blob: f76db9dcaa05e592f3138f5389ae4c22827b58d9 [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
Max Krasnyanskycf417142008-08-11 14:33:53 -070017 * 2008 Rework of the scheduler domains and CPU hotplug handling
18 * by Max Krasnyansky
Linus Torvalds1da177e2005-04-16 15:20:36 -070019 *
20 * This file is subject to the terms and conditions of the GNU General Public
21 * License. See the file COPYING in the main directory of the Linux
22 * distribution for more details.
23 */
24
Linus Torvalds1da177e2005-04-16 15:20:36 -070025#include <linux/cpu.h>
26#include <linux/cpumask.h>
27#include <linux/cpuset.h>
28#include <linux/err.h>
29#include <linux/errno.h>
30#include <linux/file.h>
31#include <linux/fs.h>
32#include <linux/init.h>
33#include <linux/interrupt.h>
34#include <linux/kernel.h>
35#include <linux/kmod.h>
36#include <linux/list.h>
Paul Jackson68860ec2005-10-30 15:02:36 -080037#include <linux/mempolicy.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070038#include <linux/mm.h>
Miao Xief4818912008-11-19 15:36:30 -080039#include <linux/memory.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070040#include <linux/module.h>
41#include <linux/mount.h>
42#include <linux/namei.h>
43#include <linux/pagemap.h>
44#include <linux/proc_fs.h>
Paul Jackson6b9c2602006-01-08 01:02:02 -080045#include <linux/rcupdate.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070046#include <linux/sched.h>
47#include <linux/seq_file.h>
David Quigley22fb52d2006-06-23 02:04:00 -070048#include <linux/security.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070049#include <linux/slab.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070050#include <linux/spinlock.h>
51#include <linux/stat.h>
52#include <linux/string.h>
53#include <linux/time.h>
54#include <linux/backing-dev.h>
55#include <linux/sort.h>
56
57#include <asm/uaccess.h>
58#include <asm/atomic.h>
Ingo Molnar3d3f26a2006-03-23 03:00:18 -080059#include <linux/mutex.h>
Cliff Wickman956db3c2008-02-07 00:14:43 -080060#include <linux/workqueue.h>
61#include <linux/cgroup.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070062
Paul Jackson202f72d2006-01-08 01:01:57 -080063/*
Miao Xief90d4112009-01-16 10:24:10 +080064 * Workqueue for cpuset related tasks.
65 *
66 * Using kevent workqueue may cause deadlock when memory_migrate
67 * is set. So we create a separate workqueue thread for cpuset.
68 */
69static struct workqueue_struct *cpuset_wq;
70
71/*
Paul Jackson202f72d2006-01-08 01:01:57 -080072 * Tracks how many cpusets are currently defined in system.
73 * When there is only one cpuset (the root cpuset) we can
74 * short circuit some hooks.
75 */
Paul Jackson7edc5962006-01-08 01:02:03 -080076int number_of_cpusets __read_mostly;
Paul Jackson202f72d2006-01-08 01:01:57 -080077
Paul Menage2df167a2008-02-07 00:14:45 -080078/* Forward declare cgroup structures */
Paul Menage8793d852007-10-18 23:39:39 -070079struct cgroup_subsys cpuset_subsys;
80struct cpuset;
81
Paul Jackson3e0d98b2006-01-08 01:01:49 -080082/* See "Frequency meter" comments, below. */
83
84struct fmeter {
85 int cnt; /* unprocessed events count */
86 int val; /* most recent output value */
87 time_t time; /* clock (secs) when val computed */
88 spinlock_t lock; /* guards read or write of above */
89};
90
Linus Torvalds1da177e2005-04-16 15:20:36 -070091struct cpuset {
Paul Menage8793d852007-10-18 23:39:39 -070092 struct cgroup_subsys_state css;
93
Linus Torvalds1da177e2005-04-16 15:20:36 -070094 unsigned long flags; /* "unsigned long" so bitops work */
Li Zefan300ed6c2009-01-07 18:08:44 -080095 cpumask_var_t cpus_allowed; /* CPUs allowed to tasks in cpuset */
Linus Torvalds1da177e2005-04-16 15:20:36 -070096 nodemask_t mems_allowed; /* Memory Nodes allowed to tasks */
97
Linus Torvalds1da177e2005-04-16 15:20:36 -070098 struct cpuset *parent; /* my parent */
Linus Torvalds1da177e2005-04-16 15:20:36 -070099
100 /*
101 * Copy of global cpuset_mems_generation as of the most
102 * recent time this cpuset changed its mems_allowed.
103 */
Paul Jackson3e0d98b2006-01-08 01:01:49 -0800104 int mems_generation;
105
106 struct fmeter fmeter; /* memory_pressure filter */
Paul Jackson029190c2007-10-18 23:40:20 -0700107
108 /* partition number for rebuild_sched_domains() */
109 int pn;
Cliff Wickman956db3c2008-02-07 00:14:43 -0800110
Hidetoshi Seto1d3504f2008-04-15 14:04:23 +0900111 /* for custom sched domain */
112 int relax_domain_level;
113
Cliff Wickman956db3c2008-02-07 00:14:43 -0800114 /* used for walking a cpuset heirarchy */
115 struct list_head stack_list;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700116};
117
Paul Menage8793d852007-10-18 23:39:39 -0700118/* Retrieve the cpuset for a cgroup */
119static inline struct cpuset *cgroup_cs(struct cgroup *cont)
120{
121 return container_of(cgroup_subsys_state(cont, cpuset_subsys_id),
122 struct cpuset, css);
123}
124
125/* Retrieve the cpuset for a task */
126static inline struct cpuset *task_cs(struct task_struct *task)
127{
128 return container_of(task_subsys_state(task, cpuset_subsys_id),
129 struct cpuset, css);
130}
Cliff Wickman956db3c2008-02-07 00:14:43 -0800131struct cpuset_hotplug_scanner {
132 struct cgroup_scanner scan;
133 struct cgroup *to;
134};
Paul Menage8793d852007-10-18 23:39:39 -0700135
Linus Torvalds1da177e2005-04-16 15:20:36 -0700136/* bits in struct cpuset flags field */
137typedef enum {
138 CS_CPU_EXCLUSIVE,
139 CS_MEM_EXCLUSIVE,
Paul Menage78608362008-04-29 01:00:26 -0700140 CS_MEM_HARDWALL,
Paul Jackson45b07ef2006-01-08 01:00:56 -0800141 CS_MEMORY_MIGRATE,
Paul Jackson029190c2007-10-18 23:40:20 -0700142 CS_SCHED_LOAD_BALANCE,
Paul Jackson825a46a2006-03-24 03:16:03 -0800143 CS_SPREAD_PAGE,
144 CS_SPREAD_SLAB,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700145} cpuset_flagbits_t;
146
147/* convenient tests for these bits */
148static inline int is_cpu_exclusive(const struct cpuset *cs)
149{
Paul Jackson7b5b9ef2006-03-24 03:16:00 -0800150 return test_bit(CS_CPU_EXCLUSIVE, &cs->flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700151}
152
153static inline int is_mem_exclusive(const struct cpuset *cs)
154{
Paul Jackson7b5b9ef2006-03-24 03:16:00 -0800155 return test_bit(CS_MEM_EXCLUSIVE, &cs->flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700156}
157
Paul Menage78608362008-04-29 01:00:26 -0700158static inline int is_mem_hardwall(const struct cpuset *cs)
159{
160 return test_bit(CS_MEM_HARDWALL, &cs->flags);
161}
162
Paul Jackson029190c2007-10-18 23:40:20 -0700163static inline int is_sched_load_balance(const struct cpuset *cs)
164{
165 return test_bit(CS_SCHED_LOAD_BALANCE, &cs->flags);
166}
167
Paul Jackson45b07ef2006-01-08 01:00:56 -0800168static inline int is_memory_migrate(const struct cpuset *cs)
169{
Paul Jackson7b5b9ef2006-03-24 03:16:00 -0800170 return test_bit(CS_MEMORY_MIGRATE, &cs->flags);
Paul Jackson45b07ef2006-01-08 01:00:56 -0800171}
172
Paul Jackson825a46a2006-03-24 03:16:03 -0800173static inline int is_spread_page(const struct cpuset *cs)
174{
175 return test_bit(CS_SPREAD_PAGE, &cs->flags);
176}
177
178static inline int is_spread_slab(const struct cpuset *cs)
179{
180 return test_bit(CS_SPREAD_SLAB, &cs->flags);
181}
182
Linus Torvalds1da177e2005-04-16 15:20:36 -0700183/*
Paul Jackson151a4422006-03-24 03:16:11 -0800184 * Increment this integer everytime any cpuset changes its
Linus Torvalds1da177e2005-04-16 15:20:36 -0700185 * mems_allowed value. Users of cpusets can track this generation
186 * number, and avoid having to lock and reload mems_allowed unless
187 * the cpuset they're using changes generation.
188 *
Paul Menage2df167a2008-02-07 00:14:45 -0800189 * A single, global generation is needed because cpuset_attach_task() could
Linus Torvalds1da177e2005-04-16 15:20:36 -0700190 * reattach a task to a different cpuset, which must not have its
191 * generation numbers aliased with those of that tasks previous cpuset.
192 *
193 * Generations are needed for mems_allowed because one task cannot
Paul Menage2df167a2008-02-07 00:14:45 -0800194 * modify another's memory placement. So we must enable every task,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700195 * on every visit to __alloc_pages(), to efficiently check whether
196 * its current->cpuset->mems_allowed has changed, requiring an update
197 * of its current->mems_allowed.
Paul Jackson151a4422006-03-24 03:16:11 -0800198 *
Paul Menage2df167a2008-02-07 00:14:45 -0800199 * Since writes to cpuset_mems_generation are guarded by the cgroup lock
Paul Jackson151a4422006-03-24 03:16:11 -0800200 * there is no need to mark it atomic.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700201 */
Paul Jackson151a4422006-03-24 03:16:11 -0800202static int cpuset_mems_generation;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700203
204static struct cpuset top_cpuset = {
205 .flags = ((1 << CS_CPU_EXCLUSIVE) | (1 << CS_MEM_EXCLUSIVE)),
Linus Torvalds1da177e2005-04-16 15:20:36 -0700206};
207
Linus Torvalds1da177e2005-04-16 15:20:36 -0700208/*
Paul Menage2df167a2008-02-07 00:14:45 -0800209 * There are two global mutexes guarding cpuset structures. The first
210 * is the main control groups cgroup_mutex, accessed via
211 * cgroup_lock()/cgroup_unlock(). The second is the cpuset-specific
212 * callback_mutex, below. They can nest. It is ok to first take
213 * cgroup_mutex, then nest callback_mutex. We also require taking
214 * task_lock() when dereferencing a task's cpuset pointer. See "The
215 * task_lock() exception", at the end of this comment.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700216 *
Ingo Molnar3d3f26a2006-03-23 03:00:18 -0800217 * A task must hold both mutexes to modify cpusets. If a task
Paul Menage2df167a2008-02-07 00:14:45 -0800218 * holds cgroup_mutex, then it blocks others wanting that mutex,
Ingo Molnar3d3f26a2006-03-23 03:00:18 -0800219 * ensuring that it is the only task able to also acquire callback_mutex
Paul Jackson053199e2005-10-30 15:02:30 -0800220 * and be able to modify cpusets. It can perform various checks on
221 * the cpuset structure first, knowing nothing will change. It can
Paul Menage2df167a2008-02-07 00:14:45 -0800222 * also allocate memory while just holding cgroup_mutex. While it is
Paul Jackson053199e2005-10-30 15:02:30 -0800223 * performing these checks, various callback routines can briefly
Ingo Molnar3d3f26a2006-03-23 03:00:18 -0800224 * acquire callback_mutex to query cpusets. Once it is ready to make
225 * the changes, it takes callback_mutex, blocking everyone else.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700226 *
Paul Jackson053199e2005-10-30 15:02:30 -0800227 * Calls to the kernel memory allocator can not be made while holding
Ingo Molnar3d3f26a2006-03-23 03:00:18 -0800228 * callback_mutex, as that would risk double tripping on callback_mutex
Paul Jackson053199e2005-10-30 15:02:30 -0800229 * from one of the callbacks into the cpuset code from within
230 * __alloc_pages().
Linus Torvalds1da177e2005-04-16 15:20:36 -0700231 *
Ingo Molnar3d3f26a2006-03-23 03:00:18 -0800232 * If a task is only holding callback_mutex, then it has read-only
Paul Jackson053199e2005-10-30 15:02:30 -0800233 * access to cpusets.
234 *
235 * The task_struct fields mems_allowed and mems_generation may only
236 * be accessed in the context of that task, so require no locks.
237 *
Ingo Molnar3d3f26a2006-03-23 03:00:18 -0800238 * The cpuset_common_file_read() handlers only hold callback_mutex across
Paul Jackson053199e2005-10-30 15:02:30 -0800239 * small pieces of code, such as when reading out possibly multi-word
240 * cpumasks and nodemasks.
241 *
Paul Menage2df167a2008-02-07 00:14:45 -0800242 * Accessing a task's cpuset should be done in accordance with the
243 * guidelines for accessing subsystem state in kernel/cgroup.c
Linus Torvalds1da177e2005-04-16 15:20:36 -0700244 */
245
Ingo Molnar3d3f26a2006-03-23 03:00:18 -0800246static DEFINE_MUTEX(callback_mutex);
Paul Jackson4247bdc2005-09-10 00:26:06 -0700247
Max Krasnyanskycf417142008-08-11 14:33:53 -0700248/*
David Rientjes75aa1992009-01-06 14:39:01 -0800249 * cpuset_buffer_lock protects both the cpuset_name and cpuset_nodelist
250 * buffers. They are statically allocated to prevent using excess stack
251 * when calling cpuset_print_task_mems_allowed().
252 */
253#define CPUSET_NAME_LEN (128)
254#define CPUSET_NODELIST_LEN (256)
255static char cpuset_name[CPUSET_NAME_LEN];
256static char cpuset_nodelist[CPUSET_NODELIST_LEN];
257static DEFINE_SPINLOCK(cpuset_buffer_lock);
258
259/*
Max Krasnyanskycf417142008-08-11 14:33:53 -0700260 * This is ugly, but preserves the userspace API for existing cpuset
Paul Menage8793d852007-10-18 23:39:39 -0700261 * users. If someone tries to mount the "cpuset" filesystem, we
Max Krasnyanskycf417142008-08-11 14:33:53 -0700262 * silently switch it to mount "cgroup" instead
263 */
David Howells454e2392006-06-23 02:02:57 -0700264static int cpuset_get_sb(struct file_system_type *fs_type,
265 int flags, const char *unused_dev_name,
266 void *data, struct vfsmount *mnt)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700267{
Paul Menage8793d852007-10-18 23:39:39 -0700268 struct file_system_type *cgroup_fs = get_fs_type("cgroup");
269 int ret = -ENODEV;
270 if (cgroup_fs) {
271 char mountopts[] =
272 "cpuset,noprefix,"
273 "release_agent=/sbin/cpuset_release_agent";
274 ret = cgroup_fs->get_sb(cgroup_fs, flags,
275 unused_dev_name, mountopts, mnt);
276 put_filesystem(cgroup_fs);
277 }
278 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700279}
280
281static struct file_system_type cpuset_fs_type = {
282 .name = "cpuset",
283 .get_sb = cpuset_get_sb,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700284};
285
Linus Torvalds1da177e2005-04-16 15:20:36 -0700286/*
Li Zefan300ed6c2009-01-07 18:08:44 -0800287 * Return in pmask the portion of a cpusets's cpus_allowed that
Linus Torvalds1da177e2005-04-16 15:20:36 -0700288 * are online. If none are online, walk up the cpuset hierarchy
289 * until we find one that does have some online cpus. If we get
290 * all the way to the top and still haven't found any online cpus,
291 * return cpu_online_map. Or if passed a NULL cs from an exit'ing
292 * task, return cpu_online_map.
293 *
294 * One way or another, we guarantee to return some non-empty subset
295 * of cpu_online_map.
296 *
Ingo Molnar3d3f26a2006-03-23 03:00:18 -0800297 * Call with callback_mutex held.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700298 */
299
Li Zefan6af866a2009-01-07 18:08:45 -0800300static void guarantee_online_cpus(const struct cpuset *cs,
301 struct cpumask *pmask)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700302{
Li Zefan300ed6c2009-01-07 18:08:44 -0800303 while (cs && !cpumask_intersects(cs->cpus_allowed, cpu_online_mask))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700304 cs = cs->parent;
305 if (cs)
Li Zefan300ed6c2009-01-07 18:08:44 -0800306 cpumask_and(pmask, cs->cpus_allowed, cpu_online_mask);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700307 else
Li Zefan300ed6c2009-01-07 18:08:44 -0800308 cpumask_copy(pmask, cpu_online_mask);
309 BUG_ON(!cpumask_intersects(pmask, cpu_online_mask));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700310}
311
312/*
313 * Return in *pmask the portion of a cpusets's mems_allowed that
Christoph Lameter0e1e7c72007-10-16 01:25:38 -0700314 * are online, with memory. If none are online with memory, walk
315 * up the cpuset hierarchy until we find one that does have some
316 * online mems. If we get all the way to the top and still haven't
317 * found any online mems, return node_states[N_HIGH_MEMORY].
Linus Torvalds1da177e2005-04-16 15:20:36 -0700318 *
319 * One way or another, we guarantee to return some non-empty subset
Christoph Lameter0e1e7c72007-10-16 01:25:38 -0700320 * of node_states[N_HIGH_MEMORY].
Linus Torvalds1da177e2005-04-16 15:20:36 -0700321 *
Ingo Molnar3d3f26a2006-03-23 03:00:18 -0800322 * Call with callback_mutex held.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700323 */
324
325static void guarantee_online_mems(const struct cpuset *cs, nodemask_t *pmask)
326{
Christoph Lameter0e1e7c72007-10-16 01:25:38 -0700327 while (cs && !nodes_intersects(cs->mems_allowed,
328 node_states[N_HIGH_MEMORY]))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700329 cs = cs->parent;
330 if (cs)
Christoph Lameter0e1e7c72007-10-16 01:25:38 -0700331 nodes_and(*pmask, cs->mems_allowed,
332 node_states[N_HIGH_MEMORY]);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700333 else
Christoph Lameter0e1e7c72007-10-16 01:25:38 -0700334 *pmask = node_states[N_HIGH_MEMORY];
335 BUG_ON(!nodes_intersects(*pmask, node_states[N_HIGH_MEMORY]));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700336}
337
Paul Jacksoncf2a473c2006-01-08 01:01:54 -0800338/**
339 * cpuset_update_task_memory_state - update task memory placement
Linus Torvalds1da177e2005-04-16 15:20:36 -0700340 *
Paul Jacksoncf2a473c2006-01-08 01:01:54 -0800341 * If the current tasks cpusets mems_allowed changed behind our
342 * backs, update current->mems_allowed, mems_generation and task NUMA
343 * mempolicy to the new value.
344 *
345 * Task mempolicy is updated by rebinding it relative to the
346 * current->cpuset if a task has its memory placement changed.
347 * Do not call this routine if in_interrupt().
348 *
Paul Jackson4a01c8d2006-03-31 02:30:50 -0800349 * Call without callback_mutex or task_lock() held. May be
Paul Menage2df167a2008-02-07 00:14:45 -0800350 * called with or without cgroup_mutex held. Thanks in part to
351 * 'the_top_cpuset_hack', the task's cpuset pointer will never
David Rientjes41f7f602008-03-04 23:32:38 -0800352 * be NULL. This routine also might acquire callback_mutex during
353 * call.
Paul Jackson5aa15b52005-10-30 15:02:28 -0800354 *
Paul Jackson6b9c2602006-01-08 01:02:02 -0800355 * Reading current->cpuset->mems_generation doesn't need task_lock
356 * to guard the current->cpuset derefence, because it is guarded
Paul Menage2df167a2008-02-07 00:14:45 -0800357 * from concurrent freeing of current->cpuset using RCU.
Paul Jackson6b9c2602006-01-08 01:02:02 -0800358 *
359 * The rcu_dereference() is technically probably not needed,
360 * as I don't actually mind if I see a new cpuset pointer but
361 * an old value of mems_generation. However this really only
362 * matters on alpha systems using cpusets heavily. If I dropped
363 * that rcu_dereference(), it would save them a memory barrier.
364 * For all other arch's, rcu_dereference is a no-op anyway, and for
365 * alpha systems not using cpusets, another planned optimization,
366 * avoiding the rcu critical section for tasks in the root cpuset
367 * which is statically allocated, so can't vanish, will make this
368 * irrelevant. Better to use RCU as intended, than to engage in
369 * some cute trick to save a memory barrier that is impossible to
370 * test, for alpha systems using cpusets heavily, which might not
371 * even exist.
Paul Jackson053199e2005-10-30 15:02:30 -0800372 *
373 * This routine is needed to update the per-task mems_allowed data,
374 * within the tasks context, when it is trying to allocate memory
375 * (in various mm/mempolicy.c routines) and notices that some other
376 * task has been modifying its cpuset.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700377 */
378
Randy Dunlapfe85a992006-02-03 03:04:23 -0800379void cpuset_update_task_memory_state(void)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700380{
Paul Jackson053199e2005-10-30 15:02:30 -0800381 int my_cpusets_mem_gen;
Paul Jacksoncf2a473c2006-01-08 01:01:54 -0800382 struct task_struct *tsk = current;
Paul Jackson6b9c2602006-01-08 01:02:02 -0800383 struct cpuset *cs;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700384
Lai Jiangshan13337712009-01-07 18:08:39 -0800385 rcu_read_lock();
386 my_cpusets_mem_gen = task_cs(tsk)->mems_generation;
387 rcu_read_unlock();
Paul Jacksoncf2a473c2006-01-08 01:01:54 -0800388
389 if (my_cpusets_mem_gen != tsk->cpuset_mems_generation) {
Ingo Molnar3d3f26a2006-03-23 03:00:18 -0800390 mutex_lock(&callback_mutex);
Paul Jacksoncf2a473c2006-01-08 01:01:54 -0800391 task_lock(tsk);
Paul Menage8793d852007-10-18 23:39:39 -0700392 cs = task_cs(tsk); /* Maybe changed when task not locked */
Paul Jacksoncf2a473c2006-01-08 01:01:54 -0800393 guarantee_online_mems(cs, &tsk->mems_allowed);
394 tsk->cpuset_mems_generation = cs->mems_generation;
Paul Jackson825a46a2006-03-24 03:16:03 -0800395 if (is_spread_page(cs))
396 tsk->flags |= PF_SPREAD_PAGE;
397 else
398 tsk->flags &= ~PF_SPREAD_PAGE;
399 if (is_spread_slab(cs))
400 tsk->flags |= PF_SPREAD_SLAB;
401 else
402 tsk->flags &= ~PF_SPREAD_SLAB;
Paul Jacksoncf2a473c2006-01-08 01:01:54 -0800403 task_unlock(tsk);
Ingo Molnar3d3f26a2006-03-23 03:00:18 -0800404 mutex_unlock(&callback_mutex);
Paul Jackson74cb2152006-01-08 01:01:56 -0800405 mpol_rebind_task(tsk, &tsk->mems_allowed);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700406 }
407}
408
409/*
410 * is_cpuset_subset(p, q) - Is cpuset p a subset of cpuset q?
411 *
412 * One cpuset is a subset of another if all its allowed CPUs and
413 * Memory Nodes are a subset of the other, and its exclusive flags
Paul Menage2df167a2008-02-07 00:14:45 -0800414 * are only set if the other's are set. Call holding cgroup_mutex.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700415 */
416
417static int is_cpuset_subset(const struct cpuset *p, const struct cpuset *q)
418{
Li Zefan300ed6c2009-01-07 18:08:44 -0800419 return cpumask_subset(p->cpus_allowed, q->cpus_allowed) &&
Linus Torvalds1da177e2005-04-16 15:20:36 -0700420 nodes_subset(p->mems_allowed, q->mems_allowed) &&
421 is_cpu_exclusive(p) <= is_cpu_exclusive(q) &&
422 is_mem_exclusive(p) <= is_mem_exclusive(q);
423}
424
Li Zefan645fcc92009-01-07 18:08:43 -0800425/**
426 * alloc_trial_cpuset - allocate a trial cpuset
427 * @cs: the cpuset that the trial cpuset duplicates
428 */
429static struct cpuset *alloc_trial_cpuset(const struct cpuset *cs)
430{
Li Zefan300ed6c2009-01-07 18:08:44 -0800431 struct cpuset *trial;
432
433 trial = kmemdup(cs, sizeof(*cs), GFP_KERNEL);
434 if (!trial)
435 return NULL;
436
437 if (!alloc_cpumask_var(&trial->cpus_allowed, GFP_KERNEL)) {
438 kfree(trial);
439 return NULL;
440 }
441 cpumask_copy(trial->cpus_allowed, cs->cpus_allowed);
442
443 return trial;
Li Zefan645fcc92009-01-07 18:08:43 -0800444}
445
446/**
447 * free_trial_cpuset - free the trial cpuset
448 * @trial: the trial cpuset to be freed
449 */
450static void free_trial_cpuset(struct cpuset *trial)
451{
Li Zefan300ed6c2009-01-07 18:08:44 -0800452 free_cpumask_var(trial->cpus_allowed);
Li Zefan645fcc92009-01-07 18:08:43 -0800453 kfree(trial);
454}
455
Linus Torvalds1da177e2005-04-16 15:20:36 -0700456/*
457 * validate_change() - Used to validate that any proposed cpuset change
458 * follows the structural rules for cpusets.
459 *
460 * If we replaced the flag and mask values of the current cpuset
461 * (cur) with those values in the trial cpuset (trial), would
462 * our various subset and exclusive rules still be valid? Presumes
Paul Menage2df167a2008-02-07 00:14:45 -0800463 * cgroup_mutex held.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700464 *
465 * 'cur' is the address of an actual, in-use cpuset. Operations
466 * such as list traversal that depend on the actual address of the
467 * cpuset in the list must use cur below, not trial.
468 *
469 * 'trial' is the address of bulk structure copy of cur, with
470 * perhaps one or more of the fields cpus_allowed, mems_allowed,
471 * or flags changed to new, trial values.
472 *
473 * Return 0 if valid, -errno if not.
474 */
475
476static int validate_change(const struct cpuset *cur, const struct cpuset *trial)
477{
Paul Menage8793d852007-10-18 23:39:39 -0700478 struct cgroup *cont;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700479 struct cpuset *c, *par;
480
481 /* Each of our child cpusets must be a subset of us */
Paul Menage8793d852007-10-18 23:39:39 -0700482 list_for_each_entry(cont, &cur->css.cgroup->children, sibling) {
483 if (!is_cpuset_subset(cgroup_cs(cont), trial))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700484 return -EBUSY;
485 }
486
487 /* Remaining checks don't apply to root cpuset */
Paul Jackson69604062006-12-06 20:36:15 -0800488 if (cur == &top_cpuset)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700489 return 0;
490
Paul Jackson69604062006-12-06 20:36:15 -0800491 par = cur->parent;
492
Linus Torvalds1da177e2005-04-16 15:20:36 -0700493 /* We must be a subset of our parent cpuset */
494 if (!is_cpuset_subset(trial, par))
495 return -EACCES;
496
Paul Menage2df167a2008-02-07 00:14:45 -0800497 /*
498 * If either I or some sibling (!= me) is exclusive, we can't
499 * overlap
500 */
Paul Menage8793d852007-10-18 23:39:39 -0700501 list_for_each_entry(cont, &par->css.cgroup->children, sibling) {
502 c = cgroup_cs(cont);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700503 if ((is_cpu_exclusive(trial) || is_cpu_exclusive(c)) &&
504 c != cur &&
Li Zefan300ed6c2009-01-07 18:08:44 -0800505 cpumask_intersects(trial->cpus_allowed, c->cpus_allowed))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700506 return -EINVAL;
507 if ((is_mem_exclusive(trial) || is_mem_exclusive(c)) &&
508 c != cur &&
509 nodes_intersects(trial->mems_allowed, c->mems_allowed))
510 return -EINVAL;
511 }
512
Paul Jackson020958b2007-10-18 23:40:21 -0700513 /* Cpusets with tasks can't have empty cpus_allowed or mems_allowed */
514 if (cgroup_task_count(cur->css.cgroup)) {
Li Zefan300ed6c2009-01-07 18:08:44 -0800515 if (cpumask_empty(trial->cpus_allowed) ||
Paul Jackson020958b2007-10-18 23:40:21 -0700516 nodes_empty(trial->mems_allowed)) {
517 return -ENOSPC;
518 }
519 }
520
Linus Torvalds1da177e2005-04-16 15:20:36 -0700521 return 0;
522}
523
Dinakar Guniguntala85d7b942005-06-25 14:57:34 -0700524/*
Max Krasnyanskycf417142008-08-11 14:33:53 -0700525 * Helper routine for generate_sched_domains().
Paul Jackson029190c2007-10-18 23:40:20 -0700526 * Do cpusets a, b have overlapping cpus_allowed masks?
527 */
Paul Jackson029190c2007-10-18 23:40:20 -0700528static int cpusets_overlap(struct cpuset *a, struct cpuset *b)
529{
Li Zefan300ed6c2009-01-07 18:08:44 -0800530 return cpumask_intersects(a->cpus_allowed, b->cpus_allowed);
Paul Jackson029190c2007-10-18 23:40:20 -0700531}
532
Hidetoshi Seto1d3504f2008-04-15 14:04:23 +0900533static void
534update_domain_attr(struct sched_domain_attr *dattr, struct cpuset *c)
535{
Hidetoshi Seto1d3504f2008-04-15 14:04:23 +0900536 if (dattr->relax_domain_level < c->relax_domain_level)
537 dattr->relax_domain_level = c->relax_domain_level;
538 return;
539}
540
Lai Jiangshanf5393692008-07-29 22:33:22 -0700541static void
542update_domain_attr_tree(struct sched_domain_attr *dattr, struct cpuset *c)
543{
544 LIST_HEAD(q);
545
546 list_add(&c->stack_list, &q);
547 while (!list_empty(&q)) {
548 struct cpuset *cp;
549 struct cgroup *cont;
550 struct cpuset *child;
551
552 cp = list_first_entry(&q, struct cpuset, stack_list);
553 list_del(q.next);
554
Li Zefan300ed6c2009-01-07 18:08:44 -0800555 if (cpumask_empty(cp->cpus_allowed))
Lai Jiangshanf5393692008-07-29 22:33:22 -0700556 continue;
557
558 if (is_sched_load_balance(cp))
559 update_domain_attr(dattr, cp);
560
561 list_for_each_entry(cont, &cp->css.cgroup->children, sibling) {
562 child = cgroup_cs(cont);
563 list_add_tail(&child->stack_list, &q);
564 }
565 }
566}
567
Paul Jackson029190c2007-10-18 23:40:20 -0700568/*
Max Krasnyanskycf417142008-08-11 14:33:53 -0700569 * generate_sched_domains()
Paul Jackson029190c2007-10-18 23:40:20 -0700570 *
Max Krasnyanskycf417142008-08-11 14:33:53 -0700571 * This function builds a partial partition of the systems CPUs
572 * A 'partial partition' is a set of non-overlapping subsets whose
573 * union is a subset of that set.
574 * The output of this function needs to be passed to kernel/sched.c
575 * partition_sched_domains() routine, which will rebuild the scheduler's
576 * load balancing domains (sched domains) as specified by that partial
577 * partition.
Paul Jackson029190c2007-10-18 23:40:20 -0700578 *
Li Zefan45ce80f2009-01-15 13:50:59 -0800579 * See "What is sched_load_balance" in Documentation/cgroups/cpusets.txt
Paul Jackson029190c2007-10-18 23:40:20 -0700580 * for a background explanation of this.
581 *
582 * Does not return errors, on the theory that the callers of this
583 * routine would rather not worry about failures to rebuild sched
584 * domains when operating in the severe memory shortage situations
585 * that could cause allocation failures below.
586 *
Max Krasnyanskycf417142008-08-11 14:33:53 -0700587 * Must be called with cgroup_lock held.
Paul Jackson029190c2007-10-18 23:40:20 -0700588 *
589 * The three key local variables below are:
Li Zefanaeed6822008-07-29 22:33:24 -0700590 * q - a linked-list queue of cpuset pointers, used to implement a
Paul Jackson029190c2007-10-18 23:40:20 -0700591 * top-down scan of all cpusets. This scan loads a pointer
592 * to each cpuset marked is_sched_load_balance into the
593 * array 'csa'. For our purposes, rebuilding the schedulers
594 * sched domains, we can ignore !is_sched_load_balance cpusets.
595 * csa - (for CpuSet Array) Array of pointers to all the cpusets
596 * that need to be load balanced, for convenient iterative
597 * access by the subsequent code that finds the best partition,
598 * i.e the set of domains (subsets) of CPUs such that the
599 * cpus_allowed of every cpuset marked is_sched_load_balance
600 * is a subset of one of these domains, while there are as
601 * many such domains as possible, each as small as possible.
602 * doms - Conversion of 'csa' to an array of cpumasks, for passing to
603 * the kernel/sched.c routine partition_sched_domains() in a
604 * convenient format, that can be easily compared to the prior
605 * value to determine what partition elements (sched domains)
606 * were changed (added or removed.)
607 *
608 * Finding the best partition (set of domains):
609 * The triple nested loops below over i, j, k scan over the
610 * load balanced cpusets (using the array of cpuset pointers in
611 * csa[]) looking for pairs of cpusets that have overlapping
612 * cpus_allowed, but which don't have the same 'pn' partition
613 * number and gives them in the same partition number. It keeps
614 * looping on the 'restart' label until it can no longer find
615 * any such pairs.
616 *
617 * The union of the cpus_allowed masks from the set of
618 * all cpusets having the same 'pn' value then form the one
619 * element of the partition (one sched domain) to be passed to
620 * partition_sched_domains().
621 */
Li Zefan6af866a2009-01-07 18:08:45 -0800622/* FIXME: see the FIXME in partition_sched_domains() */
623static int generate_sched_domains(struct cpumask **domains,
Max Krasnyanskycf417142008-08-11 14:33:53 -0700624 struct sched_domain_attr **attributes)
Paul Jackson029190c2007-10-18 23:40:20 -0700625{
Max Krasnyanskycf417142008-08-11 14:33:53 -0700626 LIST_HEAD(q); /* queue of cpusets to be scanned */
Paul Jackson029190c2007-10-18 23:40:20 -0700627 struct cpuset *cp; /* scans q */
628 struct cpuset **csa; /* array of all cpuset ptrs */
629 int csn; /* how many cpuset ptrs in csa so far */
630 int i, j, k; /* indices for partition finding loops */
Li Zefan6af866a2009-01-07 18:08:45 -0800631 struct cpumask *doms; /* resulting partition; i.e. sched domains */
Hidetoshi Seto1d3504f2008-04-15 14:04:23 +0900632 struct sched_domain_attr *dattr; /* attributes for custom domains */
Ingo Molnar15837152008-11-25 10:27:49 +0100633 int ndoms = 0; /* number of sched domains in result */
Li Zefan6af866a2009-01-07 18:08:45 -0800634 int nslot; /* next empty doms[] struct cpumask slot */
Paul Jackson029190c2007-10-18 23:40:20 -0700635
Paul Jackson029190c2007-10-18 23:40:20 -0700636 doms = NULL;
Hidetoshi Seto1d3504f2008-04-15 14:04:23 +0900637 dattr = NULL;
Max Krasnyanskycf417142008-08-11 14:33:53 -0700638 csa = NULL;
Paul Jackson029190c2007-10-18 23:40:20 -0700639
640 /* Special case for the 99% of systems with one, full, sched domain */
641 if (is_sched_load_balance(&top_cpuset)) {
Li Zefan6af866a2009-01-07 18:08:45 -0800642 doms = kmalloc(cpumask_size(), GFP_KERNEL);
Paul Jackson029190c2007-10-18 23:40:20 -0700643 if (!doms)
Max Krasnyanskycf417142008-08-11 14:33:53 -0700644 goto done;
645
Hidetoshi Seto1d3504f2008-04-15 14:04:23 +0900646 dattr = kmalloc(sizeof(struct sched_domain_attr), GFP_KERNEL);
647 if (dattr) {
648 *dattr = SD_ATTR_INIT;
Li Zefan93a65572008-07-29 22:33:23 -0700649 update_domain_attr_tree(dattr, &top_cpuset);
Hidetoshi Seto1d3504f2008-04-15 14:04:23 +0900650 }
Li Zefan300ed6c2009-01-07 18:08:44 -0800651 cpumask_copy(doms, top_cpuset.cpus_allowed);
Max Krasnyanskycf417142008-08-11 14:33:53 -0700652
653 ndoms = 1;
654 goto done;
Paul Jackson029190c2007-10-18 23:40:20 -0700655 }
656
Paul Jackson029190c2007-10-18 23:40:20 -0700657 csa = kmalloc(number_of_cpusets * sizeof(cp), GFP_KERNEL);
658 if (!csa)
659 goto done;
660 csn = 0;
661
Li Zefanaeed6822008-07-29 22:33:24 -0700662 list_add(&top_cpuset.stack_list, &q);
663 while (!list_empty(&q)) {
Paul Jackson029190c2007-10-18 23:40:20 -0700664 struct cgroup *cont;
665 struct cpuset *child; /* scans child cpusets of cp */
Lai Jiangshan489a5392008-07-25 01:47:23 -0700666
Li Zefanaeed6822008-07-29 22:33:24 -0700667 cp = list_first_entry(&q, struct cpuset, stack_list);
668 list_del(q.next);
669
Li Zefan300ed6c2009-01-07 18:08:44 -0800670 if (cpumask_empty(cp->cpus_allowed))
Lai Jiangshan489a5392008-07-25 01:47:23 -0700671 continue;
672
Lai Jiangshanf5393692008-07-29 22:33:22 -0700673 /*
674 * All child cpusets contain a subset of the parent's cpus, so
675 * just skip them, and then we call update_domain_attr_tree()
676 * to calc relax_domain_level of the corresponding sched
677 * domain.
678 */
679 if (is_sched_load_balance(cp)) {
Paul Jackson029190c2007-10-18 23:40:20 -0700680 csa[csn++] = cp;
Lai Jiangshanf5393692008-07-29 22:33:22 -0700681 continue;
682 }
Lai Jiangshan489a5392008-07-25 01:47:23 -0700683
Paul Jackson029190c2007-10-18 23:40:20 -0700684 list_for_each_entry(cont, &cp->css.cgroup->children, sibling) {
685 child = cgroup_cs(cont);
Li Zefanaeed6822008-07-29 22:33:24 -0700686 list_add_tail(&child->stack_list, &q);
Paul Jackson029190c2007-10-18 23:40:20 -0700687 }
688 }
689
690 for (i = 0; i < csn; i++)
691 csa[i]->pn = i;
692 ndoms = csn;
693
694restart:
695 /* Find the best partition (set of sched domains) */
696 for (i = 0; i < csn; i++) {
697 struct cpuset *a = csa[i];
698 int apn = a->pn;
699
700 for (j = 0; j < csn; j++) {
701 struct cpuset *b = csa[j];
702 int bpn = b->pn;
703
704 if (apn != bpn && cpusets_overlap(a, b)) {
705 for (k = 0; k < csn; k++) {
706 struct cpuset *c = csa[k];
707
708 if (c->pn == bpn)
709 c->pn = apn;
710 }
711 ndoms--; /* one less element */
712 goto restart;
713 }
714 }
715 }
716
Max Krasnyanskycf417142008-08-11 14:33:53 -0700717 /*
718 * Now we know how many domains to create.
719 * Convert <csn, csa> to <ndoms, doms> and populate cpu masks.
720 */
Li Zefan6af866a2009-01-07 18:08:45 -0800721 doms = kmalloc(ndoms * cpumask_size(), GFP_KERNEL);
Li Zefan700018e2008-11-18 14:02:03 +0800722 if (!doms)
Max Krasnyanskycf417142008-08-11 14:33:53 -0700723 goto done;
Max Krasnyanskycf417142008-08-11 14:33:53 -0700724
725 /*
726 * The rest of the code, including the scheduler, can deal with
727 * dattr==NULL case. No need to abort if alloc fails.
728 */
Hidetoshi Seto1d3504f2008-04-15 14:04:23 +0900729 dattr = kmalloc(ndoms * sizeof(struct sched_domain_attr), GFP_KERNEL);
Paul Jackson029190c2007-10-18 23:40:20 -0700730
731 for (nslot = 0, i = 0; i < csn; i++) {
732 struct cpuset *a = csa[i];
Li Zefan6af866a2009-01-07 18:08:45 -0800733 struct cpumask *dp;
Paul Jackson029190c2007-10-18 23:40:20 -0700734 int apn = a->pn;
735
Max Krasnyanskycf417142008-08-11 14:33:53 -0700736 if (apn < 0) {
737 /* Skip completed partitions */
738 continue;
Paul Jackson029190c2007-10-18 23:40:20 -0700739 }
Max Krasnyanskycf417142008-08-11 14:33:53 -0700740
741 dp = doms + nslot;
742
743 if (nslot == ndoms) {
744 static int warnings = 10;
745 if (warnings) {
746 printk(KERN_WARNING
747 "rebuild_sched_domains confused:"
748 " nslot %d, ndoms %d, csn %d, i %d,"
749 " apn %d\n",
750 nslot, ndoms, csn, i, apn);
751 warnings--;
752 }
753 continue;
754 }
755
Li Zefan6af866a2009-01-07 18:08:45 -0800756 cpumask_clear(dp);
Max Krasnyanskycf417142008-08-11 14:33:53 -0700757 if (dattr)
758 *(dattr + nslot) = SD_ATTR_INIT;
759 for (j = i; j < csn; j++) {
760 struct cpuset *b = csa[j];
761
762 if (apn == b->pn) {
Li Zefan300ed6c2009-01-07 18:08:44 -0800763 cpumask_or(dp, dp, b->cpus_allowed);
Max Krasnyanskycf417142008-08-11 14:33:53 -0700764 if (dattr)
765 update_domain_attr_tree(dattr + nslot, b);
766
767 /* Done with this partition */
768 b->pn = -1;
769 }
770 }
771 nslot++;
Paul Jackson029190c2007-10-18 23:40:20 -0700772 }
773 BUG_ON(nslot != ndoms);
774
Paul Jackson029190c2007-10-18 23:40:20 -0700775done:
Paul Jackson029190c2007-10-18 23:40:20 -0700776 kfree(csa);
Max Krasnyanskycf417142008-08-11 14:33:53 -0700777
Li Zefan700018e2008-11-18 14:02:03 +0800778 /*
779 * Fallback to the default domain if kmalloc() failed.
780 * See comments in partition_sched_domains().
781 */
782 if (doms == NULL)
783 ndoms = 1;
784
Max Krasnyanskycf417142008-08-11 14:33:53 -0700785 *domains = doms;
786 *attributes = dattr;
787 return ndoms;
788}
789
790/*
791 * Rebuild scheduler domains.
792 *
793 * Call with neither cgroup_mutex held nor within get_online_cpus().
794 * Takes both cgroup_mutex and get_online_cpus().
795 *
796 * Cannot be directly called from cpuset code handling changes
797 * to the cpuset pseudo-filesystem, because it cannot be called
798 * from code that already holds cgroup_mutex.
799 */
800static void do_rebuild_sched_domains(struct work_struct *unused)
801{
802 struct sched_domain_attr *attr;
Li Zefan6af866a2009-01-07 18:08:45 -0800803 struct cpumask *doms;
Max Krasnyanskycf417142008-08-11 14:33:53 -0700804 int ndoms;
805
806 get_online_cpus();
807
808 /* Generate domain masks and attrs */
809 cgroup_lock();
810 ndoms = generate_sched_domains(&doms, &attr);
811 cgroup_unlock();
812
813 /* Have scheduler rebuild the domains */
814 partition_sched_domains(ndoms, doms, attr);
815
816 put_online_cpus();
817}
818
819static DECLARE_WORK(rebuild_sched_domains_work, do_rebuild_sched_domains);
820
821/*
822 * Rebuild scheduler domains, asynchronously via workqueue.
823 *
824 * If the flag 'sched_load_balance' of any cpuset with non-empty
825 * 'cpus' changes, or if the 'cpus' allowed changes in any cpuset
826 * which has that flag enabled, or if any cpuset with a non-empty
827 * 'cpus' is removed, then call this routine to rebuild the
828 * scheduler's dynamic sched domains.
829 *
830 * The rebuild_sched_domains() and partition_sched_domains()
831 * routines must nest cgroup_lock() inside get_online_cpus(),
832 * but such cpuset changes as these must nest that locking the
833 * other way, holding cgroup_lock() for much of the code.
834 *
835 * So in order to avoid an ABBA deadlock, the cpuset code handling
836 * these user changes delegates the actual sched domain rebuilding
837 * to a separate workqueue thread, which ends up processing the
838 * above do_rebuild_sched_domains() function.
839 */
840static void async_rebuild_sched_domains(void)
841{
Miao Xief90d4112009-01-16 10:24:10 +0800842 queue_work(cpuset_wq, &rebuild_sched_domains_work);
Max Krasnyanskycf417142008-08-11 14:33:53 -0700843}
844
845/*
846 * Accomplishes the same scheduler domain rebuild as the above
847 * async_rebuild_sched_domains(), however it directly calls the
848 * rebuild routine synchronously rather than calling it via an
849 * asynchronous work thread.
850 *
851 * This can only be called from code that is not holding
852 * cgroup_mutex (not nested in a cgroup_lock() call.)
853 */
854void rebuild_sched_domains(void)
855{
856 do_rebuild_sched_domains(NULL);
Paul Jackson029190c2007-10-18 23:40:20 -0700857}
858
Cliff Wickman58f47902008-02-07 00:14:44 -0800859/**
860 * cpuset_test_cpumask - test a task's cpus_allowed versus its cpuset's
861 * @tsk: task to test
862 * @scan: struct cgroup_scanner contained in its struct cpuset_hotplug_scanner
863 *
Paul Menage2df167a2008-02-07 00:14:45 -0800864 * Call with cgroup_mutex held. May take callback_mutex during call.
Cliff Wickman58f47902008-02-07 00:14:44 -0800865 * Called for each task in a cgroup by cgroup_scan_tasks().
866 * Return nonzero if this tasks's cpus_allowed mask should be changed (in other
867 * words, if its mask is not equal to its cpuset's mask).
Paul Jackson053199e2005-10-30 15:02:30 -0800868 */
Adrian Bunk9e0c9142008-04-29 01:00:25 -0700869static int cpuset_test_cpumask(struct task_struct *tsk,
870 struct cgroup_scanner *scan)
Cliff Wickman58f47902008-02-07 00:14:44 -0800871{
Li Zefan300ed6c2009-01-07 18:08:44 -0800872 return !cpumask_equal(&tsk->cpus_allowed,
Cliff Wickman58f47902008-02-07 00:14:44 -0800873 (cgroup_cs(scan->cg))->cpus_allowed);
874}
Paul Jackson053199e2005-10-30 15:02:30 -0800875
Cliff Wickman58f47902008-02-07 00:14:44 -0800876/**
877 * cpuset_change_cpumask - make a task's cpus_allowed the same as its cpuset's
878 * @tsk: task to test
879 * @scan: struct cgroup_scanner containing the cgroup of the task
880 *
881 * Called by cgroup_scan_tasks() for each task in a cgroup whose
882 * cpus_allowed mask needs to be changed.
883 *
884 * We don't need to re-check for the cgroup/cpuset membership, since we're
885 * holding cgroup_lock() at this point.
886 */
Adrian Bunk9e0c9142008-04-29 01:00:25 -0700887static void cpuset_change_cpumask(struct task_struct *tsk,
888 struct cgroup_scanner *scan)
Cliff Wickman58f47902008-02-07 00:14:44 -0800889{
Li Zefan300ed6c2009-01-07 18:08:44 -0800890 set_cpus_allowed_ptr(tsk, ((cgroup_cs(scan->cg))->cpus_allowed));
Cliff Wickman58f47902008-02-07 00:14:44 -0800891}
892
893/**
Miao Xie0b2f6302008-07-25 01:47:21 -0700894 * update_tasks_cpumask - Update the cpumasks of tasks in the cpuset.
895 * @cs: the cpuset in which each task's cpus_allowed mask needs to be changed
Li Zefan4e743392008-09-13 02:33:08 -0700896 * @heap: if NULL, defer allocating heap memory to cgroup_scan_tasks()
Miao Xie0b2f6302008-07-25 01:47:21 -0700897 *
898 * Called with cgroup_mutex held
899 *
900 * The cgroup_scan_tasks() function will scan all the tasks in a cgroup,
901 * calling callback functions for each.
902 *
Li Zefan4e743392008-09-13 02:33:08 -0700903 * No return value. It's guaranteed that cgroup_scan_tasks() always returns 0
904 * if @heap != NULL.
Miao Xie0b2f6302008-07-25 01:47:21 -0700905 */
Li Zefan4e743392008-09-13 02:33:08 -0700906static void update_tasks_cpumask(struct cpuset *cs, struct ptr_heap *heap)
Miao Xie0b2f6302008-07-25 01:47:21 -0700907{
908 struct cgroup_scanner scan;
Miao Xie0b2f6302008-07-25 01:47:21 -0700909
910 scan.cg = cs->css.cgroup;
911 scan.test_task = cpuset_test_cpumask;
912 scan.process_task = cpuset_change_cpumask;
Li Zefan4e743392008-09-13 02:33:08 -0700913 scan.heap = heap;
914 cgroup_scan_tasks(&scan);
Miao Xie0b2f6302008-07-25 01:47:21 -0700915}
916
917/**
Cliff Wickman58f47902008-02-07 00:14:44 -0800918 * update_cpumask - update the cpus_allowed mask of a cpuset and all tasks in it
919 * @cs: the cpuset to consider
920 * @buf: buffer of cpu numbers written to this cpuset
921 */
Li Zefan645fcc92009-01-07 18:08:43 -0800922static int update_cpumask(struct cpuset *cs, struct cpuset *trialcs,
923 const char *buf)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700924{
Li Zefan4e743392008-09-13 02:33:08 -0700925 struct ptr_heap heap;
Cliff Wickman58f47902008-02-07 00:14:44 -0800926 int retval;
927 int is_load_balanced;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700928
Paul Jackson4c4d50f2006-08-27 01:23:51 -0700929 /* top_cpuset.cpus_allowed tracks cpu_online_map; it's read-only */
930 if (cs == &top_cpuset)
931 return -EACCES;
932
David Rientjes6f7f02e2007-05-08 00:31:43 -0700933 /*
Paul Jacksonc8d9c902008-02-07 00:14:46 -0800934 * An empty cpus_allowed is ok only if the cpuset has no tasks.
Paul Jackson020958b2007-10-18 23:40:21 -0700935 * Since cpulist_parse() fails on an empty mask, we special case
936 * that parsing. The validate_change() call ensures that cpusets
937 * with tasks have cpus.
David Rientjes6f7f02e2007-05-08 00:31:43 -0700938 */
Paul Jackson020958b2007-10-18 23:40:21 -0700939 if (!*buf) {
Li Zefan300ed6c2009-01-07 18:08:44 -0800940 cpumask_clear(trialcs->cpus_allowed);
David Rientjes6f7f02e2007-05-08 00:31:43 -0700941 } else {
Li Zefan300ed6c2009-01-07 18:08:44 -0800942 retval = cpulist_parse(buf, trialcs->cpus_allowed);
David Rientjes6f7f02e2007-05-08 00:31:43 -0700943 if (retval < 0)
944 return retval;
Lai Jiangshan37340742008-06-05 22:46:32 -0700945
Li Zefan300ed6c2009-01-07 18:08:44 -0800946 if (!cpumask_subset(trialcs->cpus_allowed, cpu_online_mask))
Lai Jiangshan37340742008-06-05 22:46:32 -0700947 return -EINVAL;
David Rientjes6f7f02e2007-05-08 00:31:43 -0700948 }
Li Zefan645fcc92009-01-07 18:08:43 -0800949 retval = validate_change(cs, trialcs);
Dinakar Guniguntala85d7b942005-06-25 14:57:34 -0700950 if (retval < 0)
951 return retval;
Paul Jackson029190c2007-10-18 23:40:20 -0700952
Paul Menage8707d8b2007-10-18 23:40:22 -0700953 /* Nothing to do if the cpus didn't change */
Li Zefan300ed6c2009-01-07 18:08:44 -0800954 if (cpumask_equal(cs->cpus_allowed, trialcs->cpus_allowed))
Paul Menage8707d8b2007-10-18 23:40:22 -0700955 return 0;
Cliff Wickman58f47902008-02-07 00:14:44 -0800956
Li Zefan4e743392008-09-13 02:33:08 -0700957 retval = heap_init(&heap, PAGE_SIZE, GFP_KERNEL, NULL);
958 if (retval)
959 return retval;
960
Li Zefan645fcc92009-01-07 18:08:43 -0800961 is_load_balanced = is_sched_load_balance(trialcs);
Paul Jackson029190c2007-10-18 23:40:20 -0700962
Ingo Molnar3d3f26a2006-03-23 03:00:18 -0800963 mutex_lock(&callback_mutex);
Li Zefan300ed6c2009-01-07 18:08:44 -0800964 cpumask_copy(cs->cpus_allowed, trialcs->cpus_allowed);
Ingo Molnar3d3f26a2006-03-23 03:00:18 -0800965 mutex_unlock(&callback_mutex);
Paul Jackson029190c2007-10-18 23:40:20 -0700966
Paul Menage8707d8b2007-10-18 23:40:22 -0700967 /*
968 * Scan tasks in the cpuset, and update the cpumasks of any
Cliff Wickman58f47902008-02-07 00:14:44 -0800969 * that need an update.
Paul Menage8707d8b2007-10-18 23:40:22 -0700970 */
Li Zefan4e743392008-09-13 02:33:08 -0700971 update_tasks_cpumask(cs, &heap);
972
973 heap_free(&heap);
Cliff Wickman58f47902008-02-07 00:14:44 -0800974
Paul Menage8707d8b2007-10-18 23:40:22 -0700975 if (is_load_balanced)
Max Krasnyanskycf417142008-08-11 14:33:53 -0700976 async_rebuild_sched_domains();
Dinakar Guniguntala85d7b942005-06-25 14:57:34 -0700977 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700978}
979
Paul Jackson053199e2005-10-30 15:02:30 -0800980/*
Paul Jacksone4e364e2006-03-31 02:30:52 -0800981 * cpuset_migrate_mm
982 *
983 * Migrate memory region from one set of nodes to another.
984 *
985 * Temporarilly set tasks mems_allowed to target nodes of migration,
986 * so that the migration code can allocate pages on these nodes.
987 *
Paul Menage2df167a2008-02-07 00:14:45 -0800988 * Call holding cgroup_mutex, so current's cpuset won't change
Paul Jacksonc8d9c902008-02-07 00:14:46 -0800989 * during this call, as manage_mutex holds off any cpuset_attach()
Paul Jacksone4e364e2006-03-31 02:30:52 -0800990 * calls. Therefore we don't need to take task_lock around the
991 * call to guarantee_online_mems(), as we know no one is changing
Paul Menage2df167a2008-02-07 00:14:45 -0800992 * our task's cpuset.
Paul Jacksone4e364e2006-03-31 02:30:52 -0800993 *
994 * Hold callback_mutex around the two modifications of our tasks
995 * mems_allowed to synchronize with cpuset_mems_allowed().
996 *
997 * While the mm_struct we are migrating is typically from some
998 * other task, the task_struct mems_allowed that we are hacking
999 * is for our current task, which must allocate new pages for that
1000 * migrating memory region.
1001 *
1002 * We call cpuset_update_task_memory_state() before hacking
1003 * our tasks mems_allowed, so that we are assured of being in
1004 * sync with our tasks cpuset, and in particular, callbacks to
1005 * cpuset_update_task_memory_state() from nested page allocations
1006 * won't see any mismatch of our cpuset and task mems_generation
1007 * values, so won't overwrite our hacked tasks mems_allowed
1008 * nodemask.
1009 */
1010
1011static void cpuset_migrate_mm(struct mm_struct *mm, const nodemask_t *from,
1012 const nodemask_t *to)
1013{
1014 struct task_struct *tsk = current;
1015
1016 cpuset_update_task_memory_state();
1017
1018 mutex_lock(&callback_mutex);
1019 tsk->mems_allowed = *to;
1020 mutex_unlock(&callback_mutex);
1021
1022 do_migrate_pages(mm, from, to, MPOL_MF_MOVE_ALL);
1023
1024 mutex_lock(&callback_mutex);
Paul Menage8793d852007-10-18 23:39:39 -07001025 guarantee_online_mems(task_cs(tsk),&tsk->mems_allowed);
Paul Jacksone4e364e2006-03-31 02:30:52 -08001026 mutex_unlock(&callback_mutex);
1027}
1028
Paul Menage8793d852007-10-18 23:39:39 -07001029static void *cpuset_being_rebound;
1030
Miao Xie0b2f6302008-07-25 01:47:21 -07001031/**
1032 * update_tasks_nodemask - Update the nodemasks of tasks in the cpuset.
1033 * @cs: the cpuset in which each task's mems_allowed mask needs to be changed
1034 * @oldmem: old mems_allowed of cpuset cs
1035 *
1036 * Called with cgroup_mutex held
1037 * Return 0 if successful, -errno if not.
1038 */
1039static int update_tasks_nodemask(struct cpuset *cs, const nodemask_t *oldmem)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001040{
Paul Menage8793d852007-10-18 23:39:39 -07001041 struct task_struct *p;
Paul Jackson42253992006-01-08 01:01:59 -08001042 struct mm_struct **mmarray;
1043 int i, n, ntasks;
Paul Jackson04c19fa2006-01-08 01:02:00 -08001044 int migrate;
Paul Jackson42253992006-01-08 01:01:59 -08001045 int fudge;
Paul Menage8793d852007-10-18 23:39:39 -07001046 struct cgroup_iter it;
Miao Xie0b2f6302008-07-25 01:47:21 -07001047 int retval;
Paul Jackson59dac162006-01-08 01:01:52 -08001048
Lee Schermerhorn846a16b2008-04-28 02:13:09 -07001049 cpuset_being_rebound = cs; /* causes mpol_dup() rebind */
Paul Jackson42253992006-01-08 01:01:59 -08001050
1051 fudge = 10; /* spare mmarray[] slots */
Li Zefan300ed6c2009-01-07 18:08:44 -08001052 fudge += cpumask_weight(cs->cpus_allowed);/* imagine 1 fork-bomb/cpu */
Paul Jackson42253992006-01-08 01:01:59 -08001053 retval = -ENOMEM;
1054
1055 /*
1056 * Allocate mmarray[] to hold mm reference for each task
1057 * in cpuset cs. Can't kmalloc GFP_KERNEL while holding
1058 * tasklist_lock. We could use GFP_ATOMIC, but with a
1059 * few more lines of code, we can retry until we get a big
1060 * enough mmarray[] w/o using GFP_ATOMIC.
1061 */
1062 while (1) {
Paul Menage8793d852007-10-18 23:39:39 -07001063 ntasks = cgroup_task_count(cs->css.cgroup); /* guess */
Paul Jackson42253992006-01-08 01:01:59 -08001064 ntasks += fudge;
1065 mmarray = kmalloc(ntasks * sizeof(*mmarray), GFP_KERNEL);
1066 if (!mmarray)
1067 goto done;
Paul Menagec2aef332007-07-15 23:40:11 -07001068 read_lock(&tasklist_lock); /* block fork */
Paul Menage8793d852007-10-18 23:39:39 -07001069 if (cgroup_task_count(cs->css.cgroup) <= ntasks)
Paul Jackson42253992006-01-08 01:01:59 -08001070 break; /* got enough */
Paul Menagec2aef332007-07-15 23:40:11 -07001071 read_unlock(&tasklist_lock); /* try again */
Paul Jackson42253992006-01-08 01:01:59 -08001072 kfree(mmarray);
1073 }
1074
1075 n = 0;
1076
1077 /* Load up mmarray[] with mm reference for each task in cpuset. */
Paul Menage8793d852007-10-18 23:39:39 -07001078 cgroup_iter_start(cs->css.cgroup, &it);
1079 while ((p = cgroup_iter_next(cs->css.cgroup, &it))) {
Paul Jackson42253992006-01-08 01:01:59 -08001080 struct mm_struct *mm;
1081
1082 if (n >= ntasks) {
1083 printk(KERN_WARNING
1084 "Cpuset mempolicy rebind incomplete.\n");
Paul Menage8793d852007-10-18 23:39:39 -07001085 break;
Paul Jackson42253992006-01-08 01:01:59 -08001086 }
Paul Jackson42253992006-01-08 01:01:59 -08001087 mm = get_task_mm(p);
1088 if (!mm)
1089 continue;
1090 mmarray[n++] = mm;
Paul Menage8793d852007-10-18 23:39:39 -07001091 }
1092 cgroup_iter_end(cs->css.cgroup, &it);
Paul Menagec2aef332007-07-15 23:40:11 -07001093 read_unlock(&tasklist_lock);
Paul Jackson42253992006-01-08 01:01:59 -08001094
1095 /*
1096 * Now that we've dropped the tasklist spinlock, we can
1097 * rebind the vma mempolicies of each mm in mmarray[] to their
1098 * new cpuset, and release that mm. The mpol_rebind_mm()
1099 * call takes mmap_sem, which we couldn't take while holding
Lee Schermerhorn846a16b2008-04-28 02:13:09 -07001100 * tasklist_lock. Forks can happen again now - the mpol_dup()
Paul Jackson42253992006-01-08 01:01:59 -08001101 * cpuset_being_rebound check will catch such forks, and rebind
1102 * their vma mempolicies too. Because we still hold the global
Paul Menage2df167a2008-02-07 00:14:45 -08001103 * cgroup_mutex, we know that no other rebind effort will
Paul Jackson42253992006-01-08 01:01:59 -08001104 * be contending for the global variable cpuset_being_rebound.
1105 * It's ok if we rebind the same mm twice; mpol_rebind_mm()
Paul Jackson04c19fa2006-01-08 01:02:00 -08001106 * is idempotent. Also migrate pages in each mm to new nodes.
Paul Jackson42253992006-01-08 01:01:59 -08001107 */
Paul Jackson04c19fa2006-01-08 01:02:00 -08001108 migrate = is_memory_migrate(cs);
Paul Jackson42253992006-01-08 01:01:59 -08001109 for (i = 0; i < n; i++) {
1110 struct mm_struct *mm = mmarray[i];
1111
1112 mpol_rebind_mm(mm, &cs->mems_allowed);
Paul Jacksone4e364e2006-03-31 02:30:52 -08001113 if (migrate)
Miao Xie0b2f6302008-07-25 01:47:21 -07001114 cpuset_migrate_mm(mm, oldmem, &cs->mems_allowed);
Paul Jackson42253992006-01-08 01:01:59 -08001115 mmput(mm);
1116 }
1117
Paul Menage2df167a2008-02-07 00:14:45 -08001118 /* We're done rebinding vmas to this cpuset's new mems_allowed. */
Paul Jackson42253992006-01-08 01:01:59 -08001119 kfree(mmarray);
Paul Menage8793d852007-10-18 23:39:39 -07001120 cpuset_being_rebound = NULL;
Paul Jackson42253992006-01-08 01:01:59 -08001121 retval = 0;
Paul Jackson59dac162006-01-08 01:01:52 -08001122done:
Linus Torvalds1da177e2005-04-16 15:20:36 -07001123 return retval;
1124}
1125
Miao Xie0b2f6302008-07-25 01:47:21 -07001126/*
1127 * Handle user request to change the 'mems' memory placement
1128 * of a cpuset. Needs to validate the request, update the
1129 * cpusets mems_allowed and mems_generation, and for each
1130 * task in the cpuset, rebind any vma mempolicies and if
1131 * the cpuset is marked 'memory_migrate', migrate the tasks
1132 * pages to the new memory.
1133 *
1134 * Call with cgroup_mutex held. May take callback_mutex during call.
1135 * Will take tasklist_lock, scan tasklist for tasks in cpuset cs,
1136 * lock each such tasks mm->mmap_sem, scan its vma's and rebind
1137 * their mempolicies to the cpusets new mems_allowed.
1138 */
Li Zefan645fcc92009-01-07 18:08:43 -08001139static int update_nodemask(struct cpuset *cs, struct cpuset *trialcs,
1140 const char *buf)
Miao Xie0b2f6302008-07-25 01:47:21 -07001141{
Miao Xie0b2f6302008-07-25 01:47:21 -07001142 nodemask_t oldmem;
1143 int retval;
1144
1145 /*
1146 * top_cpuset.mems_allowed tracks node_stats[N_HIGH_MEMORY];
1147 * it's read-only
1148 */
1149 if (cs == &top_cpuset)
1150 return -EACCES;
1151
Miao Xie0b2f6302008-07-25 01:47:21 -07001152 /*
1153 * An empty mems_allowed is ok iff there are no tasks in the cpuset.
1154 * Since nodelist_parse() fails on an empty mask, we special case
1155 * that parsing. The validate_change() call ensures that cpusets
1156 * with tasks have memory.
1157 */
1158 if (!*buf) {
Li Zefan645fcc92009-01-07 18:08:43 -08001159 nodes_clear(trialcs->mems_allowed);
Miao Xie0b2f6302008-07-25 01:47:21 -07001160 } else {
Li Zefan645fcc92009-01-07 18:08:43 -08001161 retval = nodelist_parse(buf, trialcs->mems_allowed);
Miao Xie0b2f6302008-07-25 01:47:21 -07001162 if (retval < 0)
1163 goto done;
1164
Li Zefan645fcc92009-01-07 18:08:43 -08001165 if (!nodes_subset(trialcs->mems_allowed,
Miao Xie0b2f6302008-07-25 01:47:21 -07001166 node_states[N_HIGH_MEMORY]))
1167 return -EINVAL;
1168 }
1169 oldmem = cs->mems_allowed;
Li Zefan645fcc92009-01-07 18:08:43 -08001170 if (nodes_equal(oldmem, trialcs->mems_allowed)) {
Miao Xie0b2f6302008-07-25 01:47:21 -07001171 retval = 0; /* Too easy - nothing to do */
1172 goto done;
1173 }
Li Zefan645fcc92009-01-07 18:08:43 -08001174 retval = validate_change(cs, trialcs);
Miao Xie0b2f6302008-07-25 01:47:21 -07001175 if (retval < 0)
1176 goto done;
1177
1178 mutex_lock(&callback_mutex);
Li Zefan645fcc92009-01-07 18:08:43 -08001179 cs->mems_allowed = trialcs->mems_allowed;
Miao Xie0b2f6302008-07-25 01:47:21 -07001180 cs->mems_generation = cpuset_mems_generation++;
1181 mutex_unlock(&callback_mutex);
1182
1183 retval = update_tasks_nodemask(cs, &oldmem);
1184done:
1185 return retval;
1186}
1187
Paul Menage8793d852007-10-18 23:39:39 -07001188int current_cpuset_is_being_rebound(void)
1189{
1190 return task_cs(current) == cpuset_being_rebound;
1191}
1192
Paul Menage5be7a472008-05-06 20:42:41 -07001193static int update_relax_domain_level(struct cpuset *cs, s64 val)
Hidetoshi Seto1d3504f2008-04-15 14:04:23 +09001194{
Li Zefan30e0e172008-05-13 10:27:17 +08001195 if (val < -1 || val >= SD_LV_MAX)
1196 return -EINVAL;
Hidetoshi Seto1d3504f2008-04-15 14:04:23 +09001197
1198 if (val != cs->relax_domain_level) {
1199 cs->relax_domain_level = val;
Li Zefan300ed6c2009-01-07 18:08:44 -08001200 if (!cpumask_empty(cs->cpus_allowed) &&
1201 is_sched_load_balance(cs))
Max Krasnyanskycf417142008-08-11 14:33:53 -07001202 async_rebuild_sched_domains();
Hidetoshi Seto1d3504f2008-04-15 14:04:23 +09001203 }
1204
1205 return 0;
1206}
1207
Paul Jackson3e0d98b2006-01-08 01:01:49 -08001208/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07001209 * update_flag - read a 0 or a 1 in a file and update associated flag
Paul Menage78608362008-04-29 01:00:26 -07001210 * bit: the bit to update (see cpuset_flagbits_t)
1211 * cs: the cpuset to update
1212 * turning_on: whether the flag is being set or cleared
Paul Jackson053199e2005-10-30 15:02:30 -08001213 *
Paul Menage2df167a2008-02-07 00:14:45 -08001214 * Call with cgroup_mutex held.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001215 */
1216
Paul Menage700fe1a2008-04-29 01:00:00 -07001217static int update_flag(cpuset_flagbits_t bit, struct cpuset *cs,
1218 int turning_on)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001219{
Li Zefan645fcc92009-01-07 18:08:43 -08001220 struct cpuset *trialcs;
Paul Jackson607717a2007-10-16 01:27:43 -07001221 int err;
Rakib Mullick40b6a762008-10-18 20:28:18 -07001222 int balance_flag_changed;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001223
Li Zefan645fcc92009-01-07 18:08:43 -08001224 trialcs = alloc_trial_cpuset(cs);
1225 if (!trialcs)
1226 return -ENOMEM;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001227
Li Zefan645fcc92009-01-07 18:08:43 -08001228 if (turning_on)
1229 set_bit(bit, &trialcs->flags);
1230 else
1231 clear_bit(bit, &trialcs->flags);
1232
1233 err = validate_change(cs, trialcs);
Dinakar Guniguntala85d7b942005-06-25 14:57:34 -07001234 if (err < 0)
Li Zefan645fcc92009-01-07 18:08:43 -08001235 goto out;
Paul Jackson029190c2007-10-18 23:40:20 -07001236
Paul Jackson029190c2007-10-18 23:40:20 -07001237 balance_flag_changed = (is_sched_load_balance(cs) !=
Li Zefan645fcc92009-01-07 18:08:43 -08001238 is_sched_load_balance(trialcs));
Paul Jackson029190c2007-10-18 23:40:20 -07001239
Ingo Molnar3d3f26a2006-03-23 03:00:18 -08001240 mutex_lock(&callback_mutex);
Li Zefan645fcc92009-01-07 18:08:43 -08001241 cs->flags = trialcs->flags;
Ingo Molnar3d3f26a2006-03-23 03:00:18 -08001242 mutex_unlock(&callback_mutex);
Dinakar Guniguntala85d7b942005-06-25 14:57:34 -07001243
Li Zefan300ed6c2009-01-07 18:08:44 -08001244 if (!cpumask_empty(trialcs->cpus_allowed) && balance_flag_changed)
Max Krasnyanskycf417142008-08-11 14:33:53 -07001245 async_rebuild_sched_domains();
Paul Jackson029190c2007-10-18 23:40:20 -07001246
Li Zefan645fcc92009-01-07 18:08:43 -08001247out:
1248 free_trial_cpuset(trialcs);
1249 return err;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001250}
1251
Paul Jackson053199e2005-10-30 15:02:30 -08001252/*
Adrian Bunk80f72282006-06-30 18:27:16 +02001253 * Frequency meter - How fast is some event occurring?
Paul Jackson3e0d98b2006-01-08 01:01:49 -08001254 *
1255 * These routines manage a digitally filtered, constant time based,
1256 * event frequency meter. There are four routines:
1257 * fmeter_init() - initialize a frequency meter.
1258 * fmeter_markevent() - called each time the event happens.
1259 * fmeter_getrate() - returns the recent rate of such events.
1260 * fmeter_update() - internal routine used to update fmeter.
1261 *
1262 * A common data structure is passed to each of these routines,
1263 * which is used to keep track of the state required to manage the
1264 * frequency meter and its digital filter.
1265 *
1266 * The filter works on the number of events marked per unit time.
1267 * The filter is single-pole low-pass recursive (IIR). The time unit
1268 * is 1 second. Arithmetic is done using 32-bit integers scaled to
1269 * simulate 3 decimal digits of precision (multiplied by 1000).
1270 *
1271 * With an FM_COEF of 933, and a time base of 1 second, the filter
1272 * has a half-life of 10 seconds, meaning that if the events quit
1273 * happening, then the rate returned from the fmeter_getrate()
1274 * will be cut in half each 10 seconds, until it converges to zero.
1275 *
1276 * It is not worth doing a real infinitely recursive filter. If more
1277 * than FM_MAXTICKS ticks have elapsed since the last filter event,
1278 * just compute FM_MAXTICKS ticks worth, by which point the level
1279 * will be stable.
1280 *
1281 * Limit the count of unprocessed events to FM_MAXCNT, so as to avoid
1282 * arithmetic overflow in the fmeter_update() routine.
1283 *
1284 * Given the simple 32 bit integer arithmetic used, this meter works
1285 * best for reporting rates between one per millisecond (msec) and
1286 * one per 32 (approx) seconds. At constant rates faster than one
1287 * per msec it maxes out at values just under 1,000,000. At constant
1288 * rates between one per msec, and one per second it will stabilize
1289 * to a value N*1000, where N is the rate of events per second.
1290 * At constant rates between one per second and one per 32 seconds,
1291 * it will be choppy, moving up on the seconds that have an event,
1292 * and then decaying until the next event. At rates slower than
1293 * about one in 32 seconds, it decays all the way back to zero between
1294 * each event.
1295 */
1296
1297#define FM_COEF 933 /* coefficient for half-life of 10 secs */
1298#define FM_MAXTICKS ((time_t)99) /* useless computing more ticks than this */
1299#define FM_MAXCNT 1000000 /* limit cnt to avoid overflow */
1300#define FM_SCALE 1000 /* faux fixed point scale */
1301
1302/* Initialize a frequency meter */
1303static void fmeter_init(struct fmeter *fmp)
1304{
1305 fmp->cnt = 0;
1306 fmp->val = 0;
1307 fmp->time = 0;
1308 spin_lock_init(&fmp->lock);
1309}
1310
1311/* Internal meter update - process cnt events and update value */
1312static void fmeter_update(struct fmeter *fmp)
1313{
1314 time_t now = get_seconds();
1315 time_t ticks = now - fmp->time;
1316
1317 if (ticks == 0)
1318 return;
1319
1320 ticks = min(FM_MAXTICKS, ticks);
1321 while (ticks-- > 0)
1322 fmp->val = (FM_COEF * fmp->val) / FM_SCALE;
1323 fmp->time = now;
1324
1325 fmp->val += ((FM_SCALE - FM_COEF) * fmp->cnt) / FM_SCALE;
1326 fmp->cnt = 0;
1327}
1328
1329/* Process any previous ticks, then bump cnt by one (times scale). */
1330static void fmeter_markevent(struct fmeter *fmp)
1331{
1332 spin_lock(&fmp->lock);
1333 fmeter_update(fmp);
1334 fmp->cnt = min(FM_MAXCNT, fmp->cnt + FM_SCALE);
1335 spin_unlock(&fmp->lock);
1336}
1337
1338/* Process any previous ticks, then return current value. */
1339static int fmeter_getrate(struct fmeter *fmp)
1340{
1341 int val;
1342
1343 spin_lock(&fmp->lock);
1344 fmeter_update(fmp);
1345 val = fmp->val;
1346 spin_unlock(&fmp->lock);
1347 return val;
1348}
1349
Li Zefan2341d1b2009-01-07 18:08:42 -08001350/* Protected by cgroup_lock */
1351static cpumask_var_t cpus_attach;
1352
Paul Menage2df167a2008-02-07 00:14:45 -08001353/* Called by cgroups to determine if a cpuset is usable; cgroup_mutex held */
Paul Menage8793d852007-10-18 23:39:39 -07001354static int cpuset_can_attach(struct cgroup_subsys *ss,
1355 struct cgroup *cont, struct task_struct *tsk)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001356{
Paul Menage8793d852007-10-18 23:39:39 -07001357 struct cpuset *cs = cgroup_cs(cont);
Li Zefan5771f0a2009-01-07 18:08:41 -08001358 int ret = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001359
Li Zefan300ed6c2009-01-07 18:08:44 -08001360 if (cpumask_empty(cs->cpus_allowed) || nodes_empty(cs->mems_allowed))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001361 return -ENOSPC;
David Rientjes9985b0b2008-06-05 12:57:11 -07001362
Li Zefan5771f0a2009-01-07 18:08:41 -08001363 if (tsk->flags & PF_THREAD_BOUND) {
David Rientjes9985b0b2008-06-05 12:57:11 -07001364 mutex_lock(&callback_mutex);
Li Zefan300ed6c2009-01-07 18:08:44 -08001365 if (!cpumask_equal(&tsk->cpus_allowed, cs->cpus_allowed))
Li Zefan5771f0a2009-01-07 18:08:41 -08001366 ret = -EINVAL;
David Rientjes9985b0b2008-06-05 12:57:11 -07001367 mutex_unlock(&callback_mutex);
David Rientjes9985b0b2008-06-05 12:57:11 -07001368 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001369
Li Zefan5771f0a2009-01-07 18:08:41 -08001370 return ret < 0 ? ret : security_task_setscheduler(tsk, 0, NULL);
Paul Menage8793d852007-10-18 23:39:39 -07001371}
Linus Torvalds1da177e2005-04-16 15:20:36 -07001372
Paul Menage8793d852007-10-18 23:39:39 -07001373static void cpuset_attach(struct cgroup_subsys *ss,
1374 struct cgroup *cont, struct cgroup *oldcont,
1375 struct task_struct *tsk)
1376{
Paul Menage8793d852007-10-18 23:39:39 -07001377 nodemask_t from, to;
1378 struct mm_struct *mm;
1379 struct cpuset *cs = cgroup_cs(cont);
1380 struct cpuset *oldcs = cgroup_cs(oldcont);
David Rientjes9985b0b2008-06-05 12:57:11 -07001381 int err;
David Quigley22fb52d2006-06-23 02:04:00 -07001382
Miao Xief5813d92009-01-07 18:08:40 -08001383 if (cs == &top_cpuset) {
Li Zefan2341d1b2009-01-07 18:08:42 -08001384 cpumask_copy(cpus_attach, cpu_possible_mask);
Miao Xief5813d92009-01-07 18:08:40 -08001385 } else {
1386 mutex_lock(&callback_mutex);
Li Zefan2341d1b2009-01-07 18:08:42 -08001387 guarantee_online_cpus(cs, cpus_attach);
Miao Xief5813d92009-01-07 18:08:40 -08001388 mutex_unlock(&callback_mutex);
1389 }
Li Zefan2341d1b2009-01-07 18:08:42 -08001390 err = set_cpus_allowed_ptr(tsk, cpus_attach);
David Rientjes9985b0b2008-06-05 12:57:11 -07001391 if (err)
1392 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001393
Paul Jackson45b07ef2006-01-08 01:00:56 -08001394 from = oldcs->mems_allowed;
1395 to = cs->mems_allowed;
Paul Jackson42253992006-01-08 01:01:59 -08001396 mm = get_task_mm(tsk);
1397 if (mm) {
1398 mpol_rebind_mm(mm, &to);
Paul Jackson2741a552006-03-31 02:30:51 -08001399 if (is_memory_migrate(cs))
Paul Jacksone4e364e2006-03-31 02:30:52 -08001400 cpuset_migrate_mm(mm, &from, &to);
Paul Jackson42253992006-01-08 01:01:59 -08001401 mmput(mm);
1402 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001403}
1404
1405/* The various types of files and directories in a cpuset file system */
1406
1407typedef enum {
Paul Jackson45b07ef2006-01-08 01:00:56 -08001408 FILE_MEMORY_MIGRATE,
Linus Torvalds1da177e2005-04-16 15:20:36 -07001409 FILE_CPULIST,
1410 FILE_MEMLIST,
1411 FILE_CPU_EXCLUSIVE,
1412 FILE_MEM_EXCLUSIVE,
Paul Menage78608362008-04-29 01:00:26 -07001413 FILE_MEM_HARDWALL,
Paul Jackson029190c2007-10-18 23:40:20 -07001414 FILE_SCHED_LOAD_BALANCE,
Hidetoshi Seto1d3504f2008-04-15 14:04:23 +09001415 FILE_SCHED_RELAX_DOMAIN_LEVEL,
Paul Jackson3e0d98b2006-01-08 01:01:49 -08001416 FILE_MEMORY_PRESSURE_ENABLED,
1417 FILE_MEMORY_PRESSURE,
Paul Jackson825a46a2006-03-24 03:16:03 -08001418 FILE_SPREAD_PAGE,
1419 FILE_SPREAD_SLAB,
Linus Torvalds1da177e2005-04-16 15:20:36 -07001420} cpuset_filetype_t;
1421
Paul Menage700fe1a2008-04-29 01:00:00 -07001422static int cpuset_write_u64(struct cgroup *cgrp, struct cftype *cft, u64 val)
1423{
1424 int retval = 0;
1425 struct cpuset *cs = cgroup_cs(cgrp);
1426 cpuset_filetype_t type = cft->private;
1427
Paul Menagee3712392008-07-25 01:47:02 -07001428 if (!cgroup_lock_live_group(cgrp))
Paul Menage700fe1a2008-04-29 01:00:00 -07001429 return -ENODEV;
Paul Menage700fe1a2008-04-29 01:00:00 -07001430
1431 switch (type) {
1432 case FILE_CPU_EXCLUSIVE:
1433 retval = update_flag(CS_CPU_EXCLUSIVE, cs, val);
1434 break;
1435 case FILE_MEM_EXCLUSIVE:
1436 retval = update_flag(CS_MEM_EXCLUSIVE, cs, val);
1437 break;
Paul Menage78608362008-04-29 01:00:26 -07001438 case FILE_MEM_HARDWALL:
1439 retval = update_flag(CS_MEM_HARDWALL, cs, val);
1440 break;
Paul Menage700fe1a2008-04-29 01:00:00 -07001441 case FILE_SCHED_LOAD_BALANCE:
1442 retval = update_flag(CS_SCHED_LOAD_BALANCE, cs, val);
1443 break;
1444 case FILE_MEMORY_MIGRATE:
1445 retval = update_flag(CS_MEMORY_MIGRATE, cs, val);
1446 break;
1447 case FILE_MEMORY_PRESSURE_ENABLED:
1448 cpuset_memory_pressure_enabled = !!val;
1449 break;
1450 case FILE_MEMORY_PRESSURE:
1451 retval = -EACCES;
1452 break;
1453 case FILE_SPREAD_PAGE:
1454 retval = update_flag(CS_SPREAD_PAGE, cs, val);
1455 cs->mems_generation = cpuset_mems_generation++;
1456 break;
1457 case FILE_SPREAD_SLAB:
1458 retval = update_flag(CS_SPREAD_SLAB, cs, val);
1459 cs->mems_generation = cpuset_mems_generation++;
1460 break;
1461 default:
1462 retval = -EINVAL;
1463 break;
1464 }
1465 cgroup_unlock();
1466 return retval;
1467}
1468
Paul Menage5be7a472008-05-06 20:42:41 -07001469static int cpuset_write_s64(struct cgroup *cgrp, struct cftype *cft, s64 val)
1470{
1471 int retval = 0;
1472 struct cpuset *cs = cgroup_cs(cgrp);
1473 cpuset_filetype_t type = cft->private;
1474
Paul Menagee3712392008-07-25 01:47:02 -07001475 if (!cgroup_lock_live_group(cgrp))
Paul Menage5be7a472008-05-06 20:42:41 -07001476 return -ENODEV;
Paul Menagee3712392008-07-25 01:47:02 -07001477
Paul Menage5be7a472008-05-06 20:42:41 -07001478 switch (type) {
1479 case FILE_SCHED_RELAX_DOMAIN_LEVEL:
1480 retval = update_relax_domain_level(cs, val);
1481 break;
1482 default:
1483 retval = -EINVAL;
1484 break;
1485 }
1486 cgroup_unlock();
1487 return retval;
1488}
1489
Linus Torvalds1da177e2005-04-16 15:20:36 -07001490/*
Paul Menagee3712392008-07-25 01:47:02 -07001491 * Common handling for a write to a "cpus" or "mems" file.
1492 */
1493static int cpuset_write_resmask(struct cgroup *cgrp, struct cftype *cft,
1494 const char *buf)
1495{
1496 int retval = 0;
Li Zefan645fcc92009-01-07 18:08:43 -08001497 struct cpuset *cs = cgroup_cs(cgrp);
1498 struct cpuset *trialcs;
Paul Menagee3712392008-07-25 01:47:02 -07001499
1500 if (!cgroup_lock_live_group(cgrp))
1501 return -ENODEV;
1502
Li Zefan645fcc92009-01-07 18:08:43 -08001503 trialcs = alloc_trial_cpuset(cs);
1504 if (!trialcs)
1505 return -ENOMEM;
1506
Paul Menagee3712392008-07-25 01:47:02 -07001507 switch (cft->private) {
1508 case FILE_CPULIST:
Li Zefan645fcc92009-01-07 18:08:43 -08001509 retval = update_cpumask(cs, trialcs, buf);
Paul Menagee3712392008-07-25 01:47:02 -07001510 break;
1511 case FILE_MEMLIST:
Li Zefan645fcc92009-01-07 18:08:43 -08001512 retval = update_nodemask(cs, trialcs, buf);
Paul Menagee3712392008-07-25 01:47:02 -07001513 break;
1514 default:
1515 retval = -EINVAL;
1516 break;
1517 }
Li Zefan645fcc92009-01-07 18:08:43 -08001518
1519 free_trial_cpuset(trialcs);
Paul Menagee3712392008-07-25 01:47:02 -07001520 cgroup_unlock();
1521 return retval;
1522}
1523
1524/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07001525 * These ascii lists should be read in a single call, by using a user
1526 * buffer large enough to hold the entire map. If read in smaller
1527 * chunks, there is no guarantee of atomicity. Since the display format
1528 * used, list of ranges of sequential numbers, is variable length,
1529 * and since these maps can change value dynamically, one could read
1530 * gibberish by doing partial reads while a list was changing.
1531 * A single large read to a buffer that crosses a page boundary is
1532 * ok, because the result being copied to user land is not recomputed
1533 * across a page fault.
1534 */
1535
1536static int cpuset_sprintf_cpulist(char *page, struct cpuset *cs)
1537{
Li Zefan5a7625d2009-01-07 18:08:41 -08001538 int ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001539
Ingo Molnar3d3f26a2006-03-23 03:00:18 -08001540 mutex_lock(&callback_mutex);
Li Zefan300ed6c2009-01-07 18:08:44 -08001541 ret = cpulist_scnprintf(page, PAGE_SIZE, cs->cpus_allowed);
Ingo Molnar3d3f26a2006-03-23 03:00:18 -08001542 mutex_unlock(&callback_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001543
Li Zefan5a7625d2009-01-07 18:08:41 -08001544 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001545}
1546
1547static int cpuset_sprintf_memlist(char *page, struct cpuset *cs)
1548{
1549 nodemask_t mask;
1550
Ingo Molnar3d3f26a2006-03-23 03:00:18 -08001551 mutex_lock(&callback_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001552 mask = cs->mems_allowed;
Ingo Molnar3d3f26a2006-03-23 03:00:18 -08001553 mutex_unlock(&callback_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001554
1555 return nodelist_scnprintf(page, PAGE_SIZE, mask);
1556}
1557
Paul Menage8793d852007-10-18 23:39:39 -07001558static ssize_t cpuset_common_file_read(struct cgroup *cont,
1559 struct cftype *cft,
1560 struct file *file,
1561 char __user *buf,
1562 size_t nbytes, loff_t *ppos)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001563{
Paul Menage8793d852007-10-18 23:39:39 -07001564 struct cpuset *cs = cgroup_cs(cont);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001565 cpuset_filetype_t type = cft->private;
1566 char *page;
1567 ssize_t retval = 0;
1568 char *s;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001569
Mel Gormane12ba742007-10-16 01:25:52 -07001570 if (!(page = (char *)__get_free_page(GFP_TEMPORARY)))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001571 return -ENOMEM;
1572
1573 s = page;
1574
1575 switch (type) {
1576 case FILE_CPULIST:
1577 s += cpuset_sprintf_cpulist(s, cs);
1578 break;
1579 case FILE_MEMLIST:
1580 s += cpuset_sprintf_memlist(s, cs);
1581 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001582 default:
1583 retval = -EINVAL;
1584 goto out;
1585 }
1586 *s++ = '\n';
Linus Torvalds1da177e2005-04-16 15:20:36 -07001587
Al Viroeacaa1f2005-09-30 03:26:43 +01001588 retval = simple_read_from_buffer(buf, nbytes, ppos, page, s - page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001589out:
1590 free_page((unsigned long)page);
1591 return retval;
1592}
1593
Paul Menage700fe1a2008-04-29 01:00:00 -07001594static u64 cpuset_read_u64(struct cgroup *cont, struct cftype *cft)
1595{
1596 struct cpuset *cs = cgroup_cs(cont);
1597 cpuset_filetype_t type = cft->private;
1598 switch (type) {
1599 case FILE_CPU_EXCLUSIVE:
1600 return is_cpu_exclusive(cs);
1601 case FILE_MEM_EXCLUSIVE:
1602 return is_mem_exclusive(cs);
Paul Menage78608362008-04-29 01:00:26 -07001603 case FILE_MEM_HARDWALL:
1604 return is_mem_hardwall(cs);
Paul Menage700fe1a2008-04-29 01:00:00 -07001605 case FILE_SCHED_LOAD_BALANCE:
1606 return is_sched_load_balance(cs);
1607 case FILE_MEMORY_MIGRATE:
1608 return is_memory_migrate(cs);
1609 case FILE_MEMORY_PRESSURE_ENABLED:
1610 return cpuset_memory_pressure_enabled;
1611 case FILE_MEMORY_PRESSURE:
1612 return fmeter_getrate(&cs->fmeter);
1613 case FILE_SPREAD_PAGE:
1614 return is_spread_page(cs);
1615 case FILE_SPREAD_SLAB:
1616 return is_spread_slab(cs);
1617 default:
1618 BUG();
1619 }
Max Krasnyanskycf417142008-08-11 14:33:53 -07001620
1621 /* Unreachable but makes gcc happy */
1622 return 0;
Paul Menage700fe1a2008-04-29 01:00:00 -07001623}
Linus Torvalds1da177e2005-04-16 15:20:36 -07001624
Paul Menage5be7a472008-05-06 20:42:41 -07001625static s64 cpuset_read_s64(struct cgroup *cont, struct cftype *cft)
1626{
1627 struct cpuset *cs = cgroup_cs(cont);
1628 cpuset_filetype_t type = cft->private;
1629 switch (type) {
1630 case FILE_SCHED_RELAX_DOMAIN_LEVEL:
1631 return cs->relax_domain_level;
1632 default:
1633 BUG();
1634 }
Max Krasnyanskycf417142008-08-11 14:33:53 -07001635
1636 /* Unrechable but makes gcc happy */
1637 return 0;
Paul Menage5be7a472008-05-06 20:42:41 -07001638}
1639
Linus Torvalds1da177e2005-04-16 15:20:36 -07001640
1641/*
1642 * for the common functions, 'private' gives the type of file
1643 */
1644
Paul Menageaddf2c72008-04-29 01:00:26 -07001645static struct cftype files[] = {
1646 {
1647 .name = "cpus",
1648 .read = cpuset_common_file_read,
Paul Menagee3712392008-07-25 01:47:02 -07001649 .write_string = cpuset_write_resmask,
1650 .max_write_len = (100U + 6 * NR_CPUS),
Paul Menageaddf2c72008-04-29 01:00:26 -07001651 .private = FILE_CPULIST,
1652 },
Linus Torvalds1da177e2005-04-16 15:20:36 -07001653
Paul Menageaddf2c72008-04-29 01:00:26 -07001654 {
1655 .name = "mems",
1656 .read = cpuset_common_file_read,
Paul Menagee3712392008-07-25 01:47:02 -07001657 .write_string = cpuset_write_resmask,
1658 .max_write_len = (100U + 6 * MAX_NUMNODES),
Paul Menageaddf2c72008-04-29 01:00:26 -07001659 .private = FILE_MEMLIST,
1660 },
Linus Torvalds1da177e2005-04-16 15:20:36 -07001661
Paul Menageaddf2c72008-04-29 01:00:26 -07001662 {
1663 .name = "cpu_exclusive",
1664 .read_u64 = cpuset_read_u64,
1665 .write_u64 = cpuset_write_u64,
1666 .private = FILE_CPU_EXCLUSIVE,
1667 },
Linus Torvalds1da177e2005-04-16 15:20:36 -07001668
Paul Menageaddf2c72008-04-29 01:00:26 -07001669 {
1670 .name = "mem_exclusive",
1671 .read_u64 = cpuset_read_u64,
1672 .write_u64 = cpuset_write_u64,
1673 .private = FILE_MEM_EXCLUSIVE,
1674 },
Linus Torvalds1da177e2005-04-16 15:20:36 -07001675
Paul Menageaddf2c72008-04-29 01:00:26 -07001676 {
Paul Menage78608362008-04-29 01:00:26 -07001677 .name = "mem_hardwall",
1678 .read_u64 = cpuset_read_u64,
1679 .write_u64 = cpuset_write_u64,
1680 .private = FILE_MEM_HARDWALL,
1681 },
1682
1683 {
Paul Menageaddf2c72008-04-29 01:00:26 -07001684 .name = "sched_load_balance",
1685 .read_u64 = cpuset_read_u64,
1686 .write_u64 = cpuset_write_u64,
1687 .private = FILE_SCHED_LOAD_BALANCE,
1688 },
Paul Jackson029190c2007-10-18 23:40:20 -07001689
Paul Menageaddf2c72008-04-29 01:00:26 -07001690 {
1691 .name = "sched_relax_domain_level",
Paul Menage5be7a472008-05-06 20:42:41 -07001692 .read_s64 = cpuset_read_s64,
1693 .write_s64 = cpuset_write_s64,
Paul Menageaddf2c72008-04-29 01:00:26 -07001694 .private = FILE_SCHED_RELAX_DOMAIN_LEVEL,
1695 },
Hidetoshi Seto1d3504f2008-04-15 14:04:23 +09001696
Paul Menageaddf2c72008-04-29 01:00:26 -07001697 {
1698 .name = "memory_migrate",
1699 .read_u64 = cpuset_read_u64,
1700 .write_u64 = cpuset_write_u64,
1701 .private = FILE_MEMORY_MIGRATE,
1702 },
1703
1704 {
1705 .name = "memory_pressure",
1706 .read_u64 = cpuset_read_u64,
1707 .write_u64 = cpuset_write_u64,
1708 .private = FILE_MEMORY_PRESSURE,
1709 },
1710
1711 {
1712 .name = "memory_spread_page",
1713 .read_u64 = cpuset_read_u64,
1714 .write_u64 = cpuset_write_u64,
1715 .private = FILE_SPREAD_PAGE,
1716 },
1717
1718 {
1719 .name = "memory_spread_slab",
1720 .read_u64 = cpuset_read_u64,
1721 .write_u64 = cpuset_write_u64,
1722 .private = FILE_SPREAD_SLAB,
1723 },
Paul Jackson45b07ef2006-01-08 01:00:56 -08001724};
1725
Paul Jackson3e0d98b2006-01-08 01:01:49 -08001726static struct cftype cft_memory_pressure_enabled = {
1727 .name = "memory_pressure_enabled",
Paul Menage700fe1a2008-04-29 01:00:00 -07001728 .read_u64 = cpuset_read_u64,
1729 .write_u64 = cpuset_write_u64,
Paul Jackson3e0d98b2006-01-08 01:01:49 -08001730 .private = FILE_MEMORY_PRESSURE_ENABLED,
1731};
1732
Paul Menage8793d852007-10-18 23:39:39 -07001733static int cpuset_populate(struct cgroup_subsys *ss, struct cgroup *cont)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001734{
1735 int err;
1736
Paul Menageaddf2c72008-04-29 01:00:26 -07001737 err = cgroup_add_files(cont, ss, files, ARRAY_SIZE(files));
1738 if (err)
Paul Jackson825a46a2006-03-24 03:16:03 -08001739 return err;
Paul Menage8793d852007-10-18 23:39:39 -07001740 /* memory_pressure_enabled is in root cpuset only */
Paul Menageaddf2c72008-04-29 01:00:26 -07001741 if (!cont->parent)
Paul Menage8793d852007-10-18 23:39:39 -07001742 err = cgroup_add_file(cont, ss,
Paul Menageaddf2c72008-04-29 01:00:26 -07001743 &cft_memory_pressure_enabled);
1744 return err;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001745}
1746
1747/*
Paul Menage8793d852007-10-18 23:39:39 -07001748 * post_clone() is called at the end of cgroup_clone().
1749 * 'cgroup' was just created automatically as a result of
1750 * a cgroup_clone(), and the current task is about to
1751 * be moved into 'cgroup'.
1752 *
1753 * Currently we refuse to set up the cgroup - thereby
1754 * refusing the task to be entered, and as a result refusing
1755 * the sys_unshare() or clone() which initiated it - if any
1756 * sibling cpusets have exclusive cpus or mem.
1757 *
1758 * If this becomes a problem for some users who wish to
1759 * allow that scenario, then cpuset_post_clone() could be
1760 * changed to grant parent->cpus_allowed-sibling_cpus_exclusive
Paul Menage2df167a2008-02-07 00:14:45 -08001761 * (and likewise for mems) to the new cgroup. Called with cgroup_mutex
1762 * held.
Paul Menage8793d852007-10-18 23:39:39 -07001763 */
1764static void cpuset_post_clone(struct cgroup_subsys *ss,
1765 struct cgroup *cgroup)
1766{
1767 struct cgroup *parent, *child;
1768 struct cpuset *cs, *parent_cs;
1769
1770 parent = cgroup->parent;
1771 list_for_each_entry(child, &parent->children, sibling) {
1772 cs = cgroup_cs(child);
1773 if (is_mem_exclusive(cs) || is_cpu_exclusive(cs))
1774 return;
1775 }
1776 cs = cgroup_cs(cgroup);
1777 parent_cs = cgroup_cs(parent);
1778
1779 cs->mems_allowed = parent_cs->mems_allowed;
Li Zefan300ed6c2009-01-07 18:08:44 -08001780 cpumask_copy(cs->cpus_allowed, parent_cs->cpus_allowed);
Paul Menage8793d852007-10-18 23:39:39 -07001781 return;
1782}
1783
1784/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07001785 * cpuset_create - create a cpuset
Paul Menage2df167a2008-02-07 00:14:45 -08001786 * ss: cpuset cgroup subsystem
1787 * cont: control group that the new cpuset will be part of
Linus Torvalds1da177e2005-04-16 15:20:36 -07001788 */
1789
Paul Menage8793d852007-10-18 23:39:39 -07001790static struct cgroup_subsys_state *cpuset_create(
1791 struct cgroup_subsys *ss,
1792 struct cgroup *cont)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001793{
1794 struct cpuset *cs;
Paul Menage8793d852007-10-18 23:39:39 -07001795 struct cpuset *parent;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001796
Paul Menage8793d852007-10-18 23:39:39 -07001797 if (!cont->parent) {
1798 /* This is early initialization for the top cgroup */
1799 top_cpuset.mems_generation = cpuset_mems_generation++;
1800 return &top_cpuset.css;
1801 }
1802 parent = cgroup_cs(cont->parent);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001803 cs = kmalloc(sizeof(*cs), GFP_KERNEL);
1804 if (!cs)
Paul Menage8793d852007-10-18 23:39:39 -07001805 return ERR_PTR(-ENOMEM);
Li Zefan300ed6c2009-01-07 18:08:44 -08001806 if (!alloc_cpumask_var(&cs->cpus_allowed, GFP_KERNEL)) {
1807 kfree(cs);
1808 return ERR_PTR(-ENOMEM);
1809 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001810
Paul Jacksoncf2a473c2006-01-08 01:01:54 -08001811 cpuset_update_task_memory_state();
Linus Torvalds1da177e2005-04-16 15:20:36 -07001812 cs->flags = 0;
Paul Jackson825a46a2006-03-24 03:16:03 -08001813 if (is_spread_page(parent))
1814 set_bit(CS_SPREAD_PAGE, &cs->flags);
1815 if (is_spread_slab(parent))
1816 set_bit(CS_SPREAD_SLAB, &cs->flags);
Paul Jackson029190c2007-10-18 23:40:20 -07001817 set_bit(CS_SCHED_LOAD_BALANCE, &cs->flags);
Li Zefan300ed6c2009-01-07 18:08:44 -08001818 cpumask_clear(cs->cpus_allowed);
Mike Travisf9a86fc2008-04-04 18:11:07 -07001819 nodes_clear(cs->mems_allowed);
Paul Jackson151a4422006-03-24 03:16:11 -08001820 cs->mems_generation = cpuset_mems_generation++;
Paul Jackson3e0d98b2006-01-08 01:01:49 -08001821 fmeter_init(&cs->fmeter);
Hidetoshi Seto1d3504f2008-04-15 14:04:23 +09001822 cs->relax_domain_level = -1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001823
1824 cs->parent = parent;
Paul Jackson202f72d2006-01-08 01:01:57 -08001825 number_of_cpusets++;
Paul Menage8793d852007-10-18 23:39:39 -07001826 return &cs->css ;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001827}
1828
Paul Jackson029190c2007-10-18 23:40:20 -07001829/*
Paul Jackson029190c2007-10-18 23:40:20 -07001830 * If the cpuset being removed has its flag 'sched_load_balance'
1831 * enabled, then simulate turning sched_load_balance off, which
Max Krasnyanskycf417142008-08-11 14:33:53 -07001832 * will call async_rebuild_sched_domains().
Paul Jackson029190c2007-10-18 23:40:20 -07001833 */
1834
Paul Menage8793d852007-10-18 23:39:39 -07001835static void cpuset_destroy(struct cgroup_subsys *ss, struct cgroup *cont)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001836{
Paul Menage8793d852007-10-18 23:39:39 -07001837 struct cpuset *cs = cgroup_cs(cont);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001838
Paul Jacksoncf2a473c2006-01-08 01:01:54 -08001839 cpuset_update_task_memory_state();
Paul Jackson029190c2007-10-18 23:40:20 -07001840
1841 if (is_sched_load_balance(cs))
Paul Menage700fe1a2008-04-29 01:00:00 -07001842 update_flag(CS_SCHED_LOAD_BALANCE, cs, 0);
Paul Jackson029190c2007-10-18 23:40:20 -07001843
Paul Jackson202f72d2006-01-08 01:01:57 -08001844 number_of_cpusets--;
Li Zefan300ed6c2009-01-07 18:08:44 -08001845 free_cpumask_var(cs->cpus_allowed);
Paul Menage8793d852007-10-18 23:39:39 -07001846 kfree(cs);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001847}
1848
Paul Menage8793d852007-10-18 23:39:39 -07001849struct cgroup_subsys cpuset_subsys = {
1850 .name = "cpuset",
1851 .create = cpuset_create,
Max Krasnyanskycf417142008-08-11 14:33:53 -07001852 .destroy = cpuset_destroy,
Paul Menage8793d852007-10-18 23:39:39 -07001853 .can_attach = cpuset_can_attach,
1854 .attach = cpuset_attach,
1855 .populate = cpuset_populate,
1856 .post_clone = cpuset_post_clone,
1857 .subsys_id = cpuset_subsys_id,
1858 .early_init = 1,
1859};
1860
Paul Jacksonc417f022006-01-08 01:02:01 -08001861/*
1862 * cpuset_init_early - just enough so that the calls to
1863 * cpuset_update_task_memory_state() in early init code
1864 * are harmless.
1865 */
1866
1867int __init cpuset_init_early(void)
1868{
Li Zefan300ed6c2009-01-07 18:08:44 -08001869 alloc_bootmem_cpumask_var(&top_cpuset.cpus_allowed);
1870
Paul Menage8793d852007-10-18 23:39:39 -07001871 top_cpuset.mems_generation = cpuset_mems_generation++;
Paul Jacksonc417f022006-01-08 01:02:01 -08001872 return 0;
1873}
1874
Paul Menage8793d852007-10-18 23:39:39 -07001875
Linus Torvalds1da177e2005-04-16 15:20:36 -07001876/**
1877 * cpuset_init - initialize cpusets at system boot
1878 *
1879 * Description: Initialize top_cpuset and the cpuset internal file system,
1880 **/
1881
1882int __init cpuset_init(void)
1883{
Paul Menage8793d852007-10-18 23:39:39 -07001884 int err = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001885
Li Zefan300ed6c2009-01-07 18:08:44 -08001886 cpumask_setall(top_cpuset.cpus_allowed);
Mike Travisf9a86fc2008-04-04 18:11:07 -07001887 nodes_setall(top_cpuset.mems_allowed);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001888
Paul Jackson3e0d98b2006-01-08 01:01:49 -08001889 fmeter_init(&top_cpuset.fmeter);
Paul Jackson151a4422006-03-24 03:16:11 -08001890 top_cpuset.mems_generation = cpuset_mems_generation++;
Paul Jackson029190c2007-10-18 23:40:20 -07001891 set_bit(CS_SCHED_LOAD_BALANCE, &top_cpuset.flags);
Hidetoshi Seto1d3504f2008-04-15 14:04:23 +09001892 top_cpuset.relax_domain_level = -1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001893
Linus Torvalds1da177e2005-04-16 15:20:36 -07001894 err = register_filesystem(&cpuset_fs_type);
1895 if (err < 0)
Paul Menage8793d852007-10-18 23:39:39 -07001896 return err;
1897
Li Zefan2341d1b2009-01-07 18:08:42 -08001898 if (!alloc_cpumask_var(&cpus_attach, GFP_KERNEL))
1899 BUG();
1900
Paul Jackson202f72d2006-01-08 01:01:57 -08001901 number_of_cpusets = 1;
Paul Menage8793d852007-10-18 23:39:39 -07001902 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001903}
1904
Cliff Wickman956db3c2008-02-07 00:14:43 -08001905/**
1906 * cpuset_do_move_task - move a given task to another cpuset
1907 * @tsk: pointer to task_struct the task to move
1908 * @scan: struct cgroup_scanner contained in its struct cpuset_hotplug_scanner
1909 *
1910 * Called by cgroup_scan_tasks() for each task in a cgroup.
1911 * Return nonzero to stop the walk through the tasks.
1912 */
Adrian Bunk9e0c9142008-04-29 01:00:25 -07001913static void cpuset_do_move_task(struct task_struct *tsk,
1914 struct cgroup_scanner *scan)
Cliff Wickman956db3c2008-02-07 00:14:43 -08001915{
1916 struct cpuset_hotplug_scanner *chsp;
1917
1918 chsp = container_of(scan, struct cpuset_hotplug_scanner, scan);
1919 cgroup_attach_task(chsp->to, tsk);
1920}
1921
1922/**
1923 * move_member_tasks_to_cpuset - move tasks from one cpuset to another
1924 * @from: cpuset in which the tasks currently reside
1925 * @to: cpuset to which the tasks will be moved
1926 *
Paul Jacksonc8d9c902008-02-07 00:14:46 -08001927 * Called with cgroup_mutex held
1928 * callback_mutex must not be held, as cpuset_attach() will take it.
Cliff Wickman956db3c2008-02-07 00:14:43 -08001929 *
1930 * The cgroup_scan_tasks() function will scan all the tasks in a cgroup,
1931 * calling callback functions for each.
1932 */
1933static void move_member_tasks_to_cpuset(struct cpuset *from, struct cpuset *to)
1934{
1935 struct cpuset_hotplug_scanner scan;
1936
1937 scan.scan.cg = from->css.cgroup;
1938 scan.scan.test_task = NULL; /* select all tasks in cgroup */
1939 scan.scan.process_task = cpuset_do_move_task;
1940 scan.scan.heap = NULL;
1941 scan.to = to->css.cgroup;
1942
Lai Jiangshanda5ef6b2008-07-25 01:47:25 -07001943 if (cgroup_scan_tasks(&scan.scan))
Cliff Wickman956db3c2008-02-07 00:14:43 -08001944 printk(KERN_ERR "move_member_tasks_to_cpuset: "
1945 "cgroup_scan_tasks failed\n");
1946}
1947
Paul Jacksonb1aac8b2006-09-29 02:01:17 -07001948/*
Max Krasnyanskycf417142008-08-11 14:33:53 -07001949 * If CPU and/or memory hotplug handlers, below, unplug any CPUs
Paul Jacksonb1aac8b2006-09-29 02:01:17 -07001950 * or memory nodes, we need to walk over the cpuset hierarchy,
1951 * removing that CPU or node from all cpusets. If this removes the
Cliff Wickman956db3c2008-02-07 00:14:43 -08001952 * last CPU or node from a cpuset, then move the tasks in the empty
1953 * cpuset to its next-highest non-empty parent.
Paul Jacksonb1aac8b2006-09-29 02:01:17 -07001954 *
Paul Jacksonc8d9c902008-02-07 00:14:46 -08001955 * Called with cgroup_mutex held
1956 * callback_mutex must not be held, as cpuset_attach() will take it.
Paul Jacksonb1aac8b2006-09-29 02:01:17 -07001957 */
Cliff Wickman956db3c2008-02-07 00:14:43 -08001958static void remove_tasks_in_empty_cpuset(struct cpuset *cs)
Paul Jacksonb1aac8b2006-09-29 02:01:17 -07001959{
Cliff Wickman956db3c2008-02-07 00:14:43 -08001960 struct cpuset *parent;
Paul Jacksonb1aac8b2006-09-29 02:01:17 -07001961
Paul Jacksonc8d9c902008-02-07 00:14:46 -08001962 /*
1963 * The cgroup's css_sets list is in use if there are tasks
1964 * in the cpuset; the list is empty if there are none;
1965 * the cs->css.refcnt seems always 0.
1966 */
Cliff Wickman956db3c2008-02-07 00:14:43 -08001967 if (list_empty(&cs->css.cgroup->css_sets))
1968 return;
1969
1970 /*
1971 * Find its next-highest non-empty parent, (top cpuset
1972 * has online cpus, so can't be empty).
1973 */
1974 parent = cs->parent;
Li Zefan300ed6c2009-01-07 18:08:44 -08001975 while (cpumask_empty(parent->cpus_allowed) ||
Paul Jacksonb4501292008-02-07 00:14:47 -08001976 nodes_empty(parent->mems_allowed))
Cliff Wickman956db3c2008-02-07 00:14:43 -08001977 parent = parent->parent;
Cliff Wickman956db3c2008-02-07 00:14:43 -08001978
1979 move_member_tasks_to_cpuset(cs, parent);
1980}
1981
1982/*
1983 * Walk the specified cpuset subtree and look for empty cpusets.
1984 * The tasks of such cpuset must be moved to a parent cpuset.
1985 *
Paul Menage2df167a2008-02-07 00:14:45 -08001986 * Called with cgroup_mutex held. We take callback_mutex to modify
Cliff Wickman956db3c2008-02-07 00:14:43 -08001987 * cpus_allowed and mems_allowed.
1988 *
1989 * This walk processes the tree from top to bottom, completing one layer
1990 * before dropping down to the next. It always processes a node before
1991 * any of its children.
1992 *
1993 * For now, since we lack memory hot unplug, we'll never see a cpuset
1994 * that has tasks along with an empty 'mems'. But if we did see such
1995 * a cpuset, we'd handle it just like we do if its 'cpus' was empty.
1996 */
Frederic Weisbeckerd294eb82008-10-03 12:10:10 +02001997static void scan_for_empty_cpusets(struct cpuset *root)
Cliff Wickman956db3c2008-02-07 00:14:43 -08001998{
Li Zefan8d1e6262008-07-29 22:33:21 -07001999 LIST_HEAD(queue);
Cliff Wickman956db3c2008-02-07 00:14:43 -08002000 struct cpuset *cp; /* scans cpusets being updated */
2001 struct cpuset *child; /* scans child cpusets of cp */
Cliff Wickman956db3c2008-02-07 00:14:43 -08002002 struct cgroup *cont;
Miao Xief9b4fb82008-07-25 01:47:22 -07002003 nodemask_t oldmems;
Cliff Wickman956db3c2008-02-07 00:14:43 -08002004
Cliff Wickman956db3c2008-02-07 00:14:43 -08002005 list_add_tail((struct list_head *)&root->stack_list, &queue);
2006
Cliff Wickman956db3c2008-02-07 00:14:43 -08002007 while (!list_empty(&queue)) {
Li Zefan8d1e6262008-07-29 22:33:21 -07002008 cp = list_first_entry(&queue, struct cpuset, stack_list);
Cliff Wickman956db3c2008-02-07 00:14:43 -08002009 list_del(queue.next);
2010 list_for_each_entry(cont, &cp->css.cgroup->children, sibling) {
2011 child = cgroup_cs(cont);
2012 list_add_tail(&child->stack_list, &queue);
2013 }
Paul Jacksonb4501292008-02-07 00:14:47 -08002014
2015 /* Continue past cpusets with all cpus, mems online */
Li Zefan300ed6c2009-01-07 18:08:44 -08002016 if (cpumask_subset(cp->cpus_allowed, cpu_online_mask) &&
Paul Jacksonb4501292008-02-07 00:14:47 -08002017 nodes_subset(cp->mems_allowed, node_states[N_HIGH_MEMORY]))
2018 continue;
2019
Miao Xief9b4fb82008-07-25 01:47:22 -07002020 oldmems = cp->mems_allowed;
2021
Cliff Wickman956db3c2008-02-07 00:14:43 -08002022 /* Remove offline cpus and mems from this cpuset. */
Paul Jacksonb4501292008-02-07 00:14:47 -08002023 mutex_lock(&callback_mutex);
Li Zefan300ed6c2009-01-07 18:08:44 -08002024 cpumask_and(cp->cpus_allowed, cp->cpus_allowed,
2025 cpu_online_mask);
Cliff Wickman956db3c2008-02-07 00:14:43 -08002026 nodes_and(cp->mems_allowed, cp->mems_allowed,
2027 node_states[N_HIGH_MEMORY]);
Paul Jacksonb4501292008-02-07 00:14:47 -08002028 mutex_unlock(&callback_mutex);
2029
2030 /* Move tasks from the empty cpuset to a parent */
Li Zefan300ed6c2009-01-07 18:08:44 -08002031 if (cpumask_empty(cp->cpus_allowed) ||
Paul Jacksonb4501292008-02-07 00:14:47 -08002032 nodes_empty(cp->mems_allowed))
Cliff Wickman956db3c2008-02-07 00:14:43 -08002033 remove_tasks_in_empty_cpuset(cp);
Miao Xief9b4fb82008-07-25 01:47:22 -07002034 else {
Li Zefan4e743392008-09-13 02:33:08 -07002035 update_tasks_cpumask(cp, NULL);
Miao Xief9b4fb82008-07-25 01:47:22 -07002036 update_tasks_nodemask(cp, &oldmems);
2037 }
Cliff Wickman956db3c2008-02-07 00:14:43 -08002038 }
Paul Jacksonb1aac8b2006-09-29 02:01:17 -07002039}
2040
2041/*
Paul Jackson4c4d50f2006-08-27 01:23:51 -07002042 * The top_cpuset tracks what CPUs and Memory Nodes are online,
2043 * period. This is necessary in order to make cpusets transparent
2044 * (of no affect) on systems that are actively using CPU hotplug
2045 * but making no active use of cpusets.
2046 *
Paul Jackson38837fc2006-09-29 02:01:16 -07002047 * This routine ensures that top_cpuset.cpus_allowed tracks
2048 * cpu_online_map on each CPU hotplug (cpuhp) event.
Max Krasnyanskycf417142008-08-11 14:33:53 -07002049 *
2050 * Called within get_online_cpus(). Needs to call cgroup_lock()
2051 * before calling generate_sched_domains().
Paul Jackson4c4d50f2006-08-27 01:23:51 -07002052 */
Max Krasnyanskycf417142008-08-11 14:33:53 -07002053static int cpuset_track_online_cpus(struct notifier_block *unused_nb,
Paul Jackson029190c2007-10-18 23:40:20 -07002054 unsigned long phase, void *unused_cpu)
Paul Jackson4c4d50f2006-08-27 01:23:51 -07002055{
Max Krasnyanskycf417142008-08-11 14:33:53 -07002056 struct sched_domain_attr *attr;
Li Zefan6af866a2009-01-07 18:08:45 -08002057 struct cpumask *doms;
Max Krasnyanskycf417142008-08-11 14:33:53 -07002058 int ndoms;
2059
Dmitry Adamushko3e840502008-07-13 02:10:29 +02002060 switch (phase) {
Dmitry Adamushko3e840502008-07-13 02:10:29 +02002061 case CPU_ONLINE:
2062 case CPU_ONLINE_FROZEN:
2063 case CPU_DEAD:
2064 case CPU_DEAD_FROZEN:
Dmitry Adamushko3e840502008-07-13 02:10:29 +02002065 break;
Max Krasnyanskycf417142008-08-11 14:33:53 -07002066
Dmitry Adamushko3e840502008-07-13 02:10:29 +02002067 default:
Avi Kivityac076752007-05-24 12:33:15 +03002068 return NOTIFY_DONE;
Dmitry Adamushko3e840502008-07-13 02:10:29 +02002069 }
Avi Kivityac076752007-05-24 12:33:15 +03002070
Max Krasnyanskycf417142008-08-11 14:33:53 -07002071 cgroup_lock();
Li Zefan300ed6c2009-01-07 18:08:44 -08002072 cpumask_copy(top_cpuset.cpus_allowed, cpu_online_mask);
Max Krasnyanskycf417142008-08-11 14:33:53 -07002073 scan_for_empty_cpusets(&top_cpuset);
2074 ndoms = generate_sched_domains(&doms, &attr);
2075 cgroup_unlock();
2076
2077 /* Have scheduler rebuild the domains */
2078 partition_sched_domains(ndoms, doms, attr);
2079
Dmitry Adamushko3e840502008-07-13 02:10:29 +02002080 return NOTIFY_OK;
Paul Jackson4c4d50f2006-08-27 01:23:51 -07002081}
Paul Jackson4c4d50f2006-08-27 01:23:51 -07002082
Paul Jacksonb1aac8b2006-09-29 02:01:17 -07002083#ifdef CONFIG_MEMORY_HOTPLUG
Paul Jackson38837fc2006-09-29 02:01:16 -07002084/*
Christoph Lameter0e1e7c72007-10-16 01:25:38 -07002085 * Keep top_cpuset.mems_allowed tracking node_states[N_HIGH_MEMORY].
Max Krasnyanskycf417142008-08-11 14:33:53 -07002086 * Call this routine anytime after node_states[N_HIGH_MEMORY] changes.
2087 * See also the previous routine cpuset_track_online_cpus().
Paul Jackson38837fc2006-09-29 02:01:16 -07002088 */
Miao Xief4818912008-11-19 15:36:30 -08002089static int cpuset_track_online_nodes(struct notifier_block *self,
2090 unsigned long action, void *arg)
Paul Jackson38837fc2006-09-29 02:01:16 -07002091{
Max Krasnyanskycf417142008-08-11 14:33:53 -07002092 cgroup_lock();
Miao Xief4818912008-11-19 15:36:30 -08002093 switch (action) {
2094 case MEM_ONLINE:
2095 top_cpuset.mems_allowed = node_states[N_HIGH_MEMORY];
2096 break;
2097 case MEM_OFFLINE:
2098 top_cpuset.mems_allowed = node_states[N_HIGH_MEMORY];
2099 scan_for_empty_cpusets(&top_cpuset);
2100 break;
2101 default:
2102 break;
2103 }
Max Krasnyanskycf417142008-08-11 14:33:53 -07002104 cgroup_unlock();
Miao Xief4818912008-11-19 15:36:30 -08002105 return NOTIFY_OK;
Paul Jackson38837fc2006-09-29 02:01:16 -07002106}
2107#endif
2108
Linus Torvalds1da177e2005-04-16 15:20:36 -07002109/**
2110 * cpuset_init_smp - initialize cpus_allowed
2111 *
2112 * Description: Finish top cpuset after cpu, node maps are initialized
2113 **/
2114
2115void __init cpuset_init_smp(void)
2116{
Li Zefan300ed6c2009-01-07 18:08:44 -08002117 cpumask_copy(top_cpuset.cpus_allowed, cpu_online_mask);
Christoph Lameter0e1e7c72007-10-16 01:25:38 -07002118 top_cpuset.mems_allowed = node_states[N_HIGH_MEMORY];
Paul Jackson4c4d50f2006-08-27 01:23:51 -07002119
Max Krasnyanskycf417142008-08-11 14:33:53 -07002120 hotcpu_notifier(cpuset_track_online_cpus, 0);
Miao Xief4818912008-11-19 15:36:30 -08002121 hotplug_memory_notifier(cpuset_track_online_nodes, 10);
Miao Xief90d4112009-01-16 10:24:10 +08002122
2123 cpuset_wq = create_singlethread_workqueue("cpuset");
2124 BUG_ON(!cpuset_wq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002125}
2126
2127/**
Linus Torvalds1da177e2005-04-16 15:20:36 -07002128 * cpuset_cpus_allowed - return cpus_allowed mask from a tasks cpuset.
2129 * @tsk: pointer to task_struct from which to obtain cpuset->cpus_allowed.
Li Zefan6af866a2009-01-07 18:08:45 -08002130 * @pmask: pointer to struct cpumask variable to receive cpus_allowed set.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002131 *
Li Zefan300ed6c2009-01-07 18:08:44 -08002132 * Description: Returns the cpumask_var_t cpus_allowed of the cpuset
Linus Torvalds1da177e2005-04-16 15:20:36 -07002133 * attached to the specified @tsk. Guaranteed to return some non-empty
2134 * subset of cpu_online_map, even if this means going outside the
2135 * tasks cpuset.
2136 **/
2137
Li Zefan6af866a2009-01-07 18:08:45 -08002138void cpuset_cpus_allowed(struct task_struct *tsk, struct cpumask *pmask)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002139{
Ingo Molnar3d3f26a2006-03-23 03:00:18 -08002140 mutex_lock(&callback_mutex);
Mike Travisf9a86fc2008-04-04 18:11:07 -07002141 cpuset_cpus_allowed_locked(tsk, pmask);
Cliff Wickman470fd642007-10-18 23:40:46 -07002142 mutex_unlock(&callback_mutex);
Cliff Wickman470fd642007-10-18 23:40:46 -07002143}
2144
2145/**
2146 * cpuset_cpus_allowed_locked - return cpus_allowed mask from a tasks cpuset.
Paul Menage2df167a2008-02-07 00:14:45 -08002147 * Must be called with callback_mutex held.
Cliff Wickman470fd642007-10-18 23:40:46 -07002148 **/
Li Zefan6af866a2009-01-07 18:08:45 -08002149void cpuset_cpus_allowed_locked(struct task_struct *tsk, struct cpumask *pmask)
Cliff Wickman470fd642007-10-18 23:40:46 -07002150{
Paul Jackson909d75a2006-01-08 01:01:55 -08002151 task_lock(tsk);
Mike Travisf9a86fc2008-04-04 18:11:07 -07002152 guarantee_online_cpus(task_cs(tsk), pmask);
Paul Jackson909d75a2006-01-08 01:01:55 -08002153 task_unlock(tsk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002154}
2155
2156void cpuset_init_current_mems_allowed(void)
2157{
Mike Travisf9a86fc2008-04-04 18:11:07 -07002158 nodes_setall(current->mems_allowed);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002159}
2160
Randy Dunlapd9fd8a62005-07-27 11:45:11 -07002161/**
Paul Jackson909d75a2006-01-08 01:01:55 -08002162 * cpuset_mems_allowed - return mems_allowed mask from a tasks cpuset.
2163 * @tsk: pointer to task_struct from which to obtain cpuset->mems_allowed.
2164 *
2165 * Description: Returns the nodemask_t mems_allowed of the cpuset
2166 * attached to the specified @tsk. Guaranteed to return some non-empty
Christoph Lameter0e1e7c72007-10-16 01:25:38 -07002167 * subset of node_states[N_HIGH_MEMORY], even if this means going outside the
Paul Jackson909d75a2006-01-08 01:01:55 -08002168 * tasks cpuset.
2169 **/
2170
2171nodemask_t cpuset_mems_allowed(struct task_struct *tsk)
2172{
2173 nodemask_t mask;
2174
Ingo Molnar3d3f26a2006-03-23 03:00:18 -08002175 mutex_lock(&callback_mutex);
Paul Jackson909d75a2006-01-08 01:01:55 -08002176 task_lock(tsk);
Paul Menage8793d852007-10-18 23:39:39 -07002177 guarantee_online_mems(task_cs(tsk), &mask);
Paul Jackson909d75a2006-01-08 01:01:55 -08002178 task_unlock(tsk);
Ingo Molnar3d3f26a2006-03-23 03:00:18 -08002179 mutex_unlock(&callback_mutex);
Paul Jackson909d75a2006-01-08 01:01:55 -08002180
2181 return mask;
2182}
2183
2184/**
Mel Gorman19770b32008-04-28 02:12:18 -07002185 * cpuset_nodemask_valid_mems_allowed - check nodemask vs. curremt mems_allowed
2186 * @nodemask: the nodemask to be checked
Randy Dunlapd9fd8a62005-07-27 11:45:11 -07002187 *
Mel Gorman19770b32008-04-28 02:12:18 -07002188 * Are any of the nodes in the nodemask allowed in current->mems_allowed?
Linus Torvalds1da177e2005-04-16 15:20:36 -07002189 */
Mel Gorman19770b32008-04-28 02:12:18 -07002190int cpuset_nodemask_valid_mems_allowed(nodemask_t *nodemask)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002191{
Mel Gorman19770b32008-04-28 02:12:18 -07002192 return nodes_intersects(*nodemask, current->mems_allowed);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002193}
2194
Paul Jackson9bf22292005-09-06 15:18:12 -07002195/*
Paul Menage78608362008-04-29 01:00:26 -07002196 * nearest_hardwall_ancestor() - Returns the nearest mem_exclusive or
2197 * mem_hardwall ancestor to the specified cpuset. Call holding
2198 * callback_mutex. If no ancestor is mem_exclusive or mem_hardwall
2199 * (an unusual configuration), then returns the root cpuset.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002200 */
Paul Menage78608362008-04-29 01:00:26 -07002201static const struct cpuset *nearest_hardwall_ancestor(const struct cpuset *cs)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002202{
Paul Menage78608362008-04-29 01:00:26 -07002203 while (!(is_mem_exclusive(cs) || is_mem_hardwall(cs)) && cs->parent)
Paul Jackson9bf22292005-09-06 15:18:12 -07002204 cs = cs->parent;
2205 return cs;
2206}
2207
2208/**
Paul Jackson02a0e532006-12-13 00:34:25 -08002209 * cpuset_zone_allowed_softwall - Can we allocate on zone z's memory node?
Paul Jackson9bf22292005-09-06 15:18:12 -07002210 * @z: is this zone on an allowed node?
Paul Jackson02a0e532006-12-13 00:34:25 -08002211 * @gfp_mask: memory allocation flags
Paul Jackson9bf22292005-09-06 15:18:12 -07002212 *
Paul Jackson02a0e532006-12-13 00:34:25 -08002213 * If we're in interrupt, yes, we can always allocate. If
2214 * __GFP_THISNODE is set, yes, we can always allocate. If zone
Paul Jackson9bf22292005-09-06 15:18:12 -07002215 * z's node is in our tasks mems_allowed, yes. If it's not a
2216 * __GFP_HARDWALL request and this zone's nodes is in the nearest
Paul Menage78608362008-04-29 01:00:26 -07002217 * hardwalled cpuset ancestor to this tasks cpuset, yes.
David Rientjesc596d9f2007-05-06 14:49:32 -07002218 * If the task has been OOM killed and has access to memory reserves
2219 * as specified by the TIF_MEMDIE flag, yes.
Paul Jackson9bf22292005-09-06 15:18:12 -07002220 * Otherwise, no.
2221 *
Paul Jackson02a0e532006-12-13 00:34:25 -08002222 * If __GFP_HARDWALL is set, cpuset_zone_allowed_softwall()
2223 * reduces to cpuset_zone_allowed_hardwall(). Otherwise,
2224 * cpuset_zone_allowed_softwall() might sleep, and might allow a zone
2225 * from an enclosing cpuset.
2226 *
2227 * cpuset_zone_allowed_hardwall() only handles the simpler case of
2228 * hardwall cpusets, and never sleeps.
2229 *
2230 * The __GFP_THISNODE placement logic is really handled elsewhere,
2231 * by forcibly using a zonelist starting at a specified node, and by
2232 * (in get_page_from_freelist()) refusing to consider the zones for
2233 * any node on the zonelist except the first. By the time any such
2234 * calls get to this routine, we should just shut up and say 'yes'.
2235 *
Paul Jackson9bf22292005-09-06 15:18:12 -07002236 * GFP_USER allocations are marked with the __GFP_HARDWALL bit,
David Rientjesc596d9f2007-05-06 14:49:32 -07002237 * and do not allow allocations outside the current tasks cpuset
2238 * unless the task has been OOM killed as is marked TIF_MEMDIE.
Paul Jackson9bf22292005-09-06 15:18:12 -07002239 * GFP_KERNEL allocations are not so marked, so can escape to the
Paul Menage78608362008-04-29 01:00:26 -07002240 * nearest enclosing hardwalled ancestor cpuset.
Paul Jackson9bf22292005-09-06 15:18:12 -07002241 *
Paul Jackson02a0e532006-12-13 00:34:25 -08002242 * Scanning up parent cpusets requires callback_mutex. The
2243 * __alloc_pages() routine only calls here with __GFP_HARDWALL bit
2244 * _not_ set if it's a GFP_KERNEL allocation, and all nodes in the
2245 * current tasks mems_allowed came up empty on the first pass over
2246 * the zonelist. So only GFP_KERNEL allocations, if all nodes in the
2247 * cpuset are short of memory, might require taking the callback_mutex
2248 * mutex.
Paul Jackson9bf22292005-09-06 15:18:12 -07002249 *
Paul Jackson36be57f2006-05-20 15:00:10 -07002250 * The first call here from mm/page_alloc:get_page_from_freelist()
Paul Jackson02a0e532006-12-13 00:34:25 -08002251 * has __GFP_HARDWALL set in gfp_mask, enforcing hardwall cpusets,
2252 * so no allocation on a node outside the cpuset is allowed (unless
2253 * in interrupt, of course).
Paul Jackson9bf22292005-09-06 15:18:12 -07002254 *
Paul Jackson36be57f2006-05-20 15:00:10 -07002255 * The second pass through get_page_from_freelist() doesn't even call
2256 * here for GFP_ATOMIC calls. For those calls, the __alloc_pages()
2257 * variable 'wait' is not set, and the bit ALLOC_CPUSET is not set
2258 * in alloc_flags. That logic and the checks below have the combined
2259 * affect that:
Paul Jackson9bf22292005-09-06 15:18:12 -07002260 * in_interrupt - any node ok (current task context irrelevant)
2261 * GFP_ATOMIC - any node ok
David Rientjesc596d9f2007-05-06 14:49:32 -07002262 * TIF_MEMDIE - any node ok
Paul Menage78608362008-04-29 01:00:26 -07002263 * GFP_KERNEL - any node in enclosing hardwalled cpuset ok
Paul Jackson9bf22292005-09-06 15:18:12 -07002264 * GFP_USER - only nodes in current tasks mems allowed ok.
Paul Jackson36be57f2006-05-20 15:00:10 -07002265 *
2266 * Rule:
Paul Jackson02a0e532006-12-13 00:34:25 -08002267 * Don't call cpuset_zone_allowed_softwall if you can't sleep, unless you
Paul Jackson36be57f2006-05-20 15:00:10 -07002268 * pass in the __GFP_HARDWALL flag set in gfp_flag, which disables
2269 * the code that might scan up ancestor cpusets and sleep.
Paul Jackson02a0e532006-12-13 00:34:25 -08002270 */
Paul Jackson9bf22292005-09-06 15:18:12 -07002271
Paul Jackson02a0e532006-12-13 00:34:25 -08002272int __cpuset_zone_allowed_softwall(struct zone *z, gfp_t gfp_mask)
Paul Jackson9bf22292005-09-06 15:18:12 -07002273{
2274 int node; /* node that zone z is on */
2275 const struct cpuset *cs; /* current cpuset ancestors */
Paul Jackson29afd492006-03-24 03:16:12 -08002276 int allowed; /* is allocation in zone z allowed? */
Paul Jackson9bf22292005-09-06 15:18:12 -07002277
Christoph Lameter9b819d22006-09-25 23:31:40 -07002278 if (in_interrupt() || (gfp_mask & __GFP_THISNODE))
Paul Jackson9bf22292005-09-06 15:18:12 -07002279 return 1;
Christoph Lameter89fa3022006-09-25 23:31:55 -07002280 node = zone_to_nid(z);
Paul Jackson92d1dbd2006-05-20 15:00:11 -07002281 might_sleep_if(!(gfp_mask & __GFP_HARDWALL));
Paul Jackson9bf22292005-09-06 15:18:12 -07002282 if (node_isset(node, current->mems_allowed))
2283 return 1;
David Rientjesc596d9f2007-05-06 14:49:32 -07002284 /*
2285 * Allow tasks that have access to memory reserves because they have
2286 * been OOM killed to get memory anywhere.
2287 */
2288 if (unlikely(test_thread_flag(TIF_MEMDIE)))
2289 return 1;
Paul Jackson9bf22292005-09-06 15:18:12 -07002290 if (gfp_mask & __GFP_HARDWALL) /* If hardwall request, stop here */
2291 return 0;
2292
Bob Picco5563e772005-11-13 16:06:35 -08002293 if (current->flags & PF_EXITING) /* Let dying task have memory */
2294 return 1;
2295
Paul Jackson9bf22292005-09-06 15:18:12 -07002296 /* Not hardwall and node outside mems_allowed: scan up cpusets */
Ingo Molnar3d3f26a2006-03-23 03:00:18 -08002297 mutex_lock(&callback_mutex);
Paul Jackson053199e2005-10-30 15:02:30 -08002298
Paul Jackson053199e2005-10-30 15:02:30 -08002299 task_lock(current);
Paul Menage78608362008-04-29 01:00:26 -07002300 cs = nearest_hardwall_ancestor(task_cs(current));
Paul Jackson053199e2005-10-30 15:02:30 -08002301 task_unlock(current);
2302
Paul Jackson9bf22292005-09-06 15:18:12 -07002303 allowed = node_isset(node, cs->mems_allowed);
Ingo Molnar3d3f26a2006-03-23 03:00:18 -08002304 mutex_unlock(&callback_mutex);
Paul Jackson9bf22292005-09-06 15:18:12 -07002305 return allowed;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002306}
2307
Paul Jackson02a0e532006-12-13 00:34:25 -08002308/*
2309 * cpuset_zone_allowed_hardwall - Can we allocate on zone z's memory node?
2310 * @z: is this zone on an allowed node?
2311 * @gfp_mask: memory allocation flags
2312 *
2313 * If we're in interrupt, yes, we can always allocate.
2314 * If __GFP_THISNODE is set, yes, we can always allocate. If zone
David Rientjesc596d9f2007-05-06 14:49:32 -07002315 * z's node is in our tasks mems_allowed, yes. If the task has been
2316 * OOM killed and has access to memory reserves as specified by the
2317 * TIF_MEMDIE flag, yes. Otherwise, no.
Paul Jackson02a0e532006-12-13 00:34:25 -08002318 *
2319 * The __GFP_THISNODE placement logic is really handled elsewhere,
2320 * by forcibly using a zonelist starting at a specified node, and by
2321 * (in get_page_from_freelist()) refusing to consider the zones for
2322 * any node on the zonelist except the first. By the time any such
2323 * calls get to this routine, we should just shut up and say 'yes'.
2324 *
2325 * Unlike the cpuset_zone_allowed_softwall() variant, above,
2326 * this variant requires that the zone be in the current tasks
2327 * mems_allowed or that we're in interrupt. It does not scan up the
2328 * cpuset hierarchy for the nearest enclosing mem_exclusive cpuset.
2329 * It never sleeps.
2330 */
2331
2332int __cpuset_zone_allowed_hardwall(struct zone *z, gfp_t gfp_mask)
2333{
2334 int node; /* node that zone z is on */
2335
2336 if (in_interrupt() || (gfp_mask & __GFP_THISNODE))
2337 return 1;
2338 node = zone_to_nid(z);
2339 if (node_isset(node, current->mems_allowed))
2340 return 1;
Daniel Walkerdedf8b72007-10-18 03:06:04 -07002341 /*
2342 * Allow tasks that have access to memory reserves because they have
2343 * been OOM killed to get memory anywhere.
2344 */
2345 if (unlikely(test_thread_flag(TIF_MEMDIE)))
2346 return 1;
Paul Jackson02a0e532006-12-13 00:34:25 -08002347 return 0;
2348}
2349
Paul Jacksonef08e3b2005-09-06 15:18:13 -07002350/**
Paul Jackson505970b2006-01-14 13:21:06 -08002351 * cpuset_lock - lock out any changes to cpuset structures
2352 *
Ingo Molnar3d3f26a2006-03-23 03:00:18 -08002353 * The out of memory (oom) code needs to mutex_lock cpusets
Paul Jackson505970b2006-01-14 13:21:06 -08002354 * from being changed while it scans the tasklist looking for a
Ingo Molnar3d3f26a2006-03-23 03:00:18 -08002355 * task in an overlapping cpuset. Expose callback_mutex via this
Paul Jackson505970b2006-01-14 13:21:06 -08002356 * cpuset_lock() routine, so the oom code can lock it, before
2357 * locking the task list. The tasklist_lock is a spinlock, so
Ingo Molnar3d3f26a2006-03-23 03:00:18 -08002358 * must be taken inside callback_mutex.
Paul Jackson505970b2006-01-14 13:21:06 -08002359 */
2360
2361void cpuset_lock(void)
2362{
Ingo Molnar3d3f26a2006-03-23 03:00:18 -08002363 mutex_lock(&callback_mutex);
Paul Jackson505970b2006-01-14 13:21:06 -08002364}
2365
2366/**
2367 * cpuset_unlock - release lock on cpuset changes
2368 *
2369 * Undo the lock taken in a previous cpuset_lock() call.
2370 */
2371
2372void cpuset_unlock(void)
2373{
Ingo Molnar3d3f26a2006-03-23 03:00:18 -08002374 mutex_unlock(&callback_mutex);
Paul Jackson505970b2006-01-14 13:21:06 -08002375}
2376
2377/**
Paul Jackson825a46a2006-03-24 03:16:03 -08002378 * cpuset_mem_spread_node() - On which node to begin search for a page
2379 *
2380 * If a task is marked PF_SPREAD_PAGE or PF_SPREAD_SLAB (as for
2381 * tasks in a cpuset with is_spread_page or is_spread_slab set),
2382 * and if the memory allocation used cpuset_mem_spread_node()
2383 * to determine on which node to start looking, as it will for
2384 * certain page cache or slab cache pages such as used for file
2385 * system buffers and inode caches, then instead of starting on the
2386 * local node to look for a free page, rather spread the starting
2387 * node around the tasks mems_allowed nodes.
2388 *
2389 * We don't have to worry about the returned node being offline
2390 * because "it can't happen", and even if it did, it would be ok.
2391 *
2392 * The routines calling guarantee_online_mems() are careful to
2393 * only set nodes in task->mems_allowed that are online. So it
2394 * should not be possible for the following code to return an
2395 * offline node. But if it did, that would be ok, as this routine
2396 * is not returning the node where the allocation must be, only
2397 * the node where the search should start. The zonelist passed to
2398 * __alloc_pages() will include all nodes. If the slab allocator
2399 * is passed an offline node, it will fall back to the local node.
2400 * See kmem_cache_alloc_node().
2401 */
2402
2403int cpuset_mem_spread_node(void)
2404{
2405 int node;
2406
2407 node = next_node(current->cpuset_mem_spread_rotor, current->mems_allowed);
2408 if (node == MAX_NUMNODES)
2409 node = first_node(current->mems_allowed);
2410 current->cpuset_mem_spread_rotor = node;
2411 return node;
2412}
2413EXPORT_SYMBOL_GPL(cpuset_mem_spread_node);
2414
2415/**
David Rientjesbbe373f2007-10-16 23:25:58 -07002416 * cpuset_mems_allowed_intersects - Does @tsk1's mems_allowed intersect @tsk2's?
2417 * @tsk1: pointer to task_struct of some task.
2418 * @tsk2: pointer to task_struct of some other task.
Paul Jacksonef08e3b2005-09-06 15:18:13 -07002419 *
David Rientjesbbe373f2007-10-16 23:25:58 -07002420 * Description: Return true if @tsk1's mems_allowed intersects the
2421 * mems_allowed of @tsk2. Used by the OOM killer to determine if
2422 * one of the task's memory usage might impact the memory available
2423 * to the other.
Paul Jacksonef08e3b2005-09-06 15:18:13 -07002424 **/
2425
David Rientjesbbe373f2007-10-16 23:25:58 -07002426int cpuset_mems_allowed_intersects(const struct task_struct *tsk1,
2427 const struct task_struct *tsk2)
Paul Jacksonef08e3b2005-09-06 15:18:13 -07002428{
David Rientjesbbe373f2007-10-16 23:25:58 -07002429 return nodes_intersects(tsk1->mems_allowed, tsk2->mems_allowed);
Paul Jacksonef08e3b2005-09-06 15:18:13 -07002430}
2431
David Rientjes75aa1992009-01-06 14:39:01 -08002432/**
2433 * cpuset_print_task_mems_allowed - prints task's cpuset and mems_allowed
2434 * @task: pointer to task_struct of some task.
2435 *
2436 * Description: Prints @task's name, cpuset name, and cached copy of its
2437 * mems_allowed to the kernel log. Must hold task_lock(task) to allow
2438 * dereferencing task_cs(task).
2439 */
2440void cpuset_print_task_mems_allowed(struct task_struct *tsk)
2441{
2442 struct dentry *dentry;
2443
2444 dentry = task_cs(tsk)->css.cgroup->dentry;
2445 spin_lock(&cpuset_buffer_lock);
2446 snprintf(cpuset_name, CPUSET_NAME_LEN,
2447 dentry ? (const char *)dentry->d_name.name : "/");
2448 nodelist_scnprintf(cpuset_nodelist, CPUSET_NODELIST_LEN,
2449 tsk->mems_allowed);
2450 printk(KERN_INFO "%s cpuset=%s mems_allowed=%s\n",
2451 tsk->comm, cpuset_name, cpuset_nodelist);
2452 spin_unlock(&cpuset_buffer_lock);
2453}
2454
Linus Torvalds1da177e2005-04-16 15:20:36 -07002455/*
Paul Jackson3e0d98b2006-01-08 01:01:49 -08002456 * Collection of memory_pressure is suppressed unless
2457 * this flag is enabled by writing "1" to the special
2458 * cpuset file 'memory_pressure_enabled' in the root cpuset.
2459 */
2460
Paul Jacksonc5b2aff2006-01-08 01:01:51 -08002461int cpuset_memory_pressure_enabled __read_mostly;
Paul Jackson3e0d98b2006-01-08 01:01:49 -08002462
2463/**
2464 * cpuset_memory_pressure_bump - keep stats of per-cpuset reclaims.
2465 *
2466 * Keep a running average of the rate of synchronous (direct)
2467 * page reclaim efforts initiated by tasks in each cpuset.
2468 *
2469 * This represents the rate at which some task in the cpuset
2470 * ran low on memory on all nodes it was allowed to use, and
2471 * had to enter the kernels page reclaim code in an effort to
2472 * create more free memory by tossing clean pages or swapping
2473 * or writing dirty pages.
2474 *
2475 * Display to user space in the per-cpuset read-only file
2476 * "memory_pressure". Value displayed is an integer
2477 * representing the recent rate of entry into the synchronous
2478 * (direct) page reclaim by any task attached to the cpuset.
2479 **/
2480
2481void __cpuset_memory_pressure_bump(void)
2482{
Paul Jackson3e0d98b2006-01-08 01:01:49 -08002483 task_lock(current);
Paul Menage8793d852007-10-18 23:39:39 -07002484 fmeter_markevent(&task_cs(current)->fmeter);
Paul Jackson3e0d98b2006-01-08 01:01:49 -08002485 task_unlock(current);
2486}
2487
Paul Menage8793d852007-10-18 23:39:39 -07002488#ifdef CONFIG_PROC_PID_CPUSET
Paul Jackson3e0d98b2006-01-08 01:01:49 -08002489/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07002490 * proc_cpuset_show()
2491 * - Print tasks cpuset path into seq_file.
2492 * - Used for /proc/<pid>/cpuset.
Paul Jackson053199e2005-10-30 15:02:30 -08002493 * - No need to task_lock(tsk) on this tsk->cpuset reference, as it
2494 * doesn't really matter if tsk->cpuset changes after we read it,
Paul Jacksonc8d9c902008-02-07 00:14:46 -08002495 * and we take cgroup_mutex, keeping cpuset_attach() from changing it
Paul Menage2df167a2008-02-07 00:14:45 -08002496 * anyway.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002497 */
Paul Jackson029190c2007-10-18 23:40:20 -07002498static int proc_cpuset_show(struct seq_file *m, void *unused_v)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002499{
Eric W. Biederman13b41b02006-06-26 00:25:56 -07002500 struct pid *pid;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002501 struct task_struct *tsk;
2502 char *buf;
Paul Menage8793d852007-10-18 23:39:39 -07002503 struct cgroup_subsys_state *css;
Eric W. Biederman99f89552006-06-26 00:25:55 -07002504 int retval;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002505
Eric W. Biederman99f89552006-06-26 00:25:55 -07002506 retval = -ENOMEM;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002507 buf = kmalloc(PAGE_SIZE, GFP_KERNEL);
2508 if (!buf)
Eric W. Biederman99f89552006-06-26 00:25:55 -07002509 goto out;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002510
Eric W. Biederman99f89552006-06-26 00:25:55 -07002511 retval = -ESRCH;
Eric W. Biederman13b41b02006-06-26 00:25:56 -07002512 pid = m->private;
2513 tsk = get_pid_task(pid, PIDTYPE_PID);
Eric W. Biederman99f89552006-06-26 00:25:55 -07002514 if (!tsk)
2515 goto out_free;
2516
2517 retval = -EINVAL;
Paul Menage8793d852007-10-18 23:39:39 -07002518 cgroup_lock();
2519 css = task_subsys_state(tsk, cpuset_subsys_id);
2520 retval = cgroup_path(css->cgroup, buf, PAGE_SIZE);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002521 if (retval < 0)
Eric W. Biederman99f89552006-06-26 00:25:55 -07002522 goto out_unlock;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002523 seq_puts(m, buf);
2524 seq_putc(m, '\n');
Eric W. Biederman99f89552006-06-26 00:25:55 -07002525out_unlock:
Paul Menage8793d852007-10-18 23:39:39 -07002526 cgroup_unlock();
Eric W. Biederman99f89552006-06-26 00:25:55 -07002527 put_task_struct(tsk);
2528out_free:
Linus Torvalds1da177e2005-04-16 15:20:36 -07002529 kfree(buf);
Eric W. Biederman99f89552006-06-26 00:25:55 -07002530out:
Linus Torvalds1da177e2005-04-16 15:20:36 -07002531 return retval;
2532}
2533
2534static int cpuset_open(struct inode *inode, struct file *file)
2535{
Eric W. Biederman13b41b02006-06-26 00:25:56 -07002536 struct pid *pid = PROC_I(inode)->pid;
2537 return single_open(file, proc_cpuset_show, pid);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002538}
2539
Arjan van de Ven9a321442007-02-12 00:55:35 -08002540const struct file_operations proc_cpuset_operations = {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002541 .open = cpuset_open,
2542 .read = seq_read,
2543 .llseek = seq_lseek,
2544 .release = single_release,
2545};
Paul Menage8793d852007-10-18 23:39:39 -07002546#endif /* CONFIG_PROC_PID_CPUSET */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002547
2548/* Display task cpus_allowed, mems_allowed in /proc/<pid>/status file. */
Eric W. Biedermandf5f8312008-02-08 04:18:33 -08002549void cpuset_task_status_allowed(struct seq_file *m, struct task_struct *task)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002550{
Eric W. Biedermandf5f8312008-02-08 04:18:33 -08002551 seq_printf(m, "Cpus_allowed:\t");
Lai Jiangshan30e8e132008-10-18 20:28:20 -07002552 seq_cpumask(m, &task->cpus_allowed);
Eric W. Biedermandf5f8312008-02-08 04:18:33 -08002553 seq_printf(m, "\n");
Mike Travis39106dc2008-04-08 11:43:03 -07002554 seq_printf(m, "Cpus_allowed_list:\t");
Lai Jiangshan30e8e132008-10-18 20:28:20 -07002555 seq_cpumask_list(m, &task->cpus_allowed);
Mike Travis39106dc2008-04-08 11:43:03 -07002556 seq_printf(m, "\n");
Eric W. Biedermandf5f8312008-02-08 04:18:33 -08002557 seq_printf(m, "Mems_allowed:\t");
Lai Jiangshan30e8e132008-10-18 20:28:20 -07002558 seq_nodemask(m, &task->mems_allowed);
Eric W. Biedermandf5f8312008-02-08 04:18:33 -08002559 seq_printf(m, "\n");
Mike Travis39106dc2008-04-08 11:43:03 -07002560 seq_printf(m, "Mems_allowed_list:\t");
Lai Jiangshan30e8e132008-10-18 20:28:20 -07002561 seq_nodemask_list(m, &task->mems_allowed);
Mike Travis39106dc2008-04-08 11:43:03 -07002562 seq_printf(m, "\n");
Linus Torvalds1da177e2005-04-16 15:20:36 -07002563}