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/sched/mm.h>
48 #include <linux/sched/task.h>
49 #include <linux/seq_file.h>
50 #include <linux/security.h>
51 #include <linux/slab.h>
52 #include <linux/spinlock.h>
53 #include <linux/stat.h>
54 #include <linux/string.h>
55 #include <linux/time.h>
56 #include <linux/time64.h>
57 #include <linux/backing-dev.h>
58 #include <linux/sort.h>
59 #include <linux/oom.h>
60 #include <linux/sched/isolation.h>
61 #include <linux/uaccess.h>
62 #include <linux/atomic.h>
63 #include <linux/mutex.h>
64 #include <linux/cgroup.h>
65 #include <linux/wait.h>
67 DEFINE_STATIC_KEY_FALSE(cpusets_pre_enable_key);
68 DEFINE_STATIC_KEY_FALSE(cpusets_enabled_key);
70 /* See "Frequency meter" comments, below. */
73 int cnt; /* unprocessed events count */
74 int val; /* most recent output value */
75 time64_t time; /* clock (secs) when val computed */
76 spinlock_t lock; /* guards read or write of above */
80 struct cgroup_subsys_state css;
82 unsigned long flags; /* "unsigned long" so bitops work */
85 * On default hierarchy:
87 * The user-configured masks can only be changed by writing to
88 * cpuset.cpus and cpuset.mems, and won't be limited by the
91 * The effective masks is the real masks that apply to the tasks
92 * in the cpuset. They may be changed if the configured masks are
93 * changed or hotplug happens.
95 * effective_mask == configured_mask & parent's effective_mask,
96 * and if it ends up empty, it will inherit the parent's mask.
101 * The user-configured masks are always the same with effective masks.
104 /* user-configured CPUs and Memory Nodes allow to tasks */
105 cpumask_var_t cpus_allowed;
106 nodemask_t mems_allowed;
108 /* effective CPUs and Memory Nodes allow to tasks */
109 cpumask_var_t effective_cpus;
110 nodemask_t effective_mems;
113 * CPUs allocated to child sub-partitions (default hierarchy only)
114 * - CPUs granted by the parent = effective_cpus U subparts_cpus
115 * - effective_cpus and subparts_cpus are mutually exclusive.
117 cpumask_var_t subparts_cpus;
120 * This is old Memory Nodes tasks took on.
122 * - top_cpuset.old_mems_allowed is initialized to mems_allowed.
123 * - A new cpuset's old_mems_allowed is initialized when some
124 * task is moved into it.
125 * - old_mems_allowed is used in cpuset_migrate_mm() when we change
126 * cpuset.mems_allowed and have tasks' nodemask updated, and
127 * then old_mems_allowed is updated to mems_allowed.
129 nodemask_t old_mems_allowed;
131 struct fmeter fmeter; /* memory_pressure filter */
134 * Tasks are being attached to this cpuset. Used to prevent
135 * zeroing cpus/mems_allowed between ->can_attach() and ->attach().
137 int attach_in_progress;
139 /* partition number for rebuild_sched_domains() */
142 /* for custom sched domain */
143 int relax_domain_level;
145 /* number of CPUs in subparts_cpus */
146 int nr_subparts_cpus;
148 /* partition root state */
149 int partition_root_state;
153 * Partition root states:
155 * 0 - not a partition root
158 #define PRS_DISABLED 0
159 #define PRS_ENABLED 1
162 * Temporary cpumasks for working with partitions that are passed among
163 * functions to avoid memory allocation in inner functions.
166 cpumask_var_t addmask, delmask; /* For partition root */
167 cpumask_var_t new_cpus; /* For update_cpumasks_hier() */
170 static inline struct cpuset *css_cs(struct cgroup_subsys_state *css)
172 return css ? container_of(css, struct cpuset, css) : NULL;
175 /* Retrieve the cpuset for a task */
176 static inline struct cpuset *task_cs(struct task_struct *task)
178 return css_cs(task_css(task, cpuset_cgrp_id));
181 static inline struct cpuset *parent_cs(struct cpuset *cs)
183 return css_cs(cs->css.parent);
187 static inline bool task_has_mempolicy(struct task_struct *task)
189 return task->mempolicy;
192 static inline bool task_has_mempolicy(struct task_struct *task)
199 /* bits in struct cpuset flags field */
206 CS_SCHED_LOAD_BALANCE,
211 /* convenient tests for these bits */
212 static inline bool is_cpuset_online(struct cpuset *cs)
214 return test_bit(CS_ONLINE, &cs->flags) && !css_is_dying(&cs->css);
217 static inline int is_cpu_exclusive(const struct cpuset *cs)
219 return test_bit(CS_CPU_EXCLUSIVE, &cs->flags);
222 static inline int is_mem_exclusive(const struct cpuset *cs)
224 return test_bit(CS_MEM_EXCLUSIVE, &cs->flags);
227 static inline int is_mem_hardwall(const struct cpuset *cs)
229 return test_bit(CS_MEM_HARDWALL, &cs->flags);
232 static inline int is_sched_load_balance(const struct cpuset *cs)
234 return test_bit(CS_SCHED_LOAD_BALANCE, &cs->flags);
237 static inline int is_memory_migrate(const struct cpuset *cs)
239 return test_bit(CS_MEMORY_MIGRATE, &cs->flags);
242 static inline int is_spread_page(const struct cpuset *cs)
244 return test_bit(CS_SPREAD_PAGE, &cs->flags);
247 static inline int is_spread_slab(const struct cpuset *cs)
249 return test_bit(CS_SPREAD_SLAB, &cs->flags);
252 static inline int is_partition_root(const struct cpuset *cs)
254 return cs->partition_root_state;
257 static struct cpuset top_cpuset = {
258 .flags = ((1 << CS_ONLINE) | (1 << CS_CPU_EXCLUSIVE) |
259 (1 << CS_MEM_EXCLUSIVE)),
260 .partition_root_state = PRS_ENABLED,
264 * cpuset_for_each_child - traverse online children of a cpuset
265 * @child_cs: loop cursor pointing to the current child
266 * @pos_css: used for iteration
267 * @parent_cs: target cpuset to walk children of
269 * Walk @child_cs through the online children of @parent_cs. Must be used
270 * with RCU read locked.
272 #define cpuset_for_each_child(child_cs, pos_css, parent_cs) \
273 css_for_each_child((pos_css), &(parent_cs)->css) \
274 if (is_cpuset_online(((child_cs) = css_cs((pos_css)))))
277 * cpuset_for_each_descendant_pre - pre-order walk of a cpuset's descendants
278 * @des_cs: loop cursor pointing to the current descendant
279 * @pos_css: used for iteration
280 * @root_cs: target cpuset to walk ancestor of
282 * Walk @des_cs through the online descendants of @root_cs. Must be used
283 * with RCU read locked. The caller may modify @pos_css by calling
284 * css_rightmost_descendant() to skip subtree. @root_cs is included in the
285 * iteration and the first node to be visited.
287 #define cpuset_for_each_descendant_pre(des_cs, pos_css, root_cs) \
288 css_for_each_descendant_pre((pos_css), &(root_cs)->css) \
289 if (is_cpuset_online(((des_cs) = css_cs((pos_css)))))
292 * There are two global locks guarding cpuset structures - cpuset_mutex and
293 * callback_lock. We also require taking task_lock() when dereferencing a
294 * task's cpuset pointer. See "The task_lock() exception", at the end of this
297 * A task must hold both locks to modify cpusets. If a task holds
298 * cpuset_mutex, then it blocks others wanting that mutex, ensuring that it
299 * is the only task able to also acquire callback_lock and be able to
300 * modify cpusets. It can perform various checks on the cpuset structure
301 * first, knowing nothing will change. It can also allocate memory while
302 * just holding cpuset_mutex. While it is performing these checks, various
303 * callback routines can briefly acquire callback_lock to query cpusets.
304 * Once it is ready to make the changes, it takes callback_lock, blocking
307 * Calls to the kernel memory allocator can not be made while holding
308 * callback_lock, as that would risk double tripping on callback_lock
309 * from one of the callbacks into the cpuset code from within
312 * If a task is only holding callback_lock, then it has read-only
315 * Now, the task_struct fields mems_allowed and mempolicy may be changed
316 * by other task, we use alloc_lock in the task_struct fields to protect
319 * The cpuset_common_file_read() handlers only hold callback_lock across
320 * small pieces of code, such as when reading out possibly multi-word
321 * cpumasks and nodemasks.
323 * Accessing a task's cpuset should be done in accordance with the
324 * guidelines for accessing subsystem state in kernel/cgroup.c
327 static DEFINE_MUTEX(cpuset_mutex);
328 static DEFINE_SPINLOCK(callback_lock);
330 static struct workqueue_struct *cpuset_migrate_mm_wq;
333 * CPU / memory hotplug is handled asynchronously.
335 static void cpuset_hotplug_workfn(struct work_struct *work);
336 static DECLARE_WORK(cpuset_hotplug_work, cpuset_hotplug_workfn);
338 static DECLARE_WAIT_QUEUE_HEAD(cpuset_attach_wq);
341 * Cgroup v2 behavior is used when on default hierarchy or the
342 * cgroup_v2_mode flag is set.
344 static inline bool is_in_v2_mode(void)
346 return cgroup_subsys_on_dfl(cpuset_cgrp_subsys) ||
347 (cpuset_cgrp_subsys.root->flags & CGRP_ROOT_CPUSET_V2_MODE);
351 * This is ugly, but preserves the userspace API for existing cpuset
352 * users. If someone tries to mount the "cpuset" filesystem, we
353 * silently switch it to mount "cgroup" instead
355 static struct dentry *cpuset_mount(struct file_system_type *fs_type,
356 int flags, const char *unused_dev_name, void *data)
358 struct file_system_type *cgroup_fs = get_fs_type("cgroup");
359 struct dentry *ret = ERR_PTR(-ENODEV);
363 "release_agent=/sbin/cpuset_release_agent";
364 ret = cgroup_fs->mount(cgroup_fs, flags,
365 unused_dev_name, mountopts);
366 put_filesystem(cgroup_fs);
371 static struct file_system_type cpuset_fs_type = {
373 .mount = cpuset_mount,
377 * Return in pmask the portion of a cpusets's cpus_allowed that
378 * are online. If none are online, walk up the cpuset hierarchy
379 * until we find one that does have some online cpus.
381 * One way or another, we guarantee to return some non-empty subset
382 * of cpu_online_mask.
384 * Call with callback_lock or cpuset_mutex held.
386 static void guarantee_online_cpus(struct cpuset *cs, struct cpumask *pmask)
388 while (!cpumask_intersects(cs->effective_cpus, cpu_online_mask)) {
392 * The top cpuset doesn't have any online cpu as a
393 * consequence of a race between cpuset_hotplug_work
394 * and cpu hotplug notifier. But we know the top
395 * cpuset's effective_cpus is on its way to to be
396 * identical to cpu_online_mask.
398 cpumask_copy(pmask, cpu_online_mask);
402 cpumask_and(pmask, cs->effective_cpus, cpu_online_mask);
406 * Return in *pmask the portion of a cpusets's mems_allowed that
407 * are online, with memory. If none are online with memory, walk
408 * up the cpuset hierarchy until we find one that does have some
409 * online mems. The top cpuset always has some mems online.
411 * One way or another, we guarantee to return some non-empty subset
412 * of node_states[N_MEMORY].
414 * Call with callback_lock or cpuset_mutex held.
416 static void guarantee_online_mems(struct cpuset *cs, nodemask_t *pmask)
418 while (!nodes_intersects(cs->effective_mems, node_states[N_MEMORY]))
420 nodes_and(*pmask, cs->effective_mems, node_states[N_MEMORY]);
424 * update task's spread flag if cpuset's page/slab spread flag is set
426 * Call with callback_lock or cpuset_mutex held.
428 static void cpuset_update_task_spread_flag(struct cpuset *cs,
429 struct task_struct *tsk)
431 if (is_spread_page(cs))
432 task_set_spread_page(tsk);
434 task_clear_spread_page(tsk);
436 if (is_spread_slab(cs))
437 task_set_spread_slab(tsk);
439 task_clear_spread_slab(tsk);
443 * is_cpuset_subset(p, q) - Is cpuset p a subset of cpuset q?
445 * One cpuset is a subset of another if all its allowed CPUs and
446 * Memory Nodes are a subset of the other, and its exclusive flags
447 * are only set if the other's are set. Call holding cpuset_mutex.
450 static int is_cpuset_subset(const struct cpuset *p, const struct cpuset *q)
452 return cpumask_subset(p->cpus_allowed, q->cpus_allowed) &&
453 nodes_subset(p->mems_allowed, q->mems_allowed) &&
454 is_cpu_exclusive(p) <= is_cpu_exclusive(q) &&
455 is_mem_exclusive(p) <= is_mem_exclusive(q);
459 * alloc_trial_cpuset - allocate a trial cpuset
460 * @cs: the cpuset that the trial cpuset duplicates
462 static struct cpuset *alloc_trial_cpuset(struct cpuset *cs)
464 struct cpuset *trial;
466 trial = kmemdup(cs, sizeof(*cs), GFP_KERNEL);
470 if (!alloc_cpumask_var(&trial->cpus_allowed, GFP_KERNEL))
472 if (!alloc_cpumask_var(&trial->effective_cpus, GFP_KERNEL))
475 cpumask_copy(trial->cpus_allowed, cs->cpus_allowed);
476 cpumask_copy(trial->effective_cpus, cs->effective_cpus);
480 free_cpumask_var(trial->cpus_allowed);
487 * free_trial_cpuset - free the trial cpuset
488 * @trial: the trial cpuset to be freed
490 static void free_trial_cpuset(struct cpuset *trial)
492 free_cpumask_var(trial->effective_cpus);
493 free_cpumask_var(trial->cpus_allowed);
498 * validate_change() - Used to validate that any proposed cpuset change
499 * follows the structural rules for cpusets.
501 * If we replaced the flag and mask values of the current cpuset
502 * (cur) with those values in the trial cpuset (trial), would
503 * our various subset and exclusive rules still be valid? Presumes
506 * 'cur' is the address of an actual, in-use cpuset. Operations
507 * such as list traversal that depend on the actual address of the
508 * cpuset in the list must use cur below, not trial.
510 * 'trial' is the address of bulk structure copy of cur, with
511 * perhaps one or more of the fields cpus_allowed, mems_allowed,
512 * or flags changed to new, trial values.
514 * Return 0 if valid, -errno if not.
517 static int validate_change(struct cpuset *cur, struct cpuset *trial)
519 struct cgroup_subsys_state *css;
520 struct cpuset *c, *par;
525 /* Each of our child cpusets must be a subset of us */
527 cpuset_for_each_child(c, css, cur)
528 if (!is_cpuset_subset(c, trial))
531 /* Remaining checks don't apply to root cpuset */
533 if (cur == &top_cpuset)
536 par = parent_cs(cur);
538 /* On legacy hiearchy, we must be a subset of our parent cpuset. */
540 if (!is_in_v2_mode() && !is_cpuset_subset(trial, par))
544 * If either I or some sibling (!= me) is exclusive, we can't
548 cpuset_for_each_child(c, css, par) {
549 if ((is_cpu_exclusive(trial) || is_cpu_exclusive(c)) &&
551 cpumask_intersects(trial->cpus_allowed, c->cpus_allowed))
553 if ((is_mem_exclusive(trial) || is_mem_exclusive(c)) &&
555 nodes_intersects(trial->mems_allowed, c->mems_allowed))
560 * Cpusets with tasks - existing or newly being attached - can't
561 * be changed to have empty cpus_allowed or mems_allowed.
564 if ((cgroup_is_populated(cur->css.cgroup) || cur->attach_in_progress)) {
565 if (!cpumask_empty(cur->cpus_allowed) &&
566 cpumask_empty(trial->cpus_allowed))
568 if (!nodes_empty(cur->mems_allowed) &&
569 nodes_empty(trial->mems_allowed))
574 * We can't shrink if we won't have enough room for SCHED_DEADLINE
578 if (is_cpu_exclusive(cur) &&
579 !cpuset_cpumask_can_shrink(cur->cpus_allowed,
580 trial->cpus_allowed))
591 * Helper routine for generate_sched_domains().
592 * Do cpusets a, b have overlapping effective cpus_allowed masks?
594 static int cpusets_overlap(struct cpuset *a, struct cpuset *b)
596 return cpumask_intersects(a->effective_cpus, b->effective_cpus);
600 update_domain_attr(struct sched_domain_attr *dattr, struct cpuset *c)
602 if (dattr->relax_domain_level < c->relax_domain_level)
603 dattr->relax_domain_level = c->relax_domain_level;
607 static void update_domain_attr_tree(struct sched_domain_attr *dattr,
608 struct cpuset *root_cs)
611 struct cgroup_subsys_state *pos_css;
614 cpuset_for_each_descendant_pre(cp, pos_css, root_cs) {
615 /* skip the whole subtree if @cp doesn't have any CPU */
616 if (cpumask_empty(cp->cpus_allowed)) {
617 pos_css = css_rightmost_descendant(pos_css);
621 if (is_sched_load_balance(cp))
622 update_domain_attr(dattr, cp);
627 /* Must be called with cpuset_mutex held. */
628 static inline int nr_cpusets(void)
630 /* jump label reference count + the top-level cpuset */
631 return static_key_count(&cpusets_enabled_key.key) + 1;
635 * generate_sched_domains()
637 * This function builds a partial partition of the systems CPUs
638 * A 'partial partition' is a set of non-overlapping subsets whose
639 * union is a subset of that set.
640 * The output of this function needs to be passed to kernel/sched/core.c
641 * partition_sched_domains() routine, which will rebuild the scheduler's
642 * load balancing domains (sched domains) as specified by that partial
645 * See "What is sched_load_balance" in Documentation/cgroup-v1/cpusets.txt
646 * for a background explanation of this.
648 * Does not return errors, on the theory that the callers of this
649 * routine would rather not worry about failures to rebuild sched
650 * domains when operating in the severe memory shortage situations
651 * that could cause allocation failures below.
653 * Must be called with cpuset_mutex held.
655 * The three key local variables below are:
656 * q - a linked-list queue of cpuset pointers, used to implement a
657 * top-down scan of all cpusets. This scan loads a pointer
658 * to each cpuset marked is_sched_load_balance into the
659 * array 'csa'. For our purposes, rebuilding the schedulers
660 * sched domains, we can ignore !is_sched_load_balance cpusets.
661 * csa - (for CpuSet Array) Array of pointers to all the cpusets
662 * that need to be load balanced, for convenient iterative
663 * access by the subsequent code that finds the best partition,
664 * i.e the set of domains (subsets) of CPUs such that the
665 * cpus_allowed of every cpuset marked is_sched_load_balance
666 * is a subset of one of these domains, while there are as
667 * many such domains as possible, each as small as possible.
668 * doms - Conversion of 'csa' to an array of cpumasks, for passing to
669 * the kernel/sched/core.c routine partition_sched_domains() in a
670 * convenient format, that can be easily compared to the prior
671 * value to determine what partition elements (sched domains)
672 * were changed (added or removed.)
674 * Finding the best partition (set of domains):
675 * The triple nested loops below over i, j, k scan over the
676 * load balanced cpusets (using the array of cpuset pointers in
677 * csa[]) looking for pairs of cpusets that have overlapping
678 * cpus_allowed, but which don't have the same 'pn' partition
679 * number and gives them in the same partition number. It keeps
680 * looping on the 'restart' label until it can no longer find
683 * The union of the cpus_allowed masks from the set of
684 * all cpusets having the same 'pn' value then form the one
685 * element of the partition (one sched domain) to be passed to
686 * partition_sched_domains().
688 static int generate_sched_domains(cpumask_var_t **domains,
689 struct sched_domain_attr **attributes)
691 struct cpuset *cp; /* scans q */
692 struct cpuset **csa; /* array of all cpuset ptrs */
693 int csn; /* how many cpuset ptrs in csa so far */
694 int i, j, k; /* indices for partition finding loops */
695 cpumask_var_t *doms; /* resulting partition; i.e. sched domains */
696 struct sched_domain_attr *dattr; /* attributes for custom domains */
697 int ndoms = 0; /* number of sched domains in result */
698 int nslot; /* next empty doms[] struct cpumask slot */
699 struct cgroup_subsys_state *pos_css;
705 /* Special case for the 99% of systems with one, full, sched domain */
706 if (is_sched_load_balance(&top_cpuset)) {
708 doms = alloc_sched_domains(ndoms);
712 dattr = kmalloc(sizeof(struct sched_domain_attr), GFP_KERNEL);
714 *dattr = SD_ATTR_INIT;
715 update_domain_attr_tree(dattr, &top_cpuset);
717 cpumask_and(doms[0], top_cpuset.effective_cpus,
718 housekeeping_cpumask(HK_FLAG_DOMAIN));
723 csa = kmalloc_array(nr_cpusets(), sizeof(cp), GFP_KERNEL);
729 cpuset_for_each_descendant_pre(cp, pos_css, &top_cpuset) {
730 if (cp == &top_cpuset)
733 * Continue traversing beyond @cp iff @cp has some CPUs and
734 * isn't load balancing. The former is obvious. The
735 * latter: All child cpusets contain a subset of the
736 * parent's cpus, so just skip them, and then we call
737 * update_domain_attr_tree() to calc relax_domain_level of
738 * the corresponding sched domain.
740 if (!cpumask_empty(cp->cpus_allowed) &&
741 !(is_sched_load_balance(cp) &&
742 cpumask_intersects(cp->cpus_allowed,
743 housekeeping_cpumask(HK_FLAG_DOMAIN))))
746 if (is_sched_load_balance(cp))
749 /* skip @cp's subtree */
750 pos_css = css_rightmost_descendant(pos_css);
754 for (i = 0; i < csn; i++)
759 /* Find the best partition (set of sched domains) */
760 for (i = 0; i < csn; i++) {
761 struct cpuset *a = csa[i];
764 for (j = 0; j < csn; j++) {
765 struct cpuset *b = csa[j];
768 if (apn != bpn && cpusets_overlap(a, b)) {
769 for (k = 0; k < csn; k++) {
770 struct cpuset *c = csa[k];
775 ndoms--; /* one less element */
782 * Now we know how many domains to create.
783 * Convert <csn, csa> to <ndoms, doms> and populate cpu masks.
785 doms = alloc_sched_domains(ndoms);
790 * The rest of the code, including the scheduler, can deal with
791 * dattr==NULL case. No need to abort if alloc fails.
793 dattr = kmalloc_array(ndoms, sizeof(struct sched_domain_attr),
796 for (nslot = 0, i = 0; i < csn; i++) {
797 struct cpuset *a = csa[i];
802 /* Skip completed partitions */
808 if (nslot == ndoms) {
809 static int warnings = 10;
811 pr_warn("rebuild_sched_domains confused: nslot %d, ndoms %d, csn %d, i %d, apn %d\n",
812 nslot, ndoms, csn, i, apn);
820 *(dattr + nslot) = SD_ATTR_INIT;
821 for (j = i; j < csn; j++) {
822 struct cpuset *b = csa[j];
825 cpumask_or(dp, dp, b->effective_cpus);
826 cpumask_and(dp, dp, housekeeping_cpumask(HK_FLAG_DOMAIN));
828 update_domain_attr_tree(dattr + nslot, b);
830 /* Done with this partition */
836 BUG_ON(nslot != ndoms);
842 * Fallback to the default domain if kmalloc() failed.
843 * See comments in partition_sched_domains().
854 * Rebuild scheduler domains.
856 * If the flag 'sched_load_balance' of any cpuset with non-empty
857 * 'cpus' changes, or if the 'cpus' allowed changes in any cpuset
858 * which has that flag enabled, or if any cpuset with a non-empty
859 * 'cpus' is removed, then call this routine to rebuild the
860 * scheduler's dynamic sched domains.
862 * Call with cpuset_mutex held. Takes get_online_cpus().
864 static void rebuild_sched_domains_locked(void)
866 struct sched_domain_attr *attr;
870 lockdep_assert_held(&cpuset_mutex);
874 * We have raced with CPU hotplug. Don't do anything to avoid
875 * passing doms with offlined cpu to partition_sched_domains().
876 * Anyways, hotplug work item will rebuild sched domains.
878 if (!cpumask_equal(top_cpuset.effective_cpus, cpu_active_mask))
881 /* Generate domain masks and attrs */
882 ndoms = generate_sched_domains(&doms, &attr);
884 /* Have scheduler rebuild the domains */
885 partition_sched_domains(ndoms, doms, attr);
889 #else /* !CONFIG_SMP */
890 static void rebuild_sched_domains_locked(void)
893 #endif /* CONFIG_SMP */
895 void rebuild_sched_domains(void)
897 mutex_lock(&cpuset_mutex);
898 rebuild_sched_domains_locked();
899 mutex_unlock(&cpuset_mutex);
903 * update_tasks_cpumask - Update the cpumasks of tasks in the cpuset.
904 * @cs: the cpuset in which each task's cpus_allowed mask needs to be changed
906 * Iterate through each task of @cs updating its cpus_allowed to the
907 * effective cpuset's. As this function is called with cpuset_mutex held,
908 * cpuset membership stays stable.
910 static void update_tasks_cpumask(struct cpuset *cs)
912 struct css_task_iter it;
913 struct task_struct *task;
915 css_task_iter_start(&cs->css, 0, &it);
916 while ((task = css_task_iter_next(&it)))
917 set_cpus_allowed_ptr(task, cs->effective_cpus);
918 css_task_iter_end(&it);
922 * update_cpumasks_hier - Update effective cpumasks and tasks in the subtree
923 * @cs: the cpuset to consider
924 * @new_cpus: temp variable for calculating new effective_cpus
926 * When congifured cpumask is changed, the effective cpumasks of this cpuset
927 * and all its descendants need to be updated.
929 * On legacy hierachy, effective_cpus will be the same with cpu_allowed.
931 * Called with cpuset_mutex held
933 static void update_cpumasks_hier(struct cpuset *cs, struct cpumask *new_cpus)
936 struct cgroup_subsys_state *pos_css;
937 bool need_rebuild_sched_domains = false;
940 cpuset_for_each_descendant_pre(cp, pos_css, cs) {
941 struct cpuset *parent = parent_cs(cp);
943 cpumask_and(new_cpus, cp->cpus_allowed, parent->effective_cpus);
946 * If it becomes empty, inherit the effective mask of the
947 * parent, which is guaranteed to have some CPUs.
949 if (is_in_v2_mode() && cpumask_empty(new_cpus))
950 cpumask_copy(new_cpus, parent->effective_cpus);
952 /* Skip the whole subtree if the cpumask remains the same. */
953 if (cpumask_equal(new_cpus, cp->effective_cpus)) {
954 pos_css = css_rightmost_descendant(pos_css);
958 if (!css_tryget_online(&cp->css))
962 spin_lock_irq(&callback_lock);
963 cpumask_copy(cp->effective_cpus, new_cpus);
964 spin_unlock_irq(&callback_lock);
966 WARN_ON(!is_in_v2_mode() &&
967 !cpumask_equal(cp->cpus_allowed, cp->effective_cpus));
969 update_tasks_cpumask(cp);
972 * If the effective cpumask of any non-empty cpuset is changed,
973 * we need to rebuild sched domains.
975 if (!cpumask_empty(cp->cpus_allowed) &&
976 is_sched_load_balance(cp))
977 need_rebuild_sched_domains = true;
984 if (need_rebuild_sched_domains)
985 rebuild_sched_domains_locked();
989 * update_cpumask - update the cpus_allowed mask of a cpuset and all tasks in it
990 * @cs: the cpuset to consider
991 * @trialcs: trial cpuset
992 * @buf: buffer of cpu numbers written to this cpuset
994 static int update_cpumask(struct cpuset *cs, struct cpuset *trialcs,
999 /* top_cpuset.cpus_allowed tracks cpu_online_mask; it's read-only */
1000 if (cs == &top_cpuset)
1004 * An empty cpus_allowed is ok only if the cpuset has no tasks.
1005 * Since cpulist_parse() fails on an empty mask, we special case
1006 * that parsing. The validate_change() call ensures that cpusets
1007 * with tasks have cpus.
1010 cpumask_clear(trialcs->cpus_allowed);
1012 retval = cpulist_parse(buf, trialcs->cpus_allowed);
1016 if (!cpumask_subset(trialcs->cpus_allowed,
1017 top_cpuset.cpus_allowed))
1021 /* Nothing to do if the cpus didn't change */
1022 if (cpumask_equal(cs->cpus_allowed, trialcs->cpus_allowed))
1025 retval = validate_change(cs, trialcs);
1029 spin_lock_irq(&callback_lock);
1030 cpumask_copy(cs->cpus_allowed, trialcs->cpus_allowed);
1031 spin_unlock_irq(&callback_lock);
1033 /* use trialcs->cpus_allowed as a temp variable */
1034 update_cpumasks_hier(cs, trialcs->cpus_allowed);
1039 * Migrate memory region from one set of nodes to another. This is
1040 * performed asynchronously as it can be called from process migration path
1041 * holding locks involved in process management. All mm migrations are
1042 * performed in the queued order and can be waited for by flushing
1043 * cpuset_migrate_mm_wq.
1046 struct cpuset_migrate_mm_work {
1047 struct work_struct work;
1048 struct mm_struct *mm;
1053 static void cpuset_migrate_mm_workfn(struct work_struct *work)
1055 struct cpuset_migrate_mm_work *mwork =
1056 container_of(work, struct cpuset_migrate_mm_work, work);
1058 /* on a wq worker, no need to worry about %current's mems_allowed */
1059 do_migrate_pages(mwork->mm, &mwork->from, &mwork->to, MPOL_MF_MOVE_ALL);
1064 static void cpuset_migrate_mm(struct mm_struct *mm, const nodemask_t *from,
1065 const nodemask_t *to)
1067 struct cpuset_migrate_mm_work *mwork;
1069 mwork = kzalloc(sizeof(*mwork), GFP_KERNEL);
1072 mwork->from = *from;
1074 INIT_WORK(&mwork->work, cpuset_migrate_mm_workfn);
1075 queue_work(cpuset_migrate_mm_wq, &mwork->work);
1081 static void cpuset_post_attach(void)
1083 flush_workqueue(cpuset_migrate_mm_wq);
1087 * cpuset_change_task_nodemask - change task's mems_allowed and mempolicy
1088 * @tsk: the task to change
1089 * @newmems: new nodes that the task will be set
1091 * We use the mems_allowed_seq seqlock to safely update both tsk->mems_allowed
1092 * and rebind an eventual tasks' mempolicy. If the task is allocating in
1093 * parallel, it might temporarily see an empty intersection, which results in
1094 * a seqlock check and retry before OOM or allocation failure.
1096 static void cpuset_change_task_nodemask(struct task_struct *tsk,
1097 nodemask_t *newmems)
1101 local_irq_disable();
1102 write_seqcount_begin(&tsk->mems_allowed_seq);
1104 nodes_or(tsk->mems_allowed, tsk->mems_allowed, *newmems);
1105 mpol_rebind_task(tsk, newmems);
1106 tsk->mems_allowed = *newmems;
1108 write_seqcount_end(&tsk->mems_allowed_seq);
1114 static void *cpuset_being_rebound;
1117 * update_tasks_nodemask - Update the nodemasks of tasks in the cpuset.
1118 * @cs: the cpuset in which each task's mems_allowed mask needs to be changed
1120 * Iterate through each task of @cs updating its mems_allowed to the
1121 * effective cpuset's. As this function is called with cpuset_mutex held,
1122 * cpuset membership stays stable.
1124 static void update_tasks_nodemask(struct cpuset *cs)
1126 static nodemask_t newmems; /* protected by cpuset_mutex */
1127 struct css_task_iter it;
1128 struct task_struct *task;
1130 cpuset_being_rebound = cs; /* causes mpol_dup() rebind */
1132 guarantee_online_mems(cs, &newmems);
1135 * The mpol_rebind_mm() call takes mmap_sem, which we couldn't
1136 * take while holding tasklist_lock. Forks can happen - the
1137 * mpol_dup() cpuset_being_rebound check will catch such forks,
1138 * and rebind their vma mempolicies too. Because we still hold
1139 * the global cpuset_mutex, we know that no other rebind effort
1140 * will be contending for the global variable cpuset_being_rebound.
1141 * It's ok if we rebind the same mm twice; mpol_rebind_mm()
1142 * is idempotent. Also migrate pages in each mm to new nodes.
1144 css_task_iter_start(&cs->css, 0, &it);
1145 while ((task = css_task_iter_next(&it))) {
1146 struct mm_struct *mm;
1149 cpuset_change_task_nodemask(task, &newmems);
1151 mm = get_task_mm(task);
1155 migrate = is_memory_migrate(cs);
1157 mpol_rebind_mm(mm, &cs->mems_allowed);
1159 cpuset_migrate_mm(mm, &cs->old_mems_allowed, &newmems);
1163 css_task_iter_end(&it);
1166 * All the tasks' nodemasks have been updated, update
1167 * cs->old_mems_allowed.
1169 cs->old_mems_allowed = newmems;
1171 /* We're done rebinding vmas to this cpuset's new mems_allowed. */
1172 cpuset_being_rebound = NULL;
1176 * update_nodemasks_hier - Update effective nodemasks and tasks in the subtree
1177 * @cs: the cpuset to consider
1178 * @new_mems: a temp variable for calculating new effective_mems
1180 * When configured nodemask is changed, the effective nodemasks of this cpuset
1181 * and all its descendants need to be updated.
1183 * On legacy hiearchy, effective_mems will be the same with mems_allowed.
1185 * Called with cpuset_mutex held
1187 static void update_nodemasks_hier(struct cpuset *cs, nodemask_t *new_mems)
1190 struct cgroup_subsys_state *pos_css;
1193 cpuset_for_each_descendant_pre(cp, pos_css, cs) {
1194 struct cpuset *parent = parent_cs(cp);
1196 nodes_and(*new_mems, cp->mems_allowed, parent->effective_mems);
1199 * If it becomes empty, inherit the effective mask of the
1200 * parent, which is guaranteed to have some MEMs.
1202 if (is_in_v2_mode() && nodes_empty(*new_mems))
1203 *new_mems = parent->effective_mems;
1205 /* Skip the whole subtree if the nodemask remains the same. */
1206 if (nodes_equal(*new_mems, cp->effective_mems)) {
1207 pos_css = css_rightmost_descendant(pos_css);
1211 if (!css_tryget_online(&cp->css))
1215 spin_lock_irq(&callback_lock);
1216 cp->effective_mems = *new_mems;
1217 spin_unlock_irq(&callback_lock);
1219 WARN_ON(!is_in_v2_mode() &&
1220 !nodes_equal(cp->mems_allowed, cp->effective_mems));
1222 update_tasks_nodemask(cp);
1231 * Handle user request to change the 'mems' memory placement
1232 * of a cpuset. Needs to validate the request, update the
1233 * cpusets mems_allowed, and for each task in the cpuset,
1234 * update mems_allowed and rebind task's mempolicy and any vma
1235 * mempolicies and if the cpuset is marked 'memory_migrate',
1236 * migrate the tasks pages to the new memory.
1238 * Call with cpuset_mutex held. May take callback_lock during call.
1239 * Will take tasklist_lock, scan tasklist for tasks in cpuset cs,
1240 * lock each such tasks mm->mmap_sem, scan its vma's and rebind
1241 * their mempolicies to the cpusets new mems_allowed.
1243 static int update_nodemask(struct cpuset *cs, struct cpuset *trialcs,
1249 * top_cpuset.mems_allowed tracks node_stats[N_MEMORY];
1252 if (cs == &top_cpuset) {
1258 * An empty mems_allowed is ok iff there are no tasks in the cpuset.
1259 * Since nodelist_parse() fails on an empty mask, we special case
1260 * that parsing. The validate_change() call ensures that cpusets
1261 * with tasks have memory.
1264 nodes_clear(trialcs->mems_allowed);
1266 retval = nodelist_parse(buf, trialcs->mems_allowed);
1270 if (!nodes_subset(trialcs->mems_allowed,
1271 top_cpuset.mems_allowed)) {
1277 if (nodes_equal(cs->mems_allowed, trialcs->mems_allowed)) {
1278 retval = 0; /* Too easy - nothing to do */
1281 retval = validate_change(cs, trialcs);
1285 spin_lock_irq(&callback_lock);
1286 cs->mems_allowed = trialcs->mems_allowed;
1287 spin_unlock_irq(&callback_lock);
1289 /* use trialcs->mems_allowed as a temp variable */
1290 update_nodemasks_hier(cs, &trialcs->mems_allowed);
1295 bool current_cpuset_is_being_rebound(void)
1300 ret = task_cs(current) == cpuset_being_rebound;
1306 static int update_relax_domain_level(struct cpuset *cs, s64 val)
1309 if (val < -1 || val >= sched_domain_level_max)
1313 if (val != cs->relax_domain_level) {
1314 cs->relax_domain_level = val;
1315 if (!cpumask_empty(cs->cpus_allowed) &&
1316 is_sched_load_balance(cs))
1317 rebuild_sched_domains_locked();
1324 * update_tasks_flags - update the spread flags of tasks in the cpuset.
1325 * @cs: the cpuset in which each task's spread flags needs to be changed
1327 * Iterate through each task of @cs updating its spread flags. As this
1328 * function is called with cpuset_mutex held, cpuset membership stays
1331 static void update_tasks_flags(struct cpuset *cs)
1333 struct css_task_iter it;
1334 struct task_struct *task;
1336 css_task_iter_start(&cs->css, 0, &it);
1337 while ((task = css_task_iter_next(&it)))
1338 cpuset_update_task_spread_flag(cs, task);
1339 css_task_iter_end(&it);
1343 * update_flag - read a 0 or a 1 in a file and update associated flag
1344 * bit: the bit to update (see cpuset_flagbits_t)
1345 * cs: the cpuset to update
1346 * turning_on: whether the flag is being set or cleared
1348 * Call with cpuset_mutex held.
1351 static int update_flag(cpuset_flagbits_t bit, struct cpuset *cs,
1354 struct cpuset *trialcs;
1355 int balance_flag_changed;
1356 int spread_flag_changed;
1359 trialcs = alloc_trial_cpuset(cs);
1364 set_bit(bit, &trialcs->flags);
1366 clear_bit(bit, &trialcs->flags);
1368 err = validate_change(cs, trialcs);
1372 balance_flag_changed = (is_sched_load_balance(cs) !=
1373 is_sched_load_balance(trialcs));
1375 spread_flag_changed = ((is_spread_slab(cs) != is_spread_slab(trialcs))
1376 || (is_spread_page(cs) != is_spread_page(trialcs)));
1378 spin_lock_irq(&callback_lock);
1379 cs->flags = trialcs->flags;
1380 spin_unlock_irq(&callback_lock);
1382 if (!cpumask_empty(trialcs->cpus_allowed) && balance_flag_changed)
1383 rebuild_sched_domains_locked();
1385 if (spread_flag_changed)
1386 update_tasks_flags(cs);
1388 free_trial_cpuset(trialcs);
1393 * Frequency meter - How fast is some event occurring?
1395 * These routines manage a digitally filtered, constant time based,
1396 * event frequency meter. There are four routines:
1397 * fmeter_init() - initialize a frequency meter.
1398 * fmeter_markevent() - called each time the event happens.
1399 * fmeter_getrate() - returns the recent rate of such events.
1400 * fmeter_update() - internal routine used to update fmeter.
1402 * A common data structure is passed to each of these routines,
1403 * which is used to keep track of the state required to manage the
1404 * frequency meter and its digital filter.
1406 * The filter works on the number of events marked per unit time.
1407 * The filter is single-pole low-pass recursive (IIR). The time unit
1408 * is 1 second. Arithmetic is done using 32-bit integers scaled to
1409 * simulate 3 decimal digits of precision (multiplied by 1000).
1411 * With an FM_COEF of 933, and a time base of 1 second, the filter
1412 * has a half-life of 10 seconds, meaning that if the events quit
1413 * happening, then the rate returned from the fmeter_getrate()
1414 * will be cut in half each 10 seconds, until it converges to zero.
1416 * It is not worth doing a real infinitely recursive filter. If more
1417 * than FM_MAXTICKS ticks have elapsed since the last filter event,
1418 * just compute FM_MAXTICKS ticks worth, by which point the level
1421 * Limit the count of unprocessed events to FM_MAXCNT, so as to avoid
1422 * arithmetic overflow in the fmeter_update() routine.
1424 * Given the simple 32 bit integer arithmetic used, this meter works
1425 * best for reporting rates between one per millisecond (msec) and
1426 * one per 32 (approx) seconds. At constant rates faster than one
1427 * per msec it maxes out at values just under 1,000,000. At constant
1428 * rates between one per msec, and one per second it will stabilize
1429 * to a value N*1000, where N is the rate of events per second.
1430 * At constant rates between one per second and one per 32 seconds,
1431 * it will be choppy, moving up on the seconds that have an event,
1432 * and then decaying until the next event. At rates slower than
1433 * about one in 32 seconds, it decays all the way back to zero between
1437 #define FM_COEF 933 /* coefficient for half-life of 10 secs */
1438 #define FM_MAXTICKS ((u32)99) /* useless computing more ticks than this */
1439 #define FM_MAXCNT 1000000 /* limit cnt to avoid overflow */
1440 #define FM_SCALE 1000 /* faux fixed point scale */
1442 /* Initialize a frequency meter */
1443 static void fmeter_init(struct fmeter *fmp)
1448 spin_lock_init(&fmp->lock);
1451 /* Internal meter update - process cnt events and update value */
1452 static void fmeter_update(struct fmeter *fmp)
1457 now = ktime_get_seconds();
1458 ticks = now - fmp->time;
1463 ticks = min(FM_MAXTICKS, ticks);
1465 fmp->val = (FM_COEF * fmp->val) / FM_SCALE;
1468 fmp->val += ((FM_SCALE - FM_COEF) * fmp->cnt) / FM_SCALE;
1472 /* Process any previous ticks, then bump cnt by one (times scale). */
1473 static void fmeter_markevent(struct fmeter *fmp)
1475 spin_lock(&fmp->lock);
1477 fmp->cnt = min(FM_MAXCNT, fmp->cnt + FM_SCALE);
1478 spin_unlock(&fmp->lock);
1481 /* Process any previous ticks, then return current value. */
1482 static int fmeter_getrate(struct fmeter *fmp)
1486 spin_lock(&fmp->lock);
1489 spin_unlock(&fmp->lock);
1493 static struct cpuset *cpuset_attach_old_cs;
1495 /* Called by cgroups to determine if a cpuset is usable; cpuset_mutex held */
1496 static int cpuset_can_attach(struct cgroup_taskset *tset)
1498 struct cgroup_subsys_state *css;
1500 struct task_struct *task;
1503 /* used later by cpuset_attach() */
1504 cpuset_attach_old_cs = task_cs(cgroup_taskset_first(tset, &css));
1507 mutex_lock(&cpuset_mutex);
1509 /* allow moving tasks into an empty cpuset if on default hierarchy */
1511 if (!is_in_v2_mode() &&
1512 (cpumask_empty(cs->cpus_allowed) || nodes_empty(cs->mems_allowed)))
1515 cgroup_taskset_for_each(task, css, tset) {
1516 ret = task_can_attach(task, cs->cpus_allowed);
1519 ret = security_task_setscheduler(task);
1525 * Mark attach is in progress. This makes validate_change() fail
1526 * changes which zero cpus/mems_allowed.
1528 cs->attach_in_progress++;
1531 mutex_unlock(&cpuset_mutex);
1535 static void cpuset_cancel_attach(struct cgroup_taskset *tset)
1537 struct cgroup_subsys_state *css;
1540 cgroup_taskset_first(tset, &css);
1543 mutex_lock(&cpuset_mutex);
1544 css_cs(css)->attach_in_progress--;
1545 mutex_unlock(&cpuset_mutex);
1549 * Protected by cpuset_mutex. cpus_attach is used only by cpuset_attach()
1550 * but we can't allocate it dynamically there. Define it global and
1551 * allocate from cpuset_init().
1553 static cpumask_var_t cpus_attach;
1555 static void cpuset_attach(struct cgroup_taskset *tset)
1557 /* static buf protected by cpuset_mutex */
1558 static nodemask_t cpuset_attach_nodemask_to;
1559 struct task_struct *task;
1560 struct task_struct *leader;
1561 struct cgroup_subsys_state *css;
1563 struct cpuset *oldcs = cpuset_attach_old_cs;
1565 cgroup_taskset_first(tset, &css);
1568 mutex_lock(&cpuset_mutex);
1570 /* prepare for attach */
1571 if (cs == &top_cpuset)
1572 cpumask_copy(cpus_attach, cpu_possible_mask);
1574 guarantee_online_cpus(cs, cpus_attach);
1576 guarantee_online_mems(cs, &cpuset_attach_nodemask_to);
1578 cgroup_taskset_for_each(task, css, tset) {
1580 * can_attach beforehand should guarantee that this doesn't
1581 * fail. TODO: have a better way to handle failure here
1583 WARN_ON_ONCE(set_cpus_allowed_ptr(task, cpus_attach));
1585 cpuset_change_task_nodemask(task, &cpuset_attach_nodemask_to);
1586 cpuset_update_task_spread_flag(cs, task);
1590 * Change mm for all threadgroup leaders. This is expensive and may
1591 * sleep and should be moved outside migration path proper.
1593 cpuset_attach_nodemask_to = cs->effective_mems;
1594 cgroup_taskset_for_each_leader(leader, css, tset) {
1595 struct mm_struct *mm = get_task_mm(leader);
1598 mpol_rebind_mm(mm, &cpuset_attach_nodemask_to);
1601 * old_mems_allowed is the same with mems_allowed
1602 * here, except if this task is being moved
1603 * automatically due to hotplug. In that case
1604 * @mems_allowed has been updated and is empty, so
1605 * @old_mems_allowed is the right nodesets that we
1608 if (is_memory_migrate(cs))
1609 cpuset_migrate_mm(mm, &oldcs->old_mems_allowed,
1610 &cpuset_attach_nodemask_to);
1616 cs->old_mems_allowed = cpuset_attach_nodemask_to;
1618 cs->attach_in_progress--;
1619 if (!cs->attach_in_progress)
1620 wake_up(&cpuset_attach_wq);
1622 mutex_unlock(&cpuset_mutex);
1625 /* The various types of files and directories in a cpuset file system */
1628 FILE_MEMORY_MIGRATE,
1631 FILE_EFFECTIVE_CPULIST,
1632 FILE_EFFECTIVE_MEMLIST,
1636 FILE_SCHED_LOAD_BALANCE,
1637 FILE_SCHED_RELAX_DOMAIN_LEVEL,
1638 FILE_MEMORY_PRESSURE_ENABLED,
1639 FILE_MEMORY_PRESSURE,
1642 } cpuset_filetype_t;
1644 static int cpuset_write_u64(struct cgroup_subsys_state *css, struct cftype *cft,
1647 struct cpuset *cs = css_cs(css);
1648 cpuset_filetype_t type = cft->private;
1651 mutex_lock(&cpuset_mutex);
1652 if (!is_cpuset_online(cs)) {
1658 case FILE_CPU_EXCLUSIVE:
1659 retval = update_flag(CS_CPU_EXCLUSIVE, cs, val);
1661 case FILE_MEM_EXCLUSIVE:
1662 retval = update_flag(CS_MEM_EXCLUSIVE, cs, val);
1664 case FILE_MEM_HARDWALL:
1665 retval = update_flag(CS_MEM_HARDWALL, cs, val);
1667 case FILE_SCHED_LOAD_BALANCE:
1668 retval = update_flag(CS_SCHED_LOAD_BALANCE, cs, val);
1670 case FILE_MEMORY_MIGRATE:
1671 retval = update_flag(CS_MEMORY_MIGRATE, cs, val);
1673 case FILE_MEMORY_PRESSURE_ENABLED:
1674 cpuset_memory_pressure_enabled = !!val;
1676 case FILE_SPREAD_PAGE:
1677 retval = update_flag(CS_SPREAD_PAGE, cs, val);
1679 case FILE_SPREAD_SLAB:
1680 retval = update_flag(CS_SPREAD_SLAB, cs, val);
1687 mutex_unlock(&cpuset_mutex);
1691 static int cpuset_write_s64(struct cgroup_subsys_state *css, struct cftype *cft,
1694 struct cpuset *cs = css_cs(css);
1695 cpuset_filetype_t type = cft->private;
1696 int retval = -ENODEV;
1698 mutex_lock(&cpuset_mutex);
1699 if (!is_cpuset_online(cs))
1703 case FILE_SCHED_RELAX_DOMAIN_LEVEL:
1704 retval = update_relax_domain_level(cs, val);
1711 mutex_unlock(&cpuset_mutex);
1716 * Common handling for a write to a "cpus" or "mems" file.
1718 static ssize_t cpuset_write_resmask(struct kernfs_open_file *of,
1719 char *buf, size_t nbytes, loff_t off)
1721 struct cpuset *cs = css_cs(of_css(of));
1722 struct cpuset *trialcs;
1723 int retval = -ENODEV;
1725 buf = strstrip(buf);
1728 * CPU or memory hotunplug may leave @cs w/o any execution
1729 * resources, in which case the hotplug code asynchronously updates
1730 * configuration and transfers all tasks to the nearest ancestor
1731 * which can execute.
1733 * As writes to "cpus" or "mems" may restore @cs's execution
1734 * resources, wait for the previously scheduled operations before
1735 * proceeding, so that we don't end up keep removing tasks added
1736 * after execution capability is restored.
1738 * cpuset_hotplug_work calls back into cgroup core via
1739 * cgroup_transfer_tasks() and waiting for it from a cgroupfs
1740 * operation like this one can lead to a deadlock through kernfs
1741 * active_ref protection. Let's break the protection. Losing the
1742 * protection is okay as we check whether @cs is online after
1743 * grabbing cpuset_mutex anyway. This only happens on the legacy
1747 kernfs_break_active_protection(of->kn);
1748 flush_work(&cpuset_hotplug_work);
1750 mutex_lock(&cpuset_mutex);
1751 if (!is_cpuset_online(cs))
1754 trialcs = alloc_trial_cpuset(cs);
1760 switch (of_cft(of)->private) {
1762 retval = update_cpumask(cs, trialcs, buf);
1765 retval = update_nodemask(cs, trialcs, buf);
1772 free_trial_cpuset(trialcs);
1774 mutex_unlock(&cpuset_mutex);
1775 kernfs_unbreak_active_protection(of->kn);
1777 flush_workqueue(cpuset_migrate_mm_wq);
1778 return retval ?: nbytes;
1782 * These ascii lists should be read in a single call, by using a user
1783 * buffer large enough to hold the entire map. If read in smaller
1784 * chunks, there is no guarantee of atomicity. Since the display format
1785 * used, list of ranges of sequential numbers, is variable length,
1786 * and since these maps can change value dynamically, one could read
1787 * gibberish by doing partial reads while a list was changing.
1789 static int cpuset_common_seq_show(struct seq_file *sf, void *v)
1791 struct cpuset *cs = css_cs(seq_css(sf));
1792 cpuset_filetype_t type = seq_cft(sf)->private;
1795 spin_lock_irq(&callback_lock);
1799 seq_printf(sf, "%*pbl\n", cpumask_pr_args(cs->cpus_allowed));
1802 seq_printf(sf, "%*pbl\n", nodemask_pr_args(&cs->mems_allowed));
1804 case FILE_EFFECTIVE_CPULIST:
1805 seq_printf(sf, "%*pbl\n", cpumask_pr_args(cs->effective_cpus));
1807 case FILE_EFFECTIVE_MEMLIST:
1808 seq_printf(sf, "%*pbl\n", nodemask_pr_args(&cs->effective_mems));
1814 spin_unlock_irq(&callback_lock);
1818 static u64 cpuset_read_u64(struct cgroup_subsys_state *css, struct cftype *cft)
1820 struct cpuset *cs = css_cs(css);
1821 cpuset_filetype_t type = cft->private;
1823 case FILE_CPU_EXCLUSIVE:
1824 return is_cpu_exclusive(cs);
1825 case FILE_MEM_EXCLUSIVE:
1826 return is_mem_exclusive(cs);
1827 case FILE_MEM_HARDWALL:
1828 return is_mem_hardwall(cs);
1829 case FILE_SCHED_LOAD_BALANCE:
1830 return is_sched_load_balance(cs);
1831 case FILE_MEMORY_MIGRATE:
1832 return is_memory_migrate(cs);
1833 case FILE_MEMORY_PRESSURE_ENABLED:
1834 return cpuset_memory_pressure_enabled;
1835 case FILE_MEMORY_PRESSURE:
1836 return fmeter_getrate(&cs->fmeter);
1837 case FILE_SPREAD_PAGE:
1838 return is_spread_page(cs);
1839 case FILE_SPREAD_SLAB:
1840 return is_spread_slab(cs);
1845 /* Unreachable but makes gcc happy */
1849 static s64 cpuset_read_s64(struct cgroup_subsys_state *css, struct cftype *cft)
1851 struct cpuset *cs = css_cs(css);
1852 cpuset_filetype_t type = cft->private;
1854 case FILE_SCHED_RELAX_DOMAIN_LEVEL:
1855 return cs->relax_domain_level;
1860 /* Unrechable but makes gcc happy */
1865 * for the common functions, 'private' gives the type of file
1868 static struct cftype legacy_files[] = {
1871 .seq_show = cpuset_common_seq_show,
1872 .write = cpuset_write_resmask,
1873 .max_write_len = (100U + 6 * NR_CPUS),
1874 .private = FILE_CPULIST,
1879 .seq_show = cpuset_common_seq_show,
1880 .write = cpuset_write_resmask,
1881 .max_write_len = (100U + 6 * MAX_NUMNODES),
1882 .private = FILE_MEMLIST,
1886 .name = "effective_cpus",
1887 .seq_show = cpuset_common_seq_show,
1888 .private = FILE_EFFECTIVE_CPULIST,
1892 .name = "effective_mems",
1893 .seq_show = cpuset_common_seq_show,
1894 .private = FILE_EFFECTIVE_MEMLIST,
1898 .name = "cpu_exclusive",
1899 .read_u64 = cpuset_read_u64,
1900 .write_u64 = cpuset_write_u64,
1901 .private = FILE_CPU_EXCLUSIVE,
1905 .name = "mem_exclusive",
1906 .read_u64 = cpuset_read_u64,
1907 .write_u64 = cpuset_write_u64,
1908 .private = FILE_MEM_EXCLUSIVE,
1912 .name = "mem_hardwall",
1913 .read_u64 = cpuset_read_u64,
1914 .write_u64 = cpuset_write_u64,
1915 .private = FILE_MEM_HARDWALL,
1919 .name = "sched_load_balance",
1920 .read_u64 = cpuset_read_u64,
1921 .write_u64 = cpuset_write_u64,
1922 .private = FILE_SCHED_LOAD_BALANCE,
1926 .name = "sched_relax_domain_level",
1927 .read_s64 = cpuset_read_s64,
1928 .write_s64 = cpuset_write_s64,
1929 .private = FILE_SCHED_RELAX_DOMAIN_LEVEL,
1933 .name = "memory_migrate",
1934 .read_u64 = cpuset_read_u64,
1935 .write_u64 = cpuset_write_u64,
1936 .private = FILE_MEMORY_MIGRATE,
1940 .name = "memory_pressure",
1941 .read_u64 = cpuset_read_u64,
1942 .private = FILE_MEMORY_PRESSURE,
1946 .name = "memory_spread_page",
1947 .read_u64 = cpuset_read_u64,
1948 .write_u64 = cpuset_write_u64,
1949 .private = FILE_SPREAD_PAGE,
1953 .name = "memory_spread_slab",
1954 .read_u64 = cpuset_read_u64,
1955 .write_u64 = cpuset_write_u64,
1956 .private = FILE_SPREAD_SLAB,
1960 .name = "memory_pressure_enabled",
1961 .flags = CFTYPE_ONLY_ON_ROOT,
1962 .read_u64 = cpuset_read_u64,
1963 .write_u64 = cpuset_write_u64,
1964 .private = FILE_MEMORY_PRESSURE_ENABLED,
1971 * This is currently a minimal set for the default hierarchy. It can be
1972 * expanded later on by migrating more features and control files from v1.
1974 static struct cftype dfl_files[] = {
1977 .seq_show = cpuset_common_seq_show,
1978 .write = cpuset_write_resmask,
1979 .max_write_len = (100U + 6 * NR_CPUS),
1980 .private = FILE_CPULIST,
1981 .flags = CFTYPE_NOT_ON_ROOT,
1986 .seq_show = cpuset_common_seq_show,
1987 .write = cpuset_write_resmask,
1988 .max_write_len = (100U + 6 * MAX_NUMNODES),
1989 .private = FILE_MEMLIST,
1990 .flags = CFTYPE_NOT_ON_ROOT,
1994 .name = "cpus.effective",
1995 .seq_show = cpuset_common_seq_show,
1996 .private = FILE_EFFECTIVE_CPULIST,
1997 .flags = CFTYPE_NOT_ON_ROOT,
2001 .name = "mems.effective",
2002 .seq_show = cpuset_common_seq_show,
2003 .private = FILE_EFFECTIVE_MEMLIST,
2004 .flags = CFTYPE_NOT_ON_ROOT,
2012 * cpuset_css_alloc - allocate a cpuset css
2013 * cgrp: control group that the new cpuset will be part of
2016 static struct cgroup_subsys_state *
2017 cpuset_css_alloc(struct cgroup_subsys_state *parent_css)
2022 return &top_cpuset.css;
2024 cs = kzalloc(sizeof(*cs), GFP_KERNEL);
2026 return ERR_PTR(-ENOMEM);
2027 if (!alloc_cpumask_var(&cs->cpus_allowed, GFP_KERNEL))
2029 if (!alloc_cpumask_var(&cs->effective_cpus, GFP_KERNEL))
2032 set_bit(CS_SCHED_LOAD_BALANCE, &cs->flags);
2033 cpumask_clear(cs->cpus_allowed);
2034 nodes_clear(cs->mems_allowed);
2035 cpumask_clear(cs->effective_cpus);
2036 nodes_clear(cs->effective_mems);
2037 fmeter_init(&cs->fmeter);
2038 cs->relax_domain_level = -1;
2043 free_cpumask_var(cs->cpus_allowed);
2046 return ERR_PTR(-ENOMEM);
2049 static int cpuset_css_online(struct cgroup_subsys_state *css)
2051 struct cpuset *cs = css_cs(css);
2052 struct cpuset *parent = parent_cs(cs);
2053 struct cpuset *tmp_cs;
2054 struct cgroup_subsys_state *pos_css;
2059 mutex_lock(&cpuset_mutex);
2061 set_bit(CS_ONLINE, &cs->flags);
2062 if (is_spread_page(parent))
2063 set_bit(CS_SPREAD_PAGE, &cs->flags);
2064 if (is_spread_slab(parent))
2065 set_bit(CS_SPREAD_SLAB, &cs->flags);
2069 spin_lock_irq(&callback_lock);
2070 if (is_in_v2_mode()) {
2071 cpumask_copy(cs->effective_cpus, parent->effective_cpus);
2072 cs->effective_mems = parent->effective_mems;
2074 spin_unlock_irq(&callback_lock);
2076 if (!test_bit(CGRP_CPUSET_CLONE_CHILDREN, &css->cgroup->flags))
2080 * Clone @parent's configuration if CGRP_CPUSET_CLONE_CHILDREN is
2081 * set. This flag handling is implemented in cgroup core for
2082 * histrical reasons - the flag may be specified during mount.
2084 * Currently, if any sibling cpusets have exclusive cpus or mem, we
2085 * refuse to clone the configuration - thereby refusing the task to
2086 * be entered, and as a result refusing the sys_unshare() or
2087 * clone() which initiated it. If this becomes a problem for some
2088 * users who wish to allow that scenario, then this could be
2089 * changed to grant parent->cpus_allowed-sibling_cpus_exclusive
2090 * (and likewise for mems) to the new cgroup.
2093 cpuset_for_each_child(tmp_cs, pos_css, parent) {
2094 if (is_mem_exclusive(tmp_cs) || is_cpu_exclusive(tmp_cs)) {
2101 spin_lock_irq(&callback_lock);
2102 cs->mems_allowed = parent->mems_allowed;
2103 cs->effective_mems = parent->mems_allowed;
2104 cpumask_copy(cs->cpus_allowed, parent->cpus_allowed);
2105 cpumask_copy(cs->effective_cpus, parent->cpus_allowed);
2106 spin_unlock_irq(&callback_lock);
2108 mutex_unlock(&cpuset_mutex);
2113 * If the cpuset being removed has its flag 'sched_load_balance'
2114 * enabled, then simulate turning sched_load_balance off, which
2115 * will call rebuild_sched_domains_locked().
2118 static void cpuset_css_offline(struct cgroup_subsys_state *css)
2120 struct cpuset *cs = css_cs(css);
2122 mutex_lock(&cpuset_mutex);
2124 if (is_sched_load_balance(cs))
2125 update_flag(CS_SCHED_LOAD_BALANCE, cs, 0);
2128 clear_bit(CS_ONLINE, &cs->flags);
2130 mutex_unlock(&cpuset_mutex);
2133 static void cpuset_css_free(struct cgroup_subsys_state *css)
2135 struct cpuset *cs = css_cs(css);
2137 free_cpumask_var(cs->effective_cpus);
2138 free_cpumask_var(cs->cpus_allowed);
2142 static void cpuset_bind(struct cgroup_subsys_state *root_css)
2144 mutex_lock(&cpuset_mutex);
2145 spin_lock_irq(&callback_lock);
2147 if (is_in_v2_mode()) {
2148 cpumask_copy(top_cpuset.cpus_allowed, cpu_possible_mask);
2149 top_cpuset.mems_allowed = node_possible_map;
2151 cpumask_copy(top_cpuset.cpus_allowed,
2152 top_cpuset.effective_cpus);
2153 top_cpuset.mems_allowed = top_cpuset.effective_mems;
2156 spin_unlock_irq(&callback_lock);
2157 mutex_unlock(&cpuset_mutex);
2161 * Make sure the new task conform to the current state of its parent,
2162 * which could have been changed by cpuset just after it inherits the
2163 * state from the parent and before it sits on the cgroup's task list.
2165 static void cpuset_fork(struct task_struct *task)
2167 if (task_css_is_root(task, cpuset_cgrp_id))
2170 set_cpus_allowed_ptr(task, ¤t->cpus_allowed);
2171 task->mems_allowed = current->mems_allowed;
2174 struct cgroup_subsys cpuset_cgrp_subsys = {
2175 .css_alloc = cpuset_css_alloc,
2176 .css_online = cpuset_css_online,
2177 .css_offline = cpuset_css_offline,
2178 .css_free = cpuset_css_free,
2179 .can_attach = cpuset_can_attach,
2180 .cancel_attach = cpuset_cancel_attach,
2181 .attach = cpuset_attach,
2182 .post_attach = cpuset_post_attach,
2183 .bind = cpuset_bind,
2184 .fork = cpuset_fork,
2185 .legacy_cftypes = legacy_files,
2186 .dfl_cftypes = dfl_files,
2192 * cpuset_init - initialize cpusets at system boot
2194 * Description: Initialize top_cpuset and the cpuset internal file system,
2197 int __init cpuset_init(void)
2201 BUG_ON(!alloc_cpumask_var(&top_cpuset.cpus_allowed, GFP_KERNEL));
2202 BUG_ON(!alloc_cpumask_var(&top_cpuset.effective_cpus, GFP_KERNEL));
2204 cpumask_setall(top_cpuset.cpus_allowed);
2205 nodes_setall(top_cpuset.mems_allowed);
2206 cpumask_setall(top_cpuset.effective_cpus);
2207 nodes_setall(top_cpuset.effective_mems);
2209 fmeter_init(&top_cpuset.fmeter);
2210 set_bit(CS_SCHED_LOAD_BALANCE, &top_cpuset.flags);
2211 top_cpuset.relax_domain_level = -1;
2213 err = register_filesystem(&cpuset_fs_type);
2217 BUG_ON(!alloc_cpumask_var(&cpus_attach, GFP_KERNEL));
2223 * If CPU and/or memory hotplug handlers, below, unplug any CPUs
2224 * or memory nodes, we need to walk over the cpuset hierarchy,
2225 * removing that CPU or node from all cpusets. If this removes the
2226 * last CPU or node from a cpuset, then move the tasks in the empty
2227 * cpuset to its next-highest non-empty parent.
2229 static void remove_tasks_in_empty_cpuset(struct cpuset *cs)
2231 struct cpuset *parent;
2234 * Find its next-highest non-empty parent, (top cpuset
2235 * has online cpus, so can't be empty).
2237 parent = parent_cs(cs);
2238 while (cpumask_empty(parent->cpus_allowed) ||
2239 nodes_empty(parent->mems_allowed))
2240 parent = parent_cs(parent);
2242 if (cgroup_transfer_tasks(parent->css.cgroup, cs->css.cgroup)) {
2243 pr_err("cpuset: failed to transfer tasks out of empty cpuset ");
2244 pr_cont_cgroup_name(cs->css.cgroup);
2250 hotplug_update_tasks_legacy(struct cpuset *cs,
2251 struct cpumask *new_cpus, nodemask_t *new_mems,
2252 bool cpus_updated, bool mems_updated)
2256 spin_lock_irq(&callback_lock);
2257 cpumask_copy(cs->cpus_allowed, new_cpus);
2258 cpumask_copy(cs->effective_cpus, new_cpus);
2259 cs->mems_allowed = *new_mems;
2260 cs->effective_mems = *new_mems;
2261 spin_unlock_irq(&callback_lock);
2264 * Don't call update_tasks_cpumask() if the cpuset becomes empty,
2265 * as the tasks will be migratecd to an ancestor.
2267 if (cpus_updated && !cpumask_empty(cs->cpus_allowed))
2268 update_tasks_cpumask(cs);
2269 if (mems_updated && !nodes_empty(cs->mems_allowed))
2270 update_tasks_nodemask(cs);
2272 is_empty = cpumask_empty(cs->cpus_allowed) ||
2273 nodes_empty(cs->mems_allowed);
2275 mutex_unlock(&cpuset_mutex);
2278 * Move tasks to the nearest ancestor with execution resources,
2279 * This is full cgroup operation which will also call back into
2280 * cpuset. Should be done outside any lock.
2283 remove_tasks_in_empty_cpuset(cs);
2285 mutex_lock(&cpuset_mutex);
2289 hotplug_update_tasks(struct cpuset *cs,
2290 struct cpumask *new_cpus, nodemask_t *new_mems,
2291 bool cpus_updated, bool mems_updated)
2293 if (cpumask_empty(new_cpus))
2294 cpumask_copy(new_cpus, parent_cs(cs)->effective_cpus);
2295 if (nodes_empty(*new_mems))
2296 *new_mems = parent_cs(cs)->effective_mems;
2298 spin_lock_irq(&callback_lock);
2299 cpumask_copy(cs->effective_cpus, new_cpus);
2300 cs->effective_mems = *new_mems;
2301 spin_unlock_irq(&callback_lock);
2304 update_tasks_cpumask(cs);
2306 update_tasks_nodemask(cs);
2310 * cpuset_hotplug_update_tasks - update tasks in a cpuset for hotunplug
2311 * @cs: cpuset in interest
2313 * Compare @cs's cpu and mem masks against top_cpuset and if some have gone
2314 * offline, update @cs accordingly. If @cs ends up with no CPU or memory,
2315 * all its tasks are moved to the nearest ancestor with both resources.
2317 static void cpuset_hotplug_update_tasks(struct cpuset *cs)
2319 static cpumask_t new_cpus;
2320 static nodemask_t new_mems;
2324 wait_event(cpuset_attach_wq, cs->attach_in_progress == 0);
2326 mutex_lock(&cpuset_mutex);
2329 * We have raced with task attaching. We wait until attaching
2330 * is finished, so we won't attach a task to an empty cpuset.
2332 if (cs->attach_in_progress) {
2333 mutex_unlock(&cpuset_mutex);
2337 cpumask_and(&new_cpus, cs->cpus_allowed, parent_cs(cs)->effective_cpus);
2338 nodes_and(new_mems, cs->mems_allowed, parent_cs(cs)->effective_mems);
2340 cpus_updated = !cpumask_equal(&new_cpus, cs->effective_cpus);
2341 mems_updated = !nodes_equal(new_mems, cs->effective_mems);
2343 if (is_in_v2_mode())
2344 hotplug_update_tasks(cs, &new_cpus, &new_mems,
2345 cpus_updated, mems_updated);
2347 hotplug_update_tasks_legacy(cs, &new_cpus, &new_mems,
2348 cpus_updated, mems_updated);
2350 mutex_unlock(&cpuset_mutex);
2353 static bool force_rebuild;
2355 void cpuset_force_rebuild(void)
2357 force_rebuild = true;
2361 * cpuset_hotplug_workfn - handle CPU/memory hotunplug for a cpuset
2363 * This function is called after either CPU or memory configuration has
2364 * changed and updates cpuset accordingly. The top_cpuset is always
2365 * synchronized to cpu_active_mask and N_MEMORY, which is necessary in
2366 * order to make cpusets transparent (of no affect) on systems that are
2367 * actively using CPU hotplug but making no active use of cpusets.
2369 * Non-root cpusets are only affected by offlining. If any CPUs or memory
2370 * nodes have been taken down, cpuset_hotplug_update_tasks() is invoked on
2373 * Note that CPU offlining during suspend is ignored. We don't modify
2374 * cpusets across suspend/resume cycles at all.
2376 static void cpuset_hotplug_workfn(struct work_struct *work)
2378 static cpumask_t new_cpus;
2379 static nodemask_t new_mems;
2380 bool cpus_updated, mems_updated;
2381 bool on_dfl = is_in_v2_mode();
2383 mutex_lock(&cpuset_mutex);
2385 /* fetch the available cpus/mems and find out which changed how */
2386 cpumask_copy(&new_cpus, cpu_active_mask);
2387 new_mems = node_states[N_MEMORY];
2389 cpus_updated = !cpumask_equal(top_cpuset.effective_cpus, &new_cpus);
2390 mems_updated = !nodes_equal(top_cpuset.effective_mems, new_mems);
2392 /* synchronize cpus_allowed to cpu_active_mask */
2394 spin_lock_irq(&callback_lock);
2396 cpumask_copy(top_cpuset.cpus_allowed, &new_cpus);
2397 cpumask_copy(top_cpuset.effective_cpus, &new_cpus);
2398 spin_unlock_irq(&callback_lock);
2399 /* we don't mess with cpumasks of tasks in top_cpuset */
2402 /* synchronize mems_allowed to N_MEMORY */
2404 spin_lock_irq(&callback_lock);
2406 top_cpuset.mems_allowed = new_mems;
2407 top_cpuset.effective_mems = new_mems;
2408 spin_unlock_irq(&callback_lock);
2409 update_tasks_nodemask(&top_cpuset);
2412 mutex_unlock(&cpuset_mutex);
2414 /* if cpus or mems changed, we need to propagate to descendants */
2415 if (cpus_updated || mems_updated) {
2417 struct cgroup_subsys_state *pos_css;
2420 cpuset_for_each_descendant_pre(cs, pos_css, &top_cpuset) {
2421 if (cs == &top_cpuset || !css_tryget_online(&cs->css))
2425 cpuset_hotplug_update_tasks(cs);
2433 /* rebuild sched domains if cpus_allowed has changed */
2434 if (cpus_updated || force_rebuild) {
2435 force_rebuild = false;
2436 rebuild_sched_domains();
2440 void cpuset_update_active_cpus(void)
2443 * We're inside cpu hotplug critical region which usually nests
2444 * inside cgroup synchronization. Bounce actual hotplug processing
2445 * to a work item to avoid reverse locking order.
2447 schedule_work(&cpuset_hotplug_work);
2450 void cpuset_wait_for_hotplug(void)
2452 flush_work(&cpuset_hotplug_work);
2456 * Keep top_cpuset.mems_allowed tracking node_states[N_MEMORY].
2457 * Call this routine anytime after node_states[N_MEMORY] changes.
2458 * See cpuset_update_active_cpus() for CPU hotplug handling.
2460 static int cpuset_track_online_nodes(struct notifier_block *self,
2461 unsigned long action, void *arg)
2463 schedule_work(&cpuset_hotplug_work);
2467 static struct notifier_block cpuset_track_online_nodes_nb = {
2468 .notifier_call = cpuset_track_online_nodes,
2469 .priority = 10, /* ??! */
2473 * cpuset_init_smp - initialize cpus_allowed
2475 * Description: Finish top cpuset after cpu, node maps are initialized
2477 void __init cpuset_init_smp(void)
2479 cpumask_copy(top_cpuset.cpus_allowed, cpu_active_mask);
2480 top_cpuset.mems_allowed = node_states[N_MEMORY];
2481 top_cpuset.old_mems_allowed = top_cpuset.mems_allowed;
2483 cpumask_copy(top_cpuset.effective_cpus, cpu_active_mask);
2484 top_cpuset.effective_mems = node_states[N_MEMORY];
2486 register_hotmemory_notifier(&cpuset_track_online_nodes_nb);
2488 cpuset_migrate_mm_wq = alloc_ordered_workqueue("cpuset_migrate_mm", 0);
2489 BUG_ON(!cpuset_migrate_mm_wq);
2493 * cpuset_cpus_allowed - return cpus_allowed mask from a tasks cpuset.
2494 * @tsk: pointer to task_struct from which to obtain cpuset->cpus_allowed.
2495 * @pmask: pointer to struct cpumask variable to receive cpus_allowed set.
2497 * Description: Returns the cpumask_var_t cpus_allowed of the cpuset
2498 * attached to the specified @tsk. Guaranteed to return some non-empty
2499 * subset of cpu_online_mask, even if this means going outside the
2503 void cpuset_cpus_allowed(struct task_struct *tsk, struct cpumask *pmask)
2505 unsigned long flags;
2507 spin_lock_irqsave(&callback_lock, flags);
2509 guarantee_online_cpus(task_cs(tsk), pmask);
2511 spin_unlock_irqrestore(&callback_lock, flags);
2514 void cpuset_cpus_allowed_fallback(struct task_struct *tsk)
2517 do_set_cpus_allowed(tsk, task_cs(tsk)->effective_cpus);
2521 * We own tsk->cpus_allowed, nobody can change it under us.
2523 * But we used cs && cs->cpus_allowed lockless and thus can
2524 * race with cgroup_attach_task() or update_cpumask() and get
2525 * the wrong tsk->cpus_allowed. However, both cases imply the
2526 * subsequent cpuset_change_cpumask()->set_cpus_allowed_ptr()
2527 * which takes task_rq_lock().
2529 * If we are called after it dropped the lock we must see all
2530 * changes in tsk_cs()->cpus_allowed. Otherwise we can temporary
2531 * set any mask even if it is not right from task_cs() pov,
2532 * the pending set_cpus_allowed_ptr() will fix things.
2534 * select_fallback_rq() will fix things ups and set cpu_possible_mask
2539 void __init cpuset_init_current_mems_allowed(void)
2541 nodes_setall(current->mems_allowed);
2545 * cpuset_mems_allowed - return mems_allowed mask from a tasks cpuset.
2546 * @tsk: pointer to task_struct from which to obtain cpuset->mems_allowed.
2548 * Description: Returns the nodemask_t mems_allowed of the cpuset
2549 * attached to the specified @tsk. Guaranteed to return some non-empty
2550 * subset of node_states[N_MEMORY], even if this means going outside the
2554 nodemask_t cpuset_mems_allowed(struct task_struct *tsk)
2557 unsigned long flags;
2559 spin_lock_irqsave(&callback_lock, flags);
2561 guarantee_online_mems(task_cs(tsk), &mask);
2563 spin_unlock_irqrestore(&callback_lock, flags);
2569 * cpuset_nodemask_valid_mems_allowed - check nodemask vs. curremt mems_allowed
2570 * @nodemask: the nodemask to be checked
2572 * Are any of the nodes in the nodemask allowed in current->mems_allowed?
2574 int cpuset_nodemask_valid_mems_allowed(nodemask_t *nodemask)
2576 return nodes_intersects(*nodemask, current->mems_allowed);
2580 * nearest_hardwall_ancestor() - Returns the nearest mem_exclusive or
2581 * mem_hardwall ancestor to the specified cpuset. Call holding
2582 * callback_lock. If no ancestor is mem_exclusive or mem_hardwall
2583 * (an unusual configuration), then returns the root cpuset.
2585 static struct cpuset *nearest_hardwall_ancestor(struct cpuset *cs)
2587 while (!(is_mem_exclusive(cs) || is_mem_hardwall(cs)) && parent_cs(cs))
2593 * cpuset_node_allowed - Can we allocate on a memory node?
2594 * @node: is this an allowed node?
2595 * @gfp_mask: memory allocation flags
2597 * If we're in interrupt, yes, we can always allocate. If @node is set in
2598 * current's mems_allowed, yes. If it's not a __GFP_HARDWALL request and this
2599 * node is set in the nearest hardwalled cpuset ancestor to current's cpuset,
2600 * yes. If current has access to memory reserves as an oom victim, yes.
2603 * GFP_USER allocations are marked with the __GFP_HARDWALL bit,
2604 * and do not allow allocations outside the current tasks cpuset
2605 * unless the task has been OOM killed.
2606 * GFP_KERNEL allocations are not so marked, so can escape to the
2607 * nearest enclosing hardwalled ancestor cpuset.
2609 * Scanning up parent cpusets requires callback_lock. The
2610 * __alloc_pages() routine only calls here with __GFP_HARDWALL bit
2611 * _not_ set if it's a GFP_KERNEL allocation, and all nodes in the
2612 * current tasks mems_allowed came up empty on the first pass over
2613 * the zonelist. So only GFP_KERNEL allocations, if all nodes in the
2614 * cpuset are short of memory, might require taking the callback_lock.
2616 * The first call here from mm/page_alloc:get_page_from_freelist()
2617 * has __GFP_HARDWALL set in gfp_mask, enforcing hardwall cpusets,
2618 * so no allocation on a node outside the cpuset is allowed (unless
2619 * in interrupt, of course).
2621 * The second pass through get_page_from_freelist() doesn't even call
2622 * here for GFP_ATOMIC calls. For those calls, the __alloc_pages()
2623 * variable 'wait' is not set, and the bit ALLOC_CPUSET is not set
2624 * in alloc_flags. That logic and the checks below have the combined
2626 * in_interrupt - any node ok (current task context irrelevant)
2627 * GFP_ATOMIC - any node ok
2628 * tsk_is_oom_victim - any node ok
2629 * GFP_KERNEL - any node in enclosing hardwalled cpuset ok
2630 * GFP_USER - only nodes in current tasks mems allowed ok.
2632 bool __cpuset_node_allowed(int node, gfp_t gfp_mask)
2634 struct cpuset *cs; /* current cpuset ancestors */
2635 int allowed; /* is allocation in zone z allowed? */
2636 unsigned long flags;
2640 if (node_isset(node, current->mems_allowed))
2643 * Allow tasks that have access to memory reserves because they have
2644 * been OOM killed to get memory anywhere.
2646 if (unlikely(tsk_is_oom_victim(current)))
2648 if (gfp_mask & __GFP_HARDWALL) /* If hardwall request, stop here */
2651 if (current->flags & PF_EXITING) /* Let dying task have memory */
2654 /* Not hardwall and node outside mems_allowed: scan up cpusets */
2655 spin_lock_irqsave(&callback_lock, flags);
2658 cs = nearest_hardwall_ancestor(task_cs(current));
2659 allowed = node_isset(node, cs->mems_allowed);
2662 spin_unlock_irqrestore(&callback_lock, flags);
2667 * cpuset_mem_spread_node() - On which node to begin search for a file page
2668 * cpuset_slab_spread_node() - On which node to begin search for a slab page
2670 * If a task is marked PF_SPREAD_PAGE or PF_SPREAD_SLAB (as for
2671 * tasks in a cpuset with is_spread_page or is_spread_slab set),
2672 * and if the memory allocation used cpuset_mem_spread_node()
2673 * to determine on which node to start looking, as it will for
2674 * certain page cache or slab cache pages such as used for file
2675 * system buffers and inode caches, then instead of starting on the
2676 * local node to look for a free page, rather spread the starting
2677 * node around the tasks mems_allowed nodes.
2679 * We don't have to worry about the returned node being offline
2680 * because "it can't happen", and even if it did, it would be ok.
2682 * The routines calling guarantee_online_mems() are careful to
2683 * only set nodes in task->mems_allowed that are online. So it
2684 * should not be possible for the following code to return an
2685 * offline node. But if it did, that would be ok, as this routine
2686 * is not returning the node where the allocation must be, only
2687 * the node where the search should start. The zonelist passed to
2688 * __alloc_pages() will include all nodes. If the slab allocator
2689 * is passed an offline node, it will fall back to the local node.
2690 * See kmem_cache_alloc_node().
2693 static int cpuset_spread_node(int *rotor)
2695 return *rotor = next_node_in(*rotor, current->mems_allowed);
2698 int cpuset_mem_spread_node(void)
2700 if (current->cpuset_mem_spread_rotor == NUMA_NO_NODE)
2701 current->cpuset_mem_spread_rotor =
2702 node_random(¤t->mems_allowed);
2704 return cpuset_spread_node(¤t->cpuset_mem_spread_rotor);
2707 int cpuset_slab_spread_node(void)
2709 if (current->cpuset_slab_spread_rotor == NUMA_NO_NODE)
2710 current->cpuset_slab_spread_rotor =
2711 node_random(¤t->mems_allowed);
2713 return cpuset_spread_node(¤t->cpuset_slab_spread_rotor);
2716 EXPORT_SYMBOL_GPL(cpuset_mem_spread_node);
2719 * cpuset_mems_allowed_intersects - Does @tsk1's mems_allowed intersect @tsk2's?
2720 * @tsk1: pointer to task_struct of some task.
2721 * @tsk2: pointer to task_struct of some other task.
2723 * Description: Return true if @tsk1's mems_allowed intersects the
2724 * mems_allowed of @tsk2. Used by the OOM killer to determine if
2725 * one of the task's memory usage might impact the memory available
2729 int cpuset_mems_allowed_intersects(const struct task_struct *tsk1,
2730 const struct task_struct *tsk2)
2732 return nodes_intersects(tsk1->mems_allowed, tsk2->mems_allowed);
2736 * cpuset_print_current_mems_allowed - prints current's cpuset and mems_allowed
2738 * Description: Prints current's name, cpuset name, and cached copy of its
2739 * mems_allowed to the kernel log.
2741 void cpuset_print_current_mems_allowed(void)
2743 struct cgroup *cgrp;
2747 cgrp = task_cs(current)->css.cgroup;
2748 pr_info("%s cpuset=", current->comm);
2749 pr_cont_cgroup_name(cgrp);
2750 pr_cont(" mems_allowed=%*pbl\n",
2751 nodemask_pr_args(¤t->mems_allowed));
2757 * Collection of memory_pressure is suppressed unless
2758 * this flag is enabled by writing "1" to the special
2759 * cpuset file 'memory_pressure_enabled' in the root cpuset.
2762 int cpuset_memory_pressure_enabled __read_mostly;
2765 * cpuset_memory_pressure_bump - keep stats of per-cpuset reclaims.
2767 * Keep a running average of the rate of synchronous (direct)
2768 * page reclaim efforts initiated by tasks in each cpuset.
2770 * This represents the rate at which some task in the cpuset
2771 * ran low on memory on all nodes it was allowed to use, and
2772 * had to enter the kernels page reclaim code in an effort to
2773 * create more free memory by tossing clean pages or swapping
2774 * or writing dirty pages.
2776 * Display to user space in the per-cpuset read-only file
2777 * "memory_pressure". Value displayed is an integer
2778 * representing the recent rate of entry into the synchronous
2779 * (direct) page reclaim by any task attached to the cpuset.
2782 void __cpuset_memory_pressure_bump(void)
2785 fmeter_markevent(&task_cs(current)->fmeter);
2789 #ifdef CONFIG_PROC_PID_CPUSET
2791 * proc_cpuset_show()
2792 * - Print tasks cpuset path into seq_file.
2793 * - Used for /proc/<pid>/cpuset.
2794 * - No need to task_lock(tsk) on this tsk->cpuset reference, as it
2795 * doesn't really matter if tsk->cpuset changes after we read it,
2796 * and we take cpuset_mutex, keeping cpuset_attach() from changing it
2799 int proc_cpuset_show(struct seq_file *m, struct pid_namespace *ns,
2800 struct pid *pid, struct task_struct *tsk)
2803 struct cgroup_subsys_state *css;
2807 buf = kmalloc(PATH_MAX, GFP_KERNEL);
2811 css = task_get_css(tsk, cpuset_cgrp_id);
2812 retval = cgroup_path_ns(css->cgroup, buf, PATH_MAX,
2813 current->nsproxy->cgroup_ns);
2815 if (retval >= PATH_MAX)
2816 retval = -ENAMETOOLONG;
2827 #endif /* CONFIG_PROC_PID_CPUSET */
2829 /* Display task mems_allowed in /proc/<pid>/status file. */
2830 void cpuset_task_status_allowed(struct seq_file *m, struct task_struct *task)
2832 seq_printf(m, "Mems_allowed:\t%*pb\n",
2833 nodemask_pr_args(&task->mems_allowed));
2834 seq_printf(m, "Mems_allowed_list:\t%*pbl\n",
2835 nodemask_pr_args(&task->mems_allowed));