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
17 * 2008 Rework of the scheduler domains and CPU hotplug handling
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.
25 #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>
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>
37 #include <linux/mempolicy.h>
39 #include <linux/memory.h>
40 #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>
45 #include <linux/rcupdate.h>
46 #include <linux/sched.h>
47 #include <linux/seq_file.h>
48 #include <linux/security.h>
49 #include <linux/slab.h>
50 #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>
57 #include <asm/uaccess.h>
58 #include <asm/atomic.h>
59 #include <linux/mutex.h>
60 #include <linux/workqueue.h>
61 #include <linux/cgroup.h>
64 * Workqueue for cpuset related tasks.
66 * Using kevent workqueue may cause deadlock when memory_migrate
67 * is set. So we create a separate workqueue thread for cpuset.
69 static struct workqueue_struct *cpuset_wq;
72 * 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.
76 int number_of_cpusets __read_mostly;
78 /* Forward declare cgroup structures */
79 struct cgroup_subsys cpuset_subsys;
82 /* See "Frequency meter" comments, below. */
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 */
92 struct cgroup_subsys_state css;
94 unsigned long flags; /* "unsigned long" so bitops work */
95 cpumask_var_t cpus_allowed; /* CPUs allowed to tasks in cpuset */
96 nodemask_t mems_allowed; /* Memory Nodes allowed to tasks */
98 struct cpuset *parent; /* my parent */
101 * Copy of global cpuset_mems_generation as of the most
102 * recent time this cpuset changed its mems_allowed.
106 struct fmeter fmeter; /* memory_pressure filter */
108 /* partition number for rebuild_sched_domains() */
111 /* for custom sched domain */
112 int relax_domain_level;
114 /* used for walking a cpuset heirarchy */
115 struct list_head stack_list;
118 /* Retrieve the cpuset for a cgroup */
119 static inline struct cpuset *cgroup_cs(struct cgroup *cont)
121 return container_of(cgroup_subsys_state(cont, cpuset_subsys_id),
125 /* Retrieve the cpuset for a task */
126 static inline struct cpuset *task_cs(struct task_struct *task)
128 return container_of(task_subsys_state(task, cpuset_subsys_id),
132 /* bits in struct cpuset flags field */
138 CS_SCHED_LOAD_BALANCE,
143 /* convenient tests for these bits */
144 static inline int is_cpu_exclusive(const struct cpuset *cs)
146 return test_bit(CS_CPU_EXCLUSIVE, &cs->flags);
149 static inline int is_mem_exclusive(const struct cpuset *cs)
151 return test_bit(CS_MEM_EXCLUSIVE, &cs->flags);
154 static inline int is_mem_hardwall(const struct cpuset *cs)
156 return test_bit(CS_MEM_HARDWALL, &cs->flags);
159 static inline int is_sched_load_balance(const struct cpuset *cs)
161 return test_bit(CS_SCHED_LOAD_BALANCE, &cs->flags);
164 static inline int is_memory_migrate(const struct cpuset *cs)
166 return test_bit(CS_MEMORY_MIGRATE, &cs->flags);
169 static inline int is_spread_page(const struct cpuset *cs)
171 return test_bit(CS_SPREAD_PAGE, &cs->flags);
174 static inline int is_spread_slab(const struct cpuset *cs)
176 return test_bit(CS_SPREAD_SLAB, &cs->flags);
180 * Increment this integer everytime any cpuset changes its
181 * mems_allowed value. Users of cpusets can track this generation
182 * number, and avoid having to lock and reload mems_allowed unless
183 * the cpuset they're using changes generation.
185 * A single, global generation is needed because cpuset_attach_task() could
186 * reattach a task to a different cpuset, which must not have its
187 * generation numbers aliased with those of that tasks previous cpuset.
189 * Generations are needed for mems_allowed because one task cannot
190 * modify another's memory placement. So we must enable every task,
191 * on every visit to __alloc_pages(), to efficiently check whether
192 * its current->cpuset->mems_allowed has changed, requiring an update
193 * of its current->mems_allowed.
195 * Since writes to cpuset_mems_generation are guarded by the cgroup lock
196 * there is no need to mark it atomic.
198 static int cpuset_mems_generation;
200 static struct cpuset top_cpuset = {
201 .flags = ((1 << CS_CPU_EXCLUSIVE) | (1 << CS_MEM_EXCLUSIVE)),
205 * There are two global mutexes guarding cpuset structures. The first
206 * is the main control groups cgroup_mutex, accessed via
207 * cgroup_lock()/cgroup_unlock(). The second is the cpuset-specific
208 * callback_mutex, below. They can nest. It is ok to first take
209 * cgroup_mutex, then nest callback_mutex. We also require taking
210 * task_lock() when dereferencing a task's cpuset pointer. See "The
211 * task_lock() exception", at the end of this comment.
213 * A task must hold both mutexes to modify cpusets. If a task
214 * holds cgroup_mutex, then it blocks others wanting that mutex,
215 * ensuring that it is the only task able to also acquire callback_mutex
216 * and be able to modify cpusets. It can perform various checks on
217 * the cpuset structure first, knowing nothing will change. It can
218 * also allocate memory while just holding cgroup_mutex. While it is
219 * performing these checks, various callback routines can briefly
220 * acquire callback_mutex to query cpusets. Once it is ready to make
221 * the changes, it takes callback_mutex, blocking everyone else.
223 * Calls to the kernel memory allocator can not be made while holding
224 * callback_mutex, as that would risk double tripping on callback_mutex
225 * from one of the callbacks into the cpuset code from within
228 * If a task is only holding callback_mutex, then it has read-only
231 * The task_struct fields mems_allowed and mems_generation may only
232 * be accessed in the context of that task, so require no locks.
234 * The cpuset_common_file_read() handlers only hold callback_mutex across
235 * small pieces of code, such as when reading out possibly multi-word
236 * cpumasks and nodemasks.
238 * Accessing a task's cpuset should be done in accordance with the
239 * guidelines for accessing subsystem state in kernel/cgroup.c
242 static DEFINE_MUTEX(callback_mutex);
245 * cpuset_buffer_lock protects both the cpuset_name and cpuset_nodelist
246 * buffers. They are statically allocated to prevent using excess stack
247 * when calling cpuset_print_task_mems_allowed().
249 #define CPUSET_NAME_LEN (128)
250 #define CPUSET_NODELIST_LEN (256)
251 static char cpuset_name[CPUSET_NAME_LEN];
252 static char cpuset_nodelist[CPUSET_NODELIST_LEN];
253 static DEFINE_SPINLOCK(cpuset_buffer_lock);
256 * This is ugly, but preserves the userspace API for existing cpuset
257 * users. If someone tries to mount the "cpuset" filesystem, we
258 * silently switch it to mount "cgroup" instead
260 static int cpuset_get_sb(struct file_system_type *fs_type,
261 int flags, const char *unused_dev_name,
262 void *data, struct vfsmount *mnt)
264 struct file_system_type *cgroup_fs = get_fs_type("cgroup");
269 "release_agent=/sbin/cpuset_release_agent";
270 ret = cgroup_fs->get_sb(cgroup_fs, flags,
271 unused_dev_name, mountopts, mnt);
272 put_filesystem(cgroup_fs);
277 static struct file_system_type cpuset_fs_type = {
279 .get_sb = cpuset_get_sb,
283 * Return in pmask the portion of a cpusets's cpus_allowed that
284 * are online. If none are online, walk up the cpuset hierarchy
285 * until we find one that does have some online cpus. If we get
286 * all the way to the top and still haven't found any online cpus,
287 * return cpu_online_map. Or if passed a NULL cs from an exit'ing
288 * task, return cpu_online_map.
290 * One way or another, we guarantee to return some non-empty subset
293 * Call with callback_mutex held.
296 static void guarantee_online_cpus(const struct cpuset *cs,
297 struct cpumask *pmask)
299 while (cs && !cpumask_intersects(cs->cpus_allowed, cpu_online_mask))
302 cpumask_and(pmask, cs->cpus_allowed, cpu_online_mask);
304 cpumask_copy(pmask, cpu_online_mask);
305 BUG_ON(!cpumask_intersects(pmask, cpu_online_mask));
309 * Return in *pmask the portion of a cpusets's mems_allowed that
310 * are online, with memory. If none are online with memory, walk
311 * up the cpuset hierarchy until we find one that does have some
312 * online mems. If we get all the way to the top and still haven't
313 * found any online mems, return node_states[N_HIGH_MEMORY].
315 * One way or another, we guarantee to return some non-empty subset
316 * of node_states[N_HIGH_MEMORY].
318 * Call with callback_mutex held.
321 static void guarantee_online_mems(const struct cpuset *cs, nodemask_t *pmask)
323 while (cs && !nodes_intersects(cs->mems_allowed,
324 node_states[N_HIGH_MEMORY]))
327 nodes_and(*pmask, cs->mems_allowed,
328 node_states[N_HIGH_MEMORY]);
330 *pmask = node_states[N_HIGH_MEMORY];
331 BUG_ON(!nodes_intersects(*pmask, node_states[N_HIGH_MEMORY]));
335 * update task's spread flag if cpuset's page/slab spread flag is set
337 * Called with callback_mutex/cgroup_mutex held
339 static void cpuset_update_task_spread_flag(struct cpuset *cs,
340 struct task_struct *tsk)
342 if (is_spread_page(cs))
343 tsk->flags |= PF_SPREAD_PAGE;
345 tsk->flags &= ~PF_SPREAD_PAGE;
346 if (is_spread_slab(cs))
347 tsk->flags |= PF_SPREAD_SLAB;
349 tsk->flags &= ~PF_SPREAD_SLAB;
353 * cpuset_update_task_memory_state - update task memory placement
355 * If the current tasks cpusets mems_allowed changed behind our
356 * backs, update current->mems_allowed, mems_generation and task NUMA
357 * mempolicy to the new value.
359 * Task mempolicy is updated by rebinding it relative to the
360 * current->cpuset if a task has its memory placement changed.
361 * Do not call this routine if in_interrupt().
363 * Call without callback_mutex or task_lock() held. May be
364 * called with or without cgroup_mutex held. Thanks in part to
365 * 'the_top_cpuset_hack', the task's cpuset pointer will never
366 * be NULL. This routine also might acquire callback_mutex during
369 * Reading current->cpuset->mems_generation doesn't need task_lock
370 * to guard the current->cpuset derefence, because it is guarded
371 * from concurrent freeing of current->cpuset using RCU.
373 * The rcu_dereference() is technically probably not needed,
374 * as I don't actually mind if I see a new cpuset pointer but
375 * an old value of mems_generation. However this really only
376 * matters on alpha systems using cpusets heavily. If I dropped
377 * that rcu_dereference(), it would save them a memory barrier.
378 * For all other arch's, rcu_dereference is a no-op anyway, and for
379 * alpha systems not using cpusets, another planned optimization,
380 * avoiding the rcu critical section for tasks in the root cpuset
381 * which is statically allocated, so can't vanish, will make this
382 * irrelevant. Better to use RCU as intended, than to engage in
383 * some cute trick to save a memory barrier that is impossible to
384 * test, for alpha systems using cpusets heavily, which might not
387 * This routine is needed to update the per-task mems_allowed data,
388 * within the tasks context, when it is trying to allocate memory
389 * (in various mm/mempolicy.c routines) and notices that some other
390 * task has been modifying its cpuset.
393 void cpuset_update_task_memory_state(void)
395 int my_cpusets_mem_gen;
396 struct task_struct *tsk = current;
400 my_cpusets_mem_gen = task_cs(tsk)->mems_generation;
403 if (my_cpusets_mem_gen != tsk->cpuset_mems_generation) {
404 mutex_lock(&callback_mutex);
406 cs = task_cs(tsk); /* Maybe changed when task not locked */
407 guarantee_online_mems(cs, &tsk->mems_allowed);
408 tsk->cpuset_mems_generation = cs->mems_generation;
409 cpuset_update_task_spread_flag(cs, tsk);
411 mutex_unlock(&callback_mutex);
412 mpol_rebind_task(tsk, &tsk->mems_allowed);
417 * is_cpuset_subset(p, q) - Is cpuset p a subset of cpuset q?
419 * One cpuset is a subset of another if all its allowed CPUs and
420 * Memory Nodes are a subset of the other, and its exclusive flags
421 * are only set if the other's are set. Call holding cgroup_mutex.
424 static int is_cpuset_subset(const struct cpuset *p, const struct cpuset *q)
426 return cpumask_subset(p->cpus_allowed, q->cpus_allowed) &&
427 nodes_subset(p->mems_allowed, q->mems_allowed) &&
428 is_cpu_exclusive(p) <= is_cpu_exclusive(q) &&
429 is_mem_exclusive(p) <= is_mem_exclusive(q);
433 * alloc_trial_cpuset - allocate a trial cpuset
434 * @cs: the cpuset that the trial cpuset duplicates
436 static struct cpuset *alloc_trial_cpuset(const struct cpuset *cs)
438 struct cpuset *trial;
440 trial = kmemdup(cs, sizeof(*cs), GFP_KERNEL);
444 if (!alloc_cpumask_var(&trial->cpus_allowed, GFP_KERNEL)) {
448 cpumask_copy(trial->cpus_allowed, cs->cpus_allowed);
454 * free_trial_cpuset - free the trial cpuset
455 * @trial: the trial cpuset to be freed
457 static void free_trial_cpuset(struct cpuset *trial)
459 free_cpumask_var(trial->cpus_allowed);
464 * validate_change() - Used to validate that any proposed cpuset change
465 * follows the structural rules for cpusets.
467 * If we replaced the flag and mask values of the current cpuset
468 * (cur) with those values in the trial cpuset (trial), would
469 * our various subset and exclusive rules still be valid? Presumes
472 * 'cur' is the address of an actual, in-use cpuset. Operations
473 * such as list traversal that depend on the actual address of the
474 * cpuset in the list must use cur below, not trial.
476 * 'trial' is the address of bulk structure copy of cur, with
477 * perhaps one or more of the fields cpus_allowed, mems_allowed,
478 * or flags changed to new, trial values.
480 * Return 0 if valid, -errno if not.
483 static int validate_change(const struct cpuset *cur, const struct cpuset *trial)
486 struct cpuset *c, *par;
488 /* Each of our child cpusets must be a subset of us */
489 list_for_each_entry(cont, &cur->css.cgroup->children, sibling) {
490 if (!is_cpuset_subset(cgroup_cs(cont), trial))
494 /* Remaining checks don't apply to root cpuset */
495 if (cur == &top_cpuset)
500 /* We must be a subset of our parent cpuset */
501 if (!is_cpuset_subset(trial, par))
505 * If either I or some sibling (!= me) is exclusive, we can't
508 list_for_each_entry(cont, &par->css.cgroup->children, sibling) {
510 if ((is_cpu_exclusive(trial) || is_cpu_exclusive(c)) &&
512 cpumask_intersects(trial->cpus_allowed, c->cpus_allowed))
514 if ((is_mem_exclusive(trial) || is_mem_exclusive(c)) &&
516 nodes_intersects(trial->mems_allowed, c->mems_allowed))
520 /* Cpusets with tasks can't have empty cpus_allowed or mems_allowed */
521 if (cgroup_task_count(cur->css.cgroup)) {
522 if (cpumask_empty(trial->cpus_allowed) ||
523 nodes_empty(trial->mems_allowed)) {
533 * Helper routine for generate_sched_domains().
534 * Do cpusets a, b have overlapping cpus_allowed masks?
536 static int cpusets_overlap(struct cpuset *a, struct cpuset *b)
538 return cpumask_intersects(a->cpus_allowed, b->cpus_allowed);
542 update_domain_attr(struct sched_domain_attr *dattr, struct cpuset *c)
544 if (dattr->relax_domain_level < c->relax_domain_level)
545 dattr->relax_domain_level = c->relax_domain_level;
550 update_domain_attr_tree(struct sched_domain_attr *dattr, struct cpuset *c)
554 list_add(&c->stack_list, &q);
555 while (!list_empty(&q)) {
558 struct cpuset *child;
560 cp = list_first_entry(&q, struct cpuset, stack_list);
563 if (cpumask_empty(cp->cpus_allowed))
566 if (is_sched_load_balance(cp))
567 update_domain_attr(dattr, cp);
569 list_for_each_entry(cont, &cp->css.cgroup->children, sibling) {
570 child = cgroup_cs(cont);
571 list_add_tail(&child->stack_list, &q);
577 * generate_sched_domains()
579 * This function builds a partial partition of the systems CPUs
580 * A 'partial partition' is a set of non-overlapping subsets whose
581 * union is a subset of that set.
582 * The output of this function needs to be passed to kernel/sched.c
583 * partition_sched_domains() routine, which will rebuild the scheduler's
584 * load balancing domains (sched domains) as specified by that partial
587 * See "What is sched_load_balance" in Documentation/cgroups/cpusets.txt
588 * for a background explanation of this.
590 * Does not return errors, on the theory that the callers of this
591 * routine would rather not worry about failures to rebuild sched
592 * domains when operating in the severe memory shortage situations
593 * that could cause allocation failures below.
595 * Must be called with cgroup_lock held.
597 * The three key local variables below are:
598 * q - a linked-list queue of cpuset pointers, used to implement a
599 * top-down scan of all cpusets. This scan loads a pointer
600 * to each cpuset marked is_sched_load_balance into the
601 * array 'csa'. For our purposes, rebuilding the schedulers
602 * sched domains, we can ignore !is_sched_load_balance cpusets.
603 * csa - (for CpuSet Array) Array of pointers to all the cpusets
604 * that need to be load balanced, for convenient iterative
605 * access by the subsequent code that finds the best partition,
606 * i.e the set of domains (subsets) of CPUs such that the
607 * cpus_allowed of every cpuset marked is_sched_load_balance
608 * is a subset of one of these domains, while there are as
609 * many such domains as possible, each as small as possible.
610 * doms - Conversion of 'csa' to an array of cpumasks, for passing to
611 * the kernel/sched.c routine partition_sched_domains() in a
612 * convenient format, that can be easily compared to the prior
613 * value to determine what partition elements (sched domains)
614 * were changed (added or removed.)
616 * Finding the best partition (set of domains):
617 * The triple nested loops below over i, j, k scan over the
618 * load balanced cpusets (using the array of cpuset pointers in
619 * csa[]) looking for pairs of cpusets that have overlapping
620 * cpus_allowed, but which don't have the same 'pn' partition
621 * number and gives them in the same partition number. It keeps
622 * looping on the 'restart' label until it can no longer find
625 * The union of the cpus_allowed masks from the set of
626 * all cpusets having the same 'pn' value then form the one
627 * element of the partition (one sched domain) to be passed to
628 * partition_sched_domains().
630 /* FIXME: see the FIXME in partition_sched_domains() */
631 static int generate_sched_domains(struct cpumask **domains,
632 struct sched_domain_attr **attributes)
634 LIST_HEAD(q); /* queue of cpusets to be scanned */
635 struct cpuset *cp; /* scans q */
636 struct cpuset **csa; /* array of all cpuset ptrs */
637 int csn; /* how many cpuset ptrs in csa so far */
638 int i, j, k; /* indices for partition finding loops */
639 struct cpumask *doms; /* resulting partition; i.e. sched domains */
640 struct sched_domain_attr *dattr; /* attributes for custom domains */
641 int ndoms = 0; /* number of sched domains in result */
642 int nslot; /* next empty doms[] struct cpumask slot */
648 /* Special case for the 99% of systems with one, full, sched domain */
649 if (is_sched_load_balance(&top_cpuset)) {
650 doms = kmalloc(cpumask_size(), GFP_KERNEL);
654 dattr = kmalloc(sizeof(struct sched_domain_attr), GFP_KERNEL);
656 *dattr = SD_ATTR_INIT;
657 update_domain_attr_tree(dattr, &top_cpuset);
659 cpumask_copy(doms, top_cpuset.cpus_allowed);
665 csa = kmalloc(number_of_cpusets * sizeof(cp), GFP_KERNEL);
670 list_add(&top_cpuset.stack_list, &q);
671 while (!list_empty(&q)) {
673 struct cpuset *child; /* scans child cpusets of cp */
675 cp = list_first_entry(&q, struct cpuset, stack_list);
678 if (cpumask_empty(cp->cpus_allowed))
682 * All child cpusets contain a subset of the parent's cpus, so
683 * just skip them, and then we call update_domain_attr_tree()
684 * to calc relax_domain_level of the corresponding sched
687 if (is_sched_load_balance(cp)) {
692 list_for_each_entry(cont, &cp->css.cgroup->children, sibling) {
693 child = cgroup_cs(cont);
694 list_add_tail(&child->stack_list, &q);
698 for (i = 0; i < csn; i++)
703 /* Find the best partition (set of sched domains) */
704 for (i = 0; i < csn; i++) {
705 struct cpuset *a = csa[i];
708 for (j = 0; j < csn; j++) {
709 struct cpuset *b = csa[j];
712 if (apn != bpn && cpusets_overlap(a, b)) {
713 for (k = 0; k < csn; k++) {
714 struct cpuset *c = csa[k];
719 ndoms--; /* one less element */
726 * Now we know how many domains to create.
727 * Convert <csn, csa> to <ndoms, doms> and populate cpu masks.
729 doms = kmalloc(ndoms * cpumask_size(), GFP_KERNEL);
734 * The rest of the code, including the scheduler, can deal with
735 * dattr==NULL case. No need to abort if alloc fails.
737 dattr = kmalloc(ndoms * sizeof(struct sched_domain_attr), GFP_KERNEL);
739 for (nslot = 0, i = 0; i < csn; i++) {
740 struct cpuset *a = csa[i];
745 /* Skip completed partitions */
751 if (nslot == ndoms) {
752 static int warnings = 10;
755 "rebuild_sched_domains confused:"
756 " nslot %d, ndoms %d, csn %d, i %d,"
758 nslot, ndoms, csn, i, apn);
766 *(dattr + nslot) = SD_ATTR_INIT;
767 for (j = i; j < csn; j++) {
768 struct cpuset *b = csa[j];
771 cpumask_or(dp, dp, b->cpus_allowed);
773 update_domain_attr_tree(dattr + nslot, b);
775 /* Done with this partition */
781 BUG_ON(nslot != ndoms);
787 * Fallback to the default domain if kmalloc() failed.
788 * See comments in partition_sched_domains().
799 * Rebuild scheduler domains.
801 * Call with neither cgroup_mutex held nor within get_online_cpus().
802 * Takes both cgroup_mutex and get_online_cpus().
804 * Cannot be directly called from cpuset code handling changes
805 * to the cpuset pseudo-filesystem, because it cannot be called
806 * from code that already holds cgroup_mutex.
808 static void do_rebuild_sched_domains(struct work_struct *unused)
810 struct sched_domain_attr *attr;
811 struct cpumask *doms;
816 /* Generate domain masks and attrs */
818 ndoms = generate_sched_domains(&doms, &attr);
821 /* Have scheduler rebuild the domains */
822 partition_sched_domains(ndoms, doms, attr);
826 #else /* !CONFIG_SMP */
827 static void do_rebuild_sched_domains(struct work_struct *unused)
831 static int generate_sched_domains(struct cpumask **domains,
832 struct sched_domain_attr **attributes)
837 #endif /* CONFIG_SMP */
839 static DECLARE_WORK(rebuild_sched_domains_work, do_rebuild_sched_domains);
842 * Rebuild scheduler domains, asynchronously via workqueue.
844 * If the flag 'sched_load_balance' of any cpuset with non-empty
845 * 'cpus' changes, or if the 'cpus' allowed changes in any cpuset
846 * which has that flag enabled, or if any cpuset with a non-empty
847 * 'cpus' is removed, then call this routine to rebuild the
848 * scheduler's dynamic sched domains.
850 * The rebuild_sched_domains() and partition_sched_domains()
851 * routines must nest cgroup_lock() inside get_online_cpus(),
852 * but such cpuset changes as these must nest that locking the
853 * other way, holding cgroup_lock() for much of the code.
855 * So in order to avoid an ABBA deadlock, the cpuset code handling
856 * these user changes delegates the actual sched domain rebuilding
857 * to a separate workqueue thread, which ends up processing the
858 * above do_rebuild_sched_domains() function.
860 static void async_rebuild_sched_domains(void)
862 queue_work(cpuset_wq, &rebuild_sched_domains_work);
866 * Accomplishes the same scheduler domain rebuild as the above
867 * async_rebuild_sched_domains(), however it directly calls the
868 * rebuild routine synchronously rather than calling it via an
869 * asynchronous work thread.
871 * This can only be called from code that is not holding
872 * cgroup_mutex (not nested in a cgroup_lock() call.)
874 void rebuild_sched_domains(void)
876 do_rebuild_sched_domains(NULL);
880 * cpuset_test_cpumask - test a task's cpus_allowed versus its cpuset's
882 * @scan: struct cgroup_scanner contained in its struct cpuset_hotplug_scanner
884 * Call with cgroup_mutex held. May take callback_mutex during call.
885 * Called for each task in a cgroup by cgroup_scan_tasks().
886 * Return nonzero if this tasks's cpus_allowed mask should be changed (in other
887 * words, if its mask is not equal to its cpuset's mask).
889 static int cpuset_test_cpumask(struct task_struct *tsk,
890 struct cgroup_scanner *scan)
892 return !cpumask_equal(&tsk->cpus_allowed,
893 (cgroup_cs(scan->cg))->cpus_allowed);
897 * cpuset_change_cpumask - make a task's cpus_allowed the same as its cpuset's
899 * @scan: struct cgroup_scanner containing the cgroup of the task
901 * Called by cgroup_scan_tasks() for each task in a cgroup whose
902 * cpus_allowed mask needs to be changed.
904 * We don't need to re-check for the cgroup/cpuset membership, since we're
905 * holding cgroup_lock() at this point.
907 static void cpuset_change_cpumask(struct task_struct *tsk,
908 struct cgroup_scanner *scan)
910 set_cpus_allowed_ptr(tsk, ((cgroup_cs(scan->cg))->cpus_allowed));
914 * update_tasks_cpumask - Update the cpumasks of tasks in the cpuset.
915 * @cs: the cpuset in which each task's cpus_allowed mask needs to be changed
916 * @heap: if NULL, defer allocating heap memory to cgroup_scan_tasks()
918 * Called with cgroup_mutex held
920 * The cgroup_scan_tasks() function will scan all the tasks in a cgroup,
921 * calling callback functions for each.
923 * No return value. It's guaranteed that cgroup_scan_tasks() always returns 0
926 static void update_tasks_cpumask(struct cpuset *cs, struct ptr_heap *heap)
928 struct cgroup_scanner scan;
930 scan.cg = cs->css.cgroup;
931 scan.test_task = cpuset_test_cpumask;
932 scan.process_task = cpuset_change_cpumask;
934 cgroup_scan_tasks(&scan);
938 * update_cpumask - update the cpus_allowed mask of a cpuset and all tasks in it
939 * @cs: the cpuset to consider
940 * @buf: buffer of cpu numbers written to this cpuset
942 static int update_cpumask(struct cpuset *cs, struct cpuset *trialcs,
945 struct ptr_heap heap;
947 int is_load_balanced;
949 /* top_cpuset.cpus_allowed tracks cpu_online_map; it's read-only */
950 if (cs == &top_cpuset)
954 * An empty cpus_allowed is ok only if the cpuset has no tasks.
955 * Since cpulist_parse() fails on an empty mask, we special case
956 * that parsing. The validate_change() call ensures that cpusets
957 * with tasks have cpus.
960 cpumask_clear(trialcs->cpus_allowed);
962 retval = cpulist_parse(buf, trialcs->cpus_allowed);
966 if (!cpumask_subset(trialcs->cpus_allowed, cpu_online_mask))
969 retval = validate_change(cs, trialcs);
973 /* Nothing to do if the cpus didn't change */
974 if (cpumask_equal(cs->cpus_allowed, trialcs->cpus_allowed))
977 retval = heap_init(&heap, PAGE_SIZE, GFP_KERNEL, NULL);
981 is_load_balanced = is_sched_load_balance(trialcs);
983 mutex_lock(&callback_mutex);
984 cpumask_copy(cs->cpus_allowed, trialcs->cpus_allowed);
985 mutex_unlock(&callback_mutex);
988 * Scan tasks in the cpuset, and update the cpumasks of any
989 * that need an update.
991 update_tasks_cpumask(cs, &heap);
995 if (is_load_balanced)
996 async_rebuild_sched_domains();
1003 * Migrate memory region from one set of nodes to another.
1005 * Temporarilly set tasks mems_allowed to target nodes of migration,
1006 * so that the migration code can allocate pages on these nodes.
1008 * Call holding cgroup_mutex, so current's cpuset won't change
1009 * during this call, as manage_mutex holds off any cpuset_attach()
1010 * calls. Therefore we don't need to take task_lock around the
1011 * call to guarantee_online_mems(), as we know no one is changing
1012 * our task's cpuset.
1014 * Hold callback_mutex around the two modifications of our tasks
1015 * mems_allowed to synchronize with cpuset_mems_allowed().
1017 * While the mm_struct we are migrating is typically from some
1018 * other task, the task_struct mems_allowed that we are hacking
1019 * is for our current task, which must allocate new pages for that
1020 * migrating memory region.
1022 * We call cpuset_update_task_memory_state() before hacking
1023 * our tasks mems_allowed, so that we are assured of being in
1024 * sync with our tasks cpuset, and in particular, callbacks to
1025 * cpuset_update_task_memory_state() from nested page allocations
1026 * won't see any mismatch of our cpuset and task mems_generation
1027 * values, so won't overwrite our hacked tasks mems_allowed
1031 static void cpuset_migrate_mm(struct mm_struct *mm, const nodemask_t *from,
1032 const nodemask_t *to)
1034 struct task_struct *tsk = current;
1036 cpuset_update_task_memory_state();
1038 mutex_lock(&callback_mutex);
1039 tsk->mems_allowed = *to;
1040 mutex_unlock(&callback_mutex);
1042 do_migrate_pages(mm, from, to, MPOL_MF_MOVE_ALL);
1044 mutex_lock(&callback_mutex);
1045 guarantee_online_mems(task_cs(tsk),&tsk->mems_allowed);
1046 mutex_unlock(&callback_mutex);
1050 * Rebind task's vmas to cpuset's new mems_allowed, and migrate pages to new
1051 * nodes if memory_migrate flag is set. Called with cgroup_mutex held.
1053 static void cpuset_change_nodemask(struct task_struct *p,
1054 struct cgroup_scanner *scan)
1056 struct mm_struct *mm;
1059 const nodemask_t *oldmem = scan->data;
1061 mm = get_task_mm(p);
1065 cs = cgroup_cs(scan->cg);
1066 migrate = is_memory_migrate(cs);
1068 mpol_rebind_mm(mm, &cs->mems_allowed);
1070 cpuset_migrate_mm(mm, oldmem, &cs->mems_allowed);
1074 static void *cpuset_being_rebound;
1077 * update_tasks_nodemask - Update the nodemasks of tasks in the cpuset.
1078 * @cs: the cpuset in which each task's mems_allowed mask needs to be changed
1079 * @oldmem: old mems_allowed of cpuset cs
1080 * @heap: if NULL, defer allocating heap memory to cgroup_scan_tasks()
1082 * Called with cgroup_mutex held
1083 * No return value. It's guaranteed that cgroup_scan_tasks() always returns 0
1086 static void update_tasks_nodemask(struct cpuset *cs, const nodemask_t *oldmem,
1087 struct ptr_heap *heap)
1089 struct cgroup_scanner scan;
1091 cpuset_being_rebound = cs; /* causes mpol_dup() rebind */
1093 scan.cg = cs->css.cgroup;
1094 scan.test_task = NULL;
1095 scan.process_task = cpuset_change_nodemask;
1097 scan.data = (nodemask_t *)oldmem;
1100 * The mpol_rebind_mm() call takes mmap_sem, which we couldn't
1101 * take while holding tasklist_lock. Forks can happen - the
1102 * mpol_dup() cpuset_being_rebound check will catch such forks,
1103 * and rebind their vma mempolicies too. Because we still hold
1104 * the global cgroup_mutex, we know that no other rebind effort
1105 * will be contending for the global variable cpuset_being_rebound.
1106 * It's ok if we rebind the same mm twice; mpol_rebind_mm()
1107 * is idempotent. Also migrate pages in each mm to new nodes.
1109 cgroup_scan_tasks(&scan);
1111 /* We're done rebinding vmas to this cpuset's new mems_allowed. */
1112 cpuset_being_rebound = NULL;
1116 * Handle user request to change the 'mems' memory placement
1117 * of a cpuset. Needs to validate the request, update the
1118 * cpusets mems_allowed and mems_generation, and for each
1119 * task in the cpuset, rebind any vma mempolicies and if
1120 * the cpuset is marked 'memory_migrate', migrate the tasks
1121 * pages to the new memory.
1123 * Call with cgroup_mutex held. May take callback_mutex during call.
1124 * Will take tasklist_lock, scan tasklist for tasks in cpuset cs,
1125 * lock each such tasks mm->mmap_sem, scan its vma's and rebind
1126 * their mempolicies to the cpusets new mems_allowed.
1128 static int update_nodemask(struct cpuset *cs, struct cpuset *trialcs,
1133 struct ptr_heap heap;
1136 * top_cpuset.mems_allowed tracks node_stats[N_HIGH_MEMORY];
1139 if (cs == &top_cpuset)
1143 * An empty mems_allowed is ok iff there are no tasks in the cpuset.
1144 * Since nodelist_parse() fails on an empty mask, we special case
1145 * that parsing. The validate_change() call ensures that cpusets
1146 * with tasks have memory.
1149 nodes_clear(trialcs->mems_allowed);
1151 retval = nodelist_parse(buf, trialcs->mems_allowed);
1155 if (!nodes_subset(trialcs->mems_allowed,
1156 node_states[N_HIGH_MEMORY]))
1159 oldmem = cs->mems_allowed;
1160 if (nodes_equal(oldmem, trialcs->mems_allowed)) {
1161 retval = 0; /* Too easy - nothing to do */
1164 retval = validate_change(cs, trialcs);
1168 retval = heap_init(&heap, PAGE_SIZE, GFP_KERNEL, NULL);
1172 mutex_lock(&callback_mutex);
1173 cs->mems_allowed = trialcs->mems_allowed;
1174 cs->mems_generation = cpuset_mems_generation++;
1175 mutex_unlock(&callback_mutex);
1177 update_tasks_nodemask(cs, &oldmem, &heap);
1184 int current_cpuset_is_being_rebound(void)
1186 return task_cs(current) == cpuset_being_rebound;
1189 static int update_relax_domain_level(struct cpuset *cs, s64 val)
1192 if (val < -1 || val >= SD_LV_MAX)
1196 if (val != cs->relax_domain_level) {
1197 cs->relax_domain_level = val;
1198 if (!cpumask_empty(cs->cpus_allowed) &&
1199 is_sched_load_balance(cs))
1200 async_rebuild_sched_domains();
1207 * update_flag - read a 0 or a 1 in a file and update associated flag
1208 * bit: the bit to update (see cpuset_flagbits_t)
1209 * cs: the cpuset to update
1210 * turning_on: whether the flag is being set or cleared
1212 * Call with cgroup_mutex held.
1215 static int update_flag(cpuset_flagbits_t bit, struct cpuset *cs,
1218 struct cpuset *trialcs;
1220 int balance_flag_changed;
1222 trialcs = alloc_trial_cpuset(cs);
1227 set_bit(bit, &trialcs->flags);
1229 clear_bit(bit, &trialcs->flags);
1231 err = validate_change(cs, trialcs);
1235 balance_flag_changed = (is_sched_load_balance(cs) !=
1236 is_sched_load_balance(trialcs));
1238 mutex_lock(&callback_mutex);
1239 cs->flags = trialcs->flags;
1240 mutex_unlock(&callback_mutex);
1242 if (!cpumask_empty(trialcs->cpus_allowed) && balance_flag_changed)
1243 async_rebuild_sched_domains();
1246 free_trial_cpuset(trialcs);
1251 * Frequency meter - How fast is some event occurring?
1253 * These routines manage a digitally filtered, constant time based,
1254 * event frequency meter. There are four routines:
1255 * fmeter_init() - initialize a frequency meter.
1256 * fmeter_markevent() - called each time the event happens.
1257 * fmeter_getrate() - returns the recent rate of such events.
1258 * fmeter_update() - internal routine used to update fmeter.
1260 * A common data structure is passed to each of these routines,
1261 * which is used to keep track of the state required to manage the
1262 * frequency meter and its digital filter.
1264 * The filter works on the number of events marked per unit time.
1265 * The filter is single-pole low-pass recursive (IIR). The time unit
1266 * is 1 second. Arithmetic is done using 32-bit integers scaled to
1267 * simulate 3 decimal digits of precision (multiplied by 1000).
1269 * With an FM_COEF of 933, and a time base of 1 second, the filter
1270 * has a half-life of 10 seconds, meaning that if the events quit
1271 * happening, then the rate returned from the fmeter_getrate()
1272 * will be cut in half each 10 seconds, until it converges to zero.
1274 * It is not worth doing a real infinitely recursive filter. If more
1275 * than FM_MAXTICKS ticks have elapsed since the last filter event,
1276 * just compute FM_MAXTICKS ticks worth, by which point the level
1279 * Limit the count of unprocessed events to FM_MAXCNT, so as to avoid
1280 * arithmetic overflow in the fmeter_update() routine.
1282 * Given the simple 32 bit integer arithmetic used, this meter works
1283 * best for reporting rates between one per millisecond (msec) and
1284 * one per 32 (approx) seconds. At constant rates faster than one
1285 * per msec it maxes out at values just under 1,000,000. At constant
1286 * rates between one per msec, and one per second it will stabilize
1287 * to a value N*1000, where N is the rate of events per second.
1288 * At constant rates between one per second and one per 32 seconds,
1289 * it will be choppy, moving up on the seconds that have an event,
1290 * and then decaying until the next event. At rates slower than
1291 * about one in 32 seconds, it decays all the way back to zero between
1295 #define FM_COEF 933 /* coefficient for half-life of 10 secs */
1296 #define FM_MAXTICKS ((time_t)99) /* useless computing more ticks than this */
1297 #define FM_MAXCNT 1000000 /* limit cnt to avoid overflow */
1298 #define FM_SCALE 1000 /* faux fixed point scale */
1300 /* Initialize a frequency meter */
1301 static void fmeter_init(struct fmeter *fmp)
1306 spin_lock_init(&fmp->lock);
1309 /* Internal meter update - process cnt events and update value */
1310 static void fmeter_update(struct fmeter *fmp)
1312 time_t now = get_seconds();
1313 time_t ticks = now - fmp->time;
1318 ticks = min(FM_MAXTICKS, ticks);
1320 fmp->val = (FM_COEF * fmp->val) / FM_SCALE;
1323 fmp->val += ((FM_SCALE - FM_COEF) * fmp->cnt) / FM_SCALE;
1327 /* Process any previous ticks, then bump cnt by one (times scale). */
1328 static void fmeter_markevent(struct fmeter *fmp)
1330 spin_lock(&fmp->lock);
1332 fmp->cnt = min(FM_MAXCNT, fmp->cnt + FM_SCALE);
1333 spin_unlock(&fmp->lock);
1336 /* Process any previous ticks, then return current value. */
1337 static int fmeter_getrate(struct fmeter *fmp)
1341 spin_lock(&fmp->lock);
1344 spin_unlock(&fmp->lock);
1348 /* Protected by cgroup_lock */
1349 static cpumask_var_t cpus_attach;
1351 /* Called by cgroups to determine if a cpuset is usable; cgroup_mutex held */
1352 static int cpuset_can_attach(struct cgroup_subsys *ss,
1353 struct cgroup *cont, struct task_struct *tsk)
1355 struct cpuset *cs = cgroup_cs(cont);
1357 if (cpumask_empty(cs->cpus_allowed) || nodes_empty(cs->mems_allowed))
1361 * Kthreads bound to specific cpus cannot be moved to a new cpuset; we
1362 * cannot change their cpu affinity and isolating such threads by their
1363 * set of allowed nodes is unnecessary. Thus, cpusets are not
1364 * applicable for such threads. This prevents checking for success of
1365 * set_cpus_allowed_ptr() on all attached tasks before cpus_allowed may
1368 if (tsk->flags & PF_THREAD_BOUND)
1371 return security_task_setscheduler(tsk, 0, NULL);
1374 static void cpuset_attach(struct cgroup_subsys *ss,
1375 struct cgroup *cont, struct cgroup *oldcont,
1376 struct task_struct *tsk)
1378 nodemask_t from, to;
1379 struct mm_struct *mm;
1380 struct cpuset *cs = cgroup_cs(cont);
1381 struct cpuset *oldcs = cgroup_cs(oldcont);
1384 if (cs == &top_cpuset) {
1385 cpumask_copy(cpus_attach, cpu_possible_mask);
1387 mutex_lock(&callback_mutex);
1388 guarantee_online_cpus(cs, cpus_attach);
1389 mutex_unlock(&callback_mutex);
1391 err = set_cpus_allowed_ptr(tsk, cpus_attach);
1395 from = oldcs->mems_allowed;
1396 to = cs->mems_allowed;
1397 mm = get_task_mm(tsk);
1399 mpol_rebind_mm(mm, &to);
1400 if (is_memory_migrate(cs))
1401 cpuset_migrate_mm(mm, &from, &to);
1406 /* The various types of files and directories in a cpuset file system */
1409 FILE_MEMORY_MIGRATE,
1415 FILE_SCHED_LOAD_BALANCE,
1416 FILE_SCHED_RELAX_DOMAIN_LEVEL,
1417 FILE_MEMORY_PRESSURE_ENABLED,
1418 FILE_MEMORY_PRESSURE,
1421 } cpuset_filetype_t;
1423 static int cpuset_write_u64(struct cgroup *cgrp, struct cftype *cft, u64 val)
1426 struct cpuset *cs = cgroup_cs(cgrp);
1427 cpuset_filetype_t type = cft->private;
1429 if (!cgroup_lock_live_group(cgrp))
1433 case FILE_CPU_EXCLUSIVE:
1434 retval = update_flag(CS_CPU_EXCLUSIVE, cs, val);
1436 case FILE_MEM_EXCLUSIVE:
1437 retval = update_flag(CS_MEM_EXCLUSIVE, cs, val);
1439 case FILE_MEM_HARDWALL:
1440 retval = update_flag(CS_MEM_HARDWALL, cs, val);
1442 case FILE_SCHED_LOAD_BALANCE:
1443 retval = update_flag(CS_SCHED_LOAD_BALANCE, cs, val);
1445 case FILE_MEMORY_MIGRATE:
1446 retval = update_flag(CS_MEMORY_MIGRATE, cs, val);
1448 case FILE_MEMORY_PRESSURE_ENABLED:
1449 cpuset_memory_pressure_enabled = !!val;
1451 case FILE_MEMORY_PRESSURE:
1454 case FILE_SPREAD_PAGE:
1455 retval = update_flag(CS_SPREAD_PAGE, cs, val);
1456 cs->mems_generation = cpuset_mems_generation++;
1458 case FILE_SPREAD_SLAB:
1459 retval = update_flag(CS_SPREAD_SLAB, cs, val);
1460 cs->mems_generation = cpuset_mems_generation++;
1470 static int cpuset_write_s64(struct cgroup *cgrp, struct cftype *cft, s64 val)
1473 struct cpuset *cs = cgroup_cs(cgrp);
1474 cpuset_filetype_t type = cft->private;
1476 if (!cgroup_lock_live_group(cgrp))
1480 case FILE_SCHED_RELAX_DOMAIN_LEVEL:
1481 retval = update_relax_domain_level(cs, val);
1492 * Common handling for a write to a "cpus" or "mems" file.
1494 static int cpuset_write_resmask(struct cgroup *cgrp, struct cftype *cft,
1498 struct cpuset *cs = cgroup_cs(cgrp);
1499 struct cpuset *trialcs;
1501 if (!cgroup_lock_live_group(cgrp))
1504 trialcs = alloc_trial_cpuset(cs);
1508 switch (cft->private) {
1510 retval = update_cpumask(cs, trialcs, buf);
1513 retval = update_nodemask(cs, trialcs, buf);
1520 free_trial_cpuset(trialcs);
1526 * These ascii lists should be read in a single call, by using a user
1527 * buffer large enough to hold the entire map. If read in smaller
1528 * chunks, there is no guarantee of atomicity. Since the display format
1529 * used, list of ranges of sequential numbers, is variable length,
1530 * and since these maps can change value dynamically, one could read
1531 * gibberish by doing partial reads while a list was changing.
1532 * A single large read to a buffer that crosses a page boundary is
1533 * ok, because the result being copied to user land is not recomputed
1534 * across a page fault.
1537 static int cpuset_sprintf_cpulist(char *page, struct cpuset *cs)
1541 mutex_lock(&callback_mutex);
1542 ret = cpulist_scnprintf(page, PAGE_SIZE, cs->cpus_allowed);
1543 mutex_unlock(&callback_mutex);
1548 static int cpuset_sprintf_memlist(char *page, struct cpuset *cs)
1552 mutex_lock(&callback_mutex);
1553 mask = cs->mems_allowed;
1554 mutex_unlock(&callback_mutex);
1556 return nodelist_scnprintf(page, PAGE_SIZE, mask);
1559 static ssize_t cpuset_common_file_read(struct cgroup *cont,
1563 size_t nbytes, loff_t *ppos)
1565 struct cpuset *cs = cgroup_cs(cont);
1566 cpuset_filetype_t type = cft->private;
1571 if (!(page = (char *)__get_free_page(GFP_TEMPORARY)))
1578 s += cpuset_sprintf_cpulist(s, cs);
1581 s += cpuset_sprintf_memlist(s, cs);
1589 retval = simple_read_from_buffer(buf, nbytes, ppos, page, s - page);
1591 free_page((unsigned long)page);
1595 static u64 cpuset_read_u64(struct cgroup *cont, struct cftype *cft)
1597 struct cpuset *cs = cgroup_cs(cont);
1598 cpuset_filetype_t type = cft->private;
1600 case FILE_CPU_EXCLUSIVE:
1601 return is_cpu_exclusive(cs);
1602 case FILE_MEM_EXCLUSIVE:
1603 return is_mem_exclusive(cs);
1604 case FILE_MEM_HARDWALL:
1605 return is_mem_hardwall(cs);
1606 case FILE_SCHED_LOAD_BALANCE:
1607 return is_sched_load_balance(cs);
1608 case FILE_MEMORY_MIGRATE:
1609 return is_memory_migrate(cs);
1610 case FILE_MEMORY_PRESSURE_ENABLED:
1611 return cpuset_memory_pressure_enabled;
1612 case FILE_MEMORY_PRESSURE:
1613 return fmeter_getrate(&cs->fmeter);
1614 case FILE_SPREAD_PAGE:
1615 return is_spread_page(cs);
1616 case FILE_SPREAD_SLAB:
1617 return is_spread_slab(cs);
1622 /* Unreachable but makes gcc happy */
1626 static s64 cpuset_read_s64(struct cgroup *cont, struct cftype *cft)
1628 struct cpuset *cs = cgroup_cs(cont);
1629 cpuset_filetype_t type = cft->private;
1631 case FILE_SCHED_RELAX_DOMAIN_LEVEL:
1632 return cs->relax_domain_level;
1637 /* Unrechable but makes gcc happy */
1643 * for the common functions, 'private' gives the type of file
1646 static struct cftype files[] = {
1649 .read = cpuset_common_file_read,
1650 .write_string = cpuset_write_resmask,
1651 .max_write_len = (100U + 6 * NR_CPUS),
1652 .private = FILE_CPULIST,
1657 .read = cpuset_common_file_read,
1658 .write_string = cpuset_write_resmask,
1659 .max_write_len = (100U + 6 * MAX_NUMNODES),
1660 .private = FILE_MEMLIST,
1664 .name = "cpu_exclusive",
1665 .read_u64 = cpuset_read_u64,
1666 .write_u64 = cpuset_write_u64,
1667 .private = FILE_CPU_EXCLUSIVE,
1671 .name = "mem_exclusive",
1672 .read_u64 = cpuset_read_u64,
1673 .write_u64 = cpuset_write_u64,
1674 .private = FILE_MEM_EXCLUSIVE,
1678 .name = "mem_hardwall",
1679 .read_u64 = cpuset_read_u64,
1680 .write_u64 = cpuset_write_u64,
1681 .private = FILE_MEM_HARDWALL,
1685 .name = "sched_load_balance",
1686 .read_u64 = cpuset_read_u64,
1687 .write_u64 = cpuset_write_u64,
1688 .private = FILE_SCHED_LOAD_BALANCE,
1692 .name = "sched_relax_domain_level",
1693 .read_s64 = cpuset_read_s64,
1694 .write_s64 = cpuset_write_s64,
1695 .private = FILE_SCHED_RELAX_DOMAIN_LEVEL,
1699 .name = "memory_migrate",
1700 .read_u64 = cpuset_read_u64,
1701 .write_u64 = cpuset_write_u64,
1702 .private = FILE_MEMORY_MIGRATE,
1706 .name = "memory_pressure",
1707 .read_u64 = cpuset_read_u64,
1708 .write_u64 = cpuset_write_u64,
1709 .private = FILE_MEMORY_PRESSURE,
1714 .name = "memory_spread_page",
1715 .read_u64 = cpuset_read_u64,
1716 .write_u64 = cpuset_write_u64,
1717 .private = FILE_SPREAD_PAGE,
1721 .name = "memory_spread_slab",
1722 .read_u64 = cpuset_read_u64,
1723 .write_u64 = cpuset_write_u64,
1724 .private = FILE_SPREAD_SLAB,
1728 static struct cftype cft_memory_pressure_enabled = {
1729 .name = "memory_pressure_enabled",
1730 .read_u64 = cpuset_read_u64,
1731 .write_u64 = cpuset_write_u64,
1732 .private = FILE_MEMORY_PRESSURE_ENABLED,
1735 static int cpuset_populate(struct cgroup_subsys *ss, struct cgroup *cont)
1739 err = cgroup_add_files(cont, ss, files, ARRAY_SIZE(files));
1742 /* memory_pressure_enabled is in root cpuset only */
1744 err = cgroup_add_file(cont, ss,
1745 &cft_memory_pressure_enabled);
1750 * post_clone() is called at the end of cgroup_clone().
1751 * 'cgroup' was just created automatically as a result of
1752 * a cgroup_clone(), and the current task is about to
1753 * be moved into 'cgroup'.
1755 * Currently we refuse to set up the cgroup - thereby
1756 * refusing the task to be entered, and as a result refusing
1757 * the sys_unshare() or clone() which initiated it - if any
1758 * sibling cpusets have exclusive cpus or mem.
1760 * If this becomes a problem for some users who wish to
1761 * allow that scenario, then cpuset_post_clone() could be
1762 * changed to grant parent->cpus_allowed-sibling_cpus_exclusive
1763 * (and likewise for mems) to the new cgroup. Called with cgroup_mutex
1766 static void cpuset_post_clone(struct cgroup_subsys *ss,
1767 struct cgroup *cgroup)
1769 struct cgroup *parent, *child;
1770 struct cpuset *cs, *parent_cs;
1772 parent = cgroup->parent;
1773 list_for_each_entry(child, &parent->children, sibling) {
1774 cs = cgroup_cs(child);
1775 if (is_mem_exclusive(cs) || is_cpu_exclusive(cs))
1778 cs = cgroup_cs(cgroup);
1779 parent_cs = cgroup_cs(parent);
1781 cs->mems_allowed = parent_cs->mems_allowed;
1782 cpumask_copy(cs->cpus_allowed, parent_cs->cpus_allowed);
1787 * cpuset_create - create a cpuset
1788 * ss: cpuset cgroup subsystem
1789 * cont: control group that the new cpuset will be part of
1792 static struct cgroup_subsys_state *cpuset_create(
1793 struct cgroup_subsys *ss,
1794 struct cgroup *cont)
1797 struct cpuset *parent;
1799 if (!cont->parent) {
1800 /* This is early initialization for the top cgroup */
1801 top_cpuset.mems_generation = cpuset_mems_generation++;
1802 return &top_cpuset.css;
1804 parent = cgroup_cs(cont->parent);
1805 cs = kmalloc(sizeof(*cs), GFP_KERNEL);
1807 return ERR_PTR(-ENOMEM);
1808 if (!alloc_cpumask_var(&cs->cpus_allowed, GFP_KERNEL)) {
1810 return ERR_PTR(-ENOMEM);
1813 cpuset_update_task_memory_state();
1815 if (is_spread_page(parent))
1816 set_bit(CS_SPREAD_PAGE, &cs->flags);
1817 if (is_spread_slab(parent))
1818 set_bit(CS_SPREAD_SLAB, &cs->flags);
1819 set_bit(CS_SCHED_LOAD_BALANCE, &cs->flags);
1820 cpumask_clear(cs->cpus_allowed);
1821 nodes_clear(cs->mems_allowed);
1822 cs->mems_generation = cpuset_mems_generation++;
1823 fmeter_init(&cs->fmeter);
1824 cs->relax_domain_level = -1;
1826 cs->parent = parent;
1827 number_of_cpusets++;
1832 * If the cpuset being removed has its flag 'sched_load_balance'
1833 * enabled, then simulate turning sched_load_balance off, which
1834 * will call async_rebuild_sched_domains().
1837 static void cpuset_destroy(struct cgroup_subsys *ss, struct cgroup *cont)
1839 struct cpuset *cs = cgroup_cs(cont);
1841 cpuset_update_task_memory_state();
1843 if (is_sched_load_balance(cs))
1844 update_flag(CS_SCHED_LOAD_BALANCE, cs, 0);
1846 number_of_cpusets--;
1847 free_cpumask_var(cs->cpus_allowed);
1851 struct cgroup_subsys cpuset_subsys = {
1853 .create = cpuset_create,
1854 .destroy = cpuset_destroy,
1855 .can_attach = cpuset_can_attach,
1856 .attach = cpuset_attach,
1857 .populate = cpuset_populate,
1858 .post_clone = cpuset_post_clone,
1859 .subsys_id = cpuset_subsys_id,
1864 * cpuset_init_early - just enough so that the calls to
1865 * cpuset_update_task_memory_state() in early init code
1869 int __init cpuset_init_early(void)
1871 alloc_cpumask_var(&top_cpuset.cpus_allowed, GFP_NOWAIT);
1873 top_cpuset.mems_generation = cpuset_mems_generation++;
1879 * cpuset_init - initialize cpusets at system boot
1881 * Description: Initialize top_cpuset and the cpuset internal file system,
1884 int __init cpuset_init(void)
1888 cpumask_setall(top_cpuset.cpus_allowed);
1889 nodes_setall(top_cpuset.mems_allowed);
1891 fmeter_init(&top_cpuset.fmeter);
1892 top_cpuset.mems_generation = cpuset_mems_generation++;
1893 set_bit(CS_SCHED_LOAD_BALANCE, &top_cpuset.flags);
1894 top_cpuset.relax_domain_level = -1;
1896 err = register_filesystem(&cpuset_fs_type);
1900 if (!alloc_cpumask_var(&cpus_attach, GFP_KERNEL))
1903 number_of_cpusets = 1;
1908 * cpuset_do_move_task - move a given task to another cpuset
1909 * @tsk: pointer to task_struct the task to move
1910 * @scan: struct cgroup_scanner contained in its struct cpuset_hotplug_scanner
1912 * Called by cgroup_scan_tasks() for each task in a cgroup.
1913 * Return nonzero to stop the walk through the tasks.
1915 static void cpuset_do_move_task(struct task_struct *tsk,
1916 struct cgroup_scanner *scan)
1918 struct cgroup *new_cgroup = scan->data;
1920 cgroup_attach_task(new_cgroup, tsk);
1924 * move_member_tasks_to_cpuset - move tasks from one cpuset to another
1925 * @from: cpuset in which the tasks currently reside
1926 * @to: cpuset to which the tasks will be moved
1928 * Called with cgroup_mutex held
1929 * callback_mutex must not be held, as cpuset_attach() will take it.
1931 * The cgroup_scan_tasks() function will scan all the tasks in a cgroup,
1932 * calling callback functions for each.
1934 static void move_member_tasks_to_cpuset(struct cpuset *from, struct cpuset *to)
1936 struct cgroup_scanner scan;
1938 scan.cg = from->css.cgroup;
1939 scan.test_task = NULL; /* select all tasks in cgroup */
1940 scan.process_task = cpuset_do_move_task;
1942 scan.data = to->css.cgroup;
1944 if (cgroup_scan_tasks(&scan))
1945 printk(KERN_ERR "move_member_tasks_to_cpuset: "
1946 "cgroup_scan_tasks failed\n");
1950 * If CPU and/or memory hotplug handlers, below, unplug any CPUs
1951 * or memory nodes, we need to walk over the cpuset hierarchy,
1952 * removing that CPU or node from all cpusets. If this removes the
1953 * last CPU or node from a cpuset, then move the tasks in the empty
1954 * cpuset to its next-highest non-empty parent.
1956 * Called with cgroup_mutex held
1957 * callback_mutex must not be held, as cpuset_attach() will take it.
1959 static void remove_tasks_in_empty_cpuset(struct cpuset *cs)
1961 struct cpuset *parent;
1964 * The cgroup's css_sets list is in use if there are tasks
1965 * in the cpuset; the list is empty if there are none;
1966 * the cs->css.refcnt seems always 0.
1968 if (list_empty(&cs->css.cgroup->css_sets))
1972 * Find its next-highest non-empty parent, (top cpuset
1973 * has online cpus, so can't be empty).
1975 parent = cs->parent;
1976 while (cpumask_empty(parent->cpus_allowed) ||
1977 nodes_empty(parent->mems_allowed))
1978 parent = parent->parent;
1980 move_member_tasks_to_cpuset(cs, parent);
1984 * Walk the specified cpuset subtree and look for empty cpusets.
1985 * The tasks of such cpuset must be moved to a parent cpuset.
1987 * Called with cgroup_mutex held. We take callback_mutex to modify
1988 * cpus_allowed and mems_allowed.
1990 * This walk processes the tree from top to bottom, completing one layer
1991 * before dropping down to the next. It always processes a node before
1992 * any of its children.
1994 * For now, since we lack memory hot unplug, we'll never see a cpuset
1995 * that has tasks along with an empty 'mems'. But if we did see such
1996 * a cpuset, we'd handle it just like we do if its 'cpus' was empty.
1998 static void scan_for_empty_cpusets(struct cpuset *root)
2001 struct cpuset *cp; /* scans cpusets being updated */
2002 struct cpuset *child; /* scans child cpusets of cp */
2003 struct cgroup *cont;
2006 list_add_tail((struct list_head *)&root->stack_list, &queue);
2008 while (!list_empty(&queue)) {
2009 cp = list_first_entry(&queue, struct cpuset, stack_list);
2010 list_del(queue.next);
2011 list_for_each_entry(cont, &cp->css.cgroup->children, sibling) {
2012 child = cgroup_cs(cont);
2013 list_add_tail(&child->stack_list, &queue);
2016 /* Continue past cpusets with all cpus, mems online */
2017 if (cpumask_subset(cp->cpus_allowed, cpu_online_mask) &&
2018 nodes_subset(cp->mems_allowed, node_states[N_HIGH_MEMORY]))
2021 oldmems = cp->mems_allowed;
2023 /* Remove offline cpus and mems from this cpuset. */
2024 mutex_lock(&callback_mutex);
2025 cpumask_and(cp->cpus_allowed, cp->cpus_allowed,
2027 nodes_and(cp->mems_allowed, cp->mems_allowed,
2028 node_states[N_HIGH_MEMORY]);
2029 mutex_unlock(&callback_mutex);
2031 /* Move tasks from the empty cpuset to a parent */
2032 if (cpumask_empty(cp->cpus_allowed) ||
2033 nodes_empty(cp->mems_allowed))
2034 remove_tasks_in_empty_cpuset(cp);
2036 update_tasks_cpumask(cp, NULL);
2037 update_tasks_nodemask(cp, &oldmems, NULL);
2043 * The top_cpuset tracks what CPUs and Memory Nodes are online,
2044 * period. This is necessary in order to make cpusets transparent
2045 * (of no affect) on systems that are actively using CPU hotplug
2046 * but making no active use of cpusets.
2048 * This routine ensures that top_cpuset.cpus_allowed tracks
2049 * cpu_online_map on each CPU hotplug (cpuhp) event.
2051 * Called within get_online_cpus(). Needs to call cgroup_lock()
2052 * before calling generate_sched_domains().
2054 static int cpuset_track_online_cpus(struct notifier_block *unused_nb,
2055 unsigned long phase, void *unused_cpu)
2057 struct sched_domain_attr *attr;
2058 struct cpumask *doms;
2063 case CPU_ONLINE_FROZEN:
2065 case CPU_DEAD_FROZEN:
2073 mutex_lock(&callback_mutex);
2074 cpumask_copy(top_cpuset.cpus_allowed, cpu_online_mask);
2075 mutex_unlock(&callback_mutex);
2076 scan_for_empty_cpusets(&top_cpuset);
2077 ndoms = generate_sched_domains(&doms, &attr);
2080 /* Have scheduler rebuild the domains */
2081 partition_sched_domains(ndoms, doms, attr);
2086 #ifdef CONFIG_MEMORY_HOTPLUG
2088 * Keep top_cpuset.mems_allowed tracking node_states[N_HIGH_MEMORY].
2089 * Call this routine anytime after node_states[N_HIGH_MEMORY] changes.
2090 * See also the previous routine cpuset_track_online_cpus().
2092 static int cpuset_track_online_nodes(struct notifier_block *self,
2093 unsigned long action, void *arg)
2099 mutex_lock(&callback_mutex);
2100 top_cpuset.mems_allowed = node_states[N_HIGH_MEMORY];
2101 mutex_unlock(&callback_mutex);
2102 if (action == MEM_OFFLINE)
2103 scan_for_empty_cpusets(&top_cpuset);
2114 * cpuset_init_smp - initialize cpus_allowed
2116 * Description: Finish top cpuset after cpu, node maps are initialized
2119 void __init cpuset_init_smp(void)
2121 cpumask_copy(top_cpuset.cpus_allowed, cpu_online_mask);
2122 top_cpuset.mems_allowed = node_states[N_HIGH_MEMORY];
2124 hotcpu_notifier(cpuset_track_online_cpus, 0);
2125 hotplug_memory_notifier(cpuset_track_online_nodes, 10);
2127 cpuset_wq = create_singlethread_workqueue("cpuset");
2132 * cpuset_cpus_allowed - return cpus_allowed mask from a tasks cpuset.
2133 * @tsk: pointer to task_struct from which to obtain cpuset->cpus_allowed.
2134 * @pmask: pointer to struct cpumask variable to receive cpus_allowed set.
2136 * Description: Returns the cpumask_var_t cpus_allowed of the cpuset
2137 * attached to the specified @tsk. Guaranteed to return some non-empty
2138 * subset of cpu_online_map, even if this means going outside the
2142 void cpuset_cpus_allowed(struct task_struct *tsk, struct cpumask *pmask)
2144 mutex_lock(&callback_mutex);
2145 cpuset_cpus_allowed_locked(tsk, pmask);
2146 mutex_unlock(&callback_mutex);
2150 * cpuset_cpus_allowed_locked - return cpus_allowed mask from a tasks cpuset.
2151 * Must be called with callback_mutex held.
2153 void cpuset_cpus_allowed_locked(struct task_struct *tsk, struct cpumask *pmask)
2156 guarantee_online_cpus(task_cs(tsk), pmask);
2160 void cpuset_init_current_mems_allowed(void)
2162 nodes_setall(current->mems_allowed);
2166 * cpuset_mems_allowed - return mems_allowed mask from a tasks cpuset.
2167 * @tsk: pointer to task_struct from which to obtain cpuset->mems_allowed.
2169 * Description: Returns the nodemask_t mems_allowed of the cpuset
2170 * attached to the specified @tsk. Guaranteed to return some non-empty
2171 * subset of node_states[N_HIGH_MEMORY], even if this means going outside the
2175 nodemask_t cpuset_mems_allowed(struct task_struct *tsk)
2179 mutex_lock(&callback_mutex);
2181 guarantee_online_mems(task_cs(tsk), &mask);
2183 mutex_unlock(&callback_mutex);
2189 * cpuset_nodemask_valid_mems_allowed - check nodemask vs. curremt mems_allowed
2190 * @nodemask: the nodemask to be checked
2192 * Are any of the nodes in the nodemask allowed in current->mems_allowed?
2194 int cpuset_nodemask_valid_mems_allowed(nodemask_t *nodemask)
2196 return nodes_intersects(*nodemask, current->mems_allowed);
2200 * nearest_hardwall_ancestor() - Returns the nearest mem_exclusive or
2201 * mem_hardwall ancestor to the specified cpuset. Call holding
2202 * callback_mutex. If no ancestor is mem_exclusive or mem_hardwall
2203 * (an unusual configuration), then returns the root cpuset.
2205 static const struct cpuset *nearest_hardwall_ancestor(const struct cpuset *cs)
2207 while (!(is_mem_exclusive(cs) || is_mem_hardwall(cs)) && cs->parent)
2213 * cpuset_node_allowed_softwall - Can we allocate on a memory node?
2214 * @node: is this an allowed node?
2215 * @gfp_mask: memory allocation flags
2217 * If we're in interrupt, yes, we can always allocate. If __GFP_THISNODE is
2218 * set, yes, we can always allocate. If node is in our task's mems_allowed,
2219 * yes. If it's not a __GFP_HARDWALL request and this node is in the nearest
2220 * hardwalled cpuset ancestor to this task's cpuset, yes. If the task has been
2221 * OOM killed and has access to memory reserves as specified by the TIF_MEMDIE
2225 * If __GFP_HARDWALL is set, cpuset_node_allowed_softwall() reduces to
2226 * cpuset_node_allowed_hardwall(). Otherwise, cpuset_node_allowed_softwall()
2227 * might sleep, and might allow a node from an enclosing cpuset.
2229 * cpuset_node_allowed_hardwall() only handles the simpler case of hardwall
2230 * cpusets, and never sleeps.
2232 * The __GFP_THISNODE placement logic is really handled elsewhere,
2233 * by forcibly using a zonelist starting at a specified node, and by
2234 * (in get_page_from_freelist()) refusing to consider the zones for
2235 * any node on the zonelist except the first. By the time any such
2236 * calls get to this routine, we should just shut up and say 'yes'.
2238 * GFP_USER allocations are marked with the __GFP_HARDWALL bit,
2239 * and do not allow allocations outside the current tasks cpuset
2240 * unless the task has been OOM killed as is marked TIF_MEMDIE.
2241 * GFP_KERNEL allocations are not so marked, so can escape to the
2242 * nearest enclosing hardwalled ancestor cpuset.
2244 * Scanning up parent cpusets requires callback_mutex. The
2245 * __alloc_pages() routine only calls here with __GFP_HARDWALL bit
2246 * _not_ set if it's a GFP_KERNEL allocation, and all nodes in the
2247 * current tasks mems_allowed came up empty on the first pass over
2248 * the zonelist. So only GFP_KERNEL allocations, if all nodes in the
2249 * cpuset are short of memory, might require taking the callback_mutex
2252 * The first call here from mm/page_alloc:get_page_from_freelist()
2253 * has __GFP_HARDWALL set in gfp_mask, enforcing hardwall cpusets,
2254 * so no allocation on a node outside the cpuset is allowed (unless
2255 * in interrupt, of course).
2257 * The second pass through get_page_from_freelist() doesn't even call
2258 * here for GFP_ATOMIC calls. For those calls, the __alloc_pages()
2259 * variable 'wait' is not set, and the bit ALLOC_CPUSET is not set
2260 * in alloc_flags. That logic and the checks below have the combined
2262 * in_interrupt - any node ok (current task context irrelevant)
2263 * GFP_ATOMIC - any node ok
2264 * TIF_MEMDIE - any node ok
2265 * GFP_KERNEL - any node in enclosing hardwalled cpuset ok
2266 * GFP_USER - only nodes in current tasks mems allowed ok.
2269 * Don't call cpuset_node_allowed_softwall if you can't sleep, unless you
2270 * pass in the __GFP_HARDWALL flag set in gfp_flag, which disables
2271 * the code that might scan up ancestor cpusets and sleep.
2273 int __cpuset_node_allowed_softwall(int node, gfp_t gfp_mask)
2275 const struct cpuset *cs; /* current cpuset ancestors */
2276 int allowed; /* is allocation in zone z allowed? */
2278 if (in_interrupt() || (gfp_mask & __GFP_THISNODE))
2280 might_sleep_if(!(gfp_mask & __GFP_HARDWALL));
2281 if (node_isset(node, current->mems_allowed))
2284 * Allow tasks that have access to memory reserves because they have
2285 * been OOM killed to get memory anywhere.
2287 if (unlikely(test_thread_flag(TIF_MEMDIE)))
2289 if (gfp_mask & __GFP_HARDWALL) /* If hardwall request, stop here */
2292 if (current->flags & PF_EXITING) /* Let dying task have memory */
2295 /* Not hardwall and node outside mems_allowed: scan up cpusets */
2296 mutex_lock(&callback_mutex);
2299 cs = nearest_hardwall_ancestor(task_cs(current));
2300 task_unlock(current);
2302 allowed = node_isset(node, cs->mems_allowed);
2303 mutex_unlock(&callback_mutex);
2308 * cpuset_node_allowed_hardwall - Can we allocate on a memory node?
2309 * @node: is this an allowed node?
2310 * @gfp_mask: memory allocation flags
2312 * If we're in interrupt, yes, we can always allocate. If __GFP_THISNODE is
2313 * set, yes, we can always allocate. If node is in our task's mems_allowed,
2314 * yes. If the task has been OOM killed and has access to memory reserves as
2315 * specified by the TIF_MEMDIE flag, yes.
2318 * The __GFP_THISNODE placement logic is really handled elsewhere,
2319 * by forcibly using a zonelist starting at a specified node, and by
2320 * (in get_page_from_freelist()) refusing to consider the zones for
2321 * any node on the zonelist except the first. By the time any such
2322 * calls get to this routine, we should just shut up and say 'yes'.
2324 * Unlike the cpuset_node_allowed_softwall() variant, above,
2325 * this variant requires that the node be in the current task's
2326 * mems_allowed or that we're in interrupt. It does not scan up the
2327 * cpuset hierarchy for the nearest enclosing mem_exclusive cpuset.
2330 int __cpuset_node_allowed_hardwall(int node, gfp_t gfp_mask)
2332 if (in_interrupt() || (gfp_mask & __GFP_THISNODE))
2334 if (node_isset(node, current->mems_allowed))
2337 * Allow tasks that have access to memory reserves because they have
2338 * been OOM killed to get memory anywhere.
2340 if (unlikely(test_thread_flag(TIF_MEMDIE)))
2346 * cpuset_lock - lock out any changes to cpuset structures
2348 * The out of memory (oom) code needs to mutex_lock cpusets
2349 * from being changed while it scans the tasklist looking for a
2350 * task in an overlapping cpuset. Expose callback_mutex via this
2351 * cpuset_lock() routine, so the oom code can lock it, before
2352 * locking the task list. The tasklist_lock is a spinlock, so
2353 * must be taken inside callback_mutex.
2356 void cpuset_lock(void)
2358 mutex_lock(&callback_mutex);
2362 * cpuset_unlock - release lock on cpuset changes
2364 * Undo the lock taken in a previous cpuset_lock() call.
2367 void cpuset_unlock(void)
2369 mutex_unlock(&callback_mutex);
2373 * cpuset_mem_spread_node() - On which node to begin search for a page
2375 * If a task is marked PF_SPREAD_PAGE or PF_SPREAD_SLAB (as for
2376 * tasks in a cpuset with is_spread_page or is_spread_slab set),
2377 * and if the memory allocation used cpuset_mem_spread_node()
2378 * to determine on which node to start looking, as it will for
2379 * certain page cache or slab cache pages such as used for file
2380 * system buffers and inode caches, then instead of starting on the
2381 * local node to look for a free page, rather spread the starting
2382 * node around the tasks mems_allowed nodes.
2384 * We don't have to worry about the returned node being offline
2385 * because "it can't happen", and even if it did, it would be ok.
2387 * The routines calling guarantee_online_mems() are careful to
2388 * only set nodes in task->mems_allowed that are online. So it
2389 * should not be possible for the following code to return an
2390 * offline node. But if it did, that would be ok, as this routine
2391 * is not returning the node where the allocation must be, only
2392 * the node where the search should start. The zonelist passed to
2393 * __alloc_pages() will include all nodes. If the slab allocator
2394 * is passed an offline node, it will fall back to the local node.
2395 * See kmem_cache_alloc_node().
2398 int cpuset_mem_spread_node(void)
2402 node = next_node(current->cpuset_mem_spread_rotor, current->mems_allowed);
2403 if (node == MAX_NUMNODES)
2404 node = first_node(current->mems_allowed);
2405 current->cpuset_mem_spread_rotor = node;
2408 EXPORT_SYMBOL_GPL(cpuset_mem_spread_node);
2411 * cpuset_mems_allowed_intersects - Does @tsk1's mems_allowed intersect @tsk2's?
2412 * @tsk1: pointer to task_struct of some task.
2413 * @tsk2: pointer to task_struct of some other task.
2415 * Description: Return true if @tsk1's mems_allowed intersects the
2416 * mems_allowed of @tsk2. Used by the OOM killer to determine if
2417 * one of the task's memory usage might impact the memory available
2421 int cpuset_mems_allowed_intersects(const struct task_struct *tsk1,
2422 const struct task_struct *tsk2)
2424 return nodes_intersects(tsk1->mems_allowed, tsk2->mems_allowed);
2428 * cpuset_print_task_mems_allowed - prints task's cpuset and mems_allowed
2429 * @task: pointer to task_struct of some task.
2431 * Description: Prints @task's name, cpuset name, and cached copy of its
2432 * mems_allowed to the kernel log. Must hold task_lock(task) to allow
2433 * dereferencing task_cs(task).
2435 void cpuset_print_task_mems_allowed(struct task_struct *tsk)
2437 struct dentry *dentry;
2439 dentry = task_cs(tsk)->css.cgroup->dentry;
2440 spin_lock(&cpuset_buffer_lock);
2441 snprintf(cpuset_name, CPUSET_NAME_LEN,
2442 dentry ? (const char *)dentry->d_name.name : "/");
2443 nodelist_scnprintf(cpuset_nodelist, CPUSET_NODELIST_LEN,
2445 printk(KERN_INFO "%s cpuset=%s mems_allowed=%s\n",
2446 tsk->comm, cpuset_name, cpuset_nodelist);
2447 spin_unlock(&cpuset_buffer_lock);
2451 * Collection of memory_pressure is suppressed unless
2452 * this flag is enabled by writing "1" to the special
2453 * cpuset file 'memory_pressure_enabled' in the root cpuset.
2456 int cpuset_memory_pressure_enabled __read_mostly;
2459 * cpuset_memory_pressure_bump - keep stats of per-cpuset reclaims.
2461 * Keep a running average of the rate of synchronous (direct)
2462 * page reclaim efforts initiated by tasks in each cpuset.
2464 * This represents the rate at which some task in the cpuset
2465 * ran low on memory on all nodes it was allowed to use, and
2466 * had to enter the kernels page reclaim code in an effort to
2467 * create more free memory by tossing clean pages or swapping
2468 * or writing dirty pages.
2470 * Display to user space in the per-cpuset read-only file
2471 * "memory_pressure". Value displayed is an integer
2472 * representing the recent rate of entry into the synchronous
2473 * (direct) page reclaim by any task attached to the cpuset.
2476 void __cpuset_memory_pressure_bump(void)
2479 fmeter_markevent(&task_cs(current)->fmeter);
2480 task_unlock(current);
2483 #ifdef CONFIG_PROC_PID_CPUSET
2485 * proc_cpuset_show()
2486 * - Print tasks cpuset path into seq_file.
2487 * - Used for /proc/<pid>/cpuset.
2488 * - No need to task_lock(tsk) on this tsk->cpuset reference, as it
2489 * doesn't really matter if tsk->cpuset changes after we read it,
2490 * and we take cgroup_mutex, keeping cpuset_attach() from changing it
2493 static int proc_cpuset_show(struct seq_file *m, void *unused_v)
2496 struct task_struct *tsk;
2498 struct cgroup_subsys_state *css;
2502 buf = kmalloc(PAGE_SIZE, GFP_KERNEL);
2508 tsk = get_pid_task(pid, PIDTYPE_PID);
2514 css = task_subsys_state(tsk, cpuset_subsys_id);
2515 retval = cgroup_path(css->cgroup, buf, PAGE_SIZE);
2522 put_task_struct(tsk);
2529 static int cpuset_open(struct inode *inode, struct file *file)
2531 struct pid *pid = PROC_I(inode)->pid;
2532 return single_open(file, proc_cpuset_show, pid);
2535 const struct file_operations proc_cpuset_operations = {
2536 .open = cpuset_open,
2538 .llseek = seq_lseek,
2539 .release = single_release,
2541 #endif /* CONFIG_PROC_PID_CPUSET */
2543 /* Display task cpus_allowed, mems_allowed in /proc/<pid>/status file. */
2544 void cpuset_task_status_allowed(struct seq_file *m, struct task_struct *task)
2546 seq_printf(m, "Cpus_allowed:\t");
2547 seq_cpumask(m, &task->cpus_allowed);
2548 seq_printf(m, "\n");
2549 seq_printf(m, "Cpus_allowed_list:\t");
2550 seq_cpumask_list(m, &task->cpus_allowed);
2551 seq_printf(m, "\n");
2552 seq_printf(m, "Mems_allowed:\t");
2553 seq_nodemask(m, &task->mems_allowed);
2554 seq_printf(m, "\n");
2555 seq_printf(m, "Mems_allowed_list:\t");
2556 seq_nodemask_list(m, &task->mems_allowed);
2557 seq_printf(m, "\n");