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/export.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 <linux/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 */
100 struct fmeter fmeter; /* memory_pressure filter */
102 /* partition number for rebuild_sched_domains() */
105 /* for custom sched domain */
106 int relax_domain_level;
108 /* used for walking a cpuset hierarchy */
109 struct list_head stack_list;
112 /* Retrieve the cpuset for a cgroup */
113 static inline struct cpuset *cgroup_cs(struct cgroup *cont)
115 return container_of(cgroup_subsys_state(cont, cpuset_subsys_id),
119 /* Retrieve the cpuset for a task */
120 static inline struct cpuset *task_cs(struct task_struct *task)
122 return container_of(task_subsys_state(task, cpuset_subsys_id),
127 static inline bool task_has_mempolicy(struct task_struct *task)
129 return task->mempolicy;
132 static inline bool task_has_mempolicy(struct task_struct *task)
139 /* bits in struct cpuset flags field */
145 CS_SCHED_LOAD_BALANCE,
150 /* convenient tests for these bits */
151 static inline int is_cpu_exclusive(const struct cpuset *cs)
153 return test_bit(CS_CPU_EXCLUSIVE, &cs->flags);
156 static inline int is_mem_exclusive(const struct cpuset *cs)
158 return test_bit(CS_MEM_EXCLUSIVE, &cs->flags);
161 static inline int is_mem_hardwall(const struct cpuset *cs)
163 return test_bit(CS_MEM_HARDWALL, &cs->flags);
166 static inline int is_sched_load_balance(const struct cpuset *cs)
168 return test_bit(CS_SCHED_LOAD_BALANCE, &cs->flags);
171 static inline int is_memory_migrate(const struct cpuset *cs)
173 return test_bit(CS_MEMORY_MIGRATE, &cs->flags);
176 static inline int is_spread_page(const struct cpuset *cs)
178 return test_bit(CS_SPREAD_PAGE, &cs->flags);
181 static inline int is_spread_slab(const struct cpuset *cs)
183 return test_bit(CS_SPREAD_SLAB, &cs->flags);
186 static struct cpuset top_cpuset = {
187 .flags = ((1 << CS_CPU_EXCLUSIVE) | (1 << CS_MEM_EXCLUSIVE)),
191 * There are two global mutexes guarding cpuset structures. The first
192 * is the main control groups cgroup_mutex, accessed via
193 * cgroup_lock()/cgroup_unlock(). The second is the cpuset-specific
194 * callback_mutex, below. They can nest. It is ok to first take
195 * cgroup_mutex, then nest callback_mutex. We also require taking
196 * task_lock() when dereferencing a task's cpuset pointer. See "The
197 * task_lock() exception", at the end of this comment.
199 * A task must hold both mutexes to modify cpusets. If a task
200 * holds cgroup_mutex, then it blocks others wanting that mutex,
201 * ensuring that it is the only task able to also acquire callback_mutex
202 * and be able to modify cpusets. It can perform various checks on
203 * the cpuset structure first, knowing nothing will change. It can
204 * also allocate memory while just holding cgroup_mutex. While it is
205 * performing these checks, various callback routines can briefly
206 * acquire callback_mutex to query cpusets. Once it is ready to make
207 * the changes, it takes callback_mutex, blocking everyone else.
209 * Calls to the kernel memory allocator can not be made while holding
210 * callback_mutex, as that would risk double tripping on callback_mutex
211 * from one of the callbacks into the cpuset code from within
214 * If a task is only holding callback_mutex, then it has read-only
217 * Now, the task_struct fields mems_allowed and mempolicy may be changed
218 * by other task, we use alloc_lock in the task_struct fields to protect
221 * The cpuset_common_file_read() handlers only hold callback_mutex across
222 * small pieces of code, such as when reading out possibly multi-word
223 * cpumasks and nodemasks.
225 * Accessing a task's cpuset should be done in accordance with the
226 * guidelines for accessing subsystem state in kernel/cgroup.c
229 static DEFINE_MUTEX(callback_mutex);
232 * cpuset_buffer_lock protects both the cpuset_name and cpuset_nodelist
233 * buffers. They are statically allocated to prevent using excess stack
234 * when calling cpuset_print_task_mems_allowed().
236 #define CPUSET_NAME_LEN (128)
237 #define CPUSET_NODELIST_LEN (256)
238 static char cpuset_name[CPUSET_NAME_LEN];
239 static char cpuset_nodelist[CPUSET_NODELIST_LEN];
240 static DEFINE_SPINLOCK(cpuset_buffer_lock);
243 * This is ugly, but preserves the userspace API for existing cpuset
244 * users. If someone tries to mount the "cpuset" filesystem, we
245 * silently switch it to mount "cgroup" instead
247 static struct dentry *cpuset_mount(struct file_system_type *fs_type,
248 int flags, const char *unused_dev_name, void *data)
250 struct file_system_type *cgroup_fs = get_fs_type("cgroup");
251 struct dentry *ret = ERR_PTR(-ENODEV);
255 "release_agent=/sbin/cpuset_release_agent";
256 ret = cgroup_fs->mount(cgroup_fs, flags,
257 unused_dev_name, mountopts);
258 put_filesystem(cgroup_fs);
263 static struct file_system_type cpuset_fs_type = {
265 .mount = cpuset_mount,
269 * Return in pmask the portion of a cpusets's cpus_allowed that
270 * are online. If none are online, walk up the cpuset hierarchy
271 * until we find one that does have some online cpus. If we get
272 * all the way to the top and still haven't found any online cpus,
273 * return cpu_online_map. Or if passed a NULL cs from an exit'ing
274 * task, return cpu_online_map.
276 * One way or another, we guarantee to return some non-empty subset
279 * Call with callback_mutex held.
282 static void guarantee_online_cpus(const struct cpuset *cs,
283 struct cpumask *pmask)
285 while (cs && !cpumask_intersects(cs->cpus_allowed, cpu_online_mask))
288 cpumask_and(pmask, cs->cpus_allowed, cpu_online_mask);
290 cpumask_copy(pmask, cpu_online_mask);
291 BUG_ON(!cpumask_intersects(pmask, cpu_online_mask));
295 * Return in *pmask the portion of a cpusets's mems_allowed that
296 * are online, with memory. If none are online with memory, walk
297 * up the cpuset hierarchy until we find one that does have some
298 * online mems. If we get all the way to the top and still haven't
299 * found any online mems, return node_states[N_HIGH_MEMORY].
301 * One way or another, we guarantee to return some non-empty subset
302 * of node_states[N_HIGH_MEMORY].
304 * Call with callback_mutex held.
307 static void guarantee_online_mems(const struct cpuset *cs, nodemask_t *pmask)
309 while (cs && !nodes_intersects(cs->mems_allowed,
310 node_states[N_HIGH_MEMORY]))
313 nodes_and(*pmask, cs->mems_allowed,
314 node_states[N_HIGH_MEMORY]);
316 *pmask = node_states[N_HIGH_MEMORY];
317 BUG_ON(!nodes_intersects(*pmask, node_states[N_HIGH_MEMORY]));
321 * update task's spread flag if cpuset's page/slab spread flag is set
323 * Called with callback_mutex/cgroup_mutex held
325 static void cpuset_update_task_spread_flag(struct cpuset *cs,
326 struct task_struct *tsk)
328 if (is_spread_page(cs))
329 tsk->flags |= PF_SPREAD_PAGE;
331 tsk->flags &= ~PF_SPREAD_PAGE;
332 if (is_spread_slab(cs))
333 tsk->flags |= PF_SPREAD_SLAB;
335 tsk->flags &= ~PF_SPREAD_SLAB;
339 * is_cpuset_subset(p, q) - Is cpuset p a subset of cpuset q?
341 * One cpuset is a subset of another if all its allowed CPUs and
342 * Memory Nodes are a subset of the other, and its exclusive flags
343 * are only set if the other's are set. Call holding cgroup_mutex.
346 static int is_cpuset_subset(const struct cpuset *p, const struct cpuset *q)
348 return cpumask_subset(p->cpus_allowed, q->cpus_allowed) &&
349 nodes_subset(p->mems_allowed, q->mems_allowed) &&
350 is_cpu_exclusive(p) <= is_cpu_exclusive(q) &&
351 is_mem_exclusive(p) <= is_mem_exclusive(q);
355 * alloc_trial_cpuset - allocate a trial cpuset
356 * @cs: the cpuset that the trial cpuset duplicates
358 static struct cpuset *alloc_trial_cpuset(const struct cpuset *cs)
360 struct cpuset *trial;
362 trial = kmemdup(cs, sizeof(*cs), GFP_KERNEL);
366 if (!alloc_cpumask_var(&trial->cpus_allowed, GFP_KERNEL)) {
370 cpumask_copy(trial->cpus_allowed, cs->cpus_allowed);
376 * free_trial_cpuset - free the trial cpuset
377 * @trial: the trial cpuset to be freed
379 static void free_trial_cpuset(struct cpuset *trial)
381 free_cpumask_var(trial->cpus_allowed);
386 * validate_change() - Used to validate that any proposed cpuset change
387 * follows the structural rules for cpusets.
389 * If we replaced the flag and mask values of the current cpuset
390 * (cur) with those values in the trial cpuset (trial), would
391 * our various subset and exclusive rules still be valid? Presumes
394 * 'cur' is the address of an actual, in-use cpuset. Operations
395 * such as list traversal that depend on the actual address of the
396 * cpuset in the list must use cur below, not trial.
398 * 'trial' is the address of bulk structure copy of cur, with
399 * perhaps one or more of the fields cpus_allowed, mems_allowed,
400 * or flags changed to new, trial values.
402 * Return 0 if valid, -errno if not.
405 static int validate_change(const struct cpuset *cur, const struct cpuset *trial)
408 struct cpuset *c, *par;
410 /* Each of our child cpusets must be a subset of us */
411 list_for_each_entry(cont, &cur->css.cgroup->children, sibling) {
412 if (!is_cpuset_subset(cgroup_cs(cont), trial))
416 /* Remaining checks don't apply to root cpuset */
417 if (cur == &top_cpuset)
422 /* We must be a subset of our parent cpuset */
423 if (!is_cpuset_subset(trial, par))
427 * If either I or some sibling (!= me) is exclusive, we can't
430 list_for_each_entry(cont, &par->css.cgroup->children, sibling) {
432 if ((is_cpu_exclusive(trial) || is_cpu_exclusive(c)) &&
434 cpumask_intersects(trial->cpus_allowed, c->cpus_allowed))
436 if ((is_mem_exclusive(trial) || is_mem_exclusive(c)) &&
438 nodes_intersects(trial->mems_allowed, c->mems_allowed))
442 /* Cpusets with tasks can't have empty cpus_allowed or mems_allowed */
443 if (cgroup_task_count(cur->css.cgroup)) {
444 if (cpumask_empty(trial->cpus_allowed) ||
445 nodes_empty(trial->mems_allowed)) {
455 * Helper routine for generate_sched_domains().
456 * Do cpusets a, b have overlapping cpus_allowed masks?
458 static int cpusets_overlap(struct cpuset *a, struct cpuset *b)
460 return cpumask_intersects(a->cpus_allowed, b->cpus_allowed);
464 update_domain_attr(struct sched_domain_attr *dattr, struct cpuset *c)
466 if (dattr->relax_domain_level < c->relax_domain_level)
467 dattr->relax_domain_level = c->relax_domain_level;
472 update_domain_attr_tree(struct sched_domain_attr *dattr, struct cpuset *c)
476 list_add(&c->stack_list, &q);
477 while (!list_empty(&q)) {
480 struct cpuset *child;
482 cp = list_first_entry(&q, struct cpuset, stack_list);
485 if (cpumask_empty(cp->cpus_allowed))
488 if (is_sched_load_balance(cp))
489 update_domain_attr(dattr, cp);
491 list_for_each_entry(cont, &cp->css.cgroup->children, sibling) {
492 child = cgroup_cs(cont);
493 list_add_tail(&child->stack_list, &q);
499 * generate_sched_domains()
501 * This function builds a partial partition of the systems CPUs
502 * A 'partial partition' is a set of non-overlapping subsets whose
503 * union is a subset of that set.
504 * The output of this function needs to be passed to kernel/sched.c
505 * partition_sched_domains() routine, which will rebuild the scheduler's
506 * load balancing domains (sched domains) as specified by that partial
509 * See "What is sched_load_balance" in Documentation/cgroups/cpusets.txt
510 * for a background explanation of this.
512 * Does not return errors, on the theory that the callers of this
513 * routine would rather not worry about failures to rebuild sched
514 * domains when operating in the severe memory shortage situations
515 * that could cause allocation failures below.
517 * Must be called with cgroup_lock held.
519 * The three key local variables below are:
520 * q - a linked-list queue of cpuset pointers, used to implement a
521 * top-down scan of all cpusets. This scan loads a pointer
522 * to each cpuset marked is_sched_load_balance into the
523 * array 'csa'. For our purposes, rebuilding the schedulers
524 * sched domains, we can ignore !is_sched_load_balance cpusets.
525 * csa - (for CpuSet Array) Array of pointers to all the cpusets
526 * that need to be load balanced, for convenient iterative
527 * access by the subsequent code that finds the best partition,
528 * i.e the set of domains (subsets) of CPUs such that the
529 * cpus_allowed of every cpuset marked is_sched_load_balance
530 * is a subset of one of these domains, while there are as
531 * many such domains as possible, each as small as possible.
532 * doms - Conversion of 'csa' to an array of cpumasks, for passing to
533 * the kernel/sched.c routine partition_sched_domains() in a
534 * convenient format, that can be easily compared to the prior
535 * value to determine what partition elements (sched domains)
536 * were changed (added or removed.)
538 * Finding the best partition (set of domains):
539 * The triple nested loops below over i, j, k scan over the
540 * load balanced cpusets (using the array of cpuset pointers in
541 * csa[]) looking for pairs of cpusets that have overlapping
542 * cpus_allowed, but which don't have the same 'pn' partition
543 * number and gives them in the same partition number. It keeps
544 * looping on the 'restart' label until it can no longer find
547 * The union of the cpus_allowed masks from the set of
548 * all cpusets having the same 'pn' value then form the one
549 * element of the partition (one sched domain) to be passed to
550 * partition_sched_domains().
552 static int generate_sched_domains(cpumask_var_t **domains,
553 struct sched_domain_attr **attributes)
555 LIST_HEAD(q); /* queue of cpusets to be scanned */
556 struct cpuset *cp; /* scans q */
557 struct cpuset **csa; /* array of all cpuset ptrs */
558 int csn; /* how many cpuset ptrs in csa so far */
559 int i, j, k; /* indices for partition finding loops */
560 cpumask_var_t *doms; /* resulting partition; i.e. sched domains */
561 struct sched_domain_attr *dattr; /* attributes for custom domains */
562 int ndoms = 0; /* number of sched domains in result */
563 int nslot; /* next empty doms[] struct cpumask slot */
569 /* Special case for the 99% of systems with one, full, sched domain */
570 if (is_sched_load_balance(&top_cpuset)) {
572 doms = alloc_sched_domains(ndoms);
576 dattr = kmalloc(sizeof(struct sched_domain_attr), GFP_KERNEL);
578 *dattr = SD_ATTR_INIT;
579 update_domain_attr_tree(dattr, &top_cpuset);
581 cpumask_copy(doms[0], top_cpuset.cpus_allowed);
586 csa = kmalloc(number_of_cpusets * sizeof(cp), GFP_KERNEL);
591 list_add(&top_cpuset.stack_list, &q);
592 while (!list_empty(&q)) {
594 struct cpuset *child; /* scans child cpusets of cp */
596 cp = list_first_entry(&q, struct cpuset, stack_list);
599 if (cpumask_empty(cp->cpus_allowed))
603 * All child cpusets contain a subset of the parent's cpus, so
604 * just skip them, and then we call update_domain_attr_tree()
605 * to calc relax_domain_level of the corresponding sched
608 if (is_sched_load_balance(cp)) {
613 list_for_each_entry(cont, &cp->css.cgroup->children, sibling) {
614 child = cgroup_cs(cont);
615 list_add_tail(&child->stack_list, &q);
619 for (i = 0; i < csn; i++)
624 /* Find the best partition (set of sched domains) */
625 for (i = 0; i < csn; i++) {
626 struct cpuset *a = csa[i];
629 for (j = 0; j < csn; j++) {
630 struct cpuset *b = csa[j];
633 if (apn != bpn && cpusets_overlap(a, b)) {
634 for (k = 0; k < csn; k++) {
635 struct cpuset *c = csa[k];
640 ndoms--; /* one less element */
647 * Now we know how many domains to create.
648 * Convert <csn, csa> to <ndoms, doms> and populate cpu masks.
650 doms = alloc_sched_domains(ndoms);
655 * The rest of the code, including the scheduler, can deal with
656 * dattr==NULL case. No need to abort if alloc fails.
658 dattr = kmalloc(ndoms * sizeof(struct sched_domain_attr), GFP_KERNEL);
660 for (nslot = 0, i = 0; i < csn; i++) {
661 struct cpuset *a = csa[i];
666 /* Skip completed partitions */
672 if (nslot == ndoms) {
673 static int warnings = 10;
676 "rebuild_sched_domains confused:"
677 " nslot %d, ndoms %d, csn %d, i %d,"
679 nslot, ndoms, csn, i, apn);
687 *(dattr + nslot) = SD_ATTR_INIT;
688 for (j = i; j < csn; j++) {
689 struct cpuset *b = csa[j];
692 cpumask_or(dp, dp, b->cpus_allowed);
694 update_domain_attr_tree(dattr + nslot, b);
696 /* Done with this partition */
702 BUG_ON(nslot != ndoms);
708 * Fallback to the default domain if kmalloc() failed.
709 * See comments in partition_sched_domains().
720 * Rebuild scheduler domains.
722 * Call with neither cgroup_mutex held nor within get_online_cpus().
723 * Takes both cgroup_mutex and get_online_cpus().
725 * Cannot be directly called from cpuset code handling changes
726 * to the cpuset pseudo-filesystem, because it cannot be called
727 * from code that already holds cgroup_mutex.
729 static void do_rebuild_sched_domains(struct work_struct *unused)
731 struct sched_domain_attr *attr;
737 /* Generate domain masks and attrs */
739 ndoms = generate_sched_domains(&doms, &attr);
742 /* Have scheduler rebuild the domains */
743 partition_sched_domains(ndoms, doms, attr);
747 #else /* !CONFIG_SMP */
748 static void do_rebuild_sched_domains(struct work_struct *unused)
752 static int generate_sched_domains(cpumask_var_t **domains,
753 struct sched_domain_attr **attributes)
758 #endif /* CONFIG_SMP */
760 static DECLARE_WORK(rebuild_sched_domains_work, do_rebuild_sched_domains);
763 * Rebuild scheduler domains, asynchronously via workqueue.
765 * If the flag 'sched_load_balance' of any cpuset with non-empty
766 * 'cpus' changes, or if the 'cpus' allowed changes in any cpuset
767 * which has that flag enabled, or if any cpuset with a non-empty
768 * 'cpus' is removed, then call this routine to rebuild the
769 * scheduler's dynamic sched domains.
771 * The rebuild_sched_domains() and partition_sched_domains()
772 * routines must nest cgroup_lock() inside get_online_cpus(),
773 * but such cpuset changes as these must nest that locking the
774 * other way, holding cgroup_lock() for much of the code.
776 * So in order to avoid an ABBA deadlock, the cpuset code handling
777 * these user changes delegates the actual sched domain rebuilding
778 * to a separate workqueue thread, which ends up processing the
779 * above do_rebuild_sched_domains() function.
781 static void async_rebuild_sched_domains(void)
783 queue_work(cpuset_wq, &rebuild_sched_domains_work);
787 * Accomplishes the same scheduler domain rebuild as the above
788 * async_rebuild_sched_domains(), however it directly calls the
789 * rebuild routine synchronously rather than calling it via an
790 * asynchronous work thread.
792 * This can only be called from code that is not holding
793 * cgroup_mutex (not nested in a cgroup_lock() call.)
795 void rebuild_sched_domains(void)
797 do_rebuild_sched_domains(NULL);
801 * cpuset_test_cpumask - test a task's cpus_allowed versus its cpuset's
803 * @scan: struct cgroup_scanner contained in its struct cpuset_hotplug_scanner
805 * Call with cgroup_mutex held. May take callback_mutex during call.
806 * Called for each task in a cgroup by cgroup_scan_tasks().
807 * Return nonzero if this tasks's cpus_allowed mask should be changed (in other
808 * words, if its mask is not equal to its cpuset's mask).
810 static int cpuset_test_cpumask(struct task_struct *tsk,
811 struct cgroup_scanner *scan)
813 return !cpumask_equal(&tsk->cpus_allowed,
814 (cgroup_cs(scan->cg))->cpus_allowed);
818 * cpuset_change_cpumask - make a task's cpus_allowed the same as its cpuset's
820 * @scan: struct cgroup_scanner containing the cgroup of the task
822 * Called by cgroup_scan_tasks() for each task in a cgroup whose
823 * cpus_allowed mask needs to be changed.
825 * We don't need to re-check for the cgroup/cpuset membership, since we're
826 * holding cgroup_lock() at this point.
828 static void cpuset_change_cpumask(struct task_struct *tsk,
829 struct cgroup_scanner *scan)
831 set_cpus_allowed_ptr(tsk, ((cgroup_cs(scan->cg))->cpus_allowed));
835 * update_tasks_cpumask - Update the cpumasks of tasks in the cpuset.
836 * @cs: the cpuset in which each task's cpus_allowed mask needs to be changed
837 * @heap: if NULL, defer allocating heap memory to cgroup_scan_tasks()
839 * Called with cgroup_mutex held
841 * The cgroup_scan_tasks() function will scan all the tasks in a cgroup,
842 * calling callback functions for each.
844 * No return value. It's guaranteed that cgroup_scan_tasks() always returns 0
847 static void update_tasks_cpumask(struct cpuset *cs, struct ptr_heap *heap)
849 struct cgroup_scanner scan;
851 scan.cg = cs->css.cgroup;
852 scan.test_task = cpuset_test_cpumask;
853 scan.process_task = cpuset_change_cpumask;
855 cgroup_scan_tasks(&scan);
859 * update_cpumask - update the cpus_allowed mask of a cpuset and all tasks in it
860 * @cs: the cpuset to consider
861 * @buf: buffer of cpu numbers written to this cpuset
863 static int update_cpumask(struct cpuset *cs, struct cpuset *trialcs,
866 struct ptr_heap heap;
868 int is_load_balanced;
870 /* top_cpuset.cpus_allowed tracks cpu_online_map; it's read-only */
871 if (cs == &top_cpuset)
875 * An empty cpus_allowed is ok only if the cpuset has no tasks.
876 * Since cpulist_parse() fails on an empty mask, we special case
877 * that parsing. The validate_change() call ensures that cpusets
878 * with tasks have cpus.
881 cpumask_clear(trialcs->cpus_allowed);
883 retval = cpulist_parse(buf, trialcs->cpus_allowed);
887 if (!cpumask_subset(trialcs->cpus_allowed, cpu_active_mask))
890 retval = validate_change(cs, trialcs);
894 /* Nothing to do if the cpus didn't change */
895 if (cpumask_equal(cs->cpus_allowed, trialcs->cpus_allowed))
898 retval = heap_init(&heap, PAGE_SIZE, GFP_KERNEL, NULL);
902 is_load_balanced = is_sched_load_balance(trialcs);
904 mutex_lock(&callback_mutex);
905 cpumask_copy(cs->cpus_allowed, trialcs->cpus_allowed);
906 mutex_unlock(&callback_mutex);
909 * Scan tasks in the cpuset, and update the cpumasks of any
910 * that need an update.
912 update_tasks_cpumask(cs, &heap);
916 if (is_load_balanced)
917 async_rebuild_sched_domains();
924 * Migrate memory region from one set of nodes to another.
926 * Temporarilly set tasks mems_allowed to target nodes of migration,
927 * so that the migration code can allocate pages on these nodes.
929 * Call holding cgroup_mutex, so current's cpuset won't change
930 * during this call, as manage_mutex holds off any cpuset_attach()
931 * calls. Therefore we don't need to take task_lock around the
932 * call to guarantee_online_mems(), as we know no one is changing
935 * While the mm_struct we are migrating is typically from some
936 * other task, the task_struct mems_allowed that we are hacking
937 * is for our current task, which must allocate new pages for that
938 * migrating memory region.
941 static void cpuset_migrate_mm(struct mm_struct *mm, const nodemask_t *from,
942 const nodemask_t *to)
944 struct task_struct *tsk = current;
946 tsk->mems_allowed = *to;
948 do_migrate_pages(mm, from, to, MPOL_MF_MOVE_ALL);
950 guarantee_online_mems(task_cs(tsk),&tsk->mems_allowed);
954 * cpuset_change_task_nodemask - change task's mems_allowed and mempolicy
955 * @tsk: the task to change
956 * @newmems: new nodes that the task will be set
958 * In order to avoid seeing no nodes if the old and new nodes are disjoint,
959 * we structure updates as setting all new allowed nodes, then clearing newly
962 static void cpuset_change_task_nodemask(struct task_struct *tsk,
968 * Allow tasks that have access to memory reserves because they have
969 * been OOM killed to get memory anywhere.
971 if (unlikely(test_thread_flag(TIF_MEMDIE)))
973 if (current->flags & PF_EXITING) /* Let dying task have memory */
978 * Determine if a loop is necessary if another thread is doing
979 * get_mems_allowed(). If at least one node remains unchanged and
980 * tsk does not have a mempolicy, then an empty nodemask will not be
981 * possible when mems_allowed is larger than a word.
983 need_loop = task_has_mempolicy(tsk) ||
984 !nodes_intersects(*newmems, tsk->mems_allowed);
988 write_seqcount_begin(&tsk->mems_allowed_seq);
991 nodes_or(tsk->mems_allowed, tsk->mems_allowed, *newmems);
992 mpol_rebind_task(tsk, newmems, MPOL_REBIND_STEP1);
994 mpol_rebind_task(tsk, newmems, MPOL_REBIND_STEP2);
995 tsk->mems_allowed = *newmems;
998 write_seqcount_end(&tsk->mems_allowed_seq);
1006 * Update task's mems_allowed and rebind its mempolicy and vmas' mempolicy
1007 * of it to cpuset's new mems_allowed, and migrate pages to new nodes if
1008 * memory_migrate flag is set. Called with cgroup_mutex held.
1010 static void cpuset_change_nodemask(struct task_struct *p,
1011 struct cgroup_scanner *scan)
1013 struct mm_struct *mm;
1016 const nodemask_t *oldmem = scan->data;
1017 static nodemask_t newmems; /* protected by cgroup_mutex */
1019 cs = cgroup_cs(scan->cg);
1020 guarantee_online_mems(cs, &newmems);
1022 cpuset_change_task_nodemask(p, &newmems);
1024 mm = get_task_mm(p);
1028 migrate = is_memory_migrate(cs);
1030 mpol_rebind_mm(mm, &cs->mems_allowed);
1032 cpuset_migrate_mm(mm, oldmem, &cs->mems_allowed);
1036 static void *cpuset_being_rebound;
1039 * update_tasks_nodemask - Update the nodemasks of tasks in the cpuset.
1040 * @cs: the cpuset in which each task's mems_allowed mask needs to be changed
1041 * @oldmem: old mems_allowed of cpuset cs
1042 * @heap: if NULL, defer allocating heap memory to cgroup_scan_tasks()
1044 * Called with cgroup_mutex held
1045 * No return value. It's guaranteed that cgroup_scan_tasks() always returns 0
1048 static void update_tasks_nodemask(struct cpuset *cs, const nodemask_t *oldmem,
1049 struct ptr_heap *heap)
1051 struct cgroup_scanner scan;
1053 cpuset_being_rebound = cs; /* causes mpol_dup() rebind */
1055 scan.cg = cs->css.cgroup;
1056 scan.test_task = NULL;
1057 scan.process_task = cpuset_change_nodemask;
1059 scan.data = (nodemask_t *)oldmem;
1062 * The mpol_rebind_mm() call takes mmap_sem, which we couldn't
1063 * take while holding tasklist_lock. Forks can happen - the
1064 * mpol_dup() cpuset_being_rebound check will catch such forks,
1065 * and rebind their vma mempolicies too. Because we still hold
1066 * the global cgroup_mutex, we know that no other rebind effort
1067 * will be contending for the global variable cpuset_being_rebound.
1068 * It's ok if we rebind the same mm twice; mpol_rebind_mm()
1069 * is idempotent. Also migrate pages in each mm to new nodes.
1071 cgroup_scan_tasks(&scan);
1073 /* We're done rebinding vmas to this cpuset's new mems_allowed. */
1074 cpuset_being_rebound = NULL;
1078 * Handle user request to change the 'mems' memory placement
1079 * of a cpuset. Needs to validate the request, update the
1080 * cpusets mems_allowed, and for each task in the cpuset,
1081 * update mems_allowed and rebind task's mempolicy and any vma
1082 * mempolicies and if the cpuset is marked 'memory_migrate',
1083 * migrate the tasks pages to the new memory.
1085 * Call with cgroup_mutex held. May take callback_mutex during call.
1086 * Will take tasklist_lock, scan tasklist for tasks in cpuset cs,
1087 * lock each such tasks mm->mmap_sem, scan its vma's and rebind
1088 * their mempolicies to the cpusets new mems_allowed.
1090 static int update_nodemask(struct cpuset *cs, struct cpuset *trialcs,
1093 NODEMASK_ALLOC(nodemask_t, oldmem, GFP_KERNEL);
1095 struct ptr_heap heap;
1101 * top_cpuset.mems_allowed tracks node_stats[N_HIGH_MEMORY];
1104 if (cs == &top_cpuset) {
1110 * An empty mems_allowed is ok iff there are no tasks in the cpuset.
1111 * Since nodelist_parse() fails on an empty mask, we special case
1112 * that parsing. The validate_change() call ensures that cpusets
1113 * with tasks have memory.
1116 nodes_clear(trialcs->mems_allowed);
1118 retval = nodelist_parse(buf, trialcs->mems_allowed);
1122 if (!nodes_subset(trialcs->mems_allowed,
1123 node_states[N_HIGH_MEMORY])) {
1128 *oldmem = cs->mems_allowed;
1129 if (nodes_equal(*oldmem, trialcs->mems_allowed)) {
1130 retval = 0; /* Too easy - nothing to do */
1133 retval = validate_change(cs, trialcs);
1137 retval = heap_init(&heap, PAGE_SIZE, GFP_KERNEL, NULL);
1141 mutex_lock(&callback_mutex);
1142 cs->mems_allowed = trialcs->mems_allowed;
1143 mutex_unlock(&callback_mutex);
1145 update_tasks_nodemask(cs, oldmem, &heap);
1149 NODEMASK_FREE(oldmem);
1153 int current_cpuset_is_being_rebound(void)
1158 ret = task_cs(current) == cpuset_being_rebound;
1164 static int update_relax_domain_level(struct cpuset *cs, s64 val)
1167 if (val < -1 || val >= sched_domain_level_max)
1171 if (val != cs->relax_domain_level) {
1172 cs->relax_domain_level = val;
1173 if (!cpumask_empty(cs->cpus_allowed) &&
1174 is_sched_load_balance(cs))
1175 async_rebuild_sched_domains();
1182 * cpuset_change_flag - make a task's spread flags the same as its cpuset's
1183 * @tsk: task to be updated
1184 * @scan: struct cgroup_scanner containing the cgroup of the task
1186 * Called by cgroup_scan_tasks() for each task in a cgroup.
1188 * We don't need to re-check for the cgroup/cpuset membership, since we're
1189 * holding cgroup_lock() at this point.
1191 static void cpuset_change_flag(struct task_struct *tsk,
1192 struct cgroup_scanner *scan)
1194 cpuset_update_task_spread_flag(cgroup_cs(scan->cg), tsk);
1198 * update_tasks_flags - update the spread flags of tasks in the cpuset.
1199 * @cs: the cpuset in which each task's spread flags needs to be changed
1200 * @heap: if NULL, defer allocating heap memory to cgroup_scan_tasks()
1202 * Called with cgroup_mutex held
1204 * The cgroup_scan_tasks() function will scan all the tasks in a cgroup,
1205 * calling callback functions for each.
1207 * No return value. It's guaranteed that cgroup_scan_tasks() always returns 0
1210 static void update_tasks_flags(struct cpuset *cs, struct ptr_heap *heap)
1212 struct cgroup_scanner scan;
1214 scan.cg = cs->css.cgroup;
1215 scan.test_task = NULL;
1216 scan.process_task = cpuset_change_flag;
1218 cgroup_scan_tasks(&scan);
1222 * update_flag - read a 0 or a 1 in a file and update associated flag
1223 * bit: the bit to update (see cpuset_flagbits_t)
1224 * cs: the cpuset to update
1225 * turning_on: whether the flag is being set or cleared
1227 * Call with cgroup_mutex held.
1230 static int update_flag(cpuset_flagbits_t bit, struct cpuset *cs,
1233 struct cpuset *trialcs;
1234 int balance_flag_changed;
1235 int spread_flag_changed;
1236 struct ptr_heap heap;
1239 trialcs = alloc_trial_cpuset(cs);
1244 set_bit(bit, &trialcs->flags);
1246 clear_bit(bit, &trialcs->flags);
1248 err = validate_change(cs, trialcs);
1252 err = heap_init(&heap, PAGE_SIZE, GFP_KERNEL, NULL);
1256 balance_flag_changed = (is_sched_load_balance(cs) !=
1257 is_sched_load_balance(trialcs));
1259 spread_flag_changed = ((is_spread_slab(cs) != is_spread_slab(trialcs))
1260 || (is_spread_page(cs) != is_spread_page(trialcs)));
1262 mutex_lock(&callback_mutex);
1263 cs->flags = trialcs->flags;
1264 mutex_unlock(&callback_mutex);
1266 if (!cpumask_empty(trialcs->cpus_allowed) && balance_flag_changed)
1267 async_rebuild_sched_domains();
1269 if (spread_flag_changed)
1270 update_tasks_flags(cs, &heap);
1273 free_trial_cpuset(trialcs);
1278 * Frequency meter - How fast is some event occurring?
1280 * These routines manage a digitally filtered, constant time based,
1281 * event frequency meter. There are four routines:
1282 * fmeter_init() - initialize a frequency meter.
1283 * fmeter_markevent() - called each time the event happens.
1284 * fmeter_getrate() - returns the recent rate of such events.
1285 * fmeter_update() - internal routine used to update fmeter.
1287 * A common data structure is passed to each of these routines,
1288 * which is used to keep track of the state required to manage the
1289 * frequency meter and its digital filter.
1291 * The filter works on the number of events marked per unit time.
1292 * The filter is single-pole low-pass recursive (IIR). The time unit
1293 * is 1 second. Arithmetic is done using 32-bit integers scaled to
1294 * simulate 3 decimal digits of precision (multiplied by 1000).
1296 * With an FM_COEF of 933, and a time base of 1 second, the filter
1297 * has a half-life of 10 seconds, meaning that if the events quit
1298 * happening, then the rate returned from the fmeter_getrate()
1299 * will be cut in half each 10 seconds, until it converges to zero.
1301 * It is not worth doing a real infinitely recursive filter. If more
1302 * than FM_MAXTICKS ticks have elapsed since the last filter event,
1303 * just compute FM_MAXTICKS ticks worth, by which point the level
1306 * Limit the count of unprocessed events to FM_MAXCNT, so as to avoid
1307 * arithmetic overflow in the fmeter_update() routine.
1309 * Given the simple 32 bit integer arithmetic used, this meter works
1310 * best for reporting rates between one per millisecond (msec) and
1311 * one per 32 (approx) seconds. At constant rates faster than one
1312 * per msec it maxes out at values just under 1,000,000. At constant
1313 * rates between one per msec, and one per second it will stabilize
1314 * to a value N*1000, where N is the rate of events per second.
1315 * At constant rates between one per second and one per 32 seconds,
1316 * it will be choppy, moving up on the seconds that have an event,
1317 * and then decaying until the next event. At rates slower than
1318 * about one in 32 seconds, it decays all the way back to zero between
1322 #define FM_COEF 933 /* coefficient for half-life of 10 secs */
1323 #define FM_MAXTICKS ((time_t)99) /* useless computing more ticks than this */
1324 #define FM_MAXCNT 1000000 /* limit cnt to avoid overflow */
1325 #define FM_SCALE 1000 /* faux fixed point scale */
1327 /* Initialize a frequency meter */
1328 static void fmeter_init(struct fmeter *fmp)
1333 spin_lock_init(&fmp->lock);
1336 /* Internal meter update - process cnt events and update value */
1337 static void fmeter_update(struct fmeter *fmp)
1339 time_t now = get_seconds();
1340 time_t ticks = now - fmp->time;
1345 ticks = min(FM_MAXTICKS, ticks);
1347 fmp->val = (FM_COEF * fmp->val) / FM_SCALE;
1350 fmp->val += ((FM_SCALE - FM_COEF) * fmp->cnt) / FM_SCALE;
1354 /* Process any previous ticks, then bump cnt by one (times scale). */
1355 static void fmeter_markevent(struct fmeter *fmp)
1357 spin_lock(&fmp->lock);
1359 fmp->cnt = min(FM_MAXCNT, fmp->cnt + FM_SCALE);
1360 spin_unlock(&fmp->lock);
1363 /* Process any previous ticks, then return current value. */
1364 static int fmeter_getrate(struct fmeter *fmp)
1368 spin_lock(&fmp->lock);
1371 spin_unlock(&fmp->lock);
1375 /* Called by cgroups to determine if a cpuset is usable; cgroup_mutex held */
1376 static int cpuset_can_attach(struct cgroup_subsys *ss, struct cgroup *cont,
1377 struct task_struct *tsk)
1379 struct cpuset *cs = cgroup_cs(cont);
1381 if (cpumask_empty(cs->cpus_allowed) || nodes_empty(cs->mems_allowed))
1385 * Kthreads bound to specific cpus cannot be moved to a new cpuset; we
1386 * cannot change their cpu affinity and isolating such threads by their
1387 * set of allowed nodes is unnecessary. Thus, cpusets are not
1388 * applicable for such threads. This prevents checking for success of
1389 * set_cpus_allowed_ptr() on all attached tasks before cpus_allowed may
1392 if (tsk->flags & PF_THREAD_BOUND)
1398 static int cpuset_can_attach_task(struct cgroup *cgrp, struct task_struct *task)
1400 return security_task_setscheduler(task);
1404 * Protected by cgroup_lock. The nodemasks must be stored globally because
1405 * dynamically allocating them is not allowed in pre_attach, and they must
1406 * persist among pre_attach, attach_task, and attach.
1408 static cpumask_var_t cpus_attach;
1409 static nodemask_t cpuset_attach_nodemask_from;
1410 static nodemask_t cpuset_attach_nodemask_to;
1412 /* Set-up work for before attaching each task. */
1413 static void cpuset_pre_attach(struct cgroup *cont)
1415 struct cpuset *cs = cgroup_cs(cont);
1417 if (cs == &top_cpuset)
1418 cpumask_copy(cpus_attach, cpu_possible_mask);
1420 guarantee_online_cpus(cs, cpus_attach);
1422 guarantee_online_mems(cs, &cpuset_attach_nodemask_to);
1425 /* Per-thread attachment work. */
1426 static void cpuset_attach_task(struct cgroup *cont, struct task_struct *tsk)
1429 struct cpuset *cs = cgroup_cs(cont);
1432 * can_attach beforehand should guarantee that this doesn't fail.
1433 * TODO: have a better way to handle failure here
1435 err = set_cpus_allowed_ptr(tsk, cpus_attach);
1438 cpuset_change_task_nodemask(tsk, &cpuset_attach_nodemask_to);
1439 cpuset_update_task_spread_flag(cs, tsk);
1442 static void cpuset_attach(struct cgroup_subsys *ss, struct cgroup *cont,
1443 struct cgroup *oldcont, struct task_struct *tsk)
1445 struct mm_struct *mm;
1446 struct cpuset *cs = cgroup_cs(cont);
1447 struct cpuset *oldcs = cgroup_cs(oldcont);
1450 * Change mm, possibly for multiple threads in a threadgroup. This is
1451 * expensive and may sleep.
1453 cpuset_attach_nodemask_from = oldcs->mems_allowed;
1454 cpuset_attach_nodemask_to = cs->mems_allowed;
1455 mm = get_task_mm(tsk);
1457 mpol_rebind_mm(mm, &cpuset_attach_nodemask_to);
1458 if (is_memory_migrate(cs))
1459 cpuset_migrate_mm(mm, &cpuset_attach_nodemask_from,
1460 &cpuset_attach_nodemask_to);
1465 /* The various types of files and directories in a cpuset file system */
1468 FILE_MEMORY_MIGRATE,
1474 FILE_SCHED_LOAD_BALANCE,
1475 FILE_SCHED_RELAX_DOMAIN_LEVEL,
1476 FILE_MEMORY_PRESSURE_ENABLED,
1477 FILE_MEMORY_PRESSURE,
1480 } cpuset_filetype_t;
1482 static int cpuset_write_u64(struct cgroup *cgrp, struct cftype *cft, u64 val)
1485 struct cpuset *cs = cgroup_cs(cgrp);
1486 cpuset_filetype_t type = cft->private;
1488 if (!cgroup_lock_live_group(cgrp))
1492 case FILE_CPU_EXCLUSIVE:
1493 retval = update_flag(CS_CPU_EXCLUSIVE, cs, val);
1495 case FILE_MEM_EXCLUSIVE:
1496 retval = update_flag(CS_MEM_EXCLUSIVE, cs, val);
1498 case FILE_MEM_HARDWALL:
1499 retval = update_flag(CS_MEM_HARDWALL, cs, val);
1501 case FILE_SCHED_LOAD_BALANCE:
1502 retval = update_flag(CS_SCHED_LOAD_BALANCE, cs, val);
1504 case FILE_MEMORY_MIGRATE:
1505 retval = update_flag(CS_MEMORY_MIGRATE, cs, val);
1507 case FILE_MEMORY_PRESSURE_ENABLED:
1508 cpuset_memory_pressure_enabled = !!val;
1510 case FILE_MEMORY_PRESSURE:
1513 case FILE_SPREAD_PAGE:
1514 retval = update_flag(CS_SPREAD_PAGE, cs, val);
1516 case FILE_SPREAD_SLAB:
1517 retval = update_flag(CS_SPREAD_SLAB, cs, val);
1527 static int cpuset_write_s64(struct cgroup *cgrp, struct cftype *cft, s64 val)
1530 struct cpuset *cs = cgroup_cs(cgrp);
1531 cpuset_filetype_t type = cft->private;
1533 if (!cgroup_lock_live_group(cgrp))
1537 case FILE_SCHED_RELAX_DOMAIN_LEVEL:
1538 retval = update_relax_domain_level(cs, val);
1549 * Common handling for a write to a "cpus" or "mems" file.
1551 static int cpuset_write_resmask(struct cgroup *cgrp, struct cftype *cft,
1555 struct cpuset *cs = cgroup_cs(cgrp);
1556 struct cpuset *trialcs;
1558 if (!cgroup_lock_live_group(cgrp))
1561 trialcs = alloc_trial_cpuset(cs);
1567 switch (cft->private) {
1569 retval = update_cpumask(cs, trialcs, buf);
1572 retval = update_nodemask(cs, trialcs, buf);
1579 free_trial_cpuset(trialcs);
1586 * These ascii lists should be read in a single call, by using a user
1587 * buffer large enough to hold the entire map. If read in smaller
1588 * chunks, there is no guarantee of atomicity. Since the display format
1589 * used, list of ranges of sequential numbers, is variable length,
1590 * and since these maps can change value dynamically, one could read
1591 * gibberish by doing partial reads while a list was changing.
1592 * A single large read to a buffer that crosses a page boundary is
1593 * ok, because the result being copied to user land is not recomputed
1594 * across a page fault.
1597 static size_t cpuset_sprintf_cpulist(char *page, struct cpuset *cs)
1601 mutex_lock(&callback_mutex);
1602 count = cpulist_scnprintf(page, PAGE_SIZE, cs->cpus_allowed);
1603 mutex_unlock(&callback_mutex);
1608 static size_t cpuset_sprintf_memlist(char *page, struct cpuset *cs)
1612 mutex_lock(&callback_mutex);
1613 count = nodelist_scnprintf(page, PAGE_SIZE, cs->mems_allowed);
1614 mutex_unlock(&callback_mutex);
1619 static ssize_t cpuset_common_file_read(struct cgroup *cont,
1623 size_t nbytes, loff_t *ppos)
1625 struct cpuset *cs = cgroup_cs(cont);
1626 cpuset_filetype_t type = cft->private;
1631 if (!(page = (char *)__get_free_page(GFP_TEMPORARY)))
1638 s += cpuset_sprintf_cpulist(s, cs);
1641 s += cpuset_sprintf_memlist(s, cs);
1649 retval = simple_read_from_buffer(buf, nbytes, ppos, page, s - page);
1651 free_page((unsigned long)page);
1655 static u64 cpuset_read_u64(struct cgroup *cont, struct cftype *cft)
1657 struct cpuset *cs = cgroup_cs(cont);
1658 cpuset_filetype_t type = cft->private;
1660 case FILE_CPU_EXCLUSIVE:
1661 return is_cpu_exclusive(cs);
1662 case FILE_MEM_EXCLUSIVE:
1663 return is_mem_exclusive(cs);
1664 case FILE_MEM_HARDWALL:
1665 return is_mem_hardwall(cs);
1666 case FILE_SCHED_LOAD_BALANCE:
1667 return is_sched_load_balance(cs);
1668 case FILE_MEMORY_MIGRATE:
1669 return is_memory_migrate(cs);
1670 case FILE_MEMORY_PRESSURE_ENABLED:
1671 return cpuset_memory_pressure_enabled;
1672 case FILE_MEMORY_PRESSURE:
1673 return fmeter_getrate(&cs->fmeter);
1674 case FILE_SPREAD_PAGE:
1675 return is_spread_page(cs);
1676 case FILE_SPREAD_SLAB:
1677 return is_spread_slab(cs);
1682 /* Unreachable but makes gcc happy */
1686 static s64 cpuset_read_s64(struct cgroup *cont, struct cftype *cft)
1688 struct cpuset *cs = cgroup_cs(cont);
1689 cpuset_filetype_t type = cft->private;
1691 case FILE_SCHED_RELAX_DOMAIN_LEVEL:
1692 return cs->relax_domain_level;
1697 /* Unrechable but makes gcc happy */
1703 * for the common functions, 'private' gives the type of file
1706 static struct cftype files[] = {
1709 .read = cpuset_common_file_read,
1710 .write_string = cpuset_write_resmask,
1711 .max_write_len = (100U + 6 * NR_CPUS),
1712 .private = FILE_CPULIST,
1717 .read = cpuset_common_file_read,
1718 .write_string = cpuset_write_resmask,
1719 .max_write_len = (100U + 6 * MAX_NUMNODES),
1720 .private = FILE_MEMLIST,
1724 .name = "cpu_exclusive",
1725 .read_u64 = cpuset_read_u64,
1726 .write_u64 = cpuset_write_u64,
1727 .private = FILE_CPU_EXCLUSIVE,
1731 .name = "mem_exclusive",
1732 .read_u64 = cpuset_read_u64,
1733 .write_u64 = cpuset_write_u64,
1734 .private = FILE_MEM_EXCLUSIVE,
1738 .name = "mem_hardwall",
1739 .read_u64 = cpuset_read_u64,
1740 .write_u64 = cpuset_write_u64,
1741 .private = FILE_MEM_HARDWALL,
1745 .name = "sched_load_balance",
1746 .read_u64 = cpuset_read_u64,
1747 .write_u64 = cpuset_write_u64,
1748 .private = FILE_SCHED_LOAD_BALANCE,
1752 .name = "sched_relax_domain_level",
1753 .read_s64 = cpuset_read_s64,
1754 .write_s64 = cpuset_write_s64,
1755 .private = FILE_SCHED_RELAX_DOMAIN_LEVEL,
1759 .name = "memory_migrate",
1760 .read_u64 = cpuset_read_u64,
1761 .write_u64 = cpuset_write_u64,
1762 .private = FILE_MEMORY_MIGRATE,
1766 .name = "memory_pressure",
1767 .read_u64 = cpuset_read_u64,
1768 .write_u64 = cpuset_write_u64,
1769 .private = FILE_MEMORY_PRESSURE,
1774 .name = "memory_spread_page",
1775 .read_u64 = cpuset_read_u64,
1776 .write_u64 = cpuset_write_u64,
1777 .private = FILE_SPREAD_PAGE,
1781 .name = "memory_spread_slab",
1782 .read_u64 = cpuset_read_u64,
1783 .write_u64 = cpuset_write_u64,
1784 .private = FILE_SPREAD_SLAB,
1788 static struct cftype cft_memory_pressure_enabled = {
1789 .name = "memory_pressure_enabled",
1790 .read_u64 = cpuset_read_u64,
1791 .write_u64 = cpuset_write_u64,
1792 .private = FILE_MEMORY_PRESSURE_ENABLED,
1795 static int cpuset_populate(struct cgroup_subsys *ss, struct cgroup *cont)
1799 err = cgroup_add_files(cont, ss, files, ARRAY_SIZE(files));
1802 /* memory_pressure_enabled is in root cpuset only */
1804 err = cgroup_add_file(cont, ss,
1805 &cft_memory_pressure_enabled);
1810 * post_clone() is called during cgroup_create() when the
1811 * clone_children mount argument was specified. The cgroup
1812 * can not yet have any tasks.
1814 * Currently we refuse to set up the cgroup - thereby
1815 * refusing the task to be entered, and as a result refusing
1816 * the sys_unshare() or clone() which initiated it - if any
1817 * sibling cpusets have exclusive cpus or mem.
1819 * If this becomes a problem for some users who wish to
1820 * allow that scenario, then cpuset_post_clone() could be
1821 * changed to grant parent->cpus_allowed-sibling_cpus_exclusive
1822 * (and likewise for mems) to the new cgroup. Called with cgroup_mutex
1825 static void cpuset_post_clone(struct cgroup_subsys *ss,
1826 struct cgroup *cgroup)
1828 struct cgroup *parent, *child;
1829 struct cpuset *cs, *parent_cs;
1831 parent = cgroup->parent;
1832 list_for_each_entry(child, &parent->children, sibling) {
1833 cs = cgroup_cs(child);
1834 if (is_mem_exclusive(cs) || is_cpu_exclusive(cs))
1837 cs = cgroup_cs(cgroup);
1838 parent_cs = cgroup_cs(parent);
1840 mutex_lock(&callback_mutex);
1841 cs->mems_allowed = parent_cs->mems_allowed;
1842 cpumask_copy(cs->cpus_allowed, parent_cs->cpus_allowed);
1843 mutex_unlock(&callback_mutex);
1848 * cpuset_create - create a cpuset
1849 * ss: cpuset cgroup subsystem
1850 * cont: control group that the new cpuset will be part of
1853 static struct cgroup_subsys_state *cpuset_create(
1854 struct cgroup_subsys *ss,
1855 struct cgroup *cont)
1858 struct cpuset *parent;
1860 if (!cont->parent) {
1861 return &top_cpuset.css;
1863 parent = cgroup_cs(cont->parent);
1864 cs = kmalloc(sizeof(*cs), GFP_KERNEL);
1866 return ERR_PTR(-ENOMEM);
1867 if (!alloc_cpumask_var(&cs->cpus_allowed, GFP_KERNEL)) {
1869 return ERR_PTR(-ENOMEM);
1873 if (is_spread_page(parent))
1874 set_bit(CS_SPREAD_PAGE, &cs->flags);
1875 if (is_spread_slab(parent))
1876 set_bit(CS_SPREAD_SLAB, &cs->flags);
1877 set_bit(CS_SCHED_LOAD_BALANCE, &cs->flags);
1878 cpumask_clear(cs->cpus_allowed);
1879 nodes_clear(cs->mems_allowed);
1880 fmeter_init(&cs->fmeter);
1881 cs->relax_domain_level = -1;
1883 cs->parent = parent;
1884 number_of_cpusets++;
1889 * If the cpuset being removed has its flag 'sched_load_balance'
1890 * enabled, then simulate turning sched_load_balance off, which
1891 * will call async_rebuild_sched_domains().
1894 static void cpuset_destroy(struct cgroup_subsys *ss, struct cgroup *cont)
1896 struct cpuset *cs = cgroup_cs(cont);
1898 if (is_sched_load_balance(cs))
1899 update_flag(CS_SCHED_LOAD_BALANCE, cs, 0);
1901 number_of_cpusets--;
1902 free_cpumask_var(cs->cpus_allowed);
1906 struct cgroup_subsys cpuset_subsys = {
1908 .create = cpuset_create,
1909 .destroy = cpuset_destroy,
1910 .can_attach = cpuset_can_attach,
1911 .can_attach_task = cpuset_can_attach_task,
1912 .pre_attach = cpuset_pre_attach,
1913 .attach_task = cpuset_attach_task,
1914 .attach = cpuset_attach,
1915 .populate = cpuset_populate,
1916 .post_clone = cpuset_post_clone,
1917 .subsys_id = cpuset_subsys_id,
1922 * cpuset_init - initialize cpusets at system boot
1924 * Description: Initialize top_cpuset and the cpuset internal file system,
1927 int __init cpuset_init(void)
1931 if (!alloc_cpumask_var(&top_cpuset.cpus_allowed, GFP_KERNEL))
1934 cpumask_setall(top_cpuset.cpus_allowed);
1935 nodes_setall(top_cpuset.mems_allowed);
1937 fmeter_init(&top_cpuset.fmeter);
1938 set_bit(CS_SCHED_LOAD_BALANCE, &top_cpuset.flags);
1939 top_cpuset.relax_domain_level = -1;
1941 err = register_filesystem(&cpuset_fs_type);
1945 if (!alloc_cpumask_var(&cpus_attach, GFP_KERNEL))
1948 number_of_cpusets = 1;
1953 * cpuset_do_move_task - move a given task to another cpuset
1954 * @tsk: pointer to task_struct the task to move
1955 * @scan: struct cgroup_scanner contained in its struct cpuset_hotplug_scanner
1957 * Called by cgroup_scan_tasks() for each task in a cgroup.
1958 * Return nonzero to stop the walk through the tasks.
1960 static void cpuset_do_move_task(struct task_struct *tsk,
1961 struct cgroup_scanner *scan)
1963 struct cgroup *new_cgroup = scan->data;
1965 cgroup_attach_task(new_cgroup, tsk);
1969 * move_member_tasks_to_cpuset - move tasks from one cpuset to another
1970 * @from: cpuset in which the tasks currently reside
1971 * @to: cpuset to which the tasks will be moved
1973 * Called with cgroup_mutex held
1974 * callback_mutex must not be held, as cpuset_attach() will take it.
1976 * The cgroup_scan_tasks() function will scan all the tasks in a cgroup,
1977 * calling callback functions for each.
1979 static void move_member_tasks_to_cpuset(struct cpuset *from, struct cpuset *to)
1981 struct cgroup_scanner scan;
1983 scan.cg = from->css.cgroup;
1984 scan.test_task = NULL; /* select all tasks in cgroup */
1985 scan.process_task = cpuset_do_move_task;
1987 scan.data = to->css.cgroup;
1989 if (cgroup_scan_tasks(&scan))
1990 printk(KERN_ERR "move_member_tasks_to_cpuset: "
1991 "cgroup_scan_tasks failed\n");
1995 * If CPU and/or memory hotplug handlers, below, unplug any CPUs
1996 * or memory nodes, we need to walk over the cpuset hierarchy,
1997 * removing that CPU or node from all cpusets. If this removes the
1998 * last CPU or node from a cpuset, then move the tasks in the empty
1999 * cpuset to its next-highest non-empty parent.
2001 * Called with cgroup_mutex held
2002 * callback_mutex must not be held, as cpuset_attach() will take it.
2004 static void remove_tasks_in_empty_cpuset(struct cpuset *cs)
2006 struct cpuset *parent;
2009 * The cgroup's css_sets list is in use if there are tasks
2010 * in the cpuset; the list is empty if there are none;
2011 * the cs->css.refcnt seems always 0.
2013 if (list_empty(&cs->css.cgroup->css_sets))
2017 * Find its next-highest non-empty parent, (top cpuset
2018 * has online cpus, so can't be empty).
2020 parent = cs->parent;
2021 while (cpumask_empty(parent->cpus_allowed) ||
2022 nodes_empty(parent->mems_allowed))
2023 parent = parent->parent;
2025 move_member_tasks_to_cpuset(cs, parent);
2029 * Walk the specified cpuset subtree and look for empty cpusets.
2030 * The tasks of such cpuset must be moved to a parent cpuset.
2032 * Called with cgroup_mutex held. We take callback_mutex to modify
2033 * cpus_allowed and mems_allowed.
2035 * This walk processes the tree from top to bottom, completing one layer
2036 * before dropping down to the next. It always processes a node before
2037 * any of its children.
2039 * For now, since we lack memory hot unplug, we'll never see a cpuset
2040 * that has tasks along with an empty 'mems'. But if we did see such
2041 * a cpuset, we'd handle it just like we do if its 'cpus' was empty.
2043 static void scan_for_empty_cpusets(struct cpuset *root)
2046 struct cpuset *cp; /* scans cpusets being updated */
2047 struct cpuset *child; /* scans child cpusets of cp */
2048 struct cgroup *cont;
2049 static nodemask_t oldmems; /* protected by cgroup_mutex */
2051 list_add_tail((struct list_head *)&root->stack_list, &queue);
2053 while (!list_empty(&queue)) {
2054 cp = list_first_entry(&queue, struct cpuset, stack_list);
2055 list_del(queue.next);
2056 list_for_each_entry(cont, &cp->css.cgroup->children, sibling) {
2057 child = cgroup_cs(cont);
2058 list_add_tail(&child->stack_list, &queue);
2061 /* Continue past cpusets with all cpus, mems online */
2062 if (cpumask_subset(cp->cpus_allowed, cpu_active_mask) &&
2063 nodes_subset(cp->mems_allowed, node_states[N_HIGH_MEMORY]))
2066 oldmems = cp->mems_allowed;
2068 /* Remove offline cpus and mems from this cpuset. */
2069 mutex_lock(&callback_mutex);
2070 cpumask_and(cp->cpus_allowed, cp->cpus_allowed,
2072 nodes_and(cp->mems_allowed, cp->mems_allowed,
2073 node_states[N_HIGH_MEMORY]);
2074 mutex_unlock(&callback_mutex);
2076 /* Move tasks from the empty cpuset to a parent */
2077 if (cpumask_empty(cp->cpus_allowed) ||
2078 nodes_empty(cp->mems_allowed))
2079 remove_tasks_in_empty_cpuset(cp);
2081 update_tasks_cpumask(cp, NULL);
2082 update_tasks_nodemask(cp, &oldmems, NULL);
2088 * The top_cpuset tracks what CPUs and Memory Nodes are online,
2089 * period. This is necessary in order to make cpusets transparent
2090 * (of no affect) on systems that are actively using CPU hotplug
2091 * but making no active use of cpusets.
2093 * The only exception to this is suspend/resume, where we don't
2094 * modify cpusets at all.
2096 * This routine ensures that top_cpuset.cpus_allowed tracks
2097 * cpu_active_mask on each CPU hotplug (cpuhp) event.
2099 * Called within get_online_cpus(). Needs to call cgroup_lock()
2100 * before calling generate_sched_domains().
2102 void cpuset_update_active_cpus(void)
2104 struct sched_domain_attr *attr;
2105 cpumask_var_t *doms;
2109 mutex_lock(&callback_mutex);
2110 cpumask_copy(top_cpuset.cpus_allowed, cpu_active_mask);
2111 mutex_unlock(&callback_mutex);
2112 scan_for_empty_cpusets(&top_cpuset);
2113 ndoms = generate_sched_domains(&doms, &attr);
2116 /* Have scheduler rebuild the domains */
2117 partition_sched_domains(ndoms, doms, attr);
2120 #ifdef CONFIG_MEMORY_HOTPLUG
2122 * Keep top_cpuset.mems_allowed tracking node_states[N_HIGH_MEMORY].
2123 * Call this routine anytime after node_states[N_HIGH_MEMORY] changes.
2124 * See also the previous routine cpuset_track_online_cpus().
2126 static int cpuset_track_online_nodes(struct notifier_block *self,
2127 unsigned long action, void *arg)
2129 static nodemask_t oldmems; /* protected by cgroup_mutex */
2134 oldmems = top_cpuset.mems_allowed;
2135 mutex_lock(&callback_mutex);
2136 top_cpuset.mems_allowed = node_states[N_HIGH_MEMORY];
2137 mutex_unlock(&callback_mutex);
2138 update_tasks_nodemask(&top_cpuset, &oldmems, NULL);
2142 * needn't update top_cpuset.mems_allowed explicitly because
2143 * scan_for_empty_cpusets() will update it.
2145 scan_for_empty_cpusets(&top_cpuset);
2157 * cpuset_init_smp - initialize cpus_allowed
2159 * Description: Finish top cpuset after cpu, node maps are initialized
2162 void __init cpuset_init_smp(void)
2164 cpumask_copy(top_cpuset.cpus_allowed, cpu_active_mask);
2165 top_cpuset.mems_allowed = node_states[N_HIGH_MEMORY];
2167 hotplug_memory_notifier(cpuset_track_online_nodes, 10);
2169 cpuset_wq = create_singlethread_workqueue("cpuset");
2174 * cpuset_cpus_allowed - return cpus_allowed mask from a tasks cpuset.
2175 * @tsk: pointer to task_struct from which to obtain cpuset->cpus_allowed.
2176 * @pmask: pointer to struct cpumask variable to receive cpus_allowed set.
2178 * Description: Returns the cpumask_var_t cpus_allowed of the cpuset
2179 * attached to the specified @tsk. Guaranteed to return some non-empty
2180 * subset of cpu_online_map, even if this means going outside the
2184 void cpuset_cpus_allowed(struct task_struct *tsk, struct cpumask *pmask)
2186 mutex_lock(&callback_mutex);
2188 guarantee_online_cpus(task_cs(tsk), pmask);
2190 mutex_unlock(&callback_mutex);
2193 int cpuset_cpus_allowed_fallback(struct task_struct *tsk)
2195 const struct cpuset *cs;
2201 do_set_cpus_allowed(tsk, cs->cpus_allowed);
2205 * We own tsk->cpus_allowed, nobody can change it under us.
2207 * But we used cs && cs->cpus_allowed lockless and thus can
2208 * race with cgroup_attach_task() or update_cpumask() and get
2209 * the wrong tsk->cpus_allowed. However, both cases imply the
2210 * subsequent cpuset_change_cpumask()->set_cpus_allowed_ptr()
2211 * which takes task_rq_lock().
2213 * If we are called after it dropped the lock we must see all
2214 * changes in tsk_cs()->cpus_allowed. Otherwise we can temporary
2215 * set any mask even if it is not right from task_cs() pov,
2216 * the pending set_cpus_allowed_ptr() will fix things.
2219 cpu = cpumask_any_and(&tsk->cpus_allowed, cpu_active_mask);
2220 if (cpu >= nr_cpu_ids) {
2222 * Either tsk->cpus_allowed is wrong (see above) or it
2223 * is actually empty. The latter case is only possible
2224 * if we are racing with remove_tasks_in_empty_cpuset().
2225 * Like above we can temporary set any mask and rely on
2226 * set_cpus_allowed_ptr() as synchronization point.
2228 do_set_cpus_allowed(tsk, cpu_possible_mask);
2229 cpu = cpumask_any(cpu_active_mask);
2235 void cpuset_init_current_mems_allowed(void)
2237 nodes_setall(current->mems_allowed);
2241 * cpuset_mems_allowed - return mems_allowed mask from a tasks cpuset.
2242 * @tsk: pointer to task_struct from which to obtain cpuset->mems_allowed.
2244 * Description: Returns the nodemask_t mems_allowed of the cpuset
2245 * attached to the specified @tsk. Guaranteed to return some non-empty
2246 * subset of node_states[N_HIGH_MEMORY], even if this means going outside the
2250 nodemask_t cpuset_mems_allowed(struct task_struct *tsk)
2254 mutex_lock(&callback_mutex);
2256 guarantee_online_mems(task_cs(tsk), &mask);
2258 mutex_unlock(&callback_mutex);
2264 * cpuset_nodemask_valid_mems_allowed - check nodemask vs. curremt mems_allowed
2265 * @nodemask: the nodemask to be checked
2267 * Are any of the nodes in the nodemask allowed in current->mems_allowed?
2269 int cpuset_nodemask_valid_mems_allowed(nodemask_t *nodemask)
2271 return nodes_intersects(*nodemask, current->mems_allowed);
2275 * nearest_hardwall_ancestor() - Returns the nearest mem_exclusive or
2276 * mem_hardwall ancestor to the specified cpuset. Call holding
2277 * callback_mutex. If no ancestor is mem_exclusive or mem_hardwall
2278 * (an unusual configuration), then returns the root cpuset.
2280 static const struct cpuset *nearest_hardwall_ancestor(const struct cpuset *cs)
2282 while (!(is_mem_exclusive(cs) || is_mem_hardwall(cs)) && cs->parent)
2288 * cpuset_node_allowed_softwall - Can we allocate on a memory node?
2289 * @node: is this an allowed node?
2290 * @gfp_mask: memory allocation flags
2292 * If we're in interrupt, yes, we can always allocate. If __GFP_THISNODE is
2293 * set, yes, we can always allocate. If node is in our task's mems_allowed,
2294 * yes. If it's not a __GFP_HARDWALL request and this node is in the nearest
2295 * hardwalled cpuset ancestor to this task's cpuset, yes. If the task has been
2296 * OOM killed and has access to memory reserves as specified by the TIF_MEMDIE
2300 * If __GFP_HARDWALL is set, cpuset_node_allowed_softwall() reduces to
2301 * cpuset_node_allowed_hardwall(). Otherwise, cpuset_node_allowed_softwall()
2302 * might sleep, and might allow a node from an enclosing cpuset.
2304 * cpuset_node_allowed_hardwall() only handles the simpler case of hardwall
2305 * cpusets, and never sleeps.
2307 * The __GFP_THISNODE placement logic is really handled elsewhere,
2308 * by forcibly using a zonelist starting at a specified node, and by
2309 * (in get_page_from_freelist()) refusing to consider the zones for
2310 * any node on the zonelist except the first. By the time any such
2311 * calls get to this routine, we should just shut up and say 'yes'.
2313 * GFP_USER allocations are marked with the __GFP_HARDWALL bit,
2314 * and do not allow allocations outside the current tasks cpuset
2315 * unless the task has been OOM killed as is marked TIF_MEMDIE.
2316 * GFP_KERNEL allocations are not so marked, so can escape to the
2317 * nearest enclosing hardwalled ancestor cpuset.
2319 * Scanning up parent cpusets requires callback_mutex. The
2320 * __alloc_pages() routine only calls here with __GFP_HARDWALL bit
2321 * _not_ set if it's a GFP_KERNEL allocation, and all nodes in the
2322 * current tasks mems_allowed came up empty on the first pass over
2323 * the zonelist. So only GFP_KERNEL allocations, if all nodes in the
2324 * cpuset are short of memory, might require taking the callback_mutex
2327 * The first call here from mm/page_alloc:get_page_from_freelist()
2328 * has __GFP_HARDWALL set in gfp_mask, enforcing hardwall cpusets,
2329 * so no allocation on a node outside the cpuset is allowed (unless
2330 * in interrupt, of course).
2332 * The second pass through get_page_from_freelist() doesn't even call
2333 * here for GFP_ATOMIC calls. For those calls, the __alloc_pages()
2334 * variable 'wait' is not set, and the bit ALLOC_CPUSET is not set
2335 * in alloc_flags. That logic and the checks below have the combined
2337 * in_interrupt - any node ok (current task context irrelevant)
2338 * GFP_ATOMIC - any node ok
2339 * TIF_MEMDIE - any node ok
2340 * GFP_KERNEL - any node in enclosing hardwalled cpuset ok
2341 * GFP_USER - only nodes in current tasks mems allowed ok.
2344 * Don't call cpuset_node_allowed_softwall if you can't sleep, unless you
2345 * pass in the __GFP_HARDWALL flag set in gfp_flag, which disables
2346 * the code that might scan up ancestor cpusets and sleep.
2348 int __cpuset_node_allowed_softwall(int node, gfp_t gfp_mask)
2350 const struct cpuset *cs; /* current cpuset ancestors */
2351 int allowed; /* is allocation in zone z allowed? */
2353 if (in_interrupt() || (gfp_mask & __GFP_THISNODE))
2355 might_sleep_if(!(gfp_mask & __GFP_HARDWALL));
2356 if (node_isset(node, current->mems_allowed))
2359 * Allow tasks that have access to memory reserves because they have
2360 * been OOM killed to get memory anywhere.
2362 if (unlikely(test_thread_flag(TIF_MEMDIE)))
2364 if (gfp_mask & __GFP_HARDWALL) /* If hardwall request, stop here */
2367 if (current->flags & PF_EXITING) /* Let dying task have memory */
2370 /* Not hardwall and node outside mems_allowed: scan up cpusets */
2371 mutex_lock(&callback_mutex);
2374 cs = nearest_hardwall_ancestor(task_cs(current));
2375 allowed = node_isset(node, cs->mems_allowed);
2376 task_unlock(current);
2378 mutex_unlock(&callback_mutex);
2383 * cpuset_node_allowed_hardwall - Can we allocate on a memory node?
2384 * @node: is this an allowed node?
2385 * @gfp_mask: memory allocation flags
2387 * If we're in interrupt, yes, we can always allocate. If __GFP_THISNODE is
2388 * set, yes, we can always allocate. If node is in our task's mems_allowed,
2389 * yes. If the task has been OOM killed and has access to memory reserves as
2390 * specified by the TIF_MEMDIE flag, yes.
2393 * The __GFP_THISNODE placement logic is really handled elsewhere,
2394 * by forcibly using a zonelist starting at a specified node, and by
2395 * (in get_page_from_freelist()) refusing to consider the zones for
2396 * any node on the zonelist except the first. By the time any such
2397 * calls get to this routine, we should just shut up and say 'yes'.
2399 * Unlike the cpuset_node_allowed_softwall() variant, above,
2400 * this variant requires that the node be in the current task's
2401 * mems_allowed or that we're in interrupt. It does not scan up the
2402 * cpuset hierarchy for the nearest enclosing mem_exclusive cpuset.
2405 int __cpuset_node_allowed_hardwall(int node, gfp_t gfp_mask)
2407 if (in_interrupt() || (gfp_mask & __GFP_THISNODE))
2409 if (node_isset(node, current->mems_allowed))
2412 * Allow tasks that have access to memory reserves because they have
2413 * been OOM killed to get memory anywhere.
2415 if (unlikely(test_thread_flag(TIF_MEMDIE)))
2421 * cpuset_unlock - release lock on cpuset changes
2423 * Undo the lock taken in a previous cpuset_lock() call.
2426 void cpuset_unlock(void)
2428 mutex_unlock(&callback_mutex);
2432 * cpuset_mem_spread_node() - On which node to begin search for a file page
2433 * cpuset_slab_spread_node() - On which node to begin search for a slab page
2435 * If a task is marked PF_SPREAD_PAGE or PF_SPREAD_SLAB (as for
2436 * tasks in a cpuset with is_spread_page or is_spread_slab set),
2437 * and if the memory allocation used cpuset_mem_spread_node()
2438 * to determine on which node to start looking, as it will for
2439 * certain page cache or slab cache pages such as used for file
2440 * system buffers and inode caches, then instead of starting on the
2441 * local node to look for a free page, rather spread the starting
2442 * node around the tasks mems_allowed nodes.
2444 * We don't have to worry about the returned node being offline
2445 * because "it can't happen", and even if it did, it would be ok.
2447 * The routines calling guarantee_online_mems() are careful to
2448 * only set nodes in task->mems_allowed that are online. So it
2449 * should not be possible for the following code to return an
2450 * offline node. But if it did, that would be ok, as this routine
2451 * is not returning the node where the allocation must be, only
2452 * the node where the search should start. The zonelist passed to
2453 * __alloc_pages() will include all nodes. If the slab allocator
2454 * is passed an offline node, it will fall back to the local node.
2455 * See kmem_cache_alloc_node().
2458 static int cpuset_spread_node(int *rotor)
2462 node = next_node(*rotor, current->mems_allowed);
2463 if (node == MAX_NUMNODES)
2464 node = first_node(current->mems_allowed);
2469 int cpuset_mem_spread_node(void)
2471 if (current->cpuset_mem_spread_rotor == NUMA_NO_NODE)
2472 current->cpuset_mem_spread_rotor =
2473 node_random(¤t->mems_allowed);
2475 return cpuset_spread_node(¤t->cpuset_mem_spread_rotor);
2478 int cpuset_slab_spread_node(void)
2480 if (current->cpuset_slab_spread_rotor == NUMA_NO_NODE)
2481 current->cpuset_slab_spread_rotor =
2482 node_random(¤t->mems_allowed);
2484 return cpuset_spread_node(¤t->cpuset_slab_spread_rotor);
2487 EXPORT_SYMBOL_GPL(cpuset_mem_spread_node);
2490 * cpuset_mems_allowed_intersects - Does @tsk1's mems_allowed intersect @tsk2's?
2491 * @tsk1: pointer to task_struct of some task.
2492 * @tsk2: pointer to task_struct of some other task.
2494 * Description: Return true if @tsk1's mems_allowed intersects the
2495 * mems_allowed of @tsk2. Used by the OOM killer to determine if
2496 * one of the task's memory usage might impact the memory available
2500 int cpuset_mems_allowed_intersects(const struct task_struct *tsk1,
2501 const struct task_struct *tsk2)
2503 return nodes_intersects(tsk1->mems_allowed, tsk2->mems_allowed);
2507 * cpuset_print_task_mems_allowed - prints task's cpuset and mems_allowed
2508 * @task: pointer to task_struct of some task.
2510 * Description: Prints @task's name, cpuset name, and cached copy of its
2511 * mems_allowed to the kernel log. Must hold task_lock(task) to allow
2512 * dereferencing task_cs(task).
2514 void cpuset_print_task_mems_allowed(struct task_struct *tsk)
2516 struct dentry *dentry;
2518 dentry = task_cs(tsk)->css.cgroup->dentry;
2519 spin_lock(&cpuset_buffer_lock);
2522 strcpy(cpuset_name, "/");
2524 spin_lock(&dentry->d_lock);
2525 strlcpy(cpuset_name, (const char *)dentry->d_name.name,
2527 spin_unlock(&dentry->d_lock);
2530 nodelist_scnprintf(cpuset_nodelist, CPUSET_NODELIST_LEN,
2532 printk(KERN_INFO "%s cpuset=%s mems_allowed=%s\n",
2533 tsk->comm, cpuset_name, cpuset_nodelist);
2534 spin_unlock(&cpuset_buffer_lock);
2538 * Collection of memory_pressure is suppressed unless
2539 * this flag is enabled by writing "1" to the special
2540 * cpuset file 'memory_pressure_enabled' in the root cpuset.
2543 int cpuset_memory_pressure_enabled __read_mostly;
2546 * cpuset_memory_pressure_bump - keep stats of per-cpuset reclaims.
2548 * Keep a running average of the rate of synchronous (direct)
2549 * page reclaim efforts initiated by tasks in each cpuset.
2551 * This represents the rate at which some task in the cpuset
2552 * ran low on memory on all nodes it was allowed to use, and
2553 * had to enter the kernels page reclaim code in an effort to
2554 * create more free memory by tossing clean pages or swapping
2555 * or writing dirty pages.
2557 * Display to user space in the per-cpuset read-only file
2558 * "memory_pressure". Value displayed is an integer
2559 * representing the recent rate of entry into the synchronous
2560 * (direct) page reclaim by any task attached to the cpuset.
2563 void __cpuset_memory_pressure_bump(void)
2566 fmeter_markevent(&task_cs(current)->fmeter);
2567 task_unlock(current);
2570 #ifdef CONFIG_PROC_PID_CPUSET
2572 * proc_cpuset_show()
2573 * - Print tasks cpuset path into seq_file.
2574 * - Used for /proc/<pid>/cpuset.
2575 * - No need to task_lock(tsk) on this tsk->cpuset reference, as it
2576 * doesn't really matter if tsk->cpuset changes after we read it,
2577 * and we take cgroup_mutex, keeping cpuset_attach() from changing it
2580 static int proc_cpuset_show(struct seq_file *m, void *unused_v)
2583 struct task_struct *tsk;
2585 struct cgroup_subsys_state *css;
2589 buf = kmalloc(PAGE_SIZE, GFP_KERNEL);
2595 tsk = get_pid_task(pid, PIDTYPE_PID);
2601 css = task_subsys_state(tsk, cpuset_subsys_id);
2602 retval = cgroup_path(css->cgroup, buf, PAGE_SIZE);
2609 put_task_struct(tsk);
2616 static int cpuset_open(struct inode *inode, struct file *file)
2618 struct pid *pid = PROC_I(inode)->pid;
2619 return single_open(file, proc_cpuset_show, pid);
2622 const struct file_operations proc_cpuset_operations = {
2623 .open = cpuset_open,
2625 .llseek = seq_lseek,
2626 .release = single_release,
2628 #endif /* CONFIG_PROC_PID_CPUSET */
2630 /* Display task mems_allowed in /proc/<pid>/status file. */
2631 void cpuset_task_status_allowed(struct seq_file *m, struct task_struct *task)
2633 seq_printf(m, "Mems_allowed:\t");
2634 seq_nodemask(m, &task->mems_allowed);
2635 seq_printf(m, "\n");
2636 seq_printf(m, "Mems_allowed_list:\t");
2637 seq_nodemask_list(m, &task->mems_allowed);
2638 seq_printf(m, "\n");