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_mutex 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
319 * A task must hold both locks to modify cpusets. If a task holds
320 * cpuset_mutex, then it blocks others wanting that mutex, ensuring that it
321 * is the only task able to also acquire callback_lock and be able to
322 * modify cpusets. It can perform various checks on the cpuset structure
323 * first, knowing nothing will change. It can also allocate memory while
324 * just holding cpuset_mutex. While it is performing these checks, various
325 * callback routines can briefly acquire callback_lock to query cpusets.
326 * Once it is ready to make the changes, it takes callback_lock, blocking
329 * Calls to the kernel memory allocator can not be made while holding
330 * callback_lock, as that would risk double tripping on callback_lock
331 * from one of the callbacks into the cpuset code from within
334 * If a task is only holding callback_lock, then it has read-only
337 * Now, the task_struct fields mems_allowed and mempolicy may be changed
338 * by other task, we use alloc_lock in the task_struct fields to protect
341 * The cpuset_common_file_read() handlers only hold callback_lock across
342 * small pieces of code, such as when reading out possibly multi-word
343 * cpumasks and nodemasks.
345 * Accessing a task's cpuset should be done in accordance with the
346 * guidelines for accessing subsystem state in kernel/cgroup.c
349 DEFINE_STATIC_PERCPU_RWSEM(cpuset_rwsem);
351 void cpuset_read_lock(void)
353 percpu_down_read(&cpuset_rwsem);
356 void cpuset_read_unlock(void)
358 percpu_up_read(&cpuset_rwsem);
361 static DEFINE_SPINLOCK(callback_lock);
363 static struct workqueue_struct *cpuset_migrate_mm_wq;
366 * CPU / memory hotplug is handled asynchronously.
368 static void cpuset_hotplug_workfn(struct work_struct *work);
369 static DECLARE_WORK(cpuset_hotplug_work, cpuset_hotplug_workfn);
371 static DECLARE_WAIT_QUEUE_HEAD(cpuset_attach_wq);
374 * Cgroup v2 behavior is used on the "cpus" and "mems" control files when
375 * on default hierarchy or when the cpuset_v2_mode flag is set by mounting
376 * the v1 cpuset cgroup filesystem with the "cpuset_v2_mode" mount option.
377 * With v2 behavior, "cpus" and "mems" are always what the users have
378 * requested and won't be changed by hotplug events. Only the effective
379 * cpus or mems will be affected.
381 static inline bool is_in_v2_mode(void)
383 return cgroup_subsys_on_dfl(cpuset_cgrp_subsys) ||
384 (cpuset_cgrp_subsys.root->flags & CGRP_ROOT_CPUSET_V2_MODE);
388 * Return in pmask the portion of a cpusets's cpus_allowed that
389 * are online. If none are online, walk up the cpuset hierarchy
390 * until we find one that does have some online cpus.
392 * One way or another, we guarantee to return some non-empty subset
393 * of cpu_online_mask.
395 * Call with callback_lock or cpuset_mutex held.
397 static void guarantee_online_cpus(struct cpuset *cs, struct cpumask *pmask)
399 while (!cpumask_intersects(cs->effective_cpus, cpu_online_mask)) {
403 * The top cpuset doesn't have any online cpu as a
404 * consequence of a race between cpuset_hotplug_work
405 * and cpu hotplug notifier. But we know the top
406 * cpuset's effective_cpus is on its way to be
407 * identical to cpu_online_mask.
409 cpumask_copy(pmask, cpu_online_mask);
413 cpumask_and(pmask, cs->effective_cpus, cpu_online_mask);
417 * Return in *pmask the portion of a cpusets's mems_allowed that
418 * are online, with memory. If none are online with memory, walk
419 * up the cpuset hierarchy until we find one that does have some
420 * online mems. The top cpuset always has some mems online.
422 * One way or another, we guarantee to return some non-empty subset
423 * of node_states[N_MEMORY].
425 * Call with callback_lock or cpuset_mutex held.
427 static void guarantee_online_mems(struct cpuset *cs, nodemask_t *pmask)
429 while (!nodes_intersects(cs->effective_mems, node_states[N_MEMORY]))
431 nodes_and(*pmask, cs->effective_mems, node_states[N_MEMORY]);
435 * update task's spread flag if cpuset's page/slab spread flag is set
437 * Call with callback_lock or cpuset_mutex held.
439 static void cpuset_update_task_spread_flag(struct cpuset *cs,
440 struct task_struct *tsk)
442 if (is_spread_page(cs))
443 task_set_spread_page(tsk);
445 task_clear_spread_page(tsk);
447 if (is_spread_slab(cs))
448 task_set_spread_slab(tsk);
450 task_clear_spread_slab(tsk);
454 * is_cpuset_subset(p, q) - Is cpuset p a subset of cpuset q?
456 * One cpuset is a subset of another if all its allowed CPUs and
457 * Memory Nodes are a subset of the other, and its exclusive flags
458 * are only set if the other's are set. Call holding cpuset_mutex.
461 static int is_cpuset_subset(const struct cpuset *p, const struct cpuset *q)
463 return cpumask_subset(p->cpus_allowed, q->cpus_allowed) &&
464 nodes_subset(p->mems_allowed, q->mems_allowed) &&
465 is_cpu_exclusive(p) <= is_cpu_exclusive(q) &&
466 is_mem_exclusive(p) <= is_mem_exclusive(q);
470 * alloc_cpumasks - allocate three cpumasks for cpuset
471 * @cs: the cpuset that have cpumasks to be allocated.
472 * @tmp: the tmpmasks structure pointer
473 * Return: 0 if successful, -ENOMEM otherwise.
475 * Only one of the two input arguments should be non-NULL.
477 static inline int alloc_cpumasks(struct cpuset *cs, struct tmpmasks *tmp)
479 cpumask_var_t *pmask1, *pmask2, *pmask3;
482 pmask1 = &cs->cpus_allowed;
483 pmask2 = &cs->effective_cpus;
484 pmask3 = &cs->subparts_cpus;
486 pmask1 = &tmp->new_cpus;
487 pmask2 = &tmp->addmask;
488 pmask3 = &tmp->delmask;
491 if (!zalloc_cpumask_var(pmask1, GFP_KERNEL))
494 if (!zalloc_cpumask_var(pmask2, GFP_KERNEL))
497 if (!zalloc_cpumask_var(pmask3, GFP_KERNEL))
503 free_cpumask_var(*pmask2);
505 free_cpumask_var(*pmask1);
510 * free_cpumasks - free cpumasks in a tmpmasks structure
511 * @cs: the cpuset that have cpumasks to be free.
512 * @tmp: the tmpmasks structure pointer
514 static inline void free_cpumasks(struct cpuset *cs, struct tmpmasks *tmp)
517 free_cpumask_var(cs->cpus_allowed);
518 free_cpumask_var(cs->effective_cpus);
519 free_cpumask_var(cs->subparts_cpus);
522 free_cpumask_var(tmp->new_cpus);
523 free_cpumask_var(tmp->addmask);
524 free_cpumask_var(tmp->delmask);
529 * alloc_trial_cpuset - allocate a trial cpuset
530 * @cs: the cpuset that the trial cpuset duplicates
532 static struct cpuset *alloc_trial_cpuset(struct cpuset *cs)
534 struct cpuset *trial;
536 trial = kmemdup(cs, sizeof(*cs), GFP_KERNEL);
540 if (alloc_cpumasks(trial, NULL)) {
545 cpumask_copy(trial->cpus_allowed, cs->cpus_allowed);
546 cpumask_copy(trial->effective_cpus, cs->effective_cpus);
551 * free_cpuset - free the cpuset
552 * @cs: the cpuset to be freed
554 static inline void free_cpuset(struct cpuset *cs)
556 free_cpumasks(cs, NULL);
561 * validate_change() - Used to validate that any proposed cpuset change
562 * follows the structural rules for cpusets.
564 * If we replaced the flag and mask values of the current cpuset
565 * (cur) with those values in the trial cpuset (trial), would
566 * our various subset and exclusive rules still be valid? Presumes
569 * 'cur' is the address of an actual, in-use cpuset. Operations
570 * such as list traversal that depend on the actual address of the
571 * cpuset in the list must use cur below, not trial.
573 * 'trial' is the address of bulk structure copy of cur, with
574 * perhaps one or more of the fields cpus_allowed, mems_allowed,
575 * or flags changed to new, trial values.
577 * Return 0 if valid, -errno if not.
580 static int validate_change(struct cpuset *cur, struct cpuset *trial)
582 struct cgroup_subsys_state *css;
583 struct cpuset *c, *par;
588 /* Each of our child cpusets must be a subset of us */
590 cpuset_for_each_child(c, css, cur)
591 if (!is_cpuset_subset(c, trial))
594 /* Remaining checks don't apply to root cpuset */
596 if (cur == &top_cpuset)
599 par = parent_cs(cur);
601 /* On legacy hierarchy, we must be a subset of our parent cpuset. */
603 if (!is_in_v2_mode() && !is_cpuset_subset(trial, par))
607 * If either I or some sibling (!= me) is exclusive, we can't
611 cpuset_for_each_child(c, css, par) {
612 if ((is_cpu_exclusive(trial) || is_cpu_exclusive(c)) &&
614 cpumask_intersects(trial->cpus_allowed, c->cpus_allowed))
616 if ((is_mem_exclusive(trial) || is_mem_exclusive(c)) &&
618 nodes_intersects(trial->mems_allowed, c->mems_allowed))
623 * Cpusets with tasks - existing or newly being attached - can't
624 * be changed to have empty cpus_allowed or mems_allowed.
627 if ((cgroup_is_populated(cur->css.cgroup) || cur->attach_in_progress)) {
628 if (!cpumask_empty(cur->cpus_allowed) &&
629 cpumask_empty(trial->cpus_allowed))
631 if (!nodes_empty(cur->mems_allowed) &&
632 nodes_empty(trial->mems_allowed))
637 * We can't shrink if we won't have enough room for SCHED_DEADLINE
641 if (is_cpu_exclusive(cur) &&
642 !cpuset_cpumask_can_shrink(cur->cpus_allowed,
643 trial->cpus_allowed))
654 * Helper routine for generate_sched_domains().
655 * Do cpusets a, b have overlapping effective cpus_allowed masks?
657 static int cpusets_overlap(struct cpuset *a, struct cpuset *b)
659 return cpumask_intersects(a->effective_cpus, b->effective_cpus);
663 update_domain_attr(struct sched_domain_attr *dattr, struct cpuset *c)
665 if (dattr->relax_domain_level < c->relax_domain_level)
666 dattr->relax_domain_level = c->relax_domain_level;
670 static void update_domain_attr_tree(struct sched_domain_attr *dattr,
671 struct cpuset *root_cs)
674 struct cgroup_subsys_state *pos_css;
677 cpuset_for_each_descendant_pre(cp, pos_css, root_cs) {
678 /* skip the whole subtree if @cp doesn't have any CPU */
679 if (cpumask_empty(cp->cpus_allowed)) {
680 pos_css = css_rightmost_descendant(pos_css);
684 if (is_sched_load_balance(cp))
685 update_domain_attr(dattr, cp);
690 /* Must be called with cpuset_mutex held. */
691 static inline int nr_cpusets(void)
693 /* jump label reference count + the top-level cpuset */
694 return static_key_count(&cpusets_enabled_key.key) + 1;
698 * generate_sched_domains()
700 * This function builds a partial partition of the systems CPUs
701 * A 'partial partition' is a set of non-overlapping subsets whose
702 * union is a subset of that set.
703 * The output of this function needs to be passed to kernel/sched/core.c
704 * partition_sched_domains() routine, which will rebuild the scheduler's
705 * load balancing domains (sched domains) as specified by that partial
708 * See "What is sched_load_balance" in Documentation/admin-guide/cgroup-v1/cpusets.rst
709 * for a background explanation of this.
711 * Does not return errors, on the theory that the callers of this
712 * routine would rather not worry about failures to rebuild sched
713 * domains when operating in the severe memory shortage situations
714 * that could cause allocation failures below.
716 * Must be called with cpuset_mutex held.
718 * The three key local variables below are:
719 * cp - cpuset pointer, used (together with pos_css) to perform a
720 * top-down scan of all cpusets. For our purposes, rebuilding
721 * the schedulers sched domains, we can ignore !is_sched_load_
723 * csa - (for CpuSet Array) Array of pointers to all the cpusets
724 * that need to be load balanced, for convenient iterative
725 * access by the subsequent code that finds the best partition,
726 * i.e the set of domains (subsets) of CPUs such that the
727 * cpus_allowed of every cpuset marked is_sched_load_balance
728 * is a subset of one of these domains, while there are as
729 * many such domains as possible, each as small as possible.
730 * doms - Conversion of 'csa' to an array of cpumasks, for passing to
731 * the kernel/sched/core.c routine partition_sched_domains() in a
732 * convenient format, that can be easily compared to the prior
733 * value to determine what partition elements (sched domains)
734 * were changed (added or removed.)
736 * Finding the best partition (set of domains):
737 * The triple nested loops below over i, j, k scan over the
738 * load balanced cpusets (using the array of cpuset pointers in
739 * csa[]) looking for pairs of cpusets that have overlapping
740 * cpus_allowed, but which don't have the same 'pn' partition
741 * number and gives them in the same partition number. It keeps
742 * looping on the 'restart' label until it can no longer find
745 * The union of the cpus_allowed masks from the set of
746 * all cpusets having the same 'pn' value then form the one
747 * element of the partition (one sched domain) to be passed to
748 * partition_sched_domains().
750 static int generate_sched_domains(cpumask_var_t **domains,
751 struct sched_domain_attr **attributes)
753 struct cpuset *cp; /* top-down scan of cpusets */
754 struct cpuset **csa; /* array of all cpuset ptrs */
755 int csn; /* how many cpuset ptrs in csa so far */
756 int i, j, k; /* indices for partition finding loops */
757 cpumask_var_t *doms; /* resulting partition; i.e. sched domains */
758 struct sched_domain_attr *dattr; /* attributes for custom domains */
759 int ndoms = 0; /* number of sched domains in result */
760 int nslot; /* next empty doms[] struct cpumask slot */
761 struct cgroup_subsys_state *pos_css;
762 bool root_load_balance = is_sched_load_balance(&top_cpuset);
768 /* Special case for the 99% of systems with one, full, sched domain */
769 if (root_load_balance && !top_cpuset.nr_subparts_cpus) {
771 doms = alloc_sched_domains(ndoms);
775 dattr = kmalloc(sizeof(struct sched_domain_attr), GFP_KERNEL);
777 *dattr = SD_ATTR_INIT;
778 update_domain_attr_tree(dattr, &top_cpuset);
780 cpumask_and(doms[0], top_cpuset.effective_cpus,
781 housekeeping_cpumask(HK_FLAG_DOMAIN));
786 csa = kmalloc_array(nr_cpusets(), sizeof(cp), GFP_KERNEL);
792 if (root_load_balance)
793 csa[csn++] = &top_cpuset;
794 cpuset_for_each_descendant_pre(cp, pos_css, &top_cpuset) {
795 if (cp == &top_cpuset)
798 * Continue traversing beyond @cp iff @cp has some CPUs and
799 * isn't load balancing. The former is obvious. The
800 * latter: All child cpusets contain a subset of the
801 * parent's cpus, so just skip them, and then we call
802 * update_domain_attr_tree() to calc relax_domain_level of
803 * the corresponding sched domain.
805 * If root is load-balancing, we can skip @cp if it
806 * is a subset of the root's effective_cpus.
808 if (!cpumask_empty(cp->cpus_allowed) &&
809 !(is_sched_load_balance(cp) &&
810 cpumask_intersects(cp->cpus_allowed,
811 housekeeping_cpumask(HK_FLAG_DOMAIN))))
814 if (root_load_balance &&
815 cpumask_subset(cp->cpus_allowed, top_cpuset.effective_cpus))
818 if (is_sched_load_balance(cp) &&
819 !cpumask_empty(cp->effective_cpus))
822 /* skip @cp's subtree if not a partition root */
823 if (!is_partition_root(cp))
824 pos_css = css_rightmost_descendant(pos_css);
828 for (i = 0; i < csn; i++)
833 /* Find the best partition (set of sched domains) */
834 for (i = 0; i < csn; i++) {
835 struct cpuset *a = csa[i];
838 for (j = 0; j < csn; j++) {
839 struct cpuset *b = csa[j];
842 if (apn != bpn && cpusets_overlap(a, b)) {
843 for (k = 0; k < csn; k++) {
844 struct cpuset *c = csa[k];
849 ndoms--; /* one less element */
856 * Now we know how many domains to create.
857 * Convert <csn, csa> to <ndoms, doms> and populate cpu masks.
859 doms = alloc_sched_domains(ndoms);
864 * The rest of the code, including the scheduler, can deal with
865 * dattr==NULL case. No need to abort if alloc fails.
867 dattr = kmalloc_array(ndoms, sizeof(struct sched_domain_attr),
870 for (nslot = 0, i = 0; i < csn; i++) {
871 struct cpuset *a = csa[i];
876 /* Skip completed partitions */
882 if (nslot == ndoms) {
883 static int warnings = 10;
885 pr_warn("rebuild_sched_domains confused: nslot %d, ndoms %d, csn %d, i %d, apn %d\n",
886 nslot, ndoms, csn, i, apn);
894 *(dattr + nslot) = SD_ATTR_INIT;
895 for (j = i; j < csn; j++) {
896 struct cpuset *b = csa[j];
899 cpumask_or(dp, dp, b->effective_cpus);
900 cpumask_and(dp, dp, housekeeping_cpumask(HK_FLAG_DOMAIN));
902 update_domain_attr_tree(dattr + nslot, b);
904 /* Done with this partition */
910 BUG_ON(nslot != ndoms);
916 * Fallback to the default domain if kmalloc() failed.
917 * See comments in partition_sched_domains().
927 static void update_tasks_root_domain(struct cpuset *cs)
929 struct css_task_iter it;
930 struct task_struct *task;
932 css_task_iter_start(&cs->css, 0, &it);
934 while ((task = css_task_iter_next(&it)))
935 dl_add_task_root_domain(task);
937 css_task_iter_end(&it);
940 static void rebuild_root_domains(void)
942 struct cpuset *cs = NULL;
943 struct cgroup_subsys_state *pos_css;
945 percpu_rwsem_assert_held(&cpuset_rwsem);
946 lockdep_assert_cpus_held();
947 lockdep_assert_held(&sched_domains_mutex);
952 * Clear default root domain DL accounting, it will be computed again
953 * if a task belongs to it.
955 dl_clear_root_domain(&def_root_domain);
957 cpuset_for_each_descendant_pre(cs, pos_css, &top_cpuset) {
959 if (cpumask_empty(cs->effective_cpus)) {
960 pos_css = css_rightmost_descendant(pos_css);
968 update_tasks_root_domain(cs);
977 partition_and_rebuild_sched_domains(int ndoms_new, cpumask_var_t doms_new[],
978 struct sched_domain_attr *dattr_new)
980 mutex_lock(&sched_domains_mutex);
981 partition_sched_domains_locked(ndoms_new, doms_new, dattr_new);
982 rebuild_root_domains();
983 mutex_unlock(&sched_domains_mutex);
987 * Rebuild scheduler domains.
989 * If the flag 'sched_load_balance' of any cpuset with non-empty
990 * 'cpus' changes, or if the 'cpus' allowed changes in any cpuset
991 * which has that flag enabled, or if any cpuset with a non-empty
992 * 'cpus' is removed, then call this routine to rebuild the
993 * scheduler's dynamic sched domains.
995 * Call with cpuset_mutex held. Takes cpus_read_lock().
997 static void rebuild_sched_domains_locked(void)
999 struct cgroup_subsys_state *pos_css;
1000 struct sched_domain_attr *attr;
1001 cpumask_var_t *doms;
1005 lockdep_assert_cpus_held();
1006 percpu_rwsem_assert_held(&cpuset_rwsem);
1009 * If we have raced with CPU hotplug, return early to avoid
1010 * passing doms with offlined cpu to partition_sched_domains().
1011 * Anyways, cpuset_hotplug_workfn() will rebuild sched domains.
1013 * With no CPUs in any subpartitions, top_cpuset's effective CPUs
1014 * should be the same as the active CPUs, so checking only top_cpuset
1015 * is enough to detect racing CPU offlines.
1017 if (!top_cpuset.nr_subparts_cpus &&
1018 !cpumask_equal(top_cpuset.effective_cpus, cpu_active_mask))
1022 * With subpartition CPUs, however, the effective CPUs of a partition
1023 * root should be only a subset of the active CPUs. Since a CPU in any
1024 * partition root could be offlined, all must be checked.
1026 if (top_cpuset.nr_subparts_cpus) {
1028 cpuset_for_each_descendant_pre(cs, pos_css, &top_cpuset) {
1029 if (!is_partition_root(cs)) {
1030 pos_css = css_rightmost_descendant(pos_css);
1033 if (!cpumask_subset(cs->effective_cpus,
1042 /* Generate domain masks and attrs */
1043 ndoms = generate_sched_domains(&doms, &attr);
1045 /* Have scheduler rebuild the domains */
1046 partition_and_rebuild_sched_domains(ndoms, doms, attr);
1048 #else /* !CONFIG_SMP */
1049 static void rebuild_sched_domains_locked(void)
1052 #endif /* CONFIG_SMP */
1054 void rebuild_sched_domains(void)
1057 percpu_down_write(&cpuset_rwsem);
1058 rebuild_sched_domains_locked();
1059 percpu_up_write(&cpuset_rwsem);
1064 * update_tasks_cpumask - Update the cpumasks of tasks in the cpuset.
1065 * @cs: the cpuset in which each task's cpus_allowed mask needs to be changed
1067 * Iterate through each task of @cs updating its cpus_allowed to the
1068 * effective cpuset's. As this function is called with cpuset_mutex held,
1069 * cpuset membership stays stable.
1071 static void update_tasks_cpumask(struct cpuset *cs)
1073 struct css_task_iter it;
1074 struct task_struct *task;
1076 css_task_iter_start(&cs->css, 0, &it);
1077 while ((task = css_task_iter_next(&it)))
1078 set_cpus_allowed_ptr(task, cs->effective_cpus);
1079 css_task_iter_end(&it);
1083 * compute_effective_cpumask - Compute the effective cpumask of the cpuset
1084 * @new_cpus: the temp variable for the new effective_cpus mask
1085 * @cs: the cpuset the need to recompute the new effective_cpus mask
1086 * @parent: the parent cpuset
1088 * If the parent has subpartition CPUs, include them in the list of
1089 * allowable CPUs in computing the new effective_cpus mask. Since offlined
1090 * CPUs are not removed from subparts_cpus, we have to use cpu_active_mask
1091 * to mask those out.
1093 static void compute_effective_cpumask(struct cpumask *new_cpus,
1094 struct cpuset *cs, struct cpuset *parent)
1096 if (parent->nr_subparts_cpus) {
1097 cpumask_or(new_cpus, parent->effective_cpus,
1098 parent->subparts_cpus);
1099 cpumask_and(new_cpus, new_cpus, cs->cpus_allowed);
1100 cpumask_and(new_cpus, new_cpus, cpu_active_mask);
1102 cpumask_and(new_cpus, cs->cpus_allowed, parent->effective_cpus);
1107 * Commands for update_parent_subparts_cpumask
1110 partcmd_enable, /* Enable partition root */
1111 partcmd_disable, /* Disable partition root */
1112 partcmd_update, /* Update parent's subparts_cpus */
1116 * update_parent_subparts_cpumask - update subparts_cpus mask of parent cpuset
1117 * @cpuset: The cpuset that requests change in partition root state
1118 * @cmd: Partition root state change command
1119 * @newmask: Optional new cpumask for partcmd_update
1120 * @tmp: Temporary addmask and delmask
1121 * Return: 0, 1 or an error code
1123 * For partcmd_enable, the cpuset is being transformed from a non-partition
1124 * root to a partition root. The cpus_allowed mask of the given cpuset will
1125 * be put into parent's subparts_cpus and taken away from parent's
1126 * effective_cpus. The function will return 0 if all the CPUs listed in
1127 * cpus_allowed can be granted or an error code will be returned.
1129 * For partcmd_disable, the cpuset is being transofrmed from a partition
1130 * root back to a non-partition root. Any CPUs in cpus_allowed that are in
1131 * parent's subparts_cpus will be taken away from that cpumask and put back
1132 * into parent's effective_cpus. 0 should always be returned.
1134 * For partcmd_update, if the optional newmask is specified, the cpu
1135 * list is to be changed from cpus_allowed to newmask. Otherwise,
1136 * cpus_allowed is assumed to remain the same. The cpuset should either
1137 * be a partition root or an invalid partition root. The partition root
1138 * state may change if newmask is NULL and none of the requested CPUs can
1139 * be granted by the parent. The function will return 1 if changes to
1140 * parent's subparts_cpus and effective_cpus happen or 0 otherwise.
1141 * Error code should only be returned when newmask is non-NULL.
1143 * The partcmd_enable and partcmd_disable commands are used by
1144 * update_prstate(). The partcmd_update command is used by
1145 * update_cpumasks_hier() with newmask NULL and update_cpumask() with
1148 * The checking is more strict when enabling partition root than the
1149 * other two commands.
1151 * Because of the implicit cpu exclusive nature of a partition root,
1152 * cpumask changes that violates the cpu exclusivity rule will not be
1153 * permitted when checked by validate_change(). The validate_change()
1154 * function will also prevent any changes to the cpu list if it is not
1155 * a superset of children's cpu lists.
1157 static int update_parent_subparts_cpumask(struct cpuset *cpuset, int cmd,
1158 struct cpumask *newmask,
1159 struct tmpmasks *tmp)
1161 struct cpuset *parent = parent_cs(cpuset);
1162 int adding; /* Moving cpus from effective_cpus to subparts_cpus */
1163 int deleting; /* Moving cpus from subparts_cpus to effective_cpus */
1164 int old_prs, new_prs;
1165 bool part_error = false; /* Partition error? */
1167 percpu_rwsem_assert_held(&cpuset_rwsem);
1170 * The parent must be a partition root.
1171 * The new cpumask, if present, or the current cpus_allowed must
1174 if (!is_partition_root(parent) ||
1175 (newmask && cpumask_empty(newmask)) ||
1176 (!newmask && cpumask_empty(cpuset->cpus_allowed)))
1180 * Enabling/disabling partition root is not allowed if there are
1183 if ((cmd != partcmd_update) && css_has_online_children(&cpuset->css))
1187 * Enabling partition root is not allowed if not all the CPUs
1188 * can be granted from parent's effective_cpus or at least one
1189 * CPU will be left after that.
1191 if ((cmd == partcmd_enable) &&
1192 (!cpumask_subset(cpuset->cpus_allowed, parent->effective_cpus) ||
1193 cpumask_equal(cpuset->cpus_allowed, parent->effective_cpus)))
1197 * A cpumask update cannot make parent's effective_cpus become empty.
1199 adding = deleting = false;
1200 old_prs = new_prs = cpuset->partition_root_state;
1201 if (cmd == partcmd_enable) {
1202 cpumask_copy(tmp->addmask, cpuset->cpus_allowed);
1204 } else if (cmd == partcmd_disable) {
1205 deleting = cpumask_and(tmp->delmask, cpuset->cpus_allowed,
1206 parent->subparts_cpus);
1207 } else if (newmask) {
1209 * partcmd_update with newmask:
1211 * delmask = cpus_allowed & ~newmask & parent->subparts_cpus
1212 * addmask = newmask & parent->effective_cpus
1213 * & ~parent->subparts_cpus
1215 cpumask_andnot(tmp->delmask, cpuset->cpus_allowed, newmask);
1216 deleting = cpumask_and(tmp->delmask, tmp->delmask,
1217 parent->subparts_cpus);
1219 cpumask_and(tmp->addmask, newmask, parent->effective_cpus);
1220 adding = cpumask_andnot(tmp->addmask, tmp->addmask,
1221 parent->subparts_cpus);
1223 * Return error if the new effective_cpus could become empty.
1226 cpumask_equal(parent->effective_cpus, tmp->addmask)) {
1230 * As some of the CPUs in subparts_cpus might have
1231 * been offlined, we need to compute the real delmask
1234 if (!cpumask_and(tmp->addmask, tmp->delmask,
1237 cpumask_copy(tmp->addmask, parent->effective_cpus);
1241 * partcmd_update w/o newmask:
1243 * addmask = cpus_allowed & parent->effective_cpus
1245 * Note that parent's subparts_cpus may have been
1246 * pre-shrunk in case there is a change in the cpu list.
1247 * So no deletion is needed.
1249 adding = cpumask_and(tmp->addmask, cpuset->cpus_allowed,
1250 parent->effective_cpus);
1251 part_error = cpumask_equal(tmp->addmask,
1252 parent->effective_cpus);
1255 if (cmd == partcmd_update) {
1256 int prev_prs = cpuset->partition_root_state;
1259 * Check for possible transition between PRS_ENABLED
1262 switch (cpuset->partition_root_state) {
1265 new_prs = PRS_ERROR;
1269 new_prs = PRS_ENABLED;
1273 * Set part_error if previously in invalid state.
1275 part_error = (prev_prs == PRS_ERROR);
1278 if (!part_error && (new_prs == PRS_ERROR))
1279 return 0; /* Nothing need to be done */
1281 if (new_prs == PRS_ERROR) {
1283 * Remove all its cpus from parent's subparts_cpus.
1286 deleting = cpumask_and(tmp->delmask, cpuset->cpus_allowed,
1287 parent->subparts_cpus);
1290 if (!adding && !deleting && (new_prs == old_prs))
1294 * Change the parent's subparts_cpus.
1295 * Newly added CPUs will be removed from effective_cpus and
1296 * newly deleted ones will be added back to effective_cpus.
1298 spin_lock_irq(&callback_lock);
1300 cpumask_or(parent->subparts_cpus,
1301 parent->subparts_cpus, tmp->addmask);
1302 cpumask_andnot(parent->effective_cpus,
1303 parent->effective_cpus, tmp->addmask);
1306 cpumask_andnot(parent->subparts_cpus,
1307 parent->subparts_cpus, tmp->delmask);
1309 * Some of the CPUs in subparts_cpus might have been offlined.
1311 cpumask_and(tmp->delmask, tmp->delmask, cpu_active_mask);
1312 cpumask_or(parent->effective_cpus,
1313 parent->effective_cpus, tmp->delmask);
1316 parent->nr_subparts_cpus = cpumask_weight(parent->subparts_cpus);
1318 if (old_prs != new_prs)
1319 cpuset->partition_root_state = new_prs;
1321 spin_unlock_irq(&callback_lock);
1322 notify_partition_change(cpuset, old_prs, new_prs);
1324 return cmd == partcmd_update;
1328 * update_cpumasks_hier - Update effective cpumasks and tasks in the subtree
1329 * @cs: the cpuset to consider
1330 * @tmp: temp variables for calculating effective_cpus & partition setup
1332 * When configured cpumask is changed, the effective cpumasks of this cpuset
1333 * and all its descendants need to be updated.
1335 * On legacy hierarchy, effective_cpus will be the same with cpu_allowed.
1337 * Called with cpuset_mutex held
1339 static void update_cpumasks_hier(struct cpuset *cs, struct tmpmasks *tmp)
1342 struct cgroup_subsys_state *pos_css;
1343 bool need_rebuild_sched_domains = false;
1344 int old_prs, new_prs;
1347 cpuset_for_each_descendant_pre(cp, pos_css, cs) {
1348 struct cpuset *parent = parent_cs(cp);
1350 compute_effective_cpumask(tmp->new_cpus, cp, parent);
1353 * If it becomes empty, inherit the effective mask of the
1354 * parent, which is guaranteed to have some CPUs.
1356 if (is_in_v2_mode() && cpumask_empty(tmp->new_cpus)) {
1357 cpumask_copy(tmp->new_cpus, parent->effective_cpus);
1358 if (!cp->use_parent_ecpus) {
1359 cp->use_parent_ecpus = true;
1360 parent->child_ecpus_count++;
1362 } else if (cp->use_parent_ecpus) {
1363 cp->use_parent_ecpus = false;
1364 WARN_ON_ONCE(!parent->child_ecpus_count);
1365 parent->child_ecpus_count--;
1369 * Skip the whole subtree if the cpumask remains the same
1370 * and has no partition root state.
1372 if (!cp->partition_root_state &&
1373 cpumask_equal(tmp->new_cpus, cp->effective_cpus)) {
1374 pos_css = css_rightmost_descendant(pos_css);
1379 * update_parent_subparts_cpumask() should have been called
1380 * for cs already in update_cpumask(). We should also call
1381 * update_tasks_cpumask() again for tasks in the parent
1382 * cpuset if the parent's subparts_cpus changes.
1384 old_prs = new_prs = cp->partition_root_state;
1385 if ((cp != cs) && old_prs) {
1386 switch (parent->partition_root_state) {
1389 * If parent is not a partition root or an
1390 * invalid partition root, clear its state
1391 * and its CS_CPU_EXCLUSIVE flag.
1393 WARN_ON_ONCE(cp->partition_root_state
1395 new_prs = PRS_DISABLED;
1398 * clear_bit() is an atomic operation and
1399 * readers aren't interested in the state
1400 * of CS_CPU_EXCLUSIVE anyway. So we can
1401 * just update the flag without holding
1402 * the callback_lock.
1404 clear_bit(CS_CPU_EXCLUSIVE, &cp->flags);
1408 if (update_parent_subparts_cpumask(cp, partcmd_update, NULL, tmp))
1409 update_tasks_cpumask(parent);
1414 * When parent is invalid, it has to be too.
1416 new_prs = PRS_ERROR;
1421 if (!css_tryget_online(&cp->css))
1425 spin_lock_irq(&callback_lock);
1427 cpumask_copy(cp->effective_cpus, tmp->new_cpus);
1428 if (cp->nr_subparts_cpus && (new_prs != PRS_ENABLED)) {
1429 cp->nr_subparts_cpus = 0;
1430 cpumask_clear(cp->subparts_cpus);
1431 } else if (cp->nr_subparts_cpus) {
1433 * Make sure that effective_cpus & subparts_cpus
1434 * are mutually exclusive.
1436 * In the unlikely event that effective_cpus
1437 * becomes empty. we clear cp->nr_subparts_cpus and
1438 * let its child partition roots to compete for
1441 cpumask_andnot(cp->effective_cpus, cp->effective_cpus,
1443 if (cpumask_empty(cp->effective_cpus)) {
1444 cpumask_copy(cp->effective_cpus, tmp->new_cpus);
1445 cpumask_clear(cp->subparts_cpus);
1446 cp->nr_subparts_cpus = 0;
1447 } else if (!cpumask_subset(cp->subparts_cpus,
1449 cpumask_andnot(cp->subparts_cpus,
1450 cp->subparts_cpus, tmp->new_cpus);
1451 cp->nr_subparts_cpus
1452 = cpumask_weight(cp->subparts_cpus);
1456 if (new_prs != old_prs)
1457 cp->partition_root_state = new_prs;
1459 spin_unlock_irq(&callback_lock);
1460 notify_partition_change(cp, old_prs, new_prs);
1462 WARN_ON(!is_in_v2_mode() &&
1463 !cpumask_equal(cp->cpus_allowed, cp->effective_cpus));
1465 update_tasks_cpumask(cp);
1468 * On legacy hierarchy, if the effective cpumask of any non-
1469 * empty cpuset is changed, we need to rebuild sched domains.
1470 * On default hierarchy, the cpuset needs to be a partition
1473 if (!cpumask_empty(cp->cpus_allowed) &&
1474 is_sched_load_balance(cp) &&
1475 (!cgroup_subsys_on_dfl(cpuset_cgrp_subsys) ||
1476 is_partition_root(cp)))
1477 need_rebuild_sched_domains = true;
1484 if (need_rebuild_sched_domains)
1485 rebuild_sched_domains_locked();
1489 * update_sibling_cpumasks - Update siblings cpumasks
1490 * @parent: Parent cpuset
1491 * @cs: Current cpuset
1492 * @tmp: Temp variables
1494 static void update_sibling_cpumasks(struct cpuset *parent, struct cpuset *cs,
1495 struct tmpmasks *tmp)
1497 struct cpuset *sibling;
1498 struct cgroup_subsys_state *pos_css;
1501 * Check all its siblings and call update_cpumasks_hier()
1502 * if their use_parent_ecpus flag is set in order for them
1503 * to use the right effective_cpus value.
1506 cpuset_for_each_child(sibling, pos_css, parent) {
1509 if (!sibling->use_parent_ecpus)
1512 update_cpumasks_hier(sibling, tmp);
1518 * update_cpumask - update the cpus_allowed mask of a cpuset and all tasks in it
1519 * @cs: the cpuset to consider
1520 * @trialcs: trial cpuset
1521 * @buf: buffer of cpu numbers written to this cpuset
1523 static int update_cpumask(struct cpuset *cs, struct cpuset *trialcs,
1527 struct tmpmasks tmp;
1529 /* top_cpuset.cpus_allowed tracks cpu_online_mask; it's read-only */
1530 if (cs == &top_cpuset)
1534 * An empty cpus_allowed is ok only if the cpuset has no tasks.
1535 * Since cpulist_parse() fails on an empty mask, we special case
1536 * that parsing. The validate_change() call ensures that cpusets
1537 * with tasks have cpus.
1540 cpumask_clear(trialcs->cpus_allowed);
1542 retval = cpulist_parse(buf, trialcs->cpus_allowed);
1546 if (!cpumask_subset(trialcs->cpus_allowed,
1547 top_cpuset.cpus_allowed))
1551 /* Nothing to do if the cpus didn't change */
1552 if (cpumask_equal(cs->cpus_allowed, trialcs->cpus_allowed))
1555 retval = validate_change(cs, trialcs);
1559 #ifdef CONFIG_CPUMASK_OFFSTACK
1561 * Use the cpumasks in trialcs for tmpmasks when they are pointers
1562 * to allocated cpumasks.
1564 tmp.addmask = trialcs->subparts_cpus;
1565 tmp.delmask = trialcs->effective_cpus;
1566 tmp.new_cpus = trialcs->cpus_allowed;
1569 if (cs->partition_root_state) {
1570 /* Cpumask of a partition root cannot be empty */
1571 if (cpumask_empty(trialcs->cpus_allowed))
1573 if (update_parent_subparts_cpumask(cs, partcmd_update,
1574 trialcs->cpus_allowed, &tmp) < 0)
1578 spin_lock_irq(&callback_lock);
1579 cpumask_copy(cs->cpus_allowed, trialcs->cpus_allowed);
1582 * Make sure that subparts_cpus is a subset of cpus_allowed.
1584 if (cs->nr_subparts_cpus) {
1585 cpumask_andnot(cs->subparts_cpus, cs->subparts_cpus,
1587 cs->nr_subparts_cpus = cpumask_weight(cs->subparts_cpus);
1589 spin_unlock_irq(&callback_lock);
1591 update_cpumasks_hier(cs, &tmp);
1593 if (cs->partition_root_state) {
1594 struct cpuset *parent = parent_cs(cs);
1597 * For partition root, update the cpumasks of sibling
1598 * cpusets if they use parent's effective_cpus.
1600 if (parent->child_ecpus_count)
1601 update_sibling_cpumasks(parent, cs, &tmp);
1607 * Migrate memory region from one set of nodes to another. This is
1608 * performed asynchronously as it can be called from process migration path
1609 * holding locks involved in process management. All mm migrations are
1610 * performed in the queued order and can be waited for by flushing
1611 * cpuset_migrate_mm_wq.
1614 struct cpuset_migrate_mm_work {
1615 struct work_struct work;
1616 struct mm_struct *mm;
1621 static void cpuset_migrate_mm_workfn(struct work_struct *work)
1623 struct cpuset_migrate_mm_work *mwork =
1624 container_of(work, struct cpuset_migrate_mm_work, work);
1626 /* on a wq worker, no need to worry about %current's mems_allowed */
1627 do_migrate_pages(mwork->mm, &mwork->from, &mwork->to, MPOL_MF_MOVE_ALL);
1632 static void cpuset_migrate_mm(struct mm_struct *mm, const nodemask_t *from,
1633 const nodemask_t *to)
1635 struct cpuset_migrate_mm_work *mwork;
1637 if (nodes_equal(*from, *to)) {
1642 mwork = kzalloc(sizeof(*mwork), GFP_KERNEL);
1645 mwork->from = *from;
1647 INIT_WORK(&mwork->work, cpuset_migrate_mm_workfn);
1648 queue_work(cpuset_migrate_mm_wq, &mwork->work);
1654 static void cpuset_post_attach(void)
1656 flush_workqueue(cpuset_migrate_mm_wq);
1660 * cpuset_change_task_nodemask - change task's mems_allowed and mempolicy
1661 * @tsk: the task to change
1662 * @newmems: new nodes that the task will be set
1664 * We use the mems_allowed_seq seqlock to safely update both tsk->mems_allowed
1665 * and rebind an eventual tasks' mempolicy. If the task is allocating in
1666 * parallel, it might temporarily see an empty intersection, which results in
1667 * a seqlock check and retry before OOM or allocation failure.
1669 static void cpuset_change_task_nodemask(struct task_struct *tsk,
1670 nodemask_t *newmems)
1674 local_irq_disable();
1675 write_seqcount_begin(&tsk->mems_allowed_seq);
1677 nodes_or(tsk->mems_allowed, tsk->mems_allowed, *newmems);
1678 mpol_rebind_task(tsk, newmems);
1679 tsk->mems_allowed = *newmems;
1681 write_seqcount_end(&tsk->mems_allowed_seq);
1687 static void *cpuset_being_rebound;
1690 * update_tasks_nodemask - Update the nodemasks of tasks in the cpuset.
1691 * @cs: the cpuset in which each task's mems_allowed mask needs to be changed
1693 * Iterate through each task of @cs updating its mems_allowed to the
1694 * effective cpuset's. As this function is called with cpuset_mutex held,
1695 * cpuset membership stays stable.
1697 static void update_tasks_nodemask(struct cpuset *cs)
1699 static nodemask_t newmems; /* protected by cpuset_mutex */
1700 struct css_task_iter it;
1701 struct task_struct *task;
1703 cpuset_being_rebound = cs; /* causes mpol_dup() rebind */
1705 guarantee_online_mems(cs, &newmems);
1708 * The mpol_rebind_mm() call takes mmap_lock, which we couldn't
1709 * take while holding tasklist_lock. Forks can happen - the
1710 * mpol_dup() cpuset_being_rebound check will catch such forks,
1711 * and rebind their vma mempolicies too. Because we still hold
1712 * the global cpuset_mutex, we know that no other rebind effort
1713 * will be contending for the global variable cpuset_being_rebound.
1714 * It's ok if we rebind the same mm twice; mpol_rebind_mm()
1715 * is idempotent. Also migrate pages in each mm to new nodes.
1717 css_task_iter_start(&cs->css, 0, &it);
1718 while ((task = css_task_iter_next(&it))) {
1719 struct mm_struct *mm;
1722 cpuset_change_task_nodemask(task, &newmems);
1724 mm = get_task_mm(task);
1728 migrate = is_memory_migrate(cs);
1730 mpol_rebind_mm(mm, &cs->mems_allowed);
1732 cpuset_migrate_mm(mm, &cs->old_mems_allowed, &newmems);
1736 css_task_iter_end(&it);
1739 * All the tasks' nodemasks have been updated, update
1740 * cs->old_mems_allowed.
1742 cs->old_mems_allowed = newmems;
1744 /* We're done rebinding vmas to this cpuset's new mems_allowed. */
1745 cpuset_being_rebound = NULL;
1749 * update_nodemasks_hier - Update effective nodemasks and tasks in the subtree
1750 * @cs: the cpuset to consider
1751 * @new_mems: a temp variable for calculating new effective_mems
1753 * When configured nodemask is changed, the effective nodemasks of this cpuset
1754 * and all its descendants need to be updated.
1756 * On legacy hierarchy, effective_mems will be the same with mems_allowed.
1758 * Called with cpuset_mutex held
1760 static void update_nodemasks_hier(struct cpuset *cs, nodemask_t *new_mems)
1763 struct cgroup_subsys_state *pos_css;
1766 cpuset_for_each_descendant_pre(cp, pos_css, cs) {
1767 struct cpuset *parent = parent_cs(cp);
1769 nodes_and(*new_mems, cp->mems_allowed, parent->effective_mems);
1772 * If it becomes empty, inherit the effective mask of the
1773 * parent, which is guaranteed to have some MEMs.
1775 if (is_in_v2_mode() && nodes_empty(*new_mems))
1776 *new_mems = parent->effective_mems;
1778 /* Skip the whole subtree if the nodemask remains the same. */
1779 if (nodes_equal(*new_mems, cp->effective_mems)) {
1780 pos_css = css_rightmost_descendant(pos_css);
1784 if (!css_tryget_online(&cp->css))
1788 spin_lock_irq(&callback_lock);
1789 cp->effective_mems = *new_mems;
1790 spin_unlock_irq(&callback_lock);
1792 WARN_ON(!is_in_v2_mode() &&
1793 !nodes_equal(cp->mems_allowed, cp->effective_mems));
1795 update_tasks_nodemask(cp);
1804 * Handle user request to change the 'mems' memory placement
1805 * of a cpuset. Needs to validate the request, update the
1806 * cpusets mems_allowed, and for each task in the cpuset,
1807 * update mems_allowed and rebind task's mempolicy and any vma
1808 * mempolicies and if the cpuset is marked 'memory_migrate',
1809 * migrate the tasks pages to the new memory.
1811 * Call with cpuset_mutex held. May take callback_lock during call.
1812 * Will take tasklist_lock, scan tasklist for tasks in cpuset cs,
1813 * lock each such tasks mm->mmap_lock, scan its vma's and rebind
1814 * their mempolicies to the cpusets new mems_allowed.
1816 static int update_nodemask(struct cpuset *cs, struct cpuset *trialcs,
1822 * top_cpuset.mems_allowed tracks node_stats[N_MEMORY];
1825 if (cs == &top_cpuset) {
1831 * An empty mems_allowed is ok iff there are no tasks in the cpuset.
1832 * Since nodelist_parse() fails on an empty mask, we special case
1833 * that parsing. The validate_change() call ensures that cpusets
1834 * with tasks have memory.
1837 nodes_clear(trialcs->mems_allowed);
1839 retval = nodelist_parse(buf, trialcs->mems_allowed);
1843 if (!nodes_subset(trialcs->mems_allowed,
1844 top_cpuset.mems_allowed)) {
1850 if (nodes_equal(cs->mems_allowed, trialcs->mems_allowed)) {
1851 retval = 0; /* Too easy - nothing to do */
1854 retval = validate_change(cs, trialcs);
1858 spin_lock_irq(&callback_lock);
1859 cs->mems_allowed = trialcs->mems_allowed;
1860 spin_unlock_irq(&callback_lock);
1862 /* use trialcs->mems_allowed as a temp variable */
1863 update_nodemasks_hier(cs, &trialcs->mems_allowed);
1868 bool current_cpuset_is_being_rebound(void)
1873 ret = task_cs(current) == cpuset_being_rebound;
1879 static int update_relax_domain_level(struct cpuset *cs, s64 val)
1882 if (val < -1 || val >= sched_domain_level_max)
1886 if (val != cs->relax_domain_level) {
1887 cs->relax_domain_level = val;
1888 if (!cpumask_empty(cs->cpus_allowed) &&
1889 is_sched_load_balance(cs))
1890 rebuild_sched_domains_locked();
1897 * update_tasks_flags - update the spread flags of tasks in the cpuset.
1898 * @cs: the cpuset in which each task's spread flags needs to be changed
1900 * Iterate through each task of @cs updating its spread flags. As this
1901 * function is called with cpuset_mutex held, cpuset membership stays
1904 static void update_tasks_flags(struct cpuset *cs)
1906 struct css_task_iter it;
1907 struct task_struct *task;
1909 css_task_iter_start(&cs->css, 0, &it);
1910 while ((task = css_task_iter_next(&it)))
1911 cpuset_update_task_spread_flag(cs, task);
1912 css_task_iter_end(&it);
1916 * update_flag - read a 0 or a 1 in a file and update associated flag
1917 * bit: the bit to update (see cpuset_flagbits_t)
1918 * cs: the cpuset to update
1919 * turning_on: whether the flag is being set or cleared
1921 * Call with cpuset_mutex held.
1924 static int update_flag(cpuset_flagbits_t bit, struct cpuset *cs,
1927 struct cpuset *trialcs;
1928 int balance_flag_changed;
1929 int spread_flag_changed;
1932 trialcs = alloc_trial_cpuset(cs);
1937 set_bit(bit, &trialcs->flags);
1939 clear_bit(bit, &trialcs->flags);
1941 err = validate_change(cs, trialcs);
1945 balance_flag_changed = (is_sched_load_balance(cs) !=
1946 is_sched_load_balance(trialcs));
1948 spread_flag_changed = ((is_spread_slab(cs) != is_spread_slab(trialcs))
1949 || (is_spread_page(cs) != is_spread_page(trialcs)));
1951 spin_lock_irq(&callback_lock);
1952 cs->flags = trialcs->flags;
1953 spin_unlock_irq(&callback_lock);
1955 if (!cpumask_empty(trialcs->cpus_allowed) && balance_flag_changed)
1956 rebuild_sched_domains_locked();
1958 if (spread_flag_changed)
1959 update_tasks_flags(cs);
1961 free_cpuset(trialcs);
1966 * update_prstate - update partititon_root_state
1967 * cs: the cpuset to update
1968 * new_prs: new partition root state
1970 * Call with cpuset_mutex held.
1972 static int update_prstate(struct cpuset *cs, int new_prs)
1974 int err, old_prs = cs->partition_root_state;
1975 struct cpuset *parent = parent_cs(cs);
1976 struct tmpmasks tmpmask;
1978 if (old_prs == new_prs)
1982 * Cannot force a partial or invalid partition root to a full
1985 if (new_prs && (old_prs == PRS_ERROR))
1988 if (alloc_cpumasks(NULL, &tmpmask))
1994 * Turning on partition root requires setting the
1995 * CS_CPU_EXCLUSIVE bit implicitly as well and cpus_allowed
1998 if (cpumask_empty(cs->cpus_allowed))
2001 err = update_flag(CS_CPU_EXCLUSIVE, cs, 1);
2005 err = update_parent_subparts_cpumask(cs, partcmd_enable,
2008 update_flag(CS_CPU_EXCLUSIVE, cs, 0);
2013 * Turning off partition root will clear the
2014 * CS_CPU_EXCLUSIVE bit.
2016 if (old_prs == PRS_ERROR) {
2017 update_flag(CS_CPU_EXCLUSIVE, cs, 0);
2022 err = update_parent_subparts_cpumask(cs, partcmd_disable,
2027 /* Turning off CS_CPU_EXCLUSIVE will not return error */
2028 update_flag(CS_CPU_EXCLUSIVE, cs, 0);
2032 * Update cpumask of parent's tasks except when it is the top
2033 * cpuset as some system daemons cannot be mapped to other CPUs.
2035 if (parent != &top_cpuset)
2036 update_tasks_cpumask(parent);
2038 if (parent->child_ecpus_count)
2039 update_sibling_cpumasks(parent, cs, &tmpmask);
2041 rebuild_sched_domains_locked();
2044 spin_lock_irq(&callback_lock);
2045 cs->partition_root_state = new_prs;
2046 spin_unlock_irq(&callback_lock);
2047 notify_partition_change(cs, old_prs, new_prs);
2050 free_cpumasks(NULL, &tmpmask);
2055 * Frequency meter - How fast is some event occurring?
2057 * These routines manage a digitally filtered, constant time based,
2058 * event frequency meter. There are four routines:
2059 * fmeter_init() - initialize a frequency meter.
2060 * fmeter_markevent() - called each time the event happens.
2061 * fmeter_getrate() - returns the recent rate of such events.
2062 * fmeter_update() - internal routine used to update fmeter.
2064 * A common data structure is passed to each of these routines,
2065 * which is used to keep track of the state required to manage the
2066 * frequency meter and its digital filter.
2068 * The filter works on the number of events marked per unit time.
2069 * The filter is single-pole low-pass recursive (IIR). The time unit
2070 * is 1 second. Arithmetic is done using 32-bit integers scaled to
2071 * simulate 3 decimal digits of precision (multiplied by 1000).
2073 * With an FM_COEF of 933, and a time base of 1 second, the filter
2074 * has a half-life of 10 seconds, meaning that if the events quit
2075 * happening, then the rate returned from the fmeter_getrate()
2076 * will be cut in half each 10 seconds, until it converges to zero.
2078 * It is not worth doing a real infinitely recursive filter. If more
2079 * than FM_MAXTICKS ticks have elapsed since the last filter event,
2080 * just compute FM_MAXTICKS ticks worth, by which point the level
2083 * Limit the count of unprocessed events to FM_MAXCNT, so as to avoid
2084 * arithmetic overflow in the fmeter_update() routine.
2086 * Given the simple 32 bit integer arithmetic used, this meter works
2087 * best for reporting rates between one per millisecond (msec) and
2088 * one per 32 (approx) seconds. At constant rates faster than one
2089 * per msec it maxes out at values just under 1,000,000. At constant
2090 * rates between one per msec, and one per second it will stabilize
2091 * to a value N*1000, where N is the rate of events per second.
2092 * At constant rates between one per second and one per 32 seconds,
2093 * it will be choppy, moving up on the seconds that have an event,
2094 * and then decaying until the next event. At rates slower than
2095 * about one in 32 seconds, it decays all the way back to zero between
2099 #define FM_COEF 933 /* coefficient for half-life of 10 secs */
2100 #define FM_MAXTICKS ((u32)99) /* useless computing more ticks than this */
2101 #define FM_MAXCNT 1000000 /* limit cnt to avoid overflow */
2102 #define FM_SCALE 1000 /* faux fixed point scale */
2104 /* Initialize a frequency meter */
2105 static void fmeter_init(struct fmeter *fmp)
2110 spin_lock_init(&fmp->lock);
2113 /* Internal meter update - process cnt events and update value */
2114 static void fmeter_update(struct fmeter *fmp)
2119 now = ktime_get_seconds();
2120 ticks = now - fmp->time;
2125 ticks = min(FM_MAXTICKS, ticks);
2127 fmp->val = (FM_COEF * fmp->val) / FM_SCALE;
2130 fmp->val += ((FM_SCALE - FM_COEF) * fmp->cnt) / FM_SCALE;
2134 /* Process any previous ticks, then bump cnt by one (times scale). */
2135 static void fmeter_markevent(struct fmeter *fmp)
2137 spin_lock(&fmp->lock);
2139 fmp->cnt = min(FM_MAXCNT, fmp->cnt + FM_SCALE);
2140 spin_unlock(&fmp->lock);
2143 /* Process any previous ticks, then return current value. */
2144 static int fmeter_getrate(struct fmeter *fmp)
2148 spin_lock(&fmp->lock);
2151 spin_unlock(&fmp->lock);
2155 static struct cpuset *cpuset_attach_old_cs;
2157 /* Called by cgroups to determine if a cpuset is usable; cpuset_mutex held */
2158 static int cpuset_can_attach(struct cgroup_taskset *tset)
2160 struct cgroup_subsys_state *css;
2162 struct task_struct *task;
2165 /* used later by cpuset_attach() */
2166 cpuset_attach_old_cs = task_cs(cgroup_taskset_first(tset, &css));
2169 percpu_down_write(&cpuset_rwsem);
2171 /* allow moving tasks into an empty cpuset if on default hierarchy */
2173 if (!is_in_v2_mode() &&
2174 (cpumask_empty(cs->cpus_allowed) || nodes_empty(cs->mems_allowed)))
2177 cgroup_taskset_for_each(task, css, tset) {
2178 ret = task_can_attach(task, cs->cpus_allowed);
2181 ret = security_task_setscheduler(task);
2187 * Mark attach is in progress. This makes validate_change() fail
2188 * changes which zero cpus/mems_allowed.
2190 cs->attach_in_progress++;
2193 percpu_up_write(&cpuset_rwsem);
2197 static void cpuset_cancel_attach(struct cgroup_taskset *tset)
2199 struct cgroup_subsys_state *css;
2201 cgroup_taskset_first(tset, &css);
2203 percpu_down_write(&cpuset_rwsem);
2204 css_cs(css)->attach_in_progress--;
2205 percpu_up_write(&cpuset_rwsem);
2209 * Protected by cpuset_mutex. cpus_attach is used only by cpuset_attach()
2210 * but we can't allocate it dynamically there. Define it global and
2211 * allocate from cpuset_init().
2213 static cpumask_var_t cpus_attach;
2215 static void cpuset_attach(struct cgroup_taskset *tset)
2217 /* static buf protected by cpuset_mutex */
2218 static nodemask_t cpuset_attach_nodemask_to;
2219 struct task_struct *task;
2220 struct task_struct *leader;
2221 struct cgroup_subsys_state *css;
2223 struct cpuset *oldcs = cpuset_attach_old_cs;
2225 cgroup_taskset_first(tset, &css);
2228 percpu_down_write(&cpuset_rwsem);
2230 /* prepare for attach */
2231 if (cs == &top_cpuset)
2232 cpumask_copy(cpus_attach, cpu_possible_mask);
2234 guarantee_online_cpus(cs, cpus_attach);
2236 guarantee_online_mems(cs, &cpuset_attach_nodemask_to);
2238 cgroup_taskset_for_each(task, css, tset) {
2240 * can_attach beforehand should guarantee that this doesn't
2241 * fail. TODO: have a better way to handle failure here
2243 WARN_ON_ONCE(set_cpus_allowed_ptr(task, cpus_attach));
2245 cpuset_change_task_nodemask(task, &cpuset_attach_nodemask_to);
2246 cpuset_update_task_spread_flag(cs, task);
2250 * Change mm for all threadgroup leaders. This is expensive and may
2251 * sleep and should be moved outside migration path proper.
2253 cpuset_attach_nodemask_to = cs->effective_mems;
2254 cgroup_taskset_for_each_leader(leader, css, tset) {
2255 struct mm_struct *mm = get_task_mm(leader);
2258 mpol_rebind_mm(mm, &cpuset_attach_nodemask_to);
2261 * old_mems_allowed is the same with mems_allowed
2262 * here, except if this task is being moved
2263 * automatically due to hotplug. In that case
2264 * @mems_allowed has been updated and is empty, so
2265 * @old_mems_allowed is the right nodesets that we
2268 if (is_memory_migrate(cs))
2269 cpuset_migrate_mm(mm, &oldcs->old_mems_allowed,
2270 &cpuset_attach_nodemask_to);
2276 cs->old_mems_allowed = cpuset_attach_nodemask_to;
2278 cs->attach_in_progress--;
2279 if (!cs->attach_in_progress)
2280 wake_up(&cpuset_attach_wq);
2282 percpu_up_write(&cpuset_rwsem);
2285 /* The various types of files and directories in a cpuset file system */
2288 FILE_MEMORY_MIGRATE,
2291 FILE_EFFECTIVE_CPULIST,
2292 FILE_EFFECTIVE_MEMLIST,
2293 FILE_SUBPARTS_CPULIST,
2297 FILE_SCHED_LOAD_BALANCE,
2298 FILE_PARTITION_ROOT,
2299 FILE_SCHED_RELAX_DOMAIN_LEVEL,
2300 FILE_MEMORY_PRESSURE_ENABLED,
2301 FILE_MEMORY_PRESSURE,
2304 } cpuset_filetype_t;
2306 static int cpuset_write_u64(struct cgroup_subsys_state *css, struct cftype *cft,
2309 struct cpuset *cs = css_cs(css);
2310 cpuset_filetype_t type = cft->private;
2314 percpu_down_write(&cpuset_rwsem);
2315 if (!is_cpuset_online(cs)) {
2321 case FILE_CPU_EXCLUSIVE:
2322 retval = update_flag(CS_CPU_EXCLUSIVE, cs, val);
2324 case FILE_MEM_EXCLUSIVE:
2325 retval = update_flag(CS_MEM_EXCLUSIVE, cs, val);
2327 case FILE_MEM_HARDWALL:
2328 retval = update_flag(CS_MEM_HARDWALL, cs, val);
2330 case FILE_SCHED_LOAD_BALANCE:
2331 retval = update_flag(CS_SCHED_LOAD_BALANCE, cs, val);
2333 case FILE_MEMORY_MIGRATE:
2334 retval = update_flag(CS_MEMORY_MIGRATE, cs, val);
2336 case FILE_MEMORY_PRESSURE_ENABLED:
2337 cpuset_memory_pressure_enabled = !!val;
2339 case FILE_SPREAD_PAGE:
2340 retval = update_flag(CS_SPREAD_PAGE, cs, val);
2342 case FILE_SPREAD_SLAB:
2343 retval = update_flag(CS_SPREAD_SLAB, cs, val);
2350 percpu_up_write(&cpuset_rwsem);
2355 static int cpuset_write_s64(struct cgroup_subsys_state *css, struct cftype *cft,
2358 struct cpuset *cs = css_cs(css);
2359 cpuset_filetype_t type = cft->private;
2360 int retval = -ENODEV;
2363 percpu_down_write(&cpuset_rwsem);
2364 if (!is_cpuset_online(cs))
2368 case FILE_SCHED_RELAX_DOMAIN_LEVEL:
2369 retval = update_relax_domain_level(cs, val);
2376 percpu_up_write(&cpuset_rwsem);
2382 * Common handling for a write to a "cpus" or "mems" file.
2384 static ssize_t cpuset_write_resmask(struct kernfs_open_file *of,
2385 char *buf, size_t nbytes, loff_t off)
2387 struct cpuset *cs = css_cs(of_css(of));
2388 struct cpuset *trialcs;
2389 int retval = -ENODEV;
2391 buf = strstrip(buf);
2394 * CPU or memory hotunplug may leave @cs w/o any execution
2395 * resources, in which case the hotplug code asynchronously updates
2396 * configuration and transfers all tasks to the nearest ancestor
2397 * which can execute.
2399 * As writes to "cpus" or "mems" may restore @cs's execution
2400 * resources, wait for the previously scheduled operations before
2401 * proceeding, so that we don't end up keep removing tasks added
2402 * after execution capability is restored.
2404 * cpuset_hotplug_work calls back into cgroup core via
2405 * cgroup_transfer_tasks() and waiting for it from a cgroupfs
2406 * operation like this one can lead to a deadlock through kernfs
2407 * active_ref protection. Let's break the protection. Losing the
2408 * protection is okay as we check whether @cs is online after
2409 * grabbing cpuset_mutex anyway. This only happens on the legacy
2413 kernfs_break_active_protection(of->kn);
2414 flush_work(&cpuset_hotplug_work);
2417 percpu_down_write(&cpuset_rwsem);
2418 if (!is_cpuset_online(cs))
2421 trialcs = alloc_trial_cpuset(cs);
2427 switch (of_cft(of)->private) {
2429 retval = update_cpumask(cs, trialcs, buf);
2432 retval = update_nodemask(cs, trialcs, buf);
2439 free_cpuset(trialcs);
2441 percpu_up_write(&cpuset_rwsem);
2443 kernfs_unbreak_active_protection(of->kn);
2445 flush_workqueue(cpuset_migrate_mm_wq);
2446 return retval ?: nbytes;
2450 * These ascii lists should be read in a single call, by using a user
2451 * buffer large enough to hold the entire map. If read in smaller
2452 * chunks, there is no guarantee of atomicity. Since the display format
2453 * used, list of ranges of sequential numbers, is variable length,
2454 * and since these maps can change value dynamically, one could read
2455 * gibberish by doing partial reads while a list was changing.
2457 static int cpuset_common_seq_show(struct seq_file *sf, void *v)
2459 struct cpuset *cs = css_cs(seq_css(sf));
2460 cpuset_filetype_t type = seq_cft(sf)->private;
2463 spin_lock_irq(&callback_lock);
2467 seq_printf(sf, "%*pbl\n", cpumask_pr_args(cs->cpus_allowed));
2470 seq_printf(sf, "%*pbl\n", nodemask_pr_args(&cs->mems_allowed));
2472 case FILE_EFFECTIVE_CPULIST:
2473 seq_printf(sf, "%*pbl\n", cpumask_pr_args(cs->effective_cpus));
2475 case FILE_EFFECTIVE_MEMLIST:
2476 seq_printf(sf, "%*pbl\n", nodemask_pr_args(&cs->effective_mems));
2478 case FILE_SUBPARTS_CPULIST:
2479 seq_printf(sf, "%*pbl\n", cpumask_pr_args(cs->subparts_cpus));
2485 spin_unlock_irq(&callback_lock);
2489 static u64 cpuset_read_u64(struct cgroup_subsys_state *css, struct cftype *cft)
2491 struct cpuset *cs = css_cs(css);
2492 cpuset_filetype_t type = cft->private;
2494 case FILE_CPU_EXCLUSIVE:
2495 return is_cpu_exclusive(cs);
2496 case FILE_MEM_EXCLUSIVE:
2497 return is_mem_exclusive(cs);
2498 case FILE_MEM_HARDWALL:
2499 return is_mem_hardwall(cs);
2500 case FILE_SCHED_LOAD_BALANCE:
2501 return is_sched_load_balance(cs);
2502 case FILE_MEMORY_MIGRATE:
2503 return is_memory_migrate(cs);
2504 case FILE_MEMORY_PRESSURE_ENABLED:
2505 return cpuset_memory_pressure_enabled;
2506 case FILE_MEMORY_PRESSURE:
2507 return fmeter_getrate(&cs->fmeter);
2508 case FILE_SPREAD_PAGE:
2509 return is_spread_page(cs);
2510 case FILE_SPREAD_SLAB:
2511 return is_spread_slab(cs);
2516 /* Unreachable but makes gcc happy */
2520 static s64 cpuset_read_s64(struct cgroup_subsys_state *css, struct cftype *cft)
2522 struct cpuset *cs = css_cs(css);
2523 cpuset_filetype_t type = cft->private;
2525 case FILE_SCHED_RELAX_DOMAIN_LEVEL:
2526 return cs->relax_domain_level;
2531 /* Unreachable but makes gcc happy */
2535 static int sched_partition_show(struct seq_file *seq, void *v)
2537 struct cpuset *cs = css_cs(seq_css(seq));
2539 switch (cs->partition_root_state) {
2541 seq_puts(seq, "root\n");
2544 seq_puts(seq, "member\n");
2547 seq_puts(seq, "root invalid\n");
2553 static ssize_t sched_partition_write(struct kernfs_open_file *of, char *buf,
2554 size_t nbytes, loff_t off)
2556 struct cpuset *cs = css_cs(of_css(of));
2558 int retval = -ENODEV;
2560 buf = strstrip(buf);
2563 * Convert "root" to ENABLED, and convert "member" to DISABLED.
2565 if (!strcmp(buf, "root"))
2567 else if (!strcmp(buf, "member"))
2574 percpu_down_write(&cpuset_rwsem);
2575 if (!is_cpuset_online(cs))
2578 retval = update_prstate(cs, val);
2580 percpu_up_write(&cpuset_rwsem);
2583 return retval ?: nbytes;
2587 * for the common functions, 'private' gives the type of file
2590 static struct cftype legacy_files[] = {
2593 .seq_show = cpuset_common_seq_show,
2594 .write = cpuset_write_resmask,
2595 .max_write_len = (100U + 6 * NR_CPUS),
2596 .private = FILE_CPULIST,
2601 .seq_show = cpuset_common_seq_show,
2602 .write = cpuset_write_resmask,
2603 .max_write_len = (100U + 6 * MAX_NUMNODES),
2604 .private = FILE_MEMLIST,
2608 .name = "effective_cpus",
2609 .seq_show = cpuset_common_seq_show,
2610 .private = FILE_EFFECTIVE_CPULIST,
2614 .name = "effective_mems",
2615 .seq_show = cpuset_common_seq_show,
2616 .private = FILE_EFFECTIVE_MEMLIST,
2620 .name = "cpu_exclusive",
2621 .read_u64 = cpuset_read_u64,
2622 .write_u64 = cpuset_write_u64,
2623 .private = FILE_CPU_EXCLUSIVE,
2627 .name = "mem_exclusive",
2628 .read_u64 = cpuset_read_u64,
2629 .write_u64 = cpuset_write_u64,
2630 .private = FILE_MEM_EXCLUSIVE,
2634 .name = "mem_hardwall",
2635 .read_u64 = cpuset_read_u64,
2636 .write_u64 = cpuset_write_u64,
2637 .private = FILE_MEM_HARDWALL,
2641 .name = "sched_load_balance",
2642 .read_u64 = cpuset_read_u64,
2643 .write_u64 = cpuset_write_u64,
2644 .private = FILE_SCHED_LOAD_BALANCE,
2648 .name = "sched_relax_domain_level",
2649 .read_s64 = cpuset_read_s64,
2650 .write_s64 = cpuset_write_s64,
2651 .private = FILE_SCHED_RELAX_DOMAIN_LEVEL,
2655 .name = "memory_migrate",
2656 .read_u64 = cpuset_read_u64,
2657 .write_u64 = cpuset_write_u64,
2658 .private = FILE_MEMORY_MIGRATE,
2662 .name = "memory_pressure",
2663 .read_u64 = cpuset_read_u64,
2664 .private = FILE_MEMORY_PRESSURE,
2668 .name = "memory_spread_page",
2669 .read_u64 = cpuset_read_u64,
2670 .write_u64 = cpuset_write_u64,
2671 .private = FILE_SPREAD_PAGE,
2675 .name = "memory_spread_slab",
2676 .read_u64 = cpuset_read_u64,
2677 .write_u64 = cpuset_write_u64,
2678 .private = FILE_SPREAD_SLAB,
2682 .name = "memory_pressure_enabled",
2683 .flags = CFTYPE_ONLY_ON_ROOT,
2684 .read_u64 = cpuset_read_u64,
2685 .write_u64 = cpuset_write_u64,
2686 .private = FILE_MEMORY_PRESSURE_ENABLED,
2693 * This is currently a minimal set for the default hierarchy. It can be
2694 * expanded later on by migrating more features and control files from v1.
2696 static struct cftype dfl_files[] = {
2699 .seq_show = cpuset_common_seq_show,
2700 .write = cpuset_write_resmask,
2701 .max_write_len = (100U + 6 * NR_CPUS),
2702 .private = FILE_CPULIST,
2703 .flags = CFTYPE_NOT_ON_ROOT,
2708 .seq_show = cpuset_common_seq_show,
2709 .write = cpuset_write_resmask,
2710 .max_write_len = (100U + 6 * MAX_NUMNODES),
2711 .private = FILE_MEMLIST,
2712 .flags = CFTYPE_NOT_ON_ROOT,
2716 .name = "cpus.effective",
2717 .seq_show = cpuset_common_seq_show,
2718 .private = FILE_EFFECTIVE_CPULIST,
2722 .name = "mems.effective",
2723 .seq_show = cpuset_common_seq_show,
2724 .private = FILE_EFFECTIVE_MEMLIST,
2728 .name = "cpus.partition",
2729 .seq_show = sched_partition_show,
2730 .write = sched_partition_write,
2731 .private = FILE_PARTITION_ROOT,
2732 .flags = CFTYPE_NOT_ON_ROOT,
2733 .file_offset = offsetof(struct cpuset, partition_file),
2737 .name = "cpus.subpartitions",
2738 .seq_show = cpuset_common_seq_show,
2739 .private = FILE_SUBPARTS_CPULIST,
2740 .flags = CFTYPE_DEBUG,
2748 * cpuset_css_alloc - allocate a cpuset css
2749 * cgrp: control group that the new cpuset will be part of
2752 static struct cgroup_subsys_state *
2753 cpuset_css_alloc(struct cgroup_subsys_state *parent_css)
2758 return &top_cpuset.css;
2760 cs = kzalloc(sizeof(*cs), GFP_KERNEL);
2762 return ERR_PTR(-ENOMEM);
2764 if (alloc_cpumasks(cs, NULL)) {
2766 return ERR_PTR(-ENOMEM);
2769 __set_bit(CS_SCHED_LOAD_BALANCE, &cs->flags);
2770 nodes_clear(cs->mems_allowed);
2771 nodes_clear(cs->effective_mems);
2772 fmeter_init(&cs->fmeter);
2773 cs->relax_domain_level = -1;
2775 /* Set CS_MEMORY_MIGRATE for default hierarchy */
2776 if (cgroup_subsys_on_dfl(cpuset_cgrp_subsys))
2777 __set_bit(CS_MEMORY_MIGRATE, &cs->flags);
2782 static int cpuset_css_online(struct cgroup_subsys_state *css)
2784 struct cpuset *cs = css_cs(css);
2785 struct cpuset *parent = parent_cs(cs);
2786 struct cpuset *tmp_cs;
2787 struct cgroup_subsys_state *pos_css;
2793 percpu_down_write(&cpuset_rwsem);
2795 set_bit(CS_ONLINE, &cs->flags);
2796 if (is_spread_page(parent))
2797 set_bit(CS_SPREAD_PAGE, &cs->flags);
2798 if (is_spread_slab(parent))
2799 set_bit(CS_SPREAD_SLAB, &cs->flags);
2803 spin_lock_irq(&callback_lock);
2804 if (is_in_v2_mode()) {
2805 cpumask_copy(cs->effective_cpus, parent->effective_cpus);
2806 cs->effective_mems = parent->effective_mems;
2807 cs->use_parent_ecpus = true;
2808 parent->child_ecpus_count++;
2810 spin_unlock_irq(&callback_lock);
2812 if (!test_bit(CGRP_CPUSET_CLONE_CHILDREN, &css->cgroup->flags))
2816 * Clone @parent's configuration if CGRP_CPUSET_CLONE_CHILDREN is
2817 * set. This flag handling is implemented in cgroup core for
2818 * histrical reasons - the flag may be specified during mount.
2820 * Currently, if any sibling cpusets have exclusive cpus or mem, we
2821 * refuse to clone the configuration - thereby refusing the task to
2822 * be entered, and as a result refusing the sys_unshare() or
2823 * clone() which initiated it. If this becomes a problem for some
2824 * users who wish to allow that scenario, then this could be
2825 * changed to grant parent->cpus_allowed-sibling_cpus_exclusive
2826 * (and likewise for mems) to the new cgroup.
2829 cpuset_for_each_child(tmp_cs, pos_css, parent) {
2830 if (is_mem_exclusive(tmp_cs) || is_cpu_exclusive(tmp_cs)) {
2837 spin_lock_irq(&callback_lock);
2838 cs->mems_allowed = parent->mems_allowed;
2839 cs->effective_mems = parent->mems_allowed;
2840 cpumask_copy(cs->cpus_allowed, parent->cpus_allowed);
2841 cpumask_copy(cs->effective_cpus, parent->cpus_allowed);
2842 spin_unlock_irq(&callback_lock);
2844 percpu_up_write(&cpuset_rwsem);
2850 * If the cpuset being removed has its flag 'sched_load_balance'
2851 * enabled, then simulate turning sched_load_balance off, which
2852 * will call rebuild_sched_domains_locked(). That is not needed
2853 * in the default hierarchy where only changes in partition
2854 * will cause repartitioning.
2856 * If the cpuset has the 'sched.partition' flag enabled, simulate
2857 * turning 'sched.partition" off.
2860 static void cpuset_css_offline(struct cgroup_subsys_state *css)
2862 struct cpuset *cs = css_cs(css);
2865 percpu_down_write(&cpuset_rwsem);
2867 if (is_partition_root(cs))
2868 update_prstate(cs, 0);
2870 if (!cgroup_subsys_on_dfl(cpuset_cgrp_subsys) &&
2871 is_sched_load_balance(cs))
2872 update_flag(CS_SCHED_LOAD_BALANCE, cs, 0);
2874 if (cs->use_parent_ecpus) {
2875 struct cpuset *parent = parent_cs(cs);
2877 cs->use_parent_ecpus = false;
2878 parent->child_ecpus_count--;
2882 clear_bit(CS_ONLINE, &cs->flags);
2884 percpu_up_write(&cpuset_rwsem);
2888 static void cpuset_css_free(struct cgroup_subsys_state *css)
2890 struct cpuset *cs = css_cs(css);
2895 static void cpuset_bind(struct cgroup_subsys_state *root_css)
2897 percpu_down_write(&cpuset_rwsem);
2898 spin_lock_irq(&callback_lock);
2900 if (is_in_v2_mode()) {
2901 cpumask_copy(top_cpuset.cpus_allowed, cpu_possible_mask);
2902 top_cpuset.mems_allowed = node_possible_map;
2904 cpumask_copy(top_cpuset.cpus_allowed,
2905 top_cpuset.effective_cpus);
2906 top_cpuset.mems_allowed = top_cpuset.effective_mems;
2909 spin_unlock_irq(&callback_lock);
2910 percpu_up_write(&cpuset_rwsem);
2914 * Make sure the new task conform to the current state of its parent,
2915 * which could have been changed by cpuset just after it inherits the
2916 * state from the parent and before it sits on the cgroup's task list.
2918 static void cpuset_fork(struct task_struct *task)
2920 if (task_css_is_root(task, cpuset_cgrp_id))
2923 set_cpus_allowed_ptr(task, current->cpus_ptr);
2924 task->mems_allowed = current->mems_allowed;
2927 struct cgroup_subsys cpuset_cgrp_subsys = {
2928 .css_alloc = cpuset_css_alloc,
2929 .css_online = cpuset_css_online,
2930 .css_offline = cpuset_css_offline,
2931 .css_free = cpuset_css_free,
2932 .can_attach = cpuset_can_attach,
2933 .cancel_attach = cpuset_cancel_attach,
2934 .attach = cpuset_attach,
2935 .post_attach = cpuset_post_attach,
2936 .bind = cpuset_bind,
2937 .fork = cpuset_fork,
2938 .legacy_cftypes = legacy_files,
2939 .dfl_cftypes = dfl_files,
2945 * cpuset_init - initialize cpusets at system boot
2947 * Description: Initialize top_cpuset
2950 int __init cpuset_init(void)
2952 BUG_ON(percpu_init_rwsem(&cpuset_rwsem));
2954 BUG_ON(!alloc_cpumask_var(&top_cpuset.cpus_allowed, GFP_KERNEL));
2955 BUG_ON(!alloc_cpumask_var(&top_cpuset.effective_cpus, GFP_KERNEL));
2956 BUG_ON(!zalloc_cpumask_var(&top_cpuset.subparts_cpus, GFP_KERNEL));
2958 cpumask_setall(top_cpuset.cpus_allowed);
2959 nodes_setall(top_cpuset.mems_allowed);
2960 cpumask_setall(top_cpuset.effective_cpus);
2961 nodes_setall(top_cpuset.effective_mems);
2963 fmeter_init(&top_cpuset.fmeter);
2964 set_bit(CS_SCHED_LOAD_BALANCE, &top_cpuset.flags);
2965 top_cpuset.relax_domain_level = -1;
2967 BUG_ON(!alloc_cpumask_var(&cpus_attach, GFP_KERNEL));
2973 * If CPU and/or memory hotplug handlers, below, unplug any CPUs
2974 * or memory nodes, we need to walk over the cpuset hierarchy,
2975 * removing that CPU or node from all cpusets. If this removes the
2976 * last CPU or node from a cpuset, then move the tasks in the empty
2977 * cpuset to its next-highest non-empty parent.
2979 static void remove_tasks_in_empty_cpuset(struct cpuset *cs)
2981 struct cpuset *parent;
2984 * Find its next-highest non-empty parent, (top cpuset
2985 * has online cpus, so can't be empty).
2987 parent = parent_cs(cs);
2988 while (cpumask_empty(parent->cpus_allowed) ||
2989 nodes_empty(parent->mems_allowed))
2990 parent = parent_cs(parent);
2992 if (cgroup_transfer_tasks(parent->css.cgroup, cs->css.cgroup)) {
2993 pr_err("cpuset: failed to transfer tasks out of empty cpuset ");
2994 pr_cont_cgroup_name(cs->css.cgroup);
3000 hotplug_update_tasks_legacy(struct cpuset *cs,
3001 struct cpumask *new_cpus, nodemask_t *new_mems,
3002 bool cpus_updated, bool mems_updated)
3006 spin_lock_irq(&callback_lock);
3007 cpumask_copy(cs->cpus_allowed, new_cpus);
3008 cpumask_copy(cs->effective_cpus, new_cpus);
3009 cs->mems_allowed = *new_mems;
3010 cs->effective_mems = *new_mems;
3011 spin_unlock_irq(&callback_lock);
3014 * Don't call update_tasks_cpumask() if the cpuset becomes empty,
3015 * as the tasks will be migratecd to an ancestor.
3017 if (cpus_updated && !cpumask_empty(cs->cpus_allowed))
3018 update_tasks_cpumask(cs);
3019 if (mems_updated && !nodes_empty(cs->mems_allowed))
3020 update_tasks_nodemask(cs);
3022 is_empty = cpumask_empty(cs->cpus_allowed) ||
3023 nodes_empty(cs->mems_allowed);
3025 percpu_up_write(&cpuset_rwsem);
3028 * Move tasks to the nearest ancestor with execution resources,
3029 * This is full cgroup operation which will also call back into
3030 * cpuset. Should be done outside any lock.
3033 remove_tasks_in_empty_cpuset(cs);
3035 percpu_down_write(&cpuset_rwsem);
3039 hotplug_update_tasks(struct cpuset *cs,
3040 struct cpumask *new_cpus, nodemask_t *new_mems,
3041 bool cpus_updated, bool mems_updated)
3043 if (cpumask_empty(new_cpus))
3044 cpumask_copy(new_cpus, parent_cs(cs)->effective_cpus);
3045 if (nodes_empty(*new_mems))
3046 *new_mems = parent_cs(cs)->effective_mems;
3048 spin_lock_irq(&callback_lock);
3049 cpumask_copy(cs->effective_cpus, new_cpus);
3050 cs->effective_mems = *new_mems;
3051 spin_unlock_irq(&callback_lock);
3054 update_tasks_cpumask(cs);
3056 update_tasks_nodemask(cs);
3059 static bool force_rebuild;
3061 void cpuset_force_rebuild(void)
3063 force_rebuild = true;
3067 * cpuset_hotplug_update_tasks - update tasks in a cpuset for hotunplug
3068 * @cs: cpuset in interest
3069 * @tmp: the tmpmasks structure pointer
3071 * Compare @cs's cpu and mem masks against top_cpuset and if some have gone
3072 * offline, update @cs accordingly. If @cs ends up with no CPU or memory,
3073 * all its tasks are moved to the nearest ancestor with both resources.
3075 static void cpuset_hotplug_update_tasks(struct cpuset *cs, struct tmpmasks *tmp)
3077 static cpumask_t new_cpus;
3078 static nodemask_t new_mems;
3081 struct cpuset *parent;
3083 wait_event(cpuset_attach_wq, cs->attach_in_progress == 0);
3085 percpu_down_write(&cpuset_rwsem);
3088 * We have raced with task attaching. We wait until attaching
3089 * is finished, so we won't attach a task to an empty cpuset.
3091 if (cs->attach_in_progress) {
3092 percpu_up_write(&cpuset_rwsem);
3096 parent = parent_cs(cs);
3097 compute_effective_cpumask(&new_cpus, cs, parent);
3098 nodes_and(new_mems, cs->mems_allowed, parent->effective_mems);
3100 if (cs->nr_subparts_cpus)
3102 * Make sure that CPUs allocated to child partitions
3103 * do not show up in effective_cpus.
3105 cpumask_andnot(&new_cpus, &new_cpus, cs->subparts_cpus);
3107 if (!tmp || !cs->partition_root_state)
3111 * In the unlikely event that a partition root has empty
3112 * effective_cpus or its parent becomes erroneous, we have to
3113 * transition it to the erroneous state.
3115 if (is_partition_root(cs) && (cpumask_empty(&new_cpus) ||
3116 (parent->partition_root_state == PRS_ERROR))) {
3117 if (cs->nr_subparts_cpus) {
3118 spin_lock_irq(&callback_lock);
3119 cs->nr_subparts_cpus = 0;
3120 cpumask_clear(cs->subparts_cpus);
3121 spin_unlock_irq(&callback_lock);
3122 compute_effective_cpumask(&new_cpus, cs, parent);
3126 * If the effective_cpus is empty because the child
3127 * partitions take away all the CPUs, we can keep
3128 * the current partition and let the child partitions
3129 * fight for available CPUs.
3131 if ((parent->partition_root_state == PRS_ERROR) ||
3132 cpumask_empty(&new_cpus)) {
3135 update_parent_subparts_cpumask(cs, partcmd_disable,
3137 old_prs = cs->partition_root_state;
3138 if (old_prs != PRS_ERROR) {
3139 spin_lock_irq(&callback_lock);
3140 cs->partition_root_state = PRS_ERROR;
3141 spin_unlock_irq(&callback_lock);
3142 notify_partition_change(cs, old_prs, PRS_ERROR);
3145 cpuset_force_rebuild();
3149 * On the other hand, an erroneous partition root may be transitioned
3150 * back to a regular one or a partition root with no CPU allocated
3151 * from the parent may change to erroneous.
3153 if (is_partition_root(parent) &&
3154 ((cs->partition_root_state == PRS_ERROR) ||
3155 !cpumask_intersects(&new_cpus, parent->subparts_cpus)) &&
3156 update_parent_subparts_cpumask(cs, partcmd_update, NULL, tmp))
3157 cpuset_force_rebuild();
3160 cpus_updated = !cpumask_equal(&new_cpus, cs->effective_cpus);
3161 mems_updated = !nodes_equal(new_mems, cs->effective_mems);
3163 if (is_in_v2_mode())
3164 hotplug_update_tasks(cs, &new_cpus, &new_mems,
3165 cpus_updated, mems_updated);
3167 hotplug_update_tasks_legacy(cs, &new_cpus, &new_mems,
3168 cpus_updated, mems_updated);
3170 percpu_up_write(&cpuset_rwsem);
3174 * cpuset_hotplug_workfn - handle CPU/memory hotunplug for a cpuset
3176 * This function is called after either CPU or memory configuration has
3177 * changed and updates cpuset accordingly. The top_cpuset is always
3178 * synchronized to cpu_active_mask and N_MEMORY, which is necessary in
3179 * order to make cpusets transparent (of no affect) on systems that are
3180 * actively using CPU hotplug but making no active use of cpusets.
3182 * Non-root cpusets are only affected by offlining. If any CPUs or memory
3183 * nodes have been taken down, cpuset_hotplug_update_tasks() is invoked on
3186 * Note that CPU offlining during suspend is ignored. We don't modify
3187 * cpusets across suspend/resume cycles at all.
3189 static void cpuset_hotplug_workfn(struct work_struct *work)
3191 static cpumask_t new_cpus;
3192 static nodemask_t new_mems;
3193 bool cpus_updated, mems_updated;
3194 bool on_dfl = is_in_v2_mode();
3195 struct tmpmasks tmp, *ptmp = NULL;
3197 if (on_dfl && !alloc_cpumasks(NULL, &tmp))
3200 percpu_down_write(&cpuset_rwsem);
3202 /* fetch the available cpus/mems and find out which changed how */
3203 cpumask_copy(&new_cpus, cpu_active_mask);
3204 new_mems = node_states[N_MEMORY];
3207 * If subparts_cpus is populated, it is likely that the check below
3208 * will produce a false positive on cpus_updated when the cpu list
3209 * isn't changed. It is extra work, but it is better to be safe.
3211 cpus_updated = !cpumask_equal(top_cpuset.effective_cpus, &new_cpus);
3212 mems_updated = !nodes_equal(top_cpuset.effective_mems, new_mems);
3215 * In the rare case that hotplug removes all the cpus in subparts_cpus,
3216 * we assumed that cpus are updated.
3218 if (!cpus_updated && top_cpuset.nr_subparts_cpus)
3219 cpus_updated = true;
3221 /* synchronize cpus_allowed to cpu_active_mask */
3223 spin_lock_irq(&callback_lock);
3225 cpumask_copy(top_cpuset.cpus_allowed, &new_cpus);
3227 * Make sure that CPUs allocated to child partitions
3228 * do not show up in effective_cpus. If no CPU is left,
3229 * we clear the subparts_cpus & let the child partitions
3230 * fight for the CPUs again.
3232 if (top_cpuset.nr_subparts_cpus) {
3233 if (cpumask_subset(&new_cpus,
3234 top_cpuset.subparts_cpus)) {
3235 top_cpuset.nr_subparts_cpus = 0;
3236 cpumask_clear(top_cpuset.subparts_cpus);
3238 cpumask_andnot(&new_cpus, &new_cpus,
3239 top_cpuset.subparts_cpus);
3242 cpumask_copy(top_cpuset.effective_cpus, &new_cpus);
3243 spin_unlock_irq(&callback_lock);
3244 /* we don't mess with cpumasks of tasks in top_cpuset */
3247 /* synchronize mems_allowed to N_MEMORY */
3249 spin_lock_irq(&callback_lock);
3251 top_cpuset.mems_allowed = new_mems;
3252 top_cpuset.effective_mems = new_mems;
3253 spin_unlock_irq(&callback_lock);
3254 update_tasks_nodemask(&top_cpuset);
3257 percpu_up_write(&cpuset_rwsem);
3259 /* if cpus or mems changed, we need to propagate to descendants */
3260 if (cpus_updated || mems_updated) {
3262 struct cgroup_subsys_state *pos_css;
3265 cpuset_for_each_descendant_pre(cs, pos_css, &top_cpuset) {
3266 if (cs == &top_cpuset || !css_tryget_online(&cs->css))
3270 cpuset_hotplug_update_tasks(cs, ptmp);
3278 /* rebuild sched domains if cpus_allowed has changed */
3279 if (cpus_updated || force_rebuild) {
3280 force_rebuild = false;
3281 rebuild_sched_domains();
3284 free_cpumasks(NULL, ptmp);
3287 void cpuset_update_active_cpus(void)
3290 * We're inside cpu hotplug critical region which usually nests
3291 * inside cgroup synchronization. Bounce actual hotplug processing
3292 * to a work item to avoid reverse locking order.
3294 schedule_work(&cpuset_hotplug_work);
3297 void cpuset_wait_for_hotplug(void)
3299 flush_work(&cpuset_hotplug_work);
3303 * Keep top_cpuset.mems_allowed tracking node_states[N_MEMORY].
3304 * Call this routine anytime after node_states[N_MEMORY] changes.
3305 * See cpuset_update_active_cpus() for CPU hotplug handling.
3307 static int cpuset_track_online_nodes(struct notifier_block *self,
3308 unsigned long action, void *arg)
3310 schedule_work(&cpuset_hotplug_work);
3314 static struct notifier_block cpuset_track_online_nodes_nb = {
3315 .notifier_call = cpuset_track_online_nodes,
3316 .priority = 10, /* ??! */
3320 * cpuset_init_smp - initialize cpus_allowed
3322 * Description: Finish top cpuset after cpu, node maps are initialized
3324 void __init cpuset_init_smp(void)
3326 cpumask_copy(top_cpuset.cpus_allowed, cpu_active_mask);
3327 top_cpuset.mems_allowed = node_states[N_MEMORY];
3328 top_cpuset.old_mems_allowed = top_cpuset.mems_allowed;
3330 cpumask_copy(top_cpuset.effective_cpus, cpu_active_mask);
3331 top_cpuset.effective_mems = node_states[N_MEMORY];
3333 register_hotmemory_notifier(&cpuset_track_online_nodes_nb);
3335 cpuset_migrate_mm_wq = alloc_ordered_workqueue("cpuset_migrate_mm", 0);
3336 BUG_ON(!cpuset_migrate_mm_wq);
3340 * cpuset_cpus_allowed - return cpus_allowed mask from a tasks cpuset.
3341 * @tsk: pointer to task_struct from which to obtain cpuset->cpus_allowed.
3342 * @pmask: pointer to struct cpumask variable to receive cpus_allowed set.
3344 * Description: Returns the cpumask_var_t cpus_allowed of the cpuset
3345 * attached to the specified @tsk. Guaranteed to return some non-empty
3346 * subset of cpu_online_mask, even if this means going outside the
3350 void cpuset_cpus_allowed(struct task_struct *tsk, struct cpumask *pmask)
3352 unsigned long flags;
3354 spin_lock_irqsave(&callback_lock, flags);
3356 guarantee_online_cpus(task_cs(tsk), pmask);
3358 spin_unlock_irqrestore(&callback_lock, flags);
3362 * cpuset_cpus_allowed_fallback - final fallback before complete catastrophe.
3363 * @tsk: pointer to task_struct with which the scheduler is struggling
3365 * Description: In the case that the scheduler cannot find an allowed cpu in
3366 * tsk->cpus_allowed, we fall back to task_cs(tsk)->cpus_allowed. In legacy
3367 * mode however, this value is the same as task_cs(tsk)->effective_cpus,
3368 * which will not contain a sane cpumask during cases such as cpu hotplugging.
3369 * This is the absolute last resort for the scheduler and it is only used if
3370 * _every_ other avenue has been traveled.
3373 void cpuset_cpus_allowed_fallback(struct task_struct *tsk)
3376 do_set_cpus_allowed(tsk, is_in_v2_mode() ?
3377 task_cs(tsk)->cpus_allowed : cpu_possible_mask);
3381 * We own tsk->cpus_allowed, nobody can change it under us.
3383 * But we used cs && cs->cpus_allowed lockless and thus can
3384 * race with cgroup_attach_task() or update_cpumask() and get
3385 * the wrong tsk->cpus_allowed. However, both cases imply the
3386 * subsequent cpuset_change_cpumask()->set_cpus_allowed_ptr()
3387 * which takes task_rq_lock().
3389 * If we are called after it dropped the lock we must see all
3390 * changes in tsk_cs()->cpus_allowed. Otherwise we can temporary
3391 * set any mask even if it is not right from task_cs() pov,
3392 * the pending set_cpus_allowed_ptr() will fix things.
3394 * select_fallback_rq() will fix things ups and set cpu_possible_mask
3399 void __init cpuset_init_current_mems_allowed(void)
3401 nodes_setall(current->mems_allowed);
3405 * cpuset_mems_allowed - return mems_allowed mask from a tasks cpuset.
3406 * @tsk: pointer to task_struct from which to obtain cpuset->mems_allowed.
3408 * Description: Returns the nodemask_t mems_allowed of the cpuset
3409 * attached to the specified @tsk. Guaranteed to return some non-empty
3410 * subset of node_states[N_MEMORY], even if this means going outside the
3414 nodemask_t cpuset_mems_allowed(struct task_struct *tsk)
3417 unsigned long flags;
3419 spin_lock_irqsave(&callback_lock, flags);
3421 guarantee_online_mems(task_cs(tsk), &mask);
3423 spin_unlock_irqrestore(&callback_lock, flags);
3429 * cpuset_nodemask_valid_mems_allowed - check nodemask vs. current mems_allowed
3430 * @nodemask: the nodemask to be checked
3432 * Are any of the nodes in the nodemask allowed in current->mems_allowed?
3434 int cpuset_nodemask_valid_mems_allowed(nodemask_t *nodemask)
3436 return nodes_intersects(*nodemask, current->mems_allowed);
3440 * nearest_hardwall_ancestor() - Returns the nearest mem_exclusive or
3441 * mem_hardwall ancestor to the specified cpuset. Call holding
3442 * callback_lock. If no ancestor is mem_exclusive or mem_hardwall
3443 * (an unusual configuration), then returns the root cpuset.
3445 static struct cpuset *nearest_hardwall_ancestor(struct cpuset *cs)
3447 while (!(is_mem_exclusive(cs) || is_mem_hardwall(cs)) && parent_cs(cs))
3453 * cpuset_node_allowed - Can we allocate on a memory node?
3454 * @node: is this an allowed node?
3455 * @gfp_mask: memory allocation flags
3457 * If we're in interrupt, yes, we can always allocate. If @node is set in
3458 * current's mems_allowed, yes. If it's not a __GFP_HARDWALL request and this
3459 * node is set in the nearest hardwalled cpuset ancestor to current's cpuset,
3460 * yes. If current has access to memory reserves as an oom victim, yes.
3463 * GFP_USER allocations are marked with the __GFP_HARDWALL bit,
3464 * and do not allow allocations outside the current tasks cpuset
3465 * unless the task has been OOM killed.
3466 * GFP_KERNEL allocations are not so marked, so can escape to the
3467 * nearest enclosing hardwalled ancestor cpuset.
3469 * Scanning up parent cpusets requires callback_lock. The
3470 * __alloc_pages() routine only calls here with __GFP_HARDWALL bit
3471 * _not_ set if it's a GFP_KERNEL allocation, and all nodes in the
3472 * current tasks mems_allowed came up empty on the first pass over
3473 * the zonelist. So only GFP_KERNEL allocations, if all nodes in the
3474 * cpuset are short of memory, might require taking the callback_lock.
3476 * The first call here from mm/page_alloc:get_page_from_freelist()
3477 * has __GFP_HARDWALL set in gfp_mask, enforcing hardwall cpusets,
3478 * so no allocation on a node outside the cpuset is allowed (unless
3479 * in interrupt, of course).
3481 * The second pass through get_page_from_freelist() doesn't even call
3482 * here for GFP_ATOMIC calls. For those calls, the __alloc_pages()
3483 * variable 'wait' is not set, and the bit ALLOC_CPUSET is not set
3484 * in alloc_flags. That logic and the checks below have the combined
3486 * in_interrupt - any node ok (current task context irrelevant)
3487 * GFP_ATOMIC - any node ok
3488 * tsk_is_oom_victim - any node ok
3489 * GFP_KERNEL - any node in enclosing hardwalled cpuset ok
3490 * GFP_USER - only nodes in current tasks mems allowed ok.
3492 bool __cpuset_node_allowed(int node, gfp_t gfp_mask)
3494 struct cpuset *cs; /* current cpuset ancestors */
3495 int allowed; /* is allocation in zone z allowed? */
3496 unsigned long flags;
3500 if (node_isset(node, current->mems_allowed))
3503 * Allow tasks that have access to memory reserves because they have
3504 * been OOM killed to get memory anywhere.
3506 if (unlikely(tsk_is_oom_victim(current)))
3508 if (gfp_mask & __GFP_HARDWALL) /* If hardwall request, stop here */
3511 if (current->flags & PF_EXITING) /* Let dying task have memory */
3514 /* Not hardwall and node outside mems_allowed: scan up cpusets */
3515 spin_lock_irqsave(&callback_lock, flags);
3518 cs = nearest_hardwall_ancestor(task_cs(current));
3519 allowed = node_isset(node, cs->mems_allowed);
3522 spin_unlock_irqrestore(&callback_lock, flags);
3527 * cpuset_mem_spread_node() - On which node to begin search for a file page
3528 * cpuset_slab_spread_node() - On which node to begin search for a slab page
3530 * If a task is marked PF_SPREAD_PAGE or PF_SPREAD_SLAB (as for
3531 * tasks in a cpuset with is_spread_page or is_spread_slab set),
3532 * and if the memory allocation used cpuset_mem_spread_node()
3533 * to determine on which node to start looking, as it will for
3534 * certain page cache or slab cache pages such as used for file
3535 * system buffers and inode caches, then instead of starting on the
3536 * local node to look for a free page, rather spread the starting
3537 * node around the tasks mems_allowed nodes.
3539 * We don't have to worry about the returned node being offline
3540 * because "it can't happen", and even if it did, it would be ok.
3542 * The routines calling guarantee_online_mems() are careful to
3543 * only set nodes in task->mems_allowed that are online. So it
3544 * should not be possible for the following code to return an
3545 * offline node. But if it did, that would be ok, as this routine
3546 * is not returning the node where the allocation must be, only
3547 * the node where the search should start. The zonelist passed to
3548 * __alloc_pages() will include all nodes. If the slab allocator
3549 * is passed an offline node, it will fall back to the local node.
3550 * See kmem_cache_alloc_node().
3553 static int cpuset_spread_node(int *rotor)
3555 return *rotor = next_node_in(*rotor, current->mems_allowed);
3558 int cpuset_mem_spread_node(void)
3560 if (current->cpuset_mem_spread_rotor == NUMA_NO_NODE)
3561 current->cpuset_mem_spread_rotor =
3562 node_random(¤t->mems_allowed);
3564 return cpuset_spread_node(¤t->cpuset_mem_spread_rotor);
3567 int cpuset_slab_spread_node(void)
3569 if (current->cpuset_slab_spread_rotor == NUMA_NO_NODE)
3570 current->cpuset_slab_spread_rotor =
3571 node_random(¤t->mems_allowed);
3573 return cpuset_spread_node(¤t->cpuset_slab_spread_rotor);
3576 EXPORT_SYMBOL_GPL(cpuset_mem_spread_node);
3579 * cpuset_mems_allowed_intersects - Does @tsk1's mems_allowed intersect @tsk2's?
3580 * @tsk1: pointer to task_struct of some task.
3581 * @tsk2: pointer to task_struct of some other task.
3583 * Description: Return true if @tsk1's mems_allowed intersects the
3584 * mems_allowed of @tsk2. Used by the OOM killer to determine if
3585 * one of the task's memory usage might impact the memory available
3589 int cpuset_mems_allowed_intersects(const struct task_struct *tsk1,
3590 const struct task_struct *tsk2)
3592 return nodes_intersects(tsk1->mems_allowed, tsk2->mems_allowed);
3596 * cpuset_print_current_mems_allowed - prints current's cpuset and mems_allowed
3598 * Description: Prints current's name, cpuset name, and cached copy of its
3599 * mems_allowed to the kernel log.
3601 void cpuset_print_current_mems_allowed(void)
3603 struct cgroup *cgrp;
3607 cgrp = task_cs(current)->css.cgroup;
3608 pr_cont(",cpuset=");
3609 pr_cont_cgroup_name(cgrp);
3610 pr_cont(",mems_allowed=%*pbl",
3611 nodemask_pr_args(¤t->mems_allowed));
3617 * Collection of memory_pressure is suppressed unless
3618 * this flag is enabled by writing "1" to the special
3619 * cpuset file 'memory_pressure_enabled' in the root cpuset.
3622 int cpuset_memory_pressure_enabled __read_mostly;
3625 * cpuset_memory_pressure_bump - keep stats of per-cpuset reclaims.
3627 * Keep a running average of the rate of synchronous (direct)
3628 * page reclaim efforts initiated by tasks in each cpuset.
3630 * This represents the rate at which some task in the cpuset
3631 * ran low on memory on all nodes it was allowed to use, and
3632 * had to enter the kernels page reclaim code in an effort to
3633 * create more free memory by tossing clean pages or swapping
3634 * or writing dirty pages.
3636 * Display to user space in the per-cpuset read-only file
3637 * "memory_pressure". Value displayed is an integer
3638 * representing the recent rate of entry into the synchronous
3639 * (direct) page reclaim by any task attached to the cpuset.
3642 void __cpuset_memory_pressure_bump(void)
3645 fmeter_markevent(&task_cs(current)->fmeter);
3649 #ifdef CONFIG_PROC_PID_CPUSET
3651 * proc_cpuset_show()
3652 * - Print tasks cpuset path into seq_file.
3653 * - Used for /proc/<pid>/cpuset.
3654 * - No need to task_lock(tsk) on this tsk->cpuset reference, as it
3655 * doesn't really matter if tsk->cpuset changes after we read it,
3656 * and we take cpuset_mutex, keeping cpuset_attach() from changing it
3659 int proc_cpuset_show(struct seq_file *m, struct pid_namespace *ns,
3660 struct pid *pid, struct task_struct *tsk)
3663 struct cgroup_subsys_state *css;
3667 buf = kmalloc(PATH_MAX, GFP_KERNEL);
3671 css = task_get_css(tsk, cpuset_cgrp_id);
3672 retval = cgroup_path_ns(css->cgroup, buf, PATH_MAX,
3673 current->nsproxy->cgroup_ns);
3675 if (retval >= PATH_MAX)
3676 retval = -ENAMETOOLONG;
3687 #endif /* CONFIG_PROC_PID_CPUSET */
3689 /* Display task mems_allowed in /proc/<pid>/status file. */
3690 void cpuset_task_status_allowed(struct seq_file *m, struct task_struct *task)
3692 seq_printf(m, "Mems_allowed:\t%*pb\n",
3693 nodemask_pr_args(&task->mems_allowed));
3694 seq_printf(m, "Mems_allowed_list:\t%*pbl\n",
3695 nodemask_pr_args(&task->mems_allowed));