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 * Tracks how many cpusets are currently defined in system.
65 * When there is only one cpuset (the root cpuset) we can
66 * short circuit some hooks.
68 int number_of_cpusets __read_mostly;
70 /* Forward declare cgroup structures */
71 struct cgroup_subsys cpuset_subsys;
74 /* See "Frequency meter" comments, below. */
77 int cnt; /* unprocessed events count */
78 int val; /* most recent output value */
79 time_t time; /* clock (secs) when val computed */
80 spinlock_t lock; /* guards read or write of above */
84 struct cgroup_subsys_state css;
86 unsigned long flags; /* "unsigned long" so bitops work */
87 cpumask_var_t cpus_allowed; /* CPUs allowed to tasks in cpuset */
88 nodemask_t mems_allowed; /* Memory Nodes allowed to tasks */
90 struct cpuset *parent; /* my parent */
93 * Copy of global cpuset_mems_generation as of the most
94 * recent time this cpuset changed its mems_allowed.
98 struct fmeter fmeter; /* memory_pressure filter */
100 /* partition number for rebuild_sched_domains() */
103 /* for custom sched domain */
104 int relax_domain_level;
106 /* used for walking a cpuset heirarchy */
107 struct list_head stack_list;
110 /* Retrieve the cpuset for a cgroup */
111 static inline struct cpuset *cgroup_cs(struct cgroup *cont)
113 return container_of(cgroup_subsys_state(cont, cpuset_subsys_id),
117 /* Retrieve the cpuset for a task */
118 static inline struct cpuset *task_cs(struct task_struct *task)
120 return container_of(task_subsys_state(task, cpuset_subsys_id),
123 struct cpuset_hotplug_scanner {
124 struct cgroup_scanner scan;
128 /* bits in struct cpuset flags field */
134 CS_SCHED_LOAD_BALANCE,
139 /* convenient tests for these bits */
140 static inline int is_cpu_exclusive(const struct cpuset *cs)
142 return test_bit(CS_CPU_EXCLUSIVE, &cs->flags);
145 static inline int is_mem_exclusive(const struct cpuset *cs)
147 return test_bit(CS_MEM_EXCLUSIVE, &cs->flags);
150 static inline int is_mem_hardwall(const struct cpuset *cs)
152 return test_bit(CS_MEM_HARDWALL, &cs->flags);
155 static inline int is_sched_load_balance(const struct cpuset *cs)
157 return test_bit(CS_SCHED_LOAD_BALANCE, &cs->flags);
160 static inline int is_memory_migrate(const struct cpuset *cs)
162 return test_bit(CS_MEMORY_MIGRATE, &cs->flags);
165 static inline int is_spread_page(const struct cpuset *cs)
167 return test_bit(CS_SPREAD_PAGE, &cs->flags);
170 static inline int is_spread_slab(const struct cpuset *cs)
172 return test_bit(CS_SPREAD_SLAB, &cs->flags);
176 * Increment this integer everytime any cpuset changes its
177 * mems_allowed value. Users of cpusets can track this generation
178 * number, and avoid having to lock and reload mems_allowed unless
179 * the cpuset they're using changes generation.
181 * A single, global generation is needed because cpuset_attach_task() could
182 * reattach a task to a different cpuset, which must not have its
183 * generation numbers aliased with those of that tasks previous cpuset.
185 * Generations are needed for mems_allowed because one task cannot
186 * modify another's memory placement. So we must enable every task,
187 * on every visit to __alloc_pages(), to efficiently check whether
188 * its current->cpuset->mems_allowed has changed, requiring an update
189 * of its current->mems_allowed.
191 * Since writes to cpuset_mems_generation are guarded by the cgroup lock
192 * there is no need to mark it atomic.
194 static int cpuset_mems_generation;
196 static struct cpuset top_cpuset = {
197 .flags = ((1 << CS_CPU_EXCLUSIVE) | (1 << CS_MEM_EXCLUSIVE)),
201 * There are two global mutexes guarding cpuset structures. The first
202 * is the main control groups cgroup_mutex, accessed via
203 * cgroup_lock()/cgroup_unlock(). The second is the cpuset-specific
204 * callback_mutex, below. They can nest. It is ok to first take
205 * cgroup_mutex, then nest callback_mutex. We also require taking
206 * task_lock() when dereferencing a task's cpuset pointer. See "The
207 * task_lock() exception", at the end of this comment.
209 * A task must hold both mutexes to modify cpusets. If a task
210 * holds cgroup_mutex, then it blocks others wanting that mutex,
211 * ensuring that it is the only task able to also acquire callback_mutex
212 * and be able to modify cpusets. It can perform various checks on
213 * the cpuset structure first, knowing nothing will change. It can
214 * also allocate memory while just holding cgroup_mutex. While it is
215 * performing these checks, various callback routines can briefly
216 * acquire callback_mutex to query cpusets. Once it is ready to make
217 * the changes, it takes callback_mutex, blocking everyone else.
219 * Calls to the kernel memory allocator can not be made while holding
220 * callback_mutex, as that would risk double tripping on callback_mutex
221 * from one of the callbacks into the cpuset code from within
224 * If a task is only holding callback_mutex, then it has read-only
227 * The task_struct fields mems_allowed and mems_generation may only
228 * be accessed in the context of that task, so require no locks.
230 * The cpuset_common_file_read() handlers only hold callback_mutex across
231 * small pieces of code, such as when reading out possibly multi-word
232 * cpumasks and nodemasks.
234 * Accessing a task's cpuset should be done in accordance with the
235 * guidelines for accessing subsystem state in kernel/cgroup.c
238 static DEFINE_MUTEX(callback_mutex);
241 * cpuset_buffer_lock protects both the cpuset_name and cpuset_nodelist
242 * buffers. They are statically allocated to prevent using excess stack
243 * when calling cpuset_print_task_mems_allowed().
245 #define CPUSET_NAME_LEN (128)
246 #define CPUSET_NODELIST_LEN (256)
247 static char cpuset_name[CPUSET_NAME_LEN];
248 static char cpuset_nodelist[CPUSET_NODELIST_LEN];
249 static DEFINE_SPINLOCK(cpuset_buffer_lock);
252 * This is ugly, but preserves the userspace API for existing cpuset
253 * users. If someone tries to mount the "cpuset" filesystem, we
254 * silently switch it to mount "cgroup" instead
256 static int cpuset_get_sb(struct file_system_type *fs_type,
257 int flags, const char *unused_dev_name,
258 void *data, struct vfsmount *mnt)
260 struct file_system_type *cgroup_fs = get_fs_type("cgroup");
265 "release_agent=/sbin/cpuset_release_agent";
266 ret = cgroup_fs->get_sb(cgroup_fs, flags,
267 unused_dev_name, mountopts, mnt);
268 put_filesystem(cgroup_fs);
273 static struct file_system_type cpuset_fs_type = {
275 .get_sb = cpuset_get_sb,
279 * Return in pmask the portion of a cpusets's cpus_allowed that
280 * are online. If none are online, walk up the cpuset hierarchy
281 * until we find one that does have some online cpus. If we get
282 * all the way to the top and still haven't found any online cpus,
283 * return cpu_online_map. Or if passed a NULL cs from an exit'ing
284 * task, return cpu_online_map.
286 * One way or another, we guarantee to return some non-empty subset
289 * Call with callback_mutex held.
292 static void guarantee_online_cpus(const struct cpuset *cs,
293 struct cpumask *pmask)
295 while (cs && !cpumask_intersects(cs->cpus_allowed, cpu_online_mask))
298 cpumask_and(pmask, cs->cpus_allowed, cpu_online_mask);
300 cpumask_copy(pmask, cpu_online_mask);
301 BUG_ON(!cpumask_intersects(pmask, cpu_online_mask));
305 * Return in *pmask the portion of a cpusets's mems_allowed that
306 * are online, with memory. If none are online with memory, walk
307 * up the cpuset hierarchy until we find one that does have some
308 * online mems. If we get all the way to the top and still haven't
309 * found any online mems, return node_states[N_HIGH_MEMORY].
311 * One way or another, we guarantee to return some non-empty subset
312 * of node_states[N_HIGH_MEMORY].
314 * Call with callback_mutex held.
317 static void guarantee_online_mems(const struct cpuset *cs, nodemask_t *pmask)
319 while (cs && !nodes_intersects(cs->mems_allowed,
320 node_states[N_HIGH_MEMORY]))
323 nodes_and(*pmask, cs->mems_allowed,
324 node_states[N_HIGH_MEMORY]);
326 *pmask = node_states[N_HIGH_MEMORY];
327 BUG_ON(!nodes_intersects(*pmask, node_states[N_HIGH_MEMORY]));
331 * cpuset_update_task_memory_state - update task memory placement
333 * If the current tasks cpusets mems_allowed changed behind our
334 * backs, update current->mems_allowed, mems_generation and task NUMA
335 * mempolicy to the new value.
337 * Task mempolicy is updated by rebinding it relative to the
338 * current->cpuset if a task has its memory placement changed.
339 * Do not call this routine if in_interrupt().
341 * Call without callback_mutex or task_lock() held. May be
342 * called with or without cgroup_mutex held. Thanks in part to
343 * 'the_top_cpuset_hack', the task's cpuset pointer will never
344 * be NULL. This routine also might acquire callback_mutex during
347 * Reading current->cpuset->mems_generation doesn't need task_lock
348 * to guard the current->cpuset derefence, because it is guarded
349 * from concurrent freeing of current->cpuset using RCU.
351 * The rcu_dereference() is technically probably not needed,
352 * as I don't actually mind if I see a new cpuset pointer but
353 * an old value of mems_generation. However this really only
354 * matters on alpha systems using cpusets heavily. If I dropped
355 * that rcu_dereference(), it would save them a memory barrier.
356 * For all other arch's, rcu_dereference is a no-op anyway, and for
357 * alpha systems not using cpusets, another planned optimization,
358 * avoiding the rcu critical section for tasks in the root cpuset
359 * which is statically allocated, so can't vanish, will make this
360 * irrelevant. Better to use RCU as intended, than to engage in
361 * some cute trick to save a memory barrier that is impossible to
362 * test, for alpha systems using cpusets heavily, which might not
365 * This routine is needed to update the per-task mems_allowed data,
366 * within the tasks context, when it is trying to allocate memory
367 * (in various mm/mempolicy.c routines) and notices that some other
368 * task has been modifying its cpuset.
371 void cpuset_update_task_memory_state(void)
373 int my_cpusets_mem_gen;
374 struct task_struct *tsk = current;
378 my_cpusets_mem_gen = task_cs(tsk)->mems_generation;
381 if (my_cpusets_mem_gen != tsk->cpuset_mems_generation) {
382 mutex_lock(&callback_mutex);
384 cs = task_cs(tsk); /* Maybe changed when task not locked */
385 guarantee_online_mems(cs, &tsk->mems_allowed);
386 tsk->cpuset_mems_generation = cs->mems_generation;
387 if (is_spread_page(cs))
388 tsk->flags |= PF_SPREAD_PAGE;
390 tsk->flags &= ~PF_SPREAD_PAGE;
391 if (is_spread_slab(cs))
392 tsk->flags |= PF_SPREAD_SLAB;
394 tsk->flags &= ~PF_SPREAD_SLAB;
396 mutex_unlock(&callback_mutex);
397 mpol_rebind_task(tsk, &tsk->mems_allowed);
402 * is_cpuset_subset(p, q) - Is cpuset p a subset of cpuset q?
404 * One cpuset is a subset of another if all its allowed CPUs and
405 * Memory Nodes are a subset of the other, and its exclusive flags
406 * are only set if the other's are set. Call holding cgroup_mutex.
409 static int is_cpuset_subset(const struct cpuset *p, const struct cpuset *q)
411 return cpumask_subset(p->cpus_allowed, q->cpus_allowed) &&
412 nodes_subset(p->mems_allowed, q->mems_allowed) &&
413 is_cpu_exclusive(p) <= is_cpu_exclusive(q) &&
414 is_mem_exclusive(p) <= is_mem_exclusive(q);
418 * alloc_trial_cpuset - allocate a trial cpuset
419 * @cs: the cpuset that the trial cpuset duplicates
421 static struct cpuset *alloc_trial_cpuset(const struct cpuset *cs)
423 struct cpuset *trial;
425 trial = kmemdup(cs, sizeof(*cs), GFP_KERNEL);
429 if (!alloc_cpumask_var(&trial->cpus_allowed, GFP_KERNEL)) {
433 cpumask_copy(trial->cpus_allowed, cs->cpus_allowed);
439 * free_trial_cpuset - free the trial cpuset
440 * @trial: the trial cpuset to be freed
442 static void free_trial_cpuset(struct cpuset *trial)
444 free_cpumask_var(trial->cpus_allowed);
449 * validate_change() - Used to validate that any proposed cpuset change
450 * follows the structural rules for cpusets.
452 * If we replaced the flag and mask values of the current cpuset
453 * (cur) with those values in the trial cpuset (trial), would
454 * our various subset and exclusive rules still be valid? Presumes
457 * 'cur' is the address of an actual, in-use cpuset. Operations
458 * such as list traversal that depend on the actual address of the
459 * cpuset in the list must use cur below, not trial.
461 * 'trial' is the address of bulk structure copy of cur, with
462 * perhaps one or more of the fields cpus_allowed, mems_allowed,
463 * or flags changed to new, trial values.
465 * Return 0 if valid, -errno if not.
468 static int validate_change(const struct cpuset *cur, const struct cpuset *trial)
471 struct cpuset *c, *par;
473 /* Each of our child cpusets must be a subset of us */
474 list_for_each_entry(cont, &cur->css.cgroup->children, sibling) {
475 if (!is_cpuset_subset(cgroup_cs(cont), trial))
479 /* Remaining checks don't apply to root cpuset */
480 if (cur == &top_cpuset)
485 /* We must be a subset of our parent cpuset */
486 if (!is_cpuset_subset(trial, par))
490 * If either I or some sibling (!= me) is exclusive, we can't
493 list_for_each_entry(cont, &par->css.cgroup->children, sibling) {
495 if ((is_cpu_exclusive(trial) || is_cpu_exclusive(c)) &&
497 cpumask_intersects(trial->cpus_allowed, c->cpus_allowed))
499 if ((is_mem_exclusive(trial) || is_mem_exclusive(c)) &&
501 nodes_intersects(trial->mems_allowed, c->mems_allowed))
505 /* Cpusets with tasks can't have empty cpus_allowed or mems_allowed */
506 if (cgroup_task_count(cur->css.cgroup)) {
507 if (cpumask_empty(trial->cpus_allowed) ||
508 nodes_empty(trial->mems_allowed)) {
517 * Helper routine for generate_sched_domains().
518 * Do cpusets a, b have overlapping cpus_allowed masks?
520 static int cpusets_overlap(struct cpuset *a, struct cpuset *b)
522 return cpumask_intersects(a->cpus_allowed, b->cpus_allowed);
526 update_domain_attr(struct sched_domain_attr *dattr, struct cpuset *c)
528 if (dattr->relax_domain_level < c->relax_domain_level)
529 dattr->relax_domain_level = c->relax_domain_level;
534 update_domain_attr_tree(struct sched_domain_attr *dattr, struct cpuset *c)
538 list_add(&c->stack_list, &q);
539 while (!list_empty(&q)) {
542 struct cpuset *child;
544 cp = list_first_entry(&q, struct cpuset, stack_list);
547 if (cpumask_empty(cp->cpus_allowed))
550 if (is_sched_load_balance(cp))
551 update_domain_attr(dattr, cp);
553 list_for_each_entry(cont, &cp->css.cgroup->children, sibling) {
554 child = cgroup_cs(cont);
555 list_add_tail(&child->stack_list, &q);
561 * generate_sched_domains()
563 * This function builds a partial partition of the systems CPUs
564 * A 'partial partition' is a set of non-overlapping subsets whose
565 * union is a subset of that set.
566 * The output of this function needs to be passed to kernel/sched.c
567 * partition_sched_domains() routine, which will rebuild the scheduler's
568 * load balancing domains (sched domains) as specified by that partial
571 * See "What is sched_load_balance" in Documentation/cpusets.txt
572 * for a background explanation of this.
574 * Does not return errors, on the theory that the callers of this
575 * routine would rather not worry about failures to rebuild sched
576 * domains when operating in the severe memory shortage situations
577 * that could cause allocation failures below.
579 * Must be called with cgroup_lock held.
581 * The three key local variables below are:
582 * q - a linked-list queue of cpuset pointers, used to implement a
583 * top-down scan of all cpusets. This scan loads a pointer
584 * to each cpuset marked is_sched_load_balance into the
585 * array 'csa'. For our purposes, rebuilding the schedulers
586 * sched domains, we can ignore !is_sched_load_balance cpusets.
587 * csa - (for CpuSet Array) Array of pointers to all the cpusets
588 * that need to be load balanced, for convenient iterative
589 * access by the subsequent code that finds the best partition,
590 * i.e the set of domains (subsets) of CPUs such that the
591 * cpus_allowed of every cpuset marked is_sched_load_balance
592 * is a subset of one of these domains, while there are as
593 * many such domains as possible, each as small as possible.
594 * doms - Conversion of 'csa' to an array of cpumasks, for passing to
595 * the kernel/sched.c routine partition_sched_domains() in a
596 * convenient format, that can be easily compared to the prior
597 * value to determine what partition elements (sched domains)
598 * were changed (added or removed.)
600 * Finding the best partition (set of domains):
601 * The triple nested loops below over i, j, k scan over the
602 * load balanced cpusets (using the array of cpuset pointers in
603 * csa[]) looking for pairs of cpusets that have overlapping
604 * cpus_allowed, but which don't have the same 'pn' partition
605 * number and gives them in the same partition number. It keeps
606 * looping on the 'restart' label until it can no longer find
609 * The union of the cpus_allowed masks from the set of
610 * all cpusets having the same 'pn' value then form the one
611 * element of the partition (one sched domain) to be passed to
612 * partition_sched_domains().
614 /* FIXME: see the FIXME in partition_sched_domains() */
615 static int generate_sched_domains(struct cpumask **domains,
616 struct sched_domain_attr **attributes)
618 LIST_HEAD(q); /* queue of cpusets to be scanned */
619 struct cpuset *cp; /* scans q */
620 struct cpuset **csa; /* array of all cpuset ptrs */
621 int csn; /* how many cpuset ptrs in csa so far */
622 int i, j, k; /* indices for partition finding loops */
623 struct cpumask *doms; /* resulting partition; i.e. sched domains */
624 struct sched_domain_attr *dattr; /* attributes for custom domains */
625 int ndoms = 0; /* number of sched domains in result */
626 int nslot; /* next empty doms[] struct cpumask slot */
632 /* Special case for the 99% of systems with one, full, sched domain */
633 if (is_sched_load_balance(&top_cpuset)) {
634 doms = kmalloc(cpumask_size(), GFP_KERNEL);
638 dattr = kmalloc(sizeof(struct sched_domain_attr), GFP_KERNEL);
640 *dattr = SD_ATTR_INIT;
641 update_domain_attr_tree(dattr, &top_cpuset);
643 cpumask_copy(doms, top_cpuset.cpus_allowed);
649 csa = kmalloc(number_of_cpusets * sizeof(cp), GFP_KERNEL);
654 list_add(&top_cpuset.stack_list, &q);
655 while (!list_empty(&q)) {
657 struct cpuset *child; /* scans child cpusets of cp */
659 cp = list_first_entry(&q, struct cpuset, stack_list);
662 if (cpumask_empty(cp->cpus_allowed))
666 * All child cpusets contain a subset of the parent's cpus, so
667 * just skip them, and then we call update_domain_attr_tree()
668 * to calc relax_domain_level of the corresponding sched
671 if (is_sched_load_balance(cp)) {
676 list_for_each_entry(cont, &cp->css.cgroup->children, sibling) {
677 child = cgroup_cs(cont);
678 list_add_tail(&child->stack_list, &q);
682 for (i = 0; i < csn; i++)
687 /* Find the best partition (set of sched domains) */
688 for (i = 0; i < csn; i++) {
689 struct cpuset *a = csa[i];
692 for (j = 0; j < csn; j++) {
693 struct cpuset *b = csa[j];
696 if (apn != bpn && cpusets_overlap(a, b)) {
697 for (k = 0; k < csn; k++) {
698 struct cpuset *c = csa[k];
703 ndoms--; /* one less element */
710 * Now we know how many domains to create.
711 * Convert <csn, csa> to <ndoms, doms> and populate cpu masks.
713 doms = kmalloc(ndoms * cpumask_size(), GFP_KERNEL);
718 * The rest of the code, including the scheduler, can deal with
719 * dattr==NULL case. No need to abort if alloc fails.
721 dattr = kmalloc(ndoms * sizeof(struct sched_domain_attr), GFP_KERNEL);
723 for (nslot = 0, i = 0; i < csn; i++) {
724 struct cpuset *a = csa[i];
729 /* Skip completed partitions */
735 if (nslot == ndoms) {
736 static int warnings = 10;
739 "rebuild_sched_domains confused:"
740 " nslot %d, ndoms %d, csn %d, i %d,"
742 nslot, ndoms, csn, i, apn);
750 *(dattr + nslot) = SD_ATTR_INIT;
751 for (j = i; j < csn; j++) {
752 struct cpuset *b = csa[j];
755 cpumask_or(dp, dp, b->cpus_allowed);
757 update_domain_attr_tree(dattr + nslot, b);
759 /* Done with this partition */
765 BUG_ON(nslot != ndoms);
771 * Fallback to the default domain if kmalloc() failed.
772 * See comments in partition_sched_domains().
783 * Rebuild scheduler domains.
785 * Call with neither cgroup_mutex held nor within get_online_cpus().
786 * Takes both cgroup_mutex and get_online_cpus().
788 * Cannot be directly called from cpuset code handling changes
789 * to the cpuset pseudo-filesystem, because it cannot be called
790 * from code that already holds cgroup_mutex.
792 static void do_rebuild_sched_domains(struct work_struct *unused)
794 struct sched_domain_attr *attr;
795 struct cpumask *doms;
800 /* Generate domain masks and attrs */
802 ndoms = generate_sched_domains(&doms, &attr);
805 /* Have scheduler rebuild the domains */
806 partition_sched_domains(ndoms, doms, attr);
811 static DECLARE_WORK(rebuild_sched_domains_work, do_rebuild_sched_domains);
814 * Rebuild scheduler domains, asynchronously via workqueue.
816 * If the flag 'sched_load_balance' of any cpuset with non-empty
817 * 'cpus' changes, or if the 'cpus' allowed changes in any cpuset
818 * which has that flag enabled, or if any cpuset with a non-empty
819 * 'cpus' is removed, then call this routine to rebuild the
820 * scheduler's dynamic sched domains.
822 * The rebuild_sched_domains() and partition_sched_domains()
823 * routines must nest cgroup_lock() inside get_online_cpus(),
824 * but such cpuset changes as these must nest that locking the
825 * other way, holding cgroup_lock() for much of the code.
827 * So in order to avoid an ABBA deadlock, the cpuset code handling
828 * these user changes delegates the actual sched domain rebuilding
829 * to a separate workqueue thread, which ends up processing the
830 * above do_rebuild_sched_domains() function.
832 static void async_rebuild_sched_domains(void)
834 schedule_work(&rebuild_sched_domains_work);
838 * Accomplishes the same scheduler domain rebuild as the above
839 * async_rebuild_sched_domains(), however it directly calls the
840 * rebuild routine synchronously rather than calling it via an
841 * asynchronous work thread.
843 * This can only be called from code that is not holding
844 * cgroup_mutex (not nested in a cgroup_lock() call.)
846 void rebuild_sched_domains(void)
848 do_rebuild_sched_domains(NULL);
852 * cpuset_test_cpumask - test a task's cpus_allowed versus its cpuset's
854 * @scan: struct cgroup_scanner contained in its struct cpuset_hotplug_scanner
856 * Call with cgroup_mutex held. May take callback_mutex during call.
857 * Called for each task in a cgroup by cgroup_scan_tasks().
858 * Return nonzero if this tasks's cpus_allowed mask should be changed (in other
859 * words, if its mask is not equal to its cpuset's mask).
861 static int cpuset_test_cpumask(struct task_struct *tsk,
862 struct cgroup_scanner *scan)
864 return !cpumask_equal(&tsk->cpus_allowed,
865 (cgroup_cs(scan->cg))->cpus_allowed);
869 * cpuset_change_cpumask - make a task's cpus_allowed the same as its cpuset's
871 * @scan: struct cgroup_scanner containing the cgroup of the task
873 * Called by cgroup_scan_tasks() for each task in a cgroup whose
874 * cpus_allowed mask needs to be changed.
876 * We don't need to re-check for the cgroup/cpuset membership, since we're
877 * holding cgroup_lock() at this point.
879 static void cpuset_change_cpumask(struct task_struct *tsk,
880 struct cgroup_scanner *scan)
882 set_cpus_allowed_ptr(tsk, ((cgroup_cs(scan->cg))->cpus_allowed));
886 * update_tasks_cpumask - Update the cpumasks of tasks in the cpuset.
887 * @cs: the cpuset in which each task's cpus_allowed mask needs to be changed
888 * @heap: if NULL, defer allocating heap memory to cgroup_scan_tasks()
890 * Called with cgroup_mutex held
892 * The cgroup_scan_tasks() function will scan all the tasks in a cgroup,
893 * calling callback functions for each.
895 * No return value. It's guaranteed that cgroup_scan_tasks() always returns 0
898 static void update_tasks_cpumask(struct cpuset *cs, struct ptr_heap *heap)
900 struct cgroup_scanner scan;
902 scan.cg = cs->css.cgroup;
903 scan.test_task = cpuset_test_cpumask;
904 scan.process_task = cpuset_change_cpumask;
906 cgroup_scan_tasks(&scan);
910 * update_cpumask - update the cpus_allowed mask of a cpuset and all tasks in it
911 * @cs: the cpuset to consider
912 * @buf: buffer of cpu numbers written to this cpuset
914 static int update_cpumask(struct cpuset *cs, struct cpuset *trialcs,
917 struct ptr_heap heap;
919 int is_load_balanced;
921 /* top_cpuset.cpus_allowed tracks cpu_online_map; it's read-only */
922 if (cs == &top_cpuset)
926 * An empty cpus_allowed is ok only if the cpuset has no tasks.
927 * Since cpulist_parse() fails on an empty mask, we special case
928 * that parsing. The validate_change() call ensures that cpusets
929 * with tasks have cpus.
932 cpumask_clear(trialcs->cpus_allowed);
934 retval = cpulist_parse(buf, trialcs->cpus_allowed);
938 if (!cpumask_subset(trialcs->cpus_allowed, cpu_online_mask))
941 retval = validate_change(cs, trialcs);
945 /* Nothing to do if the cpus didn't change */
946 if (cpumask_equal(cs->cpus_allowed, trialcs->cpus_allowed))
949 retval = heap_init(&heap, PAGE_SIZE, GFP_KERNEL, NULL);
953 is_load_balanced = is_sched_load_balance(trialcs);
955 mutex_lock(&callback_mutex);
956 cpumask_copy(cs->cpus_allowed, trialcs->cpus_allowed);
957 mutex_unlock(&callback_mutex);
960 * Scan tasks in the cpuset, and update the cpumasks of any
961 * that need an update.
963 update_tasks_cpumask(cs, &heap);
967 if (is_load_balanced)
968 async_rebuild_sched_domains();
975 * Migrate memory region from one set of nodes to another.
977 * Temporarilly set tasks mems_allowed to target nodes of migration,
978 * so that the migration code can allocate pages on these nodes.
980 * Call holding cgroup_mutex, so current's cpuset won't change
981 * during this call, as manage_mutex holds off any cpuset_attach()
982 * calls. Therefore we don't need to take task_lock around the
983 * call to guarantee_online_mems(), as we know no one is changing
986 * Hold callback_mutex around the two modifications of our tasks
987 * mems_allowed to synchronize with cpuset_mems_allowed().
989 * While the mm_struct we are migrating is typically from some
990 * other task, the task_struct mems_allowed that we are hacking
991 * is for our current task, which must allocate new pages for that
992 * migrating memory region.
994 * We call cpuset_update_task_memory_state() before hacking
995 * our tasks mems_allowed, so that we are assured of being in
996 * sync with our tasks cpuset, and in particular, callbacks to
997 * cpuset_update_task_memory_state() from nested page allocations
998 * won't see any mismatch of our cpuset and task mems_generation
999 * values, so won't overwrite our hacked tasks mems_allowed
1003 static void cpuset_migrate_mm(struct mm_struct *mm, const nodemask_t *from,
1004 const nodemask_t *to)
1006 struct task_struct *tsk = current;
1008 cpuset_update_task_memory_state();
1010 mutex_lock(&callback_mutex);
1011 tsk->mems_allowed = *to;
1012 mutex_unlock(&callback_mutex);
1014 do_migrate_pages(mm, from, to, MPOL_MF_MOVE_ALL);
1016 mutex_lock(&callback_mutex);
1017 guarantee_online_mems(task_cs(tsk),&tsk->mems_allowed);
1018 mutex_unlock(&callback_mutex);
1021 static void *cpuset_being_rebound;
1024 * update_tasks_nodemask - Update the nodemasks of tasks in the cpuset.
1025 * @cs: the cpuset in which each task's mems_allowed mask needs to be changed
1026 * @oldmem: old mems_allowed of cpuset cs
1028 * Called with cgroup_mutex held
1029 * Return 0 if successful, -errno if not.
1031 static int update_tasks_nodemask(struct cpuset *cs, const nodemask_t *oldmem)
1033 struct task_struct *p;
1034 struct mm_struct **mmarray;
1038 struct cgroup_iter it;
1041 cpuset_being_rebound = cs; /* causes mpol_dup() rebind */
1043 fudge = 10; /* spare mmarray[] slots */
1044 fudge += cpumask_weight(cs->cpus_allowed);/* imagine 1 fork-bomb/cpu */
1048 * Allocate mmarray[] to hold mm reference for each task
1049 * in cpuset cs. Can't kmalloc GFP_KERNEL while holding
1050 * tasklist_lock. We could use GFP_ATOMIC, but with a
1051 * few more lines of code, we can retry until we get a big
1052 * enough mmarray[] w/o using GFP_ATOMIC.
1055 ntasks = cgroup_task_count(cs->css.cgroup); /* guess */
1057 mmarray = kmalloc(ntasks * sizeof(*mmarray), GFP_KERNEL);
1060 read_lock(&tasklist_lock); /* block fork */
1061 if (cgroup_task_count(cs->css.cgroup) <= ntasks)
1062 break; /* got enough */
1063 read_unlock(&tasklist_lock); /* try again */
1069 /* Load up mmarray[] with mm reference for each task in cpuset. */
1070 cgroup_iter_start(cs->css.cgroup, &it);
1071 while ((p = cgroup_iter_next(cs->css.cgroup, &it))) {
1072 struct mm_struct *mm;
1076 "Cpuset mempolicy rebind incomplete.\n");
1079 mm = get_task_mm(p);
1084 cgroup_iter_end(cs->css.cgroup, &it);
1085 read_unlock(&tasklist_lock);
1088 * Now that we've dropped the tasklist spinlock, we can
1089 * rebind the vma mempolicies of each mm in mmarray[] to their
1090 * new cpuset, and release that mm. The mpol_rebind_mm()
1091 * call takes mmap_sem, which we couldn't take while holding
1092 * tasklist_lock. Forks can happen again now - the mpol_dup()
1093 * cpuset_being_rebound check will catch such forks, and rebind
1094 * their vma mempolicies too. Because we still hold the global
1095 * cgroup_mutex, we know that no other rebind effort will
1096 * be contending for the global variable cpuset_being_rebound.
1097 * It's ok if we rebind the same mm twice; mpol_rebind_mm()
1098 * is idempotent. Also migrate pages in each mm to new nodes.
1100 migrate = is_memory_migrate(cs);
1101 for (i = 0; i < n; i++) {
1102 struct mm_struct *mm = mmarray[i];
1104 mpol_rebind_mm(mm, &cs->mems_allowed);
1106 cpuset_migrate_mm(mm, oldmem, &cs->mems_allowed);
1110 /* We're done rebinding vmas to this cpuset's new mems_allowed. */
1112 cpuset_being_rebound = NULL;
1119 * Handle user request to change the 'mems' memory placement
1120 * of a cpuset. Needs to validate the request, update the
1121 * cpusets mems_allowed and mems_generation, and for each
1122 * task in the cpuset, rebind any vma mempolicies and if
1123 * the cpuset is marked 'memory_migrate', migrate the tasks
1124 * pages to the new memory.
1126 * Call with cgroup_mutex held. May take callback_mutex during call.
1127 * Will take tasklist_lock, scan tasklist for tasks in cpuset cs,
1128 * lock each such tasks mm->mmap_sem, scan its vma's and rebind
1129 * their mempolicies to the cpusets new mems_allowed.
1131 static int update_nodemask(struct cpuset *cs, struct cpuset *trialcs,
1138 * top_cpuset.mems_allowed tracks node_stats[N_HIGH_MEMORY];
1141 if (cs == &top_cpuset)
1145 * An empty mems_allowed is ok iff there are no tasks in the cpuset.
1146 * Since nodelist_parse() fails on an empty mask, we special case
1147 * that parsing. The validate_change() call ensures that cpusets
1148 * with tasks have memory.
1151 nodes_clear(trialcs->mems_allowed);
1153 retval = nodelist_parse(buf, trialcs->mems_allowed);
1157 if (!nodes_subset(trialcs->mems_allowed,
1158 node_states[N_HIGH_MEMORY]))
1161 oldmem = cs->mems_allowed;
1162 if (nodes_equal(oldmem, trialcs->mems_allowed)) {
1163 retval = 0; /* Too easy - nothing to do */
1166 retval = validate_change(cs, trialcs);
1170 mutex_lock(&callback_mutex);
1171 cs->mems_allowed = trialcs->mems_allowed;
1172 cs->mems_generation = cpuset_mems_generation++;
1173 mutex_unlock(&callback_mutex);
1175 retval = update_tasks_nodemask(cs, &oldmem);
1180 int current_cpuset_is_being_rebound(void)
1182 return task_cs(current) == cpuset_being_rebound;
1185 static int update_relax_domain_level(struct cpuset *cs, s64 val)
1187 if (val < -1 || val >= SD_LV_MAX)
1190 if (val != cs->relax_domain_level) {
1191 cs->relax_domain_level = val;
1192 if (!cpumask_empty(cs->cpus_allowed) &&
1193 is_sched_load_balance(cs))
1194 async_rebuild_sched_domains();
1201 * update_flag - read a 0 or a 1 in a file and update associated flag
1202 * bit: the bit to update (see cpuset_flagbits_t)
1203 * cs: the cpuset to update
1204 * turning_on: whether the flag is being set or cleared
1206 * Call with cgroup_mutex held.
1209 static int update_flag(cpuset_flagbits_t bit, struct cpuset *cs,
1212 struct cpuset *trialcs;
1214 int balance_flag_changed;
1216 trialcs = alloc_trial_cpuset(cs);
1221 set_bit(bit, &trialcs->flags);
1223 clear_bit(bit, &trialcs->flags);
1225 err = validate_change(cs, trialcs);
1229 balance_flag_changed = (is_sched_load_balance(cs) !=
1230 is_sched_load_balance(trialcs));
1232 mutex_lock(&callback_mutex);
1233 cs->flags = trialcs->flags;
1234 mutex_unlock(&callback_mutex);
1236 if (!cpumask_empty(trialcs->cpus_allowed) && balance_flag_changed)
1237 async_rebuild_sched_domains();
1240 free_trial_cpuset(trialcs);
1245 * Frequency meter - How fast is some event occurring?
1247 * These routines manage a digitally filtered, constant time based,
1248 * event frequency meter. There are four routines:
1249 * fmeter_init() - initialize a frequency meter.
1250 * fmeter_markevent() - called each time the event happens.
1251 * fmeter_getrate() - returns the recent rate of such events.
1252 * fmeter_update() - internal routine used to update fmeter.
1254 * A common data structure is passed to each of these routines,
1255 * which is used to keep track of the state required to manage the
1256 * frequency meter and its digital filter.
1258 * The filter works on the number of events marked per unit time.
1259 * The filter is single-pole low-pass recursive (IIR). The time unit
1260 * is 1 second. Arithmetic is done using 32-bit integers scaled to
1261 * simulate 3 decimal digits of precision (multiplied by 1000).
1263 * With an FM_COEF of 933, and a time base of 1 second, the filter
1264 * has a half-life of 10 seconds, meaning that if the events quit
1265 * happening, then the rate returned from the fmeter_getrate()
1266 * will be cut in half each 10 seconds, until it converges to zero.
1268 * It is not worth doing a real infinitely recursive filter. If more
1269 * than FM_MAXTICKS ticks have elapsed since the last filter event,
1270 * just compute FM_MAXTICKS ticks worth, by which point the level
1273 * Limit the count of unprocessed events to FM_MAXCNT, so as to avoid
1274 * arithmetic overflow in the fmeter_update() routine.
1276 * Given the simple 32 bit integer arithmetic used, this meter works
1277 * best for reporting rates between one per millisecond (msec) and
1278 * one per 32 (approx) seconds. At constant rates faster than one
1279 * per msec it maxes out at values just under 1,000,000. At constant
1280 * rates between one per msec, and one per second it will stabilize
1281 * to a value N*1000, where N is the rate of events per second.
1282 * At constant rates between one per second and one per 32 seconds,
1283 * it will be choppy, moving up on the seconds that have an event,
1284 * and then decaying until the next event. At rates slower than
1285 * about one in 32 seconds, it decays all the way back to zero between
1289 #define FM_COEF 933 /* coefficient for half-life of 10 secs */
1290 #define FM_MAXTICKS ((time_t)99) /* useless computing more ticks than this */
1291 #define FM_MAXCNT 1000000 /* limit cnt to avoid overflow */
1292 #define FM_SCALE 1000 /* faux fixed point scale */
1294 /* Initialize a frequency meter */
1295 static void fmeter_init(struct fmeter *fmp)
1300 spin_lock_init(&fmp->lock);
1303 /* Internal meter update - process cnt events and update value */
1304 static void fmeter_update(struct fmeter *fmp)
1306 time_t now = get_seconds();
1307 time_t ticks = now - fmp->time;
1312 ticks = min(FM_MAXTICKS, ticks);
1314 fmp->val = (FM_COEF * fmp->val) / FM_SCALE;
1317 fmp->val += ((FM_SCALE - FM_COEF) * fmp->cnt) / FM_SCALE;
1321 /* Process any previous ticks, then bump cnt by one (times scale). */
1322 static void fmeter_markevent(struct fmeter *fmp)
1324 spin_lock(&fmp->lock);
1326 fmp->cnt = min(FM_MAXCNT, fmp->cnt + FM_SCALE);
1327 spin_unlock(&fmp->lock);
1330 /* Process any previous ticks, then return current value. */
1331 static int fmeter_getrate(struct fmeter *fmp)
1335 spin_lock(&fmp->lock);
1338 spin_unlock(&fmp->lock);
1342 /* Protected by cgroup_lock */
1343 static cpumask_var_t cpus_attach;
1345 /* Called by cgroups to determine if a cpuset is usable; cgroup_mutex held */
1346 static int cpuset_can_attach(struct cgroup_subsys *ss,
1347 struct cgroup *cont, struct task_struct *tsk)
1349 struct cpuset *cs = cgroup_cs(cont);
1352 if (cpumask_empty(cs->cpus_allowed) || nodes_empty(cs->mems_allowed))
1355 if (tsk->flags & PF_THREAD_BOUND) {
1356 mutex_lock(&callback_mutex);
1357 if (!cpumask_equal(&tsk->cpus_allowed, cs->cpus_allowed))
1359 mutex_unlock(&callback_mutex);
1362 return ret < 0 ? ret : security_task_setscheduler(tsk, 0, NULL);
1365 static void cpuset_attach(struct cgroup_subsys *ss,
1366 struct cgroup *cont, struct cgroup *oldcont,
1367 struct task_struct *tsk)
1369 nodemask_t from, to;
1370 struct mm_struct *mm;
1371 struct cpuset *cs = cgroup_cs(cont);
1372 struct cpuset *oldcs = cgroup_cs(oldcont);
1375 if (cs == &top_cpuset) {
1376 cpumask_copy(cpus_attach, cpu_possible_mask);
1378 mutex_lock(&callback_mutex);
1379 guarantee_online_cpus(cs, cpus_attach);
1380 mutex_unlock(&callback_mutex);
1382 err = set_cpus_allowed_ptr(tsk, cpus_attach);
1386 from = oldcs->mems_allowed;
1387 to = cs->mems_allowed;
1388 mm = get_task_mm(tsk);
1390 mpol_rebind_mm(mm, &to);
1391 if (is_memory_migrate(cs))
1392 cpuset_migrate_mm(mm, &from, &to);
1397 /* The various types of files and directories in a cpuset file system */
1400 FILE_MEMORY_MIGRATE,
1406 FILE_SCHED_LOAD_BALANCE,
1407 FILE_SCHED_RELAX_DOMAIN_LEVEL,
1408 FILE_MEMORY_PRESSURE_ENABLED,
1409 FILE_MEMORY_PRESSURE,
1412 } cpuset_filetype_t;
1414 static int cpuset_write_u64(struct cgroup *cgrp, struct cftype *cft, u64 val)
1417 struct cpuset *cs = cgroup_cs(cgrp);
1418 cpuset_filetype_t type = cft->private;
1420 if (!cgroup_lock_live_group(cgrp))
1424 case FILE_CPU_EXCLUSIVE:
1425 retval = update_flag(CS_CPU_EXCLUSIVE, cs, val);
1427 case FILE_MEM_EXCLUSIVE:
1428 retval = update_flag(CS_MEM_EXCLUSIVE, cs, val);
1430 case FILE_MEM_HARDWALL:
1431 retval = update_flag(CS_MEM_HARDWALL, cs, val);
1433 case FILE_SCHED_LOAD_BALANCE:
1434 retval = update_flag(CS_SCHED_LOAD_BALANCE, cs, val);
1436 case FILE_MEMORY_MIGRATE:
1437 retval = update_flag(CS_MEMORY_MIGRATE, cs, val);
1439 case FILE_MEMORY_PRESSURE_ENABLED:
1440 cpuset_memory_pressure_enabled = !!val;
1442 case FILE_MEMORY_PRESSURE:
1445 case FILE_SPREAD_PAGE:
1446 retval = update_flag(CS_SPREAD_PAGE, cs, val);
1447 cs->mems_generation = cpuset_mems_generation++;
1449 case FILE_SPREAD_SLAB:
1450 retval = update_flag(CS_SPREAD_SLAB, cs, val);
1451 cs->mems_generation = cpuset_mems_generation++;
1461 static int cpuset_write_s64(struct cgroup *cgrp, struct cftype *cft, s64 val)
1464 struct cpuset *cs = cgroup_cs(cgrp);
1465 cpuset_filetype_t type = cft->private;
1467 if (!cgroup_lock_live_group(cgrp))
1471 case FILE_SCHED_RELAX_DOMAIN_LEVEL:
1472 retval = update_relax_domain_level(cs, val);
1483 * Common handling for a write to a "cpus" or "mems" file.
1485 static int cpuset_write_resmask(struct cgroup *cgrp, struct cftype *cft,
1489 struct cpuset *cs = cgroup_cs(cgrp);
1490 struct cpuset *trialcs;
1492 if (!cgroup_lock_live_group(cgrp))
1495 trialcs = alloc_trial_cpuset(cs);
1499 switch (cft->private) {
1501 retval = update_cpumask(cs, trialcs, buf);
1504 retval = update_nodemask(cs, trialcs, buf);
1511 free_trial_cpuset(trialcs);
1517 * These ascii lists should be read in a single call, by using a user
1518 * buffer large enough to hold the entire map. If read in smaller
1519 * chunks, there is no guarantee of atomicity. Since the display format
1520 * used, list of ranges of sequential numbers, is variable length,
1521 * and since these maps can change value dynamically, one could read
1522 * gibberish by doing partial reads while a list was changing.
1523 * A single large read to a buffer that crosses a page boundary is
1524 * ok, because the result being copied to user land is not recomputed
1525 * across a page fault.
1528 static int cpuset_sprintf_cpulist(char *page, struct cpuset *cs)
1532 mutex_lock(&callback_mutex);
1533 ret = cpulist_scnprintf(page, PAGE_SIZE, cs->cpus_allowed);
1534 mutex_unlock(&callback_mutex);
1539 static int cpuset_sprintf_memlist(char *page, struct cpuset *cs)
1543 mutex_lock(&callback_mutex);
1544 mask = cs->mems_allowed;
1545 mutex_unlock(&callback_mutex);
1547 return nodelist_scnprintf(page, PAGE_SIZE, mask);
1550 static ssize_t cpuset_common_file_read(struct cgroup *cont,
1554 size_t nbytes, loff_t *ppos)
1556 struct cpuset *cs = cgroup_cs(cont);
1557 cpuset_filetype_t type = cft->private;
1562 if (!(page = (char *)__get_free_page(GFP_TEMPORARY)))
1569 s += cpuset_sprintf_cpulist(s, cs);
1572 s += cpuset_sprintf_memlist(s, cs);
1580 retval = simple_read_from_buffer(buf, nbytes, ppos, page, s - page);
1582 free_page((unsigned long)page);
1586 static u64 cpuset_read_u64(struct cgroup *cont, struct cftype *cft)
1588 struct cpuset *cs = cgroup_cs(cont);
1589 cpuset_filetype_t type = cft->private;
1591 case FILE_CPU_EXCLUSIVE:
1592 return is_cpu_exclusive(cs);
1593 case FILE_MEM_EXCLUSIVE:
1594 return is_mem_exclusive(cs);
1595 case FILE_MEM_HARDWALL:
1596 return is_mem_hardwall(cs);
1597 case FILE_SCHED_LOAD_BALANCE:
1598 return is_sched_load_balance(cs);
1599 case FILE_MEMORY_MIGRATE:
1600 return is_memory_migrate(cs);
1601 case FILE_MEMORY_PRESSURE_ENABLED:
1602 return cpuset_memory_pressure_enabled;
1603 case FILE_MEMORY_PRESSURE:
1604 return fmeter_getrate(&cs->fmeter);
1605 case FILE_SPREAD_PAGE:
1606 return is_spread_page(cs);
1607 case FILE_SPREAD_SLAB:
1608 return is_spread_slab(cs);
1613 /* Unreachable but makes gcc happy */
1617 static s64 cpuset_read_s64(struct cgroup *cont, struct cftype *cft)
1619 struct cpuset *cs = cgroup_cs(cont);
1620 cpuset_filetype_t type = cft->private;
1622 case FILE_SCHED_RELAX_DOMAIN_LEVEL:
1623 return cs->relax_domain_level;
1628 /* Unrechable but makes gcc happy */
1634 * for the common functions, 'private' gives the type of file
1637 static struct cftype files[] = {
1640 .read = cpuset_common_file_read,
1641 .write_string = cpuset_write_resmask,
1642 .max_write_len = (100U + 6 * NR_CPUS),
1643 .private = FILE_CPULIST,
1648 .read = cpuset_common_file_read,
1649 .write_string = cpuset_write_resmask,
1650 .max_write_len = (100U + 6 * MAX_NUMNODES),
1651 .private = FILE_MEMLIST,
1655 .name = "cpu_exclusive",
1656 .read_u64 = cpuset_read_u64,
1657 .write_u64 = cpuset_write_u64,
1658 .private = FILE_CPU_EXCLUSIVE,
1662 .name = "mem_exclusive",
1663 .read_u64 = cpuset_read_u64,
1664 .write_u64 = cpuset_write_u64,
1665 .private = FILE_MEM_EXCLUSIVE,
1669 .name = "mem_hardwall",
1670 .read_u64 = cpuset_read_u64,
1671 .write_u64 = cpuset_write_u64,
1672 .private = FILE_MEM_HARDWALL,
1676 .name = "sched_load_balance",
1677 .read_u64 = cpuset_read_u64,
1678 .write_u64 = cpuset_write_u64,
1679 .private = FILE_SCHED_LOAD_BALANCE,
1683 .name = "sched_relax_domain_level",
1684 .read_s64 = cpuset_read_s64,
1685 .write_s64 = cpuset_write_s64,
1686 .private = FILE_SCHED_RELAX_DOMAIN_LEVEL,
1690 .name = "memory_migrate",
1691 .read_u64 = cpuset_read_u64,
1692 .write_u64 = cpuset_write_u64,
1693 .private = FILE_MEMORY_MIGRATE,
1697 .name = "memory_pressure",
1698 .read_u64 = cpuset_read_u64,
1699 .write_u64 = cpuset_write_u64,
1700 .private = FILE_MEMORY_PRESSURE,
1704 .name = "memory_spread_page",
1705 .read_u64 = cpuset_read_u64,
1706 .write_u64 = cpuset_write_u64,
1707 .private = FILE_SPREAD_PAGE,
1711 .name = "memory_spread_slab",
1712 .read_u64 = cpuset_read_u64,
1713 .write_u64 = cpuset_write_u64,
1714 .private = FILE_SPREAD_SLAB,
1718 static struct cftype cft_memory_pressure_enabled = {
1719 .name = "memory_pressure_enabled",
1720 .read_u64 = cpuset_read_u64,
1721 .write_u64 = cpuset_write_u64,
1722 .private = FILE_MEMORY_PRESSURE_ENABLED,
1725 static int cpuset_populate(struct cgroup_subsys *ss, struct cgroup *cont)
1729 err = cgroup_add_files(cont, ss, files, ARRAY_SIZE(files));
1732 /* memory_pressure_enabled is in root cpuset only */
1734 err = cgroup_add_file(cont, ss,
1735 &cft_memory_pressure_enabled);
1740 * post_clone() is called at the end of cgroup_clone().
1741 * 'cgroup' was just created automatically as a result of
1742 * a cgroup_clone(), and the current task is about to
1743 * be moved into 'cgroup'.
1745 * Currently we refuse to set up the cgroup - thereby
1746 * refusing the task to be entered, and as a result refusing
1747 * the sys_unshare() or clone() which initiated it - if any
1748 * sibling cpusets have exclusive cpus or mem.
1750 * If this becomes a problem for some users who wish to
1751 * allow that scenario, then cpuset_post_clone() could be
1752 * changed to grant parent->cpus_allowed-sibling_cpus_exclusive
1753 * (and likewise for mems) to the new cgroup. Called with cgroup_mutex
1756 static void cpuset_post_clone(struct cgroup_subsys *ss,
1757 struct cgroup *cgroup)
1759 struct cgroup *parent, *child;
1760 struct cpuset *cs, *parent_cs;
1762 parent = cgroup->parent;
1763 list_for_each_entry(child, &parent->children, sibling) {
1764 cs = cgroup_cs(child);
1765 if (is_mem_exclusive(cs) || is_cpu_exclusive(cs))
1768 cs = cgroup_cs(cgroup);
1769 parent_cs = cgroup_cs(parent);
1771 cs->mems_allowed = parent_cs->mems_allowed;
1772 cpumask_copy(cs->cpus_allowed, parent_cs->cpus_allowed);
1777 * cpuset_create - create a cpuset
1778 * ss: cpuset cgroup subsystem
1779 * cont: control group that the new cpuset will be part of
1782 static struct cgroup_subsys_state *cpuset_create(
1783 struct cgroup_subsys *ss,
1784 struct cgroup *cont)
1787 struct cpuset *parent;
1789 if (!cont->parent) {
1790 /* This is early initialization for the top cgroup */
1791 top_cpuset.mems_generation = cpuset_mems_generation++;
1792 return &top_cpuset.css;
1794 parent = cgroup_cs(cont->parent);
1795 cs = kmalloc(sizeof(*cs), GFP_KERNEL);
1797 return ERR_PTR(-ENOMEM);
1798 if (!alloc_cpumask_var(&cs->cpus_allowed, GFP_KERNEL)) {
1800 return ERR_PTR(-ENOMEM);
1803 cpuset_update_task_memory_state();
1805 if (is_spread_page(parent))
1806 set_bit(CS_SPREAD_PAGE, &cs->flags);
1807 if (is_spread_slab(parent))
1808 set_bit(CS_SPREAD_SLAB, &cs->flags);
1809 set_bit(CS_SCHED_LOAD_BALANCE, &cs->flags);
1810 cpumask_clear(cs->cpus_allowed);
1811 nodes_clear(cs->mems_allowed);
1812 cs->mems_generation = cpuset_mems_generation++;
1813 fmeter_init(&cs->fmeter);
1814 cs->relax_domain_level = -1;
1816 cs->parent = parent;
1817 number_of_cpusets++;
1822 * If the cpuset being removed has its flag 'sched_load_balance'
1823 * enabled, then simulate turning sched_load_balance off, which
1824 * will call async_rebuild_sched_domains().
1827 static void cpuset_destroy(struct cgroup_subsys *ss, struct cgroup *cont)
1829 struct cpuset *cs = cgroup_cs(cont);
1831 cpuset_update_task_memory_state();
1833 if (is_sched_load_balance(cs))
1834 update_flag(CS_SCHED_LOAD_BALANCE, cs, 0);
1836 number_of_cpusets--;
1837 free_cpumask_var(cs->cpus_allowed);
1841 struct cgroup_subsys cpuset_subsys = {
1843 .create = cpuset_create,
1844 .destroy = cpuset_destroy,
1845 .can_attach = cpuset_can_attach,
1846 .attach = cpuset_attach,
1847 .populate = cpuset_populate,
1848 .post_clone = cpuset_post_clone,
1849 .subsys_id = cpuset_subsys_id,
1854 * cpuset_init_early - just enough so that the calls to
1855 * cpuset_update_task_memory_state() in early init code
1859 int __init cpuset_init_early(void)
1861 alloc_bootmem_cpumask_var(&top_cpuset.cpus_allowed);
1863 top_cpuset.mems_generation = cpuset_mems_generation++;
1869 * cpuset_init - initialize cpusets at system boot
1871 * Description: Initialize top_cpuset and the cpuset internal file system,
1874 int __init cpuset_init(void)
1878 cpumask_setall(top_cpuset.cpus_allowed);
1879 nodes_setall(top_cpuset.mems_allowed);
1881 fmeter_init(&top_cpuset.fmeter);
1882 top_cpuset.mems_generation = cpuset_mems_generation++;
1883 set_bit(CS_SCHED_LOAD_BALANCE, &top_cpuset.flags);
1884 top_cpuset.relax_domain_level = -1;
1886 err = register_filesystem(&cpuset_fs_type);
1890 if (!alloc_cpumask_var(&cpus_attach, GFP_KERNEL))
1893 number_of_cpusets = 1;
1898 * cpuset_do_move_task - move a given task to another cpuset
1899 * @tsk: pointer to task_struct the task to move
1900 * @scan: struct cgroup_scanner contained in its struct cpuset_hotplug_scanner
1902 * Called by cgroup_scan_tasks() for each task in a cgroup.
1903 * Return nonzero to stop the walk through the tasks.
1905 static void cpuset_do_move_task(struct task_struct *tsk,
1906 struct cgroup_scanner *scan)
1908 struct cpuset_hotplug_scanner *chsp;
1910 chsp = container_of(scan, struct cpuset_hotplug_scanner, scan);
1911 cgroup_attach_task(chsp->to, tsk);
1915 * move_member_tasks_to_cpuset - move tasks from one cpuset to another
1916 * @from: cpuset in which the tasks currently reside
1917 * @to: cpuset to which the tasks will be moved
1919 * Called with cgroup_mutex held
1920 * callback_mutex must not be held, as cpuset_attach() will take it.
1922 * The cgroup_scan_tasks() function will scan all the tasks in a cgroup,
1923 * calling callback functions for each.
1925 static void move_member_tasks_to_cpuset(struct cpuset *from, struct cpuset *to)
1927 struct cpuset_hotplug_scanner scan;
1929 scan.scan.cg = from->css.cgroup;
1930 scan.scan.test_task = NULL; /* select all tasks in cgroup */
1931 scan.scan.process_task = cpuset_do_move_task;
1932 scan.scan.heap = NULL;
1933 scan.to = to->css.cgroup;
1935 if (cgroup_scan_tasks(&scan.scan))
1936 printk(KERN_ERR "move_member_tasks_to_cpuset: "
1937 "cgroup_scan_tasks failed\n");
1941 * If CPU and/or memory hotplug handlers, below, unplug any CPUs
1942 * or memory nodes, we need to walk over the cpuset hierarchy,
1943 * removing that CPU or node from all cpusets. If this removes the
1944 * last CPU or node from a cpuset, then move the tasks in the empty
1945 * cpuset to its next-highest non-empty parent.
1947 * Called with cgroup_mutex held
1948 * callback_mutex must not be held, as cpuset_attach() will take it.
1950 static void remove_tasks_in_empty_cpuset(struct cpuset *cs)
1952 struct cpuset *parent;
1955 * The cgroup's css_sets list is in use if there are tasks
1956 * in the cpuset; the list is empty if there are none;
1957 * the cs->css.refcnt seems always 0.
1959 if (list_empty(&cs->css.cgroup->css_sets))
1963 * Find its next-highest non-empty parent, (top cpuset
1964 * has online cpus, so can't be empty).
1966 parent = cs->parent;
1967 while (cpumask_empty(parent->cpus_allowed) ||
1968 nodes_empty(parent->mems_allowed))
1969 parent = parent->parent;
1971 move_member_tasks_to_cpuset(cs, parent);
1975 * Walk the specified cpuset subtree and look for empty cpusets.
1976 * The tasks of such cpuset must be moved to a parent cpuset.
1978 * Called with cgroup_mutex held. We take callback_mutex to modify
1979 * cpus_allowed and mems_allowed.
1981 * This walk processes the tree from top to bottom, completing one layer
1982 * before dropping down to the next. It always processes a node before
1983 * any of its children.
1985 * For now, since we lack memory hot unplug, we'll never see a cpuset
1986 * that has tasks along with an empty 'mems'. But if we did see such
1987 * a cpuset, we'd handle it just like we do if its 'cpus' was empty.
1989 static void scan_for_empty_cpusets(struct cpuset *root)
1992 struct cpuset *cp; /* scans cpusets being updated */
1993 struct cpuset *child; /* scans child cpusets of cp */
1994 struct cgroup *cont;
1997 list_add_tail((struct list_head *)&root->stack_list, &queue);
1999 while (!list_empty(&queue)) {
2000 cp = list_first_entry(&queue, struct cpuset, stack_list);
2001 list_del(queue.next);
2002 list_for_each_entry(cont, &cp->css.cgroup->children, sibling) {
2003 child = cgroup_cs(cont);
2004 list_add_tail(&child->stack_list, &queue);
2007 /* Continue past cpusets with all cpus, mems online */
2008 if (cpumask_subset(cp->cpus_allowed, cpu_online_mask) &&
2009 nodes_subset(cp->mems_allowed, node_states[N_HIGH_MEMORY]))
2012 oldmems = cp->mems_allowed;
2014 /* Remove offline cpus and mems from this cpuset. */
2015 mutex_lock(&callback_mutex);
2016 cpumask_and(cp->cpus_allowed, cp->cpus_allowed,
2018 nodes_and(cp->mems_allowed, cp->mems_allowed,
2019 node_states[N_HIGH_MEMORY]);
2020 mutex_unlock(&callback_mutex);
2022 /* Move tasks from the empty cpuset to a parent */
2023 if (cpumask_empty(cp->cpus_allowed) ||
2024 nodes_empty(cp->mems_allowed))
2025 remove_tasks_in_empty_cpuset(cp);
2027 update_tasks_cpumask(cp, NULL);
2028 update_tasks_nodemask(cp, &oldmems);
2034 * The top_cpuset tracks what CPUs and Memory Nodes are online,
2035 * period. This is necessary in order to make cpusets transparent
2036 * (of no affect) on systems that are actively using CPU hotplug
2037 * but making no active use of cpusets.
2039 * This routine ensures that top_cpuset.cpus_allowed tracks
2040 * cpu_online_map on each CPU hotplug (cpuhp) event.
2042 * Called within get_online_cpus(). Needs to call cgroup_lock()
2043 * before calling generate_sched_domains().
2045 static int cpuset_track_online_cpus(struct notifier_block *unused_nb,
2046 unsigned long phase, void *unused_cpu)
2048 struct sched_domain_attr *attr;
2049 struct cpumask *doms;
2054 case CPU_ONLINE_FROZEN:
2056 case CPU_DEAD_FROZEN:
2064 cpumask_copy(top_cpuset.cpus_allowed, cpu_online_mask);
2065 scan_for_empty_cpusets(&top_cpuset);
2066 ndoms = generate_sched_domains(&doms, &attr);
2069 /* Have scheduler rebuild the domains */
2070 partition_sched_domains(ndoms, doms, attr);
2075 #ifdef CONFIG_MEMORY_HOTPLUG
2077 * Keep top_cpuset.mems_allowed tracking node_states[N_HIGH_MEMORY].
2078 * Call this routine anytime after node_states[N_HIGH_MEMORY] changes.
2079 * See also the previous routine cpuset_track_online_cpus().
2081 static int cpuset_track_online_nodes(struct notifier_block *self,
2082 unsigned long action, void *arg)
2087 top_cpuset.mems_allowed = node_states[N_HIGH_MEMORY];
2090 top_cpuset.mems_allowed = node_states[N_HIGH_MEMORY];
2091 scan_for_empty_cpusets(&top_cpuset);
2102 * cpuset_init_smp - initialize cpus_allowed
2104 * Description: Finish top cpuset after cpu, node maps are initialized
2107 void __init cpuset_init_smp(void)
2109 cpumask_copy(top_cpuset.cpus_allowed, cpu_online_mask);
2110 top_cpuset.mems_allowed = node_states[N_HIGH_MEMORY];
2112 hotcpu_notifier(cpuset_track_online_cpus, 0);
2113 hotplug_memory_notifier(cpuset_track_online_nodes, 10);
2117 * cpuset_cpus_allowed - return cpus_allowed mask from a tasks cpuset.
2118 * @tsk: pointer to task_struct from which to obtain cpuset->cpus_allowed.
2119 * @pmask: pointer to struct cpumask variable to receive cpus_allowed set.
2121 * Description: Returns the cpumask_var_t cpus_allowed of the cpuset
2122 * attached to the specified @tsk. Guaranteed to return some non-empty
2123 * subset of cpu_online_map, even if this means going outside the
2127 void cpuset_cpus_allowed(struct task_struct *tsk, struct cpumask *pmask)
2129 mutex_lock(&callback_mutex);
2130 cpuset_cpus_allowed_locked(tsk, pmask);
2131 mutex_unlock(&callback_mutex);
2135 * cpuset_cpus_allowed_locked - return cpus_allowed mask from a tasks cpuset.
2136 * Must be called with callback_mutex held.
2138 void cpuset_cpus_allowed_locked(struct task_struct *tsk, struct cpumask *pmask)
2141 guarantee_online_cpus(task_cs(tsk), pmask);
2145 void cpuset_init_current_mems_allowed(void)
2147 nodes_setall(current->mems_allowed);
2151 * cpuset_mems_allowed - return mems_allowed mask from a tasks cpuset.
2152 * @tsk: pointer to task_struct from which to obtain cpuset->mems_allowed.
2154 * Description: Returns the nodemask_t mems_allowed of the cpuset
2155 * attached to the specified @tsk. Guaranteed to return some non-empty
2156 * subset of node_states[N_HIGH_MEMORY], even if this means going outside the
2160 nodemask_t cpuset_mems_allowed(struct task_struct *tsk)
2164 mutex_lock(&callback_mutex);
2166 guarantee_online_mems(task_cs(tsk), &mask);
2168 mutex_unlock(&callback_mutex);
2174 * cpuset_nodemask_valid_mems_allowed - check nodemask vs. curremt mems_allowed
2175 * @nodemask: the nodemask to be checked
2177 * Are any of the nodes in the nodemask allowed in current->mems_allowed?
2179 int cpuset_nodemask_valid_mems_allowed(nodemask_t *nodemask)
2181 return nodes_intersects(*nodemask, current->mems_allowed);
2185 * nearest_hardwall_ancestor() - Returns the nearest mem_exclusive or
2186 * mem_hardwall ancestor to the specified cpuset. Call holding
2187 * callback_mutex. If no ancestor is mem_exclusive or mem_hardwall
2188 * (an unusual configuration), then returns the root cpuset.
2190 static const struct cpuset *nearest_hardwall_ancestor(const struct cpuset *cs)
2192 while (!(is_mem_exclusive(cs) || is_mem_hardwall(cs)) && cs->parent)
2198 * cpuset_zone_allowed_softwall - Can we allocate on zone z's memory node?
2199 * @z: is this zone on an allowed node?
2200 * @gfp_mask: memory allocation flags
2202 * If we're in interrupt, yes, we can always allocate. If
2203 * __GFP_THISNODE is set, yes, we can always allocate. If zone
2204 * z's node is in our tasks mems_allowed, yes. If it's not a
2205 * __GFP_HARDWALL request and this zone's nodes is in the nearest
2206 * hardwalled cpuset ancestor to this tasks cpuset, yes.
2207 * If the task has been OOM killed and has access to memory reserves
2208 * as specified by the TIF_MEMDIE flag, yes.
2211 * If __GFP_HARDWALL is set, cpuset_zone_allowed_softwall()
2212 * reduces to cpuset_zone_allowed_hardwall(). Otherwise,
2213 * cpuset_zone_allowed_softwall() might sleep, and might allow a zone
2214 * from an enclosing cpuset.
2216 * cpuset_zone_allowed_hardwall() only handles the simpler case of
2217 * hardwall cpusets, and never sleeps.
2219 * The __GFP_THISNODE placement logic is really handled elsewhere,
2220 * by forcibly using a zonelist starting at a specified node, and by
2221 * (in get_page_from_freelist()) refusing to consider the zones for
2222 * any node on the zonelist except the first. By the time any such
2223 * calls get to this routine, we should just shut up and say 'yes'.
2225 * GFP_USER allocations are marked with the __GFP_HARDWALL bit,
2226 * and do not allow allocations outside the current tasks cpuset
2227 * unless the task has been OOM killed as is marked TIF_MEMDIE.
2228 * GFP_KERNEL allocations are not so marked, so can escape to the
2229 * nearest enclosing hardwalled ancestor cpuset.
2231 * Scanning up parent cpusets requires callback_mutex. The
2232 * __alloc_pages() routine only calls here with __GFP_HARDWALL bit
2233 * _not_ set if it's a GFP_KERNEL allocation, and all nodes in the
2234 * current tasks mems_allowed came up empty on the first pass over
2235 * the zonelist. So only GFP_KERNEL allocations, if all nodes in the
2236 * cpuset are short of memory, might require taking the callback_mutex
2239 * The first call here from mm/page_alloc:get_page_from_freelist()
2240 * has __GFP_HARDWALL set in gfp_mask, enforcing hardwall cpusets,
2241 * so no allocation on a node outside the cpuset is allowed (unless
2242 * in interrupt, of course).
2244 * The second pass through get_page_from_freelist() doesn't even call
2245 * here for GFP_ATOMIC calls. For those calls, the __alloc_pages()
2246 * variable 'wait' is not set, and the bit ALLOC_CPUSET is not set
2247 * in alloc_flags. That logic and the checks below have the combined
2249 * in_interrupt - any node ok (current task context irrelevant)
2250 * GFP_ATOMIC - any node ok
2251 * TIF_MEMDIE - any node ok
2252 * GFP_KERNEL - any node in enclosing hardwalled cpuset ok
2253 * GFP_USER - only nodes in current tasks mems allowed ok.
2256 * Don't call cpuset_zone_allowed_softwall if you can't sleep, unless you
2257 * pass in the __GFP_HARDWALL flag set in gfp_flag, which disables
2258 * the code that might scan up ancestor cpusets and sleep.
2261 int __cpuset_zone_allowed_softwall(struct zone *z, gfp_t gfp_mask)
2263 int node; /* node that zone z is on */
2264 const struct cpuset *cs; /* current cpuset ancestors */
2265 int allowed; /* is allocation in zone z allowed? */
2267 if (in_interrupt() || (gfp_mask & __GFP_THISNODE))
2269 node = zone_to_nid(z);
2270 might_sleep_if(!(gfp_mask & __GFP_HARDWALL));
2271 if (node_isset(node, current->mems_allowed))
2274 * Allow tasks that have access to memory reserves because they have
2275 * been OOM killed to get memory anywhere.
2277 if (unlikely(test_thread_flag(TIF_MEMDIE)))
2279 if (gfp_mask & __GFP_HARDWALL) /* If hardwall request, stop here */
2282 if (current->flags & PF_EXITING) /* Let dying task have memory */
2285 /* Not hardwall and node outside mems_allowed: scan up cpusets */
2286 mutex_lock(&callback_mutex);
2289 cs = nearest_hardwall_ancestor(task_cs(current));
2290 task_unlock(current);
2292 allowed = node_isset(node, cs->mems_allowed);
2293 mutex_unlock(&callback_mutex);
2298 * cpuset_zone_allowed_hardwall - Can we allocate on zone z's memory node?
2299 * @z: is this zone on an allowed node?
2300 * @gfp_mask: memory allocation flags
2302 * If we're in interrupt, yes, we can always allocate.
2303 * If __GFP_THISNODE is set, yes, we can always allocate. If zone
2304 * z's node is in our tasks mems_allowed, yes. If the task has been
2305 * OOM killed and has access to memory reserves as specified by the
2306 * TIF_MEMDIE flag, yes. Otherwise, no.
2308 * The __GFP_THISNODE placement logic is really handled elsewhere,
2309 * by forcibly using a zonelist starting at a specified node, and by
2310 * (in get_page_from_freelist()) refusing to consider the zones for
2311 * any node on the zonelist except the first. By the time any such
2312 * calls get to this routine, we should just shut up and say 'yes'.
2314 * Unlike the cpuset_zone_allowed_softwall() variant, above,
2315 * this variant requires that the zone be in the current tasks
2316 * mems_allowed or that we're in interrupt. It does not scan up the
2317 * cpuset hierarchy for the nearest enclosing mem_exclusive cpuset.
2321 int __cpuset_zone_allowed_hardwall(struct zone *z, gfp_t gfp_mask)
2323 int node; /* node that zone z is on */
2325 if (in_interrupt() || (gfp_mask & __GFP_THISNODE))
2327 node = zone_to_nid(z);
2328 if (node_isset(node, current->mems_allowed))
2331 * Allow tasks that have access to memory reserves because they have
2332 * been OOM killed to get memory anywhere.
2334 if (unlikely(test_thread_flag(TIF_MEMDIE)))
2340 * cpuset_lock - lock out any changes to cpuset structures
2342 * The out of memory (oom) code needs to mutex_lock cpusets
2343 * from being changed while it scans the tasklist looking for a
2344 * task in an overlapping cpuset. Expose callback_mutex via this
2345 * cpuset_lock() routine, so the oom code can lock it, before
2346 * locking the task list. The tasklist_lock is a spinlock, so
2347 * must be taken inside callback_mutex.
2350 void cpuset_lock(void)
2352 mutex_lock(&callback_mutex);
2356 * cpuset_unlock - release lock on cpuset changes
2358 * Undo the lock taken in a previous cpuset_lock() call.
2361 void cpuset_unlock(void)
2363 mutex_unlock(&callback_mutex);
2367 * cpuset_mem_spread_node() - On which node to begin search for a page
2369 * If a task is marked PF_SPREAD_PAGE or PF_SPREAD_SLAB (as for
2370 * tasks in a cpuset with is_spread_page or is_spread_slab set),
2371 * and if the memory allocation used cpuset_mem_spread_node()
2372 * to determine on which node to start looking, as it will for
2373 * certain page cache or slab cache pages such as used for file
2374 * system buffers and inode caches, then instead of starting on the
2375 * local node to look for a free page, rather spread the starting
2376 * node around the tasks mems_allowed nodes.
2378 * We don't have to worry about the returned node being offline
2379 * because "it can't happen", and even if it did, it would be ok.
2381 * The routines calling guarantee_online_mems() are careful to
2382 * only set nodes in task->mems_allowed that are online. So it
2383 * should not be possible for the following code to return an
2384 * offline node. But if it did, that would be ok, as this routine
2385 * is not returning the node where the allocation must be, only
2386 * the node where the search should start. The zonelist passed to
2387 * __alloc_pages() will include all nodes. If the slab allocator
2388 * is passed an offline node, it will fall back to the local node.
2389 * See kmem_cache_alloc_node().
2392 int cpuset_mem_spread_node(void)
2396 node = next_node(current->cpuset_mem_spread_rotor, current->mems_allowed);
2397 if (node == MAX_NUMNODES)
2398 node = first_node(current->mems_allowed);
2399 current->cpuset_mem_spread_rotor = node;
2402 EXPORT_SYMBOL_GPL(cpuset_mem_spread_node);
2405 * cpuset_mems_allowed_intersects - Does @tsk1's mems_allowed intersect @tsk2's?
2406 * @tsk1: pointer to task_struct of some task.
2407 * @tsk2: pointer to task_struct of some other task.
2409 * Description: Return true if @tsk1's mems_allowed intersects the
2410 * mems_allowed of @tsk2. Used by the OOM killer to determine if
2411 * one of the task's memory usage might impact the memory available
2415 int cpuset_mems_allowed_intersects(const struct task_struct *tsk1,
2416 const struct task_struct *tsk2)
2418 return nodes_intersects(tsk1->mems_allowed, tsk2->mems_allowed);
2422 * cpuset_print_task_mems_allowed - prints task's cpuset and mems_allowed
2423 * @task: pointer to task_struct of some task.
2425 * Description: Prints @task's name, cpuset name, and cached copy of its
2426 * mems_allowed to the kernel log. Must hold task_lock(task) to allow
2427 * dereferencing task_cs(task).
2429 void cpuset_print_task_mems_allowed(struct task_struct *tsk)
2431 struct dentry *dentry;
2433 dentry = task_cs(tsk)->css.cgroup->dentry;
2434 spin_lock(&cpuset_buffer_lock);
2435 snprintf(cpuset_name, CPUSET_NAME_LEN,
2436 dentry ? (const char *)dentry->d_name.name : "/");
2437 nodelist_scnprintf(cpuset_nodelist, CPUSET_NODELIST_LEN,
2439 printk(KERN_INFO "%s cpuset=%s mems_allowed=%s\n",
2440 tsk->comm, cpuset_name, cpuset_nodelist);
2441 spin_unlock(&cpuset_buffer_lock);
2445 * Collection of memory_pressure is suppressed unless
2446 * this flag is enabled by writing "1" to the special
2447 * cpuset file 'memory_pressure_enabled' in the root cpuset.
2450 int cpuset_memory_pressure_enabled __read_mostly;
2453 * cpuset_memory_pressure_bump - keep stats of per-cpuset reclaims.
2455 * Keep a running average of the rate of synchronous (direct)
2456 * page reclaim efforts initiated by tasks in each cpuset.
2458 * This represents the rate at which some task in the cpuset
2459 * ran low on memory on all nodes it was allowed to use, and
2460 * had to enter the kernels page reclaim code in an effort to
2461 * create more free memory by tossing clean pages or swapping
2462 * or writing dirty pages.
2464 * Display to user space in the per-cpuset read-only file
2465 * "memory_pressure". Value displayed is an integer
2466 * representing the recent rate of entry into the synchronous
2467 * (direct) page reclaim by any task attached to the cpuset.
2470 void __cpuset_memory_pressure_bump(void)
2473 fmeter_markevent(&task_cs(current)->fmeter);
2474 task_unlock(current);
2477 #ifdef CONFIG_PROC_PID_CPUSET
2479 * proc_cpuset_show()
2480 * - Print tasks cpuset path into seq_file.
2481 * - Used for /proc/<pid>/cpuset.
2482 * - No need to task_lock(tsk) on this tsk->cpuset reference, as it
2483 * doesn't really matter if tsk->cpuset changes after we read it,
2484 * and we take cgroup_mutex, keeping cpuset_attach() from changing it
2487 static int proc_cpuset_show(struct seq_file *m, void *unused_v)
2490 struct task_struct *tsk;
2492 struct cgroup_subsys_state *css;
2496 buf = kmalloc(PAGE_SIZE, GFP_KERNEL);
2502 tsk = get_pid_task(pid, PIDTYPE_PID);
2508 css = task_subsys_state(tsk, cpuset_subsys_id);
2509 retval = cgroup_path(css->cgroup, buf, PAGE_SIZE);
2516 put_task_struct(tsk);
2523 static int cpuset_open(struct inode *inode, struct file *file)
2525 struct pid *pid = PROC_I(inode)->pid;
2526 return single_open(file, proc_cpuset_show, pid);
2529 const struct file_operations proc_cpuset_operations = {
2530 .open = cpuset_open,
2532 .llseek = seq_lseek,
2533 .release = single_release,
2535 #endif /* CONFIG_PROC_PID_CPUSET */
2537 /* Display task cpus_allowed, mems_allowed in /proc/<pid>/status file. */
2538 void cpuset_task_status_allowed(struct seq_file *m, struct task_struct *task)
2540 seq_printf(m, "Cpus_allowed:\t");
2541 seq_cpumask(m, &task->cpus_allowed);
2542 seq_printf(m, "\n");
2543 seq_printf(m, "Cpus_allowed_list:\t");
2544 seq_cpumask_list(m, &task->cpus_allowed);
2545 seq_printf(m, "\n");
2546 seq_printf(m, "Mems_allowed:\t");
2547 seq_nodemask(m, &task->mems_allowed);
2548 seq_printf(m, "\n");
2549 seq_printf(m, "Mems_allowed_list:\t");
2550 seq_nodemask_list(m, &task->mems_allowed);
2551 seq_printf(m, "\n");