222
7223 /* Move init over to a non-isolated CPU */
Rusty Russelldcc30a32008-11-25 02:35:12 +10307224 if (set_cpus_allowed_ptr(current, non_isolated_cpus) < 0)
Nick Piggin5c1e1762006-10-03 01:14:04 -07007225 BUG();
Ingo Molnar19978ca2007-11-09 22:39:38 +01007226 sched_init_granularity();
Rusty Russelldcc30a32008-11-25 02:35:12 +10307227 free_cpumask_var(non_isolated_cpus);
Rusty Russell42128232008-11-25 02:35:12 +10307228
Rusty Russell0e3900e2008-11-25 02:35:13 +10307229 init_sched_rt_class();
Juri Lelli1baca4c2013-11-07 14:43:38 +01007230 init_sched_dl_class();
Linus Torvalds1da177e2005-04-16 15:20:36 -07007231}
7232#else
7233void __init sched_init_smp(void)
7234{
Ingo Molnar19978ca2007-11-09 22:39:38 +01007235 sched_init_granularity();
Linus Torvalds1da177e2005-04-16 15:20:36 -07007236}
7237#endif /* CONFIG_SMP */
7238
7239int in_sched_functions(unsigned long addr)
7240{
Linus Torvalds1da177e2005-04-16 15:20:36 -07007241 return in_lock_functions(addr) ||
7242 (addr >= (unsigned long)__sched_text_start
7243 && addr < (unsigned long)__sched_text_end);
7244}
7245
Peter Zijlstra029632f2011-10-25 10:00:11 +02007246#ifdef CONFIG_CGROUP_SCHED
Li Zefan27b4b932013-03-05 16:07:52 +08007247/*
7248 * Default task group.
7249 * Every task in system belongs to this group at bootup.
7250 */
Peter Zijlstra029632f2011-10-25 10:00:11 +02007251struct task_group root_task_group;
Mike Galbraith35cf4e52012-08-07 05:00:13 +02007252LIST_HEAD(task_groups);
Peter Zijlstrafa85ae22008-01-25 21:08:29 +01007253#endif
7254
Joonsoo Kime6252c32013-04-23 17:27:41 +09007255DECLARE_PER_CPU(cpumask_var_t, load_balance_mask);
Peter Zijlstra6f505b12008-01-25 21:08:30 +01007256
Linus Torvalds1da177e2005-04-16 15:20:36 -07007257void __init sched_init(void)
7258{
Ingo Molnardd41f592007-07-09 18:51:59 +02007259 int i, j;
Mike Travis434d53b2008-04-04 18:11:04 -07007260 unsigned long alloc_size = 0, ptr;
7261
7262#ifdef CONFIG_FAIR_GROUP_SCHED
7263 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
7264#endif
7265#ifdef CONFIG_RT_GROUP_SCHED
7266 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
7267#endif
Mike Travis434d53b2008-04-04 18:11:04 -07007268 if (alloc_size) {
Pekka Enberg36b7b6d2009-06-10 23:42:36 +03007269 ptr = (unsigned long)kzalloc(alloc_size, GFP_NOWAIT);
Mike Travis434d53b2008-04-04 18:11:04 -07007270
7271#ifdef CONFIG_FAIR_GROUP_SCHED
Yong Zhang07e06b02011-01-07 15:17:36 +08007272 root_task_group.se = (struct sched_entity **)ptr;
Mike Travis434d53b2008-04-04 18:11:04 -07007273 ptr += nr_cpu_ids * sizeof(void **);
7274
Yong Zhang07e06b02011-01-07 15:17:36 +08007275 root_task_group.cfs_rq = (struct cfs_rq **)ptr;
Mike Travis434d53b2008-04-04 18:11:04 -07007276 ptr += nr_cpu_ids * sizeof(void **);
Peter Zijlstraeff766a2008-04-19 19:45:00 +02007277
Dhaval Giani6d6bc0a2008-05-30 14:23:45 +02007278#endif /* CONFIG_FAIR_GROUP_SCHED */
Mike Travis434d53b2008-04-04 18:11:04 -07007279#ifdef CONFIG_RT_GROUP_SCHED
Yong Zhang07e06b02011-01-07 15:17:36 +08007280 root_task_group.rt_se = (struct sched_rt_entity **)ptr;
Mike Travis434d53b2008-04-04 18:11:04 -07007281 ptr += nr_cpu_ids * sizeof(void **);
7282
Yong Zhang07e06b02011-01-07 15:17:36 +08007283 root_task_group.rt_rq = (struct rt_rq **)ptr;
Peter Zijlstraeff766a2008-04-19 19:45:00 +02007284 ptr += nr_cpu_ids * sizeof(void **);
7285
Dhaval Giani6d6bc0a2008-05-30 14:23:45 +02007286#endif /* CONFIG_RT_GROUP_SCHED */
Mike Travis434d53b2008-04-04 18:11:04 -07007287 }
Alex Thorltonb74e6272014-12-18 12:44:30 -06007288#ifdef CONFIG_CPUMASK_OFFSTACK
7289 for_each_possible_cpu(i) {
7290 per_cpu(load_balance_mask, i) = (cpumask_var_t)kzalloc_node(
7291 cpumask_size(), GFP_KERNEL, cpu_to_node(i));
7292 }
7293#endif /* CONFIG_CPUMASK_OFFSTACK */
Ingo Molnardd41f592007-07-09 18:51:59 +02007294
Dario Faggioli332ac172013-11-07 14:43:45 +01007295 init_rt_bandwidth(&def_rt_bandwidth,
7296 global_rt_period(), global_rt_runtime());
7297 init_dl_bandwidth(&def_dl_bandwidth,
Peter Zijlstra17248132013-12-17 12:44:49 +01007298 global_rt_period(), global_rt_runtime());
Dario Faggioli332ac172013-11-07 14:43:45 +01007299
Gregory Haskins57d885f2008-01-25 21:08:18 +01007300#ifdef CONFIG_SMP
7301 init_defrootdomain();
7302#endif
7303
Peter Zijlstrad0b27fa2008-04-19 19:44:57 +02007304#ifdef CONFIG_RT_GROUP_SCHED
Yong Zhang07e06b02011-01-07 15:17:36 +08007305 init_rt_bandwidth(&root_task_group.rt_bandwidth,
Peter Zijlstrad0b27fa2008-04-19 19:44:57 +02007306 global_rt_period(), global_rt_runtime());
Dhaval Giani6d6bc0a2008-05-30 14:23:45 +02007307#endif /* CONFIG_RT_GROUP_SCHED */
Peter Zijlstrad0b27fa2008-04-19 19:44:57 +02007308
Dhaval Giani7c941432010-01-20 13:26:18 +01007309#ifdef CONFIG_CGROUP_SCHED
Yong Zhang07e06b02011-01-07 15:17:36 +08007310 list_add(&root_task_group.list, &task_groups);
7311 INIT_LIST_HEAD(&root_task_group.children);
Glauber Costaf4d6f6c2011-11-01 19:19:07 -02007312 INIT_LIST_HEAD(&root_task_group.siblings);
Mike Galbraith5091faa2010-11-30 14:18:03 +01007313 autogroup_init(&init_task);
Glauber Costa54c707e2011-11-28 14:45:19 -02007314
Dhaval Giani7c941432010-01-20 13:26:18 +01007315#endif /* CONFIG_CGROUP_SCHED */
Peter Zijlstra6f505b12008-01-25 21:08:30 +01007316
KAMEZAWA Hiroyuki0a945022006-03-28 01:56:37 -08007317 for_each_possible_cpu(i) {
Ingo Molnar70b97a72006-07-03 00:25:42 -07007318 struct rq *rq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07007319
7320 rq = cpu_rq(i);
Thomas Gleixner05fa7852009-11-17 14:28:38 +01007321 raw_spin_lock_init(&rq->lock);
Nick Piggin78979862005-06-25 14:57:13 -07007322 rq->nr_running = 0;
Thomas Gleixnerdce48a82009-04-11 10:43:41 +02007323 rq->calc_load_active = 0;
7324 rq->calc_load_update = jiffies + LOAD_FREQ;
Jan H. Schönherracb5a9b2011-07-14 18:32:43 +02007325 init_cfs_rq(&rq->cfs);
Abel Vesa07c54f72015-03-03 13:50:27 +02007326 init_rt_rq(&rq->rt);
7327 init_dl_rq(&rq->dl);
Peter Zijlstra6f505b12008-01-25 21:08:30 +01007328#ifdef CONFIG_FAIR_GROUP_SCHED
Peter Zijlstra029632f2011-10-25 10:00:11 +02007329 root_task_group.shares = ROOT_TASK_GROUP_LOAD;
Peter Zijlstra6f505b12008-01-25 21:08:30 +01007330 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
Dhaval Giani354d60c2008-04-19 19:44:59 +02007331 /*
Yong Zhang07e06b02011-01-07 15:17:36 +08007332 * How much cpu bandwidth does root_task_group get?
Dhaval Giani354d60c2008-04-19 19:44:59 +02007333 *
7334 * In case of task-groups formed thr' the cgroup filesystem, it
7335 * gets 100% of the cpu resources in the system. This overall
7336 * system cpu resource is divided among the tasks of
Yong Zhang07e06b02011-01-07 15:17:36 +08007337 * root_task_group and its child task-groups in a fair manner,
Dhaval Giani354d60c2008-04-19 19:44:59 +02007338 * based on each entity's (task or task-group's) weight
7339 * (se->load.weight).
7340 *
Yong Zhang07e06b02011-01-07 15:17:36 +08007341 * In other words, if root_task_group has 10 tasks of weight
Dhaval Giani354d60c2008-04-19 19:44:59 +02007342 * 1024) and two child groups A0 and A1 (of weight 1024 each),
7343 * then A0's share of the cpu resource is:
7344 *
Ingo Molnar0d905bc2009-05-04 19:13:30 +02007345 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
Dhaval Giani354d60c2008-04-19 19:44:59 +02007346 *
Yong Zhang07e06b02011-01-07 15:17:36 +08007347 * We achieve this by letting root_task_group's tasks sit
7348 * directly in rq->cfs (i.e root_task_group->se[] = NULL).
Dhaval Giani354d60c2008-04-19 19:44:59 +02007349 */
Paul Turnerab84d312011-07-21 09:43:28 -07007350 init_cfs_bandwidth(&root_task_group.cfs_bandwidth);
Yong Zhang07e06b02011-01-07 15:17:36 +08007351 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, NULL);
Dhaval Giani354d60c2008-04-19 19:44:59 +02007352#endif /* CONFIG_FAIR_GROUP_SCHED */
7353
7354 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
Peter Zijlstra052f1dc2008-02-13 15:45:40 +01007355#ifdef CONFIG_RT_GROUP_SCHED
Yong Zhang07e06b02011-01-07 15:17:36 +08007356 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, NULL);
Peter Zijlstra6f505b12008-01-25 21:08:30 +01007357#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07007358
Ingo Molnardd41f592007-07-09 18:51:59 +02007359 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
7360 rq->cpu_load[j] = 0;
Venkatesh Pallipadifdf3e952010-05-17 18:14:43 -07007361
7362 rq->last_load_update_tick = jiffies;
7363
Linus Torvalds1da177e2005-04-16 15:20:36 -07007364#ifdef CONFIG_SMP
Nick Piggin41c7ce92005-06-25 14:57:24 -07007365 rq->sd = NULL;
Gregory Haskins57d885f2008-01-25 21:08:18 +01007366 rq->rd = NULL;
Vincent Guittotca6d75e2015-02-27 16:54:09 +01007367 rq->cpu_capacity = rq->cpu_capacity_orig = SCHED_CAPACITY_SCALE;
Peter Zijlstrae3fca9e2015-06-11 14:46:37 +02007368 rq->balance_callback = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07007369 rq->active_balance = 0;
Ingo Molnardd41f592007-07-09 18:51:59 +02007370 rq->next_balance = jiffies;
Linus Torvalds1da177e2005-04-16 15:20:36 -07007371 rq->push_cpu = 0;
Christoph Lameter0a2966b2006-09-25 23:30:51 -07007372 rq->cpu = i;
Gregory Haskins1f11eb62008-06-04 15:04:05 -04007373 rq->online = 0;
Mike Galbraitheae0c9d2009-11-10 03:50:02 +01007374 rq->idle_stamp = 0;
7375 rq->avg_idle = 2*sysctl_sched_migration_cost;
Jason Low9bd721c2013-09-13 11:26:52 -07007376 rq->max_idle_balance_cost = sysctl_sched_migration_cost;
Peter Zijlstra367456c2012-02-20 21:49:09 +01007377
7378 INIT_LIST_HEAD(&rq->cfs_tasks);
7379
Gregory Haskinsdc938522008-01-25 21:08:26 +01007380 rq_attach_root(rq, &def_root_domain);
Frederic Weisbecker3451d022011-08-10 23:21:01 +02007381#ifdef CONFIG_NO_HZ_COMMON
Suresh Siddha1c792db2011-12-01 17:07:32 -08007382 rq->nohz_flags = 0;
Venkatesh Pallipadi83cd4fe2010-05-21 17:09:41 -07007383#endif
Frederic Weisbecker265f22a2013-05-03 03:39:05 +02007384#ifdef CONFIG_NO_HZ_FULL
7385 rq->last_sched_tick = 0;
7386#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07007387#endif
Peter Zijlstra8f4d37e2008-01-25 21:08:29 +01007388 init_rq_hrtick(rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07007389 atomic_set(&rq->nr_iowait, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07007390 }
7391
Peter Williams2dd73a42006-06-27 02:54:34 -07007392 set_load_weight(&init_task);
Heiko Carstensb50f60c2006-07-30 03:03:52 -07007393
Avi Kivitye107be32007-07-26 13:40:43 +02007394#ifdef CONFIG_PREEMPT_NOTIFIERS
7395 INIT_HLIST_HEAD(&init_task.preempt_notifiers);
7396#endif
7397
Linus Torvalds1da177e2005-04-16 15:20:36 -07007398 /*
7399 * The boot idle thread does lazy MMU switching as well:
7400 */
7401 atomic_inc(&init_mm.mm_count);
7402 enter_lazy_tlb(&init_mm, current);
7403
7404 /*
Yao Dongdong1b537c72014-12-29 14:41:43 +08007405 * During early bootup we pretend to be a normal task:
7406 */
7407 current->sched_class = &fair_sched_class;
7408
7409 /*
Linus Torvalds1da177e2005-04-16 15:20:36 -07007410 * Make us the idle thread. Technically, schedule() should not be
7411 * called from this thread, however somewhere below it might be,
7412 * but because we are the idle thread, we just pick up running again
7413 * when this runqueue becomes "idle".
7414 */
7415 init_idle(current, smp_processor_id());
Thomas Gleixnerdce48a82009-04-11 10:43:41 +02007416
7417 calc_load_update = jiffies + LOAD_FREQ;
7418
Rusty Russellbf4d83f2008-11-25 09:57:51 +10307419#ifdef CONFIG_SMP
Peter Zijlstra4cb98832011-04-07 14:09:58 +02007420 zalloc_cpumask_var(&sched_domains_tmpmask, GFP_NOWAIT);
Rusty Russellbdddd292009-12-02 14:09:16 +10307421 /* May be allocated at isolcpus cmdline parse time */
7422 if (cpu_isolated_map == NULL)
7423 zalloc_cpumask_var(&cpu_isolated_map, GFP_NOWAIT);
Thomas Gleixner29d5e042012-04-20 13:05:45 +00007424 idle_thread_set_boot_cpu();
Corey Minyarda803f022014-05-08 13:47:39 -05007425 set_cpu_rq_start_time();
Peter Zijlstra029632f2011-10-25 10:00:11 +02007426#endif
7427 init_sched_fair_class();
Rusty Russell6a7b3dc2008-11-25 02:35:04 +10307428
Ingo Molnar6892b752008-02-13 14:02:36 +01007429 scheduler_running = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07007430}
7431
Frederic Weisbeckerd902db12011-06-08 19:31:56 +02007432#ifdef CONFIG_DEBUG_ATOMIC_SLEEP
Frederic Weisbeckere4aafea2009-07-16 15:44:29 +02007433static inline int preempt_count_equals(int preempt_offset)
7434{
Frederic Weisbecker234da7b2009-12-16 20:21:05 +01007435 int nested = (preempt_count() & ~PREEMPT_ACTIVE) + rcu_preempt_depth();
Frederic Weisbeckere4aafea2009-07-16 15:44:29 +02007436
Arnd Bergmann4ba82162011-01-25 22:52:22 +01007437 return (nested == preempt_offset);
Frederic Weisbeckere4aafea2009-07-16 15:44:29 +02007438}
7439
Simon Kagstromd8948372009-12-23 11:08:18 +01007440void __might_sleep(const char *file, int line, int preempt_offset)
Linus Torvalds1da177e2005-04-16 15:20:36 -07007441{
Peter Zijlstra8eb23b92014-09-24 10:18:55 +02007442 /*
7443 * Blocking primitives will set (and therefore destroy) current->state,
7444 * since we will exit with TASK_RUNNING make sure we enter with it,
7445 * otherwise we will destroy state.
7446 */
Linus Torvalds00845eb2015-02-01 12:23:32 -08007447 WARN_ONCE(current->state != TASK_RUNNING && current->task_state_change,
Peter Zijlstra8eb23b92014-09-24 10:18:55 +02007448 "do not call blocking ops when !TASK_RUNNING; "
7449 "state=%lx set at [<%p>] %pS\n",
7450 current->state,
7451 (void *)current->task_state_change,
Linus Torvalds00845eb2015-02-01 12:23:32 -08007452 (void *)current->task_state_change);
Peter Zijlstra8eb23b92014-09-24 10:18:55 +02007453
Peter Zijlstra34274452014-09-24 10:18:56 +02007454 ___might_sleep(file, line, preempt_offset);
7455}
7456EXPORT_SYMBOL(__might_sleep);
7457
7458void ___might_sleep(const char *file, int line, int preempt_offset)
7459{
Linus Torvalds1da177e2005-04-16 15:20:36 -07007460 static unsigned long prev_jiffy; /* ratelimiting */
7461
Paul E. McKenneyb3fbab02011-05-24 08:31:09 -07007462 rcu_sleep_check(); /* WARN_ON_ONCE() by default, no rate limit reqd. */
Thomas Gleixnerdb273be2014-02-07 20:58:38 +01007463 if ((preempt_count_equals(preempt_offset) && !irqs_disabled() &&
7464 !is_idle_task(current)) ||
Frederic Weisbeckere4aafea2009-07-16 15:44:29 +02007465 system_state != SYSTEM_RUNNING || oops_in_progress)
Ingo Molnaraef745f2008-08-28 11:34:43 +02007466 return;
7467 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
7468 return;
7469 prev_jiffy = jiffies;
7470
Peter Zijlstra3df0fc52009-12-20 14:23:57 +01007471 printk(KERN_ERR
7472 "BUG: sleeping function called from invalid context at %s:%d\n",
7473 file, line);
7474 printk(KERN_ERR
7475 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
7476 in_atomic(), irqs_disabled(),
7477 current->pid, current->comm);
Ingo Molnaraef745f2008-08-28 11:34:43 +02007478
Eric Sandeena8b686b2014-12-16 16:25:28 -06007479 if (task_stack_end_corrupted(current))
7480 printk(KERN_EMERG "Thread overran stack, or stack corrupted\n");
7481
Ingo Molnaraef745f2008-08-28 11:34:43 +02007482 debug_show_held_locks(current);
7483 if (irqs_disabled())
7484 print_irqtrace_events(current);
Thomas Gleixner8f47b182014-02-07 20:58:39 +01007485#ifdef CONFIG_DEBUG_PREEMPT
7486 if (!preempt_count_equals(preempt_offset)) {
7487 pr_err("Preemption disabled at:");
7488 print_ip_sym(current->preempt_disable_ip);
7489 pr_cont("\n");
7490 }
7491#endif
Ingo Molnaraef745f2008-08-28 11:34:43 +02007492 dump_stack();
Linus Torvalds1da177e2005-04-16 15:20:36 -07007493}
Peter Zijlstra34274452014-09-24 10:18:56 +02007494EXPORT_SYMBOL(___might_sleep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07007495#endif
7496
7497#ifdef CONFIG_MAGIC_SYSRQ
7498void normalize_rt_tasks(void)
7499{
Ingo Molnara0f98a12007-06-17 18:37:45 +02007500 struct task_struct *g, *p;
Peter Zijlstradbc7f062015-06-11 14:46:38 +02007501 struct sched_attr attr = {
7502 .sched_policy = SCHED_NORMAL,
7503 };
Linus Torvalds1da177e2005-04-16 15:20:36 -07007504
Oleg Nesterov3472eaa2014-09-21 21:33:38 +02007505 read_lock(&tasklist_lock);
Oleg Nesterov5d07f422014-08-13 21:19:53 +02007506 for_each_process_thread(g, p) {
Ingo Molnar178be792007-10-15 17:00:18 +02007507 /*
7508 * Only normalize user tasks:
7509 */
Oleg Nesterov3472eaa2014-09-21 21:33:38 +02007510 if (p->flags & PF_KTHREAD)
Ingo Molnar178be792007-10-15 17:00:18 +02007511 continue;
7512
Ingo Molnardd41f592007-07-09 18:51:59 +02007513 p->se.exec_start = 0;
Ingo Molnar6cfb0d52007-08-02 17:41:40 +02007514#ifdef CONFIG_SCHEDSTATS
Lucas De Marchi41acab82010-03-10 23:37:45 -03007515 p->se.statistics.wait_start = 0;
7516 p->se.statistics.sleep_start = 0;
7517 p->se.statistics.block_start = 0;
Ingo Molnar6cfb0d52007-08-02 17:41:40 +02007518#endif
Ingo Molnardd41f592007-07-09 18:51:59 +02007519
Dario Faggioliaab03e02013-11-28 11:14:43 +01007520 if (!dl_task(p) && !rt_task(p)) {
Ingo Molnardd41f592007-07-09 18:51:59 +02007521 /*
7522 * Renice negative nice level userspace
7523 * tasks back to 0:
7524 */
Oleg Nesterov3472eaa2014-09-21 21:33:38 +02007525 if (task_nice(p) < 0)
Ingo Molnardd41f592007-07-09 18:51:59 +02007526 set_user_nice(p, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07007527 continue;
Ingo Molnardd41f592007-07-09 18:51:59 +02007528 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07007529
Peter Zijlstradbc7f062015-06-11 14:46:38 +02007530 __sched_setscheduler(p, &attr, false, false);
Oleg Nesterov5d07f422014-08-13 21:19:53 +02007531 }
Oleg Nesterov3472eaa2014-09-21 21:33:38 +02007532 read_unlock(&tasklist_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07007533}
7534
7535#endif /* CONFIG_MAGIC_SYSRQ */
Linus Torvalds1df5c102005-09-12 07:59:21 -07007536
Jason Wessel67fc4e02010-05-20 21:04:21 -05007537#if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB)
Linus Torvalds1df5c102005-09-12 07:59:21 -07007538/*
Jason Wessel67fc4e02010-05-20 21:04:21 -05007539 * These functions are only useful for the IA64 MCA handling, or kdb.
Linus Torvalds1df5c102005-09-12 07:59:21 -07007540 *
7541 * They can only be called when the whole system has been
7542 * stopped - every CPU needs to be quiescent, and no scheduling
7543 * activity can take place. Using them for anything else would
7544 * be a serious bug, and as a result, they aren't even visible
7545 * under any other configuration.
7546 */
7547
7548/**
7549 * curr_task - return the current task for a given cpu.
7550 * @cpu: the processor in question.
7551 *
7552 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
Yacine Belkadie69f6182013-07-12 20:45:47 +02007553 *
7554 * Return: The current task for @cpu.
Linus Torvalds1df5c102005-09-12 07:59:21 -07007555 */
Ingo Molnar36c8b582006-07-03 00:25:41 -07007556struct task_struct *curr_task(int cpu)
Linus Torvalds1df5c102005-09-12 07:59:21 -07007557{
7558 return cpu_curr(cpu);
7559}
7560
Jason Wessel67fc4e02010-05-20 21:04:21 -05007561#endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */
7562
7563#ifdef CONFIG_IA64
Linus Torvalds1df5c102005-09-12 07:59:21 -07007564/**
7565 * set_curr_task - set the current task for a given cpu.
7566 * @cpu: the processor in question.
7567 * @p: the task pointer to set.
7568 *
7569 * Description: This function must only be used when non-maskable interrupts
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01007570 * are serviced on a separate stack. It allows the architecture to switch the
7571 * notion of the current task on a cpu in a non-blocking manner. This function
Linus Torvalds1df5c102005-09-12 07:59:21 -07007572 * must be called with all CPU's synchronized, and interrupts disabled, the
7573 * and caller must save the original value of the current task (see
7574 * curr_task() above) and restore that value before reenabling interrupts and
7575 * re-starting the system.
7576 *
7577 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7578 */
Ingo Molnar36c8b582006-07-03 00:25:41 -07007579void set_curr_task(int cpu, struct task_struct *p)
Linus Torvalds1df5c102005-09-12 07:59:21 -07007580{
7581 cpu_curr(cpu) = p;
7582}
7583
7584#endif
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02007585
Dhaval Giani7c941432010-01-20 13:26:18 +01007586#ifdef CONFIG_CGROUP_SCHED
Peter Zijlstra029632f2011-10-25 10:00:11 +02007587/* task_group_lock serializes the addition/removal of task groups */
7588static DEFINE_SPINLOCK(task_group_lock);
7589
Peter Zijlstrabccbe082008-02-13 15:45:40 +01007590static void free_sched_group(struct task_group *tg)
7591{
7592 free_fair_sched_group(tg);
7593 free_rt_sched_group(tg);
Mike Galbraithe9aa1dd2011-01-05 11:11:25 +01007594 autogroup_free(tg);
Peter Zijlstrabccbe082008-02-13 15:45:40 +01007595 kfree(tg);
7596}
7597
7598/* allocate runqueue etc for a new task group */
Dhaval Gianiec7dc8a2008-04-19 19:44:59 +02007599struct task_group *sched_create_group(struct task_group *parent)
Peter Zijlstrabccbe082008-02-13 15:45:40 +01007600{
7601 struct task_group *tg;
Peter Zijlstrabccbe082008-02-13 15:45:40 +01007602
7603 tg = kzalloc(sizeof(*tg), GFP_KERNEL);
7604 if (!tg)
7605 return ERR_PTR(-ENOMEM);
7606
Dhaval Gianiec7dc8a2008-04-19 19:44:59 +02007607 if (!alloc_fair_sched_group(tg, parent))
Peter Zijlstrabccbe082008-02-13 15:45:40 +01007608 goto err;
7609
Dhaval Gianiec7dc8a2008-04-19 19:44:59 +02007610 if (!alloc_rt_sched_group(tg, parent))
Peter Zijlstrabccbe082008-02-13 15:45:40 +01007611 goto err;
7612
Li Zefanace783b2013-01-24 14:30:48 +08007613 return tg;
7614
7615err:
7616 free_sched_group(tg);
7617 return ERR_PTR(-ENOMEM);
7618}
7619
7620void sched_online_group(struct task_group *tg, struct task_group *parent)
7621{
7622 unsigned long flags;
7623
Peter Zijlstra8ed36992008-02-13 15:45:39 +01007624 spin_lock_irqsave(&task_group_lock, flags);
Peter Zijlstra6f505b12008-01-25 21:08:30 +01007625 list_add_rcu(&tg->list, &task_groups);
Peter Zijlstraf473aa52008-04-19 19:45:00 +02007626
7627 WARN_ON(!parent); /* root should already exist */
7628
7629 tg->parent = parent;
Peter Zijlstraf473aa52008-04-19 19:45:00 +02007630 INIT_LIST_HEAD(&tg->children);
Zhang, Yanmin09f27242030-08-14 15:56:40 +08007631 list_add_rcu(&tg->siblings, &parent->children);
Peter Zijlstra8ed36992008-02-13 15:45:39 +01007632 spin_unlock_irqrestore(&task_group_lock, flags);
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02007633}
7634
Srivatsa Vaddagiri9b5b7752007-10-15 17:00:09 +02007635/* rcu callback to free various structures associated with a task group */
Peter Zijlstra6f505b12008-01-25 21:08:30 +01007636static void free_sched_group_rcu(struct rcu_head *rhp)
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02007637{
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02007638 /* now it should be safe to free those cfs_rqs */
Peter Zijlstra6f505b12008-01-25 21:08:30 +01007639 free_sched_group(container_of(rhp, struct task_group, rcu));
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02007640}
7641
Srivatsa Vaddagiri9b5b7752007-10-15 17:00:09 +02007642/* Destroy runqueue etc associated with a task group */
Ingo Molnar4cf86d72007-10-15 17:00:14 +02007643void sched_destroy_group(struct task_group *tg)
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02007644{
Li Zefanace783b2013-01-24 14:30:48 +08007645 /* wait for possible concurrent references to cfs_rqs complete */
7646 call_rcu(&tg->rcu, free_sched_group_rcu);
7647}
7648
7649void sched_offline_group(struct task_group *tg)
7650{
Peter Zijlstra8ed36992008-02-13 15:45:39 +01007651 unsigned long flags;
Srivatsa Vaddagiri9b5b7752007-10-15 17:00:09 +02007652 int i;
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02007653
Peter Zijlstra3d4b47b2010-11-15 15:47:01 -08007654 /* end participation in shares distribution */
7655 for_each_possible_cpu(i)
Peter Zijlstrabccbe082008-02-13 15:45:40 +01007656 unregister_fair_sched_group(tg, i);
Peter Zijlstra3d4b47b2010-11-15 15:47:01 -08007657
7658 spin_lock_irqsave(&task_group_lock, flags);
Peter Zijlstra6f505b12008-01-25 21:08:30 +01007659 list_del_rcu(&tg->list);
Peter Zijlstraf473aa52008-04-19 19:45:00 +02007660 list_del_rcu(&tg->siblings);
Peter Zijlstra8ed36992008-02-13 15:45:39 +01007661 spin_unlock_irqrestore(&task_group_lock, flags);
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02007662}
7663
Srivatsa Vaddagiri9b5b7752007-10-15 17:00:09 +02007664/* change task's runqueue when it moves between groups.
Ingo Molnar3a252012007-10-15 17:00:12 +02007665 * The caller of this function should have put the task in its new group
7666 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
7667 * reflect its new group.
Srivatsa Vaddagiri9b5b7752007-10-15 17:00:09 +02007668 */
7669void sched_move_task(struct task_struct *tsk)
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02007670{
Peter Zijlstra8323f262012-06-22 13:36:05 +02007671 struct task_group *tg;
Kirill Tkhaida0c1e62014-08-20 13:47:32 +04007672 int queued, running;
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02007673 unsigned long flags;
7674 struct rq *rq;
7675
7676 rq = task_rq_lock(tsk, &flags);
7677
Dmitry Adamushko051a1d12007-12-18 15:21:13 +01007678 running = task_current(rq, tsk);
Kirill Tkhaida0c1e62014-08-20 13:47:32 +04007679 queued = task_on_rq_queued(tsk);
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02007680
Kirill Tkhaida0c1e62014-08-20 13:47:32 +04007681 if (queued)
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02007682 dequeue_task(rq, tsk, 0);
Hiroshi Shimamoto0e1f3482008-03-10 11:01:20 -07007683 if (unlikely(running))
Kirill Tkhaif3cd1c42014-09-12 17:41:40 +04007684 put_prev_task(rq, tsk);
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02007685
Kirill Tkhaif7b8a472014-10-28 08:24:34 +03007686 /*
7687 * All callers are synchronized by task_rq_lock(); we do not use RCU
7688 * which is pointless here. Thus, we pass "true" to task_css_check()
7689 * to prevent lockdep warnings.
7690 */
7691 tg = container_of(task_css_check(tsk, cpu_cgrp_id, true),
Peter Zijlstra8323f262012-06-22 13:36:05 +02007692 struct task_group, css);
7693 tg = autogroup_task_group(tsk, tg);
7694 tsk->sched_task_group = tg;
7695
Peter Zijlstra810b3812008-02-29 15:21:01 -05007696#ifdef CONFIG_FAIR_GROUP_SCHED
Peter Zijlstrab2b5ce02010-10-15 15:24:15 +02007697 if (tsk->sched_class->task_move_group)
Kirill Tkhaida0c1e62014-08-20 13:47:32 +04007698 tsk->sched_class->task_move_group(tsk, queued);
Peter Zijlstrab2b5ce02010-10-15 15:24:15 +02007699 else
Peter Zijlstra810b3812008-02-29 15:21:01 -05007700#endif
Peter Zijlstrab2b5ce02010-10-15 15:24:15 +02007701 set_task_rq(tsk, task_cpu(tsk));
Peter Zijlstra810b3812008-02-29 15:21:01 -05007702
Hiroshi Shimamoto0e1f3482008-03-10 11:01:20 -07007703 if (unlikely(running))
7704 tsk->sched_class->set_curr_task(rq);
Kirill Tkhaida0c1e62014-08-20 13:47:32 +04007705 if (queued)
Peter Zijlstra371fd7e2010-03-24 16:38:48 +01007706 enqueue_task(rq, tsk, 0);
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02007707
Peter Zijlstra0122ec52011-04-05 17:23:51 +02007708 task_rq_unlock(rq, tsk, &flags);
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02007709}
Dhaval Giani7c941432010-01-20 13:26:18 +01007710#endif /* CONFIG_CGROUP_SCHED */
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02007711
Paul Turnera790de92011-07-21 09:43:29 -07007712#ifdef CONFIG_RT_GROUP_SCHED
7713/*
7714 * Ensure that the real time constraints are schedulable.
7715 */
7716static DEFINE_MUTEX(rt_constraints_mutex);
Peter Zijlstra9f0c1e52008-02-13 15:45:39 +01007717
Dhaval Giani521f1a242008-02-28 15:21:56 +05307718/* Must be called with tasklist_lock held */
7719static inline int tg_has_rt_tasks(struct task_group *tg)
7720{
7721 struct task_struct *g, *p;
Peter Zijlstra9a7e0b12008-08-19 12:33:06 +02007722
Peter Zijlstra1fe89e12015-02-09 11:53:18 +01007723 /*
7724 * Autogroups do not have RT tasks; see autogroup_create().
7725 */
7726 if (task_group_is_autogroup(tg))
7727 return 0;
7728
Oleg Nesterov5d07f422014-08-13 21:19:53 +02007729 for_each_process_thread(g, p) {
Oleg Nesterov8651c652014-09-21 21:33:36 +02007730 if (rt_task(p) && task_group(p) == tg)
Dhaval Giani521f1a242008-02-28 15:21:56 +05307731 return 1;
Oleg Nesterov5d07f422014-08-13 21:19:53 +02007732 }
Peter Zijlstra9a7e0b12008-08-19 12:33:06 +02007733
Dhaval Giani521f1a242008-02-28 15:21:56 +05307734 return 0;
7735}
7736
Peter Zijlstra9a7e0b12008-08-19 12:33:06 +02007737struct rt_schedulable_data {
7738 struct task_group *tg;
7739 u64 rt_period;
7740 u64 rt_runtime;
7741};
7742
Paul Turnera790de92011-07-21 09:43:29 -07007743static int tg_rt_schedulable(struct task_group *tg, void *data)
Peter Zijlstra9a7e0b12008-08-19 12:33:06 +02007744{
7745 struct rt_schedulable_data *d = data;
7746 struct task_group *child;
7747 unsigned long total, sum = 0;
7748 u64 period, runtime;
7749
7750 period = ktime_to_ns(tg->rt_bandwidth.rt_period);
7751 runtime = tg->rt_bandwidth.rt_runtime;
7752
7753 if (tg == d->tg) {
7754 period = d->rt_period;
7755 runtime = d->rt_runtime;
7756 }
7757
Peter Zijlstra4653f802008-09-23 15:33:44 +02007758 /*
7759 * Cannot have more runtime than the period.
7760 */
7761 if (runtime > period && runtime != RUNTIME_INF)
7762 return -EINVAL;
7763
7764 /*
7765 * Ensure we don't starve existing RT tasks.
7766 */
Peter Zijlstra9a7e0b12008-08-19 12:33:06 +02007767 if (rt_bandwidth_enabled() && !runtime && tg_has_rt_tasks(tg))
7768 return -EBUSY;
7769
7770 total = to_ratio(period, runtime);
7771
Peter Zijlstra4653f802008-09-23 15:33:44 +02007772 /*
7773 * Nobody can have more than the global setting allows.
7774 */
7775 if (total > to_ratio(global_rt_period(), global_rt_runtime()))
7776 return -EINVAL;
7777
7778 /*
7779 * The sum of our children's runtime should not exceed our own.
7780 */
Peter Zijlstra9a7e0b12008-08-19 12:33:06 +02007781 list_for_each_entry_rcu(child, &tg->children, siblings) {
7782 period = ktime_to_ns(child->rt_bandwidth.rt_period);
7783 runtime = child->rt_bandwidth.rt_runtime;
7784
7785 if (child == d->tg) {
7786 period = d->rt_period;
7787 runtime = d->rt_runtime;
7788 }
7789
7790 sum += to_ratio(period, runtime);
7791 }
7792
7793 if (sum > total)
7794 return -EINVAL;
7795
7796 return 0;
7797}
7798
7799static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
7800{
Paul Turner82774342011-07-21 09:43:35 -07007801 int ret;
7802
Peter Zijlstra9a7e0b12008-08-19 12:33:06 +02007803 struct rt_schedulable_data data = {
7804 .tg = tg,
7805 .rt_period = period,
7806 .rt_runtime = runtime,
7807 };
7808
Paul Turner82774342011-07-21 09:43:35 -07007809 rcu_read_lock();
7810 ret = walk_tg_tree(tg_rt_schedulable, tg_nop, &data);
7811 rcu_read_unlock();
7812
7813 return ret;
Peter Zijlstra9a7e0b12008-08-19 12:33:06 +02007814}
7815
Paul Turnerab84d312011-07-21 09:43:28 -07007816static int tg_set_rt_bandwidth(struct task_group *tg,
Peter Zijlstrad0b27fa2008-04-19 19:44:57 +02007817 u64 rt_period, u64 rt_runtime)
Peter Zijlstra6f505b12008-01-25 21:08:30 +01007818{
Peter Zijlstraac086bc2008-04-19 19:44:58 +02007819 int i, err = 0;
Peter Zijlstra9f0c1e52008-02-13 15:45:39 +01007820
Peter Zijlstra2636ed52015-02-09 12:23:20 +01007821 /*
7822 * Disallowing the root group RT runtime is BAD, it would disallow the
7823 * kernel creating (and or operating) RT threads.
7824 */
7825 if (tg == &root_task_group && rt_runtime == 0)
7826 return -EINVAL;
7827
7828 /* No period doesn't make any sense. */
7829 if (rt_period == 0)
7830 return -EINVAL;
7831
Peter Zijlstra9f0c1e52008-02-13 15:45:39 +01007832 mutex_lock(&rt_constraints_mutex);
Dhaval Giani521f1a242008-02-28 15:21:56 +05307833 read_lock(&tasklist_lock);
Peter Zijlstra9a7e0b12008-08-19 12:33:06 +02007834 err = __rt_schedulable(tg, rt_period, rt_runtime);
7835 if (err)
Dhaval Giani521f1a242008-02-28 15:21:56 +05307836 goto unlock;
Peter Zijlstraac086bc2008-04-19 19:44:58 +02007837
Thomas Gleixner0986b112009-11-17 15:32:06 +01007838 raw_spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock);
Peter Zijlstrad0b27fa2008-04-19 19:44:57 +02007839 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period);
7840 tg->rt_bandwidth.rt_runtime = rt_runtime;
Peter Zijlstraac086bc2008-04-19 19:44:58 +02007841
7842 for_each_possible_cpu(i) {
7843 struct rt_rq *rt_rq = tg->rt_rq[i];
7844
Thomas Gleixner0986b112009-11-17 15:32:06 +01007845 raw_spin_lock(&rt_rq->rt_runtime_lock);
Peter Zijlstraac086bc2008-04-19 19:44:58 +02007846 rt_rq->rt_runtime = rt_runtime;
Thomas Gleixner0986b112009-11-17 15:32:06 +01007847 raw_spin_unlock(&rt_rq->rt_runtime_lock);
Peter Zijlstraac086bc2008-04-19 19:44:58 +02007848 }
Thomas Gleixner0986b112009-11-17 15:32:06 +01007849 raw_spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock);
Peter Zijlstra49246272010-10-17 21:46:10 +02007850unlock:
Dhaval Giani521f1a242008-02-28 15:21:56 +05307851 read_unlock(&tasklist_lock);
Peter Zijlstra9f0c1e52008-02-13 15:45:39 +01007852 mutex_unlock(&rt_constraints_mutex);
7853
7854 return err;
Peter Zijlstra6f505b12008-01-25 21:08:30 +01007855}
7856
Li Zefan25cc7da2013-03-05 16:07:33 +08007857static int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us)
Peter Zijlstrad0b27fa2008-04-19 19:44:57 +02007858{
7859 u64 rt_runtime, rt_period;
7860
7861 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
7862 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC;
7863 if (rt_runtime_us < 0)
7864 rt_runtime = RUNTIME_INF;
7865
Paul Turnerab84d312011-07-21 09:43:28 -07007866 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime);
Peter Zijlstrad0b27fa2008-04-19 19:44:57 +02007867}
7868
Li Zefan25cc7da2013-03-05 16:07:33 +08007869static long sched_group_rt_runtime(struct task_group *tg)
Peter Zijlstra9f0c1e52008-02-13 15:45:39 +01007870{
7871 u64 rt_runtime_us;
7872
Peter Zijlstrad0b27fa2008-04-19 19:44:57 +02007873 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF)
Peter Zijlstra9f0c1e52008-02-13 15:45:39 +01007874 return -1;
7875
Peter Zijlstrad0b27fa2008-04-19 19:44:57 +02007876 rt_runtime_us = tg->rt_bandwidth.rt_runtime;
Peter Zijlstra9f0c1e52008-02-13 15:45:39 +01007877 do_div(rt_runtime_us, NSEC_PER_USEC);
7878 return rt_runtime_us;
7879}
Peter Zijlstrad0b27fa2008-04-19 19:44:57 +02007880
Nicholas Mc Guirece2f5fe2015-05-03 10:51:56 +02007881static int sched_group_set_rt_period(struct task_group *tg, u64 rt_period_us)
Peter Zijlstrad0b27fa2008-04-19 19:44:57 +02007882{
7883 u64 rt_runtime, rt_period;
7884
Nicholas Mc Guirece2f5fe2015-05-03 10:51:56 +02007885 rt_period = rt_period_us * NSEC_PER_USEC;
Peter Zijlstrad0b27fa2008-04-19 19:44:57 +02007886 rt_runtime = tg->rt_bandwidth.rt_runtime;
7887
Paul Turnerab84d312011-07-21 09:43:28 -07007888 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime);
Peter Zijlstrad0b27fa2008-04-19 19:44:57 +02007889}
7890
Li Zefan25cc7da2013-03-05 16:07:33 +08007891static long sched_group_rt_period(struct task_group *tg)
Peter Zijlstrad0b27fa2008-04-19 19:44:57 +02007892{
7893 u64 rt_period_us;
7894
7895 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period);
7896 do_div(rt_period_us, NSEC_PER_USEC);
7897 return rt_period_us;
7898}
Dario Faggioli332ac172013-11-07 14:43:45 +01007899#endif /* CONFIG_RT_GROUP_SCHED */
Peter Zijlstrad0b27fa2008-04-19 19:44:57 +02007900
Dario Faggioli332ac172013-11-07 14:43:45 +01007901#ifdef CONFIG_RT_GROUP_SCHED
Peter Zijlstrad0b27fa2008-04-19 19:44:57 +02007902static int sched_rt_global_constraints(void)
7903{
7904 int ret = 0;
7905
7906 mutex_lock(&rt_constraints_mutex);
Peter Zijlstra9a7e0b12008-08-19 12:33:06 +02007907 read_lock(&tasklist_lock);
Peter Zijlstra4653f802008-09-23 15:33:44 +02007908 ret = __rt_schedulable(NULL, 0, 0);
Peter Zijlstra9a7e0b12008-08-19 12:33:06 +02007909 read_unlock(&tasklist_lock);
Peter Zijlstrad0b27fa2008-04-19 19:44:57 +02007910 mutex_unlock(&rt_constraints_mutex);
7911
7912 return ret;
7913}
Dhaval Giani54e99122009-02-27 15:13:54 +05307914
Li Zefan25cc7da2013-03-05 16:07:33 +08007915static int sched_rt_can_attach(struct task_group *tg, struct task_struct *tsk)
Dhaval Giani54e99122009-02-27 15:13:54 +05307916{
7917 /* Don't accept realtime tasks when there is no way for them to run */
7918 if (rt_task(tsk) && tg->rt_bandwidth.rt_runtime == 0)
7919 return 0;
7920
7921 return 1;
7922}
7923
Dhaval Giani6d6bc0a2008-05-30 14:23:45 +02007924#else /* !CONFIG_RT_GROUP_SCHED */
Peter Zijlstrad0b27fa2008-04-19 19:44:57 +02007925static int sched_rt_global_constraints(void)
7926{
Peter Zijlstraac086bc2008-04-19 19:44:58 +02007927 unsigned long flags;
Dario Faggioli332ac172013-11-07 14:43:45 +01007928 int i, ret = 0;
Hiroshi Shimamotoec5d4982008-09-10 17:00:19 -07007929
Thomas Gleixner0986b112009-11-17 15:32:06 +01007930 raw_spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags);
Peter Zijlstraac086bc2008-04-19 19:44:58 +02007931 for_each_possible_cpu(i) {
7932 struct rt_rq *rt_rq = &cpu_rq(i)->rt;
7933
Thomas Gleixner0986b112009-11-17 15:32:06 +01007934 raw_spin_lock(&rt_rq->rt_runtime_lock);
Peter Zijlstraac086bc2008-04-19 19:44:58 +02007935 rt_rq->rt_runtime = global_rt_runtime();
Thomas Gleixner0986b112009-11-17 15:32:06 +01007936 raw_spin_unlock(&rt_rq->rt_runtime_lock);
Peter Zijlstraac086bc2008-04-19 19:44:58 +02007937 }
Thomas Gleixner0986b112009-11-17 15:32:06 +01007938 raw_spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags);
Peter Zijlstraac086bc2008-04-19 19:44:58 +02007939
Dario Faggioli332ac172013-11-07 14:43:45 +01007940 return ret;
7941}
7942#endif /* CONFIG_RT_GROUP_SCHED */
7943
Wanpeng Lia1963b82015-03-17 19:15:31 +08007944static int sched_dl_global_validate(void)
Dario Faggioli332ac172013-11-07 14:43:45 +01007945{
Peter Zijlstra17248132013-12-17 12:44:49 +01007946 u64 runtime = global_rt_runtime();
7947 u64 period = global_rt_period();
Dario Faggioli332ac172013-11-07 14:43:45 +01007948 u64 new_bw = to_ratio(period, runtime);
Kirill Tkhaif10e00f2014-09-30 12:23:37 +04007949 struct dl_bw *dl_b;
Peter Zijlstra17248132013-12-17 12:44:49 +01007950 int cpu, ret = 0;
Juri Lelli49516342014-02-11 09:24:27 +01007951 unsigned long flags;
Dario Faggioli332ac172013-11-07 14:43:45 +01007952
7953 /*
7954 * Here we want to check the bandwidth not being set to some
7955 * value smaller than the currently allocated bandwidth in
7956 * any of the root_domains.
7957 *
7958 * FIXME: Cycling on all the CPUs is overdoing, but simpler than
7959 * cycling on root_domains... Discussion on different/better
7960 * solutions is welcome!
7961 */
Peter Zijlstra17248132013-12-17 12:44:49 +01007962 for_each_possible_cpu(cpu) {
Kirill Tkhaif10e00f2014-09-30 12:23:37 +04007963 rcu_read_lock_sched();
7964 dl_b = dl_bw_of(cpu);
Dario Faggioli332ac172013-11-07 14:43:45 +01007965
Juri Lelli49516342014-02-11 09:24:27 +01007966 raw_spin_lock_irqsave(&dl_b->lock, flags);
Peter Zijlstra17248132013-12-17 12:44:49 +01007967 if (new_bw < dl_b->total_bw)
7968 ret = -EBUSY;
Juri Lelli49516342014-02-11 09:24:27 +01007969 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
Peter Zijlstra17248132013-12-17 12:44:49 +01007970
Kirill Tkhaif10e00f2014-09-30 12:23:37 +04007971 rcu_read_unlock_sched();
7972
Peter Zijlstra17248132013-12-17 12:44:49 +01007973 if (ret)
7974 break;
Dario Faggioli332ac172013-11-07 14:43:45 +01007975 }
7976
Peter Zijlstra17248132013-12-17 12:44:49 +01007977 return ret;
7978}
7979
7980static void sched_dl_do_global(void)
7981{
7982 u64 new_bw = -1;
Kirill Tkhaif10e00f2014-09-30 12:23:37 +04007983 struct dl_bw *dl_b;
Peter Zijlstra17248132013-12-17 12:44:49 +01007984 int cpu;
Juri Lelli49516342014-02-11 09:24:27 +01007985 unsigned long flags;
Peter Zijlstra17248132013-12-17 12:44:49 +01007986
7987 def_dl_bandwidth.dl_period = global_rt_period();
7988 def_dl_bandwidth.dl_runtime = global_rt_runtime();
7989
7990 if (global_rt_runtime() != RUNTIME_INF)
7991 new_bw = to_ratio(global_rt_period(), global_rt_runtime());
7992
7993 /*
7994 * FIXME: As above...
7995 */
7996 for_each_possible_cpu(cpu) {
Kirill Tkhaif10e00f2014-09-30 12:23:37 +04007997 rcu_read_lock_sched();
7998 dl_b = dl_bw_of(cpu);
Peter Zijlstra17248132013-12-17 12:44:49 +01007999
Juri Lelli49516342014-02-11 09:24:27 +01008000 raw_spin_lock_irqsave(&dl_b->lock, flags);
Peter Zijlstra17248132013-12-17 12:44:49 +01008001 dl_b->bw = new_bw;
Juri Lelli49516342014-02-11 09:24:27 +01008002 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
Kirill Tkhaif10e00f2014-09-30 12:23:37 +04008003
8004 rcu_read_unlock_sched();
Peter Zijlstra17248132013-12-17 12:44:49 +01008005 }
8006}
8007
8008static int sched_rt_global_validate(void)
8009{
8010 if (sysctl_sched_rt_period <= 0)
8011 return -EINVAL;
8012
Juri Lellie9e7cb32014-02-11 09:24:26 +01008013 if ((sysctl_sched_rt_runtime != RUNTIME_INF) &&
8014 (sysctl_sched_rt_runtime > sysctl_sched_rt_period))
Peter Zijlstra17248132013-12-17 12:44:49 +01008015 return -EINVAL;
8016
Dario Faggioli332ac172013-11-07 14:43:45 +01008017 return 0;
8018}
Peter Zijlstrad0b27fa2008-04-19 19:44:57 +02008019
Peter Zijlstra17248132013-12-17 12:44:49 +01008020static void sched_rt_do_global(void)
8021{
8022 def_rt_bandwidth.rt_runtime = global_rt_runtime();
8023 def_rt_bandwidth.rt_period = ns_to_ktime(global_rt_period());
8024}
8025
8026int sched_rt_handler(struct ctl_table *table, int write,
8027 void __user *buffer, size_t *lenp,
8028 loff_t *ppos)
8029{
8030 int old_period, old_runtime;
8031 static DEFINE_MUTEX(mutex);
8032 int ret;
8033
8034 mutex_lock(&mutex);
8035 old_period = sysctl_sched_rt_period;
8036 old_runtime = sysctl_sched_rt_runtime;
8037
8038 ret = proc_dointvec(table, write, buffer, lenp, ppos);
8039
8040 if (!ret && write) {
8041 ret = sched_rt_global_validate();
8042 if (ret)
8043 goto undo;
8044
Wanpeng Lia1963b82015-03-17 19:15:31 +08008045 ret = sched_dl_global_validate();
Peter Zijlstra17248132013-12-17 12:44:49 +01008046 if (ret)
8047 goto undo;
8048
Wanpeng Lia1963b82015-03-17 19:15:31 +08008049 ret = sched_rt_global_constraints();
Peter Zijlstra17248132013-12-17 12:44:49 +01008050 if (ret)
8051 goto undo;
8052
8053 sched_rt_do_global();
8054 sched_dl_do_global();
8055 }
8056 if (0) {
8057undo:
8058 sysctl_sched_rt_period = old_period;
8059 sysctl_sched_rt_runtime = old_runtime;
8060 }
8061 mutex_unlock(&mutex);
8062
8063 return ret;
8064}
8065
Clark Williamsce0dbbb2013-02-07 09:47:04 -06008066int sched_rr_handler(struct ctl_table *table, int write,
8067 void __user *buffer, size_t *lenp,
8068 loff_t *ppos)
8069{
8070 int ret;
8071 static DEFINE_MUTEX(mutex);
8072
8073 mutex_lock(&mutex);
8074 ret = proc_dointvec(table, write, buffer, lenp, ppos);
8075 /* make sure that internally we keep jiffies */
8076 /* also, writing zero resets timeslice to default */
8077 if (!ret && write) {
8078 sched_rr_timeslice = sched_rr_timeslice <= 0 ?
8079 RR_TIMESLICE : msecs_to_jiffies(sched_rr_timeslice);
8080 }
8081 mutex_unlock(&mutex);
8082 return ret;
8083}
8084
Peter Zijlstra052f1dc2008-02-13 15:45:40 +01008085#ifdef CONFIG_CGROUP_SCHED
Srivatsa Vaddagiri68318b82007-10-18 23:41:03 -07008086
Tejun Heoa7c6d552013-08-08 20:11:23 -04008087static inline struct task_group *css_tg(struct cgroup_subsys_state *css)
Srivatsa Vaddagiri68318b82007-10-18 23:41:03 -07008088{
Tejun Heoa7c6d552013-08-08 20:11:23 -04008089 return css ? container_of(css, struct task_group, css) : NULL;
Srivatsa Vaddagiri68318b82007-10-18 23:41:03 -07008090}
8091
Tejun Heoeb954192013-08-08 20:11:23 -04008092static struct cgroup_subsys_state *
8093cpu_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
Srivatsa Vaddagiri68318b82007-10-18 23:41:03 -07008094{
Tejun Heoeb954192013-08-08 20:11:23 -04008095 struct task_group *parent = css_tg(parent_css);
8096 struct task_group *tg;
Srivatsa Vaddagiri68318b82007-10-18 23:41:03 -07008097
Tejun Heoeb954192013-08-08 20:11:23 -04008098 if (!parent) {
Srivatsa Vaddagiri68318b82007-10-18 23:41:03 -07008099 /* This is early initialization for the top cgroup */
Yong Zhang07e06b02011-01-07 15:17:36 +08008100 return &root_task_group.css;
Srivatsa Vaddagiri68318b82007-10-18 23:41:03 -07008101 }
8102
Dhaval Gianiec7dc8a2008-04-19 19:44:59 +02008103 tg = sched_create_group(parent);
Srivatsa Vaddagiri68318b82007-10-18 23:41:03 -07008104 if (IS_ERR(tg))
8105 return ERR_PTR(-ENOMEM);
8106
Srivatsa Vaddagiri68318b82007-10-18 23:41:03 -07008107 return &tg->css;
8108}
8109
Tejun Heoeb954192013-08-08 20:11:23 -04008110static int cpu_cgroup_css_online(struct cgroup_subsys_state *css)
Li Zefanace783b2013-01-24 14:30:48 +08008111{
Tejun Heoeb954192013-08-08 20:11:23 -04008112 struct task_group *tg = css_tg(css);
Tejun Heo5c9d5352014-05-16 13:22:48 -04008113 struct task_group *parent = css_tg(css->parent);
Li Zefanace783b2013-01-24 14:30:48 +08008114
Tejun Heo63876982013-08-08 20:11:23 -04008115 if (parent)
8116 sched_online_group(tg, parent);
Li Zefanace783b2013-01-24 14:30:48 +08008117 return 0;
8118}
8119
Tejun Heoeb954192013-08-08 20:11:23 -04008120static void cpu_cgroup_css_free(struct cgroup_subsys_state *css)
Srivatsa Vaddagiri68318b82007-10-18 23:41:03 -07008121{
Tejun Heoeb954192013-08-08 20:11:23 -04008122 struct task_group *tg = css_tg(css);
Srivatsa Vaddagiri68318b82007-10-18 23:41:03 -07008123
8124 sched_destroy_group(tg);
8125}
8126
Tejun Heoeb954192013-08-08 20:11:23 -04008127static void cpu_cgroup_css_offline(struct cgroup_subsys_state *css)
Li Zefanace783b2013-01-24 14:30:48 +08008128{
Tejun Heoeb954192013-08-08 20:11:23 -04008129 struct task_group *tg = css_tg(css);
Li Zefanace783b2013-01-24 14:30:48 +08008130
8131 sched_offline_group(tg);
8132}
8133
Aleksa Sarai7e476822015-06-09 21:32:09 +10008134static void cpu_cgroup_fork(struct task_struct *task, void *private)
Kirill Tkhaieeb61e52014-10-27 14:18:25 +04008135{
8136 sched_move_task(task);
8137}
8138
Tejun Heoeb954192013-08-08 20:11:23 -04008139static int cpu_cgroup_can_attach(struct cgroup_subsys_state *css,
Tejun Heobb9d97b2011-12-12 18:12:21 -08008140 struct cgroup_taskset *tset)
Srivatsa Vaddagiri68318b82007-10-18 23:41:03 -07008141{
Tejun Heobb9d97b2011-12-12 18:12:21 -08008142 struct task_struct *task;
8143
Tejun Heo924f0d92014-02-13 06:58:41 -05008144 cgroup_taskset_for_each(task, tset) {
Peter Zijlstrab68aa232008-02-13 15:45:40 +01008145#ifdef CONFIG_RT_GROUP_SCHED
Tejun Heoeb954192013-08-08 20:11:23 -04008146 if (!sched_rt_can_attach(css_tg(css), task))
Tejun Heobb9d97b2011-12-12 18:12:21 -08008147 return -EINVAL;
Peter Zijlstrab68aa232008-02-13 15:45:40 +01008148#else
Tejun Heobb9d97b2011-12-12 18:12:21 -08008149 /* We don't support RT-tasks being in separate groups */
8150 if (task->sched_class != &fair_sched_class)
8151 return -EINVAL;
Peter Zijlstrab68aa232008-02-13 15:45:40 +01008152#endif
Tejun Heobb9d97b2011-12-12 18:12:21 -08008153 }
Ben Blumbe367d02009-09-23 15:56:31 -07008154 return 0;
8155}
Srivatsa Vaddagiri68318b82007-10-18 23:41:03 -07008156
Tejun Heoeb954192013-08-08 20:11:23 -04008157static void cpu_cgroup_attach(struct cgroup_subsys_state *css,
Tejun Heobb9d97b2011-12-12 18:12:21 -08008158 struct cgroup_taskset *tset)
Srivatsa Vaddagiri68318b82007-10-18 23:41:03 -07008159{
Tejun Heobb9d97b2011-12-12 18:12:21 -08008160 struct task_struct *task;
8161
Tejun Heo924f0d92014-02-13 06:58:41 -05008162 cgroup_taskset_for_each(task, tset)
Tejun Heobb9d97b2011-12-12 18:12:21 -08008163 sched_move_task(task);
Srivatsa Vaddagiri68318b82007-10-18 23:41:03 -07008164}
8165
Tejun Heoeb954192013-08-08 20:11:23 -04008166static void cpu_cgroup_exit(struct cgroup_subsys_state *css,
8167 struct cgroup_subsys_state *old_css,
8168 struct task_struct *task)
Peter Zijlstra068c5cc2011-01-19 12:26:11 +01008169{
8170 /*
8171 * cgroup_exit() is called in the copy_process() failure path.
8172 * Ignore this case since the task hasn't ran yet, this avoids
8173 * trying to poke a half freed task state from generic code.
8174 */
8175 if (!(task->flags & PF_EXITING))
8176 return;
8177
8178 sched_move_task(task);
8179}
8180
Peter Zijlstra052f1dc2008-02-13 15:45:40 +01008181#ifdef CONFIG_FAIR_GROUP_SCHED
Tejun Heo182446d2013-08-08 20:11:24 -04008182static int cpu_shares_write_u64(struct cgroup_subsys_state *css,
8183 struct cftype *cftype, u64 shareval)
Srivatsa Vaddagiri68318b82007-10-18 23:41:03 -07008184{
Tejun Heo182446d2013-08-08 20:11:24 -04008185 return sched_group_set_shares(css_tg(css), scale_load(shareval));
Srivatsa Vaddagiri68318b82007-10-18 23:41:03 -07008186}
8187
Tejun Heo182446d2013-08-08 20:11:24 -04008188static u64 cpu_shares_read_u64(struct cgroup_subsys_state *css,
8189 struct cftype *cft)
Srivatsa Vaddagiri68318b82007-10-18 23:41:03 -07008190{
Tejun Heo182446d2013-08-08 20:11:24 -04008191 struct task_group *tg = css_tg(css);
Srivatsa Vaddagiri68318b82007-10-18 23:41:03 -07008192
Nikhil Raoc8b28112011-05-18 14:37:48 -07008193 return (u64) scale_load_down(tg->shares);
Srivatsa Vaddagiri68318b82007-10-18 23:41:03 -07008194}
Paul Turnerab84d312011-07-21 09:43:28 -07008195
8196#ifdef CONFIG_CFS_BANDWIDTH
Paul Turnera790de92011-07-21 09:43:29 -07008197static DEFINE_MUTEX(cfs_constraints_mutex);
8198
Paul Turnerab84d312011-07-21 09:43:28 -07008199const u64 max_cfs_quota_period = 1 * NSEC_PER_SEC; /* 1s */
8200const u64 min_cfs_quota_period = 1 * NSEC_PER_MSEC; /* 1ms */
8201
Paul Turnera790de92011-07-21 09:43:29 -07008202static int __cfs_schedulable(struct task_group *tg, u64 period, u64 runtime);
8203
Paul Turnerab84d312011-07-21 09:43:28 -07008204static int tg_set_cfs_bandwidth(struct task_group *tg, u64 period, u64 quota)
8205{
Paul Turner56f570e2011-11-07 20:26:33 -08008206 int i, ret = 0, runtime_enabled, runtime_was_enabled;
Peter Zijlstra029632f2011-10-25 10:00:11 +02008207 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
Paul Turnerab84d312011-07-21 09:43:28 -07008208
8209 if (tg == &root_task_group)
8210 return -EINVAL;
8211
8212 /*
8213 * Ensure we have at some amount of bandwidth every period. This is
8214 * to prevent reaching a state of large arrears when throttled via
8215 * entity_tick() resulting in prolonged exit starvation.
8216 */
8217 if (quota < min_cfs_quota_period || period < min_cfs_quota_period)
8218 return -EINVAL;
8219
8220 /*
8221 * Likewise, bound things on the otherside by preventing insane quota
8222 * periods. This also allows us to normalize in computing quota
8223 * feasibility.
8224 */
8225 if (period > max_cfs_quota_period)
8226 return -EINVAL;
8227
Kirill Tkhai0e59bda2014-06-25 12:19:42 +04008228 /*
8229 * Prevent race between setting of cfs_rq->runtime_enabled and
8230 * unthrottle_offline_cfs_rqs().
8231 */
8232 get_online_cpus();
Paul Turnera790de92011-07-21 09:43:29 -07008233 mutex_lock(&cfs_constraints_mutex);
8234 ret = __cfs_schedulable(tg, period, quota);
8235 if (ret)
8236 goto out_unlock;
8237
Paul Turner58088ad2011-07-21 09:43:31 -07008238 runtime_enabled = quota != RUNTIME_INF;
Paul Turner56f570e2011-11-07 20:26:33 -08008239 runtime_was_enabled = cfs_b->quota != RUNTIME_INF;
Ben Segall1ee14e62013-10-16 11:16:12 -07008240 /*
8241 * If we need to toggle cfs_bandwidth_used, off->on must occur
8242 * before making related changes, and on->off must occur afterwards
8243 */
8244 if (runtime_enabled && !runtime_was_enabled)
8245 cfs_bandwidth_usage_inc();
Paul Turnerab84d312011-07-21 09:43:28 -07008246 raw_spin_lock_irq(&cfs_b->lock);
8247 cfs_b->period = ns_to_ktime(period);
8248 cfs_b->quota = quota;
Paul Turner58088ad2011-07-21 09:43:31 -07008249
Paul Turnera9cf55b2011-07-21 09:43:32 -07008250 __refill_cfs_bandwidth_runtime(cfs_b);
Paul Turner58088ad2011-07-21 09:43:31 -07008251 /* restart the period timer (if active) to handle new period expiry */
Peter Zijlstra77a4d1a2015-04-15 11:41:57 +02008252 if (runtime_enabled)
8253 start_cfs_bandwidth(cfs_b);
Paul Turnerab84d312011-07-21 09:43:28 -07008254 raw_spin_unlock_irq(&cfs_b->lock);
8255
Kirill Tkhai0e59bda2014-06-25 12:19:42 +04008256 for_each_online_cpu(i) {
Paul Turnerab84d312011-07-21 09:43:28 -07008257 struct cfs_rq *cfs_rq = tg->cfs_rq[i];
Peter Zijlstra029632f2011-10-25 10:00:11 +02008258 struct rq *rq = cfs_rq->rq;
Paul Turnerab84d312011-07-21 09:43:28 -07008259
8260 raw_spin_lock_irq(&rq->lock);
Paul Turner58088ad2011-07-21 09:43:31 -07008261 cfs_rq->runtime_enabled = runtime_enabled;
Paul Turnerab84d312011-07-21 09:43:28 -07008262 cfs_rq->runtime_remaining = 0;
Paul Turner671fd9d2011-07-21 09:43:34 -07008263
Peter Zijlstra029632f2011-10-25 10:00:11 +02008264 if (cfs_rq->throttled)
Paul Turner671fd9d2011-07-21 09:43:34 -07008265 unthrottle_cfs_rq(cfs_rq);
Paul Turnerab84d312011-07-21 09:43:28 -07008266 raw_spin_unlock_irq(&rq->lock);
8267 }
Ben Segall1ee14e62013-10-16 11:16:12 -07008268 if (runtime_was_enabled && !runtime_enabled)
8269 cfs_bandwidth_usage_dec();
Paul Turnera790de92011-07-21 09:43:29 -07008270out_unlock:
8271 mutex_unlock(&cfs_constraints_mutex);
Kirill Tkhai0e59bda2014-06-25 12:19:42 +04008272 put_online_cpus();
Paul Turnerab84d312011-07-21 09:43:28 -07008273
Paul Turnera790de92011-07-21 09:43:29 -07008274 return ret;
Paul Turnerab84d312011-07-21 09:43:28 -07008275}
8276
8277int tg_set_cfs_quota(struct task_group *tg, long cfs_quota_us)
8278{
8279 u64 quota, period;
8280
Peter Zijlstra029632f2011-10-25 10:00:11 +02008281 period = ktime_to_ns(tg->cfs_bandwidth.period);
Paul Turnerab84d312011-07-21 09:43:28 -07008282 if (cfs_quota_us < 0)
8283 quota = RUNTIME_INF;
8284 else
8285 quota = (u64)cfs_quota_us * NSEC_PER_USEC;
8286
8287 return tg_set_cfs_bandwidth(tg, period, quota);
8288}
8289
8290long tg_get_cfs_quota(struct task_group *tg)
8291{
8292 u64 quota_us;
8293
Peter Zijlstra029632f2011-10-25 10:00:11 +02008294 if (tg->cfs_bandwidth.quota == RUNTIME_INF)
Paul Turnerab84d312011-07-21 09:43:28 -07008295 return -1;
8296
Peter Zijlstra029632f2011-10-25 10:00:11 +02008297 quota_us = tg->cfs_bandwidth.quota;
Paul Turnerab84d312011-07-21 09:43:28 -07008298 do_div(quota_us, NSEC_PER_USEC);
8299
8300 return quota_us;
8301}
8302
8303int tg_set_cfs_period(struct task_group *tg, long cfs_period_us)
8304{
8305 u64 quota, period;
8306
8307 period = (u64)cfs_period_us * NSEC_PER_USEC;
Peter Zijlstra029632f2011-10-25 10:00:11 +02008308 quota = tg->cfs_bandwidth.quota;
Paul Turnerab84d312011-07-21 09:43:28 -07008309
Paul Turnerab84d312011-07-21 09:43:28 -07008310 return tg_set_cfs_bandwidth(tg, period, quota);
8311}
8312
8313long tg_get_cfs_period(struct task_group *tg)
8314{
8315 u64 cfs_period_us;
8316
Peter Zijlstra029632f2011-10-25 10:00:11 +02008317 cfs_period_us = ktime_to_ns(tg->cfs_bandwidth.period);
Paul Turnerab84d312011-07-21 09:43:28 -07008318 do_div(cfs_period_us, NSEC_PER_USEC);
8319
8320 return cfs_period_us;
8321}
8322
Tejun Heo182446d2013-08-08 20:11:24 -04008323static s64 cpu_cfs_quota_read_s64(struct cgroup_subsys_state *css,
8324 struct cftype *cft)
Paul Turnerab84d312011-07-21 09:43:28 -07008325{
Tejun Heo182446d2013-08-08 20:11:24 -04008326 return tg_get_cfs_quota(css_tg(css));
Paul Turnerab84d312011-07-21 09:43:28 -07008327}
8328
Tejun Heo182446d2013-08-08 20:11:24 -04008329static int cpu_cfs_quota_write_s64(struct cgroup_subsys_state *css,
8330 struct cftype *cftype, s64 cfs_quota_us)
Paul Turnerab84d312011-07-21 09:43:28 -07008331{
Tejun Heo182446d2013-08-08 20:11:24 -04008332 return tg_set_cfs_quota(css_tg(css), cfs_quota_us);
Paul Turnerab84d312011-07-21 09:43:28 -07008333}
8334
Tejun Heo182446d2013-08-08 20:11:24 -04008335static u64 cpu_cfs_period_read_u64(struct cgroup_subsys_state *css,
8336 struct cftype *cft)
Paul Turnerab84d312011-07-21 09:43:28 -07008337{
Tejun Heo182446d2013-08-08 20:11:24 -04008338 return tg_get_cfs_period(css_tg(css));
Paul Turnerab84d312011-07-21 09:43:28 -07008339}
8340
Tejun Heo182446d2013-08-08 20:11:24 -04008341static int cpu_cfs_period_write_u64(struct cgroup_subsys_state *css,
8342 struct cftype *cftype, u64 cfs_period_us)
Paul Turnerab84d312011-07-21 09:43:28 -07008343{
Tejun Heo182446d2013-08-08 20:11:24 -04008344 return tg_set_cfs_period(css_tg(css), cfs_period_us);
Paul Turnerab84d312011-07-21 09:43:28 -07008345}
8346
Paul Turnera790de92011-07-21 09:43:29 -07008347struct cfs_schedulable_data {
8348 struct task_group *tg;
8349 u64 period, quota;
8350};
8351
8352/*
8353 * normalize group quota/period to be quota/max_period
8354 * note: units are usecs
8355 */
8356static u64 normalize_cfs_quota(struct task_group *tg,
8357 struct cfs_schedulable_data *d)
8358{
8359 u64 quota, period;
8360
8361 if (tg == d->tg) {
8362 period = d->period;
8363 quota = d->quota;
8364 } else {
8365 period = tg_get_cfs_period(tg);
8366 quota = tg_get_cfs_quota(tg);
8367 }
8368
8369 /* note: these should typically be equivalent */
8370 if (quota == RUNTIME_INF || quota == -1)
8371 return RUNTIME_INF;
8372
8373 return to_ratio(period, quota);
8374}
8375
8376static int tg_cfs_schedulable_down(struct task_group *tg, void *data)
8377{
8378 struct cfs_schedulable_data *d = data;
Peter Zijlstra029632f2011-10-25 10:00:11 +02008379 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
Paul Turnera790de92011-07-21 09:43:29 -07008380 s64 quota = 0, parent_quota = -1;
8381
8382 if (!tg->parent) {
8383 quota = RUNTIME_INF;
8384 } else {
Peter Zijlstra029632f2011-10-25 10:00:11 +02008385 struct cfs_bandwidth *parent_b = &tg->parent->cfs_bandwidth;
Paul Turnera790de92011-07-21 09:43:29 -07008386
8387 quota = normalize_cfs_quota(tg, d);
Zhihui Zhang9c58c792014-09-20 21:24:36 -04008388 parent_quota = parent_b->hierarchical_quota;
Paul Turnera790de92011-07-21 09:43:29 -07008389
8390 /*
8391 * ensure max(child_quota) <= parent_quota, inherit when no
8392 * limit is set
8393 */
8394 if (quota == RUNTIME_INF)
8395 quota = parent_quota;
8396 else if (parent_quota != RUNTIME_INF && quota > parent_quota)
8397 return -EINVAL;
8398 }
Zhihui Zhang9c58c792014-09-20 21:24:36 -04008399 cfs_b->hierarchical_quota = quota;
Paul Turnera790de92011-07-21 09:43:29 -07008400
8401 return 0;
8402}
8403
8404static int __cfs_schedulable(struct task_group *tg, u64 period, u64 quota)
8405{
Paul Turner82774342011-07-21 09:43:35 -07008406 int ret;
Paul Turnera790de92011-07-21 09:43:29 -07008407 struct cfs_schedulable_data data = {
8408 .tg = tg,
8409 .period = period,
8410 .quota = quota,
8411 };
8412
8413 if (quota != RUNTIME_INF) {
8414 do_div(data.period, NSEC_PER_USEC);
8415 do_div(data.quota, NSEC_PER_USEC);
8416 }
8417
Paul Turner82774342011-07-21 09:43:35 -07008418 rcu_read_lock();
8419 ret = walk_tg_tree(tg_cfs_schedulable_down, tg_nop, &data);
8420 rcu_read_unlock();
8421
8422 return ret;
Paul Turnera790de92011-07-21 09:43:29 -07008423}
Nikhil Raoe8da1b12011-07-21 09:43:40 -07008424
Tejun Heo2da8ca82013-12-05 12:28:04 -05008425static int cpu_stats_show(struct seq_file *sf, void *v)
Nikhil Raoe8da1b12011-07-21 09:43:40 -07008426{
Tejun Heo2da8ca82013-12-05 12:28:04 -05008427 struct task_group *tg = css_tg(seq_css(sf));
Peter Zijlstra029632f2011-10-25 10:00:11 +02008428 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
Nikhil Raoe8da1b12011-07-21 09:43:40 -07008429
Tejun Heo44ffc752013-12-05 12:28:01 -05008430 seq_printf(sf, "nr_periods %d\n", cfs_b->nr_periods);
8431 seq_printf(sf, "nr_throttled %d\n", cfs_b->nr_throttled);
8432 seq_printf(sf, "throttled_time %llu\n", cfs_b->throttled_time);
Nikhil Raoe8da1b12011-07-21 09:43:40 -07008433
8434 return 0;
8435}
Paul Turnerab84d312011-07-21 09:43:28 -07008436#endif /* CONFIG_CFS_BANDWIDTH */
Dhaval Giani6d6bc0a2008-05-30 14:23:45 +02008437#endif /* CONFIG_FAIR_GROUP_SCHED */
Srivatsa Vaddagiri68318b82007-10-18 23:41:03 -07008438
Peter Zijlstra052f1dc2008-02-13 15:45:40 +01008439#ifdef CONFIG_RT_GROUP_SCHED
Tejun Heo182446d2013-08-08 20:11:24 -04008440static int cpu_rt_runtime_write(struct cgroup_subsys_state *css,
8441 struct cftype *cft, s64 val)
Peter Zijlstra6f505b12008-01-25 21:08:30 +01008442{
Tejun Heo182446d2013-08-08 20:11:24 -04008443 return sched_group_set_rt_runtime(css_tg(css), val);
Peter Zijlstra6f505b12008-01-25 21:08:30 +01008444}
8445
Tejun Heo182446d2013-08-08 20:11:24 -04008446static s64 cpu_rt_runtime_read(struct cgroup_subsys_state *css,
8447 struct cftype *cft)
Peter Zijlstra6f505b12008-01-25 21:08:30 +01008448{
Tejun Heo182446d2013-08-08 20:11:24 -04008449 return sched_group_rt_runtime(css_tg(css));
Peter Zijlstra6f505b12008-01-25 21:08:30 +01008450}
Peter Zijlstrad0b27fa2008-04-19 19:44:57 +02008451
Tejun Heo182446d2013-08-08 20:11:24 -04008452static int cpu_rt_period_write_uint(struct cgroup_subsys_state *css,
8453 struct cftype *cftype, u64 rt_period_us)
Peter Zijlstrad0b27fa2008-04-19 19:44:57 +02008454{
Tejun Heo182446d2013-08-08 20:11:24 -04008455 return sched_group_set_rt_period(css_tg(css), rt_period_us);
Peter Zijlstrad0b27fa2008-04-19 19:44:57 +02008456}
8457
Tejun Heo182446d2013-08-08 20:11:24 -04008458static u64 cpu_rt_period_read_uint(struct cgroup_subsys_state *css,
8459 struct cftype *cft)
Peter Zijlstrad0b27fa2008-04-19 19:44:57 +02008460{
Tejun Heo182446d2013-08-08 20:11:24 -04008461 return sched_group_rt_period(css_tg(css));
Peter Zijlstrad0b27fa2008-04-19 19:44:57 +02008462}
Dhaval Giani6d6bc0a2008-05-30 14:23:45 +02008463#endif /* CONFIG_RT_GROUP_SCHED */
Peter Zijlstra6f505b12008-01-25 21:08:30 +01008464
Paul Menagefe5c7cc2007-10-29 21:18:11 +01008465static struct cftype cpu_files[] = {
Peter Zijlstra052f1dc2008-02-13 15:45:40 +01008466#ifdef CONFIG_FAIR_GROUP_SCHED
Paul Menagefe5c7cc2007-10-29 21:18:11 +01008467 {
8468 .name = "shares",
Paul Menagef4c753b2008-04-29 00:59:56 -07008469 .read_u64 = cpu_shares_read_u64,
8470 .write_u64 = cpu_shares_write_u64,
Paul Menagefe5c7cc2007-10-29 21:18:11 +01008471 },
Peter Zijlstra052f1dc2008-02-13 15:45:40 +01008472#endif
Paul Turnerab84d312011-07-21 09:43:28 -07008473#ifdef CONFIG_CFS_BANDWIDTH
8474 {
8475 .name = "cfs_quota_us",
8476 .read_s64 = cpu_cfs_quota_read_s64,
8477 .write_s64 = cpu_cfs_quota_write_s64,
8478 },
8479 {
8480 .name = "cfs_period_us",
8481 .read_u64 = cpu_cfs_period_read_u64,
8482 .write_u64 = cpu_cfs_period_write_u64,
8483 },
Nikhil Raoe8da1b12011-07-21 09:43:40 -07008484 {
8485 .name = "stat",
Tejun Heo2da8ca82013-12-05 12:28:04 -05008486 .seq_show = cpu_stats_show,
Nikhil Raoe8da1b12011-07-21 09:43:40 -07008487 },
Paul Turnerab84d312011-07-21 09:43:28 -07008488#endif
Peter Zijlstra052f1dc2008-02-13 15:45:40 +01008489#ifdef CONFIG_RT_GROUP_SCHED
Peter Zijlstra6f505b12008-01-25 21:08:30 +01008490 {
Peter Zijlstra9f0c1e52008-02-13 15:45:39 +01008491 .name = "rt_runtime_us",
Paul Menage06ecb272008-04-29 01:00:06 -07008492 .read_s64 = cpu_rt_runtime_read,
8493 .write_s64 = cpu_rt_runtime_write,
Peter Zijlstra6f505b12008-01-25 21:08:30 +01008494 },
Peter Zijlstrad0b27fa2008-04-19 19:44:57 +02008495 {
8496 .name = "rt_period_us",
Paul Menagef4c753b2008-04-29 00:59:56 -07008497 .read_u64 = cpu_rt_period_read_uint,
8498 .write_u64 = cpu_rt_period_write_uint,
Peter Zijlstrad0b27fa2008-04-19 19:44:57 +02008499 },
Peter Zijlstra052f1dc2008-02-13 15:45:40 +01008500#endif
Tejun Heo4baf6e32012-04-01 12:09:55 -07008501 { } /* terminate */
Srivatsa Vaddagiri68318b82007-10-18 23:41:03 -07008502};
8503
Tejun Heo073219e2014-02-08 10:36:58 -05008504struct cgroup_subsys cpu_cgrp_subsys = {
Tejun Heo92fb9742012-11-19 08:13:38 -08008505 .css_alloc = cpu_cgroup_css_alloc,
8506 .css_free = cpu_cgroup_css_free,
Li Zefanace783b2013-01-24 14:30:48 +08008507 .css_online = cpu_cgroup_css_online,
8508 .css_offline = cpu_cgroup_css_offline,
Kirill Tkhaieeb61e52014-10-27 14:18:25 +04008509 .fork = cpu_cgroup_fork,
Tejun Heobb9d97b2011-12-12 18:12:21 -08008510 .can_attach = cpu_cgroup_can_attach,
8511 .attach = cpu_cgroup_attach,
Peter Zijlstra068c5cc2011-01-19 12:26:11 +01008512 .exit = cpu_cgroup_exit,
Tejun Heo55779642014-07-15 11:05:09 -04008513 .legacy_cftypes = cpu_files,
Srivatsa Vaddagiri68318b82007-10-18 23:41:03 -07008514 .early_init = 1,
8515};
8516
Peter Zijlstra052f1dc2008-02-13 15:45:40 +01008517#endif /* CONFIG_CGROUP_SCHED */
Srivatsa Vaddagirid842de82007-12-02 20:04:49 +01008518
Paul E. McKenneyb637a322012-09-19 16:58:38 -07008519void dump_cpu_task(int cpu)
8520{
8521 pr_info("Task dump for CPU %d:\n", cpu);
8522 sched_show_task(cpu_curr(cpu));
8523}