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>
60 #include <linux/uaccess.h>
61 #include <linux/atomic.h>
62 #include <linux/mutex.h>
63 #include <linux/cgroup.h>
64 #include <linux/wait.h>
66 DEFINE_STATIC_KEY_FALSE(cpusets_enabled_key);
68 /* See "Frequency meter" comments, below. */
71 int cnt; /* unprocessed events count */
72 int val; /* most recent output value */
73 time64_t time; /* clock (secs) when val computed */
74 spinlock_t lock; /* guards read or write of above */
78 struct cgroup_subsys_state css;
80 unsigned long flags; /* "unsigned long" so bitops work */
83 * On default hierarchy:
85 * The user-configured masks can only be changed by writing to
86 * cpuset.cpus and cpuset.mems, and won't be limited by the
89 * The effective masks is the real masks that apply to the tasks
90 * in the cpuset. They may be changed if the configured masks are
91 * changed or hotplug happens.
93 * effective_mask == configured_mask & parent's effective_mask,
94 * and if it ends up empty, it will inherit the parent's mask.
99 * The user-configured masks are always the same with effective masks.
102 /* user-configured CPUs and Memory Nodes allow to tasks */
103 cpumask_var_t cpus_allowed;
104 nodemask_t mems_allowed;
106 /* effective CPUs and Memory Nodes allow to tasks */
107 cpumask_var_t effective_cpus;
108 nodemask_t effective_mems;
111 * This is old Memory Nodes tasks took on.
113 * - top_cpuset.old_mems_allowed is initialized to mems_allowed.
114 * - A new cpuset's old_mems_allowed is initialized when some
115 * task is moved into it.
116 * - old_mems_allowed is used in cpuset_migrate_mm() when we change
117 * cpuset.mems_allowed and have tasks' nodemask updated, and
118 * then old_mems_allowed is updated to mems_allowed.
120 nodemask_t old_mems_allowed;
122 struct fmeter fmeter; /* memory_pressure filter */
125 * Tasks are being attached to this cpuset. Used to prevent
126 * zeroing cpus/mems_allowed between ->can_attach() and ->attach().
128 int attach_in_progress;
130 /* partition number for rebuild_sched_domains() */
133 /* for custom sched domain */
134 int relax_domain_level;
137 static inline struct cpuset *css_cs(struct cgroup_subsys_state *css)
139 return css ? container_of(css, struct cpuset, css) : NULL;
142 /* Retrieve the cpuset for a task */
143 static inline struct cpuset *task_cs(struct task_struct *task)
145 return css_cs(task_css(task, cpuset_cgrp_id));
148 static inline struct cpuset *parent_cs(struct cpuset *cs)
150 return css_cs(cs->css.parent);
154 static inline bool task_has_mempolicy(struct task_struct *task)
156 return task->mempolicy;
159 static inline bool task_has_mempolicy(struct task_struct *task)
166 /* bits in struct cpuset flags field */
173 CS_SCHED_LOAD_BALANCE,
178 /* convenient tests for these bits */
179 static inline bool is_cpuset_online(struct cpuset *cs)
181 return test_bit(CS_ONLINE, &cs->flags) && !css_is_dying(&cs->css);
184 static inline int is_cpu_exclusive(const struct cpuset *cs)
186 return test_bit(CS_CPU_EXCLUSIVE, &cs->flags);
189 static inline int is_mem_exclusive(const struct cpuset *cs)
191 return test_bit(CS_MEM_EXCLUSIVE, &cs->flags);
194 static inline int is_mem_hardwall(const struct cpuset *cs)
196 return test_bit(CS_MEM_HARDWALL, &cs->flags);
199 static inline int is_sched_load_balance(const struct cpuset *cs)
201 return test_bit(CS_SCHED_LOAD_BALANCE, &cs->flags);
204 static inline int is_memory_migrate(const struct cpuset *cs)
206 return test_bit(CS_MEMORY_MIGRATE, &cs->flags);
209 static inline int is_spread_page(const struct cpuset *cs)
211 return test_bit(CS_SPREAD_PAGE, &cs->flags);
214 static inline int is_spread_slab(const struct cpuset *cs)
216 return test_bit(CS_SPREAD_SLAB, &cs->flags);
219 static struct cpuset top_cpuset = {
220 .flags = ((1 << CS_ONLINE) | (1 << CS_CPU_EXCLUSIVE) |
221 (1 << CS_MEM_EXCLUSIVE)),
225 * cpuset_for_each_child - traverse online children of a cpuset
226 * @child_cs: loop cursor pointing to the current child
227 * @pos_css: used for iteration
228 * @parent_cs: target cpuset to walk children of
230 * Walk @child_cs through the online children of @parent_cs. Must be used
231 * with RCU read locked.
233 #define cpuset_for_each_child(child_cs, pos_css, parent_cs) \
234 css_for_each_child((pos_css), &(parent_cs)->css) \
235 if (is_cpuset_online(((child_cs) = css_cs((pos_css)))))
238 * cpuset_for_each_descendant_pre - pre-order walk of a cpuset's descendants
239 * @des_cs: loop cursor pointing to the current descendant
240 * @pos_css: used for iteration
241 * @root_cs: target cpuset to walk ancestor of
243 * Walk @des_cs through the online descendants of @root_cs. Must be used
244 * with RCU read locked. The caller may modify @pos_css by calling
245 * css_rightmost_descendant() to skip subtree. @root_cs is included in the
246 * iteration and the first node to be visited.
248 #define cpuset_for_each_descendant_pre(des_cs, pos_css, root_cs) \
249 css_for_each_descendant_pre((pos_css), &(root_cs)->css) \
250 if (is_cpuset_online(((des_cs) = css_cs((pos_css)))))
253 * There are two global locks guarding cpuset structures - cpuset_mutex and
254 * callback_lock. We also require taking task_lock() when dereferencing a
255 * task's cpuset pointer. See "The task_lock() exception", at the end of this
258 * A task must hold both locks to modify cpusets. If a task holds
259 * cpuset_mutex, then it blocks others wanting that mutex, ensuring that it
260 * is the only task able to also acquire callback_lock and be able to
261 * modify cpusets. It can perform various checks on the cpuset structure
262 * first, knowing nothing will change. It can also allocate memory while
263 * just holding cpuset_mutex. While it is performing these checks, various
264 * callback routines can briefly acquire callback_lock to query cpusets.
265 * Once it is ready to make the changes, it takes callback_lock, blocking
268 * Calls to the kernel memory allocator can not be made while holding
269 * callback_lock, as that would risk double tripping on callback_lock
270 * from one of the callbacks into the cpuset code from within
273 * If a task is only holding callback_lock, then it has read-only
276 * Now, the task_struct fields mems_allowed and mempolicy may be changed
277 * by other task, we use alloc_lock in the task_struct fields to protect
280 * The cpuset_common_file_read() handlers only hold callback_lock across
281 * small pieces of code, such as when reading out possibly multi-word
282 * cpumasks and nodemasks.
284 * Accessing a task's cpuset should be done in accordance with the
285 * guidelines for accessing subsystem state in kernel/cgroup.c
288 static DEFINE_MUTEX(cpuset_mutex);
289 static DEFINE_SPINLOCK(callback_lock);
291 static struct workqueue_struct *cpuset_migrate_mm_wq;
294 * CPU / memory hotplug is handled asynchronously.
296 static void cpuset_hotplug_workfn(struct work_struct *work);
297 static DECLARE_WORK(cpuset_hotplug_work, cpuset_hotplug_workfn);
299 static DECLARE_WAIT_QUEUE_HEAD(cpuset_attach_wq);
302 * This is ugly, but preserves the userspace API for existing cpuset
303 * users. If someone tries to mount the "cpuset" filesystem, we
304 * silently switch it to mount "cgroup" instead
306 static struct dentry *cpuset_mount(struct file_system_type *fs_type,
307 int flags, const char *unused_dev_name, void *data)
309 struct file_system_type *cgroup_fs = get_fs_type("cgroup");
310 struct dentry *ret = ERR_PTR(-ENODEV);
314 "release_agent=/sbin/cpuset_release_agent";
315 ret = cgroup_fs->mount(cgroup_fs, flags,
316 unused_dev_name, mountopts);
317 put_filesystem(cgroup_fs);
322 static struct file_system_type cpuset_fs_type = {
324 .mount = cpuset_mount,
328 * Return in pmask the portion of a cpusets's cpus_allowed that
329 * are online. If none are online, walk up the cpuset hierarchy
330 * until we find one that does have some online cpus.
332 * One way or another, we guarantee to return some non-empty subset
333 * of cpu_online_mask.
335 * Call with callback_lock or cpuset_mutex held.
337 static void guarantee_online_cpus(struct cpuset *cs, struct cpumask *pmask)
339 while (!cpumask_intersects(cs->effective_cpus, cpu_online_mask)) {
343 * The top cpuset doesn't have any online cpu as a
344 * consequence of a race between cpuset_hotplug_work
345 * and cpu hotplug notifier. But we know the top
346 * cpuset's effective_cpus is on its way to to be
347 * identical to cpu_online_mask.
349 cpumask_copy(pmask, cpu_online_mask);
353 cpumask_and(pmask, cs->effective_cpus, cpu_online_mask);
357 * Return in *pmask the portion of a cpusets's mems_allowed that
358 * are online, with memory. If none are online with memory, walk
359 * up the cpuset hierarchy until we find one that does have some
360 * online mems. The top cpuset always has some mems online.
362 * One way or another, we guarantee to return some non-empty subset
363 * of node_states[N_MEMORY].
365 * Call with callback_lock or cpuset_mutex held.
367 static void guarantee_online_mems(struct cpuset *cs, nodemask_t *pmask)
369 while (!nodes_intersects(cs->effective_mems, node_states[N_MEMORY]))
371 nodes_and(*pmask, cs->effective_mems, node_states[N_MEMORY]);
375 * update task's spread flag if cpuset's page/slab spread flag is set
377 * Call with callback_lock or cpuset_mutex held.
379 static void cpuset_update_task_spread_flag(struct cpuset *cs,
380 struct task_struct *tsk)
382 if (is_spread_page(cs))
383 task_set_spread_page(tsk);
385 task_clear_spread_page(tsk);
387 if (is_spread_slab(cs))
388 task_set_spread_slab(tsk);
390 task_clear_spread_slab(tsk);
394 * is_cpuset_subset(p, q) - Is cpuset p a subset of cpuset q?
396 * One cpuset is a subset of another if all its allowed CPUs and
397 * Memory Nodes are a subset of the other, and its exclusive flags
398 * are only set if the other's are set. Call holding cpuset_mutex.
401 static int is_cpuset_subset(const struct cpuset *p, const struct cpuset *q)
403 return cpumask_subset(p->cpus_allowed, q->cpus_allowed) &&
404 nodes_subset(p->mems_allowed, q->mems_allowed) &&
405 is_cpu_exclusive(p) <= is_cpu_exclusive(q) &&
406 is_mem_exclusive(p) <= is_mem_exclusive(q);
410 * alloc_trial_cpuset - allocate a trial cpuset
411 * @cs: the cpuset that the trial cpuset duplicates
413 static struct cpuset *alloc_trial_cpuset(struct cpuset *cs)
415 struct cpuset *trial;
417 trial = kmemdup(cs, sizeof(*cs), GFP_KERNEL);
421 if (!alloc_cpumask_var(&trial->cpus_allowed, GFP_KERNEL))
423 if (!alloc_cpumask_var(&trial->effective_cpus, GFP_KERNEL))
426 cpumask_copy(trial->cpus_allowed, cs->cpus_allowed);
427 cpumask_copy(trial->effective_cpus, cs->effective_cpus);
431 free_cpumask_var(trial->cpus_allowed);
438 * free_trial_cpuset - free the trial cpuset
439 * @trial: the trial cpuset to be freed
441 static void free_trial_cpuset(struct cpuset *trial)
443 free_cpumask_var(trial->effective_cpus);
444 free_cpumask_var(trial->cpus_allowed);
449 * validate_change() - Used to validate that any proposed cpuset change
450 * follows the structural rules for cpusets.
452 * If we replaced the flag and mask values of the current cpuset
453 * (cur) with those values in the trial cpuset (trial), would
454 * our various subset and exclusive rules still be valid? Presumes
457 * 'cur' is the address of an actual, in-use cpuset. Operations
458 * such as list traversal that depend on the actual address of the
459 * cpuset in the list must use cur below, not trial.
461 * 'trial' is the address of bulk structure copy of cur, with
462 * perhaps one or more of the fields cpus_allowed, mems_allowed,
463 * or flags changed to new, trial values.
465 * Return 0 if valid, -errno if not.
468 static int validate_change(struct cpuset *cur, struct cpuset *trial)
470 struct cgroup_subsys_state *css;
471 struct cpuset *c, *par;
476 /* Each of our child cpusets must be a subset of us */
478 cpuset_for_each_child(c, css, cur)
479 if (!is_cpuset_subset(c, trial))
482 /* Remaining checks don't apply to root cpuset */
484 if (cur == &top_cpuset)
487 par = parent_cs(cur);
489 /* On legacy hiearchy, we must be a subset of our parent cpuset. */
491 if (!cgroup_subsys_on_dfl(cpuset_cgrp_subsys) &&
492 !is_cpuset_subset(trial, par))
496 * If either I or some sibling (!= me) is exclusive, we can't
500 cpuset_for_each_child(c, css, par) {
501 if ((is_cpu_exclusive(trial) || is_cpu_exclusive(c)) &&
503 cpumask_intersects(trial->cpus_allowed, c->cpus_allowed))
505 if ((is_mem_exclusive(trial) || is_mem_exclusive(c)) &&
507 nodes_intersects(trial->mems_allowed, c->mems_allowed))
512 * Cpusets with tasks - existing or newly being attached - can't
513 * be changed to have empty cpus_allowed or mems_allowed.
516 if ((cgroup_is_populated(cur->css.cgroup) || cur->attach_in_progress)) {
517 if (!cpumask_empty(cur->cpus_allowed) &&
518 cpumask_empty(trial->cpus_allowed))
520 if (!nodes_empty(cur->mems_allowed) &&
521 nodes_empty(trial->mems_allowed))
526 * We can't shrink if we won't have enough room for SCHED_DEADLINE
530 if (is_cpu_exclusive(cur) &&
531 !cpuset_cpumask_can_shrink(cur->cpus_allowed,
532 trial->cpus_allowed))
543 * Helper routine for generate_sched_domains().
544 * Do cpusets a, b have overlapping effective cpus_allowed masks?
546 static int cpusets_overlap(struct cpuset *a, struct cpuset *b)
548 return cpumask_intersects(a->effective_cpus, b->effective_cpus);
552 update_domain_attr(struct sched_domain_attr *dattr, struct cpuset *c)
554 if (dattr->relax_domain_level < c->relax_domain_level)
555 dattr->relax_domain_level = c->relax_domain_level;
559 static void update_domain_attr_tree(struct sched_domain_attr *dattr,
560 struct cpuset *root_cs)
563 struct cgroup_subsys_state *pos_css;
566 cpuset_for_each_descendant_pre(cp, pos_css, root_cs) {
567 /* skip the whole subtree if @cp doesn't have any CPU */
568 if (cpumask_empty(cp->cpus_allowed)) {
569 pos_css = css_rightmost_descendant(pos_css);
573 if (is_sched_load_balance(cp))
574 update_domain_attr(dattr, cp);
580 * generate_sched_domains()
582 * This function builds a partial partition of the systems CPUs
583 * A 'partial partition' is a set of non-overlapping subsets whose
584 * union is a subset of that set.
585 * The output of this function needs to be passed to kernel/sched/core.c
586 * partition_sched_domains() routine, which will rebuild the scheduler's
587 * load balancing domains (sched domains) as specified by that partial
590 * See "What is sched_load_balance" in Documentation/cgroups/cpusets.txt
591 * for a background explanation of this.
593 * Does not return errors, on the theory that the callers of this
594 * routine would rather not worry about failures to rebuild sched
595 * domains when operating in the severe memory shortage situations
596 * that could cause allocation failures below.
598 * Must be called with cpuset_mutex held.
600 * The three key local variables below are:
601 * q - a linked-list queue of cpuset pointers, used to implement a
602 * top-down scan of all cpusets. This scan loads a pointer
603 * to each cpuset marked is_sched_load_balance into the
604 * array 'csa'. For our purposes, rebuilding the schedulers
605 * sched domains, we can ignore !is_sched_load_balance cpusets.
606 * csa - (for CpuSet Array) Array of pointers to all the cpusets
607 * that need to be load balanced, for convenient iterative
608 * access by the subsequent code that finds the best partition,
609 * i.e the set of domains (subsets) of CPUs such that the
610 * cpus_allowed of every cpuset marked is_sched_load_balance
611 * is a subset of one of these domains, while there are as
612 * many such domains as possible, each as small as possible.
613 * doms - Conversion of 'csa' to an array of cpumasks, for passing to
614 * the kernel/sched/core.c routine partition_sched_domains() in a
615 * convenient format, that can be easily compared to the prior
616 * value to determine what partition elements (sched domains)
617 * were changed (added or removed.)
619 * Finding the best partition (set of domains):
620 * The triple nested loops below over i, j, k scan over the
621 * load balanced cpusets (using the array of cpuset pointers in
622 * csa[]) looking for pairs of cpusets that have overlapping
623 * cpus_allowed, but which don't have the same 'pn' partition
624 * number and gives them in the same partition number. It keeps
625 * looping on the 'restart' label until it can no longer find
628 * The union of the cpus_allowed masks from the set of
629 * all cpusets having the same 'pn' value then form the one
630 * element of the partition (one sched domain) to be passed to
631 * partition_sched_domains().
633 static int generate_sched_domains(cpumask_var_t **domains,
634 struct sched_domain_attr **attributes)
636 struct cpuset *cp; /* scans q */
637 struct cpuset **csa; /* array of all cpuset ptrs */
638 int csn; /* how many cpuset ptrs in csa so far */
639 int i, j, k; /* indices for partition finding loops */
640 cpumask_var_t *doms; /* resulting partition; i.e. sched domains */
641 cpumask_var_t non_isolated_cpus; /* load balanced CPUs */
642 struct sched_domain_attr *dattr; /* attributes for custom domains */
643 int ndoms = 0; /* number of sched domains in result */
644 int nslot; /* next empty doms[] struct cpumask slot */
645 struct cgroup_subsys_state *pos_css;
651 if (!alloc_cpumask_var(&non_isolated_cpus, GFP_KERNEL))
653 cpumask_andnot(non_isolated_cpus, cpu_possible_mask, cpu_isolated_map);
655 /* Special case for the 99% of systems with one, full, sched domain */
656 if (is_sched_load_balance(&top_cpuset)) {
658 doms = alloc_sched_domains(ndoms);
662 dattr = kmalloc(sizeof(struct sched_domain_attr), GFP_KERNEL);
664 *dattr = SD_ATTR_INIT;
665 update_domain_attr_tree(dattr, &top_cpuset);
667 cpumask_and(doms[0], top_cpuset.effective_cpus,
673 csa = kmalloc(nr_cpusets() * sizeof(cp), GFP_KERNEL);
679 cpuset_for_each_descendant_pre(cp, pos_css, &top_cpuset) {
680 if (cp == &top_cpuset)
683 * Continue traversing beyond @cp iff @cp has some CPUs and
684 * isn't load balancing. The former is obvious. The
685 * latter: All child cpusets contain a subset of the
686 * parent's cpus, so just skip them, and then we call
687 * update_domain_attr_tree() to calc relax_domain_level of
688 * the corresponding sched domain.
690 if (!cpumask_empty(cp->cpus_allowed) &&
691 !(is_sched_load_balance(cp) &&
692 cpumask_intersects(cp->cpus_allowed, non_isolated_cpus)))
695 if (is_sched_load_balance(cp))
698 /* skip @cp's subtree */
699 pos_css = css_rightmost_descendant(pos_css);
703 for (i = 0; i < csn; i++)
708 /* Find the best partition (set of sched domains) */
709 for (i = 0; i < csn; i++) {
710 struct cpuset *a = csa[i];
713 for (j = 0; j < csn; j++) {
714 struct cpuset *b = csa[j];
717 if (apn != bpn && cpusets_overlap(a, b)) {
718 for (k = 0; k < csn; k++) {
719 struct cpuset *c = csa[k];
724 ndoms--; /* one less element */
731 * Now we know how many domains to create.
732 * Convert <csn, csa> to <ndoms, doms> and populate cpu masks.
734 doms = alloc_sched_domains(ndoms);
739 * The rest of the code, including the scheduler, can deal with
740 * dattr==NULL case. No need to abort if alloc fails.
742 dattr = kmalloc(ndoms * sizeof(struct sched_domain_attr), GFP_KERNEL);
744 for (nslot = 0, i = 0; i < csn; i++) {
745 struct cpuset *a = csa[i];
750 /* Skip completed partitions */
756 if (nslot == ndoms) {
757 static int warnings = 10;
759 pr_warn("rebuild_sched_domains confused: nslot %d, ndoms %d, csn %d, i %d, apn %d\n",
760 nslot, ndoms, csn, i, apn);
768 *(dattr + nslot) = SD_ATTR_INIT;
769 for (j = i; j < csn; j++) {
770 struct cpuset *b = csa[j];
773 cpumask_or(dp, dp, b->effective_cpus);
774 cpumask_and(dp, dp, non_isolated_cpus);
776 update_domain_attr_tree(dattr + nslot, b);
778 /* Done with this partition */
784 BUG_ON(nslot != ndoms);
787 free_cpumask_var(non_isolated_cpus);
791 * Fallback to the default domain if kmalloc() failed.
792 * See comments in partition_sched_domains().
803 * Rebuild scheduler domains.
805 * If the flag 'sched_load_balance' of any cpuset with non-empty
806 * 'cpus' changes, or if the 'cpus' allowed changes in any cpuset
807 * which has that flag enabled, or if any cpuset with a non-empty
808 * 'cpus' is removed, then call this routine to rebuild the
809 * scheduler's dynamic sched domains.
811 * Call with cpuset_mutex held. Takes get_online_cpus().
813 static void rebuild_sched_domains_locked(void)
815 struct sched_domain_attr *attr;
819 lockdep_assert_held(&cpuset_mutex);
823 * We have raced with CPU hotplug. Don't do anything to avoid
824 * passing doms with offlined cpu to partition_sched_domains().
825 * Anyways, hotplug work item will rebuild sched domains.
827 if (!cpumask_equal(top_cpuset.effective_cpus, cpu_active_mask))
830 /* Generate domain masks and attrs */
831 ndoms = generate_sched_domains(&doms, &attr);
833 /* Have scheduler rebuild the domains */
834 partition_sched_domains(ndoms, doms, attr);
838 #else /* !CONFIG_SMP */
839 static void rebuild_sched_domains_locked(void)
842 #endif /* CONFIG_SMP */
844 void rebuild_sched_domains(void)
846 mutex_lock(&cpuset_mutex);
847 rebuild_sched_domains_locked();
848 mutex_unlock(&cpuset_mutex);
852 * update_tasks_cpumask - Update the cpumasks of tasks in the cpuset.
853 * @cs: the cpuset in which each task's cpus_allowed mask needs to be changed
855 * Iterate through each task of @cs updating its cpus_allowed to the
856 * effective cpuset's. As this function is called with cpuset_mutex held,
857 * cpuset membership stays stable.
859 static void update_tasks_cpumask(struct cpuset *cs)
861 struct css_task_iter it;
862 struct task_struct *task;
864 css_task_iter_start(&cs->css, &it);
865 while ((task = css_task_iter_next(&it)))
866 set_cpus_allowed_ptr(task, cs->effective_cpus);
867 css_task_iter_end(&it);
871 * update_cpumasks_hier - Update effective cpumasks and tasks in the subtree
872 * @cs: the cpuset to consider
873 * @new_cpus: temp variable for calculating new effective_cpus
875 * When congifured cpumask is changed, the effective cpumasks of this cpuset
876 * and all its descendants need to be updated.
878 * On legacy hierachy, effective_cpus will be the same with cpu_allowed.
880 * Called with cpuset_mutex held
882 static void update_cpumasks_hier(struct cpuset *cs, struct cpumask *new_cpus)
885 struct cgroup_subsys_state *pos_css;
886 bool need_rebuild_sched_domains = false;
889 cpuset_for_each_descendant_pre(cp, pos_css, cs) {
890 struct cpuset *parent = parent_cs(cp);
892 cpumask_and(new_cpus, cp->cpus_allowed, parent->effective_cpus);
895 * If it becomes empty, inherit the effective mask of the
896 * parent, which is guaranteed to have some CPUs.
898 if (cgroup_subsys_on_dfl(cpuset_cgrp_subsys) &&
899 cpumask_empty(new_cpus))
900 cpumask_copy(new_cpus, parent->effective_cpus);
902 /* Skip the whole subtree if the cpumask remains the same. */
903 if (cpumask_equal(new_cpus, cp->effective_cpus)) {
904 pos_css = css_rightmost_descendant(pos_css);
908 if (!css_tryget_online(&cp->css))
912 spin_lock_irq(&callback_lock);
913 cpumask_copy(cp->effective_cpus, new_cpus);
914 spin_unlock_irq(&callback_lock);
916 WARN_ON(!cgroup_subsys_on_dfl(cpuset_cgrp_subsys) &&
917 !cpumask_equal(cp->cpus_allowed, cp->effective_cpus));
919 update_tasks_cpumask(cp);
922 * If the effective cpumask of any non-empty cpuset is changed,
923 * we need to rebuild sched domains.
925 if (!cpumask_empty(cp->cpus_allowed) &&
926 is_sched_load_balance(cp))
927 need_rebuild_sched_domains = true;
934 if (need_rebuild_sched_domains)
935 rebuild_sched_domains_locked();
939 * update_cpumask - update the cpus_allowed mask of a cpuset and all tasks in it
940 * @cs: the cpuset to consider
941 * @trialcs: trial cpuset
942 * @buf: buffer of cpu numbers written to this cpuset
944 static int update_cpumask(struct cpuset *cs, struct cpuset *trialcs,
949 /* top_cpuset.cpus_allowed tracks cpu_online_mask; it's read-only */
950 if (cs == &top_cpuset)
954 * An empty cpus_allowed is ok only if the cpuset has no tasks.
955 * Since cpulist_parse() fails on an empty mask, we special case
956 * that parsing. The validate_change() call ensures that cpusets
957 * with tasks have cpus.
960 cpumask_clear(trialcs->cpus_allowed);
962 retval = cpulist_parse(buf, trialcs->cpus_allowed);
966 if (!cpumask_subset(trialcs->cpus_allowed,
967 top_cpuset.cpus_allowed))
971 /* Nothing to do if the cpus didn't change */
972 if (cpumask_equal(cs->cpus_allowed, trialcs->cpus_allowed))
975 retval = validate_change(cs, trialcs);
979 spin_lock_irq(&callback_lock);
980 cpumask_copy(cs->cpus_allowed, trialcs->cpus_allowed);
981 spin_unlock_irq(&callback_lock);
983 /* use trialcs->cpus_allowed as a temp variable */
984 update_cpumasks_hier(cs, trialcs->cpus_allowed);
989 * Migrate memory region from one set of nodes to another. This is
990 * performed asynchronously as it can be called from process migration path
991 * holding locks involved in process management. All mm migrations are
992 * performed in the queued order and can be waited for by flushing
993 * cpuset_migrate_mm_wq.
996 struct cpuset_migrate_mm_work {
997 struct work_struct work;
998 struct mm_struct *mm;
1003 static void cpuset_migrate_mm_workfn(struct work_struct *work)
1005 struct cpuset_migrate_mm_work *mwork =
1006 container_of(work, struct cpuset_migrate_mm_work, work);
1008 /* on a wq worker, no need to worry about %current's mems_allowed */
1009 do_migrate_pages(mwork->mm, &mwork->from, &mwork->to, MPOL_MF_MOVE_ALL);
1014 static void cpuset_migrate_mm(struct mm_struct *mm, const nodemask_t *from,
1015 const nodemask_t *to)
1017 struct cpuset_migrate_mm_work *mwork;
1019 mwork = kzalloc(sizeof(*mwork), GFP_KERNEL);
1022 mwork->from = *from;
1024 INIT_WORK(&mwork->work, cpuset_migrate_mm_workfn);
1025 queue_work(cpuset_migrate_mm_wq, &mwork->work);
1031 static void cpuset_post_attach(void)
1033 flush_workqueue(cpuset_migrate_mm_wq);
1037 * cpuset_change_task_nodemask - change task's mems_allowed and mempolicy
1038 * @tsk: the task to change
1039 * @newmems: new nodes that the task will be set
1041 * In order to avoid seeing no nodes if the old and new nodes are disjoint,
1042 * we structure updates as setting all new allowed nodes, then clearing newly
1045 static void cpuset_change_task_nodemask(struct task_struct *tsk,
1046 nodemask_t *newmems)
1052 * Determine if a loop is necessary if another thread is doing
1053 * read_mems_allowed_begin(). If at least one node remains unchanged and
1054 * tsk does not have a mempolicy, then an empty nodemask will not be
1055 * possible when mems_allowed is larger than a word.
1057 need_loop = task_has_mempolicy(tsk) ||
1058 !nodes_intersects(*newmems, tsk->mems_allowed);
1061 local_irq_disable();
1062 write_seqcount_begin(&tsk->mems_allowed_seq);
1065 nodes_or(tsk->mems_allowed, tsk->mems_allowed, *newmems);
1066 mpol_rebind_task(tsk, newmems, MPOL_REBIND_STEP1);
1068 mpol_rebind_task(tsk, newmems, MPOL_REBIND_STEP2);
1069 tsk->mems_allowed = *newmems;
1072 write_seqcount_end(&tsk->mems_allowed_seq);
1079 static void *cpuset_being_rebound;
1082 * update_tasks_nodemask - Update the nodemasks of tasks in the cpuset.
1083 * @cs: the cpuset in which each task's mems_allowed mask needs to be changed
1085 * Iterate through each task of @cs updating its mems_allowed to the
1086 * effective cpuset's. As this function is called with cpuset_mutex held,
1087 * cpuset membership stays stable.
1089 static void update_tasks_nodemask(struct cpuset *cs)
1091 static nodemask_t newmems; /* protected by cpuset_mutex */
1092 struct css_task_iter it;
1093 struct task_struct *task;
1095 cpuset_being_rebound = cs; /* causes mpol_dup() rebind */
1097 guarantee_online_mems(cs, &newmems);
1100 * The mpol_rebind_mm() call takes mmap_sem, which we couldn't
1101 * take while holding tasklist_lock. Forks can happen - the
1102 * mpol_dup() cpuset_being_rebound check will catch such forks,
1103 * and rebind their vma mempolicies too. Because we still hold
1104 * the global cpuset_mutex, we know that no other rebind effort
1105 * will be contending for the global variable cpuset_being_rebound.
1106 * It's ok if we rebind the same mm twice; mpol_rebind_mm()
1107 * is idempotent. Also migrate pages in each mm to new nodes.
1109 css_task_iter_start(&cs->css, &it);
1110 while ((task = css_task_iter_next(&it))) {
1111 struct mm_struct *mm;
1114 cpuset_change_task_nodemask(task, &newmems);
1116 mm = get_task_mm(task);
1120 migrate = is_memory_migrate(cs);
1122 mpol_rebind_mm(mm, &cs->mems_allowed);
1124 cpuset_migrate_mm(mm, &cs->old_mems_allowed, &newmems);
1128 css_task_iter_end(&it);
1131 * All the tasks' nodemasks have been updated, update
1132 * cs->old_mems_allowed.
1134 cs->old_mems_allowed = newmems;
1136 /* We're done rebinding vmas to this cpuset's new mems_allowed. */
1137 cpuset_being_rebound = NULL;
1141 * update_nodemasks_hier - Update effective nodemasks and tasks in the subtree
1142 * @cs: the cpuset to consider
1143 * @new_mems: a temp variable for calculating new effective_mems
1145 * When configured nodemask is changed, the effective nodemasks of this cpuset
1146 * and all its descendants need to be updated.
1148 * On legacy hiearchy, effective_mems will be the same with mems_allowed.
1150 * Called with cpuset_mutex held
1152 static void update_nodemasks_hier(struct cpuset *cs, nodemask_t *new_mems)
1155 struct cgroup_subsys_state *pos_css;
1158 cpuset_for_each_descendant_pre(cp, pos_css, cs) {
1159 struct cpuset *parent = parent_cs(cp);
1161 nodes_and(*new_mems, cp->mems_allowed, parent->effective_mems);
1164 * If it becomes empty, inherit the effective mask of the
1165 * parent, which is guaranteed to have some MEMs.
1167 if (cgroup_subsys_on_dfl(cpuset_cgrp_subsys) &&
1168 nodes_empty(*new_mems))
1169 *new_mems = parent->effective_mems;
1171 /* Skip the whole subtree if the nodemask remains the same. */
1172 if (nodes_equal(*new_mems, cp->effective_mems)) {
1173 pos_css = css_rightmost_descendant(pos_css);
1177 if (!css_tryget_online(&cp->css))
1181 spin_lock_irq(&callback_lock);
1182 cp->effective_mems = *new_mems;
1183 spin_unlock_irq(&callback_lock);
1185 WARN_ON(!cgroup_subsys_on_dfl(cpuset_cgrp_subsys) &&
1186 !nodes_equal(cp->mems_allowed, cp->effective_mems));
1188 update_tasks_nodemask(cp);
1197 * Handle user request to change the 'mems' memory placement
1198 * of a cpuset. Needs to validate the request, update the
1199 * cpusets mems_allowed, and for each task in the cpuset,
1200 * update mems_allowed and rebind task's mempolicy and any vma
1201 * mempolicies and if the cpuset is marked 'memory_migrate',
1202 * migrate the tasks pages to the new memory.
1204 * Call with cpuset_mutex held. May take callback_lock during call.
1205 * Will take tasklist_lock, scan tasklist for tasks in cpuset cs,
1206 * lock each such tasks mm->mmap_sem, scan its vma's and rebind
1207 * their mempolicies to the cpusets new mems_allowed.
1209 static int update_nodemask(struct cpuset *cs, struct cpuset *trialcs,
1215 * top_cpuset.mems_allowed tracks node_stats[N_MEMORY];
1218 if (cs == &top_cpuset) {
1224 * An empty mems_allowed is ok iff there are no tasks in the cpuset.
1225 * Since nodelist_parse() fails on an empty mask, we special case
1226 * that parsing. The validate_change() call ensures that cpusets
1227 * with tasks have memory.
1230 nodes_clear(trialcs->mems_allowed);
1232 retval = nodelist_parse(buf, trialcs->mems_allowed);
1236 if (!nodes_subset(trialcs->mems_allowed,
1237 top_cpuset.mems_allowed)) {
1243 if (nodes_equal(cs->mems_allowed, trialcs->mems_allowed)) {
1244 retval = 0; /* Too easy - nothing to do */
1247 retval = validate_change(cs, trialcs);
1251 spin_lock_irq(&callback_lock);
1252 cs->mems_allowed = trialcs->mems_allowed;
1253 spin_unlock_irq(&callback_lock);
1255 /* use trialcs->mems_allowed as a temp variable */
1256 update_nodemasks_hier(cs, &trialcs->mems_allowed);
1261 int current_cpuset_is_being_rebound(void)
1266 ret = task_cs(current) == cpuset_being_rebound;
1272 static int update_relax_domain_level(struct cpuset *cs, s64 val)
1275 if (val < -1 || val >= sched_domain_level_max)
1279 if (val != cs->relax_domain_level) {
1280 cs->relax_domain_level = val;
1281 if (!cpumask_empty(cs->cpus_allowed) &&
1282 is_sched_load_balance(cs))
1283 rebuild_sched_domains_locked();
1290 * update_tasks_flags - update the spread flags of tasks in the cpuset.
1291 * @cs: the cpuset in which each task's spread flags needs to be changed
1293 * Iterate through each task of @cs updating its spread flags. As this
1294 * function is called with cpuset_mutex held, cpuset membership stays
1297 static void update_tasks_flags(struct cpuset *cs)
1299 struct css_task_iter it;
1300 struct task_struct *task;
1302 css_task_iter_start(&cs->css, &it);
1303 while ((task = css_task_iter_next(&it)))
1304 cpuset_update_task_spread_flag(cs, task);
1305 css_task_iter_end(&it);
1309 * update_flag - read a 0 or a 1 in a file and update associated flag
1310 * bit: the bit to update (see cpuset_flagbits_t)
1311 * cs: the cpuset to update
1312 * turning_on: whether the flag is being set or cleared
1314 * Call with cpuset_mutex held.
1317 static int update_flag(cpuset_flagbits_t bit, struct cpuset *cs,
1320 struct cpuset *trialcs;
1321 int balance_flag_changed;
1322 int spread_flag_changed;
1325 trialcs = alloc_trial_cpuset(cs);
1330 set_bit(bit, &trialcs->flags);
1332 clear_bit(bit, &trialcs->flags);
1334 err = validate_change(cs, trialcs);
1338 balance_flag_changed = (is_sched_load_balance(cs) !=
1339 is_sched_load_balance(trialcs));
1341 spread_flag_changed = ((is_spread_slab(cs) != is_spread_slab(trialcs))
1342 || (is_spread_page(cs) != is_spread_page(trialcs)));
1344 spin_lock_irq(&callback_lock);
1345 cs->flags = trialcs->flags;
1346 spin_unlock_irq(&callback_lock);
1348 if (!cpumask_empty(trialcs->cpus_allowed) && balance_flag_changed)
1349 rebuild_sched_domains_locked();
1351 if (spread_flag_changed)
1352 update_tasks_flags(cs);
1354 free_trial_cpuset(trialcs);
1359 * Frequency meter - How fast is some event occurring?
1361 * These routines manage a digitally filtered, constant time based,
1362 * event frequency meter. There are four routines:
1363 * fmeter_init() - initialize a frequency meter.
1364 * fmeter_markevent() - called each time the event happens.
1365 * fmeter_getrate() - returns the recent rate of such events.
1366 * fmeter_update() - internal routine used to update fmeter.
1368 * A common data structure is passed to each of these routines,
1369 * which is used to keep track of the state required to manage the
1370 * frequency meter and its digital filter.
1372 * The filter works on the number of events marked per unit time.
1373 * The filter is single-pole low-pass recursive (IIR). The time unit
1374 * is 1 second. Arithmetic is done using 32-bit integers scaled to
1375 * simulate 3 decimal digits of precision (multiplied by 1000).
1377 * With an FM_COEF of 933, and a time base of 1 second, the filter
1378 * has a half-life of 10 seconds, meaning that if the events quit
1379 * happening, then the rate returned from the fmeter_getrate()
1380 * will be cut in half each 10 seconds, until it converges to zero.
1382 * It is not worth doing a real infinitely recursive filter. If more
1383 * than FM_MAXTICKS ticks have elapsed since the last filter event,
1384 * just compute FM_MAXTICKS ticks worth, by which point the level
1387 * Limit the count of unprocessed events to FM_MAXCNT, so as to avoid
1388 * arithmetic overflow in the fmeter_update() routine.
1390 * Given the simple 32 bit integer arithmetic used, this meter works
1391 * best for reporting rates between one per millisecond (msec) and
1392 * one per 32 (approx) seconds. At constant rates faster than one
1393 * per msec it maxes out at values just under 1,000,000. At constant
1394 * rates between one per msec, and one per second it will stabilize
1395 * to a value N*1000, where N is the rate of events per second.
1396 * At constant rates between one per second and one per 32 seconds,
1397 * it will be choppy, moving up on the seconds that have an event,
1398 * and then decaying until the next event. At rates slower than
1399 * about one in 32 seconds, it decays all the way back to zero between
1403 #define FM_COEF 933 /* coefficient for half-life of 10 secs */
1404 #define FM_MAXTICKS ((u32)99) /* useless computing more ticks than this */
1405 #define FM_MAXCNT 1000000 /* limit cnt to avoid overflow */
1406 #define FM_SCALE 1000 /* faux fixed point scale */
1408 /* Initialize a frequency meter */
1409 static void fmeter_init(struct fmeter *fmp)
1414 spin_lock_init(&fmp->lock);
1417 /* Internal meter update - process cnt events and update value */
1418 static void fmeter_update(struct fmeter *fmp)
1423 now = ktime_get_seconds();
1424 ticks = now - fmp->time;
1429 ticks = min(FM_MAXTICKS, ticks);
1431 fmp->val = (FM_COEF * fmp->val) / FM_SCALE;
1434 fmp->val += ((FM_SCALE - FM_COEF) * fmp->cnt) / FM_SCALE;
1438 /* Process any previous ticks, then bump cnt by one (times scale). */
1439 static void fmeter_markevent(struct fmeter *fmp)
1441 spin_lock(&fmp->lock);
1443 fmp->cnt = min(FM_MAXCNT, fmp->cnt + FM_SCALE);
1444 spin_unlock(&fmp->lock);
1447 /* Process any previous ticks, then return current value. */
1448 static int fmeter_getrate(struct fmeter *fmp)
1452 spin_lock(&fmp->lock);
1455 spin_unlock(&fmp->lock);
1459 static struct cpuset *cpuset_attach_old_cs;
1461 /* Called by cgroups to determine if a cpuset is usable; cpuset_mutex held */
1462 static int cpuset_can_attach(struct cgroup_taskset *tset)
1464 struct cgroup_subsys_state *css;
1466 struct task_struct *task;
1469 /* used later by cpuset_attach() */
1470 cpuset_attach_old_cs = task_cs(cgroup_taskset_first(tset, &css));
1473 mutex_lock(&cpuset_mutex);
1475 /* allow moving tasks into an empty cpuset if on default hierarchy */
1477 if (!cgroup_subsys_on_dfl(cpuset_cgrp_subsys) &&
1478 (cpumask_empty(cs->cpus_allowed) || nodes_empty(cs->mems_allowed)))
1481 cgroup_taskset_for_each(task, css, tset) {
1482 ret = task_can_attach(task, cs->cpus_allowed);
1485 ret = security_task_setscheduler(task);
1491 * Mark attach is in progress. This makes validate_change() fail
1492 * changes which zero cpus/mems_allowed.
1494 cs->attach_in_progress++;
1497 mutex_unlock(&cpuset_mutex);
1501 static void cpuset_cancel_attach(struct cgroup_taskset *tset)
1503 struct cgroup_subsys_state *css;
1506 cgroup_taskset_first(tset, &css);
1509 mutex_lock(&cpuset_mutex);
1510 css_cs(css)->attach_in_progress--;
1511 mutex_unlock(&cpuset_mutex);
1515 * Protected by cpuset_mutex. cpus_attach is used only by cpuset_attach()
1516 * but we can't allocate it dynamically there. Define it global and
1517 * allocate from cpuset_init().
1519 static cpumask_var_t cpus_attach;
1521 static void cpuset_attach(struct cgroup_taskset *tset)
1523 /* static buf protected by cpuset_mutex */
1524 static nodemask_t cpuset_attach_nodemask_to;
1525 struct task_struct *task;
1526 struct task_struct *leader;
1527 struct cgroup_subsys_state *css;
1529 struct cpuset *oldcs = cpuset_attach_old_cs;
1531 cgroup_taskset_first(tset, &css);
1534 mutex_lock(&cpuset_mutex);
1536 /* prepare for attach */
1537 if (cs == &top_cpuset)
1538 cpumask_copy(cpus_attach, cpu_possible_mask);
1540 guarantee_online_cpus(cs, cpus_attach);
1542 guarantee_online_mems(cs, &cpuset_attach_nodemask_to);
1544 cgroup_taskset_for_each(task, css, tset) {
1546 * can_attach beforehand should guarantee that this doesn't
1547 * fail. TODO: have a better way to handle failure here
1549 WARN_ON_ONCE(set_cpus_allowed_ptr(task, cpus_attach));
1551 cpuset_change_task_nodemask(task, &cpuset_attach_nodemask_to);
1552 cpuset_update_task_spread_flag(cs, task);
1556 * Change mm for all threadgroup leaders. This is expensive and may
1557 * sleep and should be moved outside migration path proper.
1559 cpuset_attach_nodemask_to = cs->effective_mems;
1560 cgroup_taskset_for_each_leader(leader, css, tset) {
1561 struct mm_struct *mm = get_task_mm(leader);
1564 mpol_rebind_mm(mm, &cpuset_attach_nodemask_to);
1567 * old_mems_allowed is the same with mems_allowed
1568 * here, except if this task is being moved
1569 * automatically due to hotplug. In that case
1570 * @mems_allowed has been updated and is empty, so
1571 * @old_mems_allowed is the right nodesets that we
1574 if (is_memory_migrate(cs))
1575 cpuset_migrate_mm(mm, &oldcs->old_mems_allowed,
1576 &cpuset_attach_nodemask_to);
1582 cs->old_mems_allowed = cpuset_attach_nodemask_to;
1584 cs->attach_in_progress--;
1585 if (!cs->attach_in_progress)
1586 wake_up(&cpuset_attach_wq);
1588 mutex_unlock(&cpuset_mutex);
1591 /* The various types of files and directories in a cpuset file system */
1594 FILE_MEMORY_MIGRATE,
1597 FILE_EFFECTIVE_CPULIST,
1598 FILE_EFFECTIVE_MEMLIST,
1602 FILE_SCHED_LOAD_BALANCE,
1603 FILE_SCHED_RELAX_DOMAIN_LEVEL,
1604 FILE_MEMORY_PRESSURE_ENABLED,
1605 FILE_MEMORY_PRESSURE,
1608 } cpuset_filetype_t;
1610 static int cpuset_write_u64(struct cgroup_subsys_state *css, struct cftype *cft,
1613 struct cpuset *cs = css_cs(css);
1614 cpuset_filetype_t type = cft->private;
1617 mutex_lock(&cpuset_mutex);
1618 if (!is_cpuset_online(cs)) {
1624 case FILE_CPU_EXCLUSIVE:
1625 retval = update_flag(CS_CPU_EXCLUSIVE, cs, val);
1627 case FILE_MEM_EXCLUSIVE:
1628 retval = update_flag(CS_MEM_EXCLUSIVE, cs, val);
1630 case FILE_MEM_HARDWALL:
1631 retval = update_flag(CS_MEM_HARDWALL, cs, val);
1633 case FILE_SCHED_LOAD_BALANCE:
1634 retval = update_flag(CS_SCHED_LOAD_BALANCE, cs, val);
1636 case FILE_MEMORY_MIGRATE:
1637 retval = update_flag(CS_MEMORY_MIGRATE, cs, val);
1639 case FILE_MEMORY_PRESSURE_ENABLED:
1640 cpuset_memory_pressure_enabled = !!val;
1642 case FILE_SPREAD_PAGE:
1643 retval = update_flag(CS_SPREAD_PAGE, cs, val);
1645 case FILE_SPREAD_SLAB:
1646 retval = update_flag(CS_SPREAD_SLAB, cs, val);
1653 mutex_unlock(&cpuset_mutex);
1657 static int cpuset_write_s64(struct cgroup_subsys_state *css, struct cftype *cft,
1660 struct cpuset *cs = css_cs(css);
1661 cpuset_filetype_t type = cft->private;
1662 int retval = -ENODEV;
1664 mutex_lock(&cpuset_mutex);
1665 if (!is_cpuset_online(cs))
1669 case FILE_SCHED_RELAX_DOMAIN_LEVEL:
1670 retval = update_relax_domain_level(cs, val);
1677 mutex_unlock(&cpuset_mutex);
1682 * Common handling for a write to a "cpus" or "mems" file.
1684 static ssize_t cpuset_write_resmask(struct kernfs_open_file *of,
1685 char *buf, size_t nbytes, loff_t off)
1687 struct cpuset *cs = css_cs(of_css(of));
1688 struct cpuset *trialcs;
1689 int retval = -ENODEV;
1691 buf = strstrip(buf);
1694 * CPU or memory hotunplug may leave @cs w/o any execution
1695 * resources, in which case the hotplug code asynchronously updates
1696 * configuration and transfers all tasks to the nearest ancestor
1697 * which can execute.
1699 * As writes to "cpus" or "mems" may restore @cs's execution
1700 * resources, wait for the previously scheduled operations before
1701 * proceeding, so that we don't end up keep removing tasks added
1702 * after execution capability is restored.
1704 * cpuset_hotplug_work calls back into cgroup core via
1705 * cgroup_transfer_tasks() and waiting for it from a cgroupfs
1706 * operation like this one can lead to a deadlock through kernfs
1707 * active_ref protection. Let's break the protection. Losing the
1708 * protection is okay as we check whether @cs is online after
1709 * grabbing cpuset_mutex anyway. This only happens on the legacy
1713 kernfs_break_active_protection(of->kn);
1714 flush_work(&cpuset_hotplug_work);
1716 mutex_lock(&cpuset_mutex);
1717 if (!is_cpuset_online(cs))
1720 trialcs = alloc_trial_cpuset(cs);
1726 switch (of_cft(of)->private) {
1728 retval = update_cpumask(cs, trialcs, buf);
1731 retval = update_nodemask(cs, trialcs, buf);
1738 free_trial_cpuset(trialcs);
1740 mutex_unlock(&cpuset_mutex);
1741 kernfs_unbreak_active_protection(of->kn);
1743 flush_workqueue(cpuset_migrate_mm_wq);
1744 return retval ?: nbytes;
1748 * These ascii lists should be read in a single call, by using a user
1749 * buffer large enough to hold the entire map. If read in smaller
1750 * chunks, there is no guarantee of atomicity. Since the display format
1751 * used, list of ranges of sequential numbers, is variable length,
1752 * and since these maps can change value dynamically, one could read
1753 * gibberish by doing partial reads while a list was changing.
1755 static int cpuset_common_seq_show(struct seq_file *sf, void *v)
1757 struct cpuset *cs = css_cs(seq_css(sf));
1758 cpuset_filetype_t type = seq_cft(sf)->private;
1761 spin_lock_irq(&callback_lock);
1765 seq_printf(sf, "%*pbl\n", cpumask_pr_args(cs->cpus_allowed));
1768 seq_printf(sf, "%*pbl\n", nodemask_pr_args(&cs->mems_allowed));
1770 case FILE_EFFECTIVE_CPULIST:
1771 seq_printf(sf, "%*pbl\n", cpumask_pr_args(cs->effective_cpus));
1773 case FILE_EFFECTIVE_MEMLIST:
1774 seq_printf(sf, "%*pbl\n", nodemask_pr_args(&cs->effective_mems));
1780 spin_unlock_irq(&callback_lock);
1784 static u64 cpuset_read_u64(struct cgroup_subsys_state *css, struct cftype *cft)
1786 struct cpuset *cs = css_cs(css);
1787 cpuset_filetype_t type = cft->private;
1789 case FILE_CPU_EXCLUSIVE:
1790 return is_cpu_exclusive(cs);
1791 case FILE_MEM_EXCLUSIVE:
1792 return is_mem_exclusive(cs);
1793 case FILE_MEM_HARDWALL:
1794 return is_mem_hardwall(cs);
1795 case FILE_SCHED_LOAD_BALANCE:
1796 return is_sched_load_balance(cs);
1797 case FILE_MEMORY_MIGRATE:
1798 return is_memory_migrate(cs);
1799 case FILE_MEMORY_PRESSURE_ENABLED:
1800 return cpuset_memory_pressure_enabled;
1801 case FILE_MEMORY_PRESSURE:
1802 return fmeter_getrate(&cs->fmeter);
1803 case FILE_SPREAD_PAGE:
1804 return is_spread_page(cs);
1805 case FILE_SPREAD_SLAB:
1806 return is_spread_slab(cs);
1811 /* Unreachable but makes gcc happy */
1815 static s64 cpuset_read_s64(struct cgroup_subsys_state *css, struct cftype *cft)
1817 struct cpuset *cs = css_cs(css);
1818 cpuset_filetype_t type = cft->private;
1820 case FILE_SCHED_RELAX_DOMAIN_LEVEL:
1821 return cs->relax_domain_level;
1826 /* Unrechable but makes gcc happy */
1832 * for the common functions, 'private' gives the type of file
1835 static struct cftype files[] = {
1838 .seq_show = cpuset_common_seq_show,
1839 .write = cpuset_write_resmask,
1840 .max_write_len = (100U + 6 * NR_CPUS),
1841 .private = FILE_CPULIST,
1846 .seq_show = cpuset_common_seq_show,
1847 .write = cpuset_write_resmask,
1848 .max_write_len = (100U + 6 * MAX_NUMNODES),
1849 .private = FILE_MEMLIST,
1853 .name = "effective_cpus",
1854 .seq_show = cpuset_common_seq_show,
1855 .private = FILE_EFFECTIVE_CPULIST,
1859 .name = "effective_mems",
1860 .seq_show = cpuset_common_seq_show,
1861 .private = FILE_EFFECTIVE_MEMLIST,
1865 .name = "cpu_exclusive",
1866 .read_u64 = cpuset_read_u64,
1867 .write_u64 = cpuset_write_u64,
1868 .private = FILE_CPU_EXCLUSIVE,
1872 .name = "mem_exclusive",
1873 .read_u64 = cpuset_read_u64,
1874 .write_u64 = cpuset_write_u64,
1875 .private = FILE_MEM_EXCLUSIVE,
1879 .name = "mem_hardwall",
1880 .read_u64 = cpuset_read_u64,
1881 .write_u64 = cpuset_write_u64,
1882 .private = FILE_MEM_HARDWALL,
1886 .name = "sched_load_balance",
1887 .read_u64 = cpuset_read_u64,
1888 .write_u64 = cpuset_write_u64,
1889 .private = FILE_SCHED_LOAD_BALANCE,
1893 .name = "sched_relax_domain_level",
1894 .read_s64 = cpuset_read_s64,
1895 .write_s64 = cpuset_write_s64,
1896 .private = FILE_SCHED_RELAX_DOMAIN_LEVEL,
1900 .name = "memory_migrate",
1901 .read_u64 = cpuset_read_u64,
1902 .write_u64 = cpuset_write_u64,
1903 .private = FILE_MEMORY_MIGRATE,
1907 .name = "memory_pressure",
1908 .read_u64 = cpuset_read_u64,
1912 .name = "memory_spread_page",
1913 .read_u64 = cpuset_read_u64,
1914 .write_u64 = cpuset_write_u64,
1915 .private = FILE_SPREAD_PAGE,
1919 .name = "memory_spread_slab",
1920 .read_u64 = cpuset_read_u64,
1921 .write_u64 = cpuset_write_u64,
1922 .private = FILE_SPREAD_SLAB,
1926 .name = "memory_pressure_enabled",
1927 .flags = CFTYPE_ONLY_ON_ROOT,
1928 .read_u64 = cpuset_read_u64,
1929 .write_u64 = cpuset_write_u64,
1930 .private = FILE_MEMORY_PRESSURE_ENABLED,
1937 * cpuset_css_alloc - allocate a cpuset css
1938 * cgrp: control group that the new cpuset will be part of
1941 static struct cgroup_subsys_state *
1942 cpuset_css_alloc(struct cgroup_subsys_state *parent_css)
1947 return &top_cpuset.css;
1949 cs = kzalloc(sizeof(*cs), GFP_KERNEL);
1951 return ERR_PTR(-ENOMEM);
1952 if (!alloc_cpumask_var(&cs->cpus_allowed, GFP_KERNEL))
1954 if (!alloc_cpumask_var(&cs->effective_cpus, GFP_KERNEL))
1957 set_bit(CS_SCHED_LOAD_BALANCE, &cs->flags);
1958 cpumask_clear(cs->cpus_allowed);
1959 nodes_clear(cs->mems_allowed);
1960 cpumask_clear(cs->effective_cpus);
1961 nodes_clear(cs->effective_mems);
1962 fmeter_init(&cs->fmeter);
1963 cs->relax_domain_level = -1;
1968 free_cpumask_var(cs->cpus_allowed);
1971 return ERR_PTR(-ENOMEM);
1974 static int cpuset_css_online(struct cgroup_subsys_state *css)
1976 struct cpuset *cs = css_cs(css);
1977 struct cpuset *parent = parent_cs(cs);
1978 struct cpuset *tmp_cs;
1979 struct cgroup_subsys_state *pos_css;
1984 mutex_lock(&cpuset_mutex);
1986 set_bit(CS_ONLINE, &cs->flags);
1987 if (is_spread_page(parent))
1988 set_bit(CS_SPREAD_PAGE, &cs->flags);
1989 if (is_spread_slab(parent))
1990 set_bit(CS_SPREAD_SLAB, &cs->flags);
1994 spin_lock_irq(&callback_lock);
1995 if (cgroup_subsys_on_dfl(cpuset_cgrp_subsys)) {
1996 cpumask_copy(cs->effective_cpus, parent->effective_cpus);
1997 cs->effective_mems = parent->effective_mems;
1999 spin_unlock_irq(&callback_lock);
2001 if (!test_bit(CGRP_CPUSET_CLONE_CHILDREN, &css->cgroup->flags))
2005 * Clone @parent's configuration if CGRP_CPUSET_CLONE_CHILDREN is
2006 * set. This flag handling is implemented in cgroup core for
2007 * histrical reasons - the flag may be specified during mount.
2009 * Currently, if any sibling cpusets have exclusive cpus or mem, we
2010 * refuse to clone the configuration - thereby refusing the task to
2011 * be entered, and as a result refusing the sys_unshare() or
2012 * clone() which initiated it. If this becomes a problem for some
2013 * users who wish to allow that scenario, then this could be
2014 * changed to grant parent->cpus_allowed-sibling_cpus_exclusive
2015 * (and likewise for mems) to the new cgroup.
2018 cpuset_for_each_child(tmp_cs, pos_css, parent) {
2019 if (is_mem_exclusive(tmp_cs) || is_cpu_exclusive(tmp_cs)) {
2026 spin_lock_irq(&callback_lock);
2027 cs->mems_allowed = parent->mems_allowed;
2028 cs->effective_mems = parent->mems_allowed;
2029 cpumask_copy(cs->cpus_allowed, parent->cpus_allowed);
2030 cpumask_copy(cs->effective_cpus, parent->cpus_allowed);
2031 spin_unlock_irq(&callback_lock);
2033 mutex_unlock(&cpuset_mutex);
2038 * If the cpuset being removed has its flag 'sched_load_balance'
2039 * enabled, then simulate turning sched_load_balance off, which
2040 * will call rebuild_sched_domains_locked().
2043 static void cpuset_css_offline(struct cgroup_subsys_state *css)
2045 struct cpuset *cs = css_cs(css);
2047 mutex_lock(&cpuset_mutex);
2049 if (is_sched_load_balance(cs))
2050 update_flag(CS_SCHED_LOAD_BALANCE, cs, 0);
2053 clear_bit(CS_ONLINE, &cs->flags);
2055 mutex_unlock(&cpuset_mutex);
2058 static void cpuset_css_free(struct cgroup_subsys_state *css)
2060 struct cpuset *cs = css_cs(css);
2062 free_cpumask_var(cs->effective_cpus);
2063 free_cpumask_var(cs->cpus_allowed);
2067 static void cpuset_bind(struct cgroup_subsys_state *root_css)
2069 mutex_lock(&cpuset_mutex);
2070 spin_lock_irq(&callback_lock);
2072 if (cgroup_subsys_on_dfl(cpuset_cgrp_subsys)) {
2073 cpumask_copy(top_cpuset.cpus_allowed, cpu_possible_mask);
2074 top_cpuset.mems_allowed = node_possible_map;
2076 cpumask_copy(top_cpuset.cpus_allowed,
2077 top_cpuset.effective_cpus);
2078 top_cpuset.mems_allowed = top_cpuset.effective_mems;
2081 spin_unlock_irq(&callback_lock);
2082 mutex_unlock(&cpuset_mutex);
2086 * Make sure the new task conform to the current state of its parent,
2087 * which could have been changed by cpuset just after it inherits the
2088 * state from the parent and before it sits on the cgroup's task list.
2090 static void cpuset_fork(struct task_struct *task)
2092 if (task_css_is_root(task, cpuset_cgrp_id))
2095 set_cpus_allowed_ptr(task, ¤t->cpus_allowed);
2096 task->mems_allowed = current->mems_allowed;
2099 struct cgroup_subsys cpuset_cgrp_subsys = {
2100 .css_alloc = cpuset_css_alloc,
2101 .css_online = cpuset_css_online,
2102 .css_offline = cpuset_css_offline,
2103 .css_free = cpuset_css_free,
2104 .can_attach = cpuset_can_attach,
2105 .cancel_attach = cpuset_cancel_attach,
2106 .attach = cpuset_attach,
2107 .post_attach = cpuset_post_attach,
2108 .bind = cpuset_bind,
2109 .fork = cpuset_fork,
2110 .legacy_cftypes = files,
2115 * cpuset_init - initialize cpusets at system boot
2117 * Description: Initialize top_cpuset and the cpuset internal file system,
2120 int __init cpuset_init(void)
2124 BUG_ON(!alloc_cpumask_var(&top_cpuset.cpus_allowed, GFP_KERNEL));
2125 BUG_ON(!alloc_cpumask_var(&top_cpuset.effective_cpus, GFP_KERNEL));
2127 cpumask_setall(top_cpuset.cpus_allowed);
2128 nodes_setall(top_cpuset.mems_allowed);
2129 cpumask_setall(top_cpuset.effective_cpus);
2130 nodes_setall(top_cpuset.effective_mems);
2132 fmeter_init(&top_cpuset.fmeter);
2133 set_bit(CS_SCHED_LOAD_BALANCE, &top_cpuset.flags);
2134 top_cpuset.relax_domain_level = -1;
2136 err = register_filesystem(&cpuset_fs_type);
2140 BUG_ON(!alloc_cpumask_var(&cpus_attach, GFP_KERNEL));
2146 * If CPU and/or memory hotplug handlers, below, unplug any CPUs
2147 * or memory nodes, we need to walk over the cpuset hierarchy,
2148 * removing that CPU or node from all cpusets. If this removes the
2149 * last CPU or node from a cpuset, then move the tasks in the empty
2150 * cpuset to its next-highest non-empty parent.
2152 static void remove_tasks_in_empty_cpuset(struct cpuset *cs)
2154 struct cpuset *parent;
2157 * Find its next-highest non-empty parent, (top cpuset
2158 * has online cpus, so can't be empty).
2160 parent = parent_cs(cs);
2161 while (cpumask_empty(parent->cpus_allowed) ||
2162 nodes_empty(parent->mems_allowed))
2163 parent = parent_cs(parent);
2165 if (cgroup_transfer_tasks(parent->css.cgroup, cs->css.cgroup)) {
2166 pr_err("cpuset: failed to transfer tasks out of empty cpuset ");
2167 pr_cont_cgroup_name(cs->css.cgroup);
2173 hotplug_update_tasks_legacy(struct cpuset *cs,
2174 struct cpumask *new_cpus, nodemask_t *new_mems,
2175 bool cpus_updated, bool mems_updated)
2179 spin_lock_irq(&callback_lock);
2180 cpumask_copy(cs->cpus_allowed, new_cpus);
2181 cpumask_copy(cs->effective_cpus, new_cpus);
2182 cs->mems_allowed = *new_mems;
2183 cs->effective_mems = *new_mems;
2184 spin_unlock_irq(&callback_lock);
2187 * Don't call update_tasks_cpumask() if the cpuset becomes empty,
2188 * as the tasks will be migratecd to an ancestor.
2190 if (cpus_updated && !cpumask_empty(cs->cpus_allowed))
2191 update_tasks_cpumask(cs);
2192 if (mems_updated && !nodes_empty(cs->mems_allowed))
2193 update_tasks_nodemask(cs);
2195 is_empty = cpumask_empty(cs->cpus_allowed) ||
2196 nodes_empty(cs->mems_allowed);
2198 mutex_unlock(&cpuset_mutex);
2201 * Move tasks to the nearest ancestor with execution resources,
2202 * This is full cgroup operation which will also call back into
2203 * cpuset. Should be done outside any lock.
2206 remove_tasks_in_empty_cpuset(cs);
2208 mutex_lock(&cpuset_mutex);
2212 hotplug_update_tasks(struct cpuset *cs,
2213 struct cpumask *new_cpus, nodemask_t *new_mems,
2214 bool cpus_updated, bool mems_updated)
2216 if (cpumask_empty(new_cpus))
2217 cpumask_copy(new_cpus, parent_cs(cs)->effective_cpus);
2218 if (nodes_empty(*new_mems))
2219 *new_mems = parent_cs(cs)->effective_mems;
2221 spin_lock_irq(&callback_lock);
2222 cpumask_copy(cs->effective_cpus, new_cpus);
2223 cs->effective_mems = *new_mems;
2224 spin_unlock_irq(&callback_lock);
2227 update_tasks_cpumask(cs);
2229 update_tasks_nodemask(cs);
2233 * cpuset_hotplug_update_tasks - update tasks in a cpuset for hotunplug
2234 * @cs: cpuset in interest
2236 * Compare @cs's cpu and mem masks against top_cpuset and if some have gone
2237 * offline, update @cs accordingly. If @cs ends up with no CPU or memory,
2238 * all its tasks are moved to the nearest ancestor with both resources.
2240 static void cpuset_hotplug_update_tasks(struct cpuset *cs)
2242 static cpumask_t new_cpus;
2243 static nodemask_t new_mems;
2247 wait_event(cpuset_attach_wq, cs->attach_in_progress == 0);
2249 mutex_lock(&cpuset_mutex);
2252 * We have raced with task attaching. We wait until attaching
2253 * is finished, so we won't attach a task to an empty cpuset.
2255 if (cs->attach_in_progress) {
2256 mutex_unlock(&cpuset_mutex);
2260 cpumask_and(&new_cpus, cs->cpus_allowed, parent_cs(cs)->effective_cpus);
2261 nodes_and(new_mems, cs->mems_allowed, parent_cs(cs)->effective_mems);
2263 cpus_updated = !cpumask_equal(&new_cpus, cs->effective_cpus);
2264 mems_updated = !nodes_equal(new_mems, cs->effective_mems);
2266 if (cgroup_subsys_on_dfl(cpuset_cgrp_subsys))
2267 hotplug_update_tasks(cs, &new_cpus, &new_mems,
2268 cpus_updated, mems_updated);
2270 hotplug_update_tasks_legacy(cs, &new_cpus, &new_mems,
2271 cpus_updated, mems_updated);
2273 mutex_unlock(&cpuset_mutex);
2277 * cpuset_hotplug_workfn - handle CPU/memory hotunplug for a cpuset
2279 * This function is called after either CPU or memory configuration has
2280 * changed and updates cpuset accordingly. The top_cpuset is always
2281 * synchronized to cpu_active_mask and N_MEMORY, which is necessary in
2282 * order to make cpusets transparent (of no affect) on systems that are
2283 * actively using CPU hotplug but making no active use of cpusets.
2285 * Non-root cpusets are only affected by offlining. If any CPUs or memory
2286 * nodes have been taken down, cpuset_hotplug_update_tasks() is invoked on
2289 * Note that CPU offlining during suspend is ignored. We don't modify
2290 * cpusets across suspend/resume cycles at all.
2292 static void cpuset_hotplug_workfn(struct work_struct *work)
2294 static cpumask_t new_cpus;
2295 static nodemask_t new_mems;
2296 bool cpus_updated, mems_updated;
2297 bool on_dfl = cgroup_subsys_on_dfl(cpuset_cgrp_subsys);
2299 mutex_lock(&cpuset_mutex);
2301 /* fetch the available cpus/mems and find out which changed how */
2302 cpumask_copy(&new_cpus, cpu_active_mask);
2303 new_mems = node_states[N_MEMORY];
2305 cpus_updated = !cpumask_equal(top_cpuset.effective_cpus, &new_cpus);
2306 mems_updated = !nodes_equal(top_cpuset.effective_mems, new_mems);
2308 /* synchronize cpus_allowed to cpu_active_mask */
2310 spin_lock_irq(&callback_lock);
2312 cpumask_copy(top_cpuset.cpus_allowed, &new_cpus);
2313 cpumask_copy(top_cpuset.effective_cpus, &new_cpus);
2314 spin_unlock_irq(&callback_lock);
2315 /* we don't mess with cpumasks of tasks in top_cpuset */
2318 /* synchronize mems_allowed to N_MEMORY */
2320 spin_lock_irq(&callback_lock);
2322 top_cpuset.mems_allowed = new_mems;
2323 top_cpuset.effective_mems = new_mems;
2324 spin_unlock_irq(&callback_lock);
2325 update_tasks_nodemask(&top_cpuset);
2328 mutex_unlock(&cpuset_mutex);
2330 /* if cpus or mems changed, we need to propagate to descendants */
2331 if (cpus_updated || mems_updated) {
2333 struct cgroup_subsys_state *pos_css;
2336 cpuset_for_each_descendant_pre(cs, pos_css, &top_cpuset) {
2337 if (cs == &top_cpuset || !css_tryget_online(&cs->css))
2341 cpuset_hotplug_update_tasks(cs);
2349 /* rebuild sched domains if cpus_allowed has changed */
2351 rebuild_sched_domains();
2354 void cpuset_update_active_cpus(void)
2357 * We're inside cpu hotplug critical region which usually nests
2358 * inside cgroup synchronization. Bounce actual hotplug processing
2359 * to a work item to avoid reverse locking order.
2361 * We still need to do partition_sched_domains() synchronously;
2362 * otherwise, the scheduler will get confused and put tasks to the
2363 * dead CPU. Fall back to the default single domain.
2364 * cpuset_hotplug_workfn() will rebuild it as necessary.
2366 partition_sched_domains(1, NULL, NULL);
2367 schedule_work(&cpuset_hotplug_work);
2371 * Keep top_cpuset.mems_allowed tracking node_states[N_MEMORY].
2372 * Call this routine anytime after node_states[N_MEMORY] changes.
2373 * See cpuset_update_active_cpus() for CPU hotplug handling.
2375 static int cpuset_track_online_nodes(struct notifier_block *self,
2376 unsigned long action, void *arg)
2378 schedule_work(&cpuset_hotplug_work);
2382 static struct notifier_block cpuset_track_online_nodes_nb = {
2383 .notifier_call = cpuset_track_online_nodes,
2384 .priority = 10, /* ??! */
2388 * cpuset_init_smp - initialize cpus_allowed
2390 * Description: Finish top cpuset after cpu, node maps are initialized
2392 void __init cpuset_init_smp(void)
2394 cpumask_copy(top_cpuset.cpus_allowed, cpu_active_mask);
2395 top_cpuset.mems_allowed = node_states[N_MEMORY];
2396 top_cpuset.old_mems_allowed = top_cpuset.mems_allowed;
2398 cpumask_copy(top_cpuset.effective_cpus, cpu_active_mask);
2399 top_cpuset.effective_mems = node_states[N_MEMORY];
2401 register_hotmemory_notifier(&cpuset_track_online_nodes_nb);
2403 cpuset_migrate_mm_wq = alloc_ordered_workqueue("cpuset_migrate_mm", 0);
2404 BUG_ON(!cpuset_migrate_mm_wq);
2408 * cpuset_cpus_allowed - return cpus_allowed mask from a tasks cpuset.
2409 * @tsk: pointer to task_struct from which to obtain cpuset->cpus_allowed.
2410 * @pmask: pointer to struct cpumask variable to receive cpus_allowed set.
2412 * Description: Returns the cpumask_var_t cpus_allowed of the cpuset
2413 * attached to the specified @tsk. Guaranteed to return some non-empty
2414 * subset of cpu_online_mask, even if this means going outside the
2418 void cpuset_cpus_allowed(struct task_struct *tsk, struct cpumask *pmask)
2420 unsigned long flags;
2422 spin_lock_irqsave(&callback_lock, flags);
2424 guarantee_online_cpus(task_cs(tsk), pmask);
2426 spin_unlock_irqrestore(&callback_lock, flags);
2429 void cpuset_cpus_allowed_fallback(struct task_struct *tsk)
2432 do_set_cpus_allowed(tsk, task_cs(tsk)->effective_cpus);
2436 * We own tsk->cpus_allowed, nobody can change it under us.
2438 * But we used cs && cs->cpus_allowed lockless and thus can
2439 * race with cgroup_attach_task() or update_cpumask() and get
2440 * the wrong tsk->cpus_allowed. However, both cases imply the
2441 * subsequent cpuset_change_cpumask()->set_cpus_allowed_ptr()
2442 * which takes task_rq_lock().
2444 * If we are called after it dropped the lock we must see all
2445 * changes in tsk_cs()->cpus_allowed. Otherwise we can temporary
2446 * set any mask even if it is not right from task_cs() pov,
2447 * the pending set_cpus_allowed_ptr() will fix things.
2449 * select_fallback_rq() will fix things ups and set cpu_possible_mask
2454 void __init cpuset_init_current_mems_allowed(void)
2456 nodes_setall(current->mems_allowed);
2460 * cpuset_mems_allowed - return mems_allowed mask from a tasks cpuset.
2461 * @tsk: pointer to task_struct from which to obtain cpuset->mems_allowed.
2463 * Description: Returns the nodemask_t mems_allowed of the cpuset
2464 * attached to the specified @tsk. Guaranteed to return some non-empty
2465 * subset of node_states[N_MEMORY], even if this means going outside the
2469 nodemask_t cpuset_mems_allowed(struct task_struct *tsk)
2472 unsigned long flags;
2474 spin_lock_irqsave(&callback_lock, flags);
2476 guarantee_online_mems(task_cs(tsk), &mask);
2478 spin_unlock_irqrestore(&callback_lock, flags);
2484 * cpuset_nodemask_valid_mems_allowed - check nodemask vs. curremt mems_allowed
2485 * @nodemask: the nodemask to be checked
2487 * Are any of the nodes in the nodemask allowed in current->mems_allowed?
2489 int cpuset_nodemask_valid_mems_allowed(nodemask_t *nodemask)
2491 return nodes_intersects(*nodemask, current->mems_allowed);
2495 * nearest_hardwall_ancestor() - Returns the nearest mem_exclusive or
2496 * mem_hardwall ancestor to the specified cpuset. Call holding
2497 * callback_lock. If no ancestor is mem_exclusive or mem_hardwall
2498 * (an unusual configuration), then returns the root cpuset.
2500 static struct cpuset *nearest_hardwall_ancestor(struct cpuset *cs)
2502 while (!(is_mem_exclusive(cs) || is_mem_hardwall(cs)) && parent_cs(cs))
2508 * cpuset_node_allowed - Can we allocate on a memory node?
2509 * @node: is this an allowed node?
2510 * @gfp_mask: memory allocation flags
2512 * If we're in interrupt, yes, we can always allocate. If @node is set in
2513 * current's mems_allowed, yes. If it's not a __GFP_HARDWALL request and this
2514 * node is set in the nearest hardwalled cpuset ancestor to current's cpuset,
2515 * yes. If current has access to memory reserves due to TIF_MEMDIE, yes.
2518 * GFP_USER allocations are marked with the __GFP_HARDWALL bit,
2519 * and do not allow allocations outside the current tasks cpuset
2520 * unless the task has been OOM killed as is marked TIF_MEMDIE.
2521 * GFP_KERNEL allocations are not so marked, so can escape to the
2522 * nearest enclosing hardwalled ancestor cpuset.
2524 * Scanning up parent cpusets requires callback_lock. The
2525 * __alloc_pages() routine only calls here with __GFP_HARDWALL bit
2526 * _not_ set if it's a GFP_KERNEL allocation, and all nodes in the
2527 * current tasks mems_allowed came up empty on the first pass over
2528 * the zonelist. So only GFP_KERNEL allocations, if all nodes in the
2529 * cpuset are short of memory, might require taking the callback_lock.
2531 * The first call here from mm/page_alloc:get_page_from_freelist()
2532 * has __GFP_HARDWALL set in gfp_mask, enforcing hardwall cpusets,
2533 * so no allocation on a node outside the cpuset is allowed (unless
2534 * in interrupt, of course).
2536 * The second pass through get_page_from_freelist() doesn't even call
2537 * here for GFP_ATOMIC calls. For those calls, the __alloc_pages()
2538 * variable 'wait' is not set, and the bit ALLOC_CPUSET is not set
2539 * in alloc_flags. That logic and the checks below have the combined
2541 * in_interrupt - any node ok (current task context irrelevant)
2542 * GFP_ATOMIC - any node ok
2543 * TIF_MEMDIE - any node ok
2544 * GFP_KERNEL - any node in enclosing hardwalled cpuset ok
2545 * GFP_USER - only nodes in current tasks mems allowed ok.
2547 bool __cpuset_node_allowed(int node, gfp_t gfp_mask)
2549 struct cpuset *cs; /* current cpuset ancestors */
2550 int allowed; /* is allocation in zone z allowed? */
2551 unsigned long flags;
2555 if (node_isset(node, current->mems_allowed))
2558 * Allow tasks that have access to memory reserves because they have
2559 * been OOM killed to get memory anywhere.
2561 if (unlikely(test_thread_flag(TIF_MEMDIE)))
2563 if (gfp_mask & __GFP_HARDWALL) /* If hardwall request, stop here */
2566 if (current->flags & PF_EXITING) /* Let dying task have memory */
2569 /* Not hardwall and node outside mems_allowed: scan up cpusets */
2570 spin_lock_irqsave(&callback_lock, flags);
2573 cs = nearest_hardwall_ancestor(task_cs(current));
2574 allowed = node_isset(node, cs->mems_allowed);
2577 spin_unlock_irqrestore(&callback_lock, flags);
2582 * cpuset_mem_spread_node() - On which node to begin search for a file page
2583 * cpuset_slab_spread_node() - On which node to begin search for a slab page
2585 * If a task is marked PF_SPREAD_PAGE or PF_SPREAD_SLAB (as for
2586 * tasks in a cpuset with is_spread_page or is_spread_slab set),
2587 * and if the memory allocation used cpuset_mem_spread_node()
2588 * to determine on which node to start looking, as it will for
2589 * certain page cache or slab cache pages such as used for file
2590 * system buffers and inode caches, then instead of starting on the
2591 * local node to look for a free page, rather spread the starting
2592 * node around the tasks mems_allowed nodes.
2594 * We don't have to worry about the returned node being offline
2595 * because "it can't happen", and even if it did, it would be ok.
2597 * The routines calling guarantee_online_mems() are careful to
2598 * only set nodes in task->mems_allowed that are online. So it
2599 * should not be possible for the following code to return an
2600 * offline node. But if it did, that would be ok, as this routine
2601 * is not returning the node where the allocation must be, only
2602 * the node where the search should start. The zonelist passed to
2603 * __alloc_pages() will include all nodes. If the slab allocator
2604 * is passed an offline node, it will fall back to the local node.
2605 * See kmem_cache_alloc_node().
2608 static int cpuset_spread_node(int *rotor)
2610 return *rotor = next_node_in(*rotor, current->mems_allowed);
2613 int cpuset_mem_spread_node(void)
2615 if (current->cpuset_mem_spread_rotor == NUMA_NO_NODE)
2616 current->cpuset_mem_spread_rotor =
2617 node_random(¤t->mems_allowed);
2619 return cpuset_spread_node(¤t->cpuset_mem_spread_rotor);
2622 int cpuset_slab_spread_node(void)
2624 if (current->cpuset_slab_spread_rotor == NUMA_NO_NODE)
2625 current->cpuset_slab_spread_rotor =
2626 node_random(¤t->mems_allowed);
2628 return cpuset_spread_node(¤t->cpuset_slab_spread_rotor);
2631 EXPORT_SYMBOL_GPL(cpuset_mem_spread_node);
2634 * cpuset_mems_allowed_intersects - Does @tsk1's mems_allowed intersect @tsk2's?
2635 * @tsk1: pointer to task_struct of some task.
2636 * @tsk2: pointer to task_struct of some other task.
2638 * Description: Return true if @tsk1's mems_allowed intersects the
2639 * mems_allowed of @tsk2. Used by the OOM killer to determine if
2640 * one of the task's memory usage might impact the memory available
2644 int cpuset_mems_allowed_intersects(const struct task_struct *tsk1,
2645 const struct task_struct *tsk2)
2647 return nodes_intersects(tsk1->mems_allowed, tsk2->mems_allowed);
2651 * cpuset_print_current_mems_allowed - prints current's cpuset and mems_allowed
2653 * Description: Prints current's name, cpuset name, and cached copy of its
2654 * mems_allowed to the kernel log.
2656 void cpuset_print_current_mems_allowed(void)
2658 struct cgroup *cgrp;
2662 cgrp = task_cs(current)->css.cgroup;
2663 pr_info("%s cpuset=", current->comm);
2664 pr_cont_cgroup_name(cgrp);
2665 pr_cont(" mems_allowed=%*pbl\n",
2666 nodemask_pr_args(¤t->mems_allowed));
2672 * Collection of memory_pressure is suppressed unless
2673 * this flag is enabled by writing "1" to the special
2674 * cpuset file 'memory_pressure_enabled' in the root cpuset.
2677 int cpuset_memory_pressure_enabled __read_mostly;
2680 * cpuset_memory_pressure_bump - keep stats of per-cpuset reclaims.
2682 * Keep a running average of the rate of synchronous (direct)
2683 * page reclaim efforts initiated by tasks in each cpuset.
2685 * This represents the rate at which some task in the cpuset
2686 * ran low on memory on all nodes it was allowed to use, and
2687 * had to enter the kernels page reclaim code in an effort to
2688 * create more free memory by tossing clean pages or swapping
2689 * or writing dirty pages.
2691 * Display to user space in the per-cpuset read-only file
2692 * "memory_pressure". Value displayed is an integer
2693 * representing the recent rate of entry into the synchronous
2694 * (direct) page reclaim by any task attached to the cpuset.
2697 void __cpuset_memory_pressure_bump(void)
2700 fmeter_markevent(&task_cs(current)->fmeter);
2704 #ifdef CONFIG_PROC_PID_CPUSET
2706 * proc_cpuset_show()
2707 * - Print tasks cpuset path into seq_file.
2708 * - Used for /proc/<pid>/cpuset.
2709 * - No need to task_lock(tsk) on this tsk->cpuset reference, as it
2710 * doesn't really matter if tsk->cpuset changes after we read it,
2711 * and we take cpuset_mutex, keeping cpuset_attach() from changing it
2714 int proc_cpuset_show(struct seq_file *m, struct pid_namespace *ns,
2715 struct pid *pid, struct task_struct *tsk)
2718 struct cgroup_subsys_state *css;
2722 buf = kmalloc(PATH_MAX, GFP_KERNEL);
2726 css = task_get_css(tsk, cpuset_cgrp_id);
2727 retval = cgroup_path_ns(css->cgroup, buf, PATH_MAX,
2728 current->nsproxy->cgroup_ns);
2730 if (retval >= PATH_MAX)
2731 retval = -ENAMETOOLONG;
2742 #endif /* CONFIG_PROC_PID_CPUSET */
2744 /* Display task mems_allowed in /proc/<pid>/status file. */
2745 void cpuset_task_status_allowed(struct seq_file *m, struct task_struct *task)
2747 seq_printf(m, "Mems_allowed:\t%*pb\n",
2748 nodemask_pr_args(&task->mems_allowed));
2749 seq_printf(m, "Mems_allowed_list:\t%*pbl\n",
2750 nodemask_pr_args(&task->mems_allowed));