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 cpumask_var_t cpus_requested;
102 nodemask_t mems_allowed;
104 /* effective CPUs and Memory Nodes allow to tasks */
105 cpumask_var_t effective_cpus;
106 nodemask_t effective_mems;
109 * This is old Memory Nodes tasks took on.
111 * - top_cpuset.old_mems_allowed is initialized to mems_allowed.
112 * - A new cpuset's old_mems_allowed is initialized when some
113 * task is moved into it.
114 * - old_mems_allowed is used in cpuset_migrate_mm() when we change
115 * cpuset.mems_allowed and have tasks' nodemask updated, and
116 * then old_mems_allowed is updated to mems_allowed.
118 nodemask_t old_mems_allowed;
120 struct fmeter fmeter; /* memory_pressure filter */
123 * Tasks are being attached to this cpuset. Used to prevent
124 * zeroing cpus/mems_allowed between ->can_attach() and ->attach().
126 int attach_in_progress;
128 /* partition number for rebuild_sched_domains() */
131 /* for custom sched domain */
132 int relax_domain_level;
135 static inline struct cpuset *css_cs(struct cgroup_subsys_state *css)
137 return css ? container_of(css, struct cpuset, css) : NULL;
140 /* Retrieve the cpuset for a task */
141 static inline struct cpuset *task_cs(struct task_struct *task)
143 return css_cs(task_css(task, cpuset_cgrp_id));
146 static inline struct cpuset *parent_cs(struct cpuset *cs)
148 return css_cs(cs->css.parent);
152 static inline bool task_has_mempolicy(struct task_struct *task)
154 return task->mempolicy;
157 static inline bool task_has_mempolicy(struct task_struct *task)
164 /* bits in struct cpuset flags field */
171 CS_SCHED_LOAD_BALANCE,
176 /* convenient tests for these bits */
177 static inline bool is_cpuset_online(const struct cpuset *cs)
179 return test_bit(CS_ONLINE, &cs->flags);
182 static inline int is_cpu_exclusive(const struct cpuset *cs)
184 return test_bit(CS_CPU_EXCLUSIVE, &cs->flags);
187 static inline int is_mem_exclusive(const struct cpuset *cs)
189 return test_bit(CS_MEM_EXCLUSIVE, &cs->flags);
192 static inline int is_mem_hardwall(const struct cpuset *cs)
194 return test_bit(CS_MEM_HARDWALL, &cs->flags);
197 static inline int is_sched_load_balance(const struct cpuset *cs)
199 return test_bit(CS_SCHED_LOAD_BALANCE, &cs->flags);
202 static inline int is_memory_migrate(const struct cpuset *cs)
204 return test_bit(CS_MEMORY_MIGRATE, &cs->flags);
207 static inline int is_spread_page(const struct cpuset *cs)
209 return test_bit(CS_SPREAD_PAGE, &cs->flags);
212 static inline int is_spread_slab(const struct cpuset *cs)
214 return test_bit(CS_SPREAD_SLAB, &cs->flags);
217 static struct cpuset top_cpuset = {
218 .flags = ((1 << CS_ONLINE) | (1 << CS_CPU_EXCLUSIVE) |
219 (1 << CS_MEM_EXCLUSIVE)),
223 * cpuset_for_each_child - traverse online children of a cpuset
224 * @child_cs: loop cursor pointing to the current child
225 * @pos_css: used for iteration
226 * @parent_cs: target cpuset to walk children of
228 * Walk @child_cs through the online children of @parent_cs. Must be used
229 * with RCU read locked.
231 #define cpuset_for_each_child(child_cs, pos_css, parent_cs) \
232 css_for_each_child((pos_css), &(parent_cs)->css) \
233 if (is_cpuset_online(((child_cs) = css_cs((pos_css)))))
236 * cpuset_for_each_descendant_pre - pre-order walk of a cpuset's descendants
237 * @des_cs: loop cursor pointing to the current descendant
238 * @pos_css: used for iteration
239 * @root_cs: target cpuset to walk ancestor of
241 * Walk @des_cs through the online descendants of @root_cs. Must be used
242 * with RCU read locked. The caller may modify @pos_css by calling
243 * css_rightmost_descendant() to skip subtree. @root_cs is included in the
244 * iteration and the first node to be visited.
246 #define cpuset_for_each_descendant_pre(des_cs, pos_css, root_cs) \
247 css_for_each_descendant_pre((pos_css), &(root_cs)->css) \
248 if (is_cpuset_online(((des_cs) = css_cs((pos_css)))))
251 * There are two global locks guarding cpuset structures - cpuset_mutex and
252 * callback_lock. We also require taking task_lock() when dereferencing a
253 * task's cpuset pointer. See "The task_lock() exception", at the end of this
256 * A task must hold both locks to modify cpusets. If a task holds
257 * cpuset_mutex, then it blocks others wanting that mutex, ensuring that it
258 * is the only task able to also acquire callback_lock and be able to
259 * modify cpusets. It can perform various checks on the cpuset structure
260 * first, knowing nothing will change. It can also allocate memory while
261 * just holding cpuset_mutex. While it is performing these checks, various
262 * callback routines can briefly acquire callback_lock to query cpusets.
263 * Once it is ready to make the changes, it takes callback_lock, blocking
266 * Calls to the kernel memory allocator can not be made while holding
267 * callback_lock, as that would risk double tripping on callback_lock
268 * from one of the callbacks into the cpuset code from within
271 * If a task is only holding callback_lock, then it has read-only
274 * Now, the task_struct fields mems_allowed and mempolicy may be changed
275 * by other task, we use alloc_lock in the task_struct fields to protect
278 * The cpuset_common_file_read() handlers only hold callback_lock across
279 * small pieces of code, such as when reading out possibly multi-word
280 * cpumasks and nodemasks.
282 * Accessing a task's cpuset should be done in accordance with the
283 * guidelines for accessing subsystem state in kernel/cgroup.c
286 static DEFINE_MUTEX(cpuset_mutex);
287 static DEFINE_SPINLOCK(callback_lock);
289 static struct workqueue_struct *cpuset_migrate_mm_wq;
292 * CPU / memory hotplug is handled asynchronously.
294 static void cpuset_hotplug_workfn(struct work_struct *work);
295 static DECLARE_WORK(cpuset_hotplug_work, cpuset_hotplug_workfn);
297 static DECLARE_WAIT_QUEUE_HEAD(cpuset_attach_wq);
300 * This is ugly, but preserves the userspace API for existing cpuset
301 * users. If someone tries to mount the "cpuset" filesystem, we
302 * silently switch it to mount "cgroup" instead
304 static struct dentry *cpuset_mount(struct file_system_type *fs_type,
305 int flags, const char *unused_dev_name, void *data)
307 struct file_system_type *cgroup_fs = get_fs_type("cgroup");
308 struct dentry *ret = ERR_PTR(-ENODEV);
312 "release_agent=/sbin/cpuset_release_agent";
313 ret = cgroup_fs->mount(cgroup_fs, flags,
314 unused_dev_name, mountopts);
315 put_filesystem(cgroup_fs);
320 static struct file_system_type cpuset_fs_type = {
322 .mount = cpuset_mount,
326 * Return in pmask the portion of a cpusets's cpus_allowed that
327 * are online. If none are online, walk up the cpuset hierarchy
328 * until we find one that does have some online cpus. The top
329 * cpuset always has some cpus online.
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))
340 cpumask_and(pmask, cs->effective_cpus, cpu_online_mask);
344 * Return in *pmask the portion of a cpusets's mems_allowed that
345 * are online, with memory. If none are online with memory, walk
346 * up the cpuset hierarchy until we find one that does have some
347 * online mems. The top cpuset always has some mems online.
349 * One way or another, we guarantee to return some non-empty subset
350 * of node_states[N_MEMORY].
352 * Call with callback_lock or cpuset_mutex held.
354 static void guarantee_online_mems(struct cpuset *cs, nodemask_t *pmask)
356 while (!nodes_intersects(cs->effective_mems, node_states[N_MEMORY]))
358 nodes_and(*pmask, cs->effective_mems, node_states[N_MEMORY]);
362 * update task's spread flag if cpuset's page/slab spread flag is set
364 * Call with callback_lock or cpuset_mutex held.
366 static void cpuset_update_task_spread_flag(struct cpuset *cs,
367 struct task_struct *tsk)
369 if (is_spread_page(cs))
370 task_set_spread_page(tsk);
372 task_clear_spread_page(tsk);
374 if (is_spread_slab(cs))
375 task_set_spread_slab(tsk);
377 task_clear_spread_slab(tsk);
381 * is_cpuset_subset(p, q) - Is cpuset p a subset of cpuset q?
383 * One cpuset is a subset of another if all its allowed CPUs and
384 * Memory Nodes are a subset of the other, and its exclusive flags
385 * are only set if the other's are set. Call holding cpuset_mutex.
388 static int is_cpuset_subset(const struct cpuset *p, const struct cpuset *q)
390 return cpumask_subset(p->cpus_requested, q->cpus_requested) &&
391 nodes_subset(p->mems_allowed, q->mems_allowed) &&
392 is_cpu_exclusive(p) <= is_cpu_exclusive(q) &&
393 is_mem_exclusive(p) <= is_mem_exclusive(q);
397 * alloc_trial_cpuset - allocate a trial cpuset
398 * @cs: the cpuset that the trial cpuset duplicates
400 static struct cpuset *alloc_trial_cpuset(struct cpuset *cs)
402 struct cpuset *trial;
404 trial = kmemdup(cs, sizeof(*cs), GFP_KERNEL);
408 if (!alloc_cpumask_var(&trial->cpus_allowed, GFP_KERNEL))
410 if (!alloc_cpumask_var(&trial->effective_cpus, GFP_KERNEL))
413 cpumask_copy(trial->cpus_allowed, cs->cpus_allowed);
414 cpumask_copy(trial->effective_cpus, cs->effective_cpus);
418 free_cpumask_var(trial->cpus_allowed);
425 * free_trial_cpuset - free the trial cpuset
426 * @trial: the trial cpuset to be freed
428 static void free_trial_cpuset(struct cpuset *trial)
430 free_cpumask_var(trial->effective_cpus);
431 free_cpumask_var(trial->cpus_allowed);
436 * validate_change() - Used to validate that any proposed cpuset change
437 * follows the structural rules for cpusets.
439 * If we replaced the flag and mask values of the current cpuset
440 * (cur) with those values in the trial cpuset (trial), would
441 * our various subset and exclusive rules still be valid? Presumes
444 * 'cur' is the address of an actual, in-use cpuset. Operations
445 * such as list traversal that depend on the actual address of the
446 * cpuset in the list must use cur below, not trial.
448 * 'trial' is the address of bulk structure copy of cur, with
449 * perhaps one or more of the fields cpus_allowed, mems_allowed,
450 * or flags changed to new, trial values.
452 * Return 0 if valid, -errno if not.
455 static int validate_change(struct cpuset *cur, struct cpuset *trial)
457 struct cgroup_subsys_state *css;
458 struct cpuset *c, *par;
463 /* Each of our child cpusets must be a subset of us */
465 cpuset_for_each_child(c, css, cur)
466 if (!is_cpuset_subset(c, trial))
469 /* Remaining checks don't apply to root cpuset */
471 if (cur == &top_cpuset)
474 par = parent_cs(cur);
476 /* On legacy hiearchy, we must be a subset of our parent cpuset. */
478 if (!cgroup_subsys_on_dfl(cpuset_cgrp_subsys) &&
479 !is_cpuset_subset(trial, par))
483 * If either I or some sibling (!= me) is exclusive, we can't
487 cpuset_for_each_child(c, css, par) {
488 if ((is_cpu_exclusive(trial) || is_cpu_exclusive(c)) &&
490 cpumask_intersects(trial->cpus_requested, c->cpus_requested))
492 if ((is_mem_exclusive(trial) || is_mem_exclusive(c)) &&
494 nodes_intersects(trial->mems_allowed, c->mems_allowed))
499 * Cpusets with tasks - existing or newly being attached - can't
500 * be changed to have empty cpus_allowed or mems_allowed.
503 if ((cgroup_is_populated(cur->css.cgroup) || cur->attach_in_progress)) {
504 if (!cpumask_empty(cur->cpus_allowed) &&
505 cpumask_empty(trial->cpus_allowed))
507 if (!nodes_empty(cur->mems_allowed) &&
508 nodes_empty(trial->mems_allowed))
513 * We can't shrink if we won't have enough room for SCHED_DEADLINE
517 if (is_cpu_exclusive(cur) &&
518 !cpuset_cpumask_can_shrink(cur->cpus_allowed,
519 trial->cpus_allowed))
530 * Helper routine for generate_sched_domains().
531 * Do cpusets a, b have overlapping effective cpus_allowed masks?
533 static int cpusets_overlap(struct cpuset *a, struct cpuset *b)
535 return cpumask_intersects(a->effective_cpus, b->effective_cpus);
539 update_domain_attr(struct sched_domain_attr *dattr, struct cpuset *c)
541 if (dattr->relax_domain_level < c->relax_domain_level)
542 dattr->relax_domain_level = c->relax_domain_level;
546 static void update_domain_attr_tree(struct sched_domain_attr *dattr,
547 struct cpuset *root_cs)
550 struct cgroup_subsys_state *pos_css;
553 cpuset_for_each_descendant_pre(cp, pos_css, root_cs) {
554 /* skip the whole subtree if @cp doesn't have any CPU */
555 if (cpumask_empty(cp->cpus_allowed)) {
556 pos_css = css_rightmost_descendant(pos_css);
560 if (is_sched_load_balance(cp))
561 update_domain_attr(dattr, cp);
567 * generate_sched_domains()
569 * This function builds a partial partition of the systems CPUs
570 * A 'partial partition' is a set of non-overlapping subsets whose
571 * union is a subset of that set.
572 * The output of this function needs to be passed to kernel/sched/core.c
573 * partition_sched_domains() routine, which will rebuild the scheduler's
574 * load balancing domains (sched domains) as specified by that partial
577 * See "What is sched_load_balance" in Documentation/cgroups/cpusets.txt
578 * for a background explanation of this.
580 * Does not return errors, on the theory that the callers of this
581 * routine would rather not worry about failures to rebuild sched
582 * domains when operating in the severe memory shortage situations
583 * that could cause allocation failures below.
585 * Must be called with cpuset_mutex held.
587 * The three key local variables below are:
588 * q - a linked-list queue of cpuset pointers, used to implement a
589 * top-down scan of all cpusets. This scan loads a pointer
590 * to each cpuset marked is_sched_load_balance into the
591 * array 'csa'. For our purposes, rebuilding the schedulers
592 * sched domains, we can ignore !is_sched_load_balance cpusets.
593 * csa - (for CpuSet Array) Array of pointers to all the cpusets
594 * that need to be load balanced, for convenient iterative
595 * access by the subsequent code that finds the best partition,
596 * i.e the set of domains (subsets) of CPUs such that the
597 * cpus_allowed of every cpuset marked is_sched_load_balance
598 * is a subset of one of these domains, while there are as
599 * many such domains as possible, each as small as possible.
600 * doms - Conversion of 'csa' to an array of cpumasks, for passing to
601 * the kernel/sched/core.c routine partition_sched_domains() in a
602 * convenient format, that can be easily compared to the prior
603 * value to determine what partition elements (sched domains)
604 * were changed (added or removed.)
606 * Finding the best partition (set of domains):
607 * The triple nested loops below over i, j, k scan over the
608 * load balanced cpusets (using the array of cpuset pointers in
609 * csa[]) looking for pairs of cpusets that have overlapping
610 * cpus_allowed, but which don't have the same 'pn' partition
611 * number and gives them in the same partition number. It keeps
612 * looping on the 'restart' label until it can no longer find
615 * The union of the cpus_allowed masks from the set of
616 * all cpusets having the same 'pn' value then form the one
617 * element of the partition (one sched domain) to be passed to
618 * partition_sched_domains().
620 static int generate_sched_domains(cpumask_var_t **domains,
621 struct sched_domain_attr **attributes)
623 struct cpuset *cp; /* scans q */
624 struct cpuset **csa; /* array of all cpuset ptrs */
625 int csn; /* how many cpuset ptrs in csa so far */
626 int i, j, k; /* indices for partition finding loops */
627 cpumask_var_t *doms; /* resulting partition; i.e. sched domains */
628 cpumask_var_t non_isolated_cpus; /* load balanced CPUs */
629 struct sched_domain_attr *dattr; /* attributes for custom domains */
630 int ndoms = 0; /* number of sched domains in result */
631 int nslot; /* next empty doms[] struct cpumask slot */
632 struct cgroup_subsys_state *pos_css;
638 if (!alloc_cpumask_var(&non_isolated_cpus, GFP_KERNEL))
640 cpumask_andnot(non_isolated_cpus, cpu_possible_mask, cpu_isolated_map);
642 /* Special case for the 99% of systems with one, full, sched domain */
643 if (is_sched_load_balance(&top_cpuset)) {
645 doms = alloc_sched_domains(ndoms);
649 dattr = kmalloc(sizeof(struct sched_domain_attr), GFP_KERNEL);
651 *dattr = SD_ATTR_INIT;
652 update_domain_attr_tree(dattr, &top_cpuset);
654 cpumask_and(doms[0], top_cpuset.effective_cpus,
660 csa = kmalloc(nr_cpusets() * sizeof(cp), GFP_KERNEL);
666 cpuset_for_each_descendant_pre(cp, pos_css, &top_cpuset) {
667 if (cp == &top_cpuset)
670 * Continue traversing beyond @cp iff @cp has some CPUs and
671 * isn't load balancing. The former is obvious. The
672 * latter: All child cpusets contain a subset of the
673 * parent's cpus, so just skip them, and then we call
674 * update_domain_attr_tree() to calc relax_domain_level of
675 * the corresponding sched domain.
677 if (!cpumask_empty(cp->cpus_allowed) &&
678 !(is_sched_load_balance(cp) &&
679 cpumask_intersects(cp->cpus_allowed, non_isolated_cpus)))
682 if (is_sched_load_balance(cp))
685 /* skip @cp's subtree */
686 pos_css = css_rightmost_descendant(pos_css);
690 for (i = 0; i < csn; i++)
695 /* Find the best partition (set of sched domains) */
696 for (i = 0; i < csn; i++) {
697 struct cpuset *a = csa[i];
700 for (j = 0; j < csn; j++) {
701 struct cpuset *b = csa[j];
704 if (apn != bpn && cpusets_overlap(a, b)) {
705 for (k = 0; k < csn; k++) {
706 struct cpuset *c = csa[k];
711 ndoms--; /* one less element */
718 * Now we know how many domains to create.
719 * Convert <csn, csa> to <ndoms, doms> and populate cpu masks.
721 doms = alloc_sched_domains(ndoms);
726 * The rest of the code, including the scheduler, can deal with
727 * dattr==NULL case. No need to abort if alloc fails.
729 dattr = kmalloc(ndoms * sizeof(struct sched_domain_attr), GFP_KERNEL);
731 for (nslot = 0, i = 0; i < csn; i++) {
732 struct cpuset *a = csa[i];
737 /* Skip completed partitions */
743 if (nslot == ndoms) {
744 static int warnings = 10;
746 pr_warn("rebuild_sched_domains confused: nslot %d, ndoms %d, csn %d, i %d, apn %d\n",
747 nslot, ndoms, csn, i, apn);
755 *(dattr + nslot) = SD_ATTR_INIT;
756 for (j = i; j < csn; j++) {
757 struct cpuset *b = csa[j];
760 cpumask_or(dp, dp, b->effective_cpus);
761 cpumask_and(dp, dp, non_isolated_cpus);
763 update_domain_attr_tree(dattr + nslot, b);
765 /* Done with this partition */
771 BUG_ON(nslot != ndoms);
774 free_cpumask_var(non_isolated_cpus);
778 * Fallback to the default domain if kmalloc() failed.
779 * See comments in partition_sched_domains().
790 * Rebuild scheduler domains.
792 * If the flag 'sched_load_balance' of any cpuset with non-empty
793 * 'cpus' changes, or if the 'cpus' allowed changes in any cpuset
794 * which has that flag enabled, or if any cpuset with a non-empty
795 * 'cpus' is removed, then call this routine to rebuild the
796 * scheduler's dynamic sched domains.
798 * Call with cpuset_mutex held. Takes get_online_cpus().
800 static void rebuild_sched_domains_locked(void)
802 struct sched_domain_attr *attr;
806 lockdep_assert_held(&cpuset_mutex);
810 * We have raced with CPU hotplug. Don't do anything to avoid
811 * passing doms with offlined cpu to partition_sched_domains().
812 * Anyways, hotplug work item will rebuild sched domains.
814 if (!cpumask_equal(top_cpuset.effective_cpus, cpu_active_mask))
817 /* Generate domain masks and attrs */
818 ndoms = generate_sched_domains(&doms, &attr);
820 /* Have scheduler rebuild the domains */
821 partition_sched_domains(ndoms, doms, attr);
825 #else /* !CONFIG_SMP */
826 static void rebuild_sched_domains_locked(void)
829 #endif /* CONFIG_SMP */
831 void rebuild_sched_domains(void)
833 mutex_lock(&cpuset_mutex);
834 rebuild_sched_domains_locked();
835 mutex_unlock(&cpuset_mutex);
839 * update_tasks_cpumask - Update the cpumasks of tasks in the cpuset.
840 * @cs: the cpuset in which each task's cpus_allowed mask needs to be changed
842 * Iterate through each task of @cs updating its cpus_allowed to the
843 * effective cpuset's. As this function is called with cpuset_mutex held,
844 * cpuset membership stays stable.
846 static void update_tasks_cpumask(struct cpuset *cs)
848 struct css_task_iter it;
849 struct task_struct *task;
851 css_task_iter_start(&cs->css, &it);
852 while ((task = css_task_iter_next(&it)))
853 set_cpus_allowed_ptr(task, cs->effective_cpus);
854 css_task_iter_end(&it);
858 * update_cpumasks_hier - Update effective cpumasks and tasks in the subtree
859 * @cs: the cpuset to consider
860 * @new_cpus: temp variable for calculating new effective_cpus
862 * When congifured cpumask is changed, the effective cpumasks of this cpuset
863 * and all its descendants need to be updated.
865 * On legacy hierachy, effective_cpus will be the same with cpu_allowed.
867 * Called with cpuset_mutex held
869 static void update_cpumasks_hier(struct cpuset *cs, struct cpumask *new_cpus)
872 struct cgroup_subsys_state *pos_css;
873 bool need_rebuild_sched_domains = false;
876 cpuset_for_each_descendant_pre(cp, pos_css, cs) {
877 struct cpuset *parent = parent_cs(cp);
879 cpumask_and(new_cpus, cp->cpus_allowed, parent->effective_cpus);
882 * If it becomes empty, inherit the effective mask of the
883 * parent, which is guaranteed to have some CPUs.
885 if (cgroup_subsys_on_dfl(cpuset_cgrp_subsys) &&
886 cpumask_empty(new_cpus))
887 cpumask_copy(new_cpus, parent->effective_cpus);
889 /* Skip the whole subtree if the cpumask remains the same. */
890 if (cpumask_equal(new_cpus, cp->effective_cpus)) {
891 pos_css = css_rightmost_descendant(pos_css);
895 if (!css_tryget_online(&cp->css))
899 spin_lock_irq(&callback_lock);
900 cpumask_copy(cp->effective_cpus, new_cpus);
901 spin_unlock_irq(&callback_lock);
903 WARN_ON(!cgroup_subsys_on_dfl(cpuset_cgrp_subsys) &&
904 !cpumask_equal(cp->cpus_allowed, cp->effective_cpus));
906 update_tasks_cpumask(cp);
909 * If the effective cpumask of any non-empty cpuset is changed,
910 * we need to rebuild sched domains.
912 if (!cpumask_empty(cp->cpus_allowed) &&
913 is_sched_load_balance(cp))
914 need_rebuild_sched_domains = true;
921 if (need_rebuild_sched_domains)
922 rebuild_sched_domains_locked();
926 * update_cpumask - update the cpus_allowed mask of a cpuset and all tasks in it
927 * @cs: the cpuset to consider
928 * @trialcs: trial cpuset
929 * @buf: buffer of cpu numbers written to this cpuset
931 static int update_cpumask(struct cpuset *cs, struct cpuset *trialcs,
936 /* top_cpuset.cpus_allowed tracks cpu_online_mask; it's read-only */
937 if (cs == &top_cpuset)
941 * An empty cpus_allowed is ok only if the cpuset has no tasks.
942 * Since cpulist_parse() fails on an empty mask, we special case
943 * that parsing. The validate_change() call ensures that cpusets
944 * with tasks have cpus.
947 cpumask_clear(trialcs->cpus_allowed);
949 retval = cpulist_parse(buf, trialcs->cpus_requested);
953 if (!cpumask_subset(trialcs->cpus_requested, cpu_present_mask))
956 cpumask_and(trialcs->cpus_allowed, trialcs->cpus_requested, cpu_active_mask);
959 /* Nothing to do if the cpus didn't change */
960 if (cpumask_equal(cs->cpus_requested, trialcs->cpus_requested))
963 retval = validate_change(cs, trialcs);
967 spin_lock_irq(&callback_lock);
968 cpumask_copy(cs->cpus_allowed, trialcs->cpus_allowed);
969 cpumask_copy(cs->cpus_requested, trialcs->cpus_requested);
970 spin_unlock_irq(&callback_lock);
972 /* use trialcs->cpus_allowed as a temp variable */
973 update_cpumasks_hier(cs, trialcs->cpus_allowed);
978 * Migrate memory region from one set of nodes to another. This is
979 * performed asynchronously as it can be called from process migration path
980 * holding locks involved in process management. All mm migrations are
981 * performed in the queued order and can be waited for by flushing
982 * cpuset_migrate_mm_wq.
985 struct cpuset_migrate_mm_work {
986 struct work_struct work;
987 struct mm_struct *mm;
992 static void cpuset_migrate_mm_workfn(struct work_struct *work)
994 struct cpuset_migrate_mm_work *mwork =
995 container_of(work, struct cpuset_migrate_mm_work, work);
997 /* on a wq worker, no need to worry about %current's mems_allowed */
998 do_migrate_pages(mwork->mm, &mwork->from, &mwork->to, MPOL_MF_MOVE_ALL);
1003 static void cpuset_migrate_mm(struct mm_struct *mm, const nodemask_t *from,
1004 const nodemask_t *to)
1006 struct cpuset_migrate_mm_work *mwork;
1008 mwork = kzalloc(sizeof(*mwork), GFP_KERNEL);
1011 mwork->from = *from;
1013 INIT_WORK(&mwork->work, cpuset_migrate_mm_workfn);
1014 queue_work(cpuset_migrate_mm_wq, &mwork->work);
1020 static void cpuset_post_attach(void)
1022 flush_workqueue(cpuset_migrate_mm_wq);
1026 * cpuset_change_task_nodemask - change task's mems_allowed and mempolicy
1027 * @tsk: the task to change
1028 * @newmems: new nodes that the task will be set
1030 * In order to avoid seeing no nodes if the old and new nodes are disjoint,
1031 * we structure updates as setting all new allowed nodes, then clearing newly
1034 static void cpuset_change_task_nodemask(struct task_struct *tsk,
1035 nodemask_t *newmems)
1040 * Allow tasks that have access to memory reserves because they have
1041 * been OOM killed to get memory anywhere.
1043 if (unlikely(test_thread_flag(TIF_MEMDIE)))
1045 if (current->flags & PF_EXITING) /* Let dying task have memory */
1050 * Determine if a loop is necessary if another thread is doing
1051 * read_mems_allowed_begin(). If at least one node remains unchanged and
1052 * tsk does not have a mempolicy, then an empty nodemask will not be
1053 * possible when mems_allowed is larger than a word.
1055 need_loop = task_has_mempolicy(tsk) ||
1056 !nodes_intersects(*newmems, tsk->mems_allowed);
1059 local_irq_disable();
1060 write_seqcount_begin(&tsk->mems_allowed_seq);
1063 nodes_or(tsk->mems_allowed, tsk->mems_allowed, *newmems);
1064 mpol_rebind_task(tsk, newmems, MPOL_REBIND_STEP1);
1066 mpol_rebind_task(tsk, newmems, MPOL_REBIND_STEP2);
1067 tsk->mems_allowed = *newmems;
1070 write_seqcount_end(&tsk->mems_allowed_seq);
1077 static void *cpuset_being_rebound;
1080 * update_tasks_nodemask - Update the nodemasks of tasks in the cpuset.
1081 * @cs: the cpuset in which each task's mems_allowed mask needs to be changed
1083 * Iterate through each task of @cs updating its mems_allowed to the
1084 * effective cpuset's. As this function is called with cpuset_mutex held,
1085 * cpuset membership stays stable.
1087 static void update_tasks_nodemask(struct cpuset *cs)
1089 static nodemask_t newmems; /* protected by cpuset_mutex */
1090 struct css_task_iter it;
1091 struct task_struct *task;
1093 cpuset_being_rebound = cs; /* causes mpol_dup() rebind */
1095 guarantee_online_mems(cs, &newmems);
1098 * The mpol_rebind_mm() call takes mmap_sem, which we couldn't
1099 * take while holding tasklist_lock. Forks can happen - the
1100 * mpol_dup() cpuset_being_rebound check will catch such forks,
1101 * and rebind their vma mempolicies too. Because we still hold
1102 * the global cpuset_mutex, we know that no other rebind effort
1103 * will be contending for the global variable cpuset_being_rebound.
1104 * It's ok if we rebind the same mm twice; mpol_rebind_mm()
1105 * is idempotent. Also migrate pages in each mm to new nodes.
1107 css_task_iter_start(&cs->css, &it);
1108 while ((task = css_task_iter_next(&it))) {
1109 struct mm_struct *mm;
1112 cpuset_change_task_nodemask(task, &newmems);
1114 mm = get_task_mm(task);
1118 migrate = is_memory_migrate(cs);
1120 mpol_rebind_mm(mm, &cs->mems_allowed);
1122 cpuset_migrate_mm(mm, &cs->old_mems_allowed, &newmems);
1126 css_task_iter_end(&it);
1129 * All the tasks' nodemasks have been updated, update
1130 * cs->old_mems_allowed.
1132 cs->old_mems_allowed = newmems;
1134 /* We're done rebinding vmas to this cpuset's new mems_allowed. */
1135 cpuset_being_rebound = NULL;
1139 * update_nodemasks_hier - Update effective nodemasks and tasks in the subtree
1140 * @cs: the cpuset to consider
1141 * @new_mems: a temp variable for calculating new effective_mems
1143 * When configured nodemask is changed, the effective nodemasks of this cpuset
1144 * and all its descendants need to be updated.
1146 * On legacy hiearchy, effective_mems will be the same with mems_allowed.
1148 * Called with cpuset_mutex held
1150 static void update_nodemasks_hier(struct cpuset *cs, nodemask_t *new_mems)
1153 struct cgroup_subsys_state *pos_css;
1156 cpuset_for_each_descendant_pre(cp, pos_css, cs) {
1157 struct cpuset *parent = parent_cs(cp);
1159 nodes_and(*new_mems, cp->mems_allowed, parent->effective_mems);
1162 * If it becomes empty, inherit the effective mask of the
1163 * parent, which is guaranteed to have some MEMs.
1165 if (cgroup_subsys_on_dfl(cpuset_cgrp_subsys) &&
1166 nodes_empty(*new_mems))
1167 *new_mems = parent->effective_mems;
1169 /* Skip the whole subtree if the nodemask remains the same. */
1170 if (nodes_equal(*new_mems, cp->effective_mems)) {
1171 pos_css = css_rightmost_descendant(pos_css);
1175 if (!css_tryget_online(&cp->css))
1179 spin_lock_irq(&callback_lock);
1180 cp->effective_mems = *new_mems;
1181 spin_unlock_irq(&callback_lock);
1183 WARN_ON(!cgroup_subsys_on_dfl(cpuset_cgrp_subsys) &&
1184 !nodes_equal(cp->mems_allowed, cp->effective_mems));
1186 update_tasks_nodemask(cp);
1195 * Handle user request to change the 'mems' memory placement
1196 * of a cpuset. Needs to validate the request, update the
1197 * cpusets mems_allowed, and for each task in the cpuset,
1198 * update mems_allowed and rebind task's mempolicy and any vma
1199 * mempolicies and if the cpuset is marked 'memory_migrate',
1200 * migrate the tasks pages to the new memory.
1202 * Call with cpuset_mutex held. May take callback_lock during call.
1203 * Will take tasklist_lock, scan tasklist for tasks in cpuset cs,
1204 * lock each such tasks mm->mmap_sem, scan its vma's and rebind
1205 * their mempolicies to the cpusets new mems_allowed.
1207 static int update_nodemask(struct cpuset *cs, struct cpuset *trialcs,
1213 * top_cpuset.mems_allowed tracks node_stats[N_MEMORY];
1216 if (cs == &top_cpuset) {
1222 * An empty mems_allowed is ok iff there are no tasks in the cpuset.
1223 * Since nodelist_parse() fails on an empty mask, we special case
1224 * that parsing. The validate_change() call ensures that cpusets
1225 * with tasks have memory.
1228 nodes_clear(trialcs->mems_allowed);
1230 retval = nodelist_parse(buf, trialcs->mems_allowed);
1234 if (!nodes_subset(trialcs->mems_allowed,
1235 top_cpuset.mems_allowed)) {
1241 if (nodes_equal(cs->mems_allowed, trialcs->mems_allowed)) {
1242 retval = 0; /* Too easy - nothing to do */
1245 retval = validate_change(cs, trialcs);
1249 spin_lock_irq(&callback_lock);
1250 cs->mems_allowed = trialcs->mems_allowed;
1251 spin_unlock_irq(&callback_lock);
1253 /* use trialcs->mems_allowed as a temp variable */
1254 update_nodemasks_hier(cs, &trialcs->mems_allowed);
1259 int current_cpuset_is_being_rebound(void)
1264 ret = task_cs(current) == cpuset_being_rebound;
1270 static int update_relax_domain_level(struct cpuset *cs, s64 val)
1273 if (val < -1 || val >= sched_domain_level_max)
1277 if (val != cs->relax_domain_level) {
1278 cs->relax_domain_level = val;
1279 if (!cpumask_empty(cs->cpus_allowed) &&
1280 is_sched_load_balance(cs))
1281 rebuild_sched_domains_locked();
1288 * update_tasks_flags - update the spread flags of tasks in the cpuset.
1289 * @cs: the cpuset in which each task's spread flags needs to be changed
1291 * Iterate through each task of @cs updating its spread flags. As this
1292 * function is called with cpuset_mutex held, cpuset membership stays
1295 static void update_tasks_flags(struct cpuset *cs)
1297 struct css_task_iter it;
1298 struct task_struct *task;
1300 css_task_iter_start(&cs->css, &it);
1301 while ((task = css_task_iter_next(&it)))
1302 cpuset_update_task_spread_flag(cs, task);
1303 css_task_iter_end(&it);
1307 * update_flag - read a 0 or a 1 in a file and update associated flag
1308 * bit: the bit to update (see cpuset_flagbits_t)
1309 * cs: the cpuset to update
1310 * turning_on: whether the flag is being set or cleared
1312 * Call with cpuset_mutex held.
1315 static int update_flag(cpuset_flagbits_t bit, struct cpuset *cs,
1318 struct cpuset *trialcs;
1319 int balance_flag_changed;
1320 int spread_flag_changed;
1323 trialcs = alloc_trial_cpuset(cs);
1328 set_bit(bit, &trialcs->flags);
1330 clear_bit(bit, &trialcs->flags);
1332 err = validate_change(cs, trialcs);
1336 balance_flag_changed = (is_sched_load_balance(cs) !=
1337 is_sched_load_balance(trialcs));
1339 spread_flag_changed = ((is_spread_slab(cs) != is_spread_slab(trialcs))
1340 || (is_spread_page(cs) != is_spread_page(trialcs)));
1342 spin_lock_irq(&callback_lock);
1343 cs->flags = trialcs->flags;
1344 spin_unlock_irq(&callback_lock);
1346 if (!cpumask_empty(trialcs->cpus_allowed) && balance_flag_changed)
1347 rebuild_sched_domains_locked();
1349 if (spread_flag_changed)
1350 update_tasks_flags(cs);
1352 free_trial_cpuset(trialcs);
1357 * Frequency meter - How fast is some event occurring?
1359 * These routines manage a digitally filtered, constant time based,
1360 * event frequency meter. There are four routines:
1361 * fmeter_init() - initialize a frequency meter.
1362 * fmeter_markevent() - called each time the event happens.
1363 * fmeter_getrate() - returns the recent rate of such events.
1364 * fmeter_update() - internal routine used to update fmeter.
1366 * A common data structure is passed to each of these routines,
1367 * which is used to keep track of the state required to manage the
1368 * frequency meter and its digital filter.
1370 * The filter works on the number of events marked per unit time.
1371 * The filter is single-pole low-pass recursive (IIR). The time unit
1372 * is 1 second. Arithmetic is done using 32-bit integers scaled to
1373 * simulate 3 decimal digits of precision (multiplied by 1000).
1375 * With an FM_COEF of 933, and a time base of 1 second, the filter
1376 * has a half-life of 10 seconds, meaning that if the events quit
1377 * happening, then the rate returned from the fmeter_getrate()
1378 * will be cut in half each 10 seconds, until it converges to zero.
1380 * It is not worth doing a real infinitely recursive filter. If more
1381 * than FM_MAXTICKS ticks have elapsed since the last filter event,
1382 * just compute FM_MAXTICKS ticks worth, by which point the level
1385 * Limit the count of unprocessed events to FM_MAXCNT, so as to avoid
1386 * arithmetic overflow in the fmeter_update() routine.
1388 * Given the simple 32 bit integer arithmetic used, this meter works
1389 * best for reporting rates between one per millisecond (msec) and
1390 * one per 32 (approx) seconds. At constant rates faster than one
1391 * per msec it maxes out at values just under 1,000,000. At constant
1392 * rates between one per msec, and one per second it will stabilize
1393 * to a value N*1000, where N is the rate of events per second.
1394 * At constant rates between one per second and one per 32 seconds,
1395 * it will be choppy, moving up on the seconds that have an event,
1396 * and then decaying until the next event. At rates slower than
1397 * about one in 32 seconds, it decays all the way back to zero between
1401 #define FM_COEF 933 /* coefficient for half-life of 10 secs */
1402 #define FM_MAXTICKS ((time_t)99) /* useless computing more ticks than this */
1403 #define FM_MAXCNT 1000000 /* limit cnt to avoid overflow */
1404 #define FM_SCALE 1000 /* faux fixed point scale */
1406 /* Initialize a frequency meter */
1407 static void fmeter_init(struct fmeter *fmp)
1412 spin_lock_init(&fmp->lock);
1415 /* Internal meter update - process cnt events and update value */
1416 static void fmeter_update(struct fmeter *fmp)
1418 time_t now = get_seconds();
1419 time_t ticks = now - fmp->time;
1424 ticks = min(FM_MAXTICKS, ticks);
1426 fmp->val = (FM_COEF * fmp->val) / FM_SCALE;
1429 fmp->val += ((FM_SCALE - FM_COEF) * fmp->cnt) / FM_SCALE;
1433 /* Process any previous ticks, then bump cnt by one (times scale). */
1434 static void fmeter_markevent(struct fmeter *fmp)
1436 spin_lock(&fmp->lock);
1438 fmp->cnt = min(FM_MAXCNT, fmp->cnt + FM_SCALE);
1439 spin_unlock(&fmp->lock);
1442 /* Process any previous ticks, then return current value. */
1443 static int fmeter_getrate(struct fmeter *fmp)
1447 spin_lock(&fmp->lock);
1450 spin_unlock(&fmp->lock);
1454 static struct cpuset *cpuset_attach_old_cs;
1456 /* Called by cgroups to determine if a cpuset is usable; cpuset_mutex held */
1457 static int cpuset_can_attach(struct cgroup_taskset *tset)
1459 struct cgroup_subsys_state *css;
1461 struct task_struct *task;
1464 /* used later by cpuset_attach() */
1465 cpuset_attach_old_cs = task_cs(cgroup_taskset_first(tset, &css));
1468 mutex_lock(&cpuset_mutex);
1470 /* allow moving tasks into an empty cpuset if on default hierarchy */
1472 if (!cgroup_subsys_on_dfl(cpuset_cgrp_subsys) &&
1473 (cpumask_empty(cs->cpus_allowed) || nodes_empty(cs->mems_allowed)))
1476 cgroup_taskset_for_each(task, css, tset) {
1477 ret = task_can_attach(task, cs->cpus_allowed);
1480 ret = security_task_setscheduler(task);
1486 * Mark attach is in progress. This makes validate_change() fail
1487 * changes which zero cpus/mems_allowed.
1489 cs->attach_in_progress++;
1492 mutex_unlock(&cpuset_mutex);
1496 static void cpuset_cancel_attach(struct cgroup_taskset *tset)
1498 struct cgroup_subsys_state *css;
1501 cgroup_taskset_first(tset, &css);
1504 mutex_lock(&cpuset_mutex);
1505 css_cs(css)->attach_in_progress--;
1506 mutex_unlock(&cpuset_mutex);
1510 * Protected by cpuset_mutex. cpus_attach is used only by cpuset_attach()
1511 * but we can't allocate it dynamically there. Define it global and
1512 * allocate from cpuset_init().
1514 static cpumask_var_t cpus_attach;
1516 static void cpuset_attach(struct cgroup_taskset *tset)
1518 /* static buf protected by cpuset_mutex */
1519 static nodemask_t cpuset_attach_nodemask_to;
1520 struct task_struct *task;
1521 struct task_struct *leader;
1522 struct cgroup_subsys_state *css;
1524 struct cpuset *oldcs = cpuset_attach_old_cs;
1526 cgroup_taskset_first(tset, &css);
1529 mutex_lock(&cpuset_mutex);
1531 /* prepare for attach */
1532 if (cs == &top_cpuset)
1533 cpumask_copy(cpus_attach, cpu_possible_mask);
1535 guarantee_online_cpus(cs, cpus_attach);
1537 guarantee_online_mems(cs, &cpuset_attach_nodemask_to);
1539 cgroup_taskset_for_each(task, css, tset) {
1541 * can_attach beforehand should guarantee that this doesn't
1542 * fail. TODO: have a better way to handle failure here
1544 WARN_ON_ONCE(set_cpus_allowed_ptr(task, cpus_attach));
1546 cpuset_change_task_nodemask(task, &cpuset_attach_nodemask_to);
1547 cpuset_update_task_spread_flag(cs, task);
1551 * Change mm for all threadgroup leaders. This is expensive and may
1552 * sleep and should be moved outside migration path proper.
1554 cpuset_attach_nodemask_to = cs->effective_mems;
1555 cgroup_taskset_for_each_leader(leader, css, tset) {
1556 struct mm_struct *mm = get_task_mm(leader);
1559 mpol_rebind_mm(mm, &cpuset_attach_nodemask_to);
1562 * old_mems_allowed is the same with mems_allowed
1563 * here, except if this task is being moved
1564 * automatically due to hotplug. In that case
1565 * @mems_allowed has been updated and is empty, so
1566 * @old_mems_allowed is the right nodesets that we
1569 if (is_memory_migrate(cs))
1570 cpuset_migrate_mm(mm, &oldcs->old_mems_allowed,
1571 &cpuset_attach_nodemask_to);
1577 cs->old_mems_allowed = cpuset_attach_nodemask_to;
1579 cs->attach_in_progress--;
1580 if (!cs->attach_in_progress)
1581 wake_up(&cpuset_attach_wq);
1583 mutex_unlock(&cpuset_mutex);
1586 /* The various types of files and directories in a cpuset file system */
1589 FILE_MEMORY_MIGRATE,
1592 FILE_EFFECTIVE_CPULIST,
1593 FILE_EFFECTIVE_MEMLIST,
1597 FILE_SCHED_LOAD_BALANCE,
1598 FILE_SCHED_RELAX_DOMAIN_LEVEL,
1599 FILE_MEMORY_PRESSURE_ENABLED,
1600 FILE_MEMORY_PRESSURE,
1603 } cpuset_filetype_t;
1605 static int cpuset_write_u64(struct cgroup_subsys_state *css, struct cftype *cft,
1608 struct cpuset *cs = css_cs(css);
1609 cpuset_filetype_t type = cft->private;
1612 mutex_lock(&cpuset_mutex);
1613 if (!is_cpuset_online(cs)) {
1619 case FILE_CPU_EXCLUSIVE:
1620 retval = update_flag(CS_CPU_EXCLUSIVE, cs, val);
1622 case FILE_MEM_EXCLUSIVE:
1623 retval = update_flag(CS_MEM_EXCLUSIVE, cs, val);
1625 case FILE_MEM_HARDWALL:
1626 retval = update_flag(CS_MEM_HARDWALL, cs, val);
1628 case FILE_SCHED_LOAD_BALANCE:
1629 retval = update_flag(CS_SCHED_LOAD_BALANCE, cs, val);
1631 case FILE_MEMORY_MIGRATE:
1632 retval = update_flag(CS_MEMORY_MIGRATE, cs, val);
1634 case FILE_MEMORY_PRESSURE_ENABLED:
1635 cpuset_memory_pressure_enabled = !!val;
1637 case FILE_SPREAD_PAGE:
1638 retval = update_flag(CS_SPREAD_PAGE, cs, val);
1640 case FILE_SPREAD_SLAB:
1641 retval = update_flag(CS_SPREAD_SLAB, cs, val);
1648 mutex_unlock(&cpuset_mutex);
1652 static int cpuset_write_s64(struct cgroup_subsys_state *css, struct cftype *cft,
1655 struct cpuset *cs = css_cs(css);
1656 cpuset_filetype_t type = cft->private;
1657 int retval = -ENODEV;
1659 mutex_lock(&cpuset_mutex);
1660 if (!is_cpuset_online(cs))
1664 case FILE_SCHED_RELAX_DOMAIN_LEVEL:
1665 retval = update_relax_domain_level(cs, val);
1672 mutex_unlock(&cpuset_mutex);
1677 * Common handling for a write to a "cpus" or "mems" file.
1679 static ssize_t cpuset_write_resmask(struct kernfs_open_file *of,
1680 char *buf, size_t nbytes, loff_t off)
1682 struct cpuset *cs = css_cs(of_css(of));
1683 struct cpuset *trialcs;
1684 int retval = -ENODEV;
1686 buf = strstrip(buf);
1689 * CPU or memory hotunplug may leave @cs w/o any execution
1690 * resources, in which case the hotplug code asynchronously updates
1691 * configuration and transfers all tasks to the nearest ancestor
1692 * which can execute.
1694 * As writes to "cpus" or "mems" may restore @cs's execution
1695 * resources, wait for the previously scheduled operations before
1696 * proceeding, so that we don't end up keep removing tasks added
1697 * after execution capability is restored.
1699 * cpuset_hotplug_work calls back into cgroup core via
1700 * cgroup_transfer_tasks() and waiting for it from a cgroupfs
1701 * operation like this one can lead to a deadlock through kernfs
1702 * active_ref protection. Let's break the protection. Losing the
1703 * protection is okay as we check whether @cs is online after
1704 * grabbing cpuset_mutex anyway. This only happens on the legacy
1708 kernfs_break_active_protection(of->kn);
1709 flush_work(&cpuset_hotplug_work);
1711 mutex_lock(&cpuset_mutex);
1712 if (!is_cpuset_online(cs))
1715 trialcs = alloc_trial_cpuset(cs);
1721 switch (of_cft(of)->private) {
1723 retval = update_cpumask(cs, trialcs, buf);
1726 retval = update_nodemask(cs, trialcs, buf);
1733 free_trial_cpuset(trialcs);
1735 mutex_unlock(&cpuset_mutex);
1736 kernfs_unbreak_active_protection(of->kn);
1738 flush_workqueue(cpuset_migrate_mm_wq);
1739 return retval ?: nbytes;
1743 * These ascii lists should be read in a single call, by using a user
1744 * buffer large enough to hold the entire map. If read in smaller
1745 * chunks, there is no guarantee of atomicity. Since the display format
1746 * used, list of ranges of sequential numbers, is variable length,
1747 * and since these maps can change value dynamically, one could read
1748 * gibberish by doing partial reads while a list was changing.
1750 static int cpuset_common_seq_show(struct seq_file *sf, void *v)
1752 struct cpuset *cs = css_cs(seq_css(sf));
1753 cpuset_filetype_t type = seq_cft(sf)->private;
1756 spin_lock_irq(&callback_lock);
1760 seq_printf(sf, "%*pbl\n", cpumask_pr_args(cs->cpus_requested));
1763 seq_printf(sf, "%*pbl\n", nodemask_pr_args(&cs->mems_allowed));
1765 case FILE_EFFECTIVE_CPULIST:
1766 seq_printf(sf, "%*pbl\n", cpumask_pr_args(cs->effective_cpus));
1768 case FILE_EFFECTIVE_MEMLIST:
1769 seq_printf(sf, "%*pbl\n", nodemask_pr_args(&cs->effective_mems));
1775 spin_unlock_irq(&callback_lock);
1779 static u64 cpuset_read_u64(struct cgroup_subsys_state *css, struct cftype *cft)
1781 struct cpuset *cs = css_cs(css);
1782 cpuset_filetype_t type = cft->private;
1784 case FILE_CPU_EXCLUSIVE:
1785 return is_cpu_exclusive(cs);
1786 case FILE_MEM_EXCLUSIVE:
1787 return is_mem_exclusive(cs);
1788 case FILE_MEM_HARDWALL:
1789 return is_mem_hardwall(cs);
1790 case FILE_SCHED_LOAD_BALANCE:
1791 return is_sched_load_balance(cs);
1792 case FILE_MEMORY_MIGRATE:
1793 return is_memory_migrate(cs);
1794 case FILE_MEMORY_PRESSURE_ENABLED:
1795 return cpuset_memory_pressure_enabled;
1796 case FILE_MEMORY_PRESSURE:
1797 return fmeter_getrate(&cs->fmeter);
1798 case FILE_SPREAD_PAGE:
1799 return is_spread_page(cs);
1800 case FILE_SPREAD_SLAB:
1801 return is_spread_slab(cs);
1806 /* Unreachable but makes gcc happy */
1810 static s64 cpuset_read_s64(struct cgroup_subsys_state *css, struct cftype *cft)
1812 struct cpuset *cs = css_cs(css);
1813 cpuset_filetype_t type = cft->private;
1815 case FILE_SCHED_RELAX_DOMAIN_LEVEL:
1816 return cs->relax_domain_level;
1821 /* Unrechable but makes gcc happy */
1827 * for the common functions, 'private' gives the type of file
1830 static struct cftype files[] = {
1833 .seq_show = cpuset_common_seq_show,
1834 .write = cpuset_write_resmask,
1835 .max_write_len = (100U + 6 * NR_CPUS),
1836 .private = FILE_CPULIST,
1841 .seq_show = cpuset_common_seq_show,
1842 .write = cpuset_write_resmask,
1843 .max_write_len = (100U + 6 * MAX_NUMNODES),
1844 .private = FILE_MEMLIST,
1848 .name = "effective_cpus",
1849 .seq_show = cpuset_common_seq_show,
1850 .private = FILE_EFFECTIVE_CPULIST,
1854 .name = "effective_mems",
1855 .seq_show = cpuset_common_seq_show,
1856 .private = FILE_EFFECTIVE_MEMLIST,
1860 .name = "cpu_exclusive",
1861 .read_u64 = cpuset_read_u64,
1862 .write_u64 = cpuset_write_u64,
1863 .private = FILE_CPU_EXCLUSIVE,
1867 .name = "mem_exclusive",
1868 .read_u64 = cpuset_read_u64,
1869 .write_u64 = cpuset_write_u64,
1870 .private = FILE_MEM_EXCLUSIVE,
1874 .name = "mem_hardwall",
1875 .read_u64 = cpuset_read_u64,
1876 .write_u64 = cpuset_write_u64,
1877 .private = FILE_MEM_HARDWALL,
1881 .name = "sched_load_balance",
1882 .read_u64 = cpuset_read_u64,
1883 .write_u64 = cpuset_write_u64,
1884 .private = FILE_SCHED_LOAD_BALANCE,
1888 .name = "sched_relax_domain_level",
1889 .read_s64 = cpuset_read_s64,
1890 .write_s64 = cpuset_write_s64,
1891 .private = FILE_SCHED_RELAX_DOMAIN_LEVEL,
1895 .name = "memory_migrate",
1896 .read_u64 = cpuset_read_u64,
1897 .write_u64 = cpuset_write_u64,
1898 .private = FILE_MEMORY_MIGRATE,
1902 .name = "memory_pressure",
1903 .read_u64 = cpuset_read_u64,
1907 .name = "memory_spread_page",
1908 .read_u64 = cpuset_read_u64,
1909 .write_u64 = cpuset_write_u64,
1910 .private = FILE_SPREAD_PAGE,
1914 .name = "memory_spread_slab",
1915 .read_u64 = cpuset_read_u64,
1916 .write_u64 = cpuset_write_u64,
1917 .private = FILE_SPREAD_SLAB,
1921 .name = "memory_pressure_enabled",
1922 .flags = CFTYPE_ONLY_ON_ROOT,
1923 .read_u64 = cpuset_read_u64,
1924 .write_u64 = cpuset_write_u64,
1925 .private = FILE_MEMORY_PRESSURE_ENABLED,
1932 * cpuset_css_alloc - allocate a cpuset css
1933 * cgrp: control group that the new cpuset will be part of
1936 static struct cgroup_subsys_state *
1937 cpuset_css_alloc(struct cgroup_subsys_state *parent_css)
1942 return &top_cpuset.css;
1944 cs = kzalloc(sizeof(*cs), GFP_KERNEL);
1946 return ERR_PTR(-ENOMEM);
1947 if (!alloc_cpumask_var(&cs->cpus_allowed, GFP_KERNEL))
1949 if (!alloc_cpumask_var(&cs->cpus_requested, GFP_KERNEL))
1951 if (!alloc_cpumask_var(&cs->effective_cpus, GFP_KERNEL))
1952 goto free_requested;
1954 set_bit(CS_SCHED_LOAD_BALANCE, &cs->flags);
1955 cpumask_clear(cs->cpus_allowed);
1956 cpumask_clear(cs->cpus_requested);
1957 nodes_clear(cs->mems_allowed);
1958 cpumask_clear(cs->effective_cpus);
1959 nodes_clear(cs->effective_mems);
1960 fmeter_init(&cs->fmeter);
1961 cs->relax_domain_level = -1;
1966 free_cpumask_var(cs->cpus_requested);
1968 free_cpumask_var(cs->cpus_allowed);
1971 return ERR_PTR(-ENOMEM);
1974 static int cpuset_css_online(struct cgroup_subsys_state *css)
1976 struct cpuset *cs = css_cs(css);
1977 struct cpuset *parent = parent_cs(cs);
1978 struct cpuset *tmp_cs;
1979 struct cgroup_subsys_state *pos_css;
1984 mutex_lock(&cpuset_mutex);
1986 set_bit(CS_ONLINE, &cs->flags);
1987 if (is_spread_page(parent))
1988 set_bit(CS_SPREAD_PAGE, &cs->flags);
1989 if (is_spread_slab(parent))
1990 set_bit(CS_SPREAD_SLAB, &cs->flags);
1994 spin_lock_irq(&callback_lock);
1995 if (cgroup_subsys_on_dfl(cpuset_cgrp_subsys)) {
1996 cpumask_copy(cs->effective_cpus, parent->effective_cpus);
1997 cs->effective_mems = parent->effective_mems;
1999 spin_unlock_irq(&callback_lock);
2001 if (!test_bit(CGRP_CPUSET_CLONE_CHILDREN, &css->cgroup->flags))
2005 * Clone @parent's configuration if CGRP_CPUSET_CLONE_CHILDREN is
2006 * set. This flag handling is implemented in cgroup core for
2007 * histrical reasons - the flag may be specified during mount.
2009 * Currently, if any sibling cpusets have exclusive cpus or mem, we
2010 * refuse to clone the configuration - thereby refusing the task to
2011 * be entered, and as a result refusing the sys_unshare() or
2012 * clone() which initiated it. If this becomes a problem for some
2013 * users who wish to allow that scenario, then this could be
2014 * changed to grant parent->cpus_allowed-sibling_cpus_exclusive
2015 * (and likewise for mems) to the new cgroup.
2018 cpuset_for_each_child(tmp_cs, pos_css, parent) {
2019 if (is_mem_exclusive(tmp_cs) || is_cpu_exclusive(tmp_cs)) {
2026 spin_lock_irq(&callback_lock);
2027 cs->mems_allowed = parent->mems_allowed;
2028 cs->effective_mems = parent->mems_allowed;
2029 cpumask_copy(cs->cpus_allowed, parent->cpus_allowed);
2030 cpumask_copy(cs->cpus_requested, parent->cpus_requested);
2031 cpumask_copy(cs->effective_cpus, parent->cpus_allowed);
2032 spin_unlock_irq(&callback_lock);
2034 mutex_unlock(&cpuset_mutex);
2039 * If the cpuset being removed has its flag 'sched_load_balance'
2040 * enabled, then simulate turning sched_load_balance off, which
2041 * will call rebuild_sched_domains_locked().
2044 static void cpuset_css_offline(struct cgroup_subsys_state *css)
2046 struct cpuset *cs = css_cs(css);
2048 mutex_lock(&cpuset_mutex);
2050 if (is_sched_load_balance(cs))
2051 update_flag(CS_SCHED_LOAD_BALANCE, cs, 0);
2054 clear_bit(CS_ONLINE, &cs->flags);
2056 mutex_unlock(&cpuset_mutex);
2059 static void cpuset_css_free(struct cgroup_subsys_state *css)
2061 struct cpuset *cs = css_cs(css);
2063 free_cpumask_var(cs->effective_cpus);
2064 free_cpumask_var(cs->cpus_allowed);
2065 free_cpumask_var(cs->cpus_requested);
2069 static void cpuset_bind(struct cgroup_subsys_state *root_css)
2071 mutex_lock(&cpuset_mutex);
2072 spin_lock_irq(&callback_lock);
2074 if (cgroup_subsys_on_dfl(cpuset_cgrp_subsys)) {
2075 cpumask_copy(top_cpuset.cpus_allowed, cpu_possible_mask);
2076 top_cpuset.mems_allowed = node_possible_map;
2078 cpumask_copy(top_cpuset.cpus_allowed,
2079 top_cpuset.effective_cpus);
2080 top_cpuset.mems_allowed = top_cpuset.effective_mems;
2083 spin_unlock_irq(&callback_lock);
2084 mutex_unlock(&cpuset_mutex);
2087 static int cpuset_allow_attach(struct cgroup_taskset *tset)
2089 const struct cred *cred = current_cred(), *tcred;
2090 struct task_struct *task;
2091 struct cgroup_subsys_state *css;
2093 cgroup_taskset_for_each(task, css, tset) {
2094 tcred = __task_cred(task);
2096 if ((current != task) && !capable(CAP_SYS_ADMIN) &&
2097 cred->euid.val != tcred->uid.val && cred->euid.val != tcred->suid.val)
2104 struct cgroup_subsys cpuset_cgrp_subsys = {
2105 .css_alloc = cpuset_css_alloc,
2106 .css_online = cpuset_css_online,
2107 .css_offline = cpuset_css_offline,
2108 .css_free = cpuset_css_free,
2109 .can_attach = cpuset_can_attach,
2110 .allow_attach = cpuset_allow_attach,
2111 .cancel_attach = cpuset_cancel_attach,
2112 .attach = cpuset_attach,
2113 .post_attach = cpuset_post_attach,
2114 .bind = cpuset_bind,
2115 .legacy_cftypes = files,
2120 * cpuset_init - initialize cpusets at system boot
2122 * Description: Initialize top_cpuset and the cpuset internal file system,
2125 int __init cpuset_init(void)
2129 if (!alloc_cpumask_var(&top_cpuset.cpus_allowed, GFP_KERNEL))
2131 if (!alloc_cpumask_var(&top_cpuset.effective_cpus, GFP_KERNEL))
2133 if (!alloc_cpumask_var(&top_cpuset.cpus_requested, GFP_KERNEL))
2136 cpumask_setall(top_cpuset.cpus_allowed);
2137 cpumask_setall(top_cpuset.cpus_requested);
2138 nodes_setall(top_cpuset.mems_allowed);
2139 cpumask_setall(top_cpuset.effective_cpus);
2140 nodes_setall(top_cpuset.effective_mems);
2142 fmeter_init(&top_cpuset.fmeter);
2143 set_bit(CS_SCHED_LOAD_BALANCE, &top_cpuset.flags);
2144 top_cpuset.relax_domain_level = -1;
2146 err = register_filesystem(&cpuset_fs_type);
2150 if (!alloc_cpumask_var(&cpus_attach, GFP_KERNEL))
2157 * If CPU and/or memory hotplug handlers, below, unplug any CPUs
2158 * or memory nodes, we need to walk over the cpuset hierarchy,
2159 * removing that CPU or node from all cpusets. If this removes the
2160 * last CPU or node from a cpuset, then move the tasks in the empty
2161 * cpuset to its next-highest non-empty parent.
2163 static void remove_tasks_in_empty_cpuset(struct cpuset *cs)
2165 struct cpuset *parent;
2168 * Find its next-highest non-empty parent, (top cpuset
2169 * has online cpus, so can't be empty).
2171 parent = parent_cs(cs);
2172 while (cpumask_empty(parent->cpus_allowed) ||
2173 nodes_empty(parent->mems_allowed))
2174 parent = parent_cs(parent);
2176 if (cgroup_transfer_tasks(parent->css.cgroup, cs->css.cgroup)) {
2177 pr_err("cpuset: failed to transfer tasks out of empty cpuset ");
2178 pr_cont_cgroup_name(cs->css.cgroup);
2184 hotplug_update_tasks_legacy(struct cpuset *cs,
2185 struct cpumask *new_cpus, nodemask_t *new_mems,
2186 bool cpus_updated, bool mems_updated)
2190 spin_lock_irq(&callback_lock);
2191 cpumask_copy(cs->cpus_allowed, new_cpus);
2192 cpumask_copy(cs->effective_cpus, new_cpus);
2193 cs->mems_allowed = *new_mems;
2194 cs->effective_mems = *new_mems;
2195 spin_unlock_irq(&callback_lock);
2198 * Don't call update_tasks_cpumask() if the cpuset becomes empty,
2199 * as the tasks will be migratecd to an ancestor.
2201 if (cpus_updated && !cpumask_empty(cs->cpus_allowed))
2202 update_tasks_cpumask(cs);
2203 if (mems_updated && !nodes_empty(cs->mems_allowed))
2204 update_tasks_nodemask(cs);
2206 is_empty = cpumask_empty(cs->cpus_allowed) ||
2207 nodes_empty(cs->mems_allowed);
2209 mutex_unlock(&cpuset_mutex);
2212 * Move tasks to the nearest ancestor with execution resources,
2213 * This is full cgroup operation which will also call back into
2214 * cpuset. Should be done outside any lock.
2217 remove_tasks_in_empty_cpuset(cs);
2219 mutex_lock(&cpuset_mutex);
2223 hotplug_update_tasks(struct cpuset *cs,
2224 struct cpumask *new_cpus, nodemask_t *new_mems,
2225 bool cpus_updated, bool mems_updated)
2227 if (cpumask_empty(new_cpus))
2228 cpumask_copy(new_cpus, parent_cs(cs)->effective_cpus);
2229 if (nodes_empty(*new_mems))
2230 *new_mems = parent_cs(cs)->effective_mems;
2232 spin_lock_irq(&callback_lock);
2233 cpumask_copy(cs->effective_cpus, new_cpus);
2234 cs->effective_mems = *new_mems;
2235 spin_unlock_irq(&callback_lock);
2238 update_tasks_cpumask(cs);
2240 update_tasks_nodemask(cs);
2244 * cpuset_hotplug_update_tasks - update tasks in a cpuset for hotunplug
2245 * @cs: cpuset in interest
2247 * Compare @cs's cpu and mem masks against top_cpuset and if some have gone
2248 * offline, update @cs accordingly. If @cs ends up with no CPU or memory,
2249 * all its tasks are moved to the nearest ancestor with both resources.
2251 static void cpuset_hotplug_update_tasks(struct cpuset *cs)
2253 static cpumask_t new_cpus;
2254 static nodemask_t new_mems;
2258 wait_event(cpuset_attach_wq, cs->attach_in_progress == 0);
2260 mutex_lock(&cpuset_mutex);
2263 * We have raced with task attaching. We wait until attaching
2264 * is finished, so we won't attach a task to an empty cpuset.
2266 if (cs->attach_in_progress) {
2267 mutex_unlock(&cpuset_mutex);
2271 cpumask_and(&new_cpus, cs->cpus_requested, parent_cs(cs)->effective_cpus);
2272 nodes_and(new_mems, cs->mems_allowed, parent_cs(cs)->effective_mems);
2274 cpus_updated = !cpumask_equal(&new_cpus, cs->effective_cpus);
2275 mems_updated = !nodes_equal(new_mems, cs->effective_mems);
2277 if (cgroup_subsys_on_dfl(cpuset_cgrp_subsys))
2278 hotplug_update_tasks(cs, &new_cpus, &new_mems,
2279 cpus_updated, mems_updated);
2281 hotplug_update_tasks_legacy(cs, &new_cpus, &new_mems,
2282 cpus_updated, mems_updated);
2284 mutex_unlock(&cpuset_mutex);
2288 * cpuset_hotplug_workfn - handle CPU/memory hotunplug for a cpuset
2290 * This function is called after either CPU or memory configuration has
2291 * changed and updates cpuset accordingly. The top_cpuset is always
2292 * synchronized to cpu_active_mask and N_MEMORY, which is necessary in
2293 * order to make cpusets transparent (of no affect) on systems that are
2294 * actively using CPU hotplug but making no active use of cpusets.
2296 * Non-root cpusets are only affected by offlining. If any CPUs or memory
2297 * nodes have been taken down, cpuset_hotplug_update_tasks() is invoked on
2300 * Note that CPU offlining during suspend is ignored. We don't modify
2301 * cpusets across suspend/resume cycles at all.
2303 static void cpuset_hotplug_workfn(struct work_struct *work)
2305 static cpumask_t new_cpus;
2306 static nodemask_t new_mems;
2307 bool cpus_updated, mems_updated;
2308 bool on_dfl = cgroup_subsys_on_dfl(cpuset_cgrp_subsys);
2310 mutex_lock(&cpuset_mutex);
2312 /* fetch the available cpus/mems and find out which changed how */
2313 cpumask_copy(&new_cpus, cpu_active_mask);
2314 new_mems = node_states[N_MEMORY];
2316 cpus_updated = !cpumask_equal(top_cpuset.effective_cpus, &new_cpus);
2317 mems_updated = !nodes_equal(top_cpuset.effective_mems, new_mems);
2319 /* synchronize cpus_allowed to cpu_active_mask */
2321 spin_lock_irq(&callback_lock);
2323 cpumask_copy(top_cpuset.cpus_allowed, &new_cpus);
2324 cpumask_copy(top_cpuset.effective_cpus, &new_cpus);
2325 spin_unlock_irq(&callback_lock);
2326 /* we don't mess with cpumasks of tasks in top_cpuset */
2329 /* synchronize mems_allowed to N_MEMORY */
2331 spin_lock_irq(&callback_lock);
2333 top_cpuset.mems_allowed = new_mems;
2334 top_cpuset.effective_mems = new_mems;
2335 spin_unlock_irq(&callback_lock);
2336 update_tasks_nodemask(&top_cpuset);
2339 mutex_unlock(&cpuset_mutex);
2341 /* if cpus or mems changed, we need to propagate to descendants */
2342 if (cpus_updated || mems_updated) {
2344 struct cgroup_subsys_state *pos_css;
2347 cpuset_for_each_descendant_pre(cs, pos_css, &top_cpuset) {
2348 if (cs == &top_cpuset || !css_tryget_online(&cs->css))
2352 cpuset_hotplug_update_tasks(cs);
2360 /* rebuild sched domains if cpus_allowed has changed */
2362 rebuild_sched_domains();
2365 void cpuset_update_active_cpus(bool cpu_online)
2368 * We're inside cpu hotplug critical region which usually nests
2369 * inside cgroup synchronization. Bounce actual hotplug processing
2370 * to a work item to avoid reverse locking order.
2372 * We still need to do partition_sched_domains() synchronously;
2373 * otherwise, the scheduler will get confused and put tasks to the
2374 * dead CPU. Fall back to the default single domain.
2375 * cpuset_hotplug_workfn() will rebuild it as necessary.
2377 partition_sched_domains(1, NULL, NULL);
2378 schedule_work(&cpuset_hotplug_work);
2382 * Keep top_cpuset.mems_allowed tracking node_states[N_MEMORY].
2383 * Call this routine anytime after node_states[N_MEMORY] changes.
2384 * See cpuset_update_active_cpus() for CPU hotplug handling.
2386 static int cpuset_track_online_nodes(struct notifier_block *self,
2387 unsigned long action, void *arg)
2389 schedule_work(&cpuset_hotplug_work);
2393 static struct notifier_block cpuset_track_online_nodes_nb = {
2394 .notifier_call = cpuset_track_online_nodes,
2395 .priority = 10, /* ??! */
2399 * cpuset_init_smp - initialize cpus_allowed
2401 * Description: Finish top cpuset after cpu, node maps are initialized
2403 void __init cpuset_init_smp(void)
2405 cpumask_copy(top_cpuset.cpus_allowed, cpu_active_mask);
2406 top_cpuset.mems_allowed = node_states[N_MEMORY];
2407 top_cpuset.old_mems_allowed = top_cpuset.mems_allowed;
2409 cpumask_copy(top_cpuset.effective_cpus, cpu_active_mask);
2410 top_cpuset.effective_mems = node_states[N_MEMORY];
2412 register_hotmemory_notifier(&cpuset_track_online_nodes_nb);
2414 cpuset_migrate_mm_wq = alloc_ordered_workqueue("cpuset_migrate_mm", 0);
2415 BUG_ON(!cpuset_migrate_mm_wq);
2419 * cpuset_cpus_allowed - return cpus_allowed mask from a tasks cpuset.
2420 * @tsk: pointer to task_struct from which to obtain cpuset->cpus_allowed.
2421 * @pmask: pointer to struct cpumask variable to receive cpus_allowed set.
2423 * Description: Returns the cpumask_var_t cpus_allowed of the cpuset
2424 * attached to the specified @tsk. Guaranteed to return some non-empty
2425 * subset of cpu_online_mask, even if this means going outside the
2429 void cpuset_cpus_allowed(struct task_struct *tsk, struct cpumask *pmask)
2431 unsigned long flags;
2433 spin_lock_irqsave(&callback_lock, flags);
2435 guarantee_online_cpus(task_cs(tsk), pmask);
2437 spin_unlock_irqrestore(&callback_lock, flags);
2440 void cpuset_cpus_allowed_fallback(struct task_struct *tsk)
2443 do_set_cpus_allowed(tsk, task_cs(tsk)->effective_cpus);
2447 * We own tsk->cpus_allowed, nobody can change it under us.
2449 * But we used cs && cs->cpus_allowed lockless and thus can
2450 * race with cgroup_attach_task() or update_cpumask() and get
2451 * the wrong tsk->cpus_allowed. However, both cases imply the
2452 * subsequent cpuset_change_cpumask()->set_cpus_allowed_ptr()
2453 * which takes task_rq_lock().
2455 * If we are called after it dropped the lock we must see all
2456 * changes in tsk_cs()->cpus_allowed. Otherwise we can temporary
2457 * set any mask even if it is not right from task_cs() pov,
2458 * the pending set_cpus_allowed_ptr() will fix things.
2460 * select_fallback_rq() will fix things ups and set cpu_possible_mask
2465 void __init cpuset_init_current_mems_allowed(void)
2467 nodes_setall(current->mems_allowed);
2471 * cpuset_mems_allowed - return mems_allowed mask from a tasks cpuset.
2472 * @tsk: pointer to task_struct from which to obtain cpuset->mems_allowed.
2474 * Description: Returns the nodemask_t mems_allowed of the cpuset
2475 * attached to the specified @tsk. Guaranteed to return some non-empty
2476 * subset of node_states[N_MEMORY], even if this means going outside the
2480 nodemask_t cpuset_mems_allowed(struct task_struct *tsk)
2483 unsigned long flags;
2485 spin_lock_irqsave(&callback_lock, flags);
2487 guarantee_online_mems(task_cs(tsk), &mask);
2489 spin_unlock_irqrestore(&callback_lock, flags);
2495 * cpuset_nodemask_valid_mems_allowed - check nodemask vs. curremt mems_allowed
2496 * @nodemask: the nodemask to be checked
2498 * Are any of the nodes in the nodemask allowed in current->mems_allowed?
2500 int cpuset_nodemask_valid_mems_allowed(nodemask_t *nodemask)
2502 return nodes_intersects(*nodemask, current->mems_allowed);
2506 * nearest_hardwall_ancestor() - Returns the nearest mem_exclusive or
2507 * mem_hardwall ancestor to the specified cpuset. Call holding
2508 * callback_lock. If no ancestor is mem_exclusive or mem_hardwall
2509 * (an unusual configuration), then returns the root cpuset.
2511 static struct cpuset *nearest_hardwall_ancestor(struct cpuset *cs)
2513 while (!(is_mem_exclusive(cs) || is_mem_hardwall(cs)) && parent_cs(cs))
2519 * cpuset_node_allowed - Can we allocate on a memory node?
2520 * @node: is this an allowed node?
2521 * @gfp_mask: memory allocation flags
2523 * If we're in interrupt, yes, we can always allocate. If @node is set in
2524 * current's mems_allowed, yes. If it's not a __GFP_HARDWALL request and this
2525 * node is set in the nearest hardwalled cpuset ancestor to current's cpuset,
2526 * yes. If current has access to memory reserves due to TIF_MEMDIE, yes.
2529 * GFP_USER allocations are marked with the __GFP_HARDWALL bit,
2530 * and do not allow allocations outside the current tasks cpuset
2531 * unless the task has been OOM killed as is marked TIF_MEMDIE.
2532 * GFP_KERNEL allocations are not so marked, so can escape to the
2533 * nearest enclosing hardwalled ancestor cpuset.
2535 * Scanning up parent cpusets requires callback_lock. The
2536 * __alloc_pages() routine only calls here with __GFP_HARDWALL bit
2537 * _not_ set if it's a GFP_KERNEL allocation, and all nodes in the
2538 * current tasks mems_allowed came up empty on the first pass over
2539 * the zonelist. So only GFP_KERNEL allocations, if all nodes in the
2540 * cpuset are short of memory, might require taking the callback_lock.
2542 * The first call here from mm/page_alloc:get_page_from_freelist()
2543 * has __GFP_HARDWALL set in gfp_mask, enforcing hardwall cpusets,
2544 * so no allocation on a node outside the cpuset is allowed (unless
2545 * in interrupt, of course).
2547 * The second pass through get_page_from_freelist() doesn't even call
2548 * here for GFP_ATOMIC calls. For those calls, the __alloc_pages()
2549 * variable 'wait' is not set, and the bit ALLOC_CPUSET is not set
2550 * in alloc_flags. That logic and the checks below have the combined
2552 * in_interrupt - any node ok (current task context irrelevant)
2553 * GFP_ATOMIC - any node ok
2554 * TIF_MEMDIE - any node ok
2555 * GFP_KERNEL - any node in enclosing hardwalled cpuset ok
2556 * GFP_USER - only nodes in current tasks mems allowed ok.
2558 int __cpuset_node_allowed(int node, gfp_t gfp_mask)
2560 struct cpuset *cs; /* current cpuset ancestors */
2561 int allowed; /* is allocation in zone z allowed? */
2562 unsigned long flags;
2566 if (node_isset(node, current->mems_allowed))
2569 * Allow tasks that have access to memory reserves because they have
2570 * been OOM killed to get memory anywhere.
2572 if (unlikely(test_thread_flag(TIF_MEMDIE)))
2574 if (gfp_mask & __GFP_HARDWALL) /* If hardwall request, stop here */
2577 if (current->flags & PF_EXITING) /* Let dying task have memory */
2580 /* Not hardwall and node outside mems_allowed: scan up cpusets */
2581 spin_lock_irqsave(&callback_lock, flags);
2584 cs = nearest_hardwall_ancestor(task_cs(current));
2585 allowed = node_isset(node, cs->mems_allowed);
2588 spin_unlock_irqrestore(&callback_lock, flags);
2593 * cpuset_mem_spread_node() - On which node to begin search for a file page
2594 * cpuset_slab_spread_node() - On which node to begin search for a slab page
2596 * If a task is marked PF_SPREAD_PAGE or PF_SPREAD_SLAB (as for
2597 * tasks in a cpuset with is_spread_page or is_spread_slab set),
2598 * and if the memory allocation used cpuset_mem_spread_node()
2599 * to determine on which node to start looking, as it will for
2600 * certain page cache or slab cache pages such as used for file
2601 * system buffers and inode caches, then instead of starting on the
2602 * local node to look for a free page, rather spread the starting
2603 * node around the tasks mems_allowed nodes.
2605 * We don't have to worry about the returned node being offline
2606 * because "it can't happen", and even if it did, it would be ok.
2608 * The routines calling guarantee_online_mems() are careful to
2609 * only set nodes in task->mems_allowed that are online. So it
2610 * should not be possible for the following code to return an
2611 * offline node. But if it did, that would be ok, as this routine
2612 * is not returning the node where the allocation must be, only
2613 * the node where the search should start. The zonelist passed to
2614 * __alloc_pages() will include all nodes. If the slab allocator
2615 * is passed an offline node, it will fall back to the local node.
2616 * See kmem_cache_alloc_node().
2619 static int cpuset_spread_node(int *rotor)
2623 node = next_node(*rotor, current->mems_allowed);
2624 if (node == MAX_NUMNODES)
2625 node = first_node(current->mems_allowed);
2630 int cpuset_mem_spread_node(void)
2632 if (current->cpuset_mem_spread_rotor == NUMA_NO_NODE)
2633 current->cpuset_mem_spread_rotor =
2634 node_random(¤t->mems_allowed);
2636 return cpuset_spread_node(¤t->cpuset_mem_spread_rotor);
2639 int cpuset_slab_spread_node(void)
2641 if (current->cpuset_slab_spread_rotor == NUMA_NO_NODE)
2642 current->cpuset_slab_spread_rotor =
2643 node_random(¤t->mems_allowed);
2645 return cpuset_spread_node(¤t->cpuset_slab_spread_rotor);
2648 EXPORT_SYMBOL_GPL(cpuset_mem_spread_node);
2651 * cpuset_mems_allowed_intersects - Does @tsk1's mems_allowed intersect @tsk2's?
2652 * @tsk1: pointer to task_struct of some task.
2653 * @tsk2: pointer to task_struct of some other task.
2655 * Description: Return true if @tsk1's mems_allowed intersects the
2656 * mems_allowed of @tsk2. Used by the OOM killer to determine if
2657 * one of the task's memory usage might impact the memory available
2661 int cpuset_mems_allowed_intersects(const struct task_struct *tsk1,
2662 const struct task_struct *tsk2)
2664 return nodes_intersects(tsk1->mems_allowed, tsk2->mems_allowed);
2668 * cpuset_print_current_mems_allowed - prints current's cpuset and mems_allowed
2670 * Description: Prints current's name, cpuset name, and cached copy of its
2671 * mems_allowed to the kernel log.
2673 void cpuset_print_current_mems_allowed(void)
2675 struct cgroup *cgrp;
2679 cgrp = task_cs(current)->css.cgroup;
2680 pr_info("%s cpuset=", current->comm);
2681 pr_cont_cgroup_name(cgrp);
2682 pr_cont(" mems_allowed=%*pbl\n",
2683 nodemask_pr_args(¤t->mems_allowed));
2689 * Collection of memory_pressure is suppressed unless
2690 * this flag is enabled by writing "1" to the special
2691 * cpuset file 'memory_pressure_enabled' in the root cpuset.
2694 int cpuset_memory_pressure_enabled __read_mostly;
2697 * cpuset_memory_pressure_bump - keep stats of per-cpuset reclaims.
2699 * Keep a running average of the rate of synchronous (direct)
2700 * page reclaim efforts initiated by tasks in each cpuset.
2702 * This represents the rate at which some task in the cpuset
2703 * ran low on memory on all nodes it was allowed to use, and
2704 * had to enter the kernels page reclaim code in an effort to
2705 * create more free memory by tossing clean pages or swapping
2706 * or writing dirty pages.
2708 * Display to user space in the per-cpuset read-only file
2709 * "memory_pressure". Value displayed is an integer
2710 * representing the recent rate of entry into the synchronous
2711 * (direct) page reclaim by any task attached to the cpuset.
2714 void __cpuset_memory_pressure_bump(void)
2717 fmeter_markevent(&task_cs(current)->fmeter);
2721 #ifdef CONFIG_PROC_PID_CPUSET
2723 * proc_cpuset_show()
2724 * - Print tasks cpuset path into seq_file.
2725 * - Used for /proc/<pid>/cpuset.
2726 * - No need to task_lock(tsk) on this tsk->cpuset reference, as it
2727 * doesn't really matter if tsk->cpuset changes after we read it,
2728 * and we take cpuset_mutex, keeping cpuset_attach() from changing it
2731 int proc_cpuset_show(struct seq_file *m, struct pid_namespace *ns,
2732 struct pid *pid, struct task_struct *tsk)
2735 struct cgroup_subsys_state *css;
2739 buf = kmalloc(PATH_MAX, GFP_KERNEL);
2743 retval = -ENAMETOOLONG;
2745 css = task_css(tsk, cpuset_cgrp_id);
2746 p = cgroup_path(css->cgroup, buf, PATH_MAX);
2758 #endif /* CONFIG_PROC_PID_CPUSET */
2760 /* Display task mems_allowed in /proc/<pid>/status file. */
2761 void cpuset_task_status_allowed(struct seq_file *m, struct task_struct *task)
2763 seq_printf(m, "Mems_allowed:\t%*pb\n",
2764 nodemask_pr_args(&task->mems_allowed));
2765 seq_printf(m, "Mems_allowed_list:\t%*pbl\n",
2766 nodemask_pr_args(&task->mems_allowed));