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 task's cpusets's cpus_allowed that
389 * are online and are capable of running the task. If none are found,
390 * walk up the cpuset hierarchy until we find one that does have some
393 * One way or another, we guarantee to return some non-empty subset
394 * of cpu_online_mask.
396 * Call with callback_lock or cpuset_mutex held.
398 static void guarantee_online_cpus(struct task_struct *tsk,
399 struct cpumask *pmask)
401 const struct cpumask *possible_mask = task_cpu_possible_mask(tsk);
404 if (WARN_ON(!cpumask_and(pmask, possible_mask, cpu_online_mask)))
405 cpumask_copy(pmask, cpu_online_mask);
410 while (!cpumask_intersects(cs->effective_cpus, pmask)) {
414 * The top cpuset doesn't have any online cpu as a
415 * consequence of a race between cpuset_hotplug_work
416 * and cpu hotplug notifier. But we know the top
417 * cpuset's effective_cpus is on its way to be
418 * identical to cpu_online_mask.
423 cpumask_and(pmask, pmask, cs->effective_cpus);
430 * Return in *pmask the portion of a cpusets's mems_allowed that
431 * are online, with memory. If none are online with memory, walk
432 * up the cpuset hierarchy until we find one that does have some
433 * online mems. The top cpuset always has some mems online.
435 * One way or another, we guarantee to return some non-empty subset
436 * of node_states[N_MEMORY].
438 * Call with callback_lock or cpuset_mutex held.
440 static void guarantee_online_mems(struct cpuset *cs, nodemask_t *pmask)
442 while (!nodes_intersects(cs->effective_mems, node_states[N_MEMORY]))
444 nodes_and(*pmask, cs->effective_mems, node_states[N_MEMORY]);
448 * update task's spread flag if cpuset's page/slab spread flag is set
450 * Call with callback_lock or cpuset_mutex held.
452 static void cpuset_update_task_spread_flag(struct cpuset *cs,
453 struct task_struct *tsk)
455 if (is_spread_page(cs))
456 task_set_spread_page(tsk);
458 task_clear_spread_page(tsk);
460 if (is_spread_slab(cs))
461 task_set_spread_slab(tsk);
463 task_clear_spread_slab(tsk);
467 * is_cpuset_subset(p, q) - Is cpuset p a subset of cpuset q?
469 * One cpuset is a subset of another if all its allowed CPUs and
470 * Memory Nodes are a subset of the other, and its exclusive flags
471 * are only set if the other's are set. Call holding cpuset_mutex.
474 static int is_cpuset_subset(const struct cpuset *p, const struct cpuset *q)
476 return cpumask_subset(p->cpus_allowed, q->cpus_allowed) &&
477 nodes_subset(p->mems_allowed, q->mems_allowed) &&
478 is_cpu_exclusive(p) <= is_cpu_exclusive(q) &&
479 is_mem_exclusive(p) <= is_mem_exclusive(q);
483 * alloc_cpumasks - allocate three cpumasks for cpuset
484 * @cs: the cpuset that have cpumasks to be allocated.
485 * @tmp: the tmpmasks structure pointer
486 * Return: 0 if successful, -ENOMEM otherwise.
488 * Only one of the two input arguments should be non-NULL.
490 static inline int alloc_cpumasks(struct cpuset *cs, struct tmpmasks *tmp)
492 cpumask_var_t *pmask1, *pmask2, *pmask3;
495 pmask1 = &cs->cpus_allowed;
496 pmask2 = &cs->effective_cpus;
497 pmask3 = &cs->subparts_cpus;
499 pmask1 = &tmp->new_cpus;
500 pmask2 = &tmp->addmask;
501 pmask3 = &tmp->delmask;
504 if (!zalloc_cpumask_var(pmask1, GFP_KERNEL))
507 if (!zalloc_cpumask_var(pmask2, GFP_KERNEL))
510 if (!zalloc_cpumask_var(pmask3, GFP_KERNEL))
516 free_cpumask_var(*pmask2);
518 free_cpumask_var(*pmask1);
523 * free_cpumasks - free cpumasks in a tmpmasks structure
524 * @cs: the cpuset that have cpumasks to be free.
525 * @tmp: the tmpmasks structure pointer
527 static inline void free_cpumasks(struct cpuset *cs, struct tmpmasks *tmp)
530 free_cpumask_var(cs->cpus_allowed);
531 free_cpumask_var(cs->effective_cpus);
532 free_cpumask_var(cs->subparts_cpus);
535 free_cpumask_var(tmp->new_cpus);
536 free_cpumask_var(tmp->addmask);
537 free_cpumask_var(tmp->delmask);
542 * alloc_trial_cpuset - allocate a trial cpuset
543 * @cs: the cpuset that the trial cpuset duplicates
545 static struct cpuset *alloc_trial_cpuset(struct cpuset *cs)
547 struct cpuset *trial;
549 trial = kmemdup(cs, sizeof(*cs), GFP_KERNEL);
553 if (alloc_cpumasks(trial, NULL)) {
558 cpumask_copy(trial->cpus_allowed, cs->cpus_allowed);
559 cpumask_copy(trial->effective_cpus, cs->effective_cpus);
564 * free_cpuset - free the cpuset
565 * @cs: the cpuset to be freed
567 static inline void free_cpuset(struct cpuset *cs)
569 free_cpumasks(cs, NULL);
574 * validate_change() - Used to validate that any proposed cpuset change
575 * follows the structural rules for cpusets.
577 * If we replaced the flag and mask values of the current cpuset
578 * (cur) with those values in the trial cpuset (trial), would
579 * our various subset and exclusive rules still be valid? Presumes
582 * 'cur' is the address of an actual, in-use cpuset. Operations
583 * such as list traversal that depend on the actual address of the
584 * cpuset in the list must use cur below, not trial.
586 * 'trial' is the address of bulk structure copy of cur, with
587 * perhaps one or more of the fields cpus_allowed, mems_allowed,
588 * or flags changed to new, trial values.
590 * Return 0 if valid, -errno if not.
593 static int validate_change(struct cpuset *cur, struct cpuset *trial)
595 struct cgroup_subsys_state *css;
596 struct cpuset *c, *par;
601 /* Each of our child cpusets must be a subset of us */
603 cpuset_for_each_child(c, css, cur)
604 if (!is_cpuset_subset(c, trial))
607 /* Remaining checks don't apply to root cpuset */
609 if (cur == &top_cpuset)
612 par = parent_cs(cur);
614 /* On legacy hierarchy, we must be a subset of our parent cpuset. */
616 if (!is_in_v2_mode() && !is_cpuset_subset(trial, par))
620 * If either I or some sibling (!= me) is exclusive, we can't
624 cpuset_for_each_child(c, css, par) {
625 if ((is_cpu_exclusive(trial) || is_cpu_exclusive(c)) &&
627 cpumask_intersects(trial->cpus_allowed, c->cpus_allowed))
629 if ((is_mem_exclusive(trial) || is_mem_exclusive(c)) &&
631 nodes_intersects(trial->mems_allowed, c->mems_allowed))
636 * Cpusets with tasks - existing or newly being attached - can't
637 * be changed to have empty cpus_allowed or mems_allowed.
640 if ((cgroup_is_populated(cur->css.cgroup) || cur->attach_in_progress)) {
641 if (!cpumask_empty(cur->cpus_allowed) &&
642 cpumask_empty(trial->cpus_allowed))
644 if (!nodes_empty(cur->mems_allowed) &&
645 nodes_empty(trial->mems_allowed))
650 * We can't shrink if we won't have enough room for SCHED_DEADLINE
654 if (is_cpu_exclusive(cur) &&
655 !cpuset_cpumask_can_shrink(cur->cpus_allowed,
656 trial->cpus_allowed))
667 * Helper routine for generate_sched_domains().
668 * Do cpusets a, b have overlapping effective cpus_allowed masks?
670 static int cpusets_overlap(struct cpuset *a, struct cpuset *b)
672 return cpumask_intersects(a->effective_cpus, b->effective_cpus);
676 update_domain_attr(struct sched_domain_attr *dattr, struct cpuset *c)
678 if (dattr->relax_domain_level < c->relax_domain_level)
679 dattr->relax_domain_level = c->relax_domain_level;
683 static void update_domain_attr_tree(struct sched_domain_attr *dattr,
684 struct cpuset *root_cs)
687 struct cgroup_subsys_state *pos_css;
690 cpuset_for_each_descendant_pre(cp, pos_css, root_cs) {
691 /* skip the whole subtree if @cp doesn't have any CPU */
692 if (cpumask_empty(cp->cpus_allowed)) {
693 pos_css = css_rightmost_descendant(pos_css);
697 if (is_sched_load_balance(cp))
698 update_domain_attr(dattr, cp);
703 /* Must be called with cpuset_mutex held. */
704 static inline int nr_cpusets(void)
706 /* jump label reference count + the top-level cpuset */
707 return static_key_count(&cpusets_enabled_key.key) + 1;
711 * generate_sched_domains()
713 * This function builds a partial partition of the systems CPUs
714 * A 'partial partition' is a set of non-overlapping subsets whose
715 * union is a subset of that set.
716 * The output of this function needs to be passed to kernel/sched/core.c
717 * partition_sched_domains() routine, which will rebuild the scheduler's
718 * load balancing domains (sched domains) as specified by that partial
721 * See "What is sched_load_balance" in Documentation/admin-guide/cgroup-v1/cpusets.rst
722 * for a background explanation of this.
724 * Does not return errors, on the theory that the callers of this
725 * routine would rather not worry about failures to rebuild sched
726 * domains when operating in the severe memory shortage situations
727 * that could cause allocation failures below.
729 * Must be called with cpuset_mutex held.
731 * The three key local variables below are:
732 * cp - cpuset pointer, used (together with pos_css) to perform a
733 * top-down scan of all cpusets. For our purposes, rebuilding
734 * the schedulers sched domains, we can ignore !is_sched_load_
736 * csa - (for CpuSet Array) Array of pointers to all the cpusets
737 * that need to be load balanced, for convenient iterative
738 * access by the subsequent code that finds the best partition,
739 * i.e the set of domains (subsets) of CPUs such that the
740 * cpus_allowed of every cpuset marked is_sched_load_balance
741 * is a subset of one of these domains, while there are as
742 * many such domains as possible, each as small as possible.
743 * doms - Conversion of 'csa' to an array of cpumasks, for passing to
744 * the kernel/sched/core.c routine partition_sched_domains() in a
745 * convenient format, that can be easily compared to the prior
746 * value to determine what partition elements (sched domains)
747 * were changed (added or removed.)
749 * Finding the best partition (set of domains):
750 * The triple nested loops below over i, j, k scan over the
751 * load balanced cpusets (using the array of cpuset pointers in
752 * csa[]) looking for pairs of cpusets that have overlapping
753 * cpus_allowed, but which don't have the same 'pn' partition
754 * number and gives them in the same partition number. It keeps
755 * looping on the 'restart' label until it can no longer find
758 * The union of the cpus_allowed masks from the set of
759 * all cpusets having the same 'pn' value then form the one
760 * element of the partition (one sched domain) to be passed to
761 * partition_sched_domains().
763 static int generate_sched_domains(cpumask_var_t **domains,
764 struct sched_domain_attr **attributes)
766 struct cpuset *cp; /* top-down scan of cpusets */
767 struct cpuset **csa; /* array of all cpuset ptrs */
768 int csn; /* how many cpuset ptrs in csa so far */
769 int i, j, k; /* indices for partition finding loops */
770 cpumask_var_t *doms; /* resulting partition; i.e. sched domains */
771 struct sched_domain_attr *dattr; /* attributes for custom domains */
772 int ndoms = 0; /* number of sched domains in result */
773 int nslot; /* next empty doms[] struct cpumask slot */
774 struct cgroup_subsys_state *pos_css;
775 bool root_load_balance = is_sched_load_balance(&top_cpuset);
781 /* Special case for the 99% of systems with one, full, sched domain */
782 if (root_load_balance && !top_cpuset.nr_subparts_cpus) {
784 doms = alloc_sched_domains(ndoms);
788 dattr = kmalloc(sizeof(struct sched_domain_attr), GFP_KERNEL);
790 *dattr = SD_ATTR_INIT;
791 update_domain_attr_tree(dattr, &top_cpuset);
793 cpumask_and(doms[0], top_cpuset.effective_cpus,
794 housekeeping_cpumask(HK_FLAG_DOMAIN));
799 csa = kmalloc_array(nr_cpusets(), sizeof(cp), GFP_KERNEL);
805 if (root_load_balance)
806 csa[csn++] = &top_cpuset;
807 cpuset_for_each_descendant_pre(cp, pos_css, &top_cpuset) {
808 if (cp == &top_cpuset)
811 * Continue traversing beyond @cp iff @cp has some CPUs and
812 * isn't load balancing. The former is obvious. The
813 * latter: All child cpusets contain a subset of the
814 * parent's cpus, so just skip them, and then we call
815 * update_domain_attr_tree() to calc relax_domain_level of
816 * the corresponding sched domain.
818 * If root is load-balancing, we can skip @cp if it
819 * is a subset of the root's effective_cpus.
821 if (!cpumask_empty(cp->cpus_allowed) &&
822 !(is_sched_load_balance(cp) &&
823 cpumask_intersects(cp->cpus_allowed,
824 housekeeping_cpumask(HK_FLAG_DOMAIN))))
827 if (root_load_balance &&
828 cpumask_subset(cp->cpus_allowed, top_cpuset.effective_cpus))
831 if (is_sched_load_balance(cp) &&
832 !cpumask_empty(cp->effective_cpus))
835 /* skip @cp's subtree if not a partition root */
836 if (!is_partition_root(cp))
837 pos_css = css_rightmost_descendant(pos_css);
841 for (i = 0; i < csn; i++)
846 /* Find the best partition (set of sched domains) */
847 for (i = 0; i < csn; i++) {
848 struct cpuset *a = csa[i];
851 for (j = 0; j < csn; j++) {
852 struct cpuset *b = csa[j];
855 if (apn != bpn && cpusets_overlap(a, b)) {
856 for (k = 0; k < csn; k++) {
857 struct cpuset *c = csa[k];
862 ndoms--; /* one less element */
869 * Now we know how many domains to create.
870 * Convert <csn, csa> to <ndoms, doms> and populate cpu masks.
872 doms = alloc_sched_domains(ndoms);
877 * The rest of the code, including the scheduler, can deal with
878 * dattr==NULL case. No need to abort if alloc fails.
880 dattr = kmalloc_array(ndoms, sizeof(struct sched_domain_attr),
883 for (nslot = 0, i = 0; i < csn; i++) {
884 struct cpuset *a = csa[i];
889 /* Skip completed partitions */
895 if (nslot == ndoms) {
896 static int warnings = 10;
898 pr_warn("rebuild_sched_domains confused: nslot %d, ndoms %d, csn %d, i %d, apn %d\n",
899 nslot, ndoms, csn, i, apn);
907 *(dattr + nslot) = SD_ATTR_INIT;
908 for (j = i; j < csn; j++) {
909 struct cpuset *b = csa[j];
912 cpumask_or(dp, dp, b->effective_cpus);
913 cpumask_and(dp, dp, housekeeping_cpumask(HK_FLAG_DOMAIN));
915 update_domain_attr_tree(dattr + nslot, b);
917 /* Done with this partition */
923 BUG_ON(nslot != ndoms);
929 * Fallback to the default domain if kmalloc() failed.
930 * See comments in partition_sched_domains().
940 static void update_tasks_root_domain(struct cpuset *cs)
942 struct css_task_iter it;
943 struct task_struct *task;
945 css_task_iter_start(&cs->css, 0, &it);
947 while ((task = css_task_iter_next(&it)))
948 dl_add_task_root_domain(task);
950 css_task_iter_end(&it);
953 static void rebuild_root_domains(void)
955 struct cpuset *cs = NULL;
956 struct cgroup_subsys_state *pos_css;
958 percpu_rwsem_assert_held(&cpuset_rwsem);
959 lockdep_assert_cpus_held();
960 lockdep_assert_held(&sched_domains_mutex);
965 * Clear default root domain DL accounting, it will be computed again
966 * if a task belongs to it.
968 dl_clear_root_domain(&def_root_domain);
970 cpuset_for_each_descendant_pre(cs, pos_css, &top_cpuset) {
972 if (cpumask_empty(cs->effective_cpus)) {
973 pos_css = css_rightmost_descendant(pos_css);
981 update_tasks_root_domain(cs);
990 partition_and_rebuild_sched_domains(int ndoms_new, cpumask_var_t doms_new[],
991 struct sched_domain_attr *dattr_new)
993 mutex_lock(&sched_domains_mutex);
994 partition_sched_domains_locked(ndoms_new, doms_new, dattr_new);
995 rebuild_root_domains();
996 mutex_unlock(&sched_domains_mutex);
1000 * Rebuild scheduler domains.
1002 * If the flag 'sched_load_balance' of any cpuset with non-empty
1003 * 'cpus' changes, or if the 'cpus' allowed changes in any cpuset
1004 * which has that flag enabled, or if any cpuset with a non-empty
1005 * 'cpus' is removed, then call this routine to rebuild the
1006 * scheduler's dynamic sched domains.
1008 * Call with cpuset_mutex held. Takes cpus_read_lock().
1010 static void rebuild_sched_domains_locked(void)
1012 struct cgroup_subsys_state *pos_css;
1013 struct sched_domain_attr *attr;
1014 cpumask_var_t *doms;
1018 lockdep_assert_cpus_held();
1019 percpu_rwsem_assert_held(&cpuset_rwsem);
1022 * If we have raced with CPU hotplug, return early to avoid
1023 * passing doms with offlined cpu to partition_sched_domains().
1024 * Anyways, cpuset_hotplug_workfn() will rebuild sched domains.
1026 * With no CPUs in any subpartitions, top_cpuset's effective CPUs
1027 * should be the same as the active CPUs, so checking only top_cpuset
1028 * is enough to detect racing CPU offlines.
1030 if (!top_cpuset.nr_subparts_cpus &&
1031 !cpumask_equal(top_cpuset.effective_cpus, cpu_active_mask))
1035 * With subpartition CPUs, however, the effective CPUs of a partition
1036 * root should be only a subset of the active CPUs. Since a CPU in any
1037 * partition root could be offlined, all must be checked.
1039 if (top_cpuset.nr_subparts_cpus) {
1041 cpuset_for_each_descendant_pre(cs, pos_css, &top_cpuset) {
1042 if (!is_partition_root(cs)) {
1043 pos_css = css_rightmost_descendant(pos_css);
1046 if (!cpumask_subset(cs->effective_cpus,
1055 /* Generate domain masks and attrs */
1056 ndoms = generate_sched_domains(&doms, &attr);
1058 /* Have scheduler rebuild the domains */
1059 partition_and_rebuild_sched_domains(ndoms, doms, attr);
1061 #else /* !CONFIG_SMP */
1062 static void rebuild_sched_domains_locked(void)
1065 #endif /* CONFIG_SMP */
1067 void rebuild_sched_domains(void)
1070 percpu_down_write(&cpuset_rwsem);
1071 rebuild_sched_domains_locked();
1072 percpu_up_write(&cpuset_rwsem);
1077 * update_tasks_cpumask - Update the cpumasks of tasks in the cpuset.
1078 * @cs: the cpuset in which each task's cpus_allowed mask needs to be changed
1080 * Iterate through each task of @cs updating its cpus_allowed to the
1081 * effective cpuset's. As this function is called with cpuset_mutex held,
1082 * cpuset membership stays stable.
1084 static void update_tasks_cpumask(struct cpuset *cs)
1086 struct css_task_iter it;
1087 struct task_struct *task;
1089 css_task_iter_start(&cs->css, 0, &it);
1090 while ((task = css_task_iter_next(&it)))
1091 set_cpus_allowed_ptr(task, cs->effective_cpus);
1092 css_task_iter_end(&it);
1096 * compute_effective_cpumask - Compute the effective cpumask of the cpuset
1097 * @new_cpus: the temp variable for the new effective_cpus mask
1098 * @cs: the cpuset the need to recompute the new effective_cpus mask
1099 * @parent: the parent cpuset
1101 * If the parent has subpartition CPUs, include them in the list of
1102 * allowable CPUs in computing the new effective_cpus mask. Since offlined
1103 * CPUs are not removed from subparts_cpus, we have to use cpu_active_mask
1104 * to mask those out.
1106 static void compute_effective_cpumask(struct cpumask *new_cpus,
1107 struct cpuset *cs, struct cpuset *parent)
1109 if (parent->nr_subparts_cpus) {
1110 cpumask_or(new_cpus, parent->effective_cpus,
1111 parent->subparts_cpus);
1112 cpumask_and(new_cpus, new_cpus, cs->cpus_allowed);
1113 cpumask_and(new_cpus, new_cpus, cpu_active_mask);
1115 cpumask_and(new_cpus, cs->cpus_allowed, parent->effective_cpus);
1120 * Commands for update_parent_subparts_cpumask
1123 partcmd_enable, /* Enable partition root */
1124 partcmd_disable, /* Disable partition root */
1125 partcmd_update, /* Update parent's subparts_cpus */
1129 * update_parent_subparts_cpumask - update subparts_cpus mask of parent cpuset
1130 * @cpuset: The cpuset that requests change in partition root state
1131 * @cmd: Partition root state change command
1132 * @newmask: Optional new cpumask for partcmd_update
1133 * @tmp: Temporary addmask and delmask
1134 * Return: 0, 1 or an error code
1136 * For partcmd_enable, the cpuset is being transformed from a non-partition
1137 * root to a partition root. The cpus_allowed mask of the given cpuset will
1138 * be put into parent's subparts_cpus and taken away from parent's
1139 * effective_cpus. The function will return 0 if all the CPUs listed in
1140 * cpus_allowed can be granted or an error code will be returned.
1142 * For partcmd_disable, the cpuset is being transofrmed from a partition
1143 * root back to a non-partition root. Any CPUs in cpus_allowed that are in
1144 * parent's subparts_cpus will be taken away from that cpumask and put back
1145 * into parent's effective_cpus. 0 should always be returned.
1147 * For partcmd_update, if the optional newmask is specified, the cpu
1148 * list is to be changed from cpus_allowed to newmask. Otherwise,
1149 * cpus_allowed is assumed to remain the same. The cpuset should either
1150 * be a partition root or an invalid partition root. The partition root
1151 * state may change if newmask is NULL and none of the requested CPUs can
1152 * be granted by the parent. The function will return 1 if changes to
1153 * parent's subparts_cpus and effective_cpus happen or 0 otherwise.
1154 * Error code should only be returned when newmask is non-NULL.
1156 * The partcmd_enable and partcmd_disable commands are used by
1157 * update_prstate(). The partcmd_update command is used by
1158 * update_cpumasks_hier() with newmask NULL and update_cpumask() with
1161 * The checking is more strict when enabling partition root than the
1162 * other two commands.
1164 * Because of the implicit cpu exclusive nature of a partition root,
1165 * cpumask changes that violates the cpu exclusivity rule will not be
1166 * permitted when checked by validate_change(). The validate_change()
1167 * function will also prevent any changes to the cpu list if it is not
1168 * a superset of children's cpu lists.
1170 static int update_parent_subparts_cpumask(struct cpuset *cpuset, int cmd,
1171 struct cpumask *newmask,
1172 struct tmpmasks *tmp)
1174 struct cpuset *parent = parent_cs(cpuset);
1175 int adding; /* Moving cpus from effective_cpus to subparts_cpus */
1176 int deleting; /* Moving cpus from subparts_cpus to effective_cpus */
1177 int old_prs, new_prs;
1178 bool part_error = false; /* Partition error? */
1180 percpu_rwsem_assert_held(&cpuset_rwsem);
1183 * The parent must be a partition root.
1184 * The new cpumask, if present, or the current cpus_allowed must
1187 if (!is_partition_root(parent) ||
1188 (newmask && cpumask_empty(newmask)) ||
1189 (!newmask && cpumask_empty(cpuset->cpus_allowed)))
1193 * Enabling/disabling partition root is not allowed if there are
1196 if ((cmd != partcmd_update) && css_has_online_children(&cpuset->css))
1200 * Enabling partition root is not allowed if not all the CPUs
1201 * can be granted from parent's effective_cpus or at least one
1202 * CPU will be left after that.
1204 if ((cmd == partcmd_enable) &&
1205 (!cpumask_subset(cpuset->cpus_allowed, parent->effective_cpus) ||
1206 cpumask_equal(cpuset->cpus_allowed, parent->effective_cpus)))
1210 * A cpumask update cannot make parent's effective_cpus become empty.
1212 adding = deleting = false;
1213 old_prs = new_prs = cpuset->partition_root_state;
1214 if (cmd == partcmd_enable) {
1215 cpumask_copy(tmp->addmask, cpuset->cpus_allowed);
1217 } else if (cmd == partcmd_disable) {
1218 deleting = cpumask_and(tmp->delmask, cpuset->cpus_allowed,
1219 parent->subparts_cpus);
1220 } else if (newmask) {
1222 * partcmd_update with newmask:
1224 * delmask = cpus_allowed & ~newmask & parent->subparts_cpus
1225 * addmask = newmask & parent->effective_cpus
1226 * & ~parent->subparts_cpus
1228 cpumask_andnot(tmp->delmask, cpuset->cpus_allowed, newmask);
1229 deleting = cpumask_and(tmp->delmask, tmp->delmask,
1230 parent->subparts_cpus);
1232 cpumask_and(tmp->addmask, newmask, parent->effective_cpus);
1233 adding = cpumask_andnot(tmp->addmask, tmp->addmask,
1234 parent->subparts_cpus);
1236 * Return error if the new effective_cpus could become empty.
1239 cpumask_equal(parent->effective_cpus, tmp->addmask)) {
1243 * As some of the CPUs in subparts_cpus might have
1244 * been offlined, we need to compute the real delmask
1247 if (!cpumask_and(tmp->addmask, tmp->delmask,
1250 cpumask_copy(tmp->addmask, parent->effective_cpus);
1254 * partcmd_update w/o newmask:
1256 * addmask = cpus_allowed & parent->effective_cpus
1258 * Note that parent's subparts_cpus may have been
1259 * pre-shrunk in case there is a change in the cpu list.
1260 * So no deletion is needed.
1262 adding = cpumask_and(tmp->addmask, cpuset->cpus_allowed,
1263 parent->effective_cpus);
1264 part_error = cpumask_equal(tmp->addmask,
1265 parent->effective_cpus);
1268 if (cmd == partcmd_update) {
1269 int prev_prs = cpuset->partition_root_state;
1272 * Check for possible transition between PRS_ENABLED
1275 switch (cpuset->partition_root_state) {
1278 new_prs = PRS_ERROR;
1282 new_prs = PRS_ENABLED;
1286 * Set part_error if previously in invalid state.
1288 part_error = (prev_prs == PRS_ERROR);
1291 if (!part_error && (new_prs == PRS_ERROR))
1292 return 0; /* Nothing need to be done */
1294 if (new_prs == PRS_ERROR) {
1296 * Remove all its cpus from parent's subparts_cpus.
1299 deleting = cpumask_and(tmp->delmask, cpuset->cpus_allowed,
1300 parent->subparts_cpus);
1303 if (!adding && !deleting && (new_prs == old_prs))
1307 * Change the parent's subparts_cpus.
1308 * Newly added CPUs will be removed from effective_cpus and
1309 * newly deleted ones will be added back to effective_cpus.
1311 spin_lock_irq(&callback_lock);
1313 cpumask_or(parent->subparts_cpus,
1314 parent->subparts_cpus, tmp->addmask);
1315 cpumask_andnot(parent->effective_cpus,
1316 parent->effective_cpus, tmp->addmask);
1319 cpumask_andnot(parent->subparts_cpus,
1320 parent->subparts_cpus, tmp->delmask);
1322 * Some of the CPUs in subparts_cpus might have been offlined.
1324 cpumask_and(tmp->delmask, tmp->delmask, cpu_active_mask);
1325 cpumask_or(parent->effective_cpus,
1326 parent->effective_cpus, tmp->delmask);
1329 parent->nr_subparts_cpus = cpumask_weight(parent->subparts_cpus);
1331 if (old_prs != new_prs)
1332 cpuset->partition_root_state = new_prs;
1334 spin_unlock_irq(&callback_lock);
1335 notify_partition_change(cpuset, old_prs, new_prs);
1337 return cmd == partcmd_update;
1341 * update_cpumasks_hier - Update effective cpumasks and tasks in the subtree
1342 * @cs: the cpuset to consider
1343 * @tmp: temp variables for calculating effective_cpus & partition setup
1345 * When configured cpumask is changed, the effective cpumasks of this cpuset
1346 * and all its descendants need to be updated.
1348 * On legacy hierarchy, effective_cpus will be the same with cpu_allowed.
1350 * Called with cpuset_mutex held
1352 static void update_cpumasks_hier(struct cpuset *cs, struct tmpmasks *tmp)
1355 struct cgroup_subsys_state *pos_css;
1356 bool need_rebuild_sched_domains = false;
1357 int old_prs, new_prs;
1360 cpuset_for_each_descendant_pre(cp, pos_css, cs) {
1361 struct cpuset *parent = parent_cs(cp);
1363 compute_effective_cpumask(tmp->new_cpus, cp, parent);
1366 * If it becomes empty, inherit the effective mask of the
1367 * parent, which is guaranteed to have some CPUs.
1369 if (is_in_v2_mode() && cpumask_empty(tmp->new_cpus)) {
1370 cpumask_copy(tmp->new_cpus, parent->effective_cpus);
1371 if (!cp->use_parent_ecpus) {
1372 cp->use_parent_ecpus = true;
1373 parent->child_ecpus_count++;
1375 } else if (cp->use_parent_ecpus) {
1376 cp->use_parent_ecpus = false;
1377 WARN_ON_ONCE(!parent->child_ecpus_count);
1378 parent->child_ecpus_count--;
1382 * Skip the whole subtree if the cpumask remains the same
1383 * and has no partition root state.
1385 if (!cp->partition_root_state &&
1386 cpumask_equal(tmp->new_cpus, cp->effective_cpus)) {
1387 pos_css = css_rightmost_descendant(pos_css);
1392 * update_parent_subparts_cpumask() should have been called
1393 * for cs already in update_cpumask(). We should also call
1394 * update_tasks_cpumask() again for tasks in the parent
1395 * cpuset if the parent's subparts_cpus changes.
1397 old_prs = new_prs = cp->partition_root_state;
1398 if ((cp != cs) && old_prs) {
1399 switch (parent->partition_root_state) {
1402 * If parent is not a partition root or an
1403 * invalid partition root, clear its state
1404 * and its CS_CPU_EXCLUSIVE flag.
1406 WARN_ON_ONCE(cp->partition_root_state
1408 new_prs = PRS_DISABLED;
1411 * clear_bit() is an atomic operation and
1412 * readers aren't interested in the state
1413 * of CS_CPU_EXCLUSIVE anyway. So we can
1414 * just update the flag without holding
1415 * the callback_lock.
1417 clear_bit(CS_CPU_EXCLUSIVE, &cp->flags);
1421 if (update_parent_subparts_cpumask(cp, partcmd_update, NULL, tmp))
1422 update_tasks_cpumask(parent);
1427 * When parent is invalid, it has to be too.
1429 new_prs = PRS_ERROR;
1434 if (!css_tryget_online(&cp->css))
1438 spin_lock_irq(&callback_lock);
1440 cpumask_copy(cp->effective_cpus, tmp->new_cpus);
1441 if (cp->nr_subparts_cpus && (new_prs != PRS_ENABLED)) {
1442 cp->nr_subparts_cpus = 0;
1443 cpumask_clear(cp->subparts_cpus);
1444 } else if (cp->nr_subparts_cpus) {
1446 * Make sure that effective_cpus & subparts_cpus
1447 * are mutually exclusive.
1449 * In the unlikely event that effective_cpus
1450 * becomes empty. we clear cp->nr_subparts_cpus and
1451 * let its child partition roots to compete for
1454 cpumask_andnot(cp->effective_cpus, cp->effective_cpus,
1456 if (cpumask_empty(cp->effective_cpus)) {
1457 cpumask_copy(cp->effective_cpus, tmp->new_cpus);
1458 cpumask_clear(cp->subparts_cpus);
1459 cp->nr_subparts_cpus = 0;
1460 } else if (!cpumask_subset(cp->subparts_cpus,
1462 cpumask_andnot(cp->subparts_cpus,
1463 cp->subparts_cpus, tmp->new_cpus);
1464 cp->nr_subparts_cpus
1465 = cpumask_weight(cp->subparts_cpus);
1469 if (new_prs != old_prs)
1470 cp->partition_root_state = new_prs;
1472 spin_unlock_irq(&callback_lock);
1473 notify_partition_change(cp, old_prs, new_prs);
1475 WARN_ON(!is_in_v2_mode() &&
1476 !cpumask_equal(cp->cpus_allowed, cp->effective_cpus));
1478 update_tasks_cpumask(cp);
1481 * On legacy hierarchy, if the effective cpumask of any non-
1482 * empty cpuset is changed, we need to rebuild sched domains.
1483 * On default hierarchy, the cpuset needs to be a partition
1486 if (!cpumask_empty(cp->cpus_allowed) &&
1487 is_sched_load_balance(cp) &&
1488 (!cgroup_subsys_on_dfl(cpuset_cgrp_subsys) ||
1489 is_partition_root(cp)))
1490 need_rebuild_sched_domains = true;
1497 if (need_rebuild_sched_domains)
1498 rebuild_sched_domains_locked();
1502 * update_sibling_cpumasks - Update siblings cpumasks
1503 * @parent: Parent cpuset
1504 * @cs: Current cpuset
1505 * @tmp: Temp variables
1507 static void update_sibling_cpumasks(struct cpuset *parent, struct cpuset *cs,
1508 struct tmpmasks *tmp)
1510 struct cpuset *sibling;
1511 struct cgroup_subsys_state *pos_css;
1514 * Check all its siblings and call update_cpumasks_hier()
1515 * if their use_parent_ecpus flag is set in order for them
1516 * to use the right effective_cpus value.
1519 cpuset_for_each_child(sibling, pos_css, parent) {
1522 if (!sibling->use_parent_ecpus)
1525 update_cpumasks_hier(sibling, tmp);
1531 * update_cpumask - update the cpus_allowed mask of a cpuset and all tasks in it
1532 * @cs: the cpuset to consider
1533 * @trialcs: trial cpuset
1534 * @buf: buffer of cpu numbers written to this cpuset
1536 static int update_cpumask(struct cpuset *cs, struct cpuset *trialcs,
1540 struct tmpmasks tmp;
1542 /* top_cpuset.cpus_allowed tracks cpu_online_mask; it's read-only */
1543 if (cs == &top_cpuset)
1547 * An empty cpus_allowed is ok only if the cpuset has no tasks.
1548 * Since cpulist_parse() fails on an empty mask, we special case
1549 * that parsing. The validate_change() call ensures that cpusets
1550 * with tasks have cpus.
1553 cpumask_clear(trialcs->cpus_allowed);
1555 retval = cpulist_parse(buf, trialcs->cpus_allowed);
1559 if (!cpumask_subset(trialcs->cpus_allowed,
1560 top_cpuset.cpus_allowed))
1564 /* Nothing to do if the cpus didn't change */
1565 if (cpumask_equal(cs->cpus_allowed, trialcs->cpus_allowed))
1568 retval = validate_change(cs, trialcs);
1572 #ifdef CONFIG_CPUMASK_OFFSTACK
1574 * Use the cpumasks in trialcs for tmpmasks when they are pointers
1575 * to allocated cpumasks.
1577 tmp.addmask = trialcs->subparts_cpus;
1578 tmp.delmask = trialcs->effective_cpus;
1579 tmp.new_cpus = trialcs->cpus_allowed;
1582 if (cs->partition_root_state) {
1583 /* Cpumask of a partition root cannot be empty */
1584 if (cpumask_empty(trialcs->cpus_allowed))
1586 if (update_parent_subparts_cpumask(cs, partcmd_update,
1587 trialcs->cpus_allowed, &tmp) < 0)
1591 spin_lock_irq(&callback_lock);
1592 cpumask_copy(cs->cpus_allowed, trialcs->cpus_allowed);
1595 * Make sure that subparts_cpus is a subset of cpus_allowed.
1597 if (cs->nr_subparts_cpus) {
1598 cpumask_andnot(cs->subparts_cpus, cs->subparts_cpus,
1600 cs->nr_subparts_cpus = cpumask_weight(cs->subparts_cpus);
1602 spin_unlock_irq(&callback_lock);
1604 update_cpumasks_hier(cs, &tmp);
1606 if (cs->partition_root_state) {
1607 struct cpuset *parent = parent_cs(cs);
1610 * For partition root, update the cpumasks of sibling
1611 * cpusets if they use parent's effective_cpus.
1613 if (parent->child_ecpus_count)
1614 update_sibling_cpumasks(parent, cs, &tmp);
1620 * Migrate memory region from one set of nodes to another. This is
1621 * performed asynchronously as it can be called from process migration path
1622 * holding locks involved in process management. All mm migrations are
1623 * performed in the queued order and can be waited for by flushing
1624 * cpuset_migrate_mm_wq.
1627 struct cpuset_migrate_mm_work {
1628 struct work_struct work;
1629 struct mm_struct *mm;
1634 static void cpuset_migrate_mm_workfn(struct work_struct *work)
1636 struct cpuset_migrate_mm_work *mwork =
1637 container_of(work, struct cpuset_migrate_mm_work, work);
1639 /* on a wq worker, no need to worry about %current's mems_allowed */
1640 do_migrate_pages(mwork->mm, &mwork->from, &mwork->to, MPOL_MF_MOVE_ALL);
1645 static void cpuset_migrate_mm(struct mm_struct *mm, const nodemask_t *from,
1646 const nodemask_t *to)
1648 struct cpuset_migrate_mm_work *mwork;
1650 if (nodes_equal(*from, *to)) {
1655 mwork = kzalloc(sizeof(*mwork), GFP_KERNEL);
1658 mwork->from = *from;
1660 INIT_WORK(&mwork->work, cpuset_migrate_mm_workfn);
1661 queue_work(cpuset_migrate_mm_wq, &mwork->work);
1667 static void cpuset_post_attach(void)
1669 flush_workqueue(cpuset_migrate_mm_wq);
1673 * cpuset_change_task_nodemask - change task's mems_allowed and mempolicy
1674 * @tsk: the task to change
1675 * @newmems: new nodes that the task will be set
1677 * We use the mems_allowed_seq seqlock to safely update both tsk->mems_allowed
1678 * and rebind an eventual tasks' mempolicy. If the task is allocating in
1679 * parallel, it might temporarily see an empty intersection, which results in
1680 * a seqlock check and retry before OOM or allocation failure.
1682 static void cpuset_change_task_nodemask(struct task_struct *tsk,
1683 nodemask_t *newmems)
1687 local_irq_disable();
1688 write_seqcount_begin(&tsk->mems_allowed_seq);
1690 nodes_or(tsk->mems_allowed, tsk->mems_allowed, *newmems);
1691 mpol_rebind_task(tsk, newmems);
1692 tsk->mems_allowed = *newmems;
1694 write_seqcount_end(&tsk->mems_allowed_seq);
1700 static void *cpuset_being_rebound;
1703 * update_tasks_nodemask - Update the nodemasks of tasks in the cpuset.
1704 * @cs: the cpuset in which each task's mems_allowed mask needs to be changed
1706 * Iterate through each task of @cs updating its mems_allowed to the
1707 * effective cpuset's. As this function is called with cpuset_mutex held,
1708 * cpuset membership stays stable.
1710 static void update_tasks_nodemask(struct cpuset *cs)
1712 static nodemask_t newmems; /* protected by cpuset_mutex */
1713 struct css_task_iter it;
1714 struct task_struct *task;
1716 cpuset_being_rebound = cs; /* causes mpol_dup() rebind */
1718 guarantee_online_mems(cs, &newmems);
1721 * The mpol_rebind_mm() call takes mmap_lock, which we couldn't
1722 * take while holding tasklist_lock. Forks can happen - the
1723 * mpol_dup() cpuset_being_rebound check will catch such forks,
1724 * and rebind their vma mempolicies too. Because we still hold
1725 * the global cpuset_mutex, we know that no other rebind effort
1726 * will be contending for the global variable cpuset_being_rebound.
1727 * It's ok if we rebind the same mm twice; mpol_rebind_mm()
1728 * is idempotent. Also migrate pages in each mm to new nodes.
1730 css_task_iter_start(&cs->css, 0, &it);
1731 while ((task = css_task_iter_next(&it))) {
1732 struct mm_struct *mm;
1735 cpuset_change_task_nodemask(task, &newmems);
1737 mm = get_task_mm(task);
1741 migrate = is_memory_migrate(cs);
1743 mpol_rebind_mm(mm, &cs->mems_allowed);
1745 cpuset_migrate_mm(mm, &cs->old_mems_allowed, &newmems);
1749 css_task_iter_end(&it);
1752 * All the tasks' nodemasks have been updated, update
1753 * cs->old_mems_allowed.
1755 cs->old_mems_allowed = newmems;
1757 /* We're done rebinding vmas to this cpuset's new mems_allowed. */
1758 cpuset_being_rebound = NULL;
1762 * update_nodemasks_hier - Update effective nodemasks and tasks in the subtree
1763 * @cs: the cpuset to consider
1764 * @new_mems: a temp variable for calculating new effective_mems
1766 * When configured nodemask is changed, the effective nodemasks of this cpuset
1767 * and all its descendants need to be updated.
1769 * On legacy hierarchy, effective_mems will be the same with mems_allowed.
1771 * Called with cpuset_mutex held
1773 static void update_nodemasks_hier(struct cpuset *cs, nodemask_t *new_mems)
1776 struct cgroup_subsys_state *pos_css;
1779 cpuset_for_each_descendant_pre(cp, pos_css, cs) {
1780 struct cpuset *parent = parent_cs(cp);
1782 nodes_and(*new_mems, cp->mems_allowed, parent->effective_mems);
1785 * If it becomes empty, inherit the effective mask of the
1786 * parent, which is guaranteed to have some MEMs.
1788 if (is_in_v2_mode() && nodes_empty(*new_mems))
1789 *new_mems = parent->effective_mems;
1791 /* Skip the whole subtree if the nodemask remains the same. */
1792 if (nodes_equal(*new_mems, cp->effective_mems)) {
1793 pos_css = css_rightmost_descendant(pos_css);
1797 if (!css_tryget_online(&cp->css))
1801 spin_lock_irq(&callback_lock);
1802 cp->effective_mems = *new_mems;
1803 spin_unlock_irq(&callback_lock);
1805 WARN_ON(!is_in_v2_mode() &&
1806 !nodes_equal(cp->mems_allowed, cp->effective_mems));
1808 update_tasks_nodemask(cp);
1817 * Handle user request to change the 'mems' memory placement
1818 * of a cpuset. Needs to validate the request, update the
1819 * cpusets mems_allowed, and for each task in the cpuset,
1820 * update mems_allowed and rebind task's mempolicy and any vma
1821 * mempolicies and if the cpuset is marked 'memory_migrate',
1822 * migrate the tasks pages to the new memory.
1824 * Call with cpuset_mutex held. May take callback_lock during call.
1825 * Will take tasklist_lock, scan tasklist for tasks in cpuset cs,
1826 * lock each such tasks mm->mmap_lock, scan its vma's and rebind
1827 * their mempolicies to the cpusets new mems_allowed.
1829 static int update_nodemask(struct cpuset *cs, struct cpuset *trialcs,
1835 * top_cpuset.mems_allowed tracks node_stats[N_MEMORY];
1838 if (cs == &top_cpuset) {
1844 * An empty mems_allowed is ok iff there are no tasks in the cpuset.
1845 * Since nodelist_parse() fails on an empty mask, we special case
1846 * that parsing. The validate_change() call ensures that cpusets
1847 * with tasks have memory.
1850 nodes_clear(trialcs->mems_allowed);
1852 retval = nodelist_parse(buf, trialcs->mems_allowed);
1856 if (!nodes_subset(trialcs->mems_allowed,
1857 top_cpuset.mems_allowed)) {
1863 if (nodes_equal(cs->mems_allowed, trialcs->mems_allowed)) {
1864 retval = 0; /* Too easy - nothing to do */
1867 retval = validate_change(cs, trialcs);
1871 spin_lock_irq(&callback_lock);
1872 cs->mems_allowed = trialcs->mems_allowed;
1873 spin_unlock_irq(&callback_lock);
1875 /* use trialcs->mems_allowed as a temp variable */
1876 update_nodemasks_hier(cs, &trialcs->mems_allowed);
1881 bool current_cpuset_is_being_rebound(void)
1886 ret = task_cs(current) == cpuset_being_rebound;
1892 static int update_relax_domain_level(struct cpuset *cs, s64 val)
1895 if (val < -1 || val >= sched_domain_level_max)
1899 if (val != cs->relax_domain_level) {
1900 cs->relax_domain_level = val;
1901 if (!cpumask_empty(cs->cpus_allowed) &&
1902 is_sched_load_balance(cs))
1903 rebuild_sched_domains_locked();
1910 * update_tasks_flags - update the spread flags of tasks in the cpuset.
1911 * @cs: the cpuset in which each task's spread flags needs to be changed
1913 * Iterate through each task of @cs updating its spread flags. As this
1914 * function is called with cpuset_mutex held, cpuset membership stays
1917 static void update_tasks_flags(struct cpuset *cs)
1919 struct css_task_iter it;
1920 struct task_struct *task;
1922 css_task_iter_start(&cs->css, 0, &it);
1923 while ((task = css_task_iter_next(&it)))
1924 cpuset_update_task_spread_flag(cs, task);
1925 css_task_iter_end(&it);
1929 * update_flag - read a 0 or a 1 in a file and update associated flag
1930 * bit: the bit to update (see cpuset_flagbits_t)
1931 * cs: the cpuset to update
1932 * turning_on: whether the flag is being set or cleared
1934 * Call with cpuset_mutex held.
1937 static int update_flag(cpuset_flagbits_t bit, struct cpuset *cs,
1940 struct cpuset *trialcs;
1941 int balance_flag_changed;
1942 int spread_flag_changed;
1945 trialcs = alloc_trial_cpuset(cs);
1950 set_bit(bit, &trialcs->flags);
1952 clear_bit(bit, &trialcs->flags);
1954 err = validate_change(cs, trialcs);
1958 balance_flag_changed = (is_sched_load_balance(cs) !=
1959 is_sched_load_balance(trialcs));
1961 spread_flag_changed = ((is_spread_slab(cs) != is_spread_slab(trialcs))
1962 || (is_spread_page(cs) != is_spread_page(trialcs)));
1964 spin_lock_irq(&callback_lock);
1965 cs->flags = trialcs->flags;
1966 spin_unlock_irq(&callback_lock);
1968 if (!cpumask_empty(trialcs->cpus_allowed) && balance_flag_changed)
1969 rebuild_sched_domains_locked();
1971 if (spread_flag_changed)
1972 update_tasks_flags(cs);
1974 free_cpuset(trialcs);
1979 * update_prstate - update partititon_root_state
1980 * cs: the cpuset to update
1981 * new_prs: new partition root state
1983 * Call with cpuset_mutex held.
1985 static int update_prstate(struct cpuset *cs, int new_prs)
1987 int err, old_prs = cs->partition_root_state;
1988 struct cpuset *parent = parent_cs(cs);
1989 struct tmpmasks tmpmask;
1991 if (old_prs == new_prs)
1995 * Cannot force a partial or invalid partition root to a full
1998 if (new_prs && (old_prs == PRS_ERROR))
2001 if (alloc_cpumasks(NULL, &tmpmask))
2007 * Turning on partition root requires setting the
2008 * CS_CPU_EXCLUSIVE bit implicitly as well and cpus_allowed
2011 if (cpumask_empty(cs->cpus_allowed))
2014 err = update_flag(CS_CPU_EXCLUSIVE, cs, 1);
2018 err = update_parent_subparts_cpumask(cs, partcmd_enable,
2021 update_flag(CS_CPU_EXCLUSIVE, cs, 0);
2026 * Turning off partition root will clear the
2027 * CS_CPU_EXCLUSIVE bit.
2029 if (old_prs == PRS_ERROR) {
2030 update_flag(CS_CPU_EXCLUSIVE, cs, 0);
2035 err = update_parent_subparts_cpumask(cs, partcmd_disable,
2040 /* Turning off CS_CPU_EXCLUSIVE will not return error */
2041 update_flag(CS_CPU_EXCLUSIVE, cs, 0);
2045 * Update cpumask of parent's tasks except when it is the top
2046 * cpuset as some system daemons cannot be mapped to other CPUs.
2048 if (parent != &top_cpuset)
2049 update_tasks_cpumask(parent);
2051 if (parent->child_ecpus_count)
2052 update_sibling_cpumasks(parent, cs, &tmpmask);
2054 rebuild_sched_domains_locked();
2057 spin_lock_irq(&callback_lock);
2058 cs->partition_root_state = new_prs;
2059 spin_unlock_irq(&callback_lock);
2060 notify_partition_change(cs, old_prs, new_prs);
2063 free_cpumasks(NULL, &tmpmask);
2068 * Frequency meter - How fast is some event occurring?
2070 * These routines manage a digitally filtered, constant time based,
2071 * event frequency meter. There are four routines:
2072 * fmeter_init() - initialize a frequency meter.
2073 * fmeter_markevent() - called each time the event happens.
2074 * fmeter_getrate() - returns the recent rate of such events.
2075 * fmeter_update() - internal routine used to update fmeter.
2077 * A common data structure is passed to each of these routines,
2078 * which is used to keep track of the state required to manage the
2079 * frequency meter and its digital filter.
2081 * The filter works on the number of events marked per unit time.
2082 * The filter is single-pole low-pass recursive (IIR). The time unit
2083 * is 1 second. Arithmetic is done using 32-bit integers scaled to
2084 * simulate 3 decimal digits of precision (multiplied by 1000).
2086 * With an FM_COEF of 933, and a time base of 1 second, the filter
2087 * has a half-life of 10 seconds, meaning that if the events quit
2088 * happening, then the rate returned from the fmeter_getrate()
2089 * will be cut in half each 10 seconds, until it converges to zero.
2091 * It is not worth doing a real infinitely recursive filter. If more
2092 * than FM_MAXTICKS ticks have elapsed since the last filter event,
2093 * just compute FM_MAXTICKS ticks worth, by which point the level
2096 * Limit the count of unprocessed events to FM_MAXCNT, so as to avoid
2097 * arithmetic overflow in the fmeter_update() routine.
2099 * Given the simple 32 bit integer arithmetic used, this meter works
2100 * best for reporting rates between one per millisecond (msec) and
2101 * one per 32 (approx) seconds. At constant rates faster than one
2102 * per msec it maxes out at values just under 1,000,000. At constant
2103 * rates between one per msec, and one per second it will stabilize
2104 * to a value N*1000, where N is the rate of events per second.
2105 * At constant rates between one per second and one per 32 seconds,
2106 * it will be choppy, moving up on the seconds that have an event,
2107 * and then decaying until the next event. At rates slower than
2108 * about one in 32 seconds, it decays all the way back to zero between
2112 #define FM_COEF 933 /* coefficient for half-life of 10 secs */
2113 #define FM_MAXTICKS ((u32)99) /* useless computing more ticks than this */
2114 #define FM_MAXCNT 1000000 /* limit cnt to avoid overflow */
2115 #define FM_SCALE 1000 /* faux fixed point scale */
2117 /* Initialize a frequency meter */
2118 static void fmeter_init(struct fmeter *fmp)
2123 spin_lock_init(&fmp->lock);
2126 /* Internal meter update - process cnt events and update value */
2127 static void fmeter_update(struct fmeter *fmp)
2132 now = ktime_get_seconds();
2133 ticks = now - fmp->time;
2138 ticks = min(FM_MAXTICKS, ticks);
2140 fmp->val = (FM_COEF * fmp->val) / FM_SCALE;
2143 fmp->val += ((FM_SCALE - FM_COEF) * fmp->cnt) / FM_SCALE;
2147 /* Process any previous ticks, then bump cnt by one (times scale). */
2148 static void fmeter_markevent(struct fmeter *fmp)
2150 spin_lock(&fmp->lock);
2152 fmp->cnt = min(FM_MAXCNT, fmp->cnt + FM_SCALE);
2153 spin_unlock(&fmp->lock);
2156 /* Process any previous ticks, then return current value. */
2157 static int fmeter_getrate(struct fmeter *fmp)
2161 spin_lock(&fmp->lock);
2164 spin_unlock(&fmp->lock);
2168 static struct cpuset *cpuset_attach_old_cs;
2170 /* Called by cgroups to determine if a cpuset is usable; cpuset_mutex held */
2171 static int cpuset_can_attach(struct cgroup_taskset *tset)
2173 struct cgroup_subsys_state *css;
2175 struct task_struct *task;
2178 /* used later by cpuset_attach() */
2179 cpuset_attach_old_cs = task_cs(cgroup_taskset_first(tset, &css));
2182 percpu_down_write(&cpuset_rwsem);
2184 /* allow moving tasks into an empty cpuset if on default hierarchy */
2186 if (!is_in_v2_mode() &&
2187 (cpumask_empty(cs->cpus_allowed) || nodes_empty(cs->mems_allowed)))
2190 cgroup_taskset_for_each(task, css, tset) {
2191 ret = task_can_attach(task, cs->cpus_allowed);
2194 ret = security_task_setscheduler(task);
2200 * Mark attach is in progress. This makes validate_change() fail
2201 * changes which zero cpus/mems_allowed.
2203 cs->attach_in_progress++;
2206 percpu_up_write(&cpuset_rwsem);
2210 static void cpuset_cancel_attach(struct cgroup_taskset *tset)
2212 struct cgroup_subsys_state *css;
2214 cgroup_taskset_first(tset, &css);
2216 percpu_down_write(&cpuset_rwsem);
2217 css_cs(css)->attach_in_progress--;
2218 percpu_up_write(&cpuset_rwsem);
2222 * Protected by cpuset_mutex. cpus_attach is used only by cpuset_attach()
2223 * but we can't allocate it dynamically there. Define it global and
2224 * allocate from cpuset_init().
2226 static cpumask_var_t cpus_attach;
2228 static void cpuset_attach(struct cgroup_taskset *tset)
2230 /* static buf protected by cpuset_mutex */
2231 static nodemask_t cpuset_attach_nodemask_to;
2232 struct task_struct *task;
2233 struct task_struct *leader;
2234 struct cgroup_subsys_state *css;
2236 struct cpuset *oldcs = cpuset_attach_old_cs;
2238 cgroup_taskset_first(tset, &css);
2241 percpu_down_write(&cpuset_rwsem);
2243 guarantee_online_mems(cs, &cpuset_attach_nodemask_to);
2245 cgroup_taskset_for_each(task, css, tset) {
2246 if (cs != &top_cpuset)
2247 guarantee_online_cpus(task, cpus_attach);
2249 cpumask_copy(cpus_attach, task_cpu_possible_mask(task));
2251 * can_attach beforehand should guarantee that this doesn't
2252 * fail. TODO: have a better way to handle failure here
2254 WARN_ON_ONCE(set_cpus_allowed_ptr(task, cpus_attach));
2256 cpuset_change_task_nodemask(task, &cpuset_attach_nodemask_to);
2257 cpuset_update_task_spread_flag(cs, task);
2261 * Change mm for all threadgroup leaders. This is expensive and may
2262 * sleep and should be moved outside migration path proper.
2264 cpuset_attach_nodemask_to = cs->effective_mems;
2265 cgroup_taskset_for_each_leader(leader, css, tset) {
2266 struct mm_struct *mm = get_task_mm(leader);
2269 mpol_rebind_mm(mm, &cpuset_attach_nodemask_to);
2272 * old_mems_allowed is the same with mems_allowed
2273 * here, except if this task is being moved
2274 * automatically due to hotplug. In that case
2275 * @mems_allowed has been updated and is empty, so
2276 * @old_mems_allowed is the right nodesets that we
2279 if (is_memory_migrate(cs))
2280 cpuset_migrate_mm(mm, &oldcs->old_mems_allowed,
2281 &cpuset_attach_nodemask_to);
2287 cs->old_mems_allowed = cpuset_attach_nodemask_to;
2289 cs->attach_in_progress--;
2290 if (!cs->attach_in_progress)
2291 wake_up(&cpuset_attach_wq);
2293 percpu_up_write(&cpuset_rwsem);
2296 /* The various types of files and directories in a cpuset file system */
2299 FILE_MEMORY_MIGRATE,
2302 FILE_EFFECTIVE_CPULIST,
2303 FILE_EFFECTIVE_MEMLIST,
2304 FILE_SUBPARTS_CPULIST,
2308 FILE_SCHED_LOAD_BALANCE,
2309 FILE_PARTITION_ROOT,
2310 FILE_SCHED_RELAX_DOMAIN_LEVEL,
2311 FILE_MEMORY_PRESSURE_ENABLED,
2312 FILE_MEMORY_PRESSURE,
2315 } cpuset_filetype_t;
2317 static int cpuset_write_u64(struct cgroup_subsys_state *css, struct cftype *cft,
2320 struct cpuset *cs = css_cs(css);
2321 cpuset_filetype_t type = cft->private;
2325 percpu_down_write(&cpuset_rwsem);
2326 if (!is_cpuset_online(cs)) {
2332 case FILE_CPU_EXCLUSIVE:
2333 retval = update_flag(CS_CPU_EXCLUSIVE, cs, val);
2335 case FILE_MEM_EXCLUSIVE:
2336 retval = update_flag(CS_MEM_EXCLUSIVE, cs, val);
2338 case FILE_MEM_HARDWALL:
2339 retval = update_flag(CS_MEM_HARDWALL, cs, val);
2341 case FILE_SCHED_LOAD_BALANCE:
2342 retval = update_flag(CS_SCHED_LOAD_BALANCE, cs, val);
2344 case FILE_MEMORY_MIGRATE:
2345 retval = update_flag(CS_MEMORY_MIGRATE, cs, val);
2347 case FILE_MEMORY_PRESSURE_ENABLED:
2348 cpuset_memory_pressure_enabled = !!val;
2350 case FILE_SPREAD_PAGE:
2351 retval = update_flag(CS_SPREAD_PAGE, cs, val);
2353 case FILE_SPREAD_SLAB:
2354 retval = update_flag(CS_SPREAD_SLAB, cs, val);
2361 percpu_up_write(&cpuset_rwsem);
2366 static int cpuset_write_s64(struct cgroup_subsys_state *css, struct cftype *cft,
2369 struct cpuset *cs = css_cs(css);
2370 cpuset_filetype_t type = cft->private;
2371 int retval = -ENODEV;
2374 percpu_down_write(&cpuset_rwsem);
2375 if (!is_cpuset_online(cs))
2379 case FILE_SCHED_RELAX_DOMAIN_LEVEL:
2380 retval = update_relax_domain_level(cs, val);
2387 percpu_up_write(&cpuset_rwsem);
2393 * Common handling for a write to a "cpus" or "mems" file.
2395 static ssize_t cpuset_write_resmask(struct kernfs_open_file *of,
2396 char *buf, size_t nbytes, loff_t off)
2398 struct cpuset *cs = css_cs(of_css(of));
2399 struct cpuset *trialcs;
2400 int retval = -ENODEV;
2402 buf = strstrip(buf);
2405 * CPU or memory hotunplug may leave @cs w/o any execution
2406 * resources, in which case the hotplug code asynchronously updates
2407 * configuration and transfers all tasks to the nearest ancestor
2408 * which can execute.
2410 * As writes to "cpus" or "mems" may restore @cs's execution
2411 * resources, wait for the previously scheduled operations before
2412 * proceeding, so that we don't end up keep removing tasks added
2413 * after execution capability is restored.
2415 * cpuset_hotplug_work calls back into cgroup core via
2416 * cgroup_transfer_tasks() and waiting for it from a cgroupfs
2417 * operation like this one can lead to a deadlock through kernfs
2418 * active_ref protection. Let's break the protection. Losing the
2419 * protection is okay as we check whether @cs is online after
2420 * grabbing cpuset_mutex anyway. This only happens on the legacy
2424 kernfs_break_active_protection(of->kn);
2425 flush_work(&cpuset_hotplug_work);
2428 percpu_down_write(&cpuset_rwsem);
2429 if (!is_cpuset_online(cs))
2432 trialcs = alloc_trial_cpuset(cs);
2438 switch (of_cft(of)->private) {
2440 retval = update_cpumask(cs, trialcs, buf);
2443 retval = update_nodemask(cs, trialcs, buf);
2450 free_cpuset(trialcs);
2452 percpu_up_write(&cpuset_rwsem);
2454 kernfs_unbreak_active_protection(of->kn);
2456 flush_workqueue(cpuset_migrate_mm_wq);
2457 return retval ?: nbytes;
2461 * These ascii lists should be read in a single call, by using a user
2462 * buffer large enough to hold the entire map. If read in smaller
2463 * chunks, there is no guarantee of atomicity. Since the display format
2464 * used, list of ranges of sequential numbers, is variable length,
2465 * and since these maps can change value dynamically, one could read
2466 * gibberish by doing partial reads while a list was changing.
2468 static int cpuset_common_seq_show(struct seq_file *sf, void *v)
2470 struct cpuset *cs = css_cs(seq_css(sf));
2471 cpuset_filetype_t type = seq_cft(sf)->private;
2474 spin_lock_irq(&callback_lock);
2478 seq_printf(sf, "%*pbl\n", cpumask_pr_args(cs->cpus_allowed));
2481 seq_printf(sf, "%*pbl\n", nodemask_pr_args(&cs->mems_allowed));
2483 case FILE_EFFECTIVE_CPULIST:
2484 seq_printf(sf, "%*pbl\n", cpumask_pr_args(cs->effective_cpus));
2486 case FILE_EFFECTIVE_MEMLIST:
2487 seq_printf(sf, "%*pbl\n", nodemask_pr_args(&cs->effective_mems));
2489 case FILE_SUBPARTS_CPULIST:
2490 seq_printf(sf, "%*pbl\n", cpumask_pr_args(cs->subparts_cpus));
2496 spin_unlock_irq(&callback_lock);
2500 static u64 cpuset_read_u64(struct cgroup_subsys_state *css, struct cftype *cft)
2502 struct cpuset *cs = css_cs(css);
2503 cpuset_filetype_t type = cft->private;
2505 case FILE_CPU_EXCLUSIVE:
2506 return is_cpu_exclusive(cs);
2507 case FILE_MEM_EXCLUSIVE:
2508 return is_mem_exclusive(cs);
2509 case FILE_MEM_HARDWALL:
2510 return is_mem_hardwall(cs);
2511 case FILE_SCHED_LOAD_BALANCE:
2512 return is_sched_load_balance(cs);
2513 case FILE_MEMORY_MIGRATE:
2514 return is_memory_migrate(cs);
2515 case FILE_MEMORY_PRESSURE_ENABLED:
2516 return cpuset_memory_pressure_enabled;
2517 case FILE_MEMORY_PRESSURE:
2518 return fmeter_getrate(&cs->fmeter);
2519 case FILE_SPREAD_PAGE:
2520 return is_spread_page(cs);
2521 case FILE_SPREAD_SLAB:
2522 return is_spread_slab(cs);
2527 /* Unreachable but makes gcc happy */
2531 static s64 cpuset_read_s64(struct cgroup_subsys_state *css, struct cftype *cft)
2533 struct cpuset *cs = css_cs(css);
2534 cpuset_filetype_t type = cft->private;
2536 case FILE_SCHED_RELAX_DOMAIN_LEVEL:
2537 return cs->relax_domain_level;
2542 /* Unreachable but makes gcc happy */
2546 static int sched_partition_show(struct seq_file *seq, void *v)
2548 struct cpuset *cs = css_cs(seq_css(seq));
2550 switch (cs->partition_root_state) {
2552 seq_puts(seq, "root\n");
2555 seq_puts(seq, "member\n");
2558 seq_puts(seq, "root invalid\n");
2564 static ssize_t sched_partition_write(struct kernfs_open_file *of, char *buf,
2565 size_t nbytes, loff_t off)
2567 struct cpuset *cs = css_cs(of_css(of));
2569 int retval = -ENODEV;
2571 buf = strstrip(buf);
2574 * Convert "root" to ENABLED, and convert "member" to DISABLED.
2576 if (!strcmp(buf, "root"))
2578 else if (!strcmp(buf, "member"))
2585 percpu_down_write(&cpuset_rwsem);
2586 if (!is_cpuset_online(cs))
2589 retval = update_prstate(cs, val);
2591 percpu_up_write(&cpuset_rwsem);
2594 return retval ?: nbytes;
2598 * for the common functions, 'private' gives the type of file
2601 static struct cftype legacy_files[] = {
2604 .seq_show = cpuset_common_seq_show,
2605 .write = cpuset_write_resmask,
2606 .max_write_len = (100U + 6 * NR_CPUS),
2607 .private = FILE_CPULIST,
2612 .seq_show = cpuset_common_seq_show,
2613 .write = cpuset_write_resmask,
2614 .max_write_len = (100U + 6 * MAX_NUMNODES),
2615 .private = FILE_MEMLIST,
2619 .name = "effective_cpus",
2620 .seq_show = cpuset_common_seq_show,
2621 .private = FILE_EFFECTIVE_CPULIST,
2625 .name = "effective_mems",
2626 .seq_show = cpuset_common_seq_show,
2627 .private = FILE_EFFECTIVE_MEMLIST,
2631 .name = "cpu_exclusive",
2632 .read_u64 = cpuset_read_u64,
2633 .write_u64 = cpuset_write_u64,
2634 .private = FILE_CPU_EXCLUSIVE,
2638 .name = "mem_exclusive",
2639 .read_u64 = cpuset_read_u64,
2640 .write_u64 = cpuset_write_u64,
2641 .private = FILE_MEM_EXCLUSIVE,
2645 .name = "mem_hardwall",
2646 .read_u64 = cpuset_read_u64,
2647 .write_u64 = cpuset_write_u64,
2648 .private = FILE_MEM_HARDWALL,
2652 .name = "sched_load_balance",
2653 .read_u64 = cpuset_read_u64,
2654 .write_u64 = cpuset_write_u64,
2655 .private = FILE_SCHED_LOAD_BALANCE,
2659 .name = "sched_relax_domain_level",
2660 .read_s64 = cpuset_read_s64,
2661 .write_s64 = cpuset_write_s64,
2662 .private = FILE_SCHED_RELAX_DOMAIN_LEVEL,
2666 .name = "memory_migrate",
2667 .read_u64 = cpuset_read_u64,
2668 .write_u64 = cpuset_write_u64,
2669 .private = FILE_MEMORY_MIGRATE,
2673 .name = "memory_pressure",
2674 .read_u64 = cpuset_read_u64,
2675 .private = FILE_MEMORY_PRESSURE,
2679 .name = "memory_spread_page",
2680 .read_u64 = cpuset_read_u64,
2681 .write_u64 = cpuset_write_u64,
2682 .private = FILE_SPREAD_PAGE,
2686 .name = "memory_spread_slab",
2687 .read_u64 = cpuset_read_u64,
2688 .write_u64 = cpuset_write_u64,
2689 .private = FILE_SPREAD_SLAB,
2693 .name = "memory_pressure_enabled",
2694 .flags = CFTYPE_ONLY_ON_ROOT,
2695 .read_u64 = cpuset_read_u64,
2696 .write_u64 = cpuset_write_u64,
2697 .private = FILE_MEMORY_PRESSURE_ENABLED,
2704 * This is currently a minimal set for the default hierarchy. It can be
2705 * expanded later on by migrating more features and control files from v1.
2707 static struct cftype dfl_files[] = {
2710 .seq_show = cpuset_common_seq_show,
2711 .write = cpuset_write_resmask,
2712 .max_write_len = (100U + 6 * NR_CPUS),
2713 .private = FILE_CPULIST,
2714 .flags = CFTYPE_NOT_ON_ROOT,
2719 .seq_show = cpuset_common_seq_show,
2720 .write = cpuset_write_resmask,
2721 .max_write_len = (100U + 6 * MAX_NUMNODES),
2722 .private = FILE_MEMLIST,
2723 .flags = CFTYPE_NOT_ON_ROOT,
2727 .name = "cpus.effective",
2728 .seq_show = cpuset_common_seq_show,
2729 .private = FILE_EFFECTIVE_CPULIST,
2733 .name = "mems.effective",
2734 .seq_show = cpuset_common_seq_show,
2735 .private = FILE_EFFECTIVE_MEMLIST,
2739 .name = "cpus.partition",
2740 .seq_show = sched_partition_show,
2741 .write = sched_partition_write,
2742 .private = FILE_PARTITION_ROOT,
2743 .flags = CFTYPE_NOT_ON_ROOT,
2744 .file_offset = offsetof(struct cpuset, partition_file),
2748 .name = "cpus.subpartitions",
2749 .seq_show = cpuset_common_seq_show,
2750 .private = FILE_SUBPARTS_CPULIST,
2751 .flags = CFTYPE_DEBUG,
2759 * cpuset_css_alloc - allocate a cpuset css
2760 * cgrp: control group that the new cpuset will be part of
2763 static struct cgroup_subsys_state *
2764 cpuset_css_alloc(struct cgroup_subsys_state *parent_css)
2769 return &top_cpuset.css;
2771 cs = kzalloc(sizeof(*cs), GFP_KERNEL);
2773 return ERR_PTR(-ENOMEM);
2775 if (alloc_cpumasks(cs, NULL)) {
2777 return ERR_PTR(-ENOMEM);
2780 __set_bit(CS_SCHED_LOAD_BALANCE, &cs->flags);
2781 nodes_clear(cs->mems_allowed);
2782 nodes_clear(cs->effective_mems);
2783 fmeter_init(&cs->fmeter);
2784 cs->relax_domain_level = -1;
2786 /* Set CS_MEMORY_MIGRATE for default hierarchy */
2787 if (cgroup_subsys_on_dfl(cpuset_cgrp_subsys))
2788 __set_bit(CS_MEMORY_MIGRATE, &cs->flags);
2793 static int cpuset_css_online(struct cgroup_subsys_state *css)
2795 struct cpuset *cs = css_cs(css);
2796 struct cpuset *parent = parent_cs(cs);
2797 struct cpuset *tmp_cs;
2798 struct cgroup_subsys_state *pos_css;
2804 percpu_down_write(&cpuset_rwsem);
2806 set_bit(CS_ONLINE, &cs->flags);
2807 if (is_spread_page(parent))
2808 set_bit(CS_SPREAD_PAGE, &cs->flags);
2809 if (is_spread_slab(parent))
2810 set_bit(CS_SPREAD_SLAB, &cs->flags);
2814 spin_lock_irq(&callback_lock);
2815 if (is_in_v2_mode()) {
2816 cpumask_copy(cs->effective_cpus, parent->effective_cpus);
2817 cs->effective_mems = parent->effective_mems;
2818 cs->use_parent_ecpus = true;
2819 parent->child_ecpus_count++;
2821 spin_unlock_irq(&callback_lock);
2823 if (!test_bit(CGRP_CPUSET_CLONE_CHILDREN, &css->cgroup->flags))
2827 * Clone @parent's configuration if CGRP_CPUSET_CLONE_CHILDREN is
2828 * set. This flag handling is implemented in cgroup core for
2829 * histrical reasons - the flag may be specified during mount.
2831 * Currently, if any sibling cpusets have exclusive cpus or mem, we
2832 * refuse to clone the configuration - thereby refusing the task to
2833 * be entered, and as a result refusing the sys_unshare() or
2834 * clone() which initiated it. If this becomes a problem for some
2835 * users who wish to allow that scenario, then this could be
2836 * changed to grant parent->cpus_allowed-sibling_cpus_exclusive
2837 * (and likewise for mems) to the new cgroup.
2840 cpuset_for_each_child(tmp_cs, pos_css, parent) {
2841 if (is_mem_exclusive(tmp_cs) || is_cpu_exclusive(tmp_cs)) {
2848 spin_lock_irq(&callback_lock);
2849 cs->mems_allowed = parent->mems_allowed;
2850 cs->effective_mems = parent->mems_allowed;
2851 cpumask_copy(cs->cpus_allowed, parent->cpus_allowed);
2852 cpumask_copy(cs->effective_cpus, parent->cpus_allowed);
2853 spin_unlock_irq(&callback_lock);
2855 percpu_up_write(&cpuset_rwsem);
2861 * If the cpuset being removed has its flag 'sched_load_balance'
2862 * enabled, then simulate turning sched_load_balance off, which
2863 * will call rebuild_sched_domains_locked(). That is not needed
2864 * in the default hierarchy where only changes in partition
2865 * will cause repartitioning.
2867 * If the cpuset has the 'sched.partition' flag enabled, simulate
2868 * turning 'sched.partition" off.
2871 static void cpuset_css_offline(struct cgroup_subsys_state *css)
2873 struct cpuset *cs = css_cs(css);
2876 percpu_down_write(&cpuset_rwsem);
2878 if (is_partition_root(cs))
2879 update_prstate(cs, 0);
2881 if (!cgroup_subsys_on_dfl(cpuset_cgrp_subsys) &&
2882 is_sched_load_balance(cs))
2883 update_flag(CS_SCHED_LOAD_BALANCE, cs, 0);
2885 if (cs->use_parent_ecpus) {
2886 struct cpuset *parent = parent_cs(cs);
2888 cs->use_parent_ecpus = false;
2889 parent->child_ecpus_count--;
2893 clear_bit(CS_ONLINE, &cs->flags);
2895 percpu_up_write(&cpuset_rwsem);
2899 static void cpuset_css_free(struct cgroup_subsys_state *css)
2901 struct cpuset *cs = css_cs(css);
2906 static void cpuset_bind(struct cgroup_subsys_state *root_css)
2908 percpu_down_write(&cpuset_rwsem);
2909 spin_lock_irq(&callback_lock);
2911 if (is_in_v2_mode()) {
2912 cpumask_copy(top_cpuset.cpus_allowed, cpu_possible_mask);
2913 top_cpuset.mems_allowed = node_possible_map;
2915 cpumask_copy(top_cpuset.cpus_allowed,
2916 top_cpuset.effective_cpus);
2917 top_cpuset.mems_allowed = top_cpuset.effective_mems;
2920 spin_unlock_irq(&callback_lock);
2921 percpu_up_write(&cpuset_rwsem);
2925 * Make sure the new task conform to the current state of its parent,
2926 * which could have been changed by cpuset just after it inherits the
2927 * state from the parent and before it sits on the cgroup's task list.
2929 static void cpuset_fork(struct task_struct *task)
2931 if (task_css_is_root(task, cpuset_cgrp_id))
2934 set_cpus_allowed_ptr(task, current->cpus_ptr);
2935 task->mems_allowed = current->mems_allowed;
2938 struct cgroup_subsys cpuset_cgrp_subsys = {
2939 .css_alloc = cpuset_css_alloc,
2940 .css_online = cpuset_css_online,
2941 .css_offline = cpuset_css_offline,
2942 .css_free = cpuset_css_free,
2943 .can_attach = cpuset_can_attach,
2944 .cancel_attach = cpuset_cancel_attach,
2945 .attach = cpuset_attach,
2946 .post_attach = cpuset_post_attach,
2947 .bind = cpuset_bind,
2948 .fork = cpuset_fork,
2949 .legacy_cftypes = legacy_files,
2950 .dfl_cftypes = dfl_files,
2956 * cpuset_init - initialize cpusets at system boot
2958 * Description: Initialize top_cpuset
2961 int __init cpuset_init(void)
2963 BUG_ON(percpu_init_rwsem(&cpuset_rwsem));
2965 BUG_ON(!alloc_cpumask_var(&top_cpuset.cpus_allowed, GFP_KERNEL));
2966 BUG_ON(!alloc_cpumask_var(&top_cpuset.effective_cpus, GFP_KERNEL));
2967 BUG_ON(!zalloc_cpumask_var(&top_cpuset.subparts_cpus, GFP_KERNEL));
2969 cpumask_setall(top_cpuset.cpus_allowed);
2970 nodes_setall(top_cpuset.mems_allowed);
2971 cpumask_setall(top_cpuset.effective_cpus);
2972 nodes_setall(top_cpuset.effective_mems);
2974 fmeter_init(&top_cpuset.fmeter);
2975 set_bit(CS_SCHED_LOAD_BALANCE, &top_cpuset.flags);
2976 top_cpuset.relax_domain_level = -1;
2978 BUG_ON(!alloc_cpumask_var(&cpus_attach, GFP_KERNEL));
2984 * If CPU and/or memory hotplug handlers, below, unplug any CPUs
2985 * or memory nodes, we need to walk over the cpuset hierarchy,
2986 * removing that CPU or node from all cpusets. If this removes the
2987 * last CPU or node from a cpuset, then move the tasks in the empty
2988 * cpuset to its next-highest non-empty parent.
2990 static void remove_tasks_in_empty_cpuset(struct cpuset *cs)
2992 struct cpuset *parent;
2995 * Find its next-highest non-empty parent, (top cpuset
2996 * has online cpus, so can't be empty).
2998 parent = parent_cs(cs);
2999 while (cpumask_empty(parent->cpus_allowed) ||
3000 nodes_empty(parent->mems_allowed))
3001 parent = parent_cs(parent);
3003 if (cgroup_transfer_tasks(parent->css.cgroup, cs->css.cgroup)) {
3004 pr_err("cpuset: failed to transfer tasks out of empty cpuset ");
3005 pr_cont_cgroup_name(cs->css.cgroup);
3011 hotplug_update_tasks_legacy(struct cpuset *cs,
3012 struct cpumask *new_cpus, nodemask_t *new_mems,
3013 bool cpus_updated, bool mems_updated)
3017 spin_lock_irq(&callback_lock);
3018 cpumask_copy(cs->cpus_allowed, new_cpus);
3019 cpumask_copy(cs->effective_cpus, new_cpus);
3020 cs->mems_allowed = *new_mems;
3021 cs->effective_mems = *new_mems;
3022 spin_unlock_irq(&callback_lock);
3025 * Don't call update_tasks_cpumask() if the cpuset becomes empty,
3026 * as the tasks will be migratecd to an ancestor.
3028 if (cpus_updated && !cpumask_empty(cs->cpus_allowed))
3029 update_tasks_cpumask(cs);
3030 if (mems_updated && !nodes_empty(cs->mems_allowed))
3031 update_tasks_nodemask(cs);
3033 is_empty = cpumask_empty(cs->cpus_allowed) ||
3034 nodes_empty(cs->mems_allowed);
3036 percpu_up_write(&cpuset_rwsem);
3039 * Move tasks to the nearest ancestor with execution resources,
3040 * This is full cgroup operation which will also call back into
3041 * cpuset. Should be done outside any lock.
3044 remove_tasks_in_empty_cpuset(cs);
3046 percpu_down_write(&cpuset_rwsem);
3050 hotplug_update_tasks(struct cpuset *cs,
3051 struct cpumask *new_cpus, nodemask_t *new_mems,
3052 bool cpus_updated, bool mems_updated)
3054 if (cpumask_empty(new_cpus))
3055 cpumask_copy(new_cpus, parent_cs(cs)->effective_cpus);
3056 if (nodes_empty(*new_mems))
3057 *new_mems = parent_cs(cs)->effective_mems;
3059 spin_lock_irq(&callback_lock);
3060 cpumask_copy(cs->effective_cpus, new_cpus);
3061 cs->effective_mems = *new_mems;
3062 spin_unlock_irq(&callback_lock);
3065 update_tasks_cpumask(cs);
3067 update_tasks_nodemask(cs);
3070 static bool force_rebuild;
3072 void cpuset_force_rebuild(void)
3074 force_rebuild = true;
3078 * cpuset_hotplug_update_tasks - update tasks in a cpuset for hotunplug
3079 * @cs: cpuset in interest
3080 * @tmp: the tmpmasks structure pointer
3082 * Compare @cs's cpu and mem masks against top_cpuset and if some have gone
3083 * offline, update @cs accordingly. If @cs ends up with no CPU or memory,
3084 * all its tasks are moved to the nearest ancestor with both resources.
3086 static void cpuset_hotplug_update_tasks(struct cpuset *cs, struct tmpmasks *tmp)
3088 static cpumask_t new_cpus;
3089 static nodemask_t new_mems;
3092 struct cpuset *parent;
3094 wait_event(cpuset_attach_wq, cs->attach_in_progress == 0);
3096 percpu_down_write(&cpuset_rwsem);
3099 * We have raced with task attaching. We wait until attaching
3100 * is finished, so we won't attach a task to an empty cpuset.
3102 if (cs->attach_in_progress) {
3103 percpu_up_write(&cpuset_rwsem);
3107 parent = parent_cs(cs);
3108 compute_effective_cpumask(&new_cpus, cs, parent);
3109 nodes_and(new_mems, cs->mems_allowed, parent->effective_mems);
3111 if (cs->nr_subparts_cpus)
3113 * Make sure that CPUs allocated to child partitions
3114 * do not show up in effective_cpus.
3116 cpumask_andnot(&new_cpus, &new_cpus, cs->subparts_cpus);
3118 if (!tmp || !cs->partition_root_state)
3122 * In the unlikely event that a partition root has empty
3123 * effective_cpus or its parent becomes erroneous, we have to
3124 * transition it to the erroneous state.
3126 if (is_partition_root(cs) && (cpumask_empty(&new_cpus) ||
3127 (parent->partition_root_state == PRS_ERROR))) {
3128 if (cs->nr_subparts_cpus) {
3129 spin_lock_irq(&callback_lock);
3130 cs->nr_subparts_cpus = 0;
3131 cpumask_clear(cs->subparts_cpus);
3132 spin_unlock_irq(&callback_lock);
3133 compute_effective_cpumask(&new_cpus, cs, parent);
3137 * If the effective_cpus is empty because the child
3138 * partitions take away all the CPUs, we can keep
3139 * the current partition and let the child partitions
3140 * fight for available CPUs.
3142 if ((parent->partition_root_state == PRS_ERROR) ||
3143 cpumask_empty(&new_cpus)) {
3146 update_parent_subparts_cpumask(cs, partcmd_disable,
3148 old_prs = cs->partition_root_state;
3149 if (old_prs != PRS_ERROR) {
3150 spin_lock_irq(&callback_lock);
3151 cs->partition_root_state = PRS_ERROR;
3152 spin_unlock_irq(&callback_lock);
3153 notify_partition_change(cs, old_prs, PRS_ERROR);
3156 cpuset_force_rebuild();
3160 * On the other hand, an erroneous partition root may be transitioned
3161 * back to a regular one or a partition root with no CPU allocated
3162 * from the parent may change to erroneous.
3164 if (is_partition_root(parent) &&
3165 ((cs->partition_root_state == PRS_ERROR) ||
3166 !cpumask_intersects(&new_cpus, parent->subparts_cpus)) &&
3167 update_parent_subparts_cpumask(cs, partcmd_update, NULL, tmp))
3168 cpuset_force_rebuild();
3171 cpus_updated = !cpumask_equal(&new_cpus, cs->effective_cpus);
3172 mems_updated = !nodes_equal(new_mems, cs->effective_mems);
3174 if (is_in_v2_mode())
3175 hotplug_update_tasks(cs, &new_cpus, &new_mems,
3176 cpus_updated, mems_updated);
3178 hotplug_update_tasks_legacy(cs, &new_cpus, &new_mems,
3179 cpus_updated, mems_updated);
3181 percpu_up_write(&cpuset_rwsem);
3185 * cpuset_hotplug_workfn - handle CPU/memory hotunplug for a cpuset
3187 * This function is called after either CPU or memory configuration has
3188 * changed and updates cpuset accordingly. The top_cpuset is always
3189 * synchronized to cpu_active_mask and N_MEMORY, which is necessary in
3190 * order to make cpusets transparent (of no affect) on systems that are
3191 * actively using CPU hotplug but making no active use of cpusets.
3193 * Non-root cpusets are only affected by offlining. If any CPUs or memory
3194 * nodes have been taken down, cpuset_hotplug_update_tasks() is invoked on
3197 * Note that CPU offlining during suspend is ignored. We don't modify
3198 * cpusets across suspend/resume cycles at all.
3200 static void cpuset_hotplug_workfn(struct work_struct *work)
3202 static cpumask_t new_cpus;
3203 static nodemask_t new_mems;
3204 bool cpus_updated, mems_updated;
3205 bool on_dfl = is_in_v2_mode();
3206 struct tmpmasks tmp, *ptmp = NULL;
3208 if (on_dfl && !alloc_cpumasks(NULL, &tmp))
3211 percpu_down_write(&cpuset_rwsem);
3213 /* fetch the available cpus/mems and find out which changed how */
3214 cpumask_copy(&new_cpus, cpu_active_mask);
3215 new_mems = node_states[N_MEMORY];
3218 * If subparts_cpus is populated, it is likely that the check below
3219 * will produce a false positive on cpus_updated when the cpu list
3220 * isn't changed. It is extra work, but it is better to be safe.
3222 cpus_updated = !cpumask_equal(top_cpuset.effective_cpus, &new_cpus);
3223 mems_updated = !nodes_equal(top_cpuset.effective_mems, new_mems);
3226 * In the rare case that hotplug removes all the cpus in subparts_cpus,
3227 * we assumed that cpus are updated.
3229 if (!cpus_updated && top_cpuset.nr_subparts_cpus)
3230 cpus_updated = true;
3232 /* synchronize cpus_allowed to cpu_active_mask */
3234 spin_lock_irq(&callback_lock);
3236 cpumask_copy(top_cpuset.cpus_allowed, &new_cpus);
3238 * Make sure that CPUs allocated to child partitions
3239 * do not show up in effective_cpus. If no CPU is left,
3240 * we clear the subparts_cpus & let the child partitions
3241 * fight for the CPUs again.
3243 if (top_cpuset.nr_subparts_cpus) {
3244 if (cpumask_subset(&new_cpus,
3245 top_cpuset.subparts_cpus)) {
3246 top_cpuset.nr_subparts_cpus = 0;
3247 cpumask_clear(top_cpuset.subparts_cpus);
3249 cpumask_andnot(&new_cpus, &new_cpus,
3250 top_cpuset.subparts_cpus);
3253 cpumask_copy(top_cpuset.effective_cpus, &new_cpus);
3254 spin_unlock_irq(&callback_lock);
3255 /* we don't mess with cpumasks of tasks in top_cpuset */
3258 /* synchronize mems_allowed to N_MEMORY */
3260 spin_lock_irq(&callback_lock);
3262 top_cpuset.mems_allowed = new_mems;
3263 top_cpuset.effective_mems = new_mems;
3264 spin_unlock_irq(&callback_lock);
3265 update_tasks_nodemask(&top_cpuset);
3268 percpu_up_write(&cpuset_rwsem);
3270 /* if cpus or mems changed, we need to propagate to descendants */
3271 if (cpus_updated || mems_updated) {
3273 struct cgroup_subsys_state *pos_css;
3276 cpuset_for_each_descendant_pre(cs, pos_css, &top_cpuset) {
3277 if (cs == &top_cpuset || !css_tryget_online(&cs->css))
3281 cpuset_hotplug_update_tasks(cs, ptmp);
3289 /* rebuild sched domains if cpus_allowed has changed */
3290 if (cpus_updated || force_rebuild) {
3291 force_rebuild = false;
3292 rebuild_sched_domains();
3295 free_cpumasks(NULL, ptmp);
3298 void cpuset_update_active_cpus(void)
3301 * We're inside cpu hotplug critical region which usually nests
3302 * inside cgroup synchronization. Bounce actual hotplug processing
3303 * to a work item to avoid reverse locking order.
3305 schedule_work(&cpuset_hotplug_work);
3308 void cpuset_wait_for_hotplug(void)
3310 flush_work(&cpuset_hotplug_work);
3314 * Keep top_cpuset.mems_allowed tracking node_states[N_MEMORY].
3315 * Call this routine anytime after node_states[N_MEMORY] changes.
3316 * See cpuset_update_active_cpus() for CPU hotplug handling.
3318 static int cpuset_track_online_nodes(struct notifier_block *self,
3319 unsigned long action, void *arg)
3321 schedule_work(&cpuset_hotplug_work);
3325 static struct notifier_block cpuset_track_online_nodes_nb = {
3326 .notifier_call = cpuset_track_online_nodes,
3327 .priority = 10, /* ??! */
3331 * cpuset_init_smp - initialize cpus_allowed
3333 * Description: Finish top cpuset after cpu, node maps are initialized
3335 void __init cpuset_init_smp(void)
3337 cpumask_copy(top_cpuset.cpus_allowed, cpu_active_mask);
3338 top_cpuset.mems_allowed = node_states[N_MEMORY];
3339 top_cpuset.old_mems_allowed = top_cpuset.mems_allowed;
3341 cpumask_copy(top_cpuset.effective_cpus, cpu_active_mask);
3342 top_cpuset.effective_mems = node_states[N_MEMORY];
3344 register_hotmemory_notifier(&cpuset_track_online_nodes_nb);
3346 cpuset_migrate_mm_wq = alloc_ordered_workqueue("cpuset_migrate_mm", 0);
3347 BUG_ON(!cpuset_migrate_mm_wq);
3351 * cpuset_cpus_allowed - return cpus_allowed mask from a tasks cpuset.
3352 * @tsk: pointer to task_struct from which to obtain cpuset->cpus_allowed.
3353 * @pmask: pointer to struct cpumask variable to receive cpus_allowed set.
3355 * Description: Returns the cpumask_var_t cpus_allowed of the cpuset
3356 * attached to the specified @tsk. Guaranteed to return some non-empty
3357 * subset of cpu_online_mask, even if this means going outside the
3361 void cpuset_cpus_allowed(struct task_struct *tsk, struct cpumask *pmask)
3363 unsigned long flags;
3365 spin_lock_irqsave(&callback_lock, flags);
3366 guarantee_online_cpus(tsk, pmask);
3367 spin_unlock_irqrestore(&callback_lock, flags);
3371 * cpuset_cpus_allowed_fallback - final fallback before complete catastrophe.
3372 * @tsk: pointer to task_struct with which the scheduler is struggling
3374 * Description: In the case that the scheduler cannot find an allowed cpu in
3375 * tsk->cpus_allowed, we fall back to task_cs(tsk)->cpus_allowed. In legacy
3376 * mode however, this value is the same as task_cs(tsk)->effective_cpus,
3377 * which will not contain a sane cpumask during cases such as cpu hotplugging.
3378 * This is the absolute last resort for the scheduler and it is only used if
3379 * _every_ other avenue has been traveled.
3381 * Returns true if the affinity of @tsk was changed, false otherwise.
3384 bool cpuset_cpus_allowed_fallback(struct task_struct *tsk)
3386 const struct cpumask *possible_mask = task_cpu_possible_mask(tsk);
3387 const struct cpumask *cs_mask;
3388 bool changed = false;
3391 cs_mask = task_cs(tsk)->cpus_allowed;
3392 if (is_in_v2_mode() && cpumask_subset(cs_mask, possible_mask)) {
3393 do_set_cpus_allowed(tsk, cs_mask);
3399 * We own tsk->cpus_allowed, nobody can change it under us.
3401 * But we used cs && cs->cpus_allowed lockless and thus can
3402 * race with cgroup_attach_task() or update_cpumask() and get
3403 * the wrong tsk->cpus_allowed. However, both cases imply the
3404 * subsequent cpuset_change_cpumask()->set_cpus_allowed_ptr()
3405 * which takes task_rq_lock().
3407 * If we are called after it dropped the lock we must see all
3408 * changes in tsk_cs()->cpus_allowed. Otherwise we can temporary
3409 * set any mask even if it is not right from task_cs() pov,
3410 * the pending set_cpus_allowed_ptr() will fix things.
3412 * select_fallback_rq() will fix things ups and set cpu_possible_mask
3418 void __init cpuset_init_current_mems_allowed(void)
3420 nodes_setall(current->mems_allowed);
3424 * cpuset_mems_allowed - return mems_allowed mask from a tasks cpuset.
3425 * @tsk: pointer to task_struct from which to obtain cpuset->mems_allowed.
3427 * Description: Returns the nodemask_t mems_allowed of the cpuset
3428 * attached to the specified @tsk. Guaranteed to return some non-empty
3429 * subset of node_states[N_MEMORY], even if this means going outside the
3433 nodemask_t cpuset_mems_allowed(struct task_struct *tsk)
3436 unsigned long flags;
3438 spin_lock_irqsave(&callback_lock, flags);
3440 guarantee_online_mems(task_cs(tsk), &mask);
3442 spin_unlock_irqrestore(&callback_lock, flags);
3448 * cpuset_nodemask_valid_mems_allowed - check nodemask vs. current mems_allowed
3449 * @nodemask: the nodemask to be checked
3451 * Are any of the nodes in the nodemask allowed in current->mems_allowed?
3453 int cpuset_nodemask_valid_mems_allowed(nodemask_t *nodemask)
3455 return nodes_intersects(*nodemask, current->mems_allowed);
3459 * nearest_hardwall_ancestor() - Returns the nearest mem_exclusive or
3460 * mem_hardwall ancestor to the specified cpuset. Call holding
3461 * callback_lock. If no ancestor is mem_exclusive or mem_hardwall
3462 * (an unusual configuration), then returns the root cpuset.
3464 static struct cpuset *nearest_hardwall_ancestor(struct cpuset *cs)
3466 while (!(is_mem_exclusive(cs) || is_mem_hardwall(cs)) && parent_cs(cs))
3472 * cpuset_node_allowed - Can we allocate on a memory node?
3473 * @node: is this an allowed node?
3474 * @gfp_mask: memory allocation flags
3476 * If we're in interrupt, yes, we can always allocate. If @node is set in
3477 * current's mems_allowed, yes. If it's not a __GFP_HARDWALL request and this
3478 * node is set in the nearest hardwalled cpuset ancestor to current's cpuset,
3479 * yes. If current has access to memory reserves as an oom victim, yes.
3482 * GFP_USER allocations are marked with the __GFP_HARDWALL bit,
3483 * and do not allow allocations outside the current tasks cpuset
3484 * unless the task has been OOM killed.
3485 * GFP_KERNEL allocations are not so marked, so can escape to the
3486 * nearest enclosing hardwalled ancestor cpuset.
3488 * Scanning up parent cpusets requires callback_lock. The
3489 * __alloc_pages() routine only calls here with __GFP_HARDWALL bit
3490 * _not_ set if it's a GFP_KERNEL allocation, and all nodes in the
3491 * current tasks mems_allowed came up empty on the first pass over
3492 * the zonelist. So only GFP_KERNEL allocations, if all nodes in the
3493 * cpuset are short of memory, might require taking the callback_lock.
3495 * The first call here from mm/page_alloc:get_page_from_freelist()
3496 * has __GFP_HARDWALL set in gfp_mask, enforcing hardwall cpusets,
3497 * so no allocation on a node outside the cpuset is allowed (unless
3498 * in interrupt, of course).
3500 * The second pass through get_page_from_freelist() doesn't even call
3501 * here for GFP_ATOMIC calls. For those calls, the __alloc_pages()
3502 * variable 'wait' is not set, and the bit ALLOC_CPUSET is not set
3503 * in alloc_flags. That logic and the checks below have the combined
3505 * in_interrupt - any node ok (current task context irrelevant)
3506 * GFP_ATOMIC - any node ok
3507 * tsk_is_oom_victim - any node ok
3508 * GFP_KERNEL - any node in enclosing hardwalled cpuset ok
3509 * GFP_USER - only nodes in current tasks mems allowed ok.
3511 bool __cpuset_node_allowed(int node, gfp_t gfp_mask)
3513 struct cpuset *cs; /* current cpuset ancestors */
3514 int allowed; /* is allocation in zone z allowed? */
3515 unsigned long flags;
3519 if (node_isset(node, current->mems_allowed))
3522 * Allow tasks that have access to memory reserves because they have
3523 * been OOM killed to get memory anywhere.
3525 if (unlikely(tsk_is_oom_victim(current)))
3527 if (gfp_mask & __GFP_HARDWALL) /* If hardwall request, stop here */
3530 if (current->flags & PF_EXITING) /* Let dying task have memory */
3533 /* Not hardwall and node outside mems_allowed: scan up cpusets */
3534 spin_lock_irqsave(&callback_lock, flags);
3537 cs = nearest_hardwall_ancestor(task_cs(current));
3538 allowed = node_isset(node, cs->mems_allowed);
3541 spin_unlock_irqrestore(&callback_lock, flags);
3546 * cpuset_mem_spread_node() - On which node to begin search for a file page
3547 * cpuset_slab_spread_node() - On which node to begin search for a slab page
3549 * If a task is marked PF_SPREAD_PAGE or PF_SPREAD_SLAB (as for
3550 * tasks in a cpuset with is_spread_page or is_spread_slab set),
3551 * and if the memory allocation used cpuset_mem_spread_node()
3552 * to determine on which node to start looking, as it will for
3553 * certain page cache or slab cache pages such as used for file
3554 * system buffers and inode caches, then instead of starting on the
3555 * local node to look for a free page, rather spread the starting
3556 * node around the tasks mems_allowed nodes.
3558 * We don't have to worry about the returned node being offline
3559 * because "it can't happen", and even if it did, it would be ok.
3561 * The routines calling guarantee_online_mems() are careful to
3562 * only set nodes in task->mems_allowed that are online. So it
3563 * should not be possible for the following code to return an
3564 * offline node. But if it did, that would be ok, as this routine
3565 * is not returning the node where the allocation must be, only
3566 * the node where the search should start. The zonelist passed to
3567 * __alloc_pages() will include all nodes. If the slab allocator
3568 * is passed an offline node, it will fall back to the local node.
3569 * See kmem_cache_alloc_node().
3572 static int cpuset_spread_node(int *rotor)
3574 return *rotor = next_node_in(*rotor, current->mems_allowed);
3577 int cpuset_mem_spread_node(void)
3579 if (current->cpuset_mem_spread_rotor == NUMA_NO_NODE)
3580 current->cpuset_mem_spread_rotor =
3581 node_random(¤t->mems_allowed);
3583 return cpuset_spread_node(¤t->cpuset_mem_spread_rotor);
3586 int cpuset_slab_spread_node(void)
3588 if (current->cpuset_slab_spread_rotor == NUMA_NO_NODE)
3589 current->cpuset_slab_spread_rotor =
3590 node_random(¤t->mems_allowed);
3592 return cpuset_spread_node(¤t->cpuset_slab_spread_rotor);
3595 EXPORT_SYMBOL_GPL(cpuset_mem_spread_node);
3598 * cpuset_mems_allowed_intersects - Does @tsk1's mems_allowed intersect @tsk2's?
3599 * @tsk1: pointer to task_struct of some task.
3600 * @tsk2: pointer to task_struct of some other task.
3602 * Description: Return true if @tsk1's mems_allowed intersects the
3603 * mems_allowed of @tsk2. Used by the OOM killer to determine if
3604 * one of the task's memory usage might impact the memory available
3608 int cpuset_mems_allowed_intersects(const struct task_struct *tsk1,
3609 const struct task_struct *tsk2)
3611 return nodes_intersects(tsk1->mems_allowed, tsk2->mems_allowed);
3615 * cpuset_print_current_mems_allowed - prints current's cpuset and mems_allowed
3617 * Description: Prints current's name, cpuset name, and cached copy of its
3618 * mems_allowed to the kernel log.
3620 void cpuset_print_current_mems_allowed(void)
3622 struct cgroup *cgrp;
3626 cgrp = task_cs(current)->css.cgroup;
3627 pr_cont(",cpuset=");
3628 pr_cont_cgroup_name(cgrp);
3629 pr_cont(",mems_allowed=%*pbl",
3630 nodemask_pr_args(¤t->mems_allowed));
3636 * Collection of memory_pressure is suppressed unless
3637 * this flag is enabled by writing "1" to the special
3638 * cpuset file 'memory_pressure_enabled' in the root cpuset.
3641 int cpuset_memory_pressure_enabled __read_mostly;
3644 * cpuset_memory_pressure_bump - keep stats of per-cpuset reclaims.
3646 * Keep a running average of the rate of synchronous (direct)
3647 * page reclaim efforts initiated by tasks in each cpuset.
3649 * This represents the rate at which some task in the cpuset
3650 * ran low on memory on all nodes it was allowed to use, and
3651 * had to enter the kernels page reclaim code in an effort to
3652 * create more free memory by tossing clean pages or swapping
3653 * or writing dirty pages.
3655 * Display to user space in the per-cpuset read-only file
3656 * "memory_pressure". Value displayed is an integer
3657 * representing the recent rate of entry into the synchronous
3658 * (direct) page reclaim by any task attached to the cpuset.
3661 void __cpuset_memory_pressure_bump(void)
3664 fmeter_markevent(&task_cs(current)->fmeter);
3668 #ifdef CONFIG_PROC_PID_CPUSET
3670 * proc_cpuset_show()
3671 * - Print tasks cpuset path into seq_file.
3672 * - Used for /proc/<pid>/cpuset.
3673 * - No need to task_lock(tsk) on this tsk->cpuset reference, as it
3674 * doesn't really matter if tsk->cpuset changes after we read it,
3675 * and we take cpuset_mutex, keeping cpuset_attach() from changing it
3678 int proc_cpuset_show(struct seq_file *m, struct pid_namespace *ns,
3679 struct pid *pid, struct task_struct *tsk)
3682 struct cgroup_subsys_state *css;
3686 buf = kmalloc(PATH_MAX, GFP_KERNEL);
3690 css = task_get_css(tsk, cpuset_cgrp_id);
3691 retval = cgroup_path_ns(css->cgroup, buf, PATH_MAX,
3692 current->nsproxy->cgroup_ns);
3694 if (retval >= PATH_MAX)
3695 retval = -ENAMETOOLONG;
3706 #endif /* CONFIG_PROC_PID_CPUSET */
3708 /* Display task mems_allowed in /proc/<pid>/status file. */
3709 void cpuset_task_status_allowed(struct seq_file *m, struct task_struct *task)
3711 seq_printf(m, "Mems_allowed:\t%*pb\n",
3712 nodemask_pr_args(&task->mems_allowed));
3713 seq_printf(m, "Mems_allowed_list:\t%*pbl\n",
3714 nodemask_pr_args(&task->mems_allowed));