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/kthread.h>
37 #include <linux/list.h>
38 #include <linux/mempolicy.h>
40 #include <linux/memory.h>
41 #include <linux/export.h>
42 #include <linux/mount.h>
43 #include <linux/fs_context.h>
44 #include <linux/namei.h>
45 #include <linux/pagemap.h>
46 #include <linux/proc_fs.h>
47 #include <linux/rcupdate.h>
48 #include <linux/sched.h>
49 #include <linux/sched/deadline.h>
50 #include <linux/sched/mm.h>
51 #include <linux/sched/task.h>
52 #include <linux/seq_file.h>
53 #include <linux/security.h>
54 #include <linux/slab.h>
55 #include <linux/spinlock.h>
56 #include <linux/stat.h>
57 #include <linux/string.h>
58 #include <linux/time.h>
59 #include <linux/time64.h>
60 #include <linux/backing-dev.h>
61 #include <linux/sort.h>
62 #include <linux/oom.h>
63 #include <linux/sched/isolation.h>
64 #include <linux/uaccess.h>
65 #include <linux/atomic.h>
66 #include <linux/mutex.h>
67 #include <linux/cgroup.h>
68 #include <linux/wait.h>
70 DEFINE_STATIC_KEY_FALSE(cpusets_pre_enable_key);
71 DEFINE_STATIC_KEY_FALSE(cpusets_enabled_key);
73 /* See "Frequency meter" comments, below. */
76 int cnt; /* unprocessed events count */
77 int val; /* most recent output value */
78 time64_t time; /* clock (secs) when val computed */
79 spinlock_t lock; /* guards read or write of above */
83 struct cgroup_subsys_state css;
85 unsigned long flags; /* "unsigned long" so bitops work */
88 * On default hierarchy:
90 * The user-configured masks can only be changed by writing to
91 * cpuset.cpus and cpuset.mems, and won't be limited by the
94 * The effective masks is the real masks that apply to the tasks
95 * in the cpuset. They may be changed if the configured masks are
96 * changed or hotplug happens.
98 * effective_mask == configured_mask & parent's effective_mask,
99 * and if it ends up empty, it will inherit the parent's mask.
102 * On legacy hierarchy:
104 * The user-configured masks are always the same with effective masks.
107 /* user-configured CPUs and Memory Nodes allow to tasks */
108 cpumask_var_t cpus_allowed;
109 nodemask_t mems_allowed;
111 /* effective CPUs and Memory Nodes allow to tasks */
112 cpumask_var_t effective_cpus;
113 nodemask_t effective_mems;
116 * CPUs allocated to child sub-partitions (default hierarchy only)
117 * - CPUs granted by the parent = effective_cpus U subparts_cpus
118 * - effective_cpus and subparts_cpus are mutually exclusive.
120 * effective_cpus contains only onlined CPUs, but subparts_cpus
121 * may have offlined ones.
123 cpumask_var_t subparts_cpus;
126 * This is old Memory Nodes tasks took on.
128 * - top_cpuset.old_mems_allowed is initialized to mems_allowed.
129 * - A new cpuset's old_mems_allowed is initialized when some
130 * task is moved into it.
131 * - old_mems_allowed is used in cpuset_migrate_mm() when we change
132 * cpuset.mems_allowed and have tasks' nodemask updated, and
133 * then old_mems_allowed is updated to mems_allowed.
135 nodemask_t old_mems_allowed;
137 struct fmeter fmeter; /* memory_pressure filter */
140 * Tasks are being attached to this cpuset. Used to prevent
141 * zeroing cpus/mems_allowed between ->can_attach() and ->attach().
143 int attach_in_progress;
145 /* partition number for rebuild_sched_domains() */
148 /* for custom sched domain */
149 int relax_domain_level;
151 /* number of CPUs in subparts_cpus */
152 int nr_subparts_cpus;
154 /* partition root state */
155 int partition_root_state;
158 * Default hierarchy only:
159 * use_parent_ecpus - set if using parent's effective_cpus
160 * child_ecpus_count - # of children with use_parent_ecpus set
162 int use_parent_ecpus;
163 int child_ecpus_count;
165 /* Handle for cpuset.cpus.partition */
166 struct cgroup_file partition_file;
170 * Partition root states:
172 * 0 - not a partition root
176 * -1 - invalid partition root
177 * None of the cpus in cpus_allowed can be put into the parent's
178 * subparts_cpus. In this case, the cpuset is not a real partition
179 * root anymore. However, the CPU_EXCLUSIVE bit will still be set
180 * and the cpuset can be restored back to a partition root if the
181 * parent cpuset can give more CPUs back to this child cpuset.
183 #define PRS_DISABLED 0
184 #define PRS_ENABLED 1
188 * Temporary cpumasks for working with partitions that are passed among
189 * functions to avoid memory allocation in inner functions.
192 cpumask_var_t addmask, delmask; /* For partition root */
193 cpumask_var_t new_cpus; /* For update_cpumasks_hier() */
196 static inline struct cpuset *css_cs(struct cgroup_subsys_state *css)
198 return css ? container_of(css, struct cpuset, css) : NULL;
201 /* Retrieve the cpuset for a task */
202 static inline struct cpuset *task_cs(struct task_struct *task)
204 return css_cs(task_css(task, cpuset_cgrp_id));
207 static inline struct cpuset *parent_cs(struct cpuset *cs)
209 return css_cs(cs->css.parent);
212 /* bits in struct cpuset flags field */
219 CS_SCHED_LOAD_BALANCE,
224 /* convenient tests for these bits */
225 static inline bool is_cpuset_online(struct cpuset *cs)
227 return test_bit(CS_ONLINE, &cs->flags) && !css_is_dying(&cs->css);
230 static inline int is_cpu_exclusive(const struct cpuset *cs)
232 return test_bit(CS_CPU_EXCLUSIVE, &cs->flags);
235 static inline int is_mem_exclusive(const struct cpuset *cs)
237 return test_bit(CS_MEM_EXCLUSIVE, &cs->flags);
240 static inline int is_mem_hardwall(const struct cpuset *cs)
242 return test_bit(CS_MEM_HARDWALL, &cs->flags);
245 static inline int is_sched_load_balance(const struct cpuset *cs)
247 return test_bit(CS_SCHED_LOAD_BALANCE, &cs->flags);
250 static inline int is_memory_migrate(const struct cpuset *cs)
252 return test_bit(CS_MEMORY_MIGRATE, &cs->flags);
255 static inline int is_spread_page(const struct cpuset *cs)
257 return test_bit(CS_SPREAD_PAGE, &cs->flags);
260 static inline int is_spread_slab(const struct cpuset *cs)
262 return test_bit(CS_SPREAD_SLAB, &cs->flags);
265 static inline int is_partition_root(const struct cpuset *cs)
267 return cs->partition_root_state > 0;
271 * Send notification event of whenever partition_root_state changes.
273 static inline void notify_partition_change(struct cpuset *cs,
274 int old_prs, int new_prs)
276 if (old_prs != new_prs)
277 cgroup_file_notify(&cs->partition_file);
280 static struct cpuset top_cpuset = {
281 .flags = ((1 << CS_ONLINE) | (1 << CS_CPU_EXCLUSIVE) |
282 (1 << CS_MEM_EXCLUSIVE)),
283 .partition_root_state = PRS_ENABLED,
287 * cpuset_for_each_child - traverse online children of a cpuset
288 * @child_cs: loop cursor pointing to the current child
289 * @pos_css: used for iteration
290 * @parent_cs: target cpuset to walk children of
292 * Walk @child_cs through the online children of @parent_cs. Must be used
293 * with RCU read locked.
295 #define cpuset_for_each_child(child_cs, pos_css, parent_cs) \
296 css_for_each_child((pos_css), &(parent_cs)->css) \
297 if (is_cpuset_online(((child_cs) = css_cs((pos_css)))))
300 * cpuset_for_each_descendant_pre - pre-order walk of a cpuset's descendants
301 * @des_cs: loop cursor pointing to the current descendant
302 * @pos_css: used for iteration
303 * @root_cs: target cpuset to walk ancestor of
305 * Walk @des_cs through the online descendants of @root_cs. Must be used
306 * with RCU read locked. The caller may modify @pos_css by calling
307 * css_rightmost_descendant() to skip subtree. @root_cs is included in the
308 * iteration and the first node to be visited.
310 #define cpuset_for_each_descendant_pre(des_cs, pos_css, root_cs) \
311 css_for_each_descendant_pre((pos_css), &(root_cs)->css) \
312 if (is_cpuset_online(((des_cs) = css_cs((pos_css)))))
315 * There are two global locks guarding cpuset structures - cpuset_rwsem and
316 * callback_lock. We also require taking task_lock() when dereferencing a
317 * task's cpuset pointer. See "The task_lock() exception", at the end of this
318 * comment. The cpuset code uses only cpuset_rwsem write lock. Other
319 * kernel subsystems can use cpuset_read_lock()/cpuset_read_unlock() to
320 * prevent change to cpuset structures.
322 * A task must hold both locks to modify cpusets. If a task holds
323 * cpuset_rwsem, it blocks others wanting that rwsem, ensuring that it
324 * is the only task able to also acquire callback_lock and be able to
325 * modify cpusets. It can perform various checks on the cpuset structure
326 * first, knowing nothing will change. It can also allocate memory while
327 * just holding cpuset_rwsem. While it is performing these checks, various
328 * callback routines can briefly acquire callback_lock to query cpusets.
329 * Once it is ready to make the changes, it takes callback_lock, blocking
332 * Calls to the kernel memory allocator can not be made while holding
333 * callback_lock, as that would risk double tripping on callback_lock
334 * from one of the callbacks into the cpuset code from within
337 * If a task is only holding callback_lock, then it has read-only
340 * Now, the task_struct fields mems_allowed and mempolicy may be changed
341 * by other task, we use alloc_lock in the task_struct fields to protect
344 * The cpuset_common_file_read() handlers only hold callback_lock across
345 * small pieces of code, such as when reading out possibly multi-word
346 * cpumasks and nodemasks.
348 * Accessing a task's cpuset should be done in accordance with the
349 * guidelines for accessing subsystem state in kernel/cgroup.c
352 DEFINE_STATIC_PERCPU_RWSEM(cpuset_rwsem);
354 void cpuset_read_lock(void)
356 percpu_down_read(&cpuset_rwsem);
359 void cpuset_read_unlock(void)
361 percpu_up_read(&cpuset_rwsem);
364 static DEFINE_SPINLOCK(callback_lock);
366 static struct workqueue_struct *cpuset_migrate_mm_wq;
369 * CPU / memory hotplug is handled asynchronously.
371 static void cpuset_hotplug_workfn(struct work_struct *work);
372 static DECLARE_WORK(cpuset_hotplug_work, cpuset_hotplug_workfn);
374 static DECLARE_WAIT_QUEUE_HEAD(cpuset_attach_wq);
377 * Cgroup v2 behavior is used on the "cpus" and "mems" control files when
378 * on default hierarchy or when the cpuset_v2_mode flag is set by mounting
379 * the v1 cpuset cgroup filesystem with the "cpuset_v2_mode" mount option.
380 * With v2 behavior, "cpus" and "mems" are always what the users have
381 * requested and won't be changed by hotplug events. Only the effective
382 * cpus or mems will be affected.
384 static inline bool is_in_v2_mode(void)
386 return cgroup_subsys_on_dfl(cpuset_cgrp_subsys) ||
387 (cpuset_cgrp_subsys.root->flags & CGRP_ROOT_CPUSET_V2_MODE);
391 * Return in pmask the portion of a task's cpusets's cpus_allowed that
392 * are online and are capable of running the task. If none are found,
393 * walk up the cpuset hierarchy until we find one that does have some
396 * One way or another, we guarantee to return some non-empty subset
397 * of cpu_online_mask.
399 * Call with callback_lock or cpuset_rwsem held.
401 static void guarantee_online_cpus(struct task_struct *tsk,
402 struct cpumask *pmask)
404 const struct cpumask *possible_mask = task_cpu_possible_mask(tsk);
407 if (WARN_ON(!cpumask_and(pmask, possible_mask, cpu_online_mask)))
408 cpumask_copy(pmask, cpu_online_mask);
413 while (!cpumask_intersects(cs->effective_cpus, pmask)) {
417 * The top cpuset doesn't have any online cpu as a
418 * consequence of a race between cpuset_hotplug_work
419 * and cpu hotplug notifier. But we know the top
420 * cpuset's effective_cpus is on its way to be
421 * identical to cpu_online_mask.
426 cpumask_and(pmask, pmask, cs->effective_cpus);
433 * Return in *pmask the portion of a cpusets's mems_allowed that
434 * are online, with memory. If none are online with memory, walk
435 * up the cpuset hierarchy until we find one that does have some
436 * online mems. The top cpuset always has some mems online.
438 * One way or another, we guarantee to return some non-empty subset
439 * of node_states[N_MEMORY].
441 * Call with callback_lock or cpuset_rwsem held.
443 static void guarantee_online_mems(struct cpuset *cs, nodemask_t *pmask)
445 while (!nodes_intersects(cs->effective_mems, node_states[N_MEMORY]))
447 nodes_and(*pmask, cs->effective_mems, node_states[N_MEMORY]);
451 * update task's spread flag if cpuset's page/slab spread flag is set
453 * Call with callback_lock or cpuset_rwsem held.
455 static void cpuset_update_task_spread_flag(struct cpuset *cs,
456 struct task_struct *tsk)
458 if (is_spread_page(cs))
459 task_set_spread_page(tsk);
461 task_clear_spread_page(tsk);
463 if (is_spread_slab(cs))
464 task_set_spread_slab(tsk);
466 task_clear_spread_slab(tsk);
470 * is_cpuset_subset(p, q) - Is cpuset p a subset of cpuset q?
472 * One cpuset is a subset of another if all its allowed CPUs and
473 * Memory Nodes are a subset of the other, and its exclusive flags
474 * are only set if the other's are set. Call holding cpuset_rwsem.
477 static int is_cpuset_subset(const struct cpuset *p, const struct cpuset *q)
479 return cpumask_subset(p->cpus_allowed, q->cpus_allowed) &&
480 nodes_subset(p->mems_allowed, q->mems_allowed) &&
481 is_cpu_exclusive(p) <= is_cpu_exclusive(q) &&
482 is_mem_exclusive(p) <= is_mem_exclusive(q);
486 * alloc_cpumasks - allocate three cpumasks for cpuset
487 * @cs: the cpuset that have cpumasks to be allocated.
488 * @tmp: the tmpmasks structure pointer
489 * Return: 0 if successful, -ENOMEM otherwise.
491 * Only one of the two input arguments should be non-NULL.
493 static inline int alloc_cpumasks(struct cpuset *cs, struct tmpmasks *tmp)
495 cpumask_var_t *pmask1, *pmask2, *pmask3;
498 pmask1 = &cs->cpus_allowed;
499 pmask2 = &cs->effective_cpus;
500 pmask3 = &cs->subparts_cpus;
502 pmask1 = &tmp->new_cpus;
503 pmask2 = &tmp->addmask;
504 pmask3 = &tmp->delmask;
507 if (!zalloc_cpumask_var(pmask1, GFP_KERNEL))
510 if (!zalloc_cpumask_var(pmask2, GFP_KERNEL))
513 if (!zalloc_cpumask_var(pmask3, GFP_KERNEL))
519 free_cpumask_var(*pmask2);
521 free_cpumask_var(*pmask1);
526 * free_cpumasks - free cpumasks in a tmpmasks structure
527 * @cs: the cpuset that have cpumasks to be free.
528 * @tmp: the tmpmasks structure pointer
530 static inline void free_cpumasks(struct cpuset *cs, struct tmpmasks *tmp)
533 free_cpumask_var(cs->cpus_allowed);
534 free_cpumask_var(cs->effective_cpus);
535 free_cpumask_var(cs->subparts_cpus);
538 free_cpumask_var(tmp->new_cpus);
539 free_cpumask_var(tmp->addmask);
540 free_cpumask_var(tmp->delmask);
545 * alloc_trial_cpuset - allocate a trial cpuset
546 * @cs: the cpuset that the trial cpuset duplicates
548 static struct cpuset *alloc_trial_cpuset(struct cpuset *cs)
550 struct cpuset *trial;
552 trial = kmemdup(cs, sizeof(*cs), GFP_KERNEL);
556 if (alloc_cpumasks(trial, NULL)) {
561 cpumask_copy(trial->cpus_allowed, cs->cpus_allowed);
562 cpumask_copy(trial->effective_cpus, cs->effective_cpus);
567 * free_cpuset - free the cpuset
568 * @cs: the cpuset to be freed
570 static inline void free_cpuset(struct cpuset *cs)
572 free_cpumasks(cs, NULL);
577 * validate_change() - Used to validate that any proposed cpuset change
578 * follows the structural rules for cpusets.
580 * If we replaced the flag and mask values of the current cpuset
581 * (cur) with those values in the trial cpuset (trial), would
582 * our various subset and exclusive rules still be valid? Presumes
585 * 'cur' is the address of an actual, in-use cpuset. Operations
586 * such as list traversal that depend on the actual address of the
587 * cpuset in the list must use cur below, not trial.
589 * 'trial' is the address of bulk structure copy of cur, with
590 * perhaps one or more of the fields cpus_allowed, mems_allowed,
591 * or flags changed to new, trial values.
593 * Return 0 if valid, -errno if not.
596 static int validate_change(struct cpuset *cur, struct cpuset *trial)
598 struct cgroup_subsys_state *css;
599 struct cpuset *c, *par;
604 /* Each of our child cpusets must be a subset of us */
606 cpuset_for_each_child(c, css, cur)
607 if (!is_cpuset_subset(c, trial))
610 /* Remaining checks don't apply to root cpuset */
612 if (cur == &top_cpuset)
615 par = parent_cs(cur);
617 /* On legacy hierarchy, we must be a subset of our parent cpuset. */
619 if (!is_in_v2_mode() && !is_cpuset_subset(trial, par))
623 * If either I or some sibling (!= me) is exclusive, we can't
627 cpuset_for_each_child(c, css, par) {
628 if ((is_cpu_exclusive(trial) || is_cpu_exclusive(c)) &&
630 cpumask_intersects(trial->cpus_allowed, c->cpus_allowed))
632 if ((is_mem_exclusive(trial) || is_mem_exclusive(c)) &&
634 nodes_intersects(trial->mems_allowed, c->mems_allowed))
639 * Cpusets with tasks - existing or newly being attached - can't
640 * be changed to have empty cpus_allowed or mems_allowed.
643 if ((cgroup_is_populated(cur->css.cgroup) || cur->attach_in_progress)) {
644 if (!cpumask_empty(cur->cpus_allowed) &&
645 cpumask_empty(trial->cpus_allowed))
647 if (!nodes_empty(cur->mems_allowed) &&
648 nodes_empty(trial->mems_allowed))
653 * We can't shrink if we won't have enough room for SCHED_DEADLINE
657 if (is_cpu_exclusive(cur) &&
658 !cpuset_cpumask_can_shrink(cur->cpus_allowed,
659 trial->cpus_allowed))
670 * Helper routine for generate_sched_domains().
671 * Do cpusets a, b have overlapping effective cpus_allowed masks?
673 static int cpusets_overlap(struct cpuset *a, struct cpuset *b)
675 return cpumask_intersects(a->effective_cpus, b->effective_cpus);
679 update_domain_attr(struct sched_domain_attr *dattr, struct cpuset *c)
681 if (dattr->relax_domain_level < c->relax_domain_level)
682 dattr->relax_domain_level = c->relax_domain_level;
686 static void update_domain_attr_tree(struct sched_domain_attr *dattr,
687 struct cpuset *root_cs)
690 struct cgroup_subsys_state *pos_css;
693 cpuset_for_each_descendant_pre(cp, pos_css, root_cs) {
694 /* skip the whole subtree if @cp doesn't have any CPU */
695 if (cpumask_empty(cp->cpus_allowed)) {
696 pos_css = css_rightmost_descendant(pos_css);
700 if (is_sched_load_balance(cp))
701 update_domain_attr(dattr, cp);
706 /* Must be called with cpuset_rwsem held. */
707 static inline int nr_cpusets(void)
709 /* jump label reference count + the top-level cpuset */
710 return static_key_count(&cpusets_enabled_key.key) + 1;
714 * generate_sched_domains()
716 * This function builds a partial partition of the systems CPUs
717 * A 'partial partition' is a set of non-overlapping subsets whose
718 * union is a subset of that set.
719 * The output of this function needs to be passed to kernel/sched/core.c
720 * partition_sched_domains() routine, which will rebuild the scheduler's
721 * load balancing domains (sched domains) as specified by that partial
724 * See "What is sched_load_balance" in Documentation/admin-guide/cgroup-v1/cpusets.rst
725 * for a background explanation of this.
727 * Does not return errors, on the theory that the callers of this
728 * routine would rather not worry about failures to rebuild sched
729 * domains when operating in the severe memory shortage situations
730 * that could cause allocation failures below.
732 * Must be called with cpuset_rwsem held.
734 * The three key local variables below are:
735 * cp - cpuset pointer, used (together with pos_css) to perform a
736 * top-down scan of all cpusets. For our purposes, rebuilding
737 * the schedulers sched domains, we can ignore !is_sched_load_
739 * csa - (for CpuSet Array) Array of pointers to all the cpusets
740 * that need to be load balanced, for convenient iterative
741 * access by the subsequent code that finds the best partition,
742 * i.e the set of domains (subsets) of CPUs such that the
743 * cpus_allowed of every cpuset marked is_sched_load_balance
744 * is a subset of one of these domains, while there are as
745 * many such domains as possible, each as small as possible.
746 * doms - Conversion of 'csa' to an array of cpumasks, for passing to
747 * the kernel/sched/core.c routine partition_sched_domains() in a
748 * convenient format, that can be easily compared to the prior
749 * value to determine what partition elements (sched domains)
750 * were changed (added or removed.)
752 * Finding the best partition (set of domains):
753 * The triple nested loops below over i, j, k scan over the
754 * load balanced cpusets (using the array of cpuset pointers in
755 * csa[]) looking for pairs of cpusets that have overlapping
756 * cpus_allowed, but which don't have the same 'pn' partition
757 * number and gives them in the same partition number. It keeps
758 * looping on the 'restart' label until it can no longer find
761 * The union of the cpus_allowed masks from the set of
762 * all cpusets having the same 'pn' value then form the one
763 * element of the partition (one sched domain) to be passed to
764 * partition_sched_domains().
766 static int generate_sched_domains(cpumask_var_t **domains,
767 struct sched_domain_attr **attributes)
769 struct cpuset *cp; /* top-down scan of cpusets */
770 struct cpuset **csa; /* array of all cpuset ptrs */
771 int csn; /* how many cpuset ptrs in csa so far */
772 int i, j, k; /* indices for partition finding loops */
773 cpumask_var_t *doms; /* resulting partition; i.e. sched domains */
774 struct sched_domain_attr *dattr; /* attributes for custom domains */
775 int ndoms = 0; /* number of sched domains in result */
776 int nslot; /* next empty doms[] struct cpumask slot */
777 struct cgroup_subsys_state *pos_css;
778 bool root_load_balance = is_sched_load_balance(&top_cpuset);
784 /* Special case for the 99% of systems with one, full, sched domain */
785 if (root_load_balance && !top_cpuset.nr_subparts_cpus) {
787 doms = alloc_sched_domains(ndoms);
791 dattr = kmalloc(sizeof(struct sched_domain_attr), GFP_KERNEL);
793 *dattr = SD_ATTR_INIT;
794 update_domain_attr_tree(dattr, &top_cpuset);
796 cpumask_and(doms[0], top_cpuset.effective_cpus,
797 housekeeping_cpumask(HK_FLAG_DOMAIN));
802 csa = kmalloc_array(nr_cpusets(), sizeof(cp), GFP_KERNEL);
808 if (root_load_balance)
809 csa[csn++] = &top_cpuset;
810 cpuset_for_each_descendant_pre(cp, pos_css, &top_cpuset) {
811 if (cp == &top_cpuset)
814 * Continue traversing beyond @cp iff @cp has some CPUs and
815 * isn't load balancing. The former is obvious. The
816 * latter: All child cpusets contain a subset of the
817 * parent's cpus, so just skip them, and then we call
818 * update_domain_attr_tree() to calc relax_domain_level of
819 * the corresponding sched domain.
821 * If root is load-balancing, we can skip @cp if it
822 * is a subset of the root's effective_cpus.
824 if (!cpumask_empty(cp->cpus_allowed) &&
825 !(is_sched_load_balance(cp) &&
826 cpumask_intersects(cp->cpus_allowed,
827 housekeeping_cpumask(HK_FLAG_DOMAIN))))
830 if (root_load_balance &&
831 cpumask_subset(cp->cpus_allowed, top_cpuset.effective_cpus))
834 if (is_sched_load_balance(cp) &&
835 !cpumask_empty(cp->effective_cpus))
838 /* skip @cp's subtree if not a partition root */
839 if (!is_partition_root(cp))
840 pos_css = css_rightmost_descendant(pos_css);
844 for (i = 0; i < csn; i++)
849 /* Find the best partition (set of sched domains) */
850 for (i = 0; i < csn; i++) {
851 struct cpuset *a = csa[i];
854 for (j = 0; j < csn; j++) {
855 struct cpuset *b = csa[j];
858 if (apn != bpn && cpusets_overlap(a, b)) {
859 for (k = 0; k < csn; k++) {
860 struct cpuset *c = csa[k];
865 ndoms--; /* one less element */
872 * Now we know how many domains to create.
873 * Convert <csn, csa> to <ndoms, doms> and populate cpu masks.
875 doms = alloc_sched_domains(ndoms);
880 * The rest of the code, including the scheduler, can deal with
881 * dattr==NULL case. No need to abort if alloc fails.
883 dattr = kmalloc_array(ndoms, sizeof(struct sched_domain_attr),
886 for (nslot = 0, i = 0; i < csn; i++) {
887 struct cpuset *a = csa[i];
892 /* Skip completed partitions */
898 if (nslot == ndoms) {
899 static int warnings = 10;
901 pr_warn("rebuild_sched_domains confused: nslot %d, ndoms %d, csn %d, i %d, apn %d\n",
902 nslot, ndoms, csn, i, apn);
910 *(dattr + nslot) = SD_ATTR_INIT;
911 for (j = i; j < csn; j++) {
912 struct cpuset *b = csa[j];
915 cpumask_or(dp, dp, b->effective_cpus);
916 cpumask_and(dp, dp, housekeeping_cpumask(HK_FLAG_DOMAIN));
918 update_domain_attr_tree(dattr + nslot, b);
920 /* Done with this partition */
926 BUG_ON(nslot != ndoms);
932 * Fallback to the default domain if kmalloc() failed.
933 * See comments in partition_sched_domains().
943 static void update_tasks_root_domain(struct cpuset *cs)
945 struct css_task_iter it;
946 struct task_struct *task;
948 css_task_iter_start(&cs->css, 0, &it);
950 while ((task = css_task_iter_next(&it)))
951 dl_add_task_root_domain(task);
953 css_task_iter_end(&it);
956 static void rebuild_root_domains(void)
958 struct cpuset *cs = NULL;
959 struct cgroup_subsys_state *pos_css;
961 percpu_rwsem_assert_held(&cpuset_rwsem);
962 lockdep_assert_cpus_held();
963 lockdep_assert_held(&sched_domains_mutex);
968 * Clear default root domain DL accounting, it will be computed again
969 * if a task belongs to it.
971 dl_clear_root_domain(&def_root_domain);
973 cpuset_for_each_descendant_pre(cs, pos_css, &top_cpuset) {
975 if (cpumask_empty(cs->effective_cpus)) {
976 pos_css = css_rightmost_descendant(pos_css);
984 update_tasks_root_domain(cs);
993 partition_and_rebuild_sched_domains(int ndoms_new, cpumask_var_t doms_new[],
994 struct sched_domain_attr *dattr_new)
996 mutex_lock(&sched_domains_mutex);
997 partition_sched_domains_locked(ndoms_new, doms_new, dattr_new);
998 rebuild_root_domains();
999 mutex_unlock(&sched_domains_mutex);
1003 * Rebuild scheduler domains.
1005 * If the flag 'sched_load_balance' of any cpuset with non-empty
1006 * 'cpus' changes, or if the 'cpus' allowed changes in any cpuset
1007 * which has that flag enabled, or if any cpuset with a non-empty
1008 * 'cpus' is removed, then call this routine to rebuild the
1009 * scheduler's dynamic sched domains.
1011 * Call with cpuset_rwsem held. Takes cpus_read_lock().
1013 static void rebuild_sched_domains_locked(void)
1015 struct cgroup_subsys_state *pos_css;
1016 struct sched_domain_attr *attr;
1017 cpumask_var_t *doms;
1021 lockdep_assert_cpus_held();
1022 percpu_rwsem_assert_held(&cpuset_rwsem);
1025 * If we have raced with CPU hotplug, return early to avoid
1026 * passing doms with offlined cpu to partition_sched_domains().
1027 * Anyways, cpuset_hotplug_workfn() will rebuild sched domains.
1029 * With no CPUs in any subpartitions, top_cpuset's effective CPUs
1030 * should be the same as the active CPUs, so checking only top_cpuset
1031 * is enough to detect racing CPU offlines.
1033 if (!top_cpuset.nr_subparts_cpus &&
1034 !cpumask_equal(top_cpuset.effective_cpus, cpu_active_mask))
1038 * With subpartition CPUs, however, the effective CPUs of a partition
1039 * root should be only a subset of the active CPUs. Since a CPU in any
1040 * partition root could be offlined, all must be checked.
1042 if (top_cpuset.nr_subparts_cpus) {
1044 cpuset_for_each_descendant_pre(cs, pos_css, &top_cpuset) {
1045 if (!is_partition_root(cs)) {
1046 pos_css = css_rightmost_descendant(pos_css);
1049 if (!cpumask_subset(cs->effective_cpus,
1058 /* Generate domain masks and attrs */
1059 ndoms = generate_sched_domains(&doms, &attr);
1061 /* Have scheduler rebuild the domains */
1062 partition_and_rebuild_sched_domains(ndoms, doms, attr);
1064 #else /* !CONFIG_SMP */
1065 static void rebuild_sched_domains_locked(void)
1068 #endif /* CONFIG_SMP */
1070 void rebuild_sched_domains(void)
1073 percpu_down_write(&cpuset_rwsem);
1074 rebuild_sched_domains_locked();
1075 percpu_up_write(&cpuset_rwsem);
1080 * update_tasks_cpumask - Update the cpumasks of tasks in the cpuset.
1081 * @cs: the cpuset in which each task's cpus_allowed mask needs to be changed
1083 * Iterate through each task of @cs updating its cpus_allowed to the
1084 * effective cpuset's. As this function is called with cpuset_rwsem held,
1085 * cpuset membership stays stable.
1087 static void update_tasks_cpumask(struct cpuset *cs)
1089 struct css_task_iter it;
1090 struct task_struct *task;
1091 bool top_cs = cs == &top_cpuset;
1093 css_task_iter_start(&cs->css, 0, &it);
1094 while ((task = css_task_iter_next(&it))) {
1096 * Percpu kthreads in top_cpuset are ignored
1098 if (top_cs && (task->flags & PF_KTHREAD) &&
1099 kthread_is_per_cpu(task))
1101 set_cpus_allowed_ptr(task, cs->effective_cpus);
1103 css_task_iter_end(&it);
1107 * compute_effective_cpumask - Compute the effective cpumask of the cpuset
1108 * @new_cpus: the temp variable for the new effective_cpus mask
1109 * @cs: the cpuset the need to recompute the new effective_cpus mask
1110 * @parent: the parent cpuset
1112 * If the parent has subpartition CPUs, include them in the list of
1113 * allowable CPUs in computing the new effective_cpus mask. Since offlined
1114 * CPUs are not removed from subparts_cpus, we have to use cpu_active_mask
1115 * to mask those out.
1117 static void compute_effective_cpumask(struct cpumask *new_cpus,
1118 struct cpuset *cs, struct cpuset *parent)
1120 if (parent->nr_subparts_cpus) {
1121 cpumask_or(new_cpus, parent->effective_cpus,
1122 parent->subparts_cpus);
1123 cpumask_and(new_cpus, new_cpus, cs->cpus_allowed);
1124 cpumask_and(new_cpus, new_cpus, cpu_active_mask);
1126 cpumask_and(new_cpus, cs->cpus_allowed, parent->effective_cpus);
1131 * Commands for update_parent_subparts_cpumask
1134 partcmd_enable, /* Enable partition root */
1135 partcmd_disable, /* Disable partition root */
1136 partcmd_update, /* Update parent's subparts_cpus */
1140 * update_parent_subparts_cpumask - update subparts_cpus mask of parent cpuset
1141 * @cpuset: The cpuset that requests change in partition root state
1142 * @cmd: Partition root state change command
1143 * @newmask: Optional new cpumask for partcmd_update
1144 * @tmp: Temporary addmask and delmask
1145 * Return: 0, 1 or an error code
1147 * For partcmd_enable, the cpuset is being transformed from a non-partition
1148 * root to a partition root. The cpus_allowed mask of the given cpuset will
1149 * be put into parent's subparts_cpus and taken away from parent's
1150 * effective_cpus. The function will return 0 if all the CPUs listed in
1151 * cpus_allowed can be granted or an error code will be returned.
1153 * For partcmd_disable, the cpuset is being transofrmed from a partition
1154 * root back to a non-partition root. Any CPUs in cpus_allowed that are in
1155 * parent's subparts_cpus will be taken away from that cpumask and put back
1156 * into parent's effective_cpus. 0 should always be returned.
1158 * For partcmd_update, if the optional newmask is specified, the cpu
1159 * list is to be changed from cpus_allowed to newmask. Otherwise,
1160 * cpus_allowed is assumed to remain the same. The cpuset should either
1161 * be a partition root or an invalid partition root. The partition root
1162 * state may change if newmask is NULL and none of the requested CPUs can
1163 * be granted by the parent. The function will return 1 if changes to
1164 * parent's subparts_cpus and effective_cpus happen or 0 otherwise.
1165 * Error code should only be returned when newmask is non-NULL.
1167 * The partcmd_enable and partcmd_disable commands are used by
1168 * update_prstate(). The partcmd_update command is used by
1169 * update_cpumasks_hier() with newmask NULL and update_cpumask() with
1172 * The checking is more strict when enabling partition root than the
1173 * other two commands.
1175 * Because of the implicit cpu exclusive nature of a partition root,
1176 * cpumask changes that violates the cpu exclusivity rule will not be
1177 * permitted when checked by validate_change(). The validate_change()
1178 * function will also prevent any changes to the cpu list if it is not
1179 * a superset of children's cpu lists.
1181 static int update_parent_subparts_cpumask(struct cpuset *cpuset, int cmd,
1182 struct cpumask *newmask,
1183 struct tmpmasks *tmp)
1185 struct cpuset *parent = parent_cs(cpuset);
1186 int adding; /* Moving cpus from effective_cpus to subparts_cpus */
1187 int deleting; /* Moving cpus from subparts_cpus to effective_cpus */
1188 int old_prs, new_prs;
1189 bool part_error = false; /* Partition error? */
1191 percpu_rwsem_assert_held(&cpuset_rwsem);
1194 * The parent must be a partition root.
1195 * The new cpumask, if present, or the current cpus_allowed must
1198 if (!is_partition_root(parent) ||
1199 (newmask && cpumask_empty(newmask)) ||
1200 (!newmask && cpumask_empty(cpuset->cpus_allowed)))
1204 * Enabling/disabling partition root is not allowed if there are
1207 if ((cmd != partcmd_update) && css_has_online_children(&cpuset->css))
1211 * Enabling partition root is not allowed if not all the CPUs
1212 * can be granted from parent's effective_cpus or at least one
1213 * CPU will be left after that.
1215 if ((cmd == partcmd_enable) &&
1216 (!cpumask_subset(cpuset->cpus_allowed, parent->effective_cpus) ||
1217 cpumask_equal(cpuset->cpus_allowed, parent->effective_cpus)))
1221 * A cpumask update cannot make parent's effective_cpus become empty.
1223 adding = deleting = false;
1224 old_prs = new_prs = cpuset->partition_root_state;
1225 if (cmd == partcmd_enable) {
1226 cpumask_copy(tmp->addmask, cpuset->cpus_allowed);
1228 } else if (cmd == partcmd_disable) {
1229 deleting = cpumask_and(tmp->delmask, cpuset->cpus_allowed,
1230 parent->subparts_cpus);
1231 } else if (newmask) {
1233 * partcmd_update with newmask:
1235 * delmask = cpus_allowed & ~newmask & parent->subparts_cpus
1236 * addmask = newmask & parent->effective_cpus
1237 * & ~parent->subparts_cpus
1239 cpumask_andnot(tmp->delmask, cpuset->cpus_allowed, newmask);
1240 deleting = cpumask_and(tmp->delmask, tmp->delmask,
1241 parent->subparts_cpus);
1243 cpumask_and(tmp->addmask, newmask, parent->effective_cpus);
1244 adding = cpumask_andnot(tmp->addmask, tmp->addmask,
1245 parent->subparts_cpus);
1247 * Return error if the new effective_cpus could become empty.
1250 cpumask_equal(parent->effective_cpus, tmp->addmask)) {
1254 * As some of the CPUs in subparts_cpus might have
1255 * been offlined, we need to compute the real delmask
1258 if (!cpumask_and(tmp->addmask, tmp->delmask,
1261 cpumask_copy(tmp->addmask, parent->effective_cpus);
1265 * partcmd_update w/o newmask:
1267 * addmask = cpus_allowed & parent->effective_cpus
1269 * Note that parent's subparts_cpus may have been
1270 * pre-shrunk in case there is a change in the cpu list.
1271 * So no deletion is needed.
1273 adding = cpumask_and(tmp->addmask, cpuset->cpus_allowed,
1274 parent->effective_cpus);
1275 part_error = cpumask_equal(tmp->addmask,
1276 parent->effective_cpus);
1279 if (cmd == partcmd_update) {
1280 int prev_prs = cpuset->partition_root_state;
1283 * Check for possible transition between PRS_ENABLED
1286 switch (cpuset->partition_root_state) {
1289 new_prs = PRS_ERROR;
1293 new_prs = PRS_ENABLED;
1297 * Set part_error if previously in invalid state.
1299 part_error = (prev_prs == PRS_ERROR);
1302 if (!part_error && (new_prs == PRS_ERROR))
1303 return 0; /* Nothing need to be done */
1305 if (new_prs == PRS_ERROR) {
1307 * Remove all its cpus from parent's subparts_cpus.
1310 deleting = cpumask_and(tmp->delmask, cpuset->cpus_allowed,
1311 parent->subparts_cpus);
1314 if (!adding && !deleting && (new_prs == old_prs))
1318 * Change the parent's subparts_cpus.
1319 * Newly added CPUs will be removed from effective_cpus and
1320 * newly deleted ones will be added back to effective_cpus.
1322 spin_lock_irq(&callback_lock);
1324 cpumask_or(parent->subparts_cpus,
1325 parent->subparts_cpus, tmp->addmask);
1326 cpumask_andnot(parent->effective_cpus,
1327 parent->effective_cpus, tmp->addmask);
1330 cpumask_andnot(parent->subparts_cpus,
1331 parent->subparts_cpus, tmp->delmask);
1333 * Some of the CPUs in subparts_cpus might have been offlined.
1335 cpumask_and(tmp->delmask, tmp->delmask, cpu_active_mask);
1336 cpumask_or(parent->effective_cpus,
1337 parent->effective_cpus, tmp->delmask);
1340 parent->nr_subparts_cpus = cpumask_weight(parent->subparts_cpus);
1342 if (old_prs != new_prs)
1343 cpuset->partition_root_state = new_prs;
1345 spin_unlock_irq(&callback_lock);
1346 notify_partition_change(cpuset, old_prs, new_prs);
1348 return cmd == partcmd_update;
1352 * update_cpumasks_hier - Update effective cpumasks and tasks in the subtree
1353 * @cs: the cpuset to consider
1354 * @tmp: temp variables for calculating effective_cpus & partition setup
1356 * When configured cpumask is changed, the effective cpumasks of this cpuset
1357 * and all its descendants need to be updated.
1359 * On legacy hierarchy, effective_cpus will be the same with cpu_allowed.
1361 * Called with cpuset_rwsem held
1363 static void update_cpumasks_hier(struct cpuset *cs, struct tmpmasks *tmp)
1366 struct cgroup_subsys_state *pos_css;
1367 bool need_rebuild_sched_domains = false;
1368 int old_prs, new_prs;
1371 cpuset_for_each_descendant_pre(cp, pos_css, cs) {
1372 struct cpuset *parent = parent_cs(cp);
1374 compute_effective_cpumask(tmp->new_cpus, cp, parent);
1377 * If it becomes empty, inherit the effective mask of the
1378 * parent, which is guaranteed to have some CPUs.
1380 if (is_in_v2_mode() && cpumask_empty(tmp->new_cpus)) {
1381 cpumask_copy(tmp->new_cpus, parent->effective_cpus);
1382 if (!cp->use_parent_ecpus) {
1383 cp->use_parent_ecpus = true;
1384 parent->child_ecpus_count++;
1386 } else if (cp->use_parent_ecpus) {
1387 cp->use_parent_ecpus = false;
1388 WARN_ON_ONCE(!parent->child_ecpus_count);
1389 parent->child_ecpus_count--;
1393 * Skip the whole subtree if the cpumask remains the same
1394 * and has no partition root state.
1396 if (!cp->partition_root_state &&
1397 cpumask_equal(tmp->new_cpus, cp->effective_cpus)) {
1398 pos_css = css_rightmost_descendant(pos_css);
1403 * update_parent_subparts_cpumask() should have been called
1404 * for cs already in update_cpumask(). We should also call
1405 * update_tasks_cpumask() again for tasks in the parent
1406 * cpuset if the parent's subparts_cpus changes.
1408 old_prs = new_prs = cp->partition_root_state;
1409 if ((cp != cs) && old_prs) {
1410 switch (parent->partition_root_state) {
1413 * If parent is not a partition root or an
1414 * invalid partition root, clear its state
1415 * and its CS_CPU_EXCLUSIVE flag.
1417 WARN_ON_ONCE(cp->partition_root_state
1419 new_prs = PRS_DISABLED;
1422 * clear_bit() is an atomic operation and
1423 * readers aren't interested in the state
1424 * of CS_CPU_EXCLUSIVE anyway. So we can
1425 * just update the flag without holding
1426 * the callback_lock.
1428 clear_bit(CS_CPU_EXCLUSIVE, &cp->flags);
1432 if (update_parent_subparts_cpumask(cp, partcmd_update, NULL, tmp))
1433 update_tasks_cpumask(parent);
1438 * When parent is invalid, it has to be too.
1440 new_prs = PRS_ERROR;
1445 if (!css_tryget_online(&cp->css))
1449 spin_lock_irq(&callback_lock);
1451 cpumask_copy(cp->effective_cpus, tmp->new_cpus);
1452 if (cp->nr_subparts_cpus && (new_prs != PRS_ENABLED)) {
1453 cp->nr_subparts_cpus = 0;
1454 cpumask_clear(cp->subparts_cpus);
1455 } else if (cp->nr_subparts_cpus) {
1457 * Make sure that effective_cpus & subparts_cpus
1458 * are mutually exclusive.
1460 * In the unlikely event that effective_cpus
1461 * becomes empty. we clear cp->nr_subparts_cpus and
1462 * let its child partition roots to compete for
1465 cpumask_andnot(cp->effective_cpus, cp->effective_cpus,
1467 if (cpumask_empty(cp->effective_cpus)) {
1468 cpumask_copy(cp->effective_cpus, tmp->new_cpus);
1469 cpumask_clear(cp->subparts_cpus);
1470 cp->nr_subparts_cpus = 0;
1471 } else if (!cpumask_subset(cp->subparts_cpus,
1473 cpumask_andnot(cp->subparts_cpus,
1474 cp->subparts_cpus, tmp->new_cpus);
1475 cp->nr_subparts_cpus
1476 = cpumask_weight(cp->subparts_cpus);
1480 if (new_prs != old_prs)
1481 cp->partition_root_state = new_prs;
1483 spin_unlock_irq(&callback_lock);
1484 notify_partition_change(cp, old_prs, new_prs);
1486 WARN_ON(!is_in_v2_mode() &&
1487 !cpumask_equal(cp->cpus_allowed, cp->effective_cpus));
1489 update_tasks_cpumask(cp);
1492 * On legacy hierarchy, if the effective cpumask of any non-
1493 * empty cpuset is changed, we need to rebuild sched domains.
1494 * On default hierarchy, the cpuset needs to be a partition
1497 if (!cpumask_empty(cp->cpus_allowed) &&
1498 is_sched_load_balance(cp) &&
1499 (!cgroup_subsys_on_dfl(cpuset_cgrp_subsys) ||
1500 is_partition_root(cp)))
1501 need_rebuild_sched_domains = true;
1508 if (need_rebuild_sched_domains)
1509 rebuild_sched_domains_locked();
1513 * update_sibling_cpumasks - Update siblings cpumasks
1514 * @parent: Parent cpuset
1515 * @cs: Current cpuset
1516 * @tmp: Temp variables
1518 static void update_sibling_cpumasks(struct cpuset *parent, struct cpuset *cs,
1519 struct tmpmasks *tmp)
1521 struct cpuset *sibling;
1522 struct cgroup_subsys_state *pos_css;
1524 percpu_rwsem_assert_held(&cpuset_rwsem);
1527 * Check all its siblings and call update_cpumasks_hier()
1528 * if their use_parent_ecpus flag is set in order for them
1529 * to use the right effective_cpus value.
1531 * The update_cpumasks_hier() function may sleep. So we have to
1532 * release the RCU read lock before calling it.
1535 cpuset_for_each_child(sibling, pos_css, parent) {
1538 if (!sibling->use_parent_ecpus)
1540 if (!css_tryget_online(&sibling->css))
1544 update_cpumasks_hier(sibling, tmp);
1546 css_put(&sibling->css);
1552 * update_cpumask - update the cpus_allowed mask of a cpuset and all tasks in it
1553 * @cs: the cpuset to consider
1554 * @trialcs: trial cpuset
1555 * @buf: buffer of cpu numbers written to this cpuset
1557 static int update_cpumask(struct cpuset *cs, struct cpuset *trialcs,
1561 struct tmpmasks tmp;
1563 /* top_cpuset.cpus_allowed tracks cpu_online_mask; it's read-only */
1564 if (cs == &top_cpuset)
1568 * An empty cpus_allowed is ok only if the cpuset has no tasks.
1569 * Since cpulist_parse() fails on an empty mask, we special case
1570 * that parsing. The validate_change() call ensures that cpusets
1571 * with tasks have cpus.
1574 cpumask_clear(trialcs->cpus_allowed);
1576 retval = cpulist_parse(buf, trialcs->cpus_allowed);
1580 if (!cpumask_subset(trialcs->cpus_allowed,
1581 top_cpuset.cpus_allowed))
1585 /* Nothing to do if the cpus didn't change */
1586 if (cpumask_equal(cs->cpus_allowed, trialcs->cpus_allowed))
1589 retval = validate_change(cs, trialcs);
1593 #ifdef CONFIG_CPUMASK_OFFSTACK
1595 * Use the cpumasks in trialcs for tmpmasks when they are pointers
1596 * to allocated cpumasks.
1598 tmp.addmask = trialcs->subparts_cpus;
1599 tmp.delmask = trialcs->effective_cpus;
1600 tmp.new_cpus = trialcs->cpus_allowed;
1603 if (cs->partition_root_state) {
1604 /* Cpumask of a partition root cannot be empty */
1605 if (cpumask_empty(trialcs->cpus_allowed))
1607 if (update_parent_subparts_cpumask(cs, partcmd_update,
1608 trialcs->cpus_allowed, &tmp) < 0)
1612 spin_lock_irq(&callback_lock);
1613 cpumask_copy(cs->cpus_allowed, trialcs->cpus_allowed);
1616 * Make sure that subparts_cpus is a subset of cpus_allowed.
1618 if (cs->nr_subparts_cpus) {
1619 cpumask_and(cs->subparts_cpus, cs->subparts_cpus, cs->cpus_allowed);
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 spin_lock_irq(&callback_lock);
1892 cs->mems_allowed = trialcs->mems_allowed;
1893 spin_unlock_irq(&callback_lock);
1895 /* use trialcs->mems_allowed as a temp variable */
1896 update_nodemasks_hier(cs, &trialcs->mems_allowed);
1901 bool current_cpuset_is_being_rebound(void)
1906 ret = task_cs(current) == cpuset_being_rebound;
1912 static int update_relax_domain_level(struct cpuset *cs, s64 val)
1915 if (val < -1 || val >= sched_domain_level_max)
1919 if (val != cs->relax_domain_level) {
1920 cs->relax_domain_level = val;
1921 if (!cpumask_empty(cs->cpus_allowed) &&
1922 is_sched_load_balance(cs))
1923 rebuild_sched_domains_locked();
1930 * update_tasks_flags - update the spread flags of tasks in the cpuset.
1931 * @cs: the cpuset in which each task's spread flags needs to be changed
1933 * Iterate through each task of @cs updating its spread flags. As this
1934 * function is called with cpuset_rwsem held, cpuset membership stays
1937 static void update_tasks_flags(struct cpuset *cs)
1939 struct css_task_iter it;
1940 struct task_struct *task;
1942 css_task_iter_start(&cs->css, 0, &it);
1943 while ((task = css_task_iter_next(&it)))
1944 cpuset_update_task_spread_flag(cs, task);
1945 css_task_iter_end(&it);
1949 * update_flag - read a 0 or a 1 in a file and update associated flag
1950 * bit: the bit to update (see cpuset_flagbits_t)
1951 * cs: the cpuset to update
1952 * turning_on: whether the flag is being set or cleared
1954 * Call with cpuset_rwsem held.
1957 static int update_flag(cpuset_flagbits_t bit, struct cpuset *cs,
1960 struct cpuset *trialcs;
1961 int balance_flag_changed;
1962 int spread_flag_changed;
1965 trialcs = alloc_trial_cpuset(cs);
1970 set_bit(bit, &trialcs->flags);
1972 clear_bit(bit, &trialcs->flags);
1974 err = validate_change(cs, trialcs);
1978 balance_flag_changed = (is_sched_load_balance(cs) !=
1979 is_sched_load_balance(trialcs));
1981 spread_flag_changed = ((is_spread_slab(cs) != is_spread_slab(trialcs))
1982 || (is_spread_page(cs) != is_spread_page(trialcs)));
1984 spin_lock_irq(&callback_lock);
1985 cs->flags = trialcs->flags;
1986 spin_unlock_irq(&callback_lock);
1988 if (!cpumask_empty(trialcs->cpus_allowed) && balance_flag_changed)
1989 rebuild_sched_domains_locked();
1991 if (spread_flag_changed)
1992 update_tasks_flags(cs);
1994 free_cpuset(trialcs);
1999 * update_prstate - update partititon_root_state
2000 * cs: the cpuset to update
2001 * new_prs: new partition root state
2003 * Call with cpuset_rwsem held.
2005 static int update_prstate(struct cpuset *cs, int new_prs)
2007 int err, old_prs = cs->partition_root_state;
2008 struct cpuset *parent = parent_cs(cs);
2009 struct tmpmasks tmpmask;
2011 if (old_prs == new_prs)
2015 * Cannot force a partial or invalid partition root to a full
2018 if (new_prs && (old_prs == PRS_ERROR))
2021 if (alloc_cpumasks(NULL, &tmpmask))
2027 * Turning on partition root requires setting the
2028 * CS_CPU_EXCLUSIVE bit implicitly as well and cpus_allowed
2031 if (cpumask_empty(cs->cpus_allowed))
2034 err = update_flag(CS_CPU_EXCLUSIVE, cs, 1);
2038 err = update_parent_subparts_cpumask(cs, partcmd_enable,
2041 update_flag(CS_CPU_EXCLUSIVE, cs, 0);
2046 * Turning off partition root will clear the
2047 * CS_CPU_EXCLUSIVE bit.
2049 if (old_prs == PRS_ERROR) {
2050 update_flag(CS_CPU_EXCLUSIVE, cs, 0);
2055 err = update_parent_subparts_cpumask(cs, partcmd_disable,
2060 /* Turning off CS_CPU_EXCLUSIVE will not return error */
2061 update_flag(CS_CPU_EXCLUSIVE, cs, 0);
2064 update_tasks_cpumask(parent);
2066 if (parent->child_ecpus_count)
2067 update_sibling_cpumasks(parent, cs, &tmpmask);
2069 rebuild_sched_domains_locked();
2072 spin_lock_irq(&callback_lock);
2073 cs->partition_root_state = new_prs;
2074 spin_unlock_irq(&callback_lock);
2075 notify_partition_change(cs, old_prs, new_prs);
2078 free_cpumasks(NULL, &tmpmask);
2083 * Frequency meter - How fast is some event occurring?
2085 * These routines manage a digitally filtered, constant time based,
2086 * event frequency meter. There are four routines:
2087 * fmeter_init() - initialize a frequency meter.
2088 * fmeter_markevent() - called each time the event happens.
2089 * fmeter_getrate() - returns the recent rate of such events.
2090 * fmeter_update() - internal routine used to update fmeter.
2092 * A common data structure is passed to each of these routines,
2093 * which is used to keep track of the state required to manage the
2094 * frequency meter and its digital filter.
2096 * The filter works on the number of events marked per unit time.
2097 * The filter is single-pole low-pass recursive (IIR). The time unit
2098 * is 1 second. Arithmetic is done using 32-bit integers scaled to
2099 * simulate 3 decimal digits of precision (multiplied by 1000).
2101 * With an FM_COEF of 933, and a time base of 1 second, the filter
2102 * has a half-life of 10 seconds, meaning that if the events quit
2103 * happening, then the rate returned from the fmeter_getrate()
2104 * will be cut in half each 10 seconds, until it converges to zero.
2106 * It is not worth doing a real infinitely recursive filter. If more
2107 * than FM_MAXTICKS ticks have elapsed since the last filter event,
2108 * just compute FM_MAXTICKS ticks worth, by which point the level
2111 * Limit the count of unprocessed events to FM_MAXCNT, so as to avoid
2112 * arithmetic overflow in the fmeter_update() routine.
2114 * Given the simple 32 bit integer arithmetic used, this meter works
2115 * best for reporting rates between one per millisecond (msec) and
2116 * one per 32 (approx) seconds. At constant rates faster than one
2117 * per msec it maxes out at values just under 1,000,000. At constant
2118 * rates between one per msec, and one per second it will stabilize
2119 * to a value N*1000, where N is the rate of events per second.
2120 * At constant rates between one per second and one per 32 seconds,
2121 * it will be choppy, moving up on the seconds that have an event,
2122 * and then decaying until the next event. At rates slower than
2123 * about one in 32 seconds, it decays all the way back to zero between
2127 #define FM_COEF 933 /* coefficient for half-life of 10 secs */
2128 #define FM_MAXTICKS ((u32)99) /* useless computing more ticks than this */
2129 #define FM_MAXCNT 1000000 /* limit cnt to avoid overflow */
2130 #define FM_SCALE 1000 /* faux fixed point scale */
2132 /* Initialize a frequency meter */
2133 static void fmeter_init(struct fmeter *fmp)
2138 spin_lock_init(&fmp->lock);
2141 /* Internal meter update - process cnt events and update value */
2142 static void fmeter_update(struct fmeter *fmp)
2147 now = ktime_get_seconds();
2148 ticks = now - fmp->time;
2153 ticks = min(FM_MAXTICKS, ticks);
2155 fmp->val = (FM_COEF * fmp->val) / FM_SCALE;
2158 fmp->val += ((FM_SCALE - FM_COEF) * fmp->cnt) / FM_SCALE;
2162 /* Process any previous ticks, then bump cnt by one (times scale). */
2163 static void fmeter_markevent(struct fmeter *fmp)
2165 spin_lock(&fmp->lock);
2167 fmp->cnt = min(FM_MAXCNT, fmp->cnt + FM_SCALE);
2168 spin_unlock(&fmp->lock);
2171 /* Process any previous ticks, then return current value. */
2172 static int fmeter_getrate(struct fmeter *fmp)
2176 spin_lock(&fmp->lock);
2179 spin_unlock(&fmp->lock);
2183 static struct cpuset *cpuset_attach_old_cs;
2185 /* Called by cgroups to determine if a cpuset is usable; cpuset_rwsem held */
2186 static int cpuset_can_attach(struct cgroup_taskset *tset)
2188 struct cgroup_subsys_state *css;
2190 struct task_struct *task;
2193 /* used later by cpuset_attach() */
2194 cpuset_attach_old_cs = task_cs(cgroup_taskset_first(tset, &css));
2197 percpu_down_write(&cpuset_rwsem);
2199 /* allow moving tasks into an empty cpuset if on default hierarchy */
2201 if (!is_in_v2_mode() &&
2202 (cpumask_empty(cs->cpus_allowed) || nodes_empty(cs->mems_allowed)))
2205 cgroup_taskset_for_each(task, css, tset) {
2206 ret = task_can_attach(task, cs->effective_cpus);
2209 ret = security_task_setscheduler(task);
2215 * Mark attach is in progress. This makes validate_change() fail
2216 * changes which zero cpus/mems_allowed.
2218 cs->attach_in_progress++;
2221 percpu_up_write(&cpuset_rwsem);
2225 static void cpuset_cancel_attach(struct cgroup_taskset *tset)
2227 struct cgroup_subsys_state *css;
2229 cgroup_taskset_first(tset, &css);
2231 percpu_down_write(&cpuset_rwsem);
2232 css_cs(css)->attach_in_progress--;
2233 percpu_up_write(&cpuset_rwsem);
2237 * Protected by cpuset_rwsem. cpus_attach is used only by cpuset_attach()
2238 * but we can't allocate it dynamically there. Define it global and
2239 * allocate from cpuset_init().
2241 static cpumask_var_t cpus_attach;
2243 static void cpuset_attach(struct cgroup_taskset *tset)
2245 /* static buf protected by cpuset_rwsem */
2246 static nodemask_t cpuset_attach_nodemask_to;
2247 struct task_struct *task;
2248 struct task_struct *leader;
2249 struct cgroup_subsys_state *css;
2251 struct cpuset *oldcs = cpuset_attach_old_cs;
2253 cgroup_taskset_first(tset, &css);
2256 lockdep_assert_cpus_held(); /* see cgroup_attach_lock() */
2257 percpu_down_write(&cpuset_rwsem);
2259 guarantee_online_mems(cs, &cpuset_attach_nodemask_to);
2261 cgroup_taskset_for_each(task, css, tset) {
2262 if (cs != &top_cpuset)
2263 guarantee_online_cpus(task, cpus_attach);
2265 cpumask_copy(cpus_attach, task_cpu_possible_mask(task));
2267 * can_attach beforehand should guarantee that this doesn't
2268 * fail. TODO: have a better way to handle failure here
2270 WARN_ON_ONCE(set_cpus_allowed_ptr(task, cpus_attach));
2272 cpuset_change_task_nodemask(task, &cpuset_attach_nodemask_to);
2273 cpuset_update_task_spread_flag(cs, task);
2277 * Change mm for all threadgroup leaders. This is expensive and may
2278 * sleep and should be moved outside migration path proper.
2280 cpuset_attach_nodemask_to = cs->effective_mems;
2281 cgroup_taskset_for_each_leader(leader, css, tset) {
2282 struct mm_struct *mm = get_task_mm(leader);
2285 mpol_rebind_mm(mm, &cpuset_attach_nodemask_to);
2288 * old_mems_allowed is the same with mems_allowed
2289 * here, except if this task is being moved
2290 * automatically due to hotplug. In that case
2291 * @mems_allowed has been updated and is empty, so
2292 * @old_mems_allowed is the right nodesets that we
2295 if (is_memory_migrate(cs))
2296 cpuset_migrate_mm(mm, &oldcs->old_mems_allowed,
2297 &cpuset_attach_nodemask_to);
2303 cs->old_mems_allowed = cpuset_attach_nodemask_to;
2305 cs->attach_in_progress--;
2306 if (!cs->attach_in_progress)
2307 wake_up(&cpuset_attach_wq);
2309 percpu_up_write(&cpuset_rwsem);
2312 /* The various types of files and directories in a cpuset file system */
2315 FILE_MEMORY_MIGRATE,
2318 FILE_EFFECTIVE_CPULIST,
2319 FILE_EFFECTIVE_MEMLIST,
2320 FILE_SUBPARTS_CPULIST,
2324 FILE_SCHED_LOAD_BALANCE,
2325 FILE_PARTITION_ROOT,
2326 FILE_SCHED_RELAX_DOMAIN_LEVEL,
2327 FILE_MEMORY_PRESSURE_ENABLED,
2328 FILE_MEMORY_PRESSURE,
2331 } cpuset_filetype_t;
2333 static int cpuset_write_u64(struct cgroup_subsys_state *css, struct cftype *cft,
2336 struct cpuset *cs = css_cs(css);
2337 cpuset_filetype_t type = cft->private;
2341 percpu_down_write(&cpuset_rwsem);
2342 if (!is_cpuset_online(cs)) {
2348 case FILE_CPU_EXCLUSIVE:
2349 retval = update_flag(CS_CPU_EXCLUSIVE, cs, val);
2351 case FILE_MEM_EXCLUSIVE:
2352 retval = update_flag(CS_MEM_EXCLUSIVE, cs, val);
2354 case FILE_MEM_HARDWALL:
2355 retval = update_flag(CS_MEM_HARDWALL, cs, val);
2357 case FILE_SCHED_LOAD_BALANCE:
2358 retval = update_flag(CS_SCHED_LOAD_BALANCE, cs, val);
2360 case FILE_MEMORY_MIGRATE:
2361 retval = update_flag(CS_MEMORY_MIGRATE, cs, val);
2363 case FILE_MEMORY_PRESSURE_ENABLED:
2364 cpuset_memory_pressure_enabled = !!val;
2366 case FILE_SPREAD_PAGE:
2367 retval = update_flag(CS_SPREAD_PAGE, cs, val);
2369 case FILE_SPREAD_SLAB:
2370 retval = update_flag(CS_SPREAD_SLAB, cs, val);
2377 percpu_up_write(&cpuset_rwsem);
2382 static int cpuset_write_s64(struct cgroup_subsys_state *css, struct cftype *cft,
2385 struct cpuset *cs = css_cs(css);
2386 cpuset_filetype_t type = cft->private;
2387 int retval = -ENODEV;
2390 percpu_down_write(&cpuset_rwsem);
2391 if (!is_cpuset_online(cs))
2395 case FILE_SCHED_RELAX_DOMAIN_LEVEL:
2396 retval = update_relax_domain_level(cs, val);
2403 percpu_up_write(&cpuset_rwsem);
2409 * Common handling for a write to a "cpus" or "mems" file.
2411 static ssize_t cpuset_write_resmask(struct kernfs_open_file *of,
2412 char *buf, size_t nbytes, loff_t off)
2414 struct cpuset *cs = css_cs(of_css(of));
2415 struct cpuset *trialcs;
2416 int retval = -ENODEV;
2418 buf = strstrip(buf);
2421 * CPU or memory hotunplug may leave @cs w/o any execution
2422 * resources, in which case the hotplug code asynchronously updates
2423 * configuration and transfers all tasks to the nearest ancestor
2424 * which can execute.
2426 * As writes to "cpus" or "mems" may restore @cs's execution
2427 * resources, wait for the previously scheduled operations before
2428 * proceeding, so that we don't end up keep removing tasks added
2429 * after execution capability is restored.
2431 * cpuset_hotplug_work calls back into cgroup core via
2432 * cgroup_transfer_tasks() and waiting for it from a cgroupfs
2433 * operation like this one can lead to a deadlock through kernfs
2434 * active_ref protection. Let's break the protection. Losing the
2435 * protection is okay as we check whether @cs is online after
2436 * grabbing cpuset_rwsem anyway. This only happens on the legacy
2440 kernfs_break_active_protection(of->kn);
2441 flush_work(&cpuset_hotplug_work);
2444 percpu_down_write(&cpuset_rwsem);
2445 if (!is_cpuset_online(cs))
2448 trialcs = alloc_trial_cpuset(cs);
2454 switch (of_cft(of)->private) {
2456 retval = update_cpumask(cs, trialcs, buf);
2459 retval = update_nodemask(cs, trialcs, buf);
2466 free_cpuset(trialcs);
2468 percpu_up_write(&cpuset_rwsem);
2470 kernfs_unbreak_active_protection(of->kn);
2472 flush_workqueue(cpuset_migrate_mm_wq);
2473 return retval ?: nbytes;
2477 * These ascii lists should be read in a single call, by using a user
2478 * buffer large enough to hold the entire map. If read in smaller
2479 * chunks, there is no guarantee of atomicity. Since the display format
2480 * used, list of ranges of sequential numbers, is variable length,
2481 * and since these maps can change value dynamically, one could read
2482 * gibberish by doing partial reads while a list was changing.
2484 static int cpuset_common_seq_show(struct seq_file *sf, void *v)
2486 struct cpuset *cs = css_cs(seq_css(sf));
2487 cpuset_filetype_t type = seq_cft(sf)->private;
2490 spin_lock_irq(&callback_lock);
2494 seq_printf(sf, "%*pbl\n", cpumask_pr_args(cs->cpus_allowed));
2497 seq_printf(sf, "%*pbl\n", nodemask_pr_args(&cs->mems_allowed));
2499 case FILE_EFFECTIVE_CPULIST:
2500 seq_printf(sf, "%*pbl\n", cpumask_pr_args(cs->effective_cpus));
2502 case FILE_EFFECTIVE_MEMLIST:
2503 seq_printf(sf, "%*pbl\n", nodemask_pr_args(&cs->effective_mems));
2505 case FILE_SUBPARTS_CPULIST:
2506 seq_printf(sf, "%*pbl\n", cpumask_pr_args(cs->subparts_cpus));
2512 spin_unlock_irq(&callback_lock);
2516 static u64 cpuset_read_u64(struct cgroup_subsys_state *css, struct cftype *cft)
2518 struct cpuset *cs = css_cs(css);
2519 cpuset_filetype_t type = cft->private;
2521 case FILE_CPU_EXCLUSIVE:
2522 return is_cpu_exclusive(cs);
2523 case FILE_MEM_EXCLUSIVE:
2524 return is_mem_exclusive(cs);
2525 case FILE_MEM_HARDWALL:
2526 return is_mem_hardwall(cs);
2527 case FILE_SCHED_LOAD_BALANCE:
2528 return is_sched_load_balance(cs);
2529 case FILE_MEMORY_MIGRATE:
2530 return is_memory_migrate(cs);
2531 case FILE_MEMORY_PRESSURE_ENABLED:
2532 return cpuset_memory_pressure_enabled;
2533 case FILE_MEMORY_PRESSURE:
2534 return fmeter_getrate(&cs->fmeter);
2535 case FILE_SPREAD_PAGE:
2536 return is_spread_page(cs);
2537 case FILE_SPREAD_SLAB:
2538 return is_spread_slab(cs);
2543 /* Unreachable but makes gcc happy */
2547 static s64 cpuset_read_s64(struct cgroup_subsys_state *css, struct cftype *cft)
2549 struct cpuset *cs = css_cs(css);
2550 cpuset_filetype_t type = cft->private;
2552 case FILE_SCHED_RELAX_DOMAIN_LEVEL:
2553 return cs->relax_domain_level;
2558 /* Unreachable but makes gcc happy */
2562 static int sched_partition_show(struct seq_file *seq, void *v)
2564 struct cpuset *cs = css_cs(seq_css(seq));
2566 switch (cs->partition_root_state) {
2568 seq_puts(seq, "root\n");
2571 seq_puts(seq, "member\n");
2574 seq_puts(seq, "root invalid\n");
2580 static ssize_t sched_partition_write(struct kernfs_open_file *of, char *buf,
2581 size_t nbytes, loff_t off)
2583 struct cpuset *cs = css_cs(of_css(of));
2585 int retval = -ENODEV;
2587 buf = strstrip(buf);
2590 * Convert "root" to ENABLED, and convert "member" to DISABLED.
2592 if (!strcmp(buf, "root"))
2594 else if (!strcmp(buf, "member"))
2601 percpu_down_write(&cpuset_rwsem);
2602 if (!is_cpuset_online(cs))
2605 retval = update_prstate(cs, val);
2607 percpu_up_write(&cpuset_rwsem);
2610 return retval ?: nbytes;
2614 * for the common functions, 'private' gives the type of file
2617 static struct cftype legacy_files[] = {
2620 .seq_show = cpuset_common_seq_show,
2621 .write = cpuset_write_resmask,
2622 .max_write_len = (100U + 6 * NR_CPUS),
2623 .private = FILE_CPULIST,
2628 .seq_show = cpuset_common_seq_show,
2629 .write = cpuset_write_resmask,
2630 .max_write_len = (100U + 6 * MAX_NUMNODES),
2631 .private = FILE_MEMLIST,
2635 .name = "effective_cpus",
2636 .seq_show = cpuset_common_seq_show,
2637 .private = FILE_EFFECTIVE_CPULIST,
2641 .name = "effective_mems",
2642 .seq_show = cpuset_common_seq_show,
2643 .private = FILE_EFFECTIVE_MEMLIST,
2647 .name = "cpu_exclusive",
2648 .read_u64 = cpuset_read_u64,
2649 .write_u64 = cpuset_write_u64,
2650 .private = FILE_CPU_EXCLUSIVE,
2654 .name = "mem_exclusive",
2655 .read_u64 = cpuset_read_u64,
2656 .write_u64 = cpuset_write_u64,
2657 .private = FILE_MEM_EXCLUSIVE,
2661 .name = "mem_hardwall",
2662 .read_u64 = cpuset_read_u64,
2663 .write_u64 = cpuset_write_u64,
2664 .private = FILE_MEM_HARDWALL,
2668 .name = "sched_load_balance",
2669 .read_u64 = cpuset_read_u64,
2670 .write_u64 = cpuset_write_u64,
2671 .private = FILE_SCHED_LOAD_BALANCE,
2675 .name = "sched_relax_domain_level",
2676 .read_s64 = cpuset_read_s64,
2677 .write_s64 = cpuset_write_s64,
2678 .private = FILE_SCHED_RELAX_DOMAIN_LEVEL,
2682 .name = "memory_migrate",
2683 .read_u64 = cpuset_read_u64,
2684 .write_u64 = cpuset_write_u64,
2685 .private = FILE_MEMORY_MIGRATE,
2689 .name = "memory_pressure",
2690 .read_u64 = cpuset_read_u64,
2691 .private = FILE_MEMORY_PRESSURE,
2695 .name = "memory_spread_page",
2696 .read_u64 = cpuset_read_u64,
2697 .write_u64 = cpuset_write_u64,
2698 .private = FILE_SPREAD_PAGE,
2702 .name = "memory_spread_slab",
2703 .read_u64 = cpuset_read_u64,
2704 .write_u64 = cpuset_write_u64,
2705 .private = FILE_SPREAD_SLAB,
2709 .name = "memory_pressure_enabled",
2710 .flags = CFTYPE_ONLY_ON_ROOT,
2711 .read_u64 = cpuset_read_u64,
2712 .write_u64 = cpuset_write_u64,
2713 .private = FILE_MEMORY_PRESSURE_ENABLED,
2720 * This is currently a minimal set for the default hierarchy. It can be
2721 * expanded later on by migrating more features and control files from v1.
2723 static struct cftype dfl_files[] = {
2726 .seq_show = cpuset_common_seq_show,
2727 .write = cpuset_write_resmask,
2728 .max_write_len = (100U + 6 * NR_CPUS),
2729 .private = FILE_CPULIST,
2730 .flags = CFTYPE_NOT_ON_ROOT,
2735 .seq_show = cpuset_common_seq_show,
2736 .write = cpuset_write_resmask,
2737 .max_write_len = (100U + 6 * MAX_NUMNODES),
2738 .private = FILE_MEMLIST,
2739 .flags = CFTYPE_NOT_ON_ROOT,
2743 .name = "cpus.effective",
2744 .seq_show = cpuset_common_seq_show,
2745 .private = FILE_EFFECTIVE_CPULIST,
2749 .name = "mems.effective",
2750 .seq_show = cpuset_common_seq_show,
2751 .private = FILE_EFFECTIVE_MEMLIST,
2755 .name = "cpus.partition",
2756 .seq_show = sched_partition_show,
2757 .write = sched_partition_write,
2758 .private = FILE_PARTITION_ROOT,
2759 .flags = CFTYPE_NOT_ON_ROOT,
2760 .file_offset = offsetof(struct cpuset, partition_file),
2764 .name = "cpus.subpartitions",
2765 .seq_show = cpuset_common_seq_show,
2766 .private = FILE_SUBPARTS_CPULIST,
2767 .flags = CFTYPE_DEBUG,
2775 * cpuset_css_alloc - allocate a cpuset css
2776 * cgrp: control group that the new cpuset will be part of
2779 static struct cgroup_subsys_state *
2780 cpuset_css_alloc(struct cgroup_subsys_state *parent_css)
2785 return &top_cpuset.css;
2787 cs = kzalloc(sizeof(*cs), GFP_KERNEL);
2789 return ERR_PTR(-ENOMEM);
2791 if (alloc_cpumasks(cs, NULL)) {
2793 return ERR_PTR(-ENOMEM);
2796 __set_bit(CS_SCHED_LOAD_BALANCE, &cs->flags);
2797 nodes_clear(cs->mems_allowed);
2798 nodes_clear(cs->effective_mems);
2799 fmeter_init(&cs->fmeter);
2800 cs->relax_domain_level = -1;
2802 /* Set CS_MEMORY_MIGRATE for default hierarchy */
2803 if (cgroup_subsys_on_dfl(cpuset_cgrp_subsys))
2804 __set_bit(CS_MEMORY_MIGRATE, &cs->flags);
2809 static int cpuset_css_online(struct cgroup_subsys_state *css)
2811 struct cpuset *cs = css_cs(css);
2812 struct cpuset *parent = parent_cs(cs);
2813 struct cpuset *tmp_cs;
2814 struct cgroup_subsys_state *pos_css;
2820 percpu_down_write(&cpuset_rwsem);
2822 set_bit(CS_ONLINE, &cs->flags);
2823 if (is_spread_page(parent))
2824 set_bit(CS_SPREAD_PAGE, &cs->flags);
2825 if (is_spread_slab(parent))
2826 set_bit(CS_SPREAD_SLAB, &cs->flags);
2830 spin_lock_irq(&callback_lock);
2831 if (is_in_v2_mode()) {
2832 cpumask_copy(cs->effective_cpus, parent->effective_cpus);
2833 cs->effective_mems = parent->effective_mems;
2834 cs->use_parent_ecpus = true;
2835 parent->child_ecpus_count++;
2837 spin_unlock_irq(&callback_lock);
2839 if (!test_bit(CGRP_CPUSET_CLONE_CHILDREN, &css->cgroup->flags))
2843 * Clone @parent's configuration if CGRP_CPUSET_CLONE_CHILDREN is
2844 * set. This flag handling is implemented in cgroup core for
2845 * histrical reasons - the flag may be specified during mount.
2847 * Currently, if any sibling cpusets have exclusive cpus or mem, we
2848 * refuse to clone the configuration - thereby refusing the task to
2849 * be entered, and as a result refusing the sys_unshare() or
2850 * clone() which initiated it. If this becomes a problem for some
2851 * users who wish to allow that scenario, then this could be
2852 * changed to grant parent->cpus_allowed-sibling_cpus_exclusive
2853 * (and likewise for mems) to the new cgroup.
2856 cpuset_for_each_child(tmp_cs, pos_css, parent) {
2857 if (is_mem_exclusive(tmp_cs) || is_cpu_exclusive(tmp_cs)) {
2864 spin_lock_irq(&callback_lock);
2865 cs->mems_allowed = parent->mems_allowed;
2866 cs->effective_mems = parent->mems_allowed;
2867 cpumask_copy(cs->cpus_allowed, parent->cpus_allowed);
2868 cpumask_copy(cs->effective_cpus, parent->cpus_allowed);
2869 spin_unlock_irq(&callback_lock);
2871 percpu_up_write(&cpuset_rwsem);
2877 * If the cpuset being removed has its flag 'sched_load_balance'
2878 * enabled, then simulate turning sched_load_balance off, which
2879 * will call rebuild_sched_domains_locked(). That is not needed
2880 * in the default hierarchy where only changes in partition
2881 * will cause repartitioning.
2883 * If the cpuset has the 'sched.partition' flag enabled, simulate
2884 * turning 'sched.partition" off.
2887 static void cpuset_css_offline(struct cgroup_subsys_state *css)
2889 struct cpuset *cs = css_cs(css);
2892 percpu_down_write(&cpuset_rwsem);
2894 if (is_partition_root(cs))
2895 update_prstate(cs, 0);
2897 if (!cgroup_subsys_on_dfl(cpuset_cgrp_subsys) &&
2898 is_sched_load_balance(cs))
2899 update_flag(CS_SCHED_LOAD_BALANCE, cs, 0);
2901 if (cs->use_parent_ecpus) {
2902 struct cpuset *parent = parent_cs(cs);
2904 cs->use_parent_ecpus = false;
2905 parent->child_ecpus_count--;
2909 clear_bit(CS_ONLINE, &cs->flags);
2911 percpu_up_write(&cpuset_rwsem);
2915 static void cpuset_css_free(struct cgroup_subsys_state *css)
2917 struct cpuset *cs = css_cs(css);
2922 static void cpuset_bind(struct cgroup_subsys_state *root_css)
2924 percpu_down_write(&cpuset_rwsem);
2925 spin_lock_irq(&callback_lock);
2927 if (is_in_v2_mode()) {
2928 cpumask_copy(top_cpuset.cpus_allowed, cpu_possible_mask);
2929 top_cpuset.mems_allowed = node_possible_map;
2931 cpumask_copy(top_cpuset.cpus_allowed,
2932 top_cpuset.effective_cpus);
2933 top_cpuset.mems_allowed = top_cpuset.effective_mems;
2936 spin_unlock_irq(&callback_lock);
2937 percpu_up_write(&cpuset_rwsem);
2941 * Make sure the new task conform to the current state of its parent,
2942 * which could have been changed by cpuset just after it inherits the
2943 * state from the parent and before it sits on the cgroup's task list.
2945 static void cpuset_fork(struct task_struct *task)
2947 if (task_css_is_root(task, cpuset_cgrp_id))
2950 set_cpus_allowed_ptr(task, current->cpus_ptr);
2951 task->mems_allowed = current->mems_allowed;
2954 struct cgroup_subsys cpuset_cgrp_subsys = {
2955 .css_alloc = cpuset_css_alloc,
2956 .css_online = cpuset_css_online,
2957 .css_offline = cpuset_css_offline,
2958 .css_free = cpuset_css_free,
2959 .can_attach = cpuset_can_attach,
2960 .cancel_attach = cpuset_cancel_attach,
2961 .attach = cpuset_attach,
2962 .post_attach = cpuset_post_attach,
2963 .bind = cpuset_bind,
2964 .fork = cpuset_fork,
2965 .legacy_cftypes = legacy_files,
2966 .dfl_cftypes = dfl_files,
2972 * cpuset_init - initialize cpusets at system boot
2974 * Description: Initialize top_cpuset
2977 int __init cpuset_init(void)
2979 BUG_ON(percpu_init_rwsem(&cpuset_rwsem));
2981 BUG_ON(!alloc_cpumask_var(&top_cpuset.cpus_allowed, GFP_KERNEL));
2982 BUG_ON(!alloc_cpumask_var(&top_cpuset.effective_cpus, GFP_KERNEL));
2983 BUG_ON(!zalloc_cpumask_var(&top_cpuset.subparts_cpus, GFP_KERNEL));
2985 cpumask_setall(top_cpuset.cpus_allowed);
2986 nodes_setall(top_cpuset.mems_allowed);
2987 cpumask_setall(top_cpuset.effective_cpus);
2988 nodes_setall(top_cpuset.effective_mems);
2990 fmeter_init(&top_cpuset.fmeter);
2991 set_bit(CS_SCHED_LOAD_BALANCE, &top_cpuset.flags);
2992 top_cpuset.relax_domain_level = -1;
2994 BUG_ON(!alloc_cpumask_var(&cpus_attach, GFP_KERNEL));
3000 * If CPU and/or memory hotplug handlers, below, unplug any CPUs
3001 * or memory nodes, we need to walk over the cpuset hierarchy,
3002 * removing that CPU or node from all cpusets. If this removes the
3003 * last CPU or node from a cpuset, then move the tasks in the empty
3004 * cpuset to its next-highest non-empty parent.
3006 static void remove_tasks_in_empty_cpuset(struct cpuset *cs)
3008 struct cpuset *parent;
3011 * Find its next-highest non-empty parent, (top cpuset
3012 * has online cpus, so can't be empty).
3014 parent = parent_cs(cs);
3015 while (cpumask_empty(parent->cpus_allowed) ||
3016 nodes_empty(parent->mems_allowed))
3017 parent = parent_cs(parent);
3019 if (cgroup_transfer_tasks(parent->css.cgroup, cs->css.cgroup)) {
3020 pr_err("cpuset: failed to transfer tasks out of empty cpuset ");
3021 pr_cont_cgroup_name(cs->css.cgroup);
3027 hotplug_update_tasks_legacy(struct cpuset *cs,
3028 struct cpumask *new_cpus, nodemask_t *new_mems,
3029 bool cpus_updated, bool mems_updated)
3033 spin_lock_irq(&callback_lock);
3034 cpumask_copy(cs->cpus_allowed, new_cpus);
3035 cpumask_copy(cs->effective_cpus, new_cpus);
3036 cs->mems_allowed = *new_mems;
3037 cs->effective_mems = *new_mems;
3038 spin_unlock_irq(&callback_lock);
3041 * Don't call update_tasks_cpumask() if the cpuset becomes empty,
3042 * as the tasks will be migratecd to an ancestor.
3044 if (cpus_updated && !cpumask_empty(cs->cpus_allowed))
3045 update_tasks_cpumask(cs);
3046 if (mems_updated && !nodes_empty(cs->mems_allowed))
3047 update_tasks_nodemask(cs);
3049 is_empty = cpumask_empty(cs->cpus_allowed) ||
3050 nodes_empty(cs->mems_allowed);
3052 percpu_up_write(&cpuset_rwsem);
3055 * Move tasks to the nearest ancestor with execution resources,
3056 * This is full cgroup operation which will also call back into
3057 * cpuset. Should be done outside any lock.
3060 remove_tasks_in_empty_cpuset(cs);
3062 percpu_down_write(&cpuset_rwsem);
3066 hotplug_update_tasks(struct cpuset *cs,
3067 struct cpumask *new_cpus, nodemask_t *new_mems,
3068 bool cpus_updated, bool mems_updated)
3070 if (cpumask_empty(new_cpus))
3071 cpumask_copy(new_cpus, parent_cs(cs)->effective_cpus);
3072 if (nodes_empty(*new_mems))
3073 *new_mems = parent_cs(cs)->effective_mems;
3075 spin_lock_irq(&callback_lock);
3076 cpumask_copy(cs->effective_cpus, new_cpus);
3077 cs->effective_mems = *new_mems;
3078 spin_unlock_irq(&callback_lock);
3081 update_tasks_cpumask(cs);
3083 update_tasks_nodemask(cs);
3086 static bool force_rebuild;
3088 void cpuset_force_rebuild(void)
3090 force_rebuild = true;
3094 * cpuset_hotplug_update_tasks - update tasks in a cpuset for hotunplug
3095 * @cs: cpuset in interest
3096 * @tmp: the tmpmasks structure pointer
3098 * Compare @cs's cpu and mem masks against top_cpuset and if some have gone
3099 * offline, update @cs accordingly. If @cs ends up with no CPU or memory,
3100 * all its tasks are moved to the nearest ancestor with both resources.
3102 static void cpuset_hotplug_update_tasks(struct cpuset *cs, struct tmpmasks *tmp)
3104 static cpumask_t new_cpus;
3105 static nodemask_t new_mems;
3108 struct cpuset *parent;
3110 wait_event(cpuset_attach_wq, cs->attach_in_progress == 0);
3112 percpu_down_write(&cpuset_rwsem);
3115 * We have raced with task attaching. We wait until attaching
3116 * is finished, so we won't attach a task to an empty cpuset.
3118 if (cs->attach_in_progress) {
3119 percpu_up_write(&cpuset_rwsem);
3123 parent = parent_cs(cs);
3124 compute_effective_cpumask(&new_cpus, cs, parent);
3125 nodes_and(new_mems, cs->mems_allowed, parent->effective_mems);
3127 if (cs->nr_subparts_cpus)
3129 * Make sure that CPUs allocated to child partitions
3130 * do not show up in effective_cpus.
3132 cpumask_andnot(&new_cpus, &new_cpus, cs->subparts_cpus);
3134 if (!tmp || !cs->partition_root_state)
3138 * In the unlikely event that a partition root has empty
3139 * effective_cpus or its parent becomes erroneous, we have to
3140 * transition it to the erroneous state.
3142 if (is_partition_root(cs) && (cpumask_empty(&new_cpus) ||
3143 (parent->partition_root_state == PRS_ERROR))) {
3144 if (cs->nr_subparts_cpus) {
3145 spin_lock_irq(&callback_lock);
3146 cs->nr_subparts_cpus = 0;
3147 cpumask_clear(cs->subparts_cpus);
3148 spin_unlock_irq(&callback_lock);
3149 compute_effective_cpumask(&new_cpus, cs, parent);
3153 * If the effective_cpus is empty because the child
3154 * partitions take away all the CPUs, we can keep
3155 * the current partition and let the child partitions
3156 * fight for available CPUs.
3158 if ((parent->partition_root_state == PRS_ERROR) ||
3159 cpumask_empty(&new_cpus)) {
3162 update_parent_subparts_cpumask(cs, partcmd_disable,
3164 old_prs = cs->partition_root_state;
3165 if (old_prs != PRS_ERROR) {
3166 spin_lock_irq(&callback_lock);
3167 cs->partition_root_state = PRS_ERROR;
3168 spin_unlock_irq(&callback_lock);
3169 notify_partition_change(cs, old_prs, PRS_ERROR);
3172 cpuset_force_rebuild();
3176 * On the other hand, an erroneous partition root may be transitioned
3177 * back to a regular one or a partition root with no CPU allocated
3178 * from the parent may change to erroneous.
3180 if (is_partition_root(parent) &&
3181 ((cs->partition_root_state == PRS_ERROR) ||
3182 !cpumask_intersects(&new_cpus, parent->subparts_cpus)) &&
3183 update_parent_subparts_cpumask(cs, partcmd_update, NULL, tmp))
3184 cpuset_force_rebuild();
3187 cpus_updated = !cpumask_equal(&new_cpus, cs->effective_cpus);
3188 mems_updated = !nodes_equal(new_mems, cs->effective_mems);
3190 if (is_in_v2_mode())
3191 hotplug_update_tasks(cs, &new_cpus, &new_mems,
3192 cpus_updated, mems_updated);
3194 hotplug_update_tasks_legacy(cs, &new_cpus, &new_mems,
3195 cpus_updated, mems_updated);
3197 percpu_up_write(&cpuset_rwsem);
3201 * cpuset_hotplug_workfn - handle CPU/memory hotunplug for a cpuset
3203 * This function is called after either CPU or memory configuration has
3204 * changed and updates cpuset accordingly. The top_cpuset is always
3205 * synchronized to cpu_active_mask and N_MEMORY, which is necessary in
3206 * order to make cpusets transparent (of no affect) on systems that are
3207 * actively using CPU hotplug but making no active use of cpusets.
3209 * Non-root cpusets are only affected by offlining. If any CPUs or memory
3210 * nodes have been taken down, cpuset_hotplug_update_tasks() is invoked on
3213 * Note that CPU offlining during suspend is ignored. We don't modify
3214 * cpusets across suspend/resume cycles at all.
3216 static void cpuset_hotplug_workfn(struct work_struct *work)
3218 static cpumask_t new_cpus;
3219 static nodemask_t new_mems;
3220 bool cpus_updated, mems_updated;
3221 bool on_dfl = is_in_v2_mode();
3222 struct tmpmasks tmp, *ptmp = NULL;
3224 if (on_dfl && !alloc_cpumasks(NULL, &tmp))
3227 percpu_down_write(&cpuset_rwsem);
3229 /* fetch the available cpus/mems and find out which changed how */
3230 cpumask_copy(&new_cpus, cpu_active_mask);
3231 new_mems = node_states[N_MEMORY];
3234 * If subparts_cpus is populated, it is likely that the check below
3235 * will produce a false positive on cpus_updated when the cpu list
3236 * isn't changed. It is extra work, but it is better to be safe.
3238 cpus_updated = !cpumask_equal(top_cpuset.effective_cpus, &new_cpus);
3239 mems_updated = !nodes_equal(top_cpuset.effective_mems, new_mems);
3242 * In the rare case that hotplug removes all the cpus in subparts_cpus,
3243 * we assumed that cpus are updated.
3245 if (!cpus_updated && top_cpuset.nr_subparts_cpus)
3246 cpus_updated = true;
3248 /* synchronize cpus_allowed to cpu_active_mask */
3250 spin_lock_irq(&callback_lock);
3252 cpumask_copy(top_cpuset.cpus_allowed, &new_cpus);
3254 * Make sure that CPUs allocated to child partitions
3255 * do not show up in effective_cpus. If no CPU is left,
3256 * we clear the subparts_cpus & let the child partitions
3257 * fight for the CPUs again.
3259 if (top_cpuset.nr_subparts_cpus) {
3260 if (cpumask_subset(&new_cpus,
3261 top_cpuset.subparts_cpus)) {
3262 top_cpuset.nr_subparts_cpus = 0;
3263 cpumask_clear(top_cpuset.subparts_cpus);
3265 cpumask_andnot(&new_cpus, &new_cpus,
3266 top_cpuset.subparts_cpus);
3269 cpumask_copy(top_cpuset.effective_cpus, &new_cpus);
3270 spin_unlock_irq(&callback_lock);
3271 /* we don't mess with cpumasks of tasks in top_cpuset */
3274 /* synchronize mems_allowed to N_MEMORY */
3276 spin_lock_irq(&callback_lock);
3278 top_cpuset.mems_allowed = new_mems;
3279 top_cpuset.effective_mems = new_mems;
3280 spin_unlock_irq(&callback_lock);
3281 update_tasks_nodemask(&top_cpuset);
3284 percpu_up_write(&cpuset_rwsem);
3286 /* if cpus or mems changed, we need to propagate to descendants */
3287 if (cpus_updated || mems_updated) {
3289 struct cgroup_subsys_state *pos_css;
3292 cpuset_for_each_descendant_pre(cs, pos_css, &top_cpuset) {
3293 if (cs == &top_cpuset || !css_tryget_online(&cs->css))
3297 cpuset_hotplug_update_tasks(cs, ptmp);
3305 /* rebuild sched domains if cpus_allowed has changed */
3306 if (cpus_updated || force_rebuild) {
3307 force_rebuild = false;
3308 rebuild_sched_domains();
3311 free_cpumasks(NULL, ptmp);
3314 void cpuset_update_active_cpus(void)
3317 * We're inside cpu hotplug critical region which usually nests
3318 * inside cgroup synchronization. Bounce actual hotplug processing
3319 * to a work item to avoid reverse locking order.
3321 schedule_work(&cpuset_hotplug_work);
3324 void cpuset_wait_for_hotplug(void)
3326 flush_work(&cpuset_hotplug_work);
3330 * Keep top_cpuset.mems_allowed tracking node_states[N_MEMORY].
3331 * Call this routine anytime after node_states[N_MEMORY] changes.
3332 * See cpuset_update_active_cpus() for CPU hotplug handling.
3334 static int cpuset_track_online_nodes(struct notifier_block *self,
3335 unsigned long action, void *arg)
3337 schedule_work(&cpuset_hotplug_work);
3341 static struct notifier_block cpuset_track_online_nodes_nb = {
3342 .notifier_call = cpuset_track_online_nodes,
3343 .priority = 10, /* ??! */
3347 * cpuset_init_smp - initialize cpus_allowed
3349 * Description: Finish top cpuset after cpu, node maps are initialized
3351 void __init cpuset_init_smp(void)
3354 * cpus_allowd/mems_allowed set to v2 values in the initial
3355 * cpuset_bind() call will be reset to v1 values in another
3356 * cpuset_bind() call when v1 cpuset is mounted.
3358 top_cpuset.old_mems_allowed = top_cpuset.mems_allowed;
3360 cpumask_copy(top_cpuset.effective_cpus, cpu_active_mask);
3361 top_cpuset.effective_mems = node_states[N_MEMORY];
3363 register_hotmemory_notifier(&cpuset_track_online_nodes_nb);
3365 cpuset_migrate_mm_wq = alloc_ordered_workqueue("cpuset_migrate_mm", 0);
3366 BUG_ON(!cpuset_migrate_mm_wq);
3370 * cpuset_cpus_allowed - return cpus_allowed mask from a tasks cpuset.
3371 * @tsk: pointer to task_struct from which to obtain cpuset->cpus_allowed.
3372 * @pmask: pointer to struct cpumask variable to receive cpus_allowed set.
3374 * Description: Returns the cpumask_var_t cpus_allowed of the cpuset
3375 * attached to the specified @tsk. Guaranteed to return some non-empty
3376 * subset of cpu_online_mask, even if this means going outside the
3380 void cpuset_cpus_allowed(struct task_struct *tsk, struct cpumask *pmask)
3382 unsigned long flags;
3384 spin_lock_irqsave(&callback_lock, flags);
3385 guarantee_online_cpus(tsk, pmask);
3386 spin_unlock_irqrestore(&callback_lock, flags);
3390 * cpuset_cpus_allowed_fallback - final fallback before complete catastrophe.
3391 * @tsk: pointer to task_struct with which the scheduler is struggling
3393 * Description: In the case that the scheduler cannot find an allowed cpu in
3394 * tsk->cpus_allowed, we fall back to task_cs(tsk)->cpus_allowed. In legacy
3395 * mode however, this value is the same as task_cs(tsk)->effective_cpus,
3396 * which will not contain a sane cpumask during cases such as cpu hotplugging.
3397 * This is the absolute last resort for the scheduler and it is only used if
3398 * _every_ other avenue has been traveled.
3400 * Returns true if the affinity of @tsk was changed, false otherwise.
3403 bool cpuset_cpus_allowed_fallback(struct task_struct *tsk)
3405 const struct cpumask *possible_mask = task_cpu_possible_mask(tsk);
3406 const struct cpumask *cs_mask;
3407 bool changed = false;
3410 cs_mask = task_cs(tsk)->cpus_allowed;
3411 if (is_in_v2_mode() && cpumask_subset(cs_mask, possible_mask)) {
3412 do_set_cpus_allowed(tsk, cs_mask);
3418 * We own tsk->cpus_allowed, nobody can change it under us.
3420 * But we used cs && cs->cpus_allowed lockless and thus can
3421 * race with cgroup_attach_task() or update_cpumask() and get
3422 * the wrong tsk->cpus_allowed. However, both cases imply the
3423 * subsequent cpuset_change_cpumask()->set_cpus_allowed_ptr()
3424 * which takes task_rq_lock().
3426 * If we are called after it dropped the lock we must see all
3427 * changes in tsk_cs()->cpus_allowed. Otherwise we can temporary
3428 * set any mask even if it is not right from task_cs() pov,
3429 * the pending set_cpus_allowed_ptr() will fix things.
3431 * select_fallback_rq() will fix things ups and set cpu_possible_mask
3437 void __init cpuset_init_current_mems_allowed(void)
3439 nodes_setall(current->mems_allowed);
3443 * cpuset_mems_allowed - return mems_allowed mask from a tasks cpuset.
3444 * @tsk: pointer to task_struct from which to obtain cpuset->mems_allowed.
3446 * Description: Returns the nodemask_t mems_allowed of the cpuset
3447 * attached to the specified @tsk. Guaranteed to return some non-empty
3448 * subset of node_states[N_MEMORY], even if this means going outside the
3452 nodemask_t cpuset_mems_allowed(struct task_struct *tsk)
3455 unsigned long flags;
3457 spin_lock_irqsave(&callback_lock, flags);
3459 guarantee_online_mems(task_cs(tsk), &mask);
3461 spin_unlock_irqrestore(&callback_lock, flags);
3467 * cpuset_nodemask_valid_mems_allowed - check nodemask vs. current mems_allowed
3468 * @nodemask: the nodemask to be checked
3470 * Are any of the nodes in the nodemask allowed in current->mems_allowed?
3472 int cpuset_nodemask_valid_mems_allowed(nodemask_t *nodemask)
3474 return nodes_intersects(*nodemask, current->mems_allowed);
3478 * nearest_hardwall_ancestor() - Returns the nearest mem_exclusive or
3479 * mem_hardwall ancestor to the specified cpuset. Call holding
3480 * callback_lock. If no ancestor is mem_exclusive or mem_hardwall
3481 * (an unusual configuration), then returns the root cpuset.
3483 static struct cpuset *nearest_hardwall_ancestor(struct cpuset *cs)
3485 while (!(is_mem_exclusive(cs) || is_mem_hardwall(cs)) && parent_cs(cs))
3491 * cpuset_node_allowed - Can we allocate on a memory node?
3492 * @node: is this an allowed node?
3493 * @gfp_mask: memory allocation flags
3495 * If we're in interrupt, yes, we can always allocate. If @node is set in
3496 * current's mems_allowed, yes. If it's not a __GFP_HARDWALL request and this
3497 * node is set in the nearest hardwalled cpuset ancestor to current's cpuset,
3498 * yes. If current has access to memory reserves as an oom victim, yes.
3501 * GFP_USER allocations are marked with the __GFP_HARDWALL bit,
3502 * and do not allow allocations outside the current tasks cpuset
3503 * unless the task has been OOM killed.
3504 * GFP_KERNEL allocations are not so marked, so can escape to the
3505 * nearest enclosing hardwalled ancestor cpuset.
3507 * Scanning up parent cpusets requires callback_lock. The
3508 * __alloc_pages() routine only calls here with __GFP_HARDWALL bit
3509 * _not_ set if it's a GFP_KERNEL allocation, and all nodes in the
3510 * current tasks mems_allowed came up empty on the first pass over
3511 * the zonelist. So only GFP_KERNEL allocations, if all nodes in the
3512 * cpuset are short of memory, might require taking the callback_lock.
3514 * The first call here from mm/page_alloc:get_page_from_freelist()
3515 * has __GFP_HARDWALL set in gfp_mask, enforcing hardwall cpusets,
3516 * so no allocation on a node outside the cpuset is allowed (unless
3517 * in interrupt, of course).
3519 * The second pass through get_page_from_freelist() doesn't even call
3520 * here for GFP_ATOMIC calls. For those calls, the __alloc_pages()
3521 * variable 'wait' is not set, and the bit ALLOC_CPUSET is not set
3522 * in alloc_flags. That logic and the checks below have the combined
3524 * in_interrupt - any node ok (current task context irrelevant)
3525 * GFP_ATOMIC - any node ok
3526 * tsk_is_oom_victim - any node ok
3527 * GFP_KERNEL - any node in enclosing hardwalled cpuset ok
3528 * GFP_USER - only nodes in current tasks mems allowed ok.
3530 bool __cpuset_node_allowed(int node, gfp_t gfp_mask)
3532 struct cpuset *cs; /* current cpuset ancestors */
3533 int allowed; /* is allocation in zone z allowed? */
3534 unsigned long flags;
3538 if (node_isset(node, current->mems_allowed))
3541 * Allow tasks that have access to memory reserves because they have
3542 * been OOM killed to get memory anywhere.
3544 if (unlikely(tsk_is_oom_victim(current)))
3546 if (gfp_mask & __GFP_HARDWALL) /* If hardwall request, stop here */
3549 if (current->flags & PF_EXITING) /* Let dying task have memory */
3552 /* Not hardwall and node outside mems_allowed: scan up cpusets */
3553 spin_lock_irqsave(&callback_lock, flags);
3556 cs = nearest_hardwall_ancestor(task_cs(current));
3557 allowed = node_isset(node, cs->mems_allowed);
3560 spin_unlock_irqrestore(&callback_lock, flags);
3565 * cpuset_mem_spread_node() - On which node to begin search for a file page
3566 * cpuset_slab_spread_node() - On which node to begin search for a slab page
3568 * If a task is marked PF_SPREAD_PAGE or PF_SPREAD_SLAB (as for
3569 * tasks in a cpuset with is_spread_page or is_spread_slab set),
3570 * and if the memory allocation used cpuset_mem_spread_node()
3571 * to determine on which node to start looking, as it will for
3572 * certain page cache or slab cache pages such as used for file
3573 * system buffers and inode caches, then instead of starting on the
3574 * local node to look for a free page, rather spread the starting
3575 * node around the tasks mems_allowed nodes.
3577 * We don't have to worry about the returned node being offline
3578 * because "it can't happen", and even if it did, it would be ok.
3580 * The routines calling guarantee_online_mems() are careful to
3581 * only set nodes in task->mems_allowed that are online. So it
3582 * should not be possible for the following code to return an
3583 * offline node. But if it did, that would be ok, as this routine
3584 * is not returning the node where the allocation must be, only
3585 * the node where the search should start. The zonelist passed to
3586 * __alloc_pages() will include all nodes. If the slab allocator
3587 * is passed an offline node, it will fall back to the local node.
3588 * See kmem_cache_alloc_node().
3591 static int cpuset_spread_node(int *rotor)
3593 return *rotor = next_node_in(*rotor, current->mems_allowed);
3596 int cpuset_mem_spread_node(void)
3598 if (current->cpuset_mem_spread_rotor == NUMA_NO_NODE)
3599 current->cpuset_mem_spread_rotor =
3600 node_random(¤t->mems_allowed);
3602 return cpuset_spread_node(¤t->cpuset_mem_spread_rotor);
3605 int cpuset_slab_spread_node(void)
3607 if (current->cpuset_slab_spread_rotor == NUMA_NO_NODE)
3608 current->cpuset_slab_spread_rotor =
3609 node_random(¤t->mems_allowed);
3611 return cpuset_spread_node(¤t->cpuset_slab_spread_rotor);
3614 EXPORT_SYMBOL_GPL(cpuset_mem_spread_node);
3617 * cpuset_mems_allowed_intersects - Does @tsk1's mems_allowed intersect @tsk2's?
3618 * @tsk1: pointer to task_struct of some task.
3619 * @tsk2: pointer to task_struct of some other task.
3621 * Description: Return true if @tsk1's mems_allowed intersects the
3622 * mems_allowed of @tsk2. Used by the OOM killer to determine if
3623 * one of the task's memory usage might impact the memory available
3627 int cpuset_mems_allowed_intersects(const struct task_struct *tsk1,
3628 const struct task_struct *tsk2)
3630 return nodes_intersects(tsk1->mems_allowed, tsk2->mems_allowed);
3634 * cpuset_print_current_mems_allowed - prints current's cpuset and mems_allowed
3636 * Description: Prints current's name, cpuset name, and cached copy of its
3637 * mems_allowed to the kernel log.
3639 void cpuset_print_current_mems_allowed(void)
3641 struct cgroup *cgrp;
3645 cgrp = task_cs(current)->css.cgroup;
3646 pr_cont(",cpuset=");
3647 pr_cont_cgroup_name(cgrp);
3648 pr_cont(",mems_allowed=%*pbl",
3649 nodemask_pr_args(¤t->mems_allowed));
3655 * Collection of memory_pressure is suppressed unless
3656 * this flag is enabled by writing "1" to the special
3657 * cpuset file 'memory_pressure_enabled' in the root cpuset.
3660 int cpuset_memory_pressure_enabled __read_mostly;
3663 * cpuset_memory_pressure_bump - keep stats of per-cpuset reclaims.
3665 * Keep a running average of the rate of synchronous (direct)
3666 * page reclaim efforts initiated by tasks in each cpuset.
3668 * This represents the rate at which some task in the cpuset
3669 * ran low on memory on all nodes it was allowed to use, and
3670 * had to enter the kernels page reclaim code in an effort to
3671 * create more free memory by tossing clean pages or swapping
3672 * or writing dirty pages.
3674 * Display to user space in the per-cpuset read-only file
3675 * "memory_pressure". Value displayed is an integer
3676 * representing the recent rate of entry into the synchronous
3677 * (direct) page reclaim by any task attached to the cpuset.
3680 void __cpuset_memory_pressure_bump(void)
3683 fmeter_markevent(&task_cs(current)->fmeter);
3687 #ifdef CONFIG_PROC_PID_CPUSET
3689 * proc_cpuset_show()
3690 * - Print tasks cpuset path into seq_file.
3691 * - Used for /proc/<pid>/cpuset.
3692 * - No need to task_lock(tsk) on this tsk->cpuset reference, as it
3693 * doesn't really matter if tsk->cpuset changes after we read it,
3694 * and we take cpuset_rwsem, keeping cpuset_attach() from changing it
3697 int proc_cpuset_show(struct seq_file *m, struct pid_namespace *ns,
3698 struct pid *pid, struct task_struct *tsk)
3701 struct cgroup_subsys_state *css;
3705 buf = kmalloc(PATH_MAX, GFP_KERNEL);
3709 css = task_get_css(tsk, cpuset_cgrp_id);
3710 retval = cgroup_path_ns(css->cgroup, buf, PATH_MAX,
3711 current->nsproxy->cgroup_ns);
3713 if (retval >= PATH_MAX)
3714 retval = -ENAMETOOLONG;
3725 #endif /* CONFIG_PROC_PID_CPUSET */
3727 /* Display task mems_allowed in /proc/<pid>/status file. */
3728 void cpuset_task_status_allowed(struct seq_file *m, struct task_struct *task)
3730 seq_printf(m, "Mems_allowed:\t%*pb\n",
3731 nodemask_pr_args(&task->mems_allowed));
3732 seq_printf(m, "Mems_allowed_list:\t%*pbl\n",
3733 nodemask_pr_args(&task->mems_allowed));