4 * Processor and Memory placement constraints for sets of tasks.
6 * Copyright (C) 2003 BULL SA.
7 * Copyright (C) 2004-2007 Silicon Graphics, Inc.
8 * Copyright (C) 2006 Google, Inc
10 * Portions derived from Patrick Mochel's sysfs code.
11 * sysfs is Copyright (c) 2001-3 Patrick Mochel
13 * 2003-10-10 Written by Simon Derr.
14 * 2003-10-22 Updates by Stephen Hemminger.
15 * 2004 May-July Rework by Paul Jackson.
16 * 2006 Rework by Paul Menage to use generic cgroups
18 * This file is subject to the terms and conditions of the GNU General Public
19 * License. See the file COPYING in the main directory of the Linux
20 * distribution for more details.
23 #include <linux/cpu.h>
24 #include <linux/cpumask.h>
25 #include <linux/cpuset.h>
26 #include <linux/err.h>
27 #include <linux/errno.h>
28 #include <linux/file.h>
30 #include <linux/init.h>
31 #include <linux/interrupt.h>
32 #include <linux/kernel.h>
33 #include <linux/kmod.h>
34 #include <linux/list.h>
35 #include <linux/mempolicy.h>
37 #include <linux/module.h>
38 #include <linux/mount.h>
39 #include <linux/namei.h>
40 #include <linux/pagemap.h>
41 #include <linux/proc_fs.h>
42 #include <linux/rcupdate.h>
43 #include <linux/sched.h>
44 #include <linux/seq_file.h>
45 #include <linux/security.h>
46 #include <linux/slab.h>
47 #include <linux/spinlock.h>
48 #include <linux/stat.h>
49 #include <linux/string.h>
50 #include <linux/time.h>
51 #include <linux/backing-dev.h>
52 #include <linux/sort.h>
54 #include <asm/uaccess.h>
55 #include <asm/atomic.h>
56 #include <linux/mutex.h>
57 #include <linux/kfifo.h>
58 #include <linux/workqueue.h>
59 #include <linux/cgroup.h>
62 * Tracks how many cpusets are currently defined in system.
63 * When there is only one cpuset (the root cpuset) we can
64 * short circuit some hooks.
66 int number_of_cpusets __read_mostly;
68 /* Forward declare cgroup structures */
69 struct cgroup_subsys cpuset_subsys;
72 /* See "Frequency meter" comments, below. */
75 int cnt; /* unprocessed events count */
76 int val; /* most recent output value */
77 time_t time; /* clock (secs) when val computed */
78 spinlock_t lock; /* guards read or write of above */
82 struct cgroup_subsys_state css;
84 unsigned long flags; /* "unsigned long" so bitops work */
85 cpumask_t cpus_allowed; /* CPUs allowed to tasks in cpuset */
86 nodemask_t mems_allowed; /* Memory Nodes allowed to tasks */
88 struct cpuset *parent; /* my parent */
91 * Copy of global cpuset_mems_generation as of the most
92 * recent time this cpuset changed its mems_allowed.
96 struct fmeter fmeter; /* memory_pressure filter */
98 /* partition number for rebuild_sched_domains() */
101 /* for custom sched domain */
102 int relax_domain_level;
104 /* used for walking a cpuset heirarchy */
105 struct list_head stack_list;
108 /* Retrieve the cpuset for a cgroup */
109 static inline struct cpuset *cgroup_cs(struct cgroup *cont)
111 return container_of(cgroup_subsys_state(cont, cpuset_subsys_id),
115 /* Retrieve the cpuset for a task */
116 static inline struct cpuset *task_cs(struct task_struct *task)
118 return container_of(task_subsys_state(task, cpuset_subsys_id),
121 struct cpuset_hotplug_scanner {
122 struct cgroup_scanner scan;
126 /* bits in struct cpuset flags field */
131 CS_SCHED_LOAD_BALANCE,
136 /* convenient tests for these bits */
137 static inline int is_cpu_exclusive(const struct cpuset *cs)
139 return test_bit(CS_CPU_EXCLUSIVE, &cs->flags);
142 static inline int is_mem_exclusive(const struct cpuset *cs)
144 return test_bit(CS_MEM_EXCLUSIVE, &cs->flags);
147 static inline int is_sched_load_balance(const struct cpuset *cs)
149 return test_bit(CS_SCHED_LOAD_BALANCE, &cs->flags);
152 static inline int is_memory_migrate(const struct cpuset *cs)
154 return test_bit(CS_MEMORY_MIGRATE, &cs->flags);
157 static inline int is_spread_page(const struct cpuset *cs)
159 return test_bit(CS_SPREAD_PAGE, &cs->flags);
162 static inline int is_spread_slab(const struct cpuset *cs)
164 return test_bit(CS_SPREAD_SLAB, &cs->flags);
168 * Increment this integer everytime any cpuset changes its
169 * mems_allowed value. Users of cpusets can track this generation
170 * number, and avoid having to lock and reload mems_allowed unless
171 * the cpuset they're using changes generation.
173 * A single, global generation is needed because cpuset_attach_task() could
174 * reattach a task to a different cpuset, which must not have its
175 * generation numbers aliased with those of that tasks previous cpuset.
177 * Generations are needed for mems_allowed because one task cannot
178 * modify another's memory placement. So we must enable every task,
179 * on every visit to __alloc_pages(), to efficiently check whether
180 * its current->cpuset->mems_allowed has changed, requiring an update
181 * of its current->mems_allowed.
183 * Since writes to cpuset_mems_generation are guarded by the cgroup lock
184 * there is no need to mark it atomic.
186 static int cpuset_mems_generation;
188 static struct cpuset top_cpuset = {
189 .flags = ((1 << CS_CPU_EXCLUSIVE) | (1 << CS_MEM_EXCLUSIVE)),
190 .cpus_allowed = CPU_MASK_ALL,
191 .mems_allowed = NODE_MASK_ALL,
195 * There are two global mutexes guarding cpuset structures. The first
196 * is the main control groups cgroup_mutex, accessed via
197 * cgroup_lock()/cgroup_unlock(). The second is the cpuset-specific
198 * callback_mutex, below. They can nest. It is ok to first take
199 * cgroup_mutex, then nest callback_mutex. We also require taking
200 * task_lock() when dereferencing a task's cpuset pointer. See "The
201 * task_lock() exception", at the end of this comment.
203 * A task must hold both mutexes to modify cpusets. If a task
204 * holds cgroup_mutex, then it blocks others wanting that mutex,
205 * ensuring that it is the only task able to also acquire callback_mutex
206 * and be able to modify cpusets. It can perform various checks on
207 * the cpuset structure first, knowing nothing will change. It can
208 * also allocate memory while just holding cgroup_mutex. While it is
209 * performing these checks, various callback routines can briefly
210 * acquire callback_mutex to query cpusets. Once it is ready to make
211 * the changes, it takes callback_mutex, blocking everyone else.
213 * Calls to the kernel memory allocator can not be made while holding
214 * callback_mutex, as that would risk double tripping on callback_mutex
215 * from one of the callbacks into the cpuset code from within
218 * If a task is only holding callback_mutex, then it has read-only
221 * The task_struct fields mems_allowed and mems_generation may only
222 * be accessed in the context of that task, so require no locks.
224 * The cpuset_common_file_write handler for operations that modify
225 * the cpuset hierarchy holds cgroup_mutex across the entire operation,
226 * single threading all such cpuset modifications across the system.
228 * The cpuset_common_file_read() handlers only hold callback_mutex across
229 * small pieces of code, such as when reading out possibly multi-word
230 * cpumasks and nodemasks.
232 * Accessing a task's cpuset should be done in accordance with the
233 * guidelines for accessing subsystem state in kernel/cgroup.c
236 static DEFINE_MUTEX(callback_mutex);
238 /* This is ugly, but preserves the userspace API for existing cpuset
239 * users. If someone tries to mount the "cpuset" filesystem, we
240 * silently switch it to mount "cgroup" instead */
241 static int cpuset_get_sb(struct file_system_type *fs_type,
242 int flags, const char *unused_dev_name,
243 void *data, struct vfsmount *mnt)
245 struct file_system_type *cgroup_fs = get_fs_type("cgroup");
250 "release_agent=/sbin/cpuset_release_agent";
251 ret = cgroup_fs->get_sb(cgroup_fs, flags,
252 unused_dev_name, mountopts, mnt);
253 put_filesystem(cgroup_fs);
258 static struct file_system_type cpuset_fs_type = {
260 .get_sb = cpuset_get_sb,
264 * Return in *pmask the portion of a cpusets's cpus_allowed that
265 * are online. If none are online, walk up the cpuset hierarchy
266 * until we find one that does have some online cpus. If we get
267 * all the way to the top and still haven't found any online cpus,
268 * return cpu_online_map. Or if passed a NULL cs from an exit'ing
269 * task, return cpu_online_map.
271 * One way or another, we guarantee to return some non-empty subset
274 * Call with callback_mutex held.
277 static void guarantee_online_cpus(const struct cpuset *cs, cpumask_t *pmask)
279 while (cs && !cpus_intersects(cs->cpus_allowed, cpu_online_map))
282 cpus_and(*pmask, cs->cpus_allowed, cpu_online_map);
284 *pmask = cpu_online_map;
285 BUG_ON(!cpus_intersects(*pmask, cpu_online_map));
289 * Return in *pmask the portion of a cpusets's mems_allowed that
290 * are online, with memory. If none are online with memory, walk
291 * up the cpuset hierarchy until we find one that does have some
292 * online mems. If we get all the way to the top and still haven't
293 * found any online mems, return node_states[N_HIGH_MEMORY].
295 * One way or another, we guarantee to return some non-empty subset
296 * of node_states[N_HIGH_MEMORY].
298 * Call with callback_mutex held.
301 static void guarantee_online_mems(const struct cpuset *cs, nodemask_t *pmask)
303 while (cs && !nodes_intersects(cs->mems_allowed,
304 node_states[N_HIGH_MEMORY]))
307 nodes_and(*pmask, cs->mems_allowed,
308 node_states[N_HIGH_MEMORY]);
310 *pmask = node_states[N_HIGH_MEMORY];
311 BUG_ON(!nodes_intersects(*pmask, node_states[N_HIGH_MEMORY]));
315 * cpuset_update_task_memory_state - update task memory placement
317 * If the current tasks cpusets mems_allowed changed behind our
318 * backs, update current->mems_allowed, mems_generation and task NUMA
319 * mempolicy to the new value.
321 * Task mempolicy is updated by rebinding it relative to the
322 * current->cpuset if a task has its memory placement changed.
323 * Do not call this routine if in_interrupt().
325 * Call without callback_mutex or task_lock() held. May be
326 * called with or without cgroup_mutex held. Thanks in part to
327 * 'the_top_cpuset_hack', the task's cpuset pointer will never
328 * be NULL. This routine also might acquire callback_mutex during
331 * Reading current->cpuset->mems_generation doesn't need task_lock
332 * to guard the current->cpuset derefence, because it is guarded
333 * from concurrent freeing of current->cpuset using RCU.
335 * The rcu_dereference() is technically probably not needed,
336 * as I don't actually mind if I see a new cpuset pointer but
337 * an old value of mems_generation. However this really only
338 * matters on alpha systems using cpusets heavily. If I dropped
339 * that rcu_dereference(), it would save them a memory barrier.
340 * For all other arch's, rcu_dereference is a no-op anyway, and for
341 * alpha systems not using cpusets, another planned optimization,
342 * avoiding the rcu critical section for tasks in the root cpuset
343 * which is statically allocated, so can't vanish, will make this
344 * irrelevant. Better to use RCU as intended, than to engage in
345 * some cute trick to save a memory barrier that is impossible to
346 * test, for alpha systems using cpusets heavily, which might not
349 * This routine is needed to update the per-task mems_allowed data,
350 * within the tasks context, when it is trying to allocate memory
351 * (in various mm/mempolicy.c routines) and notices that some other
352 * task has been modifying its cpuset.
355 void cpuset_update_task_memory_state(void)
357 int my_cpusets_mem_gen;
358 struct task_struct *tsk = current;
361 if (task_cs(tsk) == &top_cpuset) {
362 /* Don't need rcu for top_cpuset. It's never freed. */
363 my_cpusets_mem_gen = top_cpuset.mems_generation;
366 my_cpusets_mem_gen = task_cs(current)->mems_generation;
370 if (my_cpusets_mem_gen != tsk->cpuset_mems_generation) {
371 mutex_lock(&callback_mutex);
373 cs = task_cs(tsk); /* Maybe changed when task not locked */
374 guarantee_online_mems(cs, &tsk->mems_allowed);
375 tsk->cpuset_mems_generation = cs->mems_generation;
376 if (is_spread_page(cs))
377 tsk->flags |= PF_SPREAD_PAGE;
379 tsk->flags &= ~PF_SPREAD_PAGE;
380 if (is_spread_slab(cs))
381 tsk->flags |= PF_SPREAD_SLAB;
383 tsk->flags &= ~PF_SPREAD_SLAB;
385 mutex_unlock(&callback_mutex);
386 mpol_rebind_task(tsk, &tsk->mems_allowed);
391 * is_cpuset_subset(p, q) - Is cpuset p a subset of cpuset q?
393 * One cpuset is a subset of another if all its allowed CPUs and
394 * Memory Nodes are a subset of the other, and its exclusive flags
395 * are only set if the other's are set. Call holding cgroup_mutex.
398 static int is_cpuset_subset(const struct cpuset *p, const struct cpuset *q)
400 return cpus_subset(p->cpus_allowed, q->cpus_allowed) &&
401 nodes_subset(p->mems_allowed, q->mems_allowed) &&
402 is_cpu_exclusive(p) <= is_cpu_exclusive(q) &&
403 is_mem_exclusive(p) <= is_mem_exclusive(q);
407 * validate_change() - Used to validate that any proposed cpuset change
408 * follows the structural rules for cpusets.
410 * If we replaced the flag and mask values of the current cpuset
411 * (cur) with those values in the trial cpuset (trial), would
412 * our various subset and exclusive rules still be valid? Presumes
415 * 'cur' is the address of an actual, in-use cpuset. Operations
416 * such as list traversal that depend on the actual address of the
417 * cpuset in the list must use cur below, not trial.
419 * 'trial' is the address of bulk structure copy of cur, with
420 * perhaps one or more of the fields cpus_allowed, mems_allowed,
421 * or flags changed to new, trial values.
423 * Return 0 if valid, -errno if not.
426 static int validate_change(const struct cpuset *cur, const struct cpuset *trial)
429 struct cpuset *c, *par;
431 /* Each of our child cpusets must be a subset of us */
432 list_for_each_entry(cont, &cur->css.cgroup->children, sibling) {
433 if (!is_cpuset_subset(cgroup_cs(cont), trial))
437 /* Remaining checks don't apply to root cpuset */
438 if (cur == &top_cpuset)
443 /* We must be a subset of our parent cpuset */
444 if (!is_cpuset_subset(trial, par))
448 * If either I or some sibling (!= me) is exclusive, we can't
451 list_for_each_entry(cont, &par->css.cgroup->children, sibling) {
453 if ((is_cpu_exclusive(trial) || is_cpu_exclusive(c)) &&
455 cpus_intersects(trial->cpus_allowed, c->cpus_allowed))
457 if ((is_mem_exclusive(trial) || is_mem_exclusive(c)) &&
459 nodes_intersects(trial->mems_allowed, c->mems_allowed))
463 /* Cpusets with tasks can't have empty cpus_allowed or mems_allowed */
464 if (cgroup_task_count(cur->css.cgroup)) {
465 if (cpus_empty(trial->cpus_allowed) ||
466 nodes_empty(trial->mems_allowed)) {
475 * Helper routine for rebuild_sched_domains().
476 * Do cpusets a, b have overlapping cpus_allowed masks?
479 static int cpusets_overlap(struct cpuset *a, struct cpuset *b)
481 return cpus_intersects(a->cpus_allowed, b->cpus_allowed);
485 update_domain_attr(struct sched_domain_attr *dattr, struct cpuset *c)
489 if (dattr->relax_domain_level < c->relax_domain_level)
490 dattr->relax_domain_level = c->relax_domain_level;
495 * rebuild_sched_domains()
497 * If the flag 'sched_load_balance' of any cpuset with non-empty
498 * 'cpus' changes, or if the 'cpus' allowed changes in any cpuset
499 * which has that flag enabled, or if any cpuset with a non-empty
500 * 'cpus' is removed, then call this routine to rebuild the
501 * scheduler's dynamic sched domains.
503 * This routine builds a partial partition of the systems CPUs
504 * (the set of non-overlappping cpumask_t's in the array 'part'
505 * below), and passes that partial partition to the kernel/sched.c
506 * partition_sched_domains() routine, which will rebuild the
507 * schedulers load balancing domains (sched domains) as specified
508 * by that partial partition. A 'partial partition' is a set of
509 * non-overlapping subsets whose union is a subset of that set.
511 * See "What is sched_load_balance" in Documentation/cpusets.txt
512 * for a background explanation of this.
514 * Does not return errors, on the theory that the callers of this
515 * routine would rather not worry about failures to rebuild sched
516 * domains when operating in the severe memory shortage situations
517 * that could cause allocation failures below.
519 * Call with cgroup_mutex held. May take callback_mutex during
520 * call due to the kfifo_alloc() and kmalloc() calls. May nest
521 * a call to the get_online_cpus()/put_online_cpus() pair.
522 * Must not be called holding callback_mutex, because we must not
523 * call get_online_cpus() while holding callback_mutex. Elsewhere
524 * the kernel nests callback_mutex inside get_online_cpus() calls.
525 * So the reverse nesting would risk an ABBA deadlock.
527 * The three key local variables below are:
528 * q - a kfifo queue of cpuset pointers, used to implement a
529 * top-down scan of all cpusets. This scan loads a pointer
530 * to each cpuset marked is_sched_load_balance into the
531 * array 'csa'. For our purposes, rebuilding the schedulers
532 * sched domains, we can ignore !is_sched_load_balance cpusets.
533 * csa - (for CpuSet Array) Array of pointers to all the cpusets
534 * that need to be load balanced, for convenient iterative
535 * access by the subsequent code that finds the best partition,
536 * i.e the set of domains (subsets) of CPUs such that the
537 * cpus_allowed of every cpuset marked is_sched_load_balance
538 * is a subset of one of these domains, while there are as
539 * many such domains as possible, each as small as possible.
540 * doms - Conversion of 'csa' to an array of cpumasks, for passing to
541 * the kernel/sched.c routine partition_sched_domains() in a
542 * convenient format, that can be easily compared to the prior
543 * value to determine what partition elements (sched domains)
544 * were changed (added or removed.)
546 * Finding the best partition (set of domains):
547 * The triple nested loops below over i, j, k scan over the
548 * load balanced cpusets (using the array of cpuset pointers in
549 * csa[]) looking for pairs of cpusets that have overlapping
550 * cpus_allowed, but which don't have the same 'pn' partition
551 * number and gives them in the same partition number. It keeps
552 * looping on the 'restart' label until it can no longer find
555 * The union of the cpus_allowed masks from the set of
556 * all cpusets having the same 'pn' value then form the one
557 * element of the partition (one sched domain) to be passed to
558 * partition_sched_domains().
561 static void rebuild_sched_domains(void)
563 struct kfifo *q; /* queue of cpusets to be scanned */
564 struct cpuset *cp; /* scans q */
565 struct cpuset **csa; /* array of all cpuset ptrs */
566 int csn; /* how many cpuset ptrs in csa so far */
567 int i, j, k; /* indices for partition finding loops */
568 cpumask_t *doms; /* resulting partition; i.e. sched domains */
569 struct sched_domain_attr *dattr; /* attributes for custom domains */
570 int ndoms; /* number of sched domains in result */
571 int nslot; /* next empty doms[] cpumask_t slot */
578 /* Special case for the 99% of systems with one, full, sched domain */
579 if (is_sched_load_balance(&top_cpuset)) {
581 doms = kmalloc(sizeof(cpumask_t), GFP_KERNEL);
584 dattr = kmalloc(sizeof(struct sched_domain_attr), GFP_KERNEL);
586 *dattr = SD_ATTR_INIT;
587 update_domain_attr(dattr, &top_cpuset);
589 *doms = top_cpuset.cpus_allowed;
593 q = kfifo_alloc(number_of_cpusets * sizeof(cp), GFP_KERNEL, NULL);
596 csa = kmalloc(number_of_cpusets * sizeof(cp), GFP_KERNEL);
602 __kfifo_put(q, (void *)&cp, sizeof(cp));
603 while (__kfifo_get(q, (void *)&cp, sizeof(cp))) {
605 struct cpuset *child; /* scans child cpusets of cp */
606 if (is_sched_load_balance(cp))
608 list_for_each_entry(cont, &cp->css.cgroup->children, sibling) {
609 child = cgroup_cs(cont);
610 __kfifo_put(q, (void *)&child, sizeof(cp));
614 for (i = 0; i < csn; i++)
619 /* Find the best partition (set of sched domains) */
620 for (i = 0; i < csn; i++) {
621 struct cpuset *a = csa[i];
624 for (j = 0; j < csn; j++) {
625 struct cpuset *b = csa[j];
628 if (apn != bpn && cpusets_overlap(a, b)) {
629 for (k = 0; k < csn; k++) {
630 struct cpuset *c = csa[k];
635 ndoms--; /* one less element */
641 /* Convert <csn, csa> to <ndoms, doms> */
642 doms = kmalloc(ndoms * sizeof(cpumask_t), GFP_KERNEL);
645 dattr = kmalloc(ndoms * sizeof(struct sched_domain_attr), GFP_KERNEL);
647 for (nslot = 0, i = 0; i < csn; i++) {
648 struct cpuset *a = csa[i];
652 cpumask_t *dp = doms + nslot;
654 if (nslot == ndoms) {
655 static int warnings = 10;
658 "rebuild_sched_domains confused:"
659 " nslot %d, ndoms %d, csn %d, i %d,"
661 nslot, ndoms, csn, i, apn);
669 *(dattr + nslot) = SD_ATTR_INIT;
670 for (j = i; j < csn; j++) {
671 struct cpuset *b = csa[j];
674 cpus_or(*dp, *dp, b->cpus_allowed);
676 update_domain_attr(dattr, b);
682 BUG_ON(nslot != ndoms);
685 /* Have scheduler rebuild sched domains */
687 partition_sched_domains(ndoms, doms, dattr);
694 /* Don't kfree(doms) -- partition_sched_domains() does that. */
695 /* Don't kfree(dattr) -- partition_sched_domains() does that. */
698 static inline int started_after_time(struct task_struct *t1,
699 struct timespec *time,
700 struct task_struct *t2)
702 int start_diff = timespec_compare(&t1->start_time, time);
703 if (start_diff > 0) {
705 } else if (start_diff < 0) {
709 * Arbitrarily, if two processes started at the same
710 * time, we'll say that the lower pointer value
711 * started first. Note that t2 may have exited by now
712 * so this may not be a valid pointer any longer, but
713 * that's fine - it still serves to distinguish
714 * between two tasks started (effectively)
721 static inline int started_after(void *p1, void *p2)
723 struct task_struct *t1 = p1;
724 struct task_struct *t2 = p2;
725 return started_after_time(t1, &t2->start_time, t2);
729 * cpuset_test_cpumask - test a task's cpus_allowed versus its cpuset's
731 * @scan: struct cgroup_scanner contained in its struct cpuset_hotplug_scanner
733 * Call with cgroup_mutex held. May take callback_mutex during call.
734 * Called for each task in a cgroup by cgroup_scan_tasks().
735 * Return nonzero if this tasks's cpus_allowed mask should be changed (in other
736 * words, if its mask is not equal to its cpuset's mask).
738 static int cpuset_test_cpumask(struct task_struct *tsk,
739 struct cgroup_scanner *scan)
741 return !cpus_equal(tsk->cpus_allowed,
742 (cgroup_cs(scan->cg))->cpus_allowed);
746 * cpuset_change_cpumask - make a task's cpus_allowed the same as its cpuset's
748 * @scan: struct cgroup_scanner containing the cgroup of the task
750 * Called by cgroup_scan_tasks() for each task in a cgroup whose
751 * cpus_allowed mask needs to be changed.
753 * We don't need to re-check for the cgroup/cpuset membership, since we're
754 * holding cgroup_lock() at this point.
756 static void cpuset_change_cpumask(struct task_struct *tsk,
757 struct cgroup_scanner *scan)
759 set_cpus_allowed_ptr(tsk, &((cgroup_cs(scan->cg))->cpus_allowed));
763 * update_cpumask - update the cpus_allowed mask of a cpuset and all tasks in it
764 * @cs: the cpuset to consider
765 * @buf: buffer of cpu numbers written to this cpuset
767 static int update_cpumask(struct cpuset *cs, char *buf)
769 struct cpuset trialcs;
770 struct cgroup_scanner scan;
771 struct ptr_heap heap;
773 int is_load_balanced;
775 /* top_cpuset.cpus_allowed tracks cpu_online_map; it's read-only */
776 if (cs == &top_cpuset)
782 * An empty cpus_allowed is ok only if the cpuset has no tasks.
783 * Since cpulist_parse() fails on an empty mask, we special case
784 * that parsing. The validate_change() call ensures that cpusets
785 * with tasks have cpus.
789 cpus_clear(trialcs.cpus_allowed);
791 retval = cpulist_parse(buf, trialcs.cpus_allowed);
795 cpus_and(trialcs.cpus_allowed, trialcs.cpus_allowed, cpu_online_map);
796 retval = validate_change(cs, &trialcs);
800 /* Nothing to do if the cpus didn't change */
801 if (cpus_equal(cs->cpus_allowed, trialcs.cpus_allowed))
804 retval = heap_init(&heap, PAGE_SIZE, GFP_KERNEL, &started_after);
808 is_load_balanced = is_sched_load_balance(&trialcs);
810 mutex_lock(&callback_mutex);
811 cs->cpus_allowed = trialcs.cpus_allowed;
812 mutex_unlock(&callback_mutex);
815 * Scan tasks in the cpuset, and update the cpumasks of any
816 * that need an update.
818 scan.cg = cs->css.cgroup;
819 scan.test_task = cpuset_test_cpumask;
820 scan.process_task = cpuset_change_cpumask;
822 cgroup_scan_tasks(&scan);
825 if (is_load_balanced)
826 rebuild_sched_domains();
833 * Migrate memory region from one set of nodes to another.
835 * Temporarilly set tasks mems_allowed to target nodes of migration,
836 * so that the migration code can allocate pages on these nodes.
838 * Call holding cgroup_mutex, so current's cpuset won't change
839 * during this call, as manage_mutex holds off any cpuset_attach()
840 * calls. Therefore we don't need to take task_lock around the
841 * call to guarantee_online_mems(), as we know no one is changing
844 * Hold callback_mutex around the two modifications of our tasks
845 * mems_allowed to synchronize with cpuset_mems_allowed().
847 * While the mm_struct we are migrating is typically from some
848 * other task, the task_struct mems_allowed that we are hacking
849 * is for our current task, which must allocate new pages for that
850 * migrating memory region.
852 * We call cpuset_update_task_memory_state() before hacking
853 * our tasks mems_allowed, so that we are assured of being in
854 * sync with our tasks cpuset, and in particular, callbacks to
855 * cpuset_update_task_memory_state() from nested page allocations
856 * won't see any mismatch of our cpuset and task mems_generation
857 * values, so won't overwrite our hacked tasks mems_allowed
861 static void cpuset_migrate_mm(struct mm_struct *mm, const nodemask_t *from,
862 const nodemask_t *to)
864 struct task_struct *tsk = current;
866 cpuset_update_task_memory_state();
868 mutex_lock(&callback_mutex);
869 tsk->mems_allowed = *to;
870 mutex_unlock(&callback_mutex);
872 do_migrate_pages(mm, from, to, MPOL_MF_MOVE_ALL);
874 mutex_lock(&callback_mutex);
875 guarantee_online_mems(task_cs(tsk),&tsk->mems_allowed);
876 mutex_unlock(&callback_mutex);
880 * Handle user request to change the 'mems' memory placement
881 * of a cpuset. Needs to validate the request, update the
882 * cpusets mems_allowed and mems_generation, and for each
883 * task in the cpuset, rebind any vma mempolicies and if
884 * the cpuset is marked 'memory_migrate', migrate the tasks
885 * pages to the new memory.
887 * Call with cgroup_mutex held. May take callback_mutex during call.
888 * Will take tasklist_lock, scan tasklist for tasks in cpuset cs,
889 * lock each such tasks mm->mmap_sem, scan its vma's and rebind
890 * their mempolicies to the cpusets new mems_allowed.
893 static void *cpuset_being_rebound;
895 static int update_nodemask(struct cpuset *cs, char *buf)
897 struct cpuset trialcs;
899 struct task_struct *p;
900 struct mm_struct **mmarray;
905 struct cgroup_iter it;
908 * top_cpuset.mems_allowed tracks node_stats[N_HIGH_MEMORY];
911 if (cs == &top_cpuset)
917 * An empty mems_allowed is ok iff there are no tasks in the cpuset.
918 * Since nodelist_parse() fails on an empty mask, we special case
919 * that parsing. The validate_change() call ensures that cpusets
920 * with tasks have memory.
924 nodes_clear(trialcs.mems_allowed);
926 retval = nodelist_parse(buf, trialcs.mems_allowed);
930 nodes_and(trialcs.mems_allowed, trialcs.mems_allowed,
931 node_states[N_HIGH_MEMORY]);
932 oldmem = cs->mems_allowed;
933 if (nodes_equal(oldmem, trialcs.mems_allowed)) {
934 retval = 0; /* Too easy - nothing to do */
937 retval = validate_change(cs, &trialcs);
941 mutex_lock(&callback_mutex);
942 cs->mems_allowed = trialcs.mems_allowed;
943 cs->mems_generation = cpuset_mems_generation++;
944 mutex_unlock(&callback_mutex);
946 cpuset_being_rebound = cs; /* causes mpol_dup() rebind */
948 fudge = 10; /* spare mmarray[] slots */
949 fudge += cpus_weight(cs->cpus_allowed); /* imagine one fork-bomb/cpu */
953 * Allocate mmarray[] to hold mm reference for each task
954 * in cpuset cs. Can't kmalloc GFP_KERNEL while holding
955 * tasklist_lock. We could use GFP_ATOMIC, but with a
956 * few more lines of code, we can retry until we get a big
957 * enough mmarray[] w/o using GFP_ATOMIC.
960 ntasks = cgroup_task_count(cs->css.cgroup); /* guess */
962 mmarray = kmalloc(ntasks * sizeof(*mmarray), GFP_KERNEL);
965 read_lock(&tasklist_lock); /* block fork */
966 if (cgroup_task_count(cs->css.cgroup) <= ntasks)
967 break; /* got enough */
968 read_unlock(&tasklist_lock); /* try again */
974 /* Load up mmarray[] with mm reference for each task in cpuset. */
975 cgroup_iter_start(cs->css.cgroup, &it);
976 while ((p = cgroup_iter_next(cs->css.cgroup, &it))) {
977 struct mm_struct *mm;
981 "Cpuset mempolicy rebind incomplete.\n");
989 cgroup_iter_end(cs->css.cgroup, &it);
990 read_unlock(&tasklist_lock);
993 * Now that we've dropped the tasklist spinlock, we can
994 * rebind the vma mempolicies of each mm in mmarray[] to their
995 * new cpuset, and release that mm. The mpol_rebind_mm()
996 * call takes mmap_sem, which we couldn't take while holding
997 * tasklist_lock. Forks can happen again now - the mpol_dup()
998 * cpuset_being_rebound check will catch such forks, and rebind
999 * their vma mempolicies too. Because we still hold the global
1000 * cgroup_mutex, we know that no other rebind effort will
1001 * be contending for the global variable cpuset_being_rebound.
1002 * It's ok if we rebind the same mm twice; mpol_rebind_mm()
1003 * is idempotent. Also migrate pages in each mm to new nodes.
1005 migrate = is_memory_migrate(cs);
1006 for (i = 0; i < n; i++) {
1007 struct mm_struct *mm = mmarray[i];
1009 mpol_rebind_mm(mm, &cs->mems_allowed);
1011 cpuset_migrate_mm(mm, &oldmem, &cs->mems_allowed);
1015 /* We're done rebinding vmas to this cpuset's new mems_allowed. */
1017 cpuset_being_rebound = NULL;
1023 int current_cpuset_is_being_rebound(void)
1025 return task_cs(current) == cpuset_being_rebound;
1028 static int update_relax_domain_level(struct cpuset *cs, char *buf)
1030 int val = simple_strtol(buf, NULL, 10);
1035 if (val != cs->relax_domain_level) {
1036 cs->relax_domain_level = val;
1037 rebuild_sched_domains();
1044 * update_flag - read a 0 or a 1 in a file and update associated flag
1045 * bit: the bit to update (CS_CPU_EXCLUSIVE, CS_MEM_EXCLUSIVE,
1046 * CS_SCHED_LOAD_BALANCE,
1047 * CS_NOTIFY_ON_RELEASE, CS_MEMORY_MIGRATE,
1048 * CS_SPREAD_PAGE, CS_SPREAD_SLAB)
1049 * cs: the cpuset to update
1050 * buf: the buffer where we read the 0 or 1
1052 * Call with cgroup_mutex held.
1055 static int update_flag(cpuset_flagbits_t bit, struct cpuset *cs,
1058 struct cpuset trialcs;
1060 int cpus_nonempty, balance_flag_changed;
1064 set_bit(bit, &trialcs.flags);
1066 clear_bit(bit, &trialcs.flags);
1068 err = validate_change(cs, &trialcs);
1072 cpus_nonempty = !cpus_empty(trialcs.cpus_allowed);
1073 balance_flag_changed = (is_sched_load_balance(cs) !=
1074 is_sched_load_balance(&trialcs));
1076 mutex_lock(&callback_mutex);
1077 cs->flags = trialcs.flags;
1078 mutex_unlock(&callback_mutex);
1080 if (cpus_nonempty && balance_flag_changed)
1081 rebuild_sched_domains();
1087 * Frequency meter - How fast is some event occurring?
1089 * These routines manage a digitally filtered, constant time based,
1090 * event frequency meter. There are four routines:
1091 * fmeter_init() - initialize a frequency meter.
1092 * fmeter_markevent() - called each time the event happens.
1093 * fmeter_getrate() - returns the recent rate of such events.
1094 * fmeter_update() - internal routine used to update fmeter.
1096 * A common data structure is passed to each of these routines,
1097 * which is used to keep track of the state required to manage the
1098 * frequency meter and its digital filter.
1100 * The filter works on the number of events marked per unit time.
1101 * The filter is single-pole low-pass recursive (IIR). The time unit
1102 * is 1 second. Arithmetic is done using 32-bit integers scaled to
1103 * simulate 3 decimal digits of precision (multiplied by 1000).
1105 * With an FM_COEF of 933, and a time base of 1 second, the filter
1106 * has a half-life of 10 seconds, meaning that if the events quit
1107 * happening, then the rate returned from the fmeter_getrate()
1108 * will be cut in half each 10 seconds, until it converges to zero.
1110 * It is not worth doing a real infinitely recursive filter. If more
1111 * than FM_MAXTICKS ticks have elapsed since the last filter event,
1112 * just compute FM_MAXTICKS ticks worth, by which point the level
1115 * Limit the count of unprocessed events to FM_MAXCNT, so as to avoid
1116 * arithmetic overflow in the fmeter_update() routine.
1118 * Given the simple 32 bit integer arithmetic used, this meter works
1119 * best for reporting rates between one per millisecond (msec) and
1120 * one per 32 (approx) seconds. At constant rates faster than one
1121 * per msec it maxes out at values just under 1,000,000. At constant
1122 * rates between one per msec, and one per second it will stabilize
1123 * to a value N*1000, where N is the rate of events per second.
1124 * At constant rates between one per second and one per 32 seconds,
1125 * it will be choppy, moving up on the seconds that have an event,
1126 * and then decaying until the next event. At rates slower than
1127 * about one in 32 seconds, it decays all the way back to zero between
1131 #define FM_COEF 933 /* coefficient for half-life of 10 secs */
1132 #define FM_MAXTICKS ((time_t)99) /* useless computing more ticks than this */
1133 #define FM_MAXCNT 1000000 /* limit cnt to avoid overflow */
1134 #define FM_SCALE 1000 /* faux fixed point scale */
1136 /* Initialize a frequency meter */
1137 static void fmeter_init(struct fmeter *fmp)
1142 spin_lock_init(&fmp->lock);
1145 /* Internal meter update - process cnt events and update value */
1146 static void fmeter_update(struct fmeter *fmp)
1148 time_t now = get_seconds();
1149 time_t ticks = now - fmp->time;
1154 ticks = min(FM_MAXTICKS, ticks);
1156 fmp->val = (FM_COEF * fmp->val) / FM_SCALE;
1159 fmp->val += ((FM_SCALE - FM_COEF) * fmp->cnt) / FM_SCALE;
1163 /* Process any previous ticks, then bump cnt by one (times scale). */
1164 static void fmeter_markevent(struct fmeter *fmp)
1166 spin_lock(&fmp->lock);
1168 fmp->cnt = min(FM_MAXCNT, fmp->cnt + FM_SCALE);
1169 spin_unlock(&fmp->lock);
1172 /* Process any previous ticks, then return current value. */
1173 static int fmeter_getrate(struct fmeter *fmp)
1177 spin_lock(&fmp->lock);
1180 spin_unlock(&fmp->lock);
1184 /* Called by cgroups to determine if a cpuset is usable; cgroup_mutex held */
1185 static int cpuset_can_attach(struct cgroup_subsys *ss,
1186 struct cgroup *cont, struct task_struct *tsk)
1188 struct cpuset *cs = cgroup_cs(cont);
1190 if (cpus_empty(cs->cpus_allowed) || nodes_empty(cs->mems_allowed))
1193 return security_task_setscheduler(tsk, 0, NULL);
1196 static void cpuset_attach(struct cgroup_subsys *ss,
1197 struct cgroup *cont, struct cgroup *oldcont,
1198 struct task_struct *tsk)
1201 nodemask_t from, to;
1202 struct mm_struct *mm;
1203 struct cpuset *cs = cgroup_cs(cont);
1204 struct cpuset *oldcs = cgroup_cs(oldcont);
1206 mutex_lock(&callback_mutex);
1207 guarantee_online_cpus(cs, &cpus);
1208 set_cpus_allowed_ptr(tsk, &cpus);
1209 mutex_unlock(&callback_mutex);
1211 from = oldcs->mems_allowed;
1212 to = cs->mems_allowed;
1213 mm = get_task_mm(tsk);
1215 mpol_rebind_mm(mm, &to);
1216 if (is_memory_migrate(cs))
1217 cpuset_migrate_mm(mm, &from, &to);
1223 /* The various types of files and directories in a cpuset file system */
1226 FILE_MEMORY_MIGRATE,
1231 FILE_SCHED_LOAD_BALANCE,
1232 FILE_SCHED_RELAX_DOMAIN_LEVEL,
1233 FILE_MEMORY_PRESSURE_ENABLED,
1234 FILE_MEMORY_PRESSURE,
1237 } cpuset_filetype_t;
1239 static ssize_t cpuset_common_file_write(struct cgroup *cont,
1242 const char __user *userbuf,
1243 size_t nbytes, loff_t *unused_ppos)
1245 struct cpuset *cs = cgroup_cs(cont);
1246 cpuset_filetype_t type = cft->private;
1250 /* Crude upper limit on largest legitimate cpulist user might write. */
1251 if (nbytes > 100U + 6 * max(NR_CPUS, MAX_NUMNODES))
1254 /* +1 for nul-terminator */
1255 buffer = kmalloc(nbytes + 1, GFP_KERNEL);
1259 if (copy_from_user(buffer, userbuf, nbytes)) {
1263 buffer[nbytes] = 0; /* nul-terminate */
1267 if (cgroup_is_removed(cont)) {
1274 retval = update_cpumask(cs, buffer);
1277 retval = update_nodemask(cs, buffer);
1279 case FILE_SCHED_RELAX_DOMAIN_LEVEL:
1280 retval = update_relax_domain_level(cs, buffer);
1296 static int cpuset_write_u64(struct cgroup *cgrp, struct cftype *cft, u64 val)
1299 struct cpuset *cs = cgroup_cs(cgrp);
1300 cpuset_filetype_t type = cft->private;
1304 if (cgroup_is_removed(cgrp)) {
1310 case FILE_CPU_EXCLUSIVE:
1311 retval = update_flag(CS_CPU_EXCLUSIVE, cs, val);
1313 case FILE_MEM_EXCLUSIVE:
1314 retval = update_flag(CS_MEM_EXCLUSIVE, cs, val);
1316 case FILE_SCHED_LOAD_BALANCE:
1317 retval = update_flag(CS_SCHED_LOAD_BALANCE, cs, val);
1319 case FILE_MEMORY_MIGRATE:
1320 retval = update_flag(CS_MEMORY_MIGRATE, cs, val);
1322 case FILE_MEMORY_PRESSURE_ENABLED:
1323 cpuset_memory_pressure_enabled = !!val;
1325 case FILE_MEMORY_PRESSURE:
1328 case FILE_SPREAD_PAGE:
1329 retval = update_flag(CS_SPREAD_PAGE, cs, val);
1330 cs->mems_generation = cpuset_mems_generation++;
1332 case FILE_SPREAD_SLAB:
1333 retval = update_flag(CS_SPREAD_SLAB, cs, val);
1334 cs->mems_generation = cpuset_mems_generation++;
1345 * These ascii lists should be read in a single call, by using a user
1346 * buffer large enough to hold the entire map. If read in smaller
1347 * chunks, there is no guarantee of atomicity. Since the display format
1348 * used, list of ranges of sequential numbers, is variable length,
1349 * and since these maps can change value dynamically, one could read
1350 * gibberish by doing partial reads while a list was changing.
1351 * A single large read to a buffer that crosses a page boundary is
1352 * ok, because the result being copied to user land is not recomputed
1353 * across a page fault.
1356 static int cpuset_sprintf_cpulist(char *page, struct cpuset *cs)
1360 mutex_lock(&callback_mutex);
1361 mask = cs->cpus_allowed;
1362 mutex_unlock(&callback_mutex);
1364 return cpulist_scnprintf(page, PAGE_SIZE, mask);
1367 static int cpuset_sprintf_memlist(char *page, struct cpuset *cs)
1371 mutex_lock(&callback_mutex);
1372 mask = cs->mems_allowed;
1373 mutex_unlock(&callback_mutex);
1375 return nodelist_scnprintf(page, PAGE_SIZE, mask);
1378 static ssize_t cpuset_common_file_read(struct cgroup *cont,
1382 size_t nbytes, loff_t *ppos)
1384 struct cpuset *cs = cgroup_cs(cont);
1385 cpuset_filetype_t type = cft->private;
1390 if (!(page = (char *)__get_free_page(GFP_TEMPORARY)))
1397 s += cpuset_sprintf_cpulist(s, cs);
1400 s += cpuset_sprintf_memlist(s, cs);
1402 case FILE_SCHED_RELAX_DOMAIN_LEVEL:
1403 s += sprintf(s, "%d", cs->relax_domain_level);
1411 retval = simple_read_from_buffer(buf, nbytes, ppos, page, s - page);
1413 free_page((unsigned long)page);
1417 static u64 cpuset_read_u64(struct cgroup *cont, struct cftype *cft)
1419 struct cpuset *cs = cgroup_cs(cont);
1420 cpuset_filetype_t type = cft->private;
1422 case FILE_CPU_EXCLUSIVE:
1423 return is_cpu_exclusive(cs);
1424 case FILE_MEM_EXCLUSIVE:
1425 return is_mem_exclusive(cs);
1426 case FILE_SCHED_LOAD_BALANCE:
1427 return is_sched_load_balance(cs);
1428 case FILE_MEMORY_MIGRATE:
1429 return is_memory_migrate(cs);
1430 case FILE_MEMORY_PRESSURE_ENABLED:
1431 return cpuset_memory_pressure_enabled;
1432 case FILE_MEMORY_PRESSURE:
1433 return fmeter_getrate(&cs->fmeter);
1434 case FILE_SPREAD_PAGE:
1435 return is_spread_page(cs);
1436 case FILE_SPREAD_SLAB:
1437 return is_spread_slab(cs);
1445 * for the common functions, 'private' gives the type of file
1448 static struct cftype files[] = {
1451 .read = cpuset_common_file_read,
1452 .write = cpuset_common_file_write,
1453 .private = FILE_CPULIST,
1458 .read = cpuset_common_file_read,
1459 .write = cpuset_common_file_write,
1460 .private = FILE_MEMLIST,
1464 .name = "cpu_exclusive",
1465 .read_u64 = cpuset_read_u64,
1466 .write_u64 = cpuset_write_u64,
1467 .private = FILE_CPU_EXCLUSIVE,
1471 .name = "mem_exclusive",
1472 .read_u64 = cpuset_read_u64,
1473 .write_u64 = cpuset_write_u64,
1474 .private = FILE_MEM_EXCLUSIVE,
1478 .name = "sched_load_balance",
1479 .read_u64 = cpuset_read_u64,
1480 .write_u64 = cpuset_write_u64,
1481 .private = FILE_SCHED_LOAD_BALANCE,
1485 .name = "sched_relax_domain_level",
1486 .read_u64 = cpuset_read_u64,
1487 .write_u64 = cpuset_write_u64,
1488 .private = FILE_SCHED_RELAX_DOMAIN_LEVEL,
1492 .name = "memory_migrate",
1493 .read_u64 = cpuset_read_u64,
1494 .write_u64 = cpuset_write_u64,
1495 .private = FILE_MEMORY_MIGRATE,
1499 .name = "memory_pressure",
1500 .read_u64 = cpuset_read_u64,
1501 .write_u64 = cpuset_write_u64,
1502 .private = FILE_MEMORY_PRESSURE,
1506 .name = "memory_spread_page",
1507 .read_u64 = cpuset_read_u64,
1508 .write_u64 = cpuset_write_u64,
1509 .private = FILE_SPREAD_PAGE,
1513 .name = "memory_spread_slab",
1514 .read_u64 = cpuset_read_u64,
1515 .write_u64 = cpuset_write_u64,
1516 .private = FILE_SPREAD_SLAB,
1520 static struct cftype cft_memory_pressure_enabled = {
1521 .name = "memory_pressure_enabled",
1522 .read_u64 = cpuset_read_u64,
1523 .write_u64 = cpuset_write_u64,
1524 .private = FILE_MEMORY_PRESSURE_ENABLED,
1527 static int cpuset_populate(struct cgroup_subsys *ss, struct cgroup *cont)
1531 err = cgroup_add_files(cont, ss, files, ARRAY_SIZE(files));
1534 /* memory_pressure_enabled is in root cpuset only */
1536 err = cgroup_add_file(cont, ss,
1537 &cft_memory_pressure_enabled);
1542 * post_clone() is called at the end of cgroup_clone().
1543 * 'cgroup' was just created automatically as a result of
1544 * a cgroup_clone(), and the current task is about to
1545 * be moved into 'cgroup'.
1547 * Currently we refuse to set up the cgroup - thereby
1548 * refusing the task to be entered, and as a result refusing
1549 * the sys_unshare() or clone() which initiated it - if any
1550 * sibling cpusets have exclusive cpus or mem.
1552 * If this becomes a problem for some users who wish to
1553 * allow that scenario, then cpuset_post_clone() could be
1554 * changed to grant parent->cpus_allowed-sibling_cpus_exclusive
1555 * (and likewise for mems) to the new cgroup. Called with cgroup_mutex
1558 static void cpuset_post_clone(struct cgroup_subsys *ss,
1559 struct cgroup *cgroup)
1561 struct cgroup *parent, *child;
1562 struct cpuset *cs, *parent_cs;
1564 parent = cgroup->parent;
1565 list_for_each_entry(child, &parent->children, sibling) {
1566 cs = cgroup_cs(child);
1567 if (is_mem_exclusive(cs) || is_cpu_exclusive(cs))
1570 cs = cgroup_cs(cgroup);
1571 parent_cs = cgroup_cs(parent);
1573 cs->mems_allowed = parent_cs->mems_allowed;
1574 cs->cpus_allowed = parent_cs->cpus_allowed;
1579 * cpuset_create - create a cpuset
1580 * ss: cpuset cgroup subsystem
1581 * cont: control group that the new cpuset will be part of
1584 static struct cgroup_subsys_state *cpuset_create(
1585 struct cgroup_subsys *ss,
1586 struct cgroup *cont)
1589 struct cpuset *parent;
1591 if (!cont->parent) {
1592 /* This is early initialization for the top cgroup */
1593 top_cpuset.mems_generation = cpuset_mems_generation++;
1594 return &top_cpuset.css;
1596 parent = cgroup_cs(cont->parent);
1597 cs = kmalloc(sizeof(*cs), GFP_KERNEL);
1599 return ERR_PTR(-ENOMEM);
1601 cpuset_update_task_memory_state();
1603 if (is_spread_page(parent))
1604 set_bit(CS_SPREAD_PAGE, &cs->flags);
1605 if (is_spread_slab(parent))
1606 set_bit(CS_SPREAD_SLAB, &cs->flags);
1607 set_bit(CS_SCHED_LOAD_BALANCE, &cs->flags);
1608 cpus_clear(cs->cpus_allowed);
1609 nodes_clear(cs->mems_allowed);
1610 cs->mems_generation = cpuset_mems_generation++;
1611 fmeter_init(&cs->fmeter);
1612 cs->relax_domain_level = -1;
1614 cs->parent = parent;
1615 number_of_cpusets++;
1620 * Locking note on the strange update_flag() call below:
1622 * If the cpuset being removed has its flag 'sched_load_balance'
1623 * enabled, then simulate turning sched_load_balance off, which
1624 * will call rebuild_sched_domains(). The get_online_cpus()
1625 * call in rebuild_sched_domains() must not be made while holding
1626 * callback_mutex. Elsewhere the kernel nests callback_mutex inside
1627 * get_online_cpus() calls. So the reverse nesting would risk an
1631 static void cpuset_destroy(struct cgroup_subsys *ss, struct cgroup *cont)
1633 struct cpuset *cs = cgroup_cs(cont);
1635 cpuset_update_task_memory_state();
1637 if (is_sched_load_balance(cs))
1638 update_flag(CS_SCHED_LOAD_BALANCE, cs, 0);
1640 number_of_cpusets--;
1644 struct cgroup_subsys cpuset_subsys = {
1646 .create = cpuset_create,
1647 .destroy = cpuset_destroy,
1648 .can_attach = cpuset_can_attach,
1649 .attach = cpuset_attach,
1650 .populate = cpuset_populate,
1651 .post_clone = cpuset_post_clone,
1652 .subsys_id = cpuset_subsys_id,
1657 * cpuset_init_early - just enough so that the calls to
1658 * cpuset_update_task_memory_state() in early init code
1662 int __init cpuset_init_early(void)
1664 top_cpuset.mems_generation = cpuset_mems_generation++;
1670 * cpuset_init - initialize cpusets at system boot
1672 * Description: Initialize top_cpuset and the cpuset internal file system,
1675 int __init cpuset_init(void)
1679 cpus_setall(top_cpuset.cpus_allowed);
1680 nodes_setall(top_cpuset.mems_allowed);
1682 fmeter_init(&top_cpuset.fmeter);
1683 top_cpuset.mems_generation = cpuset_mems_generation++;
1684 set_bit(CS_SCHED_LOAD_BALANCE, &top_cpuset.flags);
1685 top_cpuset.relax_domain_level = -1;
1687 err = register_filesystem(&cpuset_fs_type);
1691 number_of_cpusets = 1;
1696 * cpuset_do_move_task - move a given task to another cpuset
1697 * @tsk: pointer to task_struct the task to move
1698 * @scan: struct cgroup_scanner contained in its struct cpuset_hotplug_scanner
1700 * Called by cgroup_scan_tasks() for each task in a cgroup.
1701 * Return nonzero to stop the walk through the tasks.
1703 static void cpuset_do_move_task(struct task_struct *tsk,
1704 struct cgroup_scanner *scan)
1706 struct cpuset_hotplug_scanner *chsp;
1708 chsp = container_of(scan, struct cpuset_hotplug_scanner, scan);
1709 cgroup_attach_task(chsp->to, tsk);
1713 * move_member_tasks_to_cpuset - move tasks from one cpuset to another
1714 * @from: cpuset in which the tasks currently reside
1715 * @to: cpuset to which the tasks will be moved
1717 * Called with cgroup_mutex held
1718 * callback_mutex must not be held, as cpuset_attach() will take it.
1720 * The cgroup_scan_tasks() function will scan all the tasks in a cgroup,
1721 * calling callback functions for each.
1723 static void move_member_tasks_to_cpuset(struct cpuset *from, struct cpuset *to)
1725 struct cpuset_hotplug_scanner scan;
1727 scan.scan.cg = from->css.cgroup;
1728 scan.scan.test_task = NULL; /* select all tasks in cgroup */
1729 scan.scan.process_task = cpuset_do_move_task;
1730 scan.scan.heap = NULL;
1731 scan.to = to->css.cgroup;
1733 if (cgroup_scan_tasks((struct cgroup_scanner *)&scan))
1734 printk(KERN_ERR "move_member_tasks_to_cpuset: "
1735 "cgroup_scan_tasks failed\n");
1739 * If common_cpu_mem_hotplug_unplug(), below, unplugs any CPUs
1740 * or memory nodes, we need to walk over the cpuset hierarchy,
1741 * removing that CPU or node from all cpusets. If this removes the
1742 * last CPU or node from a cpuset, then move the tasks in the empty
1743 * cpuset to its next-highest non-empty parent.
1745 * Called with cgroup_mutex held
1746 * callback_mutex must not be held, as cpuset_attach() will take it.
1748 static void remove_tasks_in_empty_cpuset(struct cpuset *cs)
1750 struct cpuset *parent;
1753 * The cgroup's css_sets list is in use if there are tasks
1754 * in the cpuset; the list is empty if there are none;
1755 * the cs->css.refcnt seems always 0.
1757 if (list_empty(&cs->css.cgroup->css_sets))
1761 * Find its next-highest non-empty parent, (top cpuset
1762 * has online cpus, so can't be empty).
1764 parent = cs->parent;
1765 while (cpus_empty(parent->cpus_allowed) ||
1766 nodes_empty(parent->mems_allowed))
1767 parent = parent->parent;
1769 move_member_tasks_to_cpuset(cs, parent);
1773 * Walk the specified cpuset subtree and look for empty cpusets.
1774 * The tasks of such cpuset must be moved to a parent cpuset.
1776 * Called with cgroup_mutex held. We take callback_mutex to modify
1777 * cpus_allowed and mems_allowed.
1779 * This walk processes the tree from top to bottom, completing one layer
1780 * before dropping down to the next. It always processes a node before
1781 * any of its children.
1783 * For now, since we lack memory hot unplug, we'll never see a cpuset
1784 * that has tasks along with an empty 'mems'. But if we did see such
1785 * a cpuset, we'd handle it just like we do if its 'cpus' was empty.
1787 static void scan_for_empty_cpusets(const struct cpuset *root)
1789 struct cpuset *cp; /* scans cpusets being updated */
1790 struct cpuset *child; /* scans child cpusets of cp */
1791 struct list_head queue;
1792 struct cgroup *cont;
1794 INIT_LIST_HEAD(&queue);
1796 list_add_tail((struct list_head *)&root->stack_list, &queue);
1798 while (!list_empty(&queue)) {
1799 cp = container_of(queue.next, struct cpuset, stack_list);
1800 list_del(queue.next);
1801 list_for_each_entry(cont, &cp->css.cgroup->children, sibling) {
1802 child = cgroup_cs(cont);
1803 list_add_tail(&child->stack_list, &queue);
1805 cont = cp->css.cgroup;
1807 /* Continue past cpusets with all cpus, mems online */
1808 if (cpus_subset(cp->cpus_allowed, cpu_online_map) &&
1809 nodes_subset(cp->mems_allowed, node_states[N_HIGH_MEMORY]))
1812 /* Remove offline cpus and mems from this cpuset. */
1813 mutex_lock(&callback_mutex);
1814 cpus_and(cp->cpus_allowed, cp->cpus_allowed, cpu_online_map);
1815 nodes_and(cp->mems_allowed, cp->mems_allowed,
1816 node_states[N_HIGH_MEMORY]);
1817 mutex_unlock(&callback_mutex);
1819 /* Move tasks from the empty cpuset to a parent */
1820 if (cpus_empty(cp->cpus_allowed) ||
1821 nodes_empty(cp->mems_allowed))
1822 remove_tasks_in_empty_cpuset(cp);
1827 * The cpus_allowed and mems_allowed nodemasks in the top_cpuset track
1828 * cpu_online_map and node_states[N_HIGH_MEMORY]. Force the top cpuset to
1829 * track what's online after any CPU or memory node hotplug or unplug event.
1831 * Since there are two callers of this routine, one for CPU hotplug
1832 * events and one for memory node hotplug events, we could have coded
1833 * two separate routines here. We code it as a single common routine
1834 * in order to minimize text size.
1837 static void common_cpu_mem_hotplug_unplug(void)
1841 top_cpuset.cpus_allowed = cpu_online_map;
1842 top_cpuset.mems_allowed = node_states[N_HIGH_MEMORY];
1843 scan_for_empty_cpusets(&top_cpuset);
1849 * The top_cpuset tracks what CPUs and Memory Nodes are online,
1850 * period. This is necessary in order to make cpusets transparent
1851 * (of no affect) on systems that are actively using CPU hotplug
1852 * but making no active use of cpusets.
1854 * This routine ensures that top_cpuset.cpus_allowed tracks
1855 * cpu_online_map on each CPU hotplug (cpuhp) event.
1858 static int cpuset_handle_cpuhp(struct notifier_block *unused_nb,
1859 unsigned long phase, void *unused_cpu)
1861 if (phase == CPU_DYING || phase == CPU_DYING_FROZEN)
1864 common_cpu_mem_hotplug_unplug();
1868 #ifdef CONFIG_MEMORY_HOTPLUG
1870 * Keep top_cpuset.mems_allowed tracking node_states[N_HIGH_MEMORY].
1871 * Call this routine anytime after you change
1872 * node_states[N_HIGH_MEMORY].
1873 * See also the previous routine cpuset_handle_cpuhp().
1876 void cpuset_track_online_nodes(void)
1878 common_cpu_mem_hotplug_unplug();
1883 * cpuset_init_smp - initialize cpus_allowed
1885 * Description: Finish top cpuset after cpu, node maps are initialized
1888 void __init cpuset_init_smp(void)
1890 top_cpuset.cpus_allowed = cpu_online_map;
1891 top_cpuset.mems_allowed = node_states[N_HIGH_MEMORY];
1893 hotcpu_notifier(cpuset_handle_cpuhp, 0);
1898 * cpuset_cpus_allowed - return cpus_allowed mask from a tasks cpuset.
1899 * @tsk: pointer to task_struct from which to obtain cpuset->cpus_allowed.
1900 * @pmask: pointer to cpumask_t variable to receive cpus_allowed set.
1902 * Description: Returns the cpumask_t cpus_allowed of the cpuset
1903 * attached to the specified @tsk. Guaranteed to return some non-empty
1904 * subset of cpu_online_map, even if this means going outside the
1908 void cpuset_cpus_allowed(struct task_struct *tsk, cpumask_t *pmask)
1910 mutex_lock(&callback_mutex);
1911 cpuset_cpus_allowed_locked(tsk, pmask);
1912 mutex_unlock(&callback_mutex);
1916 * cpuset_cpus_allowed_locked - return cpus_allowed mask from a tasks cpuset.
1917 * Must be called with callback_mutex held.
1919 void cpuset_cpus_allowed_locked(struct task_struct *tsk, cpumask_t *pmask)
1922 guarantee_online_cpus(task_cs(tsk), pmask);
1926 void cpuset_init_current_mems_allowed(void)
1928 nodes_setall(current->mems_allowed);
1932 * cpuset_mems_allowed - return mems_allowed mask from a tasks cpuset.
1933 * @tsk: pointer to task_struct from which to obtain cpuset->mems_allowed.
1935 * Description: Returns the nodemask_t mems_allowed of the cpuset
1936 * attached to the specified @tsk. Guaranteed to return some non-empty
1937 * subset of node_states[N_HIGH_MEMORY], even if this means going outside the
1941 nodemask_t cpuset_mems_allowed(struct task_struct *tsk)
1945 mutex_lock(&callback_mutex);
1947 guarantee_online_mems(task_cs(tsk), &mask);
1949 mutex_unlock(&callback_mutex);
1955 * cpuset_nodemask_valid_mems_allowed - check nodemask vs. curremt mems_allowed
1956 * @nodemask: the nodemask to be checked
1958 * Are any of the nodes in the nodemask allowed in current->mems_allowed?
1960 int cpuset_nodemask_valid_mems_allowed(nodemask_t *nodemask)
1962 return nodes_intersects(*nodemask, current->mems_allowed);
1966 * nearest_exclusive_ancestor() - Returns the nearest mem_exclusive
1967 * ancestor to the specified cpuset. Call holding callback_mutex.
1968 * If no ancestor is mem_exclusive (an unusual configuration), then
1969 * returns the root cpuset.
1971 static const struct cpuset *nearest_exclusive_ancestor(const struct cpuset *cs)
1973 while (!is_mem_exclusive(cs) && cs->parent)
1979 * cpuset_zone_allowed_softwall - Can we allocate on zone z's memory node?
1980 * @z: is this zone on an allowed node?
1981 * @gfp_mask: memory allocation flags
1983 * If we're in interrupt, yes, we can always allocate. If
1984 * __GFP_THISNODE is set, yes, we can always allocate. If zone
1985 * z's node is in our tasks mems_allowed, yes. If it's not a
1986 * __GFP_HARDWALL request and this zone's nodes is in the nearest
1987 * mem_exclusive cpuset ancestor to this tasks cpuset, yes.
1988 * If the task has been OOM killed and has access to memory reserves
1989 * as specified by the TIF_MEMDIE flag, yes.
1992 * If __GFP_HARDWALL is set, cpuset_zone_allowed_softwall()
1993 * reduces to cpuset_zone_allowed_hardwall(). Otherwise,
1994 * cpuset_zone_allowed_softwall() might sleep, and might allow a zone
1995 * from an enclosing cpuset.
1997 * cpuset_zone_allowed_hardwall() only handles the simpler case of
1998 * hardwall cpusets, and never sleeps.
2000 * The __GFP_THISNODE placement logic is really handled elsewhere,
2001 * by forcibly using a zonelist starting at a specified node, and by
2002 * (in get_page_from_freelist()) refusing to consider the zones for
2003 * any node on the zonelist except the first. By the time any such
2004 * calls get to this routine, we should just shut up and say 'yes'.
2006 * GFP_USER allocations are marked with the __GFP_HARDWALL bit,
2007 * and do not allow allocations outside the current tasks cpuset
2008 * unless the task has been OOM killed as is marked TIF_MEMDIE.
2009 * GFP_KERNEL allocations are not so marked, so can escape to the
2010 * nearest enclosing mem_exclusive ancestor cpuset.
2012 * Scanning up parent cpusets requires callback_mutex. The
2013 * __alloc_pages() routine only calls here with __GFP_HARDWALL bit
2014 * _not_ set if it's a GFP_KERNEL allocation, and all nodes in the
2015 * current tasks mems_allowed came up empty on the first pass over
2016 * the zonelist. So only GFP_KERNEL allocations, if all nodes in the
2017 * cpuset are short of memory, might require taking the callback_mutex
2020 * The first call here from mm/page_alloc:get_page_from_freelist()
2021 * has __GFP_HARDWALL set in gfp_mask, enforcing hardwall cpusets,
2022 * so no allocation on a node outside the cpuset is allowed (unless
2023 * in interrupt, of course).
2025 * The second pass through get_page_from_freelist() doesn't even call
2026 * here for GFP_ATOMIC calls. For those calls, the __alloc_pages()
2027 * variable 'wait' is not set, and the bit ALLOC_CPUSET is not set
2028 * in alloc_flags. That logic and the checks below have the combined
2030 * in_interrupt - any node ok (current task context irrelevant)
2031 * GFP_ATOMIC - any node ok
2032 * TIF_MEMDIE - any node ok
2033 * GFP_KERNEL - any node in enclosing mem_exclusive cpuset ok
2034 * GFP_USER - only nodes in current tasks mems allowed ok.
2037 * Don't call cpuset_zone_allowed_softwall if you can't sleep, unless you
2038 * pass in the __GFP_HARDWALL flag set in gfp_flag, which disables
2039 * the code that might scan up ancestor cpusets and sleep.
2042 int __cpuset_zone_allowed_softwall(struct zone *z, gfp_t gfp_mask)
2044 int node; /* node that zone z is on */
2045 const struct cpuset *cs; /* current cpuset ancestors */
2046 int allowed; /* is allocation in zone z allowed? */
2048 if (in_interrupt() || (gfp_mask & __GFP_THISNODE))
2050 node = zone_to_nid(z);
2051 might_sleep_if(!(gfp_mask & __GFP_HARDWALL));
2052 if (node_isset(node, current->mems_allowed))
2055 * Allow tasks that have access to memory reserves because they have
2056 * been OOM killed to get memory anywhere.
2058 if (unlikely(test_thread_flag(TIF_MEMDIE)))
2060 if (gfp_mask & __GFP_HARDWALL) /* If hardwall request, stop here */
2063 if (current->flags & PF_EXITING) /* Let dying task have memory */
2066 /* Not hardwall and node outside mems_allowed: scan up cpusets */
2067 mutex_lock(&callback_mutex);
2070 cs = nearest_exclusive_ancestor(task_cs(current));
2071 task_unlock(current);
2073 allowed = node_isset(node, cs->mems_allowed);
2074 mutex_unlock(&callback_mutex);
2079 * cpuset_zone_allowed_hardwall - Can we allocate on zone z's memory node?
2080 * @z: is this zone on an allowed node?
2081 * @gfp_mask: memory allocation flags
2083 * If we're in interrupt, yes, we can always allocate.
2084 * If __GFP_THISNODE is set, yes, we can always allocate. If zone
2085 * z's node is in our tasks mems_allowed, yes. If the task has been
2086 * OOM killed and has access to memory reserves as specified by the
2087 * TIF_MEMDIE flag, yes. Otherwise, no.
2089 * The __GFP_THISNODE placement logic is really handled elsewhere,
2090 * by forcibly using a zonelist starting at a specified node, and by
2091 * (in get_page_from_freelist()) refusing to consider the zones for
2092 * any node on the zonelist except the first. By the time any such
2093 * calls get to this routine, we should just shut up and say 'yes'.
2095 * Unlike the cpuset_zone_allowed_softwall() variant, above,
2096 * this variant requires that the zone be in the current tasks
2097 * mems_allowed or that we're in interrupt. It does not scan up the
2098 * cpuset hierarchy for the nearest enclosing mem_exclusive cpuset.
2102 int __cpuset_zone_allowed_hardwall(struct zone *z, gfp_t gfp_mask)
2104 int node; /* node that zone z is on */
2106 if (in_interrupt() || (gfp_mask & __GFP_THISNODE))
2108 node = zone_to_nid(z);
2109 if (node_isset(node, current->mems_allowed))
2112 * Allow tasks that have access to memory reserves because they have
2113 * been OOM killed to get memory anywhere.
2115 if (unlikely(test_thread_flag(TIF_MEMDIE)))
2121 * cpuset_lock - lock out any changes to cpuset structures
2123 * The out of memory (oom) code needs to mutex_lock cpusets
2124 * from being changed while it scans the tasklist looking for a
2125 * task in an overlapping cpuset. Expose callback_mutex via this
2126 * cpuset_lock() routine, so the oom code can lock it, before
2127 * locking the task list. The tasklist_lock is a spinlock, so
2128 * must be taken inside callback_mutex.
2131 void cpuset_lock(void)
2133 mutex_lock(&callback_mutex);
2137 * cpuset_unlock - release lock on cpuset changes
2139 * Undo the lock taken in a previous cpuset_lock() call.
2142 void cpuset_unlock(void)
2144 mutex_unlock(&callback_mutex);
2148 * cpuset_mem_spread_node() - On which node to begin search for a page
2150 * If a task is marked PF_SPREAD_PAGE or PF_SPREAD_SLAB (as for
2151 * tasks in a cpuset with is_spread_page or is_spread_slab set),
2152 * and if the memory allocation used cpuset_mem_spread_node()
2153 * to determine on which node to start looking, as it will for
2154 * certain page cache or slab cache pages such as used for file
2155 * system buffers and inode caches, then instead of starting on the
2156 * local node to look for a free page, rather spread the starting
2157 * node around the tasks mems_allowed nodes.
2159 * We don't have to worry about the returned node being offline
2160 * because "it can't happen", and even if it did, it would be ok.
2162 * The routines calling guarantee_online_mems() are careful to
2163 * only set nodes in task->mems_allowed that are online. So it
2164 * should not be possible for the following code to return an
2165 * offline node. But if it did, that would be ok, as this routine
2166 * is not returning the node where the allocation must be, only
2167 * the node where the search should start. The zonelist passed to
2168 * __alloc_pages() will include all nodes. If the slab allocator
2169 * is passed an offline node, it will fall back to the local node.
2170 * See kmem_cache_alloc_node().
2173 int cpuset_mem_spread_node(void)
2177 node = next_node(current->cpuset_mem_spread_rotor, current->mems_allowed);
2178 if (node == MAX_NUMNODES)
2179 node = first_node(current->mems_allowed);
2180 current->cpuset_mem_spread_rotor = node;
2183 EXPORT_SYMBOL_GPL(cpuset_mem_spread_node);
2186 * cpuset_mems_allowed_intersects - Does @tsk1's mems_allowed intersect @tsk2's?
2187 * @tsk1: pointer to task_struct of some task.
2188 * @tsk2: pointer to task_struct of some other task.
2190 * Description: Return true if @tsk1's mems_allowed intersects the
2191 * mems_allowed of @tsk2. Used by the OOM killer to determine if
2192 * one of the task's memory usage might impact the memory available
2196 int cpuset_mems_allowed_intersects(const struct task_struct *tsk1,
2197 const struct task_struct *tsk2)
2199 return nodes_intersects(tsk1->mems_allowed, tsk2->mems_allowed);
2203 * Collection of memory_pressure is suppressed unless
2204 * this flag is enabled by writing "1" to the special
2205 * cpuset file 'memory_pressure_enabled' in the root cpuset.
2208 int cpuset_memory_pressure_enabled __read_mostly;
2211 * cpuset_memory_pressure_bump - keep stats of per-cpuset reclaims.
2213 * Keep a running average of the rate of synchronous (direct)
2214 * page reclaim efforts initiated by tasks in each cpuset.
2216 * This represents the rate at which some task in the cpuset
2217 * ran low on memory on all nodes it was allowed to use, and
2218 * had to enter the kernels page reclaim code in an effort to
2219 * create more free memory by tossing clean pages or swapping
2220 * or writing dirty pages.
2222 * Display to user space in the per-cpuset read-only file
2223 * "memory_pressure". Value displayed is an integer
2224 * representing the recent rate of entry into the synchronous
2225 * (direct) page reclaim by any task attached to the cpuset.
2228 void __cpuset_memory_pressure_bump(void)
2231 fmeter_markevent(&task_cs(current)->fmeter);
2232 task_unlock(current);
2235 #ifdef CONFIG_PROC_PID_CPUSET
2237 * proc_cpuset_show()
2238 * - Print tasks cpuset path into seq_file.
2239 * - Used for /proc/<pid>/cpuset.
2240 * - No need to task_lock(tsk) on this tsk->cpuset reference, as it
2241 * doesn't really matter if tsk->cpuset changes after we read it,
2242 * and we take cgroup_mutex, keeping cpuset_attach() from changing it
2245 static int proc_cpuset_show(struct seq_file *m, void *unused_v)
2248 struct task_struct *tsk;
2250 struct cgroup_subsys_state *css;
2254 buf = kmalloc(PAGE_SIZE, GFP_KERNEL);
2260 tsk = get_pid_task(pid, PIDTYPE_PID);
2266 css = task_subsys_state(tsk, cpuset_subsys_id);
2267 retval = cgroup_path(css->cgroup, buf, PAGE_SIZE);
2274 put_task_struct(tsk);
2281 static int cpuset_open(struct inode *inode, struct file *file)
2283 struct pid *pid = PROC_I(inode)->pid;
2284 return single_open(file, proc_cpuset_show, pid);
2287 const struct file_operations proc_cpuset_operations = {
2288 .open = cpuset_open,
2290 .llseek = seq_lseek,
2291 .release = single_release,
2293 #endif /* CONFIG_PROC_PID_CPUSET */
2295 /* Display task cpus_allowed, mems_allowed in /proc/<pid>/status file. */
2296 void cpuset_task_status_allowed(struct seq_file *m, struct task_struct *task)
2298 seq_printf(m, "Cpus_allowed:\t");
2299 m->count += cpumask_scnprintf(m->buf + m->count, m->size - m->count,
2300 task->cpus_allowed);
2301 seq_printf(m, "\n");
2302 seq_printf(m, "Cpus_allowed_list:\t");
2303 m->count += cpulist_scnprintf(m->buf + m->count, m->size - m->count,
2304 task->cpus_allowed);
2305 seq_printf(m, "\n");
2306 seq_printf(m, "Mems_allowed:\t");
2307 m->count += nodemask_scnprintf(m->buf + m->count, m->size - m->count,
2308 task->mems_allowed);
2309 seq_printf(m, "\n");
2310 seq_printf(m, "Mems_allowed_list:\t");
2311 m->count += nodelist_scnprintf(m->buf + m->count, m->size - m->count,
2312 task->mems_allowed);
2313 seq_printf(m, "\n");