4 * Processor and Memory placement constraints for sets of tasks.
6 * Copyright (C) 2003 BULL SA.
7 * Copyright (C) 2004-2007 Silicon Graphics, Inc.
8 * Copyright (C) 2006 Google, Inc
10 * Portions derived from Patrick Mochel's sysfs code.
11 * sysfs is Copyright (c) 2001-3 Patrick Mochel
13 * 2003-10-10 Written by Simon Derr.
14 * 2003-10-22 Updates by Stephen Hemminger.
15 * 2004 May-July Rework by Paul Jackson.
16 * 2006 Rework by Paul Menage to use generic cgroups
17 * 2008 Rework of the scheduler domains and CPU hotplug handling
20 * This file is subject to the terms and conditions of the GNU General Public
21 * License. See the file COPYING in the main directory of the Linux
22 * distribution for more details.
25 #include <linux/cpu.h>
26 #include <linux/cpumask.h>
27 #include <linux/cpuset.h>
28 #include <linux/err.h>
29 #include <linux/errno.h>
30 #include <linux/file.h>
32 #include <linux/init.h>
33 #include <linux/interrupt.h>
34 #include <linux/kernel.h>
35 #include <linux/kmod.h>
36 #include <linux/list.h>
37 #include <linux/mempolicy.h>
39 #include <linux/memory.h>
40 #include <linux/export.h>
41 #include <linux/mount.h>
42 #include <linux/namei.h>
43 #include <linux/pagemap.h>
44 #include <linux/proc_fs.h>
45 #include <linux/rcupdate.h>
46 #include <linux/sched.h>
47 #include <linux/seq_file.h>
48 #include <linux/security.h>
49 #include <linux/slab.h>
50 #include <linux/spinlock.h>
51 #include <linux/stat.h>
52 #include <linux/string.h>
53 #include <linux/time.h>
54 #include <linux/backing-dev.h>
55 #include <linux/sort.h>
57 #include <asm/uaccess.h>
58 #include <linux/atomic.h>
59 #include <linux/mutex.h>
60 #include <linux/cgroup.h>
61 #include <linux/wait.h>
63 struct static_key cpusets_pre_enable_key __read_mostly = STATIC_KEY_INIT_FALSE;
64 struct static_key cpusets_enabled_key __read_mostly = STATIC_KEY_INIT_FALSE;
66 /* See "Frequency meter" comments, below. */
69 int cnt; /* unprocessed events count */
70 int val; /* most recent output value */
71 time_t time; /* clock (secs) when val computed */
72 spinlock_t lock; /* guards read or write of above */
76 struct cgroup_subsys_state css;
78 unsigned long flags; /* "unsigned long" so bitops work */
81 * On default hierarchy:
83 * The user-configured masks can only be changed by writing to
84 * cpuset.cpus and cpuset.mems, and won't be limited by the
87 * The effective masks is the real masks that apply to the tasks
88 * in the cpuset. They may be changed if the configured masks are
89 * changed or hotplug happens.
91 * effective_mask == configured_mask & parent's effective_mask,
92 * and if it ends up empty, it will inherit the parent's mask.
97 * The user-configured masks are always the same with effective masks.
100 /* user-configured CPUs and Memory Nodes allow to tasks */
101 cpumask_var_t cpus_allowed;
102 cpumask_var_t cpus_requested; /* CPUS requested, but not used because of hotplug */
103 nodemask_t mems_allowed;
105 /* effective CPUs and Memory Nodes allow to tasks */
106 cpumask_var_t effective_cpus;
107 nodemask_t effective_mems;
110 * This is old Memory Nodes tasks took on.
112 * - top_cpuset.old_mems_allowed is initialized to mems_allowed.
113 * - A new cpuset's old_mems_allowed is initialized when some
114 * task is moved into it.
115 * - old_mems_allowed is used in cpuset_migrate_mm() when we change
116 * cpuset.mems_allowed and have tasks' nodemask updated, and
117 * then old_mems_allowed is updated to mems_allowed.
119 nodemask_t old_mems_allowed;
121 struct fmeter fmeter; /* memory_pressure filter */
124 * Tasks are being attached to this cpuset. Used to prevent
125 * zeroing cpus/mems_allowed between ->can_attach() and ->attach().
127 int attach_in_progress;
129 /* partition number for rebuild_sched_domains() */
132 /* for custom sched domain */
133 int relax_domain_level;
136 static inline struct cpuset *css_cs(struct cgroup_subsys_state *css)
138 return css ? container_of(css, struct cpuset, css) : NULL;
141 /* Retrieve the cpuset for a task */
142 static inline struct cpuset *task_cs(struct task_struct *task)
144 return css_cs(task_css(task, cpuset_cgrp_id));
147 static inline struct cpuset *parent_cs(struct cpuset *cs)
149 return css_cs(cs->css.parent);
153 static inline bool task_has_mempolicy(struct task_struct *task)
155 return task->mempolicy;
158 static inline bool task_has_mempolicy(struct task_struct *task)
165 /* bits in struct cpuset flags field */
172 CS_SCHED_LOAD_BALANCE,
177 /* convenient tests for these bits */
178 static inline bool is_cpuset_online(struct cpuset *cs)
180 return test_bit(CS_ONLINE, &cs->flags) && !css_is_dying(&cs->css);
183 static inline int is_cpu_exclusive(const struct cpuset *cs)
185 return test_bit(CS_CPU_EXCLUSIVE, &cs->flags);
188 static inline int is_mem_exclusive(const struct cpuset *cs)
190 return test_bit(CS_MEM_EXCLUSIVE, &cs->flags);
193 static inline int is_mem_hardwall(const struct cpuset *cs)
195 return test_bit(CS_MEM_HARDWALL, &cs->flags);
198 static inline int is_sched_load_balance(const struct cpuset *cs)
200 return test_bit(CS_SCHED_LOAD_BALANCE, &cs->flags);
203 static inline int is_memory_migrate(const struct cpuset *cs)
205 return test_bit(CS_MEMORY_MIGRATE, &cs->flags);
208 static inline int is_spread_page(const struct cpuset *cs)
210 return test_bit(CS_SPREAD_PAGE, &cs->flags);
213 static inline int is_spread_slab(const struct cpuset *cs)
215 return test_bit(CS_SPREAD_SLAB, &cs->flags);
218 static struct cpuset top_cpuset = {
219 .flags = ((1 << CS_ONLINE) | (1 << CS_CPU_EXCLUSIVE) |
220 (1 << CS_MEM_EXCLUSIVE)),
224 * cpuset_for_each_child - traverse online children of a cpuset
225 * @child_cs: loop cursor pointing to the current child
226 * @pos_css: used for iteration
227 * @parent_cs: target cpuset to walk children of
229 * Walk @child_cs through the online children of @parent_cs. Must be used
230 * with RCU read locked.
232 #define cpuset_for_each_child(child_cs, pos_css, parent_cs) \
233 css_for_each_child((pos_css), &(parent_cs)->css) \
234 if (is_cpuset_online(((child_cs) = css_cs((pos_css)))))
237 * cpuset_for_each_descendant_pre - pre-order walk of a cpuset's descendants
238 * @des_cs: loop cursor pointing to the current descendant
239 * @pos_css: used for iteration
240 * @root_cs: target cpuset to walk ancestor of
242 * Walk @des_cs through the online descendants of @root_cs. Must be used
243 * with RCU read locked. The caller may modify @pos_css by calling
244 * css_rightmost_descendant() to skip subtree. @root_cs is included in the
245 * iteration and the first node to be visited.
247 #define cpuset_for_each_descendant_pre(des_cs, pos_css, root_cs) \
248 css_for_each_descendant_pre((pos_css), &(root_cs)->css) \
249 if (is_cpuset_online(((des_cs) = css_cs((pos_css)))))
252 * There are two global locks guarding cpuset structures - cpuset_mutex and
253 * callback_lock. We also require taking task_lock() when dereferencing a
254 * task's cpuset pointer. See "The task_lock() exception", at the end of this
257 * A task must hold both locks to modify cpusets. If a task holds
258 * cpuset_mutex, then it blocks others wanting that mutex, ensuring that it
259 * is the only task able to also acquire callback_lock and be able to
260 * modify cpusets. It can perform various checks on the cpuset structure
261 * first, knowing nothing will change. It can also allocate memory while
262 * just holding cpuset_mutex. While it is performing these checks, various
263 * callback routines can briefly acquire callback_lock to query cpusets.
264 * Once it is ready to make the changes, it takes callback_lock, blocking
267 * Calls to the kernel memory allocator can not be made while holding
268 * callback_lock, as that would risk double tripping on callback_lock
269 * from one of the callbacks into the cpuset code from within
272 * If a task is only holding callback_lock, then it has read-only
275 * Now, the task_struct fields mems_allowed and mempolicy may be changed
276 * by other task, we use alloc_lock in the task_struct fields to protect
279 * The cpuset_common_file_read() handlers only hold callback_lock across
280 * small pieces of code, such as when reading out possibly multi-word
281 * cpumasks and nodemasks.
283 * Accessing a task's cpuset should be done in accordance with the
284 * guidelines for accessing subsystem state in kernel/cgroup.c
287 static DEFINE_MUTEX(cpuset_mutex);
288 static DEFINE_SPINLOCK(callback_lock);
290 static struct workqueue_struct *cpuset_migrate_mm_wq;
293 * CPU / memory hotplug is handled asynchronously.
295 static void cpuset_hotplug_workfn(struct work_struct *work);
296 static DECLARE_WORK(cpuset_hotplug_work, cpuset_hotplug_workfn);
298 static DECLARE_WAIT_QUEUE_HEAD(cpuset_attach_wq);
301 * This is ugly, but preserves the userspace API for existing cpuset
302 * users. If someone tries to mount the "cpuset" filesystem, we
303 * silently switch it to mount "cgroup" instead
305 static struct dentry *cpuset_mount(struct file_system_type *fs_type,
306 int flags, const char *unused_dev_name, void *data)
308 struct file_system_type *cgroup_fs = get_fs_type("cgroup");
309 struct dentry *ret = ERR_PTR(-ENODEV);
313 "release_agent=/sbin/cpuset_release_agent";
314 ret = cgroup_fs->mount(cgroup_fs, flags,
315 unused_dev_name, mountopts);
316 put_filesystem(cgroup_fs);
321 static struct file_system_type cpuset_fs_type = {
323 .mount = cpuset_mount,
327 * Return in pmask the portion of a cpusets's cpus_allowed that
328 * are online. If none are online, walk up the cpuset hierarchy
329 * until we find one that does have some online cpus.
331 * One way or another, we guarantee to return some non-empty subset
332 * of cpu_online_mask.
334 * Call with callback_lock or cpuset_mutex held.
336 static void guarantee_online_cpus(struct cpuset *cs, struct cpumask *pmask)
338 while (!cpumask_intersects(cs->effective_cpus, cpu_online_mask)) {
342 * The top cpuset doesn't have any online cpu as a
343 * consequence of a race between cpuset_hotplug_work
344 * and cpu hotplug notifier. But we know the top
345 * cpuset's effective_cpus is on its way to to be
346 * identical to cpu_online_mask.
348 cpumask_copy(pmask, cpu_online_mask);
352 cpumask_and(pmask, cs->effective_cpus, cpu_online_mask);
356 * Return in *pmask the portion of a cpusets's mems_allowed that
357 * are online, with memory. If none are online with memory, walk
358 * up the cpuset hierarchy until we find one that does have some
359 * online mems. The top cpuset always has some mems online.
361 * One way or another, we guarantee to return some non-empty subset
362 * of node_states[N_MEMORY].
364 * Call with callback_lock or cpuset_mutex held.
366 static void guarantee_online_mems(struct cpuset *cs, nodemask_t *pmask)
368 while (!nodes_intersects(cs->effective_mems, node_states[N_MEMORY]))
370 nodes_and(*pmask, cs->effective_mems, node_states[N_MEMORY]);
374 * update task's spread flag if cpuset's page/slab spread flag is set
376 * Call with callback_lock or cpuset_mutex held.
378 static void cpuset_update_task_spread_flag(struct cpuset *cs,
379 struct task_struct *tsk)
381 if (is_spread_page(cs))
382 task_set_spread_page(tsk);
384 task_clear_spread_page(tsk);
386 if (is_spread_slab(cs))
387 task_set_spread_slab(tsk);
389 task_clear_spread_slab(tsk);
393 * is_cpuset_subset(p, q) - Is cpuset p a subset of cpuset q?
395 * One cpuset is a subset of another if all its allowed CPUs and
396 * Memory Nodes are a subset of the other, and its exclusive flags
397 * are only set if the other's are set. Call holding cpuset_mutex.
400 static int is_cpuset_subset(const struct cpuset *p, const struct cpuset *q)
402 return cpumask_subset(p->cpus_requested, q->cpus_requested) &&
403 nodes_subset(p->mems_allowed, q->mems_allowed) &&
404 is_cpu_exclusive(p) <= is_cpu_exclusive(q) &&
405 is_mem_exclusive(p) <= is_mem_exclusive(q);
409 * alloc_trial_cpuset - allocate a trial cpuset
410 * @cs: the cpuset that the trial cpuset duplicates
412 static struct cpuset *alloc_trial_cpuset(struct cpuset *cs)
414 struct cpuset *trial;
416 trial = kmemdup(cs, sizeof(*cs), GFP_KERNEL);
420 if (!alloc_cpumask_var(&trial->cpus_allowed, GFP_KERNEL))
422 if (!alloc_cpumask_var(&trial->effective_cpus, GFP_KERNEL))
425 cpumask_copy(trial->cpus_allowed, cs->cpus_allowed);
426 cpumask_copy(trial->effective_cpus, cs->effective_cpus);
430 free_cpumask_var(trial->cpus_allowed);
437 * free_trial_cpuset - free the trial cpuset
438 * @trial: the trial cpuset to be freed
440 static void free_trial_cpuset(struct cpuset *trial)
442 free_cpumask_var(trial->effective_cpus);
443 free_cpumask_var(trial->cpus_allowed);
448 * validate_change() - Used to validate that any proposed cpuset change
449 * follows the structural rules for cpusets.
451 * If we replaced the flag and mask values of the current cpuset
452 * (cur) with those values in the trial cpuset (trial), would
453 * our various subset and exclusive rules still be valid? Presumes
456 * 'cur' is the address of an actual, in-use cpuset. Operations
457 * such as list traversal that depend on the actual address of the
458 * cpuset in the list must use cur below, not trial.
460 * 'trial' is the address of bulk structure copy of cur, with
461 * perhaps one or more of the fields cpus_allowed, mems_allowed,
462 * or flags changed to new, trial values.
464 * Return 0 if valid, -errno if not.
467 static int validate_change(struct cpuset *cur, struct cpuset *trial)
469 struct cgroup_subsys_state *css;
470 struct cpuset *c, *par;
475 /* Each of our child cpusets must be a subset of us */
477 cpuset_for_each_child(c, css, cur)
478 if (!is_cpuset_subset(c, trial))
481 /* Remaining checks don't apply to root cpuset */
483 if (cur == &top_cpuset)
486 par = parent_cs(cur);
488 /* On legacy hiearchy, we must be a subset of our parent cpuset. */
490 if (!cgroup_subsys_on_dfl(cpuset_cgrp_subsys) &&
491 !is_cpuset_subset(trial, par))
495 * If either I or some sibling (!= me) is exclusive, we can't
499 cpuset_for_each_child(c, css, par) {
500 if ((is_cpu_exclusive(trial) || is_cpu_exclusive(c)) &&
502 cpumask_intersects(trial->cpus_requested, c->cpus_requested))
504 if ((is_mem_exclusive(trial) || is_mem_exclusive(c)) &&
506 nodes_intersects(trial->mems_allowed, c->mems_allowed))
511 * Cpusets with tasks - existing or newly being attached - can't
512 * be changed to have empty cpus_allowed or mems_allowed.
515 if ((cgroup_is_populated(cur->css.cgroup) || cur->attach_in_progress)) {
516 if (!cpumask_empty(cur->cpus_allowed) &&
517 cpumask_empty(trial->cpus_allowed))
519 if (!nodes_empty(cur->mems_allowed) &&
520 nodes_empty(trial->mems_allowed))
525 * We can't shrink if we won't have enough room for SCHED_DEADLINE
529 if (is_cpu_exclusive(cur) &&
530 !cpuset_cpumask_can_shrink(cur->cpus_allowed,
531 trial->cpus_allowed))
542 * Helper routine for generate_sched_domains().
543 * Do cpusets a, b have overlapping effective cpus_allowed masks?
545 static int cpusets_overlap(struct cpuset *a, struct cpuset *b)
547 return cpumask_intersects(a->effective_cpus, b->effective_cpus);
551 update_domain_attr(struct sched_domain_attr *dattr, struct cpuset *c)
553 if (dattr->relax_domain_level < c->relax_domain_level)
554 dattr->relax_domain_level = c->relax_domain_level;
558 static void update_domain_attr_tree(struct sched_domain_attr *dattr,
559 struct cpuset *root_cs)
562 struct cgroup_subsys_state *pos_css;
565 cpuset_for_each_descendant_pre(cp, pos_css, root_cs) {
566 /* skip the whole subtree if @cp doesn't have any CPU */
567 if (cpumask_empty(cp->cpus_allowed)) {
568 pos_css = css_rightmost_descendant(pos_css);
572 if (is_sched_load_balance(cp))
573 update_domain_attr(dattr, cp);
579 * generate_sched_domains()
581 * This function builds a partial partition of the systems CPUs
582 * A 'partial partition' is a set of non-overlapping subsets whose
583 * union is a subset of that set.
584 * The output of this function needs to be passed to kernel/sched/core.c
585 * partition_sched_domains() routine, which will rebuild the scheduler's
586 * load balancing domains (sched domains) as specified by that partial
589 * See "What is sched_load_balance" in Documentation/cgroups/cpusets.txt
590 * for a background explanation of this.
592 * Does not return errors, on the theory that the callers of this
593 * routine would rather not worry about failures to rebuild sched
594 * domains when operating in the severe memory shortage situations
595 * that could cause allocation failures below.
597 * Must be called with cpuset_mutex held.
599 * The three key local variables below are:
600 * q - a linked-list queue of cpuset pointers, used to implement a
601 * top-down scan of all cpusets. This scan loads a pointer
602 * to each cpuset marked is_sched_load_balance into the
603 * array 'csa'. For our purposes, rebuilding the schedulers
604 * sched domains, we can ignore !is_sched_load_balance cpusets.
605 * csa - (for CpuSet Array) Array of pointers to all the cpusets
606 * that need to be load balanced, for convenient iterative
607 * access by the subsequent code that finds the best partition,
608 * i.e the set of domains (subsets) of CPUs such that the
609 * cpus_allowed of every cpuset marked is_sched_load_balance
610 * is a subset of one of these domains, while there are as
611 * many such domains as possible, each as small as possible.
612 * doms - Conversion of 'csa' to an array of cpumasks, for passing to
613 * the kernel/sched/core.c routine partition_sched_domains() in a
614 * convenient format, that can be easily compared to the prior
615 * value to determine what partition elements (sched domains)
616 * were changed (added or removed.)
618 * Finding the best partition (set of domains):
619 * The triple nested loops below over i, j, k scan over the
620 * load balanced cpusets (using the array of cpuset pointers in
621 * csa[]) looking for pairs of cpusets that have overlapping
622 * cpus_allowed, but which don't have the same 'pn' partition
623 * number and gives them in the same partition number. It keeps
624 * looping on the 'restart' label until it can no longer find
627 * The union of the cpus_allowed masks from the set of
628 * all cpusets having the same 'pn' value then form the one
629 * element of the partition (one sched domain) to be passed to
630 * partition_sched_domains().
632 static int generate_sched_domains(cpumask_var_t **domains,
633 struct sched_domain_attr **attributes)
635 struct cpuset *cp; /* scans q */
636 struct cpuset **csa; /* array of all cpuset ptrs */
637 int csn; /* how many cpuset ptrs in csa so far */
638 int i, j, k; /* indices for partition finding loops */
639 cpumask_var_t *doms; /* resulting partition; i.e. sched domains */
640 cpumask_var_t non_isolated_cpus; /* load balanced CPUs */
641 struct sched_domain_attr *dattr; /* attributes for custom domains */
642 int ndoms = 0; /* number of sched domains in result */
643 int nslot; /* next empty doms[] struct cpumask slot */
644 struct cgroup_subsys_state *pos_css;
650 if (!alloc_cpumask_var(&non_isolated_cpus, GFP_KERNEL))
652 cpumask_andnot(non_isolated_cpus, cpu_possible_mask, cpu_isolated_map);
654 /* Special case for the 99% of systems with one, full, sched domain */
655 if (is_sched_load_balance(&top_cpuset)) {
657 doms = alloc_sched_domains(ndoms);
661 dattr = kmalloc(sizeof(struct sched_domain_attr), GFP_KERNEL);
663 *dattr = SD_ATTR_INIT;
664 update_domain_attr_tree(dattr, &top_cpuset);
666 cpumask_and(doms[0], top_cpuset.effective_cpus,
672 csa = kmalloc(nr_cpusets() * sizeof(cp), GFP_KERNEL);
678 cpuset_for_each_descendant_pre(cp, pos_css, &top_cpuset) {
679 if (cp == &top_cpuset)
682 * Continue traversing beyond @cp iff @cp has some CPUs and
683 * isn't load balancing. The former is obvious. The
684 * latter: All child cpusets contain a subset of the
685 * parent's cpus, so just skip them, and then we call
686 * update_domain_attr_tree() to calc relax_domain_level of
687 * the corresponding sched domain.
689 if (!cpumask_empty(cp->cpus_allowed) &&
690 !(is_sched_load_balance(cp) &&
691 cpumask_intersects(cp->cpus_allowed, non_isolated_cpus)))
694 if (is_sched_load_balance(cp))
697 /* skip @cp's subtree */
698 pos_css = css_rightmost_descendant(pos_css);
702 for (i = 0; i < csn; i++)
707 /* Find the best partition (set of sched domains) */
708 for (i = 0; i < csn; i++) {
709 struct cpuset *a = csa[i];
712 for (j = 0; j < csn; j++) {
713 struct cpuset *b = csa[j];
716 if (apn != bpn && cpusets_overlap(a, b)) {
717 for (k = 0; k < csn; k++) {
718 struct cpuset *c = csa[k];
723 ndoms--; /* one less element */
730 * Now we know how many domains to create.
731 * Convert <csn, csa> to <ndoms, doms> and populate cpu masks.
733 doms = alloc_sched_domains(ndoms);
738 * The rest of the code, including the scheduler, can deal with
739 * dattr==NULL case. No need to abort if alloc fails.
741 dattr = kmalloc(ndoms * sizeof(struct sched_domain_attr), GFP_KERNEL);
743 for (nslot = 0, i = 0; i < csn; i++) {
744 struct cpuset *a = csa[i];
749 /* Skip completed partitions */
755 if (nslot == ndoms) {
756 static int warnings = 10;
758 pr_warn("rebuild_sched_domains confused: nslot %d, ndoms %d, csn %d, i %d, apn %d\n",
759 nslot, ndoms, csn, i, apn);
767 *(dattr + nslot) = SD_ATTR_INIT;
768 for (j = i; j < csn; j++) {
769 struct cpuset *b = csa[j];
772 cpumask_or(dp, dp, b->effective_cpus);
773 cpumask_and(dp, dp, non_isolated_cpus);
775 update_domain_attr_tree(dattr + nslot, b);
777 /* Done with this partition */
783 BUG_ON(nslot != ndoms);
786 free_cpumask_var(non_isolated_cpus);
790 * Fallback to the default domain if kmalloc() failed.
791 * See comments in partition_sched_domains().
802 * Rebuild scheduler domains.
804 * If the flag 'sched_load_balance' of any cpuset with non-empty
805 * 'cpus' changes, or if the 'cpus' allowed changes in any cpuset
806 * which has that flag enabled, or if any cpuset with a non-empty
807 * 'cpus' is removed, then call this routine to rebuild the
808 * scheduler's dynamic sched domains.
811 static void rebuild_sched_domains_unlocked(void)
813 struct sched_domain_attr *attr;
817 cpu_hotplug_mutex_held();
818 lockdep_assert_held(&cpuset_mutex);
821 * We have raced with CPU hotplug. Don't do anything to avoid
822 * passing doms with offlined cpu to partition_sched_domains().
823 * Anyways, hotplug work item will rebuild sched domains.
825 if (!cpumask_equal(top_cpuset.effective_cpus, cpu_active_mask))
828 /* Generate domain masks and attrs */
829 ndoms = generate_sched_domains(&doms, &attr);
831 /* Have scheduler rebuild the domains */
832 partition_sched_domains(ndoms, doms, attr);
834 #else /* !CONFIG_SMP */
835 static void rebuild_sched_domains_unlocked(void)
838 #endif /* CONFIG_SMP */
840 void rebuild_sched_domains(void)
843 mutex_lock(&cpuset_mutex);
844 rebuild_sched_domains_unlocked();
845 mutex_unlock(&cpuset_mutex);
850 * update_tasks_cpumask - Update the cpumasks of tasks in the cpuset.
851 * @cs: the cpuset in which each task's cpus_allowed mask needs to be changed
853 * Iterate through each task of @cs updating its cpus_allowed to the
854 * effective cpuset's. As this function is called with cpuset_mutex held,
855 * cpuset membership stays stable.
857 static void update_tasks_cpumask(struct cpuset *cs)
859 struct css_task_iter it;
860 struct task_struct *task;
862 css_task_iter_start(&cs->css, &it);
863 while ((task = css_task_iter_next(&it)))
864 set_cpus_allowed_ptr(task, cs->effective_cpus);
865 css_task_iter_end(&it);
869 * update_cpumasks_hier - Update effective cpumasks and tasks in the subtree
870 * @cs: the cpuset to consider
871 * @new_cpus: temp variable for calculating new effective_cpus
873 * When congifured cpumask is changed, the effective cpumasks of this cpuset
874 * and all its descendants need to be updated.
876 * On legacy hierachy, effective_cpus will be the same with cpu_allowed.
879 static void update_cpumasks_hier(struct cpuset *cs, struct cpumask *new_cpus)
882 struct cgroup_subsys_state *pos_css;
883 bool need_rebuild_sched_domains = false;
886 cpuset_for_each_descendant_pre(cp, pos_css, cs) {
887 struct cpuset *parent = parent_cs(cp);
889 cpumask_and(new_cpus, cp->cpus_allowed, parent->effective_cpus);
892 * If it becomes empty, inherit the effective mask of the
893 * parent, which is guaranteed to have some CPUs.
895 if (cgroup_subsys_on_dfl(cpuset_cgrp_subsys) &&
896 cpumask_empty(new_cpus))
897 cpumask_copy(new_cpus, parent->effective_cpus);
899 /* Skip the whole subtree if the cpumask remains the same. */
900 if (cpumask_equal(new_cpus, cp->effective_cpus)) {
901 pos_css = css_rightmost_descendant(pos_css);
905 if (!css_tryget_online(&cp->css))
909 spin_lock_irq(&callback_lock);
910 cpumask_copy(cp->effective_cpus, new_cpus);
911 spin_unlock_irq(&callback_lock);
913 WARN_ON(!cgroup_subsys_on_dfl(cpuset_cgrp_subsys) &&
914 !cpumask_equal(cp->cpus_allowed, cp->effective_cpus));
916 update_tasks_cpumask(cp);
919 * If the effective cpumask of any non-empty cpuset is changed,
920 * we need to rebuild sched domains.
922 if (!cpumask_empty(cp->cpus_allowed) &&
923 is_sched_load_balance(cp))
924 need_rebuild_sched_domains = true;
931 if (need_rebuild_sched_domains)
932 rebuild_sched_domains_unlocked();
936 * update_cpumask - update the cpus_allowed mask of a cpuset and all tasks in it
937 * @cs: the cpuset to consider
938 * @trialcs: trial cpuset
939 * @buf: buffer of cpu numbers written to this cpuset
941 static int update_cpumask(struct cpuset *cs, struct cpuset *trialcs,
946 /* top_cpuset.cpus_allowed tracks cpu_online_mask; it's read-only */
947 if (cs == &top_cpuset)
951 * An empty cpus_allowed is ok only if the cpuset has no tasks.
952 * Since cpulist_parse() fails on an empty mask, we special case
953 * that parsing. The validate_change() call ensures that cpusets
954 * with tasks have cpus.
957 cpumask_clear(trialcs->cpus_allowed);
959 retval = cpulist_parse(buf, trialcs->cpus_requested);
963 if (!cpumask_subset(trialcs->cpus_requested, cpu_present_mask))
966 cpumask_and(trialcs->cpus_allowed, trialcs->cpus_requested, cpu_active_mask);
969 /* Nothing to do if the cpus didn't change */
970 if (cpumask_equal(cs->cpus_requested, trialcs->cpus_requested))
973 retval = validate_change(cs, trialcs);
977 spin_lock_irq(&callback_lock);
978 cpumask_copy(cs->cpus_allowed, trialcs->cpus_allowed);
979 cpumask_copy(cs->cpus_requested, trialcs->cpus_requested);
980 spin_unlock_irq(&callback_lock);
982 /* use trialcs->cpus_allowed as a temp variable */
983 update_cpumasks_hier(cs, trialcs->cpus_allowed);
988 * Migrate memory region from one set of nodes to another. This is
989 * performed asynchronously as it can be called from process migration path
990 * holding locks involved in process management. All mm migrations are
991 * performed in the queued order and can be waited for by flushing
992 * cpuset_migrate_mm_wq.
995 struct cpuset_migrate_mm_work {
996 struct work_struct work;
997 struct mm_struct *mm;
1002 static void cpuset_migrate_mm_workfn(struct work_struct *work)
1004 struct cpuset_migrate_mm_work *mwork =
1005 container_of(work, struct cpuset_migrate_mm_work, work);
1007 /* on a wq worker, no need to worry about %current's mems_allowed */
1008 do_migrate_pages(mwork->mm, &mwork->from, &mwork->to, MPOL_MF_MOVE_ALL);
1013 static void cpuset_migrate_mm(struct mm_struct *mm, const nodemask_t *from,
1014 const nodemask_t *to)
1016 struct cpuset_migrate_mm_work *mwork;
1018 mwork = kzalloc(sizeof(*mwork), GFP_KERNEL);
1021 mwork->from = *from;
1023 INIT_WORK(&mwork->work, cpuset_migrate_mm_workfn);
1024 queue_work(cpuset_migrate_mm_wq, &mwork->work);
1030 static void cpuset_post_attach(void)
1032 flush_workqueue(cpuset_migrate_mm_wq);
1036 * cpuset_change_task_nodemask - change task's mems_allowed and mempolicy
1037 * @tsk: the task to change
1038 * @newmems: new nodes that the task will be set
1040 * In order to avoid seeing no nodes if the old and new nodes are disjoint,
1041 * we structure updates as setting all new allowed nodes, then clearing newly
1044 static void cpuset_change_task_nodemask(struct task_struct *tsk,
1045 nodemask_t *newmems)
1050 * Allow tasks that have access to memory reserves because they have
1051 * been OOM killed to get memory anywhere.
1053 if (unlikely(test_thread_flag(TIF_MEMDIE)))
1055 if (current->flags & PF_EXITING) /* Let dying task have memory */
1060 * Determine if a loop is necessary if another thread is doing
1061 * read_mems_allowed_begin(). If at least one node remains unchanged and
1062 * tsk does not have a mempolicy, then an empty nodemask will not be
1063 * possible when mems_allowed is larger than a word.
1065 need_loop = task_has_mempolicy(tsk) ||
1066 !nodes_intersects(*newmems, tsk->mems_allowed);
1069 local_irq_disable();
1070 write_seqcount_begin(&tsk->mems_allowed_seq);
1073 nodes_or(tsk->mems_allowed, tsk->mems_allowed, *newmems);
1074 mpol_rebind_task(tsk, newmems, MPOL_REBIND_STEP1);
1076 mpol_rebind_task(tsk, newmems, MPOL_REBIND_STEP2);
1077 tsk->mems_allowed = *newmems;
1080 write_seqcount_end(&tsk->mems_allowed_seq);
1087 static void *cpuset_being_rebound;
1090 * update_tasks_nodemask - Update the nodemasks of tasks in the cpuset.
1091 * @cs: the cpuset in which each task's mems_allowed mask needs to be changed
1093 * Iterate through each task of @cs updating its mems_allowed to the
1094 * effective cpuset's. As this function is called with cpuset_mutex held,
1095 * cpuset membership stays stable.
1097 static void update_tasks_nodemask(struct cpuset *cs)
1099 static nodemask_t newmems; /* protected by cpuset_mutex */
1100 struct css_task_iter it;
1101 struct task_struct *task;
1103 cpuset_being_rebound = cs; /* causes mpol_dup() rebind */
1105 guarantee_online_mems(cs, &newmems);
1108 * The mpol_rebind_mm() call takes mmap_sem, which we couldn't
1109 * take while holding tasklist_lock. Forks can happen - the
1110 * mpol_dup() cpuset_being_rebound check will catch such forks,
1111 * and rebind their vma mempolicies too. Because we still hold
1112 * the global cpuset_mutex, we know that no other rebind effort
1113 * will be contending for the global variable cpuset_being_rebound.
1114 * It's ok if we rebind the same mm twice; mpol_rebind_mm()
1115 * is idempotent. Also migrate pages in each mm to new nodes.
1117 css_task_iter_start(&cs->css, &it);
1118 while ((task = css_task_iter_next(&it))) {
1119 struct mm_struct *mm;
1122 cpuset_change_task_nodemask(task, &newmems);
1124 mm = get_task_mm(task);
1128 migrate = is_memory_migrate(cs);
1130 mpol_rebind_mm(mm, &cs->mems_allowed);
1132 cpuset_migrate_mm(mm, &cs->old_mems_allowed, &newmems);
1136 css_task_iter_end(&it);
1139 * All the tasks' nodemasks have been updated, update
1140 * cs->old_mems_allowed.
1142 cs->old_mems_allowed = newmems;
1144 /* We're done rebinding vmas to this cpuset's new mems_allowed. */
1145 cpuset_being_rebound = NULL;
1149 * update_nodemasks_hier - Update effective nodemasks and tasks in the subtree
1150 * @cs: the cpuset to consider
1151 * @new_mems: a temp variable for calculating new effective_mems
1153 * When configured nodemask is changed, the effective nodemasks of this cpuset
1154 * and all its descendants need to be updated.
1156 * On legacy hiearchy, effective_mems will be the same with mems_allowed.
1158 * Called with cpuset_mutex held
1160 static void update_nodemasks_hier(struct cpuset *cs, nodemask_t *new_mems)
1163 struct cgroup_subsys_state *pos_css;
1166 cpuset_for_each_descendant_pre(cp, pos_css, cs) {
1167 struct cpuset *parent = parent_cs(cp);
1169 nodes_and(*new_mems, cp->mems_allowed, parent->effective_mems);
1172 * If it becomes empty, inherit the effective mask of the
1173 * parent, which is guaranteed to have some MEMs.
1175 if (cgroup_subsys_on_dfl(cpuset_cgrp_subsys) &&
1176 nodes_empty(*new_mems))
1177 *new_mems = parent->effective_mems;
1179 /* Skip the whole subtree if the nodemask remains the same. */
1180 if (nodes_equal(*new_mems, cp->effective_mems)) {
1181 pos_css = css_rightmost_descendant(pos_css);
1185 if (!css_tryget_online(&cp->css))
1189 spin_lock_irq(&callback_lock);
1190 cp->effective_mems = *new_mems;
1191 spin_unlock_irq(&callback_lock);
1193 WARN_ON(!cgroup_subsys_on_dfl(cpuset_cgrp_subsys) &&
1194 !nodes_equal(cp->mems_allowed, cp->effective_mems));
1196 update_tasks_nodemask(cp);
1205 * Handle user request to change the 'mems' memory placement
1206 * of a cpuset. Needs to validate the request, update the
1207 * cpusets mems_allowed, and for each task in the cpuset,
1208 * update mems_allowed and rebind task's mempolicy and any vma
1209 * mempolicies and if the cpuset is marked 'memory_migrate',
1210 * migrate the tasks pages to the new memory.
1212 * Call with cpuset_mutex held. May take callback_lock during call.
1213 * Will take tasklist_lock, scan tasklist for tasks in cpuset cs,
1214 * lock each such tasks mm->mmap_sem, scan its vma's and rebind
1215 * their mempolicies to the cpusets new mems_allowed.
1217 static int update_nodemask(struct cpuset *cs, struct cpuset *trialcs,
1223 * top_cpuset.mems_allowed tracks node_stats[N_MEMORY];
1226 if (cs == &top_cpuset) {
1232 * An empty mems_allowed is ok iff there are no tasks in the cpuset.
1233 * Since nodelist_parse() fails on an empty mask, we special case
1234 * that parsing. The validate_change() call ensures that cpusets
1235 * with tasks have memory.
1238 nodes_clear(trialcs->mems_allowed);
1240 retval = nodelist_parse(buf, trialcs->mems_allowed);
1244 if (!nodes_subset(trialcs->mems_allowed,
1245 top_cpuset.mems_allowed)) {
1251 if (nodes_equal(cs->mems_allowed, trialcs->mems_allowed)) {
1252 retval = 0; /* Too easy - nothing to do */
1255 retval = validate_change(cs, trialcs);
1259 spin_lock_irq(&callback_lock);
1260 cs->mems_allowed = trialcs->mems_allowed;
1261 spin_unlock_irq(&callback_lock);
1263 /* use trialcs->mems_allowed as a temp variable */
1264 update_nodemasks_hier(cs, &trialcs->mems_allowed);
1269 int current_cpuset_is_being_rebound(void)
1274 ret = task_cs(current) == cpuset_being_rebound;
1280 static int update_relax_domain_level(struct cpuset *cs, s64 val)
1283 if (val < -1 || val >= sched_domain_level_max)
1287 if (val != cs->relax_domain_level) {
1288 cs->relax_domain_level = val;
1289 if (!cpumask_empty(cs->cpus_allowed) &&
1290 is_sched_load_balance(cs))
1291 rebuild_sched_domains_unlocked();
1298 * update_tasks_flags - update the spread flags of tasks in the cpuset.
1299 * @cs: the cpuset in which each task's spread flags needs to be changed
1301 * Iterate through each task of @cs updating its spread flags. As this
1302 * function is called with cpuset_mutex held, cpuset membership stays
1305 static void update_tasks_flags(struct cpuset *cs)
1307 struct css_task_iter it;
1308 struct task_struct *task;
1310 css_task_iter_start(&cs->css, &it);
1311 while ((task = css_task_iter_next(&it)))
1312 cpuset_update_task_spread_flag(cs, task);
1313 css_task_iter_end(&it);
1317 * update_flag - read a 0 or a 1 in a file and update associated flag
1318 * bit: the bit to update (see cpuset_flagbits_t)
1319 * cs: the cpuset to update
1320 * turning_on: whether the flag is being set or cleared
1324 static int update_flag(cpuset_flagbits_t bit, struct cpuset *cs,
1327 struct cpuset *trialcs;
1328 int balance_flag_changed;
1329 int spread_flag_changed;
1332 trialcs = alloc_trial_cpuset(cs);
1337 set_bit(bit, &trialcs->flags);
1339 clear_bit(bit, &trialcs->flags);
1341 err = validate_change(cs, trialcs);
1345 balance_flag_changed = (is_sched_load_balance(cs) !=
1346 is_sched_load_balance(trialcs));
1348 spread_flag_changed = ((is_spread_slab(cs) != is_spread_slab(trialcs))
1349 || (is_spread_page(cs) != is_spread_page(trialcs)));
1351 spin_lock_irq(&callback_lock);
1352 cs->flags = trialcs->flags;
1353 spin_unlock_irq(&callback_lock);
1355 if (!cpumask_empty(trialcs->cpus_allowed) && balance_flag_changed)
1356 rebuild_sched_domains_unlocked();
1358 if (spread_flag_changed)
1359 update_tasks_flags(cs);
1361 free_trial_cpuset(trialcs);
1366 * Frequency meter - How fast is some event occurring?
1368 * These routines manage a digitally filtered, constant time based,
1369 * event frequency meter. There are four routines:
1370 * fmeter_init() - initialize a frequency meter.
1371 * fmeter_markevent() - called each time the event happens.
1372 * fmeter_getrate() - returns the recent rate of such events.
1373 * fmeter_update() - internal routine used to update fmeter.
1375 * A common data structure is passed to each of these routines,
1376 * which is used to keep track of the state required to manage the
1377 * frequency meter and its digital filter.
1379 * The filter works on the number of events marked per unit time.
1380 * The filter is single-pole low-pass recursive (IIR). The time unit
1381 * is 1 second. Arithmetic is done using 32-bit integers scaled to
1382 * simulate 3 decimal digits of precision (multiplied by 1000).
1384 * With an FM_COEF of 933, and a time base of 1 second, the filter
1385 * has a half-life of 10 seconds, meaning that if the events quit
1386 * happening, then the rate returned from the fmeter_getrate()
1387 * will be cut in half each 10 seconds, until it converges to zero.
1389 * It is not worth doing a real infinitely recursive filter. If more
1390 * than FM_MAXTICKS ticks have elapsed since the last filter event,
1391 * just compute FM_MAXTICKS ticks worth, by which point the level
1394 * Limit the count of unprocessed events to FM_MAXCNT, so as to avoid
1395 * arithmetic overflow in the fmeter_update() routine.
1397 * Given the simple 32 bit integer arithmetic used, this meter works
1398 * best for reporting rates between one per millisecond (msec) and
1399 * one per 32 (approx) seconds. At constant rates faster than one
1400 * per msec it maxes out at values just under 1,000,000. At constant
1401 * rates between one per msec, and one per second it will stabilize
1402 * to a value N*1000, where N is the rate of events per second.
1403 * At constant rates between one per second and one per 32 seconds,
1404 * it will be choppy, moving up on the seconds that have an event,
1405 * and then decaying until the next event. At rates slower than
1406 * about one in 32 seconds, it decays all the way back to zero between
1410 #define FM_COEF 933 /* coefficient for half-life of 10 secs */
1411 #define FM_MAXTICKS ((time_t)99) /* useless computing more ticks than this */
1412 #define FM_MAXCNT 1000000 /* limit cnt to avoid overflow */
1413 #define FM_SCALE 1000 /* faux fixed point scale */
1415 /* Initialize a frequency meter */
1416 static void fmeter_init(struct fmeter *fmp)
1421 spin_lock_init(&fmp->lock);
1424 /* Internal meter update - process cnt events and update value */
1425 static void fmeter_update(struct fmeter *fmp)
1427 time_t now = get_seconds();
1428 time_t ticks = now - fmp->time;
1433 ticks = min(FM_MAXTICKS, ticks);
1435 fmp->val = (FM_COEF * fmp->val) / FM_SCALE;
1438 fmp->val += ((FM_SCALE - FM_COEF) * fmp->cnt) / FM_SCALE;
1442 /* Process any previous ticks, then bump cnt by one (times scale). */
1443 static void fmeter_markevent(struct fmeter *fmp)
1445 spin_lock(&fmp->lock);
1447 fmp->cnt = min(FM_MAXCNT, fmp->cnt + FM_SCALE);
1448 spin_unlock(&fmp->lock);
1451 /* Process any previous ticks, then return current value. */
1452 static int fmeter_getrate(struct fmeter *fmp)
1456 spin_lock(&fmp->lock);
1459 spin_unlock(&fmp->lock);
1463 static struct cpuset *cpuset_attach_old_cs;
1465 /* Called by cgroups to determine if a cpuset is usable; cpuset_mutex held */
1466 static int cpuset_can_attach(struct cgroup_taskset *tset)
1468 struct cgroup_subsys_state *css;
1470 struct task_struct *task;
1473 /* used later by cpuset_attach() */
1474 cpuset_attach_old_cs = task_cs(cgroup_taskset_first(tset, &css));
1477 mutex_lock(&cpuset_mutex);
1479 /* allow moving tasks into an empty cpuset if on default hierarchy */
1481 if (!cgroup_subsys_on_dfl(cpuset_cgrp_subsys) &&
1482 (cpumask_empty(cs->cpus_allowed) || nodes_empty(cs->mems_allowed)))
1485 cgroup_taskset_for_each(task, css, tset) {
1486 ret = task_can_attach(task, cs->cpus_allowed);
1489 ret = security_task_setscheduler(task);
1495 * Mark attach is in progress. This makes validate_change() fail
1496 * changes which zero cpus/mems_allowed.
1498 cs->attach_in_progress++;
1501 mutex_unlock(&cpuset_mutex);
1505 static void cpuset_cancel_attach(struct cgroup_taskset *tset)
1507 struct cgroup_subsys_state *css;
1510 cgroup_taskset_first(tset, &css);
1513 mutex_lock(&cpuset_mutex);
1514 css_cs(css)->attach_in_progress--;
1515 mutex_unlock(&cpuset_mutex);
1519 * Protected by cpuset_mutex. cpus_attach is used only by cpuset_attach()
1520 * but we can't allocate it dynamically there. Define it global and
1521 * allocate from cpuset_init().
1523 static cpumask_var_t cpus_attach;
1525 static void cpuset_attach(struct cgroup_taskset *tset)
1527 /* static buf protected by cpuset_mutex */
1528 static nodemask_t cpuset_attach_nodemask_to;
1529 struct task_struct *task;
1530 struct task_struct *leader;
1531 struct cgroup_subsys_state *css;
1533 struct cpuset *oldcs = cpuset_attach_old_cs;
1535 cgroup_taskset_first(tset, &css);
1538 mutex_lock(&cpuset_mutex);
1540 /* prepare for attach */
1541 if (cs == &top_cpuset)
1542 cpumask_copy(cpus_attach, cpu_possible_mask);
1544 guarantee_online_cpus(cs, cpus_attach);
1546 guarantee_online_mems(cs, &cpuset_attach_nodemask_to);
1548 cgroup_taskset_for_each(task, css, tset) {
1550 * can_attach beforehand should guarantee that this doesn't
1551 * fail. TODO: have a better way to handle failure here
1553 WARN_ON_ONCE(set_cpus_allowed_ptr(task, cpus_attach));
1555 cpuset_change_task_nodemask(task, &cpuset_attach_nodemask_to);
1556 cpuset_update_task_spread_flag(cs, task);
1560 * Change mm for all threadgroup leaders. This is expensive and may
1561 * sleep and should be moved outside migration path proper.
1563 cpuset_attach_nodemask_to = cs->effective_mems;
1564 cgroup_taskset_for_each_leader(leader, css, tset) {
1565 struct mm_struct *mm = get_task_mm(leader);
1568 mpol_rebind_mm(mm, &cpuset_attach_nodemask_to);
1571 * old_mems_allowed is the same with mems_allowed
1572 * here, except if this task is being moved
1573 * automatically due to hotplug. In that case
1574 * @mems_allowed has been updated and is empty, so
1575 * @old_mems_allowed is the right nodesets that we
1578 if (is_memory_migrate(cs))
1579 cpuset_migrate_mm(mm, &oldcs->old_mems_allowed,
1580 &cpuset_attach_nodemask_to);
1586 cs->old_mems_allowed = cpuset_attach_nodemask_to;
1588 cs->attach_in_progress--;
1589 if (!cs->attach_in_progress)
1590 wake_up(&cpuset_attach_wq);
1592 mutex_unlock(&cpuset_mutex);
1595 /* The various types of files and directories in a cpuset file system */
1598 FILE_MEMORY_MIGRATE,
1601 FILE_EFFECTIVE_CPULIST,
1602 FILE_EFFECTIVE_MEMLIST,
1606 FILE_SCHED_LOAD_BALANCE,
1607 FILE_SCHED_RELAX_DOMAIN_LEVEL,
1608 FILE_MEMORY_PRESSURE_ENABLED,
1609 FILE_MEMORY_PRESSURE,
1612 } cpuset_filetype_t;
1614 static int cpuset_write_u64(struct cgroup_subsys_state *css, struct cftype *cft,
1617 struct cpuset *cs = css_cs(css);
1618 cpuset_filetype_t type = cft->private;
1622 mutex_lock(&cpuset_mutex);
1623 if (!is_cpuset_online(cs)) {
1629 case FILE_CPU_EXCLUSIVE:
1630 retval = update_flag(CS_CPU_EXCLUSIVE, cs, val);
1632 case FILE_MEM_EXCLUSIVE:
1633 retval = update_flag(CS_MEM_EXCLUSIVE, cs, val);
1635 case FILE_MEM_HARDWALL:
1636 retval = update_flag(CS_MEM_HARDWALL, cs, val);
1638 case FILE_SCHED_LOAD_BALANCE:
1639 retval = update_flag(CS_SCHED_LOAD_BALANCE, cs, val);
1641 case FILE_MEMORY_MIGRATE:
1642 retval = update_flag(CS_MEMORY_MIGRATE, cs, val);
1644 case FILE_MEMORY_PRESSURE_ENABLED:
1645 cpuset_memory_pressure_enabled = !!val;
1647 case FILE_SPREAD_PAGE:
1648 retval = update_flag(CS_SPREAD_PAGE, cs, val);
1650 case FILE_SPREAD_SLAB:
1651 retval = update_flag(CS_SPREAD_SLAB, cs, val);
1658 mutex_unlock(&cpuset_mutex);
1663 static int cpuset_write_s64(struct cgroup_subsys_state *css, struct cftype *cft,
1666 struct cpuset *cs = css_cs(css);
1667 cpuset_filetype_t type = cft->private;
1668 int retval = -ENODEV;
1671 mutex_lock(&cpuset_mutex);
1672 if (!is_cpuset_online(cs))
1676 case FILE_SCHED_RELAX_DOMAIN_LEVEL:
1677 retval = update_relax_domain_level(cs, val);
1684 mutex_unlock(&cpuset_mutex);
1690 * Common handling for a write to a "cpus" or "mems" file.
1692 static ssize_t cpuset_write_resmask(struct kernfs_open_file *of,
1693 char *buf, size_t nbytes, loff_t off)
1695 struct cpuset *cs = css_cs(of_css(of));
1696 struct cpuset *trialcs;
1697 int retval = -ENODEV;
1699 buf = strstrip(buf);
1702 * CPU or memory hotunplug may leave @cs w/o any execution
1703 * resources, in which case the hotplug code asynchronously updates
1704 * configuration and transfers all tasks to the nearest ancestor
1705 * which can execute.
1707 * As writes to "cpus" or "mems" may restore @cs's execution
1708 * resources, wait for the previously scheduled operations before
1709 * proceeding, so that we don't end up keep removing tasks added
1710 * after execution capability is restored.
1712 * cpuset_hotplug_work calls back into cgroup core via
1713 * cgroup_transfer_tasks() and waiting for it from a cgroupfs
1714 * operation like this one can lead to a deadlock through kernfs
1715 * active_ref protection. Let's break the protection. Losing the
1716 * protection is okay as we check whether @cs is online after
1717 * grabbing cpuset_mutex anyway. This only happens on the legacy
1721 kernfs_break_active_protection(of->kn);
1722 flush_work(&cpuset_hotplug_work);
1725 mutex_lock(&cpuset_mutex);
1726 if (!is_cpuset_online(cs))
1729 trialcs = alloc_trial_cpuset(cs);
1735 switch (of_cft(of)->private) {
1737 retval = update_cpumask(cs, trialcs, buf);
1740 retval = update_nodemask(cs, trialcs, buf);
1747 free_trial_cpuset(trialcs);
1749 mutex_unlock(&cpuset_mutex);
1751 kernfs_unbreak_active_protection(of->kn);
1753 flush_workqueue(cpuset_migrate_mm_wq);
1754 return retval ?: nbytes;
1758 * These ascii lists should be read in a single call, by using a user
1759 * buffer large enough to hold the entire map. If read in smaller
1760 * chunks, there is no guarantee of atomicity. Since the display format
1761 * used, list of ranges of sequential numbers, is variable length,
1762 * and since these maps can change value dynamically, one could read
1763 * gibberish by doing partial reads while a list was changing.
1765 static int cpuset_common_seq_show(struct seq_file *sf, void *v)
1767 struct cpuset *cs = css_cs(seq_css(sf));
1768 cpuset_filetype_t type = seq_cft(sf)->private;
1771 spin_lock_irq(&callback_lock);
1775 seq_printf(sf, "%*pbl\n", cpumask_pr_args(cs->cpus_requested));
1778 seq_printf(sf, "%*pbl\n", nodemask_pr_args(&cs->mems_allowed));
1780 case FILE_EFFECTIVE_CPULIST:
1781 seq_printf(sf, "%*pbl\n", cpumask_pr_args(cs->effective_cpus));
1783 case FILE_EFFECTIVE_MEMLIST:
1784 seq_printf(sf, "%*pbl\n", nodemask_pr_args(&cs->effective_mems));
1790 spin_unlock_irq(&callback_lock);
1794 static u64 cpuset_read_u64(struct cgroup_subsys_state *css, struct cftype *cft)
1796 struct cpuset *cs = css_cs(css);
1797 cpuset_filetype_t type = cft->private;
1799 case FILE_CPU_EXCLUSIVE:
1800 return is_cpu_exclusive(cs);
1801 case FILE_MEM_EXCLUSIVE:
1802 return is_mem_exclusive(cs);
1803 case FILE_MEM_HARDWALL:
1804 return is_mem_hardwall(cs);
1805 case FILE_SCHED_LOAD_BALANCE:
1806 return is_sched_load_balance(cs);
1807 case FILE_MEMORY_MIGRATE:
1808 return is_memory_migrate(cs);
1809 case FILE_MEMORY_PRESSURE_ENABLED:
1810 return cpuset_memory_pressure_enabled;
1811 case FILE_MEMORY_PRESSURE:
1812 return fmeter_getrate(&cs->fmeter);
1813 case FILE_SPREAD_PAGE:
1814 return is_spread_page(cs);
1815 case FILE_SPREAD_SLAB:
1816 return is_spread_slab(cs);
1821 /* Unreachable but makes gcc happy */
1825 static s64 cpuset_read_s64(struct cgroup_subsys_state *css, struct cftype *cft)
1827 struct cpuset *cs = css_cs(css);
1828 cpuset_filetype_t type = cft->private;
1830 case FILE_SCHED_RELAX_DOMAIN_LEVEL:
1831 return cs->relax_domain_level;
1836 /* Unrechable but makes gcc happy */
1842 * for the common functions, 'private' gives the type of file
1845 static struct cftype files[] = {
1848 .seq_show = cpuset_common_seq_show,
1849 .write = cpuset_write_resmask,
1850 .max_write_len = (100U + 6 * NR_CPUS),
1851 .private = FILE_CPULIST,
1856 .seq_show = cpuset_common_seq_show,
1857 .write = cpuset_write_resmask,
1858 .max_write_len = (100U + 6 * MAX_NUMNODES),
1859 .private = FILE_MEMLIST,
1863 .name = "effective_cpus",
1864 .seq_show = cpuset_common_seq_show,
1865 .private = FILE_EFFECTIVE_CPULIST,
1869 .name = "effective_mems",
1870 .seq_show = cpuset_common_seq_show,
1871 .private = FILE_EFFECTIVE_MEMLIST,
1875 .name = "cpu_exclusive",
1876 .read_u64 = cpuset_read_u64,
1877 .write_u64 = cpuset_write_u64,
1878 .private = FILE_CPU_EXCLUSIVE,
1882 .name = "mem_exclusive",
1883 .read_u64 = cpuset_read_u64,
1884 .write_u64 = cpuset_write_u64,
1885 .private = FILE_MEM_EXCLUSIVE,
1889 .name = "mem_hardwall",
1890 .read_u64 = cpuset_read_u64,
1891 .write_u64 = cpuset_write_u64,
1892 .private = FILE_MEM_HARDWALL,
1896 .name = "sched_load_balance",
1897 .read_u64 = cpuset_read_u64,
1898 .write_u64 = cpuset_write_u64,
1899 .private = FILE_SCHED_LOAD_BALANCE,
1903 .name = "sched_relax_domain_level",
1904 .read_s64 = cpuset_read_s64,
1905 .write_s64 = cpuset_write_s64,
1906 .private = FILE_SCHED_RELAX_DOMAIN_LEVEL,
1910 .name = "memory_migrate",
1911 .read_u64 = cpuset_read_u64,
1912 .write_u64 = cpuset_write_u64,
1913 .private = FILE_MEMORY_MIGRATE,
1917 .name = "memory_pressure",
1918 .read_u64 = cpuset_read_u64,
1919 .private = FILE_MEMORY_PRESSURE,
1923 .name = "memory_spread_page",
1924 .read_u64 = cpuset_read_u64,
1925 .write_u64 = cpuset_write_u64,
1926 .private = FILE_SPREAD_PAGE,
1930 .name = "memory_spread_slab",
1931 .read_u64 = cpuset_read_u64,
1932 .write_u64 = cpuset_write_u64,
1933 .private = FILE_SPREAD_SLAB,
1937 .name = "memory_pressure_enabled",
1938 .flags = CFTYPE_ONLY_ON_ROOT,
1939 .read_u64 = cpuset_read_u64,
1940 .write_u64 = cpuset_write_u64,
1941 .private = FILE_MEMORY_PRESSURE_ENABLED,
1948 * cpuset_css_alloc - allocate a cpuset css
1949 * cgrp: control group that the new cpuset will be part of
1952 static struct cgroup_subsys_state *
1953 cpuset_css_alloc(struct cgroup_subsys_state *parent_css)
1958 return &top_cpuset.css;
1960 cs = kzalloc(sizeof(*cs), GFP_KERNEL);
1962 return ERR_PTR(-ENOMEM);
1963 if (!alloc_cpumask_var(&cs->cpus_allowed, GFP_KERNEL))
1965 if (!alloc_cpumask_var(&cs->effective_cpus, GFP_KERNEL))
1966 goto error_effective;
1967 if (!alloc_cpumask_var(&cs->cpus_requested, GFP_KERNEL))
1968 goto error_requested;
1970 set_bit(CS_SCHED_LOAD_BALANCE, &cs->flags);
1971 cpumask_clear(cs->cpus_allowed);
1972 cpumask_clear(cs->cpus_requested);
1973 nodes_clear(cs->mems_allowed);
1974 cpumask_clear(cs->effective_cpus);
1975 nodes_clear(cs->effective_mems);
1976 fmeter_init(&cs->fmeter);
1977 cs->relax_domain_level = -1;
1982 free_cpumask_var(cs->effective_cpus);
1984 free_cpumask_var(cs->cpus_allowed);
1987 return ERR_PTR(-ENOMEM);
1990 static int cpuset_css_online(struct cgroup_subsys_state *css)
1992 struct cpuset *cs = css_cs(css);
1993 struct cpuset *parent = parent_cs(cs);
1994 struct cpuset *tmp_cs;
1995 struct cgroup_subsys_state *pos_css;
2000 mutex_lock(&cpuset_mutex);
2002 set_bit(CS_ONLINE, &cs->flags);
2003 if (is_spread_page(parent))
2004 set_bit(CS_SPREAD_PAGE, &cs->flags);
2005 if (is_spread_slab(parent))
2006 set_bit(CS_SPREAD_SLAB, &cs->flags);
2010 spin_lock_irq(&callback_lock);
2011 if (cgroup_subsys_on_dfl(cpuset_cgrp_subsys)) {
2012 cpumask_copy(cs->effective_cpus, parent->effective_cpus);
2013 cs->effective_mems = parent->effective_mems;
2015 spin_unlock_irq(&callback_lock);
2017 if (!test_bit(CGRP_CPUSET_CLONE_CHILDREN, &css->cgroup->flags))
2021 * Clone @parent's configuration if CGRP_CPUSET_CLONE_CHILDREN is
2022 * set. This flag handling is implemented in cgroup core for
2023 * histrical reasons - the flag may be specified during mount.
2025 * Currently, if any sibling cpusets have exclusive cpus or mem, we
2026 * refuse to clone the configuration - thereby refusing the task to
2027 * be entered, and as a result refusing the sys_unshare() or
2028 * clone() which initiated it. If this becomes a problem for some
2029 * users who wish to allow that scenario, then this could be
2030 * changed to grant parent->cpus_allowed-sibling_cpus_exclusive
2031 * (and likewise for mems) to the new cgroup.
2034 cpuset_for_each_child(tmp_cs, pos_css, parent) {
2035 if (is_mem_exclusive(tmp_cs) || is_cpu_exclusive(tmp_cs)) {
2042 spin_lock_irq(&callback_lock);
2043 cs->mems_allowed = parent->mems_allowed;
2044 cs->effective_mems = parent->mems_allowed;
2045 cpumask_copy(cs->cpus_allowed, parent->cpus_allowed);
2046 cpumask_copy(cs->cpus_requested, parent->cpus_requested);
2047 cpumask_copy(cs->effective_cpus, parent->cpus_allowed);
2048 spin_unlock_irq(&callback_lock);
2050 mutex_unlock(&cpuset_mutex);
2055 * If the cpuset being removed has its flag 'sched_load_balance'
2056 * enabled, then simulate turning sched_load_balance off, which
2057 * will call rebuild_sched_domains_unlocked().
2060 static void cpuset_css_offline(struct cgroup_subsys_state *css)
2062 struct cpuset *cs = css_cs(css);
2065 mutex_lock(&cpuset_mutex);
2067 if (is_sched_load_balance(cs))
2068 update_flag(CS_SCHED_LOAD_BALANCE, cs, 0);
2071 clear_bit(CS_ONLINE, &cs->flags);
2073 mutex_unlock(&cpuset_mutex);
2077 static void cpuset_css_free(struct cgroup_subsys_state *css)
2079 struct cpuset *cs = css_cs(css);
2081 free_cpumask_var(cs->effective_cpus);
2082 free_cpumask_var(cs->cpus_allowed);
2083 free_cpumask_var(cs->cpus_requested);
2087 static void cpuset_bind(struct cgroup_subsys_state *root_css)
2089 mutex_lock(&cpuset_mutex);
2090 spin_lock_irq(&callback_lock);
2092 if (cgroup_subsys_on_dfl(cpuset_cgrp_subsys)) {
2093 cpumask_copy(top_cpuset.cpus_allowed, cpu_possible_mask);
2094 top_cpuset.mems_allowed = node_possible_map;
2096 cpumask_copy(top_cpuset.cpus_allowed,
2097 top_cpuset.effective_cpus);
2098 top_cpuset.mems_allowed = top_cpuset.effective_mems;
2101 spin_unlock_irq(&callback_lock);
2102 mutex_unlock(&cpuset_mutex);
2106 * Make sure the new task conform to the current state of its parent,
2107 * which could have been changed by cpuset just after it inherits the
2108 * state from the parent and before it sits on the cgroup's task list.
2110 void cpuset_fork(struct task_struct *task, void *priv)
2112 if (task_css_is_root(task, cpuset_cgrp_id))
2115 set_cpus_allowed_ptr(task, ¤t->cpus_allowed);
2116 task->mems_allowed = current->mems_allowed;
2119 struct cgroup_subsys cpuset_cgrp_subsys = {
2120 .css_alloc = cpuset_css_alloc,
2121 .css_online = cpuset_css_online,
2122 .css_offline = cpuset_css_offline,
2123 .css_free = cpuset_css_free,
2124 .can_attach = cpuset_can_attach,
2125 .cancel_attach = cpuset_cancel_attach,
2126 .attach = cpuset_attach,
2127 .post_attach = cpuset_post_attach,
2128 .bind = cpuset_bind,
2129 .fork = cpuset_fork,
2130 .legacy_cftypes = files,
2135 * cpuset_init - initialize cpusets at system boot
2137 * Description: Initialize top_cpuset and the cpuset internal file system,
2140 int __init cpuset_init(void)
2144 if (!alloc_cpumask_var(&top_cpuset.cpus_allowed, GFP_KERNEL))
2146 if (!alloc_cpumask_var(&top_cpuset.effective_cpus, GFP_KERNEL))
2148 if (!alloc_cpumask_var(&top_cpuset.cpus_requested, GFP_KERNEL))
2151 cpumask_setall(top_cpuset.cpus_allowed);
2152 cpumask_setall(top_cpuset.cpus_requested);
2153 nodes_setall(top_cpuset.mems_allowed);
2154 cpumask_setall(top_cpuset.effective_cpus);
2155 nodes_setall(top_cpuset.effective_mems);
2157 fmeter_init(&top_cpuset.fmeter);
2158 set_bit(CS_SCHED_LOAD_BALANCE, &top_cpuset.flags);
2159 top_cpuset.relax_domain_level = -1;
2161 err = register_filesystem(&cpuset_fs_type);
2165 if (!alloc_cpumask_var(&cpus_attach, GFP_KERNEL))
2172 * If CPU and/or memory hotplug handlers, below, unplug any CPUs
2173 * or memory nodes, we need to walk over the cpuset hierarchy,
2174 * removing that CPU or node from all cpusets. If this removes the
2175 * last CPU or node from a cpuset, then move the tasks in the empty
2176 * cpuset to its next-highest non-empty parent.
2178 static void remove_tasks_in_empty_cpuset(struct cpuset *cs)
2180 struct cpuset *parent;
2183 * Find its next-highest non-empty parent, (top cpuset
2184 * has online cpus, so can't be empty).
2186 parent = parent_cs(cs);
2187 while (cpumask_empty(parent->cpus_allowed) ||
2188 nodes_empty(parent->mems_allowed))
2189 parent = parent_cs(parent);
2191 if (cgroup_transfer_tasks(parent->css.cgroup, cs->css.cgroup)) {
2192 pr_err("cpuset: failed to transfer tasks out of empty cpuset ");
2193 pr_cont_cgroup_name(cs->css.cgroup);
2199 hotplug_update_tasks_legacy(struct cpuset *cs,
2200 struct cpumask *new_cpus, nodemask_t *new_mems,
2201 bool cpus_updated, bool mems_updated)
2205 spin_lock_irq(&callback_lock);
2206 cpumask_copy(cs->cpus_allowed, new_cpus);
2207 cpumask_copy(cs->effective_cpus, new_cpus);
2208 cs->mems_allowed = *new_mems;
2209 cs->effective_mems = *new_mems;
2210 spin_unlock_irq(&callback_lock);
2213 * Don't call update_tasks_cpumask() if the cpuset becomes empty,
2214 * as the tasks will be migratecd to an ancestor.
2216 if (cpus_updated && !cpumask_empty(cs->cpus_allowed))
2217 update_tasks_cpumask(cs);
2218 if (mems_updated && !nodes_empty(cs->mems_allowed))
2219 update_tasks_nodemask(cs);
2221 is_empty = cpumask_empty(cs->cpus_allowed) ||
2222 nodes_empty(cs->mems_allowed);
2224 mutex_unlock(&cpuset_mutex);
2227 * Move tasks to the nearest ancestor with execution resources,
2228 * This is full cgroup operation which will also call back into
2229 * cpuset. Should be done outside any lock.
2232 remove_tasks_in_empty_cpuset(cs);
2234 mutex_lock(&cpuset_mutex);
2238 hotplug_update_tasks(struct cpuset *cs,
2239 struct cpumask *new_cpus, nodemask_t *new_mems,
2240 bool cpus_updated, bool mems_updated)
2242 if (cpumask_empty(new_cpus))
2243 cpumask_copy(new_cpus, parent_cs(cs)->effective_cpus);
2244 if (nodes_empty(*new_mems))
2245 *new_mems = parent_cs(cs)->effective_mems;
2247 spin_lock_irq(&callback_lock);
2248 cpumask_copy(cs->effective_cpus, new_cpus);
2249 cs->effective_mems = *new_mems;
2250 spin_unlock_irq(&callback_lock);
2253 update_tasks_cpumask(cs);
2255 update_tasks_nodemask(cs);
2259 * cpuset_hotplug_update_tasks - update tasks in a cpuset for hotunplug
2260 * @cs: cpuset in interest
2262 * Compare @cs's cpu and mem masks against top_cpuset and if some have gone
2263 * offline, update @cs accordingly. If @cs ends up with no CPU or memory,
2264 * all its tasks are moved to the nearest ancestor with both resources.
2266 static void cpuset_hotplug_update_tasks(struct cpuset *cs)
2268 static cpumask_t new_cpus;
2269 static nodemask_t new_mems;
2273 wait_event(cpuset_attach_wq, cs->attach_in_progress == 0);
2275 mutex_lock(&cpuset_mutex);
2278 * We have raced with task attaching. We wait until attaching
2279 * is finished, so we won't attach a task to an empty cpuset.
2281 if (cs->attach_in_progress) {
2282 mutex_unlock(&cpuset_mutex);
2286 cpumask_and(&new_cpus, cs->cpus_requested,
2287 parent_cs(cs)->effective_cpus);
2288 nodes_and(new_mems, cs->mems_allowed, parent_cs(cs)->effective_mems);
2290 cpus_updated = !cpumask_equal(&new_cpus, cs->effective_cpus);
2291 mems_updated = !nodes_equal(new_mems, cs->effective_mems);
2293 if (cgroup_subsys_on_dfl(cpuset_cgrp_subsys))
2294 hotplug_update_tasks(cs, &new_cpus, &new_mems,
2295 cpus_updated, mems_updated);
2297 hotplug_update_tasks_legacy(cs, &new_cpus, &new_mems,
2298 cpus_updated, mems_updated);
2300 mutex_unlock(&cpuset_mutex);
2303 static bool force_rebuild;
2305 void cpuset_force_rebuild(void)
2307 force_rebuild = true;
2311 * cpuset_hotplug_workfn - handle CPU/memory hotunplug for a cpuset
2313 * This function is called after either CPU or memory configuration has
2314 * changed and updates cpuset accordingly. The top_cpuset is always
2315 * synchronized to cpu_active_mask and N_MEMORY, which is necessary in
2316 * order to make cpusets transparent (of no affect) on systems that are
2317 * actively using CPU hotplug but making no active use of cpusets.
2319 * Non-root cpusets are only affected by offlining. If any CPUs or memory
2320 * nodes have been taken down, cpuset_hotplug_update_tasks() is invoked on
2323 * Note that CPU offlining during suspend is ignored. We don't modify
2324 * cpusets across suspend/resume cycles at all.
2326 static void cpuset_hotplug_workfn(struct work_struct *work)
2328 static cpumask_t new_cpus;
2329 static nodemask_t new_mems;
2330 bool cpus_updated, mems_updated;
2331 bool on_dfl = cgroup_subsys_on_dfl(cpuset_cgrp_subsys);
2333 mutex_lock(&cpuset_mutex);
2335 /* fetch the available cpus/mems and find out which changed how */
2336 cpumask_copy(&new_cpus, cpu_active_mask);
2337 new_mems = node_states[N_MEMORY];
2339 cpus_updated = !cpumask_equal(top_cpuset.effective_cpus, &new_cpus);
2340 mems_updated = !nodes_equal(top_cpuset.effective_mems, new_mems);
2342 /* synchronize cpus_allowed to cpu_active_mask */
2344 spin_lock_irq(&callback_lock);
2346 cpumask_copy(top_cpuset.cpus_allowed, &new_cpus);
2347 cpumask_copy(top_cpuset.effective_cpus, &new_cpus);
2348 spin_unlock_irq(&callback_lock);
2349 /* we don't mess with cpumasks of tasks in top_cpuset */
2352 /* synchronize mems_allowed to N_MEMORY */
2354 spin_lock_irq(&callback_lock);
2356 top_cpuset.mems_allowed = new_mems;
2357 top_cpuset.effective_mems = new_mems;
2358 spin_unlock_irq(&callback_lock);
2359 update_tasks_nodemask(&top_cpuset);
2362 mutex_unlock(&cpuset_mutex);
2364 /* if cpus or mems changed, we need to propagate to descendants */
2365 if (cpus_updated || mems_updated) {
2367 struct cgroup_subsys_state *pos_css;
2370 cpuset_for_each_descendant_pre(cs, pos_css, &top_cpuset) {
2371 if (cs == &top_cpuset || !css_tryget_online(&cs->css))
2375 cpuset_hotplug_update_tasks(cs);
2383 /* rebuild sched domains if cpus_allowed has changed */
2384 if (cpus_updated || force_rebuild) {
2385 force_rebuild = false;
2386 rebuild_sched_domains();
2390 void cpuset_update_active_cpus(bool cpu_online)
2393 * We're inside cpu hotplug critical region which usually nests
2394 * inside cgroup synchronization. Bounce actual hotplug processing
2395 * to a work item to avoid reverse locking order.
2397 * We still need to do partition_sched_domains() synchronously;
2398 * otherwise, the scheduler will get confused and put tasks to the
2399 * dead CPU. Fall back to the default single domain.
2400 * cpuset_hotplug_workfn() will rebuild it as necessary.
2402 partition_sched_domains(1, NULL, NULL);
2403 schedule_work(&cpuset_hotplug_work);
2406 void cpuset_wait_for_hotplug(void)
2408 flush_work(&cpuset_hotplug_work);
2412 * Keep top_cpuset.mems_allowed tracking node_states[N_MEMORY].
2413 * Call this routine anytime after node_states[N_MEMORY] changes.
2414 * See cpuset_update_active_cpus() for CPU hotplug handling.
2416 static int cpuset_track_online_nodes(struct notifier_block *self,
2417 unsigned long action, void *arg)
2419 schedule_work(&cpuset_hotplug_work);
2423 static struct notifier_block cpuset_track_online_nodes_nb = {
2424 .notifier_call = cpuset_track_online_nodes,
2425 .priority = 10, /* ??! */
2429 * cpuset_init_smp - initialize cpus_allowed
2431 * Description: Finish top cpuset after cpu, node maps are initialized
2433 void __init cpuset_init_smp(void)
2435 cpumask_copy(top_cpuset.cpus_allowed, cpu_active_mask);
2436 top_cpuset.mems_allowed = node_states[N_MEMORY];
2437 top_cpuset.old_mems_allowed = top_cpuset.mems_allowed;
2439 cpumask_copy(top_cpuset.effective_cpus, cpu_active_mask);
2440 top_cpuset.effective_mems = node_states[N_MEMORY];
2442 register_hotmemory_notifier(&cpuset_track_online_nodes_nb);
2444 cpuset_migrate_mm_wq = alloc_ordered_workqueue("cpuset_migrate_mm", 0);
2445 BUG_ON(!cpuset_migrate_mm_wq);
2449 * cpuset_cpus_allowed - return cpus_allowed mask from a tasks cpuset.
2450 * @tsk: pointer to task_struct from which to obtain cpuset->cpus_allowed.
2451 * @pmask: pointer to struct cpumask variable to receive cpus_allowed set.
2453 * Description: Returns the cpumask_var_t cpus_allowed of the cpuset
2454 * attached to the specified @tsk. Guaranteed to return some non-empty
2455 * subset of cpu_online_mask, even if this means going outside the
2459 void cpuset_cpus_allowed(struct task_struct *tsk, struct cpumask *pmask)
2461 unsigned long flags;
2463 spin_lock_irqsave(&callback_lock, flags);
2465 guarantee_online_cpus(task_cs(tsk), pmask);
2467 spin_unlock_irqrestore(&callback_lock, flags);
2470 void cpuset_cpus_allowed_fallback(struct task_struct *tsk)
2473 do_set_cpus_allowed(tsk, task_cs(tsk)->effective_cpus);
2477 * We own tsk->cpus_allowed, nobody can change it under us.
2479 * But we used cs && cs->cpus_allowed lockless and thus can
2480 * race with cgroup_attach_task() or update_cpumask() and get
2481 * the wrong tsk->cpus_allowed. However, both cases imply the
2482 * subsequent cpuset_change_cpumask()->set_cpus_allowed_ptr()
2483 * which takes task_rq_lock().
2485 * If we are called after it dropped the lock we must see all
2486 * changes in tsk_cs()->cpus_allowed. Otherwise we can temporary
2487 * set any mask even if it is not right from task_cs() pov,
2488 * the pending set_cpus_allowed_ptr() will fix things.
2490 * select_fallback_rq() will fix things ups and set cpu_possible_mask
2495 void __init cpuset_init_current_mems_allowed(void)
2497 nodes_setall(current->mems_allowed);
2501 * cpuset_mems_allowed - return mems_allowed mask from a tasks cpuset.
2502 * @tsk: pointer to task_struct from which to obtain cpuset->mems_allowed.
2504 * Description: Returns the nodemask_t mems_allowed of the cpuset
2505 * attached to the specified @tsk. Guaranteed to return some non-empty
2506 * subset of node_states[N_MEMORY], even if this means going outside the
2510 nodemask_t cpuset_mems_allowed(struct task_struct *tsk)
2513 unsigned long flags;
2515 spin_lock_irqsave(&callback_lock, flags);
2517 guarantee_online_mems(task_cs(tsk), &mask);
2519 spin_unlock_irqrestore(&callback_lock, flags);
2525 * cpuset_nodemask_valid_mems_allowed - check nodemask vs. curremt mems_allowed
2526 * @nodemask: the nodemask to be checked
2528 * Are any of the nodes in the nodemask allowed in current->mems_allowed?
2530 int cpuset_nodemask_valid_mems_allowed(nodemask_t *nodemask)
2532 return nodes_intersects(*nodemask, current->mems_allowed);
2536 * nearest_hardwall_ancestor() - Returns the nearest mem_exclusive or
2537 * mem_hardwall ancestor to the specified cpuset. Call holding
2538 * callback_lock. If no ancestor is mem_exclusive or mem_hardwall
2539 * (an unusual configuration), then returns the root cpuset.
2541 static struct cpuset *nearest_hardwall_ancestor(struct cpuset *cs)
2543 while (!(is_mem_exclusive(cs) || is_mem_hardwall(cs)) && parent_cs(cs))
2549 * cpuset_node_allowed - Can we allocate on a memory node?
2550 * @node: is this an allowed node?
2551 * @gfp_mask: memory allocation flags
2553 * If we're in interrupt, yes, we can always allocate. If @node is set in
2554 * current's mems_allowed, yes. If it's not a __GFP_HARDWALL request and this
2555 * node is set in the nearest hardwalled cpuset ancestor to current's cpuset,
2556 * yes. If current has access to memory reserves due to TIF_MEMDIE, yes.
2559 * GFP_USER allocations are marked with the __GFP_HARDWALL bit,
2560 * and do not allow allocations outside the current tasks cpuset
2561 * unless the task has been OOM killed as is marked TIF_MEMDIE.
2562 * GFP_KERNEL allocations are not so marked, so can escape to the
2563 * nearest enclosing hardwalled ancestor cpuset.
2565 * Scanning up parent cpusets requires callback_lock. The
2566 * __alloc_pages() routine only calls here with __GFP_HARDWALL bit
2567 * _not_ set if it's a GFP_KERNEL allocation, and all nodes in the
2568 * current tasks mems_allowed came up empty on the first pass over
2569 * the zonelist. So only GFP_KERNEL allocations, if all nodes in the
2570 * cpuset are short of memory, might require taking the callback_lock.
2572 * The first call here from mm/page_alloc:get_page_from_freelist()
2573 * has __GFP_HARDWALL set in gfp_mask, enforcing hardwall cpusets,
2574 * so no allocation on a node outside the cpuset is allowed (unless
2575 * in interrupt, of course).
2577 * The second pass through get_page_from_freelist() doesn't even call
2578 * here for GFP_ATOMIC calls. For those calls, the __alloc_pages()
2579 * variable 'wait' is not set, and the bit ALLOC_CPUSET is not set
2580 * in alloc_flags. That logic and the checks below have the combined
2582 * in_interrupt - any node ok (current task context irrelevant)
2583 * GFP_ATOMIC - any node ok
2584 * TIF_MEMDIE - any node ok
2585 * GFP_KERNEL - any node in enclosing hardwalled cpuset ok
2586 * GFP_USER - only nodes in current tasks mems allowed ok.
2588 int __cpuset_node_allowed(int node, gfp_t gfp_mask)
2590 struct cpuset *cs; /* current cpuset ancestors */
2591 int allowed; /* is allocation in zone z allowed? */
2592 unsigned long flags;
2596 if (node_isset(node, current->mems_allowed))
2599 * Allow tasks that have access to memory reserves because they have
2600 * been OOM killed to get memory anywhere.
2602 if (unlikely(test_thread_flag(TIF_MEMDIE)))
2604 if (gfp_mask & __GFP_HARDWALL) /* If hardwall request, stop here */
2607 if (current->flags & PF_EXITING) /* Let dying task have memory */
2610 /* Not hardwall and node outside mems_allowed: scan up cpusets */
2611 spin_lock_irqsave(&callback_lock, flags);
2614 cs = nearest_hardwall_ancestor(task_cs(current));
2615 allowed = node_isset(node, cs->mems_allowed);
2618 spin_unlock_irqrestore(&callback_lock, flags);
2623 * cpuset_mem_spread_node() - On which node to begin search for a file page
2624 * cpuset_slab_spread_node() - On which node to begin search for a slab page
2626 * If a task is marked PF_SPREAD_PAGE or PF_SPREAD_SLAB (as for
2627 * tasks in a cpuset with is_spread_page or is_spread_slab set),
2628 * and if the memory allocation used cpuset_mem_spread_node()
2629 * to determine on which node to start looking, as it will for
2630 * certain page cache or slab cache pages such as used for file
2631 * system buffers and inode caches, then instead of starting on the
2632 * local node to look for a free page, rather spread the starting
2633 * node around the tasks mems_allowed nodes.
2635 * We don't have to worry about the returned node being offline
2636 * because "it can't happen", and even if it did, it would be ok.
2638 * The routines calling guarantee_online_mems() are careful to
2639 * only set nodes in task->mems_allowed that are online. So it
2640 * should not be possible for the following code to return an
2641 * offline node. But if it did, that would be ok, as this routine
2642 * is not returning the node where the allocation must be, only
2643 * the node where the search should start. The zonelist passed to
2644 * __alloc_pages() will include all nodes. If the slab allocator
2645 * is passed an offline node, it will fall back to the local node.
2646 * See kmem_cache_alloc_node().
2649 static int cpuset_spread_node(int *rotor)
2653 node = next_node(*rotor, current->mems_allowed);
2654 if (node == MAX_NUMNODES)
2655 node = first_node(current->mems_allowed);
2660 int cpuset_mem_spread_node(void)
2662 if (current->cpuset_mem_spread_rotor == NUMA_NO_NODE)
2663 current->cpuset_mem_spread_rotor =
2664 node_random(¤t->mems_allowed);
2666 return cpuset_spread_node(¤t->cpuset_mem_spread_rotor);
2669 int cpuset_slab_spread_node(void)
2671 if (current->cpuset_slab_spread_rotor == NUMA_NO_NODE)
2672 current->cpuset_slab_spread_rotor =
2673 node_random(¤t->mems_allowed);
2675 return cpuset_spread_node(¤t->cpuset_slab_spread_rotor);
2678 EXPORT_SYMBOL_GPL(cpuset_mem_spread_node);
2681 * cpuset_mems_allowed_intersects - Does @tsk1's mems_allowed intersect @tsk2's?
2682 * @tsk1: pointer to task_struct of some task.
2683 * @tsk2: pointer to task_struct of some other task.
2685 * Description: Return true if @tsk1's mems_allowed intersects the
2686 * mems_allowed of @tsk2. Used by the OOM killer to determine if
2687 * one of the task's memory usage might impact the memory available
2691 int cpuset_mems_allowed_intersects(const struct task_struct *tsk1,
2692 const struct task_struct *tsk2)
2694 return nodes_intersects(tsk1->mems_allowed, tsk2->mems_allowed);
2698 * cpuset_print_current_mems_allowed - prints current's cpuset and mems_allowed
2700 * Description: Prints current's name, cpuset name, and cached copy of its
2701 * mems_allowed to the kernel log.
2703 void cpuset_print_current_mems_allowed(void)
2705 struct cgroup *cgrp;
2709 cgrp = task_cs(current)->css.cgroup;
2710 pr_info("%s cpuset=", current->comm);
2711 pr_cont_cgroup_name(cgrp);
2712 pr_cont(" mems_allowed=%*pbl\n",
2713 nodemask_pr_args(¤t->mems_allowed));
2719 * Collection of memory_pressure is suppressed unless
2720 * this flag is enabled by writing "1" to the special
2721 * cpuset file 'memory_pressure_enabled' in the root cpuset.
2724 int cpuset_memory_pressure_enabled __read_mostly;
2727 * cpuset_memory_pressure_bump - keep stats of per-cpuset reclaims.
2729 * Keep a running average of the rate of synchronous (direct)
2730 * page reclaim efforts initiated by tasks in each cpuset.
2732 * This represents the rate at which some task in the cpuset
2733 * ran low on memory on all nodes it was allowed to use, and
2734 * had to enter the kernels page reclaim code in an effort to
2735 * create more free memory by tossing clean pages or swapping
2736 * or writing dirty pages.
2738 * Display to user space in the per-cpuset read-only file
2739 * "memory_pressure". Value displayed is an integer
2740 * representing the recent rate of entry into the synchronous
2741 * (direct) page reclaim by any task attached to the cpuset.
2744 void __cpuset_memory_pressure_bump(void)
2747 fmeter_markevent(&task_cs(current)->fmeter);
2751 #ifdef CONFIG_PROC_PID_CPUSET
2753 * proc_cpuset_show()
2754 * - Print tasks cpuset path into seq_file.
2755 * - Used for /proc/<pid>/cpuset.
2756 * - No need to task_lock(tsk) on this tsk->cpuset reference, as it
2757 * doesn't really matter if tsk->cpuset changes after we read it,
2758 * and we take cpuset_mutex, keeping cpuset_attach() from changing it
2761 int proc_cpuset_show(struct seq_file *m, struct pid_namespace *ns,
2762 struct pid *pid, struct task_struct *tsk)
2765 struct cgroup_subsys_state *css;
2769 buf = kmalloc(PATH_MAX, GFP_KERNEL);
2773 retval = -ENAMETOOLONG;
2775 css = task_css(tsk, cpuset_cgrp_id);
2776 p = cgroup_path(css->cgroup, buf, PATH_MAX);
2788 #endif /* CONFIG_PROC_PID_CPUSET */
2790 /* Display task mems_allowed in /proc/<pid>/status file. */
2791 void cpuset_task_status_allowed(struct seq_file *m, struct task_struct *task)
2793 seq_printf(m, "Mems_allowed:\t%*pb\n",
2794 nodemask_pr_args(&task->mems_allowed));
2795 seq_printf(m, "Mems_allowed_list:\t%*pbl\n",
2796 nodemask_pr_args(&task->mems_allowed));