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 * Tracks how many cpusets are currently defined in system.
65 * When there is only one cpuset (the root cpuset) we can
66 * short circuit some hooks.
68 int number_of_cpusets __read_mostly;
70 /* Forward declare cgroup structures */
71 struct cgroup_subsys cpuset_subsys;
74 /* See "Frequency meter" comments, below. */
77 int cnt; /* unprocessed events count */
78 int val; /* most recent output value */
79 time_t time; /* clock (secs) when val computed */
80 spinlock_t lock; /* guards read or write of above */
84 struct cgroup_subsys_state css;
86 unsigned long flags; /* "unsigned long" so bitops work */
87 cpumask_var_t cpus_allowed; /* CPUs allowed to tasks in cpuset */
88 nodemask_t mems_allowed; /* Memory Nodes allowed to tasks */
90 struct cpuset *parent; /* my parent */
92 struct fmeter fmeter; /* memory_pressure filter */
94 /* partition number for rebuild_sched_domains() */
97 /* for custom sched domain */
98 int relax_domain_level;
100 /* used for walking a cpuset hierarchy */
101 struct list_head stack_list;
104 /* Retrieve the cpuset for a cgroup */
105 static inline struct cpuset *cgroup_cs(struct cgroup *cont)
107 return container_of(cgroup_subsys_state(cont, cpuset_subsys_id),
111 /* Retrieve the cpuset for a task */
112 static inline struct cpuset *task_cs(struct task_struct *task)
114 return container_of(task_subsys_state(task, cpuset_subsys_id),
119 static inline bool task_has_mempolicy(struct task_struct *task)
121 return task->mempolicy;
124 static inline bool task_has_mempolicy(struct task_struct *task)
131 /* bits in struct cpuset flags field */
138 CS_SCHED_LOAD_BALANCE,
143 /* convenient tests for these bits */
144 static inline bool is_cpuset_online(const struct cpuset *cs)
146 return test_bit(CS_ONLINE, &cs->flags);
149 static inline int is_cpu_exclusive(const struct cpuset *cs)
151 return test_bit(CS_CPU_EXCLUSIVE, &cs->flags);
154 static inline int is_mem_exclusive(const struct cpuset *cs)
156 return test_bit(CS_MEM_EXCLUSIVE, &cs->flags);
159 static inline int is_mem_hardwall(const struct cpuset *cs)
161 return test_bit(CS_MEM_HARDWALL, &cs->flags);
164 static inline int is_sched_load_balance(const struct cpuset *cs)
166 return test_bit(CS_SCHED_LOAD_BALANCE, &cs->flags);
169 static inline int is_memory_migrate(const struct cpuset *cs)
171 return test_bit(CS_MEMORY_MIGRATE, &cs->flags);
174 static inline int is_spread_page(const struct cpuset *cs)
176 return test_bit(CS_SPREAD_PAGE, &cs->flags);
179 static inline int is_spread_slab(const struct cpuset *cs)
181 return test_bit(CS_SPREAD_SLAB, &cs->flags);
184 static struct cpuset top_cpuset = {
185 .flags = ((1 << CS_ONLINE) | (1 << CS_CPU_EXCLUSIVE) |
186 (1 << CS_MEM_EXCLUSIVE)),
190 * cpuset_for_each_child - traverse online children of a cpuset
191 * @child_cs: loop cursor pointing to the current child
192 * @pos_cgrp: used for iteration
193 * @parent_cs: target cpuset to walk children of
195 * Walk @child_cs through the online children of @parent_cs. Must be used
196 * with RCU read locked.
198 #define cpuset_for_each_child(child_cs, pos_cgrp, parent_cs) \
199 cgroup_for_each_child((pos_cgrp), (parent_cs)->css.cgroup) \
200 if (is_cpuset_online(((child_cs) = cgroup_cs((pos_cgrp)))))
203 * There are two global mutexes guarding cpuset structures. The first
204 * is the main control groups cgroup_mutex, accessed via
205 * cgroup_lock()/cgroup_unlock(). The second is the cpuset-specific
206 * callback_mutex, below. They can nest. It is ok to first take
207 * cgroup_mutex, then nest callback_mutex. We also require taking
208 * task_lock() when dereferencing a task's cpuset pointer. See "The
209 * task_lock() exception", at the end of this comment.
211 * A task must hold both mutexes to modify cpusets. If a task
212 * holds cgroup_mutex, then it blocks others wanting that mutex,
213 * ensuring that it is the only task able to also acquire callback_mutex
214 * and be able to modify cpusets. It can perform various checks on
215 * the cpuset structure first, knowing nothing will change. It can
216 * also allocate memory while just holding cgroup_mutex. While it is
217 * performing these checks, various callback routines can briefly
218 * acquire callback_mutex to query cpusets. Once it is ready to make
219 * the changes, it takes callback_mutex, blocking everyone else.
221 * Calls to the kernel memory allocator can not be made while holding
222 * callback_mutex, as that would risk double tripping on callback_mutex
223 * from one of the callbacks into the cpuset code from within
226 * If a task is only holding callback_mutex, then it has read-only
229 * Now, the task_struct fields mems_allowed and mempolicy may be changed
230 * by other task, we use alloc_lock in the task_struct fields to protect
233 * The cpuset_common_file_read() handlers only hold callback_mutex across
234 * small pieces of code, such as when reading out possibly multi-word
235 * cpumasks and nodemasks.
237 * Accessing a task's cpuset should be done in accordance with the
238 * guidelines for accessing subsystem state in kernel/cgroup.c
241 static DEFINE_MUTEX(callback_mutex);
244 * cpuset_buffer_lock protects both the cpuset_name and cpuset_nodelist
245 * buffers. They are statically allocated to prevent using excess stack
246 * when calling cpuset_print_task_mems_allowed().
248 #define CPUSET_NAME_LEN (128)
249 #define CPUSET_NODELIST_LEN (256)
250 static char cpuset_name[CPUSET_NAME_LEN];
251 static char cpuset_nodelist[CPUSET_NODELIST_LEN];
252 static DEFINE_SPINLOCK(cpuset_buffer_lock);
255 * CPU / memory hotplug is handled asynchronously.
257 static void cpuset_hotplug_workfn(struct work_struct *work);
259 static DECLARE_WORK(cpuset_hotplug_work, cpuset_hotplug_workfn);
262 * This is ugly, but preserves the userspace API for existing cpuset
263 * users. If someone tries to mount the "cpuset" filesystem, we
264 * silently switch it to mount "cgroup" instead
266 static struct dentry *cpuset_mount(struct file_system_type *fs_type,
267 int flags, const char *unused_dev_name, void *data)
269 struct file_system_type *cgroup_fs = get_fs_type("cgroup");
270 struct dentry *ret = ERR_PTR(-ENODEV);
274 "release_agent=/sbin/cpuset_release_agent";
275 ret = cgroup_fs->mount(cgroup_fs, flags,
276 unused_dev_name, mountopts);
277 put_filesystem(cgroup_fs);
282 static struct file_system_type cpuset_fs_type = {
284 .mount = cpuset_mount,
288 * Return in pmask the portion of a cpusets's cpus_allowed that
289 * are online. If none are online, walk up the cpuset hierarchy
290 * until we find one that does have some online cpus. If we get
291 * all the way to the top and still haven't found any online cpus,
292 * return cpu_online_mask. Or if passed a NULL cs from an exit'ing
293 * task, return cpu_online_mask.
295 * One way or another, we guarantee to return some non-empty subset
296 * of cpu_online_mask.
298 * Call with callback_mutex held.
301 static void guarantee_online_cpus(const struct cpuset *cs,
302 struct cpumask *pmask)
304 while (cs && !cpumask_intersects(cs->cpus_allowed, cpu_online_mask))
307 cpumask_and(pmask, cs->cpus_allowed, cpu_online_mask);
309 cpumask_copy(pmask, cpu_online_mask);
310 BUG_ON(!cpumask_intersects(pmask, cpu_online_mask));
314 * Return in *pmask the portion of a cpusets's mems_allowed that
315 * are online, with memory. If none are online with memory, walk
316 * up the cpuset hierarchy until we find one that does have some
317 * online mems. If we get all the way to the top and still haven't
318 * found any online mems, return node_states[N_MEMORY].
320 * One way or another, we guarantee to return some non-empty subset
321 * of node_states[N_MEMORY].
323 * Call with callback_mutex held.
326 static void guarantee_online_mems(const struct cpuset *cs, nodemask_t *pmask)
328 while (cs && !nodes_intersects(cs->mems_allowed,
329 node_states[N_MEMORY]))
332 nodes_and(*pmask, cs->mems_allowed,
333 node_states[N_MEMORY]);
335 *pmask = node_states[N_MEMORY];
336 BUG_ON(!nodes_intersects(*pmask, node_states[N_MEMORY]));
340 * update task's spread flag if cpuset's page/slab spread flag is set
342 * Called with callback_mutex/cgroup_mutex held
344 static void cpuset_update_task_spread_flag(struct cpuset *cs,
345 struct task_struct *tsk)
347 if (is_spread_page(cs))
348 tsk->flags |= PF_SPREAD_PAGE;
350 tsk->flags &= ~PF_SPREAD_PAGE;
351 if (is_spread_slab(cs))
352 tsk->flags |= PF_SPREAD_SLAB;
354 tsk->flags &= ~PF_SPREAD_SLAB;
358 * is_cpuset_subset(p, q) - Is cpuset p a subset of cpuset q?
360 * One cpuset is a subset of another if all its allowed CPUs and
361 * Memory Nodes are a subset of the other, and its exclusive flags
362 * are only set if the other's are set. Call holding cgroup_mutex.
365 static int is_cpuset_subset(const struct cpuset *p, const struct cpuset *q)
367 return cpumask_subset(p->cpus_allowed, q->cpus_allowed) &&
368 nodes_subset(p->mems_allowed, q->mems_allowed) &&
369 is_cpu_exclusive(p) <= is_cpu_exclusive(q) &&
370 is_mem_exclusive(p) <= is_mem_exclusive(q);
374 * alloc_trial_cpuset - allocate a trial cpuset
375 * @cs: the cpuset that the trial cpuset duplicates
377 static struct cpuset *alloc_trial_cpuset(const struct cpuset *cs)
379 struct cpuset *trial;
381 trial = kmemdup(cs, sizeof(*cs), GFP_KERNEL);
385 if (!alloc_cpumask_var(&trial->cpus_allowed, GFP_KERNEL)) {
389 cpumask_copy(trial->cpus_allowed, cs->cpus_allowed);
395 * free_trial_cpuset - free the trial cpuset
396 * @trial: the trial cpuset to be freed
398 static void free_trial_cpuset(struct cpuset *trial)
400 free_cpumask_var(trial->cpus_allowed);
405 * validate_change() - Used to validate that any proposed cpuset change
406 * follows the structural rules for cpusets.
408 * If we replaced the flag and mask values of the current cpuset
409 * (cur) with those values in the trial cpuset (trial), would
410 * our various subset and exclusive rules still be valid? Presumes
413 * 'cur' is the address of an actual, in-use cpuset. Operations
414 * such as list traversal that depend on the actual address of the
415 * cpuset in the list must use cur below, not trial.
417 * 'trial' is the address of bulk structure copy of cur, with
418 * perhaps one or more of the fields cpus_allowed, mems_allowed,
419 * or flags changed to new, trial values.
421 * Return 0 if valid, -errno if not.
424 static int validate_change(const struct cpuset *cur, const struct cpuset *trial)
427 struct cpuset *c, *par;
432 /* Each of our child cpusets must be a subset of us */
434 cpuset_for_each_child(c, cont, cur)
435 if (!is_cpuset_subset(c, trial))
438 /* Remaining checks don't apply to root cpuset */
440 if (cur == &top_cpuset)
445 /* We must be a subset of our parent cpuset */
447 if (!is_cpuset_subset(trial, par))
451 * If either I or some sibling (!= me) is exclusive, we can't
455 cpuset_for_each_child(c, cont, par) {
456 if ((is_cpu_exclusive(trial) || is_cpu_exclusive(c)) &&
458 cpumask_intersects(trial->cpus_allowed, c->cpus_allowed))
460 if ((is_mem_exclusive(trial) || is_mem_exclusive(c)) &&
462 nodes_intersects(trial->mems_allowed, c->mems_allowed))
466 /* Cpusets with tasks can't have empty cpus_allowed or mems_allowed */
468 if (cgroup_task_count(cur->css.cgroup) &&
469 (cpumask_empty(trial->cpus_allowed) ||
470 nodes_empty(trial->mems_allowed)))
481 * Helper routine for generate_sched_domains().
482 * Do cpusets a, b have overlapping cpus_allowed masks?
484 static int cpusets_overlap(struct cpuset *a, struct cpuset *b)
486 return cpumask_intersects(a->cpus_allowed, b->cpus_allowed);
490 update_domain_attr(struct sched_domain_attr *dattr, struct cpuset *c)
492 if (dattr->relax_domain_level < c->relax_domain_level)
493 dattr->relax_domain_level = c->relax_domain_level;
498 update_domain_attr_tree(struct sched_domain_attr *dattr, struct cpuset *c)
502 list_add(&c->stack_list, &q);
503 while (!list_empty(&q)) {
506 struct cpuset *child;
508 cp = list_first_entry(&q, struct cpuset, stack_list);
511 if (cpumask_empty(cp->cpus_allowed))
514 if (is_sched_load_balance(cp))
515 update_domain_attr(dattr, cp);
518 cpuset_for_each_child(child, cont, cp)
519 list_add_tail(&child->stack_list, &q);
525 * generate_sched_domains()
527 * This function builds a partial partition of the systems CPUs
528 * A 'partial partition' is a set of non-overlapping subsets whose
529 * union is a subset of that set.
530 * The output of this function needs to be passed to kernel/sched.c
531 * partition_sched_domains() routine, which will rebuild the scheduler's
532 * load balancing domains (sched domains) as specified by that partial
535 * See "What is sched_load_balance" in Documentation/cgroups/cpusets.txt
536 * for a background explanation of this.
538 * Does not return errors, on the theory that the callers of this
539 * routine would rather not worry about failures to rebuild sched
540 * domains when operating in the severe memory shortage situations
541 * that could cause allocation failures below.
543 * Must be called with cgroup_lock held.
545 * The three key local variables below are:
546 * q - a linked-list queue of cpuset pointers, used to implement a
547 * top-down scan of all cpusets. This scan loads a pointer
548 * to each cpuset marked is_sched_load_balance into the
549 * array 'csa'. For our purposes, rebuilding the schedulers
550 * sched domains, we can ignore !is_sched_load_balance cpusets.
551 * csa - (for CpuSet Array) Array of pointers to all the cpusets
552 * that need to be load balanced, for convenient iterative
553 * access by the subsequent code that finds the best partition,
554 * i.e the set of domains (subsets) of CPUs such that the
555 * cpus_allowed of every cpuset marked is_sched_load_balance
556 * is a subset of one of these domains, while there are as
557 * many such domains as possible, each as small as possible.
558 * doms - Conversion of 'csa' to an array of cpumasks, for passing to
559 * the kernel/sched.c routine partition_sched_domains() in a
560 * convenient format, that can be easily compared to the prior
561 * value to determine what partition elements (sched domains)
562 * were changed (added or removed.)
564 * Finding the best partition (set of domains):
565 * The triple nested loops below over i, j, k scan over the
566 * load balanced cpusets (using the array of cpuset pointers in
567 * csa[]) looking for pairs of cpusets that have overlapping
568 * cpus_allowed, but which don't have the same 'pn' partition
569 * number and gives them in the same partition number. It keeps
570 * looping on the 'restart' label until it can no longer find
573 * The union of the cpus_allowed masks from the set of
574 * all cpusets having the same 'pn' value then form the one
575 * element of the partition (one sched domain) to be passed to
576 * partition_sched_domains().
578 static int generate_sched_domains(cpumask_var_t **domains,
579 struct sched_domain_attr **attributes)
581 LIST_HEAD(q); /* queue of cpusets to be scanned */
582 struct cpuset *cp; /* scans q */
583 struct cpuset **csa; /* array of all cpuset ptrs */
584 int csn; /* how many cpuset ptrs in csa so far */
585 int i, j, k; /* indices for partition finding loops */
586 cpumask_var_t *doms; /* resulting partition; i.e. sched domains */
587 struct sched_domain_attr *dattr; /* attributes for custom domains */
588 int ndoms = 0; /* number of sched domains in result */
589 int nslot; /* next empty doms[] struct cpumask slot */
595 /* Special case for the 99% of systems with one, full, sched domain */
596 if (is_sched_load_balance(&top_cpuset)) {
598 doms = alloc_sched_domains(ndoms);
602 dattr = kmalloc(sizeof(struct sched_domain_attr), GFP_KERNEL);
604 *dattr = SD_ATTR_INIT;
605 update_domain_attr_tree(dattr, &top_cpuset);
607 cpumask_copy(doms[0], top_cpuset.cpus_allowed);
612 csa = kmalloc(number_of_cpusets * sizeof(cp), GFP_KERNEL);
617 list_add(&top_cpuset.stack_list, &q);
618 while (!list_empty(&q)) {
620 struct cpuset *child; /* scans child cpusets of cp */
622 cp = list_first_entry(&q, struct cpuset, stack_list);
625 if (cpumask_empty(cp->cpus_allowed))
629 * All child cpusets contain a subset of the parent's cpus, so
630 * just skip them, and then we call update_domain_attr_tree()
631 * to calc relax_domain_level of the corresponding sched
634 if (is_sched_load_balance(cp)) {
640 cpuset_for_each_child(child, cont, cp)
641 list_add_tail(&child->stack_list, &q);
645 for (i = 0; i < csn; i++)
650 /* Find the best partition (set of sched domains) */
651 for (i = 0; i < csn; i++) {
652 struct cpuset *a = csa[i];
655 for (j = 0; j < csn; j++) {
656 struct cpuset *b = csa[j];
659 if (apn != bpn && cpusets_overlap(a, b)) {
660 for (k = 0; k < csn; k++) {
661 struct cpuset *c = csa[k];
666 ndoms--; /* one less element */
673 * Now we know how many domains to create.
674 * Convert <csn, csa> to <ndoms, doms> and populate cpu masks.
676 doms = alloc_sched_domains(ndoms);
681 * The rest of the code, including the scheduler, can deal with
682 * dattr==NULL case. No need to abort if alloc fails.
684 dattr = kmalloc(ndoms * sizeof(struct sched_domain_attr), GFP_KERNEL);
686 for (nslot = 0, i = 0; i < csn; i++) {
687 struct cpuset *a = csa[i];
692 /* Skip completed partitions */
698 if (nslot == ndoms) {
699 static int warnings = 10;
702 "rebuild_sched_domains confused:"
703 " nslot %d, ndoms %d, csn %d, i %d,"
705 nslot, ndoms, csn, i, apn);
713 *(dattr + nslot) = SD_ATTR_INIT;
714 for (j = i; j < csn; j++) {
715 struct cpuset *b = csa[j];
718 cpumask_or(dp, dp, b->cpus_allowed);
720 update_domain_attr_tree(dattr + nslot, b);
722 /* Done with this partition */
728 BUG_ON(nslot != ndoms);
734 * Fallback to the default domain if kmalloc() failed.
735 * See comments in partition_sched_domains().
746 * Rebuild scheduler domains.
748 * If the flag 'sched_load_balance' of any cpuset with non-empty
749 * 'cpus' changes, or if the 'cpus' allowed changes in any cpuset
750 * which has that flag enabled, or if any cpuset with a non-empty
751 * 'cpus' is removed, then call this routine to rebuild the
752 * scheduler's dynamic sched domains.
754 * Call with cgroup_mutex held. Takes get_online_cpus().
756 static void rebuild_sched_domains_locked(void)
758 struct sched_domain_attr *attr;
762 WARN_ON_ONCE(!cgroup_lock_is_held());
765 /* Generate domain masks and attrs */
766 ndoms = generate_sched_domains(&doms, &attr);
768 /* Have scheduler rebuild the domains */
769 partition_sched_domains(ndoms, doms, attr);
773 #else /* !CONFIG_SMP */
774 static void rebuild_sched_domains_locked(void)
778 static int generate_sched_domains(cpumask_var_t **domains,
779 struct sched_domain_attr **attributes)
784 #endif /* CONFIG_SMP */
786 void rebuild_sched_domains(void)
789 rebuild_sched_domains_locked();
794 * cpuset_test_cpumask - test a task's cpus_allowed versus its cpuset's
796 * @scan: struct cgroup_scanner contained in its struct cpuset_hotplug_scanner
798 * Call with cgroup_mutex held. May take callback_mutex during call.
799 * Called for each task in a cgroup by cgroup_scan_tasks().
800 * Return nonzero if this tasks's cpus_allowed mask should be changed (in other
801 * words, if its mask is not equal to its cpuset's mask).
803 static int cpuset_test_cpumask(struct task_struct *tsk,
804 struct cgroup_scanner *scan)
806 return !cpumask_equal(&tsk->cpus_allowed,
807 (cgroup_cs(scan->cg))->cpus_allowed);
811 * cpuset_change_cpumask - make a task's cpus_allowed the same as its cpuset's
813 * @scan: struct cgroup_scanner containing the cgroup of the task
815 * Called by cgroup_scan_tasks() for each task in a cgroup whose
816 * cpus_allowed mask needs to be changed.
818 * We don't need to re-check for the cgroup/cpuset membership, since we're
819 * holding cgroup_lock() at this point.
821 static void cpuset_change_cpumask(struct task_struct *tsk,
822 struct cgroup_scanner *scan)
824 set_cpus_allowed_ptr(tsk, ((cgroup_cs(scan->cg))->cpus_allowed));
828 * update_tasks_cpumask - Update the cpumasks of tasks in the cpuset.
829 * @cs: the cpuset in which each task's cpus_allowed mask needs to be changed
830 * @heap: if NULL, defer allocating heap memory to cgroup_scan_tasks()
832 * Called with cgroup_mutex held
834 * The cgroup_scan_tasks() function will scan all the tasks in a cgroup,
835 * calling callback functions for each.
837 * No return value. It's guaranteed that cgroup_scan_tasks() always returns 0
840 static void update_tasks_cpumask(struct cpuset *cs, struct ptr_heap *heap)
842 struct cgroup_scanner scan;
844 scan.cg = cs->css.cgroup;
845 scan.test_task = cpuset_test_cpumask;
846 scan.process_task = cpuset_change_cpumask;
848 cgroup_scan_tasks(&scan);
852 * update_cpumask - update the cpus_allowed mask of a cpuset and all tasks in it
853 * @cs: the cpuset to consider
854 * @buf: buffer of cpu numbers written to this cpuset
856 static int update_cpumask(struct cpuset *cs, struct cpuset *trialcs,
859 struct ptr_heap heap;
861 int is_load_balanced;
863 /* top_cpuset.cpus_allowed tracks cpu_online_mask; it's read-only */
864 if (cs == &top_cpuset)
868 * An empty cpus_allowed is ok only if the cpuset has no tasks.
869 * Since cpulist_parse() fails on an empty mask, we special case
870 * that parsing. The validate_change() call ensures that cpusets
871 * with tasks have cpus.
874 cpumask_clear(trialcs->cpus_allowed);
876 retval = cpulist_parse(buf, trialcs->cpus_allowed);
880 if (!cpumask_subset(trialcs->cpus_allowed, cpu_active_mask))
883 retval = validate_change(cs, trialcs);
887 /* Nothing to do if the cpus didn't change */
888 if (cpumask_equal(cs->cpus_allowed, trialcs->cpus_allowed))
891 retval = heap_init(&heap, PAGE_SIZE, GFP_KERNEL, NULL);
895 is_load_balanced = is_sched_load_balance(trialcs);
897 mutex_lock(&callback_mutex);
898 cpumask_copy(cs->cpus_allowed, trialcs->cpus_allowed);
899 mutex_unlock(&callback_mutex);
902 * Scan tasks in the cpuset, and update the cpumasks of any
903 * that need an update.
905 update_tasks_cpumask(cs, &heap);
909 if (is_load_balanced)
910 rebuild_sched_domains_locked();
917 * Migrate memory region from one set of nodes to another.
919 * Temporarilly set tasks mems_allowed to target nodes of migration,
920 * so that the migration code can allocate pages on these nodes.
922 * Call holding cgroup_mutex, so current's cpuset won't change
923 * during this call, as manage_mutex holds off any cpuset_attach()
924 * calls. Therefore we don't need to take task_lock around the
925 * call to guarantee_online_mems(), as we know no one is changing
928 * While the mm_struct we are migrating is typically from some
929 * other task, the task_struct mems_allowed that we are hacking
930 * is for our current task, which must allocate new pages for that
931 * migrating memory region.
934 static void cpuset_migrate_mm(struct mm_struct *mm, const nodemask_t *from,
935 const nodemask_t *to)
937 struct task_struct *tsk = current;
939 tsk->mems_allowed = *to;
941 do_migrate_pages(mm, from, to, MPOL_MF_MOVE_ALL);
943 guarantee_online_mems(task_cs(tsk),&tsk->mems_allowed);
947 * cpuset_change_task_nodemask - change task's mems_allowed and mempolicy
948 * @tsk: the task to change
949 * @newmems: new nodes that the task will be set
951 * In order to avoid seeing no nodes if the old and new nodes are disjoint,
952 * we structure updates as setting all new allowed nodes, then clearing newly
955 static void cpuset_change_task_nodemask(struct task_struct *tsk,
961 * Allow tasks that have access to memory reserves because they have
962 * been OOM killed to get memory anywhere.
964 if (unlikely(test_thread_flag(TIF_MEMDIE)))
966 if (current->flags & PF_EXITING) /* Let dying task have memory */
971 * Determine if a loop is necessary if another thread is doing
972 * get_mems_allowed(). If at least one node remains unchanged and
973 * tsk does not have a mempolicy, then an empty nodemask will not be
974 * possible when mems_allowed is larger than a word.
976 need_loop = task_has_mempolicy(tsk) ||
977 !nodes_intersects(*newmems, tsk->mems_allowed);
980 write_seqcount_begin(&tsk->mems_allowed_seq);
982 nodes_or(tsk->mems_allowed, tsk->mems_allowed, *newmems);
983 mpol_rebind_task(tsk, newmems, MPOL_REBIND_STEP1);
985 mpol_rebind_task(tsk, newmems, MPOL_REBIND_STEP2);
986 tsk->mems_allowed = *newmems;
989 write_seqcount_end(&tsk->mems_allowed_seq);
995 * Update task's mems_allowed and rebind its mempolicy and vmas' mempolicy
996 * of it to cpuset's new mems_allowed, and migrate pages to new nodes if
997 * memory_migrate flag is set. Called with cgroup_mutex held.
999 static void cpuset_change_nodemask(struct task_struct *p,
1000 struct cgroup_scanner *scan)
1002 struct mm_struct *mm;
1005 const nodemask_t *oldmem = scan->data;
1006 static nodemask_t newmems; /* protected by cgroup_mutex */
1008 cs = cgroup_cs(scan->cg);
1009 guarantee_online_mems(cs, &newmems);
1011 cpuset_change_task_nodemask(p, &newmems);
1013 mm = get_task_mm(p);
1017 migrate = is_memory_migrate(cs);
1019 mpol_rebind_mm(mm, &cs->mems_allowed);
1021 cpuset_migrate_mm(mm, oldmem, &cs->mems_allowed);
1025 static void *cpuset_being_rebound;
1028 * update_tasks_nodemask - Update the nodemasks of tasks in the cpuset.
1029 * @cs: the cpuset in which each task's mems_allowed mask needs to be changed
1030 * @oldmem: old mems_allowed of cpuset cs
1031 * @heap: if NULL, defer allocating heap memory to cgroup_scan_tasks()
1033 * Called with cgroup_mutex held
1034 * No return value. It's guaranteed that cgroup_scan_tasks() always returns 0
1037 static void update_tasks_nodemask(struct cpuset *cs, const nodemask_t *oldmem,
1038 struct ptr_heap *heap)
1040 struct cgroup_scanner scan;
1042 cpuset_being_rebound = cs; /* causes mpol_dup() rebind */
1044 scan.cg = cs->css.cgroup;
1045 scan.test_task = NULL;
1046 scan.process_task = cpuset_change_nodemask;
1048 scan.data = (nodemask_t *)oldmem;
1051 * The mpol_rebind_mm() call takes mmap_sem, which we couldn't
1052 * take while holding tasklist_lock. Forks can happen - the
1053 * mpol_dup() cpuset_being_rebound check will catch such forks,
1054 * and rebind their vma mempolicies too. Because we still hold
1055 * the global cgroup_mutex, we know that no other rebind effort
1056 * will be contending for the global variable cpuset_being_rebound.
1057 * It's ok if we rebind the same mm twice; mpol_rebind_mm()
1058 * is idempotent. Also migrate pages in each mm to new nodes.
1060 cgroup_scan_tasks(&scan);
1062 /* We're done rebinding vmas to this cpuset's new mems_allowed. */
1063 cpuset_being_rebound = NULL;
1067 * Handle user request to change the 'mems' memory placement
1068 * of a cpuset. Needs to validate the request, update the
1069 * cpusets mems_allowed, and for each task in the cpuset,
1070 * update mems_allowed and rebind task's mempolicy and any vma
1071 * mempolicies and if the cpuset is marked 'memory_migrate',
1072 * migrate the tasks pages to the new memory.
1074 * Call with cgroup_mutex held. May take callback_mutex during call.
1075 * Will take tasklist_lock, scan tasklist for tasks in cpuset cs,
1076 * lock each such tasks mm->mmap_sem, scan its vma's and rebind
1077 * their mempolicies to the cpusets new mems_allowed.
1079 static int update_nodemask(struct cpuset *cs, struct cpuset *trialcs,
1082 NODEMASK_ALLOC(nodemask_t, oldmem, GFP_KERNEL);
1084 struct ptr_heap heap;
1090 * top_cpuset.mems_allowed tracks node_stats[N_MEMORY];
1093 if (cs == &top_cpuset) {
1099 * An empty mems_allowed is ok iff there are no tasks in the cpuset.
1100 * Since nodelist_parse() fails on an empty mask, we special case
1101 * that parsing. The validate_change() call ensures that cpusets
1102 * with tasks have memory.
1105 nodes_clear(trialcs->mems_allowed);
1107 retval = nodelist_parse(buf, trialcs->mems_allowed);
1111 if (!nodes_subset(trialcs->mems_allowed,
1112 node_states[N_MEMORY])) {
1117 *oldmem = cs->mems_allowed;
1118 if (nodes_equal(*oldmem, trialcs->mems_allowed)) {
1119 retval = 0; /* Too easy - nothing to do */
1122 retval = validate_change(cs, trialcs);
1126 retval = heap_init(&heap, PAGE_SIZE, GFP_KERNEL, NULL);
1130 mutex_lock(&callback_mutex);
1131 cs->mems_allowed = trialcs->mems_allowed;
1132 mutex_unlock(&callback_mutex);
1134 update_tasks_nodemask(cs, oldmem, &heap);
1138 NODEMASK_FREE(oldmem);
1142 int current_cpuset_is_being_rebound(void)
1144 return task_cs(current) == cpuset_being_rebound;
1147 static int update_relax_domain_level(struct cpuset *cs, s64 val)
1150 if (val < -1 || val >= sched_domain_level_max)
1154 if (val != cs->relax_domain_level) {
1155 cs->relax_domain_level = val;
1156 if (!cpumask_empty(cs->cpus_allowed) &&
1157 is_sched_load_balance(cs))
1158 rebuild_sched_domains_locked();
1165 * cpuset_change_flag - make a task's spread flags the same as its cpuset's
1166 * @tsk: task to be updated
1167 * @scan: struct cgroup_scanner containing the cgroup of the task
1169 * Called by cgroup_scan_tasks() for each task in a cgroup.
1171 * We don't need to re-check for the cgroup/cpuset membership, since we're
1172 * holding cgroup_lock() at this point.
1174 static void cpuset_change_flag(struct task_struct *tsk,
1175 struct cgroup_scanner *scan)
1177 cpuset_update_task_spread_flag(cgroup_cs(scan->cg), tsk);
1181 * update_tasks_flags - update the spread flags of tasks in the cpuset.
1182 * @cs: the cpuset in which each task's spread flags needs to be changed
1183 * @heap: if NULL, defer allocating heap memory to cgroup_scan_tasks()
1185 * Called with cgroup_mutex held
1187 * The cgroup_scan_tasks() function will scan all the tasks in a cgroup,
1188 * calling callback functions for each.
1190 * No return value. It's guaranteed that cgroup_scan_tasks() always returns 0
1193 static void update_tasks_flags(struct cpuset *cs, struct ptr_heap *heap)
1195 struct cgroup_scanner scan;
1197 scan.cg = cs->css.cgroup;
1198 scan.test_task = NULL;
1199 scan.process_task = cpuset_change_flag;
1201 cgroup_scan_tasks(&scan);
1205 * update_flag - read a 0 or a 1 in a file and update associated flag
1206 * bit: the bit to update (see cpuset_flagbits_t)
1207 * cs: the cpuset to update
1208 * turning_on: whether the flag is being set or cleared
1210 * Call with cgroup_mutex held.
1213 static int update_flag(cpuset_flagbits_t bit, struct cpuset *cs,
1216 struct cpuset *trialcs;
1217 int balance_flag_changed;
1218 int spread_flag_changed;
1219 struct ptr_heap heap;
1222 trialcs = alloc_trial_cpuset(cs);
1227 set_bit(bit, &trialcs->flags);
1229 clear_bit(bit, &trialcs->flags);
1231 err = validate_change(cs, trialcs);
1235 err = heap_init(&heap, PAGE_SIZE, GFP_KERNEL, NULL);
1239 balance_flag_changed = (is_sched_load_balance(cs) !=
1240 is_sched_load_balance(trialcs));
1242 spread_flag_changed = ((is_spread_slab(cs) != is_spread_slab(trialcs))
1243 || (is_spread_page(cs) != is_spread_page(trialcs)));
1245 mutex_lock(&callback_mutex);
1246 cs->flags = trialcs->flags;
1247 mutex_unlock(&callback_mutex);
1249 if (!cpumask_empty(trialcs->cpus_allowed) && balance_flag_changed)
1250 rebuild_sched_domains_locked();
1252 if (spread_flag_changed)
1253 update_tasks_flags(cs, &heap);
1256 free_trial_cpuset(trialcs);
1261 * Frequency meter - How fast is some event occurring?
1263 * These routines manage a digitally filtered, constant time based,
1264 * event frequency meter. There are four routines:
1265 * fmeter_init() - initialize a frequency meter.
1266 * fmeter_markevent() - called each time the event happens.
1267 * fmeter_getrate() - returns the recent rate of such events.
1268 * fmeter_update() - internal routine used to update fmeter.
1270 * A common data structure is passed to each of these routines,
1271 * which is used to keep track of the state required to manage the
1272 * frequency meter and its digital filter.
1274 * The filter works on the number of events marked per unit time.
1275 * The filter is single-pole low-pass recursive (IIR). The time unit
1276 * is 1 second. Arithmetic is done using 32-bit integers scaled to
1277 * simulate 3 decimal digits of precision (multiplied by 1000).
1279 * With an FM_COEF of 933, and a time base of 1 second, the filter
1280 * has a half-life of 10 seconds, meaning that if the events quit
1281 * happening, then the rate returned from the fmeter_getrate()
1282 * will be cut in half each 10 seconds, until it converges to zero.
1284 * It is not worth doing a real infinitely recursive filter. If more
1285 * than FM_MAXTICKS ticks have elapsed since the last filter event,
1286 * just compute FM_MAXTICKS ticks worth, by which point the level
1289 * Limit the count of unprocessed events to FM_MAXCNT, so as to avoid
1290 * arithmetic overflow in the fmeter_update() routine.
1292 * Given the simple 32 bit integer arithmetic used, this meter works
1293 * best for reporting rates between one per millisecond (msec) and
1294 * one per 32 (approx) seconds. At constant rates faster than one
1295 * per msec it maxes out at values just under 1,000,000. At constant
1296 * rates between one per msec, and one per second it will stabilize
1297 * to a value N*1000, where N is the rate of events per second.
1298 * At constant rates between one per second and one per 32 seconds,
1299 * it will be choppy, moving up on the seconds that have an event,
1300 * and then decaying until the next event. At rates slower than
1301 * about one in 32 seconds, it decays all the way back to zero between
1305 #define FM_COEF 933 /* coefficient for half-life of 10 secs */
1306 #define FM_MAXTICKS ((time_t)99) /* useless computing more ticks than this */
1307 #define FM_MAXCNT 1000000 /* limit cnt to avoid overflow */
1308 #define FM_SCALE 1000 /* faux fixed point scale */
1310 /* Initialize a frequency meter */
1311 static void fmeter_init(struct fmeter *fmp)
1316 spin_lock_init(&fmp->lock);
1319 /* Internal meter update - process cnt events and update value */
1320 static void fmeter_update(struct fmeter *fmp)
1322 time_t now = get_seconds();
1323 time_t ticks = now - fmp->time;
1328 ticks = min(FM_MAXTICKS, ticks);
1330 fmp->val = (FM_COEF * fmp->val) / FM_SCALE;
1333 fmp->val += ((FM_SCALE - FM_COEF) * fmp->cnt) / FM_SCALE;
1337 /* Process any previous ticks, then bump cnt by one (times scale). */
1338 static void fmeter_markevent(struct fmeter *fmp)
1340 spin_lock(&fmp->lock);
1342 fmp->cnt = min(FM_MAXCNT, fmp->cnt + FM_SCALE);
1343 spin_unlock(&fmp->lock);
1346 /* Process any previous ticks, then return current value. */
1347 static int fmeter_getrate(struct fmeter *fmp)
1351 spin_lock(&fmp->lock);
1354 spin_unlock(&fmp->lock);
1358 /* Called by cgroups to determine if a cpuset is usable; cgroup_mutex held */
1359 static int cpuset_can_attach(struct cgroup *cgrp, struct cgroup_taskset *tset)
1361 struct cpuset *cs = cgroup_cs(cgrp);
1362 struct task_struct *task;
1365 if (cpumask_empty(cs->cpus_allowed) || nodes_empty(cs->mems_allowed))
1368 cgroup_taskset_for_each(task, cgrp, tset) {
1370 * Kthreads bound to specific cpus cannot be moved to a new
1371 * cpuset; we cannot change their cpu affinity and
1372 * isolating such threads by their set of allowed nodes is
1373 * unnecessary. Thus, cpusets are not applicable for such
1374 * threads. This prevents checking for success of
1375 * set_cpus_allowed_ptr() on all attached tasks before
1376 * cpus_allowed may be changed.
1378 if (task->flags & PF_THREAD_BOUND)
1380 if ((ret = security_task_setscheduler(task)))
1388 * Protected by cgroup_mutex. cpus_attach is used only by cpuset_attach()
1389 * but we can't allocate it dynamically there. Define it global and
1390 * allocate from cpuset_init().
1392 static cpumask_var_t cpus_attach;
1394 static void cpuset_attach(struct cgroup *cgrp, struct cgroup_taskset *tset)
1396 /* static bufs protected by cgroup_mutex */
1397 static nodemask_t cpuset_attach_nodemask_from;
1398 static nodemask_t cpuset_attach_nodemask_to;
1399 struct mm_struct *mm;
1400 struct task_struct *task;
1401 struct task_struct *leader = cgroup_taskset_first(tset);
1402 struct cgroup *oldcgrp = cgroup_taskset_cur_cgroup(tset);
1403 struct cpuset *cs = cgroup_cs(cgrp);
1404 struct cpuset *oldcs = cgroup_cs(oldcgrp);
1406 /* prepare for attach */
1407 if (cs == &top_cpuset)
1408 cpumask_copy(cpus_attach, cpu_possible_mask);
1410 guarantee_online_cpus(cs, cpus_attach);
1412 guarantee_online_mems(cs, &cpuset_attach_nodemask_to);
1414 cgroup_taskset_for_each(task, cgrp, tset) {
1416 * can_attach beforehand should guarantee that this doesn't
1417 * fail. TODO: have a better way to handle failure here
1419 WARN_ON_ONCE(set_cpus_allowed_ptr(task, cpus_attach));
1421 cpuset_change_task_nodemask(task, &cpuset_attach_nodemask_to);
1422 cpuset_update_task_spread_flag(cs, task);
1426 * Change mm, possibly for multiple threads in a threadgroup. This is
1427 * expensive and may sleep.
1429 cpuset_attach_nodemask_from = oldcs->mems_allowed;
1430 cpuset_attach_nodemask_to = cs->mems_allowed;
1431 mm = get_task_mm(leader);
1433 mpol_rebind_mm(mm, &cpuset_attach_nodemask_to);
1434 if (is_memory_migrate(cs))
1435 cpuset_migrate_mm(mm, &cpuset_attach_nodemask_from,
1436 &cpuset_attach_nodemask_to);
1441 /* The various types of files and directories in a cpuset file system */
1444 FILE_MEMORY_MIGRATE,
1450 FILE_SCHED_LOAD_BALANCE,
1451 FILE_SCHED_RELAX_DOMAIN_LEVEL,
1452 FILE_MEMORY_PRESSURE_ENABLED,
1453 FILE_MEMORY_PRESSURE,
1456 } cpuset_filetype_t;
1458 static int cpuset_write_u64(struct cgroup *cgrp, struct cftype *cft, u64 val)
1461 struct cpuset *cs = cgroup_cs(cgrp);
1462 cpuset_filetype_t type = cft->private;
1464 if (!cgroup_lock_live_group(cgrp))
1468 case FILE_CPU_EXCLUSIVE:
1469 retval = update_flag(CS_CPU_EXCLUSIVE, cs, val);
1471 case FILE_MEM_EXCLUSIVE:
1472 retval = update_flag(CS_MEM_EXCLUSIVE, cs, val);
1474 case FILE_MEM_HARDWALL:
1475 retval = update_flag(CS_MEM_HARDWALL, cs, val);
1477 case FILE_SCHED_LOAD_BALANCE:
1478 retval = update_flag(CS_SCHED_LOAD_BALANCE, cs, val);
1480 case FILE_MEMORY_MIGRATE:
1481 retval = update_flag(CS_MEMORY_MIGRATE, cs, val);
1483 case FILE_MEMORY_PRESSURE_ENABLED:
1484 cpuset_memory_pressure_enabled = !!val;
1486 case FILE_MEMORY_PRESSURE:
1489 case FILE_SPREAD_PAGE:
1490 retval = update_flag(CS_SPREAD_PAGE, cs, val);
1492 case FILE_SPREAD_SLAB:
1493 retval = update_flag(CS_SPREAD_SLAB, cs, val);
1503 static int cpuset_write_s64(struct cgroup *cgrp, struct cftype *cft, s64 val)
1506 struct cpuset *cs = cgroup_cs(cgrp);
1507 cpuset_filetype_t type = cft->private;
1509 if (!cgroup_lock_live_group(cgrp))
1513 case FILE_SCHED_RELAX_DOMAIN_LEVEL:
1514 retval = update_relax_domain_level(cs, val);
1525 * Common handling for a write to a "cpus" or "mems" file.
1527 static int cpuset_write_resmask(struct cgroup *cgrp, struct cftype *cft,
1531 struct cpuset *cs = cgroup_cs(cgrp);
1532 struct cpuset *trialcs;
1535 * CPU or memory hotunplug may leave @cs w/o any execution
1536 * resources, in which case the hotplug code asynchronously updates
1537 * configuration and transfers all tasks to the nearest ancestor
1538 * which can execute.
1540 * As writes to "cpus" or "mems" may restore @cs's execution
1541 * resources, wait for the previously scheduled operations before
1542 * proceeding, so that we don't end up keep removing tasks added
1543 * after execution capability is restored.
1545 flush_work(&cpuset_hotplug_work);
1547 if (!cgroup_lock_live_group(cgrp))
1550 trialcs = alloc_trial_cpuset(cs);
1556 switch (cft->private) {
1558 retval = update_cpumask(cs, trialcs, buf);
1561 retval = update_nodemask(cs, trialcs, buf);
1568 free_trial_cpuset(trialcs);
1575 * These ascii lists should be read in a single call, by using a user
1576 * buffer large enough to hold the entire map. If read in smaller
1577 * chunks, there is no guarantee of atomicity. Since the display format
1578 * used, list of ranges of sequential numbers, is variable length,
1579 * and since these maps can change value dynamically, one could read
1580 * gibberish by doing partial reads while a list was changing.
1581 * A single large read to a buffer that crosses a page boundary is
1582 * ok, because the result being copied to user land is not recomputed
1583 * across a page fault.
1586 static size_t cpuset_sprintf_cpulist(char *page, struct cpuset *cs)
1590 mutex_lock(&callback_mutex);
1591 count = cpulist_scnprintf(page, PAGE_SIZE, cs->cpus_allowed);
1592 mutex_unlock(&callback_mutex);
1597 static size_t cpuset_sprintf_memlist(char *page, struct cpuset *cs)
1601 mutex_lock(&callback_mutex);
1602 count = nodelist_scnprintf(page, PAGE_SIZE, cs->mems_allowed);
1603 mutex_unlock(&callback_mutex);
1608 static ssize_t cpuset_common_file_read(struct cgroup *cont,
1612 size_t nbytes, loff_t *ppos)
1614 struct cpuset *cs = cgroup_cs(cont);
1615 cpuset_filetype_t type = cft->private;
1620 if (!(page = (char *)__get_free_page(GFP_TEMPORARY)))
1627 s += cpuset_sprintf_cpulist(s, cs);
1630 s += cpuset_sprintf_memlist(s, cs);
1638 retval = simple_read_from_buffer(buf, nbytes, ppos, page, s - page);
1640 free_page((unsigned long)page);
1644 static u64 cpuset_read_u64(struct cgroup *cont, struct cftype *cft)
1646 struct cpuset *cs = cgroup_cs(cont);
1647 cpuset_filetype_t type = cft->private;
1649 case FILE_CPU_EXCLUSIVE:
1650 return is_cpu_exclusive(cs);
1651 case FILE_MEM_EXCLUSIVE:
1652 return is_mem_exclusive(cs);
1653 case FILE_MEM_HARDWALL:
1654 return is_mem_hardwall(cs);
1655 case FILE_SCHED_LOAD_BALANCE:
1656 return is_sched_load_balance(cs);
1657 case FILE_MEMORY_MIGRATE:
1658 return is_memory_migrate(cs);
1659 case FILE_MEMORY_PRESSURE_ENABLED:
1660 return cpuset_memory_pressure_enabled;
1661 case FILE_MEMORY_PRESSURE:
1662 return fmeter_getrate(&cs->fmeter);
1663 case FILE_SPREAD_PAGE:
1664 return is_spread_page(cs);
1665 case FILE_SPREAD_SLAB:
1666 return is_spread_slab(cs);
1671 /* Unreachable but makes gcc happy */
1675 static s64 cpuset_read_s64(struct cgroup *cont, struct cftype *cft)
1677 struct cpuset *cs = cgroup_cs(cont);
1678 cpuset_filetype_t type = cft->private;
1680 case FILE_SCHED_RELAX_DOMAIN_LEVEL:
1681 return cs->relax_domain_level;
1686 /* Unrechable but makes gcc happy */
1692 * for the common functions, 'private' gives the type of file
1695 static struct cftype files[] = {
1698 .read = cpuset_common_file_read,
1699 .write_string = cpuset_write_resmask,
1700 .max_write_len = (100U + 6 * NR_CPUS),
1701 .private = FILE_CPULIST,
1706 .read = cpuset_common_file_read,
1707 .write_string = cpuset_write_resmask,
1708 .max_write_len = (100U + 6 * MAX_NUMNODES),
1709 .private = FILE_MEMLIST,
1713 .name = "cpu_exclusive",
1714 .read_u64 = cpuset_read_u64,
1715 .write_u64 = cpuset_write_u64,
1716 .private = FILE_CPU_EXCLUSIVE,
1720 .name = "mem_exclusive",
1721 .read_u64 = cpuset_read_u64,
1722 .write_u64 = cpuset_write_u64,
1723 .private = FILE_MEM_EXCLUSIVE,
1727 .name = "mem_hardwall",
1728 .read_u64 = cpuset_read_u64,
1729 .write_u64 = cpuset_write_u64,
1730 .private = FILE_MEM_HARDWALL,
1734 .name = "sched_load_balance",
1735 .read_u64 = cpuset_read_u64,
1736 .write_u64 = cpuset_write_u64,
1737 .private = FILE_SCHED_LOAD_BALANCE,
1741 .name = "sched_relax_domain_level",
1742 .read_s64 = cpuset_read_s64,
1743 .write_s64 = cpuset_write_s64,
1744 .private = FILE_SCHED_RELAX_DOMAIN_LEVEL,
1748 .name = "memory_migrate",
1749 .read_u64 = cpuset_read_u64,
1750 .write_u64 = cpuset_write_u64,
1751 .private = FILE_MEMORY_MIGRATE,
1755 .name = "memory_pressure",
1756 .read_u64 = cpuset_read_u64,
1757 .write_u64 = cpuset_write_u64,
1758 .private = FILE_MEMORY_PRESSURE,
1763 .name = "memory_spread_page",
1764 .read_u64 = cpuset_read_u64,
1765 .write_u64 = cpuset_write_u64,
1766 .private = FILE_SPREAD_PAGE,
1770 .name = "memory_spread_slab",
1771 .read_u64 = cpuset_read_u64,
1772 .write_u64 = cpuset_write_u64,
1773 .private = FILE_SPREAD_SLAB,
1777 .name = "memory_pressure_enabled",
1778 .flags = CFTYPE_ONLY_ON_ROOT,
1779 .read_u64 = cpuset_read_u64,
1780 .write_u64 = cpuset_write_u64,
1781 .private = FILE_MEMORY_PRESSURE_ENABLED,
1788 * cpuset_css_alloc - allocate a cpuset css
1789 * cont: control group that the new cpuset will be part of
1792 static struct cgroup_subsys_state *cpuset_css_alloc(struct cgroup *cont)
1797 return &top_cpuset.css;
1799 cs = kzalloc(sizeof(*cs), GFP_KERNEL);
1801 return ERR_PTR(-ENOMEM);
1802 if (!alloc_cpumask_var(&cs->cpus_allowed, GFP_KERNEL)) {
1804 return ERR_PTR(-ENOMEM);
1807 set_bit(CS_SCHED_LOAD_BALANCE, &cs->flags);
1808 cpumask_clear(cs->cpus_allowed);
1809 nodes_clear(cs->mems_allowed);
1810 fmeter_init(&cs->fmeter);
1811 cs->relax_domain_level = -1;
1812 cs->parent = cgroup_cs(cont->parent);
1817 static int cpuset_css_online(struct cgroup *cgrp)
1819 struct cpuset *cs = cgroup_cs(cgrp);
1820 struct cpuset *parent = cs->parent;
1821 struct cpuset *tmp_cs;
1822 struct cgroup *pos_cg;
1827 set_bit(CS_ONLINE, &cs->flags);
1828 if (is_spread_page(parent))
1829 set_bit(CS_SPREAD_PAGE, &cs->flags);
1830 if (is_spread_slab(parent))
1831 set_bit(CS_SPREAD_SLAB, &cs->flags);
1833 number_of_cpusets++;
1835 if (!test_bit(CGRP_CPUSET_CLONE_CHILDREN, &cgrp->flags))
1839 * Clone @parent's configuration if CGRP_CPUSET_CLONE_CHILDREN is
1840 * set. This flag handling is implemented in cgroup core for
1841 * histrical reasons - the flag may be specified during mount.
1843 * Currently, if any sibling cpusets have exclusive cpus or mem, we
1844 * refuse to clone the configuration - thereby refusing the task to
1845 * be entered, and as a result refusing the sys_unshare() or
1846 * clone() which initiated it. If this becomes a problem for some
1847 * users who wish to allow that scenario, then this could be
1848 * changed to grant parent->cpus_allowed-sibling_cpus_exclusive
1849 * (and likewise for mems) to the new cgroup.
1852 cpuset_for_each_child(tmp_cs, pos_cg, parent) {
1853 if (is_mem_exclusive(tmp_cs) || is_cpu_exclusive(tmp_cs)) {
1860 mutex_lock(&callback_mutex);
1861 cs->mems_allowed = parent->mems_allowed;
1862 cpumask_copy(cs->cpus_allowed, parent->cpus_allowed);
1863 mutex_unlock(&callback_mutex);
1868 static void cpuset_css_offline(struct cgroup *cgrp)
1870 struct cpuset *cs = cgroup_cs(cgrp);
1872 /* css_offline is called w/o cgroup_mutex, grab it */
1875 if (is_sched_load_balance(cs))
1876 update_flag(CS_SCHED_LOAD_BALANCE, cs, 0);
1878 number_of_cpusets--;
1879 clear_bit(CS_ONLINE, &cs->flags);
1885 * If the cpuset being removed has its flag 'sched_load_balance'
1886 * enabled, then simulate turning sched_load_balance off, which
1887 * will call rebuild_sched_domains_locked().
1890 static void cpuset_css_free(struct cgroup *cont)
1892 struct cpuset *cs = cgroup_cs(cont);
1894 free_cpumask_var(cs->cpus_allowed);
1898 struct cgroup_subsys cpuset_subsys = {
1900 .css_alloc = cpuset_css_alloc,
1901 .css_online = cpuset_css_online,
1902 .css_offline = cpuset_css_offline,
1903 .css_free = cpuset_css_free,
1904 .can_attach = cpuset_can_attach,
1905 .attach = cpuset_attach,
1906 .subsys_id = cpuset_subsys_id,
1907 .base_cftypes = files,
1912 * cpuset_init - initialize cpusets at system boot
1914 * Description: Initialize top_cpuset and the cpuset internal file system,
1917 int __init cpuset_init(void)
1921 if (!alloc_cpumask_var(&top_cpuset.cpus_allowed, GFP_KERNEL))
1924 cpumask_setall(top_cpuset.cpus_allowed);
1925 nodes_setall(top_cpuset.mems_allowed);
1927 fmeter_init(&top_cpuset.fmeter);
1928 set_bit(CS_SCHED_LOAD_BALANCE, &top_cpuset.flags);
1929 top_cpuset.relax_domain_level = -1;
1931 err = register_filesystem(&cpuset_fs_type);
1935 if (!alloc_cpumask_var(&cpus_attach, GFP_KERNEL))
1938 number_of_cpusets = 1;
1943 * cpuset_do_move_task - move a given task to another cpuset
1944 * @tsk: pointer to task_struct the task to move
1945 * @scan: struct cgroup_scanner contained in its struct cpuset_hotplug_scanner
1947 * Called by cgroup_scan_tasks() for each task in a cgroup.
1948 * Return nonzero to stop the walk through the tasks.
1950 static void cpuset_do_move_task(struct task_struct *tsk,
1951 struct cgroup_scanner *scan)
1953 struct cgroup *new_cgroup = scan->data;
1955 cgroup_attach_task(new_cgroup, tsk);
1959 * move_member_tasks_to_cpuset - move tasks from one cpuset to another
1960 * @from: cpuset in which the tasks currently reside
1961 * @to: cpuset to which the tasks will be moved
1963 * Called with cgroup_mutex held
1964 * callback_mutex must not be held, as cpuset_attach() will take it.
1966 * The cgroup_scan_tasks() function will scan all the tasks in a cgroup,
1967 * calling callback functions for each.
1969 static void move_member_tasks_to_cpuset(struct cpuset *from, struct cpuset *to)
1971 struct cgroup_scanner scan;
1973 scan.cg = from->css.cgroup;
1974 scan.test_task = NULL; /* select all tasks in cgroup */
1975 scan.process_task = cpuset_do_move_task;
1977 scan.data = to->css.cgroup;
1979 if (cgroup_scan_tasks(&scan))
1980 printk(KERN_ERR "move_member_tasks_to_cpuset: "
1981 "cgroup_scan_tasks failed\n");
1985 * If CPU and/or memory hotplug handlers, below, unplug any CPUs
1986 * or memory nodes, we need to walk over the cpuset hierarchy,
1987 * removing that CPU or node from all cpusets. If this removes the
1988 * last CPU or node from a cpuset, then move the tasks in the empty
1989 * cpuset to its next-highest non-empty parent.
1991 * Called with cgroup_mutex held
1992 * callback_mutex must not be held, as cpuset_attach() will take it.
1994 static void remove_tasks_in_empty_cpuset(struct cpuset *cs)
1996 struct cpuset *parent;
1999 * Find its next-highest non-empty parent, (top cpuset
2000 * has online cpus, so can't be empty).
2002 parent = cs->parent;
2003 while (cpumask_empty(parent->cpus_allowed) ||
2004 nodes_empty(parent->mems_allowed))
2005 parent = parent->parent;
2007 move_member_tasks_to_cpuset(cs, parent);
2011 * Helper function to traverse cpusets.
2012 * It can be used to walk the cpuset tree from top to bottom, completing
2013 * one layer before dropping down to the next (thus always processing a
2014 * node before any of its children).
2016 static struct cpuset *cpuset_next(struct list_head *queue)
2019 struct cpuset *child; /* scans child cpusets of cp */
2020 struct cgroup *cont;
2022 if (list_empty(queue))
2025 cp = list_first_entry(queue, struct cpuset, stack_list);
2026 list_del(queue->next);
2028 cpuset_for_each_child(child, cont, cp)
2029 list_add_tail(&child->stack_list, queue);
2036 * cpuset_propagate_hotplug - propagate CPU/memory hotplug to a cpuset
2037 * @cs: cpuset in interest
2039 * Compare @cs's cpu and mem masks against top_cpuset and if some have gone
2040 * offline, update @cs accordingly. If @cs ends up with no CPU or memory,
2041 * all its tasks are moved to the nearest ancestor with both resources.
2043 * Should be called with cgroup_mutex held.
2045 static void cpuset_propagate_hotplug(struct cpuset *cs)
2047 static cpumask_t off_cpus;
2048 static nodemask_t off_mems, tmp_mems;
2050 WARN_ON_ONCE(!cgroup_lock_is_held());
2052 cpumask_andnot(&off_cpus, cs->cpus_allowed, top_cpuset.cpus_allowed);
2053 nodes_andnot(off_mems, cs->mems_allowed, top_cpuset.mems_allowed);
2055 /* remove offline cpus from @cs */
2056 if (!cpumask_empty(&off_cpus)) {
2057 mutex_lock(&callback_mutex);
2058 cpumask_andnot(cs->cpus_allowed, cs->cpus_allowed, &off_cpus);
2059 mutex_unlock(&callback_mutex);
2060 update_tasks_cpumask(cs, NULL);
2063 /* remove offline mems from @cs */
2064 if (!nodes_empty(off_mems)) {
2065 tmp_mems = cs->mems_allowed;
2066 mutex_lock(&callback_mutex);
2067 nodes_andnot(cs->mems_allowed, cs->mems_allowed, off_mems);
2068 mutex_unlock(&callback_mutex);
2069 update_tasks_nodemask(cs, &tmp_mems, NULL);
2072 if (cpumask_empty(cs->cpus_allowed) || nodes_empty(cs->mems_allowed))
2073 remove_tasks_in_empty_cpuset(cs);
2077 * cpuset_hotplug_workfn - handle CPU/memory hotunplug for a cpuset
2079 * This function is called after either CPU or memory configuration has
2080 * changed and updates cpuset accordingly. The top_cpuset is always
2081 * synchronized to cpu_active_mask and N_MEMORY, which is necessary in
2082 * order to make cpusets transparent (of no affect) on systems that are
2083 * actively using CPU hotplug but making no active use of cpusets.
2085 * Non-root cpusets are only affected by offlining. If any CPUs or memory
2086 * nodes have been taken down, cpuset_propagate_hotplug() is invoked on all
2089 * Note that CPU offlining during suspend is ignored. We don't modify
2090 * cpusets across suspend/resume cycles at all.
2092 static void cpuset_hotplug_workfn(struct work_struct *work)
2094 static cpumask_t new_cpus, tmp_cpus;
2095 static nodemask_t new_mems, tmp_mems;
2096 bool cpus_updated, mems_updated;
2097 bool cpus_offlined, mems_offlined;
2101 /* fetch the available cpus/mems and find out which changed how */
2102 cpumask_copy(&new_cpus, cpu_active_mask);
2103 new_mems = node_states[N_MEMORY];
2105 cpus_updated = !cpumask_equal(top_cpuset.cpus_allowed, &new_cpus);
2106 cpus_offlined = cpumask_andnot(&tmp_cpus, top_cpuset.cpus_allowed,
2109 mems_updated = !nodes_equal(top_cpuset.mems_allowed, new_mems);
2110 nodes_andnot(tmp_mems, top_cpuset.mems_allowed, new_mems);
2111 mems_offlined = !nodes_empty(tmp_mems);
2113 /* synchronize cpus_allowed to cpu_active_mask */
2115 mutex_lock(&callback_mutex);
2116 cpumask_copy(top_cpuset.cpus_allowed, &new_cpus);
2117 mutex_unlock(&callback_mutex);
2118 /* we don't mess with cpumasks of tasks in top_cpuset */
2121 /* synchronize mems_allowed to N_MEMORY */
2123 tmp_mems = top_cpuset.mems_allowed;
2124 mutex_lock(&callback_mutex);
2125 top_cpuset.mems_allowed = new_mems;
2126 mutex_unlock(&callback_mutex);
2127 update_tasks_nodemask(&top_cpuset, &tmp_mems, NULL);
2130 /* if cpus or mems went down, we need to propagate to descendants */
2131 if (cpus_offlined || mems_offlined) {
2135 list_add_tail(&top_cpuset.stack_list, &queue);
2136 while ((cs = cpuset_next(&queue)))
2137 if (cs != &top_cpuset)
2138 cpuset_propagate_hotplug(cs);
2143 /* rebuild sched domains if cpus_allowed has changed */
2145 struct sched_domain_attr *attr;
2146 cpumask_var_t *doms;
2150 ndoms = generate_sched_domains(&doms, &attr);
2153 partition_sched_domains(ndoms, doms, attr);
2157 void cpuset_update_active_cpus(bool cpu_online)
2160 * We're inside cpu hotplug critical region which usually nests
2161 * inside cgroup synchronization. Bounce actual hotplug processing
2162 * to a work item to avoid reverse locking order.
2164 * We still need to do partition_sched_domains() synchronously;
2165 * otherwise, the scheduler will get confused and put tasks to the
2166 * dead CPU. Fall back to the default single domain.
2167 * cpuset_hotplug_workfn() will rebuild it as necessary.
2169 partition_sched_domains(1, NULL, NULL);
2170 schedule_work(&cpuset_hotplug_work);
2173 #ifdef CONFIG_MEMORY_HOTPLUG
2175 * Keep top_cpuset.mems_allowed tracking node_states[N_MEMORY].
2176 * Call this routine anytime after node_states[N_MEMORY] changes.
2177 * See cpuset_update_active_cpus() for CPU hotplug handling.
2179 static int cpuset_track_online_nodes(struct notifier_block *self,
2180 unsigned long action, void *arg)
2182 schedule_work(&cpuset_hotplug_work);
2188 * cpuset_init_smp - initialize cpus_allowed
2190 * Description: Finish top cpuset after cpu, node maps are initialized
2193 void __init cpuset_init_smp(void)
2195 cpumask_copy(top_cpuset.cpus_allowed, cpu_active_mask);
2196 top_cpuset.mems_allowed = node_states[N_MEMORY];
2198 hotplug_memory_notifier(cpuset_track_online_nodes, 10);
2202 * cpuset_cpus_allowed - return cpus_allowed mask from a tasks cpuset.
2203 * @tsk: pointer to task_struct from which to obtain cpuset->cpus_allowed.
2204 * @pmask: pointer to struct cpumask variable to receive cpus_allowed set.
2206 * Description: Returns the cpumask_var_t cpus_allowed of the cpuset
2207 * attached to the specified @tsk. Guaranteed to return some non-empty
2208 * subset of cpu_online_mask, even if this means going outside the
2212 void cpuset_cpus_allowed(struct task_struct *tsk, struct cpumask *pmask)
2214 mutex_lock(&callback_mutex);
2216 guarantee_online_cpus(task_cs(tsk), pmask);
2218 mutex_unlock(&callback_mutex);
2221 void cpuset_cpus_allowed_fallback(struct task_struct *tsk)
2223 const struct cpuset *cs;
2228 do_set_cpus_allowed(tsk, cs->cpus_allowed);
2232 * We own tsk->cpus_allowed, nobody can change it under us.
2234 * But we used cs && cs->cpus_allowed lockless and thus can
2235 * race with cgroup_attach_task() or update_cpumask() and get
2236 * the wrong tsk->cpus_allowed. However, both cases imply the
2237 * subsequent cpuset_change_cpumask()->set_cpus_allowed_ptr()
2238 * which takes task_rq_lock().
2240 * If we are called after it dropped the lock we must see all
2241 * changes in tsk_cs()->cpus_allowed. Otherwise we can temporary
2242 * set any mask even if it is not right from task_cs() pov,
2243 * the pending set_cpus_allowed_ptr() will fix things.
2245 * select_fallback_rq() will fix things ups and set cpu_possible_mask
2250 void cpuset_init_current_mems_allowed(void)
2252 nodes_setall(current->mems_allowed);
2256 * cpuset_mems_allowed - return mems_allowed mask from a tasks cpuset.
2257 * @tsk: pointer to task_struct from which to obtain cpuset->mems_allowed.
2259 * Description: Returns the nodemask_t mems_allowed of the cpuset
2260 * attached to the specified @tsk. Guaranteed to return some non-empty
2261 * subset of node_states[N_MEMORY], even if this means going outside the
2265 nodemask_t cpuset_mems_allowed(struct task_struct *tsk)
2269 mutex_lock(&callback_mutex);
2271 guarantee_online_mems(task_cs(tsk), &mask);
2273 mutex_unlock(&callback_mutex);
2279 * cpuset_nodemask_valid_mems_allowed - check nodemask vs. curremt mems_allowed
2280 * @nodemask: the nodemask to be checked
2282 * Are any of the nodes in the nodemask allowed in current->mems_allowed?
2284 int cpuset_nodemask_valid_mems_allowed(nodemask_t *nodemask)
2286 return nodes_intersects(*nodemask, current->mems_allowed);
2290 * nearest_hardwall_ancestor() - Returns the nearest mem_exclusive or
2291 * mem_hardwall ancestor to the specified cpuset. Call holding
2292 * callback_mutex. If no ancestor is mem_exclusive or mem_hardwall
2293 * (an unusual configuration), then returns the root cpuset.
2295 static const struct cpuset *nearest_hardwall_ancestor(const struct cpuset *cs)
2297 while (!(is_mem_exclusive(cs) || is_mem_hardwall(cs)) && cs->parent)
2303 * cpuset_node_allowed_softwall - Can we allocate on a memory node?
2304 * @node: is this an allowed node?
2305 * @gfp_mask: memory allocation flags
2307 * If we're in interrupt, yes, we can always allocate. If __GFP_THISNODE is
2308 * set, yes, we can always allocate. If node is in our task's mems_allowed,
2309 * yes. If it's not a __GFP_HARDWALL request and this node is in the nearest
2310 * hardwalled cpuset ancestor to this task's cpuset, yes. If the task has been
2311 * OOM killed and has access to memory reserves as specified by the TIF_MEMDIE
2315 * If __GFP_HARDWALL is set, cpuset_node_allowed_softwall() reduces to
2316 * cpuset_node_allowed_hardwall(). Otherwise, cpuset_node_allowed_softwall()
2317 * might sleep, and might allow a node from an enclosing cpuset.
2319 * cpuset_node_allowed_hardwall() only handles the simpler case of hardwall
2320 * cpusets, and never sleeps.
2322 * The __GFP_THISNODE placement logic is really handled elsewhere,
2323 * by forcibly using a zonelist starting at a specified node, and by
2324 * (in get_page_from_freelist()) refusing to consider the zones for
2325 * any node on the zonelist except the first. By the time any such
2326 * calls get to this routine, we should just shut up and say 'yes'.
2328 * GFP_USER allocations are marked with the __GFP_HARDWALL bit,
2329 * and do not allow allocations outside the current tasks cpuset
2330 * unless the task has been OOM killed as is marked TIF_MEMDIE.
2331 * GFP_KERNEL allocations are not so marked, so can escape to the
2332 * nearest enclosing hardwalled ancestor cpuset.
2334 * Scanning up parent cpusets requires callback_mutex. The
2335 * __alloc_pages() routine only calls here with __GFP_HARDWALL bit
2336 * _not_ set if it's a GFP_KERNEL allocation, and all nodes in the
2337 * current tasks mems_allowed came up empty on the first pass over
2338 * the zonelist. So only GFP_KERNEL allocations, if all nodes in the
2339 * cpuset are short of memory, might require taking the callback_mutex
2342 * The first call here from mm/page_alloc:get_page_from_freelist()
2343 * has __GFP_HARDWALL set in gfp_mask, enforcing hardwall cpusets,
2344 * so no allocation on a node outside the cpuset is allowed (unless
2345 * in interrupt, of course).
2347 * The second pass through get_page_from_freelist() doesn't even call
2348 * here for GFP_ATOMIC calls. For those calls, the __alloc_pages()
2349 * variable 'wait' is not set, and the bit ALLOC_CPUSET is not set
2350 * in alloc_flags. That logic and the checks below have the combined
2352 * in_interrupt - any node ok (current task context irrelevant)
2353 * GFP_ATOMIC - any node ok
2354 * TIF_MEMDIE - any node ok
2355 * GFP_KERNEL - any node in enclosing hardwalled cpuset ok
2356 * GFP_USER - only nodes in current tasks mems allowed ok.
2359 * Don't call cpuset_node_allowed_softwall if you can't sleep, unless you
2360 * pass in the __GFP_HARDWALL flag set in gfp_flag, which disables
2361 * the code that might scan up ancestor cpusets and sleep.
2363 int __cpuset_node_allowed_softwall(int node, gfp_t gfp_mask)
2365 const struct cpuset *cs; /* current cpuset ancestors */
2366 int allowed; /* is allocation in zone z allowed? */
2368 if (in_interrupt() || (gfp_mask & __GFP_THISNODE))
2370 might_sleep_if(!(gfp_mask & __GFP_HARDWALL));
2371 if (node_isset(node, current->mems_allowed))
2374 * Allow tasks that have access to memory reserves because they have
2375 * been OOM killed to get memory anywhere.
2377 if (unlikely(test_thread_flag(TIF_MEMDIE)))
2379 if (gfp_mask & __GFP_HARDWALL) /* If hardwall request, stop here */
2382 if (current->flags & PF_EXITING) /* Let dying task have memory */
2385 /* Not hardwall and node outside mems_allowed: scan up cpusets */
2386 mutex_lock(&callback_mutex);
2389 cs = nearest_hardwall_ancestor(task_cs(current));
2390 task_unlock(current);
2392 allowed = node_isset(node, cs->mems_allowed);
2393 mutex_unlock(&callback_mutex);
2398 * cpuset_node_allowed_hardwall - Can we allocate on a memory node?
2399 * @node: is this an allowed node?
2400 * @gfp_mask: memory allocation flags
2402 * If we're in interrupt, yes, we can always allocate. If __GFP_THISNODE is
2403 * set, yes, we can always allocate. If node is in our task's mems_allowed,
2404 * yes. If the task has been OOM killed and has access to memory reserves as
2405 * specified by the TIF_MEMDIE flag, yes.
2408 * The __GFP_THISNODE placement logic is really handled elsewhere,
2409 * by forcibly using a zonelist starting at a specified node, and by
2410 * (in get_page_from_freelist()) refusing to consider the zones for
2411 * any node on the zonelist except the first. By the time any such
2412 * calls get to this routine, we should just shut up and say 'yes'.
2414 * Unlike the cpuset_node_allowed_softwall() variant, above,
2415 * this variant requires that the node be in the current task's
2416 * mems_allowed or that we're in interrupt. It does not scan up the
2417 * cpuset hierarchy for the nearest enclosing mem_exclusive cpuset.
2420 int __cpuset_node_allowed_hardwall(int node, gfp_t gfp_mask)
2422 if (in_interrupt() || (gfp_mask & __GFP_THISNODE))
2424 if (node_isset(node, current->mems_allowed))
2427 * Allow tasks that have access to memory reserves because they have
2428 * been OOM killed to get memory anywhere.
2430 if (unlikely(test_thread_flag(TIF_MEMDIE)))
2436 * cpuset_mem_spread_node() - On which node to begin search for a file page
2437 * cpuset_slab_spread_node() - On which node to begin search for a slab page
2439 * If a task is marked PF_SPREAD_PAGE or PF_SPREAD_SLAB (as for
2440 * tasks in a cpuset with is_spread_page or is_spread_slab set),
2441 * and if the memory allocation used cpuset_mem_spread_node()
2442 * to determine on which node to start looking, as it will for
2443 * certain page cache or slab cache pages such as used for file
2444 * system buffers and inode caches, then instead of starting on the
2445 * local node to look for a free page, rather spread the starting
2446 * node around the tasks mems_allowed nodes.
2448 * We don't have to worry about the returned node being offline
2449 * because "it can't happen", and even if it did, it would be ok.
2451 * The routines calling guarantee_online_mems() are careful to
2452 * only set nodes in task->mems_allowed that are online. So it
2453 * should not be possible for the following code to return an
2454 * offline node. But if it did, that would be ok, as this routine
2455 * is not returning the node where the allocation must be, only
2456 * the node where the search should start. The zonelist passed to
2457 * __alloc_pages() will include all nodes. If the slab allocator
2458 * is passed an offline node, it will fall back to the local node.
2459 * See kmem_cache_alloc_node().
2462 static int cpuset_spread_node(int *rotor)
2466 node = next_node(*rotor, current->mems_allowed);
2467 if (node == MAX_NUMNODES)
2468 node = first_node(current->mems_allowed);
2473 int cpuset_mem_spread_node(void)
2475 if (current->cpuset_mem_spread_rotor == NUMA_NO_NODE)
2476 current->cpuset_mem_spread_rotor =
2477 node_random(¤t->mems_allowed);
2479 return cpuset_spread_node(¤t->cpuset_mem_spread_rotor);
2482 int cpuset_slab_spread_node(void)
2484 if (current->cpuset_slab_spread_rotor == NUMA_NO_NODE)
2485 current->cpuset_slab_spread_rotor =
2486 node_random(¤t->mems_allowed);
2488 return cpuset_spread_node(¤t->cpuset_slab_spread_rotor);
2491 EXPORT_SYMBOL_GPL(cpuset_mem_spread_node);
2494 * cpuset_mems_allowed_intersects - Does @tsk1's mems_allowed intersect @tsk2's?
2495 * @tsk1: pointer to task_struct of some task.
2496 * @tsk2: pointer to task_struct of some other task.
2498 * Description: Return true if @tsk1's mems_allowed intersects the
2499 * mems_allowed of @tsk2. Used by the OOM killer to determine if
2500 * one of the task's memory usage might impact the memory available
2504 int cpuset_mems_allowed_intersects(const struct task_struct *tsk1,
2505 const struct task_struct *tsk2)
2507 return nodes_intersects(tsk1->mems_allowed, tsk2->mems_allowed);
2511 * cpuset_print_task_mems_allowed - prints task's cpuset and mems_allowed
2512 * @task: pointer to task_struct of some task.
2514 * Description: Prints @task's name, cpuset name, and cached copy of its
2515 * mems_allowed to the kernel log. Must hold task_lock(task) to allow
2516 * dereferencing task_cs(task).
2518 void cpuset_print_task_mems_allowed(struct task_struct *tsk)
2520 struct dentry *dentry;
2522 dentry = task_cs(tsk)->css.cgroup->dentry;
2523 spin_lock(&cpuset_buffer_lock);
2524 snprintf(cpuset_name, CPUSET_NAME_LEN,
2525 dentry ? (const char *)dentry->d_name.name : "/");
2526 nodelist_scnprintf(cpuset_nodelist, CPUSET_NODELIST_LEN,
2528 printk(KERN_INFO "%s cpuset=%s mems_allowed=%s\n",
2529 tsk->comm, cpuset_name, cpuset_nodelist);
2530 spin_unlock(&cpuset_buffer_lock);
2534 * Collection of memory_pressure is suppressed unless
2535 * this flag is enabled by writing "1" to the special
2536 * cpuset file 'memory_pressure_enabled' in the root cpuset.
2539 int cpuset_memory_pressure_enabled __read_mostly;
2542 * cpuset_memory_pressure_bump - keep stats of per-cpuset reclaims.
2544 * Keep a running average of the rate of synchronous (direct)
2545 * page reclaim efforts initiated by tasks in each cpuset.
2547 * This represents the rate at which some task in the cpuset
2548 * ran low on memory on all nodes it was allowed to use, and
2549 * had to enter the kernels page reclaim code in an effort to
2550 * create more free memory by tossing clean pages or swapping
2551 * or writing dirty pages.
2553 * Display to user space in the per-cpuset read-only file
2554 * "memory_pressure". Value displayed is an integer
2555 * representing the recent rate of entry into the synchronous
2556 * (direct) page reclaim by any task attached to the cpuset.
2559 void __cpuset_memory_pressure_bump(void)
2562 fmeter_markevent(&task_cs(current)->fmeter);
2563 task_unlock(current);
2566 #ifdef CONFIG_PROC_PID_CPUSET
2568 * proc_cpuset_show()
2569 * - Print tasks cpuset path into seq_file.
2570 * - Used for /proc/<pid>/cpuset.
2571 * - No need to task_lock(tsk) on this tsk->cpuset reference, as it
2572 * doesn't really matter if tsk->cpuset changes after we read it,
2573 * and we take cgroup_mutex, keeping cpuset_attach() from changing it
2576 static int proc_cpuset_show(struct seq_file *m, void *unused_v)
2579 struct task_struct *tsk;
2581 struct cgroup_subsys_state *css;
2585 buf = kmalloc(PAGE_SIZE, GFP_KERNEL);
2591 tsk = get_pid_task(pid, PIDTYPE_PID);
2597 css = task_subsys_state(tsk, cpuset_subsys_id);
2598 retval = cgroup_path(css->cgroup, buf, PAGE_SIZE);
2605 put_task_struct(tsk);
2612 static int cpuset_open(struct inode *inode, struct file *file)
2614 struct pid *pid = PROC_I(inode)->pid;
2615 return single_open(file, proc_cpuset_show, pid);
2618 const struct file_operations proc_cpuset_operations = {
2619 .open = cpuset_open,
2621 .llseek = seq_lseek,
2622 .release = single_release,
2624 #endif /* CONFIG_PROC_PID_CPUSET */
2626 /* Display task mems_allowed in /proc/<pid>/status file. */
2627 void cpuset_task_status_allowed(struct seq_file *m, struct task_struct *task)
2629 seq_printf(m, "Mems_allowed:\t");
2630 seq_nodemask(m, &task->mems_allowed);
2631 seq_printf(m, "\n");
2632 seq_printf(m, "Mems_allowed_list:\t");
2633 seq_nodemask_list(m, &task->mems_allowed);
2634 seq_printf(m, "\n");