4 * Processor and Memory placement constraints for sets of tasks.
6 * Copyright (C) 2003 BULL SA.
7 * Copyright (C) 2004-2007 Silicon Graphics, Inc.
8 * Copyright (C) 2006 Google, Inc
10 * Portions derived from Patrick Mochel's sysfs code.
11 * sysfs is Copyright (c) 2001-3 Patrick Mochel
13 * 2003-10-10 Written by Simon Derr.
14 * 2003-10-22 Updates by Stephen Hemminger.
15 * 2004 May-July Rework by Paul Jackson.
16 * 2006 Rework by Paul Menage to use generic cgroups
17 * 2008 Rework of the scheduler domains and CPU hotplug handling
20 * This file is subject to the terms and conditions of the GNU General Public
21 * License. See the file COPYING in the main directory of the Linux
22 * distribution for more details.
25 #include <linux/cpu.h>
26 #include <linux/cpumask.h>
27 #include <linux/cpuset.h>
28 #include <linux/err.h>
29 #include <linux/errno.h>
30 #include <linux/file.h>
32 #include <linux/init.h>
33 #include <linux/interrupt.h>
34 #include <linux/kernel.h>
35 #include <linux/kmod.h>
36 #include <linux/list.h>
37 #include <linux/mempolicy.h>
39 #include <linux/memory.h>
40 #include <linux/export.h>
41 #include <linux/mount.h>
42 #include <linux/namei.h>
43 #include <linux/pagemap.h>
44 #include <linux/proc_fs.h>
45 #include <linux/rcupdate.h>
46 #include <linux/sched.h>
47 #include <linux/seq_file.h>
48 #include <linux/security.h>
49 #include <linux/slab.h>
50 #include <linux/spinlock.h>
51 #include <linux/stat.h>
52 #include <linux/string.h>
53 #include <linux/time.h>
54 #include <linux/backing-dev.h>
55 #include <linux/sort.h>
57 #include <asm/uaccess.h>
58 #include <linux/atomic.h>
59 #include <linux/mutex.h>
60 #include <linux/cgroup.h>
61 #include <linux/wait.h>
63 struct static_key cpusets_enabled_key __read_mostly = STATIC_KEY_INIT_FALSE;
65 /* See "Frequency meter" comments, below. */
68 int cnt; /* unprocessed events count */
69 int val; /* most recent output value */
70 time_t time; /* clock (secs) when val computed */
71 spinlock_t lock; /* guards read or write of above */
75 struct cgroup_subsys_state css;
77 unsigned long flags; /* "unsigned long" so bitops work */
80 * On default hierarchy:
82 * The user-configured masks can only be changed by writing to
83 * cpuset.cpus and cpuset.mems, and won't be limited by the
86 * The effective masks is the real masks that apply to the tasks
87 * in the cpuset. They may be changed if the configured masks are
88 * changed or hotplug happens.
90 * effective_mask == configured_mask & parent's effective_mask,
91 * and if it ends up empty, it will inherit the parent's mask.
96 * The user-configured masks are always the same with effective masks.
99 /* user-configured CPUs and Memory Nodes allow to tasks */
100 cpumask_var_t cpus_allowed;
101 nodemask_t mems_allowed;
103 /* effective CPUs and Memory Nodes allow to tasks */
104 cpumask_var_t effective_cpus;
105 nodemask_t effective_mems;
108 * This is old Memory Nodes tasks took on.
110 * - top_cpuset.old_mems_allowed is initialized to mems_allowed.
111 * - A new cpuset's old_mems_allowed is initialized when some
112 * task is moved into it.
113 * - old_mems_allowed is used in cpuset_migrate_mm() when we change
114 * cpuset.mems_allowed and have tasks' nodemask updated, and
115 * then old_mems_allowed is updated to mems_allowed.
117 nodemask_t old_mems_allowed;
119 struct fmeter fmeter; /* memory_pressure filter */
122 * Tasks are being attached to this cpuset. Used to prevent
123 * zeroing cpus/mems_allowed between ->can_attach() and ->attach().
125 int attach_in_progress;
127 /* partition number for rebuild_sched_domains() */
130 /* for custom sched domain */
131 int relax_domain_level;
134 static inline struct cpuset *css_cs(struct cgroup_subsys_state *css)
136 return css ? container_of(css, struct cpuset, css) : NULL;
139 /* Retrieve the cpuset for a task */
140 static inline struct cpuset *task_cs(struct task_struct *task)
142 return css_cs(task_css(task, cpuset_cgrp_id));
145 static inline struct cpuset *parent_cs(struct cpuset *cs)
147 return css_cs(cs->css.parent);
151 static inline bool task_has_mempolicy(struct task_struct *task)
153 return task->mempolicy;
156 static inline bool task_has_mempolicy(struct task_struct *task)
163 /* bits in struct cpuset flags field */
170 CS_SCHED_LOAD_BALANCE,
175 /* convenient tests for these bits */
176 static inline bool is_cpuset_online(const struct cpuset *cs)
178 return test_bit(CS_ONLINE, &cs->flags);
181 static inline int is_cpu_exclusive(const struct cpuset *cs)
183 return test_bit(CS_CPU_EXCLUSIVE, &cs->flags);
186 static inline int is_mem_exclusive(const struct cpuset *cs)
188 return test_bit(CS_MEM_EXCLUSIVE, &cs->flags);
191 static inline int is_mem_hardwall(const struct cpuset *cs)
193 return test_bit(CS_MEM_HARDWALL, &cs->flags);
196 static inline int is_sched_load_balance(const struct cpuset *cs)
198 return test_bit(CS_SCHED_LOAD_BALANCE, &cs->flags);
201 static inline int is_memory_migrate(const struct cpuset *cs)
203 return test_bit(CS_MEMORY_MIGRATE, &cs->flags);
206 static inline int is_spread_page(const struct cpuset *cs)
208 return test_bit(CS_SPREAD_PAGE, &cs->flags);
211 static inline int is_spread_slab(const struct cpuset *cs)
213 return test_bit(CS_SPREAD_SLAB, &cs->flags);
216 static struct cpuset top_cpuset = {
217 .flags = ((1 << CS_ONLINE) | (1 << CS_CPU_EXCLUSIVE) |
218 (1 << CS_MEM_EXCLUSIVE)),
222 * cpuset_for_each_child - traverse online children of a cpuset
223 * @child_cs: loop cursor pointing to the current child
224 * @pos_css: used for iteration
225 * @parent_cs: target cpuset to walk children of
227 * Walk @child_cs through the online children of @parent_cs. Must be used
228 * with RCU read locked.
230 #define cpuset_for_each_child(child_cs, pos_css, parent_cs) \
231 css_for_each_child((pos_css), &(parent_cs)->css) \
232 if (is_cpuset_online(((child_cs) = css_cs((pos_css)))))
235 * cpuset_for_each_descendant_pre - pre-order walk of a cpuset's descendants
236 * @des_cs: loop cursor pointing to the current descendant
237 * @pos_css: used for iteration
238 * @root_cs: target cpuset to walk ancestor of
240 * Walk @des_cs through the online descendants of @root_cs. Must be used
241 * with RCU read locked. The caller may modify @pos_css by calling
242 * css_rightmost_descendant() to skip subtree. @root_cs is included in the
243 * iteration and the first node to be visited.
245 #define cpuset_for_each_descendant_pre(des_cs, pos_css, root_cs) \
246 css_for_each_descendant_pre((pos_css), &(root_cs)->css) \
247 if (is_cpuset_online(((des_cs) = css_cs((pos_css)))))
250 * There are two global locks guarding cpuset structures - cpuset_mutex and
251 * callback_lock. We also require taking task_lock() when dereferencing a
252 * task's cpuset pointer. See "The task_lock() exception", at the end of this
255 * A task must hold both locks to modify cpusets. If a task holds
256 * cpuset_mutex, then it blocks others wanting that mutex, ensuring that it
257 * is the only task able to also acquire callback_lock and be able to
258 * modify cpusets. It can perform various checks on the cpuset structure
259 * first, knowing nothing will change. It can also allocate memory while
260 * just holding cpuset_mutex. While it is performing these checks, various
261 * callback routines can briefly acquire callback_lock to query cpusets.
262 * Once it is ready to make the changes, it takes callback_lock, blocking
265 * Calls to the kernel memory allocator can not be made while holding
266 * callback_lock, as that would risk double tripping on callback_lock
267 * from one of the callbacks into the cpuset code from within
270 * If a task is only holding callback_lock, then it has read-only
273 * Now, the task_struct fields mems_allowed and mempolicy may be changed
274 * by other task, we use alloc_lock in the task_struct fields to protect
277 * The cpuset_common_file_read() handlers only hold callback_lock across
278 * small pieces of code, such as when reading out possibly multi-word
279 * cpumasks and nodemasks.
281 * Accessing a task's cpuset should be done in accordance with the
282 * guidelines for accessing subsystem state in kernel/cgroup.c
285 static DEFINE_MUTEX(cpuset_mutex);
286 static DEFINE_SPINLOCK(callback_lock);
288 static struct workqueue_struct *cpuset_migrate_mm_wq;
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_subsys_on_dfl(cpuset_cgrp_subsys) &&
478 !is_cpuset_subset(trial, par))
482 * If either I or some sibling (!= me) is exclusive, we can't
486 cpuset_for_each_child(c, css, par) {
487 if ((is_cpu_exclusive(trial) || is_cpu_exclusive(c)) &&
489 cpumask_intersects(trial->cpus_allowed, c->cpus_allowed))
491 if ((is_mem_exclusive(trial) || is_mem_exclusive(c)) &&
493 nodes_intersects(trial->mems_allowed, c->mems_allowed))
498 * Cpusets with tasks - existing or newly being attached - can't
499 * be changed to have empty cpus_allowed or mems_allowed.
502 if ((cgroup_is_populated(cur->css.cgroup) || cur->attach_in_progress)) {
503 if (!cpumask_empty(cur->cpus_allowed) &&
504 cpumask_empty(trial->cpus_allowed))
506 if (!nodes_empty(cur->mems_allowed) &&
507 nodes_empty(trial->mems_allowed))
512 * We can't shrink if we won't have enough room for SCHED_DEADLINE
516 if (is_cpu_exclusive(cur) &&
517 !cpuset_cpumask_can_shrink(cur->cpus_allowed,
518 trial->cpus_allowed))
529 * Helper routine for generate_sched_domains().
530 * Do cpusets a, b have overlapping effective cpus_allowed masks?
532 static int cpusets_overlap(struct cpuset *a, struct cpuset *b)
534 return cpumask_intersects(a->effective_cpus, b->effective_cpus);
538 update_domain_attr(struct sched_domain_attr *dattr, struct cpuset *c)
540 if (dattr->relax_domain_level < c->relax_domain_level)
541 dattr->relax_domain_level = c->relax_domain_level;
545 static void update_domain_attr_tree(struct sched_domain_attr *dattr,
546 struct cpuset *root_cs)
549 struct cgroup_subsys_state *pos_css;
552 cpuset_for_each_descendant_pre(cp, pos_css, root_cs) {
553 /* skip the whole subtree if @cp doesn't have any CPU */
554 if (cpumask_empty(cp->cpus_allowed)) {
555 pos_css = css_rightmost_descendant(pos_css);
559 if (is_sched_load_balance(cp))
560 update_domain_attr(dattr, cp);
566 * generate_sched_domains()
568 * This function builds a partial partition of the systems CPUs
569 * A 'partial partition' is a set of non-overlapping subsets whose
570 * union is a subset of that set.
571 * The output of this function needs to be passed to kernel/sched/core.c
572 * partition_sched_domains() routine, which will rebuild the scheduler's
573 * load balancing domains (sched domains) as specified by that partial
576 * See "What is sched_load_balance" in Documentation/cgroups/cpusets.txt
577 * for a background explanation of this.
579 * Does not return errors, on the theory that the callers of this
580 * routine would rather not worry about failures to rebuild sched
581 * domains when operating in the severe memory shortage situations
582 * that could cause allocation failures below.
584 * Must be called with cpuset_mutex held.
586 * The three key local variables below are:
587 * q - a linked-list queue of cpuset pointers, used to implement a
588 * top-down scan of all cpusets. This scan loads a pointer
589 * to each cpuset marked is_sched_load_balance into the
590 * array 'csa'. For our purposes, rebuilding the schedulers
591 * sched domains, we can ignore !is_sched_load_balance cpusets.
592 * csa - (for CpuSet Array) Array of pointers to all the cpusets
593 * that need to be load balanced, for convenient iterative
594 * access by the subsequent code that finds the best partition,
595 * i.e the set of domains (subsets) of CPUs such that the
596 * cpus_allowed of every cpuset marked is_sched_load_balance
597 * is a subset of one of these domains, while there are as
598 * many such domains as possible, each as small as possible.
599 * doms - Conversion of 'csa' to an array of cpumasks, for passing to
600 * the kernel/sched/core.c routine partition_sched_domains() in a
601 * convenient format, that can be easily compared to the prior
602 * value to determine what partition elements (sched domains)
603 * were changed (added or removed.)
605 * Finding the best partition (set of domains):
606 * The triple nested loops below over i, j, k scan over the
607 * load balanced cpusets (using the array of cpuset pointers in
608 * csa[]) looking for pairs of cpusets that have overlapping
609 * cpus_allowed, but which don't have the same 'pn' partition
610 * number and gives them in the same partition number. It keeps
611 * looping on the 'restart' label until it can no longer find
614 * The union of the cpus_allowed masks from the set of
615 * all cpusets having the same 'pn' value then form the one
616 * element of the partition (one sched domain) to be passed to
617 * partition_sched_domains().
619 static int generate_sched_domains(cpumask_var_t **domains,
620 struct sched_domain_attr **attributes)
622 struct cpuset *cp; /* scans q */
623 struct cpuset **csa; /* array of all cpuset ptrs */
624 int csn; /* how many cpuset ptrs in csa so far */
625 int i, j, k; /* indices for partition finding loops */
626 cpumask_var_t *doms; /* resulting partition; i.e. sched domains */
627 cpumask_var_t non_isolated_cpus; /* load balanced CPUs */
628 struct sched_domain_attr *dattr; /* attributes for custom domains */
629 int ndoms = 0; /* number of sched domains in result */
630 int nslot; /* next empty doms[] struct cpumask slot */
631 struct cgroup_subsys_state *pos_css;
637 if (!alloc_cpumask_var(&non_isolated_cpus, GFP_KERNEL))
639 cpumask_andnot(non_isolated_cpus, cpu_possible_mask, cpu_isolated_map);
641 /* Special case for the 99% of systems with one, full, sched domain */
642 if (is_sched_load_balance(&top_cpuset)) {
644 doms = alloc_sched_domains(ndoms);
648 dattr = kmalloc(sizeof(struct sched_domain_attr), GFP_KERNEL);
650 *dattr = SD_ATTR_INIT;
651 update_domain_attr_tree(dattr, &top_cpuset);
653 cpumask_and(doms[0], top_cpuset.effective_cpus,
659 csa = kmalloc(nr_cpusets() * sizeof(cp), GFP_KERNEL);
665 cpuset_for_each_descendant_pre(cp, pos_css, &top_cpuset) {
666 if (cp == &top_cpuset)
669 * Continue traversing beyond @cp iff @cp has some CPUs and
670 * isn't load balancing. The former is obvious. The
671 * latter: All child cpusets contain a subset of the
672 * parent's cpus, so just skip them, and then we call
673 * update_domain_attr_tree() to calc relax_domain_level of
674 * the corresponding sched domain.
676 if (!cpumask_empty(cp->cpus_allowed) &&
677 !(is_sched_load_balance(cp) &&
678 cpumask_intersects(cp->cpus_allowed, non_isolated_cpus)))
681 if (is_sched_load_balance(cp))
684 /* skip @cp's subtree */
685 pos_css = css_rightmost_descendant(pos_css);
689 for (i = 0; i < csn; i++)
694 /* Find the best partition (set of sched domains) */
695 for (i = 0; i < csn; i++) {
696 struct cpuset *a = csa[i];
699 for (j = 0; j < csn; j++) {
700 struct cpuset *b = csa[j];
703 if (apn != bpn && cpusets_overlap(a, b)) {
704 for (k = 0; k < csn; k++) {
705 struct cpuset *c = csa[k];
710 ndoms--; /* one less element */
717 * Now we know how many domains to create.
718 * Convert <csn, csa> to <ndoms, doms> and populate cpu masks.
720 doms = alloc_sched_domains(ndoms);
725 * The rest of the code, including the scheduler, can deal with
726 * dattr==NULL case. No need to abort if alloc fails.
728 dattr = kmalloc(ndoms * sizeof(struct sched_domain_attr), GFP_KERNEL);
730 for (nslot = 0, i = 0; i < csn; i++) {
731 struct cpuset *a = csa[i];
736 /* Skip completed partitions */
742 if (nslot == ndoms) {
743 static int warnings = 10;
745 pr_warn("rebuild_sched_domains confused: nslot %d, ndoms %d, csn %d, i %d, apn %d\n",
746 nslot, ndoms, csn, i, apn);
754 *(dattr + nslot) = SD_ATTR_INIT;
755 for (j = i; j < csn; j++) {
756 struct cpuset *b = csa[j];
759 cpumask_or(dp, dp, b->effective_cpus);
760 cpumask_and(dp, dp, non_isolated_cpus);
762 update_domain_attr_tree(dattr + nslot, b);
764 /* Done with this partition */
770 BUG_ON(nslot != ndoms);
773 free_cpumask_var(non_isolated_cpus);
777 * Fallback to the default domain if kmalloc() failed.
778 * See comments in partition_sched_domains().
789 * Rebuild scheduler domains.
791 * If the flag 'sched_load_balance' of any cpuset with non-empty
792 * 'cpus' changes, or if the 'cpus' allowed changes in any cpuset
793 * which has that flag enabled, or if any cpuset with a non-empty
794 * 'cpus' is removed, then call this routine to rebuild the
795 * scheduler's dynamic sched domains.
797 * Call with cpuset_mutex held. Takes get_online_cpus().
799 static void rebuild_sched_domains_locked(void)
801 struct sched_domain_attr *attr;
805 lockdep_assert_held(&cpuset_mutex);
809 * We have raced with CPU hotplug. Don't do anything to avoid
810 * passing doms with offlined cpu to partition_sched_domains().
811 * Anyways, hotplug work item will rebuild sched domains.
813 if (!cpumask_equal(top_cpuset.effective_cpus, cpu_active_mask))
816 /* Generate domain masks and attrs */
817 ndoms = generate_sched_domains(&doms, &attr);
819 /* Have scheduler rebuild the domains */
820 partition_sched_domains(ndoms, doms, attr);
824 #else /* !CONFIG_SMP */
825 static void rebuild_sched_domains_locked(void)
828 #endif /* CONFIG_SMP */
830 void rebuild_sched_domains(void)
832 mutex_lock(&cpuset_mutex);
833 rebuild_sched_domains_locked();
834 mutex_unlock(&cpuset_mutex);
838 * update_tasks_cpumask - Update the cpumasks of tasks in the cpuset.
839 * @cs: the cpuset in which each task's cpus_allowed mask needs to be changed
841 * Iterate through each task of @cs updating its cpus_allowed to the
842 * effective cpuset's. As this function is called with cpuset_mutex held,
843 * cpuset membership stays stable.
845 static void update_tasks_cpumask(struct cpuset *cs)
847 struct css_task_iter it;
848 struct task_struct *task;
850 css_task_iter_start(&cs->css, &it);
851 while ((task = css_task_iter_next(&it)))
852 set_cpus_allowed_ptr(task, cs->effective_cpus);
853 css_task_iter_end(&it);
857 * update_cpumasks_hier - Update effective cpumasks and tasks in the subtree
858 * @cs: the cpuset to consider
859 * @new_cpus: temp variable for calculating new effective_cpus
861 * When congifured cpumask is changed, the effective cpumasks of this cpuset
862 * and all its descendants need to be updated.
864 * On legacy hierachy, effective_cpus will be the same with cpu_allowed.
866 * Called with cpuset_mutex held
868 static void update_cpumasks_hier(struct cpuset *cs, struct cpumask *new_cpus)
871 struct cgroup_subsys_state *pos_css;
872 bool need_rebuild_sched_domains = false;
875 cpuset_for_each_descendant_pre(cp, pos_css, cs) {
876 struct cpuset *parent = parent_cs(cp);
878 cpumask_and(new_cpus, cp->cpus_allowed, parent->effective_cpus);
881 * If it becomes empty, inherit the effective mask of the
882 * parent, which is guaranteed to have some CPUs.
884 if (cgroup_subsys_on_dfl(cpuset_cgrp_subsys) &&
885 cpumask_empty(new_cpus))
886 cpumask_copy(new_cpus, parent->effective_cpus);
888 /* Skip the whole subtree if the cpumask remains the same. */
889 if (cpumask_equal(new_cpus, cp->effective_cpus)) {
890 pos_css = css_rightmost_descendant(pos_css);
894 if (!css_tryget_online(&cp->css))
898 spin_lock_irq(&callback_lock);
899 cpumask_copy(cp->effective_cpus, new_cpus);
900 spin_unlock_irq(&callback_lock);
902 WARN_ON(!cgroup_subsys_on_dfl(cpuset_cgrp_subsys) &&
903 !cpumask_equal(cp->cpus_allowed, cp->effective_cpus));
905 update_tasks_cpumask(cp);
908 * If the effective cpumask of any non-empty cpuset is changed,
909 * we need to rebuild sched domains.
911 if (!cpumask_empty(cp->cpus_allowed) &&
912 is_sched_load_balance(cp))
913 need_rebuild_sched_domains = true;
920 if (need_rebuild_sched_domains)
921 rebuild_sched_domains_locked();
925 * update_cpumask - update the cpus_allowed mask of a cpuset and all tasks in it
926 * @cs: the cpuset to consider
927 * @trialcs: trial cpuset
928 * @buf: buffer of cpu numbers written to this cpuset
930 static int update_cpumask(struct cpuset *cs, struct cpuset *trialcs,
935 /* top_cpuset.cpus_allowed tracks cpu_online_mask; it's read-only */
936 if (cs == &top_cpuset)
940 * An empty cpus_allowed is ok only if the cpuset has no tasks.
941 * Since cpulist_parse() fails on an empty mask, we special case
942 * that parsing. The validate_change() call ensures that cpusets
943 * with tasks have cpus.
946 cpumask_clear(trialcs->cpus_allowed);
948 retval = cpulist_parse(buf, trialcs->cpus_allowed);
952 if (!cpumask_subset(trialcs->cpus_allowed,
953 top_cpuset.cpus_allowed))
957 /* Nothing to do if the cpus didn't change */
958 if (cpumask_equal(cs->cpus_allowed, trialcs->cpus_allowed))
961 retval = validate_change(cs, trialcs);
965 spin_lock_irq(&callback_lock);
966 cpumask_copy(cs->cpus_allowed, trialcs->cpus_allowed);
967 spin_unlock_irq(&callback_lock);
969 /* use trialcs->cpus_allowed as a temp variable */
970 update_cpumasks_hier(cs, trialcs->cpus_allowed);
975 * Migrate memory region from one set of nodes to another. This is
976 * performed asynchronously as it can be called from process migration path
977 * holding locks involved in process management. All mm migrations are
978 * performed in the queued order and can be waited for by flushing
979 * cpuset_migrate_mm_wq.
982 struct cpuset_migrate_mm_work {
983 struct work_struct work;
984 struct mm_struct *mm;
989 static void cpuset_migrate_mm_workfn(struct work_struct *work)
991 struct cpuset_migrate_mm_work *mwork =
992 container_of(work, struct cpuset_migrate_mm_work, work);
994 /* on a wq worker, no need to worry about %current's mems_allowed */
995 do_migrate_pages(mwork->mm, &mwork->from, &mwork->to, MPOL_MF_MOVE_ALL);
1000 static void cpuset_migrate_mm(struct mm_struct *mm, const nodemask_t *from,
1001 const nodemask_t *to)
1003 struct cpuset_migrate_mm_work *mwork;
1005 mwork = kzalloc(sizeof(*mwork), GFP_KERNEL);
1008 mwork->from = *from;
1010 INIT_WORK(&mwork->work, cpuset_migrate_mm_workfn);
1011 queue_work(cpuset_migrate_mm_wq, &mwork->work);
1017 static void cpuset_post_attach(void)
1019 flush_workqueue(cpuset_migrate_mm_wq);
1023 * cpuset_change_task_nodemask - change task's mems_allowed and mempolicy
1024 * @tsk: the task to change
1025 * @newmems: new nodes that the task will be set
1027 * In order to avoid seeing no nodes if the old and new nodes are disjoint,
1028 * we structure updates as setting all new allowed nodes, then clearing newly
1031 static void cpuset_change_task_nodemask(struct task_struct *tsk,
1032 nodemask_t *newmems)
1037 * Allow tasks that have access to memory reserves because they have
1038 * been OOM killed to get memory anywhere.
1040 if (unlikely(test_thread_flag(TIF_MEMDIE)))
1042 if (current->flags & PF_EXITING) /* Let dying task have memory */
1047 * Determine if a loop is necessary if another thread is doing
1048 * read_mems_allowed_begin(). If at least one node remains unchanged and
1049 * tsk does not have a mempolicy, then an empty nodemask will not be
1050 * possible when mems_allowed is larger than a word.
1052 need_loop = task_has_mempolicy(tsk) ||
1053 !nodes_intersects(*newmems, tsk->mems_allowed);
1056 local_irq_disable();
1057 write_seqcount_begin(&tsk->mems_allowed_seq);
1060 nodes_or(tsk->mems_allowed, tsk->mems_allowed, *newmems);
1061 mpol_rebind_task(tsk, newmems, MPOL_REBIND_STEP1);
1063 mpol_rebind_task(tsk, newmems, MPOL_REBIND_STEP2);
1064 tsk->mems_allowed = *newmems;
1067 write_seqcount_end(&tsk->mems_allowed_seq);
1074 static void *cpuset_being_rebound;
1077 * update_tasks_nodemask - Update the nodemasks of tasks in the cpuset.
1078 * @cs: the cpuset in which each task's mems_allowed mask needs to be changed
1080 * Iterate through each task of @cs updating its mems_allowed to the
1081 * effective cpuset's. As this function is called with cpuset_mutex held,
1082 * cpuset membership stays stable.
1084 static void update_tasks_nodemask(struct cpuset *cs)
1086 static nodemask_t newmems; /* protected by cpuset_mutex */
1087 struct css_task_iter it;
1088 struct task_struct *task;
1090 cpuset_being_rebound = cs; /* causes mpol_dup() rebind */
1092 guarantee_online_mems(cs, &newmems);
1095 * The mpol_rebind_mm() call takes mmap_sem, which we couldn't
1096 * take while holding tasklist_lock. Forks can happen - the
1097 * mpol_dup() cpuset_being_rebound check will catch such forks,
1098 * and rebind their vma mempolicies too. Because we still hold
1099 * the global cpuset_mutex, we know that no other rebind effort
1100 * will be contending for the global variable cpuset_being_rebound.
1101 * It's ok if we rebind the same mm twice; mpol_rebind_mm()
1102 * is idempotent. Also migrate pages in each mm to new nodes.
1104 css_task_iter_start(&cs->css, &it);
1105 while ((task = css_task_iter_next(&it))) {
1106 struct mm_struct *mm;
1109 cpuset_change_task_nodemask(task, &newmems);
1111 mm = get_task_mm(task);
1115 migrate = is_memory_migrate(cs);
1117 mpol_rebind_mm(mm, &cs->mems_allowed);
1119 cpuset_migrate_mm(mm, &cs->old_mems_allowed, &newmems);
1123 css_task_iter_end(&it);
1126 * All the tasks' nodemasks have been updated, update
1127 * cs->old_mems_allowed.
1129 cs->old_mems_allowed = newmems;
1131 /* We're done rebinding vmas to this cpuset's new mems_allowed. */
1132 cpuset_being_rebound = NULL;
1136 * update_nodemasks_hier - Update effective nodemasks and tasks in the subtree
1137 * @cs: the cpuset to consider
1138 * @new_mems: a temp variable for calculating new effective_mems
1140 * When configured nodemask is changed, the effective nodemasks of this cpuset
1141 * and all its descendants need to be updated.
1143 * On legacy hiearchy, effective_mems will be the same with mems_allowed.
1145 * Called with cpuset_mutex held
1147 static void update_nodemasks_hier(struct cpuset *cs, nodemask_t *new_mems)
1150 struct cgroup_subsys_state *pos_css;
1153 cpuset_for_each_descendant_pre(cp, pos_css, cs) {
1154 struct cpuset *parent = parent_cs(cp);
1156 nodes_and(*new_mems, cp->mems_allowed, parent->effective_mems);
1159 * If it becomes empty, inherit the effective mask of the
1160 * parent, which is guaranteed to have some MEMs.
1162 if (cgroup_subsys_on_dfl(cpuset_cgrp_subsys) &&
1163 nodes_empty(*new_mems))
1164 *new_mems = parent->effective_mems;
1166 /* Skip the whole subtree if the nodemask remains the same. */
1167 if (nodes_equal(*new_mems, cp->effective_mems)) {
1168 pos_css = css_rightmost_descendant(pos_css);
1172 if (!css_tryget_online(&cp->css))
1176 spin_lock_irq(&callback_lock);
1177 cp->effective_mems = *new_mems;
1178 spin_unlock_irq(&callback_lock);
1180 WARN_ON(!cgroup_subsys_on_dfl(cpuset_cgrp_subsys) &&
1181 !nodes_equal(cp->mems_allowed, cp->effective_mems));
1183 update_tasks_nodemask(cp);
1192 * Handle user request to change the 'mems' memory placement
1193 * of a cpuset. Needs to validate the request, update the
1194 * cpusets mems_allowed, and for each task in the cpuset,
1195 * update mems_allowed and rebind task's mempolicy and any vma
1196 * mempolicies and if the cpuset is marked 'memory_migrate',
1197 * migrate the tasks pages to the new memory.
1199 * Call with cpuset_mutex held. May take callback_lock during call.
1200 * Will take tasklist_lock, scan tasklist for tasks in cpuset cs,
1201 * lock each such tasks mm->mmap_sem, scan its vma's and rebind
1202 * their mempolicies to the cpusets new mems_allowed.
1204 static int update_nodemask(struct cpuset *cs, struct cpuset *trialcs,
1210 * top_cpuset.mems_allowed tracks node_stats[N_MEMORY];
1213 if (cs == &top_cpuset) {
1219 * An empty mems_allowed is ok iff there are no tasks in the cpuset.
1220 * Since nodelist_parse() fails on an empty mask, we special case
1221 * that parsing. The validate_change() call ensures that cpusets
1222 * with tasks have memory.
1225 nodes_clear(trialcs->mems_allowed);
1227 retval = nodelist_parse(buf, trialcs->mems_allowed);
1231 if (!nodes_subset(trialcs->mems_allowed,
1232 top_cpuset.mems_allowed)) {
1238 if (nodes_equal(cs->mems_allowed, trialcs->mems_allowed)) {
1239 retval = 0; /* Too easy - nothing to do */
1242 retval = validate_change(cs, trialcs);
1246 spin_lock_irq(&callback_lock);
1247 cs->mems_allowed = trialcs->mems_allowed;
1248 spin_unlock_irq(&callback_lock);
1250 /* use trialcs->mems_allowed as a temp variable */
1251 update_nodemasks_hier(cs, &trialcs->mems_allowed);
1256 int current_cpuset_is_being_rebound(void)
1261 ret = task_cs(current) == cpuset_being_rebound;
1267 static int update_relax_domain_level(struct cpuset *cs, s64 val)
1270 if (val < -1 || val >= sched_domain_level_max)
1274 if (val != cs->relax_domain_level) {
1275 cs->relax_domain_level = val;
1276 if (!cpumask_empty(cs->cpus_allowed) &&
1277 is_sched_load_balance(cs))
1278 rebuild_sched_domains_locked();
1285 * update_tasks_flags - update the spread flags of tasks in the cpuset.
1286 * @cs: the cpuset in which each task's spread flags needs to be changed
1288 * Iterate through each task of @cs updating its spread flags. As this
1289 * function is called with cpuset_mutex held, cpuset membership stays
1292 static void update_tasks_flags(struct cpuset *cs)
1294 struct css_task_iter it;
1295 struct task_struct *task;
1297 css_task_iter_start(&cs->css, &it);
1298 while ((task = css_task_iter_next(&it)))
1299 cpuset_update_task_spread_flag(cs, task);
1300 css_task_iter_end(&it);
1304 * update_flag - read a 0 or a 1 in a file and update associated flag
1305 * bit: the bit to update (see cpuset_flagbits_t)
1306 * cs: the cpuset to update
1307 * turning_on: whether the flag is being set or cleared
1309 * Call with cpuset_mutex held.
1312 static int update_flag(cpuset_flagbits_t bit, struct cpuset *cs,
1315 struct cpuset *trialcs;
1316 int balance_flag_changed;
1317 int spread_flag_changed;
1320 trialcs = alloc_trial_cpuset(cs);
1325 set_bit(bit, &trialcs->flags);
1327 clear_bit(bit, &trialcs->flags);
1329 err = validate_change(cs, trialcs);
1333 balance_flag_changed = (is_sched_load_balance(cs) !=
1334 is_sched_load_balance(trialcs));
1336 spread_flag_changed = ((is_spread_slab(cs) != is_spread_slab(trialcs))
1337 || (is_spread_page(cs) != is_spread_page(trialcs)));
1339 spin_lock_irq(&callback_lock);
1340 cs->flags = trialcs->flags;
1341 spin_unlock_irq(&callback_lock);
1343 if (!cpumask_empty(trialcs->cpus_allowed) && balance_flag_changed)
1344 rebuild_sched_domains_locked();
1346 if (spread_flag_changed)
1347 update_tasks_flags(cs);
1349 free_trial_cpuset(trialcs);
1354 * Frequency meter - How fast is some event occurring?
1356 * These routines manage a digitally filtered, constant time based,
1357 * event frequency meter. There are four routines:
1358 * fmeter_init() - initialize a frequency meter.
1359 * fmeter_markevent() - called each time the event happens.
1360 * fmeter_getrate() - returns the recent rate of such events.
1361 * fmeter_update() - internal routine used to update fmeter.
1363 * A common data structure is passed to each of these routines,
1364 * which is used to keep track of the state required to manage the
1365 * frequency meter and its digital filter.
1367 * The filter works on the number of events marked per unit time.
1368 * The filter is single-pole low-pass recursive (IIR). The time unit
1369 * is 1 second. Arithmetic is done using 32-bit integers scaled to
1370 * simulate 3 decimal digits of precision (multiplied by 1000).
1372 * With an FM_COEF of 933, and a time base of 1 second, the filter
1373 * has a half-life of 10 seconds, meaning that if the events quit
1374 * happening, then the rate returned from the fmeter_getrate()
1375 * will be cut in half each 10 seconds, until it converges to zero.
1377 * It is not worth doing a real infinitely recursive filter. If more
1378 * than FM_MAXTICKS ticks have elapsed since the last filter event,
1379 * just compute FM_MAXTICKS ticks worth, by which point the level
1382 * Limit the count of unprocessed events to FM_MAXCNT, so as to avoid
1383 * arithmetic overflow in the fmeter_update() routine.
1385 * Given the simple 32 bit integer arithmetic used, this meter works
1386 * best for reporting rates between one per millisecond (msec) and
1387 * one per 32 (approx) seconds. At constant rates faster than one
1388 * per msec it maxes out at values just under 1,000,000. At constant
1389 * rates between one per msec, and one per second it will stabilize
1390 * to a value N*1000, where N is the rate of events per second.
1391 * At constant rates between one per second and one per 32 seconds,
1392 * it will be choppy, moving up on the seconds that have an event,
1393 * and then decaying until the next event. At rates slower than
1394 * about one in 32 seconds, it decays all the way back to zero between
1398 #define FM_COEF 933 /* coefficient for half-life of 10 secs */
1399 #define FM_MAXTICKS ((time_t)99) /* useless computing more ticks than this */
1400 #define FM_MAXCNT 1000000 /* limit cnt to avoid overflow */
1401 #define FM_SCALE 1000 /* faux fixed point scale */
1403 /* Initialize a frequency meter */
1404 static void fmeter_init(struct fmeter *fmp)
1409 spin_lock_init(&fmp->lock);
1412 /* Internal meter update - process cnt events and update value */
1413 static void fmeter_update(struct fmeter *fmp)
1415 time_t now = get_seconds();
1416 time_t ticks = now - fmp->time;
1421 ticks = min(FM_MAXTICKS, ticks);
1423 fmp->val = (FM_COEF * fmp->val) / FM_SCALE;
1426 fmp->val += ((FM_SCALE - FM_COEF) * fmp->cnt) / FM_SCALE;
1430 /* Process any previous ticks, then bump cnt by one (times scale). */
1431 static void fmeter_markevent(struct fmeter *fmp)
1433 spin_lock(&fmp->lock);
1435 fmp->cnt = min(FM_MAXCNT, fmp->cnt + FM_SCALE);
1436 spin_unlock(&fmp->lock);
1439 /* Process any previous ticks, then return current value. */
1440 static int fmeter_getrate(struct fmeter *fmp)
1444 spin_lock(&fmp->lock);
1447 spin_unlock(&fmp->lock);
1451 static struct cpuset *cpuset_attach_old_cs;
1453 /* Called by cgroups to determine if a cpuset is usable; cpuset_mutex held */
1454 static int cpuset_can_attach(struct cgroup_taskset *tset)
1456 struct cgroup_subsys_state *css;
1458 struct task_struct *task;
1461 /* used later by cpuset_attach() */
1462 cpuset_attach_old_cs = task_cs(cgroup_taskset_first(tset, &css));
1465 mutex_lock(&cpuset_mutex);
1467 /* allow moving tasks into an empty cpuset if on default hierarchy */
1469 if (!cgroup_subsys_on_dfl(cpuset_cgrp_subsys) &&
1470 (cpumask_empty(cs->cpus_allowed) || nodes_empty(cs->mems_allowed)))
1473 cgroup_taskset_for_each(task, css, tset) {
1474 ret = task_can_attach(task, cs->cpus_allowed);
1477 ret = security_task_setscheduler(task);
1483 * Mark attach is in progress. This makes validate_change() fail
1484 * changes which zero cpus/mems_allowed.
1486 cs->attach_in_progress++;
1489 mutex_unlock(&cpuset_mutex);
1493 static void cpuset_cancel_attach(struct cgroup_taskset *tset)
1495 struct cgroup_subsys_state *css;
1498 cgroup_taskset_first(tset, &css);
1501 mutex_lock(&cpuset_mutex);
1502 css_cs(css)->attach_in_progress--;
1503 mutex_unlock(&cpuset_mutex);
1507 * Protected by cpuset_mutex. cpus_attach is used only by cpuset_attach()
1508 * but we can't allocate it dynamically there. Define it global and
1509 * allocate from cpuset_init().
1511 static cpumask_var_t cpus_attach;
1513 static void cpuset_attach(struct cgroup_taskset *tset)
1515 /* static buf protected by cpuset_mutex */
1516 static nodemask_t cpuset_attach_nodemask_to;
1517 struct task_struct *task;
1518 struct task_struct *leader;
1519 struct cgroup_subsys_state *css;
1521 struct cpuset *oldcs = cpuset_attach_old_cs;
1523 cgroup_taskset_first(tset, &css);
1526 mutex_lock(&cpuset_mutex);
1528 /* prepare for attach */
1529 if (cs == &top_cpuset)
1530 cpumask_copy(cpus_attach, cpu_possible_mask);
1532 guarantee_online_cpus(cs, cpus_attach);
1534 guarantee_online_mems(cs, &cpuset_attach_nodemask_to);
1536 cgroup_taskset_for_each(task, css, tset) {
1538 * can_attach beforehand should guarantee that this doesn't
1539 * fail. TODO: have a better way to handle failure here
1541 WARN_ON_ONCE(set_cpus_allowed_ptr(task, cpus_attach));
1543 cpuset_change_task_nodemask(task, &cpuset_attach_nodemask_to);
1544 cpuset_update_task_spread_flag(cs, task);
1548 * Change mm for all threadgroup leaders. This is expensive and may
1549 * sleep and should be moved outside migration path proper.
1551 cpuset_attach_nodemask_to = cs->effective_mems;
1552 cgroup_taskset_for_each_leader(leader, css, tset) {
1553 struct mm_struct *mm = get_task_mm(leader);
1556 mpol_rebind_mm(mm, &cpuset_attach_nodemask_to);
1559 * old_mems_allowed is the same with mems_allowed
1560 * here, except if this task is being moved
1561 * automatically due to hotplug. In that case
1562 * @mems_allowed has been updated and is empty, so
1563 * @old_mems_allowed is the right nodesets that we
1566 if (is_memory_migrate(cs))
1567 cpuset_migrate_mm(mm, &oldcs->old_mems_allowed,
1568 &cpuset_attach_nodemask_to);
1574 cs->old_mems_allowed = cpuset_attach_nodemask_to;
1576 cs->attach_in_progress--;
1577 if (!cs->attach_in_progress)
1578 wake_up(&cpuset_attach_wq);
1580 mutex_unlock(&cpuset_mutex);
1583 /* The various types of files and directories in a cpuset file system */
1586 FILE_MEMORY_MIGRATE,
1589 FILE_EFFECTIVE_CPULIST,
1590 FILE_EFFECTIVE_MEMLIST,
1594 FILE_SCHED_LOAD_BALANCE,
1595 FILE_SCHED_RELAX_DOMAIN_LEVEL,
1596 FILE_MEMORY_PRESSURE_ENABLED,
1597 FILE_MEMORY_PRESSURE,
1600 } cpuset_filetype_t;
1602 static int cpuset_write_u64(struct cgroup_subsys_state *css, struct cftype *cft,
1605 struct cpuset *cs = css_cs(css);
1606 cpuset_filetype_t type = cft->private;
1609 mutex_lock(&cpuset_mutex);
1610 if (!is_cpuset_online(cs)) {
1616 case FILE_CPU_EXCLUSIVE:
1617 retval = update_flag(CS_CPU_EXCLUSIVE, cs, val);
1619 case FILE_MEM_EXCLUSIVE:
1620 retval = update_flag(CS_MEM_EXCLUSIVE, cs, val);
1622 case FILE_MEM_HARDWALL:
1623 retval = update_flag(CS_MEM_HARDWALL, cs, val);
1625 case FILE_SCHED_LOAD_BALANCE:
1626 retval = update_flag(CS_SCHED_LOAD_BALANCE, cs, val);
1628 case FILE_MEMORY_MIGRATE:
1629 retval = update_flag(CS_MEMORY_MIGRATE, cs, val);
1631 case FILE_MEMORY_PRESSURE_ENABLED:
1632 cpuset_memory_pressure_enabled = !!val;
1634 case FILE_SPREAD_PAGE:
1635 retval = update_flag(CS_SPREAD_PAGE, cs, val);
1637 case FILE_SPREAD_SLAB:
1638 retval = update_flag(CS_SPREAD_SLAB, cs, val);
1645 mutex_unlock(&cpuset_mutex);
1649 static int cpuset_write_s64(struct cgroup_subsys_state *css, struct cftype *cft,
1652 struct cpuset *cs = css_cs(css);
1653 cpuset_filetype_t type = cft->private;
1654 int retval = -ENODEV;
1656 mutex_lock(&cpuset_mutex);
1657 if (!is_cpuset_online(cs))
1661 case FILE_SCHED_RELAX_DOMAIN_LEVEL:
1662 retval = update_relax_domain_level(cs, val);
1669 mutex_unlock(&cpuset_mutex);
1674 * Common handling for a write to a "cpus" or "mems" file.
1676 static ssize_t cpuset_write_resmask(struct kernfs_open_file *of,
1677 char *buf, size_t nbytes, loff_t off)
1679 struct cpuset *cs = css_cs(of_css(of));
1680 struct cpuset *trialcs;
1681 int retval = -ENODEV;
1683 buf = strstrip(buf);
1686 * CPU or memory hotunplug may leave @cs w/o any execution
1687 * resources, in which case the hotplug code asynchronously updates
1688 * configuration and transfers all tasks to the nearest ancestor
1689 * which can execute.
1691 * As writes to "cpus" or "mems" may restore @cs's execution
1692 * resources, wait for the previously scheduled operations before
1693 * proceeding, so that we don't end up keep removing tasks added
1694 * after execution capability is restored.
1696 * cpuset_hotplug_work calls back into cgroup core via
1697 * cgroup_transfer_tasks() and waiting for it from a cgroupfs
1698 * operation like this one can lead to a deadlock through kernfs
1699 * active_ref protection. Let's break the protection. Losing the
1700 * protection is okay as we check whether @cs is online after
1701 * grabbing cpuset_mutex anyway. This only happens on the legacy
1705 kernfs_break_active_protection(of->kn);
1706 flush_work(&cpuset_hotplug_work);
1708 mutex_lock(&cpuset_mutex);
1709 if (!is_cpuset_online(cs))
1712 trialcs = alloc_trial_cpuset(cs);
1718 switch (of_cft(of)->private) {
1720 retval = update_cpumask(cs, trialcs, buf);
1723 retval = update_nodemask(cs, trialcs, buf);
1730 free_trial_cpuset(trialcs);
1732 mutex_unlock(&cpuset_mutex);
1733 kernfs_unbreak_active_protection(of->kn);
1735 flush_workqueue(cpuset_migrate_mm_wq);
1736 return retval ?: nbytes;
1740 * These ascii lists should be read in a single call, by using a user
1741 * buffer large enough to hold the entire map. If read in smaller
1742 * chunks, there is no guarantee of atomicity. Since the display format
1743 * used, list of ranges of sequential numbers, is variable length,
1744 * and since these maps can change value dynamically, one could read
1745 * gibberish by doing partial reads while a list was changing.
1747 static int cpuset_common_seq_show(struct seq_file *sf, void *v)
1749 struct cpuset *cs = css_cs(seq_css(sf));
1750 cpuset_filetype_t type = seq_cft(sf)->private;
1753 spin_lock_irq(&callback_lock);
1757 seq_printf(sf, "%*pbl\n", cpumask_pr_args(cs->cpus_allowed));
1760 seq_printf(sf, "%*pbl\n", nodemask_pr_args(&cs->mems_allowed));
1762 case FILE_EFFECTIVE_CPULIST:
1763 seq_printf(sf, "%*pbl\n", cpumask_pr_args(cs->effective_cpus));
1765 case FILE_EFFECTIVE_MEMLIST:
1766 seq_printf(sf, "%*pbl\n", nodemask_pr_args(&cs->effective_mems));
1772 spin_unlock_irq(&callback_lock);
1776 static u64 cpuset_read_u64(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_CPU_EXCLUSIVE:
1782 return is_cpu_exclusive(cs);
1783 case FILE_MEM_EXCLUSIVE:
1784 return is_mem_exclusive(cs);
1785 case FILE_MEM_HARDWALL:
1786 return is_mem_hardwall(cs);
1787 case FILE_SCHED_LOAD_BALANCE:
1788 return is_sched_load_balance(cs);
1789 case FILE_MEMORY_MIGRATE:
1790 return is_memory_migrate(cs);
1791 case FILE_MEMORY_PRESSURE_ENABLED:
1792 return cpuset_memory_pressure_enabled;
1793 case FILE_MEMORY_PRESSURE:
1794 return fmeter_getrate(&cs->fmeter);
1795 case FILE_SPREAD_PAGE:
1796 return is_spread_page(cs);
1797 case FILE_SPREAD_SLAB:
1798 return is_spread_slab(cs);
1803 /* Unreachable but makes gcc happy */
1807 static s64 cpuset_read_s64(struct cgroup_subsys_state *css, struct cftype *cft)
1809 struct cpuset *cs = css_cs(css);
1810 cpuset_filetype_t type = cft->private;
1812 case FILE_SCHED_RELAX_DOMAIN_LEVEL:
1813 return cs->relax_domain_level;
1818 /* Unrechable but makes gcc happy */
1824 * for the common functions, 'private' gives the type of file
1827 static struct cftype files[] = {
1830 .seq_show = cpuset_common_seq_show,
1831 .write = cpuset_write_resmask,
1832 .max_write_len = (100U + 6 * NR_CPUS),
1833 .private = FILE_CPULIST,
1838 .seq_show = cpuset_common_seq_show,
1839 .write = cpuset_write_resmask,
1840 .max_write_len = (100U + 6 * MAX_NUMNODES),
1841 .private = FILE_MEMLIST,
1845 .name = "effective_cpus",
1846 .seq_show = cpuset_common_seq_show,
1847 .private = FILE_EFFECTIVE_CPULIST,
1851 .name = "effective_mems",
1852 .seq_show = cpuset_common_seq_show,
1853 .private = FILE_EFFECTIVE_MEMLIST,
1857 .name = "cpu_exclusive",
1858 .read_u64 = cpuset_read_u64,
1859 .write_u64 = cpuset_write_u64,
1860 .private = FILE_CPU_EXCLUSIVE,
1864 .name = "mem_exclusive",
1865 .read_u64 = cpuset_read_u64,
1866 .write_u64 = cpuset_write_u64,
1867 .private = FILE_MEM_EXCLUSIVE,
1871 .name = "mem_hardwall",
1872 .read_u64 = cpuset_read_u64,
1873 .write_u64 = cpuset_write_u64,
1874 .private = FILE_MEM_HARDWALL,
1878 .name = "sched_load_balance",
1879 .read_u64 = cpuset_read_u64,
1880 .write_u64 = cpuset_write_u64,
1881 .private = FILE_SCHED_LOAD_BALANCE,
1885 .name = "sched_relax_domain_level",
1886 .read_s64 = cpuset_read_s64,
1887 .write_s64 = cpuset_write_s64,
1888 .private = FILE_SCHED_RELAX_DOMAIN_LEVEL,
1892 .name = "memory_migrate",
1893 .read_u64 = cpuset_read_u64,
1894 .write_u64 = cpuset_write_u64,
1895 .private = FILE_MEMORY_MIGRATE,
1899 .name = "memory_pressure",
1900 .read_u64 = cpuset_read_u64,
1904 .name = "memory_spread_page",
1905 .read_u64 = cpuset_read_u64,
1906 .write_u64 = cpuset_write_u64,
1907 .private = FILE_SPREAD_PAGE,
1911 .name = "memory_spread_slab",
1912 .read_u64 = cpuset_read_u64,
1913 .write_u64 = cpuset_write_u64,
1914 .private = FILE_SPREAD_SLAB,
1918 .name = "memory_pressure_enabled",
1919 .flags = CFTYPE_ONLY_ON_ROOT,
1920 .read_u64 = cpuset_read_u64,
1921 .write_u64 = cpuset_write_u64,
1922 .private = FILE_MEMORY_PRESSURE_ENABLED,
1929 * cpuset_css_alloc - allocate a cpuset css
1930 * cgrp: control group that the new cpuset will be part of
1933 static struct cgroup_subsys_state *
1934 cpuset_css_alloc(struct cgroup_subsys_state *parent_css)
1939 return &top_cpuset.css;
1941 cs = kzalloc(sizeof(*cs), GFP_KERNEL);
1943 return ERR_PTR(-ENOMEM);
1944 if (!alloc_cpumask_var(&cs->cpus_allowed, GFP_KERNEL))
1946 if (!alloc_cpumask_var(&cs->effective_cpus, GFP_KERNEL))
1949 set_bit(CS_SCHED_LOAD_BALANCE, &cs->flags);
1950 cpumask_clear(cs->cpus_allowed);
1951 nodes_clear(cs->mems_allowed);
1952 cpumask_clear(cs->effective_cpus);
1953 nodes_clear(cs->effective_mems);
1954 fmeter_init(&cs->fmeter);
1955 cs->relax_domain_level = -1;
1960 free_cpumask_var(cs->cpus_allowed);
1963 return ERR_PTR(-ENOMEM);
1966 static int cpuset_css_online(struct cgroup_subsys_state *css)
1968 struct cpuset *cs = css_cs(css);
1969 struct cpuset *parent = parent_cs(cs);
1970 struct cpuset *tmp_cs;
1971 struct cgroup_subsys_state *pos_css;
1976 mutex_lock(&cpuset_mutex);
1978 set_bit(CS_ONLINE, &cs->flags);
1979 if (is_spread_page(parent))
1980 set_bit(CS_SPREAD_PAGE, &cs->flags);
1981 if (is_spread_slab(parent))
1982 set_bit(CS_SPREAD_SLAB, &cs->flags);
1986 spin_lock_irq(&callback_lock);
1987 if (cgroup_subsys_on_dfl(cpuset_cgrp_subsys)) {
1988 cpumask_copy(cs->effective_cpus, parent->effective_cpus);
1989 cs->effective_mems = parent->effective_mems;
1991 spin_unlock_irq(&callback_lock);
1993 if (!test_bit(CGRP_CPUSET_CLONE_CHILDREN, &css->cgroup->flags))
1997 * Clone @parent's configuration if CGRP_CPUSET_CLONE_CHILDREN is
1998 * set. This flag handling is implemented in cgroup core for
1999 * histrical reasons - the flag may be specified during mount.
2001 * Currently, if any sibling cpusets have exclusive cpus or mem, we
2002 * refuse to clone the configuration - thereby refusing the task to
2003 * be entered, and as a result refusing the sys_unshare() or
2004 * clone() which initiated it. If this becomes a problem for some
2005 * users who wish to allow that scenario, then this could be
2006 * changed to grant parent->cpus_allowed-sibling_cpus_exclusive
2007 * (and likewise for mems) to the new cgroup.
2010 cpuset_for_each_child(tmp_cs, pos_css, parent) {
2011 if (is_mem_exclusive(tmp_cs) || is_cpu_exclusive(tmp_cs)) {
2018 spin_lock_irq(&callback_lock);
2019 cs->mems_allowed = parent->mems_allowed;
2020 cs->effective_mems = parent->mems_allowed;
2021 cpumask_copy(cs->cpus_allowed, parent->cpus_allowed);
2022 cpumask_copy(cs->effective_cpus, parent->cpus_allowed);
2023 spin_unlock_irq(&callback_lock);
2025 mutex_unlock(&cpuset_mutex);
2030 * If the cpuset being removed has its flag 'sched_load_balance'
2031 * enabled, then simulate turning sched_load_balance off, which
2032 * will call rebuild_sched_domains_locked().
2035 static void cpuset_css_offline(struct cgroup_subsys_state *css)
2037 struct cpuset *cs = css_cs(css);
2039 mutex_lock(&cpuset_mutex);
2041 if (is_sched_load_balance(cs))
2042 update_flag(CS_SCHED_LOAD_BALANCE, cs, 0);
2045 clear_bit(CS_ONLINE, &cs->flags);
2047 mutex_unlock(&cpuset_mutex);
2050 static void cpuset_css_free(struct cgroup_subsys_state *css)
2052 struct cpuset *cs = css_cs(css);
2054 free_cpumask_var(cs->effective_cpus);
2055 free_cpumask_var(cs->cpus_allowed);
2059 static void cpuset_bind(struct cgroup_subsys_state *root_css)
2061 mutex_lock(&cpuset_mutex);
2062 spin_lock_irq(&callback_lock);
2064 if (cgroup_subsys_on_dfl(cpuset_cgrp_subsys)) {
2065 cpumask_copy(top_cpuset.cpus_allowed, cpu_possible_mask);
2066 top_cpuset.mems_allowed = node_possible_map;
2068 cpumask_copy(top_cpuset.cpus_allowed,
2069 top_cpuset.effective_cpus);
2070 top_cpuset.mems_allowed = top_cpuset.effective_mems;
2073 spin_unlock_irq(&callback_lock);
2074 mutex_unlock(&cpuset_mutex);
2077 struct cgroup_subsys cpuset_cgrp_subsys = {
2078 .css_alloc = cpuset_css_alloc,
2079 .css_online = cpuset_css_online,
2080 .css_offline = cpuset_css_offline,
2081 .css_free = cpuset_css_free,
2082 .can_attach = cpuset_can_attach,
2083 .cancel_attach = cpuset_cancel_attach,
2084 .attach = cpuset_attach,
2085 .post_attach = cpuset_post_attach,
2086 .bind = cpuset_bind,
2087 .legacy_cftypes = files,
2092 * cpuset_init - initialize cpusets at system boot
2094 * Description: Initialize top_cpuset and the cpuset internal file system,
2097 int __init cpuset_init(void)
2101 if (!alloc_cpumask_var(&top_cpuset.cpus_allowed, GFP_KERNEL))
2103 if (!alloc_cpumask_var(&top_cpuset.effective_cpus, GFP_KERNEL))
2106 cpumask_setall(top_cpuset.cpus_allowed);
2107 nodes_setall(top_cpuset.mems_allowed);
2108 cpumask_setall(top_cpuset.effective_cpus);
2109 nodes_setall(top_cpuset.effective_mems);
2111 fmeter_init(&top_cpuset.fmeter);
2112 set_bit(CS_SCHED_LOAD_BALANCE, &top_cpuset.flags);
2113 top_cpuset.relax_domain_level = -1;
2115 err = register_filesystem(&cpuset_fs_type);
2119 if (!alloc_cpumask_var(&cpus_attach, GFP_KERNEL))
2126 * If CPU and/or memory hotplug handlers, below, unplug any CPUs
2127 * or memory nodes, we need to walk over the cpuset hierarchy,
2128 * removing that CPU or node from all cpusets. If this removes the
2129 * last CPU or node from a cpuset, then move the tasks in the empty
2130 * cpuset to its next-highest non-empty parent.
2132 static void remove_tasks_in_empty_cpuset(struct cpuset *cs)
2134 struct cpuset *parent;
2137 * Find its next-highest non-empty parent, (top cpuset
2138 * has online cpus, so can't be empty).
2140 parent = parent_cs(cs);
2141 while (cpumask_empty(parent->cpus_allowed) ||
2142 nodes_empty(parent->mems_allowed))
2143 parent = parent_cs(parent);
2145 if (cgroup_transfer_tasks(parent->css.cgroup, cs->css.cgroup)) {
2146 pr_err("cpuset: failed to transfer tasks out of empty cpuset ");
2147 pr_cont_cgroup_name(cs->css.cgroup);
2153 hotplug_update_tasks_legacy(struct cpuset *cs,
2154 struct cpumask *new_cpus, nodemask_t *new_mems,
2155 bool cpus_updated, bool mems_updated)
2159 spin_lock_irq(&callback_lock);
2160 cpumask_copy(cs->cpus_allowed, new_cpus);
2161 cpumask_copy(cs->effective_cpus, new_cpus);
2162 cs->mems_allowed = *new_mems;
2163 cs->effective_mems = *new_mems;
2164 spin_unlock_irq(&callback_lock);
2167 * Don't call update_tasks_cpumask() if the cpuset becomes empty,
2168 * as the tasks will be migratecd to an ancestor.
2170 if (cpus_updated && !cpumask_empty(cs->cpus_allowed))
2171 update_tasks_cpumask(cs);
2172 if (mems_updated && !nodes_empty(cs->mems_allowed))
2173 update_tasks_nodemask(cs);
2175 is_empty = cpumask_empty(cs->cpus_allowed) ||
2176 nodes_empty(cs->mems_allowed);
2178 mutex_unlock(&cpuset_mutex);
2181 * Move tasks to the nearest ancestor with execution resources,
2182 * This is full cgroup operation which will also call back into
2183 * cpuset. Should be done outside any lock.
2186 remove_tasks_in_empty_cpuset(cs);
2188 mutex_lock(&cpuset_mutex);
2192 hotplug_update_tasks(struct cpuset *cs,
2193 struct cpumask *new_cpus, nodemask_t *new_mems,
2194 bool cpus_updated, bool mems_updated)
2196 if (cpumask_empty(new_cpus))
2197 cpumask_copy(new_cpus, parent_cs(cs)->effective_cpus);
2198 if (nodes_empty(*new_mems))
2199 *new_mems = parent_cs(cs)->effective_mems;
2201 spin_lock_irq(&callback_lock);
2202 cpumask_copy(cs->effective_cpus, new_cpus);
2203 cs->effective_mems = *new_mems;
2204 spin_unlock_irq(&callback_lock);
2207 update_tasks_cpumask(cs);
2209 update_tasks_nodemask(cs);
2213 * cpuset_hotplug_update_tasks - update tasks in a cpuset for hotunplug
2214 * @cs: cpuset in interest
2216 * Compare @cs's cpu and mem masks against top_cpuset and if some have gone
2217 * offline, update @cs accordingly. If @cs ends up with no CPU or memory,
2218 * all its tasks are moved to the nearest ancestor with both resources.
2220 static void cpuset_hotplug_update_tasks(struct cpuset *cs)
2222 static cpumask_t new_cpus;
2223 static nodemask_t new_mems;
2227 wait_event(cpuset_attach_wq, cs->attach_in_progress == 0);
2229 mutex_lock(&cpuset_mutex);
2232 * We have raced with task attaching. We wait until attaching
2233 * is finished, so we won't attach a task to an empty cpuset.
2235 if (cs->attach_in_progress) {
2236 mutex_unlock(&cpuset_mutex);
2240 cpumask_and(&new_cpus, cs->cpus_allowed, parent_cs(cs)->effective_cpus);
2241 nodes_and(new_mems, cs->mems_allowed, parent_cs(cs)->effective_mems);
2243 cpus_updated = !cpumask_equal(&new_cpus, cs->effective_cpus);
2244 mems_updated = !nodes_equal(new_mems, cs->effective_mems);
2246 if (cgroup_subsys_on_dfl(cpuset_cgrp_subsys))
2247 hotplug_update_tasks(cs, &new_cpus, &new_mems,
2248 cpus_updated, mems_updated);
2250 hotplug_update_tasks_legacy(cs, &new_cpus, &new_mems,
2251 cpus_updated, mems_updated);
2253 mutex_unlock(&cpuset_mutex);
2257 * cpuset_hotplug_workfn - handle CPU/memory hotunplug for a cpuset
2259 * This function is called after either CPU or memory configuration has
2260 * changed and updates cpuset accordingly. The top_cpuset is always
2261 * synchronized to cpu_active_mask and N_MEMORY, which is necessary in
2262 * order to make cpusets transparent (of no affect) on systems that are
2263 * actively using CPU hotplug but making no active use of cpusets.
2265 * Non-root cpusets are only affected by offlining. If any CPUs or memory
2266 * nodes have been taken down, cpuset_hotplug_update_tasks() is invoked on
2269 * Note that CPU offlining during suspend is ignored. We don't modify
2270 * cpusets across suspend/resume cycles at all.
2272 static void cpuset_hotplug_workfn(struct work_struct *work)
2274 static cpumask_t new_cpus;
2275 static nodemask_t new_mems;
2276 bool cpus_updated, mems_updated;
2277 bool on_dfl = cgroup_subsys_on_dfl(cpuset_cgrp_subsys);
2279 mutex_lock(&cpuset_mutex);
2281 /* fetch the available cpus/mems and find out which changed how */
2282 cpumask_copy(&new_cpus, cpu_active_mask);
2283 new_mems = node_states[N_MEMORY];
2285 cpus_updated = !cpumask_equal(top_cpuset.effective_cpus, &new_cpus);
2286 mems_updated = !nodes_equal(top_cpuset.effective_mems, new_mems);
2288 /* synchronize cpus_allowed to cpu_active_mask */
2290 spin_lock_irq(&callback_lock);
2292 cpumask_copy(top_cpuset.cpus_allowed, &new_cpus);
2293 cpumask_copy(top_cpuset.effective_cpus, &new_cpus);
2294 spin_unlock_irq(&callback_lock);
2295 /* we don't mess with cpumasks of tasks in top_cpuset */
2298 /* synchronize mems_allowed to N_MEMORY */
2300 spin_lock_irq(&callback_lock);
2302 top_cpuset.mems_allowed = new_mems;
2303 top_cpuset.effective_mems = new_mems;
2304 spin_unlock_irq(&callback_lock);
2305 update_tasks_nodemask(&top_cpuset);
2308 mutex_unlock(&cpuset_mutex);
2310 /* if cpus or mems changed, we need to propagate to descendants */
2311 if (cpus_updated || mems_updated) {
2313 struct cgroup_subsys_state *pos_css;
2316 cpuset_for_each_descendant_pre(cs, pos_css, &top_cpuset) {
2317 if (cs == &top_cpuset || !css_tryget_online(&cs->css))
2321 cpuset_hotplug_update_tasks(cs);
2329 /* rebuild sched domains if cpus_allowed has changed */
2331 rebuild_sched_domains();
2334 void cpuset_update_active_cpus(bool cpu_online)
2337 * We're inside cpu hotplug critical region which usually nests
2338 * inside cgroup synchronization. Bounce actual hotplug processing
2339 * to a work item to avoid reverse locking order.
2341 * We still need to do partition_sched_domains() synchronously;
2342 * otherwise, the scheduler will get confused and put tasks to the
2343 * dead CPU. Fall back to the default single domain.
2344 * cpuset_hotplug_workfn() will rebuild it as necessary.
2346 partition_sched_domains(1, NULL, NULL);
2347 schedule_work(&cpuset_hotplug_work);
2351 * Keep top_cpuset.mems_allowed tracking node_states[N_MEMORY].
2352 * Call this routine anytime after node_states[N_MEMORY] changes.
2353 * See cpuset_update_active_cpus() for CPU hotplug handling.
2355 static int cpuset_track_online_nodes(struct notifier_block *self,
2356 unsigned long action, void *arg)
2358 schedule_work(&cpuset_hotplug_work);
2362 static struct notifier_block cpuset_track_online_nodes_nb = {
2363 .notifier_call = cpuset_track_online_nodes,
2364 .priority = 10, /* ??! */
2368 * cpuset_init_smp - initialize cpus_allowed
2370 * Description: Finish top cpuset after cpu, node maps are initialized
2372 void __init cpuset_init_smp(void)
2374 cpumask_copy(top_cpuset.cpus_allowed, cpu_active_mask);
2375 top_cpuset.mems_allowed = node_states[N_MEMORY];
2376 top_cpuset.old_mems_allowed = top_cpuset.mems_allowed;
2378 cpumask_copy(top_cpuset.effective_cpus, cpu_active_mask);
2379 top_cpuset.effective_mems = node_states[N_MEMORY];
2381 register_hotmemory_notifier(&cpuset_track_online_nodes_nb);
2383 cpuset_migrate_mm_wq = alloc_ordered_workqueue("cpuset_migrate_mm", 0);
2384 BUG_ON(!cpuset_migrate_mm_wq);
2388 * cpuset_cpus_allowed - return cpus_allowed mask from a tasks cpuset.
2389 * @tsk: pointer to task_struct from which to obtain cpuset->cpus_allowed.
2390 * @pmask: pointer to struct cpumask variable to receive cpus_allowed set.
2392 * Description: Returns the cpumask_var_t cpus_allowed of the cpuset
2393 * attached to the specified @tsk. Guaranteed to return some non-empty
2394 * subset of cpu_online_mask, even if this means going outside the
2398 void cpuset_cpus_allowed(struct task_struct *tsk, struct cpumask *pmask)
2400 unsigned long flags;
2402 spin_lock_irqsave(&callback_lock, flags);
2404 guarantee_online_cpus(task_cs(tsk), pmask);
2406 spin_unlock_irqrestore(&callback_lock, flags);
2409 void cpuset_cpus_allowed_fallback(struct task_struct *tsk)
2412 do_set_cpus_allowed(tsk, task_cs(tsk)->effective_cpus);
2416 * We own tsk->cpus_allowed, nobody can change it under us.
2418 * But we used cs && cs->cpus_allowed lockless and thus can
2419 * race with cgroup_attach_task() or update_cpumask() and get
2420 * the wrong tsk->cpus_allowed. However, both cases imply the
2421 * subsequent cpuset_change_cpumask()->set_cpus_allowed_ptr()
2422 * which takes task_rq_lock().
2424 * If we are called after it dropped the lock we must see all
2425 * changes in tsk_cs()->cpus_allowed. Otherwise we can temporary
2426 * set any mask even if it is not right from task_cs() pov,
2427 * the pending set_cpus_allowed_ptr() will fix things.
2429 * select_fallback_rq() will fix things ups and set cpu_possible_mask
2434 void __init cpuset_init_current_mems_allowed(void)
2436 nodes_setall(current->mems_allowed);
2440 * cpuset_mems_allowed - return mems_allowed mask from a tasks cpuset.
2441 * @tsk: pointer to task_struct from which to obtain cpuset->mems_allowed.
2443 * Description: Returns the nodemask_t mems_allowed of the cpuset
2444 * attached to the specified @tsk. Guaranteed to return some non-empty
2445 * subset of node_states[N_MEMORY], even if this means going outside the
2449 nodemask_t cpuset_mems_allowed(struct task_struct *tsk)
2452 unsigned long flags;
2454 spin_lock_irqsave(&callback_lock, flags);
2456 guarantee_online_mems(task_cs(tsk), &mask);
2458 spin_unlock_irqrestore(&callback_lock, flags);
2464 * cpuset_nodemask_valid_mems_allowed - check nodemask vs. curremt mems_allowed
2465 * @nodemask: the nodemask to be checked
2467 * Are any of the nodes in the nodemask allowed in current->mems_allowed?
2469 int cpuset_nodemask_valid_mems_allowed(nodemask_t *nodemask)
2471 return nodes_intersects(*nodemask, current->mems_allowed);
2475 * nearest_hardwall_ancestor() - Returns the nearest mem_exclusive or
2476 * mem_hardwall ancestor to the specified cpuset. Call holding
2477 * callback_lock. If no ancestor is mem_exclusive or mem_hardwall
2478 * (an unusual configuration), then returns the root cpuset.
2480 static struct cpuset *nearest_hardwall_ancestor(struct cpuset *cs)
2482 while (!(is_mem_exclusive(cs) || is_mem_hardwall(cs)) && parent_cs(cs))
2488 * cpuset_node_allowed - Can we allocate on a memory node?
2489 * @node: is this an allowed node?
2490 * @gfp_mask: memory allocation flags
2492 * If we're in interrupt, yes, we can always allocate. If @node is set in
2493 * current's mems_allowed, yes. If it's not a __GFP_HARDWALL request and this
2494 * node is set in the nearest hardwalled cpuset ancestor to current's cpuset,
2495 * yes. If current has access to memory reserves due to TIF_MEMDIE, yes.
2498 * GFP_USER allocations are marked with the __GFP_HARDWALL bit,
2499 * and do not allow allocations outside the current tasks cpuset
2500 * unless the task has been OOM killed as is marked TIF_MEMDIE.
2501 * GFP_KERNEL allocations are not so marked, so can escape to the
2502 * nearest enclosing hardwalled ancestor cpuset.
2504 * Scanning up parent cpusets requires callback_lock. The
2505 * __alloc_pages() routine only calls here with __GFP_HARDWALL bit
2506 * _not_ set if it's a GFP_KERNEL allocation, and all nodes in the
2507 * current tasks mems_allowed came up empty on the first pass over
2508 * the zonelist. So only GFP_KERNEL allocations, if all nodes in the
2509 * cpuset are short of memory, might require taking the callback_lock.
2511 * The first call here from mm/page_alloc:get_page_from_freelist()
2512 * has __GFP_HARDWALL set in gfp_mask, enforcing hardwall cpusets,
2513 * so no allocation on a node outside the cpuset is allowed (unless
2514 * in interrupt, of course).
2516 * The second pass through get_page_from_freelist() doesn't even call
2517 * here for GFP_ATOMIC calls. For those calls, the __alloc_pages()
2518 * variable 'wait' is not set, and the bit ALLOC_CPUSET is not set
2519 * in alloc_flags. That logic and the checks below have the combined
2521 * in_interrupt - any node ok (current task context irrelevant)
2522 * GFP_ATOMIC - any node ok
2523 * TIF_MEMDIE - any node ok
2524 * GFP_KERNEL - any node in enclosing hardwalled cpuset ok
2525 * GFP_USER - only nodes in current tasks mems allowed ok.
2527 int __cpuset_node_allowed(int node, gfp_t gfp_mask)
2529 struct cpuset *cs; /* current cpuset ancestors */
2530 int allowed; /* is allocation in zone z allowed? */
2531 unsigned long flags;
2535 if (node_isset(node, current->mems_allowed))
2538 * Allow tasks that have access to memory reserves because they have
2539 * been OOM killed to get memory anywhere.
2541 if (unlikely(test_thread_flag(TIF_MEMDIE)))
2543 if (gfp_mask & __GFP_HARDWALL) /* If hardwall request, stop here */
2546 if (current->flags & PF_EXITING) /* Let dying task have memory */
2549 /* Not hardwall and node outside mems_allowed: scan up cpusets */
2550 spin_lock_irqsave(&callback_lock, flags);
2553 cs = nearest_hardwall_ancestor(task_cs(current));
2554 allowed = node_isset(node, cs->mems_allowed);
2557 spin_unlock_irqrestore(&callback_lock, flags);
2562 * cpuset_mem_spread_node() - On which node to begin search for a file page
2563 * cpuset_slab_spread_node() - On which node to begin search for a slab page
2565 * If a task is marked PF_SPREAD_PAGE or PF_SPREAD_SLAB (as for
2566 * tasks in a cpuset with is_spread_page or is_spread_slab set),
2567 * and if the memory allocation used cpuset_mem_spread_node()
2568 * to determine on which node to start looking, as it will for
2569 * certain page cache or slab cache pages such as used for file
2570 * system buffers and inode caches, then instead of starting on the
2571 * local node to look for a free page, rather spread the starting
2572 * node around the tasks mems_allowed nodes.
2574 * We don't have to worry about the returned node being offline
2575 * because "it can't happen", and even if it did, it would be ok.
2577 * The routines calling guarantee_online_mems() are careful to
2578 * only set nodes in task->mems_allowed that are online. So it
2579 * should not be possible for the following code to return an
2580 * offline node. But if it did, that would be ok, as this routine
2581 * is not returning the node where the allocation must be, only
2582 * the node where the search should start. The zonelist passed to
2583 * __alloc_pages() will include all nodes. If the slab allocator
2584 * is passed an offline node, it will fall back to the local node.
2585 * See kmem_cache_alloc_node().
2588 static int cpuset_spread_node(int *rotor)
2592 node = next_node(*rotor, current->mems_allowed);
2593 if (node == MAX_NUMNODES)
2594 node = first_node(current->mems_allowed);
2599 int cpuset_mem_spread_node(void)
2601 if (current->cpuset_mem_spread_rotor == NUMA_NO_NODE)
2602 current->cpuset_mem_spread_rotor =
2603 node_random(¤t->mems_allowed);
2605 return cpuset_spread_node(¤t->cpuset_mem_spread_rotor);
2608 int cpuset_slab_spread_node(void)
2610 if (current->cpuset_slab_spread_rotor == NUMA_NO_NODE)
2611 current->cpuset_slab_spread_rotor =
2612 node_random(¤t->mems_allowed);
2614 return cpuset_spread_node(¤t->cpuset_slab_spread_rotor);
2617 EXPORT_SYMBOL_GPL(cpuset_mem_spread_node);
2620 * cpuset_mems_allowed_intersects - Does @tsk1's mems_allowed intersect @tsk2's?
2621 * @tsk1: pointer to task_struct of some task.
2622 * @tsk2: pointer to task_struct of some other task.
2624 * Description: Return true if @tsk1's mems_allowed intersects the
2625 * mems_allowed of @tsk2. Used by the OOM killer to determine if
2626 * one of the task's memory usage might impact the memory available
2630 int cpuset_mems_allowed_intersects(const struct task_struct *tsk1,
2631 const struct task_struct *tsk2)
2633 return nodes_intersects(tsk1->mems_allowed, tsk2->mems_allowed);
2637 * cpuset_print_current_mems_allowed - prints current's cpuset and mems_allowed
2639 * Description: Prints current's name, cpuset name, and cached copy of its
2640 * mems_allowed to the kernel log.
2642 void cpuset_print_current_mems_allowed(void)
2644 struct cgroup *cgrp;
2648 cgrp = task_cs(current)->css.cgroup;
2649 pr_info("%s cpuset=", current->comm);
2650 pr_cont_cgroup_name(cgrp);
2651 pr_cont(" mems_allowed=%*pbl\n",
2652 nodemask_pr_args(¤t->mems_allowed));
2658 * Collection of memory_pressure is suppressed unless
2659 * this flag is enabled by writing "1" to the special
2660 * cpuset file 'memory_pressure_enabled' in the root cpuset.
2663 int cpuset_memory_pressure_enabled __read_mostly;
2666 * cpuset_memory_pressure_bump - keep stats of per-cpuset reclaims.
2668 * Keep a running average of the rate of synchronous (direct)
2669 * page reclaim efforts initiated by tasks in each cpuset.
2671 * This represents the rate at which some task in the cpuset
2672 * ran low on memory on all nodes it was allowed to use, and
2673 * had to enter the kernels page reclaim code in an effort to
2674 * create more free memory by tossing clean pages or swapping
2675 * or writing dirty pages.
2677 * Display to user space in the per-cpuset read-only file
2678 * "memory_pressure". Value displayed is an integer
2679 * representing the recent rate of entry into the synchronous
2680 * (direct) page reclaim by any task attached to the cpuset.
2683 void __cpuset_memory_pressure_bump(void)
2686 fmeter_markevent(&task_cs(current)->fmeter);
2690 #ifdef CONFIG_PROC_PID_CPUSET
2692 * proc_cpuset_show()
2693 * - Print tasks cpuset path into seq_file.
2694 * - Used for /proc/<pid>/cpuset.
2695 * - No need to task_lock(tsk) on this tsk->cpuset reference, as it
2696 * doesn't really matter if tsk->cpuset changes after we read it,
2697 * and we take cpuset_mutex, keeping cpuset_attach() from changing it
2700 int proc_cpuset_show(struct seq_file *m, struct pid_namespace *ns,
2701 struct pid *pid, struct task_struct *tsk)
2704 struct cgroup_subsys_state *css;
2708 buf = kmalloc(PATH_MAX, GFP_KERNEL);
2712 retval = -ENAMETOOLONG;
2714 css = task_css(tsk, cpuset_cgrp_id);
2715 p = cgroup_path(css->cgroup, buf, PATH_MAX);
2727 #endif /* CONFIG_PROC_PID_CPUSET */
2729 /* Display task mems_allowed in /proc/<pid>/status file. */
2730 void cpuset_task_status_allowed(struct seq_file *m, struct task_struct *task)
2732 seq_printf(m, "Mems_allowed:\t%*pb\n",
2733 nodemask_pr_args(&task->mems_allowed));
2734 seq_printf(m, "Mems_allowed_list:\t%*pbl\n",
2735 nodemask_pr_args(&task->mems_allowed));