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;
621 /* Each of our child cpusets must be a subset of us */
623 cpuset_for_each_child(c, css, cur)
624 if (!is_cpuset_subset(c, trial))
627 /* Remaining checks don't apply to root cpuset */
629 if (cur == &top_cpuset)
632 par = parent_cs(cur);
634 /* On legacy hierarchy, we must be a subset of our parent cpuset. */
636 if (!is_in_v2_mode() && !is_cpuset_subset(trial, par))
640 * If either I or some sibling (!= me) is exclusive, we can't
644 cpuset_for_each_child(c, css, par) {
645 if ((is_cpu_exclusive(trial) || is_cpu_exclusive(c)) &&
647 cpumask_intersects(trial->cpus_allowed, c->cpus_allowed))
649 if ((is_mem_exclusive(trial) || is_mem_exclusive(c)) &&
651 nodes_intersects(trial->mems_allowed, c->mems_allowed))
656 * Cpusets with tasks - existing or newly being attached - can't
657 * be changed to have empty cpus_allowed or mems_allowed.
660 if ((cgroup_is_populated(cur->css.cgroup) || cur->attach_in_progress)) {
661 if (!cpumask_empty(cur->cpus_allowed) &&
662 cpumask_empty(trial->cpus_allowed))
664 if (!nodes_empty(cur->mems_allowed) &&
665 nodes_empty(trial->mems_allowed))
670 * We can't shrink if we won't have enough room for SCHED_DEADLINE
674 if (is_cpu_exclusive(cur) &&
675 !cpuset_cpumask_can_shrink(cur->cpus_allowed,
676 trial->cpus_allowed))
687 * Helper routine for generate_sched_domains().
688 * Do cpusets a, b have overlapping effective cpus_allowed masks?
690 static int cpusets_overlap(struct cpuset *a, struct cpuset *b)
692 return cpumask_intersects(a->effective_cpus, b->effective_cpus);
696 update_domain_attr(struct sched_domain_attr *dattr, struct cpuset *c)
698 if (dattr->relax_domain_level < c->relax_domain_level)
699 dattr->relax_domain_level = c->relax_domain_level;
703 static void update_domain_attr_tree(struct sched_domain_attr *dattr,
704 struct cpuset *root_cs)
707 struct cgroup_subsys_state *pos_css;
710 cpuset_for_each_descendant_pre(cp, pos_css, root_cs) {
711 /* skip the whole subtree if @cp doesn't have any CPU */
712 if (cpumask_empty(cp->cpus_allowed)) {
713 pos_css = css_rightmost_descendant(pos_css);
717 if (is_sched_load_balance(cp))
718 update_domain_attr(dattr, cp);
723 /* Must be called with cpuset_rwsem held. */
724 static inline int nr_cpusets(void)
726 /* jump label reference count + the top-level cpuset */
727 return static_key_count(&cpusets_enabled_key.key) + 1;
731 * generate_sched_domains()
733 * This function builds a partial partition of the systems CPUs
734 * A 'partial partition' is a set of non-overlapping subsets whose
735 * union is a subset of that set.
736 * The output of this function needs to be passed to kernel/sched/core.c
737 * partition_sched_domains() routine, which will rebuild the scheduler's
738 * load balancing domains (sched domains) as specified by that partial
741 * See "What is sched_load_balance" in Documentation/admin-guide/cgroup-v1/cpusets.rst
742 * for a background explanation of this.
744 * Does not return errors, on the theory that the callers of this
745 * routine would rather not worry about failures to rebuild sched
746 * domains when operating in the severe memory shortage situations
747 * that could cause allocation failures below.
749 * Must be called with cpuset_rwsem held.
751 * The three key local variables below are:
752 * cp - cpuset pointer, used (together with pos_css) to perform a
753 * top-down scan of all cpusets. For our purposes, rebuilding
754 * the schedulers sched domains, we can ignore !is_sched_load_
756 * csa - (for CpuSet Array) Array of pointers to all the cpusets
757 * that need to be load balanced, for convenient iterative
758 * access by the subsequent code that finds the best partition,
759 * i.e the set of domains (subsets) of CPUs such that the
760 * cpus_allowed of every cpuset marked is_sched_load_balance
761 * is a subset of one of these domains, while there are as
762 * many such domains as possible, each as small as possible.
763 * doms - Conversion of 'csa' to an array of cpumasks, for passing to
764 * the kernel/sched/core.c routine partition_sched_domains() in a
765 * convenient format, that can be easily compared to the prior
766 * value to determine what partition elements (sched domains)
767 * were changed (added or removed.)
769 * Finding the best partition (set of domains):
770 * The triple nested loops below over i, j, k scan over the
771 * load balanced cpusets (using the array of cpuset pointers in
772 * csa[]) looking for pairs of cpusets that have overlapping
773 * cpus_allowed, but which don't have the same 'pn' partition
774 * number and gives them in the same partition number. It keeps
775 * looping on the 'restart' label until it can no longer find
778 * The union of the cpus_allowed masks from the set of
779 * all cpusets having the same 'pn' value then form the one
780 * element of the partition (one sched domain) to be passed to
781 * partition_sched_domains().
783 static int generate_sched_domains(cpumask_var_t **domains,
784 struct sched_domain_attr **attributes)
786 struct cpuset *cp; /* top-down scan of cpusets */
787 struct cpuset **csa; /* array of all cpuset ptrs */
788 int csn; /* how many cpuset ptrs in csa so far */
789 int i, j, k; /* indices for partition finding loops */
790 cpumask_var_t *doms; /* resulting partition; i.e. sched domains */
791 struct sched_domain_attr *dattr; /* attributes for custom domains */
792 int ndoms = 0; /* number of sched domains in result */
793 int nslot; /* next empty doms[] struct cpumask slot */
794 struct cgroup_subsys_state *pos_css;
795 bool root_load_balance = is_sched_load_balance(&top_cpuset);
801 /* Special case for the 99% of systems with one, full, sched domain */
802 if (root_load_balance && !top_cpuset.nr_subparts_cpus) {
804 doms = alloc_sched_domains(ndoms);
808 dattr = kmalloc(sizeof(struct sched_domain_attr), GFP_KERNEL);
810 *dattr = SD_ATTR_INIT;
811 update_domain_attr_tree(dattr, &top_cpuset);
813 cpumask_and(doms[0], top_cpuset.effective_cpus,
814 housekeeping_cpumask(HK_FLAG_DOMAIN));
819 csa = kmalloc_array(nr_cpusets(), sizeof(cp), GFP_KERNEL);
825 if (root_load_balance)
826 csa[csn++] = &top_cpuset;
827 cpuset_for_each_descendant_pre(cp, pos_css, &top_cpuset) {
828 if (cp == &top_cpuset)
831 * Continue traversing beyond @cp iff @cp has some CPUs and
832 * isn't load balancing. The former is obvious. The
833 * latter: All child cpusets contain a subset of the
834 * parent's cpus, so just skip them, and then we call
835 * update_domain_attr_tree() to calc relax_domain_level of
836 * the corresponding sched domain.
838 * If root is load-balancing, we can skip @cp if it
839 * is a subset of the root's effective_cpus.
841 if (!cpumask_empty(cp->cpus_allowed) &&
842 !(is_sched_load_balance(cp) &&
843 cpumask_intersects(cp->cpus_allowed,
844 housekeeping_cpumask(HK_FLAG_DOMAIN))))
847 if (root_load_balance &&
848 cpumask_subset(cp->cpus_allowed, top_cpuset.effective_cpus))
851 if (is_sched_load_balance(cp) &&
852 !cpumask_empty(cp->effective_cpus))
855 /* skip @cp's subtree if not a partition root */
856 if (!is_partition_root(cp))
857 pos_css = css_rightmost_descendant(pos_css);
861 for (i = 0; i < csn; i++)
866 /* Find the best partition (set of sched domains) */
867 for (i = 0; i < csn; i++) {
868 struct cpuset *a = csa[i];
871 for (j = 0; j < csn; j++) {
872 struct cpuset *b = csa[j];
875 if (apn != bpn && cpusets_overlap(a, b)) {
876 for (k = 0; k < csn; k++) {
877 struct cpuset *c = csa[k];
882 ndoms--; /* one less element */
889 * Now we know how many domains to create.
890 * Convert <csn, csa> to <ndoms, doms> and populate cpu masks.
892 doms = alloc_sched_domains(ndoms);
897 * The rest of the code, including the scheduler, can deal with
898 * dattr==NULL case. No need to abort if alloc fails.
900 dattr = kmalloc_array(ndoms, sizeof(struct sched_domain_attr),
903 for (nslot = 0, i = 0; i < csn; i++) {
904 struct cpuset *a = csa[i];
909 /* Skip completed partitions */
915 if (nslot == ndoms) {
916 static int warnings = 10;
918 pr_warn("rebuild_sched_domains confused: nslot %d, ndoms %d, csn %d, i %d, apn %d\n",
919 nslot, ndoms, csn, i, apn);
927 *(dattr + nslot) = SD_ATTR_INIT;
928 for (j = i; j < csn; j++) {
929 struct cpuset *b = csa[j];
932 cpumask_or(dp, dp, b->effective_cpus);
933 cpumask_and(dp, dp, housekeeping_cpumask(HK_FLAG_DOMAIN));
935 update_domain_attr_tree(dattr + nslot, b);
937 /* Done with this partition */
943 BUG_ON(nslot != ndoms);
949 * Fallback to the default domain if kmalloc() failed.
950 * See comments in partition_sched_domains().
960 static void update_tasks_root_domain(struct cpuset *cs)
962 struct css_task_iter it;
963 struct task_struct *task;
965 css_task_iter_start(&cs->css, 0, &it);
967 while ((task = css_task_iter_next(&it)))
968 dl_add_task_root_domain(task);
970 css_task_iter_end(&it);
973 static void rebuild_root_domains(void)
975 struct cpuset *cs = NULL;
976 struct cgroup_subsys_state *pos_css;
978 percpu_rwsem_assert_held(&cpuset_rwsem);
979 lockdep_assert_cpus_held();
980 lockdep_assert_held(&sched_domains_mutex);
985 * Clear default root domain DL accounting, it will be computed again
986 * if a task belongs to it.
988 dl_clear_root_domain(&def_root_domain);
990 cpuset_for_each_descendant_pre(cs, pos_css, &top_cpuset) {
992 if (cpumask_empty(cs->effective_cpus)) {
993 pos_css = css_rightmost_descendant(pos_css);
1001 update_tasks_root_domain(cs);
1010 partition_and_rebuild_sched_domains(int ndoms_new, cpumask_var_t doms_new[],
1011 struct sched_domain_attr *dattr_new)
1013 mutex_lock(&sched_domains_mutex);
1014 partition_sched_domains_locked(ndoms_new, doms_new, dattr_new);
1015 rebuild_root_domains();
1016 mutex_unlock(&sched_domains_mutex);
1020 * Rebuild scheduler domains.
1022 * If the flag 'sched_load_balance' of any cpuset with non-empty
1023 * 'cpus' changes, or if the 'cpus' allowed changes in any cpuset
1024 * which has that flag enabled, or if any cpuset with a non-empty
1025 * 'cpus' is removed, then call this routine to rebuild the
1026 * scheduler's dynamic sched domains.
1028 * Call with cpuset_rwsem held. Takes cpus_read_lock().
1030 static void rebuild_sched_domains_locked(void)
1032 struct cgroup_subsys_state *pos_css;
1033 struct sched_domain_attr *attr;
1034 cpumask_var_t *doms;
1038 lockdep_assert_cpus_held();
1039 percpu_rwsem_assert_held(&cpuset_rwsem);
1042 * If we have raced with CPU hotplug, return early to avoid
1043 * passing doms with offlined cpu to partition_sched_domains().
1044 * Anyways, cpuset_hotplug_workfn() will rebuild sched domains.
1046 * With no CPUs in any subpartitions, top_cpuset's effective CPUs
1047 * should be the same as the active CPUs, so checking only top_cpuset
1048 * is enough to detect racing CPU offlines.
1050 if (!top_cpuset.nr_subparts_cpus &&
1051 !cpumask_equal(top_cpuset.effective_cpus, cpu_active_mask))
1055 * With subpartition CPUs, however, the effective CPUs of a partition
1056 * root should be only a subset of the active CPUs. Since a CPU in any
1057 * partition root could be offlined, all must be checked.
1059 if (top_cpuset.nr_subparts_cpus) {
1061 cpuset_for_each_descendant_pre(cs, pos_css, &top_cpuset) {
1062 if (!is_partition_root(cs)) {
1063 pos_css = css_rightmost_descendant(pos_css);
1066 if (!cpumask_subset(cs->effective_cpus,
1075 /* Generate domain masks and attrs */
1076 ndoms = generate_sched_domains(&doms, &attr);
1078 /* Have scheduler rebuild the domains */
1079 partition_and_rebuild_sched_domains(ndoms, doms, attr);
1081 #else /* !CONFIG_SMP */
1082 static void rebuild_sched_domains_locked(void)
1085 #endif /* CONFIG_SMP */
1087 void rebuild_sched_domains(void)
1090 percpu_down_write(&cpuset_rwsem);
1091 rebuild_sched_domains_locked();
1092 percpu_up_write(&cpuset_rwsem);
1097 * update_tasks_cpumask - Update the cpumasks of tasks in the cpuset.
1098 * @cs: the cpuset in which each task's cpus_allowed mask needs to be changed
1100 * Iterate through each task of @cs updating its cpus_allowed to the
1101 * effective cpuset's. As this function is called with cpuset_rwsem held,
1102 * cpuset membership stays stable.
1104 static void update_tasks_cpumask(struct cpuset *cs)
1106 struct css_task_iter it;
1107 struct task_struct *task;
1109 css_task_iter_start(&cs->css, 0, &it);
1110 while ((task = css_task_iter_next(&it)))
1111 set_cpus_allowed_ptr(task, cs->effective_cpus);
1112 css_task_iter_end(&it);
1116 * compute_effective_cpumask - Compute the effective cpumask of the cpuset
1117 * @new_cpus: the temp variable for the new effective_cpus mask
1118 * @cs: the cpuset the need to recompute the new effective_cpus mask
1119 * @parent: the parent cpuset
1121 * If the parent has subpartition CPUs, include them in the list of
1122 * allowable CPUs in computing the new effective_cpus mask. Since offlined
1123 * CPUs are not removed from subparts_cpus, we have to use cpu_active_mask
1124 * to mask those out.
1126 static void compute_effective_cpumask(struct cpumask *new_cpus,
1127 struct cpuset *cs, struct cpuset *parent)
1129 if (parent->nr_subparts_cpus) {
1130 cpumask_or(new_cpus, parent->effective_cpus,
1131 parent->subparts_cpus);
1132 cpumask_and(new_cpus, new_cpus, cs->cpus_allowed);
1133 cpumask_and(new_cpus, new_cpus, cpu_active_mask);
1135 cpumask_and(new_cpus, cs->cpus_allowed, parent->effective_cpus);
1140 * Commands for update_parent_subparts_cpumask
1143 partcmd_enable, /* Enable partition root */
1144 partcmd_disable, /* Disable partition root */
1145 partcmd_update, /* Update parent's subparts_cpus */
1149 * update_parent_subparts_cpumask - update subparts_cpus mask of parent cpuset
1150 * @cpuset: The cpuset that requests change in partition root state
1151 * @cmd: Partition root state change command
1152 * @newmask: Optional new cpumask for partcmd_update
1153 * @tmp: Temporary addmask and delmask
1154 * Return: 0, 1 or an error code
1156 * For partcmd_enable, the cpuset is being transformed from a non-partition
1157 * root to a partition root. The cpus_allowed mask of the given cpuset will
1158 * be put into parent's subparts_cpus and taken away from parent's
1159 * effective_cpus. The function will return 0 if all the CPUs listed in
1160 * cpus_allowed can be granted or an error code will be returned.
1162 * For partcmd_disable, the cpuset is being transofrmed from a partition
1163 * root back to a non-partition root. Any CPUs in cpus_allowed that are in
1164 * parent's subparts_cpus will be taken away from that cpumask and put back
1165 * into parent's effective_cpus. 0 should always be returned.
1167 * For partcmd_update, if the optional newmask is specified, the cpu
1168 * list is to be changed from cpus_allowed to newmask. Otherwise,
1169 * cpus_allowed is assumed to remain the same. The cpuset should either
1170 * be a partition root or an invalid partition root. The partition root
1171 * state may change if newmask is NULL and none of the requested CPUs can
1172 * be granted by the parent. The function will return 1 if changes to
1173 * parent's subparts_cpus and effective_cpus happen or 0 otherwise.
1174 * Error code should only be returned when newmask is non-NULL.
1176 * The partcmd_enable and partcmd_disable commands are used by
1177 * update_prstate(). The partcmd_update command is used by
1178 * update_cpumasks_hier() with newmask NULL and update_cpumask() with
1181 * The checking is more strict when enabling partition root than the
1182 * other two commands.
1184 * Because of the implicit cpu exclusive nature of a partition root,
1185 * cpumask changes that violates the cpu exclusivity rule will not be
1186 * permitted when checked by validate_change(). The validate_change()
1187 * function will also prevent any changes to the cpu list if it is not
1188 * a superset of children's cpu lists.
1190 static int update_parent_subparts_cpumask(struct cpuset *cpuset, int cmd,
1191 struct cpumask *newmask,
1192 struct tmpmasks *tmp)
1194 struct cpuset *parent = parent_cs(cpuset);
1195 int adding; /* Moving cpus from effective_cpus to subparts_cpus */
1196 int deleting; /* Moving cpus from subparts_cpus to effective_cpus */
1197 int old_prs, new_prs;
1198 bool part_error = false; /* Partition error? */
1200 percpu_rwsem_assert_held(&cpuset_rwsem);
1203 * The parent must be a partition root.
1204 * The new cpumask, if present, or the current cpus_allowed must
1207 if (!is_partition_root(parent) ||
1208 (newmask && cpumask_empty(newmask)) ||
1209 (!newmask && cpumask_empty(cpuset->cpus_allowed)))
1213 * Enabling/disabling partition root is not allowed if there are
1216 if ((cmd != partcmd_update) && css_has_online_children(&cpuset->css))
1220 * Enabling partition root is not allowed if not all the CPUs
1221 * can be granted from parent's effective_cpus or at least one
1222 * CPU will be left after that.
1224 if ((cmd == partcmd_enable) &&
1225 (!cpumask_subset(cpuset->cpus_allowed, parent->effective_cpus) ||
1226 cpumask_equal(cpuset->cpus_allowed, parent->effective_cpus)))
1230 * A cpumask update cannot make parent's effective_cpus become empty.
1232 adding = deleting = false;
1233 old_prs = new_prs = cpuset->partition_root_state;
1234 if (cmd == partcmd_enable) {
1235 cpumask_copy(tmp->addmask, cpuset->cpus_allowed);
1237 } else if (cmd == partcmd_disable) {
1238 deleting = cpumask_and(tmp->delmask, cpuset->cpus_allowed,
1239 parent->subparts_cpus);
1240 } else if (newmask) {
1242 * partcmd_update with newmask:
1244 * delmask = cpus_allowed & ~newmask & parent->subparts_cpus
1245 * addmask = newmask & parent->effective_cpus
1246 * & ~parent->subparts_cpus
1248 cpumask_andnot(tmp->delmask, cpuset->cpus_allowed, newmask);
1249 deleting = cpumask_and(tmp->delmask, tmp->delmask,
1250 parent->subparts_cpus);
1252 cpumask_and(tmp->addmask, newmask, parent->effective_cpus);
1253 adding = cpumask_andnot(tmp->addmask, tmp->addmask,
1254 parent->subparts_cpus);
1256 * Return error if the new effective_cpus could become empty.
1259 cpumask_equal(parent->effective_cpus, tmp->addmask)) {
1263 * As some of the CPUs in subparts_cpus might have
1264 * been offlined, we need to compute the real delmask
1267 if (!cpumask_and(tmp->addmask, tmp->delmask,
1270 cpumask_copy(tmp->addmask, parent->effective_cpus);
1274 * partcmd_update w/o newmask:
1276 * addmask = cpus_allowed & parent->effective_cpus
1278 * Note that parent's subparts_cpus may have been
1279 * pre-shrunk in case there is a change in the cpu list.
1280 * So no deletion is needed.
1282 adding = cpumask_and(tmp->addmask, cpuset->cpus_allowed,
1283 parent->effective_cpus);
1284 part_error = cpumask_equal(tmp->addmask,
1285 parent->effective_cpus);
1288 if (cmd == partcmd_update) {
1289 int prev_prs = cpuset->partition_root_state;
1292 * Check for possible transition between PRS_ENABLED
1295 switch (cpuset->partition_root_state) {
1298 new_prs = PRS_ERROR;
1302 new_prs = PRS_ENABLED;
1306 * Set part_error if previously in invalid state.
1308 part_error = (prev_prs == PRS_ERROR);
1311 if (!part_error && (new_prs == PRS_ERROR))
1312 return 0; /* Nothing need to be done */
1314 if (new_prs == PRS_ERROR) {
1316 * Remove all its cpus from parent's subparts_cpus.
1319 deleting = cpumask_and(tmp->delmask, cpuset->cpus_allowed,
1320 parent->subparts_cpus);
1323 if (!adding && !deleting && (new_prs == old_prs))
1327 * Change the parent's subparts_cpus.
1328 * Newly added CPUs will be removed from effective_cpus and
1329 * newly deleted ones will be added back to effective_cpus.
1331 spin_lock_irq(&callback_lock);
1333 cpumask_or(parent->subparts_cpus,
1334 parent->subparts_cpus, tmp->addmask);
1335 cpumask_andnot(parent->effective_cpus,
1336 parent->effective_cpus, tmp->addmask);
1339 cpumask_andnot(parent->subparts_cpus,
1340 parent->subparts_cpus, tmp->delmask);
1342 * Some of the CPUs in subparts_cpus might have been offlined.
1344 cpumask_and(tmp->delmask, tmp->delmask, cpu_active_mask);
1345 cpumask_or(parent->effective_cpus,
1346 parent->effective_cpus, tmp->delmask);
1349 parent->nr_subparts_cpus = cpumask_weight(parent->subparts_cpus);
1351 if (old_prs != new_prs)
1352 cpuset->partition_root_state = new_prs;
1354 spin_unlock_irq(&callback_lock);
1355 notify_partition_change(cpuset, old_prs, new_prs);
1357 return cmd == partcmd_update;
1361 * update_cpumasks_hier - Update effective cpumasks and tasks in the subtree
1362 * @cs: the cpuset to consider
1363 * @tmp: temp variables for calculating effective_cpus & partition setup
1365 * When configured cpumask is changed, the effective cpumasks of this cpuset
1366 * and all its descendants need to be updated.
1368 * On legacy hierarchy, effective_cpus will be the same with cpu_allowed.
1370 * Called with cpuset_rwsem held
1372 static void update_cpumasks_hier(struct cpuset *cs, struct tmpmasks *tmp)
1375 struct cgroup_subsys_state *pos_css;
1376 bool need_rebuild_sched_domains = false;
1377 int old_prs, new_prs;
1380 cpuset_for_each_descendant_pre(cp, pos_css, cs) {
1381 struct cpuset *parent = parent_cs(cp);
1383 compute_effective_cpumask(tmp->new_cpus, cp, parent);
1386 * If it becomes empty, inherit the effective mask of the
1387 * parent, which is guaranteed to have some CPUs.
1389 if (is_in_v2_mode() && cpumask_empty(tmp->new_cpus)) {
1390 cpumask_copy(tmp->new_cpus, parent->effective_cpus);
1391 if (!cp->use_parent_ecpus) {
1392 cp->use_parent_ecpus = true;
1393 parent->child_ecpus_count++;
1395 } else if (cp->use_parent_ecpus) {
1396 cp->use_parent_ecpus = false;
1397 WARN_ON_ONCE(!parent->child_ecpus_count);
1398 parent->child_ecpus_count--;
1402 * Skip the whole subtree if the cpumask remains the same
1403 * and has no partition root state.
1405 if (!cp->partition_root_state &&
1406 cpumask_equal(tmp->new_cpus, cp->effective_cpus)) {
1407 pos_css = css_rightmost_descendant(pos_css);
1412 * update_parent_subparts_cpumask() should have been called
1413 * for cs already in update_cpumask(). We should also call
1414 * update_tasks_cpumask() again for tasks in the parent
1415 * cpuset if the parent's subparts_cpus changes.
1417 old_prs = new_prs = cp->partition_root_state;
1418 if ((cp != cs) && old_prs) {
1419 switch (parent->partition_root_state) {
1422 * If parent is not a partition root or an
1423 * invalid partition root, clear its state
1424 * and its CS_CPU_EXCLUSIVE flag.
1426 WARN_ON_ONCE(cp->partition_root_state
1428 new_prs = PRS_DISABLED;
1431 * clear_bit() is an atomic operation and
1432 * readers aren't interested in the state
1433 * of CS_CPU_EXCLUSIVE anyway. So we can
1434 * just update the flag without holding
1435 * the callback_lock.
1437 clear_bit(CS_CPU_EXCLUSIVE, &cp->flags);
1441 if (update_parent_subparts_cpumask(cp, partcmd_update, NULL, tmp))
1442 update_tasks_cpumask(parent);
1447 * When parent is invalid, it has to be too.
1449 new_prs = PRS_ERROR;
1454 if (!css_tryget_online(&cp->css))
1458 spin_lock_irq(&callback_lock);
1460 cpumask_copy(cp->effective_cpus, tmp->new_cpus);
1461 if (cp->nr_subparts_cpus && (new_prs != PRS_ENABLED)) {
1462 cp->nr_subparts_cpus = 0;
1463 cpumask_clear(cp->subparts_cpus);
1464 } else if (cp->nr_subparts_cpus) {
1466 * Make sure that effective_cpus & subparts_cpus
1467 * are mutually exclusive.
1469 * In the unlikely event that effective_cpus
1470 * becomes empty. we clear cp->nr_subparts_cpus and
1471 * let its child partition roots to compete for
1474 cpumask_andnot(cp->effective_cpus, cp->effective_cpus,
1476 if (cpumask_empty(cp->effective_cpus)) {
1477 cpumask_copy(cp->effective_cpus, tmp->new_cpus);
1478 cpumask_clear(cp->subparts_cpus);
1479 cp->nr_subparts_cpus = 0;
1480 } else if (!cpumask_subset(cp->subparts_cpus,
1482 cpumask_andnot(cp->subparts_cpus,
1483 cp->subparts_cpus, tmp->new_cpus);
1484 cp->nr_subparts_cpus
1485 = cpumask_weight(cp->subparts_cpus);
1489 if (new_prs != old_prs)
1490 cp->partition_root_state = new_prs;
1492 spin_unlock_irq(&callback_lock);
1493 notify_partition_change(cp, old_prs, new_prs);
1495 WARN_ON(!is_in_v2_mode() &&
1496 !cpumask_equal(cp->cpus_allowed, cp->effective_cpus));
1498 update_tasks_cpumask(cp);
1501 * On legacy hierarchy, if the effective cpumask of any non-
1502 * empty cpuset is changed, we need to rebuild sched domains.
1503 * On default hierarchy, the cpuset needs to be a partition
1506 if (!cpumask_empty(cp->cpus_allowed) &&
1507 is_sched_load_balance(cp) &&
1508 (!cgroup_subsys_on_dfl(cpuset_cgrp_subsys) ||
1509 is_partition_root(cp)))
1510 need_rebuild_sched_domains = true;
1517 if (need_rebuild_sched_domains)
1518 rebuild_sched_domains_locked();
1522 * update_sibling_cpumasks - Update siblings cpumasks
1523 * @parent: Parent cpuset
1524 * @cs: Current cpuset
1525 * @tmp: Temp variables
1527 static void update_sibling_cpumasks(struct cpuset *parent, struct cpuset *cs,
1528 struct tmpmasks *tmp)
1530 struct cpuset *sibling;
1531 struct cgroup_subsys_state *pos_css;
1534 * Check all its siblings and call update_cpumasks_hier()
1535 * if their use_parent_ecpus flag is set in order for them
1536 * to use the right effective_cpus value.
1539 cpuset_for_each_child(sibling, pos_css, parent) {
1542 if (!sibling->use_parent_ecpus)
1545 update_cpumasks_hier(sibling, tmp);
1551 * update_cpumask - update the cpus_allowed mask of a cpuset and all tasks in it
1552 * @cs: the cpuset to consider
1553 * @trialcs: trial cpuset
1554 * @buf: buffer of cpu numbers written to this cpuset
1556 static int update_cpumask(struct cpuset *cs, struct cpuset *trialcs,
1560 struct tmpmasks tmp;
1562 /* top_cpuset.cpus_allowed tracks cpu_online_mask; it's read-only */
1563 if (cs == &top_cpuset)
1567 * An empty cpus_allowed is ok only if the cpuset has no tasks.
1568 * Since cpulist_parse() fails on an empty mask, we special case
1569 * that parsing. The validate_change() call ensures that cpusets
1570 * with tasks have cpus.
1573 cpumask_clear(trialcs->cpus_allowed);
1575 retval = cpulist_parse(buf, trialcs->cpus_allowed);
1579 if (!cpumask_subset(trialcs->cpus_allowed,
1580 top_cpuset.cpus_allowed))
1584 /* Nothing to do if the cpus didn't change */
1585 if (cpumask_equal(cs->cpus_allowed, trialcs->cpus_allowed))
1588 retval = validate_change(cs, trialcs);
1592 #ifdef CONFIG_CPUMASK_OFFSTACK
1594 * Use the cpumasks in trialcs for tmpmasks when they are pointers
1595 * to allocated cpumasks.
1597 tmp.addmask = trialcs->subparts_cpus;
1598 tmp.delmask = trialcs->effective_cpus;
1599 tmp.new_cpus = trialcs->cpus_allowed;
1602 if (cs->partition_root_state) {
1603 /* Cpumask of a partition root cannot be empty */
1604 if (cpumask_empty(trialcs->cpus_allowed))
1606 if (update_parent_subparts_cpumask(cs, partcmd_update,
1607 trialcs->cpus_allowed, &tmp) < 0)
1611 spin_lock_irq(&callback_lock);
1612 cpumask_copy(cs->cpus_allowed, trialcs->cpus_allowed);
1615 * Make sure that subparts_cpus is a subset of cpus_allowed.
1617 if (cs->nr_subparts_cpus) {
1618 cpumask_andnot(cs->subparts_cpus, cs->subparts_cpus,
1620 cs->nr_subparts_cpus = cpumask_weight(cs->subparts_cpus);
1622 spin_unlock_irq(&callback_lock);
1624 update_cpumasks_hier(cs, &tmp);
1626 if (cs->partition_root_state) {
1627 struct cpuset *parent = parent_cs(cs);
1630 * For partition root, update the cpumasks of sibling
1631 * cpusets if they use parent's effective_cpus.
1633 if (parent->child_ecpus_count)
1634 update_sibling_cpumasks(parent, cs, &tmp);
1640 * Migrate memory region from one set of nodes to another. This is
1641 * performed asynchronously as it can be called from process migration path
1642 * holding locks involved in process management. All mm migrations are
1643 * performed in the queued order and can be waited for by flushing
1644 * cpuset_migrate_mm_wq.
1647 struct cpuset_migrate_mm_work {
1648 struct work_struct work;
1649 struct mm_struct *mm;
1654 static void cpuset_migrate_mm_workfn(struct work_struct *work)
1656 struct cpuset_migrate_mm_work *mwork =
1657 container_of(work, struct cpuset_migrate_mm_work, work);
1659 /* on a wq worker, no need to worry about %current's mems_allowed */
1660 do_migrate_pages(mwork->mm, &mwork->from, &mwork->to, MPOL_MF_MOVE_ALL);
1665 static void cpuset_migrate_mm(struct mm_struct *mm, const nodemask_t *from,
1666 const nodemask_t *to)
1668 struct cpuset_migrate_mm_work *mwork;
1670 if (nodes_equal(*from, *to)) {
1675 mwork = kzalloc(sizeof(*mwork), GFP_KERNEL);
1678 mwork->from = *from;
1680 INIT_WORK(&mwork->work, cpuset_migrate_mm_workfn);
1681 queue_work(cpuset_migrate_mm_wq, &mwork->work);
1687 static void cpuset_post_attach(void)
1689 flush_workqueue(cpuset_migrate_mm_wq);
1693 * cpuset_change_task_nodemask - change task's mems_allowed and mempolicy
1694 * @tsk: the task to change
1695 * @newmems: new nodes that the task will be set
1697 * We use the mems_allowed_seq seqlock to safely update both tsk->mems_allowed
1698 * and rebind an eventual tasks' mempolicy. If the task is allocating in
1699 * parallel, it might temporarily see an empty intersection, which results in
1700 * a seqlock check and retry before OOM or allocation failure.
1702 static void cpuset_change_task_nodemask(struct task_struct *tsk,
1703 nodemask_t *newmems)
1707 local_irq_disable();
1708 write_seqcount_begin(&tsk->mems_allowed_seq);
1710 nodes_or(tsk->mems_allowed, tsk->mems_allowed, *newmems);
1711 mpol_rebind_task(tsk, newmems);
1712 tsk->mems_allowed = *newmems;
1714 write_seqcount_end(&tsk->mems_allowed_seq);
1720 static void *cpuset_being_rebound;
1723 * update_tasks_nodemask - Update the nodemasks of tasks in the cpuset.
1724 * @cs: the cpuset in which each task's mems_allowed mask needs to be changed
1726 * Iterate through each task of @cs updating its mems_allowed to the
1727 * effective cpuset's. As this function is called with cpuset_rwsem held,
1728 * cpuset membership stays stable.
1730 static void update_tasks_nodemask(struct cpuset *cs)
1732 static nodemask_t newmems; /* protected by cpuset_rwsem */
1733 struct css_task_iter it;
1734 struct task_struct *task;
1736 cpuset_being_rebound = cs; /* causes mpol_dup() rebind */
1738 guarantee_online_mems(cs, &newmems);
1741 * The mpol_rebind_mm() call takes mmap_lock, which we couldn't
1742 * take while holding tasklist_lock. Forks can happen - the
1743 * mpol_dup() cpuset_being_rebound check will catch such forks,
1744 * and rebind their vma mempolicies too. Because we still hold
1745 * the global cpuset_rwsem, we know that no other rebind effort
1746 * will be contending for the global variable cpuset_being_rebound.
1747 * It's ok if we rebind the same mm twice; mpol_rebind_mm()
1748 * is idempotent. Also migrate pages in each mm to new nodes.
1750 css_task_iter_start(&cs->css, 0, &it);
1751 while ((task = css_task_iter_next(&it))) {
1752 struct mm_struct *mm;
1755 cpuset_change_task_nodemask(task, &newmems);
1757 mm = get_task_mm(task);
1761 migrate = is_memory_migrate(cs);
1763 mpol_rebind_mm(mm, &cs->mems_allowed);
1765 cpuset_migrate_mm(mm, &cs->old_mems_allowed, &newmems);
1769 css_task_iter_end(&it);
1772 * All the tasks' nodemasks have been updated, update
1773 * cs->old_mems_allowed.
1775 cs->old_mems_allowed = newmems;
1777 /* We're done rebinding vmas to this cpuset's new mems_allowed. */
1778 cpuset_being_rebound = NULL;
1782 * update_nodemasks_hier - Update effective nodemasks and tasks in the subtree
1783 * @cs: the cpuset to consider
1784 * @new_mems: a temp variable for calculating new effective_mems
1786 * When configured nodemask is changed, the effective nodemasks of this cpuset
1787 * and all its descendants need to be updated.
1789 * On legacy hierarchy, effective_mems will be the same with mems_allowed.
1791 * Called with cpuset_rwsem held
1793 static void update_nodemasks_hier(struct cpuset *cs, nodemask_t *new_mems)
1796 struct cgroup_subsys_state *pos_css;
1799 cpuset_for_each_descendant_pre(cp, pos_css, cs) {
1800 struct cpuset *parent = parent_cs(cp);
1802 nodes_and(*new_mems, cp->mems_allowed, parent->effective_mems);
1805 * If it becomes empty, inherit the effective mask of the
1806 * parent, which is guaranteed to have some MEMs.
1808 if (is_in_v2_mode() && nodes_empty(*new_mems))
1809 *new_mems = parent->effective_mems;
1811 /* Skip the whole subtree if the nodemask remains the same. */
1812 if (nodes_equal(*new_mems, cp->effective_mems)) {
1813 pos_css = css_rightmost_descendant(pos_css);
1817 if (!css_tryget_online(&cp->css))
1821 spin_lock_irq(&callback_lock);
1822 cp->effective_mems = *new_mems;
1823 spin_unlock_irq(&callback_lock);
1825 WARN_ON(!is_in_v2_mode() &&
1826 !nodes_equal(cp->mems_allowed, cp->effective_mems));
1828 update_tasks_nodemask(cp);
1837 * Handle user request to change the 'mems' memory placement
1838 * of a cpuset. Needs to validate the request, update the
1839 * cpusets mems_allowed, and for each task in the cpuset,
1840 * update mems_allowed and rebind task's mempolicy and any vma
1841 * mempolicies and if the cpuset is marked 'memory_migrate',
1842 * migrate the tasks pages to the new memory.
1844 * Call with cpuset_rwsem held. May take callback_lock during call.
1845 * Will take tasklist_lock, scan tasklist for tasks in cpuset cs,
1846 * lock each such tasks mm->mmap_lock, scan its vma's and rebind
1847 * their mempolicies to the cpusets new mems_allowed.
1849 static int update_nodemask(struct cpuset *cs, struct cpuset *trialcs,
1855 * top_cpuset.mems_allowed tracks node_stats[N_MEMORY];
1858 if (cs == &top_cpuset) {
1864 * An empty mems_allowed is ok iff there are no tasks in the cpuset.
1865 * Since nodelist_parse() fails on an empty mask, we special case
1866 * that parsing. The validate_change() call ensures that cpusets
1867 * with tasks have memory.
1870 nodes_clear(trialcs->mems_allowed);
1872 retval = nodelist_parse(buf, trialcs->mems_allowed);
1876 if (!nodes_subset(trialcs->mems_allowed,
1877 top_cpuset.mems_allowed)) {
1883 if (nodes_equal(cs->mems_allowed, trialcs->mems_allowed)) {
1884 retval = 0; /* Too easy - nothing to do */
1887 retval = validate_change(cs, trialcs);
1891 check_insane_mems_config(&trialcs->mems_allowed);
1893 spin_lock_irq(&callback_lock);
1894 cs->mems_allowed = trialcs->mems_allowed;
1895 spin_unlock_irq(&callback_lock);
1897 /* use trialcs->mems_allowed as a temp variable */
1898 update_nodemasks_hier(cs, &trialcs->mems_allowed);
1903 bool current_cpuset_is_being_rebound(void)
1908 ret = task_cs(current) == cpuset_being_rebound;
1914 static int update_relax_domain_level(struct cpuset *cs, s64 val)
1917 if (val < -1 || val >= sched_domain_level_max)
1921 if (val != cs->relax_domain_level) {
1922 cs->relax_domain_level = val;
1923 if (!cpumask_empty(cs->cpus_allowed) &&
1924 is_sched_load_balance(cs))
1925 rebuild_sched_domains_locked();
1932 * update_tasks_flags - update the spread flags of tasks in the cpuset.
1933 * @cs: the cpuset in which each task's spread flags needs to be changed
1935 * Iterate through each task of @cs updating its spread flags. As this
1936 * function is called with cpuset_rwsem held, cpuset membership stays
1939 static void update_tasks_flags(struct cpuset *cs)
1941 struct css_task_iter it;
1942 struct task_struct *task;
1944 css_task_iter_start(&cs->css, 0, &it);
1945 while ((task = css_task_iter_next(&it)))
1946 cpuset_update_task_spread_flag(cs, task);
1947 css_task_iter_end(&it);
1951 * update_flag - read a 0 or a 1 in a file and update associated flag
1952 * bit: the bit to update (see cpuset_flagbits_t)
1953 * cs: the cpuset to update
1954 * turning_on: whether the flag is being set or cleared
1956 * Call with cpuset_rwsem held.
1959 static int update_flag(cpuset_flagbits_t bit, struct cpuset *cs,
1962 struct cpuset *trialcs;
1963 int balance_flag_changed;
1964 int spread_flag_changed;
1967 trialcs = alloc_trial_cpuset(cs);
1972 set_bit(bit, &trialcs->flags);
1974 clear_bit(bit, &trialcs->flags);
1976 err = validate_change(cs, trialcs);
1980 balance_flag_changed = (is_sched_load_balance(cs) !=
1981 is_sched_load_balance(trialcs));
1983 spread_flag_changed = ((is_spread_slab(cs) != is_spread_slab(trialcs))
1984 || (is_spread_page(cs) != is_spread_page(trialcs)));
1986 spin_lock_irq(&callback_lock);
1987 cs->flags = trialcs->flags;
1988 spin_unlock_irq(&callback_lock);
1990 if (!cpumask_empty(trialcs->cpus_allowed) && balance_flag_changed)
1991 rebuild_sched_domains_locked();
1993 if (spread_flag_changed)
1994 update_tasks_flags(cs);
1996 free_cpuset(trialcs);
2001 * update_prstate - update partititon_root_state
2002 * cs: the cpuset to update
2003 * new_prs: new partition root state
2005 * Call with cpuset_rwsem held.
2007 static int update_prstate(struct cpuset *cs, int new_prs)
2009 int err, old_prs = cs->partition_root_state;
2010 struct cpuset *parent = parent_cs(cs);
2011 struct tmpmasks tmpmask;
2013 if (old_prs == new_prs)
2017 * Cannot force a partial or invalid partition root to a full
2020 if (new_prs && (old_prs == PRS_ERROR))
2023 if (alloc_cpumasks(NULL, &tmpmask))
2029 * Turning on partition root requires setting the
2030 * CS_CPU_EXCLUSIVE bit implicitly as well and cpus_allowed
2033 if (cpumask_empty(cs->cpus_allowed))
2036 err = update_flag(CS_CPU_EXCLUSIVE, cs, 1);
2040 err = update_parent_subparts_cpumask(cs, partcmd_enable,
2043 update_flag(CS_CPU_EXCLUSIVE, cs, 0);
2048 * Turning off partition root will clear the
2049 * CS_CPU_EXCLUSIVE bit.
2051 if (old_prs == PRS_ERROR) {
2052 update_flag(CS_CPU_EXCLUSIVE, cs, 0);
2057 err = update_parent_subparts_cpumask(cs, partcmd_disable,
2062 /* Turning off CS_CPU_EXCLUSIVE will not return error */
2063 update_flag(CS_CPU_EXCLUSIVE, cs, 0);
2067 * Update cpumask of parent's tasks except when it is the top
2068 * cpuset as some system daemons cannot be mapped to other CPUs.
2070 if (parent != &top_cpuset)
2071 update_tasks_cpumask(parent);
2073 if (parent->child_ecpus_count)
2074 update_sibling_cpumasks(parent, cs, &tmpmask);
2076 rebuild_sched_domains_locked();
2079 spin_lock_irq(&callback_lock);
2080 cs->partition_root_state = new_prs;
2081 spin_unlock_irq(&callback_lock);
2082 notify_partition_change(cs, old_prs, new_prs);
2085 free_cpumasks(NULL, &tmpmask);
2090 * Frequency meter - How fast is some event occurring?
2092 * These routines manage a digitally filtered, constant time based,
2093 * event frequency meter. There are four routines:
2094 * fmeter_init() - initialize a frequency meter.
2095 * fmeter_markevent() - called each time the event happens.
2096 * fmeter_getrate() - returns the recent rate of such events.
2097 * fmeter_update() - internal routine used to update fmeter.
2099 * A common data structure is passed to each of these routines,
2100 * which is used to keep track of the state required to manage the
2101 * frequency meter and its digital filter.
2103 * The filter works on the number of events marked per unit time.
2104 * The filter is single-pole low-pass recursive (IIR). The time unit
2105 * is 1 second. Arithmetic is done using 32-bit integers scaled to
2106 * simulate 3 decimal digits of precision (multiplied by 1000).
2108 * With an FM_COEF of 933, and a time base of 1 second, the filter
2109 * has a half-life of 10 seconds, meaning that if the events quit
2110 * happening, then the rate returned from the fmeter_getrate()
2111 * will be cut in half each 10 seconds, until it converges to zero.
2113 * It is not worth doing a real infinitely recursive filter. If more
2114 * than FM_MAXTICKS ticks have elapsed since the last filter event,
2115 * just compute FM_MAXTICKS ticks worth, by which point the level
2118 * Limit the count of unprocessed events to FM_MAXCNT, so as to avoid
2119 * arithmetic overflow in the fmeter_update() routine.
2121 * Given the simple 32 bit integer arithmetic used, this meter works
2122 * best for reporting rates between one per millisecond (msec) and
2123 * one per 32 (approx) seconds. At constant rates faster than one
2124 * per msec it maxes out at values just under 1,000,000. At constant
2125 * rates between one per msec, and one per second it will stabilize
2126 * to a value N*1000, where N is the rate of events per second.
2127 * At constant rates between one per second and one per 32 seconds,
2128 * it will be choppy, moving up on the seconds that have an event,
2129 * and then decaying until the next event. At rates slower than
2130 * about one in 32 seconds, it decays all the way back to zero between
2134 #define FM_COEF 933 /* coefficient for half-life of 10 secs */
2135 #define FM_MAXTICKS ((u32)99) /* useless computing more ticks than this */
2136 #define FM_MAXCNT 1000000 /* limit cnt to avoid overflow */
2137 #define FM_SCALE 1000 /* faux fixed point scale */
2139 /* Initialize a frequency meter */
2140 static void fmeter_init(struct fmeter *fmp)
2145 spin_lock_init(&fmp->lock);
2148 /* Internal meter update - process cnt events and update value */
2149 static void fmeter_update(struct fmeter *fmp)
2154 now = ktime_get_seconds();
2155 ticks = now - fmp->time;
2160 ticks = min(FM_MAXTICKS, ticks);
2162 fmp->val = (FM_COEF * fmp->val) / FM_SCALE;
2165 fmp->val += ((FM_SCALE - FM_COEF) * fmp->cnt) / FM_SCALE;
2169 /* Process any previous ticks, then bump cnt by one (times scale). */
2170 static void fmeter_markevent(struct fmeter *fmp)
2172 spin_lock(&fmp->lock);
2174 fmp->cnt = min(FM_MAXCNT, fmp->cnt + FM_SCALE);
2175 spin_unlock(&fmp->lock);
2178 /* Process any previous ticks, then return current value. */
2179 static int fmeter_getrate(struct fmeter *fmp)
2183 spin_lock(&fmp->lock);
2186 spin_unlock(&fmp->lock);
2190 static struct cpuset *cpuset_attach_old_cs;
2192 /* Called by cgroups to determine if a cpuset is usable; cpuset_rwsem held */
2193 static int cpuset_can_attach(struct cgroup_taskset *tset)
2195 struct cgroup_subsys_state *css;
2197 struct task_struct *task;
2200 /* used later by cpuset_attach() */
2201 cpuset_attach_old_cs = task_cs(cgroup_taskset_first(tset, &css));
2204 percpu_down_write(&cpuset_rwsem);
2206 /* allow moving tasks into an empty cpuset if on default hierarchy */
2208 if (!is_in_v2_mode() &&
2209 (cpumask_empty(cs->cpus_allowed) || nodes_empty(cs->mems_allowed)))
2212 cgroup_taskset_for_each(task, css, tset) {
2213 ret = task_can_attach(task, cs->cpus_allowed);
2216 ret = security_task_setscheduler(task);
2222 * Mark attach is in progress. This makes validate_change() fail
2223 * changes which zero cpus/mems_allowed.
2225 cs->attach_in_progress++;
2228 percpu_up_write(&cpuset_rwsem);
2232 static void cpuset_cancel_attach(struct cgroup_taskset *tset)
2234 struct cgroup_subsys_state *css;
2236 cgroup_taskset_first(tset, &css);
2238 percpu_down_write(&cpuset_rwsem);
2239 css_cs(css)->attach_in_progress--;
2240 percpu_up_write(&cpuset_rwsem);
2244 * Protected by cpuset_rwsem. cpus_attach is used only by cpuset_attach()
2245 * but we can't allocate it dynamically there. Define it global and
2246 * allocate from cpuset_init().
2248 static cpumask_var_t cpus_attach;
2250 static void cpuset_attach(struct cgroup_taskset *tset)
2252 /* static buf protected by cpuset_rwsem */
2253 static nodemask_t cpuset_attach_nodemask_to;
2254 struct task_struct *task;
2255 struct task_struct *leader;
2256 struct cgroup_subsys_state *css;
2258 struct cpuset *oldcs = cpuset_attach_old_cs;
2260 cgroup_taskset_first(tset, &css);
2263 percpu_down_write(&cpuset_rwsem);
2265 guarantee_online_mems(cs, &cpuset_attach_nodemask_to);
2267 cgroup_taskset_for_each(task, css, tset) {
2268 if (cs != &top_cpuset)
2269 guarantee_online_cpus(task, cpus_attach);
2271 cpumask_copy(cpus_attach, task_cpu_possible_mask(task));
2273 * can_attach beforehand should guarantee that this doesn't
2274 * fail. TODO: have a better way to handle failure here
2276 WARN_ON_ONCE(set_cpus_allowed_ptr(task, cpus_attach));
2278 cpuset_change_task_nodemask(task, &cpuset_attach_nodemask_to);
2279 cpuset_update_task_spread_flag(cs, task);
2283 * Change mm for all threadgroup leaders. This is expensive and may
2284 * sleep and should be moved outside migration path proper.
2286 cpuset_attach_nodemask_to = cs->effective_mems;
2287 cgroup_taskset_for_each_leader(leader, css, tset) {
2288 struct mm_struct *mm = get_task_mm(leader);
2291 mpol_rebind_mm(mm, &cpuset_attach_nodemask_to);
2294 * old_mems_allowed is the same with mems_allowed
2295 * here, except if this task is being moved
2296 * automatically due to hotplug. In that case
2297 * @mems_allowed has been updated and is empty, so
2298 * @old_mems_allowed is the right nodesets that we
2301 if (is_memory_migrate(cs))
2302 cpuset_migrate_mm(mm, &oldcs->old_mems_allowed,
2303 &cpuset_attach_nodemask_to);
2309 cs->old_mems_allowed = cpuset_attach_nodemask_to;
2311 cs->attach_in_progress--;
2312 if (!cs->attach_in_progress)
2313 wake_up(&cpuset_attach_wq);
2315 percpu_up_write(&cpuset_rwsem);
2318 /* The various types of files and directories in a cpuset file system */
2321 FILE_MEMORY_MIGRATE,
2324 FILE_EFFECTIVE_CPULIST,
2325 FILE_EFFECTIVE_MEMLIST,
2326 FILE_SUBPARTS_CPULIST,
2330 FILE_SCHED_LOAD_BALANCE,
2331 FILE_PARTITION_ROOT,
2332 FILE_SCHED_RELAX_DOMAIN_LEVEL,
2333 FILE_MEMORY_PRESSURE_ENABLED,
2334 FILE_MEMORY_PRESSURE,
2337 } cpuset_filetype_t;
2339 static int cpuset_write_u64(struct cgroup_subsys_state *css, struct cftype *cft,
2342 struct cpuset *cs = css_cs(css);
2343 cpuset_filetype_t type = cft->private;
2347 percpu_down_write(&cpuset_rwsem);
2348 if (!is_cpuset_online(cs)) {
2354 case FILE_CPU_EXCLUSIVE:
2355 retval = update_flag(CS_CPU_EXCLUSIVE, cs, val);
2357 case FILE_MEM_EXCLUSIVE:
2358 retval = update_flag(CS_MEM_EXCLUSIVE, cs, val);
2360 case FILE_MEM_HARDWALL:
2361 retval = update_flag(CS_MEM_HARDWALL, cs, val);
2363 case FILE_SCHED_LOAD_BALANCE:
2364 retval = update_flag(CS_SCHED_LOAD_BALANCE, cs, val);
2366 case FILE_MEMORY_MIGRATE:
2367 retval = update_flag(CS_MEMORY_MIGRATE, cs, val);
2369 case FILE_MEMORY_PRESSURE_ENABLED:
2370 cpuset_memory_pressure_enabled = !!val;
2372 case FILE_SPREAD_PAGE:
2373 retval = update_flag(CS_SPREAD_PAGE, cs, val);
2375 case FILE_SPREAD_SLAB:
2376 retval = update_flag(CS_SPREAD_SLAB, cs, val);
2383 percpu_up_write(&cpuset_rwsem);
2388 static int cpuset_write_s64(struct cgroup_subsys_state *css, struct cftype *cft,
2391 struct cpuset *cs = css_cs(css);
2392 cpuset_filetype_t type = cft->private;
2393 int retval = -ENODEV;
2396 percpu_down_write(&cpuset_rwsem);
2397 if (!is_cpuset_online(cs))
2401 case FILE_SCHED_RELAX_DOMAIN_LEVEL:
2402 retval = update_relax_domain_level(cs, val);
2409 percpu_up_write(&cpuset_rwsem);
2415 * Common handling for a write to a "cpus" or "mems" file.
2417 static ssize_t cpuset_write_resmask(struct kernfs_open_file *of,
2418 char *buf, size_t nbytes, loff_t off)
2420 struct cpuset *cs = css_cs(of_css(of));
2421 struct cpuset *trialcs;
2422 int retval = -ENODEV;
2424 buf = strstrip(buf);
2427 * CPU or memory hotunplug may leave @cs w/o any execution
2428 * resources, in which case the hotplug code asynchronously updates
2429 * configuration and transfers all tasks to the nearest ancestor
2430 * which can execute.
2432 * As writes to "cpus" or "mems" may restore @cs's execution
2433 * resources, wait for the previously scheduled operations before
2434 * proceeding, so that we don't end up keep removing tasks added
2435 * after execution capability is restored.
2437 * cpuset_hotplug_work calls back into cgroup core via
2438 * cgroup_transfer_tasks() and waiting for it from a cgroupfs
2439 * operation like this one can lead to a deadlock through kernfs
2440 * active_ref protection. Let's break the protection. Losing the
2441 * protection is okay as we check whether @cs is online after
2442 * grabbing cpuset_rwsem anyway. This only happens on the legacy
2446 kernfs_break_active_protection(of->kn);
2447 flush_work(&cpuset_hotplug_work);
2450 percpu_down_write(&cpuset_rwsem);
2451 if (!is_cpuset_online(cs))
2454 trialcs = alloc_trial_cpuset(cs);
2460 switch (of_cft(of)->private) {
2462 retval = update_cpumask(cs, trialcs, buf);
2465 retval = update_nodemask(cs, trialcs, buf);
2472 free_cpuset(trialcs);
2474 percpu_up_write(&cpuset_rwsem);
2476 kernfs_unbreak_active_protection(of->kn);
2478 flush_workqueue(cpuset_migrate_mm_wq);
2479 return retval ?: nbytes;
2483 * These ascii lists should be read in a single call, by using a user
2484 * buffer large enough to hold the entire map. If read in smaller
2485 * chunks, there is no guarantee of atomicity. Since the display format
2486 * used, list of ranges of sequential numbers, is variable length,
2487 * and since these maps can change value dynamically, one could read
2488 * gibberish by doing partial reads while a list was changing.
2490 static int cpuset_common_seq_show(struct seq_file *sf, void *v)
2492 struct cpuset *cs = css_cs(seq_css(sf));
2493 cpuset_filetype_t type = seq_cft(sf)->private;
2496 spin_lock_irq(&callback_lock);
2500 seq_printf(sf, "%*pbl\n", cpumask_pr_args(cs->cpus_allowed));
2503 seq_printf(sf, "%*pbl\n", nodemask_pr_args(&cs->mems_allowed));
2505 case FILE_EFFECTIVE_CPULIST:
2506 seq_printf(sf, "%*pbl\n", cpumask_pr_args(cs->effective_cpus));
2508 case FILE_EFFECTIVE_MEMLIST:
2509 seq_printf(sf, "%*pbl\n", nodemask_pr_args(&cs->effective_mems));
2511 case FILE_SUBPARTS_CPULIST:
2512 seq_printf(sf, "%*pbl\n", cpumask_pr_args(cs->subparts_cpus));
2518 spin_unlock_irq(&callback_lock);
2522 static u64 cpuset_read_u64(struct cgroup_subsys_state *css, struct cftype *cft)
2524 struct cpuset *cs = css_cs(css);
2525 cpuset_filetype_t type = cft->private;
2527 case FILE_CPU_EXCLUSIVE:
2528 return is_cpu_exclusive(cs);
2529 case FILE_MEM_EXCLUSIVE:
2530 return is_mem_exclusive(cs);
2531 case FILE_MEM_HARDWALL:
2532 return is_mem_hardwall(cs);
2533 case FILE_SCHED_LOAD_BALANCE:
2534 return is_sched_load_balance(cs);
2535 case FILE_MEMORY_MIGRATE:
2536 return is_memory_migrate(cs);
2537 case FILE_MEMORY_PRESSURE_ENABLED:
2538 return cpuset_memory_pressure_enabled;
2539 case FILE_MEMORY_PRESSURE:
2540 return fmeter_getrate(&cs->fmeter);
2541 case FILE_SPREAD_PAGE:
2542 return is_spread_page(cs);
2543 case FILE_SPREAD_SLAB:
2544 return is_spread_slab(cs);
2549 /* Unreachable but makes gcc happy */
2553 static s64 cpuset_read_s64(struct cgroup_subsys_state *css, struct cftype *cft)
2555 struct cpuset *cs = css_cs(css);
2556 cpuset_filetype_t type = cft->private;
2558 case FILE_SCHED_RELAX_DOMAIN_LEVEL:
2559 return cs->relax_domain_level;
2564 /* Unreachable but makes gcc happy */
2568 static int sched_partition_show(struct seq_file *seq, void *v)
2570 struct cpuset *cs = css_cs(seq_css(seq));
2572 switch (cs->partition_root_state) {
2574 seq_puts(seq, "root\n");
2577 seq_puts(seq, "member\n");
2580 seq_puts(seq, "root invalid\n");
2586 static ssize_t sched_partition_write(struct kernfs_open_file *of, char *buf,
2587 size_t nbytes, loff_t off)
2589 struct cpuset *cs = css_cs(of_css(of));
2591 int retval = -ENODEV;
2593 buf = strstrip(buf);
2596 * Convert "root" to ENABLED, and convert "member" to DISABLED.
2598 if (!strcmp(buf, "root"))
2600 else if (!strcmp(buf, "member"))
2607 percpu_down_write(&cpuset_rwsem);
2608 if (!is_cpuset_online(cs))
2611 retval = update_prstate(cs, val);
2613 percpu_up_write(&cpuset_rwsem);
2616 return retval ?: nbytes;
2620 * for the common functions, 'private' gives the type of file
2623 static struct cftype legacy_files[] = {
2626 .seq_show = cpuset_common_seq_show,
2627 .write = cpuset_write_resmask,
2628 .max_write_len = (100U + 6 * NR_CPUS),
2629 .private = FILE_CPULIST,
2634 .seq_show = cpuset_common_seq_show,
2635 .write = cpuset_write_resmask,
2636 .max_write_len = (100U + 6 * MAX_NUMNODES),
2637 .private = FILE_MEMLIST,
2641 .name = "effective_cpus",
2642 .seq_show = cpuset_common_seq_show,
2643 .private = FILE_EFFECTIVE_CPULIST,
2647 .name = "effective_mems",
2648 .seq_show = cpuset_common_seq_show,
2649 .private = FILE_EFFECTIVE_MEMLIST,
2653 .name = "cpu_exclusive",
2654 .read_u64 = cpuset_read_u64,
2655 .write_u64 = cpuset_write_u64,
2656 .private = FILE_CPU_EXCLUSIVE,
2660 .name = "mem_exclusive",
2661 .read_u64 = cpuset_read_u64,
2662 .write_u64 = cpuset_write_u64,
2663 .private = FILE_MEM_EXCLUSIVE,
2667 .name = "mem_hardwall",
2668 .read_u64 = cpuset_read_u64,
2669 .write_u64 = cpuset_write_u64,
2670 .private = FILE_MEM_HARDWALL,
2674 .name = "sched_load_balance",
2675 .read_u64 = cpuset_read_u64,
2676 .write_u64 = cpuset_write_u64,
2677 .private = FILE_SCHED_LOAD_BALANCE,
2681 .name = "sched_relax_domain_level",
2682 .read_s64 = cpuset_read_s64,
2683 .write_s64 = cpuset_write_s64,
2684 .private = FILE_SCHED_RELAX_DOMAIN_LEVEL,
2688 .name = "memory_migrate",
2689 .read_u64 = cpuset_read_u64,
2690 .write_u64 = cpuset_write_u64,
2691 .private = FILE_MEMORY_MIGRATE,
2695 .name = "memory_pressure",
2696 .read_u64 = cpuset_read_u64,
2697 .private = FILE_MEMORY_PRESSURE,
2701 .name = "memory_spread_page",
2702 .read_u64 = cpuset_read_u64,
2703 .write_u64 = cpuset_write_u64,
2704 .private = FILE_SPREAD_PAGE,
2708 .name = "memory_spread_slab",
2709 .read_u64 = cpuset_read_u64,
2710 .write_u64 = cpuset_write_u64,
2711 .private = FILE_SPREAD_SLAB,
2715 .name = "memory_pressure_enabled",
2716 .flags = CFTYPE_ONLY_ON_ROOT,
2717 .read_u64 = cpuset_read_u64,
2718 .write_u64 = cpuset_write_u64,
2719 .private = FILE_MEMORY_PRESSURE_ENABLED,
2726 * This is currently a minimal set for the default hierarchy. It can be
2727 * expanded later on by migrating more features and control files from v1.
2729 static struct cftype dfl_files[] = {
2732 .seq_show = cpuset_common_seq_show,
2733 .write = cpuset_write_resmask,
2734 .max_write_len = (100U + 6 * NR_CPUS),
2735 .private = FILE_CPULIST,
2736 .flags = CFTYPE_NOT_ON_ROOT,
2741 .seq_show = cpuset_common_seq_show,
2742 .write = cpuset_write_resmask,
2743 .max_write_len = (100U + 6 * MAX_NUMNODES),
2744 .private = FILE_MEMLIST,
2745 .flags = CFTYPE_NOT_ON_ROOT,
2749 .name = "cpus.effective",
2750 .seq_show = cpuset_common_seq_show,
2751 .private = FILE_EFFECTIVE_CPULIST,
2755 .name = "mems.effective",
2756 .seq_show = cpuset_common_seq_show,
2757 .private = FILE_EFFECTIVE_MEMLIST,
2761 .name = "cpus.partition",
2762 .seq_show = sched_partition_show,
2763 .write = sched_partition_write,
2764 .private = FILE_PARTITION_ROOT,
2765 .flags = CFTYPE_NOT_ON_ROOT,
2766 .file_offset = offsetof(struct cpuset, partition_file),
2770 .name = "cpus.subpartitions",
2771 .seq_show = cpuset_common_seq_show,
2772 .private = FILE_SUBPARTS_CPULIST,
2773 .flags = CFTYPE_DEBUG,
2781 * cpuset_css_alloc - allocate a cpuset css
2782 * cgrp: control group that the new cpuset will be part of
2785 static struct cgroup_subsys_state *
2786 cpuset_css_alloc(struct cgroup_subsys_state *parent_css)
2791 return &top_cpuset.css;
2793 cs = kzalloc(sizeof(*cs), GFP_KERNEL);
2795 return ERR_PTR(-ENOMEM);
2797 if (alloc_cpumasks(cs, NULL)) {
2799 return ERR_PTR(-ENOMEM);
2802 __set_bit(CS_SCHED_LOAD_BALANCE, &cs->flags);
2803 nodes_clear(cs->mems_allowed);
2804 nodes_clear(cs->effective_mems);
2805 fmeter_init(&cs->fmeter);
2806 cs->relax_domain_level = -1;
2808 /* Set CS_MEMORY_MIGRATE for default hierarchy */
2809 if (cgroup_subsys_on_dfl(cpuset_cgrp_subsys))
2810 __set_bit(CS_MEMORY_MIGRATE, &cs->flags);
2815 static int cpuset_css_online(struct cgroup_subsys_state *css)
2817 struct cpuset *cs = css_cs(css);
2818 struct cpuset *parent = parent_cs(cs);
2819 struct cpuset *tmp_cs;
2820 struct cgroup_subsys_state *pos_css;
2826 percpu_down_write(&cpuset_rwsem);
2828 set_bit(CS_ONLINE, &cs->flags);
2829 if (is_spread_page(parent))
2830 set_bit(CS_SPREAD_PAGE, &cs->flags);
2831 if (is_spread_slab(parent))
2832 set_bit(CS_SPREAD_SLAB, &cs->flags);
2836 spin_lock_irq(&callback_lock);
2837 if (is_in_v2_mode()) {
2838 cpumask_copy(cs->effective_cpus, parent->effective_cpus);
2839 cs->effective_mems = parent->effective_mems;
2840 cs->use_parent_ecpus = true;
2841 parent->child_ecpus_count++;
2843 spin_unlock_irq(&callback_lock);
2845 if (!test_bit(CGRP_CPUSET_CLONE_CHILDREN, &css->cgroup->flags))
2849 * Clone @parent's configuration if CGRP_CPUSET_CLONE_CHILDREN is
2850 * set. This flag handling is implemented in cgroup core for
2851 * histrical reasons - the flag may be specified during mount.
2853 * Currently, if any sibling cpusets have exclusive cpus or mem, we
2854 * refuse to clone the configuration - thereby refusing the task to
2855 * be entered, and as a result refusing the sys_unshare() or
2856 * clone() which initiated it. If this becomes a problem for some
2857 * users who wish to allow that scenario, then this could be
2858 * changed to grant parent->cpus_allowed-sibling_cpus_exclusive
2859 * (and likewise for mems) to the new cgroup.
2862 cpuset_for_each_child(tmp_cs, pos_css, parent) {
2863 if (is_mem_exclusive(tmp_cs) || is_cpu_exclusive(tmp_cs)) {
2870 spin_lock_irq(&callback_lock);
2871 cs->mems_allowed = parent->mems_allowed;
2872 cs->effective_mems = parent->mems_allowed;
2873 cpumask_copy(cs->cpus_allowed, parent->cpus_allowed);
2874 cpumask_copy(cs->effective_cpus, parent->cpus_allowed);
2875 spin_unlock_irq(&callback_lock);
2877 percpu_up_write(&cpuset_rwsem);
2883 * If the cpuset being removed has its flag 'sched_load_balance'
2884 * enabled, then simulate turning sched_load_balance off, which
2885 * will call rebuild_sched_domains_locked(). That is not needed
2886 * in the default hierarchy where only changes in partition
2887 * will cause repartitioning.
2889 * If the cpuset has the 'sched.partition' flag enabled, simulate
2890 * turning 'sched.partition" off.
2893 static void cpuset_css_offline(struct cgroup_subsys_state *css)
2895 struct cpuset *cs = css_cs(css);
2898 percpu_down_write(&cpuset_rwsem);
2900 if (is_partition_root(cs))
2901 update_prstate(cs, 0);
2903 if (!cgroup_subsys_on_dfl(cpuset_cgrp_subsys) &&
2904 is_sched_load_balance(cs))
2905 update_flag(CS_SCHED_LOAD_BALANCE, cs, 0);
2907 if (cs->use_parent_ecpus) {
2908 struct cpuset *parent = parent_cs(cs);
2910 cs->use_parent_ecpus = false;
2911 parent->child_ecpus_count--;
2915 clear_bit(CS_ONLINE, &cs->flags);
2917 percpu_up_write(&cpuset_rwsem);
2921 static void cpuset_css_free(struct cgroup_subsys_state *css)
2923 struct cpuset *cs = css_cs(css);
2928 static void cpuset_bind(struct cgroup_subsys_state *root_css)
2930 percpu_down_write(&cpuset_rwsem);
2931 spin_lock_irq(&callback_lock);
2933 if (is_in_v2_mode()) {
2934 cpumask_copy(top_cpuset.cpus_allowed, cpu_possible_mask);
2935 top_cpuset.mems_allowed = node_possible_map;
2937 cpumask_copy(top_cpuset.cpus_allowed,
2938 top_cpuset.effective_cpus);
2939 top_cpuset.mems_allowed = top_cpuset.effective_mems;
2942 spin_unlock_irq(&callback_lock);
2943 percpu_up_write(&cpuset_rwsem);
2947 * Make sure the new task conform to the current state of its parent,
2948 * which could have been changed by cpuset just after it inherits the
2949 * state from the parent and before it sits on the cgroup's task list.
2951 static void cpuset_fork(struct task_struct *task)
2953 if (task_css_is_root(task, cpuset_cgrp_id))
2956 set_cpus_allowed_ptr(task, current->cpus_ptr);
2957 task->mems_allowed = current->mems_allowed;
2960 struct cgroup_subsys cpuset_cgrp_subsys = {
2961 .css_alloc = cpuset_css_alloc,
2962 .css_online = cpuset_css_online,
2963 .css_offline = cpuset_css_offline,
2964 .css_free = cpuset_css_free,
2965 .can_attach = cpuset_can_attach,
2966 .cancel_attach = cpuset_cancel_attach,
2967 .attach = cpuset_attach,
2968 .post_attach = cpuset_post_attach,
2969 .bind = cpuset_bind,
2970 .fork = cpuset_fork,
2971 .legacy_cftypes = legacy_files,
2972 .dfl_cftypes = dfl_files,
2978 * cpuset_init - initialize cpusets at system boot
2980 * Description: Initialize top_cpuset
2983 int __init cpuset_init(void)
2985 BUG_ON(percpu_init_rwsem(&cpuset_rwsem));
2987 BUG_ON(!alloc_cpumask_var(&top_cpuset.cpus_allowed, GFP_KERNEL));
2988 BUG_ON(!alloc_cpumask_var(&top_cpuset.effective_cpus, GFP_KERNEL));
2989 BUG_ON(!zalloc_cpumask_var(&top_cpuset.subparts_cpus, GFP_KERNEL));
2991 cpumask_setall(top_cpuset.cpus_allowed);
2992 nodes_setall(top_cpuset.mems_allowed);
2993 cpumask_setall(top_cpuset.effective_cpus);
2994 nodes_setall(top_cpuset.effective_mems);
2996 fmeter_init(&top_cpuset.fmeter);
2997 set_bit(CS_SCHED_LOAD_BALANCE, &top_cpuset.flags);
2998 top_cpuset.relax_domain_level = -1;
3000 BUG_ON(!alloc_cpumask_var(&cpus_attach, GFP_KERNEL));
3006 * If CPU and/or memory hotplug handlers, below, unplug any CPUs
3007 * or memory nodes, we need to walk over the cpuset hierarchy,
3008 * removing that CPU or node from all cpusets. If this removes the
3009 * last CPU or node from a cpuset, then move the tasks in the empty
3010 * cpuset to its next-highest non-empty parent.
3012 static void remove_tasks_in_empty_cpuset(struct cpuset *cs)
3014 struct cpuset *parent;
3017 * Find its next-highest non-empty parent, (top cpuset
3018 * has online cpus, so can't be empty).
3020 parent = parent_cs(cs);
3021 while (cpumask_empty(parent->cpus_allowed) ||
3022 nodes_empty(parent->mems_allowed))
3023 parent = parent_cs(parent);
3025 if (cgroup_transfer_tasks(parent->css.cgroup, cs->css.cgroup)) {
3026 pr_err("cpuset: failed to transfer tasks out of empty cpuset ");
3027 pr_cont_cgroup_name(cs->css.cgroup);
3033 hotplug_update_tasks_legacy(struct cpuset *cs,
3034 struct cpumask *new_cpus, nodemask_t *new_mems,
3035 bool cpus_updated, bool mems_updated)
3039 spin_lock_irq(&callback_lock);
3040 cpumask_copy(cs->cpus_allowed, new_cpus);
3041 cpumask_copy(cs->effective_cpus, new_cpus);
3042 cs->mems_allowed = *new_mems;
3043 cs->effective_mems = *new_mems;
3044 spin_unlock_irq(&callback_lock);
3047 * Don't call update_tasks_cpumask() if the cpuset becomes empty,
3048 * as the tasks will be migratecd to an ancestor.
3050 if (cpus_updated && !cpumask_empty(cs->cpus_allowed))
3051 update_tasks_cpumask(cs);
3052 if (mems_updated && !nodes_empty(cs->mems_allowed))
3053 update_tasks_nodemask(cs);
3055 is_empty = cpumask_empty(cs->cpus_allowed) ||
3056 nodes_empty(cs->mems_allowed);
3058 percpu_up_write(&cpuset_rwsem);
3061 * Move tasks to the nearest ancestor with execution resources,
3062 * This is full cgroup operation which will also call back into
3063 * cpuset. Should be done outside any lock.
3066 remove_tasks_in_empty_cpuset(cs);
3068 percpu_down_write(&cpuset_rwsem);
3072 hotplug_update_tasks(struct cpuset *cs,
3073 struct cpumask *new_cpus, nodemask_t *new_mems,
3074 bool cpus_updated, bool mems_updated)
3076 if (cpumask_empty(new_cpus))
3077 cpumask_copy(new_cpus, parent_cs(cs)->effective_cpus);
3078 if (nodes_empty(*new_mems))
3079 *new_mems = parent_cs(cs)->effective_mems;
3081 spin_lock_irq(&callback_lock);
3082 cpumask_copy(cs->effective_cpus, new_cpus);
3083 cs->effective_mems = *new_mems;
3084 spin_unlock_irq(&callback_lock);
3087 update_tasks_cpumask(cs);
3089 update_tasks_nodemask(cs);
3092 static bool force_rebuild;
3094 void cpuset_force_rebuild(void)
3096 force_rebuild = true;
3100 * cpuset_hotplug_update_tasks - update tasks in a cpuset for hotunplug
3101 * @cs: cpuset in interest
3102 * @tmp: the tmpmasks structure pointer
3104 * Compare @cs's cpu and mem masks against top_cpuset and if some have gone
3105 * offline, update @cs accordingly. If @cs ends up with no CPU or memory,
3106 * all its tasks are moved to the nearest ancestor with both resources.
3108 static void cpuset_hotplug_update_tasks(struct cpuset *cs, struct tmpmasks *tmp)
3110 static cpumask_t new_cpus;
3111 static nodemask_t new_mems;
3114 struct cpuset *parent;
3116 wait_event(cpuset_attach_wq, cs->attach_in_progress == 0);
3118 percpu_down_write(&cpuset_rwsem);
3121 * We have raced with task attaching. We wait until attaching
3122 * is finished, so we won't attach a task to an empty cpuset.
3124 if (cs->attach_in_progress) {
3125 percpu_up_write(&cpuset_rwsem);
3129 parent = parent_cs(cs);
3130 compute_effective_cpumask(&new_cpus, cs, parent);
3131 nodes_and(new_mems, cs->mems_allowed, parent->effective_mems);
3133 if (cs->nr_subparts_cpus)
3135 * Make sure that CPUs allocated to child partitions
3136 * do not show up in effective_cpus.
3138 cpumask_andnot(&new_cpus, &new_cpus, cs->subparts_cpus);
3140 if (!tmp || !cs->partition_root_state)
3144 * In the unlikely event that a partition root has empty
3145 * effective_cpus or its parent becomes erroneous, we have to
3146 * transition it to the erroneous state.
3148 if (is_partition_root(cs) && (cpumask_empty(&new_cpus) ||
3149 (parent->partition_root_state == PRS_ERROR))) {
3150 if (cs->nr_subparts_cpus) {
3151 spin_lock_irq(&callback_lock);
3152 cs->nr_subparts_cpus = 0;
3153 cpumask_clear(cs->subparts_cpus);
3154 spin_unlock_irq(&callback_lock);
3155 compute_effective_cpumask(&new_cpus, cs, parent);
3159 * If the effective_cpus is empty because the child
3160 * partitions take away all the CPUs, we can keep
3161 * the current partition and let the child partitions
3162 * fight for available CPUs.
3164 if ((parent->partition_root_state == PRS_ERROR) ||
3165 cpumask_empty(&new_cpus)) {
3168 update_parent_subparts_cpumask(cs, partcmd_disable,
3170 old_prs = cs->partition_root_state;
3171 if (old_prs != PRS_ERROR) {
3172 spin_lock_irq(&callback_lock);
3173 cs->partition_root_state = PRS_ERROR;
3174 spin_unlock_irq(&callback_lock);
3175 notify_partition_change(cs, old_prs, PRS_ERROR);
3178 cpuset_force_rebuild();
3182 * On the other hand, an erroneous partition root may be transitioned
3183 * back to a regular one or a partition root with no CPU allocated
3184 * from the parent may change to erroneous.
3186 if (is_partition_root(parent) &&
3187 ((cs->partition_root_state == PRS_ERROR) ||
3188 !cpumask_intersects(&new_cpus, parent->subparts_cpus)) &&
3189 update_parent_subparts_cpumask(cs, partcmd_update, NULL, tmp))
3190 cpuset_force_rebuild();
3193 cpus_updated = !cpumask_equal(&new_cpus, cs->effective_cpus);
3194 mems_updated = !nodes_equal(new_mems, cs->effective_mems);
3197 check_insane_mems_config(&new_mems);
3199 if (is_in_v2_mode())
3200 hotplug_update_tasks(cs, &new_cpus, &new_mems,
3201 cpus_updated, mems_updated);
3203 hotplug_update_tasks_legacy(cs, &new_cpus, &new_mems,
3204 cpus_updated, mems_updated);
3206 percpu_up_write(&cpuset_rwsem);
3210 * cpuset_hotplug_workfn - handle CPU/memory hotunplug for a cpuset
3212 * This function is called after either CPU or memory configuration has
3213 * changed and updates cpuset accordingly. The top_cpuset is always
3214 * synchronized to cpu_active_mask and N_MEMORY, which is necessary in
3215 * order to make cpusets transparent (of no affect) on systems that are
3216 * actively using CPU hotplug but making no active use of cpusets.
3218 * Non-root cpusets are only affected by offlining. If any CPUs or memory
3219 * nodes have been taken down, cpuset_hotplug_update_tasks() is invoked on
3222 * Note that CPU offlining during suspend is ignored. We don't modify
3223 * cpusets across suspend/resume cycles at all.
3225 static void cpuset_hotplug_workfn(struct work_struct *work)
3227 static cpumask_t new_cpus;
3228 static nodemask_t new_mems;
3229 bool cpus_updated, mems_updated;
3230 bool on_dfl = is_in_v2_mode();
3231 struct tmpmasks tmp, *ptmp = NULL;
3233 if (on_dfl && !alloc_cpumasks(NULL, &tmp))
3236 percpu_down_write(&cpuset_rwsem);
3238 /* fetch the available cpus/mems and find out which changed how */
3239 cpumask_copy(&new_cpus, cpu_active_mask);
3240 new_mems = node_states[N_MEMORY];
3243 * If subparts_cpus is populated, it is likely that the check below
3244 * will produce a false positive on cpus_updated when the cpu list
3245 * isn't changed. It is extra work, but it is better to be safe.
3247 cpus_updated = !cpumask_equal(top_cpuset.effective_cpus, &new_cpus);
3248 mems_updated = !nodes_equal(top_cpuset.effective_mems, new_mems);
3251 * In the rare case that hotplug removes all the cpus in subparts_cpus,
3252 * we assumed that cpus are updated.
3254 if (!cpus_updated && top_cpuset.nr_subparts_cpus)
3255 cpus_updated = true;
3257 /* synchronize cpus_allowed to cpu_active_mask */
3259 spin_lock_irq(&callback_lock);
3261 cpumask_copy(top_cpuset.cpus_allowed, &new_cpus);
3263 * Make sure that CPUs allocated to child partitions
3264 * do not show up in effective_cpus. If no CPU is left,
3265 * we clear the subparts_cpus & let the child partitions
3266 * fight for the CPUs again.
3268 if (top_cpuset.nr_subparts_cpus) {
3269 if (cpumask_subset(&new_cpus,
3270 top_cpuset.subparts_cpus)) {
3271 top_cpuset.nr_subparts_cpus = 0;
3272 cpumask_clear(top_cpuset.subparts_cpus);
3274 cpumask_andnot(&new_cpus, &new_cpus,
3275 top_cpuset.subparts_cpus);
3278 cpumask_copy(top_cpuset.effective_cpus, &new_cpus);
3279 spin_unlock_irq(&callback_lock);
3280 /* we don't mess with cpumasks of tasks in top_cpuset */
3283 /* synchronize mems_allowed to N_MEMORY */
3285 spin_lock_irq(&callback_lock);
3287 top_cpuset.mems_allowed = new_mems;
3288 top_cpuset.effective_mems = new_mems;
3289 spin_unlock_irq(&callback_lock);
3290 update_tasks_nodemask(&top_cpuset);
3293 percpu_up_write(&cpuset_rwsem);
3295 /* if cpus or mems changed, we need to propagate to descendants */
3296 if (cpus_updated || mems_updated) {
3298 struct cgroup_subsys_state *pos_css;
3301 cpuset_for_each_descendant_pre(cs, pos_css, &top_cpuset) {
3302 if (cs == &top_cpuset || !css_tryget_online(&cs->css))
3306 cpuset_hotplug_update_tasks(cs, ptmp);
3314 /* rebuild sched domains if cpus_allowed has changed */
3315 if (cpus_updated || force_rebuild) {
3316 force_rebuild = false;
3317 rebuild_sched_domains();
3320 free_cpumasks(NULL, ptmp);
3323 void cpuset_update_active_cpus(void)
3326 * We're inside cpu hotplug critical region which usually nests
3327 * inside cgroup synchronization. Bounce actual hotplug processing
3328 * to a work item to avoid reverse locking order.
3330 schedule_work(&cpuset_hotplug_work);
3333 void cpuset_wait_for_hotplug(void)
3335 flush_work(&cpuset_hotplug_work);
3339 * Keep top_cpuset.mems_allowed tracking node_states[N_MEMORY].
3340 * Call this routine anytime after node_states[N_MEMORY] changes.
3341 * See cpuset_update_active_cpus() for CPU hotplug handling.
3343 static int cpuset_track_online_nodes(struct notifier_block *self,
3344 unsigned long action, void *arg)
3346 schedule_work(&cpuset_hotplug_work);
3350 static struct notifier_block cpuset_track_online_nodes_nb = {
3351 .notifier_call = cpuset_track_online_nodes,
3352 .priority = 10, /* ??! */
3356 * cpuset_init_smp - initialize cpus_allowed
3358 * Description: Finish top cpuset after cpu, node maps are initialized
3360 void __init cpuset_init_smp(void)
3362 cpumask_copy(top_cpuset.cpus_allowed, cpu_active_mask);
3363 top_cpuset.mems_allowed = node_states[N_MEMORY];
3364 top_cpuset.old_mems_allowed = top_cpuset.mems_allowed;
3366 cpumask_copy(top_cpuset.effective_cpus, cpu_active_mask);
3367 top_cpuset.effective_mems = node_states[N_MEMORY];
3369 register_hotmemory_notifier(&cpuset_track_online_nodes_nb);
3371 cpuset_migrate_mm_wq = alloc_ordered_workqueue("cpuset_migrate_mm", 0);
3372 BUG_ON(!cpuset_migrate_mm_wq);
3376 * cpuset_cpus_allowed - return cpus_allowed mask from a tasks cpuset.
3377 * @tsk: pointer to task_struct from which to obtain cpuset->cpus_allowed.
3378 * @pmask: pointer to struct cpumask variable to receive cpus_allowed set.
3380 * Description: Returns the cpumask_var_t cpus_allowed of the cpuset
3381 * attached to the specified @tsk. Guaranteed to return some non-empty
3382 * subset of cpu_online_mask, even if this means going outside the
3386 void cpuset_cpus_allowed(struct task_struct *tsk, struct cpumask *pmask)
3388 unsigned long flags;
3390 spin_lock_irqsave(&callback_lock, flags);
3391 guarantee_online_cpus(tsk, pmask);
3392 spin_unlock_irqrestore(&callback_lock, flags);
3396 * cpuset_cpus_allowed_fallback - final fallback before complete catastrophe.
3397 * @tsk: pointer to task_struct with which the scheduler is struggling
3399 * Description: In the case that the scheduler cannot find an allowed cpu in
3400 * tsk->cpus_allowed, we fall back to task_cs(tsk)->cpus_allowed. In legacy
3401 * mode however, this value is the same as task_cs(tsk)->effective_cpus,
3402 * which will not contain a sane cpumask during cases such as cpu hotplugging.
3403 * This is the absolute last resort for the scheduler and it is only used if
3404 * _every_ other avenue has been traveled.
3406 * Returns true if the affinity of @tsk was changed, false otherwise.
3409 bool cpuset_cpus_allowed_fallback(struct task_struct *tsk)
3411 const struct cpumask *possible_mask = task_cpu_possible_mask(tsk);
3412 const struct cpumask *cs_mask;
3413 bool changed = false;
3416 cs_mask = task_cs(tsk)->cpus_allowed;
3417 if (is_in_v2_mode() && cpumask_subset(cs_mask, possible_mask)) {
3418 do_set_cpus_allowed(tsk, cs_mask);
3424 * We own tsk->cpus_allowed, nobody can change it under us.
3426 * But we used cs && cs->cpus_allowed lockless and thus can
3427 * race with cgroup_attach_task() or update_cpumask() and get
3428 * the wrong tsk->cpus_allowed. However, both cases imply the
3429 * subsequent cpuset_change_cpumask()->set_cpus_allowed_ptr()
3430 * which takes task_rq_lock().
3432 * If we are called after it dropped the lock we must see all
3433 * changes in tsk_cs()->cpus_allowed. Otherwise we can temporary
3434 * set any mask even if it is not right from task_cs() pov,
3435 * the pending set_cpus_allowed_ptr() will fix things.
3437 * select_fallback_rq() will fix things ups and set cpu_possible_mask
3443 void __init cpuset_init_current_mems_allowed(void)
3445 nodes_setall(current->mems_allowed);
3449 * cpuset_mems_allowed - return mems_allowed mask from a tasks cpuset.
3450 * @tsk: pointer to task_struct from which to obtain cpuset->mems_allowed.
3452 * Description: Returns the nodemask_t mems_allowed of the cpuset
3453 * attached to the specified @tsk. Guaranteed to return some non-empty
3454 * subset of node_states[N_MEMORY], even if this means going outside the
3458 nodemask_t cpuset_mems_allowed(struct task_struct *tsk)
3461 unsigned long flags;
3463 spin_lock_irqsave(&callback_lock, flags);
3465 guarantee_online_mems(task_cs(tsk), &mask);
3467 spin_unlock_irqrestore(&callback_lock, flags);
3473 * cpuset_nodemask_valid_mems_allowed - check nodemask vs. current mems_allowed
3474 * @nodemask: the nodemask to be checked
3476 * Are any of the nodes in the nodemask allowed in current->mems_allowed?
3478 int cpuset_nodemask_valid_mems_allowed(nodemask_t *nodemask)
3480 return nodes_intersects(*nodemask, current->mems_allowed);
3484 * nearest_hardwall_ancestor() - Returns the nearest mem_exclusive or
3485 * mem_hardwall ancestor to the specified cpuset. Call holding
3486 * callback_lock. If no ancestor is mem_exclusive or mem_hardwall
3487 * (an unusual configuration), then returns the root cpuset.
3489 static struct cpuset *nearest_hardwall_ancestor(struct cpuset *cs)
3491 while (!(is_mem_exclusive(cs) || is_mem_hardwall(cs)) && parent_cs(cs))
3497 * cpuset_node_allowed - Can we allocate on a memory node?
3498 * @node: is this an allowed node?
3499 * @gfp_mask: memory allocation flags
3501 * If we're in interrupt, yes, we can always allocate. If @node is set in
3502 * current's mems_allowed, yes. If it's not a __GFP_HARDWALL request and this
3503 * node is set in the nearest hardwalled cpuset ancestor to current's cpuset,
3504 * yes. If current has access to memory reserves as an oom victim, yes.
3507 * GFP_USER allocations are marked with the __GFP_HARDWALL bit,
3508 * and do not allow allocations outside the current tasks cpuset
3509 * unless the task has been OOM killed.
3510 * GFP_KERNEL allocations are not so marked, so can escape to the
3511 * nearest enclosing hardwalled ancestor cpuset.
3513 * Scanning up parent cpusets requires callback_lock. The
3514 * __alloc_pages() routine only calls here with __GFP_HARDWALL bit
3515 * _not_ set if it's a GFP_KERNEL allocation, and all nodes in the
3516 * current tasks mems_allowed came up empty on the first pass over
3517 * the zonelist. So only GFP_KERNEL allocations, if all nodes in the
3518 * cpuset are short of memory, might require taking the callback_lock.
3520 * The first call here from mm/page_alloc:get_page_from_freelist()
3521 * has __GFP_HARDWALL set in gfp_mask, enforcing hardwall cpusets,
3522 * so no allocation on a node outside the cpuset is allowed (unless
3523 * in interrupt, of course).
3525 * The second pass through get_page_from_freelist() doesn't even call
3526 * here for GFP_ATOMIC calls. For those calls, the __alloc_pages()
3527 * variable 'wait' is not set, and the bit ALLOC_CPUSET is not set
3528 * in alloc_flags. That logic and the checks below have the combined
3530 * in_interrupt - any node ok (current task context irrelevant)
3531 * GFP_ATOMIC - any node ok
3532 * tsk_is_oom_victim - any node ok
3533 * GFP_KERNEL - any node in enclosing hardwalled cpuset ok
3534 * GFP_USER - only nodes in current tasks mems allowed ok.
3536 bool __cpuset_node_allowed(int node, gfp_t gfp_mask)
3538 struct cpuset *cs; /* current cpuset ancestors */
3539 int allowed; /* is allocation in zone z allowed? */
3540 unsigned long flags;
3544 if (node_isset(node, current->mems_allowed))
3547 * Allow tasks that have access to memory reserves because they have
3548 * been OOM killed to get memory anywhere.
3550 if (unlikely(tsk_is_oom_victim(current)))
3552 if (gfp_mask & __GFP_HARDWALL) /* If hardwall request, stop here */
3555 if (current->flags & PF_EXITING) /* Let dying task have memory */
3558 /* Not hardwall and node outside mems_allowed: scan up cpusets */
3559 spin_lock_irqsave(&callback_lock, flags);
3562 cs = nearest_hardwall_ancestor(task_cs(current));
3563 allowed = node_isset(node, cs->mems_allowed);
3566 spin_unlock_irqrestore(&callback_lock, flags);
3571 * cpuset_mem_spread_node() - On which node to begin search for a file page
3572 * cpuset_slab_spread_node() - On which node to begin search for a slab page
3574 * If a task is marked PF_SPREAD_PAGE or PF_SPREAD_SLAB (as for
3575 * tasks in a cpuset with is_spread_page or is_spread_slab set),
3576 * and if the memory allocation used cpuset_mem_spread_node()
3577 * to determine on which node to start looking, as it will for
3578 * certain page cache or slab cache pages such as used for file
3579 * system buffers and inode caches, then instead of starting on the
3580 * local node to look for a free page, rather spread the starting
3581 * node around the tasks mems_allowed nodes.
3583 * We don't have to worry about the returned node being offline
3584 * because "it can't happen", and even if it did, it would be ok.
3586 * The routines calling guarantee_online_mems() are careful to
3587 * only set nodes in task->mems_allowed that are online. So it
3588 * should not be possible for the following code to return an
3589 * offline node. But if it did, that would be ok, as this routine
3590 * is not returning the node where the allocation must be, only
3591 * the node where the search should start. The zonelist passed to
3592 * __alloc_pages() will include all nodes. If the slab allocator
3593 * is passed an offline node, it will fall back to the local node.
3594 * See kmem_cache_alloc_node().
3597 static int cpuset_spread_node(int *rotor)
3599 return *rotor = next_node_in(*rotor, current->mems_allowed);
3602 int cpuset_mem_spread_node(void)
3604 if (current->cpuset_mem_spread_rotor == NUMA_NO_NODE)
3605 current->cpuset_mem_spread_rotor =
3606 node_random(¤t->mems_allowed);
3608 return cpuset_spread_node(¤t->cpuset_mem_spread_rotor);
3611 int cpuset_slab_spread_node(void)
3613 if (current->cpuset_slab_spread_rotor == NUMA_NO_NODE)
3614 current->cpuset_slab_spread_rotor =
3615 node_random(¤t->mems_allowed);
3617 return cpuset_spread_node(¤t->cpuset_slab_spread_rotor);
3620 EXPORT_SYMBOL_GPL(cpuset_mem_spread_node);
3623 * cpuset_mems_allowed_intersects - Does @tsk1's mems_allowed intersect @tsk2's?
3624 * @tsk1: pointer to task_struct of some task.
3625 * @tsk2: pointer to task_struct of some other task.
3627 * Description: Return true if @tsk1's mems_allowed intersects the
3628 * mems_allowed of @tsk2. Used by the OOM killer to determine if
3629 * one of the task's memory usage might impact the memory available
3633 int cpuset_mems_allowed_intersects(const struct task_struct *tsk1,
3634 const struct task_struct *tsk2)
3636 return nodes_intersects(tsk1->mems_allowed, tsk2->mems_allowed);
3640 * cpuset_print_current_mems_allowed - prints current's cpuset and mems_allowed
3642 * Description: Prints current's name, cpuset name, and cached copy of its
3643 * mems_allowed to the kernel log.
3645 void cpuset_print_current_mems_allowed(void)
3647 struct cgroup *cgrp;
3651 cgrp = task_cs(current)->css.cgroup;
3652 pr_cont(",cpuset=");
3653 pr_cont_cgroup_name(cgrp);
3654 pr_cont(",mems_allowed=%*pbl",
3655 nodemask_pr_args(¤t->mems_allowed));
3661 * Collection of memory_pressure is suppressed unless
3662 * this flag is enabled by writing "1" to the special
3663 * cpuset file 'memory_pressure_enabled' in the root cpuset.
3666 int cpuset_memory_pressure_enabled __read_mostly;
3669 * cpuset_memory_pressure_bump - keep stats of per-cpuset reclaims.
3671 * Keep a running average of the rate of synchronous (direct)
3672 * page reclaim efforts initiated by tasks in each cpuset.
3674 * This represents the rate at which some task in the cpuset
3675 * ran low on memory on all nodes it was allowed to use, and
3676 * had to enter the kernels page reclaim code in an effort to
3677 * create more free memory by tossing clean pages or swapping
3678 * or writing dirty pages.
3680 * Display to user space in the per-cpuset read-only file
3681 * "memory_pressure". Value displayed is an integer
3682 * representing the recent rate of entry into the synchronous
3683 * (direct) page reclaim by any task attached to the cpuset.
3686 void __cpuset_memory_pressure_bump(void)
3689 fmeter_markevent(&task_cs(current)->fmeter);
3693 #ifdef CONFIG_PROC_PID_CPUSET
3695 * proc_cpuset_show()
3696 * - Print tasks cpuset path into seq_file.
3697 * - Used for /proc/<pid>/cpuset.
3698 * - No need to task_lock(tsk) on this tsk->cpuset reference, as it
3699 * doesn't really matter if tsk->cpuset changes after we read it,
3700 * and we take cpuset_rwsem, keeping cpuset_attach() from changing it
3703 int proc_cpuset_show(struct seq_file *m, struct pid_namespace *ns,
3704 struct pid *pid, struct task_struct *tsk)
3707 struct cgroup_subsys_state *css;
3711 buf = kmalloc(PATH_MAX, GFP_KERNEL);
3715 css = task_get_css(tsk, cpuset_cgrp_id);
3716 retval = cgroup_path_ns(css->cgroup, buf, PATH_MAX,
3717 current->nsproxy->cgroup_ns);
3719 if (retval >= PATH_MAX)
3720 retval = -ENAMETOOLONG;
3731 #endif /* CONFIG_PROC_PID_CPUSET */
3733 /* Display task mems_allowed in /proc/<pid>/status file. */
3734 void cpuset_task_status_allowed(struct seq_file *m, struct task_struct *task)
3736 seq_printf(m, "Mems_allowed:\t%*pb\n",
3737 nodemask_pr_args(&task->mems_allowed));
3738 seq_printf(m, "Mems_allowed_list:\t%*pbl\n",
3739 nodemask_pr_args(&task->mems_allowed));