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/cgroup.h>
61 #include <linux/wait.h>
63 struct static_key cpusets_pre_enable_key __read_mostly = STATIC_KEY_INIT_FALSE;
64 struct static_key cpusets_enabled_key __read_mostly = STATIC_KEY_INIT_FALSE;
66 /* See "Frequency meter" comments, below. */
69 int cnt; /* unprocessed events count */
70 int val; /* most recent output value */
71 time_t time; /* clock (secs) when val computed */
72 spinlock_t lock; /* guards read or write of above */
76 struct cgroup_subsys_state css;
78 unsigned long flags; /* "unsigned long" so bitops work */
81 * On default hierarchy:
83 * The user-configured masks can only be changed by writing to
84 * cpuset.cpus and cpuset.mems, and won't be limited by the
87 * The effective masks is the real masks that apply to the tasks
88 * in the cpuset. They may be changed if the configured masks are
89 * changed or hotplug happens.
91 * effective_mask == configured_mask & parent's effective_mask,
92 * and if it ends up empty, it will inherit the parent's mask.
97 * The user-configured masks are always the same with effective masks.
100 /* user-configured CPUs and Memory Nodes allow to tasks */
101 cpumask_var_t cpus_allowed;
102 cpumask_var_t cpus_requested; /* CPUS requested, but not used because of hotplug */
103 nodemask_t mems_allowed;
105 /* effective CPUs and Memory Nodes allow to tasks */
106 cpumask_var_t effective_cpus;
107 nodemask_t effective_mems;
110 * This is old Memory Nodes tasks took on.
112 * - top_cpuset.old_mems_allowed is initialized to mems_allowed.
113 * - A new cpuset's old_mems_allowed is initialized when some
114 * task is moved into it.
115 * - old_mems_allowed is used in cpuset_migrate_mm() when we change
116 * cpuset.mems_allowed and have tasks' nodemask updated, and
117 * then old_mems_allowed is updated to mems_allowed.
119 nodemask_t old_mems_allowed;
121 struct fmeter fmeter; /* memory_pressure filter */
124 * Tasks are being attached to this cpuset. Used to prevent
125 * zeroing cpus/mems_allowed between ->can_attach() and ->attach().
127 int attach_in_progress;
129 /* partition number for rebuild_sched_domains() */
132 /* for custom sched domain */
133 int relax_domain_level;
136 static inline struct cpuset *css_cs(struct cgroup_subsys_state *css)
138 return css ? container_of(css, struct cpuset, css) : NULL;
141 /* Retrieve the cpuset for a task */
142 static inline struct cpuset *task_cs(struct task_struct *task)
144 return css_cs(task_css(task, cpuset_cgrp_id));
147 static inline struct cpuset *parent_cs(struct cpuset *cs)
149 return css_cs(cs->css.parent);
153 static inline bool task_has_mempolicy(struct task_struct *task)
155 return task->mempolicy;
158 static inline bool task_has_mempolicy(struct task_struct *task)
165 /* bits in struct cpuset flags field */
172 CS_SCHED_LOAD_BALANCE,
177 /* convenient tests for these bits */
178 static inline bool is_cpuset_online(struct cpuset *cs)
180 return test_bit(CS_ONLINE, &cs->flags) && !css_is_dying(&cs->css);
183 static inline int is_cpu_exclusive(const struct cpuset *cs)
185 return test_bit(CS_CPU_EXCLUSIVE, &cs->flags);
188 static inline int is_mem_exclusive(const struct cpuset *cs)
190 return test_bit(CS_MEM_EXCLUSIVE, &cs->flags);
193 static inline int is_mem_hardwall(const struct cpuset *cs)
195 return test_bit(CS_MEM_HARDWALL, &cs->flags);
198 static inline int is_sched_load_balance(const struct cpuset *cs)
200 return test_bit(CS_SCHED_LOAD_BALANCE, &cs->flags);
203 static inline int is_memory_migrate(const struct cpuset *cs)
205 return test_bit(CS_MEMORY_MIGRATE, &cs->flags);
208 static inline int is_spread_page(const struct cpuset *cs)
210 return test_bit(CS_SPREAD_PAGE, &cs->flags);
213 static inline int is_spread_slab(const struct cpuset *cs)
215 return test_bit(CS_SPREAD_SLAB, &cs->flags);
218 static struct cpuset top_cpuset = {
219 .flags = ((1 << CS_ONLINE) | (1 << CS_CPU_EXCLUSIVE) |
220 (1 << CS_MEM_EXCLUSIVE)),
224 * cpuset_for_each_child - traverse online children of a cpuset
225 * @child_cs: loop cursor pointing to the current child
226 * @pos_css: used for iteration
227 * @parent_cs: target cpuset to walk children of
229 * Walk @child_cs through the online children of @parent_cs. Must be used
230 * with RCU read locked.
232 #define cpuset_for_each_child(child_cs, pos_css, parent_cs) \
233 css_for_each_child((pos_css), &(parent_cs)->css) \
234 if (is_cpuset_online(((child_cs) = css_cs((pos_css)))))
237 * cpuset_for_each_descendant_pre - pre-order walk of a cpuset's descendants
238 * @des_cs: loop cursor pointing to the current descendant
239 * @pos_css: used for iteration
240 * @root_cs: target cpuset to walk ancestor of
242 * Walk @des_cs through the online descendants of @root_cs. Must be used
243 * with RCU read locked. The caller may modify @pos_css by calling
244 * css_rightmost_descendant() to skip subtree. @root_cs is included in the
245 * iteration and the first node to be visited.
247 #define cpuset_for_each_descendant_pre(des_cs, pos_css, root_cs) \
248 css_for_each_descendant_pre((pos_css), &(root_cs)->css) \
249 if (is_cpuset_online(((des_cs) = css_cs((pos_css)))))
252 * There are two global locks guarding cpuset structures - cpuset_mutex and
253 * callback_lock. We also require taking task_lock() when dereferencing a
254 * task's cpuset pointer. See "The task_lock() exception", at the end of this
257 * A task must hold both locks to modify cpusets. If a task holds
258 * cpuset_mutex, then it blocks others wanting that mutex, ensuring that it
259 * is the only task able to also acquire callback_lock and be able to
260 * modify cpusets. It can perform various checks on the cpuset structure
261 * first, knowing nothing will change. It can also allocate memory while
262 * just holding cpuset_mutex. While it is performing these checks, various
263 * callback routines can briefly acquire callback_lock to query cpusets.
264 * Once it is ready to make the changes, it takes callback_lock, blocking
267 * Calls to the kernel memory allocator can not be made while holding
268 * callback_lock, as that would risk double tripping on callback_lock
269 * from one of the callbacks into the cpuset code from within
272 * If a task is only holding callback_lock, then it has read-only
275 * Now, the task_struct fields mems_allowed and mempolicy may be changed
276 * by other task, we use alloc_lock in the task_struct fields to protect
279 * The cpuset_common_file_read() handlers only hold callback_lock across
280 * small pieces of code, such as when reading out possibly multi-word
281 * cpumasks and nodemasks.
283 * Accessing a task's cpuset should be done in accordance with the
284 * guidelines for accessing subsystem state in kernel/cgroup.c
287 static DEFINE_MUTEX(cpuset_mutex);
288 static DEFINE_SPINLOCK(callback_lock);
290 static struct workqueue_struct *cpuset_migrate_mm_wq;
293 * CPU / memory hotplug is handled asynchronously.
295 static void cpuset_hotplug_workfn(struct work_struct *work);
296 static DECLARE_WORK(cpuset_hotplug_work, cpuset_hotplug_workfn);
298 static DECLARE_WAIT_QUEUE_HEAD(cpuset_attach_wq);
301 * This is ugly, but preserves the userspace API for existing cpuset
302 * users. If someone tries to mount the "cpuset" filesystem, we
303 * silently switch it to mount "cgroup" instead
305 static struct dentry *cpuset_mount(struct file_system_type *fs_type,
306 int flags, const char *unused_dev_name, void *data)
308 struct file_system_type *cgroup_fs = get_fs_type("cgroup");
309 struct dentry *ret = ERR_PTR(-ENODEV);
313 "release_agent=/sbin/cpuset_release_agent";
314 ret = cgroup_fs->mount(cgroup_fs, flags,
315 unused_dev_name, mountopts);
316 put_filesystem(cgroup_fs);
321 static struct file_system_type cpuset_fs_type = {
323 .mount = cpuset_mount,
327 * Return in pmask the portion of a cpusets's cpus_allowed that
328 * are online. If none are online, walk up the cpuset hierarchy
329 * until we find one that does have some online cpus.
331 * One way or another, we guarantee to return some non-empty subset
332 * of cpu_online_mask.
334 * Call with callback_lock or cpuset_mutex held.
336 static void guarantee_online_cpus(struct cpuset *cs, struct cpumask *pmask)
338 while (!cpumask_intersects(cs->effective_cpus, cpu_online_mask)) {
342 * The top cpuset doesn't have any online cpu as a
343 * consequence of a race between cpuset_hotplug_work
344 * and cpu hotplug notifier. But we know the top
345 * cpuset's effective_cpus is on its way to to be
346 * identical to cpu_online_mask.
348 cpumask_copy(pmask, cpu_online_mask);
352 cpumask_and(pmask, cs->effective_cpus, cpu_online_mask);
356 * Return in *pmask the portion of a cpusets's mems_allowed that
357 * are online, with memory. If none are online with memory, walk
358 * up the cpuset hierarchy until we find one that does have some
359 * online mems. The top cpuset always has some mems online.
361 * One way or another, we guarantee to return some non-empty subset
362 * of node_states[N_MEMORY].
364 * Call with callback_lock or cpuset_mutex held.
366 static void guarantee_online_mems(struct cpuset *cs, nodemask_t *pmask)
368 while (!nodes_intersects(cs->effective_mems, node_states[N_MEMORY]))
370 nodes_and(*pmask, cs->effective_mems, node_states[N_MEMORY]);
374 * update task's spread flag if cpuset's page/slab spread flag is set
376 * Call with callback_lock or cpuset_mutex held.
378 static void cpuset_update_task_spread_flag(struct cpuset *cs,
379 struct task_struct *tsk)
381 if (is_spread_page(cs))
382 task_set_spread_page(tsk);
384 task_clear_spread_page(tsk);
386 if (is_spread_slab(cs))
387 task_set_spread_slab(tsk);
389 task_clear_spread_slab(tsk);
393 * is_cpuset_subset(p, q) - Is cpuset p a subset of cpuset q?
395 * One cpuset is a subset of another if all its allowed CPUs and
396 * Memory Nodes are a subset of the other, and its exclusive flags
397 * are only set if the other's are set. Call holding cpuset_mutex.
400 static int is_cpuset_subset(const struct cpuset *p, const struct cpuset *q)
402 return cpumask_subset(p->cpus_requested, q->cpus_requested) &&
403 nodes_subset(p->mems_allowed, q->mems_allowed) &&
404 is_cpu_exclusive(p) <= is_cpu_exclusive(q) &&
405 is_mem_exclusive(p) <= is_mem_exclusive(q);
409 * alloc_trial_cpuset - allocate a trial cpuset
410 * @cs: the cpuset that the trial cpuset duplicates
412 static struct cpuset *alloc_trial_cpuset(struct cpuset *cs)
414 struct cpuset *trial;
416 trial = kmemdup(cs, sizeof(*cs), GFP_KERNEL);
420 if (!alloc_cpumask_var(&trial->cpus_allowed, GFP_KERNEL))
422 if (!alloc_cpumask_var(&trial->cpus_requested, GFP_KERNEL))
424 if (!alloc_cpumask_var(&trial->effective_cpus, GFP_KERNEL))
427 cpumask_copy(trial->cpus_allowed, cs->cpus_allowed);
428 cpumask_copy(trial->cpus_requested, cs->cpus_requested);
429 cpumask_copy(trial->effective_cpus, cs->effective_cpus);
433 free_cpumask_var(trial->cpus_requested);
435 free_cpumask_var(trial->cpus_allowed);
442 * free_trial_cpuset - free the trial cpuset
443 * @trial: the trial cpuset to be freed
445 static void free_trial_cpuset(struct cpuset *trial)
447 free_cpumask_var(trial->effective_cpus);
448 free_cpumask_var(trial->cpus_requested);
449 free_cpumask_var(trial->cpus_allowed);
454 * validate_change() - Used to validate that any proposed cpuset change
455 * follows the structural rules for cpusets.
457 * If we replaced the flag and mask values of the current cpuset
458 * (cur) with those values in the trial cpuset (trial), would
459 * our various subset and exclusive rules still be valid? Presumes
462 * 'cur' is the address of an actual, in-use cpuset. Operations
463 * such as list traversal that depend on the actual address of the
464 * cpuset in the list must use cur below, not trial.
466 * 'trial' is the address of bulk structure copy of cur, with
467 * perhaps one or more of the fields cpus_allowed, mems_allowed,
468 * or flags changed to new, trial values.
470 * Return 0 if valid, -errno if not.
473 static int validate_change(struct cpuset *cur, struct cpuset *trial)
475 struct cgroup_subsys_state *css;
476 struct cpuset *c, *par;
481 /* Each of our child cpusets must be a subset of us */
483 cpuset_for_each_child(c, css, cur)
484 if (!is_cpuset_subset(c, trial))
487 /* Remaining checks don't apply to root cpuset */
489 if (cur == &top_cpuset)
492 par = parent_cs(cur);
494 /* On legacy hiearchy, we must be a subset of our parent cpuset. */
496 if (!cgroup_subsys_on_dfl(cpuset_cgrp_subsys) &&
497 !is_cpuset_subset(trial, par))
501 * If either I or some sibling (!= me) is exclusive, we can't
505 cpuset_for_each_child(c, css, par) {
506 if ((is_cpu_exclusive(trial) || is_cpu_exclusive(c)) &&
508 cpumask_intersects(trial->cpus_requested, c->cpus_requested))
510 if ((is_mem_exclusive(trial) || is_mem_exclusive(c)) &&
512 nodes_intersects(trial->mems_allowed, c->mems_allowed))
517 * Cpusets with tasks - existing or newly being attached - can't
518 * be changed to have empty cpus_allowed or mems_allowed.
521 if ((cgroup_is_populated(cur->css.cgroup) || cur->attach_in_progress)) {
522 if (!cpumask_empty(cur->cpus_allowed) &&
523 cpumask_empty(trial->cpus_allowed))
525 if (!nodes_empty(cur->mems_allowed) &&
526 nodes_empty(trial->mems_allowed))
531 * We can't shrink if we won't have enough room for SCHED_DEADLINE
535 if (is_cpu_exclusive(cur) &&
536 !cpuset_cpumask_can_shrink(cur->cpus_allowed,
537 trial->cpus_allowed))
548 * Helper routine for generate_sched_domains().
549 * Do cpusets a, b have overlapping effective cpus_allowed masks?
551 static int cpusets_overlap(struct cpuset *a, struct cpuset *b)
553 return cpumask_intersects(a->effective_cpus, b->effective_cpus);
557 update_domain_attr(struct sched_domain_attr *dattr, struct cpuset *c)
559 if (dattr->relax_domain_level < c->relax_domain_level)
560 dattr->relax_domain_level = c->relax_domain_level;
564 static void update_domain_attr_tree(struct sched_domain_attr *dattr,
565 struct cpuset *root_cs)
568 struct cgroup_subsys_state *pos_css;
571 cpuset_for_each_descendant_pre(cp, pos_css, root_cs) {
572 /* skip the whole subtree if @cp doesn't have any CPU */
573 if (cpumask_empty(cp->cpus_allowed)) {
574 pos_css = css_rightmost_descendant(pos_css);
578 if (is_sched_load_balance(cp))
579 update_domain_attr(dattr, cp);
585 * generate_sched_domains()
587 * This function builds a partial partition of the systems CPUs
588 * A 'partial partition' is a set of non-overlapping subsets whose
589 * union is a subset of that set.
590 * The output of this function needs to be passed to kernel/sched/core.c
591 * partition_sched_domains() routine, which will rebuild the scheduler's
592 * load balancing domains (sched domains) as specified by that partial
595 * See "What is sched_load_balance" in Documentation/cgroups/cpusets.txt
596 * for a background explanation of this.
598 * Does not return errors, on the theory that the callers of this
599 * routine would rather not worry about failures to rebuild sched
600 * domains when operating in the severe memory shortage situations
601 * that could cause allocation failures below.
603 * Must be called with cpuset_mutex held.
605 * The three key local variables below are:
606 * q - a linked-list queue of cpuset pointers, used to implement a
607 * top-down scan of all cpusets. This scan loads a pointer
608 * to each cpuset marked is_sched_load_balance into the
609 * array 'csa'. For our purposes, rebuilding the schedulers
610 * sched domains, we can ignore !is_sched_load_balance cpusets.
611 * csa - (for CpuSet Array) Array of pointers to all the cpusets
612 * that need to be load balanced, for convenient iterative
613 * access by the subsequent code that finds the best partition,
614 * i.e the set of domains (subsets) of CPUs such that the
615 * cpus_allowed of every cpuset marked is_sched_load_balance
616 * is a subset of one of these domains, while there are as
617 * many such domains as possible, each as small as possible.
618 * doms - Conversion of 'csa' to an array of cpumasks, for passing to
619 * the kernel/sched/core.c routine partition_sched_domains() in a
620 * convenient format, that can be easily compared to the prior
621 * value to determine what partition elements (sched domains)
622 * were changed (added or removed.)
624 * Finding the best partition (set of domains):
625 * The triple nested loops below over i, j, k scan over the
626 * load balanced cpusets (using the array of cpuset pointers in
627 * csa[]) looking for pairs of cpusets that have overlapping
628 * cpus_allowed, but which don't have the same 'pn' partition
629 * number and gives them in the same partition number. It keeps
630 * looping on the 'restart' label until it can no longer find
633 * The union of the cpus_allowed masks from the set of
634 * all cpusets having the same 'pn' value then form the one
635 * element of the partition (one sched domain) to be passed to
636 * partition_sched_domains().
638 static int generate_sched_domains(cpumask_var_t **domains,
639 struct sched_domain_attr **attributes)
641 struct cpuset *cp; /* scans q */
642 struct cpuset **csa; /* array of all cpuset ptrs */
643 int csn; /* how many cpuset ptrs in csa so far */
644 int i, j, k; /* indices for partition finding loops */
645 cpumask_var_t *doms; /* resulting partition; i.e. sched domains */
646 cpumask_var_t non_isolated_cpus; /* load balanced CPUs */
647 struct sched_domain_attr *dattr; /* attributes for custom domains */
648 int ndoms = 0; /* number of sched domains in result */
649 int nslot; /* next empty doms[] struct cpumask slot */
650 struct cgroup_subsys_state *pos_css;
656 if (!alloc_cpumask_var(&non_isolated_cpus, GFP_KERNEL))
658 cpumask_andnot(non_isolated_cpus, cpu_possible_mask, cpu_isolated_map);
660 /* Special case for the 99% of systems with one, full, sched domain */
661 if (is_sched_load_balance(&top_cpuset)) {
663 doms = alloc_sched_domains(ndoms);
667 dattr = kmalloc(sizeof(struct sched_domain_attr), GFP_KERNEL);
669 *dattr = SD_ATTR_INIT;
670 update_domain_attr_tree(dattr, &top_cpuset);
672 cpumask_and(doms[0], top_cpuset.effective_cpus,
678 csa = kmalloc(nr_cpusets() * sizeof(cp), GFP_KERNEL);
684 cpuset_for_each_descendant_pre(cp, pos_css, &top_cpuset) {
685 if (cp == &top_cpuset)
688 * Continue traversing beyond @cp iff @cp has some CPUs and
689 * isn't load balancing. The former is obvious. The
690 * latter: All child cpusets contain a subset of the
691 * parent's cpus, so just skip them, and then we call
692 * update_domain_attr_tree() to calc relax_domain_level of
693 * the corresponding sched domain.
695 if (!cpumask_empty(cp->cpus_allowed) &&
696 !(is_sched_load_balance(cp) &&
697 cpumask_intersects(cp->cpus_allowed, non_isolated_cpus)))
700 if (is_sched_load_balance(cp))
703 /* skip @cp's subtree */
704 pos_css = css_rightmost_descendant(pos_css);
708 for (i = 0; i < csn; i++)
713 /* Find the best partition (set of sched domains) */
714 for (i = 0; i < csn; i++) {
715 struct cpuset *a = csa[i];
718 for (j = 0; j < csn; j++) {
719 struct cpuset *b = csa[j];
722 if (apn != bpn && cpusets_overlap(a, b)) {
723 for (k = 0; k < csn; k++) {
724 struct cpuset *c = csa[k];
729 ndoms--; /* one less element */
736 * Now we know how many domains to create.
737 * Convert <csn, csa> to <ndoms, doms> and populate cpu masks.
739 doms = alloc_sched_domains(ndoms);
744 * The rest of the code, including the scheduler, can deal with
745 * dattr==NULL case. No need to abort if alloc fails.
747 dattr = kmalloc(ndoms * sizeof(struct sched_domain_attr), GFP_KERNEL);
749 for (nslot = 0, i = 0; i < csn; i++) {
750 struct cpuset *a = csa[i];
755 /* Skip completed partitions */
761 if (nslot == ndoms) {
762 static int warnings = 10;
764 pr_warn("rebuild_sched_domains confused: nslot %d, ndoms %d, csn %d, i %d, apn %d\n",
765 nslot, ndoms, csn, i, apn);
773 *(dattr + nslot) = SD_ATTR_INIT;
774 for (j = i; j < csn; j++) {
775 struct cpuset *b = csa[j];
778 cpumask_or(dp, dp, b->effective_cpus);
779 cpumask_and(dp, dp, non_isolated_cpus);
781 update_domain_attr_tree(dattr + nslot, b);
783 /* Done with this partition */
789 BUG_ON(nslot != ndoms);
792 free_cpumask_var(non_isolated_cpus);
796 * Fallback to the default domain if kmalloc() failed.
797 * See comments in partition_sched_domains().
808 * Rebuild scheduler domains.
810 * If the flag 'sched_load_balance' of any cpuset with non-empty
811 * 'cpus' changes, or if the 'cpus' allowed changes in any cpuset
812 * which has that flag enabled, or if any cpuset with a non-empty
813 * 'cpus' is removed, then call this routine to rebuild the
814 * scheduler's dynamic sched domains.
817 static void rebuild_sched_domains_unlocked(void)
819 struct sched_domain_attr *attr;
823 cpu_hotplug_mutex_held();
824 lockdep_assert_held(&cpuset_mutex);
827 * We have raced with CPU hotplug. Don't do anything to avoid
828 * passing doms with offlined cpu to partition_sched_domains().
829 * Anyways, hotplug work item will rebuild sched domains.
831 if (!cpumask_equal(top_cpuset.effective_cpus, cpu_active_mask))
834 /* Generate domain masks and attrs */
835 ndoms = generate_sched_domains(&doms, &attr);
837 /* Have scheduler rebuild the domains */
838 partition_sched_domains(ndoms, doms, attr);
840 #else /* !CONFIG_SMP */
841 static void rebuild_sched_domains_unlocked(void)
844 #endif /* CONFIG_SMP */
846 void rebuild_sched_domains(void)
849 mutex_lock(&cpuset_mutex);
850 rebuild_sched_domains_unlocked();
851 mutex_unlock(&cpuset_mutex);
856 * update_tasks_cpumask - Update the cpumasks of tasks in the cpuset.
857 * @cs: the cpuset in which each task's cpus_allowed mask needs to be changed
859 * Iterate through each task of @cs updating its cpus_allowed to the
860 * effective cpuset's. As this function is called with cpuset_mutex held,
861 * cpuset membership stays stable.
863 static void update_tasks_cpumask(struct cpuset *cs)
865 struct css_task_iter it;
866 struct task_struct *task;
868 css_task_iter_start(&cs->css, &it);
869 while ((task = css_task_iter_next(&it)))
870 set_cpus_allowed_ptr(task, cs->effective_cpus);
871 css_task_iter_end(&it);
875 * update_cpumasks_hier - Update effective cpumasks and tasks in the subtree
876 * @cs: the cpuset to consider
877 * @new_cpus: temp variable for calculating new effective_cpus
879 * When congifured cpumask is changed, the effective cpumasks of this cpuset
880 * and all its descendants need to be updated.
882 * On legacy hierachy, effective_cpus will be the same with cpu_allowed.
885 static void update_cpumasks_hier(struct cpuset *cs, struct cpumask *new_cpus)
888 struct cgroup_subsys_state *pos_css;
889 bool need_rebuild_sched_domains = false;
892 cpuset_for_each_descendant_pre(cp, pos_css, cs) {
893 struct cpuset *parent = parent_cs(cp);
895 cpumask_and(new_cpus, cp->cpus_allowed, parent->effective_cpus);
898 * If it becomes empty, inherit the effective mask of the
899 * parent, which is guaranteed to have some CPUs.
901 if (cgroup_subsys_on_dfl(cpuset_cgrp_subsys) &&
902 cpumask_empty(new_cpus))
903 cpumask_copy(new_cpus, parent->effective_cpus);
905 /* Skip the whole subtree if the cpumask remains the same. */
906 if (cpumask_equal(new_cpus, cp->effective_cpus)) {
907 pos_css = css_rightmost_descendant(pos_css);
911 if (!css_tryget_online(&cp->css))
915 spin_lock_irq(&callback_lock);
916 cpumask_copy(cp->effective_cpus, new_cpus);
917 spin_unlock_irq(&callback_lock);
919 WARN_ON(!cgroup_subsys_on_dfl(cpuset_cgrp_subsys) &&
920 !cpumask_equal(cp->cpus_allowed, cp->effective_cpus));
922 update_tasks_cpumask(cp);
925 * If the effective cpumask of any non-empty cpuset is changed,
926 * we need to rebuild sched domains.
928 if (!cpumask_empty(cp->cpus_allowed) &&
929 is_sched_load_balance(cp))
930 need_rebuild_sched_domains = true;
937 if (need_rebuild_sched_domains)
938 rebuild_sched_domains_unlocked();
942 * update_cpumask - update the cpus_allowed mask of a cpuset and all tasks in it
943 * @cs: the cpuset to consider
944 * @trialcs: trial cpuset
945 * @buf: buffer of cpu numbers written to this cpuset
947 static int update_cpumask(struct cpuset *cs, struct cpuset *trialcs,
952 /* top_cpuset.cpus_allowed tracks cpu_online_mask; it's read-only */
953 if (cs == &top_cpuset)
957 * An empty cpus_requested is ok only if the cpuset has no tasks.
958 * Since cpulist_parse() fails on an empty mask, we special case
959 * that parsing. The validate_change() call ensures that cpusets
960 * with tasks have cpus.
963 cpumask_clear(trialcs->cpus_requested);
965 retval = cpulist_parse(buf, trialcs->cpus_requested);
970 if (!cpumask_subset(trialcs->cpus_requested, cpu_present_mask))
973 cpumask_and(trialcs->cpus_allowed, trialcs->cpus_requested, cpu_active_mask);
975 /* Nothing to do if the cpus didn't change */
976 if (cpumask_equal(cs->cpus_requested, trialcs->cpus_requested))
979 retval = validate_change(cs, trialcs);
983 spin_lock_irq(&callback_lock);
984 cpumask_copy(cs->cpus_allowed, trialcs->cpus_allowed);
985 cpumask_copy(cs->cpus_requested, trialcs->cpus_requested);
986 spin_unlock_irq(&callback_lock);
988 /* use trialcs->cpus_allowed as a temp variable */
989 update_cpumasks_hier(cs, trialcs->cpus_allowed);
994 * Migrate memory region from one set of nodes to another. This is
995 * performed asynchronously as it can be called from process migration path
996 * holding locks involved in process management. All mm migrations are
997 * performed in the queued order and can be waited for by flushing
998 * cpuset_migrate_mm_wq.
1001 struct cpuset_migrate_mm_work {
1002 struct work_struct work;
1003 struct mm_struct *mm;
1008 static void cpuset_migrate_mm_workfn(struct work_struct *work)
1010 struct cpuset_migrate_mm_work *mwork =
1011 container_of(work, struct cpuset_migrate_mm_work, work);
1013 /* on a wq worker, no need to worry about %current's mems_allowed */
1014 do_migrate_pages(mwork->mm, &mwork->from, &mwork->to, MPOL_MF_MOVE_ALL);
1019 static void cpuset_migrate_mm(struct mm_struct *mm, const nodemask_t *from,
1020 const nodemask_t *to)
1022 struct cpuset_migrate_mm_work *mwork;
1024 mwork = kzalloc(sizeof(*mwork), GFP_KERNEL);
1027 mwork->from = *from;
1029 INIT_WORK(&mwork->work, cpuset_migrate_mm_workfn);
1030 queue_work(cpuset_migrate_mm_wq, &mwork->work);
1036 static void cpuset_post_attach(void)
1038 flush_workqueue(cpuset_migrate_mm_wq);
1042 * cpuset_change_task_nodemask - change task's mems_allowed and mempolicy
1043 * @tsk: the task to change
1044 * @newmems: new nodes that the task will be set
1046 * In order to avoid seeing no nodes if the old and new nodes are disjoint,
1047 * we structure updates as setting all new allowed nodes, then clearing newly
1050 static void cpuset_change_task_nodemask(struct task_struct *tsk,
1051 nodemask_t *newmems)
1056 * Allow tasks that have access to memory reserves because they have
1057 * been OOM killed to get memory anywhere.
1059 if (unlikely(test_thread_flag(TIF_MEMDIE)))
1061 if (current->flags & PF_EXITING) /* Let dying task have memory */
1066 * Determine if a loop is necessary if another thread is doing
1067 * read_mems_allowed_begin(). If at least one node remains unchanged and
1068 * tsk does not have a mempolicy, then an empty nodemask will not be
1069 * possible when mems_allowed is larger than a word.
1071 need_loop = task_has_mempolicy(tsk) ||
1072 !nodes_intersects(*newmems, tsk->mems_allowed);
1075 local_irq_disable();
1076 write_seqcount_begin(&tsk->mems_allowed_seq);
1079 nodes_or(tsk->mems_allowed, tsk->mems_allowed, *newmems);
1080 mpol_rebind_task(tsk, newmems, MPOL_REBIND_STEP1);
1082 mpol_rebind_task(tsk, newmems, MPOL_REBIND_STEP2);
1083 tsk->mems_allowed = *newmems;
1086 write_seqcount_end(&tsk->mems_allowed_seq);
1093 static void *cpuset_being_rebound;
1096 * update_tasks_nodemask - Update the nodemasks of tasks in the cpuset.
1097 * @cs: the cpuset in which each task's mems_allowed mask needs to be changed
1099 * Iterate through each task of @cs updating its mems_allowed to the
1100 * effective cpuset's. As this function is called with cpuset_mutex held,
1101 * cpuset membership stays stable.
1103 static void update_tasks_nodemask(struct cpuset *cs)
1105 static nodemask_t newmems; /* protected by cpuset_mutex */
1106 struct css_task_iter it;
1107 struct task_struct *task;
1109 cpuset_being_rebound = cs; /* causes mpol_dup() rebind */
1111 guarantee_online_mems(cs, &newmems);
1114 * The mpol_rebind_mm() call takes mmap_sem, which we couldn't
1115 * take while holding tasklist_lock. Forks can happen - the
1116 * mpol_dup() cpuset_being_rebound check will catch such forks,
1117 * and rebind their vma mempolicies too. Because we still hold
1118 * the global cpuset_mutex, we know that no other rebind effort
1119 * will be contending for the global variable cpuset_being_rebound.
1120 * It's ok if we rebind the same mm twice; mpol_rebind_mm()
1121 * is idempotent. Also migrate pages in each mm to new nodes.
1123 css_task_iter_start(&cs->css, &it);
1124 while ((task = css_task_iter_next(&it))) {
1125 struct mm_struct *mm;
1128 cpuset_change_task_nodemask(task, &newmems);
1130 mm = get_task_mm(task);
1134 migrate = is_memory_migrate(cs);
1136 mpol_rebind_mm(mm, &cs->mems_allowed);
1138 cpuset_migrate_mm(mm, &cs->old_mems_allowed, &newmems);
1142 css_task_iter_end(&it);
1145 * All the tasks' nodemasks have been updated, update
1146 * cs->old_mems_allowed.
1148 cs->old_mems_allowed = newmems;
1150 /* We're done rebinding vmas to this cpuset's new mems_allowed. */
1151 cpuset_being_rebound = NULL;
1155 * update_nodemasks_hier - Update effective nodemasks and tasks in the subtree
1156 * @cs: the cpuset to consider
1157 * @new_mems: a temp variable for calculating new effective_mems
1159 * When configured nodemask is changed, the effective nodemasks of this cpuset
1160 * and all its descendants need to be updated.
1162 * On legacy hiearchy, effective_mems will be the same with mems_allowed.
1164 * Called with cpuset_mutex held
1166 static void update_nodemasks_hier(struct cpuset *cs, nodemask_t *new_mems)
1169 struct cgroup_subsys_state *pos_css;
1172 cpuset_for_each_descendant_pre(cp, pos_css, cs) {
1173 struct cpuset *parent = parent_cs(cp);
1175 nodes_and(*new_mems, cp->mems_allowed, parent->effective_mems);
1178 * If it becomes empty, inherit the effective mask of the
1179 * parent, which is guaranteed to have some MEMs.
1181 if (cgroup_subsys_on_dfl(cpuset_cgrp_subsys) &&
1182 nodes_empty(*new_mems))
1183 *new_mems = parent->effective_mems;
1185 /* Skip the whole subtree if the nodemask remains the same. */
1186 if (nodes_equal(*new_mems, cp->effective_mems)) {
1187 pos_css = css_rightmost_descendant(pos_css);
1191 if (!css_tryget_online(&cp->css))
1195 spin_lock_irq(&callback_lock);
1196 cp->effective_mems = *new_mems;
1197 spin_unlock_irq(&callback_lock);
1199 WARN_ON(!cgroup_subsys_on_dfl(cpuset_cgrp_subsys) &&
1200 !nodes_equal(cp->mems_allowed, cp->effective_mems));
1202 update_tasks_nodemask(cp);
1211 * Handle user request to change the 'mems' memory placement
1212 * of a cpuset. Needs to validate the request, update the
1213 * cpusets mems_allowed, and for each task in the cpuset,
1214 * update mems_allowed and rebind task's mempolicy and any vma
1215 * mempolicies and if the cpuset is marked 'memory_migrate',
1216 * migrate the tasks pages to the new memory.
1218 * Call with cpuset_mutex held. May take callback_lock during call.
1219 * Will take tasklist_lock, scan tasklist for tasks in cpuset cs,
1220 * lock each such tasks mm->mmap_sem, scan its vma's and rebind
1221 * their mempolicies to the cpusets new mems_allowed.
1223 static int update_nodemask(struct cpuset *cs, struct cpuset *trialcs,
1229 * top_cpuset.mems_allowed tracks node_stats[N_MEMORY];
1232 if (cs == &top_cpuset) {
1238 * An empty mems_allowed is ok iff there are no tasks in the cpuset.
1239 * Since nodelist_parse() fails on an empty mask, we special case
1240 * that parsing. The validate_change() call ensures that cpusets
1241 * with tasks have memory.
1244 nodes_clear(trialcs->mems_allowed);
1246 retval = nodelist_parse(buf, trialcs->mems_allowed);
1250 if (!nodes_subset(trialcs->mems_allowed,
1251 top_cpuset.mems_allowed)) {
1257 if (nodes_equal(cs->mems_allowed, trialcs->mems_allowed)) {
1258 retval = 0; /* Too easy - nothing to do */
1261 retval = validate_change(cs, trialcs);
1265 spin_lock_irq(&callback_lock);
1266 cs->mems_allowed = trialcs->mems_allowed;
1267 spin_unlock_irq(&callback_lock);
1269 /* use trialcs->mems_allowed as a temp variable */
1270 update_nodemasks_hier(cs, &trialcs->mems_allowed);
1275 int current_cpuset_is_being_rebound(void)
1280 ret = task_cs(current) == cpuset_being_rebound;
1286 static int update_relax_domain_level(struct cpuset *cs, s64 val)
1289 if (val < -1 || val >= sched_domain_level_max)
1293 if (val != cs->relax_domain_level) {
1294 cs->relax_domain_level = val;
1295 if (!cpumask_empty(cs->cpus_allowed) &&
1296 is_sched_load_balance(cs))
1297 rebuild_sched_domains_unlocked();
1304 * update_tasks_flags - update the spread flags of tasks in the cpuset.
1305 * @cs: the cpuset in which each task's spread flags needs to be changed
1307 * Iterate through each task of @cs updating its spread flags. As this
1308 * function is called with cpuset_mutex held, cpuset membership stays
1311 static void update_tasks_flags(struct cpuset *cs)
1313 struct css_task_iter it;
1314 struct task_struct *task;
1316 css_task_iter_start(&cs->css, &it);
1317 while ((task = css_task_iter_next(&it)))
1318 cpuset_update_task_spread_flag(cs, task);
1319 css_task_iter_end(&it);
1323 * update_flag - read a 0 or a 1 in a file and update associated flag
1324 * bit: the bit to update (see cpuset_flagbits_t)
1325 * cs: the cpuset to update
1326 * turning_on: whether the flag is being set or cleared
1330 static int update_flag(cpuset_flagbits_t bit, struct cpuset *cs,
1333 struct cpuset *trialcs;
1334 int balance_flag_changed;
1335 int spread_flag_changed;
1338 trialcs = alloc_trial_cpuset(cs);
1343 set_bit(bit, &trialcs->flags);
1345 clear_bit(bit, &trialcs->flags);
1347 err = validate_change(cs, trialcs);
1351 balance_flag_changed = (is_sched_load_balance(cs) !=
1352 is_sched_load_balance(trialcs));
1354 spread_flag_changed = ((is_spread_slab(cs) != is_spread_slab(trialcs))
1355 || (is_spread_page(cs) != is_spread_page(trialcs)));
1357 spin_lock_irq(&callback_lock);
1358 cs->flags = trialcs->flags;
1359 spin_unlock_irq(&callback_lock);
1361 if (!cpumask_empty(trialcs->cpus_allowed) && balance_flag_changed)
1362 rebuild_sched_domains_unlocked();
1364 if (spread_flag_changed)
1365 update_tasks_flags(cs);
1367 free_trial_cpuset(trialcs);
1372 * Frequency meter - How fast is some event occurring?
1374 * These routines manage a digitally filtered, constant time based,
1375 * event frequency meter. There are four routines:
1376 * fmeter_init() - initialize a frequency meter.
1377 * fmeter_markevent() - called each time the event happens.
1378 * fmeter_getrate() - returns the recent rate of such events.
1379 * fmeter_update() - internal routine used to update fmeter.
1381 * A common data structure is passed to each of these routines,
1382 * which is used to keep track of the state required to manage the
1383 * frequency meter and its digital filter.
1385 * The filter works on the number of events marked per unit time.
1386 * The filter is single-pole low-pass recursive (IIR). The time unit
1387 * is 1 second. Arithmetic is done using 32-bit integers scaled to
1388 * simulate 3 decimal digits of precision (multiplied by 1000).
1390 * With an FM_COEF of 933, and a time base of 1 second, the filter
1391 * has a half-life of 10 seconds, meaning that if the events quit
1392 * happening, then the rate returned from the fmeter_getrate()
1393 * will be cut in half each 10 seconds, until it converges to zero.
1395 * It is not worth doing a real infinitely recursive filter. If more
1396 * than FM_MAXTICKS ticks have elapsed since the last filter event,
1397 * just compute FM_MAXTICKS ticks worth, by which point the level
1400 * Limit the count of unprocessed events to FM_MAXCNT, so as to avoid
1401 * arithmetic overflow in the fmeter_update() routine.
1403 * Given the simple 32 bit integer arithmetic used, this meter works
1404 * best for reporting rates between one per millisecond (msec) and
1405 * one per 32 (approx) seconds. At constant rates faster than one
1406 * per msec it maxes out at values just under 1,000,000. At constant
1407 * rates between one per msec, and one per second it will stabilize
1408 * to a value N*1000, where N is the rate of events per second.
1409 * At constant rates between one per second and one per 32 seconds,
1410 * it will be choppy, moving up on the seconds that have an event,
1411 * and then decaying until the next event. At rates slower than
1412 * about one in 32 seconds, it decays all the way back to zero between
1416 #define FM_COEF 933 /* coefficient for half-life of 10 secs */
1417 #define FM_MAXTICKS ((time_t)99) /* useless computing more ticks than this */
1418 #define FM_MAXCNT 1000000 /* limit cnt to avoid overflow */
1419 #define FM_SCALE 1000 /* faux fixed point scale */
1421 /* Initialize a frequency meter */
1422 static void fmeter_init(struct fmeter *fmp)
1427 spin_lock_init(&fmp->lock);
1430 /* Internal meter update - process cnt events and update value */
1431 static void fmeter_update(struct fmeter *fmp)
1433 time_t now = get_seconds();
1434 time_t ticks = now - fmp->time;
1439 ticks = min(FM_MAXTICKS, ticks);
1441 fmp->val = (FM_COEF * fmp->val) / FM_SCALE;
1444 fmp->val += ((FM_SCALE - FM_COEF) * fmp->cnt) / FM_SCALE;
1448 /* Process any previous ticks, then bump cnt by one (times scale). */
1449 static void fmeter_markevent(struct fmeter *fmp)
1451 spin_lock(&fmp->lock);
1453 fmp->cnt = min(FM_MAXCNT, fmp->cnt + FM_SCALE);
1454 spin_unlock(&fmp->lock);
1457 /* Process any previous ticks, then return current value. */
1458 static int fmeter_getrate(struct fmeter *fmp)
1462 spin_lock(&fmp->lock);
1465 spin_unlock(&fmp->lock);
1469 static struct cpuset *cpuset_attach_old_cs;
1471 /* Called by cgroups to determine if a cpuset is usable; cpuset_mutex held */
1472 static int cpuset_can_attach(struct cgroup_taskset *tset)
1474 struct cgroup_subsys_state *css;
1476 struct task_struct *task;
1479 /* used later by cpuset_attach() */
1480 cpuset_attach_old_cs = task_cs(cgroup_taskset_first(tset, &css));
1483 mutex_lock(&cpuset_mutex);
1485 /* allow moving tasks into an empty cpuset if on default hierarchy */
1487 if (!cgroup_subsys_on_dfl(cpuset_cgrp_subsys) &&
1488 (cpumask_empty(cs->cpus_allowed) || nodes_empty(cs->mems_allowed)))
1491 cgroup_taskset_for_each(task, css, tset) {
1492 ret = task_can_attach(task, cs->cpus_allowed);
1495 ret = security_task_setscheduler(task);
1501 * Mark attach is in progress. This makes validate_change() fail
1502 * changes which zero cpus/mems_allowed.
1504 cs->attach_in_progress++;
1507 mutex_unlock(&cpuset_mutex);
1511 static void cpuset_cancel_attach(struct cgroup_taskset *tset)
1513 struct cgroup_subsys_state *css;
1516 cgroup_taskset_first(tset, &css);
1519 mutex_lock(&cpuset_mutex);
1520 css_cs(css)->attach_in_progress--;
1521 mutex_unlock(&cpuset_mutex);
1525 * Protected by cpuset_mutex. cpus_attach is used only by cpuset_attach()
1526 * but we can't allocate it dynamically there. Define it global and
1527 * allocate from cpuset_init().
1529 static cpumask_var_t cpus_attach;
1531 static void cpuset_attach(struct cgroup_taskset *tset)
1533 /* static buf protected by cpuset_mutex */
1534 static nodemask_t cpuset_attach_nodemask_to;
1535 struct task_struct *task;
1536 struct task_struct *leader;
1537 struct cgroup_subsys_state *css;
1539 struct cpuset *oldcs = cpuset_attach_old_cs;
1541 cgroup_taskset_first(tset, &css);
1544 mutex_lock(&cpuset_mutex);
1546 /* prepare for attach */
1547 if (cs == &top_cpuset)
1548 cpumask_copy(cpus_attach, cpu_possible_mask);
1550 guarantee_online_cpus(cs, cpus_attach);
1552 guarantee_online_mems(cs, &cpuset_attach_nodemask_to);
1554 cgroup_taskset_for_each(task, css, tset) {
1556 * can_attach beforehand should guarantee that this doesn't
1557 * fail. TODO: have a better way to handle failure here
1559 WARN_ON_ONCE(set_cpus_allowed_ptr(task, cpus_attach));
1561 cpuset_change_task_nodemask(task, &cpuset_attach_nodemask_to);
1562 cpuset_update_task_spread_flag(cs, task);
1566 * Change mm for all threadgroup leaders. This is expensive and may
1567 * sleep and should be moved outside migration path proper.
1569 cpuset_attach_nodemask_to = cs->effective_mems;
1570 cgroup_taskset_for_each_leader(leader, css, tset) {
1571 struct mm_struct *mm = get_task_mm(leader);
1574 mpol_rebind_mm(mm, &cpuset_attach_nodemask_to);
1577 * old_mems_allowed is the same with mems_allowed
1578 * here, except if this task is being moved
1579 * automatically due to hotplug. In that case
1580 * @mems_allowed has been updated and is empty, so
1581 * @old_mems_allowed is the right nodesets that we
1584 if (is_memory_migrate(cs))
1585 cpuset_migrate_mm(mm, &oldcs->old_mems_allowed,
1586 &cpuset_attach_nodemask_to);
1592 cs->old_mems_allowed = cpuset_attach_nodemask_to;
1594 cs->attach_in_progress--;
1595 if (!cs->attach_in_progress)
1596 wake_up(&cpuset_attach_wq);
1598 mutex_unlock(&cpuset_mutex);
1601 /* The various types of files and directories in a cpuset file system */
1604 FILE_MEMORY_MIGRATE,
1607 FILE_EFFECTIVE_CPULIST,
1608 FILE_EFFECTIVE_MEMLIST,
1612 FILE_SCHED_LOAD_BALANCE,
1613 FILE_SCHED_RELAX_DOMAIN_LEVEL,
1614 FILE_MEMORY_PRESSURE_ENABLED,
1615 FILE_MEMORY_PRESSURE,
1618 } cpuset_filetype_t;
1620 static int cpuset_write_u64(struct cgroup_subsys_state *css, struct cftype *cft,
1623 struct cpuset *cs = css_cs(css);
1624 cpuset_filetype_t type = cft->private;
1628 mutex_lock(&cpuset_mutex);
1629 if (!is_cpuset_online(cs)) {
1635 case FILE_CPU_EXCLUSIVE:
1636 retval = update_flag(CS_CPU_EXCLUSIVE, cs, val);
1638 case FILE_MEM_EXCLUSIVE:
1639 retval = update_flag(CS_MEM_EXCLUSIVE, cs, val);
1641 case FILE_MEM_HARDWALL:
1642 retval = update_flag(CS_MEM_HARDWALL, cs, val);
1644 case FILE_SCHED_LOAD_BALANCE:
1645 retval = update_flag(CS_SCHED_LOAD_BALANCE, cs, val);
1647 case FILE_MEMORY_MIGRATE:
1648 retval = update_flag(CS_MEMORY_MIGRATE, cs, val);
1650 case FILE_MEMORY_PRESSURE_ENABLED:
1651 cpuset_memory_pressure_enabled = !!val;
1653 case FILE_SPREAD_PAGE:
1654 retval = update_flag(CS_SPREAD_PAGE, cs, val);
1656 case FILE_SPREAD_SLAB:
1657 retval = update_flag(CS_SPREAD_SLAB, cs, val);
1664 mutex_unlock(&cpuset_mutex);
1669 static int cpuset_write_s64(struct cgroup_subsys_state *css, struct cftype *cft,
1672 struct cpuset *cs = css_cs(css);
1673 cpuset_filetype_t type = cft->private;
1674 int retval = -ENODEV;
1677 mutex_lock(&cpuset_mutex);
1678 if (!is_cpuset_online(cs))
1682 case FILE_SCHED_RELAX_DOMAIN_LEVEL:
1683 retval = update_relax_domain_level(cs, val);
1690 mutex_unlock(&cpuset_mutex);
1696 * Common handling for a write to a "cpus" or "mems" file.
1698 static ssize_t cpuset_write_resmask(struct kernfs_open_file *of,
1699 char *buf, size_t nbytes, loff_t off)
1701 struct cpuset *cs = css_cs(of_css(of));
1702 struct cpuset *trialcs;
1703 int retval = -ENODEV;
1705 buf = strstrip(buf);
1708 * CPU or memory hotunplug may leave @cs w/o any execution
1709 * resources, in which case the hotplug code asynchronously updates
1710 * configuration and transfers all tasks to the nearest ancestor
1711 * which can execute.
1713 * As writes to "cpus" or "mems" may restore @cs's execution
1714 * resources, wait for the previously scheduled operations before
1715 * proceeding, so that we don't end up keep removing tasks added
1716 * after execution capability is restored.
1718 * cpuset_hotplug_work calls back into cgroup core via
1719 * cgroup_transfer_tasks() and waiting for it from a cgroupfs
1720 * operation like this one can lead to a deadlock through kernfs
1721 * active_ref protection. Let's break the protection. Losing the
1722 * protection is okay as we check whether @cs is online after
1723 * grabbing cpuset_mutex anyway. This only happens on the legacy
1727 kernfs_break_active_protection(of->kn);
1728 flush_work(&cpuset_hotplug_work);
1731 mutex_lock(&cpuset_mutex);
1732 if (!is_cpuset_online(cs))
1735 trialcs = alloc_trial_cpuset(cs);
1741 switch (of_cft(of)->private) {
1743 retval = update_cpumask(cs, trialcs, buf);
1746 retval = update_nodemask(cs, trialcs, buf);
1753 free_trial_cpuset(trialcs);
1755 mutex_unlock(&cpuset_mutex);
1757 kernfs_unbreak_active_protection(of->kn);
1759 flush_workqueue(cpuset_migrate_mm_wq);
1760 return retval ?: nbytes;
1764 * These ascii lists should be read in a single call, by using a user
1765 * buffer large enough to hold the entire map. If read in smaller
1766 * chunks, there is no guarantee of atomicity. Since the display format
1767 * used, list of ranges of sequential numbers, is variable length,
1768 * and since these maps can change value dynamically, one could read
1769 * gibberish by doing partial reads while a list was changing.
1771 static int cpuset_common_seq_show(struct seq_file *sf, void *v)
1773 struct cpuset *cs = css_cs(seq_css(sf));
1774 cpuset_filetype_t type = seq_cft(sf)->private;
1777 spin_lock_irq(&callback_lock);
1781 seq_printf(sf, "%*pbl\n", cpumask_pr_args(cs->cpus_requested));
1784 seq_printf(sf, "%*pbl\n", nodemask_pr_args(&cs->mems_allowed));
1786 case FILE_EFFECTIVE_CPULIST:
1787 seq_printf(sf, "%*pbl\n", cpumask_pr_args(cs->effective_cpus));
1789 case FILE_EFFECTIVE_MEMLIST:
1790 seq_printf(sf, "%*pbl\n", nodemask_pr_args(&cs->effective_mems));
1796 spin_unlock_irq(&callback_lock);
1800 static u64 cpuset_read_u64(struct cgroup_subsys_state *css, struct cftype *cft)
1802 struct cpuset *cs = css_cs(css);
1803 cpuset_filetype_t type = cft->private;
1805 case FILE_CPU_EXCLUSIVE:
1806 return is_cpu_exclusive(cs);
1807 case FILE_MEM_EXCLUSIVE:
1808 return is_mem_exclusive(cs);
1809 case FILE_MEM_HARDWALL:
1810 return is_mem_hardwall(cs);
1811 case FILE_SCHED_LOAD_BALANCE:
1812 return is_sched_load_balance(cs);
1813 case FILE_MEMORY_MIGRATE:
1814 return is_memory_migrate(cs);
1815 case FILE_MEMORY_PRESSURE_ENABLED:
1816 return cpuset_memory_pressure_enabled;
1817 case FILE_MEMORY_PRESSURE:
1818 return fmeter_getrate(&cs->fmeter);
1819 case FILE_SPREAD_PAGE:
1820 return is_spread_page(cs);
1821 case FILE_SPREAD_SLAB:
1822 return is_spread_slab(cs);
1827 /* Unreachable but makes gcc happy */
1831 static s64 cpuset_read_s64(struct cgroup_subsys_state *css, struct cftype *cft)
1833 struct cpuset *cs = css_cs(css);
1834 cpuset_filetype_t type = cft->private;
1836 case FILE_SCHED_RELAX_DOMAIN_LEVEL:
1837 return cs->relax_domain_level;
1842 /* Unrechable but makes gcc happy */
1848 * for the common functions, 'private' gives the type of file
1851 static struct cftype files[] = {
1854 .seq_show = cpuset_common_seq_show,
1855 .write = cpuset_write_resmask,
1856 .max_write_len = (100U + 6 * NR_CPUS),
1857 .private = FILE_CPULIST,
1862 .seq_show = cpuset_common_seq_show,
1863 .write = cpuset_write_resmask,
1864 .max_write_len = (100U + 6 * MAX_NUMNODES),
1865 .private = FILE_MEMLIST,
1869 .name = "effective_cpus",
1870 .seq_show = cpuset_common_seq_show,
1871 .private = FILE_EFFECTIVE_CPULIST,
1875 .name = "effective_mems",
1876 .seq_show = cpuset_common_seq_show,
1877 .private = FILE_EFFECTIVE_MEMLIST,
1881 .name = "cpu_exclusive",
1882 .read_u64 = cpuset_read_u64,
1883 .write_u64 = cpuset_write_u64,
1884 .private = FILE_CPU_EXCLUSIVE,
1888 .name = "mem_exclusive",
1889 .read_u64 = cpuset_read_u64,
1890 .write_u64 = cpuset_write_u64,
1891 .private = FILE_MEM_EXCLUSIVE,
1895 .name = "mem_hardwall",
1896 .read_u64 = cpuset_read_u64,
1897 .write_u64 = cpuset_write_u64,
1898 .private = FILE_MEM_HARDWALL,
1902 .name = "sched_load_balance",
1903 .read_u64 = cpuset_read_u64,
1904 .write_u64 = cpuset_write_u64,
1905 .private = FILE_SCHED_LOAD_BALANCE,
1909 .name = "sched_relax_domain_level",
1910 .read_s64 = cpuset_read_s64,
1911 .write_s64 = cpuset_write_s64,
1912 .private = FILE_SCHED_RELAX_DOMAIN_LEVEL,
1916 .name = "memory_migrate",
1917 .read_u64 = cpuset_read_u64,
1918 .write_u64 = cpuset_write_u64,
1919 .private = FILE_MEMORY_MIGRATE,
1923 .name = "memory_pressure",
1924 .read_u64 = cpuset_read_u64,
1925 .private = FILE_MEMORY_PRESSURE,
1929 .name = "memory_spread_page",
1930 .read_u64 = cpuset_read_u64,
1931 .write_u64 = cpuset_write_u64,
1932 .private = FILE_SPREAD_PAGE,
1936 .name = "memory_spread_slab",
1937 .read_u64 = cpuset_read_u64,
1938 .write_u64 = cpuset_write_u64,
1939 .private = FILE_SPREAD_SLAB,
1943 .name = "memory_pressure_enabled",
1944 .flags = CFTYPE_ONLY_ON_ROOT,
1945 .read_u64 = cpuset_read_u64,
1946 .write_u64 = cpuset_write_u64,
1947 .private = FILE_MEMORY_PRESSURE_ENABLED,
1954 * cpuset_css_alloc - allocate a cpuset css
1955 * cgrp: control group that the new cpuset will be part of
1958 static struct cgroup_subsys_state *
1959 cpuset_css_alloc(struct cgroup_subsys_state *parent_css)
1964 return &top_cpuset.css;
1966 cs = kzalloc(sizeof(*cs), GFP_KERNEL);
1968 return ERR_PTR(-ENOMEM);
1969 if (!alloc_cpumask_var(&cs->cpus_allowed, GFP_KERNEL))
1971 if (!alloc_cpumask_var(&cs->effective_cpus, GFP_KERNEL))
1972 goto error_effective;
1973 if (!alloc_cpumask_var(&cs->cpus_requested, GFP_KERNEL))
1974 goto error_requested;
1976 set_bit(CS_SCHED_LOAD_BALANCE, &cs->flags);
1977 cpumask_clear(cs->cpus_allowed);
1978 cpumask_clear(cs->cpus_requested);
1979 nodes_clear(cs->mems_allowed);
1980 cpumask_clear(cs->effective_cpus);
1981 nodes_clear(cs->effective_mems);
1982 fmeter_init(&cs->fmeter);
1983 cs->relax_domain_level = -1;
1988 free_cpumask_var(cs->effective_cpus);
1990 free_cpumask_var(cs->cpus_allowed);
1993 return ERR_PTR(-ENOMEM);
1996 static int cpuset_css_online(struct cgroup_subsys_state *css)
1998 struct cpuset *cs = css_cs(css);
1999 struct cpuset *parent = parent_cs(cs);
2000 struct cpuset *tmp_cs;
2001 struct cgroup_subsys_state *pos_css;
2006 mutex_lock(&cpuset_mutex);
2008 set_bit(CS_ONLINE, &cs->flags);
2009 if (is_spread_page(parent))
2010 set_bit(CS_SPREAD_PAGE, &cs->flags);
2011 if (is_spread_slab(parent))
2012 set_bit(CS_SPREAD_SLAB, &cs->flags);
2016 spin_lock_irq(&callback_lock);
2017 if (cgroup_subsys_on_dfl(cpuset_cgrp_subsys)) {
2018 cpumask_copy(cs->effective_cpus, parent->effective_cpus);
2019 cs->effective_mems = parent->effective_mems;
2021 spin_unlock_irq(&callback_lock);
2023 if (!test_bit(CGRP_CPUSET_CLONE_CHILDREN, &css->cgroup->flags))
2027 * Clone @parent's configuration if CGRP_CPUSET_CLONE_CHILDREN is
2028 * set. This flag handling is implemented in cgroup core for
2029 * histrical reasons - the flag may be specified during mount.
2031 * Currently, if any sibling cpusets have exclusive cpus or mem, we
2032 * refuse to clone the configuration - thereby refusing the task to
2033 * be entered, and as a result refusing the sys_unshare() or
2034 * clone() which initiated it. If this becomes a problem for some
2035 * users who wish to allow that scenario, then this could be
2036 * changed to grant parent->cpus_allowed-sibling_cpus_exclusive
2037 * (and likewise for mems) to the new cgroup.
2040 cpuset_for_each_child(tmp_cs, pos_css, parent) {
2041 if (is_mem_exclusive(tmp_cs) || is_cpu_exclusive(tmp_cs)) {
2048 spin_lock_irq(&callback_lock);
2049 cs->mems_allowed = parent->mems_allowed;
2050 cs->effective_mems = parent->mems_allowed;
2051 cpumask_copy(cs->cpus_allowed, parent->cpus_allowed);
2052 cpumask_copy(cs->cpus_requested, parent->cpus_requested);
2053 cpumask_copy(cs->effective_cpus, parent->cpus_allowed);
2054 spin_unlock_irq(&callback_lock);
2056 mutex_unlock(&cpuset_mutex);
2061 * If the cpuset being removed has its flag 'sched_load_balance'
2062 * enabled, then simulate turning sched_load_balance off, which
2063 * will call rebuild_sched_domains_unlocked().
2066 static void cpuset_css_offline(struct cgroup_subsys_state *css)
2068 struct cpuset *cs = css_cs(css);
2071 mutex_lock(&cpuset_mutex);
2073 if (is_sched_load_balance(cs))
2074 update_flag(CS_SCHED_LOAD_BALANCE, cs, 0);
2077 clear_bit(CS_ONLINE, &cs->flags);
2079 mutex_unlock(&cpuset_mutex);
2083 static void cpuset_css_free(struct cgroup_subsys_state *css)
2085 struct cpuset *cs = css_cs(css);
2087 free_cpumask_var(cs->effective_cpus);
2088 free_cpumask_var(cs->cpus_allowed);
2089 free_cpumask_var(cs->cpus_requested);
2093 static void cpuset_bind(struct cgroup_subsys_state *root_css)
2095 mutex_lock(&cpuset_mutex);
2096 spin_lock_irq(&callback_lock);
2098 if (cgroup_subsys_on_dfl(cpuset_cgrp_subsys)) {
2099 cpumask_copy(top_cpuset.cpus_allowed, cpu_possible_mask);
2100 top_cpuset.mems_allowed = node_possible_map;
2102 cpumask_copy(top_cpuset.cpus_allowed,
2103 top_cpuset.effective_cpus);
2104 top_cpuset.mems_allowed = top_cpuset.effective_mems;
2107 spin_unlock_irq(&callback_lock);
2108 mutex_unlock(&cpuset_mutex);
2112 * Make sure the new task conform to the current state of its parent,
2113 * which could have been changed by cpuset just after it inherits the
2114 * state from the parent and before it sits on the cgroup's task list.
2116 void cpuset_fork(struct task_struct *task, void *priv)
2118 if (task_css_is_root(task, cpuset_cgrp_id))
2121 set_cpus_allowed_ptr(task, ¤t->cpus_allowed);
2122 task->mems_allowed = current->mems_allowed;
2125 struct cgroup_subsys cpuset_cgrp_subsys = {
2126 .css_alloc = cpuset_css_alloc,
2127 .css_online = cpuset_css_online,
2128 .css_offline = cpuset_css_offline,
2129 .css_free = cpuset_css_free,
2130 .can_attach = cpuset_can_attach,
2131 .cancel_attach = cpuset_cancel_attach,
2132 .attach = cpuset_attach,
2133 .post_attach = cpuset_post_attach,
2134 .bind = cpuset_bind,
2135 .fork = cpuset_fork,
2136 .legacy_cftypes = files,
2141 * cpuset_init - initialize cpusets at system boot
2143 * Description: Initialize top_cpuset and the cpuset internal file system,
2146 int __init cpuset_init(void)
2150 if (!alloc_cpumask_var(&top_cpuset.cpus_allowed, GFP_KERNEL))
2152 if (!alloc_cpumask_var(&top_cpuset.effective_cpus, GFP_KERNEL))
2154 if (!alloc_cpumask_var(&top_cpuset.cpus_requested, GFP_KERNEL))
2157 cpumask_setall(top_cpuset.cpus_allowed);
2158 cpumask_setall(top_cpuset.cpus_requested);
2159 nodes_setall(top_cpuset.mems_allowed);
2160 cpumask_setall(top_cpuset.effective_cpus);
2161 nodes_setall(top_cpuset.effective_mems);
2163 fmeter_init(&top_cpuset.fmeter);
2164 set_bit(CS_SCHED_LOAD_BALANCE, &top_cpuset.flags);
2165 top_cpuset.relax_domain_level = -1;
2167 err = register_filesystem(&cpuset_fs_type);
2171 if (!alloc_cpumask_var(&cpus_attach, GFP_KERNEL))
2178 * If CPU and/or memory hotplug handlers, below, unplug any CPUs
2179 * or memory nodes, we need to walk over the cpuset hierarchy,
2180 * removing that CPU or node from all cpusets. If this removes the
2181 * last CPU or node from a cpuset, then move the tasks in the empty
2182 * cpuset to its next-highest non-empty parent.
2184 static void remove_tasks_in_empty_cpuset(struct cpuset *cs)
2186 struct cpuset *parent;
2189 * Find its next-highest non-empty parent, (top cpuset
2190 * has online cpus, so can't be empty).
2192 parent = parent_cs(cs);
2193 while (cpumask_empty(parent->cpus_allowed) ||
2194 nodes_empty(parent->mems_allowed))
2195 parent = parent_cs(parent);
2197 if (cgroup_transfer_tasks(parent->css.cgroup, cs->css.cgroup)) {
2198 pr_err("cpuset: failed to transfer tasks out of empty cpuset ");
2199 pr_cont_cgroup_name(cs->css.cgroup);
2205 hotplug_update_tasks_legacy(struct cpuset *cs,
2206 struct cpumask *new_cpus, nodemask_t *new_mems,
2207 bool cpus_updated, bool mems_updated)
2211 spin_lock_irq(&callback_lock);
2212 cpumask_copy(cs->cpus_allowed, new_cpus);
2213 cpumask_copy(cs->effective_cpus, new_cpus);
2214 cs->mems_allowed = *new_mems;
2215 cs->effective_mems = *new_mems;
2216 spin_unlock_irq(&callback_lock);
2219 * Don't call update_tasks_cpumask() if the cpuset becomes empty,
2220 * as the tasks will be migratecd to an ancestor.
2222 if (cpus_updated && !cpumask_empty(cs->cpus_allowed))
2223 update_tasks_cpumask(cs);
2224 if (mems_updated && !nodes_empty(cs->mems_allowed))
2225 update_tasks_nodemask(cs);
2227 is_empty = cpumask_empty(cs->cpus_allowed) ||
2228 nodes_empty(cs->mems_allowed);
2230 mutex_unlock(&cpuset_mutex);
2233 * Move tasks to the nearest ancestor with execution resources,
2234 * This is full cgroup operation which will also call back into
2235 * cpuset. Should be done outside any lock.
2238 remove_tasks_in_empty_cpuset(cs);
2240 mutex_lock(&cpuset_mutex);
2244 hotplug_update_tasks(struct cpuset *cs,
2245 struct cpumask *new_cpus, nodemask_t *new_mems,
2246 bool cpus_updated, bool mems_updated)
2248 if (cpumask_empty(new_cpus))
2249 cpumask_copy(new_cpus, parent_cs(cs)->effective_cpus);
2250 if (nodes_empty(*new_mems))
2251 *new_mems = parent_cs(cs)->effective_mems;
2253 spin_lock_irq(&callback_lock);
2254 cpumask_copy(cs->effective_cpus, new_cpus);
2255 cs->effective_mems = *new_mems;
2256 spin_unlock_irq(&callback_lock);
2259 update_tasks_cpumask(cs);
2261 update_tasks_nodemask(cs);
2265 * cpuset_hotplug_update_tasks - update tasks in a cpuset for hotunplug
2266 * @cs: cpuset in interest
2268 * Compare @cs's cpu and mem masks against top_cpuset and if some have gone
2269 * offline, update @cs accordingly. If @cs ends up with no CPU or memory,
2270 * all its tasks are moved to the nearest ancestor with both resources.
2272 static void cpuset_hotplug_update_tasks(struct cpuset *cs)
2274 static cpumask_t new_cpus;
2275 static nodemask_t new_mems;
2279 wait_event(cpuset_attach_wq, cs->attach_in_progress == 0);
2281 mutex_lock(&cpuset_mutex);
2284 * We have raced with task attaching. We wait until attaching
2285 * is finished, so we won't attach a task to an empty cpuset.
2287 if (cs->attach_in_progress) {
2288 mutex_unlock(&cpuset_mutex);
2292 cpumask_and(&new_cpus, cs->cpus_requested,
2293 parent_cs(cs)->effective_cpus);
2294 nodes_and(new_mems, cs->mems_allowed, parent_cs(cs)->effective_mems);
2296 cpus_updated = !cpumask_equal(&new_cpus, cs->effective_cpus);
2297 mems_updated = !nodes_equal(new_mems, cs->effective_mems);
2299 if (cgroup_subsys_on_dfl(cpuset_cgrp_subsys))
2300 hotplug_update_tasks(cs, &new_cpus, &new_mems,
2301 cpus_updated, mems_updated);
2303 hotplug_update_tasks_legacy(cs, &new_cpus, &new_mems,
2304 cpus_updated, mems_updated);
2306 mutex_unlock(&cpuset_mutex);
2309 static bool force_rebuild;
2311 void cpuset_force_rebuild(void)
2313 force_rebuild = true;
2317 * cpuset_hotplug_workfn - handle CPU/memory hotunplug for a cpuset
2319 * This function is called after either CPU or memory configuration has
2320 * changed and updates cpuset accordingly. The top_cpuset is always
2321 * synchronized to cpu_active_mask and N_MEMORY, which is necessary in
2322 * order to make cpusets transparent (of no affect) on systems that are
2323 * actively using CPU hotplug but making no active use of cpusets.
2325 * Non-root cpusets are only affected by offlining. If any CPUs or memory
2326 * nodes have been taken down, cpuset_hotplug_update_tasks() is invoked on
2329 * Note that CPU offlining during suspend is ignored. We don't modify
2330 * cpusets across suspend/resume cycles at all.
2332 static void cpuset_hotplug_workfn(struct work_struct *work)
2334 static cpumask_t new_cpus;
2335 static nodemask_t new_mems;
2336 bool cpus_updated, mems_updated;
2337 bool on_dfl = cgroup_subsys_on_dfl(cpuset_cgrp_subsys);
2339 mutex_lock(&cpuset_mutex);
2341 /* fetch the available cpus/mems and find out which changed how */
2342 cpumask_copy(&new_cpus, cpu_active_mask);
2343 new_mems = node_states[N_MEMORY];
2345 cpus_updated = !cpumask_equal(top_cpuset.effective_cpus, &new_cpus);
2346 mems_updated = !nodes_equal(top_cpuset.effective_mems, new_mems);
2348 /* synchronize cpus_allowed to cpu_active_mask */
2350 spin_lock_irq(&callback_lock);
2352 cpumask_copy(top_cpuset.cpus_allowed, &new_cpus);
2353 cpumask_copy(top_cpuset.effective_cpus, &new_cpus);
2354 spin_unlock_irq(&callback_lock);
2355 /* we don't mess with cpumasks of tasks in top_cpuset */
2358 /* synchronize mems_allowed to N_MEMORY */
2360 spin_lock_irq(&callback_lock);
2362 top_cpuset.mems_allowed = new_mems;
2363 top_cpuset.effective_mems = new_mems;
2364 spin_unlock_irq(&callback_lock);
2365 update_tasks_nodemask(&top_cpuset);
2368 mutex_unlock(&cpuset_mutex);
2370 /* if cpus or mems changed, we need to propagate to descendants */
2371 if (cpus_updated || mems_updated) {
2373 struct cgroup_subsys_state *pos_css;
2376 cpuset_for_each_descendant_pre(cs, pos_css, &top_cpuset) {
2377 if (cs == &top_cpuset || !css_tryget_online(&cs->css))
2381 cpuset_hotplug_update_tasks(cs);
2389 /* rebuild sched domains if cpus_allowed has changed */
2390 if (cpus_updated || force_rebuild) {
2391 force_rebuild = false;
2392 rebuild_sched_domains();
2396 void cpuset_update_active_cpus(bool cpu_online)
2399 * We're inside cpu hotplug critical region which usually nests
2400 * inside cgroup synchronization. Bounce actual hotplug processing
2401 * to a work item to avoid reverse locking order.
2403 * We still need to do partition_sched_domains() synchronously;
2404 * otherwise, the scheduler will get confused and put tasks to the
2405 * dead CPU. Fall back to the default single domain.
2406 * cpuset_hotplug_workfn() will rebuild it as necessary.
2408 partition_sched_domains(1, NULL, NULL);
2409 schedule_work(&cpuset_hotplug_work);
2412 void cpuset_wait_for_hotplug(void)
2414 flush_work(&cpuset_hotplug_work);
2418 * Keep top_cpuset.mems_allowed tracking node_states[N_MEMORY].
2419 * Call this routine anytime after node_states[N_MEMORY] changes.
2420 * See cpuset_update_active_cpus() for CPU hotplug handling.
2422 static int cpuset_track_online_nodes(struct notifier_block *self,
2423 unsigned long action, void *arg)
2425 schedule_work(&cpuset_hotplug_work);
2429 static struct notifier_block cpuset_track_online_nodes_nb = {
2430 .notifier_call = cpuset_track_online_nodes,
2431 .priority = 10, /* ??! */
2435 * cpuset_init_smp - initialize cpus_allowed
2437 * Description: Finish top cpuset after cpu, node maps are initialized
2439 void __init cpuset_init_smp(void)
2441 cpumask_copy(top_cpuset.cpus_allowed, cpu_active_mask);
2442 top_cpuset.mems_allowed = node_states[N_MEMORY];
2443 top_cpuset.old_mems_allowed = top_cpuset.mems_allowed;
2445 cpumask_copy(top_cpuset.effective_cpus, cpu_active_mask);
2446 top_cpuset.effective_mems = node_states[N_MEMORY];
2448 register_hotmemory_notifier(&cpuset_track_online_nodes_nb);
2450 cpuset_migrate_mm_wq = alloc_ordered_workqueue("cpuset_migrate_mm", 0);
2451 BUG_ON(!cpuset_migrate_mm_wq);
2455 * cpuset_cpus_allowed - return cpus_allowed mask from a tasks cpuset.
2456 * @tsk: pointer to task_struct from which to obtain cpuset->cpus_allowed.
2457 * @pmask: pointer to struct cpumask variable to receive cpus_allowed set.
2459 * Description: Returns the cpumask_var_t cpus_allowed of the cpuset
2460 * attached to the specified @tsk. Guaranteed to return some non-empty
2461 * subset of cpu_online_mask, even if this means going outside the
2465 void cpuset_cpus_allowed(struct task_struct *tsk, struct cpumask *pmask)
2467 unsigned long flags;
2469 spin_lock_irqsave(&callback_lock, flags);
2471 guarantee_online_cpus(task_cs(tsk), pmask);
2473 spin_unlock_irqrestore(&callback_lock, flags);
2476 void cpuset_cpus_allowed_fallback(struct task_struct *tsk)
2479 do_set_cpus_allowed(tsk, task_cs(tsk)->effective_cpus);
2483 * We own tsk->cpus_allowed, nobody can change it under us.
2485 * But we used cs && cs->cpus_allowed lockless and thus can
2486 * race with cgroup_attach_task() or update_cpumask() and get
2487 * the wrong tsk->cpus_allowed. However, both cases imply the
2488 * subsequent cpuset_change_cpumask()->set_cpus_allowed_ptr()
2489 * which takes task_rq_lock().
2491 * If we are called after it dropped the lock we must see all
2492 * changes in tsk_cs()->cpus_allowed. Otherwise we can temporary
2493 * set any mask even if it is not right from task_cs() pov,
2494 * the pending set_cpus_allowed_ptr() will fix things.
2496 * select_fallback_rq() will fix things ups and set cpu_possible_mask
2501 void __init cpuset_init_current_mems_allowed(void)
2503 nodes_setall(current->mems_allowed);
2507 * cpuset_mems_allowed - return mems_allowed mask from a tasks cpuset.
2508 * @tsk: pointer to task_struct from which to obtain cpuset->mems_allowed.
2510 * Description: Returns the nodemask_t mems_allowed of the cpuset
2511 * attached to the specified @tsk. Guaranteed to return some non-empty
2512 * subset of node_states[N_MEMORY], even if this means going outside the
2516 nodemask_t cpuset_mems_allowed(struct task_struct *tsk)
2519 unsigned long flags;
2521 spin_lock_irqsave(&callback_lock, flags);
2523 guarantee_online_mems(task_cs(tsk), &mask);
2525 spin_unlock_irqrestore(&callback_lock, flags);
2531 * cpuset_nodemask_valid_mems_allowed - check nodemask vs. curremt mems_allowed
2532 * @nodemask: the nodemask to be checked
2534 * Are any of the nodes in the nodemask allowed in current->mems_allowed?
2536 int cpuset_nodemask_valid_mems_allowed(nodemask_t *nodemask)
2538 return nodes_intersects(*nodemask, current->mems_allowed);
2542 * nearest_hardwall_ancestor() - Returns the nearest mem_exclusive or
2543 * mem_hardwall ancestor to the specified cpuset. Call holding
2544 * callback_lock. If no ancestor is mem_exclusive or mem_hardwall
2545 * (an unusual configuration), then returns the root cpuset.
2547 static struct cpuset *nearest_hardwall_ancestor(struct cpuset *cs)
2549 while (!(is_mem_exclusive(cs) || is_mem_hardwall(cs)) && parent_cs(cs))
2555 * cpuset_node_allowed - Can we allocate on a memory node?
2556 * @node: is this an allowed node?
2557 * @gfp_mask: memory allocation flags
2559 * If we're in interrupt, yes, we can always allocate. If @node is set in
2560 * current's mems_allowed, yes. If it's not a __GFP_HARDWALL request and this
2561 * node is set in the nearest hardwalled cpuset ancestor to current's cpuset,
2562 * yes. If current has access to memory reserves due to TIF_MEMDIE, yes.
2565 * GFP_USER allocations are marked with the __GFP_HARDWALL bit,
2566 * and do not allow allocations outside the current tasks cpuset
2567 * unless the task has been OOM killed as is marked TIF_MEMDIE.
2568 * GFP_KERNEL allocations are not so marked, so can escape to the
2569 * nearest enclosing hardwalled ancestor cpuset.
2571 * Scanning up parent cpusets requires callback_lock. The
2572 * __alloc_pages() routine only calls here with __GFP_HARDWALL bit
2573 * _not_ set if it's a GFP_KERNEL allocation, and all nodes in the
2574 * current tasks mems_allowed came up empty on the first pass over
2575 * the zonelist. So only GFP_KERNEL allocations, if all nodes in the
2576 * cpuset are short of memory, might require taking the callback_lock.
2578 * The first call here from mm/page_alloc:get_page_from_freelist()
2579 * has __GFP_HARDWALL set in gfp_mask, enforcing hardwall cpusets,
2580 * so no allocation on a node outside the cpuset is allowed (unless
2581 * in interrupt, of course).
2583 * The second pass through get_page_from_freelist() doesn't even call
2584 * here for GFP_ATOMIC calls. For those calls, the __alloc_pages()
2585 * variable 'wait' is not set, and the bit ALLOC_CPUSET is not set
2586 * in alloc_flags. That logic and the checks below have the combined
2588 * in_interrupt - any node ok (current task context irrelevant)
2589 * GFP_ATOMIC - any node ok
2590 * TIF_MEMDIE - any node ok
2591 * GFP_KERNEL - any node in enclosing hardwalled cpuset ok
2592 * GFP_USER - only nodes in current tasks mems allowed ok.
2594 int __cpuset_node_allowed(int node, gfp_t gfp_mask)
2596 struct cpuset *cs; /* current cpuset ancestors */
2597 int allowed; /* is allocation in zone z allowed? */
2598 unsigned long flags;
2602 if (node_isset(node, current->mems_allowed))
2605 * Allow tasks that have access to memory reserves because they have
2606 * been OOM killed to get memory anywhere.
2608 if (unlikely(test_thread_flag(TIF_MEMDIE)))
2610 if (gfp_mask & __GFP_HARDWALL) /* If hardwall request, stop here */
2613 if (current->flags & PF_EXITING) /* Let dying task have memory */
2616 /* Not hardwall and node outside mems_allowed: scan up cpusets */
2617 spin_lock_irqsave(&callback_lock, flags);
2620 cs = nearest_hardwall_ancestor(task_cs(current));
2621 allowed = node_isset(node, cs->mems_allowed);
2624 spin_unlock_irqrestore(&callback_lock, flags);
2629 * cpuset_mem_spread_node() - On which node to begin search for a file page
2630 * cpuset_slab_spread_node() - On which node to begin search for a slab page
2632 * If a task is marked PF_SPREAD_PAGE or PF_SPREAD_SLAB (as for
2633 * tasks in a cpuset with is_spread_page or is_spread_slab set),
2634 * and if the memory allocation used cpuset_mem_spread_node()
2635 * to determine on which node to start looking, as it will for
2636 * certain page cache or slab cache pages such as used for file
2637 * system buffers and inode caches, then instead of starting on the
2638 * local node to look for a free page, rather spread the starting
2639 * node around the tasks mems_allowed nodes.
2641 * We don't have to worry about the returned node being offline
2642 * because "it can't happen", and even if it did, it would be ok.
2644 * The routines calling guarantee_online_mems() are careful to
2645 * only set nodes in task->mems_allowed that are online. So it
2646 * should not be possible for the following code to return an
2647 * offline node. But if it did, that would be ok, as this routine
2648 * is not returning the node where the allocation must be, only
2649 * the node where the search should start. The zonelist passed to
2650 * __alloc_pages() will include all nodes. If the slab allocator
2651 * is passed an offline node, it will fall back to the local node.
2652 * See kmem_cache_alloc_node().
2655 static int cpuset_spread_node(int *rotor)
2659 node = next_node(*rotor, current->mems_allowed);
2660 if (node == MAX_NUMNODES)
2661 node = first_node(current->mems_allowed);
2666 int cpuset_mem_spread_node(void)
2668 if (current->cpuset_mem_spread_rotor == NUMA_NO_NODE)
2669 current->cpuset_mem_spread_rotor =
2670 node_random(¤t->mems_allowed);
2672 return cpuset_spread_node(¤t->cpuset_mem_spread_rotor);
2675 int cpuset_slab_spread_node(void)
2677 if (current->cpuset_slab_spread_rotor == NUMA_NO_NODE)
2678 current->cpuset_slab_spread_rotor =
2679 node_random(¤t->mems_allowed);
2681 return cpuset_spread_node(¤t->cpuset_slab_spread_rotor);
2684 EXPORT_SYMBOL_GPL(cpuset_mem_spread_node);
2687 * cpuset_mems_allowed_intersects - Does @tsk1's mems_allowed intersect @tsk2's?
2688 * @tsk1: pointer to task_struct of some task.
2689 * @tsk2: pointer to task_struct of some other task.
2691 * Description: Return true if @tsk1's mems_allowed intersects the
2692 * mems_allowed of @tsk2. Used by the OOM killer to determine if
2693 * one of the task's memory usage might impact the memory available
2697 int cpuset_mems_allowed_intersects(const struct task_struct *tsk1,
2698 const struct task_struct *tsk2)
2700 return nodes_intersects(tsk1->mems_allowed, tsk2->mems_allowed);
2704 * cpuset_print_current_mems_allowed - prints current's cpuset and mems_allowed
2706 * Description: Prints current's name, cpuset name, and cached copy of its
2707 * mems_allowed to the kernel log.
2709 void cpuset_print_current_mems_allowed(void)
2711 struct cgroup *cgrp;
2715 cgrp = task_cs(current)->css.cgroup;
2716 pr_info("%s cpuset=", current->comm);
2717 pr_cont_cgroup_name(cgrp);
2718 pr_cont(" mems_allowed=%*pbl\n",
2719 nodemask_pr_args(¤t->mems_allowed));
2725 * Collection of memory_pressure is suppressed unless
2726 * this flag is enabled by writing "1" to the special
2727 * cpuset file 'memory_pressure_enabled' in the root cpuset.
2730 int cpuset_memory_pressure_enabled __read_mostly;
2733 * cpuset_memory_pressure_bump - keep stats of per-cpuset reclaims.
2735 * Keep a running average of the rate of synchronous (direct)
2736 * page reclaim efforts initiated by tasks in each cpuset.
2738 * This represents the rate at which some task in the cpuset
2739 * ran low on memory on all nodes it was allowed to use, and
2740 * had to enter the kernels page reclaim code in an effort to
2741 * create more free memory by tossing clean pages or swapping
2742 * or writing dirty pages.
2744 * Display to user space in the per-cpuset read-only file
2745 * "memory_pressure". Value displayed is an integer
2746 * representing the recent rate of entry into the synchronous
2747 * (direct) page reclaim by any task attached to the cpuset.
2750 void __cpuset_memory_pressure_bump(void)
2753 fmeter_markevent(&task_cs(current)->fmeter);
2757 #ifdef CONFIG_PROC_PID_CPUSET
2759 * proc_cpuset_show()
2760 * - Print tasks cpuset path into seq_file.
2761 * - Used for /proc/<pid>/cpuset.
2762 * - No need to task_lock(tsk) on this tsk->cpuset reference, as it
2763 * doesn't really matter if tsk->cpuset changes after we read it,
2764 * and we take cpuset_mutex, keeping cpuset_attach() from changing it
2767 int proc_cpuset_show(struct seq_file *m, struct pid_namespace *ns,
2768 struct pid *pid, struct task_struct *tsk)
2771 struct cgroup_subsys_state *css;
2775 buf = kmalloc(PATH_MAX, GFP_KERNEL);
2779 retval = -ENAMETOOLONG;
2781 css = task_css(tsk, cpuset_cgrp_id);
2782 p = cgroup_path(css->cgroup, buf, PATH_MAX);
2794 #endif /* CONFIG_PROC_PID_CPUSET */
2796 /* Display task mems_allowed in /proc/<pid>/status file. */
2797 void cpuset_task_status_allowed(struct seq_file *m, struct task_struct *task)
2799 seq_printf(m, "Mems_allowed:\t%*pb\n",
2800 nodemask_pr_args(&task->mems_allowed));
2801 seq_printf(m, "Mems_allowed_list:\t%*pbl\n",
2802 nodemask_pr_args(&task->mems_allowed));