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/fs_context.h>
43 #include <linux/namei.h>
44 #include <linux/pagemap.h>
45 #include <linux/proc_fs.h>
46 #include <linux/rcupdate.h>
47 #include <linux/sched.h>
48 #include <linux/sched/deadline.h>
49 #include <linux/sched/mm.h>
50 #include <linux/sched/task.h>
51 #include <linux/seq_file.h>
52 #include <linux/security.h>
53 #include <linux/slab.h>
54 #include <linux/spinlock.h>
55 #include <linux/stat.h>
56 #include <linux/string.h>
57 #include <linux/time.h>
58 #include <linux/time64.h>
59 #include <linux/backing-dev.h>
60 #include <linux/sort.h>
61 #include <linux/oom.h>
62 #include <linux/sched/isolation.h>
63 #include <linux/uaccess.h>
64 #include <linux/atomic.h>
65 #include <linux/mutex.h>
66 #include <linux/cgroup.h>
67 #include <linux/wait.h>
69 DEFINE_STATIC_KEY_FALSE(cpusets_pre_enable_key);
70 DEFINE_STATIC_KEY_FALSE(cpusets_enabled_key);
73 * There could be abnormal cpuset configurations for cpu or memory
74 * node binding, add this key to provide a quick low-cost judgement
77 DEFINE_STATIC_KEY_FALSE(cpusets_insane_config_key);
79 /* See "Frequency meter" comments, below. */
82 int cnt; /* unprocessed events count */
83 int val; /* most recent output value */
84 time64_t time; /* clock (secs) when val computed */
85 spinlock_t lock; /* guards read or write of above */
89 struct cgroup_subsys_state css;
91 unsigned long flags; /* "unsigned long" so bitops work */
94 * On default hierarchy:
96 * The user-configured masks can only be changed by writing to
97 * cpuset.cpus and cpuset.mems, and won't be limited by the
100 * The effective masks is the real masks that apply to the tasks
101 * in the cpuset. They may be changed if the configured masks are
102 * changed or hotplug happens.
104 * effective_mask == configured_mask & parent's effective_mask,
105 * and if it ends up empty, it will inherit the parent's mask.
108 * On legacy hierarchy:
110 * The user-configured masks are always the same with effective masks.
113 /* user-configured CPUs and Memory Nodes allow to tasks */
114 cpumask_var_t cpus_allowed;
115 nodemask_t mems_allowed;
117 /* effective CPUs and Memory Nodes allow to tasks */
118 cpumask_var_t effective_cpus;
119 nodemask_t effective_mems;
122 * CPUs allocated to child sub-partitions (default hierarchy only)
123 * - CPUs granted by the parent = effective_cpus U subparts_cpus
124 * - effective_cpus and subparts_cpus are mutually exclusive.
126 * effective_cpus contains only onlined CPUs, but subparts_cpus
127 * may have offlined ones.
129 cpumask_var_t subparts_cpus;
132 * This is old Memory Nodes tasks took on.
134 * - top_cpuset.old_mems_allowed is initialized to mems_allowed.
135 * - A new cpuset's old_mems_allowed is initialized when some
136 * task is moved into it.
137 * - old_mems_allowed is used in cpuset_migrate_mm() when we change
138 * cpuset.mems_allowed and have tasks' nodemask updated, and
139 * then old_mems_allowed is updated to mems_allowed.
141 nodemask_t old_mems_allowed;
143 struct fmeter fmeter; /* memory_pressure filter */
146 * Tasks are being attached to this cpuset. Used to prevent
147 * zeroing cpus/mems_allowed between ->can_attach() and ->attach().
149 int attach_in_progress;
151 /* partition number for rebuild_sched_domains() */
154 /* for custom sched domain */
155 int relax_domain_level;
157 /* number of CPUs in subparts_cpus */
158 int nr_subparts_cpus;
160 /* partition root state */
161 int partition_root_state;
164 * Default hierarchy only:
165 * use_parent_ecpus - set if using parent's effective_cpus
166 * child_ecpus_count - # of children with use_parent_ecpus set
168 int use_parent_ecpus;
169 int child_ecpus_count;
171 /* Handle for cpuset.cpus.partition */
172 struct cgroup_file partition_file;
176 * Partition root states:
178 * 0 - not a partition root
182 * -1 - invalid partition root
183 * None of the cpus in cpus_allowed can be put into the parent's
184 * subparts_cpus. In this case, the cpuset is not a real partition
185 * root anymore. However, the CPU_EXCLUSIVE bit will still be set
186 * and the cpuset can be restored back to a partition root if the
187 * parent cpuset can give more CPUs back to this child cpuset.
189 #define PRS_DISABLED 0
190 #define PRS_ENABLED 1
194 * Temporary cpumasks for working with partitions that are passed among
195 * functions to avoid memory allocation in inner functions.
198 cpumask_var_t addmask, delmask; /* For partition root */
199 cpumask_var_t new_cpus; /* For update_cpumasks_hier() */
202 static inline struct cpuset *css_cs(struct cgroup_subsys_state *css)
204 return css ? container_of(css, struct cpuset, css) : NULL;
207 /* Retrieve the cpuset for a task */
208 static inline struct cpuset *task_cs(struct task_struct *task)
210 return css_cs(task_css(task, cpuset_cgrp_id));
213 static inline struct cpuset *parent_cs(struct cpuset *cs)
215 return css_cs(cs->css.parent);
218 /* bits in struct cpuset flags field */
225 CS_SCHED_LOAD_BALANCE,
230 /* convenient tests for these bits */
231 static inline bool is_cpuset_online(struct cpuset *cs)
233 return test_bit(CS_ONLINE, &cs->flags) && !css_is_dying(&cs->css);
236 static inline int is_cpu_exclusive(const struct cpuset *cs)
238 return test_bit(CS_CPU_EXCLUSIVE, &cs->flags);
241 static inline int is_mem_exclusive(const struct cpuset *cs)
243 return test_bit(CS_MEM_EXCLUSIVE, &cs->flags);
246 static inline int is_mem_hardwall(const struct cpuset *cs)
248 return test_bit(CS_MEM_HARDWALL, &cs->flags);
251 static inline int is_sched_load_balance(const struct cpuset *cs)
253 return test_bit(CS_SCHED_LOAD_BALANCE, &cs->flags);
256 static inline int is_memory_migrate(const struct cpuset *cs)
258 return test_bit(CS_MEMORY_MIGRATE, &cs->flags);
261 static inline int is_spread_page(const struct cpuset *cs)
263 return test_bit(CS_SPREAD_PAGE, &cs->flags);
266 static inline int is_spread_slab(const struct cpuset *cs)
268 return test_bit(CS_SPREAD_SLAB, &cs->flags);
271 static inline int is_partition_root(const struct cpuset *cs)
273 return cs->partition_root_state > 0;
277 * Send notification event of whenever partition_root_state changes.
279 static inline void notify_partition_change(struct cpuset *cs,
280 int old_prs, int new_prs)
282 if (old_prs != new_prs)
283 cgroup_file_notify(&cs->partition_file);
286 static struct cpuset top_cpuset = {
287 .flags = ((1 << CS_ONLINE) | (1 << CS_CPU_EXCLUSIVE) |
288 (1 << CS_MEM_EXCLUSIVE)),
289 .partition_root_state = PRS_ENABLED,
293 * cpuset_for_each_child - traverse online children of a cpuset
294 * @child_cs: loop cursor pointing to the current child
295 * @pos_css: used for iteration
296 * @parent_cs: target cpuset to walk children of
298 * Walk @child_cs through the online children of @parent_cs. Must be used
299 * with RCU read locked.
301 #define cpuset_for_each_child(child_cs, pos_css, parent_cs) \
302 css_for_each_child((pos_css), &(parent_cs)->css) \
303 if (is_cpuset_online(((child_cs) = css_cs((pos_css)))))
306 * cpuset_for_each_descendant_pre - pre-order walk of a cpuset's descendants
307 * @des_cs: loop cursor pointing to the current descendant
308 * @pos_css: used for iteration
309 * @root_cs: target cpuset to walk ancestor of
311 * Walk @des_cs through the online descendants of @root_cs. Must be used
312 * with RCU read locked. The caller may modify @pos_css by calling
313 * css_rightmost_descendant() to skip subtree. @root_cs is included in the
314 * iteration and the first node to be visited.
316 #define cpuset_for_each_descendant_pre(des_cs, pos_css, root_cs) \
317 css_for_each_descendant_pre((pos_css), &(root_cs)->css) \
318 if (is_cpuset_online(((des_cs) = css_cs((pos_css)))))
321 * There are two global locks guarding cpuset structures - cpuset_rwsem and
322 * callback_lock. We also require taking task_lock() when dereferencing a
323 * task's cpuset pointer. See "The task_lock() exception", at the end of this
324 * comment. The cpuset code uses only cpuset_rwsem write lock. Other
325 * kernel subsystems can use cpuset_read_lock()/cpuset_read_unlock() to
326 * prevent change to cpuset structures.
328 * A task must hold both locks to modify cpusets. If a task holds
329 * cpuset_rwsem, it blocks others wanting that rwsem, ensuring that it
330 * is the only task able to also acquire callback_lock and be able to
331 * modify cpusets. It can perform various checks on the cpuset structure
332 * first, knowing nothing will change. It can also allocate memory while
333 * just holding cpuset_rwsem. While it is performing these checks, various
334 * callback routines can briefly acquire callback_lock to query cpusets.
335 * Once it is ready to make the changes, it takes callback_lock, blocking
338 * Calls to the kernel memory allocator can not be made while holding
339 * callback_lock, as that would risk double tripping on callback_lock
340 * from one of the callbacks into the cpuset code from within
343 * If a task is only holding callback_lock, then it has read-only
346 * Now, the task_struct fields mems_allowed and mempolicy may be changed
347 * by other task, we use alloc_lock in the task_struct fields to protect
350 * The cpuset_common_file_read() handlers only hold callback_lock across
351 * small pieces of code, such as when reading out possibly multi-word
352 * cpumasks and nodemasks.
354 * Accessing a task's cpuset should be done in accordance with the
355 * guidelines for accessing subsystem state in kernel/cgroup.c
358 DEFINE_STATIC_PERCPU_RWSEM(cpuset_rwsem);
360 void cpuset_read_lock(void)
362 percpu_down_read(&cpuset_rwsem);
365 void cpuset_read_unlock(void)
367 percpu_up_read(&cpuset_rwsem);
370 static DEFINE_SPINLOCK(callback_lock);
372 static struct workqueue_struct *cpuset_migrate_mm_wq;
375 * CPU / memory hotplug is handled asynchronously.
377 static void cpuset_hotplug_workfn(struct work_struct *work);
378 static DECLARE_WORK(cpuset_hotplug_work, cpuset_hotplug_workfn);
380 static DECLARE_WAIT_QUEUE_HEAD(cpuset_attach_wq);
382 static inline void check_insane_mems_config(nodemask_t *nodes)
384 if (!cpusets_insane_config() &&
385 movable_only_nodes(nodes)) {
386 static_branch_enable(&cpusets_insane_config_key);
387 pr_info("Unsupported (movable nodes only) cpuset configuration detected (nmask=%*pbl)!\n"
388 "Cpuset allocations might fail even with a lot of memory available.\n",
389 nodemask_pr_args(nodes));
394 * Cgroup v2 behavior is used on the "cpus" and "mems" control files when
395 * on default hierarchy or when the cpuset_v2_mode flag is set by mounting
396 * the v1 cpuset cgroup filesystem with the "cpuset_v2_mode" mount option.
397 * With v2 behavior, "cpus" and "mems" are always what the users have
398 * requested and won't be changed by hotplug events. Only the effective
399 * cpus or mems will be affected.
401 static inline bool is_in_v2_mode(void)
403 return cgroup_subsys_on_dfl(cpuset_cgrp_subsys) ||
404 (cpuset_cgrp_subsys.root->flags & CGRP_ROOT_CPUSET_V2_MODE);
408 * Return in pmask the portion of a task's cpusets's cpus_allowed that
409 * are online and are capable of running the task. If none are found,
410 * walk up the cpuset hierarchy until we find one that does have some
413 * One way or another, we guarantee to return some non-empty subset
414 * of cpu_online_mask.
416 * Call with callback_lock or cpuset_rwsem held.
418 static void guarantee_online_cpus(struct task_struct *tsk,
419 struct cpumask *pmask)
421 const struct cpumask *possible_mask = task_cpu_possible_mask(tsk);
424 if (WARN_ON(!cpumask_and(pmask, possible_mask, cpu_online_mask)))
425 cpumask_copy(pmask, cpu_online_mask);
430 while (!cpumask_intersects(cs->effective_cpus, pmask)) {
434 * The top cpuset doesn't have any online cpu as a
435 * consequence of a race between cpuset_hotplug_work
436 * and cpu hotplug notifier. But we know the top
437 * cpuset's effective_cpus is on its way to be
438 * identical to cpu_online_mask.
443 cpumask_and(pmask, pmask, cs->effective_cpus);
450 * Return in *pmask the portion of a cpusets's mems_allowed that
451 * are online, with memory. If none are online with memory, walk
452 * up the cpuset hierarchy until we find one that does have some
453 * online mems. The top cpuset always has some mems online.
455 * One way or another, we guarantee to return some non-empty subset
456 * of node_states[N_MEMORY].
458 * Call with callback_lock or cpuset_rwsem held.
460 static void guarantee_online_mems(struct cpuset *cs, nodemask_t *pmask)
462 while (!nodes_intersects(cs->effective_mems, node_states[N_MEMORY]))
464 nodes_and(*pmask, cs->effective_mems, node_states[N_MEMORY]);
468 * update task's spread flag if cpuset's page/slab spread flag is set
470 * Call with callback_lock or cpuset_rwsem held.
472 static void cpuset_update_task_spread_flag(struct cpuset *cs,
473 struct task_struct *tsk)
475 if (is_spread_page(cs))
476 task_set_spread_page(tsk);
478 task_clear_spread_page(tsk);
480 if (is_spread_slab(cs))
481 task_set_spread_slab(tsk);
483 task_clear_spread_slab(tsk);
487 * is_cpuset_subset(p, q) - Is cpuset p a subset of cpuset q?
489 * One cpuset is a subset of another if all its allowed CPUs and
490 * Memory Nodes are a subset of the other, and its exclusive flags
491 * are only set if the other's are set. Call holding cpuset_rwsem.
494 static int is_cpuset_subset(const struct cpuset *p, const struct cpuset *q)
496 return cpumask_subset(p->cpus_allowed, q->cpus_allowed) &&
497 nodes_subset(p->mems_allowed, q->mems_allowed) &&
498 is_cpu_exclusive(p) <= is_cpu_exclusive(q) &&
499 is_mem_exclusive(p) <= is_mem_exclusive(q);
503 * alloc_cpumasks - allocate three cpumasks for cpuset
504 * @cs: the cpuset that have cpumasks to be allocated.
505 * @tmp: the tmpmasks structure pointer
506 * Return: 0 if successful, -ENOMEM otherwise.
508 * Only one of the two input arguments should be non-NULL.
510 static inline int alloc_cpumasks(struct cpuset *cs, struct tmpmasks *tmp)
512 cpumask_var_t *pmask1, *pmask2, *pmask3;
515 pmask1 = &cs->cpus_allowed;
516 pmask2 = &cs->effective_cpus;
517 pmask3 = &cs->subparts_cpus;
519 pmask1 = &tmp->new_cpus;
520 pmask2 = &tmp->addmask;
521 pmask3 = &tmp->delmask;
524 if (!zalloc_cpumask_var(pmask1, GFP_KERNEL))
527 if (!zalloc_cpumask_var(pmask2, GFP_KERNEL))
530 if (!zalloc_cpumask_var(pmask3, GFP_KERNEL))
536 free_cpumask_var(*pmask2);
538 free_cpumask_var(*pmask1);
543 * free_cpumasks - free cpumasks in a tmpmasks structure
544 * @cs: the cpuset that have cpumasks to be free.
545 * @tmp: the tmpmasks structure pointer
547 static inline void free_cpumasks(struct cpuset *cs, struct tmpmasks *tmp)
550 free_cpumask_var(cs->cpus_allowed);
551 free_cpumask_var(cs->effective_cpus);
552 free_cpumask_var(cs->subparts_cpus);
555 free_cpumask_var(tmp->new_cpus);
556 free_cpumask_var(tmp->addmask);
557 free_cpumask_var(tmp->delmask);
562 * alloc_trial_cpuset - allocate a trial cpuset
563 * @cs: the cpuset that the trial cpuset duplicates
565 static struct cpuset *alloc_trial_cpuset(struct cpuset *cs)
567 struct cpuset *trial;
569 trial = kmemdup(cs, sizeof(*cs), GFP_KERNEL);
573 if (alloc_cpumasks(trial, NULL)) {
578 cpumask_copy(trial->cpus_allowed, cs->cpus_allowed);
579 cpumask_copy(trial->effective_cpus, cs->effective_cpus);
584 * free_cpuset - free the cpuset
585 * @cs: the cpuset to be freed
587 static inline void free_cpuset(struct cpuset *cs)
589 free_cpumasks(cs, NULL);
594 * validate_change() - Used to validate that any proposed cpuset change
595 * follows the structural rules for cpusets.
597 * If we replaced the flag and mask values of the current cpuset
598 * (cur) with those values in the trial cpuset (trial), would
599 * our various subset and exclusive rules still be valid? Presumes
602 * 'cur' is the address of an actual, in-use cpuset. Operations
603 * such as list traversal that depend on the actual address of the
604 * cpuset in the list must use cur below, not trial.
606 * 'trial' is the address of bulk structure copy of cur, with
607 * perhaps one or more of the fields cpus_allowed, mems_allowed,
608 * or flags changed to new, trial values.
610 * Return 0 if valid, -errno if not.
613 static int validate_change(struct cpuset *cur, struct cpuset *trial)
615 struct cgroup_subsys_state *css;
616 struct cpuset *c, *par;
619 /* The checks don't apply to root cpuset */
620 if (cur == &top_cpuset)
624 par = parent_cs(cur);
626 /* On legacy hierarchy, we must be a subset of our parent cpuset. */
628 if (!is_in_v2_mode() && !is_cpuset_subset(trial, par))
632 * If either I or some sibling (!= me) is exclusive, we can't
636 cpuset_for_each_child(c, css, par) {
637 if ((is_cpu_exclusive(trial) || is_cpu_exclusive(c)) &&
639 cpumask_intersects(trial->cpus_allowed, c->cpus_allowed))
641 if ((is_mem_exclusive(trial) || is_mem_exclusive(c)) &&
643 nodes_intersects(trial->mems_allowed, c->mems_allowed))
648 * Cpusets with tasks - existing or newly being attached - can't
649 * be changed to have empty cpus_allowed or mems_allowed.
652 if ((cgroup_is_populated(cur->css.cgroup) || cur->attach_in_progress)) {
653 if (!cpumask_empty(cur->cpus_allowed) &&
654 cpumask_empty(trial->cpus_allowed))
656 if (!nodes_empty(cur->mems_allowed) &&
657 nodes_empty(trial->mems_allowed))
662 * We can't shrink if we won't have enough room for SCHED_DEADLINE
666 if (is_cpu_exclusive(cur) &&
667 !cpuset_cpumask_can_shrink(cur->cpus_allowed,
668 trial->cpus_allowed))
679 * Helper routine for generate_sched_domains().
680 * Do cpusets a, b have overlapping effective cpus_allowed masks?
682 static int cpusets_overlap(struct cpuset *a, struct cpuset *b)
684 return cpumask_intersects(a->effective_cpus, b->effective_cpus);
688 update_domain_attr(struct sched_domain_attr *dattr, struct cpuset *c)
690 if (dattr->relax_domain_level < c->relax_domain_level)
691 dattr->relax_domain_level = c->relax_domain_level;
695 static void update_domain_attr_tree(struct sched_domain_attr *dattr,
696 struct cpuset *root_cs)
699 struct cgroup_subsys_state *pos_css;
702 cpuset_for_each_descendant_pre(cp, pos_css, root_cs) {
703 /* skip the whole subtree if @cp doesn't have any CPU */
704 if (cpumask_empty(cp->cpus_allowed)) {
705 pos_css = css_rightmost_descendant(pos_css);
709 if (is_sched_load_balance(cp))
710 update_domain_attr(dattr, cp);
715 /* Must be called with cpuset_rwsem held. */
716 static inline int nr_cpusets(void)
718 /* jump label reference count + the top-level cpuset */
719 return static_key_count(&cpusets_enabled_key.key) + 1;
723 * generate_sched_domains()
725 * This function builds a partial partition of the systems CPUs
726 * A 'partial partition' is a set of non-overlapping subsets whose
727 * union is a subset of that set.
728 * The output of this function needs to be passed to kernel/sched/core.c
729 * partition_sched_domains() routine, which will rebuild the scheduler's
730 * load balancing domains (sched domains) as specified by that partial
733 * See "What is sched_load_balance" in Documentation/admin-guide/cgroup-v1/cpusets.rst
734 * for a background explanation of this.
736 * Does not return errors, on the theory that the callers of this
737 * routine would rather not worry about failures to rebuild sched
738 * domains when operating in the severe memory shortage situations
739 * that could cause allocation failures below.
741 * Must be called with cpuset_rwsem held.
743 * The three key local variables below are:
744 * cp - cpuset pointer, used (together with pos_css) to perform a
745 * top-down scan of all cpusets. For our purposes, rebuilding
746 * the schedulers sched domains, we can ignore !is_sched_load_
748 * csa - (for CpuSet Array) Array of pointers to all the cpusets
749 * that need to be load balanced, for convenient iterative
750 * access by the subsequent code that finds the best partition,
751 * i.e the set of domains (subsets) of CPUs such that the
752 * cpus_allowed of every cpuset marked is_sched_load_balance
753 * is a subset of one of these domains, while there are as
754 * many such domains as possible, each as small as possible.
755 * doms - Conversion of 'csa' to an array of cpumasks, for passing to
756 * the kernel/sched/core.c routine partition_sched_domains() in a
757 * convenient format, that can be easily compared to the prior
758 * value to determine what partition elements (sched domains)
759 * were changed (added or removed.)
761 * Finding the best partition (set of domains):
762 * The triple nested loops below over i, j, k scan over the
763 * load balanced cpusets (using the array of cpuset pointers in
764 * csa[]) looking for pairs of cpusets that have overlapping
765 * cpus_allowed, but which don't have the same 'pn' partition
766 * number and gives them in the same partition number. It keeps
767 * looping on the 'restart' label until it can no longer find
770 * The union of the cpus_allowed masks from the set of
771 * all cpusets having the same 'pn' value then form the one
772 * element of the partition (one sched domain) to be passed to
773 * partition_sched_domains().
775 static int generate_sched_domains(cpumask_var_t **domains,
776 struct sched_domain_attr **attributes)
778 struct cpuset *cp; /* top-down scan of cpusets */
779 struct cpuset **csa; /* array of all cpuset ptrs */
780 int csn; /* how many cpuset ptrs in csa so far */
781 int i, j, k; /* indices for partition finding loops */
782 cpumask_var_t *doms; /* resulting partition; i.e. sched domains */
783 struct sched_domain_attr *dattr; /* attributes for custom domains */
784 int ndoms = 0; /* number of sched domains in result */
785 int nslot; /* next empty doms[] struct cpumask slot */
786 struct cgroup_subsys_state *pos_css;
787 bool root_load_balance = is_sched_load_balance(&top_cpuset);
793 /* Special case for the 99% of systems with one, full, sched domain */
794 if (root_load_balance && !top_cpuset.nr_subparts_cpus) {
796 doms = alloc_sched_domains(ndoms);
800 dattr = kmalloc(sizeof(struct sched_domain_attr), GFP_KERNEL);
802 *dattr = SD_ATTR_INIT;
803 update_domain_attr_tree(dattr, &top_cpuset);
805 cpumask_and(doms[0], top_cpuset.effective_cpus,
806 housekeeping_cpumask(HK_FLAG_DOMAIN));
811 csa = kmalloc_array(nr_cpusets(), sizeof(cp), GFP_KERNEL);
817 if (root_load_balance)
818 csa[csn++] = &top_cpuset;
819 cpuset_for_each_descendant_pre(cp, pos_css, &top_cpuset) {
820 if (cp == &top_cpuset)
823 * Continue traversing beyond @cp iff @cp has some CPUs and
824 * isn't load balancing. The former is obvious. The
825 * latter: All child cpusets contain a subset of the
826 * parent's cpus, so just skip them, and then we call
827 * update_domain_attr_tree() to calc relax_domain_level of
828 * the corresponding sched domain.
830 * If root is load-balancing, we can skip @cp if it
831 * is a subset of the root's effective_cpus.
833 if (!cpumask_empty(cp->cpus_allowed) &&
834 !(is_sched_load_balance(cp) &&
835 cpumask_intersects(cp->cpus_allowed,
836 housekeeping_cpumask(HK_FLAG_DOMAIN))))
839 if (root_load_balance &&
840 cpumask_subset(cp->cpus_allowed, top_cpuset.effective_cpus))
843 if (is_sched_load_balance(cp) &&
844 !cpumask_empty(cp->effective_cpus))
847 /* skip @cp's subtree if not a partition root */
848 if (!is_partition_root(cp))
849 pos_css = css_rightmost_descendant(pos_css);
853 for (i = 0; i < csn; i++)
858 /* Find the best partition (set of sched domains) */
859 for (i = 0; i < csn; i++) {
860 struct cpuset *a = csa[i];
863 for (j = 0; j < csn; j++) {
864 struct cpuset *b = csa[j];
867 if (apn != bpn && cpusets_overlap(a, b)) {
868 for (k = 0; k < csn; k++) {
869 struct cpuset *c = csa[k];
874 ndoms--; /* one less element */
881 * Now we know how many domains to create.
882 * Convert <csn, csa> to <ndoms, doms> and populate cpu masks.
884 doms = alloc_sched_domains(ndoms);
889 * The rest of the code, including the scheduler, can deal with
890 * dattr==NULL case. No need to abort if alloc fails.
892 dattr = kmalloc_array(ndoms, sizeof(struct sched_domain_attr),
895 for (nslot = 0, i = 0; i < csn; i++) {
896 struct cpuset *a = csa[i];
901 /* Skip completed partitions */
907 if (nslot == ndoms) {
908 static int warnings = 10;
910 pr_warn("rebuild_sched_domains confused: nslot %d, ndoms %d, csn %d, i %d, apn %d\n",
911 nslot, ndoms, csn, i, apn);
919 *(dattr + nslot) = SD_ATTR_INIT;
920 for (j = i; j < csn; j++) {
921 struct cpuset *b = csa[j];
924 cpumask_or(dp, dp, b->effective_cpus);
925 cpumask_and(dp, dp, housekeeping_cpumask(HK_FLAG_DOMAIN));
927 update_domain_attr_tree(dattr + nslot, b);
929 /* Done with this partition */
935 BUG_ON(nslot != ndoms);
941 * Fallback to the default domain if kmalloc() failed.
942 * See comments in partition_sched_domains().
952 static void update_tasks_root_domain(struct cpuset *cs)
954 struct css_task_iter it;
955 struct task_struct *task;
957 css_task_iter_start(&cs->css, 0, &it);
959 while ((task = css_task_iter_next(&it)))
960 dl_add_task_root_domain(task);
962 css_task_iter_end(&it);
965 static void rebuild_root_domains(void)
967 struct cpuset *cs = NULL;
968 struct cgroup_subsys_state *pos_css;
970 percpu_rwsem_assert_held(&cpuset_rwsem);
971 lockdep_assert_cpus_held();
972 lockdep_assert_held(&sched_domains_mutex);
977 * Clear default root domain DL accounting, it will be computed again
978 * if a task belongs to it.
980 dl_clear_root_domain(&def_root_domain);
982 cpuset_for_each_descendant_pre(cs, pos_css, &top_cpuset) {
984 if (cpumask_empty(cs->effective_cpus)) {
985 pos_css = css_rightmost_descendant(pos_css);
993 update_tasks_root_domain(cs);
1002 partition_and_rebuild_sched_domains(int ndoms_new, cpumask_var_t doms_new[],
1003 struct sched_domain_attr *dattr_new)
1005 mutex_lock(&sched_domains_mutex);
1006 partition_sched_domains_locked(ndoms_new, doms_new, dattr_new);
1007 rebuild_root_domains();
1008 mutex_unlock(&sched_domains_mutex);
1012 * Rebuild scheduler domains.
1014 * If the flag 'sched_load_balance' of any cpuset with non-empty
1015 * 'cpus' changes, or if the 'cpus' allowed changes in any cpuset
1016 * which has that flag enabled, or if any cpuset with a non-empty
1017 * 'cpus' is removed, then call this routine to rebuild the
1018 * scheduler's dynamic sched domains.
1020 * Call with cpuset_rwsem held. Takes cpus_read_lock().
1022 static void rebuild_sched_domains_locked(void)
1024 struct cgroup_subsys_state *pos_css;
1025 struct sched_domain_attr *attr;
1026 cpumask_var_t *doms;
1030 lockdep_assert_cpus_held();
1031 percpu_rwsem_assert_held(&cpuset_rwsem);
1034 * If we have raced with CPU hotplug, return early to avoid
1035 * passing doms with offlined cpu to partition_sched_domains().
1036 * Anyways, cpuset_hotplug_workfn() will rebuild sched domains.
1038 * With no CPUs in any subpartitions, top_cpuset's effective CPUs
1039 * should be the same as the active CPUs, so checking only top_cpuset
1040 * is enough to detect racing CPU offlines.
1042 if (!top_cpuset.nr_subparts_cpus &&
1043 !cpumask_equal(top_cpuset.effective_cpus, cpu_active_mask))
1047 * With subpartition CPUs, however, the effective CPUs of a partition
1048 * root should be only a subset of the active CPUs. Since a CPU in any
1049 * partition root could be offlined, all must be checked.
1051 if (top_cpuset.nr_subparts_cpus) {
1053 cpuset_for_each_descendant_pre(cs, pos_css, &top_cpuset) {
1054 if (!is_partition_root(cs)) {
1055 pos_css = css_rightmost_descendant(pos_css);
1058 if (!cpumask_subset(cs->effective_cpus,
1067 /* Generate domain masks and attrs */
1068 ndoms = generate_sched_domains(&doms, &attr);
1070 /* Have scheduler rebuild the domains */
1071 partition_and_rebuild_sched_domains(ndoms, doms, attr);
1073 #else /* !CONFIG_SMP */
1074 static void rebuild_sched_domains_locked(void)
1077 #endif /* CONFIG_SMP */
1079 void rebuild_sched_domains(void)
1082 percpu_down_write(&cpuset_rwsem);
1083 rebuild_sched_domains_locked();
1084 percpu_up_write(&cpuset_rwsem);
1089 * update_tasks_cpumask - Update the cpumasks of tasks in the cpuset.
1090 * @cs: the cpuset in which each task's cpus_allowed mask needs to be changed
1092 * Iterate through each task of @cs updating its cpus_allowed to the
1093 * effective cpuset's. As this function is called with cpuset_rwsem held,
1094 * cpuset membership stays stable.
1096 static void update_tasks_cpumask(struct cpuset *cs)
1098 struct css_task_iter it;
1099 struct task_struct *task;
1101 css_task_iter_start(&cs->css, 0, &it);
1102 while ((task = css_task_iter_next(&it)))
1103 set_cpus_allowed_ptr(task, cs->effective_cpus);
1104 css_task_iter_end(&it);
1108 * compute_effective_cpumask - Compute the effective cpumask of the cpuset
1109 * @new_cpus: the temp variable for the new effective_cpus mask
1110 * @cs: the cpuset the need to recompute the new effective_cpus mask
1111 * @parent: the parent cpuset
1113 * If the parent has subpartition CPUs, include them in the list of
1114 * allowable CPUs in computing the new effective_cpus mask. Since offlined
1115 * CPUs are not removed from subparts_cpus, we have to use cpu_active_mask
1116 * to mask those out.
1118 static void compute_effective_cpumask(struct cpumask *new_cpus,
1119 struct cpuset *cs, struct cpuset *parent)
1121 if (parent->nr_subparts_cpus) {
1122 cpumask_or(new_cpus, parent->effective_cpus,
1123 parent->subparts_cpus);
1124 cpumask_and(new_cpus, new_cpus, cs->cpus_allowed);
1125 cpumask_and(new_cpus, new_cpus, cpu_active_mask);
1127 cpumask_and(new_cpus, cs->cpus_allowed, parent->effective_cpus);
1132 * Commands for update_parent_subparts_cpumask
1135 partcmd_enable, /* Enable partition root */
1136 partcmd_disable, /* Disable partition root */
1137 partcmd_update, /* Update parent's subparts_cpus */
1141 * update_parent_subparts_cpumask - update subparts_cpus mask of parent cpuset
1142 * @cpuset: The cpuset that requests change in partition root state
1143 * @cmd: Partition root state change command
1144 * @newmask: Optional new cpumask for partcmd_update
1145 * @tmp: Temporary addmask and delmask
1146 * Return: 0, 1 or an error code
1148 * For partcmd_enable, the cpuset is being transformed from a non-partition
1149 * root to a partition root. The cpus_allowed mask of the given cpuset will
1150 * be put into parent's subparts_cpus and taken away from parent's
1151 * effective_cpus. The function will return 0 if all the CPUs listed in
1152 * cpus_allowed can be granted or an error code will be returned.
1154 * For partcmd_disable, the cpuset is being transofrmed from a partition
1155 * root back to a non-partition root. Any CPUs in cpus_allowed that are in
1156 * parent's subparts_cpus will be taken away from that cpumask and put back
1157 * into parent's effective_cpus. 0 should always be returned.
1159 * For partcmd_update, if the optional newmask is specified, the cpu
1160 * list is to be changed from cpus_allowed to newmask. Otherwise,
1161 * cpus_allowed is assumed to remain the same. The cpuset should either
1162 * be a partition root or an invalid partition root. The partition root
1163 * state may change if newmask is NULL and none of the requested CPUs can
1164 * be granted by the parent. The function will return 1 if changes to
1165 * parent's subparts_cpus and effective_cpus happen or 0 otherwise.
1166 * Error code should only be returned when newmask is non-NULL.
1168 * The partcmd_enable and partcmd_disable commands are used by
1169 * update_prstate(). The partcmd_update command is used by
1170 * update_cpumasks_hier() with newmask NULL and update_cpumask() with
1173 * The checking is more strict when enabling partition root than the
1174 * other two commands.
1176 * Because of the implicit cpu exclusive nature of a partition root,
1177 * cpumask changes that violates the cpu exclusivity rule will not be
1178 * permitted when checked by validate_change(). The validate_change()
1179 * function will also prevent any changes to the cpu list if it is not
1180 * a superset of children's cpu lists.
1182 static int update_parent_subparts_cpumask(struct cpuset *cpuset, int cmd,
1183 struct cpumask *newmask,
1184 struct tmpmasks *tmp)
1186 struct cpuset *parent = parent_cs(cpuset);
1187 int adding; /* Moving cpus from effective_cpus to subparts_cpus */
1188 int deleting; /* Moving cpus from subparts_cpus to effective_cpus */
1189 int old_prs, new_prs;
1190 bool part_error = false; /* Partition error? */
1192 percpu_rwsem_assert_held(&cpuset_rwsem);
1195 * The parent must be a partition root.
1196 * The new cpumask, if present, or the current cpus_allowed must
1199 if (!is_partition_root(parent) ||
1200 (newmask && cpumask_empty(newmask)) ||
1201 (!newmask && cpumask_empty(cpuset->cpus_allowed)))
1205 * Enabling/disabling partition root is not allowed if there are
1208 if ((cmd != partcmd_update) && css_has_online_children(&cpuset->css))
1212 * Enabling partition root is not allowed if not all the CPUs
1213 * can be granted from parent's effective_cpus or at least one
1214 * CPU will be left after that.
1216 if ((cmd == partcmd_enable) &&
1217 (!cpumask_subset(cpuset->cpus_allowed, parent->effective_cpus) ||
1218 cpumask_equal(cpuset->cpus_allowed, parent->effective_cpus)))
1222 * A cpumask update cannot make parent's effective_cpus become empty.
1224 adding = deleting = false;
1225 old_prs = new_prs = cpuset->partition_root_state;
1226 if (cmd == partcmd_enable) {
1227 cpumask_copy(tmp->addmask, cpuset->cpus_allowed);
1229 } else if (cmd == partcmd_disable) {
1230 deleting = cpumask_and(tmp->delmask, cpuset->cpus_allowed,
1231 parent->subparts_cpus);
1232 } else if (newmask) {
1234 * partcmd_update with newmask:
1236 * delmask = cpus_allowed & ~newmask & parent->subparts_cpus
1237 * addmask = newmask & parent->effective_cpus
1238 * & ~parent->subparts_cpus
1240 cpumask_andnot(tmp->delmask, cpuset->cpus_allowed, newmask);
1241 deleting = cpumask_and(tmp->delmask, tmp->delmask,
1242 parent->subparts_cpus);
1244 cpumask_and(tmp->addmask, newmask, parent->effective_cpus);
1245 adding = cpumask_andnot(tmp->addmask, tmp->addmask,
1246 parent->subparts_cpus);
1248 * Return error if the new effective_cpus could become empty.
1251 cpumask_equal(parent->effective_cpus, tmp->addmask)) {
1255 * As some of the CPUs in subparts_cpus might have
1256 * been offlined, we need to compute the real delmask
1259 if (!cpumask_and(tmp->addmask, tmp->delmask,
1262 cpumask_copy(tmp->addmask, parent->effective_cpus);
1266 * partcmd_update w/o newmask:
1268 * addmask = cpus_allowed & parent->effective_cpus
1270 * Note that parent's subparts_cpus may have been
1271 * pre-shrunk in case there is a change in the cpu list.
1272 * So no deletion is needed.
1274 adding = cpumask_and(tmp->addmask, cpuset->cpus_allowed,
1275 parent->effective_cpus);
1276 part_error = cpumask_equal(tmp->addmask,
1277 parent->effective_cpus);
1280 if (cmd == partcmd_update) {
1281 int prev_prs = cpuset->partition_root_state;
1284 * Check for possible transition between PRS_ENABLED
1287 switch (cpuset->partition_root_state) {
1290 new_prs = PRS_ERROR;
1294 new_prs = PRS_ENABLED;
1298 * Set part_error if previously in invalid state.
1300 part_error = (prev_prs == PRS_ERROR);
1303 if (!part_error && (new_prs == PRS_ERROR))
1304 return 0; /* Nothing need to be done */
1306 if (new_prs == PRS_ERROR) {
1308 * Remove all its cpus from parent's subparts_cpus.
1311 deleting = cpumask_and(tmp->delmask, cpuset->cpus_allowed,
1312 parent->subparts_cpus);
1315 if (!adding && !deleting && (new_prs == old_prs))
1319 * Change the parent's subparts_cpus.
1320 * Newly added CPUs will be removed from effective_cpus and
1321 * newly deleted ones will be added back to effective_cpus.
1323 spin_lock_irq(&callback_lock);
1325 cpumask_or(parent->subparts_cpus,
1326 parent->subparts_cpus, tmp->addmask);
1327 cpumask_andnot(parent->effective_cpus,
1328 parent->effective_cpus, tmp->addmask);
1331 cpumask_andnot(parent->subparts_cpus,
1332 parent->subparts_cpus, tmp->delmask);
1334 * Some of the CPUs in subparts_cpus might have been offlined.
1336 cpumask_and(tmp->delmask, tmp->delmask, cpu_active_mask);
1337 cpumask_or(parent->effective_cpus,
1338 parent->effective_cpus, tmp->delmask);
1341 parent->nr_subparts_cpus = cpumask_weight(parent->subparts_cpus);
1343 if (old_prs != new_prs)
1344 cpuset->partition_root_state = new_prs;
1346 spin_unlock_irq(&callback_lock);
1347 notify_partition_change(cpuset, old_prs, new_prs);
1349 return cmd == partcmd_update;
1353 * update_cpumasks_hier - Update effective cpumasks and tasks in the subtree
1354 * @cs: the cpuset to consider
1355 * @tmp: temp variables for calculating effective_cpus & partition setup
1357 * When configured cpumask is changed, the effective cpumasks of this cpuset
1358 * and all its descendants need to be updated.
1360 * On legacy hierarchy, effective_cpus will be the same with cpu_allowed.
1362 * Called with cpuset_rwsem held
1364 static void update_cpumasks_hier(struct cpuset *cs, struct tmpmasks *tmp)
1367 struct cgroup_subsys_state *pos_css;
1368 bool need_rebuild_sched_domains = false;
1369 int old_prs, new_prs;
1372 cpuset_for_each_descendant_pre(cp, pos_css, cs) {
1373 struct cpuset *parent = parent_cs(cp);
1375 compute_effective_cpumask(tmp->new_cpus, cp, parent);
1378 * If it becomes empty, inherit the effective mask of the
1379 * parent, which is guaranteed to have some CPUs.
1381 if (is_in_v2_mode() && cpumask_empty(tmp->new_cpus)) {
1382 cpumask_copy(tmp->new_cpus, parent->effective_cpus);
1383 if (!cp->use_parent_ecpus) {
1384 cp->use_parent_ecpus = true;
1385 parent->child_ecpus_count++;
1387 } else if (cp->use_parent_ecpus) {
1388 cp->use_parent_ecpus = false;
1389 WARN_ON_ONCE(!parent->child_ecpus_count);
1390 parent->child_ecpus_count--;
1394 * Skip the whole subtree if the cpumask remains the same
1395 * and has no partition root state.
1397 if (!cp->partition_root_state &&
1398 cpumask_equal(tmp->new_cpus, cp->effective_cpus)) {
1399 pos_css = css_rightmost_descendant(pos_css);
1404 * update_parent_subparts_cpumask() should have been called
1405 * for cs already in update_cpumask(). We should also call
1406 * update_tasks_cpumask() again for tasks in the parent
1407 * cpuset if the parent's subparts_cpus changes.
1409 old_prs = new_prs = cp->partition_root_state;
1410 if ((cp != cs) && old_prs) {
1411 switch (parent->partition_root_state) {
1414 * If parent is not a partition root or an
1415 * invalid partition root, clear its state
1416 * and its CS_CPU_EXCLUSIVE flag.
1418 WARN_ON_ONCE(cp->partition_root_state
1420 new_prs = PRS_DISABLED;
1423 * clear_bit() is an atomic operation and
1424 * readers aren't interested in the state
1425 * of CS_CPU_EXCLUSIVE anyway. So we can
1426 * just update the flag without holding
1427 * the callback_lock.
1429 clear_bit(CS_CPU_EXCLUSIVE, &cp->flags);
1433 if (update_parent_subparts_cpumask(cp, partcmd_update, NULL, tmp))
1434 update_tasks_cpumask(parent);
1439 * When parent is invalid, it has to be too.
1441 new_prs = PRS_ERROR;
1446 if (!css_tryget_online(&cp->css))
1450 spin_lock_irq(&callback_lock);
1452 cpumask_copy(cp->effective_cpus, tmp->new_cpus);
1453 if (cp->nr_subparts_cpus && (new_prs != PRS_ENABLED)) {
1454 cp->nr_subparts_cpus = 0;
1455 cpumask_clear(cp->subparts_cpus);
1456 } else if (cp->nr_subparts_cpus) {
1458 * Make sure that effective_cpus & subparts_cpus
1459 * are mutually exclusive.
1461 * In the unlikely event that effective_cpus
1462 * becomes empty. we clear cp->nr_subparts_cpus and
1463 * let its child partition roots to compete for
1466 cpumask_andnot(cp->effective_cpus, cp->effective_cpus,
1468 if (cpumask_empty(cp->effective_cpus)) {
1469 cpumask_copy(cp->effective_cpus, tmp->new_cpus);
1470 cpumask_clear(cp->subparts_cpus);
1471 cp->nr_subparts_cpus = 0;
1472 } else if (!cpumask_subset(cp->subparts_cpus,
1474 cpumask_andnot(cp->subparts_cpus,
1475 cp->subparts_cpus, tmp->new_cpus);
1476 cp->nr_subparts_cpus
1477 = cpumask_weight(cp->subparts_cpus);
1481 if (new_prs != old_prs)
1482 cp->partition_root_state = new_prs;
1484 spin_unlock_irq(&callback_lock);
1485 notify_partition_change(cp, old_prs, new_prs);
1487 WARN_ON(!is_in_v2_mode() &&
1488 !cpumask_equal(cp->cpus_allowed, cp->effective_cpus));
1490 update_tasks_cpumask(cp);
1493 * On legacy hierarchy, if the effective cpumask of any non-
1494 * empty cpuset is changed, we need to rebuild sched domains.
1495 * On default hierarchy, the cpuset needs to be a partition
1498 if (!cpumask_empty(cp->cpus_allowed) &&
1499 is_sched_load_balance(cp) &&
1500 (!cgroup_subsys_on_dfl(cpuset_cgrp_subsys) ||
1501 is_partition_root(cp)))
1502 need_rebuild_sched_domains = true;
1509 if (need_rebuild_sched_domains)
1510 rebuild_sched_domains_locked();
1514 * update_sibling_cpumasks - Update siblings cpumasks
1515 * @parent: Parent cpuset
1516 * @cs: Current cpuset
1517 * @tmp: Temp variables
1519 static void update_sibling_cpumasks(struct cpuset *parent, struct cpuset *cs,
1520 struct tmpmasks *tmp)
1522 struct cpuset *sibling;
1523 struct cgroup_subsys_state *pos_css;
1526 * Check all its siblings and call update_cpumasks_hier()
1527 * if their use_parent_ecpus flag is set in order for them
1528 * to use the right effective_cpus value.
1531 cpuset_for_each_child(sibling, pos_css, parent) {
1534 if (!sibling->use_parent_ecpus)
1537 update_cpumasks_hier(sibling, tmp);
1543 * update_cpumask - update the cpus_allowed mask of a cpuset and all tasks in it
1544 * @cs: the cpuset to consider
1545 * @trialcs: trial cpuset
1546 * @buf: buffer of cpu numbers written to this cpuset
1548 static int update_cpumask(struct cpuset *cs, struct cpuset *trialcs,
1552 struct tmpmasks tmp;
1554 /* top_cpuset.cpus_allowed tracks cpu_online_mask; it's read-only */
1555 if (cs == &top_cpuset)
1559 * An empty cpus_allowed is ok only if the cpuset has no tasks.
1560 * Since cpulist_parse() fails on an empty mask, we special case
1561 * that parsing. The validate_change() call ensures that cpusets
1562 * with tasks have cpus.
1565 cpumask_clear(trialcs->cpus_allowed);
1567 retval = cpulist_parse(buf, trialcs->cpus_allowed);
1571 if (!cpumask_subset(trialcs->cpus_allowed,
1572 top_cpuset.cpus_allowed))
1576 /* Nothing to do if the cpus didn't change */
1577 if (cpumask_equal(cs->cpus_allowed, trialcs->cpus_allowed))
1580 retval = validate_change(cs, trialcs);
1584 #ifdef CONFIG_CPUMASK_OFFSTACK
1586 * Use the cpumasks in trialcs for tmpmasks when they are pointers
1587 * to allocated cpumasks.
1589 tmp.addmask = trialcs->subparts_cpus;
1590 tmp.delmask = trialcs->effective_cpus;
1591 tmp.new_cpus = trialcs->cpus_allowed;
1594 if (cs->partition_root_state) {
1595 /* Cpumask of a partition root cannot be empty */
1596 if (cpumask_empty(trialcs->cpus_allowed))
1598 if (update_parent_subparts_cpumask(cs, partcmd_update,
1599 trialcs->cpus_allowed, &tmp) < 0)
1603 spin_lock_irq(&callback_lock);
1604 cpumask_copy(cs->cpus_allowed, trialcs->cpus_allowed);
1607 * Make sure that subparts_cpus is a subset of cpus_allowed.
1609 if (cs->nr_subparts_cpus) {
1610 cpumask_andnot(cs->subparts_cpus, cs->subparts_cpus,
1612 cs->nr_subparts_cpus = cpumask_weight(cs->subparts_cpus);
1614 spin_unlock_irq(&callback_lock);
1616 update_cpumasks_hier(cs, &tmp);
1618 if (cs->partition_root_state) {
1619 struct cpuset *parent = parent_cs(cs);
1622 * For partition root, update the cpumasks of sibling
1623 * cpusets if they use parent's effective_cpus.
1625 if (parent->child_ecpus_count)
1626 update_sibling_cpumasks(parent, cs, &tmp);
1632 * Migrate memory region from one set of nodes to another. This is
1633 * performed asynchronously as it can be called from process migration path
1634 * holding locks involved in process management. All mm migrations are
1635 * performed in the queued order and can be waited for by flushing
1636 * cpuset_migrate_mm_wq.
1639 struct cpuset_migrate_mm_work {
1640 struct work_struct work;
1641 struct mm_struct *mm;
1646 static void cpuset_migrate_mm_workfn(struct work_struct *work)
1648 struct cpuset_migrate_mm_work *mwork =
1649 container_of(work, struct cpuset_migrate_mm_work, work);
1651 /* on a wq worker, no need to worry about %current's mems_allowed */
1652 do_migrate_pages(mwork->mm, &mwork->from, &mwork->to, MPOL_MF_MOVE_ALL);
1657 static void cpuset_migrate_mm(struct mm_struct *mm, const nodemask_t *from,
1658 const nodemask_t *to)
1660 struct cpuset_migrate_mm_work *mwork;
1662 if (nodes_equal(*from, *to)) {
1667 mwork = kzalloc(sizeof(*mwork), GFP_KERNEL);
1670 mwork->from = *from;
1672 INIT_WORK(&mwork->work, cpuset_migrate_mm_workfn);
1673 queue_work(cpuset_migrate_mm_wq, &mwork->work);
1679 static void cpuset_post_attach(void)
1681 flush_workqueue(cpuset_migrate_mm_wq);
1685 * cpuset_change_task_nodemask - change task's mems_allowed and mempolicy
1686 * @tsk: the task to change
1687 * @newmems: new nodes that the task will be set
1689 * We use the mems_allowed_seq seqlock to safely update both tsk->mems_allowed
1690 * and rebind an eventual tasks' mempolicy. If the task is allocating in
1691 * parallel, it might temporarily see an empty intersection, which results in
1692 * a seqlock check and retry before OOM or allocation failure.
1694 static void cpuset_change_task_nodemask(struct task_struct *tsk,
1695 nodemask_t *newmems)
1699 local_irq_disable();
1700 write_seqcount_begin(&tsk->mems_allowed_seq);
1702 nodes_or(tsk->mems_allowed, tsk->mems_allowed, *newmems);
1703 mpol_rebind_task(tsk, newmems);
1704 tsk->mems_allowed = *newmems;
1706 write_seqcount_end(&tsk->mems_allowed_seq);
1712 static void *cpuset_being_rebound;
1715 * update_tasks_nodemask - Update the nodemasks of tasks in the cpuset.
1716 * @cs: the cpuset in which each task's mems_allowed mask needs to be changed
1718 * Iterate through each task of @cs updating its mems_allowed to the
1719 * effective cpuset's. As this function is called with cpuset_rwsem held,
1720 * cpuset membership stays stable.
1722 static void update_tasks_nodemask(struct cpuset *cs)
1724 static nodemask_t newmems; /* protected by cpuset_rwsem */
1725 struct css_task_iter it;
1726 struct task_struct *task;
1728 cpuset_being_rebound = cs; /* causes mpol_dup() rebind */
1730 guarantee_online_mems(cs, &newmems);
1733 * The mpol_rebind_mm() call takes mmap_lock, which we couldn't
1734 * take while holding tasklist_lock. Forks can happen - the
1735 * mpol_dup() cpuset_being_rebound check will catch such forks,
1736 * and rebind their vma mempolicies too. Because we still hold
1737 * the global cpuset_rwsem, we know that no other rebind effort
1738 * will be contending for the global variable cpuset_being_rebound.
1739 * It's ok if we rebind the same mm twice; mpol_rebind_mm()
1740 * is idempotent. Also migrate pages in each mm to new nodes.
1742 css_task_iter_start(&cs->css, 0, &it);
1743 while ((task = css_task_iter_next(&it))) {
1744 struct mm_struct *mm;
1747 cpuset_change_task_nodemask(task, &newmems);
1749 mm = get_task_mm(task);
1753 migrate = is_memory_migrate(cs);
1755 mpol_rebind_mm(mm, &cs->mems_allowed);
1757 cpuset_migrate_mm(mm, &cs->old_mems_allowed, &newmems);
1761 css_task_iter_end(&it);
1764 * All the tasks' nodemasks have been updated, update
1765 * cs->old_mems_allowed.
1767 cs->old_mems_allowed = newmems;
1769 /* We're done rebinding vmas to this cpuset's new mems_allowed. */
1770 cpuset_being_rebound = NULL;
1774 * update_nodemasks_hier - Update effective nodemasks and tasks in the subtree
1775 * @cs: the cpuset to consider
1776 * @new_mems: a temp variable for calculating new effective_mems
1778 * When configured nodemask is changed, the effective nodemasks of this cpuset
1779 * and all its descendants need to be updated.
1781 * On legacy hierarchy, effective_mems will be the same with mems_allowed.
1783 * Called with cpuset_rwsem held
1785 static void update_nodemasks_hier(struct cpuset *cs, nodemask_t *new_mems)
1788 struct cgroup_subsys_state *pos_css;
1791 cpuset_for_each_descendant_pre(cp, pos_css, cs) {
1792 struct cpuset *parent = parent_cs(cp);
1794 nodes_and(*new_mems, cp->mems_allowed, parent->effective_mems);
1797 * If it becomes empty, inherit the effective mask of the
1798 * parent, which is guaranteed to have some MEMs.
1800 if (is_in_v2_mode() && nodes_empty(*new_mems))
1801 *new_mems = parent->effective_mems;
1803 /* Skip the whole subtree if the nodemask remains the same. */
1804 if (nodes_equal(*new_mems, cp->effective_mems)) {
1805 pos_css = css_rightmost_descendant(pos_css);
1809 if (!css_tryget_online(&cp->css))
1813 spin_lock_irq(&callback_lock);
1814 cp->effective_mems = *new_mems;
1815 spin_unlock_irq(&callback_lock);
1817 WARN_ON(!is_in_v2_mode() &&
1818 !nodes_equal(cp->mems_allowed, cp->effective_mems));
1820 update_tasks_nodemask(cp);
1829 * Handle user request to change the 'mems' memory placement
1830 * of a cpuset. Needs to validate the request, update the
1831 * cpusets mems_allowed, and for each task in the cpuset,
1832 * update mems_allowed and rebind task's mempolicy and any vma
1833 * mempolicies and if the cpuset is marked 'memory_migrate',
1834 * migrate the tasks pages to the new memory.
1836 * Call with cpuset_rwsem held. May take callback_lock during call.
1837 * Will take tasklist_lock, scan tasklist for tasks in cpuset cs,
1838 * lock each such tasks mm->mmap_lock, scan its vma's and rebind
1839 * their mempolicies to the cpusets new mems_allowed.
1841 static int update_nodemask(struct cpuset *cs, struct cpuset *trialcs,
1847 * top_cpuset.mems_allowed tracks node_stats[N_MEMORY];
1850 if (cs == &top_cpuset) {
1856 * An empty mems_allowed is ok iff there are no tasks in the cpuset.
1857 * Since nodelist_parse() fails on an empty mask, we special case
1858 * that parsing. The validate_change() call ensures that cpusets
1859 * with tasks have memory.
1862 nodes_clear(trialcs->mems_allowed);
1864 retval = nodelist_parse(buf, trialcs->mems_allowed);
1868 if (!nodes_subset(trialcs->mems_allowed,
1869 top_cpuset.mems_allowed)) {
1875 if (nodes_equal(cs->mems_allowed, trialcs->mems_allowed)) {
1876 retval = 0; /* Too easy - nothing to do */
1879 retval = validate_change(cs, trialcs);
1883 check_insane_mems_config(&trialcs->mems_allowed);
1885 spin_lock_irq(&callback_lock);
1886 cs->mems_allowed = trialcs->mems_allowed;
1887 spin_unlock_irq(&callback_lock);
1889 /* use trialcs->mems_allowed as a temp variable */
1890 update_nodemasks_hier(cs, &trialcs->mems_allowed);
1895 bool current_cpuset_is_being_rebound(void)
1900 ret = task_cs(current) == cpuset_being_rebound;
1906 static int update_relax_domain_level(struct cpuset *cs, s64 val)
1909 if (val < -1 || val >= sched_domain_level_max)
1913 if (val != cs->relax_domain_level) {
1914 cs->relax_domain_level = val;
1915 if (!cpumask_empty(cs->cpus_allowed) &&
1916 is_sched_load_balance(cs))
1917 rebuild_sched_domains_locked();
1924 * update_tasks_flags - update the spread flags of tasks in the cpuset.
1925 * @cs: the cpuset in which each task's spread flags needs to be changed
1927 * Iterate through each task of @cs updating its spread flags. As this
1928 * function is called with cpuset_rwsem held, cpuset membership stays
1931 static void update_tasks_flags(struct cpuset *cs)
1933 struct css_task_iter it;
1934 struct task_struct *task;
1936 css_task_iter_start(&cs->css, 0, &it);
1937 while ((task = css_task_iter_next(&it)))
1938 cpuset_update_task_spread_flag(cs, task);
1939 css_task_iter_end(&it);
1943 * update_flag - read a 0 or a 1 in a file and update associated flag
1944 * bit: the bit to update (see cpuset_flagbits_t)
1945 * cs: the cpuset to update
1946 * turning_on: whether the flag is being set or cleared
1948 * Call with cpuset_rwsem held.
1951 static int update_flag(cpuset_flagbits_t bit, struct cpuset *cs,
1954 struct cpuset *trialcs;
1955 int balance_flag_changed;
1956 int spread_flag_changed;
1959 trialcs = alloc_trial_cpuset(cs);
1964 set_bit(bit, &trialcs->flags);
1966 clear_bit(bit, &trialcs->flags);
1968 err = validate_change(cs, trialcs);
1972 balance_flag_changed = (is_sched_load_balance(cs) !=
1973 is_sched_load_balance(trialcs));
1975 spread_flag_changed = ((is_spread_slab(cs) != is_spread_slab(trialcs))
1976 || (is_spread_page(cs) != is_spread_page(trialcs)));
1978 spin_lock_irq(&callback_lock);
1979 cs->flags = trialcs->flags;
1980 spin_unlock_irq(&callback_lock);
1982 if (!cpumask_empty(trialcs->cpus_allowed) && balance_flag_changed)
1983 rebuild_sched_domains_locked();
1985 if (spread_flag_changed)
1986 update_tasks_flags(cs);
1988 free_cpuset(trialcs);
1993 * update_prstate - update partititon_root_state
1994 * cs: the cpuset to update
1995 * new_prs: new partition root state
1997 * Call with cpuset_rwsem held.
1999 static int update_prstate(struct cpuset *cs, int new_prs)
2001 int err, old_prs = cs->partition_root_state;
2002 struct cpuset *parent = parent_cs(cs);
2003 struct tmpmasks tmpmask;
2005 if (old_prs == new_prs)
2009 * Cannot force a partial or invalid partition root to a full
2012 if (new_prs && (old_prs == PRS_ERROR))
2015 if (alloc_cpumasks(NULL, &tmpmask))
2021 * Turning on partition root requires setting the
2022 * CS_CPU_EXCLUSIVE bit implicitly as well and cpus_allowed
2025 if (cpumask_empty(cs->cpus_allowed))
2028 err = update_flag(CS_CPU_EXCLUSIVE, cs, 1);
2032 err = update_parent_subparts_cpumask(cs, partcmd_enable,
2035 update_flag(CS_CPU_EXCLUSIVE, cs, 0);
2040 * Turning off partition root will clear the
2041 * CS_CPU_EXCLUSIVE bit.
2043 if (old_prs == PRS_ERROR) {
2044 update_flag(CS_CPU_EXCLUSIVE, cs, 0);
2049 err = update_parent_subparts_cpumask(cs, partcmd_disable,
2054 /* Turning off CS_CPU_EXCLUSIVE will not return error */
2055 update_flag(CS_CPU_EXCLUSIVE, cs, 0);
2059 * Update cpumask of parent's tasks except when it is the top
2060 * cpuset as some system daemons cannot be mapped to other CPUs.
2062 if (parent != &top_cpuset)
2063 update_tasks_cpumask(parent);
2065 if (parent->child_ecpus_count)
2066 update_sibling_cpumasks(parent, cs, &tmpmask);
2068 rebuild_sched_domains_locked();
2071 spin_lock_irq(&callback_lock);
2072 cs->partition_root_state = new_prs;
2073 spin_unlock_irq(&callback_lock);
2074 notify_partition_change(cs, old_prs, new_prs);
2077 free_cpumasks(NULL, &tmpmask);
2082 * Frequency meter - How fast is some event occurring?
2084 * These routines manage a digitally filtered, constant time based,
2085 * event frequency meter. There are four routines:
2086 * fmeter_init() - initialize a frequency meter.
2087 * fmeter_markevent() - called each time the event happens.
2088 * fmeter_getrate() - returns the recent rate of such events.
2089 * fmeter_update() - internal routine used to update fmeter.
2091 * A common data structure is passed to each of these routines,
2092 * which is used to keep track of the state required to manage the
2093 * frequency meter and its digital filter.
2095 * The filter works on the number of events marked per unit time.
2096 * The filter is single-pole low-pass recursive (IIR). The time unit
2097 * is 1 second. Arithmetic is done using 32-bit integers scaled to
2098 * simulate 3 decimal digits of precision (multiplied by 1000).
2100 * With an FM_COEF of 933, and a time base of 1 second, the filter
2101 * has a half-life of 10 seconds, meaning that if the events quit
2102 * happening, then the rate returned from the fmeter_getrate()
2103 * will be cut in half each 10 seconds, until it converges to zero.
2105 * It is not worth doing a real infinitely recursive filter. If more
2106 * than FM_MAXTICKS ticks have elapsed since the last filter event,
2107 * just compute FM_MAXTICKS ticks worth, by which point the level
2110 * Limit the count of unprocessed events to FM_MAXCNT, so as to avoid
2111 * arithmetic overflow in the fmeter_update() routine.
2113 * Given the simple 32 bit integer arithmetic used, this meter works
2114 * best for reporting rates between one per millisecond (msec) and
2115 * one per 32 (approx) seconds. At constant rates faster than one
2116 * per msec it maxes out at values just under 1,000,000. At constant
2117 * rates between one per msec, and one per second it will stabilize
2118 * to a value N*1000, where N is the rate of events per second.
2119 * At constant rates between one per second and one per 32 seconds,
2120 * it will be choppy, moving up on the seconds that have an event,
2121 * and then decaying until the next event. At rates slower than
2122 * about one in 32 seconds, it decays all the way back to zero between
2126 #define FM_COEF 933 /* coefficient for half-life of 10 secs */
2127 #define FM_MAXTICKS ((u32)99) /* useless computing more ticks than this */
2128 #define FM_MAXCNT 1000000 /* limit cnt to avoid overflow */
2129 #define FM_SCALE 1000 /* faux fixed point scale */
2131 /* Initialize a frequency meter */
2132 static void fmeter_init(struct fmeter *fmp)
2137 spin_lock_init(&fmp->lock);
2140 /* Internal meter update - process cnt events and update value */
2141 static void fmeter_update(struct fmeter *fmp)
2146 now = ktime_get_seconds();
2147 ticks = now - fmp->time;
2152 ticks = min(FM_MAXTICKS, ticks);
2154 fmp->val = (FM_COEF * fmp->val) / FM_SCALE;
2157 fmp->val += ((FM_SCALE - FM_COEF) * fmp->cnt) / FM_SCALE;
2161 /* Process any previous ticks, then bump cnt by one (times scale). */
2162 static void fmeter_markevent(struct fmeter *fmp)
2164 spin_lock(&fmp->lock);
2166 fmp->cnt = min(FM_MAXCNT, fmp->cnt + FM_SCALE);
2167 spin_unlock(&fmp->lock);
2170 /* Process any previous ticks, then return current value. */
2171 static int fmeter_getrate(struct fmeter *fmp)
2175 spin_lock(&fmp->lock);
2178 spin_unlock(&fmp->lock);
2182 static struct cpuset *cpuset_attach_old_cs;
2184 /* Called by cgroups to determine if a cpuset is usable; cpuset_rwsem held */
2185 static int cpuset_can_attach(struct cgroup_taskset *tset)
2187 struct cgroup_subsys_state *css;
2189 struct task_struct *task;
2192 /* used later by cpuset_attach() */
2193 cpuset_attach_old_cs = task_cs(cgroup_taskset_first(tset, &css));
2196 percpu_down_write(&cpuset_rwsem);
2198 /* allow moving tasks into an empty cpuset if on default hierarchy */
2200 if (!is_in_v2_mode() &&
2201 (cpumask_empty(cs->cpus_allowed) || nodes_empty(cs->mems_allowed)))
2204 cgroup_taskset_for_each(task, css, tset) {
2205 ret = task_can_attach(task, cs->cpus_allowed);
2208 ret = security_task_setscheduler(task);
2214 * Mark attach is in progress. This makes validate_change() fail
2215 * changes which zero cpus/mems_allowed.
2217 cs->attach_in_progress++;
2220 percpu_up_write(&cpuset_rwsem);
2224 static void cpuset_cancel_attach(struct cgroup_taskset *tset)
2226 struct cgroup_subsys_state *css;
2228 cgroup_taskset_first(tset, &css);
2230 percpu_down_write(&cpuset_rwsem);
2231 css_cs(css)->attach_in_progress--;
2232 percpu_up_write(&cpuset_rwsem);
2236 * Protected by cpuset_rwsem. cpus_attach is used only by cpuset_attach()
2237 * but we can't allocate it dynamically there. Define it global and
2238 * allocate from cpuset_init().
2240 static cpumask_var_t cpus_attach;
2242 static void cpuset_attach(struct cgroup_taskset *tset)
2244 /* static buf protected by cpuset_rwsem */
2245 static nodemask_t cpuset_attach_nodemask_to;
2246 struct task_struct *task;
2247 struct task_struct *leader;
2248 struct cgroup_subsys_state *css;
2250 struct cpuset *oldcs = cpuset_attach_old_cs;
2252 cgroup_taskset_first(tset, &css);
2255 percpu_down_write(&cpuset_rwsem);
2257 guarantee_online_mems(cs, &cpuset_attach_nodemask_to);
2259 cgroup_taskset_for_each(task, css, tset) {
2260 if (cs != &top_cpuset)
2261 guarantee_online_cpus(task, cpus_attach);
2263 cpumask_copy(cpus_attach, task_cpu_possible_mask(task));
2265 * can_attach beforehand should guarantee that this doesn't
2266 * fail. TODO: have a better way to handle failure here
2268 WARN_ON_ONCE(set_cpus_allowed_ptr(task, cpus_attach));
2270 cpuset_change_task_nodemask(task, &cpuset_attach_nodemask_to);
2271 cpuset_update_task_spread_flag(cs, task);
2275 * Change mm for all threadgroup leaders. This is expensive and may
2276 * sleep and should be moved outside migration path proper.
2278 cpuset_attach_nodemask_to = cs->effective_mems;
2279 cgroup_taskset_for_each_leader(leader, css, tset) {
2280 struct mm_struct *mm = get_task_mm(leader);
2283 mpol_rebind_mm(mm, &cpuset_attach_nodemask_to);
2286 * old_mems_allowed is the same with mems_allowed
2287 * here, except if this task is being moved
2288 * automatically due to hotplug. In that case
2289 * @mems_allowed has been updated and is empty, so
2290 * @old_mems_allowed is the right nodesets that we
2293 if (is_memory_migrate(cs))
2294 cpuset_migrate_mm(mm, &oldcs->old_mems_allowed,
2295 &cpuset_attach_nodemask_to);
2301 cs->old_mems_allowed = cpuset_attach_nodemask_to;
2303 cs->attach_in_progress--;
2304 if (!cs->attach_in_progress)
2305 wake_up(&cpuset_attach_wq);
2307 percpu_up_write(&cpuset_rwsem);
2310 /* The various types of files and directories in a cpuset file system */
2313 FILE_MEMORY_MIGRATE,
2316 FILE_EFFECTIVE_CPULIST,
2317 FILE_EFFECTIVE_MEMLIST,
2318 FILE_SUBPARTS_CPULIST,
2322 FILE_SCHED_LOAD_BALANCE,
2323 FILE_PARTITION_ROOT,
2324 FILE_SCHED_RELAX_DOMAIN_LEVEL,
2325 FILE_MEMORY_PRESSURE_ENABLED,
2326 FILE_MEMORY_PRESSURE,
2329 } cpuset_filetype_t;
2331 static int cpuset_write_u64(struct cgroup_subsys_state *css, struct cftype *cft,
2334 struct cpuset *cs = css_cs(css);
2335 cpuset_filetype_t type = cft->private;
2339 percpu_down_write(&cpuset_rwsem);
2340 if (!is_cpuset_online(cs)) {
2346 case FILE_CPU_EXCLUSIVE:
2347 retval = update_flag(CS_CPU_EXCLUSIVE, cs, val);
2349 case FILE_MEM_EXCLUSIVE:
2350 retval = update_flag(CS_MEM_EXCLUSIVE, cs, val);
2352 case FILE_MEM_HARDWALL:
2353 retval = update_flag(CS_MEM_HARDWALL, cs, val);
2355 case FILE_SCHED_LOAD_BALANCE:
2356 retval = update_flag(CS_SCHED_LOAD_BALANCE, cs, val);
2358 case FILE_MEMORY_MIGRATE:
2359 retval = update_flag(CS_MEMORY_MIGRATE, cs, val);
2361 case FILE_MEMORY_PRESSURE_ENABLED:
2362 cpuset_memory_pressure_enabled = !!val;
2364 case FILE_SPREAD_PAGE:
2365 retval = update_flag(CS_SPREAD_PAGE, cs, val);
2367 case FILE_SPREAD_SLAB:
2368 retval = update_flag(CS_SPREAD_SLAB, cs, val);
2375 percpu_up_write(&cpuset_rwsem);
2380 static int cpuset_write_s64(struct cgroup_subsys_state *css, struct cftype *cft,
2383 struct cpuset *cs = css_cs(css);
2384 cpuset_filetype_t type = cft->private;
2385 int retval = -ENODEV;
2388 percpu_down_write(&cpuset_rwsem);
2389 if (!is_cpuset_online(cs))
2393 case FILE_SCHED_RELAX_DOMAIN_LEVEL:
2394 retval = update_relax_domain_level(cs, val);
2401 percpu_up_write(&cpuset_rwsem);
2407 * Common handling for a write to a "cpus" or "mems" file.
2409 static ssize_t cpuset_write_resmask(struct kernfs_open_file *of,
2410 char *buf, size_t nbytes, loff_t off)
2412 struct cpuset *cs = css_cs(of_css(of));
2413 struct cpuset *trialcs;
2414 int retval = -ENODEV;
2416 buf = strstrip(buf);
2419 * CPU or memory hotunplug may leave @cs w/o any execution
2420 * resources, in which case the hotplug code asynchronously updates
2421 * configuration and transfers all tasks to the nearest ancestor
2422 * which can execute.
2424 * As writes to "cpus" or "mems" may restore @cs's execution
2425 * resources, wait for the previously scheduled operations before
2426 * proceeding, so that we don't end up keep removing tasks added
2427 * after execution capability is restored.
2429 * cpuset_hotplug_work calls back into cgroup core via
2430 * cgroup_transfer_tasks() and waiting for it from a cgroupfs
2431 * operation like this one can lead to a deadlock through kernfs
2432 * active_ref protection. Let's break the protection. Losing the
2433 * protection is okay as we check whether @cs is online after
2434 * grabbing cpuset_rwsem anyway. This only happens on the legacy
2438 kernfs_break_active_protection(of->kn);
2439 flush_work(&cpuset_hotplug_work);
2442 percpu_down_write(&cpuset_rwsem);
2443 if (!is_cpuset_online(cs))
2446 trialcs = alloc_trial_cpuset(cs);
2452 switch (of_cft(of)->private) {
2454 retval = update_cpumask(cs, trialcs, buf);
2457 retval = update_nodemask(cs, trialcs, buf);
2464 free_cpuset(trialcs);
2466 percpu_up_write(&cpuset_rwsem);
2468 kernfs_unbreak_active_protection(of->kn);
2470 flush_workqueue(cpuset_migrate_mm_wq);
2471 return retval ?: nbytes;
2475 * These ascii lists should be read in a single call, by using a user
2476 * buffer large enough to hold the entire map. If read in smaller
2477 * chunks, there is no guarantee of atomicity. Since the display format
2478 * used, list of ranges of sequential numbers, is variable length,
2479 * and since these maps can change value dynamically, one could read
2480 * gibberish by doing partial reads while a list was changing.
2482 static int cpuset_common_seq_show(struct seq_file *sf, void *v)
2484 struct cpuset *cs = css_cs(seq_css(sf));
2485 cpuset_filetype_t type = seq_cft(sf)->private;
2488 spin_lock_irq(&callback_lock);
2492 seq_printf(sf, "%*pbl\n", cpumask_pr_args(cs->cpus_allowed));
2495 seq_printf(sf, "%*pbl\n", nodemask_pr_args(&cs->mems_allowed));
2497 case FILE_EFFECTIVE_CPULIST:
2498 seq_printf(sf, "%*pbl\n", cpumask_pr_args(cs->effective_cpus));
2500 case FILE_EFFECTIVE_MEMLIST:
2501 seq_printf(sf, "%*pbl\n", nodemask_pr_args(&cs->effective_mems));
2503 case FILE_SUBPARTS_CPULIST:
2504 seq_printf(sf, "%*pbl\n", cpumask_pr_args(cs->subparts_cpus));
2510 spin_unlock_irq(&callback_lock);
2514 static u64 cpuset_read_u64(struct cgroup_subsys_state *css, struct cftype *cft)
2516 struct cpuset *cs = css_cs(css);
2517 cpuset_filetype_t type = cft->private;
2519 case FILE_CPU_EXCLUSIVE:
2520 return is_cpu_exclusive(cs);
2521 case FILE_MEM_EXCLUSIVE:
2522 return is_mem_exclusive(cs);
2523 case FILE_MEM_HARDWALL:
2524 return is_mem_hardwall(cs);
2525 case FILE_SCHED_LOAD_BALANCE:
2526 return is_sched_load_balance(cs);
2527 case FILE_MEMORY_MIGRATE:
2528 return is_memory_migrate(cs);
2529 case FILE_MEMORY_PRESSURE_ENABLED:
2530 return cpuset_memory_pressure_enabled;
2531 case FILE_MEMORY_PRESSURE:
2532 return fmeter_getrate(&cs->fmeter);
2533 case FILE_SPREAD_PAGE:
2534 return is_spread_page(cs);
2535 case FILE_SPREAD_SLAB:
2536 return is_spread_slab(cs);
2541 /* Unreachable but makes gcc happy */
2545 static s64 cpuset_read_s64(struct cgroup_subsys_state *css, struct cftype *cft)
2547 struct cpuset *cs = css_cs(css);
2548 cpuset_filetype_t type = cft->private;
2550 case FILE_SCHED_RELAX_DOMAIN_LEVEL:
2551 return cs->relax_domain_level;
2556 /* Unreachable but makes gcc happy */
2560 static int sched_partition_show(struct seq_file *seq, void *v)
2562 struct cpuset *cs = css_cs(seq_css(seq));
2564 switch (cs->partition_root_state) {
2566 seq_puts(seq, "root\n");
2569 seq_puts(seq, "member\n");
2572 seq_puts(seq, "root invalid\n");
2578 static ssize_t sched_partition_write(struct kernfs_open_file *of, char *buf,
2579 size_t nbytes, loff_t off)
2581 struct cpuset *cs = css_cs(of_css(of));
2583 int retval = -ENODEV;
2585 buf = strstrip(buf);
2588 * Convert "root" to ENABLED, and convert "member" to DISABLED.
2590 if (!strcmp(buf, "root"))
2592 else if (!strcmp(buf, "member"))
2599 percpu_down_write(&cpuset_rwsem);
2600 if (!is_cpuset_online(cs))
2603 retval = update_prstate(cs, val);
2605 percpu_up_write(&cpuset_rwsem);
2608 return retval ?: nbytes;
2612 * for the common functions, 'private' gives the type of file
2615 static struct cftype legacy_files[] = {
2618 .seq_show = cpuset_common_seq_show,
2619 .write = cpuset_write_resmask,
2620 .max_write_len = (100U + 6 * NR_CPUS),
2621 .private = FILE_CPULIST,
2626 .seq_show = cpuset_common_seq_show,
2627 .write = cpuset_write_resmask,
2628 .max_write_len = (100U + 6 * MAX_NUMNODES),
2629 .private = FILE_MEMLIST,
2633 .name = "effective_cpus",
2634 .seq_show = cpuset_common_seq_show,
2635 .private = FILE_EFFECTIVE_CPULIST,
2639 .name = "effective_mems",
2640 .seq_show = cpuset_common_seq_show,
2641 .private = FILE_EFFECTIVE_MEMLIST,
2645 .name = "cpu_exclusive",
2646 .read_u64 = cpuset_read_u64,
2647 .write_u64 = cpuset_write_u64,
2648 .private = FILE_CPU_EXCLUSIVE,
2652 .name = "mem_exclusive",
2653 .read_u64 = cpuset_read_u64,
2654 .write_u64 = cpuset_write_u64,
2655 .private = FILE_MEM_EXCLUSIVE,
2659 .name = "mem_hardwall",
2660 .read_u64 = cpuset_read_u64,
2661 .write_u64 = cpuset_write_u64,
2662 .private = FILE_MEM_HARDWALL,
2666 .name = "sched_load_balance",
2667 .read_u64 = cpuset_read_u64,
2668 .write_u64 = cpuset_write_u64,
2669 .private = FILE_SCHED_LOAD_BALANCE,
2673 .name = "sched_relax_domain_level",
2674 .read_s64 = cpuset_read_s64,
2675 .write_s64 = cpuset_write_s64,
2676 .private = FILE_SCHED_RELAX_DOMAIN_LEVEL,
2680 .name = "memory_migrate",
2681 .read_u64 = cpuset_read_u64,
2682 .write_u64 = cpuset_write_u64,
2683 .private = FILE_MEMORY_MIGRATE,
2687 .name = "memory_pressure",
2688 .read_u64 = cpuset_read_u64,
2689 .private = FILE_MEMORY_PRESSURE,
2693 .name = "memory_spread_page",
2694 .read_u64 = cpuset_read_u64,
2695 .write_u64 = cpuset_write_u64,
2696 .private = FILE_SPREAD_PAGE,
2700 .name = "memory_spread_slab",
2701 .read_u64 = cpuset_read_u64,
2702 .write_u64 = cpuset_write_u64,
2703 .private = FILE_SPREAD_SLAB,
2707 .name = "memory_pressure_enabled",
2708 .flags = CFTYPE_ONLY_ON_ROOT,
2709 .read_u64 = cpuset_read_u64,
2710 .write_u64 = cpuset_write_u64,
2711 .private = FILE_MEMORY_PRESSURE_ENABLED,
2718 * This is currently a minimal set for the default hierarchy. It can be
2719 * expanded later on by migrating more features and control files from v1.
2721 static struct cftype dfl_files[] = {
2724 .seq_show = cpuset_common_seq_show,
2725 .write = cpuset_write_resmask,
2726 .max_write_len = (100U + 6 * NR_CPUS),
2727 .private = FILE_CPULIST,
2728 .flags = CFTYPE_NOT_ON_ROOT,
2733 .seq_show = cpuset_common_seq_show,
2734 .write = cpuset_write_resmask,
2735 .max_write_len = (100U + 6 * MAX_NUMNODES),
2736 .private = FILE_MEMLIST,
2737 .flags = CFTYPE_NOT_ON_ROOT,
2741 .name = "cpus.effective",
2742 .seq_show = cpuset_common_seq_show,
2743 .private = FILE_EFFECTIVE_CPULIST,
2747 .name = "mems.effective",
2748 .seq_show = cpuset_common_seq_show,
2749 .private = FILE_EFFECTIVE_MEMLIST,
2753 .name = "cpus.partition",
2754 .seq_show = sched_partition_show,
2755 .write = sched_partition_write,
2756 .private = FILE_PARTITION_ROOT,
2757 .flags = CFTYPE_NOT_ON_ROOT,
2758 .file_offset = offsetof(struct cpuset, partition_file),
2762 .name = "cpus.subpartitions",
2763 .seq_show = cpuset_common_seq_show,
2764 .private = FILE_SUBPARTS_CPULIST,
2765 .flags = CFTYPE_DEBUG,
2773 * cpuset_css_alloc - allocate a cpuset css
2774 * cgrp: control group that the new cpuset will be part of
2777 static struct cgroup_subsys_state *
2778 cpuset_css_alloc(struct cgroup_subsys_state *parent_css)
2783 return &top_cpuset.css;
2785 cs = kzalloc(sizeof(*cs), GFP_KERNEL);
2787 return ERR_PTR(-ENOMEM);
2789 if (alloc_cpumasks(cs, NULL)) {
2791 return ERR_PTR(-ENOMEM);
2794 __set_bit(CS_SCHED_LOAD_BALANCE, &cs->flags);
2795 nodes_clear(cs->mems_allowed);
2796 nodes_clear(cs->effective_mems);
2797 fmeter_init(&cs->fmeter);
2798 cs->relax_domain_level = -1;
2800 /* Set CS_MEMORY_MIGRATE for default hierarchy */
2801 if (cgroup_subsys_on_dfl(cpuset_cgrp_subsys))
2802 __set_bit(CS_MEMORY_MIGRATE, &cs->flags);
2807 static int cpuset_css_online(struct cgroup_subsys_state *css)
2809 struct cpuset *cs = css_cs(css);
2810 struct cpuset *parent = parent_cs(cs);
2811 struct cpuset *tmp_cs;
2812 struct cgroup_subsys_state *pos_css;
2818 percpu_down_write(&cpuset_rwsem);
2820 set_bit(CS_ONLINE, &cs->flags);
2821 if (is_spread_page(parent))
2822 set_bit(CS_SPREAD_PAGE, &cs->flags);
2823 if (is_spread_slab(parent))
2824 set_bit(CS_SPREAD_SLAB, &cs->flags);
2828 spin_lock_irq(&callback_lock);
2829 if (is_in_v2_mode()) {
2830 cpumask_copy(cs->effective_cpus, parent->effective_cpus);
2831 cs->effective_mems = parent->effective_mems;
2832 cs->use_parent_ecpus = true;
2833 parent->child_ecpus_count++;
2835 spin_unlock_irq(&callback_lock);
2837 if (!test_bit(CGRP_CPUSET_CLONE_CHILDREN, &css->cgroup->flags))
2841 * Clone @parent's configuration if CGRP_CPUSET_CLONE_CHILDREN is
2842 * set. This flag handling is implemented in cgroup core for
2843 * histrical reasons - the flag may be specified during mount.
2845 * Currently, if any sibling cpusets have exclusive cpus or mem, we
2846 * refuse to clone the configuration - thereby refusing the task to
2847 * be entered, and as a result refusing the sys_unshare() or
2848 * clone() which initiated it. If this becomes a problem for some
2849 * users who wish to allow that scenario, then this could be
2850 * changed to grant parent->cpus_allowed-sibling_cpus_exclusive
2851 * (and likewise for mems) to the new cgroup.
2854 cpuset_for_each_child(tmp_cs, pos_css, parent) {
2855 if (is_mem_exclusive(tmp_cs) || is_cpu_exclusive(tmp_cs)) {
2862 spin_lock_irq(&callback_lock);
2863 cs->mems_allowed = parent->mems_allowed;
2864 cs->effective_mems = parent->mems_allowed;
2865 cpumask_copy(cs->cpus_allowed, parent->cpus_allowed);
2866 cpumask_copy(cs->effective_cpus, parent->cpus_allowed);
2867 spin_unlock_irq(&callback_lock);
2869 percpu_up_write(&cpuset_rwsem);
2875 * If the cpuset being removed has its flag 'sched_load_balance'
2876 * enabled, then simulate turning sched_load_balance off, which
2877 * will call rebuild_sched_domains_locked(). That is not needed
2878 * in the default hierarchy where only changes in partition
2879 * will cause repartitioning.
2881 * If the cpuset has the 'sched.partition' flag enabled, simulate
2882 * turning 'sched.partition" off.
2885 static void cpuset_css_offline(struct cgroup_subsys_state *css)
2887 struct cpuset *cs = css_cs(css);
2890 percpu_down_write(&cpuset_rwsem);
2892 if (is_partition_root(cs))
2893 update_prstate(cs, 0);
2895 if (!cgroup_subsys_on_dfl(cpuset_cgrp_subsys) &&
2896 is_sched_load_balance(cs))
2897 update_flag(CS_SCHED_LOAD_BALANCE, cs, 0);
2899 if (cs->use_parent_ecpus) {
2900 struct cpuset *parent = parent_cs(cs);
2902 cs->use_parent_ecpus = false;
2903 parent->child_ecpus_count--;
2907 clear_bit(CS_ONLINE, &cs->flags);
2909 percpu_up_write(&cpuset_rwsem);
2913 static void cpuset_css_free(struct cgroup_subsys_state *css)
2915 struct cpuset *cs = css_cs(css);
2920 static void cpuset_bind(struct cgroup_subsys_state *root_css)
2922 percpu_down_write(&cpuset_rwsem);
2923 spin_lock_irq(&callback_lock);
2925 if (is_in_v2_mode()) {
2926 cpumask_copy(top_cpuset.cpus_allowed, cpu_possible_mask);
2927 top_cpuset.mems_allowed = node_possible_map;
2929 cpumask_copy(top_cpuset.cpus_allowed,
2930 top_cpuset.effective_cpus);
2931 top_cpuset.mems_allowed = top_cpuset.effective_mems;
2934 spin_unlock_irq(&callback_lock);
2935 percpu_up_write(&cpuset_rwsem);
2939 * Make sure the new task conform to the current state of its parent,
2940 * which could have been changed by cpuset just after it inherits the
2941 * state from the parent and before it sits on the cgroup's task list.
2943 static void cpuset_fork(struct task_struct *task)
2945 if (task_css_is_root(task, cpuset_cgrp_id))
2948 set_cpus_allowed_ptr(task, current->cpus_ptr);
2949 task->mems_allowed = current->mems_allowed;
2952 struct cgroup_subsys cpuset_cgrp_subsys = {
2953 .css_alloc = cpuset_css_alloc,
2954 .css_online = cpuset_css_online,
2955 .css_offline = cpuset_css_offline,
2956 .css_free = cpuset_css_free,
2957 .can_attach = cpuset_can_attach,
2958 .cancel_attach = cpuset_cancel_attach,
2959 .attach = cpuset_attach,
2960 .post_attach = cpuset_post_attach,
2961 .bind = cpuset_bind,
2962 .fork = cpuset_fork,
2963 .legacy_cftypes = legacy_files,
2964 .dfl_cftypes = dfl_files,
2970 * cpuset_init - initialize cpusets at system boot
2972 * Description: Initialize top_cpuset
2975 int __init cpuset_init(void)
2977 BUG_ON(percpu_init_rwsem(&cpuset_rwsem));
2979 BUG_ON(!alloc_cpumask_var(&top_cpuset.cpus_allowed, GFP_KERNEL));
2980 BUG_ON(!alloc_cpumask_var(&top_cpuset.effective_cpus, GFP_KERNEL));
2981 BUG_ON(!zalloc_cpumask_var(&top_cpuset.subparts_cpus, GFP_KERNEL));
2983 cpumask_setall(top_cpuset.cpus_allowed);
2984 nodes_setall(top_cpuset.mems_allowed);
2985 cpumask_setall(top_cpuset.effective_cpus);
2986 nodes_setall(top_cpuset.effective_mems);
2988 fmeter_init(&top_cpuset.fmeter);
2989 set_bit(CS_SCHED_LOAD_BALANCE, &top_cpuset.flags);
2990 top_cpuset.relax_domain_level = -1;
2992 BUG_ON(!alloc_cpumask_var(&cpus_attach, GFP_KERNEL));
2998 * If CPU and/or memory hotplug handlers, below, unplug any CPUs
2999 * or memory nodes, we need to walk over the cpuset hierarchy,
3000 * removing that CPU or node from all cpusets. If this removes the
3001 * last CPU or node from a cpuset, then move the tasks in the empty
3002 * cpuset to its next-highest non-empty parent.
3004 static void remove_tasks_in_empty_cpuset(struct cpuset *cs)
3006 struct cpuset *parent;
3009 * Find its next-highest non-empty parent, (top cpuset
3010 * has online cpus, so can't be empty).
3012 parent = parent_cs(cs);
3013 while (cpumask_empty(parent->cpus_allowed) ||
3014 nodes_empty(parent->mems_allowed))
3015 parent = parent_cs(parent);
3017 if (cgroup_transfer_tasks(parent->css.cgroup, cs->css.cgroup)) {
3018 pr_err("cpuset: failed to transfer tasks out of empty cpuset ");
3019 pr_cont_cgroup_name(cs->css.cgroup);
3025 hotplug_update_tasks_legacy(struct cpuset *cs,
3026 struct cpumask *new_cpus, nodemask_t *new_mems,
3027 bool cpus_updated, bool mems_updated)
3031 spin_lock_irq(&callback_lock);
3032 cpumask_copy(cs->cpus_allowed, new_cpus);
3033 cpumask_copy(cs->effective_cpus, new_cpus);
3034 cs->mems_allowed = *new_mems;
3035 cs->effective_mems = *new_mems;
3036 spin_unlock_irq(&callback_lock);
3039 * Don't call update_tasks_cpumask() if the cpuset becomes empty,
3040 * as the tasks will be migratecd to an ancestor.
3042 if (cpus_updated && !cpumask_empty(cs->cpus_allowed))
3043 update_tasks_cpumask(cs);
3044 if (mems_updated && !nodes_empty(cs->mems_allowed))
3045 update_tasks_nodemask(cs);
3047 is_empty = cpumask_empty(cs->cpus_allowed) ||
3048 nodes_empty(cs->mems_allowed);
3050 percpu_up_write(&cpuset_rwsem);
3053 * Move tasks to the nearest ancestor with execution resources,
3054 * This is full cgroup operation which will also call back into
3055 * cpuset. Should be done outside any lock.
3058 remove_tasks_in_empty_cpuset(cs);
3060 percpu_down_write(&cpuset_rwsem);
3064 hotplug_update_tasks(struct cpuset *cs,
3065 struct cpumask *new_cpus, nodemask_t *new_mems,
3066 bool cpus_updated, bool mems_updated)
3068 if (cpumask_empty(new_cpus))
3069 cpumask_copy(new_cpus, parent_cs(cs)->effective_cpus);
3070 if (nodes_empty(*new_mems))
3071 *new_mems = parent_cs(cs)->effective_mems;
3073 spin_lock_irq(&callback_lock);
3074 cpumask_copy(cs->effective_cpus, new_cpus);
3075 cs->effective_mems = *new_mems;
3076 spin_unlock_irq(&callback_lock);
3079 update_tasks_cpumask(cs);
3081 update_tasks_nodemask(cs);
3084 static bool force_rebuild;
3086 void cpuset_force_rebuild(void)
3088 force_rebuild = true;
3092 * cpuset_hotplug_update_tasks - update tasks in a cpuset for hotunplug
3093 * @cs: cpuset in interest
3094 * @tmp: the tmpmasks structure pointer
3096 * Compare @cs's cpu and mem masks against top_cpuset and if some have gone
3097 * offline, update @cs accordingly. If @cs ends up with no CPU or memory,
3098 * all its tasks are moved to the nearest ancestor with both resources.
3100 static void cpuset_hotplug_update_tasks(struct cpuset *cs, struct tmpmasks *tmp)
3102 static cpumask_t new_cpus;
3103 static nodemask_t new_mems;
3106 struct cpuset *parent;
3108 wait_event(cpuset_attach_wq, cs->attach_in_progress == 0);
3110 percpu_down_write(&cpuset_rwsem);
3113 * We have raced with task attaching. We wait until attaching
3114 * is finished, so we won't attach a task to an empty cpuset.
3116 if (cs->attach_in_progress) {
3117 percpu_up_write(&cpuset_rwsem);
3121 parent = parent_cs(cs);
3122 compute_effective_cpumask(&new_cpus, cs, parent);
3123 nodes_and(new_mems, cs->mems_allowed, parent->effective_mems);
3125 if (cs->nr_subparts_cpus)
3127 * Make sure that CPUs allocated to child partitions
3128 * do not show up in effective_cpus.
3130 cpumask_andnot(&new_cpus, &new_cpus, cs->subparts_cpus);
3132 if (!tmp || !cs->partition_root_state)
3136 * In the unlikely event that a partition root has empty
3137 * effective_cpus or its parent becomes erroneous, we have to
3138 * transition it to the erroneous state.
3140 if (is_partition_root(cs) && (cpumask_empty(&new_cpus) ||
3141 (parent->partition_root_state == PRS_ERROR))) {
3142 if (cs->nr_subparts_cpus) {
3143 spin_lock_irq(&callback_lock);
3144 cs->nr_subparts_cpus = 0;
3145 cpumask_clear(cs->subparts_cpus);
3146 spin_unlock_irq(&callback_lock);
3147 compute_effective_cpumask(&new_cpus, cs, parent);
3151 * If the effective_cpus is empty because the child
3152 * partitions take away all the CPUs, we can keep
3153 * the current partition and let the child partitions
3154 * fight for available CPUs.
3156 if ((parent->partition_root_state == PRS_ERROR) ||
3157 cpumask_empty(&new_cpus)) {
3160 update_parent_subparts_cpumask(cs, partcmd_disable,
3162 old_prs = cs->partition_root_state;
3163 if (old_prs != PRS_ERROR) {
3164 spin_lock_irq(&callback_lock);
3165 cs->partition_root_state = PRS_ERROR;
3166 spin_unlock_irq(&callback_lock);
3167 notify_partition_change(cs, old_prs, PRS_ERROR);
3170 cpuset_force_rebuild();
3174 * On the other hand, an erroneous partition root may be transitioned
3175 * back to a regular one or a partition root with no CPU allocated
3176 * from the parent may change to erroneous.
3178 if (is_partition_root(parent) &&
3179 ((cs->partition_root_state == PRS_ERROR) ||
3180 !cpumask_intersects(&new_cpus, parent->subparts_cpus)) &&
3181 update_parent_subparts_cpumask(cs, partcmd_update, NULL, tmp))
3182 cpuset_force_rebuild();
3185 cpus_updated = !cpumask_equal(&new_cpus, cs->effective_cpus);
3186 mems_updated = !nodes_equal(new_mems, cs->effective_mems);
3189 check_insane_mems_config(&new_mems);
3191 if (is_in_v2_mode())
3192 hotplug_update_tasks(cs, &new_cpus, &new_mems,
3193 cpus_updated, mems_updated);
3195 hotplug_update_tasks_legacy(cs, &new_cpus, &new_mems,
3196 cpus_updated, mems_updated);
3198 percpu_up_write(&cpuset_rwsem);
3202 * cpuset_hotplug_workfn - handle CPU/memory hotunplug for a cpuset
3204 * This function is called after either CPU or memory configuration has
3205 * changed and updates cpuset accordingly. The top_cpuset is always
3206 * synchronized to cpu_active_mask and N_MEMORY, which is necessary in
3207 * order to make cpusets transparent (of no affect) on systems that are
3208 * actively using CPU hotplug but making no active use of cpusets.
3210 * Non-root cpusets are only affected by offlining. If any CPUs or memory
3211 * nodes have been taken down, cpuset_hotplug_update_tasks() is invoked on
3214 * Note that CPU offlining during suspend is ignored. We don't modify
3215 * cpusets across suspend/resume cycles at all.
3217 static void cpuset_hotplug_workfn(struct work_struct *work)
3219 static cpumask_t new_cpus;
3220 static nodemask_t new_mems;
3221 bool cpus_updated, mems_updated;
3222 bool on_dfl = is_in_v2_mode();
3223 struct tmpmasks tmp, *ptmp = NULL;
3225 if (on_dfl && !alloc_cpumasks(NULL, &tmp))
3228 percpu_down_write(&cpuset_rwsem);
3230 /* fetch the available cpus/mems and find out which changed how */
3231 cpumask_copy(&new_cpus, cpu_active_mask);
3232 new_mems = node_states[N_MEMORY];
3235 * If subparts_cpus is populated, it is likely that the check below
3236 * will produce a false positive on cpus_updated when the cpu list
3237 * isn't changed. It is extra work, but it is better to be safe.
3239 cpus_updated = !cpumask_equal(top_cpuset.effective_cpus, &new_cpus);
3240 mems_updated = !nodes_equal(top_cpuset.effective_mems, new_mems);
3243 * In the rare case that hotplug removes all the cpus in subparts_cpus,
3244 * we assumed that cpus are updated.
3246 if (!cpus_updated && top_cpuset.nr_subparts_cpus)
3247 cpus_updated = true;
3249 /* synchronize cpus_allowed to cpu_active_mask */
3251 spin_lock_irq(&callback_lock);
3253 cpumask_copy(top_cpuset.cpus_allowed, &new_cpus);
3255 * Make sure that CPUs allocated to child partitions
3256 * do not show up in effective_cpus. If no CPU is left,
3257 * we clear the subparts_cpus & let the child partitions
3258 * fight for the CPUs again.
3260 if (top_cpuset.nr_subparts_cpus) {
3261 if (cpumask_subset(&new_cpus,
3262 top_cpuset.subparts_cpus)) {
3263 top_cpuset.nr_subparts_cpus = 0;
3264 cpumask_clear(top_cpuset.subparts_cpus);
3266 cpumask_andnot(&new_cpus, &new_cpus,
3267 top_cpuset.subparts_cpus);
3270 cpumask_copy(top_cpuset.effective_cpus, &new_cpus);
3271 spin_unlock_irq(&callback_lock);
3272 /* we don't mess with cpumasks of tasks in top_cpuset */
3275 /* synchronize mems_allowed to N_MEMORY */
3277 spin_lock_irq(&callback_lock);
3279 top_cpuset.mems_allowed = new_mems;
3280 top_cpuset.effective_mems = new_mems;
3281 spin_unlock_irq(&callback_lock);
3282 update_tasks_nodemask(&top_cpuset);
3285 percpu_up_write(&cpuset_rwsem);
3287 /* if cpus or mems changed, we need to propagate to descendants */
3288 if (cpus_updated || mems_updated) {
3290 struct cgroup_subsys_state *pos_css;
3293 cpuset_for_each_descendant_pre(cs, pos_css, &top_cpuset) {
3294 if (cs == &top_cpuset || !css_tryget_online(&cs->css))
3298 cpuset_hotplug_update_tasks(cs, ptmp);
3306 /* rebuild sched domains if cpus_allowed has changed */
3307 if (cpus_updated || force_rebuild) {
3308 force_rebuild = false;
3309 rebuild_sched_domains();
3312 free_cpumasks(NULL, ptmp);
3315 void cpuset_update_active_cpus(void)
3318 * We're inside cpu hotplug critical region which usually nests
3319 * inside cgroup synchronization. Bounce actual hotplug processing
3320 * to a work item to avoid reverse locking order.
3322 schedule_work(&cpuset_hotplug_work);
3325 void cpuset_wait_for_hotplug(void)
3327 flush_work(&cpuset_hotplug_work);
3331 * Keep top_cpuset.mems_allowed tracking node_states[N_MEMORY].
3332 * Call this routine anytime after node_states[N_MEMORY] changes.
3333 * See cpuset_update_active_cpus() for CPU hotplug handling.
3335 static int cpuset_track_online_nodes(struct notifier_block *self,
3336 unsigned long action, void *arg)
3338 schedule_work(&cpuset_hotplug_work);
3342 static struct notifier_block cpuset_track_online_nodes_nb = {
3343 .notifier_call = cpuset_track_online_nodes,
3344 .priority = 10, /* ??! */
3348 * cpuset_init_smp - initialize cpus_allowed
3350 * Description: Finish top cpuset after cpu, node maps are initialized
3352 void __init cpuset_init_smp(void)
3354 cpumask_copy(top_cpuset.cpus_allowed, cpu_active_mask);
3355 top_cpuset.mems_allowed = node_states[N_MEMORY];
3356 top_cpuset.old_mems_allowed = top_cpuset.mems_allowed;
3358 cpumask_copy(top_cpuset.effective_cpus, cpu_active_mask);
3359 top_cpuset.effective_mems = node_states[N_MEMORY];
3361 register_hotmemory_notifier(&cpuset_track_online_nodes_nb);
3363 cpuset_migrate_mm_wq = alloc_ordered_workqueue("cpuset_migrate_mm", 0);
3364 BUG_ON(!cpuset_migrate_mm_wq);
3368 * cpuset_cpus_allowed - return cpus_allowed mask from a tasks cpuset.
3369 * @tsk: pointer to task_struct from which to obtain cpuset->cpus_allowed.
3370 * @pmask: pointer to struct cpumask variable to receive cpus_allowed set.
3372 * Description: Returns the cpumask_var_t cpus_allowed of the cpuset
3373 * attached to the specified @tsk. Guaranteed to return some non-empty
3374 * subset of cpu_online_mask, even if this means going outside the
3378 void cpuset_cpus_allowed(struct task_struct *tsk, struct cpumask *pmask)
3380 unsigned long flags;
3382 spin_lock_irqsave(&callback_lock, flags);
3383 guarantee_online_cpus(tsk, pmask);
3384 spin_unlock_irqrestore(&callback_lock, flags);
3388 * cpuset_cpus_allowed_fallback - final fallback before complete catastrophe.
3389 * @tsk: pointer to task_struct with which the scheduler is struggling
3391 * Description: In the case that the scheduler cannot find an allowed cpu in
3392 * tsk->cpus_allowed, we fall back to task_cs(tsk)->cpus_allowed. In legacy
3393 * mode however, this value is the same as task_cs(tsk)->effective_cpus,
3394 * which will not contain a sane cpumask during cases such as cpu hotplugging.
3395 * This is the absolute last resort for the scheduler and it is only used if
3396 * _every_ other avenue has been traveled.
3398 * Returns true if the affinity of @tsk was changed, false otherwise.
3401 bool cpuset_cpus_allowed_fallback(struct task_struct *tsk)
3403 const struct cpumask *possible_mask = task_cpu_possible_mask(tsk);
3404 const struct cpumask *cs_mask;
3405 bool changed = false;
3408 cs_mask = task_cs(tsk)->cpus_allowed;
3409 if (is_in_v2_mode() && cpumask_subset(cs_mask, possible_mask)) {
3410 do_set_cpus_allowed(tsk, cs_mask);
3416 * We own tsk->cpus_allowed, nobody can change it under us.
3418 * But we used cs && cs->cpus_allowed lockless and thus can
3419 * race with cgroup_attach_task() or update_cpumask() and get
3420 * the wrong tsk->cpus_allowed. However, both cases imply the
3421 * subsequent cpuset_change_cpumask()->set_cpus_allowed_ptr()
3422 * which takes task_rq_lock().
3424 * If we are called after it dropped the lock we must see all
3425 * changes in tsk_cs()->cpus_allowed. Otherwise we can temporary
3426 * set any mask even if it is not right from task_cs() pov,
3427 * the pending set_cpus_allowed_ptr() will fix things.
3429 * select_fallback_rq() will fix things ups and set cpu_possible_mask
3435 void __init cpuset_init_current_mems_allowed(void)
3437 nodes_setall(current->mems_allowed);
3441 * cpuset_mems_allowed - return mems_allowed mask from a tasks cpuset.
3442 * @tsk: pointer to task_struct from which to obtain cpuset->mems_allowed.
3444 * Description: Returns the nodemask_t mems_allowed of the cpuset
3445 * attached to the specified @tsk. Guaranteed to return some non-empty
3446 * subset of node_states[N_MEMORY], even if this means going outside the
3450 nodemask_t cpuset_mems_allowed(struct task_struct *tsk)
3453 unsigned long flags;
3455 spin_lock_irqsave(&callback_lock, flags);
3457 guarantee_online_mems(task_cs(tsk), &mask);
3459 spin_unlock_irqrestore(&callback_lock, flags);
3465 * cpuset_nodemask_valid_mems_allowed - check nodemask vs. current mems_allowed
3466 * @nodemask: the nodemask to be checked
3468 * Are any of the nodes in the nodemask allowed in current->mems_allowed?
3470 int cpuset_nodemask_valid_mems_allowed(nodemask_t *nodemask)
3472 return nodes_intersects(*nodemask, current->mems_allowed);
3476 * nearest_hardwall_ancestor() - Returns the nearest mem_exclusive or
3477 * mem_hardwall ancestor to the specified cpuset. Call holding
3478 * callback_lock. If no ancestor is mem_exclusive or mem_hardwall
3479 * (an unusual configuration), then returns the root cpuset.
3481 static struct cpuset *nearest_hardwall_ancestor(struct cpuset *cs)
3483 while (!(is_mem_exclusive(cs) || is_mem_hardwall(cs)) && parent_cs(cs))
3489 * cpuset_node_allowed - Can we allocate on a memory node?
3490 * @node: is this an allowed node?
3491 * @gfp_mask: memory allocation flags
3493 * If we're in interrupt, yes, we can always allocate. If @node is set in
3494 * current's mems_allowed, yes. If it's not a __GFP_HARDWALL request and this
3495 * node is set in the nearest hardwalled cpuset ancestor to current's cpuset,
3496 * yes. If current has access to memory reserves as an oom victim, yes.
3499 * GFP_USER allocations are marked with the __GFP_HARDWALL bit,
3500 * and do not allow allocations outside the current tasks cpuset
3501 * unless the task has been OOM killed.
3502 * GFP_KERNEL allocations are not so marked, so can escape to the
3503 * nearest enclosing hardwalled ancestor cpuset.
3505 * Scanning up parent cpusets requires callback_lock. The
3506 * __alloc_pages() routine only calls here with __GFP_HARDWALL bit
3507 * _not_ set if it's a GFP_KERNEL allocation, and all nodes in the
3508 * current tasks mems_allowed came up empty on the first pass over
3509 * the zonelist. So only GFP_KERNEL allocations, if all nodes in the
3510 * cpuset are short of memory, might require taking the callback_lock.
3512 * The first call here from mm/page_alloc:get_page_from_freelist()
3513 * has __GFP_HARDWALL set in gfp_mask, enforcing hardwall cpusets,
3514 * so no allocation on a node outside the cpuset is allowed (unless
3515 * in interrupt, of course).
3517 * The second pass through get_page_from_freelist() doesn't even call
3518 * here for GFP_ATOMIC calls. For those calls, the __alloc_pages()
3519 * variable 'wait' is not set, and the bit ALLOC_CPUSET is not set
3520 * in alloc_flags. That logic and the checks below have the combined
3522 * in_interrupt - any node ok (current task context irrelevant)
3523 * GFP_ATOMIC - any node ok
3524 * tsk_is_oom_victim - any node ok
3525 * GFP_KERNEL - any node in enclosing hardwalled cpuset ok
3526 * GFP_USER - only nodes in current tasks mems allowed ok.
3528 bool __cpuset_node_allowed(int node, gfp_t gfp_mask)
3530 struct cpuset *cs; /* current cpuset ancestors */
3531 bool allowed; /* is allocation in zone z allowed? */
3532 unsigned long flags;
3536 if (node_isset(node, current->mems_allowed))
3539 * Allow tasks that have access to memory reserves because they have
3540 * been OOM killed to get memory anywhere.
3542 if (unlikely(tsk_is_oom_victim(current)))
3544 if (gfp_mask & __GFP_HARDWALL) /* If hardwall request, stop here */
3547 if (current->flags & PF_EXITING) /* Let dying task have memory */
3550 /* Not hardwall and node outside mems_allowed: scan up cpusets */
3551 spin_lock_irqsave(&callback_lock, flags);
3554 cs = nearest_hardwall_ancestor(task_cs(current));
3555 allowed = node_isset(node, cs->mems_allowed);
3558 spin_unlock_irqrestore(&callback_lock, flags);
3563 * cpuset_mem_spread_node() - On which node to begin search for a file page
3564 * cpuset_slab_spread_node() - On which node to begin search for a slab page
3566 * If a task is marked PF_SPREAD_PAGE or PF_SPREAD_SLAB (as for
3567 * tasks in a cpuset with is_spread_page or is_spread_slab set),
3568 * and if the memory allocation used cpuset_mem_spread_node()
3569 * to determine on which node to start looking, as it will for
3570 * certain page cache or slab cache pages such as used for file
3571 * system buffers and inode caches, then instead of starting on the
3572 * local node to look for a free page, rather spread the starting
3573 * node around the tasks mems_allowed nodes.
3575 * We don't have to worry about the returned node being offline
3576 * because "it can't happen", and even if it did, it would be ok.
3578 * The routines calling guarantee_online_mems() are careful to
3579 * only set nodes in task->mems_allowed that are online. So it
3580 * should not be possible for the following code to return an
3581 * offline node. But if it did, that would be ok, as this routine
3582 * is not returning the node where the allocation must be, only
3583 * the node where the search should start. The zonelist passed to
3584 * __alloc_pages() will include all nodes. If the slab allocator
3585 * is passed an offline node, it will fall back to the local node.
3586 * See kmem_cache_alloc_node().
3589 static int cpuset_spread_node(int *rotor)
3591 return *rotor = next_node_in(*rotor, current->mems_allowed);
3594 int cpuset_mem_spread_node(void)
3596 if (current->cpuset_mem_spread_rotor == NUMA_NO_NODE)
3597 current->cpuset_mem_spread_rotor =
3598 node_random(¤t->mems_allowed);
3600 return cpuset_spread_node(¤t->cpuset_mem_spread_rotor);
3603 int cpuset_slab_spread_node(void)
3605 if (current->cpuset_slab_spread_rotor == NUMA_NO_NODE)
3606 current->cpuset_slab_spread_rotor =
3607 node_random(¤t->mems_allowed);
3609 return cpuset_spread_node(¤t->cpuset_slab_spread_rotor);
3612 EXPORT_SYMBOL_GPL(cpuset_mem_spread_node);
3615 * cpuset_mems_allowed_intersects - Does @tsk1's mems_allowed intersect @tsk2's?
3616 * @tsk1: pointer to task_struct of some task.
3617 * @tsk2: pointer to task_struct of some other task.
3619 * Description: Return true if @tsk1's mems_allowed intersects the
3620 * mems_allowed of @tsk2. Used by the OOM killer to determine if
3621 * one of the task's memory usage might impact the memory available
3625 int cpuset_mems_allowed_intersects(const struct task_struct *tsk1,
3626 const struct task_struct *tsk2)
3628 return nodes_intersects(tsk1->mems_allowed, tsk2->mems_allowed);
3632 * cpuset_print_current_mems_allowed - prints current's cpuset and mems_allowed
3634 * Description: Prints current's name, cpuset name, and cached copy of its
3635 * mems_allowed to the kernel log.
3637 void cpuset_print_current_mems_allowed(void)
3639 struct cgroup *cgrp;
3643 cgrp = task_cs(current)->css.cgroup;
3644 pr_cont(",cpuset=");
3645 pr_cont_cgroup_name(cgrp);
3646 pr_cont(",mems_allowed=%*pbl",
3647 nodemask_pr_args(¤t->mems_allowed));
3653 * Collection of memory_pressure is suppressed unless
3654 * this flag is enabled by writing "1" to the special
3655 * cpuset file 'memory_pressure_enabled' in the root cpuset.
3658 int cpuset_memory_pressure_enabled __read_mostly;
3661 * cpuset_memory_pressure_bump - keep stats of per-cpuset reclaims.
3663 * Keep a running average of the rate of synchronous (direct)
3664 * page reclaim efforts initiated by tasks in each cpuset.
3666 * This represents the rate at which some task in the cpuset
3667 * ran low on memory on all nodes it was allowed to use, and
3668 * had to enter the kernels page reclaim code in an effort to
3669 * create more free memory by tossing clean pages or swapping
3670 * or writing dirty pages.
3672 * Display to user space in the per-cpuset read-only file
3673 * "memory_pressure". Value displayed is an integer
3674 * representing the recent rate of entry into the synchronous
3675 * (direct) page reclaim by any task attached to the cpuset.
3678 void __cpuset_memory_pressure_bump(void)
3681 fmeter_markevent(&task_cs(current)->fmeter);
3685 #ifdef CONFIG_PROC_PID_CPUSET
3687 * proc_cpuset_show()
3688 * - Print tasks cpuset path into seq_file.
3689 * - Used for /proc/<pid>/cpuset.
3690 * - No need to task_lock(tsk) on this tsk->cpuset reference, as it
3691 * doesn't really matter if tsk->cpuset changes after we read it,
3692 * and we take cpuset_rwsem, keeping cpuset_attach() from changing it
3695 int proc_cpuset_show(struct seq_file *m, struct pid_namespace *ns,
3696 struct pid *pid, struct task_struct *tsk)
3699 struct cgroup_subsys_state *css;
3703 buf = kmalloc(PATH_MAX, GFP_KERNEL);
3707 css = task_get_css(tsk, cpuset_cgrp_id);
3708 retval = cgroup_path_ns(css->cgroup, buf, PATH_MAX,
3709 current->nsproxy->cgroup_ns);
3711 if (retval >= PATH_MAX)
3712 retval = -ENAMETOOLONG;
3723 #endif /* CONFIG_PROC_PID_CPUSET */
3725 /* Display task mems_allowed in /proc/<pid>/status file. */
3726 void cpuset_task_status_allowed(struct seq_file *m, struct task_struct *task)
3728 seq_printf(m, "Mems_allowed:\t%*pb\n",
3729 nodemask_pr_args(&task->mems_allowed));
3730 seq_printf(m, "Mems_allowed_list:\t%*pbl\n",
3731 nodemask_pr_args(&task->mems_allowed));