blob: 459d601947a82a29693bbaa7444c74542311919e [file] [log] [blame]
Linus Torvalds1da177e2005-04-16 15:20:36 -07001/*
2 * kernel/cpuset.c
3 *
4 * Processor and Memory placement constraints for sets of tasks.
5 *
6 * Copyright (C) 2003 BULL SA.
Paul Jackson029190c2007-10-18 23:40:20 -07007 * Copyright (C) 2004-2007 Silicon Graphics, Inc.
Paul Menage8793d852007-10-18 23:39:39 -07008 * Copyright (C) 2006 Google, Inc
Linus Torvalds1da177e2005-04-16 15:20:36 -07009 *
10 * Portions derived from Patrick Mochel's sysfs code.
11 * sysfs is Copyright (c) 2001-3 Patrick Mochel
Linus Torvalds1da177e2005-04-16 15:20:36 -070012 *
Paul Jackson825a46a2006-03-24 03:16:03 -080013 * 2003-10-10 Written by Simon Derr.
Linus Torvalds1da177e2005-04-16 15:20:36 -070014 * 2003-10-22 Updates by Stephen Hemminger.
Paul Jackson825a46a2006-03-24 03:16:03 -080015 * 2004 May-July Rework by Paul Jackson.
Paul Menage8793d852007-10-18 23:39:39 -070016 * 2006 Rework by Paul Menage to use generic cgroups
Linus Torvalds1da177e2005-04-16 15:20:36 -070017 *
18 * This file is subject to the terms and conditions of the GNU General Public
19 * License. See the file COPYING in the main directory of the Linux
20 * distribution for more details.
21 */
22
Linus Torvalds1da177e2005-04-16 15:20:36 -070023#include <linux/cpu.h>
24#include <linux/cpumask.h>
25#include <linux/cpuset.h>
26#include <linux/err.h>
27#include <linux/errno.h>
28#include <linux/file.h>
29#include <linux/fs.h>
30#include <linux/init.h>
31#include <linux/interrupt.h>
32#include <linux/kernel.h>
33#include <linux/kmod.h>
34#include <linux/list.h>
Paul Jackson68860ec2005-10-30 15:02:36 -080035#include <linux/mempolicy.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070036#include <linux/mm.h>
37#include <linux/module.h>
38#include <linux/mount.h>
39#include <linux/namei.h>
40#include <linux/pagemap.h>
41#include <linux/proc_fs.h>
Paul Jackson6b9c2602006-01-08 01:02:02 -080042#include <linux/rcupdate.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070043#include <linux/sched.h>
44#include <linux/seq_file.h>
David Quigley22fb52d2006-06-23 02:04:00 -070045#include <linux/security.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070046#include <linux/slab.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070047#include <linux/spinlock.h>
48#include <linux/stat.h>
49#include <linux/string.h>
50#include <linux/time.h>
51#include <linux/backing-dev.h>
52#include <linux/sort.h>
53
54#include <asm/uaccess.h>
55#include <asm/atomic.h>
Ingo Molnar3d3f26a2006-03-23 03:00:18 -080056#include <linux/mutex.h>
Paul Jackson029190c2007-10-18 23:40:20 -070057#include <linux/kfifo.h>
Cliff Wickman956db3c2008-02-07 00:14:43 -080058#include <linux/workqueue.h>
59#include <linux/cgroup.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070060
Paul Jackson202f72d2006-01-08 01:01:57 -080061/*
62 * Tracks how many cpusets are currently defined in system.
63 * When there is only one cpuset (the root cpuset) we can
64 * short circuit some hooks.
65 */
Paul Jackson7edc5962006-01-08 01:02:03 -080066int number_of_cpusets __read_mostly;
Paul Jackson202f72d2006-01-08 01:01:57 -080067
Paul Menage2df167a2008-02-07 00:14:45 -080068/* Forward declare cgroup structures */
Paul Menage8793d852007-10-18 23:39:39 -070069struct cgroup_subsys cpuset_subsys;
70struct cpuset;
71
Paul Jackson3e0d98b2006-01-08 01:01:49 -080072/* See "Frequency meter" comments, below. */
73
74struct fmeter {
75 int cnt; /* unprocessed events count */
76 int val; /* most recent output value */
77 time_t time; /* clock (secs) when val computed */
78 spinlock_t lock; /* guards read or write of above */
79};
80
Linus Torvalds1da177e2005-04-16 15:20:36 -070081struct cpuset {
Paul Menage8793d852007-10-18 23:39:39 -070082 struct cgroup_subsys_state css;
83
Linus Torvalds1da177e2005-04-16 15:20:36 -070084 unsigned long flags; /* "unsigned long" so bitops work */
85 cpumask_t cpus_allowed; /* CPUs allowed to tasks in cpuset */
86 nodemask_t mems_allowed; /* Memory Nodes allowed to tasks */
87
Linus Torvalds1da177e2005-04-16 15:20:36 -070088 struct cpuset *parent; /* my parent */
Linus Torvalds1da177e2005-04-16 15:20:36 -070089
90 /*
91 * Copy of global cpuset_mems_generation as of the most
92 * recent time this cpuset changed its mems_allowed.
93 */
Paul Jackson3e0d98b2006-01-08 01:01:49 -080094 int mems_generation;
95
96 struct fmeter fmeter; /* memory_pressure filter */
Paul Jackson029190c2007-10-18 23:40:20 -070097
98 /* partition number for rebuild_sched_domains() */
99 int pn;
Cliff Wickman956db3c2008-02-07 00:14:43 -0800100
Hidetoshi Seto1d3504f2008-04-15 14:04:23 +0900101 /* for custom sched domain */
102 int relax_domain_level;
103
Cliff Wickman956db3c2008-02-07 00:14:43 -0800104 /* used for walking a cpuset heirarchy */
105 struct list_head stack_list;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700106};
107
Paul Menage8793d852007-10-18 23:39:39 -0700108/* Retrieve the cpuset for a cgroup */
109static inline struct cpuset *cgroup_cs(struct cgroup *cont)
110{
111 return container_of(cgroup_subsys_state(cont, cpuset_subsys_id),
112 struct cpuset, css);
113}
114
115/* Retrieve the cpuset for a task */
116static inline struct cpuset *task_cs(struct task_struct *task)
117{
118 return container_of(task_subsys_state(task, cpuset_subsys_id),
119 struct cpuset, css);
120}
Cliff Wickman956db3c2008-02-07 00:14:43 -0800121struct cpuset_hotplug_scanner {
122 struct cgroup_scanner scan;
123 struct cgroup *to;
124};
Paul Menage8793d852007-10-18 23:39:39 -0700125
Linus Torvalds1da177e2005-04-16 15:20:36 -0700126/* bits in struct cpuset flags field */
127typedef enum {
128 CS_CPU_EXCLUSIVE,
129 CS_MEM_EXCLUSIVE,
Paul Menage78608362008-04-29 01:00:26 -0700130 CS_MEM_HARDWALL,
Paul Jackson45b07ef2006-01-08 01:00:56 -0800131 CS_MEMORY_MIGRATE,
Paul Jackson029190c2007-10-18 23:40:20 -0700132 CS_SCHED_LOAD_BALANCE,
Paul Jackson825a46a2006-03-24 03:16:03 -0800133 CS_SPREAD_PAGE,
134 CS_SPREAD_SLAB,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700135} cpuset_flagbits_t;
136
137/* convenient tests for these bits */
138static inline int is_cpu_exclusive(const struct cpuset *cs)
139{
Paul Jackson7b5b9ef2006-03-24 03:16:00 -0800140 return test_bit(CS_CPU_EXCLUSIVE, &cs->flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700141}
142
143static inline int is_mem_exclusive(const struct cpuset *cs)
144{
Paul Jackson7b5b9ef2006-03-24 03:16:00 -0800145 return test_bit(CS_MEM_EXCLUSIVE, &cs->flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700146}
147
Paul Menage78608362008-04-29 01:00:26 -0700148static inline int is_mem_hardwall(const struct cpuset *cs)
149{
150 return test_bit(CS_MEM_HARDWALL, &cs->flags);
151}
152
Paul Jackson029190c2007-10-18 23:40:20 -0700153static inline int is_sched_load_balance(const struct cpuset *cs)
154{
155 return test_bit(CS_SCHED_LOAD_BALANCE, &cs->flags);
156}
157
Paul Jackson45b07ef2006-01-08 01:00:56 -0800158static inline int is_memory_migrate(const struct cpuset *cs)
159{
Paul Jackson7b5b9ef2006-03-24 03:16:00 -0800160 return test_bit(CS_MEMORY_MIGRATE, &cs->flags);
Paul Jackson45b07ef2006-01-08 01:00:56 -0800161}
162
Paul Jackson825a46a2006-03-24 03:16:03 -0800163static inline int is_spread_page(const struct cpuset *cs)
164{
165 return test_bit(CS_SPREAD_PAGE, &cs->flags);
166}
167
168static inline int is_spread_slab(const struct cpuset *cs)
169{
170 return test_bit(CS_SPREAD_SLAB, &cs->flags);
171}
172
Linus Torvalds1da177e2005-04-16 15:20:36 -0700173/*
Paul Jackson151a4422006-03-24 03:16:11 -0800174 * Increment this integer everytime any cpuset changes its
Linus Torvalds1da177e2005-04-16 15:20:36 -0700175 * mems_allowed value. Users of cpusets can track this generation
176 * number, and avoid having to lock and reload mems_allowed unless
177 * the cpuset they're using changes generation.
178 *
Paul Menage2df167a2008-02-07 00:14:45 -0800179 * A single, global generation is needed because cpuset_attach_task() could
Linus Torvalds1da177e2005-04-16 15:20:36 -0700180 * reattach a task to a different cpuset, which must not have its
181 * generation numbers aliased with those of that tasks previous cpuset.
182 *
183 * Generations are needed for mems_allowed because one task cannot
Paul Menage2df167a2008-02-07 00:14:45 -0800184 * modify another's memory placement. So we must enable every task,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700185 * on every visit to __alloc_pages(), to efficiently check whether
186 * its current->cpuset->mems_allowed has changed, requiring an update
187 * of its current->mems_allowed.
Paul Jackson151a4422006-03-24 03:16:11 -0800188 *
Paul Menage2df167a2008-02-07 00:14:45 -0800189 * Since writes to cpuset_mems_generation are guarded by the cgroup lock
Paul Jackson151a4422006-03-24 03:16:11 -0800190 * there is no need to mark it atomic.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700191 */
Paul Jackson151a4422006-03-24 03:16:11 -0800192static int cpuset_mems_generation;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700193
194static struct cpuset top_cpuset = {
195 .flags = ((1 << CS_CPU_EXCLUSIVE) | (1 << CS_MEM_EXCLUSIVE)),
196 .cpus_allowed = CPU_MASK_ALL,
197 .mems_allowed = NODE_MASK_ALL,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700198};
199
Linus Torvalds1da177e2005-04-16 15:20:36 -0700200/*
Paul Menage2df167a2008-02-07 00:14:45 -0800201 * There are two global mutexes guarding cpuset structures. The first
202 * is the main control groups cgroup_mutex, accessed via
203 * cgroup_lock()/cgroup_unlock(). The second is the cpuset-specific
204 * callback_mutex, below. They can nest. It is ok to first take
205 * cgroup_mutex, then nest callback_mutex. We also require taking
206 * task_lock() when dereferencing a task's cpuset pointer. See "The
207 * task_lock() exception", at the end of this comment.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700208 *
Ingo Molnar3d3f26a2006-03-23 03:00:18 -0800209 * A task must hold both mutexes to modify cpusets. If a task
Paul Menage2df167a2008-02-07 00:14:45 -0800210 * holds cgroup_mutex, then it blocks others wanting that mutex,
Ingo Molnar3d3f26a2006-03-23 03:00:18 -0800211 * ensuring that it is the only task able to also acquire callback_mutex
Paul Jackson053199e2005-10-30 15:02:30 -0800212 * and be able to modify cpusets. It can perform various checks on
213 * the cpuset structure first, knowing nothing will change. It can
Paul Menage2df167a2008-02-07 00:14:45 -0800214 * also allocate memory while just holding cgroup_mutex. While it is
Paul Jackson053199e2005-10-30 15:02:30 -0800215 * performing these checks, various callback routines can briefly
Ingo Molnar3d3f26a2006-03-23 03:00:18 -0800216 * acquire callback_mutex to query cpusets. Once it is ready to make
217 * the changes, it takes callback_mutex, blocking everyone else.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700218 *
Paul Jackson053199e2005-10-30 15:02:30 -0800219 * Calls to the kernel memory allocator can not be made while holding
Ingo Molnar3d3f26a2006-03-23 03:00:18 -0800220 * callback_mutex, as that would risk double tripping on callback_mutex
Paul Jackson053199e2005-10-30 15:02:30 -0800221 * from one of the callbacks into the cpuset code from within
222 * __alloc_pages().
Linus Torvalds1da177e2005-04-16 15:20:36 -0700223 *
Ingo Molnar3d3f26a2006-03-23 03:00:18 -0800224 * If a task is only holding callback_mutex, then it has read-only
Paul Jackson053199e2005-10-30 15:02:30 -0800225 * access to cpusets.
226 *
227 * The task_struct fields mems_allowed and mems_generation may only
228 * be accessed in the context of that task, so require no locks.
229 *
Paul Jackson053199e2005-10-30 15:02:30 -0800230 * The cpuset_common_file_write handler for operations that modify
Paul Menage2df167a2008-02-07 00:14:45 -0800231 * the cpuset hierarchy holds cgroup_mutex across the entire operation,
Paul Jackson053199e2005-10-30 15:02:30 -0800232 * single threading all such cpuset modifications across the system.
233 *
Ingo Molnar3d3f26a2006-03-23 03:00:18 -0800234 * The cpuset_common_file_read() handlers only hold callback_mutex across
Paul Jackson053199e2005-10-30 15:02:30 -0800235 * small pieces of code, such as when reading out possibly multi-word
236 * cpumasks and nodemasks.
237 *
Paul Menage2df167a2008-02-07 00:14:45 -0800238 * Accessing a task's cpuset should be done in accordance with the
239 * guidelines for accessing subsystem state in kernel/cgroup.c
Linus Torvalds1da177e2005-04-16 15:20:36 -0700240 */
241
Ingo Molnar3d3f26a2006-03-23 03:00:18 -0800242static DEFINE_MUTEX(callback_mutex);
Paul Jackson4247bdc2005-09-10 00:26:06 -0700243
Paul Menage8793d852007-10-18 23:39:39 -0700244/* This is ugly, but preserves the userspace API for existing cpuset
245 * users. If someone tries to mount the "cpuset" filesystem, we
246 * silently switch it to mount "cgroup" instead */
David Howells454e2392006-06-23 02:02:57 -0700247static int cpuset_get_sb(struct file_system_type *fs_type,
248 int flags, const char *unused_dev_name,
249 void *data, struct vfsmount *mnt)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700250{
Paul Menage8793d852007-10-18 23:39:39 -0700251 struct file_system_type *cgroup_fs = get_fs_type("cgroup");
252 int ret = -ENODEV;
253 if (cgroup_fs) {
254 char mountopts[] =
255 "cpuset,noprefix,"
256 "release_agent=/sbin/cpuset_release_agent";
257 ret = cgroup_fs->get_sb(cgroup_fs, flags,
258 unused_dev_name, mountopts, mnt);
259 put_filesystem(cgroup_fs);
260 }
261 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700262}
263
264static struct file_system_type cpuset_fs_type = {
265 .name = "cpuset",
266 .get_sb = cpuset_get_sb,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700267};
268
Linus Torvalds1da177e2005-04-16 15:20:36 -0700269/*
270 * Return in *pmask the portion of a cpusets's cpus_allowed that
271 * are online. If none are online, walk up the cpuset hierarchy
272 * until we find one that does have some online cpus. If we get
273 * all the way to the top and still haven't found any online cpus,
274 * return cpu_online_map. Or if passed a NULL cs from an exit'ing
275 * task, return cpu_online_map.
276 *
277 * One way or another, we guarantee to return some non-empty subset
278 * of cpu_online_map.
279 *
Ingo Molnar3d3f26a2006-03-23 03:00:18 -0800280 * Call with callback_mutex held.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700281 */
282
283static void guarantee_online_cpus(const struct cpuset *cs, cpumask_t *pmask)
284{
285 while (cs && !cpus_intersects(cs->cpus_allowed, cpu_online_map))
286 cs = cs->parent;
287 if (cs)
288 cpus_and(*pmask, cs->cpus_allowed, cpu_online_map);
289 else
290 *pmask = cpu_online_map;
291 BUG_ON(!cpus_intersects(*pmask, cpu_online_map));
292}
293
294/*
295 * Return in *pmask the portion of a cpusets's mems_allowed that
Christoph Lameter0e1e7c72007-10-16 01:25:38 -0700296 * are online, with memory. If none are online with memory, walk
297 * up the cpuset hierarchy until we find one that does have some
298 * online mems. If we get all the way to the top and still haven't
299 * found any online mems, return node_states[N_HIGH_MEMORY].
Linus Torvalds1da177e2005-04-16 15:20:36 -0700300 *
301 * One way or another, we guarantee to return some non-empty subset
Christoph Lameter0e1e7c72007-10-16 01:25:38 -0700302 * of node_states[N_HIGH_MEMORY].
Linus Torvalds1da177e2005-04-16 15:20:36 -0700303 *
Ingo Molnar3d3f26a2006-03-23 03:00:18 -0800304 * Call with callback_mutex held.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700305 */
306
307static void guarantee_online_mems(const struct cpuset *cs, nodemask_t *pmask)
308{
Christoph Lameter0e1e7c72007-10-16 01:25:38 -0700309 while (cs && !nodes_intersects(cs->mems_allowed,
310 node_states[N_HIGH_MEMORY]))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700311 cs = cs->parent;
312 if (cs)
Christoph Lameter0e1e7c72007-10-16 01:25:38 -0700313 nodes_and(*pmask, cs->mems_allowed,
314 node_states[N_HIGH_MEMORY]);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700315 else
Christoph Lameter0e1e7c72007-10-16 01:25:38 -0700316 *pmask = node_states[N_HIGH_MEMORY];
317 BUG_ON(!nodes_intersects(*pmask, node_states[N_HIGH_MEMORY]));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700318}
319
Paul Jacksoncf2a473c2006-01-08 01:01:54 -0800320/**
321 * cpuset_update_task_memory_state - update task memory placement
Linus Torvalds1da177e2005-04-16 15:20:36 -0700322 *
Paul Jacksoncf2a473c2006-01-08 01:01:54 -0800323 * If the current tasks cpusets mems_allowed changed behind our
324 * backs, update current->mems_allowed, mems_generation and task NUMA
325 * mempolicy to the new value.
326 *
327 * Task mempolicy is updated by rebinding it relative to the
328 * current->cpuset if a task has its memory placement changed.
329 * Do not call this routine if in_interrupt().
330 *
Paul Jackson4a01c8d2006-03-31 02:30:50 -0800331 * Call without callback_mutex or task_lock() held. May be
Paul Menage2df167a2008-02-07 00:14:45 -0800332 * called with or without cgroup_mutex held. Thanks in part to
333 * 'the_top_cpuset_hack', the task's cpuset pointer will never
David Rientjes41f7f602008-03-04 23:32:38 -0800334 * be NULL. This routine also might acquire callback_mutex during
335 * call.
Paul Jackson5aa15b52005-10-30 15:02:28 -0800336 *
Paul Jackson6b9c2602006-01-08 01:02:02 -0800337 * Reading current->cpuset->mems_generation doesn't need task_lock
338 * to guard the current->cpuset derefence, because it is guarded
Paul Menage2df167a2008-02-07 00:14:45 -0800339 * from concurrent freeing of current->cpuset using RCU.
Paul Jackson6b9c2602006-01-08 01:02:02 -0800340 *
341 * The rcu_dereference() is technically probably not needed,
342 * as I don't actually mind if I see a new cpuset pointer but
343 * an old value of mems_generation. However this really only
344 * matters on alpha systems using cpusets heavily. If I dropped
345 * that rcu_dereference(), it would save them a memory barrier.
346 * For all other arch's, rcu_dereference is a no-op anyway, and for
347 * alpha systems not using cpusets, another planned optimization,
348 * avoiding the rcu critical section for tasks in the root cpuset
349 * which is statically allocated, so can't vanish, will make this
350 * irrelevant. Better to use RCU as intended, than to engage in
351 * some cute trick to save a memory barrier that is impossible to
352 * test, for alpha systems using cpusets heavily, which might not
353 * even exist.
Paul Jackson053199e2005-10-30 15:02:30 -0800354 *
355 * This routine is needed to update the per-task mems_allowed data,
356 * within the tasks context, when it is trying to allocate memory
357 * (in various mm/mempolicy.c routines) and notices that some other
358 * task has been modifying its cpuset.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700359 */
360
Randy Dunlapfe85a992006-02-03 03:04:23 -0800361void cpuset_update_task_memory_state(void)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700362{
Paul Jackson053199e2005-10-30 15:02:30 -0800363 int my_cpusets_mem_gen;
Paul Jacksoncf2a473c2006-01-08 01:01:54 -0800364 struct task_struct *tsk = current;
Paul Jackson6b9c2602006-01-08 01:02:02 -0800365 struct cpuset *cs;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700366
Paul Menage8793d852007-10-18 23:39:39 -0700367 if (task_cs(tsk) == &top_cpuset) {
Paul Jackson03a285f2006-01-08 01:02:04 -0800368 /* Don't need rcu for top_cpuset. It's never freed. */
369 my_cpusets_mem_gen = top_cpuset.mems_generation;
370 } else {
371 rcu_read_lock();
Paul Menage8793d852007-10-18 23:39:39 -0700372 my_cpusets_mem_gen = task_cs(current)->mems_generation;
Paul Jackson03a285f2006-01-08 01:02:04 -0800373 rcu_read_unlock();
374 }
Paul Jacksoncf2a473c2006-01-08 01:01:54 -0800375
376 if (my_cpusets_mem_gen != tsk->cpuset_mems_generation) {
Ingo Molnar3d3f26a2006-03-23 03:00:18 -0800377 mutex_lock(&callback_mutex);
Paul Jacksoncf2a473c2006-01-08 01:01:54 -0800378 task_lock(tsk);
Paul Menage8793d852007-10-18 23:39:39 -0700379 cs = task_cs(tsk); /* Maybe changed when task not locked */
Paul Jacksoncf2a473c2006-01-08 01:01:54 -0800380 guarantee_online_mems(cs, &tsk->mems_allowed);
381 tsk->cpuset_mems_generation = cs->mems_generation;
Paul Jackson825a46a2006-03-24 03:16:03 -0800382 if (is_spread_page(cs))
383 tsk->flags |= PF_SPREAD_PAGE;
384 else
385 tsk->flags &= ~PF_SPREAD_PAGE;
386 if (is_spread_slab(cs))
387 tsk->flags |= PF_SPREAD_SLAB;
388 else
389 tsk->flags &= ~PF_SPREAD_SLAB;
Paul Jacksoncf2a473c2006-01-08 01:01:54 -0800390 task_unlock(tsk);
Ingo Molnar3d3f26a2006-03-23 03:00:18 -0800391 mutex_unlock(&callback_mutex);
Paul Jackson74cb2152006-01-08 01:01:56 -0800392 mpol_rebind_task(tsk, &tsk->mems_allowed);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700393 }
394}
395
396/*
397 * is_cpuset_subset(p, q) - Is cpuset p a subset of cpuset q?
398 *
399 * One cpuset is a subset of another if all its allowed CPUs and
400 * Memory Nodes are a subset of the other, and its exclusive flags
Paul Menage2df167a2008-02-07 00:14:45 -0800401 * are only set if the other's are set. Call holding cgroup_mutex.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700402 */
403
404static int is_cpuset_subset(const struct cpuset *p, const struct cpuset *q)
405{
406 return cpus_subset(p->cpus_allowed, q->cpus_allowed) &&
407 nodes_subset(p->mems_allowed, q->mems_allowed) &&
408 is_cpu_exclusive(p) <= is_cpu_exclusive(q) &&
409 is_mem_exclusive(p) <= is_mem_exclusive(q);
410}
411
412/*
413 * validate_change() - Used to validate that any proposed cpuset change
414 * follows the structural rules for cpusets.
415 *
416 * If we replaced the flag and mask values of the current cpuset
417 * (cur) with those values in the trial cpuset (trial), would
418 * our various subset and exclusive rules still be valid? Presumes
Paul Menage2df167a2008-02-07 00:14:45 -0800419 * cgroup_mutex held.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700420 *
421 * 'cur' is the address of an actual, in-use cpuset. Operations
422 * such as list traversal that depend on the actual address of the
423 * cpuset in the list must use cur below, not trial.
424 *
425 * 'trial' is the address of bulk structure copy of cur, with
426 * perhaps one or more of the fields cpus_allowed, mems_allowed,
427 * or flags changed to new, trial values.
428 *
429 * Return 0 if valid, -errno if not.
430 */
431
432static int validate_change(const struct cpuset *cur, const struct cpuset *trial)
433{
Paul Menage8793d852007-10-18 23:39:39 -0700434 struct cgroup *cont;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700435 struct cpuset *c, *par;
436
437 /* Each of our child cpusets must be a subset of us */
Paul Menage8793d852007-10-18 23:39:39 -0700438 list_for_each_entry(cont, &cur->css.cgroup->children, sibling) {
439 if (!is_cpuset_subset(cgroup_cs(cont), trial))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700440 return -EBUSY;
441 }
442
443 /* Remaining checks don't apply to root cpuset */
Paul Jackson69604062006-12-06 20:36:15 -0800444 if (cur == &top_cpuset)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700445 return 0;
446
Paul Jackson69604062006-12-06 20:36:15 -0800447 par = cur->parent;
448
Linus Torvalds1da177e2005-04-16 15:20:36 -0700449 /* We must be a subset of our parent cpuset */
450 if (!is_cpuset_subset(trial, par))
451 return -EACCES;
452
Paul Menage2df167a2008-02-07 00:14:45 -0800453 /*
454 * If either I or some sibling (!= me) is exclusive, we can't
455 * overlap
456 */
Paul Menage8793d852007-10-18 23:39:39 -0700457 list_for_each_entry(cont, &par->css.cgroup->children, sibling) {
458 c = cgroup_cs(cont);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700459 if ((is_cpu_exclusive(trial) || is_cpu_exclusive(c)) &&
460 c != cur &&
461 cpus_intersects(trial->cpus_allowed, c->cpus_allowed))
462 return -EINVAL;
463 if ((is_mem_exclusive(trial) || is_mem_exclusive(c)) &&
464 c != cur &&
465 nodes_intersects(trial->mems_allowed, c->mems_allowed))
466 return -EINVAL;
467 }
468
Paul Jackson020958b2007-10-18 23:40:21 -0700469 /* Cpusets with tasks can't have empty cpus_allowed or mems_allowed */
470 if (cgroup_task_count(cur->css.cgroup)) {
471 if (cpus_empty(trial->cpus_allowed) ||
472 nodes_empty(trial->mems_allowed)) {
473 return -ENOSPC;
474 }
475 }
476
Linus Torvalds1da177e2005-04-16 15:20:36 -0700477 return 0;
478}
479
Dinakar Guniguntala85d7b942005-06-25 14:57:34 -0700480/*
Paul Jackson029190c2007-10-18 23:40:20 -0700481 * Helper routine for rebuild_sched_domains().
482 * Do cpusets a, b have overlapping cpus_allowed masks?
483 */
484
485static int cpusets_overlap(struct cpuset *a, struct cpuset *b)
486{
487 return cpus_intersects(a->cpus_allowed, b->cpus_allowed);
488}
489
Hidetoshi Seto1d3504f2008-04-15 14:04:23 +0900490static void
491update_domain_attr(struct sched_domain_attr *dattr, struct cpuset *c)
492{
493 if (!dattr)
494 return;
495 if (dattr->relax_domain_level < c->relax_domain_level)
496 dattr->relax_domain_level = c->relax_domain_level;
497 return;
498}
499
Paul Jackson029190c2007-10-18 23:40:20 -0700500/*
501 * rebuild_sched_domains()
502 *
503 * If the flag 'sched_load_balance' of any cpuset with non-empty
504 * 'cpus' changes, or if the 'cpus' allowed changes in any cpuset
505 * which has that flag enabled, or if any cpuset with a non-empty
506 * 'cpus' is removed, then call this routine to rebuild the
507 * scheduler's dynamic sched domains.
508 *
509 * This routine builds a partial partition of the systems CPUs
510 * (the set of non-overlappping cpumask_t's in the array 'part'
511 * below), and passes that partial partition to the kernel/sched.c
512 * partition_sched_domains() routine, which will rebuild the
513 * schedulers load balancing domains (sched domains) as specified
514 * by that partial partition. A 'partial partition' is a set of
515 * non-overlapping subsets whose union is a subset of that set.
516 *
517 * See "What is sched_load_balance" in Documentation/cpusets.txt
518 * for a background explanation of this.
519 *
520 * Does not return errors, on the theory that the callers of this
521 * routine would rather not worry about failures to rebuild sched
522 * domains when operating in the severe memory shortage situations
523 * that could cause allocation failures below.
524 *
525 * Call with cgroup_mutex held. May take callback_mutex during
526 * call due to the kfifo_alloc() and kmalloc() calls. May nest
Gautham R Shenoy86ef5c92008-01-25 21:08:02 +0100527 * a call to the get_online_cpus()/put_online_cpus() pair.
Paul Jackson029190c2007-10-18 23:40:20 -0700528 * Must not be called holding callback_mutex, because we must not
Gautham R Shenoy86ef5c92008-01-25 21:08:02 +0100529 * call get_online_cpus() while holding callback_mutex. Elsewhere
530 * the kernel nests callback_mutex inside get_online_cpus() calls.
Paul Jackson029190c2007-10-18 23:40:20 -0700531 * So the reverse nesting would risk an ABBA deadlock.
532 *
533 * The three key local variables below are:
534 * q - a kfifo queue of cpuset pointers, used to implement a
535 * top-down scan of all cpusets. This scan loads a pointer
536 * to each cpuset marked is_sched_load_balance into the
537 * array 'csa'. For our purposes, rebuilding the schedulers
538 * sched domains, we can ignore !is_sched_load_balance cpusets.
539 * csa - (for CpuSet Array) Array of pointers to all the cpusets
540 * that need to be load balanced, for convenient iterative
541 * access by the subsequent code that finds the best partition,
542 * i.e the set of domains (subsets) of CPUs such that the
543 * cpus_allowed of every cpuset marked is_sched_load_balance
544 * is a subset of one of these domains, while there are as
545 * many such domains as possible, each as small as possible.
546 * doms - Conversion of 'csa' to an array of cpumasks, for passing to
547 * the kernel/sched.c routine partition_sched_domains() in a
548 * convenient format, that can be easily compared to the prior
549 * value to determine what partition elements (sched domains)
550 * were changed (added or removed.)
551 *
552 * Finding the best partition (set of domains):
553 * The triple nested loops below over i, j, k scan over the
554 * load balanced cpusets (using the array of cpuset pointers in
555 * csa[]) looking for pairs of cpusets that have overlapping
556 * cpus_allowed, but which don't have the same 'pn' partition
557 * number and gives them in the same partition number. It keeps
558 * looping on the 'restart' label until it can no longer find
559 * any such pairs.
560 *
561 * The union of the cpus_allowed masks from the set of
562 * all cpusets having the same 'pn' value then form the one
563 * element of the partition (one sched domain) to be passed to
564 * partition_sched_domains().
565 */
566
567static void rebuild_sched_domains(void)
568{
569 struct kfifo *q; /* queue of cpusets to be scanned */
570 struct cpuset *cp; /* scans q */
571 struct cpuset **csa; /* array of all cpuset ptrs */
572 int csn; /* how many cpuset ptrs in csa so far */
573 int i, j, k; /* indices for partition finding loops */
574 cpumask_t *doms; /* resulting partition; i.e. sched domains */
Hidetoshi Seto1d3504f2008-04-15 14:04:23 +0900575 struct sched_domain_attr *dattr; /* attributes for custom domains */
Paul Jackson029190c2007-10-18 23:40:20 -0700576 int ndoms; /* number of sched domains in result */
577 int nslot; /* next empty doms[] cpumask_t slot */
578
579 q = NULL;
580 csa = NULL;
581 doms = NULL;
Hidetoshi Seto1d3504f2008-04-15 14:04:23 +0900582 dattr = NULL;
Paul Jackson029190c2007-10-18 23:40:20 -0700583
584 /* Special case for the 99% of systems with one, full, sched domain */
585 if (is_sched_load_balance(&top_cpuset)) {
586 ndoms = 1;
587 doms = kmalloc(sizeof(cpumask_t), GFP_KERNEL);
588 if (!doms)
589 goto rebuild;
Hidetoshi Seto1d3504f2008-04-15 14:04:23 +0900590 dattr = kmalloc(sizeof(struct sched_domain_attr), GFP_KERNEL);
591 if (dattr) {
592 *dattr = SD_ATTR_INIT;
593 update_domain_attr(dattr, &top_cpuset);
594 }
Paul Jackson029190c2007-10-18 23:40:20 -0700595 *doms = top_cpuset.cpus_allowed;
596 goto rebuild;
597 }
598
599 q = kfifo_alloc(number_of_cpusets * sizeof(cp), GFP_KERNEL, NULL);
600 if (IS_ERR(q))
601 goto done;
602 csa = kmalloc(number_of_cpusets * sizeof(cp), GFP_KERNEL);
603 if (!csa)
604 goto done;
605 csn = 0;
606
607 cp = &top_cpuset;
608 __kfifo_put(q, (void *)&cp, sizeof(cp));
609 while (__kfifo_get(q, (void *)&cp, sizeof(cp))) {
610 struct cgroup *cont;
611 struct cpuset *child; /* scans child cpusets of cp */
612 if (is_sched_load_balance(cp))
613 csa[csn++] = cp;
614 list_for_each_entry(cont, &cp->css.cgroup->children, sibling) {
615 child = cgroup_cs(cont);
616 __kfifo_put(q, (void *)&child, sizeof(cp));
617 }
618 }
619
620 for (i = 0; i < csn; i++)
621 csa[i]->pn = i;
622 ndoms = csn;
623
624restart:
625 /* Find the best partition (set of sched domains) */
626 for (i = 0; i < csn; i++) {
627 struct cpuset *a = csa[i];
628 int apn = a->pn;
629
630 for (j = 0; j < csn; j++) {
631 struct cpuset *b = csa[j];
632 int bpn = b->pn;
633
634 if (apn != bpn && cpusets_overlap(a, b)) {
635 for (k = 0; k < csn; k++) {
636 struct cpuset *c = csa[k];
637
638 if (c->pn == bpn)
639 c->pn = apn;
640 }
641 ndoms--; /* one less element */
642 goto restart;
643 }
644 }
645 }
646
647 /* Convert <csn, csa> to <ndoms, doms> */
648 doms = kmalloc(ndoms * sizeof(cpumask_t), GFP_KERNEL);
649 if (!doms)
650 goto rebuild;
Hidetoshi Seto1d3504f2008-04-15 14:04:23 +0900651 dattr = kmalloc(ndoms * sizeof(struct sched_domain_attr), GFP_KERNEL);
Paul Jackson029190c2007-10-18 23:40:20 -0700652
653 for (nslot = 0, i = 0; i < csn; i++) {
654 struct cpuset *a = csa[i];
655 int apn = a->pn;
656
657 if (apn >= 0) {
658 cpumask_t *dp = doms + nslot;
659
660 if (nslot == ndoms) {
661 static int warnings = 10;
662 if (warnings) {
663 printk(KERN_WARNING
664 "rebuild_sched_domains confused:"
665 " nslot %d, ndoms %d, csn %d, i %d,"
666 " apn %d\n",
667 nslot, ndoms, csn, i, apn);
668 warnings--;
669 }
670 continue;
671 }
672
673 cpus_clear(*dp);
Hidetoshi Seto1d3504f2008-04-15 14:04:23 +0900674 if (dattr)
675 *(dattr + nslot) = SD_ATTR_INIT;
Paul Jackson029190c2007-10-18 23:40:20 -0700676 for (j = i; j < csn; j++) {
677 struct cpuset *b = csa[j];
678
679 if (apn == b->pn) {
680 cpus_or(*dp, *dp, b->cpus_allowed);
681 b->pn = -1;
Hidetoshi Seto1d3504f2008-04-15 14:04:23 +0900682 update_domain_attr(dattr, b);
Paul Jackson029190c2007-10-18 23:40:20 -0700683 }
684 }
685 nslot++;
686 }
687 }
688 BUG_ON(nslot != ndoms);
689
690rebuild:
691 /* Have scheduler rebuild sched domains */
Gautham R Shenoy86ef5c92008-01-25 21:08:02 +0100692 get_online_cpus();
Hidetoshi Seto1d3504f2008-04-15 14:04:23 +0900693 partition_sched_domains(ndoms, doms, dattr);
Gautham R Shenoy86ef5c92008-01-25 21:08:02 +0100694 put_online_cpus();
Paul Jackson029190c2007-10-18 23:40:20 -0700695
696done:
697 if (q && !IS_ERR(q))
698 kfifo_free(q);
699 kfree(csa);
700 /* Don't kfree(doms) -- partition_sched_domains() does that. */
Hidetoshi Seto1d3504f2008-04-15 14:04:23 +0900701 /* Don't kfree(dattr) -- partition_sched_domains() does that. */
Paul Jackson029190c2007-10-18 23:40:20 -0700702}
703
Paul Menage8707d8b2007-10-18 23:40:22 -0700704static inline int started_after_time(struct task_struct *t1,
705 struct timespec *time,
706 struct task_struct *t2)
707{
708 int start_diff = timespec_compare(&t1->start_time, time);
709 if (start_diff > 0) {
710 return 1;
711 } else if (start_diff < 0) {
712 return 0;
713 } else {
714 /*
715 * Arbitrarily, if two processes started at the same
716 * time, we'll say that the lower pointer value
717 * started first. Note that t2 may have exited by now
718 * so this may not be a valid pointer any longer, but
719 * that's fine - it still serves to distinguish
720 * between two tasks started (effectively)
721 * simultaneously.
722 */
723 return t1 > t2;
724 }
725}
726
727static inline int started_after(void *p1, void *p2)
728{
729 struct task_struct *t1 = p1;
730 struct task_struct *t2 = p2;
731 return started_after_time(t1, &t2->start_time, t2);
732}
733
Cliff Wickman58f47902008-02-07 00:14:44 -0800734/**
735 * cpuset_test_cpumask - test a task's cpus_allowed versus its cpuset's
736 * @tsk: task to test
737 * @scan: struct cgroup_scanner contained in its struct cpuset_hotplug_scanner
738 *
Paul Menage2df167a2008-02-07 00:14:45 -0800739 * Call with cgroup_mutex held. May take callback_mutex during call.
Cliff Wickman58f47902008-02-07 00:14:44 -0800740 * Called for each task in a cgroup by cgroup_scan_tasks().
741 * Return nonzero if this tasks's cpus_allowed mask should be changed (in other
742 * words, if its mask is not equal to its cpuset's mask).
Paul Jackson053199e2005-10-30 15:02:30 -0800743 */
Adrian Bunk9e0c9142008-04-29 01:00:25 -0700744static int cpuset_test_cpumask(struct task_struct *tsk,
745 struct cgroup_scanner *scan)
Cliff Wickman58f47902008-02-07 00:14:44 -0800746{
747 return !cpus_equal(tsk->cpus_allowed,
748 (cgroup_cs(scan->cg))->cpus_allowed);
749}
Paul Jackson053199e2005-10-30 15:02:30 -0800750
Cliff Wickman58f47902008-02-07 00:14:44 -0800751/**
752 * cpuset_change_cpumask - make a task's cpus_allowed the same as its cpuset's
753 * @tsk: task to test
754 * @scan: struct cgroup_scanner containing the cgroup of the task
755 *
756 * Called by cgroup_scan_tasks() for each task in a cgroup whose
757 * cpus_allowed mask needs to be changed.
758 *
759 * We don't need to re-check for the cgroup/cpuset membership, since we're
760 * holding cgroup_lock() at this point.
761 */
Adrian Bunk9e0c9142008-04-29 01:00:25 -0700762static void cpuset_change_cpumask(struct task_struct *tsk,
763 struct cgroup_scanner *scan)
Cliff Wickman58f47902008-02-07 00:14:44 -0800764{
Mike Travisf9a86fc2008-04-04 18:11:07 -0700765 set_cpus_allowed_ptr(tsk, &((cgroup_cs(scan->cg))->cpus_allowed));
Cliff Wickman58f47902008-02-07 00:14:44 -0800766}
767
768/**
769 * update_cpumask - update the cpus_allowed mask of a cpuset and all tasks in it
770 * @cs: the cpuset to consider
771 * @buf: buffer of cpu numbers written to this cpuset
772 */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700773static int update_cpumask(struct cpuset *cs, char *buf)
774{
775 struct cpuset trialcs;
Cliff Wickman58f47902008-02-07 00:14:44 -0800776 struct cgroup_scanner scan;
Paul Menage8707d8b2007-10-18 23:40:22 -0700777 struct ptr_heap heap;
Cliff Wickman58f47902008-02-07 00:14:44 -0800778 int retval;
779 int is_load_balanced;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700780
Paul Jackson4c4d50f2006-08-27 01:23:51 -0700781 /* top_cpuset.cpus_allowed tracks cpu_online_map; it's read-only */
782 if (cs == &top_cpuset)
783 return -EACCES;
784
Linus Torvalds1da177e2005-04-16 15:20:36 -0700785 trialcs = *cs;
David Rientjes6f7f02e2007-05-08 00:31:43 -0700786
787 /*
Paul Jacksonc8d9c902008-02-07 00:14:46 -0800788 * An empty cpus_allowed is ok only if the cpuset has no tasks.
Paul Jackson020958b2007-10-18 23:40:21 -0700789 * Since cpulist_parse() fails on an empty mask, we special case
790 * that parsing. The validate_change() call ensures that cpusets
791 * with tasks have cpus.
David Rientjes6f7f02e2007-05-08 00:31:43 -0700792 */
Paul Jackson020958b2007-10-18 23:40:21 -0700793 buf = strstrip(buf);
794 if (!*buf) {
David Rientjes6f7f02e2007-05-08 00:31:43 -0700795 cpus_clear(trialcs.cpus_allowed);
796 } else {
797 retval = cpulist_parse(buf, trialcs.cpus_allowed);
798 if (retval < 0)
799 return retval;
Lai Jiangshan37340742008-06-05 22:46:32 -0700800
801 if (!cpus_subset(trialcs.cpus_allowed, cpu_online_map))
802 return -EINVAL;
David Rientjes6f7f02e2007-05-08 00:31:43 -0700803 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700804 retval = validate_change(cs, &trialcs);
Dinakar Guniguntala85d7b942005-06-25 14:57:34 -0700805 if (retval < 0)
806 return retval;
Paul Jackson029190c2007-10-18 23:40:20 -0700807
Paul Menage8707d8b2007-10-18 23:40:22 -0700808 /* Nothing to do if the cpus didn't change */
809 if (cpus_equal(cs->cpus_allowed, trialcs.cpus_allowed))
810 return 0;
Cliff Wickman58f47902008-02-07 00:14:44 -0800811
Paul Menage8707d8b2007-10-18 23:40:22 -0700812 retval = heap_init(&heap, PAGE_SIZE, GFP_KERNEL, &started_after);
813 if (retval)
814 return retval;
815
Paul Jackson029190c2007-10-18 23:40:20 -0700816 is_load_balanced = is_sched_load_balance(&trialcs);
817
Ingo Molnar3d3f26a2006-03-23 03:00:18 -0800818 mutex_lock(&callback_mutex);
Dinakar Guniguntala85d7b942005-06-25 14:57:34 -0700819 cs->cpus_allowed = trialcs.cpus_allowed;
Ingo Molnar3d3f26a2006-03-23 03:00:18 -0800820 mutex_unlock(&callback_mutex);
Paul Jackson029190c2007-10-18 23:40:20 -0700821
Paul Menage8707d8b2007-10-18 23:40:22 -0700822 /*
823 * Scan tasks in the cpuset, and update the cpumasks of any
Cliff Wickman58f47902008-02-07 00:14:44 -0800824 * that need an update.
Paul Menage8707d8b2007-10-18 23:40:22 -0700825 */
Cliff Wickman58f47902008-02-07 00:14:44 -0800826 scan.cg = cs->css.cgroup;
827 scan.test_task = cpuset_test_cpumask;
828 scan.process_task = cpuset_change_cpumask;
829 scan.heap = &heap;
830 cgroup_scan_tasks(&scan);
Paul Menage8707d8b2007-10-18 23:40:22 -0700831 heap_free(&heap);
Cliff Wickman58f47902008-02-07 00:14:44 -0800832
Paul Menage8707d8b2007-10-18 23:40:22 -0700833 if (is_load_balanced)
Paul Jackson029190c2007-10-18 23:40:20 -0700834 rebuild_sched_domains();
Dinakar Guniguntala85d7b942005-06-25 14:57:34 -0700835 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700836}
837
Paul Jackson053199e2005-10-30 15:02:30 -0800838/*
Paul Jacksone4e364e2006-03-31 02:30:52 -0800839 * cpuset_migrate_mm
840 *
841 * Migrate memory region from one set of nodes to another.
842 *
843 * Temporarilly set tasks mems_allowed to target nodes of migration,
844 * so that the migration code can allocate pages on these nodes.
845 *
Paul Menage2df167a2008-02-07 00:14:45 -0800846 * Call holding cgroup_mutex, so current's cpuset won't change
Paul Jacksonc8d9c902008-02-07 00:14:46 -0800847 * during this call, as manage_mutex holds off any cpuset_attach()
Paul Jacksone4e364e2006-03-31 02:30:52 -0800848 * calls. Therefore we don't need to take task_lock around the
849 * call to guarantee_online_mems(), as we know no one is changing
Paul Menage2df167a2008-02-07 00:14:45 -0800850 * our task's cpuset.
Paul Jacksone4e364e2006-03-31 02:30:52 -0800851 *
852 * Hold callback_mutex around the two modifications of our tasks
853 * mems_allowed to synchronize with cpuset_mems_allowed().
854 *
855 * While the mm_struct we are migrating is typically from some
856 * other task, the task_struct mems_allowed that we are hacking
857 * is for our current task, which must allocate new pages for that
858 * migrating memory region.
859 *
860 * We call cpuset_update_task_memory_state() before hacking
861 * our tasks mems_allowed, so that we are assured of being in
862 * sync with our tasks cpuset, and in particular, callbacks to
863 * cpuset_update_task_memory_state() from nested page allocations
864 * won't see any mismatch of our cpuset and task mems_generation
865 * values, so won't overwrite our hacked tasks mems_allowed
866 * nodemask.
867 */
868
869static void cpuset_migrate_mm(struct mm_struct *mm, const nodemask_t *from,
870 const nodemask_t *to)
871{
872 struct task_struct *tsk = current;
873
874 cpuset_update_task_memory_state();
875
876 mutex_lock(&callback_mutex);
877 tsk->mems_allowed = *to;
878 mutex_unlock(&callback_mutex);
879
880 do_migrate_pages(mm, from, to, MPOL_MF_MOVE_ALL);
881
882 mutex_lock(&callback_mutex);
Paul Menage8793d852007-10-18 23:39:39 -0700883 guarantee_online_mems(task_cs(tsk),&tsk->mems_allowed);
Paul Jacksone4e364e2006-03-31 02:30:52 -0800884 mutex_unlock(&callback_mutex);
885}
886
887/*
Paul Jackson42253992006-01-08 01:01:59 -0800888 * Handle user request to change the 'mems' memory placement
889 * of a cpuset. Needs to validate the request, update the
890 * cpusets mems_allowed and mems_generation, and for each
Paul Jackson04c19fa2006-01-08 01:02:00 -0800891 * task in the cpuset, rebind any vma mempolicies and if
892 * the cpuset is marked 'memory_migrate', migrate the tasks
893 * pages to the new memory.
Paul Jackson42253992006-01-08 01:01:59 -0800894 *
Paul Menage2df167a2008-02-07 00:14:45 -0800895 * Call with cgroup_mutex held. May take callback_mutex during call.
Paul Jackson42253992006-01-08 01:01:59 -0800896 * Will take tasklist_lock, scan tasklist for tasks in cpuset cs,
897 * lock each such tasks mm->mmap_sem, scan its vma's and rebind
898 * their mempolicies to the cpusets new mems_allowed.
Paul Jackson053199e2005-10-30 15:02:30 -0800899 */
900
Paul Menage8793d852007-10-18 23:39:39 -0700901static void *cpuset_being_rebound;
902
Linus Torvalds1da177e2005-04-16 15:20:36 -0700903static int update_nodemask(struct cpuset *cs, char *buf)
904{
905 struct cpuset trialcs;
Paul Jackson04c19fa2006-01-08 01:02:00 -0800906 nodemask_t oldmem;
Paul Menage8793d852007-10-18 23:39:39 -0700907 struct task_struct *p;
Paul Jackson42253992006-01-08 01:01:59 -0800908 struct mm_struct **mmarray;
909 int i, n, ntasks;
Paul Jackson04c19fa2006-01-08 01:02:00 -0800910 int migrate;
Paul Jackson42253992006-01-08 01:01:59 -0800911 int fudge;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700912 int retval;
Paul Menage8793d852007-10-18 23:39:39 -0700913 struct cgroup_iter it;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700914
Christoph Lameter0e1e7c72007-10-16 01:25:38 -0700915 /*
916 * top_cpuset.mems_allowed tracks node_stats[N_HIGH_MEMORY];
917 * it's read-only
918 */
Paul Jackson38837fc2006-09-29 02:01:16 -0700919 if (cs == &top_cpuset)
920 return -EACCES;
921
Linus Torvalds1da177e2005-04-16 15:20:36 -0700922 trialcs = *cs;
David Rientjes6f7f02e2007-05-08 00:31:43 -0700923
924 /*
Paul Jackson020958b2007-10-18 23:40:21 -0700925 * An empty mems_allowed is ok iff there are no tasks in the cpuset.
926 * Since nodelist_parse() fails on an empty mask, we special case
927 * that parsing. The validate_change() call ensures that cpusets
928 * with tasks have memory.
David Rientjes6f7f02e2007-05-08 00:31:43 -0700929 */
Paul Jackson020958b2007-10-18 23:40:21 -0700930 buf = strstrip(buf);
931 if (!*buf) {
David Rientjes6f7f02e2007-05-08 00:31:43 -0700932 nodes_clear(trialcs.mems_allowed);
933 } else {
934 retval = nodelist_parse(buf, trialcs.mems_allowed);
935 if (retval < 0)
936 goto done;
Lai Jiangshan37340742008-06-05 22:46:32 -0700937
938 if (!nodes_subset(trialcs.mems_allowed,
939 node_states[N_HIGH_MEMORY]))
940 return -EINVAL;
David Rientjes6f7f02e2007-05-08 00:31:43 -0700941 }
Paul Jackson04c19fa2006-01-08 01:02:00 -0800942 oldmem = cs->mems_allowed;
943 if (nodes_equal(oldmem, trialcs.mems_allowed)) {
944 retval = 0; /* Too easy - nothing to do */
945 goto done;
946 }
Paul Jackson59dac162006-01-08 01:01:52 -0800947 retval = validate_change(cs, &trialcs);
948 if (retval < 0)
949 goto done;
950
Ingo Molnar3d3f26a2006-03-23 03:00:18 -0800951 mutex_lock(&callback_mutex);
Paul Jackson59dac162006-01-08 01:01:52 -0800952 cs->mems_allowed = trialcs.mems_allowed;
Paul Jackson151a4422006-03-24 03:16:11 -0800953 cs->mems_generation = cpuset_mems_generation++;
Ingo Molnar3d3f26a2006-03-23 03:00:18 -0800954 mutex_unlock(&callback_mutex);
Paul Jackson59dac162006-01-08 01:01:52 -0800955
Lee Schermerhorn846a16b2008-04-28 02:13:09 -0700956 cpuset_being_rebound = cs; /* causes mpol_dup() rebind */
Paul Jackson42253992006-01-08 01:01:59 -0800957
958 fudge = 10; /* spare mmarray[] slots */
959 fudge += cpus_weight(cs->cpus_allowed); /* imagine one fork-bomb/cpu */
960 retval = -ENOMEM;
961
962 /*
963 * Allocate mmarray[] to hold mm reference for each task
964 * in cpuset cs. Can't kmalloc GFP_KERNEL while holding
965 * tasklist_lock. We could use GFP_ATOMIC, but with a
966 * few more lines of code, we can retry until we get a big
967 * enough mmarray[] w/o using GFP_ATOMIC.
968 */
969 while (1) {
Paul Menage8793d852007-10-18 23:39:39 -0700970 ntasks = cgroup_task_count(cs->css.cgroup); /* guess */
Paul Jackson42253992006-01-08 01:01:59 -0800971 ntasks += fudge;
972 mmarray = kmalloc(ntasks * sizeof(*mmarray), GFP_KERNEL);
973 if (!mmarray)
974 goto done;
Paul Menagec2aef332007-07-15 23:40:11 -0700975 read_lock(&tasklist_lock); /* block fork */
Paul Menage8793d852007-10-18 23:39:39 -0700976 if (cgroup_task_count(cs->css.cgroup) <= ntasks)
Paul Jackson42253992006-01-08 01:01:59 -0800977 break; /* got enough */
Paul Menagec2aef332007-07-15 23:40:11 -0700978 read_unlock(&tasklist_lock); /* try again */
Paul Jackson42253992006-01-08 01:01:59 -0800979 kfree(mmarray);
980 }
981
982 n = 0;
983
984 /* Load up mmarray[] with mm reference for each task in cpuset. */
Paul Menage8793d852007-10-18 23:39:39 -0700985 cgroup_iter_start(cs->css.cgroup, &it);
986 while ((p = cgroup_iter_next(cs->css.cgroup, &it))) {
Paul Jackson42253992006-01-08 01:01:59 -0800987 struct mm_struct *mm;
988
989 if (n >= ntasks) {
990 printk(KERN_WARNING
991 "Cpuset mempolicy rebind incomplete.\n");
Paul Menage8793d852007-10-18 23:39:39 -0700992 break;
Paul Jackson42253992006-01-08 01:01:59 -0800993 }
Paul Jackson42253992006-01-08 01:01:59 -0800994 mm = get_task_mm(p);
995 if (!mm)
996 continue;
997 mmarray[n++] = mm;
Paul Menage8793d852007-10-18 23:39:39 -0700998 }
999 cgroup_iter_end(cs->css.cgroup, &it);
Paul Menagec2aef332007-07-15 23:40:11 -07001000 read_unlock(&tasklist_lock);
Paul Jackson42253992006-01-08 01:01:59 -08001001
1002 /*
1003 * Now that we've dropped the tasklist spinlock, we can
1004 * rebind the vma mempolicies of each mm in mmarray[] to their
1005 * new cpuset, and release that mm. The mpol_rebind_mm()
1006 * call takes mmap_sem, which we couldn't take while holding
Lee Schermerhorn846a16b2008-04-28 02:13:09 -07001007 * tasklist_lock. Forks can happen again now - the mpol_dup()
Paul Jackson42253992006-01-08 01:01:59 -08001008 * cpuset_being_rebound check will catch such forks, and rebind
1009 * their vma mempolicies too. Because we still hold the global
Paul Menage2df167a2008-02-07 00:14:45 -08001010 * cgroup_mutex, we know that no other rebind effort will
Paul Jackson42253992006-01-08 01:01:59 -08001011 * be contending for the global variable cpuset_being_rebound.
1012 * It's ok if we rebind the same mm twice; mpol_rebind_mm()
Paul Jackson04c19fa2006-01-08 01:02:00 -08001013 * is idempotent. Also migrate pages in each mm to new nodes.
Paul Jackson42253992006-01-08 01:01:59 -08001014 */
Paul Jackson04c19fa2006-01-08 01:02:00 -08001015 migrate = is_memory_migrate(cs);
Paul Jackson42253992006-01-08 01:01:59 -08001016 for (i = 0; i < n; i++) {
1017 struct mm_struct *mm = mmarray[i];
1018
1019 mpol_rebind_mm(mm, &cs->mems_allowed);
Paul Jacksone4e364e2006-03-31 02:30:52 -08001020 if (migrate)
1021 cpuset_migrate_mm(mm, &oldmem, &cs->mems_allowed);
Paul Jackson42253992006-01-08 01:01:59 -08001022 mmput(mm);
1023 }
1024
Paul Menage2df167a2008-02-07 00:14:45 -08001025 /* We're done rebinding vmas to this cpuset's new mems_allowed. */
Paul Jackson42253992006-01-08 01:01:59 -08001026 kfree(mmarray);
Paul Menage8793d852007-10-18 23:39:39 -07001027 cpuset_being_rebound = NULL;
Paul Jackson42253992006-01-08 01:01:59 -08001028 retval = 0;
Paul Jackson59dac162006-01-08 01:01:52 -08001029done:
Linus Torvalds1da177e2005-04-16 15:20:36 -07001030 return retval;
1031}
1032
Paul Menage8793d852007-10-18 23:39:39 -07001033int current_cpuset_is_being_rebound(void)
1034{
1035 return task_cs(current) == cpuset_being_rebound;
1036}
1037
Paul Menage5be7a472008-05-06 20:42:41 -07001038static int update_relax_domain_level(struct cpuset *cs, s64 val)
Hidetoshi Seto1d3504f2008-04-15 14:04:23 +09001039{
Li Zefan30e0e172008-05-13 10:27:17 +08001040 if (val < -1 || val >= SD_LV_MAX)
1041 return -EINVAL;
Hidetoshi Seto1d3504f2008-04-15 14:04:23 +09001042
1043 if (val != cs->relax_domain_level) {
1044 cs->relax_domain_level = val;
1045 rebuild_sched_domains();
1046 }
1047
1048 return 0;
1049}
1050
Paul Jackson3e0d98b2006-01-08 01:01:49 -08001051/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07001052 * update_flag - read a 0 or a 1 in a file and update associated flag
Paul Menage78608362008-04-29 01:00:26 -07001053 * bit: the bit to update (see cpuset_flagbits_t)
1054 * cs: the cpuset to update
1055 * turning_on: whether the flag is being set or cleared
Paul Jackson053199e2005-10-30 15:02:30 -08001056 *
Paul Menage2df167a2008-02-07 00:14:45 -08001057 * Call with cgroup_mutex held.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001058 */
1059
Paul Menage700fe1a2008-04-29 01:00:00 -07001060static int update_flag(cpuset_flagbits_t bit, struct cpuset *cs,
1061 int turning_on)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001062{
Linus Torvalds1da177e2005-04-16 15:20:36 -07001063 struct cpuset trialcs;
Paul Jackson607717a2007-10-16 01:27:43 -07001064 int err;
Paul Jackson029190c2007-10-18 23:40:20 -07001065 int cpus_nonempty, balance_flag_changed;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001066
Linus Torvalds1da177e2005-04-16 15:20:36 -07001067 trialcs = *cs;
1068 if (turning_on)
1069 set_bit(bit, &trialcs.flags);
1070 else
1071 clear_bit(bit, &trialcs.flags);
1072
1073 err = validate_change(cs, &trialcs);
Dinakar Guniguntala85d7b942005-06-25 14:57:34 -07001074 if (err < 0)
1075 return err;
Paul Jackson029190c2007-10-18 23:40:20 -07001076
1077 cpus_nonempty = !cpus_empty(trialcs.cpus_allowed);
1078 balance_flag_changed = (is_sched_load_balance(cs) !=
1079 is_sched_load_balance(&trialcs));
1080
Ingo Molnar3d3f26a2006-03-23 03:00:18 -08001081 mutex_lock(&callback_mutex);
Paul Jackson69604062006-12-06 20:36:15 -08001082 cs->flags = trialcs.flags;
Ingo Molnar3d3f26a2006-03-23 03:00:18 -08001083 mutex_unlock(&callback_mutex);
Dinakar Guniguntala85d7b942005-06-25 14:57:34 -07001084
Paul Jackson029190c2007-10-18 23:40:20 -07001085 if (cpus_nonempty && balance_flag_changed)
1086 rebuild_sched_domains();
1087
Dinakar Guniguntala85d7b942005-06-25 14:57:34 -07001088 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001089}
1090
Paul Jackson053199e2005-10-30 15:02:30 -08001091/*
Adrian Bunk80f72282006-06-30 18:27:16 +02001092 * Frequency meter - How fast is some event occurring?
Paul Jackson3e0d98b2006-01-08 01:01:49 -08001093 *
1094 * These routines manage a digitally filtered, constant time based,
1095 * event frequency meter. There are four routines:
1096 * fmeter_init() - initialize a frequency meter.
1097 * fmeter_markevent() - called each time the event happens.
1098 * fmeter_getrate() - returns the recent rate of such events.
1099 * fmeter_update() - internal routine used to update fmeter.
1100 *
1101 * A common data structure is passed to each of these routines,
1102 * which is used to keep track of the state required to manage the
1103 * frequency meter and its digital filter.
1104 *
1105 * The filter works on the number of events marked per unit time.
1106 * The filter is single-pole low-pass recursive (IIR). The time unit
1107 * is 1 second. Arithmetic is done using 32-bit integers scaled to
1108 * simulate 3 decimal digits of precision (multiplied by 1000).
1109 *
1110 * With an FM_COEF of 933, and a time base of 1 second, the filter
1111 * has a half-life of 10 seconds, meaning that if the events quit
1112 * happening, then the rate returned from the fmeter_getrate()
1113 * will be cut in half each 10 seconds, until it converges to zero.
1114 *
1115 * It is not worth doing a real infinitely recursive filter. If more
1116 * than FM_MAXTICKS ticks have elapsed since the last filter event,
1117 * just compute FM_MAXTICKS ticks worth, by which point the level
1118 * will be stable.
1119 *
1120 * Limit the count of unprocessed events to FM_MAXCNT, so as to avoid
1121 * arithmetic overflow in the fmeter_update() routine.
1122 *
1123 * Given the simple 32 bit integer arithmetic used, this meter works
1124 * best for reporting rates between one per millisecond (msec) and
1125 * one per 32 (approx) seconds. At constant rates faster than one
1126 * per msec it maxes out at values just under 1,000,000. At constant
1127 * rates between one per msec, and one per second it will stabilize
1128 * to a value N*1000, where N is the rate of events per second.
1129 * At constant rates between one per second and one per 32 seconds,
1130 * it will be choppy, moving up on the seconds that have an event,
1131 * and then decaying until the next event. At rates slower than
1132 * about one in 32 seconds, it decays all the way back to zero between
1133 * each event.
1134 */
1135
1136#define FM_COEF 933 /* coefficient for half-life of 10 secs */
1137#define FM_MAXTICKS ((time_t)99) /* useless computing more ticks than this */
1138#define FM_MAXCNT 1000000 /* limit cnt to avoid overflow */
1139#define FM_SCALE 1000 /* faux fixed point scale */
1140
1141/* Initialize a frequency meter */
1142static void fmeter_init(struct fmeter *fmp)
1143{
1144 fmp->cnt = 0;
1145 fmp->val = 0;
1146 fmp->time = 0;
1147 spin_lock_init(&fmp->lock);
1148}
1149
1150/* Internal meter update - process cnt events and update value */
1151static void fmeter_update(struct fmeter *fmp)
1152{
1153 time_t now = get_seconds();
1154 time_t ticks = now - fmp->time;
1155
1156 if (ticks == 0)
1157 return;
1158
1159 ticks = min(FM_MAXTICKS, ticks);
1160 while (ticks-- > 0)
1161 fmp->val = (FM_COEF * fmp->val) / FM_SCALE;
1162 fmp->time = now;
1163
1164 fmp->val += ((FM_SCALE - FM_COEF) * fmp->cnt) / FM_SCALE;
1165 fmp->cnt = 0;
1166}
1167
1168/* Process any previous ticks, then bump cnt by one (times scale). */
1169static void fmeter_markevent(struct fmeter *fmp)
1170{
1171 spin_lock(&fmp->lock);
1172 fmeter_update(fmp);
1173 fmp->cnt = min(FM_MAXCNT, fmp->cnt + FM_SCALE);
1174 spin_unlock(&fmp->lock);
1175}
1176
1177/* Process any previous ticks, then return current value. */
1178static int fmeter_getrate(struct fmeter *fmp)
1179{
1180 int val;
1181
1182 spin_lock(&fmp->lock);
1183 fmeter_update(fmp);
1184 val = fmp->val;
1185 spin_unlock(&fmp->lock);
1186 return val;
1187}
1188
Paul Menage2df167a2008-02-07 00:14:45 -08001189/* Called by cgroups to determine if a cpuset is usable; cgroup_mutex held */
Paul Menage8793d852007-10-18 23:39:39 -07001190static int cpuset_can_attach(struct cgroup_subsys *ss,
1191 struct cgroup *cont, struct task_struct *tsk)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001192{
Paul Menage8793d852007-10-18 23:39:39 -07001193 struct cpuset *cs = cgroup_cs(cont);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001194
Linus Torvalds1da177e2005-04-16 15:20:36 -07001195 if (cpus_empty(cs->cpus_allowed) || nodes_empty(cs->mems_allowed))
1196 return -ENOSPC;
David Rientjes9985b0b2008-06-05 12:57:11 -07001197 if (tsk->flags & PF_THREAD_BOUND) {
1198 cpumask_t mask;
1199
1200 mutex_lock(&callback_mutex);
1201 mask = cs->cpus_allowed;
1202 mutex_unlock(&callback_mutex);
1203 if (!cpus_equal(tsk->cpus_allowed, mask))
1204 return -EINVAL;
1205 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001206
Paul Menage8793d852007-10-18 23:39:39 -07001207 return security_task_setscheduler(tsk, 0, NULL);
1208}
Linus Torvalds1da177e2005-04-16 15:20:36 -07001209
Paul Menage8793d852007-10-18 23:39:39 -07001210static void cpuset_attach(struct cgroup_subsys *ss,
1211 struct cgroup *cont, struct cgroup *oldcont,
1212 struct task_struct *tsk)
1213{
1214 cpumask_t cpus;
1215 nodemask_t from, to;
1216 struct mm_struct *mm;
1217 struct cpuset *cs = cgroup_cs(cont);
1218 struct cpuset *oldcs = cgroup_cs(oldcont);
David Rientjes9985b0b2008-06-05 12:57:11 -07001219 int err;
David Quigley22fb52d2006-06-23 02:04:00 -07001220
Ingo Molnar3d3f26a2006-03-23 03:00:18 -08001221 mutex_lock(&callback_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001222 guarantee_online_cpus(cs, &cpus);
David Rientjes9985b0b2008-06-05 12:57:11 -07001223 err = set_cpus_allowed_ptr(tsk, &cpus);
Paul Menage8793d852007-10-18 23:39:39 -07001224 mutex_unlock(&callback_mutex);
David Rientjes9985b0b2008-06-05 12:57:11 -07001225 if (err)
1226 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001227
Paul Jackson45b07ef2006-01-08 01:00:56 -08001228 from = oldcs->mems_allowed;
1229 to = cs->mems_allowed;
Paul Jackson42253992006-01-08 01:01:59 -08001230 mm = get_task_mm(tsk);
1231 if (mm) {
1232 mpol_rebind_mm(mm, &to);
Paul Jackson2741a552006-03-31 02:30:51 -08001233 if (is_memory_migrate(cs))
Paul Jacksone4e364e2006-03-31 02:30:52 -08001234 cpuset_migrate_mm(mm, &from, &to);
Paul Jackson42253992006-01-08 01:01:59 -08001235 mmput(mm);
1236 }
1237
Linus Torvalds1da177e2005-04-16 15:20:36 -07001238}
1239
1240/* The various types of files and directories in a cpuset file system */
1241
1242typedef enum {
Paul Jackson45b07ef2006-01-08 01:00:56 -08001243 FILE_MEMORY_MIGRATE,
Linus Torvalds1da177e2005-04-16 15:20:36 -07001244 FILE_CPULIST,
1245 FILE_MEMLIST,
1246 FILE_CPU_EXCLUSIVE,
1247 FILE_MEM_EXCLUSIVE,
Paul Menage78608362008-04-29 01:00:26 -07001248 FILE_MEM_HARDWALL,
Paul Jackson029190c2007-10-18 23:40:20 -07001249 FILE_SCHED_LOAD_BALANCE,
Hidetoshi Seto1d3504f2008-04-15 14:04:23 +09001250 FILE_SCHED_RELAX_DOMAIN_LEVEL,
Paul Jackson3e0d98b2006-01-08 01:01:49 -08001251 FILE_MEMORY_PRESSURE_ENABLED,
1252 FILE_MEMORY_PRESSURE,
Paul Jackson825a46a2006-03-24 03:16:03 -08001253 FILE_SPREAD_PAGE,
1254 FILE_SPREAD_SLAB,
Linus Torvalds1da177e2005-04-16 15:20:36 -07001255} cpuset_filetype_t;
1256
Paul Menage8793d852007-10-18 23:39:39 -07001257static ssize_t cpuset_common_file_write(struct cgroup *cont,
1258 struct cftype *cft,
1259 struct file *file,
Paul Menaged3ed11c2006-12-06 20:41:37 -08001260 const char __user *userbuf,
Linus Torvalds1da177e2005-04-16 15:20:36 -07001261 size_t nbytes, loff_t *unused_ppos)
1262{
Paul Menage8793d852007-10-18 23:39:39 -07001263 struct cpuset *cs = cgroup_cs(cont);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001264 cpuset_filetype_t type = cft->private;
1265 char *buffer;
1266 int retval = 0;
1267
1268 /* Crude upper limit on largest legitimate cpulist user might write. */
Paul Jackson029190c2007-10-18 23:40:20 -07001269 if (nbytes > 100U + 6 * max(NR_CPUS, MAX_NUMNODES))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001270 return -E2BIG;
1271
1272 /* +1 for nul-terminator */
Harvey Harrisonb331d252008-04-28 14:13:19 -07001273 buffer = kmalloc(nbytes + 1, GFP_KERNEL);
1274 if (!buffer)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001275 return -ENOMEM;
1276
1277 if (copy_from_user(buffer, userbuf, nbytes)) {
1278 retval = -EFAULT;
1279 goto out1;
1280 }
1281 buffer[nbytes] = 0; /* nul-terminate */
1282
Paul Menage8793d852007-10-18 23:39:39 -07001283 cgroup_lock();
Linus Torvalds1da177e2005-04-16 15:20:36 -07001284
Paul Menage8793d852007-10-18 23:39:39 -07001285 if (cgroup_is_removed(cont)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001286 retval = -ENODEV;
1287 goto out2;
1288 }
1289
1290 switch (type) {
1291 case FILE_CPULIST:
1292 retval = update_cpumask(cs, buffer);
1293 break;
1294 case FILE_MEMLIST:
1295 retval = update_nodemask(cs, buffer);
1296 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001297 default:
1298 retval = -EINVAL;
1299 goto out2;
1300 }
1301
1302 if (retval == 0)
1303 retval = nbytes;
1304out2:
Paul Menage8793d852007-10-18 23:39:39 -07001305 cgroup_unlock();
Linus Torvalds1da177e2005-04-16 15:20:36 -07001306out1:
1307 kfree(buffer);
1308 return retval;
1309}
1310
Paul Menage700fe1a2008-04-29 01:00:00 -07001311static int cpuset_write_u64(struct cgroup *cgrp, struct cftype *cft, u64 val)
1312{
1313 int retval = 0;
1314 struct cpuset *cs = cgroup_cs(cgrp);
1315 cpuset_filetype_t type = cft->private;
1316
1317 cgroup_lock();
1318
1319 if (cgroup_is_removed(cgrp)) {
1320 cgroup_unlock();
1321 return -ENODEV;
1322 }
1323
1324 switch (type) {
1325 case FILE_CPU_EXCLUSIVE:
1326 retval = update_flag(CS_CPU_EXCLUSIVE, cs, val);
1327 break;
1328 case FILE_MEM_EXCLUSIVE:
1329 retval = update_flag(CS_MEM_EXCLUSIVE, cs, val);
1330 break;
Paul Menage78608362008-04-29 01:00:26 -07001331 case FILE_MEM_HARDWALL:
1332 retval = update_flag(CS_MEM_HARDWALL, cs, val);
1333 break;
Paul Menage700fe1a2008-04-29 01:00:00 -07001334 case FILE_SCHED_LOAD_BALANCE:
1335 retval = update_flag(CS_SCHED_LOAD_BALANCE, cs, val);
1336 break;
1337 case FILE_MEMORY_MIGRATE:
1338 retval = update_flag(CS_MEMORY_MIGRATE, cs, val);
1339 break;
1340 case FILE_MEMORY_PRESSURE_ENABLED:
1341 cpuset_memory_pressure_enabled = !!val;
1342 break;
1343 case FILE_MEMORY_PRESSURE:
1344 retval = -EACCES;
1345 break;
1346 case FILE_SPREAD_PAGE:
1347 retval = update_flag(CS_SPREAD_PAGE, cs, val);
1348 cs->mems_generation = cpuset_mems_generation++;
1349 break;
1350 case FILE_SPREAD_SLAB:
1351 retval = update_flag(CS_SPREAD_SLAB, cs, val);
1352 cs->mems_generation = cpuset_mems_generation++;
1353 break;
1354 default:
1355 retval = -EINVAL;
1356 break;
1357 }
1358 cgroup_unlock();
1359 return retval;
1360}
1361
Paul Menage5be7a472008-05-06 20:42:41 -07001362static int cpuset_write_s64(struct cgroup *cgrp, struct cftype *cft, s64 val)
1363{
1364 int retval = 0;
1365 struct cpuset *cs = cgroup_cs(cgrp);
1366 cpuset_filetype_t type = cft->private;
1367
1368 cgroup_lock();
1369
1370 if (cgroup_is_removed(cgrp)) {
1371 cgroup_unlock();
1372 return -ENODEV;
1373 }
1374 switch (type) {
1375 case FILE_SCHED_RELAX_DOMAIN_LEVEL:
1376 retval = update_relax_domain_level(cs, val);
1377 break;
1378 default:
1379 retval = -EINVAL;
1380 break;
1381 }
1382 cgroup_unlock();
1383 return retval;
1384}
1385
Linus Torvalds1da177e2005-04-16 15:20:36 -07001386/*
1387 * These ascii lists should be read in a single call, by using a user
1388 * buffer large enough to hold the entire map. If read in smaller
1389 * chunks, there is no guarantee of atomicity. Since the display format
1390 * used, list of ranges of sequential numbers, is variable length,
1391 * and since these maps can change value dynamically, one could read
1392 * gibberish by doing partial reads while a list was changing.
1393 * A single large read to a buffer that crosses a page boundary is
1394 * ok, because the result being copied to user land is not recomputed
1395 * across a page fault.
1396 */
1397
1398static int cpuset_sprintf_cpulist(char *page, struct cpuset *cs)
1399{
1400 cpumask_t mask;
1401
Ingo Molnar3d3f26a2006-03-23 03:00:18 -08001402 mutex_lock(&callback_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001403 mask = cs->cpus_allowed;
Ingo Molnar3d3f26a2006-03-23 03:00:18 -08001404 mutex_unlock(&callback_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001405
1406 return cpulist_scnprintf(page, PAGE_SIZE, mask);
1407}
1408
1409static int cpuset_sprintf_memlist(char *page, struct cpuset *cs)
1410{
1411 nodemask_t mask;
1412
Ingo Molnar3d3f26a2006-03-23 03:00:18 -08001413 mutex_lock(&callback_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001414 mask = cs->mems_allowed;
Ingo Molnar3d3f26a2006-03-23 03:00:18 -08001415 mutex_unlock(&callback_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001416
1417 return nodelist_scnprintf(page, PAGE_SIZE, mask);
1418}
1419
Paul Menage8793d852007-10-18 23:39:39 -07001420static ssize_t cpuset_common_file_read(struct cgroup *cont,
1421 struct cftype *cft,
1422 struct file *file,
1423 char __user *buf,
1424 size_t nbytes, loff_t *ppos)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001425{
Paul Menage8793d852007-10-18 23:39:39 -07001426 struct cpuset *cs = cgroup_cs(cont);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001427 cpuset_filetype_t type = cft->private;
1428 char *page;
1429 ssize_t retval = 0;
1430 char *s;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001431
Mel Gormane12ba742007-10-16 01:25:52 -07001432 if (!(page = (char *)__get_free_page(GFP_TEMPORARY)))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001433 return -ENOMEM;
1434
1435 s = page;
1436
1437 switch (type) {
1438 case FILE_CPULIST:
1439 s += cpuset_sprintf_cpulist(s, cs);
1440 break;
1441 case FILE_MEMLIST:
1442 s += cpuset_sprintf_memlist(s, cs);
1443 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001444 default:
1445 retval = -EINVAL;
1446 goto out;
1447 }
1448 *s++ = '\n';
Linus Torvalds1da177e2005-04-16 15:20:36 -07001449
Al Viroeacaa1f2005-09-30 03:26:43 +01001450 retval = simple_read_from_buffer(buf, nbytes, ppos, page, s - page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001451out:
1452 free_page((unsigned long)page);
1453 return retval;
1454}
1455
Paul Menage700fe1a2008-04-29 01:00:00 -07001456static u64 cpuset_read_u64(struct cgroup *cont, struct cftype *cft)
1457{
1458 struct cpuset *cs = cgroup_cs(cont);
1459 cpuset_filetype_t type = cft->private;
1460 switch (type) {
1461 case FILE_CPU_EXCLUSIVE:
1462 return is_cpu_exclusive(cs);
1463 case FILE_MEM_EXCLUSIVE:
1464 return is_mem_exclusive(cs);
Paul Menage78608362008-04-29 01:00:26 -07001465 case FILE_MEM_HARDWALL:
1466 return is_mem_hardwall(cs);
Paul Menage700fe1a2008-04-29 01:00:00 -07001467 case FILE_SCHED_LOAD_BALANCE:
1468 return is_sched_load_balance(cs);
1469 case FILE_MEMORY_MIGRATE:
1470 return is_memory_migrate(cs);
1471 case FILE_MEMORY_PRESSURE_ENABLED:
1472 return cpuset_memory_pressure_enabled;
1473 case FILE_MEMORY_PRESSURE:
1474 return fmeter_getrate(&cs->fmeter);
1475 case FILE_SPREAD_PAGE:
1476 return is_spread_page(cs);
1477 case FILE_SPREAD_SLAB:
1478 return is_spread_slab(cs);
1479 default:
1480 BUG();
1481 }
1482}
Linus Torvalds1da177e2005-04-16 15:20:36 -07001483
Paul Menage5be7a472008-05-06 20:42:41 -07001484static s64 cpuset_read_s64(struct cgroup *cont, struct cftype *cft)
1485{
1486 struct cpuset *cs = cgroup_cs(cont);
1487 cpuset_filetype_t type = cft->private;
1488 switch (type) {
1489 case FILE_SCHED_RELAX_DOMAIN_LEVEL:
1490 return cs->relax_domain_level;
1491 default:
1492 BUG();
1493 }
1494}
1495
Linus Torvalds1da177e2005-04-16 15:20:36 -07001496
1497/*
1498 * for the common functions, 'private' gives the type of file
1499 */
1500
Paul Menageaddf2c72008-04-29 01:00:26 -07001501static struct cftype files[] = {
1502 {
1503 .name = "cpus",
1504 .read = cpuset_common_file_read,
1505 .write = cpuset_common_file_write,
1506 .private = FILE_CPULIST,
1507 },
Linus Torvalds1da177e2005-04-16 15:20:36 -07001508
Paul Menageaddf2c72008-04-29 01:00:26 -07001509 {
1510 .name = "mems",
1511 .read = cpuset_common_file_read,
1512 .write = cpuset_common_file_write,
1513 .private = FILE_MEMLIST,
1514 },
Linus Torvalds1da177e2005-04-16 15:20:36 -07001515
Paul Menageaddf2c72008-04-29 01:00:26 -07001516 {
1517 .name = "cpu_exclusive",
1518 .read_u64 = cpuset_read_u64,
1519 .write_u64 = cpuset_write_u64,
1520 .private = FILE_CPU_EXCLUSIVE,
1521 },
Linus Torvalds1da177e2005-04-16 15:20:36 -07001522
Paul Menageaddf2c72008-04-29 01:00:26 -07001523 {
1524 .name = "mem_exclusive",
1525 .read_u64 = cpuset_read_u64,
1526 .write_u64 = cpuset_write_u64,
1527 .private = FILE_MEM_EXCLUSIVE,
1528 },
Linus Torvalds1da177e2005-04-16 15:20:36 -07001529
Paul Menageaddf2c72008-04-29 01:00:26 -07001530 {
Paul Menage78608362008-04-29 01:00:26 -07001531 .name = "mem_hardwall",
1532 .read_u64 = cpuset_read_u64,
1533 .write_u64 = cpuset_write_u64,
1534 .private = FILE_MEM_HARDWALL,
1535 },
1536
1537 {
Paul Menageaddf2c72008-04-29 01:00:26 -07001538 .name = "sched_load_balance",
1539 .read_u64 = cpuset_read_u64,
1540 .write_u64 = cpuset_write_u64,
1541 .private = FILE_SCHED_LOAD_BALANCE,
1542 },
Paul Jackson029190c2007-10-18 23:40:20 -07001543
Paul Menageaddf2c72008-04-29 01:00:26 -07001544 {
1545 .name = "sched_relax_domain_level",
Paul Menage5be7a472008-05-06 20:42:41 -07001546 .read_s64 = cpuset_read_s64,
1547 .write_s64 = cpuset_write_s64,
Paul Menageaddf2c72008-04-29 01:00:26 -07001548 .private = FILE_SCHED_RELAX_DOMAIN_LEVEL,
1549 },
Hidetoshi Seto1d3504f2008-04-15 14:04:23 +09001550
Paul Menageaddf2c72008-04-29 01:00:26 -07001551 {
1552 .name = "memory_migrate",
1553 .read_u64 = cpuset_read_u64,
1554 .write_u64 = cpuset_write_u64,
1555 .private = FILE_MEMORY_MIGRATE,
1556 },
1557
1558 {
1559 .name = "memory_pressure",
1560 .read_u64 = cpuset_read_u64,
1561 .write_u64 = cpuset_write_u64,
1562 .private = FILE_MEMORY_PRESSURE,
1563 },
1564
1565 {
1566 .name = "memory_spread_page",
1567 .read_u64 = cpuset_read_u64,
1568 .write_u64 = cpuset_write_u64,
1569 .private = FILE_SPREAD_PAGE,
1570 },
1571
1572 {
1573 .name = "memory_spread_slab",
1574 .read_u64 = cpuset_read_u64,
1575 .write_u64 = cpuset_write_u64,
1576 .private = FILE_SPREAD_SLAB,
1577 },
Paul Jackson45b07ef2006-01-08 01:00:56 -08001578};
1579
Paul Jackson3e0d98b2006-01-08 01:01:49 -08001580static struct cftype cft_memory_pressure_enabled = {
1581 .name = "memory_pressure_enabled",
Paul Menage700fe1a2008-04-29 01:00:00 -07001582 .read_u64 = cpuset_read_u64,
1583 .write_u64 = cpuset_write_u64,
Paul Jackson3e0d98b2006-01-08 01:01:49 -08001584 .private = FILE_MEMORY_PRESSURE_ENABLED,
1585};
1586
Paul Menage8793d852007-10-18 23:39:39 -07001587static int cpuset_populate(struct cgroup_subsys *ss, struct cgroup *cont)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001588{
1589 int err;
1590
Paul Menageaddf2c72008-04-29 01:00:26 -07001591 err = cgroup_add_files(cont, ss, files, ARRAY_SIZE(files));
1592 if (err)
Paul Jackson825a46a2006-03-24 03:16:03 -08001593 return err;
Paul Menage8793d852007-10-18 23:39:39 -07001594 /* memory_pressure_enabled is in root cpuset only */
Paul Menageaddf2c72008-04-29 01:00:26 -07001595 if (!cont->parent)
Paul Menage8793d852007-10-18 23:39:39 -07001596 err = cgroup_add_file(cont, ss,
Paul Menageaddf2c72008-04-29 01:00:26 -07001597 &cft_memory_pressure_enabled);
1598 return err;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001599}
1600
1601/*
Paul Menage8793d852007-10-18 23:39:39 -07001602 * post_clone() is called at the end of cgroup_clone().
1603 * 'cgroup' was just created automatically as a result of
1604 * a cgroup_clone(), and the current task is about to
1605 * be moved into 'cgroup'.
1606 *
1607 * Currently we refuse to set up the cgroup - thereby
1608 * refusing the task to be entered, and as a result refusing
1609 * the sys_unshare() or clone() which initiated it - if any
1610 * sibling cpusets have exclusive cpus or mem.
1611 *
1612 * If this becomes a problem for some users who wish to
1613 * allow that scenario, then cpuset_post_clone() could be
1614 * changed to grant parent->cpus_allowed-sibling_cpus_exclusive
Paul Menage2df167a2008-02-07 00:14:45 -08001615 * (and likewise for mems) to the new cgroup. Called with cgroup_mutex
1616 * held.
Paul Menage8793d852007-10-18 23:39:39 -07001617 */
1618static void cpuset_post_clone(struct cgroup_subsys *ss,
1619 struct cgroup *cgroup)
1620{
1621 struct cgroup *parent, *child;
1622 struct cpuset *cs, *parent_cs;
1623
1624 parent = cgroup->parent;
1625 list_for_each_entry(child, &parent->children, sibling) {
1626 cs = cgroup_cs(child);
1627 if (is_mem_exclusive(cs) || is_cpu_exclusive(cs))
1628 return;
1629 }
1630 cs = cgroup_cs(cgroup);
1631 parent_cs = cgroup_cs(parent);
1632
1633 cs->mems_allowed = parent_cs->mems_allowed;
1634 cs->cpus_allowed = parent_cs->cpus_allowed;
1635 return;
1636}
1637
1638/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07001639 * cpuset_create - create a cpuset
Paul Menage2df167a2008-02-07 00:14:45 -08001640 * ss: cpuset cgroup subsystem
1641 * cont: control group that the new cpuset will be part of
Linus Torvalds1da177e2005-04-16 15:20:36 -07001642 */
1643
Paul Menage8793d852007-10-18 23:39:39 -07001644static struct cgroup_subsys_state *cpuset_create(
1645 struct cgroup_subsys *ss,
1646 struct cgroup *cont)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001647{
1648 struct cpuset *cs;
Paul Menage8793d852007-10-18 23:39:39 -07001649 struct cpuset *parent;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001650
Paul Menage8793d852007-10-18 23:39:39 -07001651 if (!cont->parent) {
1652 /* This is early initialization for the top cgroup */
1653 top_cpuset.mems_generation = cpuset_mems_generation++;
1654 return &top_cpuset.css;
1655 }
1656 parent = cgroup_cs(cont->parent);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001657 cs = kmalloc(sizeof(*cs), GFP_KERNEL);
1658 if (!cs)
Paul Menage8793d852007-10-18 23:39:39 -07001659 return ERR_PTR(-ENOMEM);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001660
Paul Jacksoncf2a473c2006-01-08 01:01:54 -08001661 cpuset_update_task_memory_state();
Linus Torvalds1da177e2005-04-16 15:20:36 -07001662 cs->flags = 0;
Paul Jackson825a46a2006-03-24 03:16:03 -08001663 if (is_spread_page(parent))
1664 set_bit(CS_SPREAD_PAGE, &cs->flags);
1665 if (is_spread_slab(parent))
1666 set_bit(CS_SPREAD_SLAB, &cs->flags);
Paul Jackson029190c2007-10-18 23:40:20 -07001667 set_bit(CS_SCHED_LOAD_BALANCE, &cs->flags);
Mike Travisf9a86fc2008-04-04 18:11:07 -07001668 cpus_clear(cs->cpus_allowed);
1669 nodes_clear(cs->mems_allowed);
Paul Jackson151a4422006-03-24 03:16:11 -08001670 cs->mems_generation = cpuset_mems_generation++;
Paul Jackson3e0d98b2006-01-08 01:01:49 -08001671 fmeter_init(&cs->fmeter);
Hidetoshi Seto1d3504f2008-04-15 14:04:23 +09001672 cs->relax_domain_level = -1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001673
1674 cs->parent = parent;
Paul Jackson202f72d2006-01-08 01:01:57 -08001675 number_of_cpusets++;
Paul Menage8793d852007-10-18 23:39:39 -07001676 return &cs->css ;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001677}
1678
Paul Jackson029190c2007-10-18 23:40:20 -07001679/*
1680 * Locking note on the strange update_flag() call below:
1681 *
1682 * If the cpuset being removed has its flag 'sched_load_balance'
1683 * enabled, then simulate turning sched_load_balance off, which
Gautham R Shenoy86ef5c92008-01-25 21:08:02 +01001684 * will call rebuild_sched_domains(). The get_online_cpus()
Paul Jackson029190c2007-10-18 23:40:20 -07001685 * call in rebuild_sched_domains() must not be made while holding
1686 * callback_mutex. Elsewhere the kernel nests callback_mutex inside
Gautham R Shenoy86ef5c92008-01-25 21:08:02 +01001687 * get_online_cpus() calls. So the reverse nesting would risk an
Paul Jackson029190c2007-10-18 23:40:20 -07001688 * ABBA deadlock.
1689 */
1690
Paul Menage8793d852007-10-18 23:39:39 -07001691static void cpuset_destroy(struct cgroup_subsys *ss, struct cgroup *cont)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001692{
Paul Menage8793d852007-10-18 23:39:39 -07001693 struct cpuset *cs = cgroup_cs(cont);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001694
Paul Jacksoncf2a473c2006-01-08 01:01:54 -08001695 cpuset_update_task_memory_state();
Paul Jackson029190c2007-10-18 23:40:20 -07001696
1697 if (is_sched_load_balance(cs))
Paul Menage700fe1a2008-04-29 01:00:00 -07001698 update_flag(CS_SCHED_LOAD_BALANCE, cs, 0);
Paul Jackson029190c2007-10-18 23:40:20 -07001699
Paul Jackson202f72d2006-01-08 01:01:57 -08001700 number_of_cpusets--;
Paul Menage8793d852007-10-18 23:39:39 -07001701 kfree(cs);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001702}
1703
Paul Menage8793d852007-10-18 23:39:39 -07001704struct cgroup_subsys cpuset_subsys = {
1705 .name = "cpuset",
1706 .create = cpuset_create,
1707 .destroy = cpuset_destroy,
1708 .can_attach = cpuset_can_attach,
1709 .attach = cpuset_attach,
1710 .populate = cpuset_populate,
1711 .post_clone = cpuset_post_clone,
1712 .subsys_id = cpuset_subsys_id,
1713 .early_init = 1,
1714};
1715
Paul Jacksonc417f022006-01-08 01:02:01 -08001716/*
1717 * cpuset_init_early - just enough so that the calls to
1718 * cpuset_update_task_memory_state() in early init code
1719 * are harmless.
1720 */
1721
1722int __init cpuset_init_early(void)
1723{
Paul Menage8793d852007-10-18 23:39:39 -07001724 top_cpuset.mems_generation = cpuset_mems_generation++;
Paul Jacksonc417f022006-01-08 01:02:01 -08001725 return 0;
1726}
1727
Paul Menage8793d852007-10-18 23:39:39 -07001728
Linus Torvalds1da177e2005-04-16 15:20:36 -07001729/**
1730 * cpuset_init - initialize cpusets at system boot
1731 *
1732 * Description: Initialize top_cpuset and the cpuset internal file system,
1733 **/
1734
1735int __init cpuset_init(void)
1736{
Paul Menage8793d852007-10-18 23:39:39 -07001737 int err = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001738
Mike Travisf9a86fc2008-04-04 18:11:07 -07001739 cpus_setall(top_cpuset.cpus_allowed);
1740 nodes_setall(top_cpuset.mems_allowed);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001741
Paul Jackson3e0d98b2006-01-08 01:01:49 -08001742 fmeter_init(&top_cpuset.fmeter);
Paul Jackson151a4422006-03-24 03:16:11 -08001743 top_cpuset.mems_generation = cpuset_mems_generation++;
Paul Jackson029190c2007-10-18 23:40:20 -07001744 set_bit(CS_SCHED_LOAD_BALANCE, &top_cpuset.flags);
Hidetoshi Seto1d3504f2008-04-15 14:04:23 +09001745 top_cpuset.relax_domain_level = -1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001746
Linus Torvalds1da177e2005-04-16 15:20:36 -07001747 err = register_filesystem(&cpuset_fs_type);
1748 if (err < 0)
Paul Menage8793d852007-10-18 23:39:39 -07001749 return err;
1750
Paul Jackson202f72d2006-01-08 01:01:57 -08001751 number_of_cpusets = 1;
Paul Menage8793d852007-10-18 23:39:39 -07001752 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001753}
1754
Cliff Wickman956db3c2008-02-07 00:14:43 -08001755/**
1756 * cpuset_do_move_task - move a given task to another cpuset
1757 * @tsk: pointer to task_struct the task to move
1758 * @scan: struct cgroup_scanner contained in its struct cpuset_hotplug_scanner
1759 *
1760 * Called by cgroup_scan_tasks() for each task in a cgroup.
1761 * Return nonzero to stop the walk through the tasks.
1762 */
Adrian Bunk9e0c9142008-04-29 01:00:25 -07001763static void cpuset_do_move_task(struct task_struct *tsk,
1764 struct cgroup_scanner *scan)
Cliff Wickman956db3c2008-02-07 00:14:43 -08001765{
1766 struct cpuset_hotplug_scanner *chsp;
1767
1768 chsp = container_of(scan, struct cpuset_hotplug_scanner, scan);
1769 cgroup_attach_task(chsp->to, tsk);
1770}
1771
1772/**
1773 * move_member_tasks_to_cpuset - move tasks from one cpuset to another
1774 * @from: cpuset in which the tasks currently reside
1775 * @to: cpuset to which the tasks will be moved
1776 *
Paul Jacksonc8d9c902008-02-07 00:14:46 -08001777 * Called with cgroup_mutex held
1778 * callback_mutex must not be held, as cpuset_attach() will take it.
Cliff Wickman956db3c2008-02-07 00:14:43 -08001779 *
1780 * The cgroup_scan_tasks() function will scan all the tasks in a cgroup,
1781 * calling callback functions for each.
1782 */
1783static void move_member_tasks_to_cpuset(struct cpuset *from, struct cpuset *to)
1784{
1785 struct cpuset_hotplug_scanner scan;
1786
1787 scan.scan.cg = from->css.cgroup;
1788 scan.scan.test_task = NULL; /* select all tasks in cgroup */
1789 scan.scan.process_task = cpuset_do_move_task;
1790 scan.scan.heap = NULL;
1791 scan.to = to->css.cgroup;
1792
1793 if (cgroup_scan_tasks((struct cgroup_scanner *)&scan))
1794 printk(KERN_ERR "move_member_tasks_to_cpuset: "
1795 "cgroup_scan_tasks failed\n");
1796}
1797
Paul Jacksonb1aac8b2006-09-29 02:01:17 -07001798/*
1799 * If common_cpu_mem_hotplug_unplug(), below, unplugs any CPUs
1800 * or memory nodes, we need to walk over the cpuset hierarchy,
1801 * removing that CPU or node from all cpusets. If this removes the
Cliff Wickman956db3c2008-02-07 00:14:43 -08001802 * last CPU or node from a cpuset, then move the tasks in the empty
1803 * cpuset to its next-highest non-empty parent.
Paul Jacksonb1aac8b2006-09-29 02:01:17 -07001804 *
Paul Jacksonc8d9c902008-02-07 00:14:46 -08001805 * Called with cgroup_mutex held
1806 * callback_mutex must not be held, as cpuset_attach() will take it.
Paul Jacksonb1aac8b2006-09-29 02:01:17 -07001807 */
Cliff Wickman956db3c2008-02-07 00:14:43 -08001808static void remove_tasks_in_empty_cpuset(struct cpuset *cs)
Paul Jacksonb1aac8b2006-09-29 02:01:17 -07001809{
Cliff Wickman956db3c2008-02-07 00:14:43 -08001810 struct cpuset *parent;
Paul Jacksonb1aac8b2006-09-29 02:01:17 -07001811
Paul Jacksonc8d9c902008-02-07 00:14:46 -08001812 /*
1813 * The cgroup's css_sets list is in use if there are tasks
1814 * in the cpuset; the list is empty if there are none;
1815 * the cs->css.refcnt seems always 0.
1816 */
Cliff Wickman956db3c2008-02-07 00:14:43 -08001817 if (list_empty(&cs->css.cgroup->css_sets))
1818 return;
1819
1820 /*
1821 * Find its next-highest non-empty parent, (top cpuset
1822 * has online cpus, so can't be empty).
1823 */
1824 parent = cs->parent;
Paul Jacksonb4501292008-02-07 00:14:47 -08001825 while (cpus_empty(parent->cpus_allowed) ||
1826 nodes_empty(parent->mems_allowed))
Cliff Wickman956db3c2008-02-07 00:14:43 -08001827 parent = parent->parent;
Cliff Wickman956db3c2008-02-07 00:14:43 -08001828
1829 move_member_tasks_to_cpuset(cs, parent);
1830}
1831
1832/*
1833 * Walk the specified cpuset subtree and look for empty cpusets.
1834 * The tasks of such cpuset must be moved to a parent cpuset.
1835 *
Paul Menage2df167a2008-02-07 00:14:45 -08001836 * Called with cgroup_mutex held. We take callback_mutex to modify
Cliff Wickman956db3c2008-02-07 00:14:43 -08001837 * cpus_allowed and mems_allowed.
1838 *
1839 * This walk processes the tree from top to bottom, completing one layer
1840 * before dropping down to the next. It always processes a node before
1841 * any of its children.
1842 *
1843 * For now, since we lack memory hot unplug, we'll never see a cpuset
1844 * that has tasks along with an empty 'mems'. But if we did see such
1845 * a cpuset, we'd handle it just like we do if its 'cpus' was empty.
1846 */
1847static void scan_for_empty_cpusets(const struct cpuset *root)
1848{
1849 struct cpuset *cp; /* scans cpusets being updated */
1850 struct cpuset *child; /* scans child cpusets of cp */
1851 struct list_head queue;
1852 struct cgroup *cont;
1853
1854 INIT_LIST_HEAD(&queue);
1855
1856 list_add_tail((struct list_head *)&root->stack_list, &queue);
1857
Cliff Wickman956db3c2008-02-07 00:14:43 -08001858 while (!list_empty(&queue)) {
1859 cp = container_of(queue.next, struct cpuset, stack_list);
1860 list_del(queue.next);
1861 list_for_each_entry(cont, &cp->css.cgroup->children, sibling) {
1862 child = cgroup_cs(cont);
1863 list_add_tail(&child->stack_list, &queue);
1864 }
1865 cont = cp->css.cgroup;
Paul Jacksonb4501292008-02-07 00:14:47 -08001866
1867 /* Continue past cpusets with all cpus, mems online */
1868 if (cpus_subset(cp->cpus_allowed, cpu_online_map) &&
1869 nodes_subset(cp->mems_allowed, node_states[N_HIGH_MEMORY]))
1870 continue;
1871
Cliff Wickman956db3c2008-02-07 00:14:43 -08001872 /* Remove offline cpus and mems from this cpuset. */
Paul Jacksonb4501292008-02-07 00:14:47 -08001873 mutex_lock(&callback_mutex);
Cliff Wickman956db3c2008-02-07 00:14:43 -08001874 cpus_and(cp->cpus_allowed, cp->cpus_allowed, cpu_online_map);
1875 nodes_and(cp->mems_allowed, cp->mems_allowed,
1876 node_states[N_HIGH_MEMORY]);
Paul Jacksonb4501292008-02-07 00:14:47 -08001877 mutex_unlock(&callback_mutex);
1878
1879 /* Move tasks from the empty cpuset to a parent */
Paul Jacksonc8d9c902008-02-07 00:14:46 -08001880 if (cpus_empty(cp->cpus_allowed) ||
Paul Jacksonb4501292008-02-07 00:14:47 -08001881 nodes_empty(cp->mems_allowed))
Cliff Wickman956db3c2008-02-07 00:14:43 -08001882 remove_tasks_in_empty_cpuset(cp);
Cliff Wickman956db3c2008-02-07 00:14:43 -08001883 }
Paul Jacksonb1aac8b2006-09-29 02:01:17 -07001884}
1885
1886/*
1887 * The cpus_allowed and mems_allowed nodemasks in the top_cpuset track
Christoph Lameter0e1e7c72007-10-16 01:25:38 -07001888 * cpu_online_map and node_states[N_HIGH_MEMORY]. Force the top cpuset to
Cliff Wickman956db3c2008-02-07 00:14:43 -08001889 * track what's online after any CPU or memory node hotplug or unplug event.
Paul Jacksonb1aac8b2006-09-29 02:01:17 -07001890 *
1891 * Since there are two callers of this routine, one for CPU hotplug
1892 * events and one for memory node hotplug events, we could have coded
1893 * two separate routines here. We code it as a single common routine
1894 * in order to minimize text size.
1895 */
1896
Dmitry Adamushko3e840502008-07-13 02:10:29 +02001897static void common_cpu_mem_hotplug_unplug(int rebuild_sd)
Paul Jacksonb1aac8b2006-09-29 02:01:17 -07001898{
Paul Menage8793d852007-10-18 23:39:39 -07001899 cgroup_lock();
Paul Jacksonb1aac8b2006-09-29 02:01:17 -07001900
Paul Jacksonb1aac8b2006-09-29 02:01:17 -07001901 top_cpuset.cpus_allowed = cpu_online_map;
Christoph Lameter0e1e7c72007-10-16 01:25:38 -07001902 top_cpuset.mems_allowed = node_states[N_HIGH_MEMORY];
Cliff Wickman956db3c2008-02-07 00:14:43 -08001903 scan_for_empty_cpusets(&top_cpuset);
Paul Jacksonb1aac8b2006-09-29 02:01:17 -07001904
Max Krasnyansky5c8e1ed2008-05-29 11:17:01 -07001905 /*
1906 * Scheduler destroys domains on hotplug events.
1907 * Rebuild them based on the current settings.
1908 */
Dmitry Adamushko3e840502008-07-13 02:10:29 +02001909 if (rebuild_sd)
1910 rebuild_sched_domains();
Max Krasnyansky5c8e1ed2008-05-29 11:17:01 -07001911
Paul Menage8793d852007-10-18 23:39:39 -07001912 cgroup_unlock();
Paul Jacksonb1aac8b2006-09-29 02:01:17 -07001913}
Paul Jacksonb1aac8b2006-09-29 02:01:17 -07001914
Paul Jackson4c4d50f2006-08-27 01:23:51 -07001915/*
1916 * The top_cpuset tracks what CPUs and Memory Nodes are online,
1917 * period. This is necessary in order to make cpusets transparent
1918 * (of no affect) on systems that are actively using CPU hotplug
1919 * but making no active use of cpusets.
1920 *
Paul Jackson38837fc2006-09-29 02:01:16 -07001921 * This routine ensures that top_cpuset.cpus_allowed tracks
1922 * cpu_online_map on each CPU hotplug (cpuhp) event.
Paul Jackson4c4d50f2006-08-27 01:23:51 -07001923 */
1924
Paul Jackson029190c2007-10-18 23:40:20 -07001925static int cpuset_handle_cpuhp(struct notifier_block *unused_nb,
1926 unsigned long phase, void *unused_cpu)
Paul Jackson4c4d50f2006-08-27 01:23:51 -07001927{
Dmitry Adamushko3e840502008-07-13 02:10:29 +02001928 switch (phase) {
1929 case CPU_UP_CANCELED:
1930 case CPU_UP_CANCELED_FROZEN:
1931 case CPU_DOWN_FAILED:
1932 case CPU_DOWN_FAILED_FROZEN:
1933 case CPU_ONLINE:
1934 case CPU_ONLINE_FROZEN:
1935 case CPU_DEAD:
1936 case CPU_DEAD_FROZEN:
1937 common_cpu_mem_hotplug_unplug(1);
1938 break;
1939 default:
Avi Kivityac076752007-05-24 12:33:15 +03001940 return NOTIFY_DONE;
Dmitry Adamushko3e840502008-07-13 02:10:29 +02001941 }
Avi Kivityac076752007-05-24 12:33:15 +03001942
Dmitry Adamushko3e840502008-07-13 02:10:29 +02001943 return NOTIFY_OK;
Paul Jackson4c4d50f2006-08-27 01:23:51 -07001944}
Paul Jackson4c4d50f2006-08-27 01:23:51 -07001945
Paul Jacksonb1aac8b2006-09-29 02:01:17 -07001946#ifdef CONFIG_MEMORY_HOTPLUG
Paul Jackson38837fc2006-09-29 02:01:16 -07001947/*
Christoph Lameter0e1e7c72007-10-16 01:25:38 -07001948 * Keep top_cpuset.mems_allowed tracking node_states[N_HIGH_MEMORY].
1949 * Call this routine anytime after you change
1950 * node_states[N_HIGH_MEMORY].
Paul Jackson38837fc2006-09-29 02:01:16 -07001951 * See also the previous routine cpuset_handle_cpuhp().
1952 */
1953
Al Viro1af98922006-10-10 22:48:57 +01001954void cpuset_track_online_nodes(void)
Paul Jackson38837fc2006-09-29 02:01:16 -07001955{
Dmitry Adamushko3e840502008-07-13 02:10:29 +02001956 common_cpu_mem_hotplug_unplug(0);
Paul Jackson38837fc2006-09-29 02:01:16 -07001957}
1958#endif
1959
Linus Torvalds1da177e2005-04-16 15:20:36 -07001960/**
1961 * cpuset_init_smp - initialize cpus_allowed
1962 *
1963 * Description: Finish top cpuset after cpu, node maps are initialized
1964 **/
1965
1966void __init cpuset_init_smp(void)
1967{
1968 top_cpuset.cpus_allowed = cpu_online_map;
Christoph Lameter0e1e7c72007-10-16 01:25:38 -07001969 top_cpuset.mems_allowed = node_states[N_HIGH_MEMORY];
Paul Jackson4c4d50f2006-08-27 01:23:51 -07001970
1971 hotcpu_notifier(cpuset_handle_cpuhp, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001972}
1973
1974/**
Linus Torvalds1da177e2005-04-16 15:20:36 -07001975
Linus Torvalds1da177e2005-04-16 15:20:36 -07001976 * cpuset_cpus_allowed - return cpus_allowed mask from a tasks cpuset.
1977 * @tsk: pointer to task_struct from which to obtain cpuset->cpus_allowed.
Mike Travisf9a86fc2008-04-04 18:11:07 -07001978 * @pmask: pointer to cpumask_t variable to receive cpus_allowed set.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001979 *
1980 * Description: Returns the cpumask_t cpus_allowed of the cpuset
1981 * attached to the specified @tsk. Guaranteed to return some non-empty
1982 * subset of cpu_online_map, even if this means going outside the
1983 * tasks cpuset.
1984 **/
1985
Mike Travisf9a86fc2008-04-04 18:11:07 -07001986void cpuset_cpus_allowed(struct task_struct *tsk, cpumask_t *pmask)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001987{
Ingo Molnar3d3f26a2006-03-23 03:00:18 -08001988 mutex_lock(&callback_mutex);
Mike Travisf9a86fc2008-04-04 18:11:07 -07001989 cpuset_cpus_allowed_locked(tsk, pmask);
Cliff Wickman470fd642007-10-18 23:40:46 -07001990 mutex_unlock(&callback_mutex);
Cliff Wickman470fd642007-10-18 23:40:46 -07001991}
1992
1993/**
1994 * cpuset_cpus_allowed_locked - return cpus_allowed mask from a tasks cpuset.
Paul Menage2df167a2008-02-07 00:14:45 -08001995 * Must be called with callback_mutex held.
Cliff Wickman470fd642007-10-18 23:40:46 -07001996 **/
Mike Travisf9a86fc2008-04-04 18:11:07 -07001997void cpuset_cpus_allowed_locked(struct task_struct *tsk, cpumask_t *pmask)
Cliff Wickman470fd642007-10-18 23:40:46 -07001998{
Paul Jackson909d75a2006-01-08 01:01:55 -08001999 task_lock(tsk);
Mike Travisf9a86fc2008-04-04 18:11:07 -07002000 guarantee_online_cpus(task_cs(tsk), pmask);
Paul Jackson909d75a2006-01-08 01:01:55 -08002001 task_unlock(tsk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002002}
2003
2004void cpuset_init_current_mems_allowed(void)
2005{
Mike Travisf9a86fc2008-04-04 18:11:07 -07002006 nodes_setall(current->mems_allowed);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002007}
2008
Randy Dunlapd9fd8a62005-07-27 11:45:11 -07002009/**
Paul Jackson909d75a2006-01-08 01:01:55 -08002010 * cpuset_mems_allowed - return mems_allowed mask from a tasks cpuset.
2011 * @tsk: pointer to task_struct from which to obtain cpuset->mems_allowed.
2012 *
2013 * Description: Returns the nodemask_t mems_allowed of the cpuset
2014 * attached to the specified @tsk. Guaranteed to return some non-empty
Christoph Lameter0e1e7c72007-10-16 01:25:38 -07002015 * subset of node_states[N_HIGH_MEMORY], even if this means going outside the
Paul Jackson909d75a2006-01-08 01:01:55 -08002016 * tasks cpuset.
2017 **/
2018
2019nodemask_t cpuset_mems_allowed(struct task_struct *tsk)
2020{
2021 nodemask_t mask;
2022
Ingo Molnar3d3f26a2006-03-23 03:00:18 -08002023 mutex_lock(&callback_mutex);
Paul Jackson909d75a2006-01-08 01:01:55 -08002024 task_lock(tsk);
Paul Menage8793d852007-10-18 23:39:39 -07002025 guarantee_online_mems(task_cs(tsk), &mask);
Paul Jackson909d75a2006-01-08 01:01:55 -08002026 task_unlock(tsk);
Ingo Molnar3d3f26a2006-03-23 03:00:18 -08002027 mutex_unlock(&callback_mutex);
Paul Jackson909d75a2006-01-08 01:01:55 -08002028
2029 return mask;
2030}
2031
2032/**
Mel Gorman19770b32008-04-28 02:12:18 -07002033 * cpuset_nodemask_valid_mems_allowed - check nodemask vs. curremt mems_allowed
2034 * @nodemask: the nodemask to be checked
Randy Dunlapd9fd8a62005-07-27 11:45:11 -07002035 *
Mel Gorman19770b32008-04-28 02:12:18 -07002036 * Are any of the nodes in the nodemask allowed in current->mems_allowed?
Linus Torvalds1da177e2005-04-16 15:20:36 -07002037 */
Mel Gorman19770b32008-04-28 02:12:18 -07002038int cpuset_nodemask_valid_mems_allowed(nodemask_t *nodemask)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002039{
Mel Gorman19770b32008-04-28 02:12:18 -07002040 return nodes_intersects(*nodemask, current->mems_allowed);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002041}
2042
Paul Jackson9bf22292005-09-06 15:18:12 -07002043/*
Paul Menage78608362008-04-29 01:00:26 -07002044 * nearest_hardwall_ancestor() - Returns the nearest mem_exclusive or
2045 * mem_hardwall ancestor to the specified cpuset. Call holding
2046 * callback_mutex. If no ancestor is mem_exclusive or mem_hardwall
2047 * (an unusual configuration), then returns the root cpuset.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002048 */
Paul Menage78608362008-04-29 01:00:26 -07002049static const struct cpuset *nearest_hardwall_ancestor(const struct cpuset *cs)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002050{
Paul Menage78608362008-04-29 01:00:26 -07002051 while (!(is_mem_exclusive(cs) || is_mem_hardwall(cs)) && cs->parent)
Paul Jackson9bf22292005-09-06 15:18:12 -07002052 cs = cs->parent;
2053 return cs;
2054}
2055
2056/**
Paul Jackson02a0e532006-12-13 00:34:25 -08002057 * cpuset_zone_allowed_softwall - Can we allocate on zone z's memory node?
Paul Jackson9bf22292005-09-06 15:18:12 -07002058 * @z: is this zone on an allowed node?
Paul Jackson02a0e532006-12-13 00:34:25 -08002059 * @gfp_mask: memory allocation flags
Paul Jackson9bf22292005-09-06 15:18:12 -07002060 *
Paul Jackson02a0e532006-12-13 00:34:25 -08002061 * If we're in interrupt, yes, we can always allocate. If
2062 * __GFP_THISNODE is set, yes, we can always allocate. If zone
Paul Jackson9bf22292005-09-06 15:18:12 -07002063 * z's node is in our tasks mems_allowed, yes. If it's not a
2064 * __GFP_HARDWALL request and this zone's nodes is in the nearest
Paul Menage78608362008-04-29 01:00:26 -07002065 * hardwalled cpuset ancestor to this tasks cpuset, yes.
David Rientjesc596d9f2007-05-06 14:49:32 -07002066 * If the task has been OOM killed and has access to memory reserves
2067 * as specified by the TIF_MEMDIE flag, yes.
Paul Jackson9bf22292005-09-06 15:18:12 -07002068 * Otherwise, no.
2069 *
Paul Jackson02a0e532006-12-13 00:34:25 -08002070 * If __GFP_HARDWALL is set, cpuset_zone_allowed_softwall()
2071 * reduces to cpuset_zone_allowed_hardwall(). Otherwise,
2072 * cpuset_zone_allowed_softwall() might sleep, and might allow a zone
2073 * from an enclosing cpuset.
2074 *
2075 * cpuset_zone_allowed_hardwall() only handles the simpler case of
2076 * hardwall cpusets, and never sleeps.
2077 *
2078 * The __GFP_THISNODE placement logic is really handled elsewhere,
2079 * by forcibly using a zonelist starting at a specified node, and by
2080 * (in get_page_from_freelist()) refusing to consider the zones for
2081 * any node on the zonelist except the first. By the time any such
2082 * calls get to this routine, we should just shut up and say 'yes'.
2083 *
Paul Jackson9bf22292005-09-06 15:18:12 -07002084 * GFP_USER allocations are marked with the __GFP_HARDWALL bit,
David Rientjesc596d9f2007-05-06 14:49:32 -07002085 * and do not allow allocations outside the current tasks cpuset
2086 * unless the task has been OOM killed as is marked TIF_MEMDIE.
Paul Jackson9bf22292005-09-06 15:18:12 -07002087 * GFP_KERNEL allocations are not so marked, so can escape to the
Paul Menage78608362008-04-29 01:00:26 -07002088 * nearest enclosing hardwalled ancestor cpuset.
Paul Jackson9bf22292005-09-06 15:18:12 -07002089 *
Paul Jackson02a0e532006-12-13 00:34:25 -08002090 * Scanning up parent cpusets requires callback_mutex. The
2091 * __alloc_pages() routine only calls here with __GFP_HARDWALL bit
2092 * _not_ set if it's a GFP_KERNEL allocation, and all nodes in the
2093 * current tasks mems_allowed came up empty on the first pass over
2094 * the zonelist. So only GFP_KERNEL allocations, if all nodes in the
2095 * cpuset are short of memory, might require taking the callback_mutex
2096 * mutex.
Paul Jackson9bf22292005-09-06 15:18:12 -07002097 *
Paul Jackson36be57f2006-05-20 15:00:10 -07002098 * The first call here from mm/page_alloc:get_page_from_freelist()
Paul Jackson02a0e532006-12-13 00:34:25 -08002099 * has __GFP_HARDWALL set in gfp_mask, enforcing hardwall cpusets,
2100 * so no allocation on a node outside the cpuset is allowed (unless
2101 * in interrupt, of course).
Paul Jackson9bf22292005-09-06 15:18:12 -07002102 *
Paul Jackson36be57f2006-05-20 15:00:10 -07002103 * The second pass through get_page_from_freelist() doesn't even call
2104 * here for GFP_ATOMIC calls. For those calls, the __alloc_pages()
2105 * variable 'wait' is not set, and the bit ALLOC_CPUSET is not set
2106 * in alloc_flags. That logic and the checks below have the combined
2107 * affect that:
Paul Jackson9bf22292005-09-06 15:18:12 -07002108 * in_interrupt - any node ok (current task context irrelevant)
2109 * GFP_ATOMIC - any node ok
David Rientjesc596d9f2007-05-06 14:49:32 -07002110 * TIF_MEMDIE - any node ok
Paul Menage78608362008-04-29 01:00:26 -07002111 * GFP_KERNEL - any node in enclosing hardwalled cpuset ok
Paul Jackson9bf22292005-09-06 15:18:12 -07002112 * GFP_USER - only nodes in current tasks mems allowed ok.
Paul Jackson36be57f2006-05-20 15:00:10 -07002113 *
2114 * Rule:
Paul Jackson02a0e532006-12-13 00:34:25 -08002115 * Don't call cpuset_zone_allowed_softwall if you can't sleep, unless you
Paul Jackson36be57f2006-05-20 15:00:10 -07002116 * pass in the __GFP_HARDWALL flag set in gfp_flag, which disables
2117 * the code that might scan up ancestor cpusets and sleep.
Paul Jackson02a0e532006-12-13 00:34:25 -08002118 */
Paul Jackson9bf22292005-09-06 15:18:12 -07002119
Paul Jackson02a0e532006-12-13 00:34:25 -08002120int __cpuset_zone_allowed_softwall(struct zone *z, gfp_t gfp_mask)
Paul Jackson9bf22292005-09-06 15:18:12 -07002121{
2122 int node; /* node that zone z is on */
2123 const struct cpuset *cs; /* current cpuset ancestors */
Paul Jackson29afd492006-03-24 03:16:12 -08002124 int allowed; /* is allocation in zone z allowed? */
Paul Jackson9bf22292005-09-06 15:18:12 -07002125
Christoph Lameter9b819d22006-09-25 23:31:40 -07002126 if (in_interrupt() || (gfp_mask & __GFP_THISNODE))
Paul Jackson9bf22292005-09-06 15:18:12 -07002127 return 1;
Christoph Lameter89fa3022006-09-25 23:31:55 -07002128 node = zone_to_nid(z);
Paul Jackson92d1dbd2006-05-20 15:00:11 -07002129 might_sleep_if(!(gfp_mask & __GFP_HARDWALL));
Paul Jackson9bf22292005-09-06 15:18:12 -07002130 if (node_isset(node, current->mems_allowed))
2131 return 1;
David Rientjesc596d9f2007-05-06 14:49:32 -07002132 /*
2133 * Allow tasks that have access to memory reserves because they have
2134 * been OOM killed to get memory anywhere.
2135 */
2136 if (unlikely(test_thread_flag(TIF_MEMDIE)))
2137 return 1;
Paul Jackson9bf22292005-09-06 15:18:12 -07002138 if (gfp_mask & __GFP_HARDWALL) /* If hardwall request, stop here */
2139 return 0;
2140
Bob Picco5563e772005-11-13 16:06:35 -08002141 if (current->flags & PF_EXITING) /* Let dying task have memory */
2142 return 1;
2143
Paul Jackson9bf22292005-09-06 15:18:12 -07002144 /* Not hardwall and node outside mems_allowed: scan up cpusets */
Ingo Molnar3d3f26a2006-03-23 03:00:18 -08002145 mutex_lock(&callback_mutex);
Paul Jackson053199e2005-10-30 15:02:30 -08002146
Paul Jackson053199e2005-10-30 15:02:30 -08002147 task_lock(current);
Paul Menage78608362008-04-29 01:00:26 -07002148 cs = nearest_hardwall_ancestor(task_cs(current));
Paul Jackson053199e2005-10-30 15:02:30 -08002149 task_unlock(current);
2150
Paul Jackson9bf22292005-09-06 15:18:12 -07002151 allowed = node_isset(node, cs->mems_allowed);
Ingo Molnar3d3f26a2006-03-23 03:00:18 -08002152 mutex_unlock(&callback_mutex);
Paul Jackson9bf22292005-09-06 15:18:12 -07002153 return allowed;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002154}
2155
Paul Jackson02a0e532006-12-13 00:34:25 -08002156/*
2157 * cpuset_zone_allowed_hardwall - Can we allocate on zone z's memory node?
2158 * @z: is this zone on an allowed node?
2159 * @gfp_mask: memory allocation flags
2160 *
2161 * If we're in interrupt, yes, we can always allocate.
2162 * If __GFP_THISNODE is set, yes, we can always allocate. If zone
David Rientjesc596d9f2007-05-06 14:49:32 -07002163 * z's node is in our tasks mems_allowed, yes. If the task has been
2164 * OOM killed and has access to memory reserves as specified by the
2165 * TIF_MEMDIE flag, yes. Otherwise, no.
Paul Jackson02a0e532006-12-13 00:34:25 -08002166 *
2167 * The __GFP_THISNODE placement logic is really handled elsewhere,
2168 * by forcibly using a zonelist starting at a specified node, and by
2169 * (in get_page_from_freelist()) refusing to consider the zones for
2170 * any node on the zonelist except the first. By the time any such
2171 * calls get to this routine, we should just shut up and say 'yes'.
2172 *
2173 * Unlike the cpuset_zone_allowed_softwall() variant, above,
2174 * this variant requires that the zone be in the current tasks
2175 * mems_allowed or that we're in interrupt. It does not scan up the
2176 * cpuset hierarchy for the nearest enclosing mem_exclusive cpuset.
2177 * It never sleeps.
2178 */
2179
2180int __cpuset_zone_allowed_hardwall(struct zone *z, gfp_t gfp_mask)
2181{
2182 int node; /* node that zone z is on */
2183
2184 if (in_interrupt() || (gfp_mask & __GFP_THISNODE))
2185 return 1;
2186 node = zone_to_nid(z);
2187 if (node_isset(node, current->mems_allowed))
2188 return 1;
Daniel Walkerdedf8b72007-10-18 03:06:04 -07002189 /*
2190 * Allow tasks that have access to memory reserves because they have
2191 * been OOM killed to get memory anywhere.
2192 */
2193 if (unlikely(test_thread_flag(TIF_MEMDIE)))
2194 return 1;
Paul Jackson02a0e532006-12-13 00:34:25 -08002195 return 0;
2196}
2197
Paul Jacksonef08e3b2005-09-06 15:18:13 -07002198/**
Paul Jackson505970b2006-01-14 13:21:06 -08002199 * cpuset_lock - lock out any changes to cpuset structures
2200 *
Ingo Molnar3d3f26a2006-03-23 03:00:18 -08002201 * The out of memory (oom) code needs to mutex_lock cpusets
Paul Jackson505970b2006-01-14 13:21:06 -08002202 * from being changed while it scans the tasklist looking for a
Ingo Molnar3d3f26a2006-03-23 03:00:18 -08002203 * task in an overlapping cpuset. Expose callback_mutex via this
Paul Jackson505970b2006-01-14 13:21:06 -08002204 * cpuset_lock() routine, so the oom code can lock it, before
2205 * locking the task list. The tasklist_lock is a spinlock, so
Ingo Molnar3d3f26a2006-03-23 03:00:18 -08002206 * must be taken inside callback_mutex.
Paul Jackson505970b2006-01-14 13:21:06 -08002207 */
2208
2209void cpuset_lock(void)
2210{
Ingo Molnar3d3f26a2006-03-23 03:00:18 -08002211 mutex_lock(&callback_mutex);
Paul Jackson505970b2006-01-14 13:21:06 -08002212}
2213
2214/**
2215 * cpuset_unlock - release lock on cpuset changes
2216 *
2217 * Undo the lock taken in a previous cpuset_lock() call.
2218 */
2219
2220void cpuset_unlock(void)
2221{
Ingo Molnar3d3f26a2006-03-23 03:00:18 -08002222 mutex_unlock(&callback_mutex);
Paul Jackson505970b2006-01-14 13:21:06 -08002223}
2224
2225/**
Paul Jackson825a46a2006-03-24 03:16:03 -08002226 * cpuset_mem_spread_node() - On which node to begin search for a page
2227 *
2228 * If a task is marked PF_SPREAD_PAGE or PF_SPREAD_SLAB (as for
2229 * tasks in a cpuset with is_spread_page or is_spread_slab set),
2230 * and if the memory allocation used cpuset_mem_spread_node()
2231 * to determine on which node to start looking, as it will for
2232 * certain page cache or slab cache pages such as used for file
2233 * system buffers and inode caches, then instead of starting on the
2234 * local node to look for a free page, rather spread the starting
2235 * node around the tasks mems_allowed nodes.
2236 *
2237 * We don't have to worry about the returned node being offline
2238 * because "it can't happen", and even if it did, it would be ok.
2239 *
2240 * The routines calling guarantee_online_mems() are careful to
2241 * only set nodes in task->mems_allowed that are online. So it
2242 * should not be possible for the following code to return an
2243 * offline node. But if it did, that would be ok, as this routine
2244 * is not returning the node where the allocation must be, only
2245 * the node where the search should start. The zonelist passed to
2246 * __alloc_pages() will include all nodes. If the slab allocator
2247 * is passed an offline node, it will fall back to the local node.
2248 * See kmem_cache_alloc_node().
2249 */
2250
2251int cpuset_mem_spread_node(void)
2252{
2253 int node;
2254
2255 node = next_node(current->cpuset_mem_spread_rotor, current->mems_allowed);
2256 if (node == MAX_NUMNODES)
2257 node = first_node(current->mems_allowed);
2258 current->cpuset_mem_spread_rotor = node;
2259 return node;
2260}
2261EXPORT_SYMBOL_GPL(cpuset_mem_spread_node);
2262
2263/**
David Rientjesbbe373f2007-10-16 23:25:58 -07002264 * cpuset_mems_allowed_intersects - Does @tsk1's mems_allowed intersect @tsk2's?
2265 * @tsk1: pointer to task_struct of some task.
2266 * @tsk2: pointer to task_struct of some other task.
Paul Jacksonef08e3b2005-09-06 15:18:13 -07002267 *
David Rientjesbbe373f2007-10-16 23:25:58 -07002268 * Description: Return true if @tsk1's mems_allowed intersects the
2269 * mems_allowed of @tsk2. Used by the OOM killer to determine if
2270 * one of the task's memory usage might impact the memory available
2271 * to the other.
Paul Jacksonef08e3b2005-09-06 15:18:13 -07002272 **/
2273
David Rientjesbbe373f2007-10-16 23:25:58 -07002274int cpuset_mems_allowed_intersects(const struct task_struct *tsk1,
2275 const struct task_struct *tsk2)
Paul Jacksonef08e3b2005-09-06 15:18:13 -07002276{
David Rientjesbbe373f2007-10-16 23:25:58 -07002277 return nodes_intersects(tsk1->mems_allowed, tsk2->mems_allowed);
Paul Jacksonef08e3b2005-09-06 15:18:13 -07002278}
2279
Linus Torvalds1da177e2005-04-16 15:20:36 -07002280/*
Paul Jackson3e0d98b2006-01-08 01:01:49 -08002281 * Collection of memory_pressure is suppressed unless
2282 * this flag is enabled by writing "1" to the special
2283 * cpuset file 'memory_pressure_enabled' in the root cpuset.
2284 */
2285
Paul Jacksonc5b2aff2006-01-08 01:01:51 -08002286int cpuset_memory_pressure_enabled __read_mostly;
Paul Jackson3e0d98b2006-01-08 01:01:49 -08002287
2288/**
2289 * cpuset_memory_pressure_bump - keep stats of per-cpuset reclaims.
2290 *
2291 * Keep a running average of the rate of synchronous (direct)
2292 * page reclaim efforts initiated by tasks in each cpuset.
2293 *
2294 * This represents the rate at which some task in the cpuset
2295 * ran low on memory on all nodes it was allowed to use, and
2296 * had to enter the kernels page reclaim code in an effort to
2297 * create more free memory by tossing clean pages or swapping
2298 * or writing dirty pages.
2299 *
2300 * Display to user space in the per-cpuset read-only file
2301 * "memory_pressure". Value displayed is an integer
2302 * representing the recent rate of entry into the synchronous
2303 * (direct) page reclaim by any task attached to the cpuset.
2304 **/
2305
2306void __cpuset_memory_pressure_bump(void)
2307{
Paul Jackson3e0d98b2006-01-08 01:01:49 -08002308 task_lock(current);
Paul Menage8793d852007-10-18 23:39:39 -07002309 fmeter_markevent(&task_cs(current)->fmeter);
Paul Jackson3e0d98b2006-01-08 01:01:49 -08002310 task_unlock(current);
2311}
2312
Paul Menage8793d852007-10-18 23:39:39 -07002313#ifdef CONFIG_PROC_PID_CPUSET
Paul Jackson3e0d98b2006-01-08 01:01:49 -08002314/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07002315 * proc_cpuset_show()
2316 * - Print tasks cpuset path into seq_file.
2317 * - Used for /proc/<pid>/cpuset.
Paul Jackson053199e2005-10-30 15:02:30 -08002318 * - No need to task_lock(tsk) on this tsk->cpuset reference, as it
2319 * doesn't really matter if tsk->cpuset changes after we read it,
Paul Jacksonc8d9c902008-02-07 00:14:46 -08002320 * and we take cgroup_mutex, keeping cpuset_attach() from changing it
Paul Menage2df167a2008-02-07 00:14:45 -08002321 * anyway.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002322 */
Paul Jackson029190c2007-10-18 23:40:20 -07002323static int proc_cpuset_show(struct seq_file *m, void *unused_v)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002324{
Eric W. Biederman13b41b02006-06-26 00:25:56 -07002325 struct pid *pid;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002326 struct task_struct *tsk;
2327 char *buf;
Paul Menage8793d852007-10-18 23:39:39 -07002328 struct cgroup_subsys_state *css;
Eric W. Biederman99f89552006-06-26 00:25:55 -07002329 int retval;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002330
Eric W. Biederman99f89552006-06-26 00:25:55 -07002331 retval = -ENOMEM;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002332 buf = kmalloc(PAGE_SIZE, GFP_KERNEL);
2333 if (!buf)
Eric W. Biederman99f89552006-06-26 00:25:55 -07002334 goto out;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002335
Eric W. Biederman99f89552006-06-26 00:25:55 -07002336 retval = -ESRCH;
Eric W. Biederman13b41b02006-06-26 00:25:56 -07002337 pid = m->private;
2338 tsk = get_pid_task(pid, PIDTYPE_PID);
Eric W. Biederman99f89552006-06-26 00:25:55 -07002339 if (!tsk)
2340 goto out_free;
2341
2342 retval = -EINVAL;
Paul Menage8793d852007-10-18 23:39:39 -07002343 cgroup_lock();
2344 css = task_subsys_state(tsk, cpuset_subsys_id);
2345 retval = cgroup_path(css->cgroup, buf, PAGE_SIZE);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002346 if (retval < 0)
Eric W. Biederman99f89552006-06-26 00:25:55 -07002347 goto out_unlock;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002348 seq_puts(m, buf);
2349 seq_putc(m, '\n');
Eric W. Biederman99f89552006-06-26 00:25:55 -07002350out_unlock:
Paul Menage8793d852007-10-18 23:39:39 -07002351 cgroup_unlock();
Eric W. Biederman99f89552006-06-26 00:25:55 -07002352 put_task_struct(tsk);
2353out_free:
Linus Torvalds1da177e2005-04-16 15:20:36 -07002354 kfree(buf);
Eric W. Biederman99f89552006-06-26 00:25:55 -07002355out:
Linus Torvalds1da177e2005-04-16 15:20:36 -07002356 return retval;
2357}
2358
2359static int cpuset_open(struct inode *inode, struct file *file)
2360{
Eric W. Biederman13b41b02006-06-26 00:25:56 -07002361 struct pid *pid = PROC_I(inode)->pid;
2362 return single_open(file, proc_cpuset_show, pid);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002363}
2364
Arjan van de Ven9a321442007-02-12 00:55:35 -08002365const struct file_operations proc_cpuset_operations = {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002366 .open = cpuset_open,
2367 .read = seq_read,
2368 .llseek = seq_lseek,
2369 .release = single_release,
2370};
Paul Menage8793d852007-10-18 23:39:39 -07002371#endif /* CONFIG_PROC_PID_CPUSET */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002372
2373/* Display task cpus_allowed, mems_allowed in /proc/<pid>/status file. */
Eric W. Biedermandf5f8312008-02-08 04:18:33 -08002374void cpuset_task_status_allowed(struct seq_file *m, struct task_struct *task)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002375{
Eric W. Biedermandf5f8312008-02-08 04:18:33 -08002376 seq_printf(m, "Cpus_allowed:\t");
2377 m->count += cpumask_scnprintf(m->buf + m->count, m->size - m->count,
2378 task->cpus_allowed);
2379 seq_printf(m, "\n");
Mike Travis39106dc2008-04-08 11:43:03 -07002380 seq_printf(m, "Cpus_allowed_list:\t");
2381 m->count += cpulist_scnprintf(m->buf + m->count, m->size - m->count,
2382 task->cpus_allowed);
2383 seq_printf(m, "\n");
Eric W. Biedermandf5f8312008-02-08 04:18:33 -08002384 seq_printf(m, "Mems_allowed:\t");
2385 m->count += nodemask_scnprintf(m->buf + m->count, m->size - m->count,
2386 task->mems_allowed);
2387 seq_printf(m, "\n");
Mike Travis39106dc2008-04-08 11:43:03 -07002388 seq_printf(m, "Mems_allowed_list:\t");
2389 m->count += nodelist_scnprintf(m->buf + m->count, m->size - m->count,
2390 task->mems_allowed);
2391 seq_printf(m, "\n");
Linus Torvalds1da177e2005-04-16 15:20:36 -07002392}