4 * Processor and Memory placement constraints for sets of tasks.
6 * Copyright (C) 2003 BULL SA.
7 * Copyright (C) 2004-2007 Silicon Graphics, Inc.
8 * Copyright (C) 2006 Google, Inc
10 * Portions derived from Patrick Mochel's sysfs code.
11 * sysfs is Copyright (c) 2001-3 Patrick Mochel
13 * 2003-10-10 Written by Simon Derr.
14 * 2003-10-22 Updates by Stephen Hemminger.
15 * 2004 May-July Rework by Paul Jackson.
16 * 2006 Rework by Paul Menage to use generic cgroups
17 * 2008 Rework of the scheduler domains and CPU hotplug handling
20 * This file is subject to the terms and conditions of the GNU General Public
21 * License. See the file COPYING in the main directory of the Linux
22 * distribution for more details.
25 #include <linux/cpu.h>
26 #include <linux/cpumask.h>
27 #include <linux/cpuset.h>
28 #include <linux/err.h>
29 #include <linux/errno.h>
30 #include <linux/file.h>
32 #include <linux/init.h>
33 #include <linux/interrupt.h>
34 #include <linux/kernel.h>
35 #include <linux/kmod.h>
36 #include <linux/list.h>
37 #include <linux/mempolicy.h>
39 #include <linux/memory.h>
40 #include <linux/module.h>
41 #include <linux/mount.h>
42 #include <linux/namei.h>
43 #include <linux/pagemap.h>
44 #include <linux/proc_fs.h>
45 #include <linux/rcupdate.h>
46 #include <linux/sched.h>
47 #include <linux/seq_file.h>
48 #include <linux/security.h>
49 #include <linux/slab.h>
50 #include <linux/spinlock.h>
51 #include <linux/stat.h>
52 #include <linux/string.h>
53 #include <linux/time.h>
54 #include <linux/backing-dev.h>
55 #include <linux/sort.h>
57 #include <asm/uaccess.h>
58 #include <asm/atomic.h>
59 #include <linux/mutex.h>
60 #include <linux/workqueue.h>
61 #include <linux/cgroup.h>
64 * Workqueue for cpuset related tasks.
66 * Using kevent workqueue may cause deadlock when memory_migrate
67 * is set. So we create a separate workqueue thread for cpuset.
69 static struct workqueue_struct *cpuset_wq;
72 * Tracks how many cpusets are currently defined in system.
73 * When there is only one cpuset (the root cpuset) we can
74 * short circuit some hooks.
76 int number_of_cpusets __read_mostly;
78 /* Forward declare cgroup structures */
79 struct cgroup_subsys cpuset_subsys;
82 /* See "Frequency meter" comments, below. */
85 int cnt; /* unprocessed events count */
86 int val; /* most recent output value */
87 time_t time; /* clock (secs) when val computed */
88 spinlock_t lock; /* guards read or write of above */
92 struct cgroup_subsys_state css;
94 unsigned long flags; /* "unsigned long" so bitops work */
95 cpumask_var_t cpus_allowed; /* CPUs allowed to tasks in cpuset */
96 nodemask_t mems_allowed; /* Memory Nodes allowed to tasks */
98 struct cpuset *parent; /* my parent */
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 heirarchy */
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),
126 /* bits in struct cpuset flags field */
132 CS_SCHED_LOAD_BALANCE,
137 /* convenient tests for these bits */
138 static inline int is_cpu_exclusive(const struct cpuset *cs)
140 return test_bit(CS_CPU_EXCLUSIVE, &cs->flags);
143 static inline int is_mem_exclusive(const struct cpuset *cs)
145 return test_bit(CS_MEM_EXCLUSIVE, &cs->flags);
148 static inline int is_mem_hardwall(const struct cpuset *cs)
150 return test_bit(CS_MEM_HARDWALL, &cs->flags);
153 static inline int is_sched_load_balance(const struct cpuset *cs)
155 return test_bit(CS_SCHED_LOAD_BALANCE, &cs->flags);
158 static inline int is_memory_migrate(const struct cpuset *cs)
160 return test_bit(CS_MEMORY_MIGRATE, &cs->flags);
163 static inline int is_spread_page(const struct cpuset *cs)
165 return test_bit(CS_SPREAD_PAGE, &cs->flags);
168 static inline int is_spread_slab(const struct cpuset *cs)
170 return test_bit(CS_SPREAD_SLAB, &cs->flags);
173 static struct cpuset top_cpuset = {
174 .flags = ((1 << CS_CPU_EXCLUSIVE) | (1 << CS_MEM_EXCLUSIVE)),
178 * There are two global mutexes guarding cpuset structures. The first
179 * is the main control groups cgroup_mutex, accessed via
180 * cgroup_lock()/cgroup_unlock(). The second is the cpuset-specific
181 * callback_mutex, below. They can nest. It is ok to first take
182 * cgroup_mutex, then nest callback_mutex. We also require taking
183 * task_lock() when dereferencing a task's cpuset pointer. See "The
184 * task_lock() exception", at the end of this comment.
186 * A task must hold both mutexes to modify cpusets. If a task
187 * holds cgroup_mutex, then it blocks others wanting that mutex,
188 * ensuring that it is the only task able to also acquire callback_mutex
189 * and be able to modify cpusets. It can perform various checks on
190 * the cpuset structure first, knowing nothing will change. It can
191 * also allocate memory while just holding cgroup_mutex. While it is
192 * performing these checks, various callback routines can briefly
193 * acquire callback_mutex to query cpusets. Once it is ready to make
194 * the changes, it takes callback_mutex, blocking everyone else.
196 * Calls to the kernel memory allocator can not be made while holding
197 * callback_mutex, as that would risk double tripping on callback_mutex
198 * from one of the callbacks into the cpuset code from within
201 * If a task is only holding callback_mutex, then it has read-only
204 * Now, the task_struct fields mems_allowed and mempolicy may be changed
205 * by other task, we use alloc_lock in the task_struct fields to protect
208 * The cpuset_common_file_read() handlers only hold callback_mutex across
209 * small pieces of code, such as when reading out possibly multi-word
210 * cpumasks and nodemasks.
212 * Accessing a task's cpuset should be done in accordance with the
213 * guidelines for accessing subsystem state in kernel/cgroup.c
216 static DEFINE_MUTEX(callback_mutex);
219 * cpuset_buffer_lock protects both the cpuset_name and cpuset_nodelist
220 * buffers. They are statically allocated to prevent using excess stack
221 * when calling cpuset_print_task_mems_allowed().
223 #define CPUSET_NAME_LEN (128)
224 #define CPUSET_NODELIST_LEN (256)
225 static char cpuset_name[CPUSET_NAME_LEN];
226 static char cpuset_nodelist[CPUSET_NODELIST_LEN];
227 static DEFINE_SPINLOCK(cpuset_buffer_lock);
230 * This is ugly, but preserves the userspace API for existing cpuset
231 * users. If someone tries to mount the "cpuset" filesystem, we
232 * silently switch it to mount "cgroup" instead
234 static int cpuset_get_sb(struct file_system_type *fs_type,
235 int flags, const char *unused_dev_name,
236 void *data, struct vfsmount *mnt)
238 struct file_system_type *cgroup_fs = get_fs_type("cgroup");
243 "release_agent=/sbin/cpuset_release_agent";
244 ret = cgroup_fs->get_sb(cgroup_fs, flags,
245 unused_dev_name, mountopts, mnt);
246 put_filesystem(cgroup_fs);
251 static struct file_system_type cpuset_fs_type = {
253 .get_sb = cpuset_get_sb,
257 * Return in pmask the portion of a cpusets's cpus_allowed that
258 * are online. If none are online, walk up the cpuset hierarchy
259 * until we find one that does have some online cpus. If we get
260 * all the way to the top and still haven't found any online cpus,
261 * return cpu_online_map. Or if passed a NULL cs from an exit'ing
262 * task, return cpu_online_map.
264 * One way or another, we guarantee to return some non-empty subset
267 * Call with callback_mutex held.
270 static void guarantee_online_cpus(const struct cpuset *cs,
271 struct cpumask *pmask)
273 while (cs && !cpumask_intersects(cs->cpus_allowed, cpu_online_mask))
276 cpumask_and(pmask, cs->cpus_allowed, cpu_online_mask);
278 cpumask_copy(pmask, cpu_online_mask);
279 BUG_ON(!cpumask_intersects(pmask, cpu_online_mask));
283 * Return in *pmask the portion of a cpusets's mems_allowed that
284 * are online, with memory. If none are online with memory, walk
285 * up the cpuset hierarchy until we find one that does have some
286 * online mems. If we get all the way to the top and still haven't
287 * found any online mems, return node_states[N_HIGH_MEMORY].
289 * One way or another, we guarantee to return some non-empty subset
290 * of node_states[N_HIGH_MEMORY].
292 * Call with callback_mutex held.
295 static void guarantee_online_mems(const struct cpuset *cs, nodemask_t *pmask)
297 while (cs && !nodes_intersects(cs->mems_allowed,
298 node_states[N_HIGH_MEMORY]))
301 nodes_and(*pmask, cs->mems_allowed,
302 node_states[N_HIGH_MEMORY]);
304 *pmask = node_states[N_HIGH_MEMORY];
305 BUG_ON(!nodes_intersects(*pmask, node_states[N_HIGH_MEMORY]));
309 * update task's spread flag if cpuset's page/slab spread flag is set
311 * Called with callback_mutex/cgroup_mutex held
313 static void cpuset_update_task_spread_flag(struct cpuset *cs,
314 struct task_struct *tsk)
316 if (is_spread_page(cs))
317 tsk->flags |= PF_SPREAD_PAGE;
319 tsk->flags &= ~PF_SPREAD_PAGE;
320 if (is_spread_slab(cs))
321 tsk->flags |= PF_SPREAD_SLAB;
323 tsk->flags &= ~PF_SPREAD_SLAB;
327 * is_cpuset_subset(p, q) - Is cpuset p a subset of cpuset q?
329 * One cpuset is a subset of another if all its allowed CPUs and
330 * Memory Nodes are a subset of the other, and its exclusive flags
331 * are only set if the other's are set. Call holding cgroup_mutex.
334 static int is_cpuset_subset(const struct cpuset *p, const struct cpuset *q)
336 return cpumask_subset(p->cpus_allowed, q->cpus_allowed) &&
337 nodes_subset(p->mems_allowed, q->mems_allowed) &&
338 is_cpu_exclusive(p) <= is_cpu_exclusive(q) &&
339 is_mem_exclusive(p) <= is_mem_exclusive(q);
343 * alloc_trial_cpuset - allocate a trial cpuset
344 * @cs: the cpuset that the trial cpuset duplicates
346 static struct cpuset *alloc_trial_cpuset(const struct cpuset *cs)
348 struct cpuset *trial;
350 trial = kmemdup(cs, sizeof(*cs), GFP_KERNEL);
354 if (!alloc_cpumask_var(&trial->cpus_allowed, GFP_KERNEL)) {
358 cpumask_copy(trial->cpus_allowed, cs->cpus_allowed);
364 * free_trial_cpuset - free the trial cpuset
365 * @trial: the trial cpuset to be freed
367 static void free_trial_cpuset(struct cpuset *trial)
369 free_cpumask_var(trial->cpus_allowed);
374 * validate_change() - Used to validate that any proposed cpuset change
375 * follows the structural rules for cpusets.
377 * If we replaced the flag and mask values of the current cpuset
378 * (cur) with those values in the trial cpuset (trial), would
379 * our various subset and exclusive rules still be valid? Presumes
382 * 'cur' is the address of an actual, in-use cpuset. Operations
383 * such as list traversal that depend on the actual address of the
384 * cpuset in the list must use cur below, not trial.
386 * 'trial' is the address of bulk structure copy of cur, with
387 * perhaps one or more of the fields cpus_allowed, mems_allowed,
388 * or flags changed to new, trial values.
390 * Return 0 if valid, -errno if not.
393 static int validate_change(const struct cpuset *cur, const struct cpuset *trial)
396 struct cpuset *c, *par;
398 /* Each of our child cpusets must be a subset of us */
399 list_for_each_entry(cont, &cur->css.cgroup->children, sibling) {
400 if (!is_cpuset_subset(cgroup_cs(cont), trial))
404 /* Remaining checks don't apply to root cpuset */
405 if (cur == &top_cpuset)
410 /* We must be a subset of our parent cpuset */
411 if (!is_cpuset_subset(trial, par))
415 * If either I or some sibling (!= me) is exclusive, we can't
418 list_for_each_entry(cont, &par->css.cgroup->children, sibling) {
420 if ((is_cpu_exclusive(trial) || is_cpu_exclusive(c)) &&
422 cpumask_intersects(trial->cpus_allowed, c->cpus_allowed))
424 if ((is_mem_exclusive(trial) || is_mem_exclusive(c)) &&
426 nodes_intersects(trial->mems_allowed, c->mems_allowed))
430 /* Cpusets with tasks can't have empty cpus_allowed or mems_allowed */
431 if (cgroup_task_count(cur->css.cgroup)) {
432 if (cpumask_empty(trial->cpus_allowed) ||
433 nodes_empty(trial->mems_allowed)) {
443 * Helper routine for generate_sched_domains().
444 * Do cpusets a, b have overlapping cpus_allowed masks?
446 static int cpusets_overlap(struct cpuset *a, struct cpuset *b)
448 return cpumask_intersects(a->cpus_allowed, b->cpus_allowed);
452 update_domain_attr(struct sched_domain_attr *dattr, struct cpuset *c)
454 if (dattr->relax_domain_level < c->relax_domain_level)
455 dattr->relax_domain_level = c->relax_domain_level;
460 update_domain_attr_tree(struct sched_domain_attr *dattr, struct cpuset *c)
464 list_add(&c->stack_list, &q);
465 while (!list_empty(&q)) {
468 struct cpuset *child;
470 cp = list_first_entry(&q, struct cpuset, stack_list);
473 if (cpumask_empty(cp->cpus_allowed))
476 if (is_sched_load_balance(cp))
477 update_domain_attr(dattr, cp);
479 list_for_each_entry(cont, &cp->css.cgroup->children, sibling) {
480 child = cgroup_cs(cont);
481 list_add_tail(&child->stack_list, &q);
487 * generate_sched_domains()
489 * This function builds a partial partition of the systems CPUs
490 * A 'partial partition' is a set of non-overlapping subsets whose
491 * union is a subset of that set.
492 * The output of this function needs to be passed to kernel/sched.c
493 * partition_sched_domains() routine, which will rebuild the scheduler's
494 * load balancing domains (sched domains) as specified by that partial
497 * See "What is sched_load_balance" in Documentation/cgroups/cpusets.txt
498 * for a background explanation of this.
500 * Does not return errors, on the theory that the callers of this
501 * routine would rather not worry about failures to rebuild sched
502 * domains when operating in the severe memory shortage situations
503 * that could cause allocation failures below.
505 * Must be called with cgroup_lock held.
507 * The three key local variables below are:
508 * q - a linked-list queue of cpuset pointers, used to implement a
509 * top-down scan of all cpusets. This scan loads a pointer
510 * to each cpuset marked is_sched_load_balance into the
511 * array 'csa'. For our purposes, rebuilding the schedulers
512 * sched domains, we can ignore !is_sched_load_balance cpusets.
513 * csa - (for CpuSet Array) Array of pointers to all the cpusets
514 * that need to be load balanced, for convenient iterative
515 * access by the subsequent code that finds the best partition,
516 * i.e the set of domains (subsets) of CPUs such that the
517 * cpus_allowed of every cpuset marked is_sched_load_balance
518 * is a subset of one of these domains, while there are as
519 * many such domains as possible, each as small as possible.
520 * doms - Conversion of 'csa' to an array of cpumasks, for passing to
521 * the kernel/sched.c routine partition_sched_domains() in a
522 * convenient format, that can be easily compared to the prior
523 * value to determine what partition elements (sched domains)
524 * were changed (added or removed.)
526 * Finding the best partition (set of domains):
527 * The triple nested loops below over i, j, k scan over the
528 * load balanced cpusets (using the array of cpuset pointers in
529 * csa[]) looking for pairs of cpusets that have overlapping
530 * cpus_allowed, but which don't have the same 'pn' partition
531 * number and gives them in the same partition number. It keeps
532 * looping on the 'restart' label until it can no longer find
535 * The union of the cpus_allowed masks from the set of
536 * all cpusets having the same 'pn' value then form the one
537 * element of the partition (one sched domain) to be passed to
538 * partition_sched_domains().
540 static int generate_sched_domains(cpumask_var_t **domains,
541 struct sched_domain_attr **attributes)
543 LIST_HEAD(q); /* queue of cpusets to be scanned */
544 struct cpuset *cp; /* scans q */
545 struct cpuset **csa; /* array of all cpuset ptrs */
546 int csn; /* how many cpuset ptrs in csa so far */
547 int i, j, k; /* indices for partition finding loops */
548 cpumask_var_t *doms; /* resulting partition; i.e. sched domains */
549 struct sched_domain_attr *dattr; /* attributes for custom domains */
550 int ndoms = 0; /* number of sched domains in result */
551 int nslot; /* next empty doms[] struct cpumask slot */
557 /* Special case for the 99% of systems with one, full, sched domain */
558 if (is_sched_load_balance(&top_cpuset)) {
560 doms = alloc_sched_domains(ndoms);
564 dattr = kmalloc(sizeof(struct sched_domain_attr), GFP_KERNEL);
566 *dattr = SD_ATTR_INIT;
567 update_domain_attr_tree(dattr, &top_cpuset);
569 cpumask_copy(doms[0], top_cpuset.cpus_allowed);
574 csa = kmalloc(number_of_cpusets * sizeof(cp), GFP_KERNEL);
579 list_add(&top_cpuset.stack_list, &q);
580 while (!list_empty(&q)) {
582 struct cpuset *child; /* scans child cpusets of cp */
584 cp = list_first_entry(&q, struct cpuset, stack_list);
587 if (cpumask_empty(cp->cpus_allowed))
591 * All child cpusets contain a subset of the parent's cpus, so
592 * just skip them, and then we call update_domain_attr_tree()
593 * to calc relax_domain_level of the corresponding sched
596 if (is_sched_load_balance(cp)) {
601 list_for_each_entry(cont, &cp->css.cgroup->children, sibling) {
602 child = cgroup_cs(cont);
603 list_add_tail(&child->stack_list, &q);
607 for (i = 0; i < csn; i++)
612 /* Find the best partition (set of sched domains) */
613 for (i = 0; i < csn; i++) {
614 struct cpuset *a = csa[i];
617 for (j = 0; j < csn; j++) {
618 struct cpuset *b = csa[j];
621 if (apn != bpn && cpusets_overlap(a, b)) {
622 for (k = 0; k < csn; k++) {
623 struct cpuset *c = csa[k];
628 ndoms--; /* one less element */
635 * Now we know how many domains to create.
636 * Convert <csn, csa> to <ndoms, doms> and populate cpu masks.
638 doms = alloc_sched_domains(ndoms);
643 * The rest of the code, including the scheduler, can deal with
644 * dattr==NULL case. No need to abort if alloc fails.
646 dattr = kmalloc(ndoms * sizeof(struct sched_domain_attr), GFP_KERNEL);
648 for (nslot = 0, i = 0; i < csn; i++) {
649 struct cpuset *a = csa[i];
654 /* Skip completed partitions */
660 if (nslot == ndoms) {
661 static int warnings = 10;
664 "rebuild_sched_domains confused:"
665 " nslot %d, ndoms %d, csn %d, i %d,"
667 nslot, ndoms, csn, i, apn);
675 *(dattr + nslot) = SD_ATTR_INIT;
676 for (j = i; j < csn; j++) {
677 struct cpuset *b = csa[j];
680 cpumask_or(dp, dp, b->cpus_allowed);
682 update_domain_attr_tree(dattr + nslot, b);
684 /* Done with this partition */
690 BUG_ON(nslot != ndoms);
696 * Fallback to the default domain if kmalloc() failed.
697 * See comments in partition_sched_domains().
708 * Rebuild scheduler domains.
710 * Call with neither cgroup_mutex held nor within get_online_cpus().
711 * Takes both cgroup_mutex and get_online_cpus().
713 * Cannot be directly called from cpuset code handling changes
714 * to the cpuset pseudo-filesystem, because it cannot be called
715 * from code that already holds cgroup_mutex.
717 static void do_rebuild_sched_domains(struct work_struct *unused)
719 struct sched_domain_attr *attr;
725 /* Generate domain masks and attrs */
727 ndoms = generate_sched_domains(&doms, &attr);
730 /* Have scheduler rebuild the domains */
731 partition_sched_domains(ndoms, doms, attr);
735 #else /* !CONFIG_SMP */
736 static void do_rebuild_sched_domains(struct work_struct *unused)
740 static int generate_sched_domains(cpumask_var_t **domains,
741 struct sched_domain_attr **attributes)
746 #endif /* CONFIG_SMP */
748 static DECLARE_WORK(rebuild_sched_domains_work, do_rebuild_sched_domains);
751 * Rebuild scheduler domains, asynchronously via workqueue.
753 * If the flag 'sched_load_balance' of any cpuset with non-empty
754 * 'cpus' changes, or if the 'cpus' allowed changes in any cpuset
755 * which has that flag enabled, or if any cpuset with a non-empty
756 * 'cpus' is removed, then call this routine to rebuild the
757 * scheduler's dynamic sched domains.
759 * The rebuild_sched_domains() and partition_sched_domains()
760 * routines must nest cgroup_lock() inside get_online_cpus(),
761 * but such cpuset changes as these must nest that locking the
762 * other way, holding cgroup_lock() for much of the code.
764 * So in order to avoid an ABBA deadlock, the cpuset code handling
765 * these user changes delegates the actual sched domain rebuilding
766 * to a separate workqueue thread, which ends up processing the
767 * above do_rebuild_sched_domains() function.
769 static void async_rebuild_sched_domains(void)
771 queue_work(cpuset_wq, &rebuild_sched_domains_work);
775 * Accomplishes the same scheduler domain rebuild as the above
776 * async_rebuild_sched_domains(), however it directly calls the
777 * rebuild routine synchronously rather than calling it via an
778 * asynchronous work thread.
780 * This can only be called from code that is not holding
781 * cgroup_mutex (not nested in a cgroup_lock() call.)
783 void rebuild_sched_domains(void)
785 do_rebuild_sched_domains(NULL);
789 * cpuset_test_cpumask - test a task's cpus_allowed versus its cpuset's
791 * @scan: struct cgroup_scanner contained in its struct cpuset_hotplug_scanner
793 * Call with cgroup_mutex held. May take callback_mutex during call.
794 * Called for each task in a cgroup by cgroup_scan_tasks().
795 * Return nonzero if this tasks's cpus_allowed mask should be changed (in other
796 * words, if its mask is not equal to its cpuset's mask).
798 static int cpuset_test_cpumask(struct task_struct *tsk,
799 struct cgroup_scanner *scan)
801 return !cpumask_equal(&tsk->cpus_allowed,
802 (cgroup_cs(scan->cg))->cpus_allowed);
806 * cpuset_change_cpumask - make a task's cpus_allowed the same as its cpuset's
808 * @scan: struct cgroup_scanner containing the cgroup of the task
810 * Called by cgroup_scan_tasks() for each task in a cgroup whose
811 * cpus_allowed mask needs to be changed.
813 * We don't need to re-check for the cgroup/cpuset membership, since we're
814 * holding cgroup_lock() at this point.
816 static void cpuset_change_cpumask(struct task_struct *tsk,
817 struct cgroup_scanner *scan)
819 set_cpus_allowed_ptr(tsk, ((cgroup_cs(scan->cg))->cpus_allowed));
823 * update_tasks_cpumask - Update the cpumasks of tasks in the cpuset.
824 * @cs: the cpuset in which each task's cpus_allowed mask needs to be changed
825 * @heap: if NULL, defer allocating heap memory to cgroup_scan_tasks()
827 * Called with cgroup_mutex held
829 * The cgroup_scan_tasks() function will scan all the tasks in a cgroup,
830 * calling callback functions for each.
832 * No return value. It's guaranteed that cgroup_scan_tasks() always returns 0
835 static void update_tasks_cpumask(struct cpuset *cs, struct ptr_heap *heap)
837 struct cgroup_scanner scan;
839 scan.cg = cs->css.cgroup;
840 scan.test_task = cpuset_test_cpumask;
841 scan.process_task = cpuset_change_cpumask;
843 cgroup_scan_tasks(&scan);
847 * update_cpumask - update the cpus_allowed mask of a cpuset and all tasks in it
848 * @cs: the cpuset to consider
849 * @buf: buffer of cpu numbers written to this cpuset
851 static int update_cpumask(struct cpuset *cs, struct cpuset *trialcs,
854 struct ptr_heap heap;
856 int is_load_balanced;
858 /* top_cpuset.cpus_allowed tracks cpu_online_map; it's read-only */
859 if (cs == &top_cpuset)
863 * An empty cpus_allowed is ok only if the cpuset has no tasks.
864 * Since cpulist_parse() fails on an empty mask, we special case
865 * that parsing. The validate_change() call ensures that cpusets
866 * with tasks have cpus.
869 cpumask_clear(trialcs->cpus_allowed);
871 retval = cpulist_parse(buf, trialcs->cpus_allowed);
875 if (!cpumask_subset(trialcs->cpus_allowed, cpu_active_mask))
878 retval = validate_change(cs, trialcs);
882 /* Nothing to do if the cpus didn't change */
883 if (cpumask_equal(cs->cpus_allowed, trialcs->cpus_allowed))
886 retval = heap_init(&heap, PAGE_SIZE, GFP_KERNEL, NULL);
890 is_load_balanced = is_sched_load_balance(trialcs);
892 mutex_lock(&callback_mutex);
893 cpumask_copy(cs->cpus_allowed, trialcs->cpus_allowed);
894 mutex_unlock(&callback_mutex);
897 * Scan tasks in the cpuset, and update the cpumasks of any
898 * that need an update.
900 update_tasks_cpumask(cs, &heap);
904 if (is_load_balanced)
905 async_rebuild_sched_domains();
912 * Migrate memory region from one set of nodes to another.
914 * Temporarilly set tasks mems_allowed to target nodes of migration,
915 * so that the migration code can allocate pages on these nodes.
917 * Call holding cgroup_mutex, so current's cpuset won't change
918 * during this call, as manage_mutex holds off any cpuset_attach()
919 * calls. Therefore we don't need to take task_lock around the
920 * call to guarantee_online_mems(), as we know no one is changing
923 * While the mm_struct we are migrating is typically from some
924 * other task, the task_struct mems_allowed that we are hacking
925 * is for our current task, which must allocate new pages for that
926 * migrating memory region.
929 static void cpuset_migrate_mm(struct mm_struct *mm, const nodemask_t *from,
930 const nodemask_t *to)
932 struct task_struct *tsk = current;
934 tsk->mems_allowed = *to;
936 do_migrate_pages(mm, from, to, MPOL_MF_MOVE_ALL);
938 guarantee_online_mems(task_cs(tsk),&tsk->mems_allowed);
942 * cpuset_change_task_nodemask - change task's mems_allowed and mempolicy
943 * @tsk: the task to change
944 * @newmems: new nodes that the task will be set
946 * In order to avoid seeing no nodes if the old and new nodes are disjoint,
947 * we structure updates as setting all new allowed nodes, then clearing newly
950 * Called with task's alloc_lock held
952 static void cpuset_change_task_nodemask(struct task_struct *tsk,
955 nodes_or(tsk->mems_allowed, tsk->mems_allowed, *newmems);
956 mpol_rebind_task(tsk, &tsk->mems_allowed);
957 mpol_rebind_task(tsk, newmems);
958 tsk->mems_allowed = *newmems;
962 * Update task's mems_allowed and rebind its mempolicy and vmas' mempolicy
963 * of it to cpuset's new mems_allowed, and migrate pages to new nodes if
964 * memory_migrate flag is set. Called with cgroup_mutex held.
966 static void cpuset_change_nodemask(struct task_struct *p,
967 struct cgroup_scanner *scan)
969 struct mm_struct *mm;
972 const nodemask_t *oldmem = scan->data;
973 NODEMASK_ALLOC(nodemask_t, newmems, GFP_KERNEL);
978 cs = cgroup_cs(scan->cg);
979 guarantee_online_mems(cs, newmems);
982 cpuset_change_task_nodemask(p, newmems);
985 NODEMASK_FREE(newmems);
991 migrate = is_memory_migrate(cs);
993 mpol_rebind_mm(mm, &cs->mems_allowed);
995 cpuset_migrate_mm(mm, oldmem, &cs->mems_allowed);
999 static void *cpuset_being_rebound;
1002 * update_tasks_nodemask - Update the nodemasks of tasks in the cpuset.
1003 * @cs: the cpuset in which each task's mems_allowed mask needs to be changed
1004 * @oldmem: old mems_allowed of cpuset cs
1005 * @heap: if NULL, defer allocating heap memory to cgroup_scan_tasks()
1007 * Called with cgroup_mutex held
1008 * No return value. It's guaranteed that cgroup_scan_tasks() always returns 0
1011 static void update_tasks_nodemask(struct cpuset *cs, const nodemask_t *oldmem,
1012 struct ptr_heap *heap)
1014 struct cgroup_scanner scan;
1016 cpuset_being_rebound = cs; /* causes mpol_dup() rebind */
1018 scan.cg = cs->css.cgroup;
1019 scan.test_task = NULL;
1020 scan.process_task = cpuset_change_nodemask;
1022 scan.data = (nodemask_t *)oldmem;
1025 * The mpol_rebind_mm() call takes mmap_sem, which we couldn't
1026 * take while holding tasklist_lock. Forks can happen - the
1027 * mpol_dup() cpuset_being_rebound check will catch such forks,
1028 * and rebind their vma mempolicies too. Because we still hold
1029 * the global cgroup_mutex, we know that no other rebind effort
1030 * will be contending for the global variable cpuset_being_rebound.
1031 * It's ok if we rebind the same mm twice; mpol_rebind_mm()
1032 * is idempotent. Also migrate pages in each mm to new nodes.
1034 cgroup_scan_tasks(&scan);
1036 /* We're done rebinding vmas to this cpuset's new mems_allowed. */
1037 cpuset_being_rebound = NULL;
1041 * Handle user request to change the 'mems' memory placement
1042 * of a cpuset. Needs to validate the request, update the
1043 * cpusets mems_allowed, and for each task in the cpuset,
1044 * update mems_allowed and rebind task's mempolicy and any vma
1045 * mempolicies and if the cpuset is marked 'memory_migrate',
1046 * migrate the tasks pages to the new memory.
1048 * Call with cgroup_mutex held. May take callback_mutex during call.
1049 * Will take tasklist_lock, scan tasklist for tasks in cpuset cs,
1050 * lock each such tasks mm->mmap_sem, scan its vma's and rebind
1051 * their mempolicies to the cpusets new mems_allowed.
1053 static int update_nodemask(struct cpuset *cs, struct cpuset *trialcs,
1056 NODEMASK_ALLOC(nodemask_t, oldmem, GFP_KERNEL);
1058 struct ptr_heap heap;
1064 * top_cpuset.mems_allowed tracks node_stats[N_HIGH_MEMORY];
1067 if (cs == &top_cpuset) {
1073 * An empty mems_allowed is ok iff there are no tasks in the cpuset.
1074 * Since nodelist_parse() fails on an empty mask, we special case
1075 * that parsing. The validate_change() call ensures that cpusets
1076 * with tasks have memory.
1079 nodes_clear(trialcs->mems_allowed);
1081 retval = nodelist_parse(buf, trialcs->mems_allowed);
1085 if (!nodes_subset(trialcs->mems_allowed,
1086 node_states[N_HIGH_MEMORY])) {
1091 *oldmem = cs->mems_allowed;
1092 if (nodes_equal(*oldmem, trialcs->mems_allowed)) {
1093 retval = 0; /* Too easy - nothing to do */
1096 retval = validate_change(cs, trialcs);
1100 retval = heap_init(&heap, PAGE_SIZE, GFP_KERNEL, NULL);
1104 mutex_lock(&callback_mutex);
1105 cs->mems_allowed = trialcs->mems_allowed;
1106 mutex_unlock(&callback_mutex);
1108 update_tasks_nodemask(cs, oldmem, &heap);
1112 NODEMASK_FREE(oldmem);
1116 int current_cpuset_is_being_rebound(void)
1118 return task_cs(current) == cpuset_being_rebound;
1121 static int update_relax_domain_level(struct cpuset *cs, s64 val)
1124 if (val < -1 || val >= SD_LV_MAX)
1128 if (val != cs->relax_domain_level) {
1129 cs->relax_domain_level = val;
1130 if (!cpumask_empty(cs->cpus_allowed) &&
1131 is_sched_load_balance(cs))
1132 async_rebuild_sched_domains();
1139 * cpuset_change_flag - make a task's spread flags the same as its cpuset's
1140 * @tsk: task to be updated
1141 * @scan: struct cgroup_scanner containing the cgroup of the task
1143 * Called by cgroup_scan_tasks() for each task in a cgroup.
1145 * We don't need to re-check for the cgroup/cpuset membership, since we're
1146 * holding cgroup_lock() at this point.
1148 static void cpuset_change_flag(struct task_struct *tsk,
1149 struct cgroup_scanner *scan)
1151 cpuset_update_task_spread_flag(cgroup_cs(scan->cg), tsk);
1155 * update_tasks_flags - update the spread flags of tasks in the cpuset.
1156 * @cs: the cpuset in which each task's spread flags needs to be changed
1157 * @heap: if NULL, defer allocating heap memory to cgroup_scan_tasks()
1159 * Called with cgroup_mutex held
1161 * The cgroup_scan_tasks() function will scan all the tasks in a cgroup,
1162 * calling callback functions for each.
1164 * No return value. It's guaranteed that cgroup_scan_tasks() always returns 0
1167 static void update_tasks_flags(struct cpuset *cs, struct ptr_heap *heap)
1169 struct cgroup_scanner scan;
1171 scan.cg = cs->css.cgroup;
1172 scan.test_task = NULL;
1173 scan.process_task = cpuset_change_flag;
1175 cgroup_scan_tasks(&scan);
1179 * update_flag - read a 0 or a 1 in a file and update associated flag
1180 * bit: the bit to update (see cpuset_flagbits_t)
1181 * cs: the cpuset to update
1182 * turning_on: whether the flag is being set or cleared
1184 * Call with cgroup_mutex held.
1187 static int update_flag(cpuset_flagbits_t bit, struct cpuset *cs,
1190 struct cpuset *trialcs;
1191 int balance_flag_changed;
1192 int spread_flag_changed;
1193 struct ptr_heap heap;
1196 trialcs = alloc_trial_cpuset(cs);
1201 set_bit(bit, &trialcs->flags);
1203 clear_bit(bit, &trialcs->flags);
1205 err = validate_change(cs, trialcs);
1209 err = heap_init(&heap, PAGE_SIZE, GFP_KERNEL, NULL);
1213 balance_flag_changed = (is_sched_load_balance(cs) !=
1214 is_sched_load_balance(trialcs));
1216 spread_flag_changed = ((is_spread_slab(cs) != is_spread_slab(trialcs))
1217 || (is_spread_page(cs) != is_spread_page(trialcs)));
1219 mutex_lock(&callback_mutex);
1220 cs->flags = trialcs->flags;
1221 mutex_unlock(&callback_mutex);
1223 if (!cpumask_empty(trialcs->cpus_allowed) && balance_flag_changed)
1224 async_rebuild_sched_domains();
1226 if (spread_flag_changed)
1227 update_tasks_flags(cs, &heap);
1230 free_trial_cpuset(trialcs);
1235 * Frequency meter - How fast is some event occurring?
1237 * These routines manage a digitally filtered, constant time based,
1238 * event frequency meter. There are four routines:
1239 * fmeter_init() - initialize a frequency meter.
1240 * fmeter_markevent() - called each time the event happens.
1241 * fmeter_getrate() - returns the recent rate of such events.
1242 * fmeter_update() - internal routine used to update fmeter.
1244 * A common data structure is passed to each of these routines,
1245 * which is used to keep track of the state required to manage the
1246 * frequency meter and its digital filter.
1248 * The filter works on the number of events marked per unit time.
1249 * The filter is single-pole low-pass recursive (IIR). The time unit
1250 * is 1 second. Arithmetic is done using 32-bit integers scaled to
1251 * simulate 3 decimal digits of precision (multiplied by 1000).
1253 * With an FM_COEF of 933, and a time base of 1 second, the filter
1254 * has a half-life of 10 seconds, meaning that if the events quit
1255 * happening, then the rate returned from the fmeter_getrate()
1256 * will be cut in half each 10 seconds, until it converges to zero.
1258 * It is not worth doing a real infinitely recursive filter. If more
1259 * than FM_MAXTICKS ticks have elapsed since the last filter event,
1260 * just compute FM_MAXTICKS ticks worth, by which point the level
1263 * Limit the count of unprocessed events to FM_MAXCNT, so as to avoid
1264 * arithmetic overflow in the fmeter_update() routine.
1266 * Given the simple 32 bit integer arithmetic used, this meter works
1267 * best for reporting rates between one per millisecond (msec) and
1268 * one per 32 (approx) seconds. At constant rates faster than one
1269 * per msec it maxes out at values just under 1,000,000. At constant
1270 * rates between one per msec, and one per second it will stabilize
1271 * to a value N*1000, where N is the rate of events per second.
1272 * At constant rates between one per second and one per 32 seconds,
1273 * it will be choppy, moving up on the seconds that have an event,
1274 * and then decaying until the next event. At rates slower than
1275 * about one in 32 seconds, it decays all the way back to zero between
1279 #define FM_COEF 933 /* coefficient for half-life of 10 secs */
1280 #define FM_MAXTICKS ((time_t)99) /* useless computing more ticks than this */
1281 #define FM_MAXCNT 1000000 /* limit cnt to avoid overflow */
1282 #define FM_SCALE 1000 /* faux fixed point scale */
1284 /* Initialize a frequency meter */
1285 static void fmeter_init(struct fmeter *fmp)
1290 spin_lock_init(&fmp->lock);
1293 /* Internal meter update - process cnt events and update value */
1294 static void fmeter_update(struct fmeter *fmp)
1296 time_t now = get_seconds();
1297 time_t ticks = now - fmp->time;
1302 ticks = min(FM_MAXTICKS, ticks);
1304 fmp->val = (FM_COEF * fmp->val) / FM_SCALE;
1307 fmp->val += ((FM_SCALE - FM_COEF) * fmp->cnt) / FM_SCALE;
1311 /* Process any previous ticks, then bump cnt by one (times scale). */
1312 static void fmeter_markevent(struct fmeter *fmp)
1314 spin_lock(&fmp->lock);
1316 fmp->cnt = min(FM_MAXCNT, fmp->cnt + FM_SCALE);
1317 spin_unlock(&fmp->lock);
1320 /* Process any previous ticks, then return current value. */
1321 static int fmeter_getrate(struct fmeter *fmp)
1325 spin_lock(&fmp->lock);
1328 spin_unlock(&fmp->lock);
1332 /* Protected by cgroup_lock */
1333 static cpumask_var_t cpus_attach;
1335 /* Called by cgroups to determine if a cpuset is usable; cgroup_mutex held */
1336 static int cpuset_can_attach(struct cgroup_subsys *ss, struct cgroup *cont,
1337 struct task_struct *tsk, bool threadgroup)
1340 struct cpuset *cs = cgroup_cs(cont);
1342 if (cpumask_empty(cs->cpus_allowed) || nodes_empty(cs->mems_allowed))
1346 * Kthreads bound to specific cpus cannot be moved to a new cpuset; we
1347 * cannot change their cpu affinity and isolating such threads by their
1348 * set of allowed nodes is unnecessary. Thus, cpusets are not
1349 * applicable for such threads. This prevents checking for success of
1350 * set_cpus_allowed_ptr() on all attached tasks before cpus_allowed may
1353 if (tsk->flags & PF_THREAD_BOUND)
1356 ret = security_task_setscheduler(tsk, 0, NULL);
1360 struct task_struct *c;
1363 list_for_each_entry_rcu(c, &tsk->thread_group, thread_group) {
1364 ret = security_task_setscheduler(c, 0, NULL);
1375 static void cpuset_attach_task(struct task_struct *tsk, nodemask_t *to,
1380 * can_attach beforehand should guarantee that this doesn't fail.
1381 * TODO: have a better way to handle failure here
1383 err = set_cpus_allowed_ptr(tsk, cpus_attach);
1387 cpuset_change_task_nodemask(tsk, to);
1389 cpuset_update_task_spread_flag(cs, tsk);
1393 static void cpuset_attach(struct cgroup_subsys *ss, struct cgroup *cont,
1394 struct cgroup *oldcont, struct task_struct *tsk,
1397 struct mm_struct *mm;
1398 struct cpuset *cs = cgroup_cs(cont);
1399 struct cpuset *oldcs = cgroup_cs(oldcont);
1400 NODEMASK_ALLOC(nodemask_t, from, GFP_KERNEL);
1401 NODEMASK_ALLOC(nodemask_t, to, GFP_KERNEL);
1403 if (from == NULL || to == NULL)
1406 if (cs == &top_cpuset) {
1407 cpumask_copy(cpus_attach, cpu_possible_mask);
1409 guarantee_online_cpus(cs, cpus_attach);
1411 guarantee_online_mems(cs, to);
1413 /* do per-task migration stuff possibly for each in the threadgroup */
1414 cpuset_attach_task(tsk, to, cs);
1416 struct task_struct *c;
1418 list_for_each_entry_rcu(c, &tsk->thread_group, thread_group) {
1419 cpuset_attach_task(c, to, cs);
1424 /* change mm; only needs to be done once even if threadgroup */
1425 *from = oldcs->mems_allowed;
1426 *to = cs->mems_allowed;
1427 mm = get_task_mm(tsk);
1429 mpol_rebind_mm(mm, to);
1430 if (is_memory_migrate(cs))
1431 cpuset_migrate_mm(mm, from, to);
1436 NODEMASK_FREE(from);
1440 /* The various types of files and directories in a cpuset file system */
1443 FILE_MEMORY_MIGRATE,
1449 FILE_SCHED_LOAD_BALANCE,
1450 FILE_SCHED_RELAX_DOMAIN_LEVEL,
1451 FILE_MEMORY_PRESSURE_ENABLED,
1452 FILE_MEMORY_PRESSURE,
1455 } cpuset_filetype_t;
1457 static int cpuset_write_u64(struct cgroup *cgrp, struct cftype *cft, u64 val)
1460 struct cpuset *cs = cgroup_cs(cgrp);
1461 cpuset_filetype_t type = cft->private;
1463 if (!cgroup_lock_live_group(cgrp))
1467 case FILE_CPU_EXCLUSIVE:
1468 retval = update_flag(CS_CPU_EXCLUSIVE, cs, val);
1470 case FILE_MEM_EXCLUSIVE:
1471 retval = update_flag(CS_MEM_EXCLUSIVE, cs, val);
1473 case FILE_MEM_HARDWALL:
1474 retval = update_flag(CS_MEM_HARDWALL, cs, val);
1476 case FILE_SCHED_LOAD_BALANCE:
1477 retval = update_flag(CS_SCHED_LOAD_BALANCE, cs, val);
1479 case FILE_MEMORY_MIGRATE:
1480 retval = update_flag(CS_MEMORY_MIGRATE, cs, val);
1482 case FILE_MEMORY_PRESSURE_ENABLED:
1483 cpuset_memory_pressure_enabled = !!val;
1485 case FILE_MEMORY_PRESSURE:
1488 case FILE_SPREAD_PAGE:
1489 retval = update_flag(CS_SPREAD_PAGE, cs, val);
1491 case FILE_SPREAD_SLAB:
1492 retval = update_flag(CS_SPREAD_SLAB, cs, val);
1502 static int cpuset_write_s64(struct cgroup *cgrp, struct cftype *cft, s64 val)
1505 struct cpuset *cs = cgroup_cs(cgrp);
1506 cpuset_filetype_t type = cft->private;
1508 if (!cgroup_lock_live_group(cgrp))
1512 case FILE_SCHED_RELAX_DOMAIN_LEVEL:
1513 retval = update_relax_domain_level(cs, val);
1524 * Common handling for a write to a "cpus" or "mems" file.
1526 static int cpuset_write_resmask(struct cgroup *cgrp, struct cftype *cft,
1530 struct cpuset *cs = cgroup_cs(cgrp);
1531 struct cpuset *trialcs;
1533 if (!cgroup_lock_live_group(cgrp))
1536 trialcs = alloc_trial_cpuset(cs);
1540 switch (cft->private) {
1542 retval = update_cpumask(cs, trialcs, buf);
1545 retval = update_nodemask(cs, trialcs, buf);
1552 free_trial_cpuset(trialcs);
1558 * These ascii lists should be read in a single call, by using a user
1559 * buffer large enough to hold the entire map. If read in smaller
1560 * chunks, there is no guarantee of atomicity. Since the display format
1561 * used, list of ranges of sequential numbers, is variable length,
1562 * and since these maps can change value dynamically, one could read
1563 * gibberish by doing partial reads while a list was changing.
1564 * A single large read to a buffer that crosses a page boundary is
1565 * ok, because the result being copied to user land is not recomputed
1566 * across a page fault.
1569 static int cpuset_sprintf_cpulist(char *page, struct cpuset *cs)
1573 mutex_lock(&callback_mutex);
1574 ret = cpulist_scnprintf(page, PAGE_SIZE, cs->cpus_allowed);
1575 mutex_unlock(&callback_mutex);
1580 static int cpuset_sprintf_memlist(char *page, struct cpuset *cs)
1582 NODEMASK_ALLOC(nodemask_t, mask, GFP_KERNEL);
1588 mutex_lock(&callback_mutex);
1589 *mask = cs->mems_allowed;
1590 mutex_unlock(&callback_mutex);
1592 retval = nodelist_scnprintf(page, PAGE_SIZE, *mask);
1594 NODEMASK_FREE(mask);
1599 static ssize_t cpuset_common_file_read(struct cgroup *cont,
1603 size_t nbytes, loff_t *ppos)
1605 struct cpuset *cs = cgroup_cs(cont);
1606 cpuset_filetype_t type = cft->private;
1611 if (!(page = (char *)__get_free_page(GFP_TEMPORARY)))
1618 s += cpuset_sprintf_cpulist(s, cs);
1621 s += cpuset_sprintf_memlist(s, cs);
1629 retval = simple_read_from_buffer(buf, nbytes, ppos, page, s - page);
1631 free_page((unsigned long)page);
1635 static u64 cpuset_read_u64(struct cgroup *cont, struct cftype *cft)
1637 struct cpuset *cs = cgroup_cs(cont);
1638 cpuset_filetype_t type = cft->private;
1640 case FILE_CPU_EXCLUSIVE:
1641 return is_cpu_exclusive(cs);
1642 case FILE_MEM_EXCLUSIVE:
1643 return is_mem_exclusive(cs);
1644 case FILE_MEM_HARDWALL:
1645 return is_mem_hardwall(cs);
1646 case FILE_SCHED_LOAD_BALANCE:
1647 return is_sched_load_balance(cs);
1648 case FILE_MEMORY_MIGRATE:
1649 return is_memory_migrate(cs);
1650 case FILE_MEMORY_PRESSURE_ENABLED:
1651 return cpuset_memory_pressure_enabled;
1652 case FILE_MEMORY_PRESSURE:
1653 return fmeter_getrate(&cs->fmeter);
1654 case FILE_SPREAD_PAGE:
1655 return is_spread_page(cs);
1656 case FILE_SPREAD_SLAB:
1657 return is_spread_slab(cs);
1662 /* Unreachable but makes gcc happy */
1666 static s64 cpuset_read_s64(struct cgroup *cont, struct cftype *cft)
1668 struct cpuset *cs = cgroup_cs(cont);
1669 cpuset_filetype_t type = cft->private;
1671 case FILE_SCHED_RELAX_DOMAIN_LEVEL:
1672 return cs->relax_domain_level;
1677 /* Unrechable but makes gcc happy */
1683 * for the common functions, 'private' gives the type of file
1686 static struct cftype files[] = {
1689 .read = cpuset_common_file_read,
1690 .write_string = cpuset_write_resmask,
1691 .max_write_len = (100U + 6 * NR_CPUS),
1692 .private = FILE_CPULIST,
1697 .read = cpuset_common_file_read,
1698 .write_string = cpuset_write_resmask,
1699 .max_write_len = (100U + 6 * MAX_NUMNODES),
1700 .private = FILE_MEMLIST,
1704 .name = "cpu_exclusive",
1705 .read_u64 = cpuset_read_u64,
1706 .write_u64 = cpuset_write_u64,
1707 .private = FILE_CPU_EXCLUSIVE,
1711 .name = "mem_exclusive",
1712 .read_u64 = cpuset_read_u64,
1713 .write_u64 = cpuset_write_u64,
1714 .private = FILE_MEM_EXCLUSIVE,
1718 .name = "mem_hardwall",
1719 .read_u64 = cpuset_read_u64,
1720 .write_u64 = cpuset_write_u64,
1721 .private = FILE_MEM_HARDWALL,
1725 .name = "sched_load_balance",
1726 .read_u64 = cpuset_read_u64,
1727 .write_u64 = cpuset_write_u64,
1728 .private = FILE_SCHED_LOAD_BALANCE,
1732 .name = "sched_relax_domain_level",
1733 .read_s64 = cpuset_read_s64,
1734 .write_s64 = cpuset_write_s64,
1735 .private = FILE_SCHED_RELAX_DOMAIN_LEVEL,
1739 .name = "memory_migrate",
1740 .read_u64 = cpuset_read_u64,
1741 .write_u64 = cpuset_write_u64,
1742 .private = FILE_MEMORY_MIGRATE,
1746 .name = "memory_pressure",
1747 .read_u64 = cpuset_read_u64,
1748 .write_u64 = cpuset_write_u64,
1749 .private = FILE_MEMORY_PRESSURE,
1754 .name = "memory_spread_page",
1755 .read_u64 = cpuset_read_u64,
1756 .write_u64 = cpuset_write_u64,
1757 .private = FILE_SPREAD_PAGE,
1761 .name = "memory_spread_slab",
1762 .read_u64 = cpuset_read_u64,
1763 .write_u64 = cpuset_write_u64,
1764 .private = FILE_SPREAD_SLAB,
1768 static struct cftype cft_memory_pressure_enabled = {
1769 .name = "memory_pressure_enabled",
1770 .read_u64 = cpuset_read_u64,
1771 .write_u64 = cpuset_write_u64,
1772 .private = FILE_MEMORY_PRESSURE_ENABLED,
1775 static int cpuset_populate(struct cgroup_subsys *ss, struct cgroup *cont)
1779 err = cgroup_add_files(cont, ss, files, ARRAY_SIZE(files));
1782 /* memory_pressure_enabled is in root cpuset only */
1784 err = cgroup_add_file(cont, ss,
1785 &cft_memory_pressure_enabled);
1790 * post_clone() is called at the end of cgroup_clone().
1791 * 'cgroup' was just created automatically as a result of
1792 * a cgroup_clone(), and the current task is about to
1793 * be moved into 'cgroup'.
1795 * Currently we refuse to set up the cgroup - thereby
1796 * refusing the task to be entered, and as a result refusing
1797 * the sys_unshare() or clone() which initiated it - if any
1798 * sibling cpusets have exclusive cpus or mem.
1800 * If this becomes a problem for some users who wish to
1801 * allow that scenario, then cpuset_post_clone() could be
1802 * changed to grant parent->cpus_allowed-sibling_cpus_exclusive
1803 * (and likewise for mems) to the new cgroup. Called with cgroup_mutex
1806 static void cpuset_post_clone(struct cgroup_subsys *ss,
1807 struct cgroup *cgroup)
1809 struct cgroup *parent, *child;
1810 struct cpuset *cs, *parent_cs;
1812 parent = cgroup->parent;
1813 list_for_each_entry(child, &parent->children, sibling) {
1814 cs = cgroup_cs(child);
1815 if (is_mem_exclusive(cs) || is_cpu_exclusive(cs))
1818 cs = cgroup_cs(cgroup);
1819 parent_cs = cgroup_cs(parent);
1821 cs->mems_allowed = parent_cs->mems_allowed;
1822 cpumask_copy(cs->cpus_allowed, parent_cs->cpus_allowed);
1827 * cpuset_create - create a cpuset
1828 * ss: cpuset cgroup subsystem
1829 * cont: control group that the new cpuset will be part of
1832 static struct cgroup_subsys_state *cpuset_create(
1833 struct cgroup_subsys *ss,
1834 struct cgroup *cont)
1837 struct cpuset *parent;
1839 if (!cont->parent) {
1840 return &top_cpuset.css;
1842 parent = cgroup_cs(cont->parent);
1843 cs = kmalloc(sizeof(*cs), GFP_KERNEL);
1845 return ERR_PTR(-ENOMEM);
1846 if (!alloc_cpumask_var(&cs->cpus_allowed, GFP_KERNEL)) {
1848 return ERR_PTR(-ENOMEM);
1852 if (is_spread_page(parent))
1853 set_bit(CS_SPREAD_PAGE, &cs->flags);
1854 if (is_spread_slab(parent))
1855 set_bit(CS_SPREAD_SLAB, &cs->flags);
1856 set_bit(CS_SCHED_LOAD_BALANCE, &cs->flags);
1857 cpumask_clear(cs->cpus_allowed);
1858 nodes_clear(cs->mems_allowed);
1859 fmeter_init(&cs->fmeter);
1860 cs->relax_domain_level = -1;
1862 cs->parent = parent;
1863 number_of_cpusets++;
1868 * If the cpuset being removed has its flag 'sched_load_balance'
1869 * enabled, then simulate turning sched_load_balance off, which
1870 * will call async_rebuild_sched_domains().
1873 static void cpuset_destroy(struct cgroup_subsys *ss, struct cgroup *cont)
1875 struct cpuset *cs = cgroup_cs(cont);
1877 if (is_sched_load_balance(cs))
1878 update_flag(CS_SCHED_LOAD_BALANCE, cs, 0);
1880 number_of_cpusets--;
1881 free_cpumask_var(cs->cpus_allowed);
1885 struct cgroup_subsys cpuset_subsys = {
1887 .create = cpuset_create,
1888 .destroy = cpuset_destroy,
1889 .can_attach = cpuset_can_attach,
1890 .attach = cpuset_attach,
1891 .populate = cpuset_populate,
1892 .post_clone = cpuset_post_clone,
1893 .subsys_id = cpuset_subsys_id,
1898 * cpuset_init - initialize cpusets at system boot
1900 * Description: Initialize top_cpuset and the cpuset internal file system,
1903 int __init cpuset_init(void)
1907 if (!alloc_cpumask_var(&top_cpuset.cpus_allowed, GFP_KERNEL))
1910 cpumask_setall(top_cpuset.cpus_allowed);
1911 nodes_setall(top_cpuset.mems_allowed);
1913 fmeter_init(&top_cpuset.fmeter);
1914 set_bit(CS_SCHED_LOAD_BALANCE, &top_cpuset.flags);
1915 top_cpuset.relax_domain_level = -1;
1917 err = register_filesystem(&cpuset_fs_type);
1921 if (!alloc_cpumask_var(&cpus_attach, GFP_KERNEL))
1924 number_of_cpusets = 1;
1929 * cpuset_do_move_task - move a given task to another cpuset
1930 * @tsk: pointer to task_struct the task to move
1931 * @scan: struct cgroup_scanner contained in its struct cpuset_hotplug_scanner
1933 * Called by cgroup_scan_tasks() for each task in a cgroup.
1934 * Return nonzero to stop the walk through the tasks.
1936 static void cpuset_do_move_task(struct task_struct *tsk,
1937 struct cgroup_scanner *scan)
1939 struct cgroup *new_cgroup = scan->data;
1941 cgroup_attach_task(new_cgroup, tsk);
1945 * move_member_tasks_to_cpuset - move tasks from one cpuset to another
1946 * @from: cpuset in which the tasks currently reside
1947 * @to: cpuset to which the tasks will be moved
1949 * Called with cgroup_mutex held
1950 * callback_mutex must not be held, as cpuset_attach() will take it.
1952 * The cgroup_scan_tasks() function will scan all the tasks in a cgroup,
1953 * calling callback functions for each.
1955 static void move_member_tasks_to_cpuset(struct cpuset *from, struct cpuset *to)
1957 struct cgroup_scanner scan;
1959 scan.cg = from->css.cgroup;
1960 scan.test_task = NULL; /* select all tasks in cgroup */
1961 scan.process_task = cpuset_do_move_task;
1963 scan.data = to->css.cgroup;
1965 if (cgroup_scan_tasks(&scan))
1966 printk(KERN_ERR "move_member_tasks_to_cpuset: "
1967 "cgroup_scan_tasks failed\n");
1971 * If CPU and/or memory hotplug handlers, below, unplug any CPUs
1972 * or memory nodes, we need to walk over the cpuset hierarchy,
1973 * removing that CPU or node from all cpusets. If this removes the
1974 * last CPU or node from a cpuset, then move the tasks in the empty
1975 * cpuset to its next-highest non-empty parent.
1977 * Called with cgroup_mutex held
1978 * callback_mutex must not be held, as cpuset_attach() will take it.
1980 static void remove_tasks_in_empty_cpuset(struct cpuset *cs)
1982 struct cpuset *parent;
1985 * The cgroup's css_sets list is in use if there are tasks
1986 * in the cpuset; the list is empty if there are none;
1987 * the cs->css.refcnt seems always 0.
1989 if (list_empty(&cs->css.cgroup->css_sets))
1993 * Find its next-highest non-empty parent, (top cpuset
1994 * has online cpus, so can't be empty).
1996 parent = cs->parent;
1997 while (cpumask_empty(parent->cpus_allowed) ||
1998 nodes_empty(parent->mems_allowed))
1999 parent = parent->parent;
2001 move_member_tasks_to_cpuset(cs, parent);
2005 * Walk the specified cpuset subtree and look for empty cpusets.
2006 * The tasks of such cpuset must be moved to a parent cpuset.
2008 * Called with cgroup_mutex held. We take callback_mutex to modify
2009 * cpus_allowed and mems_allowed.
2011 * This walk processes the tree from top to bottom, completing one layer
2012 * before dropping down to the next. It always processes a node before
2013 * any of its children.
2015 * For now, since we lack memory hot unplug, we'll never see a cpuset
2016 * that has tasks along with an empty 'mems'. But if we did see such
2017 * a cpuset, we'd handle it just like we do if its 'cpus' was empty.
2019 static void scan_for_empty_cpusets(struct cpuset *root)
2022 struct cpuset *cp; /* scans cpusets being updated */
2023 struct cpuset *child; /* scans child cpusets of cp */
2024 struct cgroup *cont;
2025 NODEMASK_ALLOC(nodemask_t, oldmems, GFP_KERNEL);
2027 if (oldmems == NULL)
2030 list_add_tail((struct list_head *)&root->stack_list, &queue);
2032 while (!list_empty(&queue)) {
2033 cp = list_first_entry(&queue, struct cpuset, stack_list);
2034 list_del(queue.next);
2035 list_for_each_entry(cont, &cp->css.cgroup->children, sibling) {
2036 child = cgroup_cs(cont);
2037 list_add_tail(&child->stack_list, &queue);
2040 /* Continue past cpusets with all cpus, mems online */
2041 if (cpumask_subset(cp->cpus_allowed, cpu_active_mask) &&
2042 nodes_subset(cp->mems_allowed, node_states[N_HIGH_MEMORY]))
2045 *oldmems = cp->mems_allowed;
2047 /* Remove offline cpus and mems from this cpuset. */
2048 mutex_lock(&callback_mutex);
2049 cpumask_and(cp->cpus_allowed, cp->cpus_allowed,
2051 nodes_and(cp->mems_allowed, cp->mems_allowed,
2052 node_states[N_HIGH_MEMORY]);
2053 mutex_unlock(&callback_mutex);
2055 /* Move tasks from the empty cpuset to a parent */
2056 if (cpumask_empty(cp->cpus_allowed) ||
2057 nodes_empty(cp->mems_allowed))
2058 remove_tasks_in_empty_cpuset(cp);
2060 update_tasks_cpumask(cp, NULL);
2061 update_tasks_nodemask(cp, oldmems, NULL);
2064 NODEMASK_FREE(oldmems);
2068 * The top_cpuset tracks what CPUs and Memory Nodes are online,
2069 * period. This is necessary in order to make cpusets transparent
2070 * (of no affect) on systems that are actively using CPU hotplug
2071 * but making no active use of cpusets.
2073 * This routine ensures that top_cpuset.cpus_allowed tracks
2074 * cpu_online_map on each CPU hotplug (cpuhp) event.
2076 * Called within get_online_cpus(). Needs to call cgroup_lock()
2077 * before calling generate_sched_domains().
2079 static int cpuset_track_online_cpus(struct notifier_block *unused_nb,
2080 unsigned long phase, void *unused_cpu)
2082 struct sched_domain_attr *attr;
2083 cpumask_var_t *doms;
2088 case CPU_ONLINE_FROZEN:
2089 case CPU_DOWN_PREPARE:
2090 case CPU_DOWN_PREPARE_FROZEN:
2091 case CPU_DOWN_FAILED:
2092 case CPU_DOWN_FAILED_FROZEN:
2100 mutex_lock(&callback_mutex);
2101 cpumask_copy(top_cpuset.cpus_allowed, cpu_active_mask);
2102 mutex_unlock(&callback_mutex);
2103 scan_for_empty_cpusets(&top_cpuset);
2104 ndoms = generate_sched_domains(&doms, &attr);
2107 /* Have scheduler rebuild the domains */
2108 partition_sched_domains(ndoms, doms, attr);
2113 #ifdef CONFIG_MEMORY_HOTPLUG
2115 * Keep top_cpuset.mems_allowed tracking node_states[N_HIGH_MEMORY].
2116 * Call this routine anytime after node_states[N_HIGH_MEMORY] changes.
2117 * See also the previous routine cpuset_track_online_cpus().
2119 static int cpuset_track_online_nodes(struct notifier_block *self,
2120 unsigned long action, void *arg)
2122 NODEMASK_ALLOC(nodemask_t, oldmems, GFP_KERNEL);
2124 if (oldmems == NULL)
2130 *oldmems = top_cpuset.mems_allowed;
2131 mutex_lock(&callback_mutex);
2132 top_cpuset.mems_allowed = node_states[N_HIGH_MEMORY];
2133 mutex_unlock(&callback_mutex);
2134 update_tasks_nodemask(&top_cpuset, oldmems, NULL);
2138 * needn't update top_cpuset.mems_allowed explicitly because
2139 * scan_for_empty_cpusets() will update it.
2141 scan_for_empty_cpusets(&top_cpuset);
2148 NODEMASK_FREE(oldmems);
2154 * cpuset_init_smp - initialize cpus_allowed
2156 * Description: Finish top cpuset after cpu, node maps are initialized
2159 void __init cpuset_init_smp(void)
2161 cpumask_copy(top_cpuset.cpus_allowed, cpu_active_mask);
2162 top_cpuset.mems_allowed = node_states[N_HIGH_MEMORY];
2164 hotcpu_notifier(cpuset_track_online_cpus, 0);
2165 hotplug_memory_notifier(cpuset_track_online_nodes, 10);
2167 cpuset_wq = create_singlethread_workqueue("cpuset");
2172 * cpuset_cpus_allowed - return cpus_allowed mask from a tasks cpuset.
2173 * @tsk: pointer to task_struct from which to obtain cpuset->cpus_allowed.
2174 * @pmask: pointer to struct cpumask variable to receive cpus_allowed set.
2176 * Description: Returns the cpumask_var_t cpus_allowed of the cpuset
2177 * attached to the specified @tsk. Guaranteed to return some non-empty
2178 * subset of cpu_online_map, even if this means going outside the
2182 void cpuset_cpus_allowed(struct task_struct *tsk, struct cpumask *pmask)
2184 mutex_lock(&callback_mutex);
2185 cpuset_cpus_allowed_locked(tsk, pmask);
2186 mutex_unlock(&callback_mutex);
2190 * cpuset_cpus_allowed_locked - return cpus_allowed mask from a tasks cpuset.
2191 * Must be called with callback_mutex held.
2193 void cpuset_cpus_allowed_locked(struct task_struct *tsk, struct cpumask *pmask)
2196 guarantee_online_cpus(task_cs(tsk), pmask);
2200 void cpuset_init_current_mems_allowed(void)
2202 nodes_setall(current->mems_allowed);
2206 * cpuset_mems_allowed - return mems_allowed mask from a tasks cpuset.
2207 * @tsk: pointer to task_struct from which to obtain cpuset->mems_allowed.
2209 * Description: Returns the nodemask_t mems_allowed of the cpuset
2210 * attached to the specified @tsk. Guaranteed to return some non-empty
2211 * subset of node_states[N_HIGH_MEMORY], even if this means going outside the
2215 nodemask_t cpuset_mems_allowed(struct task_struct *tsk)
2219 mutex_lock(&callback_mutex);
2221 guarantee_online_mems(task_cs(tsk), &mask);
2223 mutex_unlock(&callback_mutex);
2229 * cpuset_nodemask_valid_mems_allowed - check nodemask vs. curremt mems_allowed
2230 * @nodemask: the nodemask to be checked
2232 * Are any of the nodes in the nodemask allowed in current->mems_allowed?
2234 int cpuset_nodemask_valid_mems_allowed(nodemask_t *nodemask)
2236 return nodes_intersects(*nodemask, current->mems_allowed);
2240 * nearest_hardwall_ancestor() - Returns the nearest mem_exclusive or
2241 * mem_hardwall ancestor to the specified cpuset. Call holding
2242 * callback_mutex. If no ancestor is mem_exclusive or mem_hardwall
2243 * (an unusual configuration), then returns the root cpuset.
2245 static const struct cpuset *nearest_hardwall_ancestor(const struct cpuset *cs)
2247 while (!(is_mem_exclusive(cs) || is_mem_hardwall(cs)) && cs->parent)
2253 * cpuset_node_allowed_softwall - Can we allocate on a memory node?
2254 * @node: is this an allowed node?
2255 * @gfp_mask: memory allocation flags
2257 * If we're in interrupt, yes, we can always allocate. If __GFP_THISNODE is
2258 * set, yes, we can always allocate. If node is in our task's mems_allowed,
2259 * yes. If it's not a __GFP_HARDWALL request and this node is in the nearest
2260 * hardwalled cpuset ancestor to this task's cpuset, yes. If the task has been
2261 * OOM killed and has access to memory reserves as specified by the TIF_MEMDIE
2265 * If __GFP_HARDWALL is set, cpuset_node_allowed_softwall() reduces to
2266 * cpuset_node_allowed_hardwall(). Otherwise, cpuset_node_allowed_softwall()
2267 * might sleep, and might allow a node from an enclosing cpuset.
2269 * cpuset_node_allowed_hardwall() only handles the simpler case of hardwall
2270 * cpusets, and never sleeps.
2272 * The __GFP_THISNODE placement logic is really handled elsewhere,
2273 * by forcibly using a zonelist starting at a specified node, and by
2274 * (in get_page_from_freelist()) refusing to consider the zones for
2275 * any node on the zonelist except the first. By the time any such
2276 * calls get to this routine, we should just shut up and say 'yes'.
2278 * GFP_USER allocations are marked with the __GFP_HARDWALL bit,
2279 * and do not allow allocations outside the current tasks cpuset
2280 * unless the task has been OOM killed as is marked TIF_MEMDIE.
2281 * GFP_KERNEL allocations are not so marked, so can escape to the
2282 * nearest enclosing hardwalled ancestor cpuset.
2284 * Scanning up parent cpusets requires callback_mutex. The
2285 * __alloc_pages() routine only calls here with __GFP_HARDWALL bit
2286 * _not_ set if it's a GFP_KERNEL allocation, and all nodes in the
2287 * current tasks mems_allowed came up empty on the first pass over
2288 * the zonelist. So only GFP_KERNEL allocations, if all nodes in the
2289 * cpuset are short of memory, might require taking the callback_mutex
2292 * The first call here from mm/page_alloc:get_page_from_freelist()
2293 * has __GFP_HARDWALL set in gfp_mask, enforcing hardwall cpusets,
2294 * so no allocation on a node outside the cpuset is allowed (unless
2295 * in interrupt, of course).
2297 * The second pass through get_page_from_freelist() doesn't even call
2298 * here for GFP_ATOMIC calls. For those calls, the __alloc_pages()
2299 * variable 'wait' is not set, and the bit ALLOC_CPUSET is not set
2300 * in alloc_flags. That logic and the checks below have the combined
2302 * in_interrupt - any node ok (current task context irrelevant)
2303 * GFP_ATOMIC - any node ok
2304 * TIF_MEMDIE - any node ok
2305 * GFP_KERNEL - any node in enclosing hardwalled cpuset ok
2306 * GFP_USER - only nodes in current tasks mems allowed ok.
2309 * Don't call cpuset_node_allowed_softwall if you can't sleep, unless you
2310 * pass in the __GFP_HARDWALL flag set in gfp_flag, which disables
2311 * the code that might scan up ancestor cpusets and sleep.
2313 int __cpuset_node_allowed_softwall(int node, gfp_t gfp_mask)
2315 const struct cpuset *cs; /* current cpuset ancestors */
2316 int allowed; /* is allocation in zone z allowed? */
2318 if (in_interrupt() || (gfp_mask & __GFP_THISNODE))
2320 might_sleep_if(!(gfp_mask & __GFP_HARDWALL));
2321 if (node_isset(node, current->mems_allowed))
2324 * Allow tasks that have access to memory reserves because they have
2325 * been OOM killed to get memory anywhere.
2327 if (unlikely(test_thread_flag(TIF_MEMDIE)))
2329 if (gfp_mask & __GFP_HARDWALL) /* If hardwall request, stop here */
2332 if (current->flags & PF_EXITING) /* Let dying task have memory */
2335 /* Not hardwall and node outside mems_allowed: scan up cpusets */
2336 mutex_lock(&callback_mutex);
2339 cs = nearest_hardwall_ancestor(task_cs(current));
2340 task_unlock(current);
2342 allowed = node_isset(node, cs->mems_allowed);
2343 mutex_unlock(&callback_mutex);
2348 * cpuset_node_allowed_hardwall - Can we allocate on a memory node?
2349 * @node: is this an allowed node?
2350 * @gfp_mask: memory allocation flags
2352 * If we're in interrupt, yes, we can always allocate. If __GFP_THISNODE is
2353 * set, yes, we can always allocate. If node is in our task's mems_allowed,
2354 * yes. If the task has been OOM killed and has access to memory reserves as
2355 * specified by the TIF_MEMDIE flag, yes.
2358 * The __GFP_THISNODE placement logic is really handled elsewhere,
2359 * by forcibly using a zonelist starting at a specified node, and by
2360 * (in get_page_from_freelist()) refusing to consider the zones for
2361 * any node on the zonelist except the first. By the time any such
2362 * calls get to this routine, we should just shut up and say 'yes'.
2364 * Unlike the cpuset_node_allowed_softwall() variant, above,
2365 * this variant requires that the node be in the current task's
2366 * mems_allowed or that we're in interrupt. It does not scan up the
2367 * cpuset hierarchy for the nearest enclosing mem_exclusive cpuset.
2370 int __cpuset_node_allowed_hardwall(int node, gfp_t gfp_mask)
2372 if (in_interrupt() || (gfp_mask & __GFP_THISNODE))
2374 if (node_isset(node, current->mems_allowed))
2377 * Allow tasks that have access to memory reserves because they have
2378 * been OOM killed to get memory anywhere.
2380 if (unlikely(test_thread_flag(TIF_MEMDIE)))
2386 * cpuset_lock - lock out any changes to cpuset structures
2388 * The out of memory (oom) code needs to mutex_lock cpusets
2389 * from being changed while it scans the tasklist looking for a
2390 * task in an overlapping cpuset. Expose callback_mutex via this
2391 * cpuset_lock() routine, so the oom code can lock it, before
2392 * locking the task list. The tasklist_lock is a spinlock, so
2393 * must be taken inside callback_mutex.
2396 void cpuset_lock(void)
2398 mutex_lock(&callback_mutex);
2402 * cpuset_unlock - release lock on cpuset changes
2404 * Undo the lock taken in a previous cpuset_lock() call.
2407 void cpuset_unlock(void)
2409 mutex_unlock(&callback_mutex);
2413 * cpuset_mem_spread_node() - On which node to begin search for a page
2415 * If a task is marked PF_SPREAD_PAGE or PF_SPREAD_SLAB (as for
2416 * tasks in a cpuset with is_spread_page or is_spread_slab set),
2417 * and if the memory allocation used cpuset_mem_spread_node()
2418 * to determine on which node to start looking, as it will for
2419 * certain page cache or slab cache pages such as used for file
2420 * system buffers and inode caches, then instead of starting on the
2421 * local node to look for a free page, rather spread the starting
2422 * node around the tasks mems_allowed nodes.
2424 * We don't have to worry about the returned node being offline
2425 * because "it can't happen", and even if it did, it would be ok.
2427 * The routines calling guarantee_online_mems() are careful to
2428 * only set nodes in task->mems_allowed that are online. So it
2429 * should not be possible for the following code to return an
2430 * offline node. But if it did, that would be ok, as this routine
2431 * is not returning the node where the allocation must be, only
2432 * the node where the search should start. The zonelist passed to
2433 * __alloc_pages() will include all nodes. If the slab allocator
2434 * is passed an offline node, it will fall back to the local node.
2435 * See kmem_cache_alloc_node().
2438 int cpuset_mem_spread_node(void)
2442 node = next_node(current->cpuset_mem_spread_rotor, current->mems_allowed);
2443 if (node == MAX_NUMNODES)
2444 node = first_node(current->mems_allowed);
2445 current->cpuset_mem_spread_rotor = node;
2448 EXPORT_SYMBOL_GPL(cpuset_mem_spread_node);
2451 * cpuset_mems_allowed_intersects - Does @tsk1's mems_allowed intersect @tsk2's?
2452 * @tsk1: pointer to task_struct of some task.
2453 * @tsk2: pointer to task_struct of some other task.
2455 * Description: Return true if @tsk1's mems_allowed intersects the
2456 * mems_allowed of @tsk2. Used by the OOM killer to determine if
2457 * one of the task's memory usage might impact the memory available
2461 int cpuset_mems_allowed_intersects(const struct task_struct *tsk1,
2462 const struct task_struct *tsk2)
2464 return nodes_intersects(tsk1->mems_allowed, tsk2->mems_allowed);
2468 * cpuset_print_task_mems_allowed - prints task's cpuset and mems_allowed
2469 * @task: pointer to task_struct of some task.
2471 * Description: Prints @task's name, cpuset name, and cached copy of its
2472 * mems_allowed to the kernel log. Must hold task_lock(task) to allow
2473 * dereferencing task_cs(task).
2475 void cpuset_print_task_mems_allowed(struct task_struct *tsk)
2477 struct dentry *dentry;
2479 dentry = task_cs(tsk)->css.cgroup->dentry;
2480 spin_lock(&cpuset_buffer_lock);
2481 snprintf(cpuset_name, CPUSET_NAME_LEN,
2482 dentry ? (const char *)dentry->d_name.name : "/");
2483 nodelist_scnprintf(cpuset_nodelist, CPUSET_NODELIST_LEN,
2485 printk(KERN_INFO "%s cpuset=%s mems_allowed=%s\n",
2486 tsk->comm, cpuset_name, cpuset_nodelist);
2487 spin_unlock(&cpuset_buffer_lock);
2491 * Collection of memory_pressure is suppressed unless
2492 * this flag is enabled by writing "1" to the special
2493 * cpuset file 'memory_pressure_enabled' in the root cpuset.
2496 int cpuset_memory_pressure_enabled __read_mostly;
2499 * cpuset_memory_pressure_bump - keep stats of per-cpuset reclaims.
2501 * Keep a running average of the rate of synchronous (direct)
2502 * page reclaim efforts initiated by tasks in each cpuset.
2504 * This represents the rate at which some task in the cpuset
2505 * ran low on memory on all nodes it was allowed to use, and
2506 * had to enter the kernels page reclaim code in an effort to
2507 * create more free memory by tossing clean pages or swapping
2508 * or writing dirty pages.
2510 * Display to user space in the per-cpuset read-only file
2511 * "memory_pressure". Value displayed is an integer
2512 * representing the recent rate of entry into the synchronous
2513 * (direct) page reclaim by any task attached to the cpuset.
2516 void __cpuset_memory_pressure_bump(void)
2519 fmeter_markevent(&task_cs(current)->fmeter);
2520 task_unlock(current);
2523 #ifdef CONFIG_PROC_PID_CPUSET
2525 * proc_cpuset_show()
2526 * - Print tasks cpuset path into seq_file.
2527 * - Used for /proc/<pid>/cpuset.
2528 * - No need to task_lock(tsk) on this tsk->cpuset reference, as it
2529 * doesn't really matter if tsk->cpuset changes after we read it,
2530 * and we take cgroup_mutex, keeping cpuset_attach() from changing it
2533 static int proc_cpuset_show(struct seq_file *m, void *unused_v)
2536 struct task_struct *tsk;
2538 struct cgroup_subsys_state *css;
2542 buf = kmalloc(PAGE_SIZE, GFP_KERNEL);
2548 tsk = get_pid_task(pid, PIDTYPE_PID);
2554 css = task_subsys_state(tsk, cpuset_subsys_id);
2555 retval = cgroup_path(css->cgroup, buf, PAGE_SIZE);
2562 put_task_struct(tsk);
2569 static int cpuset_open(struct inode *inode, struct file *file)
2571 struct pid *pid = PROC_I(inode)->pid;
2572 return single_open(file, proc_cpuset_show, pid);
2575 const struct file_operations proc_cpuset_operations = {
2576 .open = cpuset_open,
2578 .llseek = seq_lseek,
2579 .release = single_release,
2581 #endif /* CONFIG_PROC_PID_CPUSET */
2583 /* Display task mems_allowed in /proc/<pid>/status file. */
2584 void cpuset_task_status_allowed(struct seq_file *m, struct task_struct *task)
2586 seq_printf(m, "Mems_allowed:\t");
2587 seq_nodemask(m, &task->mems_allowed);
2588 seq_printf(m, "\n");
2589 seq_printf(m, "Mems_allowed_list:\t");
2590 seq_nodemask_list(m, &task->mems_allowed);
2591 seq_printf(m, "\n");