4 * Processor and Memory placement constraints for sets of tasks.
6 * Copyright (C) 2003 BULL SA.
7 * Copyright (C) 2004-2007 Silicon Graphics, Inc.
8 * Copyright (C) 2006 Google, Inc
10 * Portions derived from Patrick Mochel's sysfs code.
11 * sysfs is Copyright (c) 2001-3 Patrick Mochel
13 * 2003-10-10 Written by Simon Derr.
14 * 2003-10-22 Updates by Stephen Hemminger.
15 * 2004 May-July Rework by Paul Jackson.
16 * 2006 Rework by Paul Menage to use generic cgroups
17 * 2008 Rework of the scheduler domains and CPU hotplug handling
20 * This file is subject to the terms and conditions of the GNU General Public
21 * License. See the file COPYING in the main directory of the Linux
22 * distribution for more details.
25 #include <linux/cpu.h>
26 #include <linux/cpumask.h>
27 #include <linux/cpuset.h>
28 #include <linux/err.h>
29 #include <linux/errno.h>
30 #include <linux/file.h>
32 #include <linux/init.h>
33 #include <linux/interrupt.h>
34 #include <linux/kernel.h>
35 #include <linux/kmod.h>
36 #include <linux/list.h>
37 #include <linux/mempolicy.h>
39 #include <linux/memory.h>
40 #include <linux/export.h>
41 #include <linux/mount.h>
42 #include <linux/namei.h>
43 #include <linux/pagemap.h>
44 #include <linux/proc_fs.h>
45 #include <linux/rcupdate.h>
46 #include <linux/sched.h>
47 #include <linux/sched/mm.h>
48 #include <linux/sched/task.h>
49 #include <linux/seq_file.h>
50 #include <linux/security.h>
51 #include <linux/slab.h>
52 #include <linux/spinlock.h>
53 #include <linux/stat.h>
54 #include <linux/string.h>
55 #include <linux/time.h>
56 #include <linux/time64.h>
57 #include <linux/backing-dev.h>
58 #include <linux/sort.h>
59 #include <linux/oom.h>
61 #include <linux/uaccess.h>
62 #include <linux/atomic.h>
63 #include <linux/mutex.h>
64 #include <linux/cgroup.h>
65 #include <linux/wait.h>
67 DEFINE_STATIC_KEY_FALSE(cpusets_pre_enable_key);
68 DEFINE_STATIC_KEY_FALSE(cpusets_enabled_key);
70 /* See "Frequency meter" comments, below. */
73 int cnt; /* unprocessed events count */
74 int val; /* most recent output value */
75 time64_t time; /* clock (secs) when val computed */
76 spinlock_t lock; /* guards read or write of above */
80 struct cgroup_subsys_state css;
82 unsigned long flags; /* "unsigned long" so bitops work */
85 * On default hierarchy:
87 * The user-configured masks can only be changed by writing to
88 * cpuset.cpus and cpuset.mems, and won't be limited by the
91 * The effective masks is the real masks that apply to the tasks
92 * in the cpuset. They may be changed if the configured masks are
93 * changed or hotplug happens.
95 * effective_mask == configured_mask & parent's effective_mask,
96 * and if it ends up empty, it will inherit the parent's mask.
101 * The user-configured masks are always the same with effective masks.
104 /* user-configured CPUs and Memory Nodes allow to tasks */
105 cpumask_var_t cpus_allowed;
106 nodemask_t mems_allowed;
108 /* effective CPUs and Memory Nodes allow to tasks */
109 cpumask_var_t effective_cpus;
110 nodemask_t effective_mems;
113 * This is old Memory Nodes tasks took on.
115 * - top_cpuset.old_mems_allowed is initialized to mems_allowed.
116 * - A new cpuset's old_mems_allowed is initialized when some
117 * task is moved into it.
118 * - old_mems_allowed is used in cpuset_migrate_mm() when we change
119 * cpuset.mems_allowed and have tasks' nodemask updated, and
120 * then old_mems_allowed is updated to mems_allowed.
122 nodemask_t old_mems_allowed;
124 struct fmeter fmeter; /* memory_pressure filter */
127 * Tasks are being attached to this cpuset. Used to prevent
128 * zeroing cpus/mems_allowed between ->can_attach() and ->attach().
130 int attach_in_progress;
132 /* partition number for rebuild_sched_domains() */
135 /* for custom sched domain */
136 int relax_domain_level;
139 static inline struct cpuset *css_cs(struct cgroup_subsys_state *css)
141 return css ? container_of(css, struct cpuset, css) : NULL;
144 /* Retrieve the cpuset for a task */
145 static inline struct cpuset *task_cs(struct task_struct *task)
147 return css_cs(task_css(task, cpuset_cgrp_id));
150 static inline struct cpuset *parent_cs(struct cpuset *cs)
152 return css_cs(cs->css.parent);
156 static inline bool task_has_mempolicy(struct task_struct *task)
158 return task->mempolicy;
161 static inline bool task_has_mempolicy(struct task_struct *task)
168 /* bits in struct cpuset flags field */
175 CS_SCHED_LOAD_BALANCE,
180 /* convenient tests for these bits */
181 static inline bool is_cpuset_online(struct cpuset *cs)
183 return test_bit(CS_ONLINE, &cs->flags) && !css_is_dying(&cs->css);
186 static inline int is_cpu_exclusive(const struct cpuset *cs)
188 return test_bit(CS_CPU_EXCLUSIVE, &cs->flags);
191 static inline int is_mem_exclusive(const struct cpuset *cs)
193 return test_bit(CS_MEM_EXCLUSIVE, &cs->flags);
196 static inline int is_mem_hardwall(const struct cpuset *cs)
198 return test_bit(CS_MEM_HARDWALL, &cs->flags);
201 static inline int is_sched_load_balance(const struct cpuset *cs)
203 return test_bit(CS_SCHED_LOAD_BALANCE, &cs->flags);
206 static inline int is_memory_migrate(const struct cpuset *cs)
208 return test_bit(CS_MEMORY_MIGRATE, &cs->flags);
211 static inline int is_spread_page(const struct cpuset *cs)
213 return test_bit(CS_SPREAD_PAGE, &cs->flags);
216 static inline int is_spread_slab(const struct cpuset *cs)
218 return test_bit(CS_SPREAD_SLAB, &cs->flags);
221 static struct cpuset top_cpuset = {
222 .flags = ((1 << CS_ONLINE) | (1 << CS_CPU_EXCLUSIVE) |
223 (1 << CS_MEM_EXCLUSIVE)),
227 * cpuset_for_each_child - traverse online children of a cpuset
228 * @child_cs: loop cursor pointing to the current child
229 * @pos_css: used for iteration
230 * @parent_cs: target cpuset to walk children of
232 * Walk @child_cs through the online children of @parent_cs. Must be used
233 * with RCU read locked.
235 #define cpuset_for_each_child(child_cs, pos_css, parent_cs) \
236 css_for_each_child((pos_css), &(parent_cs)->css) \
237 if (is_cpuset_online(((child_cs) = css_cs((pos_css)))))
240 * cpuset_for_each_descendant_pre - pre-order walk of a cpuset's descendants
241 * @des_cs: loop cursor pointing to the current descendant
242 * @pos_css: used for iteration
243 * @root_cs: target cpuset to walk ancestor of
245 * Walk @des_cs through the online descendants of @root_cs. Must be used
246 * with RCU read locked. The caller may modify @pos_css by calling
247 * css_rightmost_descendant() to skip subtree. @root_cs is included in the
248 * iteration and the first node to be visited.
250 #define cpuset_for_each_descendant_pre(des_cs, pos_css, root_cs) \
251 css_for_each_descendant_pre((pos_css), &(root_cs)->css) \
252 if (is_cpuset_online(((des_cs) = css_cs((pos_css)))))
255 * There are two global locks guarding cpuset structures - cpuset_mutex and
256 * callback_lock. We also require taking task_lock() when dereferencing a
257 * task's cpuset pointer. See "The task_lock() exception", at the end of this
260 * A task must hold both locks to modify cpusets. If a task holds
261 * cpuset_mutex, then it blocks others wanting that mutex, ensuring that it
262 * is the only task able to also acquire callback_lock and be able to
263 * modify cpusets. It can perform various checks on the cpuset structure
264 * first, knowing nothing will change. It can also allocate memory while
265 * just holding cpuset_mutex. While it is performing these checks, various
266 * callback routines can briefly acquire callback_lock to query cpusets.
267 * Once it is ready to make the changes, it takes callback_lock, blocking
270 * Calls to the kernel memory allocator can not be made while holding
271 * callback_lock, as that would risk double tripping on callback_lock
272 * from one of the callbacks into the cpuset code from within
275 * If a task is only holding callback_lock, then it has read-only
278 * Now, the task_struct fields mems_allowed and mempolicy may be changed
279 * by other task, we use alloc_lock in the task_struct fields to protect
282 * The cpuset_common_file_read() handlers only hold callback_lock across
283 * small pieces of code, such as when reading out possibly multi-word
284 * cpumasks and nodemasks.
286 * Accessing a task's cpuset should be done in accordance with the
287 * guidelines for accessing subsystem state in kernel/cgroup.c
290 static DEFINE_MUTEX(cpuset_mutex);
291 static DEFINE_SPINLOCK(callback_lock);
293 static struct workqueue_struct *cpuset_migrate_mm_wq;
296 * CPU / memory hotplug is handled asynchronously.
298 static void cpuset_hotplug_workfn(struct work_struct *work);
299 static DECLARE_WORK(cpuset_hotplug_work, cpuset_hotplug_workfn);
301 static DECLARE_WAIT_QUEUE_HEAD(cpuset_attach_wq);
304 * Cgroup v2 behavior is used when on default hierarchy or the
305 * cgroup_v2_mode flag is set.
307 static inline bool is_in_v2_mode(void)
309 return cgroup_subsys_on_dfl(cpuset_cgrp_subsys) ||
310 (cpuset_cgrp_subsys.root->flags & CGRP_ROOT_CPUSET_V2_MODE);
314 * This is ugly, but preserves the userspace API for existing cpuset
315 * users. If someone tries to mount the "cpuset" filesystem, we
316 * silently switch it to mount "cgroup" instead
318 static struct dentry *cpuset_mount(struct file_system_type *fs_type,
319 int flags, const char *unused_dev_name, void *data)
321 struct file_system_type *cgroup_fs = get_fs_type("cgroup");
322 struct dentry *ret = ERR_PTR(-ENODEV);
326 "release_agent=/sbin/cpuset_release_agent";
327 ret = cgroup_fs->mount(cgroup_fs, flags,
328 unused_dev_name, mountopts);
329 put_filesystem(cgroup_fs);
334 static struct file_system_type cpuset_fs_type = {
336 .mount = cpuset_mount,
340 * Return in pmask the portion of a cpusets's cpus_allowed that
341 * are online. If none are online, walk up the cpuset hierarchy
342 * until we find one that does have some online cpus.
344 * One way or another, we guarantee to return some non-empty subset
345 * of cpu_online_mask.
347 * Call with callback_lock or cpuset_mutex held.
349 static void guarantee_online_cpus(struct cpuset *cs, struct cpumask *pmask)
351 while (!cpumask_intersects(cs->effective_cpus, cpu_online_mask)) {
355 * The top cpuset doesn't have any online cpu as a
356 * consequence of a race between cpuset_hotplug_work
357 * and cpu hotplug notifier. But we know the top
358 * cpuset's effective_cpus is on its way to to be
359 * identical to cpu_online_mask.
361 cpumask_copy(pmask, cpu_online_mask);
365 cpumask_and(pmask, cs->effective_cpus, cpu_online_mask);
369 * Return in *pmask the portion of a cpusets's mems_allowed that
370 * are online, with memory. If none are online with memory, walk
371 * up the cpuset hierarchy until we find one that does have some
372 * online mems. The top cpuset always has some mems online.
374 * One way or another, we guarantee to return some non-empty subset
375 * of node_states[N_MEMORY].
377 * Call with callback_lock or cpuset_mutex held.
379 static void guarantee_online_mems(struct cpuset *cs, nodemask_t *pmask)
381 while (!nodes_intersects(cs->effective_mems, node_states[N_MEMORY]))
383 nodes_and(*pmask, cs->effective_mems, node_states[N_MEMORY]);
387 * update task's spread flag if cpuset's page/slab spread flag is set
389 * Call with callback_lock or cpuset_mutex held.
391 static void cpuset_update_task_spread_flag(struct cpuset *cs,
392 struct task_struct *tsk)
394 if (is_spread_page(cs))
395 task_set_spread_page(tsk);
397 task_clear_spread_page(tsk);
399 if (is_spread_slab(cs))
400 task_set_spread_slab(tsk);
402 task_clear_spread_slab(tsk);
406 * is_cpuset_subset(p, q) - Is cpuset p a subset of cpuset q?
408 * One cpuset is a subset of another if all its allowed CPUs and
409 * Memory Nodes are a subset of the other, and its exclusive flags
410 * are only set if the other's are set. Call holding cpuset_mutex.
413 static int is_cpuset_subset(const struct cpuset *p, const struct cpuset *q)
415 return cpumask_subset(p->cpus_allowed, q->cpus_allowed) &&
416 nodes_subset(p->mems_allowed, q->mems_allowed) &&
417 is_cpu_exclusive(p) <= is_cpu_exclusive(q) &&
418 is_mem_exclusive(p) <= is_mem_exclusive(q);
422 * alloc_trial_cpuset - allocate a trial cpuset
423 * @cs: the cpuset that the trial cpuset duplicates
425 static struct cpuset *alloc_trial_cpuset(struct cpuset *cs)
427 struct cpuset *trial;
429 trial = kmemdup(cs, sizeof(*cs), GFP_KERNEL);
433 if (!alloc_cpumask_var(&trial->cpus_allowed, GFP_KERNEL))
435 if (!alloc_cpumask_var(&trial->effective_cpus, GFP_KERNEL))
438 cpumask_copy(trial->cpus_allowed, cs->cpus_allowed);
439 cpumask_copy(trial->effective_cpus, cs->effective_cpus);
443 free_cpumask_var(trial->cpus_allowed);
450 * free_trial_cpuset - free the trial cpuset
451 * @trial: the trial cpuset to be freed
453 static void free_trial_cpuset(struct cpuset *trial)
455 free_cpumask_var(trial->effective_cpus);
456 free_cpumask_var(trial->cpus_allowed);
461 * validate_change() - Used to validate that any proposed cpuset change
462 * follows the structural rules for cpusets.
464 * If we replaced the flag and mask values of the current cpuset
465 * (cur) with those values in the trial cpuset (trial), would
466 * our various subset and exclusive rules still be valid? Presumes
469 * 'cur' is the address of an actual, in-use cpuset. Operations
470 * such as list traversal that depend on the actual address of the
471 * cpuset in the list must use cur below, not trial.
473 * 'trial' is the address of bulk structure copy of cur, with
474 * perhaps one or more of the fields cpus_allowed, mems_allowed,
475 * or flags changed to new, trial values.
477 * Return 0 if valid, -errno if not.
480 static int validate_change(struct cpuset *cur, struct cpuset *trial)
482 struct cgroup_subsys_state *css;
483 struct cpuset *c, *par;
488 /* Each of our child cpusets must be a subset of us */
490 cpuset_for_each_child(c, css, cur)
491 if (!is_cpuset_subset(c, trial))
494 /* Remaining checks don't apply to root cpuset */
496 if (cur == &top_cpuset)
499 par = parent_cs(cur);
501 /* On legacy hiearchy, we must be a subset of our parent cpuset. */
503 if (!is_in_v2_mode() && !is_cpuset_subset(trial, par))
507 * If either I or some sibling (!= me) is exclusive, we can't
511 cpuset_for_each_child(c, css, par) {
512 if ((is_cpu_exclusive(trial) || is_cpu_exclusive(c)) &&
514 cpumask_intersects(trial->cpus_allowed, c->cpus_allowed))
516 if ((is_mem_exclusive(trial) || is_mem_exclusive(c)) &&
518 nodes_intersects(trial->mems_allowed, c->mems_allowed))
523 * Cpusets with tasks - existing or newly being attached - can't
524 * be changed to have empty cpus_allowed or mems_allowed.
527 if ((cgroup_is_populated(cur->css.cgroup) || cur->attach_in_progress)) {
528 if (!cpumask_empty(cur->cpus_allowed) &&
529 cpumask_empty(trial->cpus_allowed))
531 if (!nodes_empty(cur->mems_allowed) &&
532 nodes_empty(trial->mems_allowed))
537 * We can't shrink if we won't have enough room for SCHED_DEADLINE
541 if (is_cpu_exclusive(cur) &&
542 !cpuset_cpumask_can_shrink(cur->cpus_allowed,
543 trial->cpus_allowed))
554 * Helper routine for generate_sched_domains().
555 * Do cpusets a, b have overlapping effective cpus_allowed masks?
557 static int cpusets_overlap(struct cpuset *a, struct cpuset *b)
559 return cpumask_intersects(a->effective_cpus, b->effective_cpus);
563 update_domain_attr(struct sched_domain_attr *dattr, struct cpuset *c)
565 if (dattr->relax_domain_level < c->relax_domain_level)
566 dattr->relax_domain_level = c->relax_domain_level;
570 static void update_domain_attr_tree(struct sched_domain_attr *dattr,
571 struct cpuset *root_cs)
574 struct cgroup_subsys_state *pos_css;
577 cpuset_for_each_descendant_pre(cp, pos_css, root_cs) {
578 /* skip the whole subtree if @cp doesn't have any CPU */
579 if (cpumask_empty(cp->cpus_allowed)) {
580 pos_css = css_rightmost_descendant(pos_css);
584 if (is_sched_load_balance(cp))
585 update_domain_attr(dattr, cp);
590 /* Must be called with cpuset_mutex held. */
591 static inline int nr_cpusets(void)
593 /* jump label reference count + the top-level cpuset */
594 return static_key_count(&cpusets_enabled_key.key) + 1;
598 * generate_sched_domains()
600 * This function builds a partial partition of the systems CPUs
601 * A 'partial partition' is a set of non-overlapping subsets whose
602 * union is a subset of that set.
603 * The output of this function needs to be passed to kernel/sched/core.c
604 * partition_sched_domains() routine, which will rebuild the scheduler's
605 * load balancing domains (sched domains) as specified by that partial
608 * See "What is sched_load_balance" in Documentation/cgroups/cpusets.txt
609 * for a background explanation of this.
611 * Does not return errors, on the theory that the callers of this
612 * routine would rather not worry about failures to rebuild sched
613 * domains when operating in the severe memory shortage situations
614 * that could cause allocation failures below.
616 * Must be called with cpuset_mutex held.
618 * The three key local variables below are:
619 * q - a linked-list queue of cpuset pointers, used to implement a
620 * top-down scan of all cpusets. This scan loads a pointer
621 * to each cpuset marked is_sched_load_balance into the
622 * array 'csa'. For our purposes, rebuilding the schedulers
623 * sched domains, we can ignore !is_sched_load_balance cpusets.
624 * csa - (for CpuSet Array) Array of pointers to all the cpusets
625 * that need to be load balanced, for convenient iterative
626 * access by the subsequent code that finds the best partition,
627 * i.e the set of domains (subsets) of CPUs such that the
628 * cpus_allowed of every cpuset marked is_sched_load_balance
629 * is a subset of one of these domains, while there are as
630 * many such domains as possible, each as small as possible.
631 * doms - Conversion of 'csa' to an array of cpumasks, for passing to
632 * the kernel/sched/core.c routine partition_sched_domains() in a
633 * convenient format, that can be easily compared to the prior
634 * value to determine what partition elements (sched domains)
635 * were changed (added or removed.)
637 * Finding the best partition (set of domains):
638 * The triple nested loops below over i, j, k scan over the
639 * load balanced cpusets (using the array of cpuset pointers in
640 * csa[]) looking for pairs of cpusets that have overlapping
641 * cpus_allowed, but which don't have the same 'pn' partition
642 * number and gives them in the same partition number. It keeps
643 * looping on the 'restart' label until it can no longer find
646 * The union of the cpus_allowed masks from the set of
647 * all cpusets having the same 'pn' value then form the one
648 * element of the partition (one sched domain) to be passed to
649 * partition_sched_domains().
651 static int generate_sched_domains(cpumask_var_t **domains,
652 struct sched_domain_attr **attributes)
654 struct cpuset *cp; /* scans q */
655 struct cpuset **csa; /* array of all cpuset ptrs */
656 int csn; /* how many cpuset ptrs in csa so far */
657 int i, j, k; /* indices for partition finding loops */
658 cpumask_var_t *doms; /* resulting partition; i.e. sched domains */
659 cpumask_var_t non_isolated_cpus; /* load balanced CPUs */
660 struct sched_domain_attr *dattr; /* attributes for custom domains */
661 int ndoms = 0; /* number of sched domains in result */
662 int nslot; /* next empty doms[] struct cpumask slot */
663 struct cgroup_subsys_state *pos_css;
669 if (!alloc_cpumask_var(&non_isolated_cpus, GFP_KERNEL))
671 cpumask_andnot(non_isolated_cpus, cpu_possible_mask, cpu_isolated_map);
673 /* Special case for the 99% of systems with one, full, sched domain */
674 if (is_sched_load_balance(&top_cpuset)) {
676 doms = alloc_sched_domains(ndoms);
680 dattr = kmalloc(sizeof(struct sched_domain_attr), GFP_KERNEL);
682 *dattr = SD_ATTR_INIT;
683 update_domain_attr_tree(dattr, &top_cpuset);
685 cpumask_and(doms[0], top_cpuset.effective_cpus,
691 csa = kmalloc(nr_cpusets() * sizeof(cp), GFP_KERNEL);
697 cpuset_for_each_descendant_pre(cp, pos_css, &top_cpuset) {
698 if (cp == &top_cpuset)
701 * Continue traversing beyond @cp iff @cp has some CPUs and
702 * isn't load balancing. The former is obvious. The
703 * latter: All child cpusets contain a subset of the
704 * parent's cpus, so just skip them, and then we call
705 * update_domain_attr_tree() to calc relax_domain_level of
706 * the corresponding sched domain.
708 if (!cpumask_empty(cp->cpus_allowed) &&
709 !(is_sched_load_balance(cp) &&
710 cpumask_intersects(cp->cpus_allowed, non_isolated_cpus)))
713 if (is_sched_load_balance(cp))
716 /* skip @cp's subtree */
717 pos_css = css_rightmost_descendant(pos_css);
721 for (i = 0; i < csn; i++)
726 /* Find the best partition (set of sched domains) */
727 for (i = 0; i < csn; i++) {
728 struct cpuset *a = csa[i];
731 for (j = 0; j < csn; j++) {
732 struct cpuset *b = csa[j];
735 if (apn != bpn && cpusets_overlap(a, b)) {
736 for (k = 0; k < csn; k++) {
737 struct cpuset *c = csa[k];
742 ndoms--; /* one less element */
749 * Now we know how many domains to create.
750 * Convert <csn, csa> to <ndoms, doms> and populate cpu masks.
752 doms = alloc_sched_domains(ndoms);
757 * The rest of the code, including the scheduler, can deal with
758 * dattr==NULL case. No need to abort if alloc fails.
760 dattr = kmalloc(ndoms * sizeof(struct sched_domain_attr), GFP_KERNEL);
762 for (nslot = 0, i = 0; i < csn; i++) {
763 struct cpuset *a = csa[i];
768 /* Skip completed partitions */
774 if (nslot == ndoms) {
775 static int warnings = 10;
777 pr_warn("rebuild_sched_domains confused: nslot %d, ndoms %d, csn %d, i %d, apn %d\n",
778 nslot, ndoms, csn, i, apn);
786 *(dattr + nslot) = SD_ATTR_INIT;
787 for (j = i; j < csn; j++) {
788 struct cpuset *b = csa[j];
791 cpumask_or(dp, dp, b->effective_cpus);
792 cpumask_and(dp, dp, non_isolated_cpus);
794 update_domain_attr_tree(dattr + nslot, b);
796 /* Done with this partition */
802 BUG_ON(nslot != ndoms);
805 free_cpumask_var(non_isolated_cpus);
809 * Fallback to the default domain if kmalloc() failed.
810 * See comments in partition_sched_domains().
821 * Rebuild scheduler domains.
823 * If the flag 'sched_load_balance' of any cpuset with non-empty
824 * 'cpus' changes, or if the 'cpus' allowed changes in any cpuset
825 * which has that flag enabled, or if any cpuset with a non-empty
826 * 'cpus' is removed, then call this routine to rebuild the
827 * scheduler's dynamic sched domains.
829 * Call with cpuset_mutex held. Takes get_online_cpus().
831 static void rebuild_sched_domains_locked(void)
833 struct sched_domain_attr *attr;
837 lockdep_assert_held(&cpuset_mutex);
841 * We have raced with CPU hotplug. Don't do anything to avoid
842 * passing doms with offlined cpu to partition_sched_domains().
843 * Anyways, hotplug work item will rebuild sched domains.
845 if (!cpumask_equal(top_cpuset.effective_cpus, cpu_active_mask))
848 /* Generate domain masks and attrs */
849 ndoms = generate_sched_domains(&doms, &attr);
851 /* Have scheduler rebuild the domains */
852 partition_sched_domains(ndoms, doms, attr);
856 #else /* !CONFIG_SMP */
857 static void rebuild_sched_domains_locked(void)
860 #endif /* CONFIG_SMP */
862 void rebuild_sched_domains(void)
864 mutex_lock(&cpuset_mutex);
865 rebuild_sched_domains_locked();
866 mutex_unlock(&cpuset_mutex);
870 * update_tasks_cpumask - Update the cpumasks of tasks in the cpuset.
871 * @cs: the cpuset in which each task's cpus_allowed mask needs to be changed
873 * Iterate through each task of @cs updating its cpus_allowed to the
874 * effective cpuset's. As this function is called with cpuset_mutex held,
875 * cpuset membership stays stable.
877 static void update_tasks_cpumask(struct cpuset *cs)
879 struct css_task_iter it;
880 struct task_struct *task;
882 css_task_iter_start(&cs->css, 0, &it);
883 while ((task = css_task_iter_next(&it)))
884 set_cpus_allowed_ptr(task, cs->effective_cpus);
885 css_task_iter_end(&it);
889 * update_cpumasks_hier - Update effective cpumasks and tasks in the subtree
890 * @cs: the cpuset to consider
891 * @new_cpus: temp variable for calculating new effective_cpus
893 * When congifured cpumask is changed, the effective cpumasks of this cpuset
894 * and all its descendants need to be updated.
896 * On legacy hierachy, effective_cpus will be the same with cpu_allowed.
898 * Called with cpuset_mutex held
900 static void update_cpumasks_hier(struct cpuset *cs, struct cpumask *new_cpus)
903 struct cgroup_subsys_state *pos_css;
904 bool need_rebuild_sched_domains = false;
907 cpuset_for_each_descendant_pre(cp, pos_css, cs) {
908 struct cpuset *parent = parent_cs(cp);
910 cpumask_and(new_cpus, cp->cpus_allowed, parent->effective_cpus);
913 * If it becomes empty, inherit the effective mask of the
914 * parent, which is guaranteed to have some CPUs.
916 if (is_in_v2_mode() && cpumask_empty(new_cpus))
917 cpumask_copy(new_cpus, parent->effective_cpus);
919 /* Skip the whole subtree if the cpumask remains the same. */
920 if (cpumask_equal(new_cpus, cp->effective_cpus)) {
921 pos_css = css_rightmost_descendant(pos_css);
925 if (!css_tryget_online(&cp->css))
929 spin_lock_irq(&callback_lock);
930 cpumask_copy(cp->effective_cpus, new_cpus);
931 spin_unlock_irq(&callback_lock);
933 WARN_ON(!is_in_v2_mode() &&
934 !cpumask_equal(cp->cpus_allowed, cp->effective_cpus));
936 update_tasks_cpumask(cp);
939 * If the effective cpumask of any non-empty cpuset is changed,
940 * we need to rebuild sched domains.
942 if (!cpumask_empty(cp->cpus_allowed) &&
943 is_sched_load_balance(cp))
944 need_rebuild_sched_domains = true;
951 if (need_rebuild_sched_domains)
952 rebuild_sched_domains_locked();
956 * update_cpumask - update the cpus_allowed mask of a cpuset and all tasks in it
957 * @cs: the cpuset to consider
958 * @trialcs: trial cpuset
959 * @buf: buffer of cpu numbers written to this cpuset
961 static int update_cpumask(struct cpuset *cs, struct cpuset *trialcs,
966 /* top_cpuset.cpus_allowed tracks cpu_online_mask; it's read-only */
967 if (cs == &top_cpuset)
971 * An empty cpus_allowed is ok only if the cpuset has no tasks.
972 * Since cpulist_parse() fails on an empty mask, we special case
973 * that parsing. The validate_change() call ensures that cpusets
974 * with tasks have cpus.
977 cpumask_clear(trialcs->cpus_allowed);
979 retval = cpulist_parse(buf, trialcs->cpus_allowed);
983 if (!cpumask_subset(trialcs->cpus_allowed,
984 top_cpuset.cpus_allowed))
988 /* Nothing to do if the cpus didn't change */
989 if (cpumask_equal(cs->cpus_allowed, trialcs->cpus_allowed))
992 retval = validate_change(cs, trialcs);
996 spin_lock_irq(&callback_lock);
997 cpumask_copy(cs->cpus_allowed, trialcs->cpus_allowed);
998 spin_unlock_irq(&callback_lock);
1000 /* use trialcs->cpus_allowed as a temp variable */
1001 update_cpumasks_hier(cs, trialcs->cpus_allowed);
1006 * Migrate memory region from one set of nodes to another. This is
1007 * performed asynchronously as it can be called from process migration path
1008 * holding locks involved in process management. All mm migrations are
1009 * performed in the queued order and can be waited for by flushing
1010 * cpuset_migrate_mm_wq.
1013 struct cpuset_migrate_mm_work {
1014 struct work_struct work;
1015 struct mm_struct *mm;
1020 static void cpuset_migrate_mm_workfn(struct work_struct *work)
1022 struct cpuset_migrate_mm_work *mwork =
1023 container_of(work, struct cpuset_migrate_mm_work, work);
1025 /* on a wq worker, no need to worry about %current's mems_allowed */
1026 do_migrate_pages(mwork->mm, &mwork->from, &mwork->to, MPOL_MF_MOVE_ALL);
1031 static void cpuset_migrate_mm(struct mm_struct *mm, const nodemask_t *from,
1032 const nodemask_t *to)
1034 struct cpuset_migrate_mm_work *mwork;
1036 mwork = kzalloc(sizeof(*mwork), GFP_KERNEL);
1039 mwork->from = *from;
1041 INIT_WORK(&mwork->work, cpuset_migrate_mm_workfn);
1042 queue_work(cpuset_migrate_mm_wq, &mwork->work);
1048 static void cpuset_post_attach(void)
1050 flush_workqueue(cpuset_migrate_mm_wq);
1054 * cpuset_change_task_nodemask - change task's mems_allowed and mempolicy
1055 * @tsk: the task to change
1056 * @newmems: new nodes that the task will be set
1058 * We use the mems_allowed_seq seqlock to safely update both tsk->mems_allowed
1059 * and rebind an eventual tasks' mempolicy. If the task is allocating in
1060 * parallel, it might temporarily see an empty intersection, which results in
1061 * a seqlock check and retry before OOM or allocation failure.
1063 static void cpuset_change_task_nodemask(struct task_struct *tsk,
1064 nodemask_t *newmems)
1068 local_irq_disable();
1069 write_seqcount_begin(&tsk->mems_allowed_seq);
1071 nodes_or(tsk->mems_allowed, tsk->mems_allowed, *newmems);
1072 mpol_rebind_task(tsk, newmems);
1073 tsk->mems_allowed = *newmems;
1075 write_seqcount_end(&tsk->mems_allowed_seq);
1081 static void *cpuset_being_rebound;
1084 * update_tasks_nodemask - Update the nodemasks of tasks in the cpuset.
1085 * @cs: the cpuset in which each task's mems_allowed mask needs to be changed
1087 * Iterate through each task of @cs updating its mems_allowed to the
1088 * effective cpuset's. As this function is called with cpuset_mutex held,
1089 * cpuset membership stays stable.
1091 static void update_tasks_nodemask(struct cpuset *cs)
1093 static nodemask_t newmems; /* protected by cpuset_mutex */
1094 struct css_task_iter it;
1095 struct task_struct *task;
1097 cpuset_being_rebound = cs; /* causes mpol_dup() rebind */
1099 guarantee_online_mems(cs, &newmems);
1102 * The mpol_rebind_mm() call takes mmap_sem, which we couldn't
1103 * take while holding tasklist_lock. Forks can happen - the
1104 * mpol_dup() cpuset_being_rebound check will catch such forks,
1105 * and rebind their vma mempolicies too. Because we still hold
1106 * the global cpuset_mutex, we know that no other rebind effort
1107 * will be contending for the global variable cpuset_being_rebound.
1108 * It's ok if we rebind the same mm twice; mpol_rebind_mm()
1109 * is idempotent. Also migrate pages in each mm to new nodes.
1111 css_task_iter_start(&cs->css, 0, &it);
1112 while ((task = css_task_iter_next(&it))) {
1113 struct mm_struct *mm;
1116 cpuset_change_task_nodemask(task, &newmems);
1118 mm = get_task_mm(task);
1122 migrate = is_memory_migrate(cs);
1124 mpol_rebind_mm(mm, &cs->mems_allowed);
1126 cpuset_migrate_mm(mm, &cs->old_mems_allowed, &newmems);
1130 css_task_iter_end(&it);
1133 * All the tasks' nodemasks have been updated, update
1134 * cs->old_mems_allowed.
1136 cs->old_mems_allowed = newmems;
1138 /* We're done rebinding vmas to this cpuset's new mems_allowed. */
1139 cpuset_being_rebound = NULL;
1143 * update_nodemasks_hier - Update effective nodemasks and tasks in the subtree
1144 * @cs: the cpuset to consider
1145 * @new_mems: a temp variable for calculating new effective_mems
1147 * When configured nodemask is changed, the effective nodemasks of this cpuset
1148 * and all its descendants need to be updated.
1150 * On legacy hiearchy, effective_mems will be the same with mems_allowed.
1152 * Called with cpuset_mutex held
1154 static void update_nodemasks_hier(struct cpuset *cs, nodemask_t *new_mems)
1157 struct cgroup_subsys_state *pos_css;
1160 cpuset_for_each_descendant_pre(cp, pos_css, cs) {
1161 struct cpuset *parent = parent_cs(cp);
1163 nodes_and(*new_mems, cp->mems_allowed, parent->effective_mems);
1166 * If it becomes empty, inherit the effective mask of the
1167 * parent, which is guaranteed to have some MEMs.
1169 if (is_in_v2_mode() && nodes_empty(*new_mems))
1170 *new_mems = parent->effective_mems;
1172 /* Skip the whole subtree if the nodemask remains the same. */
1173 if (nodes_equal(*new_mems, cp->effective_mems)) {
1174 pos_css = css_rightmost_descendant(pos_css);
1178 if (!css_tryget_online(&cp->css))
1182 spin_lock_irq(&callback_lock);
1183 cp->effective_mems = *new_mems;
1184 spin_unlock_irq(&callback_lock);
1186 WARN_ON(!is_in_v2_mode() &&
1187 !nodes_equal(cp->mems_allowed, cp->effective_mems));
1189 update_tasks_nodemask(cp);
1198 * Handle user request to change the 'mems' memory placement
1199 * of a cpuset. Needs to validate the request, update the
1200 * cpusets mems_allowed, and for each task in the cpuset,
1201 * update mems_allowed and rebind task's mempolicy and any vma
1202 * mempolicies and if the cpuset is marked 'memory_migrate',
1203 * migrate the tasks pages to the new memory.
1205 * Call with cpuset_mutex held. May take callback_lock during call.
1206 * Will take tasklist_lock, scan tasklist for tasks in cpuset cs,
1207 * lock each such tasks mm->mmap_sem, scan its vma's and rebind
1208 * their mempolicies to the cpusets new mems_allowed.
1210 static int update_nodemask(struct cpuset *cs, struct cpuset *trialcs,
1216 * top_cpuset.mems_allowed tracks node_stats[N_MEMORY];
1219 if (cs == &top_cpuset) {
1225 * An empty mems_allowed is ok iff there are no tasks in the cpuset.
1226 * Since nodelist_parse() fails on an empty mask, we special case
1227 * that parsing. The validate_change() call ensures that cpusets
1228 * with tasks have memory.
1231 nodes_clear(trialcs->mems_allowed);
1233 retval = nodelist_parse(buf, trialcs->mems_allowed);
1237 if (!nodes_subset(trialcs->mems_allowed,
1238 top_cpuset.mems_allowed)) {
1244 if (nodes_equal(cs->mems_allowed, trialcs->mems_allowed)) {
1245 retval = 0; /* Too easy - nothing to do */
1248 retval = validate_change(cs, trialcs);
1252 spin_lock_irq(&callback_lock);
1253 cs->mems_allowed = trialcs->mems_allowed;
1254 spin_unlock_irq(&callback_lock);
1256 /* use trialcs->mems_allowed as a temp variable */
1257 update_nodemasks_hier(cs, &trialcs->mems_allowed);
1262 int current_cpuset_is_being_rebound(void)
1267 ret = task_cs(current) == cpuset_being_rebound;
1273 static int update_relax_domain_level(struct cpuset *cs, s64 val)
1276 if (val < -1 || val >= sched_domain_level_max)
1280 if (val != cs->relax_domain_level) {
1281 cs->relax_domain_level = val;
1282 if (!cpumask_empty(cs->cpus_allowed) &&
1283 is_sched_load_balance(cs))
1284 rebuild_sched_domains_locked();
1291 * update_tasks_flags - update the spread flags of tasks in the cpuset.
1292 * @cs: the cpuset in which each task's spread flags needs to be changed
1294 * Iterate through each task of @cs updating its spread flags. As this
1295 * function is called with cpuset_mutex held, cpuset membership stays
1298 static void update_tasks_flags(struct cpuset *cs)
1300 struct css_task_iter it;
1301 struct task_struct *task;
1303 css_task_iter_start(&cs->css, 0, &it);
1304 while ((task = css_task_iter_next(&it)))
1305 cpuset_update_task_spread_flag(cs, task);
1306 css_task_iter_end(&it);
1310 * update_flag - read a 0 or a 1 in a file and update associated flag
1311 * bit: the bit to update (see cpuset_flagbits_t)
1312 * cs: the cpuset to update
1313 * turning_on: whether the flag is being set or cleared
1315 * Call with cpuset_mutex held.
1318 static int update_flag(cpuset_flagbits_t bit, struct cpuset *cs,
1321 struct cpuset *trialcs;
1322 int balance_flag_changed;
1323 int spread_flag_changed;
1326 trialcs = alloc_trial_cpuset(cs);
1331 set_bit(bit, &trialcs->flags);
1333 clear_bit(bit, &trialcs->flags);
1335 err = validate_change(cs, trialcs);
1339 balance_flag_changed = (is_sched_load_balance(cs) !=
1340 is_sched_load_balance(trialcs));
1342 spread_flag_changed = ((is_spread_slab(cs) != is_spread_slab(trialcs))
1343 || (is_spread_page(cs) != is_spread_page(trialcs)));
1345 spin_lock_irq(&callback_lock);
1346 cs->flags = trialcs->flags;
1347 spin_unlock_irq(&callback_lock);
1349 if (!cpumask_empty(trialcs->cpus_allowed) && balance_flag_changed)
1350 rebuild_sched_domains_locked();
1352 if (spread_flag_changed)
1353 update_tasks_flags(cs);
1355 free_trial_cpuset(trialcs);
1360 * Frequency meter - How fast is some event occurring?
1362 * These routines manage a digitally filtered, constant time based,
1363 * event frequency meter. There are four routines:
1364 * fmeter_init() - initialize a frequency meter.
1365 * fmeter_markevent() - called each time the event happens.
1366 * fmeter_getrate() - returns the recent rate of such events.
1367 * fmeter_update() - internal routine used to update fmeter.
1369 * A common data structure is passed to each of these routines,
1370 * which is used to keep track of the state required to manage the
1371 * frequency meter and its digital filter.
1373 * The filter works on the number of events marked per unit time.
1374 * The filter is single-pole low-pass recursive (IIR). The time unit
1375 * is 1 second. Arithmetic is done using 32-bit integers scaled to
1376 * simulate 3 decimal digits of precision (multiplied by 1000).
1378 * With an FM_COEF of 933, and a time base of 1 second, the filter
1379 * has a half-life of 10 seconds, meaning that if the events quit
1380 * happening, then the rate returned from the fmeter_getrate()
1381 * will be cut in half each 10 seconds, until it converges to zero.
1383 * It is not worth doing a real infinitely recursive filter. If more
1384 * than FM_MAXTICKS ticks have elapsed since the last filter event,
1385 * just compute FM_MAXTICKS ticks worth, by which point the level
1388 * Limit the count of unprocessed events to FM_MAXCNT, so as to avoid
1389 * arithmetic overflow in the fmeter_update() routine.
1391 * Given the simple 32 bit integer arithmetic used, this meter works
1392 * best for reporting rates between one per millisecond (msec) and
1393 * one per 32 (approx) seconds. At constant rates faster than one
1394 * per msec it maxes out at values just under 1,000,000. At constant
1395 * rates between one per msec, and one per second it will stabilize
1396 * to a value N*1000, where N is the rate of events per second.
1397 * At constant rates between one per second and one per 32 seconds,
1398 * it will be choppy, moving up on the seconds that have an event,
1399 * and then decaying until the next event. At rates slower than
1400 * about one in 32 seconds, it decays all the way back to zero between
1404 #define FM_COEF 933 /* coefficient for half-life of 10 secs */
1405 #define FM_MAXTICKS ((u32)99) /* useless computing more ticks than this */
1406 #define FM_MAXCNT 1000000 /* limit cnt to avoid overflow */
1407 #define FM_SCALE 1000 /* faux fixed point scale */
1409 /* Initialize a frequency meter */
1410 static void fmeter_init(struct fmeter *fmp)
1415 spin_lock_init(&fmp->lock);
1418 /* Internal meter update - process cnt events and update value */
1419 static void fmeter_update(struct fmeter *fmp)
1424 now = ktime_get_seconds();
1425 ticks = now - fmp->time;
1430 ticks = min(FM_MAXTICKS, ticks);
1432 fmp->val = (FM_COEF * fmp->val) / FM_SCALE;
1435 fmp->val += ((FM_SCALE - FM_COEF) * fmp->cnt) / FM_SCALE;
1439 /* Process any previous ticks, then bump cnt by one (times scale). */
1440 static void fmeter_markevent(struct fmeter *fmp)
1442 spin_lock(&fmp->lock);
1444 fmp->cnt = min(FM_MAXCNT, fmp->cnt + FM_SCALE);
1445 spin_unlock(&fmp->lock);
1448 /* Process any previous ticks, then return current value. */
1449 static int fmeter_getrate(struct fmeter *fmp)
1453 spin_lock(&fmp->lock);
1456 spin_unlock(&fmp->lock);
1460 static struct cpuset *cpuset_attach_old_cs;
1462 /* Called by cgroups to determine if a cpuset is usable; cpuset_mutex held */
1463 static int cpuset_can_attach(struct cgroup_taskset *tset)
1465 struct cgroup_subsys_state *css;
1467 struct task_struct *task;
1470 /* used later by cpuset_attach() */
1471 cpuset_attach_old_cs = task_cs(cgroup_taskset_first(tset, &css));
1474 mutex_lock(&cpuset_mutex);
1476 /* allow moving tasks into an empty cpuset if on default hierarchy */
1478 if (!is_in_v2_mode() &&
1479 (cpumask_empty(cs->cpus_allowed) || nodes_empty(cs->mems_allowed)))
1482 cgroup_taskset_for_each(task, css, tset) {
1483 ret = task_can_attach(task, cs->cpus_allowed);
1486 ret = security_task_setscheduler(task);
1492 * Mark attach is in progress. This makes validate_change() fail
1493 * changes which zero cpus/mems_allowed.
1495 cs->attach_in_progress++;
1498 mutex_unlock(&cpuset_mutex);
1502 static void cpuset_cancel_attach(struct cgroup_taskset *tset)
1504 struct cgroup_subsys_state *css;
1507 cgroup_taskset_first(tset, &css);
1510 mutex_lock(&cpuset_mutex);
1511 css_cs(css)->attach_in_progress--;
1512 mutex_unlock(&cpuset_mutex);
1516 * Protected by cpuset_mutex. cpus_attach is used only by cpuset_attach()
1517 * but we can't allocate it dynamically there. Define it global and
1518 * allocate from cpuset_init().
1520 static cpumask_var_t cpus_attach;
1522 static void cpuset_attach(struct cgroup_taskset *tset)
1524 /* static buf protected by cpuset_mutex */
1525 static nodemask_t cpuset_attach_nodemask_to;
1526 struct task_struct *task;
1527 struct task_struct *leader;
1528 struct cgroup_subsys_state *css;
1530 struct cpuset *oldcs = cpuset_attach_old_cs;
1532 cgroup_taskset_first(tset, &css);
1535 mutex_lock(&cpuset_mutex);
1537 /* prepare for attach */
1538 if (cs == &top_cpuset)
1539 cpumask_copy(cpus_attach, cpu_possible_mask);
1541 guarantee_online_cpus(cs, cpus_attach);
1543 guarantee_online_mems(cs, &cpuset_attach_nodemask_to);
1545 cgroup_taskset_for_each(task, css, tset) {
1547 * can_attach beforehand should guarantee that this doesn't
1548 * fail. TODO: have a better way to handle failure here
1550 WARN_ON_ONCE(set_cpus_allowed_ptr(task, cpus_attach));
1552 cpuset_change_task_nodemask(task, &cpuset_attach_nodemask_to);
1553 cpuset_update_task_spread_flag(cs, task);
1557 * Change mm for all threadgroup leaders. This is expensive and may
1558 * sleep and should be moved outside migration path proper.
1560 cpuset_attach_nodemask_to = cs->effective_mems;
1561 cgroup_taskset_for_each_leader(leader, css, tset) {
1562 struct mm_struct *mm = get_task_mm(leader);
1565 mpol_rebind_mm(mm, &cpuset_attach_nodemask_to);
1568 * old_mems_allowed is the same with mems_allowed
1569 * here, except if this task is being moved
1570 * automatically due to hotplug. In that case
1571 * @mems_allowed has been updated and is empty, so
1572 * @old_mems_allowed is the right nodesets that we
1575 if (is_memory_migrate(cs))
1576 cpuset_migrate_mm(mm, &oldcs->old_mems_allowed,
1577 &cpuset_attach_nodemask_to);
1583 cs->old_mems_allowed = cpuset_attach_nodemask_to;
1585 cs->attach_in_progress--;
1586 if (!cs->attach_in_progress)
1587 wake_up(&cpuset_attach_wq);
1589 mutex_unlock(&cpuset_mutex);
1592 /* The various types of files and directories in a cpuset file system */
1595 FILE_MEMORY_MIGRATE,
1598 FILE_EFFECTIVE_CPULIST,
1599 FILE_EFFECTIVE_MEMLIST,
1603 FILE_SCHED_LOAD_BALANCE,
1604 FILE_SCHED_RELAX_DOMAIN_LEVEL,
1605 FILE_MEMORY_PRESSURE_ENABLED,
1606 FILE_MEMORY_PRESSURE,
1609 } cpuset_filetype_t;
1611 static int cpuset_write_u64(struct cgroup_subsys_state *css, struct cftype *cft,
1614 struct cpuset *cs = css_cs(css);
1615 cpuset_filetype_t type = cft->private;
1618 mutex_lock(&cpuset_mutex);
1619 if (!is_cpuset_online(cs)) {
1625 case FILE_CPU_EXCLUSIVE:
1626 retval = update_flag(CS_CPU_EXCLUSIVE, cs, val);
1628 case FILE_MEM_EXCLUSIVE:
1629 retval = update_flag(CS_MEM_EXCLUSIVE, cs, val);
1631 case FILE_MEM_HARDWALL:
1632 retval = update_flag(CS_MEM_HARDWALL, cs, val);
1634 case FILE_SCHED_LOAD_BALANCE:
1635 retval = update_flag(CS_SCHED_LOAD_BALANCE, cs, val);
1637 case FILE_MEMORY_MIGRATE:
1638 retval = update_flag(CS_MEMORY_MIGRATE, cs, val);
1640 case FILE_MEMORY_PRESSURE_ENABLED:
1641 cpuset_memory_pressure_enabled = !!val;
1643 case FILE_SPREAD_PAGE:
1644 retval = update_flag(CS_SPREAD_PAGE, cs, val);
1646 case FILE_SPREAD_SLAB:
1647 retval = update_flag(CS_SPREAD_SLAB, cs, val);
1654 mutex_unlock(&cpuset_mutex);
1658 static int cpuset_write_s64(struct cgroup_subsys_state *css, struct cftype *cft,
1661 struct cpuset *cs = css_cs(css);
1662 cpuset_filetype_t type = cft->private;
1663 int retval = -ENODEV;
1665 mutex_lock(&cpuset_mutex);
1666 if (!is_cpuset_online(cs))
1670 case FILE_SCHED_RELAX_DOMAIN_LEVEL:
1671 retval = update_relax_domain_level(cs, val);
1678 mutex_unlock(&cpuset_mutex);
1683 * Common handling for a write to a "cpus" or "mems" file.
1685 static ssize_t cpuset_write_resmask(struct kernfs_open_file *of,
1686 char *buf, size_t nbytes, loff_t off)
1688 struct cpuset *cs = css_cs(of_css(of));
1689 struct cpuset *trialcs;
1690 int retval = -ENODEV;
1692 buf = strstrip(buf);
1695 * CPU or memory hotunplug may leave @cs w/o any execution
1696 * resources, in which case the hotplug code asynchronously updates
1697 * configuration and transfers all tasks to the nearest ancestor
1698 * which can execute.
1700 * As writes to "cpus" or "mems" may restore @cs's execution
1701 * resources, wait for the previously scheduled operations before
1702 * proceeding, so that we don't end up keep removing tasks added
1703 * after execution capability is restored.
1705 * cpuset_hotplug_work calls back into cgroup core via
1706 * cgroup_transfer_tasks() and waiting for it from a cgroupfs
1707 * operation like this one can lead to a deadlock through kernfs
1708 * active_ref protection. Let's break the protection. Losing the
1709 * protection is okay as we check whether @cs is online after
1710 * grabbing cpuset_mutex anyway. This only happens on the legacy
1714 kernfs_break_active_protection(of->kn);
1715 flush_work(&cpuset_hotplug_work);
1717 mutex_lock(&cpuset_mutex);
1718 if (!is_cpuset_online(cs))
1721 trialcs = alloc_trial_cpuset(cs);
1727 switch (of_cft(of)->private) {
1729 retval = update_cpumask(cs, trialcs, buf);
1732 retval = update_nodemask(cs, trialcs, buf);
1739 free_trial_cpuset(trialcs);
1741 mutex_unlock(&cpuset_mutex);
1742 kernfs_unbreak_active_protection(of->kn);
1744 flush_workqueue(cpuset_migrate_mm_wq);
1745 return retval ?: nbytes;
1749 * These ascii lists should be read in a single call, by using a user
1750 * buffer large enough to hold the entire map. If read in smaller
1751 * chunks, there is no guarantee of atomicity. Since the display format
1752 * used, list of ranges of sequential numbers, is variable length,
1753 * and since these maps can change value dynamically, one could read
1754 * gibberish by doing partial reads while a list was changing.
1756 static int cpuset_common_seq_show(struct seq_file *sf, void *v)
1758 struct cpuset *cs = css_cs(seq_css(sf));
1759 cpuset_filetype_t type = seq_cft(sf)->private;
1762 spin_lock_irq(&callback_lock);
1766 seq_printf(sf, "%*pbl\n", cpumask_pr_args(cs->cpus_allowed));
1769 seq_printf(sf, "%*pbl\n", nodemask_pr_args(&cs->mems_allowed));
1771 case FILE_EFFECTIVE_CPULIST:
1772 seq_printf(sf, "%*pbl\n", cpumask_pr_args(cs->effective_cpus));
1774 case FILE_EFFECTIVE_MEMLIST:
1775 seq_printf(sf, "%*pbl\n", nodemask_pr_args(&cs->effective_mems));
1781 spin_unlock_irq(&callback_lock);
1785 static u64 cpuset_read_u64(struct cgroup_subsys_state *css, struct cftype *cft)
1787 struct cpuset *cs = css_cs(css);
1788 cpuset_filetype_t type = cft->private;
1790 case FILE_CPU_EXCLUSIVE:
1791 return is_cpu_exclusive(cs);
1792 case FILE_MEM_EXCLUSIVE:
1793 return is_mem_exclusive(cs);
1794 case FILE_MEM_HARDWALL:
1795 return is_mem_hardwall(cs);
1796 case FILE_SCHED_LOAD_BALANCE:
1797 return is_sched_load_balance(cs);
1798 case FILE_MEMORY_MIGRATE:
1799 return is_memory_migrate(cs);
1800 case FILE_MEMORY_PRESSURE_ENABLED:
1801 return cpuset_memory_pressure_enabled;
1802 case FILE_MEMORY_PRESSURE:
1803 return fmeter_getrate(&cs->fmeter);
1804 case FILE_SPREAD_PAGE:
1805 return is_spread_page(cs);
1806 case FILE_SPREAD_SLAB:
1807 return is_spread_slab(cs);
1812 /* Unreachable but makes gcc happy */
1816 static s64 cpuset_read_s64(struct cgroup_subsys_state *css, struct cftype *cft)
1818 struct cpuset *cs = css_cs(css);
1819 cpuset_filetype_t type = cft->private;
1821 case FILE_SCHED_RELAX_DOMAIN_LEVEL:
1822 return cs->relax_domain_level;
1827 /* Unrechable but makes gcc happy */
1833 * for the common functions, 'private' gives the type of file
1836 static struct cftype files[] = {
1839 .seq_show = cpuset_common_seq_show,
1840 .write = cpuset_write_resmask,
1841 .max_write_len = (100U + 6 * NR_CPUS),
1842 .private = FILE_CPULIST,
1847 .seq_show = cpuset_common_seq_show,
1848 .write = cpuset_write_resmask,
1849 .max_write_len = (100U + 6 * MAX_NUMNODES),
1850 .private = FILE_MEMLIST,
1854 .name = "effective_cpus",
1855 .seq_show = cpuset_common_seq_show,
1856 .private = FILE_EFFECTIVE_CPULIST,
1860 .name = "effective_mems",
1861 .seq_show = cpuset_common_seq_show,
1862 .private = FILE_EFFECTIVE_MEMLIST,
1866 .name = "cpu_exclusive",
1867 .read_u64 = cpuset_read_u64,
1868 .write_u64 = cpuset_write_u64,
1869 .private = FILE_CPU_EXCLUSIVE,
1873 .name = "mem_exclusive",
1874 .read_u64 = cpuset_read_u64,
1875 .write_u64 = cpuset_write_u64,
1876 .private = FILE_MEM_EXCLUSIVE,
1880 .name = "mem_hardwall",
1881 .read_u64 = cpuset_read_u64,
1882 .write_u64 = cpuset_write_u64,
1883 .private = FILE_MEM_HARDWALL,
1887 .name = "sched_load_balance",
1888 .read_u64 = cpuset_read_u64,
1889 .write_u64 = cpuset_write_u64,
1890 .private = FILE_SCHED_LOAD_BALANCE,
1894 .name = "sched_relax_domain_level",
1895 .read_s64 = cpuset_read_s64,
1896 .write_s64 = cpuset_write_s64,
1897 .private = FILE_SCHED_RELAX_DOMAIN_LEVEL,
1901 .name = "memory_migrate",
1902 .read_u64 = cpuset_read_u64,
1903 .write_u64 = cpuset_write_u64,
1904 .private = FILE_MEMORY_MIGRATE,
1908 .name = "memory_pressure",
1909 .read_u64 = cpuset_read_u64,
1910 .private = FILE_MEMORY_PRESSURE,
1914 .name = "memory_spread_page",
1915 .read_u64 = cpuset_read_u64,
1916 .write_u64 = cpuset_write_u64,
1917 .private = FILE_SPREAD_PAGE,
1921 .name = "memory_spread_slab",
1922 .read_u64 = cpuset_read_u64,
1923 .write_u64 = cpuset_write_u64,
1924 .private = FILE_SPREAD_SLAB,
1928 .name = "memory_pressure_enabled",
1929 .flags = CFTYPE_ONLY_ON_ROOT,
1930 .read_u64 = cpuset_read_u64,
1931 .write_u64 = cpuset_write_u64,
1932 .private = FILE_MEMORY_PRESSURE_ENABLED,
1939 * cpuset_css_alloc - allocate a cpuset css
1940 * cgrp: control group that the new cpuset will be part of
1943 static struct cgroup_subsys_state *
1944 cpuset_css_alloc(struct cgroup_subsys_state *parent_css)
1949 return &top_cpuset.css;
1951 cs = kzalloc(sizeof(*cs), GFP_KERNEL);
1953 return ERR_PTR(-ENOMEM);
1954 if (!alloc_cpumask_var(&cs->cpus_allowed, GFP_KERNEL))
1956 if (!alloc_cpumask_var(&cs->effective_cpus, GFP_KERNEL))
1959 set_bit(CS_SCHED_LOAD_BALANCE, &cs->flags);
1960 cpumask_clear(cs->cpus_allowed);
1961 nodes_clear(cs->mems_allowed);
1962 cpumask_clear(cs->effective_cpus);
1963 nodes_clear(cs->effective_mems);
1964 fmeter_init(&cs->fmeter);
1965 cs->relax_domain_level = -1;
1970 free_cpumask_var(cs->cpus_allowed);
1973 return ERR_PTR(-ENOMEM);
1976 static int cpuset_css_online(struct cgroup_subsys_state *css)
1978 struct cpuset *cs = css_cs(css);
1979 struct cpuset *parent = parent_cs(cs);
1980 struct cpuset *tmp_cs;
1981 struct cgroup_subsys_state *pos_css;
1986 mutex_lock(&cpuset_mutex);
1988 set_bit(CS_ONLINE, &cs->flags);
1989 if (is_spread_page(parent))
1990 set_bit(CS_SPREAD_PAGE, &cs->flags);
1991 if (is_spread_slab(parent))
1992 set_bit(CS_SPREAD_SLAB, &cs->flags);
1996 spin_lock_irq(&callback_lock);
1997 if (is_in_v2_mode()) {
1998 cpumask_copy(cs->effective_cpus, parent->effective_cpus);
1999 cs->effective_mems = parent->effective_mems;
2001 spin_unlock_irq(&callback_lock);
2003 if (!test_bit(CGRP_CPUSET_CLONE_CHILDREN, &css->cgroup->flags))
2007 * Clone @parent's configuration if CGRP_CPUSET_CLONE_CHILDREN is
2008 * set. This flag handling is implemented in cgroup core for
2009 * histrical reasons - the flag may be specified during mount.
2011 * Currently, if any sibling cpusets have exclusive cpus or mem, we
2012 * refuse to clone the configuration - thereby refusing the task to
2013 * be entered, and as a result refusing the sys_unshare() or
2014 * clone() which initiated it. If this becomes a problem for some
2015 * users who wish to allow that scenario, then this could be
2016 * changed to grant parent->cpus_allowed-sibling_cpus_exclusive
2017 * (and likewise for mems) to the new cgroup.
2020 cpuset_for_each_child(tmp_cs, pos_css, parent) {
2021 if (is_mem_exclusive(tmp_cs) || is_cpu_exclusive(tmp_cs)) {
2028 spin_lock_irq(&callback_lock);
2029 cs->mems_allowed = parent->mems_allowed;
2030 cs->effective_mems = parent->mems_allowed;
2031 cpumask_copy(cs->cpus_allowed, parent->cpus_allowed);
2032 cpumask_copy(cs->effective_cpus, parent->cpus_allowed);
2033 spin_unlock_irq(&callback_lock);
2035 mutex_unlock(&cpuset_mutex);
2040 * If the cpuset being removed has its flag 'sched_load_balance'
2041 * enabled, then simulate turning sched_load_balance off, which
2042 * will call rebuild_sched_domains_locked().
2045 static void cpuset_css_offline(struct cgroup_subsys_state *css)
2047 struct cpuset *cs = css_cs(css);
2049 mutex_lock(&cpuset_mutex);
2051 if (is_sched_load_balance(cs))
2052 update_flag(CS_SCHED_LOAD_BALANCE, cs, 0);
2055 clear_bit(CS_ONLINE, &cs->flags);
2057 mutex_unlock(&cpuset_mutex);
2060 static void cpuset_css_free(struct cgroup_subsys_state *css)
2062 struct cpuset *cs = css_cs(css);
2064 free_cpumask_var(cs->effective_cpus);
2065 free_cpumask_var(cs->cpus_allowed);
2069 static void cpuset_bind(struct cgroup_subsys_state *root_css)
2071 mutex_lock(&cpuset_mutex);
2072 spin_lock_irq(&callback_lock);
2074 if (is_in_v2_mode()) {
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);
2088 * Make sure the new task conform to the current state of its parent,
2089 * which could have been changed by cpuset just after it inherits the
2090 * state from the parent and before it sits on the cgroup's task list.
2092 static void cpuset_fork(struct task_struct *task)
2094 if (task_css_is_root(task, cpuset_cgrp_id))
2097 set_cpus_allowed_ptr(task, ¤t->cpus_allowed);
2098 task->mems_allowed = current->mems_allowed;
2101 struct cgroup_subsys cpuset_cgrp_subsys = {
2102 .css_alloc = cpuset_css_alloc,
2103 .css_online = cpuset_css_online,
2104 .css_offline = cpuset_css_offline,
2105 .css_free = cpuset_css_free,
2106 .can_attach = cpuset_can_attach,
2107 .cancel_attach = cpuset_cancel_attach,
2108 .attach = cpuset_attach,
2109 .post_attach = cpuset_post_attach,
2110 .bind = cpuset_bind,
2111 .fork = cpuset_fork,
2112 .legacy_cftypes = files,
2117 * cpuset_init - initialize cpusets at system boot
2119 * Description: Initialize top_cpuset and the cpuset internal file system,
2122 int __init cpuset_init(void)
2126 BUG_ON(!alloc_cpumask_var(&top_cpuset.cpus_allowed, GFP_KERNEL));
2127 BUG_ON(!alloc_cpumask_var(&top_cpuset.effective_cpus, GFP_KERNEL));
2129 cpumask_setall(top_cpuset.cpus_allowed);
2130 nodes_setall(top_cpuset.mems_allowed);
2131 cpumask_setall(top_cpuset.effective_cpus);
2132 nodes_setall(top_cpuset.effective_mems);
2134 fmeter_init(&top_cpuset.fmeter);
2135 set_bit(CS_SCHED_LOAD_BALANCE, &top_cpuset.flags);
2136 top_cpuset.relax_domain_level = -1;
2138 err = register_filesystem(&cpuset_fs_type);
2142 BUG_ON(!alloc_cpumask_var(&cpus_attach, GFP_KERNEL));
2148 * If CPU and/or memory hotplug handlers, below, unplug any CPUs
2149 * or memory nodes, we need to walk over the cpuset hierarchy,
2150 * removing that CPU or node from all cpusets. If this removes the
2151 * last CPU or node from a cpuset, then move the tasks in the empty
2152 * cpuset to its next-highest non-empty parent.
2154 static void remove_tasks_in_empty_cpuset(struct cpuset *cs)
2156 struct cpuset *parent;
2159 * Find its next-highest non-empty parent, (top cpuset
2160 * has online cpus, so can't be empty).
2162 parent = parent_cs(cs);
2163 while (cpumask_empty(parent->cpus_allowed) ||
2164 nodes_empty(parent->mems_allowed))
2165 parent = parent_cs(parent);
2167 if (cgroup_transfer_tasks(parent->css.cgroup, cs->css.cgroup)) {
2168 pr_err("cpuset: failed to transfer tasks out of empty cpuset ");
2169 pr_cont_cgroup_name(cs->css.cgroup);
2175 hotplug_update_tasks_legacy(struct cpuset *cs,
2176 struct cpumask *new_cpus, nodemask_t *new_mems,
2177 bool cpus_updated, bool mems_updated)
2181 spin_lock_irq(&callback_lock);
2182 cpumask_copy(cs->cpus_allowed, new_cpus);
2183 cpumask_copy(cs->effective_cpus, new_cpus);
2184 cs->mems_allowed = *new_mems;
2185 cs->effective_mems = *new_mems;
2186 spin_unlock_irq(&callback_lock);
2189 * Don't call update_tasks_cpumask() if the cpuset becomes empty,
2190 * as the tasks will be migratecd to an ancestor.
2192 if (cpus_updated && !cpumask_empty(cs->cpus_allowed))
2193 update_tasks_cpumask(cs);
2194 if (mems_updated && !nodes_empty(cs->mems_allowed))
2195 update_tasks_nodemask(cs);
2197 is_empty = cpumask_empty(cs->cpus_allowed) ||
2198 nodes_empty(cs->mems_allowed);
2200 mutex_unlock(&cpuset_mutex);
2203 * Move tasks to the nearest ancestor with execution resources,
2204 * This is full cgroup operation which will also call back into
2205 * cpuset. Should be done outside any lock.
2208 remove_tasks_in_empty_cpuset(cs);
2210 mutex_lock(&cpuset_mutex);
2214 hotplug_update_tasks(struct cpuset *cs,
2215 struct cpumask *new_cpus, nodemask_t *new_mems,
2216 bool cpus_updated, bool mems_updated)
2218 if (cpumask_empty(new_cpus))
2219 cpumask_copy(new_cpus, parent_cs(cs)->effective_cpus);
2220 if (nodes_empty(*new_mems))
2221 *new_mems = parent_cs(cs)->effective_mems;
2223 spin_lock_irq(&callback_lock);
2224 cpumask_copy(cs->effective_cpus, new_cpus);
2225 cs->effective_mems = *new_mems;
2226 spin_unlock_irq(&callback_lock);
2229 update_tasks_cpumask(cs);
2231 update_tasks_nodemask(cs);
2235 * cpuset_hotplug_update_tasks - update tasks in a cpuset for hotunplug
2236 * @cs: cpuset in interest
2238 * Compare @cs's cpu and mem masks against top_cpuset and if some have gone
2239 * offline, update @cs accordingly. If @cs ends up with no CPU or memory,
2240 * all its tasks are moved to the nearest ancestor with both resources.
2242 static void cpuset_hotplug_update_tasks(struct cpuset *cs)
2244 static cpumask_t new_cpus;
2245 static nodemask_t new_mems;
2249 wait_event(cpuset_attach_wq, cs->attach_in_progress == 0);
2251 mutex_lock(&cpuset_mutex);
2254 * We have raced with task attaching. We wait until attaching
2255 * is finished, so we won't attach a task to an empty cpuset.
2257 if (cs->attach_in_progress) {
2258 mutex_unlock(&cpuset_mutex);
2262 cpumask_and(&new_cpus, cs->cpus_allowed, parent_cs(cs)->effective_cpus);
2263 nodes_and(new_mems, cs->mems_allowed, parent_cs(cs)->effective_mems);
2265 cpus_updated = !cpumask_equal(&new_cpus, cs->effective_cpus);
2266 mems_updated = !nodes_equal(new_mems, cs->effective_mems);
2268 if (is_in_v2_mode())
2269 hotplug_update_tasks(cs, &new_cpus, &new_mems,
2270 cpus_updated, mems_updated);
2272 hotplug_update_tasks_legacy(cs, &new_cpus, &new_mems,
2273 cpus_updated, mems_updated);
2275 mutex_unlock(&cpuset_mutex);
2278 static bool force_rebuild;
2280 void cpuset_force_rebuild(void)
2282 force_rebuild = true;
2286 * cpuset_hotplug_workfn - handle CPU/memory hotunplug for a cpuset
2288 * This function is called after either CPU or memory configuration has
2289 * changed and updates cpuset accordingly. The top_cpuset is always
2290 * synchronized to cpu_active_mask and N_MEMORY, which is necessary in
2291 * order to make cpusets transparent (of no affect) on systems that are
2292 * actively using CPU hotplug but making no active use of cpusets.
2294 * Non-root cpusets are only affected by offlining. If any CPUs or memory
2295 * nodes have been taken down, cpuset_hotplug_update_tasks() is invoked on
2298 * Note that CPU offlining during suspend is ignored. We don't modify
2299 * cpusets across suspend/resume cycles at all.
2301 static void cpuset_hotplug_workfn(struct work_struct *work)
2303 static cpumask_t new_cpus;
2304 static nodemask_t new_mems;
2305 bool cpus_updated, mems_updated;
2306 bool on_dfl = is_in_v2_mode();
2308 mutex_lock(&cpuset_mutex);
2310 /* fetch the available cpus/mems and find out which changed how */
2311 cpumask_copy(&new_cpus, cpu_active_mask);
2312 new_mems = node_states[N_MEMORY];
2314 cpus_updated = !cpumask_equal(top_cpuset.effective_cpus, &new_cpus);
2315 mems_updated = !nodes_equal(top_cpuset.effective_mems, new_mems);
2317 /* synchronize cpus_allowed to cpu_active_mask */
2319 spin_lock_irq(&callback_lock);
2321 cpumask_copy(top_cpuset.cpus_allowed, &new_cpus);
2322 cpumask_copy(top_cpuset.effective_cpus, &new_cpus);
2323 spin_unlock_irq(&callback_lock);
2324 /* we don't mess with cpumasks of tasks in top_cpuset */
2327 /* synchronize mems_allowed to N_MEMORY */
2329 spin_lock_irq(&callback_lock);
2331 top_cpuset.mems_allowed = new_mems;
2332 top_cpuset.effective_mems = new_mems;
2333 spin_unlock_irq(&callback_lock);
2334 update_tasks_nodemask(&top_cpuset);
2337 mutex_unlock(&cpuset_mutex);
2339 /* if cpus or mems changed, we need to propagate to descendants */
2340 if (cpus_updated || mems_updated) {
2342 struct cgroup_subsys_state *pos_css;
2345 cpuset_for_each_descendant_pre(cs, pos_css, &top_cpuset) {
2346 if (cs == &top_cpuset || !css_tryget_online(&cs->css))
2350 cpuset_hotplug_update_tasks(cs);
2358 /* rebuild sched domains if cpus_allowed has changed */
2359 if (cpus_updated || force_rebuild) {
2360 force_rebuild = false;
2361 rebuild_sched_domains();
2365 void cpuset_update_active_cpus(void)
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 schedule_work(&cpuset_hotplug_work);
2375 void cpuset_wait_for_hotplug(void)
2377 flush_work(&cpuset_hotplug_work);
2381 * Keep top_cpuset.mems_allowed tracking node_states[N_MEMORY].
2382 * Call this routine anytime after node_states[N_MEMORY] changes.
2383 * See cpuset_update_active_cpus() for CPU hotplug handling.
2385 static int cpuset_track_online_nodes(struct notifier_block *self,
2386 unsigned long action, void *arg)
2388 schedule_work(&cpuset_hotplug_work);
2392 static struct notifier_block cpuset_track_online_nodes_nb = {
2393 .notifier_call = cpuset_track_online_nodes,
2394 .priority = 10, /* ??! */
2398 * cpuset_init_smp - initialize cpus_allowed
2400 * Description: Finish top cpuset after cpu, node maps are initialized
2402 void __init cpuset_init_smp(void)
2404 cpumask_copy(top_cpuset.cpus_allowed, cpu_active_mask);
2405 top_cpuset.mems_allowed = node_states[N_MEMORY];
2406 top_cpuset.old_mems_allowed = top_cpuset.mems_allowed;
2408 cpumask_copy(top_cpuset.effective_cpus, cpu_active_mask);
2409 top_cpuset.effective_mems = node_states[N_MEMORY];
2411 register_hotmemory_notifier(&cpuset_track_online_nodes_nb);
2413 cpuset_migrate_mm_wq = alloc_ordered_workqueue("cpuset_migrate_mm", 0);
2414 BUG_ON(!cpuset_migrate_mm_wq);
2418 * cpuset_cpus_allowed - return cpus_allowed mask from a tasks cpuset.
2419 * @tsk: pointer to task_struct from which to obtain cpuset->cpus_allowed.
2420 * @pmask: pointer to struct cpumask variable to receive cpus_allowed set.
2422 * Description: Returns the cpumask_var_t cpus_allowed of the cpuset
2423 * attached to the specified @tsk. Guaranteed to return some non-empty
2424 * subset of cpu_online_mask, even if this means going outside the
2428 void cpuset_cpus_allowed(struct task_struct *tsk, struct cpumask *pmask)
2430 unsigned long flags;
2432 spin_lock_irqsave(&callback_lock, flags);
2434 guarantee_online_cpus(task_cs(tsk), pmask);
2436 spin_unlock_irqrestore(&callback_lock, flags);
2440 * cpuset_cpus_allowed_fallback - final fallback before complete catastrophe.
2441 * @tsk: pointer to task_struct with which the scheduler is struggling
2443 * Description: In the case that the scheduler cannot find an allowed cpu in
2444 * tsk->cpus_allowed, we fall back to task_cs(tsk)->cpus_allowed. In legacy
2445 * mode however, this value is the same as task_cs(tsk)->effective_cpus,
2446 * which will not contain a sane cpumask during cases such as cpu hotplugging.
2447 * This is the absolute last resort for the scheduler and it is only used if
2448 * _every_ other avenue has been traveled.
2451 void cpuset_cpus_allowed_fallback(struct task_struct *tsk)
2454 do_set_cpus_allowed(tsk, is_in_v2_mode() ?
2455 task_cs(tsk)->cpus_allowed : cpu_possible_mask);
2459 * We own tsk->cpus_allowed, nobody can change it under us.
2461 * But we used cs && cs->cpus_allowed lockless and thus can
2462 * race with cgroup_attach_task() or update_cpumask() and get
2463 * the wrong tsk->cpus_allowed. However, both cases imply the
2464 * subsequent cpuset_change_cpumask()->set_cpus_allowed_ptr()
2465 * which takes task_rq_lock().
2467 * If we are called after it dropped the lock we must see all
2468 * changes in tsk_cs()->cpus_allowed. Otherwise we can temporary
2469 * set any mask even if it is not right from task_cs() pov,
2470 * the pending set_cpus_allowed_ptr() will fix things.
2472 * select_fallback_rq() will fix things ups and set cpu_possible_mask
2477 void __init cpuset_init_current_mems_allowed(void)
2479 nodes_setall(current->mems_allowed);
2483 * cpuset_mems_allowed - return mems_allowed mask from a tasks cpuset.
2484 * @tsk: pointer to task_struct from which to obtain cpuset->mems_allowed.
2486 * Description: Returns the nodemask_t mems_allowed of the cpuset
2487 * attached to the specified @tsk. Guaranteed to return some non-empty
2488 * subset of node_states[N_MEMORY], even if this means going outside the
2492 nodemask_t cpuset_mems_allowed(struct task_struct *tsk)
2495 unsigned long flags;
2497 spin_lock_irqsave(&callback_lock, flags);
2499 guarantee_online_mems(task_cs(tsk), &mask);
2501 spin_unlock_irqrestore(&callback_lock, flags);
2507 * cpuset_nodemask_valid_mems_allowed - check nodemask vs. curremt mems_allowed
2508 * @nodemask: the nodemask to be checked
2510 * Are any of the nodes in the nodemask allowed in current->mems_allowed?
2512 int cpuset_nodemask_valid_mems_allowed(nodemask_t *nodemask)
2514 return nodes_intersects(*nodemask, current->mems_allowed);
2518 * nearest_hardwall_ancestor() - Returns the nearest mem_exclusive or
2519 * mem_hardwall ancestor to the specified cpuset. Call holding
2520 * callback_lock. If no ancestor is mem_exclusive or mem_hardwall
2521 * (an unusual configuration), then returns the root cpuset.
2523 static struct cpuset *nearest_hardwall_ancestor(struct cpuset *cs)
2525 while (!(is_mem_exclusive(cs) || is_mem_hardwall(cs)) && parent_cs(cs))
2531 * cpuset_node_allowed - Can we allocate on a memory node?
2532 * @node: is this an allowed node?
2533 * @gfp_mask: memory allocation flags
2535 * If we're in interrupt, yes, we can always allocate. If @node is set in
2536 * current's mems_allowed, yes. If it's not a __GFP_HARDWALL request and this
2537 * node is set in the nearest hardwalled cpuset ancestor to current's cpuset,
2538 * yes. If current has access to memory reserves as an oom victim, yes.
2541 * GFP_USER allocations are marked with the __GFP_HARDWALL bit,
2542 * and do not allow allocations outside the current tasks cpuset
2543 * unless the task has been OOM killed.
2544 * GFP_KERNEL allocations are not so marked, so can escape to the
2545 * nearest enclosing hardwalled ancestor cpuset.
2547 * Scanning up parent cpusets requires callback_lock. The
2548 * __alloc_pages() routine only calls here with __GFP_HARDWALL bit
2549 * _not_ set if it's a GFP_KERNEL allocation, and all nodes in the
2550 * current tasks mems_allowed came up empty on the first pass over
2551 * the zonelist. So only GFP_KERNEL allocations, if all nodes in the
2552 * cpuset are short of memory, might require taking the callback_lock.
2554 * The first call here from mm/page_alloc:get_page_from_freelist()
2555 * has __GFP_HARDWALL set in gfp_mask, enforcing hardwall cpusets,
2556 * so no allocation on a node outside the cpuset is allowed (unless
2557 * in interrupt, of course).
2559 * The second pass through get_page_from_freelist() doesn't even call
2560 * here for GFP_ATOMIC calls. For those calls, the __alloc_pages()
2561 * variable 'wait' is not set, and the bit ALLOC_CPUSET is not set
2562 * in alloc_flags. That logic and the checks below have the combined
2564 * in_interrupt - any node ok (current task context irrelevant)
2565 * GFP_ATOMIC - any node ok
2566 * tsk_is_oom_victim - any node ok
2567 * GFP_KERNEL - any node in enclosing hardwalled cpuset ok
2568 * GFP_USER - only nodes in current tasks mems allowed ok.
2570 bool __cpuset_node_allowed(int node, gfp_t gfp_mask)
2572 struct cpuset *cs; /* current cpuset ancestors */
2573 int allowed; /* is allocation in zone z allowed? */
2574 unsigned long flags;
2578 if (node_isset(node, current->mems_allowed))
2581 * Allow tasks that have access to memory reserves because they have
2582 * been OOM killed to get memory anywhere.
2584 if (unlikely(tsk_is_oom_victim(current)))
2586 if (gfp_mask & __GFP_HARDWALL) /* If hardwall request, stop here */
2589 if (current->flags & PF_EXITING) /* Let dying task have memory */
2592 /* Not hardwall and node outside mems_allowed: scan up cpusets */
2593 spin_lock_irqsave(&callback_lock, flags);
2596 cs = nearest_hardwall_ancestor(task_cs(current));
2597 allowed = node_isset(node, cs->mems_allowed);
2600 spin_unlock_irqrestore(&callback_lock, flags);
2605 * cpuset_mem_spread_node() - On which node to begin search for a file page
2606 * cpuset_slab_spread_node() - On which node to begin search for a slab page
2608 * If a task is marked PF_SPREAD_PAGE or PF_SPREAD_SLAB (as for
2609 * tasks in a cpuset with is_spread_page or is_spread_slab set),
2610 * and if the memory allocation used cpuset_mem_spread_node()
2611 * to determine on which node to start looking, as it will for
2612 * certain page cache or slab cache pages such as used for file
2613 * system buffers and inode caches, then instead of starting on the
2614 * local node to look for a free page, rather spread the starting
2615 * node around the tasks mems_allowed nodes.
2617 * We don't have to worry about the returned node being offline
2618 * because "it can't happen", and even if it did, it would be ok.
2620 * The routines calling guarantee_online_mems() are careful to
2621 * only set nodes in task->mems_allowed that are online. So it
2622 * should not be possible for the following code to return an
2623 * offline node. But if it did, that would be ok, as this routine
2624 * is not returning the node where the allocation must be, only
2625 * the node where the search should start. The zonelist passed to
2626 * __alloc_pages() will include all nodes. If the slab allocator
2627 * is passed an offline node, it will fall back to the local node.
2628 * See kmem_cache_alloc_node().
2631 static int cpuset_spread_node(int *rotor)
2633 return *rotor = next_node_in(*rotor, current->mems_allowed);
2636 int cpuset_mem_spread_node(void)
2638 if (current->cpuset_mem_spread_rotor == NUMA_NO_NODE)
2639 current->cpuset_mem_spread_rotor =
2640 node_random(¤t->mems_allowed);
2642 return cpuset_spread_node(¤t->cpuset_mem_spread_rotor);
2645 int cpuset_slab_spread_node(void)
2647 if (current->cpuset_slab_spread_rotor == NUMA_NO_NODE)
2648 current->cpuset_slab_spread_rotor =
2649 node_random(¤t->mems_allowed);
2651 return cpuset_spread_node(¤t->cpuset_slab_spread_rotor);
2654 EXPORT_SYMBOL_GPL(cpuset_mem_spread_node);
2657 * cpuset_mems_allowed_intersects - Does @tsk1's mems_allowed intersect @tsk2's?
2658 * @tsk1: pointer to task_struct of some task.
2659 * @tsk2: pointer to task_struct of some other task.
2661 * Description: Return true if @tsk1's mems_allowed intersects the
2662 * mems_allowed of @tsk2. Used by the OOM killer to determine if
2663 * one of the task's memory usage might impact the memory available
2667 int cpuset_mems_allowed_intersects(const struct task_struct *tsk1,
2668 const struct task_struct *tsk2)
2670 return nodes_intersects(tsk1->mems_allowed, tsk2->mems_allowed);
2674 * cpuset_print_current_mems_allowed - prints current's cpuset and mems_allowed
2676 * Description: Prints current's name, cpuset name, and cached copy of its
2677 * mems_allowed to the kernel log.
2679 void cpuset_print_current_mems_allowed(void)
2681 struct cgroup *cgrp;
2685 cgrp = task_cs(current)->css.cgroup;
2686 pr_info("%s cpuset=", current->comm);
2687 pr_cont_cgroup_name(cgrp);
2688 pr_cont(" mems_allowed=%*pbl\n",
2689 nodemask_pr_args(¤t->mems_allowed));
2695 * Collection of memory_pressure is suppressed unless
2696 * this flag is enabled by writing "1" to the special
2697 * cpuset file 'memory_pressure_enabled' in the root cpuset.
2700 int cpuset_memory_pressure_enabled __read_mostly;
2703 * cpuset_memory_pressure_bump - keep stats of per-cpuset reclaims.
2705 * Keep a running average of the rate of synchronous (direct)
2706 * page reclaim efforts initiated by tasks in each cpuset.
2708 * This represents the rate at which some task in the cpuset
2709 * ran low on memory on all nodes it was allowed to use, and
2710 * had to enter the kernels page reclaim code in an effort to
2711 * create more free memory by tossing clean pages or swapping
2712 * or writing dirty pages.
2714 * Display to user space in the per-cpuset read-only file
2715 * "memory_pressure". Value displayed is an integer
2716 * representing the recent rate of entry into the synchronous
2717 * (direct) page reclaim by any task attached to the cpuset.
2720 void __cpuset_memory_pressure_bump(void)
2723 fmeter_markevent(&task_cs(current)->fmeter);
2727 #ifdef CONFIG_PROC_PID_CPUSET
2729 * proc_cpuset_show()
2730 * - Print tasks cpuset path into seq_file.
2731 * - Used for /proc/<pid>/cpuset.
2732 * - No need to task_lock(tsk) on this tsk->cpuset reference, as it
2733 * doesn't really matter if tsk->cpuset changes after we read it,
2734 * and we take cpuset_mutex, keeping cpuset_attach() from changing it
2737 int proc_cpuset_show(struct seq_file *m, struct pid_namespace *ns,
2738 struct pid *pid, struct task_struct *tsk)
2741 struct cgroup_subsys_state *css;
2745 buf = kmalloc(PATH_MAX, GFP_KERNEL);
2749 css = task_get_css(tsk, cpuset_cgrp_id);
2750 retval = cgroup_path_ns(css->cgroup, buf, PATH_MAX,
2751 current->nsproxy->cgroup_ns);
2753 if (retval >= PATH_MAX)
2754 retval = -ENAMETOOLONG;
2765 #endif /* CONFIG_PROC_PID_CPUSET */
2767 /* Display task mems_allowed in /proc/<pid>/status file. */
2768 void cpuset_task_status_allowed(struct seq_file *m, struct task_struct *task)
2770 seq_printf(m, "Mems_allowed:\t%*pb\n",
2771 nodemask_pr_args(&task->mems_allowed));
2772 seq_printf(m, "Mems_allowed_list:\t%*pbl\n",
2773 nodemask_pr_args(&task->mems_allowed));