4 * Processor and Memory placement constraints for sets of tasks.
6 * Copyright (C) 2003 BULL SA.
7 * Copyright (C) 2004-2007 Silicon Graphics, Inc.
8 * Copyright (C) 2006 Google, Inc
10 * Portions derived from Patrick Mochel's sysfs code.
11 * sysfs is Copyright (c) 2001-3 Patrick Mochel
13 * 2003-10-10 Written by Simon Derr.
14 * 2003-10-22 Updates by Stephen Hemminger.
15 * 2004 May-July Rework by Paul Jackson.
16 * 2006 Rework by Paul Menage to use generic cgroups
17 * 2008 Rework of the scheduler domains and CPU hotplug handling
20 * This file is subject to the terms and conditions of the GNU General Public
21 * License. See the file COPYING in the main directory of the Linux
22 * distribution for more details.
25 #include <linux/cpu.h>
26 #include <linux/cpumask.h>
27 #include <linux/cpuset.h>
28 #include <linux/err.h>
29 #include <linux/errno.h>
30 #include <linux/file.h>
32 #include <linux/init.h>
33 #include <linux/interrupt.h>
34 #include <linux/kernel.h>
35 #include <linux/kmod.h>
36 #include <linux/list.h>
37 #include <linux/mempolicy.h>
39 #include <linux/memory.h>
40 #include <linux/export.h>
41 #include <linux/mount.h>
42 #include <linux/fs_context.h>
43 #include <linux/namei.h>
44 #include <linux/pagemap.h>
45 #include <linux/proc_fs.h>
46 #include <linux/rcupdate.h>
47 #include <linux/sched.h>
48 #include <linux/sched/deadline.h>
49 #include <linux/sched/mm.h>
50 #include <linux/sched/task.h>
51 #include <linux/seq_file.h>
52 #include <linux/security.h>
53 #include <linux/slab.h>
54 #include <linux/spinlock.h>
55 #include <linux/stat.h>
56 #include <linux/string.h>
57 #include <linux/time.h>
58 #include <linux/time64.h>
59 #include <linux/backing-dev.h>
60 #include <linux/sort.h>
61 #include <linux/oom.h>
62 #include <linux/sched/isolation.h>
63 #include <linux/uaccess.h>
64 #include <linux/atomic.h>
65 #include <linux/mutex.h>
66 #include <linux/cgroup.h>
67 #include <linux/wait.h>
69 DEFINE_STATIC_KEY_FALSE(cpusets_pre_enable_key);
70 DEFINE_STATIC_KEY_FALSE(cpusets_enabled_key);
72 /* See "Frequency meter" comments, below. */
75 int cnt; /* unprocessed events count */
76 int val; /* most recent output value */
77 time64_t time; /* clock (secs) when val computed */
78 spinlock_t lock; /* guards read or write of above */
82 struct cgroup_subsys_state css;
84 unsigned long flags; /* "unsigned long" so bitops work */
87 * On default hierarchy:
89 * The user-configured masks can only be changed by writing to
90 * cpuset.cpus and cpuset.mems, and won't be limited by the
93 * The effective masks is the real masks that apply to the tasks
94 * in the cpuset. They may be changed if the configured masks are
95 * changed or hotplug happens.
97 * effective_mask == configured_mask & parent's effective_mask,
98 * and if it ends up empty, it will inherit the parent's mask.
101 * On legacy hierarchy:
103 * The user-configured masks are always the same with effective masks.
106 /* user-configured CPUs and Memory Nodes allow to tasks */
107 cpumask_var_t cpus_allowed;
108 nodemask_t mems_allowed;
110 /* effective CPUs and Memory Nodes allow to tasks */
111 cpumask_var_t effective_cpus;
112 nodemask_t effective_mems;
115 * CPUs allocated to child sub-partitions (default hierarchy only)
116 * - CPUs granted by the parent = effective_cpus U subparts_cpus
117 * - effective_cpus and subparts_cpus are mutually exclusive.
119 * effective_cpus contains only onlined CPUs, but subparts_cpus
120 * may have offlined ones.
122 cpumask_var_t subparts_cpus;
125 * This is old Memory Nodes tasks took on.
127 * - top_cpuset.old_mems_allowed is initialized to mems_allowed.
128 * - A new cpuset's old_mems_allowed is initialized when some
129 * task is moved into it.
130 * - old_mems_allowed is used in cpuset_migrate_mm() when we change
131 * cpuset.mems_allowed and have tasks' nodemask updated, and
132 * then old_mems_allowed is updated to mems_allowed.
134 nodemask_t old_mems_allowed;
136 struct fmeter fmeter; /* memory_pressure filter */
139 * Tasks are being attached to this cpuset. Used to prevent
140 * zeroing cpus/mems_allowed between ->can_attach() and ->attach().
142 int attach_in_progress;
144 /* partition number for rebuild_sched_domains() */
147 /* for custom sched domain */
148 int relax_domain_level;
150 /* number of CPUs in subparts_cpus */
151 int nr_subparts_cpus;
153 /* partition root state */
154 int partition_root_state;
157 * Default hierarchy only:
158 * use_parent_ecpus - set if using parent's effective_cpus
159 * child_ecpus_count - # of children with use_parent_ecpus set
161 int use_parent_ecpus;
162 int child_ecpus_count;
164 /* Handle for cpuset.cpus.partition */
165 struct cgroup_file partition_file;
169 * Partition root states:
171 * 0 - not a partition root
175 * -1 - invalid partition root
176 * None of the cpus in cpus_allowed can be put into the parent's
177 * subparts_cpus. In this case, the cpuset is not a real partition
178 * root anymore. However, the CPU_EXCLUSIVE bit will still be set
179 * and the cpuset can be restored back to a partition root if the
180 * parent cpuset can give more CPUs back to this child cpuset.
182 #define PRS_DISABLED 0
183 #define PRS_ENABLED 1
187 * Temporary cpumasks for working with partitions that are passed among
188 * functions to avoid memory allocation in inner functions.
191 cpumask_var_t addmask, delmask; /* For partition root */
192 cpumask_var_t new_cpus; /* For update_cpumasks_hier() */
195 static inline struct cpuset *css_cs(struct cgroup_subsys_state *css)
197 return css ? container_of(css, struct cpuset, css) : NULL;
200 /* Retrieve the cpuset for a task */
201 static inline struct cpuset *task_cs(struct task_struct *task)
203 return css_cs(task_css(task, cpuset_cgrp_id));
206 static inline struct cpuset *parent_cs(struct cpuset *cs)
208 return css_cs(cs->css.parent);
211 /* bits in struct cpuset flags field */
218 CS_SCHED_LOAD_BALANCE,
223 /* convenient tests for these bits */
224 static inline bool is_cpuset_online(struct cpuset *cs)
226 return test_bit(CS_ONLINE, &cs->flags) && !css_is_dying(&cs->css);
229 static inline int is_cpu_exclusive(const struct cpuset *cs)
231 return test_bit(CS_CPU_EXCLUSIVE, &cs->flags);
234 static inline int is_mem_exclusive(const struct cpuset *cs)
236 return test_bit(CS_MEM_EXCLUSIVE, &cs->flags);
239 static inline int is_mem_hardwall(const struct cpuset *cs)
241 return test_bit(CS_MEM_HARDWALL, &cs->flags);
244 static inline int is_sched_load_balance(const struct cpuset *cs)
246 return test_bit(CS_SCHED_LOAD_BALANCE, &cs->flags);
249 static inline int is_memory_migrate(const struct cpuset *cs)
251 return test_bit(CS_MEMORY_MIGRATE, &cs->flags);
254 static inline int is_spread_page(const struct cpuset *cs)
256 return test_bit(CS_SPREAD_PAGE, &cs->flags);
259 static inline int is_spread_slab(const struct cpuset *cs)
261 return test_bit(CS_SPREAD_SLAB, &cs->flags);
264 static inline int is_partition_root(const struct cpuset *cs)
266 return cs->partition_root_state > 0;
270 * Send notification event of whenever partition_root_state changes.
272 static inline void notify_partition_change(struct cpuset *cs,
273 int old_prs, int new_prs)
275 if (old_prs != new_prs)
276 cgroup_file_notify(&cs->partition_file);
279 static struct cpuset top_cpuset = {
280 .flags = ((1 << CS_ONLINE) | (1 << CS_CPU_EXCLUSIVE) |
281 (1 << CS_MEM_EXCLUSIVE)),
282 .partition_root_state = PRS_ENABLED,
286 * cpuset_for_each_child - traverse online children of a cpuset
287 * @child_cs: loop cursor pointing to the current child
288 * @pos_css: used for iteration
289 * @parent_cs: target cpuset to walk children of
291 * Walk @child_cs through the online children of @parent_cs. Must be used
292 * with RCU read locked.
294 #define cpuset_for_each_child(child_cs, pos_css, parent_cs) \
295 css_for_each_child((pos_css), &(parent_cs)->css) \
296 if (is_cpuset_online(((child_cs) = css_cs((pos_css)))))
299 * cpuset_for_each_descendant_pre - pre-order walk of a cpuset's descendants
300 * @des_cs: loop cursor pointing to the current descendant
301 * @pos_css: used for iteration
302 * @root_cs: target cpuset to walk ancestor of
304 * Walk @des_cs through the online descendants of @root_cs. Must be used
305 * with RCU read locked. The caller may modify @pos_css by calling
306 * css_rightmost_descendant() to skip subtree. @root_cs is included in the
307 * iteration and the first node to be visited.
309 #define cpuset_for_each_descendant_pre(des_cs, pos_css, root_cs) \
310 css_for_each_descendant_pre((pos_css), &(root_cs)->css) \
311 if (is_cpuset_online(((des_cs) = css_cs((pos_css)))))
314 * There are two global locks guarding cpuset structures - cpuset_rwsem and
315 * callback_lock. We also require taking task_lock() when dereferencing a
316 * task's cpuset pointer. See "The task_lock() exception", at the end of this
317 * comment. The cpuset code uses only cpuset_rwsem write lock. Other
318 * kernel subsystems can use cpuset_read_lock()/cpuset_read_unlock() to
319 * prevent change to cpuset structures.
321 * A task must hold both locks to modify cpusets. If a task holds
322 * cpuset_rwsem, it blocks others wanting that rwsem, ensuring that it
323 * is the only task able to also acquire callback_lock and be able to
324 * modify cpusets. It can perform various checks on the cpuset structure
325 * first, knowing nothing will change. It can also allocate memory while
326 * just holding cpuset_rwsem. While it is performing these checks, various
327 * callback routines can briefly acquire callback_lock to query cpusets.
328 * Once it is ready to make the changes, it takes callback_lock, blocking
331 * Calls to the kernel memory allocator can not be made while holding
332 * callback_lock, as that would risk double tripping on callback_lock
333 * from one of the callbacks into the cpuset code from within
336 * If a task is only holding callback_lock, then it has read-only
339 * Now, the task_struct fields mems_allowed and mempolicy may be changed
340 * by other task, we use alloc_lock in the task_struct fields to protect
343 * The cpuset_common_file_read() handlers only hold callback_lock across
344 * small pieces of code, such as when reading out possibly multi-word
345 * cpumasks and nodemasks.
347 * Accessing a task's cpuset should be done in accordance with the
348 * guidelines for accessing subsystem state in kernel/cgroup.c
351 DEFINE_STATIC_PERCPU_RWSEM(cpuset_rwsem);
353 void cpuset_read_lock(void)
355 percpu_down_read(&cpuset_rwsem);
358 void cpuset_read_unlock(void)
360 percpu_up_read(&cpuset_rwsem);
363 static DEFINE_SPINLOCK(callback_lock);
365 static struct workqueue_struct *cpuset_migrate_mm_wq;
368 * CPU / memory hotplug is handled asynchronously.
370 static void cpuset_hotplug_workfn(struct work_struct *work);
371 static DECLARE_WORK(cpuset_hotplug_work, cpuset_hotplug_workfn);
373 static DECLARE_WAIT_QUEUE_HEAD(cpuset_attach_wq);
376 * Cgroup v2 behavior is used on the "cpus" and "mems" control files when
377 * on default hierarchy or when the cpuset_v2_mode flag is set by mounting
378 * the v1 cpuset cgroup filesystem with the "cpuset_v2_mode" mount option.
379 * With v2 behavior, "cpus" and "mems" are always what the users have
380 * requested and won't be changed by hotplug events. Only the effective
381 * cpus or mems will be affected.
383 static inline bool is_in_v2_mode(void)
385 return cgroup_subsys_on_dfl(cpuset_cgrp_subsys) ||
386 (cpuset_cgrp_subsys.root->flags & CGRP_ROOT_CPUSET_V2_MODE);
390 * Return in pmask the portion of a task's cpusets's cpus_allowed that
391 * are online and are capable of running the task. If none are found,
392 * walk up the cpuset hierarchy until we find one that does have some
395 * One way or another, we guarantee to return some non-empty subset
396 * of cpu_online_mask.
398 * Call with callback_lock or cpuset_rwsem held.
400 static void guarantee_online_cpus(struct task_struct *tsk,
401 struct cpumask *pmask)
403 const struct cpumask *possible_mask = task_cpu_possible_mask(tsk);
406 if (WARN_ON(!cpumask_and(pmask, possible_mask, cpu_online_mask)))
407 cpumask_copy(pmask, cpu_online_mask);
412 while (!cpumask_intersects(cs->effective_cpus, pmask)) {
416 * The top cpuset doesn't have any online cpu as a
417 * consequence of a race between cpuset_hotplug_work
418 * and cpu hotplug notifier. But we know the top
419 * cpuset's effective_cpus is on its way to be
420 * identical to cpu_online_mask.
425 cpumask_and(pmask, pmask, cs->effective_cpus);
432 * Return in *pmask the portion of a cpusets's mems_allowed that
433 * are online, with memory. If none are online with memory, walk
434 * up the cpuset hierarchy until we find one that does have some
435 * online mems. The top cpuset always has some mems online.
437 * One way or another, we guarantee to return some non-empty subset
438 * of node_states[N_MEMORY].
440 * Call with callback_lock or cpuset_rwsem held.
442 static void guarantee_online_mems(struct cpuset *cs, nodemask_t *pmask)
444 while (!nodes_intersects(cs->effective_mems, node_states[N_MEMORY]))
446 nodes_and(*pmask, cs->effective_mems, node_states[N_MEMORY]);
450 * update task's spread flag if cpuset's page/slab spread flag is set
452 * Call with callback_lock or cpuset_rwsem held.
454 static void cpuset_update_task_spread_flag(struct cpuset *cs,
455 struct task_struct *tsk)
457 if (is_spread_page(cs))
458 task_set_spread_page(tsk);
460 task_clear_spread_page(tsk);
462 if (is_spread_slab(cs))
463 task_set_spread_slab(tsk);
465 task_clear_spread_slab(tsk);
469 * is_cpuset_subset(p, q) - Is cpuset p a subset of cpuset q?
471 * One cpuset is a subset of another if all its allowed CPUs and
472 * Memory Nodes are a subset of the other, and its exclusive flags
473 * are only set if the other's are set. Call holding cpuset_rwsem.
476 static int is_cpuset_subset(const struct cpuset *p, const struct cpuset *q)
478 return cpumask_subset(p->cpus_allowed, q->cpus_allowed) &&
479 nodes_subset(p->mems_allowed, q->mems_allowed) &&
480 is_cpu_exclusive(p) <= is_cpu_exclusive(q) &&
481 is_mem_exclusive(p) <= is_mem_exclusive(q);
485 * alloc_cpumasks - allocate three cpumasks for cpuset
486 * @cs: the cpuset that have cpumasks to be allocated.
487 * @tmp: the tmpmasks structure pointer
488 * Return: 0 if successful, -ENOMEM otherwise.
490 * Only one of the two input arguments should be non-NULL.
492 static inline int alloc_cpumasks(struct cpuset *cs, struct tmpmasks *tmp)
494 cpumask_var_t *pmask1, *pmask2, *pmask3;
497 pmask1 = &cs->cpus_allowed;
498 pmask2 = &cs->effective_cpus;
499 pmask3 = &cs->subparts_cpus;
501 pmask1 = &tmp->new_cpus;
502 pmask2 = &tmp->addmask;
503 pmask3 = &tmp->delmask;
506 if (!zalloc_cpumask_var(pmask1, GFP_KERNEL))
509 if (!zalloc_cpumask_var(pmask2, GFP_KERNEL))
512 if (!zalloc_cpumask_var(pmask3, GFP_KERNEL))
518 free_cpumask_var(*pmask2);
520 free_cpumask_var(*pmask1);
525 * free_cpumasks - free cpumasks in a tmpmasks structure
526 * @cs: the cpuset that have cpumasks to be free.
527 * @tmp: the tmpmasks structure pointer
529 static inline void free_cpumasks(struct cpuset *cs, struct tmpmasks *tmp)
532 free_cpumask_var(cs->cpus_allowed);
533 free_cpumask_var(cs->effective_cpus);
534 free_cpumask_var(cs->subparts_cpus);
537 free_cpumask_var(tmp->new_cpus);
538 free_cpumask_var(tmp->addmask);
539 free_cpumask_var(tmp->delmask);
544 * alloc_trial_cpuset - allocate a trial cpuset
545 * @cs: the cpuset that the trial cpuset duplicates
547 static struct cpuset *alloc_trial_cpuset(struct cpuset *cs)
549 struct cpuset *trial;
551 trial = kmemdup(cs, sizeof(*cs), GFP_KERNEL);
555 if (alloc_cpumasks(trial, NULL)) {
560 cpumask_copy(trial->cpus_allowed, cs->cpus_allowed);
561 cpumask_copy(trial->effective_cpus, cs->effective_cpus);
566 * free_cpuset - free the cpuset
567 * @cs: the cpuset to be freed
569 static inline void free_cpuset(struct cpuset *cs)
571 free_cpumasks(cs, NULL);
576 * validate_change() - Used to validate that any proposed cpuset change
577 * follows the structural rules for cpusets.
579 * If we replaced the flag and mask values of the current cpuset
580 * (cur) with those values in the trial cpuset (trial), would
581 * our various subset and exclusive rules still be valid? Presumes
584 * 'cur' is the address of an actual, in-use cpuset. Operations
585 * such as list traversal that depend on the actual address of the
586 * cpuset in the list must use cur below, not trial.
588 * 'trial' is the address of bulk structure copy of cur, with
589 * perhaps one or more of the fields cpus_allowed, mems_allowed,
590 * or flags changed to new, trial values.
592 * Return 0 if valid, -errno if not.
595 static int validate_change(struct cpuset *cur, struct cpuset *trial)
597 struct cgroup_subsys_state *css;
598 struct cpuset *c, *par;
603 /* Each of our child cpusets must be a subset of us */
605 cpuset_for_each_child(c, css, cur)
606 if (!is_cpuset_subset(c, trial))
609 /* Remaining checks don't apply to root cpuset */
611 if (cur == &top_cpuset)
614 par = parent_cs(cur);
616 /* On legacy hierarchy, we must be a subset of our parent cpuset. */
618 if (!is_in_v2_mode() && !is_cpuset_subset(trial, par))
622 * If either I or some sibling (!= me) is exclusive, we can't
626 cpuset_for_each_child(c, css, par) {
627 if ((is_cpu_exclusive(trial) || is_cpu_exclusive(c)) &&
629 cpumask_intersects(trial->cpus_allowed, c->cpus_allowed))
631 if ((is_mem_exclusive(trial) || is_mem_exclusive(c)) &&
633 nodes_intersects(trial->mems_allowed, c->mems_allowed))
638 * Cpusets with tasks - existing or newly being attached - can't
639 * be changed to have empty cpus_allowed or mems_allowed.
642 if ((cgroup_is_populated(cur->css.cgroup) || cur->attach_in_progress)) {
643 if (!cpumask_empty(cur->cpus_allowed) &&
644 cpumask_empty(trial->cpus_allowed))
646 if (!nodes_empty(cur->mems_allowed) &&
647 nodes_empty(trial->mems_allowed))
652 * We can't shrink if we won't have enough room for SCHED_DEADLINE
656 if (is_cpu_exclusive(cur) &&
657 !cpuset_cpumask_can_shrink(cur->cpus_allowed,
658 trial->cpus_allowed))
669 * Helper routine for generate_sched_domains().
670 * Do cpusets a, b have overlapping effective cpus_allowed masks?
672 static int cpusets_overlap(struct cpuset *a, struct cpuset *b)
674 return cpumask_intersects(a->effective_cpus, b->effective_cpus);
678 update_domain_attr(struct sched_domain_attr *dattr, struct cpuset *c)
680 if (dattr->relax_domain_level < c->relax_domain_level)
681 dattr->relax_domain_level = c->relax_domain_level;
685 static void update_domain_attr_tree(struct sched_domain_attr *dattr,
686 struct cpuset *root_cs)
689 struct cgroup_subsys_state *pos_css;
692 cpuset_for_each_descendant_pre(cp, pos_css, root_cs) {
693 /* skip the whole subtree if @cp doesn't have any CPU */
694 if (cpumask_empty(cp->cpus_allowed)) {
695 pos_css = css_rightmost_descendant(pos_css);
699 if (is_sched_load_balance(cp))
700 update_domain_attr(dattr, cp);
705 /* Must be called with cpuset_rwsem held. */
706 static inline int nr_cpusets(void)
708 /* jump label reference count + the top-level cpuset */
709 return static_key_count(&cpusets_enabled_key.key) + 1;
713 * generate_sched_domains()
715 * This function builds a partial partition of the systems CPUs
716 * A 'partial partition' is a set of non-overlapping subsets whose
717 * union is a subset of that set.
718 * The output of this function needs to be passed to kernel/sched/core.c
719 * partition_sched_domains() routine, which will rebuild the scheduler's
720 * load balancing domains (sched domains) as specified by that partial
723 * See "What is sched_load_balance" in Documentation/admin-guide/cgroup-v1/cpusets.rst
724 * for a background explanation of this.
726 * Does not return errors, on the theory that the callers of this
727 * routine would rather not worry about failures to rebuild sched
728 * domains when operating in the severe memory shortage situations
729 * that could cause allocation failures below.
731 * Must be called with cpuset_rwsem held.
733 * The three key local variables below are:
734 * cp - cpuset pointer, used (together with pos_css) to perform a
735 * top-down scan of all cpusets. For our purposes, rebuilding
736 * the schedulers sched domains, we can ignore !is_sched_load_
738 * csa - (for CpuSet Array) Array of pointers to all the cpusets
739 * that need to be load balanced, for convenient iterative
740 * access by the subsequent code that finds the best partition,
741 * i.e the set of domains (subsets) of CPUs such that the
742 * cpus_allowed of every cpuset marked is_sched_load_balance
743 * is a subset of one of these domains, while there are as
744 * many such domains as possible, each as small as possible.
745 * doms - Conversion of 'csa' to an array of cpumasks, for passing to
746 * the kernel/sched/core.c routine partition_sched_domains() in a
747 * convenient format, that can be easily compared to the prior
748 * value to determine what partition elements (sched domains)
749 * were changed (added or removed.)
751 * Finding the best partition (set of domains):
752 * The triple nested loops below over i, j, k scan over the
753 * load balanced cpusets (using the array of cpuset pointers in
754 * csa[]) looking for pairs of cpusets that have overlapping
755 * cpus_allowed, but which don't have the same 'pn' partition
756 * number and gives them in the same partition number. It keeps
757 * looping on the 'restart' label until it can no longer find
760 * The union of the cpus_allowed masks from the set of
761 * all cpusets having the same 'pn' value then form the one
762 * element of the partition (one sched domain) to be passed to
763 * partition_sched_domains().
765 static int generate_sched_domains(cpumask_var_t **domains,
766 struct sched_domain_attr **attributes)
768 struct cpuset *cp; /* top-down scan of cpusets */
769 struct cpuset **csa; /* array of all cpuset ptrs */
770 int csn; /* how many cpuset ptrs in csa so far */
771 int i, j, k; /* indices for partition finding loops */
772 cpumask_var_t *doms; /* resulting partition; i.e. sched domains */
773 struct sched_domain_attr *dattr; /* attributes for custom domains */
774 int ndoms = 0; /* number of sched domains in result */
775 int nslot; /* next empty doms[] struct cpumask slot */
776 struct cgroup_subsys_state *pos_css;
777 bool root_load_balance = is_sched_load_balance(&top_cpuset);
783 /* Special case for the 99% of systems with one, full, sched domain */
784 if (root_load_balance && !top_cpuset.nr_subparts_cpus) {
786 doms = alloc_sched_domains(ndoms);
790 dattr = kmalloc(sizeof(struct sched_domain_attr), GFP_KERNEL);
792 *dattr = SD_ATTR_INIT;
793 update_domain_attr_tree(dattr, &top_cpuset);
795 cpumask_and(doms[0], top_cpuset.effective_cpus,
796 housekeeping_cpumask(HK_FLAG_DOMAIN));
801 csa = kmalloc_array(nr_cpusets(), sizeof(cp), GFP_KERNEL);
807 if (root_load_balance)
808 csa[csn++] = &top_cpuset;
809 cpuset_for_each_descendant_pre(cp, pos_css, &top_cpuset) {
810 if (cp == &top_cpuset)
813 * Continue traversing beyond @cp iff @cp has some CPUs and
814 * isn't load balancing. The former is obvious. The
815 * latter: All child cpusets contain a subset of the
816 * parent's cpus, so just skip them, and then we call
817 * update_domain_attr_tree() to calc relax_domain_level of
818 * the corresponding sched domain.
820 * If root is load-balancing, we can skip @cp if it
821 * is a subset of the root's effective_cpus.
823 if (!cpumask_empty(cp->cpus_allowed) &&
824 !(is_sched_load_balance(cp) &&
825 cpumask_intersects(cp->cpus_allowed,
826 housekeeping_cpumask(HK_FLAG_DOMAIN))))
829 if (root_load_balance &&
830 cpumask_subset(cp->cpus_allowed, top_cpuset.effective_cpus))
833 if (is_sched_load_balance(cp) &&
834 !cpumask_empty(cp->effective_cpus))
837 /* skip @cp's subtree if not a partition root */
838 if (!is_partition_root(cp))
839 pos_css = css_rightmost_descendant(pos_css);
843 for (i = 0; i < csn; i++)
848 /* Find the best partition (set of sched domains) */
849 for (i = 0; i < csn; i++) {
850 struct cpuset *a = csa[i];
853 for (j = 0; j < csn; j++) {
854 struct cpuset *b = csa[j];
857 if (apn != bpn && cpusets_overlap(a, b)) {
858 for (k = 0; k < csn; k++) {
859 struct cpuset *c = csa[k];
864 ndoms--; /* one less element */
871 * Now we know how many domains to create.
872 * Convert <csn, csa> to <ndoms, doms> and populate cpu masks.
874 doms = alloc_sched_domains(ndoms);
879 * The rest of the code, including the scheduler, can deal with
880 * dattr==NULL case. No need to abort if alloc fails.
882 dattr = kmalloc_array(ndoms, sizeof(struct sched_domain_attr),
885 for (nslot = 0, i = 0; i < csn; i++) {
886 struct cpuset *a = csa[i];
891 /* Skip completed partitions */
897 if (nslot == ndoms) {
898 static int warnings = 10;
900 pr_warn("rebuild_sched_domains confused: nslot %d, ndoms %d, csn %d, i %d, apn %d\n",
901 nslot, ndoms, csn, i, apn);
909 *(dattr + nslot) = SD_ATTR_INIT;
910 for (j = i; j < csn; j++) {
911 struct cpuset *b = csa[j];
914 cpumask_or(dp, dp, b->effective_cpus);
915 cpumask_and(dp, dp, housekeeping_cpumask(HK_FLAG_DOMAIN));
917 update_domain_attr_tree(dattr + nslot, b);
919 /* Done with this partition */
925 BUG_ON(nslot != ndoms);
931 * Fallback to the default domain if kmalloc() failed.
932 * See comments in partition_sched_domains().
942 static void update_tasks_root_domain(struct cpuset *cs)
944 struct css_task_iter it;
945 struct task_struct *task;
947 css_task_iter_start(&cs->css, 0, &it);
949 while ((task = css_task_iter_next(&it)))
950 dl_add_task_root_domain(task);
952 css_task_iter_end(&it);
955 static void rebuild_root_domains(void)
957 struct cpuset *cs = NULL;
958 struct cgroup_subsys_state *pos_css;
960 percpu_rwsem_assert_held(&cpuset_rwsem);
961 lockdep_assert_cpus_held();
962 lockdep_assert_held(&sched_domains_mutex);
967 * Clear default root domain DL accounting, it will be computed again
968 * if a task belongs to it.
970 dl_clear_root_domain(&def_root_domain);
972 cpuset_for_each_descendant_pre(cs, pos_css, &top_cpuset) {
974 if (cpumask_empty(cs->effective_cpus)) {
975 pos_css = css_rightmost_descendant(pos_css);
983 update_tasks_root_domain(cs);
992 partition_and_rebuild_sched_domains(int ndoms_new, cpumask_var_t doms_new[],
993 struct sched_domain_attr *dattr_new)
995 mutex_lock(&sched_domains_mutex);
996 partition_sched_domains_locked(ndoms_new, doms_new, dattr_new);
997 rebuild_root_domains();
998 mutex_unlock(&sched_domains_mutex);
1002 * Rebuild scheduler domains.
1004 * If the flag 'sched_load_balance' of any cpuset with non-empty
1005 * 'cpus' changes, or if the 'cpus' allowed changes in any cpuset
1006 * which has that flag enabled, or if any cpuset with a non-empty
1007 * 'cpus' is removed, then call this routine to rebuild the
1008 * scheduler's dynamic sched domains.
1010 * Call with cpuset_rwsem held. Takes cpus_read_lock().
1012 static void rebuild_sched_domains_locked(void)
1014 struct cgroup_subsys_state *pos_css;
1015 struct sched_domain_attr *attr;
1016 cpumask_var_t *doms;
1020 lockdep_assert_cpus_held();
1021 percpu_rwsem_assert_held(&cpuset_rwsem);
1024 * If we have raced with CPU hotplug, return early to avoid
1025 * passing doms with offlined cpu to partition_sched_domains().
1026 * Anyways, cpuset_hotplug_workfn() will rebuild sched domains.
1028 * With no CPUs in any subpartitions, top_cpuset's effective CPUs
1029 * should be the same as the active CPUs, so checking only top_cpuset
1030 * is enough to detect racing CPU offlines.
1032 if (!top_cpuset.nr_subparts_cpus &&
1033 !cpumask_equal(top_cpuset.effective_cpus, cpu_active_mask))
1037 * With subpartition CPUs, however, the effective CPUs of a partition
1038 * root should be only a subset of the active CPUs. Since a CPU in any
1039 * partition root could be offlined, all must be checked.
1041 if (top_cpuset.nr_subparts_cpus) {
1043 cpuset_for_each_descendant_pre(cs, pos_css, &top_cpuset) {
1044 if (!is_partition_root(cs)) {
1045 pos_css = css_rightmost_descendant(pos_css);
1048 if (!cpumask_subset(cs->effective_cpus,
1057 /* Generate domain masks and attrs */
1058 ndoms = generate_sched_domains(&doms, &attr);
1060 /* Have scheduler rebuild the domains */
1061 partition_and_rebuild_sched_domains(ndoms, doms, attr);
1063 #else /* !CONFIG_SMP */
1064 static void rebuild_sched_domains_locked(void)
1067 #endif /* CONFIG_SMP */
1069 void rebuild_sched_domains(void)
1072 percpu_down_write(&cpuset_rwsem);
1073 rebuild_sched_domains_locked();
1074 percpu_up_write(&cpuset_rwsem);
1079 * update_tasks_cpumask - Update the cpumasks of tasks in the cpuset.
1080 * @cs: the cpuset in which each task's cpus_allowed mask needs to be changed
1082 * Iterate through each task of @cs updating its cpus_allowed to the
1083 * effective cpuset's. As this function is called with cpuset_rwsem held,
1084 * cpuset membership stays stable.
1086 static void update_tasks_cpumask(struct cpuset *cs)
1088 struct css_task_iter it;
1089 struct task_struct *task;
1091 css_task_iter_start(&cs->css, 0, &it);
1092 while ((task = css_task_iter_next(&it)))
1093 set_cpus_allowed_ptr(task, cs->effective_cpus);
1094 css_task_iter_end(&it);
1098 * compute_effective_cpumask - Compute the effective cpumask of the cpuset
1099 * @new_cpus: the temp variable for the new effective_cpus mask
1100 * @cs: the cpuset the need to recompute the new effective_cpus mask
1101 * @parent: the parent cpuset
1103 * If the parent has subpartition CPUs, include them in the list of
1104 * allowable CPUs in computing the new effective_cpus mask. Since offlined
1105 * CPUs are not removed from subparts_cpus, we have to use cpu_active_mask
1106 * to mask those out.
1108 static void compute_effective_cpumask(struct cpumask *new_cpus,
1109 struct cpuset *cs, struct cpuset *parent)
1111 if (parent->nr_subparts_cpus) {
1112 cpumask_or(new_cpus, parent->effective_cpus,
1113 parent->subparts_cpus);
1114 cpumask_and(new_cpus, new_cpus, cs->cpus_allowed);
1115 cpumask_and(new_cpus, new_cpus, cpu_active_mask);
1117 cpumask_and(new_cpus, cs->cpus_allowed, parent->effective_cpus);
1122 * Commands for update_parent_subparts_cpumask
1125 partcmd_enable, /* Enable partition root */
1126 partcmd_disable, /* Disable partition root */
1127 partcmd_update, /* Update parent's subparts_cpus */
1131 * update_parent_subparts_cpumask - update subparts_cpus mask of parent cpuset
1132 * @cpuset: The cpuset that requests change in partition root state
1133 * @cmd: Partition root state change command
1134 * @newmask: Optional new cpumask for partcmd_update
1135 * @tmp: Temporary addmask and delmask
1136 * Return: 0, 1 or an error code
1138 * For partcmd_enable, the cpuset is being transformed from a non-partition
1139 * root to a partition root. The cpus_allowed mask of the given cpuset will
1140 * be put into parent's subparts_cpus and taken away from parent's
1141 * effective_cpus. The function will return 0 if all the CPUs listed in
1142 * cpus_allowed can be granted or an error code will be returned.
1144 * For partcmd_disable, the cpuset is being transofrmed from a partition
1145 * root back to a non-partition root. Any CPUs in cpus_allowed that are in
1146 * parent's subparts_cpus will be taken away from that cpumask and put back
1147 * into parent's effective_cpus. 0 should always be returned.
1149 * For partcmd_update, if the optional newmask is specified, the cpu
1150 * list is to be changed from cpus_allowed to newmask. Otherwise,
1151 * cpus_allowed is assumed to remain the same. The cpuset should either
1152 * be a partition root or an invalid partition root. The partition root
1153 * state may change if newmask is NULL and none of the requested CPUs can
1154 * be granted by the parent. The function will return 1 if changes to
1155 * parent's subparts_cpus and effective_cpus happen or 0 otherwise.
1156 * Error code should only be returned when newmask is non-NULL.
1158 * The partcmd_enable and partcmd_disable commands are used by
1159 * update_prstate(). The partcmd_update command is used by
1160 * update_cpumasks_hier() with newmask NULL and update_cpumask() with
1163 * The checking is more strict when enabling partition root than the
1164 * other two commands.
1166 * Because of the implicit cpu exclusive nature of a partition root,
1167 * cpumask changes that violates the cpu exclusivity rule will not be
1168 * permitted when checked by validate_change(). The validate_change()
1169 * function will also prevent any changes to the cpu list if it is not
1170 * a superset of children's cpu lists.
1172 static int update_parent_subparts_cpumask(struct cpuset *cpuset, int cmd,
1173 struct cpumask *newmask,
1174 struct tmpmasks *tmp)
1176 struct cpuset *parent = parent_cs(cpuset);
1177 int adding; /* Moving cpus from effective_cpus to subparts_cpus */
1178 int deleting; /* Moving cpus from subparts_cpus to effective_cpus */
1179 int old_prs, new_prs;
1180 bool part_error = false; /* Partition error? */
1182 percpu_rwsem_assert_held(&cpuset_rwsem);
1185 * The parent must be a partition root.
1186 * The new cpumask, if present, or the current cpus_allowed must
1189 if (!is_partition_root(parent) ||
1190 (newmask && cpumask_empty(newmask)) ||
1191 (!newmask && cpumask_empty(cpuset->cpus_allowed)))
1195 * Enabling/disabling partition root is not allowed if there are
1198 if ((cmd != partcmd_update) && css_has_online_children(&cpuset->css))
1202 * Enabling partition root is not allowed if not all the CPUs
1203 * can be granted from parent's effective_cpus or at least one
1204 * CPU will be left after that.
1206 if ((cmd == partcmd_enable) &&
1207 (!cpumask_subset(cpuset->cpus_allowed, parent->effective_cpus) ||
1208 cpumask_equal(cpuset->cpus_allowed, parent->effective_cpus)))
1212 * A cpumask update cannot make parent's effective_cpus become empty.
1214 adding = deleting = false;
1215 old_prs = new_prs = cpuset->partition_root_state;
1216 if (cmd == partcmd_enable) {
1217 cpumask_copy(tmp->addmask, cpuset->cpus_allowed);
1219 } else if (cmd == partcmd_disable) {
1220 deleting = cpumask_and(tmp->delmask, cpuset->cpus_allowed,
1221 parent->subparts_cpus);
1222 } else if (newmask) {
1224 * partcmd_update with newmask:
1226 * delmask = cpus_allowed & ~newmask & parent->subparts_cpus
1227 * addmask = newmask & parent->effective_cpus
1228 * & ~parent->subparts_cpus
1230 cpumask_andnot(tmp->delmask, cpuset->cpus_allowed, newmask);
1231 deleting = cpumask_and(tmp->delmask, tmp->delmask,
1232 parent->subparts_cpus);
1234 cpumask_and(tmp->addmask, newmask, parent->effective_cpus);
1235 adding = cpumask_andnot(tmp->addmask, tmp->addmask,
1236 parent->subparts_cpus);
1238 * Return error if the new effective_cpus could become empty.
1241 cpumask_equal(parent->effective_cpus, tmp->addmask)) {
1245 * As some of the CPUs in subparts_cpus might have
1246 * been offlined, we need to compute the real delmask
1249 if (!cpumask_and(tmp->addmask, tmp->delmask,
1252 cpumask_copy(tmp->addmask, parent->effective_cpus);
1256 * partcmd_update w/o newmask:
1258 * addmask = cpus_allowed & parent->effective_cpus
1260 * Note that parent's subparts_cpus may have been
1261 * pre-shrunk in case there is a change in the cpu list.
1262 * So no deletion is needed.
1264 adding = cpumask_and(tmp->addmask, cpuset->cpus_allowed,
1265 parent->effective_cpus);
1266 part_error = cpumask_equal(tmp->addmask,
1267 parent->effective_cpus);
1270 if (cmd == partcmd_update) {
1271 int prev_prs = cpuset->partition_root_state;
1274 * Check for possible transition between PRS_ENABLED
1277 switch (cpuset->partition_root_state) {
1280 new_prs = PRS_ERROR;
1284 new_prs = PRS_ENABLED;
1288 * Set part_error if previously in invalid state.
1290 part_error = (prev_prs == PRS_ERROR);
1293 if (!part_error && (new_prs == PRS_ERROR))
1294 return 0; /* Nothing need to be done */
1296 if (new_prs == PRS_ERROR) {
1298 * Remove all its cpus from parent's subparts_cpus.
1301 deleting = cpumask_and(tmp->delmask, cpuset->cpus_allowed,
1302 parent->subparts_cpus);
1305 if (!adding && !deleting && (new_prs == old_prs))
1309 * Change the parent's subparts_cpus.
1310 * Newly added CPUs will be removed from effective_cpus and
1311 * newly deleted ones will be added back to effective_cpus.
1313 spin_lock_irq(&callback_lock);
1315 cpumask_or(parent->subparts_cpus,
1316 parent->subparts_cpus, tmp->addmask);
1317 cpumask_andnot(parent->effective_cpus,
1318 parent->effective_cpus, tmp->addmask);
1321 cpumask_andnot(parent->subparts_cpus,
1322 parent->subparts_cpus, tmp->delmask);
1324 * Some of the CPUs in subparts_cpus might have been offlined.
1326 cpumask_and(tmp->delmask, tmp->delmask, cpu_active_mask);
1327 cpumask_or(parent->effective_cpus,
1328 parent->effective_cpus, tmp->delmask);
1331 parent->nr_subparts_cpus = cpumask_weight(parent->subparts_cpus);
1333 if (old_prs != new_prs)
1334 cpuset->partition_root_state = new_prs;
1336 spin_unlock_irq(&callback_lock);
1337 notify_partition_change(cpuset, old_prs, new_prs);
1339 return cmd == partcmd_update;
1343 * update_cpumasks_hier - Update effective cpumasks and tasks in the subtree
1344 * @cs: the cpuset to consider
1345 * @tmp: temp variables for calculating effective_cpus & partition setup
1347 * When configured cpumask is changed, the effective cpumasks of this cpuset
1348 * and all its descendants need to be updated.
1350 * On legacy hierarchy, effective_cpus will be the same with cpu_allowed.
1352 * Called with cpuset_rwsem held
1354 static void update_cpumasks_hier(struct cpuset *cs, struct tmpmasks *tmp)
1357 struct cgroup_subsys_state *pos_css;
1358 bool need_rebuild_sched_domains = false;
1359 int old_prs, new_prs;
1362 cpuset_for_each_descendant_pre(cp, pos_css, cs) {
1363 struct cpuset *parent = parent_cs(cp);
1365 compute_effective_cpumask(tmp->new_cpus, cp, parent);
1368 * If it becomes empty, inherit the effective mask of the
1369 * parent, which is guaranteed to have some CPUs.
1371 if (is_in_v2_mode() && cpumask_empty(tmp->new_cpus)) {
1372 cpumask_copy(tmp->new_cpus, parent->effective_cpus);
1373 if (!cp->use_parent_ecpus) {
1374 cp->use_parent_ecpus = true;
1375 parent->child_ecpus_count++;
1377 } else if (cp->use_parent_ecpus) {
1378 cp->use_parent_ecpus = false;
1379 WARN_ON_ONCE(!parent->child_ecpus_count);
1380 parent->child_ecpus_count--;
1384 * Skip the whole subtree if the cpumask remains the same
1385 * and has no partition root state.
1387 if (!cp->partition_root_state &&
1388 cpumask_equal(tmp->new_cpus, cp->effective_cpus)) {
1389 pos_css = css_rightmost_descendant(pos_css);
1394 * update_parent_subparts_cpumask() should have been called
1395 * for cs already in update_cpumask(). We should also call
1396 * update_tasks_cpumask() again for tasks in the parent
1397 * cpuset if the parent's subparts_cpus changes.
1399 old_prs = new_prs = cp->partition_root_state;
1400 if ((cp != cs) && old_prs) {
1401 switch (parent->partition_root_state) {
1404 * If parent is not a partition root or an
1405 * invalid partition root, clear its state
1406 * and its CS_CPU_EXCLUSIVE flag.
1408 WARN_ON_ONCE(cp->partition_root_state
1410 new_prs = PRS_DISABLED;
1413 * clear_bit() is an atomic operation and
1414 * readers aren't interested in the state
1415 * of CS_CPU_EXCLUSIVE anyway. So we can
1416 * just update the flag without holding
1417 * the callback_lock.
1419 clear_bit(CS_CPU_EXCLUSIVE, &cp->flags);
1423 if (update_parent_subparts_cpumask(cp, partcmd_update, NULL, tmp))
1424 update_tasks_cpumask(parent);
1429 * When parent is invalid, it has to be too.
1431 new_prs = PRS_ERROR;
1436 if (!css_tryget_online(&cp->css))
1440 spin_lock_irq(&callback_lock);
1442 cpumask_copy(cp->effective_cpus, tmp->new_cpus);
1443 if (cp->nr_subparts_cpus && (new_prs != PRS_ENABLED)) {
1444 cp->nr_subparts_cpus = 0;
1445 cpumask_clear(cp->subparts_cpus);
1446 } else if (cp->nr_subparts_cpus) {
1448 * Make sure that effective_cpus & subparts_cpus
1449 * are mutually exclusive.
1451 * In the unlikely event that effective_cpus
1452 * becomes empty. we clear cp->nr_subparts_cpus and
1453 * let its child partition roots to compete for
1456 cpumask_andnot(cp->effective_cpus, cp->effective_cpus,
1458 if (cpumask_empty(cp->effective_cpus)) {
1459 cpumask_copy(cp->effective_cpus, tmp->new_cpus);
1460 cpumask_clear(cp->subparts_cpus);
1461 cp->nr_subparts_cpus = 0;
1462 } else if (!cpumask_subset(cp->subparts_cpus,
1464 cpumask_andnot(cp->subparts_cpus,
1465 cp->subparts_cpus, tmp->new_cpus);
1466 cp->nr_subparts_cpus
1467 = cpumask_weight(cp->subparts_cpus);
1471 if (new_prs != old_prs)
1472 cp->partition_root_state = new_prs;
1474 spin_unlock_irq(&callback_lock);
1475 notify_partition_change(cp, old_prs, new_prs);
1477 WARN_ON(!is_in_v2_mode() &&
1478 !cpumask_equal(cp->cpus_allowed, cp->effective_cpus));
1480 update_tasks_cpumask(cp);
1483 * On legacy hierarchy, if the effective cpumask of any non-
1484 * empty cpuset is changed, we need to rebuild sched domains.
1485 * On default hierarchy, the cpuset needs to be a partition
1488 if (!cpumask_empty(cp->cpus_allowed) &&
1489 is_sched_load_balance(cp) &&
1490 (!cgroup_subsys_on_dfl(cpuset_cgrp_subsys) ||
1491 is_partition_root(cp)))
1492 need_rebuild_sched_domains = true;
1499 if (need_rebuild_sched_domains)
1500 rebuild_sched_domains_locked();
1504 * update_sibling_cpumasks - Update siblings cpumasks
1505 * @parent: Parent cpuset
1506 * @cs: Current cpuset
1507 * @tmp: Temp variables
1509 static void update_sibling_cpumasks(struct cpuset *parent, struct cpuset *cs,
1510 struct tmpmasks *tmp)
1512 struct cpuset *sibling;
1513 struct cgroup_subsys_state *pos_css;
1515 percpu_rwsem_assert_held(&cpuset_rwsem);
1518 * Check all its siblings and call update_cpumasks_hier()
1519 * if their use_parent_ecpus flag is set in order for them
1520 * to use the right effective_cpus value.
1522 * The update_cpumasks_hier() function may sleep. So we have to
1523 * release the RCU read lock before calling it.
1526 cpuset_for_each_child(sibling, pos_css, parent) {
1529 if (!sibling->use_parent_ecpus)
1531 if (!css_tryget_online(&sibling->css))
1535 update_cpumasks_hier(sibling, tmp);
1537 css_put(&sibling->css);
1543 * update_cpumask - update the cpus_allowed mask of a cpuset and all tasks in it
1544 * @cs: the cpuset to consider
1545 * @trialcs: trial cpuset
1546 * @buf: buffer of cpu numbers written to this cpuset
1548 static int update_cpumask(struct cpuset *cs, struct cpuset *trialcs,
1552 struct tmpmasks tmp;
1554 /* top_cpuset.cpus_allowed tracks cpu_online_mask; it's read-only */
1555 if (cs == &top_cpuset)
1559 * An empty cpus_allowed is ok only if the cpuset has no tasks.
1560 * Since cpulist_parse() fails on an empty mask, we special case
1561 * that parsing. The validate_change() call ensures that cpusets
1562 * with tasks have cpus.
1565 cpumask_clear(trialcs->cpus_allowed);
1567 retval = cpulist_parse(buf, trialcs->cpus_allowed);
1571 if (!cpumask_subset(trialcs->cpus_allowed,
1572 top_cpuset.cpus_allowed))
1576 /* Nothing to do if the cpus didn't change */
1577 if (cpumask_equal(cs->cpus_allowed, trialcs->cpus_allowed))
1580 retval = validate_change(cs, trialcs);
1584 #ifdef CONFIG_CPUMASK_OFFSTACK
1586 * Use the cpumasks in trialcs for tmpmasks when they are pointers
1587 * to allocated cpumasks.
1589 tmp.addmask = trialcs->subparts_cpus;
1590 tmp.delmask = trialcs->effective_cpus;
1591 tmp.new_cpus = trialcs->cpus_allowed;
1594 if (cs->partition_root_state) {
1595 /* Cpumask of a partition root cannot be empty */
1596 if (cpumask_empty(trialcs->cpus_allowed))
1598 if (update_parent_subparts_cpumask(cs, partcmd_update,
1599 trialcs->cpus_allowed, &tmp) < 0)
1603 spin_lock_irq(&callback_lock);
1604 cpumask_copy(cs->cpus_allowed, trialcs->cpus_allowed);
1607 * Make sure that subparts_cpus is a subset of cpus_allowed.
1609 if (cs->nr_subparts_cpus) {
1610 cpumask_and(cs->subparts_cpus, cs->subparts_cpus, cs->cpus_allowed);
1611 cs->nr_subparts_cpus = cpumask_weight(cs->subparts_cpus);
1613 spin_unlock_irq(&callback_lock);
1615 update_cpumasks_hier(cs, &tmp);
1617 if (cs->partition_root_state) {
1618 struct cpuset *parent = parent_cs(cs);
1621 * For partition root, update the cpumasks of sibling
1622 * cpusets if they use parent's effective_cpus.
1624 if (parent->child_ecpus_count)
1625 update_sibling_cpumasks(parent, cs, &tmp);
1631 * Migrate memory region from one set of nodes to another. This is
1632 * performed asynchronously as it can be called from process migration path
1633 * holding locks involved in process management. All mm migrations are
1634 * performed in the queued order and can be waited for by flushing
1635 * cpuset_migrate_mm_wq.
1638 struct cpuset_migrate_mm_work {
1639 struct work_struct work;
1640 struct mm_struct *mm;
1645 static void cpuset_migrate_mm_workfn(struct work_struct *work)
1647 struct cpuset_migrate_mm_work *mwork =
1648 container_of(work, struct cpuset_migrate_mm_work, work);
1650 /* on a wq worker, no need to worry about %current's mems_allowed */
1651 do_migrate_pages(mwork->mm, &mwork->from, &mwork->to, MPOL_MF_MOVE_ALL);
1656 static void cpuset_migrate_mm(struct mm_struct *mm, const nodemask_t *from,
1657 const nodemask_t *to)
1659 struct cpuset_migrate_mm_work *mwork;
1661 if (nodes_equal(*from, *to)) {
1666 mwork = kzalloc(sizeof(*mwork), GFP_KERNEL);
1669 mwork->from = *from;
1671 INIT_WORK(&mwork->work, cpuset_migrate_mm_workfn);
1672 queue_work(cpuset_migrate_mm_wq, &mwork->work);
1678 static void cpuset_post_attach(void)
1680 flush_workqueue(cpuset_migrate_mm_wq);
1684 * cpuset_change_task_nodemask - change task's mems_allowed and mempolicy
1685 * @tsk: the task to change
1686 * @newmems: new nodes that the task will be set
1688 * We use the mems_allowed_seq seqlock to safely update both tsk->mems_allowed
1689 * and rebind an eventual tasks' mempolicy. If the task is allocating in
1690 * parallel, it might temporarily see an empty intersection, which results in
1691 * a seqlock check and retry before OOM or allocation failure.
1693 static void cpuset_change_task_nodemask(struct task_struct *tsk,
1694 nodemask_t *newmems)
1698 local_irq_disable();
1699 write_seqcount_begin(&tsk->mems_allowed_seq);
1701 nodes_or(tsk->mems_allowed, tsk->mems_allowed, *newmems);
1702 mpol_rebind_task(tsk, newmems);
1703 tsk->mems_allowed = *newmems;
1705 write_seqcount_end(&tsk->mems_allowed_seq);
1711 static void *cpuset_being_rebound;
1714 * update_tasks_nodemask - Update the nodemasks of tasks in the cpuset.
1715 * @cs: the cpuset in which each task's mems_allowed mask needs to be changed
1717 * Iterate through each task of @cs updating its mems_allowed to the
1718 * effective cpuset's. As this function is called with cpuset_rwsem held,
1719 * cpuset membership stays stable.
1721 static void update_tasks_nodemask(struct cpuset *cs)
1723 static nodemask_t newmems; /* protected by cpuset_rwsem */
1724 struct css_task_iter it;
1725 struct task_struct *task;
1727 cpuset_being_rebound = cs; /* causes mpol_dup() rebind */
1729 guarantee_online_mems(cs, &newmems);
1732 * The mpol_rebind_mm() call takes mmap_lock, which we couldn't
1733 * take while holding tasklist_lock. Forks can happen - the
1734 * mpol_dup() cpuset_being_rebound check will catch such forks,
1735 * and rebind their vma mempolicies too. Because we still hold
1736 * the global cpuset_rwsem, we know that no other rebind effort
1737 * will be contending for the global variable cpuset_being_rebound.
1738 * It's ok if we rebind the same mm twice; mpol_rebind_mm()
1739 * is idempotent. Also migrate pages in each mm to new nodes.
1741 css_task_iter_start(&cs->css, 0, &it);
1742 while ((task = css_task_iter_next(&it))) {
1743 struct mm_struct *mm;
1746 cpuset_change_task_nodemask(task, &newmems);
1748 mm = get_task_mm(task);
1752 migrate = is_memory_migrate(cs);
1754 mpol_rebind_mm(mm, &cs->mems_allowed);
1756 cpuset_migrate_mm(mm, &cs->old_mems_allowed, &newmems);
1760 css_task_iter_end(&it);
1763 * All the tasks' nodemasks have been updated, update
1764 * cs->old_mems_allowed.
1766 cs->old_mems_allowed = newmems;
1768 /* We're done rebinding vmas to this cpuset's new mems_allowed. */
1769 cpuset_being_rebound = NULL;
1773 * update_nodemasks_hier - Update effective nodemasks and tasks in the subtree
1774 * @cs: the cpuset to consider
1775 * @new_mems: a temp variable for calculating new effective_mems
1777 * When configured nodemask is changed, the effective nodemasks of this cpuset
1778 * and all its descendants need to be updated.
1780 * On legacy hierarchy, effective_mems will be the same with mems_allowed.
1782 * Called with cpuset_rwsem held
1784 static void update_nodemasks_hier(struct cpuset *cs, nodemask_t *new_mems)
1787 struct cgroup_subsys_state *pos_css;
1790 cpuset_for_each_descendant_pre(cp, pos_css, cs) {
1791 struct cpuset *parent = parent_cs(cp);
1793 nodes_and(*new_mems, cp->mems_allowed, parent->effective_mems);
1796 * If it becomes empty, inherit the effective mask of the
1797 * parent, which is guaranteed to have some MEMs.
1799 if (is_in_v2_mode() && nodes_empty(*new_mems))
1800 *new_mems = parent->effective_mems;
1802 /* Skip the whole subtree if the nodemask remains the same. */
1803 if (nodes_equal(*new_mems, cp->effective_mems)) {
1804 pos_css = css_rightmost_descendant(pos_css);
1808 if (!css_tryget_online(&cp->css))
1812 spin_lock_irq(&callback_lock);
1813 cp->effective_mems = *new_mems;
1814 spin_unlock_irq(&callback_lock);
1816 WARN_ON(!is_in_v2_mode() &&
1817 !nodes_equal(cp->mems_allowed, cp->effective_mems));
1819 update_tasks_nodemask(cp);
1828 * Handle user request to change the 'mems' memory placement
1829 * of a cpuset. Needs to validate the request, update the
1830 * cpusets mems_allowed, and for each task in the cpuset,
1831 * update mems_allowed and rebind task's mempolicy and any vma
1832 * mempolicies and if the cpuset is marked 'memory_migrate',
1833 * migrate the tasks pages to the new memory.
1835 * Call with cpuset_rwsem held. May take callback_lock during call.
1836 * Will take tasklist_lock, scan tasklist for tasks in cpuset cs,
1837 * lock each such tasks mm->mmap_lock, scan its vma's and rebind
1838 * their mempolicies to the cpusets new mems_allowed.
1840 static int update_nodemask(struct cpuset *cs, struct cpuset *trialcs,
1846 * top_cpuset.mems_allowed tracks node_stats[N_MEMORY];
1849 if (cs == &top_cpuset) {
1855 * An empty mems_allowed is ok iff there are no tasks in the cpuset.
1856 * Since nodelist_parse() fails on an empty mask, we special case
1857 * that parsing. The validate_change() call ensures that cpusets
1858 * with tasks have memory.
1861 nodes_clear(trialcs->mems_allowed);
1863 retval = nodelist_parse(buf, trialcs->mems_allowed);
1867 if (!nodes_subset(trialcs->mems_allowed,
1868 top_cpuset.mems_allowed)) {
1874 if (nodes_equal(cs->mems_allowed, trialcs->mems_allowed)) {
1875 retval = 0; /* Too easy - nothing to do */
1878 retval = validate_change(cs, trialcs);
1882 spin_lock_irq(&callback_lock);
1883 cs->mems_allowed = trialcs->mems_allowed;
1884 spin_unlock_irq(&callback_lock);
1886 /* use trialcs->mems_allowed as a temp variable */
1887 update_nodemasks_hier(cs, &trialcs->mems_allowed);
1892 bool current_cpuset_is_being_rebound(void)
1897 ret = task_cs(current) == cpuset_being_rebound;
1903 static int update_relax_domain_level(struct cpuset *cs, s64 val)
1906 if (val < -1 || val >= sched_domain_level_max)
1910 if (val != cs->relax_domain_level) {
1911 cs->relax_domain_level = val;
1912 if (!cpumask_empty(cs->cpus_allowed) &&
1913 is_sched_load_balance(cs))
1914 rebuild_sched_domains_locked();
1921 * update_tasks_flags - update the spread flags of tasks in the cpuset.
1922 * @cs: the cpuset in which each task's spread flags needs to be changed
1924 * Iterate through each task of @cs updating its spread flags. As this
1925 * function is called with cpuset_rwsem held, cpuset membership stays
1928 static void update_tasks_flags(struct cpuset *cs)
1930 struct css_task_iter it;
1931 struct task_struct *task;
1933 css_task_iter_start(&cs->css, 0, &it);
1934 while ((task = css_task_iter_next(&it)))
1935 cpuset_update_task_spread_flag(cs, task);
1936 css_task_iter_end(&it);
1940 * update_flag - read a 0 or a 1 in a file and update associated flag
1941 * bit: the bit to update (see cpuset_flagbits_t)
1942 * cs: the cpuset to update
1943 * turning_on: whether the flag is being set or cleared
1945 * Call with cpuset_rwsem held.
1948 static int update_flag(cpuset_flagbits_t bit, struct cpuset *cs,
1951 struct cpuset *trialcs;
1952 int balance_flag_changed;
1953 int spread_flag_changed;
1956 trialcs = alloc_trial_cpuset(cs);
1961 set_bit(bit, &trialcs->flags);
1963 clear_bit(bit, &trialcs->flags);
1965 err = validate_change(cs, trialcs);
1969 balance_flag_changed = (is_sched_load_balance(cs) !=
1970 is_sched_load_balance(trialcs));
1972 spread_flag_changed = ((is_spread_slab(cs) != is_spread_slab(trialcs))
1973 || (is_spread_page(cs) != is_spread_page(trialcs)));
1975 spin_lock_irq(&callback_lock);
1976 cs->flags = trialcs->flags;
1977 spin_unlock_irq(&callback_lock);
1979 if (!cpumask_empty(trialcs->cpus_allowed) && balance_flag_changed)
1980 rebuild_sched_domains_locked();
1982 if (spread_flag_changed)
1983 update_tasks_flags(cs);
1985 free_cpuset(trialcs);
1990 * update_prstate - update partititon_root_state
1991 * cs: the cpuset to update
1992 * new_prs: new partition root state
1994 * Call with cpuset_rwsem held.
1996 static int update_prstate(struct cpuset *cs, int new_prs)
1998 int err, old_prs = cs->partition_root_state;
1999 struct cpuset *parent = parent_cs(cs);
2000 struct tmpmasks tmpmask;
2002 if (old_prs == new_prs)
2006 * Cannot force a partial or invalid partition root to a full
2009 if (new_prs && (old_prs == PRS_ERROR))
2012 if (alloc_cpumasks(NULL, &tmpmask))
2018 * Turning on partition root requires setting the
2019 * CS_CPU_EXCLUSIVE bit implicitly as well and cpus_allowed
2022 if (cpumask_empty(cs->cpus_allowed))
2025 err = update_flag(CS_CPU_EXCLUSIVE, cs, 1);
2029 err = update_parent_subparts_cpumask(cs, partcmd_enable,
2032 update_flag(CS_CPU_EXCLUSIVE, cs, 0);
2037 * Turning off partition root will clear the
2038 * CS_CPU_EXCLUSIVE bit.
2040 if (old_prs == PRS_ERROR) {
2041 update_flag(CS_CPU_EXCLUSIVE, cs, 0);
2046 err = update_parent_subparts_cpumask(cs, partcmd_disable,
2051 /* Turning off CS_CPU_EXCLUSIVE will not return error */
2052 update_flag(CS_CPU_EXCLUSIVE, cs, 0);
2056 * Update cpumask of parent's tasks except when it is the top
2057 * cpuset as some system daemons cannot be mapped to other CPUs.
2059 if (parent != &top_cpuset)
2060 update_tasks_cpumask(parent);
2062 if (parent->child_ecpus_count)
2063 update_sibling_cpumasks(parent, cs, &tmpmask);
2065 rebuild_sched_domains_locked();
2068 spin_lock_irq(&callback_lock);
2069 cs->partition_root_state = new_prs;
2070 spin_unlock_irq(&callback_lock);
2071 notify_partition_change(cs, old_prs, new_prs);
2074 free_cpumasks(NULL, &tmpmask);
2079 * Frequency meter - How fast is some event occurring?
2081 * These routines manage a digitally filtered, constant time based,
2082 * event frequency meter. There are four routines:
2083 * fmeter_init() - initialize a frequency meter.
2084 * fmeter_markevent() - called each time the event happens.
2085 * fmeter_getrate() - returns the recent rate of such events.
2086 * fmeter_update() - internal routine used to update fmeter.
2088 * A common data structure is passed to each of these routines,
2089 * which is used to keep track of the state required to manage the
2090 * frequency meter and its digital filter.
2092 * The filter works on the number of events marked per unit time.
2093 * The filter is single-pole low-pass recursive (IIR). The time unit
2094 * is 1 second. Arithmetic is done using 32-bit integers scaled to
2095 * simulate 3 decimal digits of precision (multiplied by 1000).
2097 * With an FM_COEF of 933, and a time base of 1 second, the filter
2098 * has a half-life of 10 seconds, meaning that if the events quit
2099 * happening, then the rate returned from the fmeter_getrate()
2100 * will be cut in half each 10 seconds, until it converges to zero.
2102 * It is not worth doing a real infinitely recursive filter. If more
2103 * than FM_MAXTICKS ticks have elapsed since the last filter event,
2104 * just compute FM_MAXTICKS ticks worth, by which point the level
2107 * Limit the count of unprocessed events to FM_MAXCNT, so as to avoid
2108 * arithmetic overflow in the fmeter_update() routine.
2110 * Given the simple 32 bit integer arithmetic used, this meter works
2111 * best for reporting rates between one per millisecond (msec) and
2112 * one per 32 (approx) seconds. At constant rates faster than one
2113 * per msec it maxes out at values just under 1,000,000. At constant
2114 * rates between one per msec, and one per second it will stabilize
2115 * to a value N*1000, where N is the rate of events per second.
2116 * At constant rates between one per second and one per 32 seconds,
2117 * it will be choppy, moving up on the seconds that have an event,
2118 * and then decaying until the next event. At rates slower than
2119 * about one in 32 seconds, it decays all the way back to zero between
2123 #define FM_COEF 933 /* coefficient for half-life of 10 secs */
2124 #define FM_MAXTICKS ((u32)99) /* useless computing more ticks than this */
2125 #define FM_MAXCNT 1000000 /* limit cnt to avoid overflow */
2126 #define FM_SCALE 1000 /* faux fixed point scale */
2128 /* Initialize a frequency meter */
2129 static void fmeter_init(struct fmeter *fmp)
2134 spin_lock_init(&fmp->lock);
2137 /* Internal meter update - process cnt events and update value */
2138 static void fmeter_update(struct fmeter *fmp)
2143 now = ktime_get_seconds();
2144 ticks = now - fmp->time;
2149 ticks = min(FM_MAXTICKS, ticks);
2151 fmp->val = (FM_COEF * fmp->val) / FM_SCALE;
2154 fmp->val += ((FM_SCALE - FM_COEF) * fmp->cnt) / FM_SCALE;
2158 /* Process any previous ticks, then bump cnt by one (times scale). */
2159 static void fmeter_markevent(struct fmeter *fmp)
2161 spin_lock(&fmp->lock);
2163 fmp->cnt = min(FM_MAXCNT, fmp->cnt + FM_SCALE);
2164 spin_unlock(&fmp->lock);
2167 /* Process any previous ticks, then return current value. */
2168 static int fmeter_getrate(struct fmeter *fmp)
2172 spin_lock(&fmp->lock);
2175 spin_unlock(&fmp->lock);
2179 static struct cpuset *cpuset_attach_old_cs;
2181 /* Called by cgroups to determine if a cpuset is usable; cpuset_rwsem held */
2182 static int cpuset_can_attach(struct cgroup_taskset *tset)
2184 struct cgroup_subsys_state *css;
2186 struct task_struct *task;
2189 /* used later by cpuset_attach() */
2190 cpuset_attach_old_cs = task_cs(cgroup_taskset_first(tset, &css));
2193 percpu_down_write(&cpuset_rwsem);
2195 /* allow moving tasks into an empty cpuset if on default hierarchy */
2197 if (!is_in_v2_mode() &&
2198 (cpumask_empty(cs->cpus_allowed) || nodes_empty(cs->mems_allowed)))
2201 cgroup_taskset_for_each(task, css, tset) {
2202 ret = task_can_attach(task, cs->effective_cpus);
2205 ret = security_task_setscheduler(task);
2211 * Mark attach is in progress. This makes validate_change() fail
2212 * changes which zero cpus/mems_allowed.
2214 cs->attach_in_progress++;
2217 percpu_up_write(&cpuset_rwsem);
2221 static void cpuset_cancel_attach(struct cgroup_taskset *tset)
2223 struct cgroup_subsys_state *css;
2225 cgroup_taskset_first(tset, &css);
2227 percpu_down_write(&cpuset_rwsem);
2228 css_cs(css)->attach_in_progress--;
2229 percpu_up_write(&cpuset_rwsem);
2233 * Protected by cpuset_rwsem. cpus_attach is used only by cpuset_attach()
2234 * but we can't allocate it dynamically there. Define it global and
2235 * allocate from cpuset_init().
2237 static cpumask_var_t cpus_attach;
2239 static void cpuset_attach(struct cgroup_taskset *tset)
2241 /* static buf protected by cpuset_rwsem */
2242 static nodemask_t cpuset_attach_nodemask_to;
2243 struct task_struct *task;
2244 struct task_struct *leader;
2245 struct cgroup_subsys_state *css;
2247 struct cpuset *oldcs = cpuset_attach_old_cs;
2249 cgroup_taskset_first(tset, &css);
2252 lockdep_assert_cpus_held(); /* see cgroup_attach_lock() */
2253 percpu_down_write(&cpuset_rwsem);
2255 guarantee_online_mems(cs, &cpuset_attach_nodemask_to);
2257 cgroup_taskset_for_each(task, css, tset) {
2258 if (cs != &top_cpuset)
2259 guarantee_online_cpus(task, cpus_attach);
2261 cpumask_copy(cpus_attach, task_cpu_possible_mask(task));
2263 * can_attach beforehand should guarantee that this doesn't
2264 * fail. TODO: have a better way to handle failure here
2266 WARN_ON_ONCE(set_cpus_allowed_ptr(task, cpus_attach));
2268 cpuset_change_task_nodemask(task, &cpuset_attach_nodemask_to);
2269 cpuset_update_task_spread_flag(cs, task);
2273 * Change mm for all threadgroup leaders. This is expensive and may
2274 * sleep and should be moved outside migration path proper.
2276 cpuset_attach_nodemask_to = cs->effective_mems;
2277 cgroup_taskset_for_each_leader(leader, css, tset) {
2278 struct mm_struct *mm = get_task_mm(leader);
2281 mpol_rebind_mm(mm, &cpuset_attach_nodemask_to);
2284 * old_mems_allowed is the same with mems_allowed
2285 * here, except if this task is being moved
2286 * automatically due to hotplug. In that case
2287 * @mems_allowed has been updated and is empty, so
2288 * @old_mems_allowed is the right nodesets that we
2291 if (is_memory_migrate(cs))
2292 cpuset_migrate_mm(mm, &oldcs->old_mems_allowed,
2293 &cpuset_attach_nodemask_to);
2299 cs->old_mems_allowed = cpuset_attach_nodemask_to;
2301 cs->attach_in_progress--;
2302 if (!cs->attach_in_progress)
2303 wake_up(&cpuset_attach_wq);
2305 percpu_up_write(&cpuset_rwsem);
2308 /* The various types of files and directories in a cpuset file system */
2311 FILE_MEMORY_MIGRATE,
2314 FILE_EFFECTIVE_CPULIST,
2315 FILE_EFFECTIVE_MEMLIST,
2316 FILE_SUBPARTS_CPULIST,
2320 FILE_SCHED_LOAD_BALANCE,
2321 FILE_PARTITION_ROOT,
2322 FILE_SCHED_RELAX_DOMAIN_LEVEL,
2323 FILE_MEMORY_PRESSURE_ENABLED,
2324 FILE_MEMORY_PRESSURE,
2327 } cpuset_filetype_t;
2329 static int cpuset_write_u64(struct cgroup_subsys_state *css, struct cftype *cft,
2332 struct cpuset *cs = css_cs(css);
2333 cpuset_filetype_t type = cft->private;
2337 percpu_down_write(&cpuset_rwsem);
2338 if (!is_cpuset_online(cs)) {
2344 case FILE_CPU_EXCLUSIVE:
2345 retval = update_flag(CS_CPU_EXCLUSIVE, cs, val);
2347 case FILE_MEM_EXCLUSIVE:
2348 retval = update_flag(CS_MEM_EXCLUSIVE, cs, val);
2350 case FILE_MEM_HARDWALL:
2351 retval = update_flag(CS_MEM_HARDWALL, cs, val);
2353 case FILE_SCHED_LOAD_BALANCE:
2354 retval = update_flag(CS_SCHED_LOAD_BALANCE, cs, val);
2356 case FILE_MEMORY_MIGRATE:
2357 retval = update_flag(CS_MEMORY_MIGRATE, cs, val);
2359 case FILE_MEMORY_PRESSURE_ENABLED:
2360 cpuset_memory_pressure_enabled = !!val;
2362 case FILE_SPREAD_PAGE:
2363 retval = update_flag(CS_SPREAD_PAGE, cs, val);
2365 case FILE_SPREAD_SLAB:
2366 retval = update_flag(CS_SPREAD_SLAB, cs, val);
2373 percpu_up_write(&cpuset_rwsem);
2378 static int cpuset_write_s64(struct cgroup_subsys_state *css, struct cftype *cft,
2381 struct cpuset *cs = css_cs(css);
2382 cpuset_filetype_t type = cft->private;
2383 int retval = -ENODEV;
2386 percpu_down_write(&cpuset_rwsem);
2387 if (!is_cpuset_online(cs))
2391 case FILE_SCHED_RELAX_DOMAIN_LEVEL:
2392 retval = update_relax_domain_level(cs, val);
2399 percpu_up_write(&cpuset_rwsem);
2405 * Common handling for a write to a "cpus" or "mems" file.
2407 static ssize_t cpuset_write_resmask(struct kernfs_open_file *of,
2408 char *buf, size_t nbytes, loff_t off)
2410 struct cpuset *cs = css_cs(of_css(of));
2411 struct cpuset *trialcs;
2412 int retval = -ENODEV;
2414 buf = strstrip(buf);
2417 * CPU or memory hotunplug may leave @cs w/o any execution
2418 * resources, in which case the hotplug code asynchronously updates
2419 * configuration and transfers all tasks to the nearest ancestor
2420 * which can execute.
2422 * As writes to "cpus" or "mems" may restore @cs's execution
2423 * resources, wait for the previously scheduled operations before
2424 * proceeding, so that we don't end up keep removing tasks added
2425 * after execution capability is restored.
2427 * cpuset_hotplug_work calls back into cgroup core via
2428 * cgroup_transfer_tasks() and waiting for it from a cgroupfs
2429 * operation like this one can lead to a deadlock through kernfs
2430 * active_ref protection. Let's break the protection. Losing the
2431 * protection is okay as we check whether @cs is online after
2432 * grabbing cpuset_rwsem anyway. This only happens on the legacy
2436 kernfs_break_active_protection(of->kn);
2437 flush_work(&cpuset_hotplug_work);
2440 percpu_down_write(&cpuset_rwsem);
2441 if (!is_cpuset_online(cs))
2444 trialcs = alloc_trial_cpuset(cs);
2450 switch (of_cft(of)->private) {
2452 retval = update_cpumask(cs, trialcs, buf);
2455 retval = update_nodemask(cs, trialcs, buf);
2462 free_cpuset(trialcs);
2464 percpu_up_write(&cpuset_rwsem);
2466 kernfs_unbreak_active_protection(of->kn);
2468 flush_workqueue(cpuset_migrate_mm_wq);
2469 return retval ?: nbytes;
2473 * These ascii lists should be read in a single call, by using a user
2474 * buffer large enough to hold the entire map. If read in smaller
2475 * chunks, there is no guarantee of atomicity. Since the display format
2476 * used, list of ranges of sequential numbers, is variable length,
2477 * and since these maps can change value dynamically, one could read
2478 * gibberish by doing partial reads while a list was changing.
2480 static int cpuset_common_seq_show(struct seq_file *sf, void *v)
2482 struct cpuset *cs = css_cs(seq_css(sf));
2483 cpuset_filetype_t type = seq_cft(sf)->private;
2486 spin_lock_irq(&callback_lock);
2490 seq_printf(sf, "%*pbl\n", cpumask_pr_args(cs->cpus_allowed));
2493 seq_printf(sf, "%*pbl\n", nodemask_pr_args(&cs->mems_allowed));
2495 case FILE_EFFECTIVE_CPULIST:
2496 seq_printf(sf, "%*pbl\n", cpumask_pr_args(cs->effective_cpus));
2498 case FILE_EFFECTIVE_MEMLIST:
2499 seq_printf(sf, "%*pbl\n", nodemask_pr_args(&cs->effective_mems));
2501 case FILE_SUBPARTS_CPULIST:
2502 seq_printf(sf, "%*pbl\n", cpumask_pr_args(cs->subparts_cpus));
2508 spin_unlock_irq(&callback_lock);
2512 static u64 cpuset_read_u64(struct cgroup_subsys_state *css, struct cftype *cft)
2514 struct cpuset *cs = css_cs(css);
2515 cpuset_filetype_t type = cft->private;
2517 case FILE_CPU_EXCLUSIVE:
2518 return is_cpu_exclusive(cs);
2519 case FILE_MEM_EXCLUSIVE:
2520 return is_mem_exclusive(cs);
2521 case FILE_MEM_HARDWALL:
2522 return is_mem_hardwall(cs);
2523 case FILE_SCHED_LOAD_BALANCE:
2524 return is_sched_load_balance(cs);
2525 case FILE_MEMORY_MIGRATE:
2526 return is_memory_migrate(cs);
2527 case FILE_MEMORY_PRESSURE_ENABLED:
2528 return cpuset_memory_pressure_enabled;
2529 case FILE_MEMORY_PRESSURE:
2530 return fmeter_getrate(&cs->fmeter);
2531 case FILE_SPREAD_PAGE:
2532 return is_spread_page(cs);
2533 case FILE_SPREAD_SLAB:
2534 return is_spread_slab(cs);
2539 /* Unreachable but makes gcc happy */
2543 static s64 cpuset_read_s64(struct cgroup_subsys_state *css, struct cftype *cft)
2545 struct cpuset *cs = css_cs(css);
2546 cpuset_filetype_t type = cft->private;
2548 case FILE_SCHED_RELAX_DOMAIN_LEVEL:
2549 return cs->relax_domain_level;
2554 /* Unreachable but makes gcc happy */
2558 static int sched_partition_show(struct seq_file *seq, void *v)
2560 struct cpuset *cs = css_cs(seq_css(seq));
2562 switch (cs->partition_root_state) {
2564 seq_puts(seq, "root\n");
2567 seq_puts(seq, "member\n");
2570 seq_puts(seq, "root invalid\n");
2576 static ssize_t sched_partition_write(struct kernfs_open_file *of, char *buf,
2577 size_t nbytes, loff_t off)
2579 struct cpuset *cs = css_cs(of_css(of));
2581 int retval = -ENODEV;
2583 buf = strstrip(buf);
2586 * Convert "root" to ENABLED, and convert "member" to DISABLED.
2588 if (!strcmp(buf, "root"))
2590 else if (!strcmp(buf, "member"))
2597 percpu_down_write(&cpuset_rwsem);
2598 if (!is_cpuset_online(cs))
2601 retval = update_prstate(cs, val);
2603 percpu_up_write(&cpuset_rwsem);
2606 return retval ?: nbytes;
2610 * for the common functions, 'private' gives the type of file
2613 static struct cftype legacy_files[] = {
2616 .seq_show = cpuset_common_seq_show,
2617 .write = cpuset_write_resmask,
2618 .max_write_len = (100U + 6 * NR_CPUS),
2619 .private = FILE_CPULIST,
2624 .seq_show = cpuset_common_seq_show,
2625 .write = cpuset_write_resmask,
2626 .max_write_len = (100U + 6 * MAX_NUMNODES),
2627 .private = FILE_MEMLIST,
2631 .name = "effective_cpus",
2632 .seq_show = cpuset_common_seq_show,
2633 .private = FILE_EFFECTIVE_CPULIST,
2637 .name = "effective_mems",
2638 .seq_show = cpuset_common_seq_show,
2639 .private = FILE_EFFECTIVE_MEMLIST,
2643 .name = "cpu_exclusive",
2644 .read_u64 = cpuset_read_u64,
2645 .write_u64 = cpuset_write_u64,
2646 .private = FILE_CPU_EXCLUSIVE,
2650 .name = "mem_exclusive",
2651 .read_u64 = cpuset_read_u64,
2652 .write_u64 = cpuset_write_u64,
2653 .private = FILE_MEM_EXCLUSIVE,
2657 .name = "mem_hardwall",
2658 .read_u64 = cpuset_read_u64,
2659 .write_u64 = cpuset_write_u64,
2660 .private = FILE_MEM_HARDWALL,
2664 .name = "sched_load_balance",
2665 .read_u64 = cpuset_read_u64,
2666 .write_u64 = cpuset_write_u64,
2667 .private = FILE_SCHED_LOAD_BALANCE,
2671 .name = "sched_relax_domain_level",
2672 .read_s64 = cpuset_read_s64,
2673 .write_s64 = cpuset_write_s64,
2674 .private = FILE_SCHED_RELAX_DOMAIN_LEVEL,
2678 .name = "memory_migrate",
2679 .read_u64 = cpuset_read_u64,
2680 .write_u64 = cpuset_write_u64,
2681 .private = FILE_MEMORY_MIGRATE,
2685 .name = "memory_pressure",
2686 .read_u64 = cpuset_read_u64,
2687 .private = FILE_MEMORY_PRESSURE,
2691 .name = "memory_spread_page",
2692 .read_u64 = cpuset_read_u64,
2693 .write_u64 = cpuset_write_u64,
2694 .private = FILE_SPREAD_PAGE,
2698 .name = "memory_spread_slab",
2699 .read_u64 = cpuset_read_u64,
2700 .write_u64 = cpuset_write_u64,
2701 .private = FILE_SPREAD_SLAB,
2705 .name = "memory_pressure_enabled",
2706 .flags = CFTYPE_ONLY_ON_ROOT,
2707 .read_u64 = cpuset_read_u64,
2708 .write_u64 = cpuset_write_u64,
2709 .private = FILE_MEMORY_PRESSURE_ENABLED,
2716 * This is currently a minimal set for the default hierarchy. It can be
2717 * expanded later on by migrating more features and control files from v1.
2719 static struct cftype dfl_files[] = {
2722 .seq_show = cpuset_common_seq_show,
2723 .write = cpuset_write_resmask,
2724 .max_write_len = (100U + 6 * NR_CPUS),
2725 .private = FILE_CPULIST,
2726 .flags = CFTYPE_NOT_ON_ROOT,
2731 .seq_show = cpuset_common_seq_show,
2732 .write = cpuset_write_resmask,
2733 .max_write_len = (100U + 6 * MAX_NUMNODES),
2734 .private = FILE_MEMLIST,
2735 .flags = CFTYPE_NOT_ON_ROOT,
2739 .name = "cpus.effective",
2740 .seq_show = cpuset_common_seq_show,
2741 .private = FILE_EFFECTIVE_CPULIST,
2745 .name = "mems.effective",
2746 .seq_show = cpuset_common_seq_show,
2747 .private = FILE_EFFECTIVE_MEMLIST,
2751 .name = "cpus.partition",
2752 .seq_show = sched_partition_show,
2753 .write = sched_partition_write,
2754 .private = FILE_PARTITION_ROOT,
2755 .flags = CFTYPE_NOT_ON_ROOT,
2756 .file_offset = offsetof(struct cpuset, partition_file),
2760 .name = "cpus.subpartitions",
2761 .seq_show = cpuset_common_seq_show,
2762 .private = FILE_SUBPARTS_CPULIST,
2763 .flags = CFTYPE_DEBUG,
2771 * cpuset_css_alloc - allocate a cpuset css
2772 * cgrp: control group that the new cpuset will be part of
2775 static struct cgroup_subsys_state *
2776 cpuset_css_alloc(struct cgroup_subsys_state *parent_css)
2781 return &top_cpuset.css;
2783 cs = kzalloc(sizeof(*cs), GFP_KERNEL);
2785 return ERR_PTR(-ENOMEM);
2787 if (alloc_cpumasks(cs, NULL)) {
2789 return ERR_PTR(-ENOMEM);
2792 __set_bit(CS_SCHED_LOAD_BALANCE, &cs->flags);
2793 nodes_clear(cs->mems_allowed);
2794 nodes_clear(cs->effective_mems);
2795 fmeter_init(&cs->fmeter);
2796 cs->relax_domain_level = -1;
2798 /* Set CS_MEMORY_MIGRATE for default hierarchy */
2799 if (cgroup_subsys_on_dfl(cpuset_cgrp_subsys))
2800 __set_bit(CS_MEMORY_MIGRATE, &cs->flags);
2805 static int cpuset_css_online(struct cgroup_subsys_state *css)
2807 struct cpuset *cs = css_cs(css);
2808 struct cpuset *parent = parent_cs(cs);
2809 struct cpuset *tmp_cs;
2810 struct cgroup_subsys_state *pos_css;
2816 percpu_down_write(&cpuset_rwsem);
2818 set_bit(CS_ONLINE, &cs->flags);
2819 if (is_spread_page(parent))
2820 set_bit(CS_SPREAD_PAGE, &cs->flags);
2821 if (is_spread_slab(parent))
2822 set_bit(CS_SPREAD_SLAB, &cs->flags);
2826 spin_lock_irq(&callback_lock);
2827 if (is_in_v2_mode()) {
2828 cpumask_copy(cs->effective_cpus, parent->effective_cpus);
2829 cs->effective_mems = parent->effective_mems;
2830 cs->use_parent_ecpus = true;
2831 parent->child_ecpus_count++;
2833 spin_unlock_irq(&callback_lock);
2835 if (!test_bit(CGRP_CPUSET_CLONE_CHILDREN, &css->cgroup->flags))
2839 * Clone @parent's configuration if CGRP_CPUSET_CLONE_CHILDREN is
2840 * set. This flag handling is implemented in cgroup core for
2841 * histrical reasons - the flag may be specified during mount.
2843 * Currently, if any sibling cpusets have exclusive cpus or mem, we
2844 * refuse to clone the configuration - thereby refusing the task to
2845 * be entered, and as a result refusing the sys_unshare() or
2846 * clone() which initiated it. If this becomes a problem for some
2847 * users who wish to allow that scenario, then this could be
2848 * changed to grant parent->cpus_allowed-sibling_cpus_exclusive
2849 * (and likewise for mems) to the new cgroup.
2852 cpuset_for_each_child(tmp_cs, pos_css, parent) {
2853 if (is_mem_exclusive(tmp_cs) || is_cpu_exclusive(tmp_cs)) {
2860 spin_lock_irq(&callback_lock);
2861 cs->mems_allowed = parent->mems_allowed;
2862 cs->effective_mems = parent->mems_allowed;
2863 cpumask_copy(cs->cpus_allowed, parent->cpus_allowed);
2864 cpumask_copy(cs->effective_cpus, parent->cpus_allowed);
2865 spin_unlock_irq(&callback_lock);
2867 percpu_up_write(&cpuset_rwsem);
2873 * If the cpuset being removed has its flag 'sched_load_balance'
2874 * enabled, then simulate turning sched_load_balance off, which
2875 * will call rebuild_sched_domains_locked(). That is not needed
2876 * in the default hierarchy where only changes in partition
2877 * will cause repartitioning.
2879 * If the cpuset has the 'sched.partition' flag enabled, simulate
2880 * turning 'sched.partition" off.
2883 static void cpuset_css_offline(struct cgroup_subsys_state *css)
2885 struct cpuset *cs = css_cs(css);
2888 percpu_down_write(&cpuset_rwsem);
2890 if (is_partition_root(cs))
2891 update_prstate(cs, 0);
2893 if (!cgroup_subsys_on_dfl(cpuset_cgrp_subsys) &&
2894 is_sched_load_balance(cs))
2895 update_flag(CS_SCHED_LOAD_BALANCE, cs, 0);
2897 if (cs->use_parent_ecpus) {
2898 struct cpuset *parent = parent_cs(cs);
2900 cs->use_parent_ecpus = false;
2901 parent->child_ecpus_count--;
2905 clear_bit(CS_ONLINE, &cs->flags);
2907 percpu_up_write(&cpuset_rwsem);
2911 static void cpuset_css_free(struct cgroup_subsys_state *css)
2913 struct cpuset *cs = css_cs(css);
2918 static void cpuset_bind(struct cgroup_subsys_state *root_css)
2920 percpu_down_write(&cpuset_rwsem);
2921 spin_lock_irq(&callback_lock);
2923 if (is_in_v2_mode()) {
2924 cpumask_copy(top_cpuset.cpus_allowed, cpu_possible_mask);
2925 top_cpuset.mems_allowed = node_possible_map;
2927 cpumask_copy(top_cpuset.cpus_allowed,
2928 top_cpuset.effective_cpus);
2929 top_cpuset.mems_allowed = top_cpuset.effective_mems;
2932 spin_unlock_irq(&callback_lock);
2933 percpu_up_write(&cpuset_rwsem);
2937 * Make sure the new task conform to the current state of its parent,
2938 * which could have been changed by cpuset just after it inherits the
2939 * state from the parent and before it sits on the cgroup's task list.
2941 static void cpuset_fork(struct task_struct *task)
2943 if (task_css_is_root(task, cpuset_cgrp_id))
2946 set_cpus_allowed_ptr(task, current->cpus_ptr);
2947 task->mems_allowed = current->mems_allowed;
2950 struct cgroup_subsys cpuset_cgrp_subsys = {
2951 .css_alloc = cpuset_css_alloc,
2952 .css_online = cpuset_css_online,
2953 .css_offline = cpuset_css_offline,
2954 .css_free = cpuset_css_free,
2955 .can_attach = cpuset_can_attach,
2956 .cancel_attach = cpuset_cancel_attach,
2957 .attach = cpuset_attach,
2958 .post_attach = cpuset_post_attach,
2959 .bind = cpuset_bind,
2960 .fork = cpuset_fork,
2961 .legacy_cftypes = legacy_files,
2962 .dfl_cftypes = dfl_files,
2968 * cpuset_init - initialize cpusets at system boot
2970 * Description: Initialize top_cpuset
2973 int __init cpuset_init(void)
2975 BUG_ON(percpu_init_rwsem(&cpuset_rwsem));
2977 BUG_ON(!alloc_cpumask_var(&top_cpuset.cpus_allowed, GFP_KERNEL));
2978 BUG_ON(!alloc_cpumask_var(&top_cpuset.effective_cpus, GFP_KERNEL));
2979 BUG_ON(!zalloc_cpumask_var(&top_cpuset.subparts_cpus, GFP_KERNEL));
2981 cpumask_setall(top_cpuset.cpus_allowed);
2982 nodes_setall(top_cpuset.mems_allowed);
2983 cpumask_setall(top_cpuset.effective_cpus);
2984 nodes_setall(top_cpuset.effective_mems);
2986 fmeter_init(&top_cpuset.fmeter);
2987 set_bit(CS_SCHED_LOAD_BALANCE, &top_cpuset.flags);
2988 top_cpuset.relax_domain_level = -1;
2990 BUG_ON(!alloc_cpumask_var(&cpus_attach, GFP_KERNEL));
2996 * If CPU and/or memory hotplug handlers, below, unplug any CPUs
2997 * or memory nodes, we need to walk over the cpuset hierarchy,
2998 * removing that CPU or node from all cpusets. If this removes the
2999 * last CPU or node from a cpuset, then move the tasks in the empty
3000 * cpuset to its next-highest non-empty parent.
3002 static void remove_tasks_in_empty_cpuset(struct cpuset *cs)
3004 struct cpuset *parent;
3007 * Find its next-highest non-empty parent, (top cpuset
3008 * has online cpus, so can't be empty).
3010 parent = parent_cs(cs);
3011 while (cpumask_empty(parent->cpus_allowed) ||
3012 nodes_empty(parent->mems_allowed))
3013 parent = parent_cs(parent);
3015 if (cgroup_transfer_tasks(parent->css.cgroup, cs->css.cgroup)) {
3016 pr_err("cpuset: failed to transfer tasks out of empty cpuset ");
3017 pr_cont_cgroup_name(cs->css.cgroup);
3023 hotplug_update_tasks_legacy(struct cpuset *cs,
3024 struct cpumask *new_cpus, nodemask_t *new_mems,
3025 bool cpus_updated, bool mems_updated)
3029 spin_lock_irq(&callback_lock);
3030 cpumask_copy(cs->cpus_allowed, new_cpus);
3031 cpumask_copy(cs->effective_cpus, new_cpus);
3032 cs->mems_allowed = *new_mems;
3033 cs->effective_mems = *new_mems;
3034 spin_unlock_irq(&callback_lock);
3037 * Don't call update_tasks_cpumask() if the cpuset becomes empty,
3038 * as the tasks will be migratecd to an ancestor.
3040 if (cpus_updated && !cpumask_empty(cs->cpus_allowed))
3041 update_tasks_cpumask(cs);
3042 if (mems_updated && !nodes_empty(cs->mems_allowed))
3043 update_tasks_nodemask(cs);
3045 is_empty = cpumask_empty(cs->cpus_allowed) ||
3046 nodes_empty(cs->mems_allowed);
3048 percpu_up_write(&cpuset_rwsem);
3051 * Move tasks to the nearest ancestor with execution resources,
3052 * This is full cgroup operation which will also call back into
3053 * cpuset. Should be done outside any lock.
3056 remove_tasks_in_empty_cpuset(cs);
3058 percpu_down_write(&cpuset_rwsem);
3062 hotplug_update_tasks(struct cpuset *cs,
3063 struct cpumask *new_cpus, nodemask_t *new_mems,
3064 bool cpus_updated, bool mems_updated)
3066 if (cpumask_empty(new_cpus))
3067 cpumask_copy(new_cpus, parent_cs(cs)->effective_cpus);
3068 if (nodes_empty(*new_mems))
3069 *new_mems = parent_cs(cs)->effective_mems;
3071 spin_lock_irq(&callback_lock);
3072 cpumask_copy(cs->effective_cpus, new_cpus);
3073 cs->effective_mems = *new_mems;
3074 spin_unlock_irq(&callback_lock);
3077 update_tasks_cpumask(cs);
3079 update_tasks_nodemask(cs);
3082 static bool force_rebuild;
3084 void cpuset_force_rebuild(void)
3086 force_rebuild = true;
3090 * cpuset_hotplug_update_tasks - update tasks in a cpuset for hotunplug
3091 * @cs: cpuset in interest
3092 * @tmp: the tmpmasks structure pointer
3094 * Compare @cs's cpu and mem masks against top_cpuset and if some have gone
3095 * offline, update @cs accordingly. If @cs ends up with no CPU or memory,
3096 * all its tasks are moved to the nearest ancestor with both resources.
3098 static void cpuset_hotplug_update_tasks(struct cpuset *cs, struct tmpmasks *tmp)
3100 static cpumask_t new_cpus;
3101 static nodemask_t new_mems;
3104 struct cpuset *parent;
3106 wait_event(cpuset_attach_wq, cs->attach_in_progress == 0);
3108 percpu_down_write(&cpuset_rwsem);
3111 * We have raced with task attaching. We wait until attaching
3112 * is finished, so we won't attach a task to an empty cpuset.
3114 if (cs->attach_in_progress) {
3115 percpu_up_write(&cpuset_rwsem);
3119 parent = parent_cs(cs);
3120 compute_effective_cpumask(&new_cpus, cs, parent);
3121 nodes_and(new_mems, cs->mems_allowed, parent->effective_mems);
3123 if (cs->nr_subparts_cpus)
3125 * Make sure that CPUs allocated to child partitions
3126 * do not show up in effective_cpus.
3128 cpumask_andnot(&new_cpus, &new_cpus, cs->subparts_cpus);
3130 if (!tmp || !cs->partition_root_state)
3134 * In the unlikely event that a partition root has empty
3135 * effective_cpus or its parent becomes erroneous, we have to
3136 * transition it to the erroneous state.
3138 if (is_partition_root(cs) && (cpumask_empty(&new_cpus) ||
3139 (parent->partition_root_state == PRS_ERROR))) {
3140 if (cs->nr_subparts_cpus) {
3141 spin_lock_irq(&callback_lock);
3142 cs->nr_subparts_cpus = 0;
3143 cpumask_clear(cs->subparts_cpus);
3144 spin_unlock_irq(&callback_lock);
3145 compute_effective_cpumask(&new_cpus, cs, parent);
3149 * If the effective_cpus is empty because the child
3150 * partitions take away all the CPUs, we can keep
3151 * the current partition and let the child partitions
3152 * fight for available CPUs.
3154 if ((parent->partition_root_state == PRS_ERROR) ||
3155 cpumask_empty(&new_cpus)) {
3158 update_parent_subparts_cpumask(cs, partcmd_disable,
3160 old_prs = cs->partition_root_state;
3161 if (old_prs != PRS_ERROR) {
3162 spin_lock_irq(&callback_lock);
3163 cs->partition_root_state = PRS_ERROR;
3164 spin_unlock_irq(&callback_lock);
3165 notify_partition_change(cs, old_prs, PRS_ERROR);
3168 cpuset_force_rebuild();
3172 * On the other hand, an erroneous partition root may be transitioned
3173 * back to a regular one or a partition root with no CPU allocated
3174 * from the parent may change to erroneous.
3176 if (is_partition_root(parent) &&
3177 ((cs->partition_root_state == PRS_ERROR) ||
3178 !cpumask_intersects(&new_cpus, parent->subparts_cpus)) &&
3179 update_parent_subparts_cpumask(cs, partcmd_update, NULL, tmp))
3180 cpuset_force_rebuild();
3183 cpus_updated = !cpumask_equal(&new_cpus, cs->effective_cpus);
3184 mems_updated = !nodes_equal(new_mems, cs->effective_mems);
3186 if (is_in_v2_mode())
3187 hotplug_update_tasks(cs, &new_cpus, &new_mems,
3188 cpus_updated, mems_updated);
3190 hotplug_update_tasks_legacy(cs, &new_cpus, &new_mems,
3191 cpus_updated, mems_updated);
3193 percpu_up_write(&cpuset_rwsem);
3197 * cpuset_hotplug_workfn - handle CPU/memory hotunplug for a cpuset
3199 * This function is called after either CPU or memory configuration has
3200 * changed and updates cpuset accordingly. The top_cpuset is always
3201 * synchronized to cpu_active_mask and N_MEMORY, which is necessary in
3202 * order to make cpusets transparent (of no affect) on systems that are
3203 * actively using CPU hotplug but making no active use of cpusets.
3205 * Non-root cpusets are only affected by offlining. If any CPUs or memory
3206 * nodes have been taken down, cpuset_hotplug_update_tasks() is invoked on
3209 * Note that CPU offlining during suspend is ignored. We don't modify
3210 * cpusets across suspend/resume cycles at all.
3212 static void cpuset_hotplug_workfn(struct work_struct *work)
3214 static cpumask_t new_cpus;
3215 static nodemask_t new_mems;
3216 bool cpus_updated, mems_updated;
3217 bool on_dfl = is_in_v2_mode();
3218 struct tmpmasks tmp, *ptmp = NULL;
3220 if (on_dfl && !alloc_cpumasks(NULL, &tmp))
3223 percpu_down_write(&cpuset_rwsem);
3225 /* fetch the available cpus/mems and find out which changed how */
3226 cpumask_copy(&new_cpus, cpu_active_mask);
3227 new_mems = node_states[N_MEMORY];
3230 * If subparts_cpus is populated, it is likely that the check below
3231 * will produce a false positive on cpus_updated when the cpu list
3232 * isn't changed. It is extra work, but it is better to be safe.
3234 cpus_updated = !cpumask_equal(top_cpuset.effective_cpus, &new_cpus);
3235 mems_updated = !nodes_equal(top_cpuset.effective_mems, new_mems);
3238 * In the rare case that hotplug removes all the cpus in subparts_cpus,
3239 * we assumed that cpus are updated.
3241 if (!cpus_updated && top_cpuset.nr_subparts_cpus)
3242 cpus_updated = true;
3244 /* synchronize cpus_allowed to cpu_active_mask */
3246 spin_lock_irq(&callback_lock);
3248 cpumask_copy(top_cpuset.cpus_allowed, &new_cpus);
3250 * Make sure that CPUs allocated to child partitions
3251 * do not show up in effective_cpus. If no CPU is left,
3252 * we clear the subparts_cpus & let the child partitions
3253 * fight for the CPUs again.
3255 if (top_cpuset.nr_subparts_cpus) {
3256 if (cpumask_subset(&new_cpus,
3257 top_cpuset.subparts_cpus)) {
3258 top_cpuset.nr_subparts_cpus = 0;
3259 cpumask_clear(top_cpuset.subparts_cpus);
3261 cpumask_andnot(&new_cpus, &new_cpus,
3262 top_cpuset.subparts_cpus);
3265 cpumask_copy(top_cpuset.effective_cpus, &new_cpus);
3266 spin_unlock_irq(&callback_lock);
3267 /* we don't mess with cpumasks of tasks in top_cpuset */
3270 /* synchronize mems_allowed to N_MEMORY */
3272 spin_lock_irq(&callback_lock);
3274 top_cpuset.mems_allowed = new_mems;
3275 top_cpuset.effective_mems = new_mems;
3276 spin_unlock_irq(&callback_lock);
3277 update_tasks_nodemask(&top_cpuset);
3280 percpu_up_write(&cpuset_rwsem);
3282 /* if cpus or mems changed, we need to propagate to descendants */
3283 if (cpus_updated || mems_updated) {
3285 struct cgroup_subsys_state *pos_css;
3288 cpuset_for_each_descendant_pre(cs, pos_css, &top_cpuset) {
3289 if (cs == &top_cpuset || !css_tryget_online(&cs->css))
3293 cpuset_hotplug_update_tasks(cs, ptmp);
3301 /* rebuild sched domains if cpus_allowed has changed */
3302 if (cpus_updated || force_rebuild) {
3303 force_rebuild = false;
3304 rebuild_sched_domains();
3307 free_cpumasks(NULL, ptmp);
3310 void cpuset_update_active_cpus(void)
3313 * We're inside cpu hotplug critical region which usually nests
3314 * inside cgroup synchronization. Bounce actual hotplug processing
3315 * to a work item to avoid reverse locking order.
3317 schedule_work(&cpuset_hotplug_work);
3320 void cpuset_wait_for_hotplug(void)
3322 flush_work(&cpuset_hotplug_work);
3326 * Keep top_cpuset.mems_allowed tracking node_states[N_MEMORY].
3327 * Call this routine anytime after node_states[N_MEMORY] changes.
3328 * See cpuset_update_active_cpus() for CPU hotplug handling.
3330 static int cpuset_track_online_nodes(struct notifier_block *self,
3331 unsigned long action, void *arg)
3333 schedule_work(&cpuset_hotplug_work);
3337 static struct notifier_block cpuset_track_online_nodes_nb = {
3338 .notifier_call = cpuset_track_online_nodes,
3339 .priority = 10, /* ??! */
3343 * cpuset_init_smp - initialize cpus_allowed
3345 * Description: Finish top cpuset after cpu, node maps are initialized
3347 void __init cpuset_init_smp(void)
3350 * cpus_allowd/mems_allowed set to v2 values in the initial
3351 * cpuset_bind() call will be reset to v1 values in another
3352 * cpuset_bind() call when v1 cpuset is mounted.
3354 top_cpuset.old_mems_allowed = top_cpuset.mems_allowed;
3356 cpumask_copy(top_cpuset.effective_cpus, cpu_active_mask);
3357 top_cpuset.effective_mems = node_states[N_MEMORY];
3359 register_hotmemory_notifier(&cpuset_track_online_nodes_nb);
3361 cpuset_migrate_mm_wq = alloc_ordered_workqueue("cpuset_migrate_mm", 0);
3362 BUG_ON(!cpuset_migrate_mm_wq);
3366 * cpuset_cpus_allowed - return cpus_allowed mask from a tasks cpuset.
3367 * @tsk: pointer to task_struct from which to obtain cpuset->cpus_allowed.
3368 * @pmask: pointer to struct cpumask variable to receive cpus_allowed set.
3370 * Description: Returns the cpumask_var_t cpus_allowed of the cpuset
3371 * attached to the specified @tsk. Guaranteed to return some non-empty
3372 * subset of cpu_online_mask, even if this means going outside the
3376 void cpuset_cpus_allowed(struct task_struct *tsk, struct cpumask *pmask)
3378 unsigned long flags;
3380 spin_lock_irqsave(&callback_lock, flags);
3381 guarantee_online_cpus(tsk, pmask);
3382 spin_unlock_irqrestore(&callback_lock, flags);
3386 * cpuset_cpus_allowed_fallback - final fallback before complete catastrophe.
3387 * @tsk: pointer to task_struct with which the scheduler is struggling
3389 * Description: In the case that the scheduler cannot find an allowed cpu in
3390 * tsk->cpus_allowed, we fall back to task_cs(tsk)->cpus_allowed. In legacy
3391 * mode however, this value is the same as task_cs(tsk)->effective_cpus,
3392 * which will not contain a sane cpumask during cases such as cpu hotplugging.
3393 * This is the absolute last resort for the scheduler and it is only used if
3394 * _every_ other avenue has been traveled.
3396 * Returns true if the affinity of @tsk was changed, false otherwise.
3399 bool cpuset_cpus_allowed_fallback(struct task_struct *tsk)
3401 const struct cpumask *possible_mask = task_cpu_possible_mask(tsk);
3402 const struct cpumask *cs_mask;
3403 bool changed = false;
3406 cs_mask = task_cs(tsk)->cpus_allowed;
3407 if (is_in_v2_mode() && cpumask_subset(cs_mask, possible_mask)) {
3408 do_set_cpus_allowed(tsk, cs_mask);
3414 * We own tsk->cpus_allowed, nobody can change it under us.
3416 * But we used cs && cs->cpus_allowed lockless and thus can
3417 * race with cgroup_attach_task() or update_cpumask() and get
3418 * the wrong tsk->cpus_allowed. However, both cases imply the
3419 * subsequent cpuset_change_cpumask()->set_cpus_allowed_ptr()
3420 * which takes task_rq_lock().
3422 * If we are called after it dropped the lock we must see all
3423 * changes in tsk_cs()->cpus_allowed. Otherwise we can temporary
3424 * set any mask even if it is not right from task_cs() pov,
3425 * the pending set_cpus_allowed_ptr() will fix things.
3427 * select_fallback_rq() will fix things ups and set cpu_possible_mask
3433 void __init cpuset_init_current_mems_allowed(void)
3435 nodes_setall(current->mems_allowed);
3439 * cpuset_mems_allowed - return mems_allowed mask from a tasks cpuset.
3440 * @tsk: pointer to task_struct from which to obtain cpuset->mems_allowed.
3442 * Description: Returns the nodemask_t mems_allowed of the cpuset
3443 * attached to the specified @tsk. Guaranteed to return some non-empty
3444 * subset of node_states[N_MEMORY], even if this means going outside the
3448 nodemask_t cpuset_mems_allowed(struct task_struct *tsk)
3451 unsigned long flags;
3453 spin_lock_irqsave(&callback_lock, flags);
3455 guarantee_online_mems(task_cs(tsk), &mask);
3457 spin_unlock_irqrestore(&callback_lock, flags);
3463 * cpuset_nodemask_valid_mems_allowed - check nodemask vs. current mems_allowed
3464 * @nodemask: the nodemask to be checked
3466 * Are any of the nodes in the nodemask allowed in current->mems_allowed?
3468 int cpuset_nodemask_valid_mems_allowed(nodemask_t *nodemask)
3470 return nodes_intersects(*nodemask, current->mems_allowed);
3474 * nearest_hardwall_ancestor() - Returns the nearest mem_exclusive or
3475 * mem_hardwall ancestor to the specified cpuset. Call holding
3476 * callback_lock. If no ancestor is mem_exclusive or mem_hardwall
3477 * (an unusual configuration), then returns the root cpuset.
3479 static struct cpuset *nearest_hardwall_ancestor(struct cpuset *cs)
3481 while (!(is_mem_exclusive(cs) || is_mem_hardwall(cs)) && parent_cs(cs))
3487 * cpuset_node_allowed - Can we allocate on a memory node?
3488 * @node: is this an allowed node?
3489 * @gfp_mask: memory allocation flags
3491 * If we're in interrupt, yes, we can always allocate. If @node is set in
3492 * current's mems_allowed, yes. If it's not a __GFP_HARDWALL request and this
3493 * node is set in the nearest hardwalled cpuset ancestor to current's cpuset,
3494 * yes. If current has access to memory reserves as an oom victim, yes.
3497 * GFP_USER allocations are marked with the __GFP_HARDWALL bit,
3498 * and do not allow allocations outside the current tasks cpuset
3499 * unless the task has been OOM killed.
3500 * GFP_KERNEL allocations are not so marked, so can escape to the
3501 * nearest enclosing hardwalled ancestor cpuset.
3503 * Scanning up parent cpusets requires callback_lock. The
3504 * __alloc_pages() routine only calls here with __GFP_HARDWALL bit
3505 * _not_ set if it's a GFP_KERNEL allocation, and all nodes in the
3506 * current tasks mems_allowed came up empty on the first pass over
3507 * the zonelist. So only GFP_KERNEL allocations, if all nodes in the
3508 * cpuset are short of memory, might require taking the callback_lock.
3510 * The first call here from mm/page_alloc:get_page_from_freelist()
3511 * has __GFP_HARDWALL set in gfp_mask, enforcing hardwall cpusets,
3512 * so no allocation on a node outside the cpuset is allowed (unless
3513 * in interrupt, of course).
3515 * The second pass through get_page_from_freelist() doesn't even call
3516 * here for GFP_ATOMIC calls. For those calls, the __alloc_pages()
3517 * variable 'wait' is not set, and the bit ALLOC_CPUSET is not set
3518 * in alloc_flags. That logic and the checks below have the combined
3520 * in_interrupt - any node ok (current task context irrelevant)
3521 * GFP_ATOMIC - any node ok
3522 * tsk_is_oom_victim - any node ok
3523 * GFP_KERNEL - any node in enclosing hardwalled cpuset ok
3524 * GFP_USER - only nodes in current tasks mems allowed ok.
3526 bool __cpuset_node_allowed(int node, gfp_t gfp_mask)
3528 struct cpuset *cs; /* current cpuset ancestors */
3529 int allowed; /* is allocation in zone z allowed? */
3530 unsigned long flags;
3534 if (node_isset(node, current->mems_allowed))
3537 * Allow tasks that have access to memory reserves because they have
3538 * been OOM killed to get memory anywhere.
3540 if (unlikely(tsk_is_oom_victim(current)))
3542 if (gfp_mask & __GFP_HARDWALL) /* If hardwall request, stop here */
3545 if (current->flags & PF_EXITING) /* Let dying task have memory */
3548 /* Not hardwall and node outside mems_allowed: scan up cpusets */
3549 spin_lock_irqsave(&callback_lock, flags);
3552 cs = nearest_hardwall_ancestor(task_cs(current));
3553 allowed = node_isset(node, cs->mems_allowed);
3556 spin_unlock_irqrestore(&callback_lock, flags);
3561 * cpuset_mem_spread_node() - On which node to begin search for a file page
3562 * cpuset_slab_spread_node() - On which node to begin search for a slab page
3564 * If a task is marked PF_SPREAD_PAGE or PF_SPREAD_SLAB (as for
3565 * tasks in a cpuset with is_spread_page or is_spread_slab set),
3566 * and if the memory allocation used cpuset_mem_spread_node()
3567 * to determine on which node to start looking, as it will for
3568 * certain page cache or slab cache pages such as used for file
3569 * system buffers and inode caches, then instead of starting on the
3570 * local node to look for a free page, rather spread the starting
3571 * node around the tasks mems_allowed nodes.
3573 * We don't have to worry about the returned node being offline
3574 * because "it can't happen", and even if it did, it would be ok.
3576 * The routines calling guarantee_online_mems() are careful to
3577 * only set nodes in task->mems_allowed that are online. So it
3578 * should not be possible for the following code to return an
3579 * offline node. But if it did, that would be ok, as this routine
3580 * is not returning the node where the allocation must be, only
3581 * the node where the search should start. The zonelist passed to
3582 * __alloc_pages() will include all nodes. If the slab allocator
3583 * is passed an offline node, it will fall back to the local node.
3584 * See kmem_cache_alloc_node().
3587 static int cpuset_spread_node(int *rotor)
3589 return *rotor = next_node_in(*rotor, current->mems_allowed);
3592 int cpuset_mem_spread_node(void)
3594 if (current->cpuset_mem_spread_rotor == NUMA_NO_NODE)
3595 current->cpuset_mem_spread_rotor =
3596 node_random(¤t->mems_allowed);
3598 return cpuset_spread_node(¤t->cpuset_mem_spread_rotor);
3601 int cpuset_slab_spread_node(void)
3603 if (current->cpuset_slab_spread_rotor == NUMA_NO_NODE)
3604 current->cpuset_slab_spread_rotor =
3605 node_random(¤t->mems_allowed);
3607 return cpuset_spread_node(¤t->cpuset_slab_spread_rotor);
3610 EXPORT_SYMBOL_GPL(cpuset_mem_spread_node);
3613 * cpuset_mems_allowed_intersects - Does @tsk1's mems_allowed intersect @tsk2's?
3614 * @tsk1: pointer to task_struct of some task.
3615 * @tsk2: pointer to task_struct of some other task.
3617 * Description: Return true if @tsk1's mems_allowed intersects the
3618 * mems_allowed of @tsk2. Used by the OOM killer to determine if
3619 * one of the task's memory usage might impact the memory available
3623 int cpuset_mems_allowed_intersects(const struct task_struct *tsk1,
3624 const struct task_struct *tsk2)
3626 return nodes_intersects(tsk1->mems_allowed, tsk2->mems_allowed);
3630 * cpuset_print_current_mems_allowed - prints current's cpuset and mems_allowed
3632 * Description: Prints current's name, cpuset name, and cached copy of its
3633 * mems_allowed to the kernel log.
3635 void cpuset_print_current_mems_allowed(void)
3637 struct cgroup *cgrp;
3641 cgrp = task_cs(current)->css.cgroup;
3642 pr_cont(",cpuset=");
3643 pr_cont_cgroup_name(cgrp);
3644 pr_cont(",mems_allowed=%*pbl",
3645 nodemask_pr_args(¤t->mems_allowed));
3651 * Collection of memory_pressure is suppressed unless
3652 * this flag is enabled by writing "1" to the special
3653 * cpuset file 'memory_pressure_enabled' in the root cpuset.
3656 int cpuset_memory_pressure_enabled __read_mostly;
3659 * cpuset_memory_pressure_bump - keep stats of per-cpuset reclaims.
3661 * Keep a running average of the rate of synchronous (direct)
3662 * page reclaim efforts initiated by tasks in each cpuset.
3664 * This represents the rate at which some task in the cpuset
3665 * ran low on memory on all nodes it was allowed to use, and
3666 * had to enter the kernels page reclaim code in an effort to
3667 * create more free memory by tossing clean pages or swapping
3668 * or writing dirty pages.
3670 * Display to user space in the per-cpuset read-only file
3671 * "memory_pressure". Value displayed is an integer
3672 * representing the recent rate of entry into the synchronous
3673 * (direct) page reclaim by any task attached to the cpuset.
3676 void __cpuset_memory_pressure_bump(void)
3679 fmeter_markevent(&task_cs(current)->fmeter);
3683 #ifdef CONFIG_PROC_PID_CPUSET
3685 * proc_cpuset_show()
3686 * - Print tasks cpuset path into seq_file.
3687 * - Used for /proc/<pid>/cpuset.
3688 * - No need to task_lock(tsk) on this tsk->cpuset reference, as it
3689 * doesn't really matter if tsk->cpuset changes after we read it,
3690 * and we take cpuset_rwsem, keeping cpuset_attach() from changing it
3693 int proc_cpuset_show(struct seq_file *m, struct pid_namespace *ns,
3694 struct pid *pid, struct task_struct *tsk)
3697 struct cgroup_subsys_state *css;
3701 buf = kmalloc(PATH_MAX, GFP_KERNEL);
3705 css = task_get_css(tsk, cpuset_cgrp_id);
3706 retval = cgroup_path_ns(css->cgroup, buf, PATH_MAX,
3707 current->nsproxy->cgroup_ns);
3709 if (retval >= PATH_MAX)
3710 retval = -ENAMETOOLONG;
3721 #endif /* CONFIG_PROC_PID_CPUSET */
3723 /* Display task mems_allowed in /proc/<pid>/status file. */
3724 void cpuset_task_status_allowed(struct seq_file *m, struct task_struct *task)
3726 seq_printf(m, "Mems_allowed:\t%*pb\n",
3727 nodemask_pr_args(&task->mems_allowed));
3728 seq_printf(m, "Mems_allowed_list:\t%*pbl\n",
3729 nodemask_pr_args(&task->mems_allowed));