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/time64.h>
55 #include <linux/backing-dev.h>
56 #include <linux/sort.h>
58 #include <asm/uaccess.h>
59 #include <linux/atomic.h>
60 #include <linux/mutex.h>
61 #include <linux/workqueue.h>
62 #include <linux/cgroup.h>
63 #include <linux/wait.h>
65 struct static_key cpusets_enabled_key __read_mostly = STATIC_KEY_INIT_FALSE;
67 /* See "Frequency meter" comments, below. */
70 int cnt; /* unprocessed events count */
71 int val; /* most recent output value */
72 time64_t time; /* clock (secs) when val computed */
73 spinlock_t lock; /* guards read or write of above */
77 struct cgroup_subsys_state css;
79 unsigned long flags; /* "unsigned long" so bitops work */
82 * On default hierarchy:
84 * The user-configured masks can only be changed by writing to
85 * cpuset.cpus and cpuset.mems, and won't be limited by the
88 * The effective masks is the real masks that apply to the tasks
89 * in the cpuset. They may be changed if the configured masks are
90 * changed or hotplug happens.
92 * effective_mask == configured_mask & parent's effective_mask,
93 * and if it ends up empty, it will inherit the parent's mask.
98 * The user-configured masks are always the same with effective masks.
101 /* user-configured CPUs and Memory Nodes allow to tasks */
102 cpumask_var_t cpus_allowed;
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(const struct cpuset *cs)
180 return test_bit(CS_ONLINE, &cs->flags);
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);
291 * CPU / memory hotplug is handled asynchronously.
293 static void cpuset_hotplug_workfn(struct work_struct *work);
294 static DECLARE_WORK(cpuset_hotplug_work, cpuset_hotplug_workfn);
296 static DECLARE_WAIT_QUEUE_HEAD(cpuset_attach_wq);
299 * This is ugly, but preserves the userspace API for existing cpuset
300 * users. If someone tries to mount the "cpuset" filesystem, we
301 * silently switch it to mount "cgroup" instead
303 static struct dentry *cpuset_mount(struct file_system_type *fs_type,
304 int flags, const char *unused_dev_name, void *data)
306 struct file_system_type *cgroup_fs = get_fs_type("cgroup");
307 struct dentry *ret = ERR_PTR(-ENODEV);
311 "release_agent=/sbin/cpuset_release_agent";
312 ret = cgroup_fs->mount(cgroup_fs, flags,
313 unused_dev_name, mountopts);
314 put_filesystem(cgroup_fs);
319 static struct file_system_type cpuset_fs_type = {
321 .mount = cpuset_mount,
325 * Return in pmask the portion of a cpusets's cpus_allowed that
326 * are online. If none are online, walk up the cpuset hierarchy
327 * until we find one that does have some online cpus. The top
328 * cpuset always has some cpus online.
330 * One way or another, we guarantee to return some non-empty subset
331 * of cpu_online_mask.
333 * Call with callback_lock or cpuset_mutex held.
335 static void guarantee_online_cpus(struct cpuset *cs, struct cpumask *pmask)
337 while (!cpumask_intersects(cs->effective_cpus, cpu_online_mask))
339 cpumask_and(pmask, cs->effective_cpus, cpu_online_mask);
343 * Return in *pmask the portion of a cpusets's mems_allowed that
344 * are online, with memory. If none are online with memory, walk
345 * up the cpuset hierarchy until we find one that does have some
346 * online mems. The top cpuset always has some mems online.
348 * One way or another, we guarantee to return some non-empty subset
349 * of node_states[N_MEMORY].
351 * Call with callback_lock or cpuset_mutex held.
353 static void guarantee_online_mems(struct cpuset *cs, nodemask_t *pmask)
355 while (!nodes_intersects(cs->effective_mems, node_states[N_MEMORY]))
357 nodes_and(*pmask, cs->effective_mems, node_states[N_MEMORY]);
361 * update task's spread flag if cpuset's page/slab spread flag is set
363 * Call with callback_lock or cpuset_mutex held.
365 static void cpuset_update_task_spread_flag(struct cpuset *cs,
366 struct task_struct *tsk)
368 if (is_spread_page(cs))
369 task_set_spread_page(tsk);
371 task_clear_spread_page(tsk);
373 if (is_spread_slab(cs))
374 task_set_spread_slab(tsk);
376 task_clear_spread_slab(tsk);
380 * is_cpuset_subset(p, q) - Is cpuset p a subset of cpuset q?
382 * One cpuset is a subset of another if all its allowed CPUs and
383 * Memory Nodes are a subset of the other, and its exclusive flags
384 * are only set if the other's are set. Call holding cpuset_mutex.
387 static int is_cpuset_subset(const struct cpuset *p, const struct cpuset *q)
389 return cpumask_subset(p->cpus_allowed, q->cpus_allowed) &&
390 nodes_subset(p->mems_allowed, q->mems_allowed) &&
391 is_cpu_exclusive(p) <= is_cpu_exclusive(q) &&
392 is_mem_exclusive(p) <= is_mem_exclusive(q);
396 * alloc_trial_cpuset - allocate a trial cpuset
397 * @cs: the cpuset that the trial cpuset duplicates
399 static struct cpuset *alloc_trial_cpuset(struct cpuset *cs)
401 struct cpuset *trial;
403 trial = kmemdup(cs, sizeof(*cs), GFP_KERNEL);
407 if (!alloc_cpumask_var(&trial->cpus_allowed, GFP_KERNEL))
409 if (!alloc_cpumask_var(&trial->effective_cpus, GFP_KERNEL))
412 cpumask_copy(trial->cpus_allowed, cs->cpus_allowed);
413 cpumask_copy(trial->effective_cpus, cs->effective_cpus);
417 free_cpumask_var(trial->cpus_allowed);
424 * free_trial_cpuset - free the trial cpuset
425 * @trial: the trial cpuset to be freed
427 static void free_trial_cpuset(struct cpuset *trial)
429 free_cpumask_var(trial->effective_cpus);
430 free_cpumask_var(trial->cpus_allowed);
435 * validate_change() - Used to validate that any proposed cpuset change
436 * follows the structural rules for cpusets.
438 * If we replaced the flag and mask values of the current cpuset
439 * (cur) with those values in the trial cpuset (trial), would
440 * our various subset and exclusive rules still be valid? Presumes
443 * 'cur' is the address of an actual, in-use cpuset. Operations
444 * such as list traversal that depend on the actual address of the
445 * cpuset in the list must use cur below, not trial.
447 * 'trial' is the address of bulk structure copy of cur, with
448 * perhaps one or more of the fields cpus_allowed, mems_allowed,
449 * or flags changed to new, trial values.
451 * Return 0 if valid, -errno if not.
454 static int validate_change(struct cpuset *cur, struct cpuset *trial)
456 struct cgroup_subsys_state *css;
457 struct cpuset *c, *par;
462 /* Each of our child cpusets must be a subset of us */
464 cpuset_for_each_child(c, css, cur)
465 if (!is_cpuset_subset(c, trial))
468 /* Remaining checks don't apply to root cpuset */
470 if (cur == &top_cpuset)
473 par = parent_cs(cur);
475 /* On legacy hiearchy, we must be a subset of our parent cpuset. */
477 if (!cgroup_on_dfl(cur->css.cgroup) && !is_cpuset_subset(trial, par))
481 * If either I or some sibling (!= me) is exclusive, we can't
485 cpuset_for_each_child(c, css, par) {
486 if ((is_cpu_exclusive(trial) || is_cpu_exclusive(c)) &&
488 cpumask_intersects(trial->cpus_allowed, c->cpus_allowed))
490 if ((is_mem_exclusive(trial) || is_mem_exclusive(c)) &&
492 nodes_intersects(trial->mems_allowed, c->mems_allowed))
497 * Cpusets with tasks - existing or newly being attached - can't
498 * be changed to have empty cpus_allowed or mems_allowed.
501 if ((cgroup_has_tasks(cur->css.cgroup) || cur->attach_in_progress)) {
502 if (!cpumask_empty(cur->cpus_allowed) &&
503 cpumask_empty(trial->cpus_allowed))
505 if (!nodes_empty(cur->mems_allowed) &&
506 nodes_empty(trial->mems_allowed))
511 * We can't shrink if we won't have enough room for SCHED_DEADLINE
515 if (is_cpu_exclusive(cur) &&
516 !cpuset_cpumask_can_shrink(cur->cpus_allowed,
517 trial->cpus_allowed))
528 * Helper routine for generate_sched_domains().
529 * Do cpusets a, b have overlapping effective cpus_allowed masks?
531 static int cpusets_overlap(struct cpuset *a, struct cpuset *b)
533 return cpumask_intersects(a->effective_cpus, b->effective_cpus);
537 update_domain_attr(struct sched_domain_attr *dattr, struct cpuset *c)
539 if (dattr->relax_domain_level < c->relax_domain_level)
540 dattr->relax_domain_level = c->relax_domain_level;
544 static void update_domain_attr_tree(struct sched_domain_attr *dattr,
545 struct cpuset *root_cs)
548 struct cgroup_subsys_state *pos_css;
551 cpuset_for_each_descendant_pre(cp, pos_css, root_cs) {
552 /* skip the whole subtree if @cp doesn't have any CPU */
553 if (cpumask_empty(cp->cpus_allowed)) {
554 pos_css = css_rightmost_descendant(pos_css);
558 if (is_sched_load_balance(cp))
559 update_domain_attr(dattr, cp);
565 * generate_sched_domains()
567 * This function builds a partial partition of the systems CPUs
568 * A 'partial partition' is a set of non-overlapping subsets whose
569 * union is a subset of that set.
570 * The output of this function needs to be passed to kernel/sched/core.c
571 * partition_sched_domains() routine, which will rebuild the scheduler's
572 * load balancing domains (sched domains) as specified by that partial
575 * See "What is sched_load_balance" in Documentation/cgroups/cpusets.txt
576 * for a background explanation of this.
578 * Does not return errors, on the theory that the callers of this
579 * routine would rather not worry about failures to rebuild sched
580 * domains when operating in the severe memory shortage situations
581 * that could cause allocation failures below.
583 * Must be called with cpuset_mutex held.
585 * The three key local variables below are:
586 * q - a linked-list queue of cpuset pointers, used to implement a
587 * top-down scan of all cpusets. This scan loads a pointer
588 * to each cpuset marked is_sched_load_balance into the
589 * array 'csa'. For our purposes, rebuilding the schedulers
590 * sched domains, we can ignore !is_sched_load_balance cpusets.
591 * csa - (for CpuSet Array) Array of pointers to all the cpusets
592 * that need to be load balanced, for convenient iterative
593 * access by the subsequent code that finds the best partition,
594 * i.e the set of domains (subsets) of CPUs such that the
595 * cpus_allowed of every cpuset marked is_sched_load_balance
596 * is a subset of one of these domains, while there are as
597 * many such domains as possible, each as small as possible.
598 * doms - Conversion of 'csa' to an array of cpumasks, for passing to
599 * the kernel/sched/core.c routine partition_sched_domains() in a
600 * convenient format, that can be easily compared to the prior
601 * value to determine what partition elements (sched domains)
602 * were changed (added or removed.)
604 * Finding the best partition (set of domains):
605 * The triple nested loops below over i, j, k scan over the
606 * load balanced cpusets (using the array of cpuset pointers in
607 * csa[]) looking for pairs of cpusets that have overlapping
608 * cpus_allowed, but which don't have the same 'pn' partition
609 * number and gives them in the same partition number. It keeps
610 * looping on the 'restart' label until it can no longer find
613 * The union of the cpus_allowed masks from the set of
614 * all cpusets having the same 'pn' value then form the one
615 * element of the partition (one sched domain) to be passed to
616 * partition_sched_domains().
618 static int generate_sched_domains(cpumask_var_t **domains,
619 struct sched_domain_attr **attributes)
621 struct cpuset *cp; /* scans q */
622 struct cpuset **csa; /* array of all cpuset ptrs */
623 int csn; /* how many cpuset ptrs in csa so far */
624 int i, j, k; /* indices for partition finding loops */
625 cpumask_var_t *doms; /* resulting partition; i.e. sched domains */
626 cpumask_var_t non_isolated_cpus; /* load balanced CPUs */
627 struct sched_domain_attr *dattr; /* attributes for custom domains */
628 int ndoms = 0; /* number of sched domains in result */
629 int nslot; /* next empty doms[] struct cpumask slot */
630 struct cgroup_subsys_state *pos_css;
636 if (!alloc_cpumask_var(&non_isolated_cpus, GFP_KERNEL))
638 cpumask_andnot(non_isolated_cpus, cpu_possible_mask, cpu_isolated_map);
640 /* Special case for the 99% of systems with one, full, sched domain */
641 if (is_sched_load_balance(&top_cpuset)) {
643 doms = alloc_sched_domains(ndoms);
647 dattr = kmalloc(sizeof(struct sched_domain_attr), GFP_KERNEL);
649 *dattr = SD_ATTR_INIT;
650 update_domain_attr_tree(dattr, &top_cpuset);
652 cpumask_and(doms[0], top_cpuset.effective_cpus,
658 csa = kmalloc(nr_cpusets() * sizeof(cp), GFP_KERNEL);
664 cpuset_for_each_descendant_pre(cp, pos_css, &top_cpuset) {
665 if (cp == &top_cpuset)
668 * Continue traversing beyond @cp iff @cp has some CPUs and
669 * isn't load balancing. The former is obvious. The
670 * latter: All child cpusets contain a subset of the
671 * parent's cpus, so just skip them, and then we call
672 * update_domain_attr_tree() to calc relax_domain_level of
673 * the corresponding sched domain.
675 if (!cpumask_empty(cp->cpus_allowed) &&
676 !(is_sched_load_balance(cp) &&
677 cpumask_intersects(cp->cpus_allowed, non_isolated_cpus)))
680 if (is_sched_load_balance(cp))
683 /* skip @cp's subtree */
684 pos_css = css_rightmost_descendant(pos_css);
688 for (i = 0; i < csn; i++)
693 /* Find the best partition (set of sched domains) */
694 for (i = 0; i < csn; i++) {
695 struct cpuset *a = csa[i];
698 for (j = 0; j < csn; j++) {
699 struct cpuset *b = csa[j];
702 if (apn != bpn && cpusets_overlap(a, b)) {
703 for (k = 0; k < csn; k++) {
704 struct cpuset *c = csa[k];
709 ndoms--; /* one less element */
716 * Now we know how many domains to create.
717 * Convert <csn, csa> to <ndoms, doms> and populate cpu masks.
719 doms = alloc_sched_domains(ndoms);
724 * The rest of the code, including the scheduler, can deal with
725 * dattr==NULL case. No need to abort if alloc fails.
727 dattr = kmalloc(ndoms * sizeof(struct sched_domain_attr), GFP_KERNEL);
729 for (nslot = 0, i = 0; i < csn; i++) {
730 struct cpuset *a = csa[i];
735 /* Skip completed partitions */
741 if (nslot == ndoms) {
742 static int warnings = 10;
744 pr_warn("rebuild_sched_domains confused: nslot %d, ndoms %d, csn %d, i %d, apn %d\n",
745 nslot, ndoms, csn, i, apn);
753 *(dattr + nslot) = SD_ATTR_INIT;
754 for (j = i; j < csn; j++) {
755 struct cpuset *b = csa[j];
758 cpumask_or(dp, dp, b->effective_cpus);
759 cpumask_and(dp, dp, non_isolated_cpus);
761 update_domain_attr_tree(dattr + nslot, b);
763 /* Done with this partition */
769 BUG_ON(nslot != ndoms);
772 free_cpumask_var(non_isolated_cpus);
776 * Fallback to the default domain if kmalloc() failed.
777 * See comments in partition_sched_domains().
788 * Rebuild scheduler domains.
790 * If the flag 'sched_load_balance' of any cpuset with non-empty
791 * 'cpus' changes, or if the 'cpus' allowed changes in any cpuset
792 * which has that flag enabled, or if any cpuset with a non-empty
793 * 'cpus' is removed, then call this routine to rebuild the
794 * scheduler's dynamic sched domains.
796 * Call with cpuset_mutex held. Takes get_online_cpus().
798 static void rebuild_sched_domains_locked(void)
800 struct sched_domain_attr *attr;
804 lockdep_assert_held(&cpuset_mutex);
808 * We have raced with CPU hotplug. Don't do anything to avoid
809 * passing doms with offlined cpu to partition_sched_domains().
810 * Anyways, hotplug work item will rebuild sched domains.
812 if (!cpumask_equal(top_cpuset.effective_cpus, cpu_active_mask))
815 /* Generate domain masks and attrs */
816 ndoms = generate_sched_domains(&doms, &attr);
818 /* Have scheduler rebuild the domains */
819 partition_sched_domains(ndoms, doms, attr);
823 #else /* !CONFIG_SMP */
824 static void rebuild_sched_domains_locked(void)
827 #endif /* CONFIG_SMP */
829 void rebuild_sched_domains(void)
831 mutex_lock(&cpuset_mutex);
832 rebuild_sched_domains_locked();
833 mutex_unlock(&cpuset_mutex);
837 * update_tasks_cpumask - Update the cpumasks of tasks in the cpuset.
838 * @cs: the cpuset in which each task's cpus_allowed mask needs to be changed
840 * Iterate through each task of @cs updating its cpus_allowed to the
841 * effective cpuset's. As this function is called with cpuset_mutex held,
842 * cpuset membership stays stable.
844 static void update_tasks_cpumask(struct cpuset *cs)
846 struct css_task_iter it;
847 struct task_struct *task;
849 css_task_iter_start(&cs->css, &it);
850 while ((task = css_task_iter_next(&it)))
851 set_cpus_allowed_ptr(task, cs->effective_cpus);
852 css_task_iter_end(&it);
856 * update_cpumasks_hier - Update effective cpumasks and tasks in the subtree
857 * @cs: the cpuset to consider
858 * @new_cpus: temp variable for calculating new effective_cpus
860 * When congifured cpumask is changed, the effective cpumasks of this cpuset
861 * and all its descendants need to be updated.
863 * On legacy hierachy, effective_cpus will be the same with cpu_allowed.
865 * Called with cpuset_mutex held
867 static void update_cpumasks_hier(struct cpuset *cs, struct cpumask *new_cpus)
870 struct cgroup_subsys_state *pos_css;
871 bool need_rebuild_sched_domains = false;
874 cpuset_for_each_descendant_pre(cp, pos_css, cs) {
875 struct cpuset *parent = parent_cs(cp);
877 cpumask_and(new_cpus, cp->cpus_allowed, parent->effective_cpus);
880 * If it becomes empty, inherit the effective mask of the
881 * parent, which is guaranteed to have some CPUs.
883 if (cgroup_on_dfl(cp->css.cgroup) && cpumask_empty(new_cpus))
884 cpumask_copy(new_cpus, parent->effective_cpus);
886 /* Skip the whole subtree if the cpumask remains the same. */
887 if (cpumask_equal(new_cpus, cp->effective_cpus)) {
888 pos_css = css_rightmost_descendant(pos_css);
892 if (!css_tryget_online(&cp->css))
896 spin_lock_irq(&callback_lock);
897 cpumask_copy(cp->effective_cpus, new_cpus);
898 spin_unlock_irq(&callback_lock);
900 WARN_ON(!cgroup_on_dfl(cp->css.cgroup) &&
901 !cpumask_equal(cp->cpus_allowed, cp->effective_cpus));
903 update_tasks_cpumask(cp);
906 * If the effective cpumask of any non-empty cpuset is changed,
907 * we need to rebuild sched domains.
909 if (!cpumask_empty(cp->cpus_allowed) &&
910 is_sched_load_balance(cp))
911 need_rebuild_sched_domains = true;
918 if (need_rebuild_sched_domains)
919 rebuild_sched_domains_locked();
923 * update_cpumask - update the cpus_allowed mask of a cpuset and all tasks in it
924 * @cs: the cpuset to consider
925 * @trialcs: trial cpuset
926 * @buf: buffer of cpu numbers written to this cpuset
928 static int update_cpumask(struct cpuset *cs, struct cpuset *trialcs,
933 /* top_cpuset.cpus_allowed tracks cpu_online_mask; it's read-only */
934 if (cs == &top_cpuset)
938 * An empty cpus_allowed is ok only if the cpuset has no tasks.
939 * Since cpulist_parse() fails on an empty mask, we special case
940 * that parsing. The validate_change() call ensures that cpusets
941 * with tasks have cpus.
944 cpumask_clear(trialcs->cpus_allowed);
946 retval = cpulist_parse(buf, trialcs->cpus_allowed);
950 if (!cpumask_subset(trialcs->cpus_allowed,
951 top_cpuset.cpus_allowed))
955 /* Nothing to do if the cpus didn't change */
956 if (cpumask_equal(cs->cpus_allowed, trialcs->cpus_allowed))
959 retval = validate_change(cs, trialcs);
963 spin_lock_irq(&callback_lock);
964 cpumask_copy(cs->cpus_allowed, trialcs->cpus_allowed);
965 spin_unlock_irq(&callback_lock);
967 /* use trialcs->cpus_allowed as a temp variable */
968 update_cpumasks_hier(cs, trialcs->cpus_allowed);
975 * Migrate memory region from one set of nodes to another.
977 * Temporarilly set tasks mems_allowed to target nodes of migration,
978 * so that the migration code can allocate pages on these nodes.
980 * While the mm_struct we are migrating is typically from some
981 * other task, the task_struct mems_allowed that we are hacking
982 * is for our current task, which must allocate new pages for that
983 * migrating memory region.
986 static void cpuset_migrate_mm(struct mm_struct *mm, const nodemask_t *from,
987 const nodemask_t *to)
989 struct task_struct *tsk = current;
991 tsk->mems_allowed = *to;
993 do_migrate_pages(mm, from, to, MPOL_MF_MOVE_ALL);
996 guarantee_online_mems(task_cs(tsk), &tsk->mems_allowed);
1001 * cpuset_change_task_nodemask - change task's mems_allowed and mempolicy
1002 * @tsk: the task to change
1003 * @newmems: new nodes that the task will be set
1005 * In order to avoid seeing no nodes if the old and new nodes are disjoint,
1006 * we structure updates as setting all new allowed nodes, then clearing newly
1009 static void cpuset_change_task_nodemask(struct task_struct *tsk,
1010 nodemask_t *newmems)
1015 * Allow tasks that have access to memory reserves because they have
1016 * been OOM killed to get memory anywhere.
1018 if (unlikely(test_thread_flag(TIF_MEMDIE)))
1020 if (current->flags & PF_EXITING) /* Let dying task have memory */
1025 * Determine if a loop is necessary if another thread is doing
1026 * read_mems_allowed_begin(). If at least one node remains unchanged and
1027 * tsk does not have a mempolicy, then an empty nodemask will not be
1028 * possible when mems_allowed is larger than a word.
1030 need_loop = task_has_mempolicy(tsk) ||
1031 !nodes_intersects(*newmems, tsk->mems_allowed);
1034 local_irq_disable();
1035 write_seqcount_begin(&tsk->mems_allowed_seq);
1038 nodes_or(tsk->mems_allowed, tsk->mems_allowed, *newmems);
1039 mpol_rebind_task(tsk, newmems, MPOL_REBIND_STEP1);
1041 mpol_rebind_task(tsk, newmems, MPOL_REBIND_STEP2);
1042 tsk->mems_allowed = *newmems;
1045 write_seqcount_end(&tsk->mems_allowed_seq);
1052 static void *cpuset_being_rebound;
1055 * update_tasks_nodemask - Update the nodemasks of tasks in the cpuset.
1056 * @cs: the cpuset in which each task's mems_allowed mask needs to be changed
1058 * Iterate through each task of @cs updating its mems_allowed to the
1059 * effective cpuset's. As this function is called with cpuset_mutex held,
1060 * cpuset membership stays stable.
1062 static void update_tasks_nodemask(struct cpuset *cs)
1064 static nodemask_t newmems; /* protected by cpuset_mutex */
1065 struct css_task_iter it;
1066 struct task_struct *task;
1068 cpuset_being_rebound = cs; /* causes mpol_dup() rebind */
1070 guarantee_online_mems(cs, &newmems);
1073 * The mpol_rebind_mm() call takes mmap_sem, which we couldn't
1074 * take while holding tasklist_lock. Forks can happen - the
1075 * mpol_dup() cpuset_being_rebound check will catch such forks,
1076 * and rebind their vma mempolicies too. Because we still hold
1077 * the global cpuset_mutex, we know that no other rebind effort
1078 * will be contending for the global variable cpuset_being_rebound.
1079 * It's ok if we rebind the same mm twice; mpol_rebind_mm()
1080 * is idempotent. Also migrate pages in each mm to new nodes.
1082 css_task_iter_start(&cs->css, &it);
1083 while ((task = css_task_iter_next(&it))) {
1084 struct mm_struct *mm;
1087 cpuset_change_task_nodemask(task, &newmems);
1089 mm = get_task_mm(task);
1093 migrate = is_memory_migrate(cs);
1095 mpol_rebind_mm(mm, &cs->mems_allowed);
1097 cpuset_migrate_mm(mm, &cs->old_mems_allowed, &newmems);
1100 css_task_iter_end(&it);
1103 * All the tasks' nodemasks have been updated, update
1104 * cs->old_mems_allowed.
1106 cs->old_mems_allowed = newmems;
1108 /* We're done rebinding vmas to this cpuset's new mems_allowed. */
1109 cpuset_being_rebound = NULL;
1113 * update_nodemasks_hier - Update effective nodemasks and tasks in the subtree
1114 * @cs: the cpuset to consider
1115 * @new_mems: a temp variable for calculating new effective_mems
1117 * When configured nodemask is changed, the effective nodemasks of this cpuset
1118 * and all its descendants need to be updated.
1120 * On legacy hiearchy, effective_mems will be the same with mems_allowed.
1122 * Called with cpuset_mutex held
1124 static void update_nodemasks_hier(struct cpuset *cs, nodemask_t *new_mems)
1127 struct cgroup_subsys_state *pos_css;
1130 cpuset_for_each_descendant_pre(cp, pos_css, cs) {
1131 struct cpuset *parent = parent_cs(cp);
1133 nodes_and(*new_mems, cp->mems_allowed, parent->effective_mems);
1136 * If it becomes empty, inherit the effective mask of the
1137 * parent, which is guaranteed to have some MEMs.
1139 if (cgroup_on_dfl(cp->css.cgroup) && nodes_empty(*new_mems))
1140 *new_mems = parent->effective_mems;
1142 /* Skip the whole subtree if the nodemask remains the same. */
1143 if (nodes_equal(*new_mems, cp->effective_mems)) {
1144 pos_css = css_rightmost_descendant(pos_css);
1148 if (!css_tryget_online(&cp->css))
1152 spin_lock_irq(&callback_lock);
1153 cp->effective_mems = *new_mems;
1154 spin_unlock_irq(&callback_lock);
1156 WARN_ON(!cgroup_on_dfl(cp->css.cgroup) &&
1157 !nodes_equal(cp->mems_allowed, cp->effective_mems));
1159 update_tasks_nodemask(cp);
1168 * Handle user request to change the 'mems' memory placement
1169 * of a cpuset. Needs to validate the request, update the
1170 * cpusets mems_allowed, and for each task in the cpuset,
1171 * update mems_allowed and rebind task's mempolicy and any vma
1172 * mempolicies and if the cpuset is marked 'memory_migrate',
1173 * migrate the tasks pages to the new memory.
1175 * Call with cpuset_mutex held. May take callback_lock during call.
1176 * Will take tasklist_lock, scan tasklist for tasks in cpuset cs,
1177 * lock each such tasks mm->mmap_sem, scan its vma's and rebind
1178 * their mempolicies to the cpusets new mems_allowed.
1180 static int update_nodemask(struct cpuset *cs, struct cpuset *trialcs,
1186 * top_cpuset.mems_allowed tracks node_stats[N_MEMORY];
1189 if (cs == &top_cpuset) {
1195 * An empty mems_allowed is ok iff there are no tasks in the cpuset.
1196 * Since nodelist_parse() fails on an empty mask, we special case
1197 * that parsing. The validate_change() call ensures that cpusets
1198 * with tasks have memory.
1201 nodes_clear(trialcs->mems_allowed);
1203 retval = nodelist_parse(buf, trialcs->mems_allowed);
1207 if (!nodes_subset(trialcs->mems_allowed,
1208 top_cpuset.mems_allowed)) {
1214 if (nodes_equal(cs->mems_allowed, trialcs->mems_allowed)) {
1215 retval = 0; /* Too easy - nothing to do */
1218 retval = validate_change(cs, trialcs);
1222 spin_lock_irq(&callback_lock);
1223 cs->mems_allowed = trialcs->mems_allowed;
1224 spin_unlock_irq(&callback_lock);
1226 /* use trialcs->mems_allowed as a temp variable */
1227 update_nodemasks_hier(cs, &cs->mems_allowed);
1232 int current_cpuset_is_being_rebound(void)
1237 ret = task_cs(current) == cpuset_being_rebound;
1243 static int update_relax_domain_level(struct cpuset *cs, s64 val)
1246 if (val < -1 || val >= sched_domain_level_max)
1250 if (val != cs->relax_domain_level) {
1251 cs->relax_domain_level = val;
1252 if (!cpumask_empty(cs->cpus_allowed) &&
1253 is_sched_load_balance(cs))
1254 rebuild_sched_domains_locked();
1261 * update_tasks_flags - update the spread flags of tasks in the cpuset.
1262 * @cs: the cpuset in which each task's spread flags needs to be changed
1264 * Iterate through each task of @cs updating its spread flags. As this
1265 * function is called with cpuset_mutex held, cpuset membership stays
1268 static void update_tasks_flags(struct cpuset *cs)
1270 struct css_task_iter it;
1271 struct task_struct *task;
1273 css_task_iter_start(&cs->css, &it);
1274 while ((task = css_task_iter_next(&it)))
1275 cpuset_update_task_spread_flag(cs, task);
1276 css_task_iter_end(&it);
1280 * update_flag - read a 0 or a 1 in a file and update associated flag
1281 * bit: the bit to update (see cpuset_flagbits_t)
1282 * cs: the cpuset to update
1283 * turning_on: whether the flag is being set or cleared
1285 * Call with cpuset_mutex held.
1288 static int update_flag(cpuset_flagbits_t bit, struct cpuset *cs,
1291 struct cpuset *trialcs;
1292 int balance_flag_changed;
1293 int spread_flag_changed;
1296 trialcs = alloc_trial_cpuset(cs);
1301 set_bit(bit, &trialcs->flags);
1303 clear_bit(bit, &trialcs->flags);
1305 err = validate_change(cs, trialcs);
1309 balance_flag_changed = (is_sched_load_balance(cs) !=
1310 is_sched_load_balance(trialcs));
1312 spread_flag_changed = ((is_spread_slab(cs) != is_spread_slab(trialcs))
1313 || (is_spread_page(cs) != is_spread_page(trialcs)));
1315 spin_lock_irq(&callback_lock);
1316 cs->flags = trialcs->flags;
1317 spin_unlock_irq(&callback_lock);
1319 if (!cpumask_empty(trialcs->cpus_allowed) && balance_flag_changed)
1320 rebuild_sched_domains_locked();
1322 if (spread_flag_changed)
1323 update_tasks_flags(cs);
1325 free_trial_cpuset(trialcs);
1330 * Frequency meter - How fast is some event occurring?
1332 * These routines manage a digitally filtered, constant time based,
1333 * event frequency meter. There are four routines:
1334 * fmeter_init() - initialize a frequency meter.
1335 * fmeter_markevent() - called each time the event happens.
1336 * fmeter_getrate() - returns the recent rate of such events.
1337 * fmeter_update() - internal routine used to update fmeter.
1339 * A common data structure is passed to each of these routines,
1340 * which is used to keep track of the state required to manage the
1341 * frequency meter and its digital filter.
1343 * The filter works on the number of events marked per unit time.
1344 * The filter is single-pole low-pass recursive (IIR). The time unit
1345 * is 1 second. Arithmetic is done using 32-bit integers scaled to
1346 * simulate 3 decimal digits of precision (multiplied by 1000).
1348 * With an FM_COEF of 933, and a time base of 1 second, the filter
1349 * has a half-life of 10 seconds, meaning that if the events quit
1350 * happening, then the rate returned from the fmeter_getrate()
1351 * will be cut in half each 10 seconds, until it converges to zero.
1353 * It is not worth doing a real infinitely recursive filter. If more
1354 * than FM_MAXTICKS ticks have elapsed since the last filter event,
1355 * just compute FM_MAXTICKS ticks worth, by which point the level
1358 * Limit the count of unprocessed events to FM_MAXCNT, so as to avoid
1359 * arithmetic overflow in the fmeter_update() routine.
1361 * Given the simple 32 bit integer arithmetic used, this meter works
1362 * best for reporting rates between one per millisecond (msec) and
1363 * one per 32 (approx) seconds. At constant rates faster than one
1364 * per msec it maxes out at values just under 1,000,000. At constant
1365 * rates between one per msec, and one per second it will stabilize
1366 * to a value N*1000, where N is the rate of events per second.
1367 * At constant rates between one per second and one per 32 seconds,
1368 * it will be choppy, moving up on the seconds that have an event,
1369 * and then decaying until the next event. At rates slower than
1370 * about one in 32 seconds, it decays all the way back to zero between
1374 #define FM_COEF 933 /* coefficient for half-life of 10 secs */
1375 #define FM_MAXTICKS ((u32)99) /* useless computing more ticks than this */
1376 #define FM_MAXCNT 1000000 /* limit cnt to avoid overflow */
1377 #define FM_SCALE 1000 /* faux fixed point scale */
1379 /* Initialize a frequency meter */
1380 static void fmeter_init(struct fmeter *fmp)
1385 spin_lock_init(&fmp->lock);
1388 /* Internal meter update - process cnt events and update value */
1389 static void fmeter_update(struct fmeter *fmp)
1394 now = ktime_get_seconds();
1395 ticks = now - fmp->time;
1400 ticks = min(FM_MAXTICKS, ticks);
1402 fmp->val = (FM_COEF * fmp->val) / FM_SCALE;
1405 fmp->val += ((FM_SCALE - FM_COEF) * fmp->cnt) / FM_SCALE;
1409 /* Process any previous ticks, then bump cnt by one (times scale). */
1410 static void fmeter_markevent(struct fmeter *fmp)
1412 spin_lock(&fmp->lock);
1414 fmp->cnt = min(FM_MAXCNT, fmp->cnt + FM_SCALE);
1415 spin_unlock(&fmp->lock);
1418 /* Process any previous ticks, then return current value. */
1419 static int fmeter_getrate(struct fmeter *fmp)
1423 spin_lock(&fmp->lock);
1426 spin_unlock(&fmp->lock);
1430 static struct cpuset *cpuset_attach_old_cs;
1432 /* Called by cgroups to determine if a cpuset is usable; cpuset_mutex held */
1433 static int cpuset_can_attach(struct cgroup_subsys_state *css,
1434 struct cgroup_taskset *tset)
1436 struct cpuset *cs = css_cs(css);
1437 struct task_struct *task;
1440 /* used later by cpuset_attach() */
1441 cpuset_attach_old_cs = task_cs(cgroup_taskset_first(tset));
1443 mutex_lock(&cpuset_mutex);
1445 /* allow moving tasks into an empty cpuset if on default hierarchy */
1447 if (!cgroup_on_dfl(css->cgroup) &&
1448 (cpumask_empty(cs->cpus_allowed) || nodes_empty(cs->mems_allowed)))
1451 cgroup_taskset_for_each(task, tset) {
1452 ret = task_can_attach(task, cs->cpus_allowed);
1455 ret = security_task_setscheduler(task);
1461 * Mark attach is in progress. This makes validate_change() fail
1462 * changes which zero cpus/mems_allowed.
1464 cs->attach_in_progress++;
1467 mutex_unlock(&cpuset_mutex);
1471 static void cpuset_cancel_attach(struct cgroup_subsys_state *css,
1472 struct cgroup_taskset *tset)
1474 mutex_lock(&cpuset_mutex);
1475 css_cs(css)->attach_in_progress--;
1476 mutex_unlock(&cpuset_mutex);
1480 * Protected by cpuset_mutex. cpus_attach is used only by cpuset_attach()
1481 * but we can't allocate it dynamically there. Define it global and
1482 * allocate from cpuset_init().
1484 static cpumask_var_t cpus_attach;
1486 static void cpuset_attach(struct cgroup_subsys_state *css,
1487 struct cgroup_taskset *tset)
1489 /* static buf protected by cpuset_mutex */
1490 static nodemask_t cpuset_attach_nodemask_to;
1491 struct mm_struct *mm;
1492 struct task_struct *task;
1493 struct task_struct *leader = cgroup_taskset_first(tset);
1494 struct cpuset *cs = css_cs(css);
1495 struct cpuset *oldcs = cpuset_attach_old_cs;
1497 mutex_lock(&cpuset_mutex);
1499 /* prepare for attach */
1500 if (cs == &top_cpuset)
1501 cpumask_copy(cpus_attach, cpu_possible_mask);
1503 guarantee_online_cpus(cs, cpus_attach);
1505 guarantee_online_mems(cs, &cpuset_attach_nodemask_to);
1507 cgroup_taskset_for_each(task, tset) {
1509 * can_attach beforehand should guarantee that this doesn't
1510 * fail. TODO: have a better way to handle failure here
1512 WARN_ON_ONCE(set_cpus_allowed_ptr(task, cpus_attach));
1514 cpuset_change_task_nodemask(task, &cpuset_attach_nodemask_to);
1515 cpuset_update_task_spread_flag(cs, task);
1519 * Change mm, possibly for multiple threads in a threadgroup. This is
1520 * expensive and may sleep.
1522 cpuset_attach_nodemask_to = cs->effective_mems;
1523 mm = get_task_mm(leader);
1525 mpol_rebind_mm(mm, &cpuset_attach_nodemask_to);
1528 * old_mems_allowed is the same with mems_allowed here, except
1529 * if this task is being moved automatically due to hotplug.
1530 * In that case @mems_allowed has been updated and is empty,
1531 * so @old_mems_allowed is the right nodesets that we migrate
1534 if (is_memory_migrate(cs)) {
1535 cpuset_migrate_mm(mm, &oldcs->old_mems_allowed,
1536 &cpuset_attach_nodemask_to);
1541 cs->old_mems_allowed = cpuset_attach_nodemask_to;
1543 cs->attach_in_progress--;
1544 if (!cs->attach_in_progress)
1545 wake_up(&cpuset_attach_wq);
1547 mutex_unlock(&cpuset_mutex);
1550 /* The various types of files and directories in a cpuset file system */
1553 FILE_MEMORY_MIGRATE,
1556 FILE_EFFECTIVE_CPULIST,
1557 FILE_EFFECTIVE_MEMLIST,
1561 FILE_SCHED_LOAD_BALANCE,
1562 FILE_SCHED_RELAX_DOMAIN_LEVEL,
1563 FILE_MEMORY_PRESSURE_ENABLED,
1564 FILE_MEMORY_PRESSURE,
1567 } cpuset_filetype_t;
1569 static int cpuset_write_u64(struct cgroup_subsys_state *css, struct cftype *cft,
1572 struct cpuset *cs = css_cs(css);
1573 cpuset_filetype_t type = cft->private;
1576 mutex_lock(&cpuset_mutex);
1577 if (!is_cpuset_online(cs)) {
1583 case FILE_CPU_EXCLUSIVE:
1584 retval = update_flag(CS_CPU_EXCLUSIVE, cs, val);
1586 case FILE_MEM_EXCLUSIVE:
1587 retval = update_flag(CS_MEM_EXCLUSIVE, cs, val);
1589 case FILE_MEM_HARDWALL:
1590 retval = update_flag(CS_MEM_HARDWALL, cs, val);
1592 case FILE_SCHED_LOAD_BALANCE:
1593 retval = update_flag(CS_SCHED_LOAD_BALANCE, cs, val);
1595 case FILE_MEMORY_MIGRATE:
1596 retval = update_flag(CS_MEMORY_MIGRATE, cs, val);
1598 case FILE_MEMORY_PRESSURE_ENABLED:
1599 cpuset_memory_pressure_enabled = !!val;
1601 case FILE_MEMORY_PRESSURE:
1604 case FILE_SPREAD_PAGE:
1605 retval = update_flag(CS_SPREAD_PAGE, cs, val);
1607 case FILE_SPREAD_SLAB:
1608 retval = update_flag(CS_SPREAD_SLAB, cs, val);
1615 mutex_unlock(&cpuset_mutex);
1619 static int cpuset_write_s64(struct cgroup_subsys_state *css, struct cftype *cft,
1622 struct cpuset *cs = css_cs(css);
1623 cpuset_filetype_t type = cft->private;
1624 int retval = -ENODEV;
1626 mutex_lock(&cpuset_mutex);
1627 if (!is_cpuset_online(cs))
1631 case FILE_SCHED_RELAX_DOMAIN_LEVEL:
1632 retval = update_relax_domain_level(cs, val);
1639 mutex_unlock(&cpuset_mutex);
1644 * Common handling for a write to a "cpus" or "mems" file.
1646 static ssize_t cpuset_write_resmask(struct kernfs_open_file *of,
1647 char *buf, size_t nbytes, loff_t off)
1649 struct cpuset *cs = css_cs(of_css(of));
1650 struct cpuset *trialcs;
1651 int retval = -ENODEV;
1653 buf = strstrip(buf);
1656 * CPU or memory hotunplug may leave @cs w/o any execution
1657 * resources, in which case the hotplug code asynchronously updates
1658 * configuration and transfers all tasks to the nearest ancestor
1659 * which can execute.
1661 * As writes to "cpus" or "mems" may restore @cs's execution
1662 * resources, wait for the previously scheduled operations before
1663 * proceeding, so that we don't end up keep removing tasks added
1664 * after execution capability is restored.
1666 * cpuset_hotplug_work calls back into cgroup core via
1667 * cgroup_transfer_tasks() and waiting for it from a cgroupfs
1668 * operation like this one can lead to a deadlock through kernfs
1669 * active_ref protection. Let's break the protection. Losing the
1670 * protection is okay as we check whether @cs is online after
1671 * grabbing cpuset_mutex anyway. This only happens on the legacy
1675 kernfs_break_active_protection(of->kn);
1676 flush_work(&cpuset_hotplug_work);
1678 mutex_lock(&cpuset_mutex);
1679 if (!is_cpuset_online(cs))
1682 trialcs = alloc_trial_cpuset(cs);
1688 switch (of_cft(of)->private) {
1690 retval = update_cpumask(cs, trialcs, buf);
1693 retval = update_nodemask(cs, trialcs, buf);
1700 free_trial_cpuset(trialcs);
1702 mutex_unlock(&cpuset_mutex);
1703 kernfs_unbreak_active_protection(of->kn);
1705 return retval ?: nbytes;
1709 * These ascii lists should be read in a single call, by using a user
1710 * buffer large enough to hold the entire map. If read in smaller
1711 * chunks, there is no guarantee of atomicity. Since the display format
1712 * used, list of ranges of sequential numbers, is variable length,
1713 * and since these maps can change value dynamically, one could read
1714 * gibberish by doing partial reads while a list was changing.
1716 static int cpuset_common_seq_show(struct seq_file *sf, void *v)
1718 struct cpuset *cs = css_cs(seq_css(sf));
1719 cpuset_filetype_t type = seq_cft(sf)->private;
1722 spin_lock_irq(&callback_lock);
1726 seq_printf(sf, "%*pbl\n", cpumask_pr_args(cs->cpus_allowed));
1729 seq_printf(sf, "%*pbl\n", nodemask_pr_args(&cs->mems_allowed));
1731 case FILE_EFFECTIVE_CPULIST:
1732 seq_printf(sf, "%*pbl\n", cpumask_pr_args(cs->effective_cpus));
1734 case FILE_EFFECTIVE_MEMLIST:
1735 seq_printf(sf, "%*pbl\n", nodemask_pr_args(&cs->effective_mems));
1741 spin_unlock_irq(&callback_lock);
1745 static u64 cpuset_read_u64(struct cgroup_subsys_state *css, struct cftype *cft)
1747 struct cpuset *cs = css_cs(css);
1748 cpuset_filetype_t type = cft->private;
1750 case FILE_CPU_EXCLUSIVE:
1751 return is_cpu_exclusive(cs);
1752 case FILE_MEM_EXCLUSIVE:
1753 return is_mem_exclusive(cs);
1754 case FILE_MEM_HARDWALL:
1755 return is_mem_hardwall(cs);
1756 case FILE_SCHED_LOAD_BALANCE:
1757 return is_sched_load_balance(cs);
1758 case FILE_MEMORY_MIGRATE:
1759 return is_memory_migrate(cs);
1760 case FILE_MEMORY_PRESSURE_ENABLED:
1761 return cpuset_memory_pressure_enabled;
1762 case FILE_MEMORY_PRESSURE:
1763 return fmeter_getrate(&cs->fmeter);
1764 case FILE_SPREAD_PAGE:
1765 return is_spread_page(cs);
1766 case FILE_SPREAD_SLAB:
1767 return is_spread_slab(cs);
1772 /* Unreachable but makes gcc happy */
1776 static s64 cpuset_read_s64(struct cgroup_subsys_state *css, struct cftype *cft)
1778 struct cpuset *cs = css_cs(css);
1779 cpuset_filetype_t type = cft->private;
1781 case FILE_SCHED_RELAX_DOMAIN_LEVEL:
1782 return cs->relax_domain_level;
1787 /* Unrechable but makes gcc happy */
1793 * for the common functions, 'private' gives the type of file
1796 static struct cftype files[] = {
1799 .seq_show = cpuset_common_seq_show,
1800 .write = cpuset_write_resmask,
1801 .max_write_len = (100U + 6 * NR_CPUS),
1802 .private = FILE_CPULIST,
1807 .seq_show = cpuset_common_seq_show,
1808 .write = cpuset_write_resmask,
1809 .max_write_len = (100U + 6 * MAX_NUMNODES),
1810 .private = FILE_MEMLIST,
1814 .name = "effective_cpus",
1815 .seq_show = cpuset_common_seq_show,
1816 .private = FILE_EFFECTIVE_CPULIST,
1820 .name = "effective_mems",
1821 .seq_show = cpuset_common_seq_show,
1822 .private = FILE_EFFECTIVE_MEMLIST,
1826 .name = "cpu_exclusive",
1827 .read_u64 = cpuset_read_u64,
1828 .write_u64 = cpuset_write_u64,
1829 .private = FILE_CPU_EXCLUSIVE,
1833 .name = "mem_exclusive",
1834 .read_u64 = cpuset_read_u64,
1835 .write_u64 = cpuset_write_u64,
1836 .private = FILE_MEM_EXCLUSIVE,
1840 .name = "mem_hardwall",
1841 .read_u64 = cpuset_read_u64,
1842 .write_u64 = cpuset_write_u64,
1843 .private = FILE_MEM_HARDWALL,
1847 .name = "sched_load_balance",
1848 .read_u64 = cpuset_read_u64,
1849 .write_u64 = cpuset_write_u64,
1850 .private = FILE_SCHED_LOAD_BALANCE,
1854 .name = "sched_relax_domain_level",
1855 .read_s64 = cpuset_read_s64,
1856 .write_s64 = cpuset_write_s64,
1857 .private = FILE_SCHED_RELAX_DOMAIN_LEVEL,
1861 .name = "memory_migrate",
1862 .read_u64 = cpuset_read_u64,
1863 .write_u64 = cpuset_write_u64,
1864 .private = FILE_MEMORY_MIGRATE,
1868 .name = "memory_pressure",
1869 .read_u64 = cpuset_read_u64,
1870 .write_u64 = cpuset_write_u64,
1871 .private = FILE_MEMORY_PRESSURE,
1876 .name = "memory_spread_page",
1877 .read_u64 = cpuset_read_u64,
1878 .write_u64 = cpuset_write_u64,
1879 .private = FILE_SPREAD_PAGE,
1883 .name = "memory_spread_slab",
1884 .read_u64 = cpuset_read_u64,
1885 .write_u64 = cpuset_write_u64,
1886 .private = FILE_SPREAD_SLAB,
1890 .name = "memory_pressure_enabled",
1891 .flags = CFTYPE_ONLY_ON_ROOT,
1892 .read_u64 = cpuset_read_u64,
1893 .write_u64 = cpuset_write_u64,
1894 .private = FILE_MEMORY_PRESSURE_ENABLED,
1901 * cpuset_css_alloc - allocate a cpuset css
1902 * cgrp: control group that the new cpuset will be part of
1905 static struct cgroup_subsys_state *
1906 cpuset_css_alloc(struct cgroup_subsys_state *parent_css)
1911 return &top_cpuset.css;
1913 cs = kzalloc(sizeof(*cs), GFP_KERNEL);
1915 return ERR_PTR(-ENOMEM);
1916 if (!alloc_cpumask_var(&cs->cpus_allowed, GFP_KERNEL))
1918 if (!alloc_cpumask_var(&cs->effective_cpus, GFP_KERNEL))
1921 set_bit(CS_SCHED_LOAD_BALANCE, &cs->flags);
1922 cpumask_clear(cs->cpus_allowed);
1923 nodes_clear(cs->mems_allowed);
1924 cpumask_clear(cs->effective_cpus);
1925 nodes_clear(cs->effective_mems);
1926 fmeter_init(&cs->fmeter);
1927 cs->relax_domain_level = -1;
1932 free_cpumask_var(cs->cpus_allowed);
1935 return ERR_PTR(-ENOMEM);
1938 static int cpuset_css_online(struct cgroup_subsys_state *css)
1940 struct cpuset *cs = css_cs(css);
1941 struct cpuset *parent = parent_cs(cs);
1942 struct cpuset *tmp_cs;
1943 struct cgroup_subsys_state *pos_css;
1948 mutex_lock(&cpuset_mutex);
1950 set_bit(CS_ONLINE, &cs->flags);
1951 if (is_spread_page(parent))
1952 set_bit(CS_SPREAD_PAGE, &cs->flags);
1953 if (is_spread_slab(parent))
1954 set_bit(CS_SPREAD_SLAB, &cs->flags);
1958 spin_lock_irq(&callback_lock);
1959 if (cgroup_on_dfl(cs->css.cgroup)) {
1960 cpumask_copy(cs->effective_cpus, parent->effective_cpus);
1961 cs->effective_mems = parent->effective_mems;
1963 spin_unlock_irq(&callback_lock);
1965 if (!test_bit(CGRP_CPUSET_CLONE_CHILDREN, &css->cgroup->flags))
1969 * Clone @parent's configuration if CGRP_CPUSET_CLONE_CHILDREN is
1970 * set. This flag handling is implemented in cgroup core for
1971 * histrical reasons - the flag may be specified during mount.
1973 * Currently, if any sibling cpusets have exclusive cpus or mem, we
1974 * refuse to clone the configuration - thereby refusing the task to
1975 * be entered, and as a result refusing the sys_unshare() or
1976 * clone() which initiated it. If this becomes a problem for some
1977 * users who wish to allow that scenario, then this could be
1978 * changed to grant parent->cpus_allowed-sibling_cpus_exclusive
1979 * (and likewise for mems) to the new cgroup.
1982 cpuset_for_each_child(tmp_cs, pos_css, parent) {
1983 if (is_mem_exclusive(tmp_cs) || is_cpu_exclusive(tmp_cs)) {
1990 spin_lock_irq(&callback_lock);
1991 cs->mems_allowed = parent->mems_allowed;
1992 cs->effective_mems = parent->mems_allowed;
1993 cpumask_copy(cs->cpus_allowed, parent->cpus_allowed);
1994 cpumask_copy(cs->effective_cpus, parent->cpus_allowed);
1995 spin_unlock_irq(&callback_lock);
1997 mutex_unlock(&cpuset_mutex);
2002 * If the cpuset being removed has its flag 'sched_load_balance'
2003 * enabled, then simulate turning sched_load_balance off, which
2004 * will call rebuild_sched_domains_locked().
2007 static void cpuset_css_offline(struct cgroup_subsys_state *css)
2009 struct cpuset *cs = css_cs(css);
2011 mutex_lock(&cpuset_mutex);
2013 if (is_sched_load_balance(cs))
2014 update_flag(CS_SCHED_LOAD_BALANCE, cs, 0);
2017 clear_bit(CS_ONLINE, &cs->flags);
2019 mutex_unlock(&cpuset_mutex);
2022 static void cpuset_css_free(struct cgroup_subsys_state *css)
2024 struct cpuset *cs = css_cs(css);
2026 free_cpumask_var(cs->effective_cpus);
2027 free_cpumask_var(cs->cpus_allowed);
2031 static void cpuset_bind(struct cgroup_subsys_state *root_css)
2033 mutex_lock(&cpuset_mutex);
2034 spin_lock_irq(&callback_lock);
2036 if (cgroup_on_dfl(root_css->cgroup)) {
2037 cpumask_copy(top_cpuset.cpus_allowed, cpu_possible_mask);
2038 top_cpuset.mems_allowed = node_possible_map;
2040 cpumask_copy(top_cpuset.cpus_allowed,
2041 top_cpuset.effective_cpus);
2042 top_cpuset.mems_allowed = top_cpuset.effective_mems;
2045 spin_unlock_irq(&callback_lock);
2046 mutex_unlock(&cpuset_mutex);
2049 struct cgroup_subsys cpuset_cgrp_subsys = {
2050 .css_alloc = cpuset_css_alloc,
2051 .css_online = cpuset_css_online,
2052 .css_offline = cpuset_css_offline,
2053 .css_free = cpuset_css_free,
2054 .can_attach = cpuset_can_attach,
2055 .cancel_attach = cpuset_cancel_attach,
2056 .attach = cpuset_attach,
2057 .bind = cpuset_bind,
2058 .legacy_cftypes = files,
2063 * cpuset_init - initialize cpusets at system boot
2065 * Description: Initialize top_cpuset and the cpuset internal file system,
2068 int __init cpuset_init(void)
2072 if (!alloc_cpumask_var(&top_cpuset.cpus_allowed, GFP_KERNEL))
2074 if (!alloc_cpumask_var(&top_cpuset.effective_cpus, GFP_KERNEL))
2077 cpumask_setall(top_cpuset.cpus_allowed);
2078 nodes_setall(top_cpuset.mems_allowed);
2079 cpumask_setall(top_cpuset.effective_cpus);
2080 nodes_setall(top_cpuset.effective_mems);
2082 fmeter_init(&top_cpuset.fmeter);
2083 set_bit(CS_SCHED_LOAD_BALANCE, &top_cpuset.flags);
2084 top_cpuset.relax_domain_level = -1;
2086 err = register_filesystem(&cpuset_fs_type);
2090 if (!alloc_cpumask_var(&cpus_attach, GFP_KERNEL))
2097 * If CPU and/or memory hotplug handlers, below, unplug any CPUs
2098 * or memory nodes, we need to walk over the cpuset hierarchy,
2099 * removing that CPU or node from all cpusets. If this removes the
2100 * last CPU or node from a cpuset, then move the tasks in the empty
2101 * cpuset to its next-highest non-empty parent.
2103 static void remove_tasks_in_empty_cpuset(struct cpuset *cs)
2105 struct cpuset *parent;
2108 * Find its next-highest non-empty parent, (top cpuset
2109 * has online cpus, so can't be empty).
2111 parent = parent_cs(cs);
2112 while (cpumask_empty(parent->cpus_allowed) ||
2113 nodes_empty(parent->mems_allowed))
2114 parent = parent_cs(parent);
2116 if (cgroup_transfer_tasks(parent->css.cgroup, cs->css.cgroup)) {
2117 pr_err("cpuset: failed to transfer tasks out of empty cpuset ");
2118 pr_cont_cgroup_name(cs->css.cgroup);
2124 hotplug_update_tasks_legacy(struct cpuset *cs,
2125 struct cpumask *new_cpus, nodemask_t *new_mems,
2126 bool cpus_updated, bool mems_updated)
2130 spin_lock_irq(&callback_lock);
2131 cpumask_copy(cs->cpus_allowed, new_cpus);
2132 cpumask_copy(cs->effective_cpus, new_cpus);
2133 cs->mems_allowed = *new_mems;
2134 cs->effective_mems = *new_mems;
2135 spin_unlock_irq(&callback_lock);
2138 * Don't call update_tasks_cpumask() if the cpuset becomes empty,
2139 * as the tasks will be migratecd to an ancestor.
2141 if (cpus_updated && !cpumask_empty(cs->cpus_allowed))
2142 update_tasks_cpumask(cs);
2143 if (mems_updated && !nodes_empty(cs->mems_allowed))
2144 update_tasks_nodemask(cs);
2146 is_empty = cpumask_empty(cs->cpus_allowed) ||
2147 nodes_empty(cs->mems_allowed);
2149 mutex_unlock(&cpuset_mutex);
2152 * Move tasks to the nearest ancestor with execution resources,
2153 * This is full cgroup operation which will also call back into
2154 * cpuset. Should be done outside any lock.
2157 remove_tasks_in_empty_cpuset(cs);
2159 mutex_lock(&cpuset_mutex);
2163 hotplug_update_tasks(struct cpuset *cs,
2164 struct cpumask *new_cpus, nodemask_t *new_mems,
2165 bool cpus_updated, bool mems_updated)
2167 if (cpumask_empty(new_cpus))
2168 cpumask_copy(new_cpus, parent_cs(cs)->effective_cpus);
2169 if (nodes_empty(*new_mems))
2170 *new_mems = parent_cs(cs)->effective_mems;
2172 spin_lock_irq(&callback_lock);
2173 cpumask_copy(cs->effective_cpus, new_cpus);
2174 cs->effective_mems = *new_mems;
2175 spin_unlock_irq(&callback_lock);
2178 update_tasks_cpumask(cs);
2180 update_tasks_nodemask(cs);
2184 * cpuset_hotplug_update_tasks - update tasks in a cpuset for hotunplug
2185 * @cs: cpuset in interest
2187 * Compare @cs's cpu and mem masks against top_cpuset and if some have gone
2188 * offline, update @cs accordingly. If @cs ends up with no CPU or memory,
2189 * all its tasks are moved to the nearest ancestor with both resources.
2191 static void cpuset_hotplug_update_tasks(struct cpuset *cs)
2193 static cpumask_t new_cpus;
2194 static nodemask_t new_mems;
2198 wait_event(cpuset_attach_wq, cs->attach_in_progress == 0);
2200 mutex_lock(&cpuset_mutex);
2203 * We have raced with task attaching. We wait until attaching
2204 * is finished, so we won't attach a task to an empty cpuset.
2206 if (cs->attach_in_progress) {
2207 mutex_unlock(&cpuset_mutex);
2211 cpumask_and(&new_cpus, cs->cpus_allowed, parent_cs(cs)->effective_cpus);
2212 nodes_and(new_mems, cs->mems_allowed, parent_cs(cs)->effective_mems);
2214 cpus_updated = !cpumask_equal(&new_cpus, cs->effective_cpus);
2215 mems_updated = !nodes_equal(new_mems, cs->effective_mems);
2217 if (cgroup_on_dfl(cs->css.cgroup))
2218 hotplug_update_tasks(cs, &new_cpus, &new_mems,
2219 cpus_updated, mems_updated);
2221 hotplug_update_tasks_legacy(cs, &new_cpus, &new_mems,
2222 cpus_updated, mems_updated);
2224 mutex_unlock(&cpuset_mutex);
2228 * cpuset_hotplug_workfn - handle CPU/memory hotunplug for a cpuset
2230 * This function is called after either CPU or memory configuration has
2231 * changed and updates cpuset accordingly. The top_cpuset is always
2232 * synchronized to cpu_active_mask and N_MEMORY, which is necessary in
2233 * order to make cpusets transparent (of no affect) on systems that are
2234 * actively using CPU hotplug but making no active use of cpusets.
2236 * Non-root cpusets are only affected by offlining. If any CPUs or memory
2237 * nodes have been taken down, cpuset_hotplug_update_tasks() is invoked on
2240 * Note that CPU offlining during suspend is ignored. We don't modify
2241 * cpusets across suspend/resume cycles at all.
2243 static void cpuset_hotplug_workfn(struct work_struct *work)
2245 static cpumask_t new_cpus;
2246 static nodemask_t new_mems;
2247 bool cpus_updated, mems_updated;
2248 bool on_dfl = cgroup_on_dfl(top_cpuset.css.cgroup);
2250 mutex_lock(&cpuset_mutex);
2252 /* fetch the available cpus/mems and find out which changed how */
2253 cpumask_copy(&new_cpus, cpu_active_mask);
2254 new_mems = node_states[N_MEMORY];
2256 cpus_updated = !cpumask_equal(top_cpuset.effective_cpus, &new_cpus);
2257 mems_updated = !nodes_equal(top_cpuset.effective_mems, new_mems);
2259 /* synchronize cpus_allowed to cpu_active_mask */
2261 spin_lock_irq(&callback_lock);
2263 cpumask_copy(top_cpuset.cpus_allowed, &new_cpus);
2264 cpumask_copy(top_cpuset.effective_cpus, &new_cpus);
2265 spin_unlock_irq(&callback_lock);
2266 /* we don't mess with cpumasks of tasks in top_cpuset */
2269 /* synchronize mems_allowed to N_MEMORY */
2271 spin_lock_irq(&callback_lock);
2273 top_cpuset.mems_allowed = new_mems;
2274 top_cpuset.effective_mems = new_mems;
2275 spin_unlock_irq(&callback_lock);
2276 update_tasks_nodemask(&top_cpuset);
2279 mutex_unlock(&cpuset_mutex);
2281 /* if cpus or mems changed, we need to propagate to descendants */
2282 if (cpus_updated || mems_updated) {
2284 struct cgroup_subsys_state *pos_css;
2287 cpuset_for_each_descendant_pre(cs, pos_css, &top_cpuset) {
2288 if (cs == &top_cpuset || !css_tryget_online(&cs->css))
2292 cpuset_hotplug_update_tasks(cs);
2300 /* rebuild sched domains if cpus_allowed has changed */
2302 rebuild_sched_domains();
2305 void cpuset_update_active_cpus(bool cpu_online)
2308 * We're inside cpu hotplug critical region which usually nests
2309 * inside cgroup synchronization. Bounce actual hotplug processing
2310 * to a work item to avoid reverse locking order.
2312 * We still need to do partition_sched_domains() synchronously;
2313 * otherwise, the scheduler will get confused and put tasks to the
2314 * dead CPU. Fall back to the default single domain.
2315 * cpuset_hotplug_workfn() will rebuild it as necessary.
2317 partition_sched_domains(1, NULL, NULL);
2318 schedule_work(&cpuset_hotplug_work);
2322 * Keep top_cpuset.mems_allowed tracking node_states[N_MEMORY].
2323 * Call this routine anytime after node_states[N_MEMORY] changes.
2324 * See cpuset_update_active_cpus() for CPU hotplug handling.
2326 static int cpuset_track_online_nodes(struct notifier_block *self,
2327 unsigned long action, void *arg)
2329 schedule_work(&cpuset_hotplug_work);
2333 static struct notifier_block cpuset_track_online_nodes_nb = {
2334 .notifier_call = cpuset_track_online_nodes,
2335 .priority = 10, /* ??! */
2339 * cpuset_init_smp - initialize cpus_allowed
2341 * Description: Finish top cpuset after cpu, node maps are initialized
2343 void __init cpuset_init_smp(void)
2345 cpumask_copy(top_cpuset.cpus_allowed, cpu_active_mask);
2346 top_cpuset.mems_allowed = node_states[N_MEMORY];
2347 top_cpuset.old_mems_allowed = top_cpuset.mems_allowed;
2349 cpumask_copy(top_cpuset.effective_cpus, cpu_active_mask);
2350 top_cpuset.effective_mems = node_states[N_MEMORY];
2352 register_hotmemory_notifier(&cpuset_track_online_nodes_nb);
2356 * cpuset_cpus_allowed - return cpus_allowed mask from a tasks cpuset.
2357 * @tsk: pointer to task_struct from which to obtain cpuset->cpus_allowed.
2358 * @pmask: pointer to struct cpumask variable to receive cpus_allowed set.
2360 * Description: Returns the cpumask_var_t cpus_allowed of the cpuset
2361 * attached to the specified @tsk. Guaranteed to return some non-empty
2362 * subset of cpu_online_mask, even if this means going outside the
2366 void cpuset_cpus_allowed(struct task_struct *tsk, struct cpumask *pmask)
2368 unsigned long flags;
2370 spin_lock_irqsave(&callback_lock, flags);
2372 guarantee_online_cpus(task_cs(tsk), pmask);
2374 spin_unlock_irqrestore(&callback_lock, flags);
2377 void cpuset_cpus_allowed_fallback(struct task_struct *tsk)
2380 do_set_cpus_allowed(tsk, task_cs(tsk)->effective_cpus);
2384 * We own tsk->cpus_allowed, nobody can change it under us.
2386 * But we used cs && cs->cpus_allowed lockless and thus can
2387 * race with cgroup_attach_task() or update_cpumask() and get
2388 * the wrong tsk->cpus_allowed. However, both cases imply the
2389 * subsequent cpuset_change_cpumask()->set_cpus_allowed_ptr()
2390 * which takes task_rq_lock().
2392 * If we are called after it dropped the lock we must see all
2393 * changes in tsk_cs()->cpus_allowed. Otherwise we can temporary
2394 * set any mask even if it is not right from task_cs() pov,
2395 * the pending set_cpus_allowed_ptr() will fix things.
2397 * select_fallback_rq() will fix things ups and set cpu_possible_mask
2402 void __init cpuset_init_current_mems_allowed(void)
2404 nodes_setall(current->mems_allowed);
2408 * cpuset_mems_allowed - return mems_allowed mask from a tasks cpuset.
2409 * @tsk: pointer to task_struct from which to obtain cpuset->mems_allowed.
2411 * Description: Returns the nodemask_t mems_allowed of the cpuset
2412 * attached to the specified @tsk. Guaranteed to return some non-empty
2413 * subset of node_states[N_MEMORY], even if this means going outside the
2417 nodemask_t cpuset_mems_allowed(struct task_struct *tsk)
2420 unsigned long flags;
2422 spin_lock_irqsave(&callback_lock, flags);
2424 guarantee_online_mems(task_cs(tsk), &mask);
2426 spin_unlock_irqrestore(&callback_lock, flags);
2432 * cpuset_nodemask_valid_mems_allowed - check nodemask vs. curremt mems_allowed
2433 * @nodemask: the nodemask to be checked
2435 * Are any of the nodes in the nodemask allowed in current->mems_allowed?
2437 int cpuset_nodemask_valid_mems_allowed(nodemask_t *nodemask)
2439 return nodes_intersects(*nodemask, current->mems_allowed);
2443 * nearest_hardwall_ancestor() - Returns the nearest mem_exclusive or
2444 * mem_hardwall ancestor to the specified cpuset. Call holding
2445 * callback_lock. If no ancestor is mem_exclusive or mem_hardwall
2446 * (an unusual configuration), then returns the root cpuset.
2448 static struct cpuset *nearest_hardwall_ancestor(struct cpuset *cs)
2450 while (!(is_mem_exclusive(cs) || is_mem_hardwall(cs)) && parent_cs(cs))
2456 * cpuset_node_allowed - Can we allocate on a memory node?
2457 * @node: is this an allowed node?
2458 * @gfp_mask: memory allocation flags
2460 * If we're in interrupt, yes, we can always allocate. If @node is set in
2461 * current's mems_allowed, yes. If it's not a __GFP_HARDWALL request and this
2462 * node is set in the nearest hardwalled cpuset ancestor to current's cpuset,
2463 * yes. If current has access to memory reserves due to TIF_MEMDIE, yes.
2466 * GFP_USER allocations are marked with the __GFP_HARDWALL bit,
2467 * and do not allow allocations outside the current tasks cpuset
2468 * unless the task has been OOM killed as is marked TIF_MEMDIE.
2469 * GFP_KERNEL allocations are not so marked, so can escape to the
2470 * nearest enclosing hardwalled ancestor cpuset.
2472 * Scanning up parent cpusets requires callback_lock. The
2473 * __alloc_pages() routine only calls here with __GFP_HARDWALL bit
2474 * _not_ set if it's a GFP_KERNEL allocation, and all nodes in the
2475 * current tasks mems_allowed came up empty on the first pass over
2476 * the zonelist. So only GFP_KERNEL allocations, if all nodes in the
2477 * cpuset are short of memory, might require taking the callback_lock.
2479 * The first call here from mm/page_alloc:get_page_from_freelist()
2480 * has __GFP_HARDWALL set in gfp_mask, enforcing hardwall cpusets,
2481 * so no allocation on a node outside the cpuset is allowed (unless
2482 * in interrupt, of course).
2484 * The second pass through get_page_from_freelist() doesn't even call
2485 * here for GFP_ATOMIC calls. For those calls, the __alloc_pages()
2486 * variable 'wait' is not set, and the bit ALLOC_CPUSET is not set
2487 * in alloc_flags. That logic and the checks below have the combined
2489 * in_interrupt - any node ok (current task context irrelevant)
2490 * GFP_ATOMIC - any node ok
2491 * TIF_MEMDIE - any node ok
2492 * GFP_KERNEL - any node in enclosing hardwalled cpuset ok
2493 * GFP_USER - only nodes in current tasks mems allowed ok.
2495 int __cpuset_node_allowed(int node, gfp_t gfp_mask)
2497 struct cpuset *cs; /* current cpuset ancestors */
2498 int allowed; /* is allocation in zone z allowed? */
2499 unsigned long flags;
2503 if (node_isset(node, current->mems_allowed))
2506 * Allow tasks that have access to memory reserves because they have
2507 * been OOM killed to get memory anywhere.
2509 if (unlikely(test_thread_flag(TIF_MEMDIE)))
2511 if (gfp_mask & __GFP_HARDWALL) /* If hardwall request, stop here */
2514 if (current->flags & PF_EXITING) /* Let dying task have memory */
2517 /* Not hardwall and node outside mems_allowed: scan up cpusets */
2518 spin_lock_irqsave(&callback_lock, flags);
2521 cs = nearest_hardwall_ancestor(task_cs(current));
2522 allowed = node_isset(node, cs->mems_allowed);
2525 spin_unlock_irqrestore(&callback_lock, flags);
2530 * cpuset_mem_spread_node() - On which node to begin search for a file page
2531 * cpuset_slab_spread_node() - On which node to begin search for a slab page
2533 * If a task is marked PF_SPREAD_PAGE or PF_SPREAD_SLAB (as for
2534 * tasks in a cpuset with is_spread_page or is_spread_slab set),
2535 * and if the memory allocation used cpuset_mem_spread_node()
2536 * to determine on which node to start looking, as it will for
2537 * certain page cache or slab cache pages such as used for file
2538 * system buffers and inode caches, then instead of starting on the
2539 * local node to look for a free page, rather spread the starting
2540 * node around the tasks mems_allowed nodes.
2542 * We don't have to worry about the returned node being offline
2543 * because "it can't happen", and even if it did, it would be ok.
2545 * The routines calling guarantee_online_mems() are careful to
2546 * only set nodes in task->mems_allowed that are online. So it
2547 * should not be possible for the following code to return an
2548 * offline node. But if it did, that would be ok, as this routine
2549 * is not returning the node where the allocation must be, only
2550 * the node where the search should start. The zonelist passed to
2551 * __alloc_pages() will include all nodes. If the slab allocator
2552 * is passed an offline node, it will fall back to the local node.
2553 * See kmem_cache_alloc_node().
2556 static int cpuset_spread_node(int *rotor)
2560 node = next_node(*rotor, current->mems_allowed);
2561 if (node == MAX_NUMNODES)
2562 node = first_node(current->mems_allowed);
2567 int cpuset_mem_spread_node(void)
2569 if (current->cpuset_mem_spread_rotor == NUMA_NO_NODE)
2570 current->cpuset_mem_spread_rotor =
2571 node_random(¤t->mems_allowed);
2573 return cpuset_spread_node(¤t->cpuset_mem_spread_rotor);
2576 int cpuset_slab_spread_node(void)
2578 if (current->cpuset_slab_spread_rotor == NUMA_NO_NODE)
2579 current->cpuset_slab_spread_rotor =
2580 node_random(¤t->mems_allowed);
2582 return cpuset_spread_node(¤t->cpuset_slab_spread_rotor);
2585 EXPORT_SYMBOL_GPL(cpuset_mem_spread_node);
2588 * cpuset_mems_allowed_intersects - Does @tsk1's mems_allowed intersect @tsk2's?
2589 * @tsk1: pointer to task_struct of some task.
2590 * @tsk2: pointer to task_struct of some other task.
2592 * Description: Return true if @tsk1's mems_allowed intersects the
2593 * mems_allowed of @tsk2. Used by the OOM killer to determine if
2594 * one of the task's memory usage might impact the memory available
2598 int cpuset_mems_allowed_intersects(const struct task_struct *tsk1,
2599 const struct task_struct *tsk2)
2601 return nodes_intersects(tsk1->mems_allowed, tsk2->mems_allowed);
2605 * cpuset_print_task_mems_allowed - prints task's cpuset and mems_allowed
2606 * @tsk: pointer to task_struct of some task.
2608 * Description: Prints @task's name, cpuset name, and cached copy of its
2609 * mems_allowed to the kernel log.
2611 void cpuset_print_task_mems_allowed(struct task_struct *tsk)
2613 struct cgroup *cgrp;
2617 cgrp = task_cs(tsk)->css.cgroup;
2618 pr_info("%s cpuset=", tsk->comm);
2619 pr_cont_cgroup_name(cgrp);
2620 pr_cont(" mems_allowed=%*pbl\n", nodemask_pr_args(&tsk->mems_allowed));
2626 * Collection of memory_pressure is suppressed unless
2627 * this flag is enabled by writing "1" to the special
2628 * cpuset file 'memory_pressure_enabled' in the root cpuset.
2631 int cpuset_memory_pressure_enabled __read_mostly;
2634 * cpuset_memory_pressure_bump - keep stats of per-cpuset reclaims.
2636 * Keep a running average of the rate of synchronous (direct)
2637 * page reclaim efforts initiated by tasks in each cpuset.
2639 * This represents the rate at which some task in the cpuset
2640 * ran low on memory on all nodes it was allowed to use, and
2641 * had to enter the kernels page reclaim code in an effort to
2642 * create more free memory by tossing clean pages or swapping
2643 * or writing dirty pages.
2645 * Display to user space in the per-cpuset read-only file
2646 * "memory_pressure". Value displayed is an integer
2647 * representing the recent rate of entry into the synchronous
2648 * (direct) page reclaim by any task attached to the cpuset.
2651 void __cpuset_memory_pressure_bump(void)
2654 fmeter_markevent(&task_cs(current)->fmeter);
2658 #ifdef CONFIG_PROC_PID_CPUSET
2660 * proc_cpuset_show()
2661 * - Print tasks cpuset path into seq_file.
2662 * - Used for /proc/<pid>/cpuset.
2663 * - No need to task_lock(tsk) on this tsk->cpuset reference, as it
2664 * doesn't really matter if tsk->cpuset changes after we read it,
2665 * and we take cpuset_mutex, keeping cpuset_attach() from changing it
2668 int proc_cpuset_show(struct seq_file *m, struct pid_namespace *ns,
2669 struct pid *pid, struct task_struct *tsk)
2672 struct cgroup_subsys_state *css;
2676 buf = kmalloc(PATH_MAX, GFP_KERNEL);
2680 retval = -ENAMETOOLONG;
2682 css = task_css(tsk, cpuset_cgrp_id);
2683 p = cgroup_path(css->cgroup, buf, PATH_MAX);
2695 #endif /* CONFIG_PROC_PID_CPUSET */
2697 /* Display task mems_allowed in /proc/<pid>/status file. */
2698 void cpuset_task_status_allowed(struct seq_file *m, struct task_struct *task)
2700 seq_printf(m, "Mems_allowed:\t%*pb\n",
2701 nodemask_pr_args(&task->mems_allowed));
2702 seq_printf(m, "Mems_allowed_list:\t%*pbl\n",
2703 nodemask_pr_args(&task->mems_allowed));