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 hierachy:
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;
166 * Partition root states:
168 * 0 - not a partition root
172 * -1 - invalid partition root
173 * None of the cpus in cpus_allowed can be put into the parent's
174 * subparts_cpus. In this case, the cpuset is not a real partition
175 * root anymore. However, the CPU_EXCLUSIVE bit will still be set
176 * and the cpuset can be restored back to a partition root if the
177 * parent cpuset can give more CPUs back to this child cpuset.
179 #define PRS_DISABLED 0
180 #define PRS_ENABLED 1
184 * Temporary cpumasks for working with partitions that are passed among
185 * functions to avoid memory allocation in inner functions.
188 cpumask_var_t addmask, delmask; /* For partition root */
189 cpumask_var_t new_cpus; /* For update_cpumasks_hier() */
192 static inline struct cpuset *css_cs(struct cgroup_subsys_state *css)
194 return css ? container_of(css, struct cpuset, css) : NULL;
197 /* Retrieve the cpuset for a task */
198 static inline struct cpuset *task_cs(struct task_struct *task)
200 return css_cs(task_css(task, cpuset_cgrp_id));
203 static inline struct cpuset *parent_cs(struct cpuset *cs)
205 return css_cs(cs->css.parent);
208 /* bits in struct cpuset flags field */
215 CS_SCHED_LOAD_BALANCE,
220 /* convenient tests for these bits */
221 static inline bool is_cpuset_online(struct cpuset *cs)
223 return test_bit(CS_ONLINE, &cs->flags) && !css_is_dying(&cs->css);
226 static inline int is_cpu_exclusive(const struct cpuset *cs)
228 return test_bit(CS_CPU_EXCLUSIVE, &cs->flags);
231 static inline int is_mem_exclusive(const struct cpuset *cs)
233 return test_bit(CS_MEM_EXCLUSIVE, &cs->flags);
236 static inline int is_mem_hardwall(const struct cpuset *cs)
238 return test_bit(CS_MEM_HARDWALL, &cs->flags);
241 static inline int is_sched_load_balance(const struct cpuset *cs)
243 return test_bit(CS_SCHED_LOAD_BALANCE, &cs->flags);
246 static inline int is_memory_migrate(const struct cpuset *cs)
248 return test_bit(CS_MEMORY_MIGRATE, &cs->flags);
251 static inline int is_spread_page(const struct cpuset *cs)
253 return test_bit(CS_SPREAD_PAGE, &cs->flags);
256 static inline int is_spread_slab(const struct cpuset *cs)
258 return test_bit(CS_SPREAD_SLAB, &cs->flags);
261 static inline int is_partition_root(const struct cpuset *cs)
263 return cs->partition_root_state > 0;
266 static struct cpuset top_cpuset = {
267 .flags = ((1 << CS_ONLINE) | (1 << CS_CPU_EXCLUSIVE) |
268 (1 << CS_MEM_EXCLUSIVE)),
269 .partition_root_state = PRS_ENABLED,
273 * cpuset_for_each_child - traverse online children of a cpuset
274 * @child_cs: loop cursor pointing to the current child
275 * @pos_css: used for iteration
276 * @parent_cs: target cpuset to walk children of
278 * Walk @child_cs through the online children of @parent_cs. Must be used
279 * with RCU read locked.
281 #define cpuset_for_each_child(child_cs, pos_css, parent_cs) \
282 css_for_each_child((pos_css), &(parent_cs)->css) \
283 if (is_cpuset_online(((child_cs) = css_cs((pos_css)))))
286 * cpuset_for_each_descendant_pre - pre-order walk of a cpuset's descendants
287 * @des_cs: loop cursor pointing to the current descendant
288 * @pos_css: used for iteration
289 * @root_cs: target cpuset to walk ancestor of
291 * Walk @des_cs through the online descendants of @root_cs. Must be used
292 * with RCU read locked. The caller may modify @pos_css by calling
293 * css_rightmost_descendant() to skip subtree. @root_cs is included in the
294 * iteration and the first node to be visited.
296 #define cpuset_for_each_descendant_pre(des_cs, pos_css, root_cs) \
297 css_for_each_descendant_pre((pos_css), &(root_cs)->css) \
298 if (is_cpuset_online(((des_cs) = css_cs((pos_css)))))
301 * There are two global locks guarding cpuset structures - cpuset_mutex and
302 * callback_lock. We also require taking task_lock() when dereferencing a
303 * task's cpuset pointer. See "The task_lock() exception", at the end of this
306 * A task must hold both locks to modify cpusets. If a task holds
307 * cpuset_mutex, then it blocks others wanting that mutex, ensuring that it
308 * is the only task able to also acquire callback_lock and be able to
309 * modify cpusets. It can perform various checks on the cpuset structure
310 * first, knowing nothing will change. It can also allocate memory while
311 * just holding cpuset_mutex. While it is performing these checks, various
312 * callback routines can briefly acquire callback_lock to query cpusets.
313 * Once it is ready to make the changes, it takes callback_lock, blocking
316 * Calls to the kernel memory allocator can not be made while holding
317 * callback_lock, as that would risk double tripping on callback_lock
318 * from one of the callbacks into the cpuset code from within
321 * If a task is only holding callback_lock, then it has read-only
324 * Now, the task_struct fields mems_allowed and mempolicy may be changed
325 * by other task, we use alloc_lock in the task_struct fields to protect
328 * The cpuset_common_file_read() handlers only hold callback_lock across
329 * small pieces of code, such as when reading out possibly multi-word
330 * cpumasks and nodemasks.
332 * Accessing a task's cpuset should be done in accordance with the
333 * guidelines for accessing subsystem state in kernel/cgroup.c
336 DEFINE_STATIC_PERCPU_RWSEM(cpuset_rwsem);
338 void cpuset_read_lock(void)
340 percpu_down_read(&cpuset_rwsem);
343 void cpuset_read_unlock(void)
345 percpu_up_read(&cpuset_rwsem);
348 static DEFINE_SPINLOCK(callback_lock);
350 static struct workqueue_struct *cpuset_migrate_mm_wq;
353 * CPU / memory hotplug is handled asynchronously.
355 static void cpuset_hotplug_workfn(struct work_struct *work);
356 static DECLARE_WORK(cpuset_hotplug_work, cpuset_hotplug_workfn);
358 static DECLARE_WAIT_QUEUE_HEAD(cpuset_attach_wq);
361 * Cgroup v2 behavior is used when on default hierarchy or the
362 * cgroup_v2_mode flag is set.
364 static inline bool is_in_v2_mode(void)
366 return cgroup_subsys_on_dfl(cpuset_cgrp_subsys) ||
367 (cpuset_cgrp_subsys.root->flags & CGRP_ROOT_CPUSET_V2_MODE);
371 * Return in pmask the portion of a cpusets's cpus_allowed that
372 * are online. If none are online, walk up the cpuset hierarchy
373 * until we find one that does have some online cpus.
375 * One way or another, we guarantee to return some non-empty subset
376 * of cpu_online_mask.
378 * Call with callback_lock or cpuset_mutex held.
380 static void guarantee_online_cpus(struct cpuset *cs, struct cpumask *pmask)
382 while (!cpumask_intersects(cs->effective_cpus, cpu_online_mask)) {
386 * The top cpuset doesn't have any online cpu as a
387 * consequence of a race between cpuset_hotplug_work
388 * and cpu hotplug notifier. But we know the top
389 * cpuset's effective_cpus is on its way to to be
390 * identical to cpu_online_mask.
392 cpumask_copy(pmask, cpu_online_mask);
396 cpumask_and(pmask, cs->effective_cpus, cpu_online_mask);
400 * Return in *pmask the portion of a cpusets's mems_allowed that
401 * are online, with memory. If none are online with memory, walk
402 * up the cpuset hierarchy until we find one that does have some
403 * online mems. The top cpuset always has some mems online.
405 * One way or another, we guarantee to return some non-empty subset
406 * of node_states[N_MEMORY].
408 * Call with callback_lock or cpuset_mutex held.
410 static void guarantee_online_mems(struct cpuset *cs, nodemask_t *pmask)
412 while (!nodes_intersects(cs->effective_mems, node_states[N_MEMORY]))
414 nodes_and(*pmask, cs->effective_mems, node_states[N_MEMORY]);
418 * update task's spread flag if cpuset's page/slab spread flag is set
420 * Call with callback_lock or cpuset_mutex held.
422 static void cpuset_update_task_spread_flag(struct cpuset *cs,
423 struct task_struct *tsk)
425 if (is_spread_page(cs))
426 task_set_spread_page(tsk);
428 task_clear_spread_page(tsk);
430 if (is_spread_slab(cs))
431 task_set_spread_slab(tsk);
433 task_clear_spread_slab(tsk);
437 * is_cpuset_subset(p, q) - Is cpuset p a subset of cpuset q?
439 * One cpuset is a subset of another if all its allowed CPUs and
440 * Memory Nodes are a subset of the other, and its exclusive flags
441 * are only set if the other's are set. Call holding cpuset_mutex.
444 static int is_cpuset_subset(const struct cpuset *p, const struct cpuset *q)
446 return cpumask_subset(p->cpus_allowed, q->cpus_allowed) &&
447 nodes_subset(p->mems_allowed, q->mems_allowed) &&
448 is_cpu_exclusive(p) <= is_cpu_exclusive(q) &&
449 is_mem_exclusive(p) <= is_mem_exclusive(q);
453 * alloc_cpumasks - allocate three cpumasks for cpuset
454 * @cs: the cpuset that have cpumasks to be allocated.
455 * @tmp: the tmpmasks structure pointer
456 * Return: 0 if successful, -ENOMEM otherwise.
458 * Only one of the two input arguments should be non-NULL.
460 static inline int alloc_cpumasks(struct cpuset *cs, struct tmpmasks *tmp)
462 cpumask_var_t *pmask1, *pmask2, *pmask3;
465 pmask1 = &cs->cpus_allowed;
466 pmask2 = &cs->effective_cpus;
467 pmask3 = &cs->subparts_cpus;
469 pmask1 = &tmp->new_cpus;
470 pmask2 = &tmp->addmask;
471 pmask3 = &tmp->delmask;
474 if (!zalloc_cpumask_var(pmask1, GFP_KERNEL))
477 if (!zalloc_cpumask_var(pmask2, GFP_KERNEL))
480 if (!zalloc_cpumask_var(pmask3, GFP_KERNEL))
486 free_cpumask_var(*pmask2);
488 free_cpumask_var(*pmask1);
493 * free_cpumasks - free cpumasks in a tmpmasks structure
494 * @cs: the cpuset that have cpumasks to be free.
495 * @tmp: the tmpmasks structure pointer
497 static inline void free_cpumasks(struct cpuset *cs, struct tmpmasks *tmp)
500 free_cpumask_var(cs->cpus_allowed);
501 free_cpumask_var(cs->effective_cpus);
502 free_cpumask_var(cs->subparts_cpus);
505 free_cpumask_var(tmp->new_cpus);
506 free_cpumask_var(tmp->addmask);
507 free_cpumask_var(tmp->delmask);
512 * alloc_trial_cpuset - allocate a trial cpuset
513 * @cs: the cpuset that the trial cpuset duplicates
515 static struct cpuset *alloc_trial_cpuset(struct cpuset *cs)
517 struct cpuset *trial;
519 trial = kmemdup(cs, sizeof(*cs), GFP_KERNEL);
523 if (alloc_cpumasks(trial, NULL)) {
528 cpumask_copy(trial->cpus_allowed, cs->cpus_allowed);
529 cpumask_copy(trial->effective_cpus, cs->effective_cpus);
534 * free_cpuset - free the cpuset
535 * @cs: the cpuset to be freed
537 static inline void free_cpuset(struct cpuset *cs)
539 free_cpumasks(cs, NULL);
544 * validate_change() - Used to validate that any proposed cpuset change
545 * follows the structural rules for cpusets.
547 * If we replaced the flag and mask values of the current cpuset
548 * (cur) with those values in the trial cpuset (trial), would
549 * our various subset and exclusive rules still be valid? Presumes
552 * 'cur' is the address of an actual, in-use cpuset. Operations
553 * such as list traversal that depend on the actual address of the
554 * cpuset in the list must use cur below, not trial.
556 * 'trial' is the address of bulk structure copy of cur, with
557 * perhaps one or more of the fields cpus_allowed, mems_allowed,
558 * or flags changed to new, trial values.
560 * Return 0 if valid, -errno if not.
563 static int validate_change(struct cpuset *cur, struct cpuset *trial)
565 struct cgroup_subsys_state *css;
566 struct cpuset *c, *par;
571 /* Each of our child cpusets must be a subset of us */
573 cpuset_for_each_child(c, css, cur)
574 if (!is_cpuset_subset(c, trial))
577 /* Remaining checks don't apply to root cpuset */
579 if (cur == &top_cpuset)
582 par = parent_cs(cur);
584 /* On legacy hiearchy, we must be a subset of our parent cpuset. */
586 if (!is_in_v2_mode() && !is_cpuset_subset(trial, par))
590 * If either I or some sibling (!= me) is exclusive, we can't
594 cpuset_for_each_child(c, css, par) {
595 if ((is_cpu_exclusive(trial) || is_cpu_exclusive(c)) &&
597 cpumask_intersects(trial->cpus_allowed, c->cpus_allowed))
599 if ((is_mem_exclusive(trial) || is_mem_exclusive(c)) &&
601 nodes_intersects(trial->mems_allowed, c->mems_allowed))
606 * Cpusets with tasks - existing or newly being attached - can't
607 * be changed to have empty cpus_allowed or mems_allowed.
610 if ((cgroup_is_populated(cur->css.cgroup) || cur->attach_in_progress)) {
611 if (!cpumask_empty(cur->cpus_allowed) &&
612 cpumask_empty(trial->cpus_allowed))
614 if (!nodes_empty(cur->mems_allowed) &&
615 nodes_empty(trial->mems_allowed))
620 * We can't shrink if we won't have enough room for SCHED_DEADLINE
624 if (is_cpu_exclusive(cur) &&
625 !cpuset_cpumask_can_shrink(cur->cpus_allowed,
626 trial->cpus_allowed))
637 * Helper routine for generate_sched_domains().
638 * Do cpusets a, b have overlapping effective cpus_allowed masks?
640 static int cpusets_overlap(struct cpuset *a, struct cpuset *b)
642 return cpumask_intersects(a->effective_cpus, b->effective_cpus);
646 update_domain_attr(struct sched_domain_attr *dattr, struct cpuset *c)
648 if (dattr->relax_domain_level < c->relax_domain_level)
649 dattr->relax_domain_level = c->relax_domain_level;
653 static void update_domain_attr_tree(struct sched_domain_attr *dattr,
654 struct cpuset *root_cs)
657 struct cgroup_subsys_state *pos_css;
660 cpuset_for_each_descendant_pre(cp, pos_css, root_cs) {
661 /* skip the whole subtree if @cp doesn't have any CPU */
662 if (cpumask_empty(cp->cpus_allowed)) {
663 pos_css = css_rightmost_descendant(pos_css);
667 if (is_sched_load_balance(cp))
668 update_domain_attr(dattr, cp);
673 /* Must be called with cpuset_mutex held. */
674 static inline int nr_cpusets(void)
676 /* jump label reference count + the top-level cpuset */
677 return static_key_count(&cpusets_enabled_key.key) + 1;
681 * generate_sched_domains()
683 * This function builds a partial partition of the systems CPUs
684 * A 'partial partition' is a set of non-overlapping subsets whose
685 * union is a subset of that set.
686 * The output of this function needs to be passed to kernel/sched/core.c
687 * partition_sched_domains() routine, which will rebuild the scheduler's
688 * load balancing domains (sched domains) as specified by that partial
691 * See "What is sched_load_balance" in Documentation/admin-guide/cgroup-v1/cpusets.rst
692 * for a background explanation of this.
694 * Does not return errors, on the theory that the callers of this
695 * routine would rather not worry about failures to rebuild sched
696 * domains when operating in the severe memory shortage situations
697 * that could cause allocation failures below.
699 * Must be called with cpuset_mutex held.
701 * The three key local variables below are:
702 * cp - cpuset pointer, used (together with pos_css) to perform a
703 * top-down scan of all cpusets. For our purposes, rebuilding
704 * the schedulers sched domains, we can ignore !is_sched_load_
706 * csa - (for CpuSet Array) Array of pointers to all the cpusets
707 * that need to be load balanced, for convenient iterative
708 * access by the subsequent code that finds the best partition,
709 * i.e the set of domains (subsets) of CPUs such that the
710 * cpus_allowed of every cpuset marked is_sched_load_balance
711 * is a subset of one of these domains, while there are as
712 * many such domains as possible, each as small as possible.
713 * doms - Conversion of 'csa' to an array of cpumasks, for passing to
714 * the kernel/sched/core.c routine partition_sched_domains() in a
715 * convenient format, that can be easily compared to the prior
716 * value to determine what partition elements (sched domains)
717 * were changed (added or removed.)
719 * Finding the best partition (set of domains):
720 * The triple nested loops below over i, j, k scan over the
721 * load balanced cpusets (using the array of cpuset pointers in
722 * csa[]) looking for pairs of cpusets that have overlapping
723 * cpus_allowed, but which don't have the same 'pn' partition
724 * number and gives them in the same partition number. It keeps
725 * looping on the 'restart' label until it can no longer find
728 * The union of the cpus_allowed masks from the set of
729 * all cpusets having the same 'pn' value then form the one
730 * element of the partition (one sched domain) to be passed to
731 * partition_sched_domains().
733 static int generate_sched_domains(cpumask_var_t **domains,
734 struct sched_domain_attr **attributes)
736 struct cpuset *cp; /* top-down scan of cpusets */
737 struct cpuset **csa; /* array of all cpuset ptrs */
738 int csn; /* how many cpuset ptrs in csa so far */
739 int i, j, k; /* indices for partition finding loops */
740 cpumask_var_t *doms; /* resulting partition; i.e. sched domains */
741 struct sched_domain_attr *dattr; /* attributes for custom domains */
742 int ndoms = 0; /* number of sched domains in result */
743 int nslot; /* next empty doms[] struct cpumask slot */
744 struct cgroup_subsys_state *pos_css;
745 bool root_load_balance = is_sched_load_balance(&top_cpuset);
751 /* Special case for the 99% of systems with one, full, sched domain */
752 if (root_load_balance && !top_cpuset.nr_subparts_cpus) {
754 doms = alloc_sched_domains(ndoms);
758 dattr = kmalloc(sizeof(struct sched_domain_attr), GFP_KERNEL);
760 *dattr = SD_ATTR_INIT;
761 update_domain_attr_tree(dattr, &top_cpuset);
763 cpumask_and(doms[0], top_cpuset.effective_cpus,
764 housekeeping_cpumask(HK_FLAG_DOMAIN));
769 csa = kmalloc_array(nr_cpusets(), sizeof(cp), GFP_KERNEL);
775 if (root_load_balance)
776 csa[csn++] = &top_cpuset;
777 cpuset_for_each_descendant_pre(cp, pos_css, &top_cpuset) {
778 if (cp == &top_cpuset)
781 * Continue traversing beyond @cp iff @cp has some CPUs and
782 * isn't load balancing. The former is obvious. The
783 * latter: All child cpusets contain a subset of the
784 * parent's cpus, so just skip them, and then we call
785 * update_domain_attr_tree() to calc relax_domain_level of
786 * the corresponding sched domain.
788 * If root is load-balancing, we can skip @cp if it
789 * is a subset of the root's effective_cpus.
791 if (!cpumask_empty(cp->cpus_allowed) &&
792 !(is_sched_load_balance(cp) &&
793 cpumask_intersects(cp->cpus_allowed,
794 housekeeping_cpumask(HK_FLAG_DOMAIN))))
797 if (root_load_balance &&
798 cpumask_subset(cp->cpus_allowed, top_cpuset.effective_cpus))
801 if (is_sched_load_balance(cp) &&
802 !cpumask_empty(cp->effective_cpus))
805 /* skip @cp's subtree if not a partition root */
806 if (!is_partition_root(cp))
807 pos_css = css_rightmost_descendant(pos_css);
811 for (i = 0; i < csn; i++)
816 /* Find the best partition (set of sched domains) */
817 for (i = 0; i < csn; i++) {
818 struct cpuset *a = csa[i];
821 for (j = 0; j < csn; j++) {
822 struct cpuset *b = csa[j];
825 if (apn != bpn && cpusets_overlap(a, b)) {
826 for (k = 0; k < csn; k++) {
827 struct cpuset *c = csa[k];
832 ndoms--; /* one less element */
839 * Now we know how many domains to create.
840 * Convert <csn, csa> to <ndoms, doms> and populate cpu masks.
842 doms = alloc_sched_domains(ndoms);
847 * The rest of the code, including the scheduler, can deal with
848 * dattr==NULL case. No need to abort if alloc fails.
850 dattr = kmalloc_array(ndoms, sizeof(struct sched_domain_attr),
853 for (nslot = 0, i = 0; i < csn; i++) {
854 struct cpuset *a = csa[i];
859 /* Skip completed partitions */
865 if (nslot == ndoms) {
866 static int warnings = 10;
868 pr_warn("rebuild_sched_domains confused: nslot %d, ndoms %d, csn %d, i %d, apn %d\n",
869 nslot, ndoms, csn, i, apn);
877 *(dattr + nslot) = SD_ATTR_INIT;
878 for (j = i; j < csn; j++) {
879 struct cpuset *b = csa[j];
882 cpumask_or(dp, dp, b->effective_cpus);
883 cpumask_and(dp, dp, housekeeping_cpumask(HK_FLAG_DOMAIN));
885 update_domain_attr_tree(dattr + nslot, b);
887 /* Done with this partition */
893 BUG_ON(nslot != ndoms);
899 * Fallback to the default domain if kmalloc() failed.
900 * See comments in partition_sched_domains().
910 static void update_tasks_root_domain(struct cpuset *cs)
912 struct css_task_iter it;
913 struct task_struct *task;
915 css_task_iter_start(&cs->css, 0, &it);
917 while ((task = css_task_iter_next(&it)))
918 dl_add_task_root_domain(task);
920 css_task_iter_end(&it);
923 static void rebuild_root_domains(void)
925 struct cpuset *cs = NULL;
926 struct cgroup_subsys_state *pos_css;
928 percpu_rwsem_assert_held(&cpuset_rwsem);
929 lockdep_assert_cpus_held();
930 lockdep_assert_held(&sched_domains_mutex);
935 * Clear default root domain DL accounting, it will be computed again
936 * if a task belongs to it.
938 dl_clear_root_domain(&def_root_domain);
940 cpuset_for_each_descendant_pre(cs, pos_css, &top_cpuset) {
942 if (cpumask_empty(cs->effective_cpus)) {
943 pos_css = css_rightmost_descendant(pos_css);
951 update_tasks_root_domain(cs);
960 partition_and_rebuild_sched_domains(int ndoms_new, cpumask_var_t doms_new[],
961 struct sched_domain_attr *dattr_new)
963 mutex_lock(&sched_domains_mutex);
964 partition_sched_domains_locked(ndoms_new, doms_new, dattr_new);
965 rebuild_root_domains();
966 mutex_unlock(&sched_domains_mutex);
970 * Rebuild scheduler domains.
972 * If the flag 'sched_load_balance' of any cpuset with non-empty
973 * 'cpus' changes, or if the 'cpus' allowed changes in any cpuset
974 * which has that flag enabled, or if any cpuset with a non-empty
975 * 'cpus' is removed, then call this routine to rebuild the
976 * scheduler's dynamic sched domains.
978 * Call with cpuset_mutex held. Takes get_online_cpus().
980 static void rebuild_sched_domains_locked(void)
982 struct sched_domain_attr *attr;
986 lockdep_assert_cpus_held();
987 percpu_rwsem_assert_held(&cpuset_rwsem);
990 * We have raced with CPU hotplug. Don't do anything to avoid
991 * passing doms with offlined cpu to partition_sched_domains().
992 * Anyways, hotplug work item will rebuild sched domains.
994 if (!top_cpuset.nr_subparts_cpus &&
995 !cpumask_equal(top_cpuset.effective_cpus, cpu_active_mask))
998 if (top_cpuset.nr_subparts_cpus &&
999 !cpumask_subset(top_cpuset.effective_cpus, cpu_active_mask))
1002 /* Generate domain masks and attrs */
1003 ndoms = generate_sched_domains(&doms, &attr);
1005 /* Have scheduler rebuild the domains */
1006 partition_and_rebuild_sched_domains(ndoms, doms, attr);
1008 #else /* !CONFIG_SMP */
1009 static void rebuild_sched_domains_locked(void)
1012 #endif /* CONFIG_SMP */
1014 void rebuild_sched_domains(void)
1017 percpu_down_write(&cpuset_rwsem);
1018 rebuild_sched_domains_locked();
1019 percpu_up_write(&cpuset_rwsem);
1024 * update_tasks_cpumask - Update the cpumasks of tasks in the cpuset.
1025 * @cs: the cpuset in which each task's cpus_allowed mask needs to be changed
1027 * Iterate through each task of @cs updating its cpus_allowed to the
1028 * effective cpuset's. As this function is called with cpuset_mutex held,
1029 * cpuset membership stays stable.
1031 static void update_tasks_cpumask(struct cpuset *cs)
1033 struct css_task_iter it;
1034 struct task_struct *task;
1036 css_task_iter_start(&cs->css, 0, &it);
1037 while ((task = css_task_iter_next(&it)))
1038 set_cpus_allowed_ptr(task, cs->effective_cpus);
1039 css_task_iter_end(&it);
1043 * compute_effective_cpumask - Compute the effective cpumask of the cpuset
1044 * @new_cpus: the temp variable for the new effective_cpus mask
1045 * @cs: the cpuset the need to recompute the new effective_cpus mask
1046 * @parent: the parent cpuset
1048 * If the parent has subpartition CPUs, include them in the list of
1049 * allowable CPUs in computing the new effective_cpus mask. Since offlined
1050 * CPUs are not removed from subparts_cpus, we have to use cpu_active_mask
1051 * to mask those out.
1053 static void compute_effective_cpumask(struct cpumask *new_cpus,
1054 struct cpuset *cs, struct cpuset *parent)
1056 if (parent->nr_subparts_cpus) {
1057 cpumask_or(new_cpus, parent->effective_cpus,
1058 parent->subparts_cpus);
1059 cpumask_and(new_cpus, new_cpus, cs->cpus_allowed);
1060 cpumask_and(new_cpus, new_cpus, cpu_active_mask);
1062 cpumask_and(new_cpus, cs->cpus_allowed, parent->effective_cpus);
1067 * Commands for update_parent_subparts_cpumask
1070 partcmd_enable, /* Enable partition root */
1071 partcmd_disable, /* Disable partition root */
1072 partcmd_update, /* Update parent's subparts_cpus */
1076 * update_parent_subparts_cpumask - update subparts_cpus mask of parent cpuset
1077 * @cpuset: The cpuset that requests change in partition root state
1078 * @cmd: Partition root state change command
1079 * @newmask: Optional new cpumask for partcmd_update
1080 * @tmp: Temporary addmask and delmask
1081 * Return: 0, 1 or an error code
1083 * For partcmd_enable, the cpuset is being transformed from a non-partition
1084 * root to a partition root. The cpus_allowed mask of the given cpuset will
1085 * be put into parent's subparts_cpus and taken away from parent's
1086 * effective_cpus. The function will return 0 if all the CPUs listed in
1087 * cpus_allowed can be granted or an error code will be returned.
1089 * For partcmd_disable, the cpuset is being transofrmed from a partition
1090 * root back to a non-partition root. any CPUs in cpus_allowed that are in
1091 * parent's subparts_cpus will be taken away from that cpumask and put back
1092 * into parent's effective_cpus. 0 should always be returned.
1094 * For partcmd_update, if the optional newmask is specified, the cpu
1095 * list is to be changed from cpus_allowed to newmask. Otherwise,
1096 * cpus_allowed is assumed to remain the same. The cpuset should either
1097 * be a partition root or an invalid partition root. The partition root
1098 * state may change if newmask is NULL and none of the requested CPUs can
1099 * be granted by the parent. The function will return 1 if changes to
1100 * parent's subparts_cpus and effective_cpus happen or 0 otherwise.
1101 * Error code should only be returned when newmask is non-NULL.
1103 * The partcmd_enable and partcmd_disable commands are used by
1104 * update_prstate(). The partcmd_update command is used by
1105 * update_cpumasks_hier() with newmask NULL and update_cpumask() with
1108 * The checking is more strict when enabling partition root than the
1109 * other two commands.
1111 * Because of the implicit cpu exclusive nature of a partition root,
1112 * cpumask changes that violates the cpu exclusivity rule will not be
1113 * permitted when checked by validate_change(). The validate_change()
1114 * function will also prevent any changes to the cpu list if it is not
1115 * a superset of children's cpu lists.
1117 static int update_parent_subparts_cpumask(struct cpuset *cpuset, int cmd,
1118 struct cpumask *newmask,
1119 struct tmpmasks *tmp)
1121 struct cpuset *parent = parent_cs(cpuset);
1122 int adding; /* Moving cpus from effective_cpus to subparts_cpus */
1123 int deleting; /* Moving cpus from subparts_cpus to effective_cpus */
1124 bool part_error = false; /* Partition error? */
1126 percpu_rwsem_assert_held(&cpuset_rwsem);
1129 * The parent must be a partition root.
1130 * The new cpumask, if present, or the current cpus_allowed must
1133 if (!is_partition_root(parent) ||
1134 (newmask && cpumask_empty(newmask)) ||
1135 (!newmask && cpumask_empty(cpuset->cpus_allowed)))
1139 * Enabling/disabling partition root is not allowed if there are
1142 if ((cmd != partcmd_update) && css_has_online_children(&cpuset->css))
1146 * Enabling partition root is not allowed if not all the CPUs
1147 * can be granted from parent's effective_cpus or at least one
1148 * CPU will be left after that.
1150 if ((cmd == partcmd_enable) &&
1151 (!cpumask_subset(cpuset->cpus_allowed, parent->effective_cpus) ||
1152 cpumask_equal(cpuset->cpus_allowed, parent->effective_cpus)))
1156 * A cpumask update cannot make parent's effective_cpus become empty.
1158 adding = deleting = false;
1159 if (cmd == partcmd_enable) {
1160 cpumask_copy(tmp->addmask, cpuset->cpus_allowed);
1162 } else if (cmd == partcmd_disable) {
1163 deleting = cpumask_and(tmp->delmask, cpuset->cpus_allowed,
1164 parent->subparts_cpus);
1165 } else if (newmask) {
1167 * partcmd_update with newmask:
1169 * delmask = cpus_allowed & ~newmask & parent->subparts_cpus
1170 * addmask = newmask & parent->effective_cpus
1171 * & ~parent->subparts_cpus
1173 cpumask_andnot(tmp->delmask, cpuset->cpus_allowed, newmask);
1174 deleting = cpumask_and(tmp->delmask, tmp->delmask,
1175 parent->subparts_cpus);
1177 cpumask_and(tmp->addmask, newmask, parent->effective_cpus);
1178 adding = cpumask_andnot(tmp->addmask, tmp->addmask,
1179 parent->subparts_cpus);
1181 * Return error if the new effective_cpus could become empty.
1184 cpumask_equal(parent->effective_cpus, tmp->addmask)) {
1188 * As some of the CPUs in subparts_cpus might have
1189 * been offlined, we need to compute the real delmask
1192 if (!cpumask_and(tmp->addmask, tmp->delmask,
1195 cpumask_copy(tmp->addmask, parent->effective_cpus);
1199 * partcmd_update w/o newmask:
1201 * addmask = cpus_allowed & parent->effectiveb_cpus
1203 * Note that parent's subparts_cpus may have been
1204 * pre-shrunk in case there is a change in the cpu list.
1205 * So no deletion is needed.
1207 adding = cpumask_and(tmp->addmask, cpuset->cpus_allowed,
1208 parent->effective_cpus);
1209 part_error = cpumask_equal(tmp->addmask,
1210 parent->effective_cpus);
1213 if (cmd == partcmd_update) {
1214 int prev_prs = cpuset->partition_root_state;
1217 * Check for possible transition between PRS_ENABLED
1220 switch (cpuset->partition_root_state) {
1223 cpuset->partition_root_state = PRS_ERROR;
1227 cpuset->partition_root_state = PRS_ENABLED;
1231 * Set part_error if previously in invalid state.
1233 part_error = (prev_prs == PRS_ERROR);
1236 if (!part_error && (cpuset->partition_root_state == PRS_ERROR))
1237 return 0; /* Nothing need to be done */
1239 if (cpuset->partition_root_state == PRS_ERROR) {
1241 * Remove all its cpus from parent's subparts_cpus.
1244 deleting = cpumask_and(tmp->delmask, cpuset->cpus_allowed,
1245 parent->subparts_cpus);
1248 if (!adding && !deleting)
1252 * Change the parent's subparts_cpus.
1253 * Newly added CPUs will be removed from effective_cpus and
1254 * newly deleted ones will be added back to effective_cpus.
1256 spin_lock_irq(&callback_lock);
1258 cpumask_or(parent->subparts_cpus,
1259 parent->subparts_cpus, tmp->addmask);
1260 cpumask_andnot(parent->effective_cpus,
1261 parent->effective_cpus, tmp->addmask);
1264 cpumask_andnot(parent->subparts_cpus,
1265 parent->subparts_cpus, tmp->delmask);
1267 * Some of the CPUs in subparts_cpus might have been offlined.
1269 cpumask_and(tmp->delmask, tmp->delmask, cpu_active_mask);
1270 cpumask_or(parent->effective_cpus,
1271 parent->effective_cpus, tmp->delmask);
1274 parent->nr_subparts_cpus = cpumask_weight(parent->subparts_cpus);
1275 spin_unlock_irq(&callback_lock);
1277 return cmd == partcmd_update;
1281 * update_cpumasks_hier - Update effective cpumasks and tasks in the subtree
1282 * @cs: the cpuset to consider
1283 * @tmp: temp variables for calculating effective_cpus & partition setup
1285 * When congifured cpumask is changed, the effective cpumasks of this cpuset
1286 * and all its descendants need to be updated.
1288 * On legacy hierachy, effective_cpus will be the same with cpu_allowed.
1290 * Called with cpuset_mutex held
1292 static void update_cpumasks_hier(struct cpuset *cs, struct tmpmasks *tmp)
1295 struct cgroup_subsys_state *pos_css;
1296 bool need_rebuild_sched_domains = false;
1299 cpuset_for_each_descendant_pre(cp, pos_css, cs) {
1300 struct cpuset *parent = parent_cs(cp);
1302 compute_effective_cpumask(tmp->new_cpus, cp, parent);
1305 * If it becomes empty, inherit the effective mask of the
1306 * parent, which is guaranteed to have some CPUs.
1308 if (is_in_v2_mode() && cpumask_empty(tmp->new_cpus)) {
1309 cpumask_copy(tmp->new_cpus, parent->effective_cpus);
1310 if (!cp->use_parent_ecpus) {
1311 cp->use_parent_ecpus = true;
1312 parent->child_ecpus_count++;
1314 } else if (cp->use_parent_ecpus) {
1315 cp->use_parent_ecpus = false;
1316 WARN_ON_ONCE(!parent->child_ecpus_count);
1317 parent->child_ecpus_count--;
1321 * Skip the whole subtree if the cpumask remains the same
1322 * and has no partition root state.
1324 if (!cp->partition_root_state &&
1325 cpumask_equal(tmp->new_cpus, cp->effective_cpus)) {
1326 pos_css = css_rightmost_descendant(pos_css);
1331 * update_parent_subparts_cpumask() should have been called
1332 * for cs already in update_cpumask(). We should also call
1333 * update_tasks_cpumask() again for tasks in the parent
1334 * cpuset if the parent's subparts_cpus changes.
1336 if ((cp != cs) && cp->partition_root_state) {
1337 switch (parent->partition_root_state) {
1340 * If parent is not a partition root or an
1341 * invalid partition root, clear the state
1342 * state and the CS_CPU_EXCLUSIVE flag.
1344 WARN_ON_ONCE(cp->partition_root_state
1346 cp->partition_root_state = 0;
1349 * clear_bit() is an atomic operation and
1350 * readers aren't interested in the state
1351 * of CS_CPU_EXCLUSIVE anyway. So we can
1352 * just update the flag without holding
1353 * the callback_lock.
1355 clear_bit(CS_CPU_EXCLUSIVE, &cp->flags);
1359 if (update_parent_subparts_cpumask(cp, partcmd_update, NULL, tmp))
1360 update_tasks_cpumask(parent);
1365 * When parent is invalid, it has to be too.
1367 cp->partition_root_state = PRS_ERROR;
1368 if (cp->nr_subparts_cpus) {
1369 cp->nr_subparts_cpus = 0;
1370 cpumask_clear(cp->subparts_cpus);
1376 if (!css_tryget_online(&cp->css))
1380 spin_lock_irq(&callback_lock);
1382 cpumask_copy(cp->effective_cpus, tmp->new_cpus);
1383 if (cp->nr_subparts_cpus &&
1384 (cp->partition_root_state != PRS_ENABLED)) {
1385 cp->nr_subparts_cpus = 0;
1386 cpumask_clear(cp->subparts_cpus);
1387 } else if (cp->nr_subparts_cpus) {
1389 * Make sure that effective_cpus & subparts_cpus
1390 * are mutually exclusive.
1392 * In the unlikely event that effective_cpus
1393 * becomes empty. we clear cp->nr_subparts_cpus and
1394 * let its child partition roots to compete for
1397 cpumask_andnot(cp->effective_cpus, cp->effective_cpus,
1399 if (cpumask_empty(cp->effective_cpus)) {
1400 cpumask_copy(cp->effective_cpus, tmp->new_cpus);
1401 cpumask_clear(cp->subparts_cpus);
1402 cp->nr_subparts_cpus = 0;
1403 } else if (!cpumask_subset(cp->subparts_cpus,
1405 cpumask_andnot(cp->subparts_cpus,
1406 cp->subparts_cpus, tmp->new_cpus);
1407 cp->nr_subparts_cpus
1408 = cpumask_weight(cp->subparts_cpus);
1411 spin_unlock_irq(&callback_lock);
1413 WARN_ON(!is_in_v2_mode() &&
1414 !cpumask_equal(cp->cpus_allowed, cp->effective_cpus));
1416 update_tasks_cpumask(cp);
1419 * On legacy hierarchy, if the effective cpumask of any non-
1420 * empty cpuset is changed, we need to rebuild sched domains.
1421 * On default hierarchy, the cpuset needs to be a partition
1424 if (!cpumask_empty(cp->cpus_allowed) &&
1425 is_sched_load_balance(cp) &&
1426 (!cgroup_subsys_on_dfl(cpuset_cgrp_subsys) ||
1427 is_partition_root(cp)))
1428 need_rebuild_sched_domains = true;
1435 if (need_rebuild_sched_domains)
1436 rebuild_sched_domains_locked();
1440 * update_sibling_cpumasks - Update siblings cpumasks
1441 * @parent: Parent cpuset
1442 * @cs: Current cpuset
1443 * @tmp: Temp variables
1445 static void update_sibling_cpumasks(struct cpuset *parent, struct cpuset *cs,
1446 struct tmpmasks *tmp)
1448 struct cpuset *sibling;
1449 struct cgroup_subsys_state *pos_css;
1452 * Check all its siblings and call update_cpumasks_hier()
1453 * if their use_parent_ecpus flag is set in order for them
1454 * to use the right effective_cpus value.
1457 cpuset_for_each_child(sibling, pos_css, parent) {
1460 if (!sibling->use_parent_ecpus)
1463 update_cpumasks_hier(sibling, tmp);
1469 * update_cpumask - update the cpus_allowed mask of a cpuset and all tasks in it
1470 * @cs: the cpuset to consider
1471 * @trialcs: trial cpuset
1472 * @buf: buffer of cpu numbers written to this cpuset
1474 static int update_cpumask(struct cpuset *cs, struct cpuset *trialcs,
1478 struct tmpmasks tmp;
1480 /* top_cpuset.cpus_allowed tracks cpu_online_mask; it's read-only */
1481 if (cs == &top_cpuset)
1485 * An empty cpus_allowed is ok only if the cpuset has no tasks.
1486 * Since cpulist_parse() fails on an empty mask, we special case
1487 * that parsing. The validate_change() call ensures that cpusets
1488 * with tasks have cpus.
1491 cpumask_clear(trialcs->cpus_allowed);
1493 retval = cpulist_parse(buf, trialcs->cpus_allowed);
1497 if (!cpumask_subset(trialcs->cpus_allowed,
1498 top_cpuset.cpus_allowed))
1502 /* Nothing to do if the cpus didn't change */
1503 if (cpumask_equal(cs->cpus_allowed, trialcs->cpus_allowed))
1506 retval = validate_change(cs, trialcs);
1510 #ifdef CONFIG_CPUMASK_OFFSTACK
1512 * Use the cpumasks in trialcs for tmpmasks when they are pointers
1513 * to allocated cpumasks.
1515 tmp.addmask = trialcs->subparts_cpus;
1516 tmp.delmask = trialcs->effective_cpus;
1517 tmp.new_cpus = trialcs->cpus_allowed;
1520 if (cs->partition_root_state) {
1521 /* Cpumask of a partition root cannot be empty */
1522 if (cpumask_empty(trialcs->cpus_allowed))
1524 if (update_parent_subparts_cpumask(cs, partcmd_update,
1525 trialcs->cpus_allowed, &tmp) < 0)
1529 spin_lock_irq(&callback_lock);
1530 cpumask_copy(cs->cpus_allowed, trialcs->cpus_allowed);
1533 * Make sure that subparts_cpus is a subset of cpus_allowed.
1535 if (cs->nr_subparts_cpus) {
1536 cpumask_andnot(cs->subparts_cpus, cs->subparts_cpus,
1538 cs->nr_subparts_cpus = cpumask_weight(cs->subparts_cpus);
1540 spin_unlock_irq(&callback_lock);
1542 update_cpumasks_hier(cs, &tmp);
1544 if (cs->partition_root_state) {
1545 struct cpuset *parent = parent_cs(cs);
1548 * For partition root, update the cpumasks of sibling
1549 * cpusets if they use parent's effective_cpus.
1551 if (parent->child_ecpus_count)
1552 update_sibling_cpumasks(parent, cs, &tmp);
1558 * Migrate memory region from one set of nodes to another. This is
1559 * performed asynchronously as it can be called from process migration path
1560 * holding locks involved in process management. All mm migrations are
1561 * performed in the queued order and can be waited for by flushing
1562 * cpuset_migrate_mm_wq.
1565 struct cpuset_migrate_mm_work {
1566 struct work_struct work;
1567 struct mm_struct *mm;
1572 static void cpuset_migrate_mm_workfn(struct work_struct *work)
1574 struct cpuset_migrate_mm_work *mwork =
1575 container_of(work, struct cpuset_migrate_mm_work, work);
1577 /* on a wq worker, no need to worry about %current's mems_allowed */
1578 do_migrate_pages(mwork->mm, &mwork->from, &mwork->to, MPOL_MF_MOVE_ALL);
1583 static void cpuset_migrate_mm(struct mm_struct *mm, const nodemask_t *from,
1584 const nodemask_t *to)
1586 struct cpuset_migrate_mm_work *mwork;
1588 mwork = kzalloc(sizeof(*mwork), GFP_KERNEL);
1591 mwork->from = *from;
1593 INIT_WORK(&mwork->work, cpuset_migrate_mm_workfn);
1594 queue_work(cpuset_migrate_mm_wq, &mwork->work);
1600 static void cpuset_post_attach(void)
1602 flush_workqueue(cpuset_migrate_mm_wq);
1606 * cpuset_change_task_nodemask - change task's mems_allowed and mempolicy
1607 * @tsk: the task to change
1608 * @newmems: new nodes that the task will be set
1610 * We use the mems_allowed_seq seqlock to safely update both tsk->mems_allowed
1611 * and rebind an eventual tasks' mempolicy. If the task is allocating in
1612 * parallel, it might temporarily see an empty intersection, which results in
1613 * a seqlock check and retry before OOM or allocation failure.
1615 static void cpuset_change_task_nodemask(struct task_struct *tsk,
1616 nodemask_t *newmems)
1620 local_irq_disable();
1621 write_seqcount_begin(&tsk->mems_allowed_seq);
1623 nodes_or(tsk->mems_allowed, tsk->mems_allowed, *newmems);
1624 mpol_rebind_task(tsk, newmems);
1625 tsk->mems_allowed = *newmems;
1627 write_seqcount_end(&tsk->mems_allowed_seq);
1633 static void *cpuset_being_rebound;
1636 * update_tasks_nodemask - Update the nodemasks of tasks in the cpuset.
1637 * @cs: the cpuset in which each task's mems_allowed mask needs to be changed
1639 * Iterate through each task of @cs updating its mems_allowed to the
1640 * effective cpuset's. As this function is called with cpuset_mutex held,
1641 * cpuset membership stays stable.
1643 static void update_tasks_nodemask(struct cpuset *cs)
1645 static nodemask_t newmems; /* protected by cpuset_mutex */
1646 struct css_task_iter it;
1647 struct task_struct *task;
1649 cpuset_being_rebound = cs; /* causes mpol_dup() rebind */
1651 guarantee_online_mems(cs, &newmems);
1654 * The mpol_rebind_mm() call takes mmap_sem, which we couldn't
1655 * take while holding tasklist_lock. Forks can happen - the
1656 * mpol_dup() cpuset_being_rebound check will catch such forks,
1657 * and rebind their vma mempolicies too. Because we still hold
1658 * the global cpuset_mutex, we know that no other rebind effort
1659 * will be contending for the global variable cpuset_being_rebound.
1660 * It's ok if we rebind the same mm twice; mpol_rebind_mm()
1661 * is idempotent. Also migrate pages in each mm to new nodes.
1663 css_task_iter_start(&cs->css, 0, &it);
1664 while ((task = css_task_iter_next(&it))) {
1665 struct mm_struct *mm;
1668 cpuset_change_task_nodemask(task, &newmems);
1670 mm = get_task_mm(task);
1674 migrate = is_memory_migrate(cs);
1676 mpol_rebind_mm(mm, &cs->mems_allowed);
1678 cpuset_migrate_mm(mm, &cs->old_mems_allowed, &newmems);
1682 css_task_iter_end(&it);
1685 * All the tasks' nodemasks have been updated, update
1686 * cs->old_mems_allowed.
1688 cs->old_mems_allowed = newmems;
1690 /* We're done rebinding vmas to this cpuset's new mems_allowed. */
1691 cpuset_being_rebound = NULL;
1695 * update_nodemasks_hier - Update effective nodemasks and tasks in the subtree
1696 * @cs: the cpuset to consider
1697 * @new_mems: a temp variable for calculating new effective_mems
1699 * When configured nodemask is changed, the effective nodemasks of this cpuset
1700 * and all its descendants need to be updated.
1702 * On legacy hiearchy, effective_mems will be the same with mems_allowed.
1704 * Called with cpuset_mutex held
1706 static void update_nodemasks_hier(struct cpuset *cs, nodemask_t *new_mems)
1709 struct cgroup_subsys_state *pos_css;
1712 cpuset_for_each_descendant_pre(cp, pos_css, cs) {
1713 struct cpuset *parent = parent_cs(cp);
1715 nodes_and(*new_mems, cp->mems_allowed, parent->effective_mems);
1718 * If it becomes empty, inherit the effective mask of the
1719 * parent, which is guaranteed to have some MEMs.
1721 if (is_in_v2_mode() && nodes_empty(*new_mems))
1722 *new_mems = parent->effective_mems;
1724 /* Skip the whole subtree if the nodemask remains the same. */
1725 if (nodes_equal(*new_mems, cp->effective_mems)) {
1726 pos_css = css_rightmost_descendant(pos_css);
1730 if (!css_tryget_online(&cp->css))
1734 spin_lock_irq(&callback_lock);
1735 cp->effective_mems = *new_mems;
1736 spin_unlock_irq(&callback_lock);
1738 WARN_ON(!is_in_v2_mode() &&
1739 !nodes_equal(cp->mems_allowed, cp->effective_mems));
1741 update_tasks_nodemask(cp);
1750 * Handle user request to change the 'mems' memory placement
1751 * of a cpuset. Needs to validate the request, update the
1752 * cpusets mems_allowed, and for each task in the cpuset,
1753 * update mems_allowed and rebind task's mempolicy and any vma
1754 * mempolicies and if the cpuset is marked 'memory_migrate',
1755 * migrate the tasks pages to the new memory.
1757 * Call with cpuset_mutex held. May take callback_lock during call.
1758 * Will take tasklist_lock, scan tasklist for tasks in cpuset cs,
1759 * lock each such tasks mm->mmap_sem, scan its vma's and rebind
1760 * their mempolicies to the cpusets new mems_allowed.
1762 static int update_nodemask(struct cpuset *cs, struct cpuset *trialcs,
1768 * top_cpuset.mems_allowed tracks node_stats[N_MEMORY];
1771 if (cs == &top_cpuset) {
1777 * An empty mems_allowed is ok iff there are no tasks in the cpuset.
1778 * Since nodelist_parse() fails on an empty mask, we special case
1779 * that parsing. The validate_change() call ensures that cpusets
1780 * with tasks have memory.
1783 nodes_clear(trialcs->mems_allowed);
1785 retval = nodelist_parse(buf, trialcs->mems_allowed);
1789 if (!nodes_subset(trialcs->mems_allowed,
1790 top_cpuset.mems_allowed)) {
1796 if (nodes_equal(cs->mems_allowed, trialcs->mems_allowed)) {
1797 retval = 0; /* Too easy - nothing to do */
1800 retval = validate_change(cs, trialcs);
1804 spin_lock_irq(&callback_lock);
1805 cs->mems_allowed = trialcs->mems_allowed;
1806 spin_unlock_irq(&callback_lock);
1808 /* use trialcs->mems_allowed as a temp variable */
1809 update_nodemasks_hier(cs, &trialcs->mems_allowed);
1814 bool current_cpuset_is_being_rebound(void)
1819 ret = task_cs(current) == cpuset_being_rebound;
1825 static int update_relax_domain_level(struct cpuset *cs, s64 val)
1828 if (val < -1 || val >= sched_domain_level_max)
1832 if (val != cs->relax_domain_level) {
1833 cs->relax_domain_level = val;
1834 if (!cpumask_empty(cs->cpus_allowed) &&
1835 is_sched_load_balance(cs))
1836 rebuild_sched_domains_locked();
1843 * update_tasks_flags - update the spread flags of tasks in the cpuset.
1844 * @cs: the cpuset in which each task's spread flags needs to be changed
1846 * Iterate through each task of @cs updating its spread flags. As this
1847 * function is called with cpuset_mutex held, cpuset membership stays
1850 static void update_tasks_flags(struct cpuset *cs)
1852 struct css_task_iter it;
1853 struct task_struct *task;
1855 css_task_iter_start(&cs->css, 0, &it);
1856 while ((task = css_task_iter_next(&it)))
1857 cpuset_update_task_spread_flag(cs, task);
1858 css_task_iter_end(&it);
1862 * update_flag - read a 0 or a 1 in a file and update associated flag
1863 * bit: the bit to update (see cpuset_flagbits_t)
1864 * cs: the cpuset to update
1865 * turning_on: whether the flag is being set or cleared
1867 * Call with cpuset_mutex held.
1870 static int update_flag(cpuset_flagbits_t bit, struct cpuset *cs,
1873 struct cpuset *trialcs;
1874 int balance_flag_changed;
1875 int spread_flag_changed;
1878 trialcs = alloc_trial_cpuset(cs);
1883 set_bit(bit, &trialcs->flags);
1885 clear_bit(bit, &trialcs->flags);
1887 err = validate_change(cs, trialcs);
1891 balance_flag_changed = (is_sched_load_balance(cs) !=
1892 is_sched_load_balance(trialcs));
1894 spread_flag_changed = ((is_spread_slab(cs) != is_spread_slab(trialcs))
1895 || (is_spread_page(cs) != is_spread_page(trialcs)));
1897 spin_lock_irq(&callback_lock);
1898 cs->flags = trialcs->flags;
1899 spin_unlock_irq(&callback_lock);
1901 if (!cpumask_empty(trialcs->cpus_allowed) && balance_flag_changed)
1902 rebuild_sched_domains_locked();
1904 if (spread_flag_changed)
1905 update_tasks_flags(cs);
1907 free_cpuset(trialcs);
1912 * update_prstate - update partititon_root_state
1913 * cs: the cpuset to update
1914 * val: 0 - disabled, 1 - enabled
1916 * Call with cpuset_mutex held.
1918 static int update_prstate(struct cpuset *cs, int val)
1921 struct cpuset *parent = parent_cs(cs);
1922 struct tmpmasks tmp;
1924 if ((val != 0) && (val != 1))
1926 if (val == cs->partition_root_state)
1930 * Cannot force a partial or invalid partition root to a full
1933 if (val && cs->partition_root_state)
1936 if (alloc_cpumasks(NULL, &tmp))
1940 if (!cs->partition_root_state) {
1942 * Turning on partition root requires setting the
1943 * CS_CPU_EXCLUSIVE bit implicitly as well and cpus_allowed
1946 if (cpumask_empty(cs->cpus_allowed))
1949 err = update_flag(CS_CPU_EXCLUSIVE, cs, 1);
1953 err = update_parent_subparts_cpumask(cs, partcmd_enable,
1956 update_flag(CS_CPU_EXCLUSIVE, cs, 0);
1959 cs->partition_root_state = PRS_ENABLED;
1962 * Turning off partition root will clear the
1963 * CS_CPU_EXCLUSIVE bit.
1965 if (cs->partition_root_state == PRS_ERROR) {
1966 cs->partition_root_state = 0;
1967 update_flag(CS_CPU_EXCLUSIVE, cs, 0);
1972 err = update_parent_subparts_cpumask(cs, partcmd_disable,
1977 cs->partition_root_state = 0;
1979 /* Turning off CS_CPU_EXCLUSIVE will not return error */
1980 update_flag(CS_CPU_EXCLUSIVE, cs, 0);
1984 * Update cpumask of parent's tasks except when it is the top
1985 * cpuset as some system daemons cannot be mapped to other CPUs.
1987 if (parent != &top_cpuset)
1988 update_tasks_cpumask(parent);
1990 if (parent->child_ecpus_count)
1991 update_sibling_cpumasks(parent, cs, &tmp);
1993 rebuild_sched_domains_locked();
1995 free_cpumasks(NULL, &tmp);
2000 * Frequency meter - How fast is some event occurring?
2002 * These routines manage a digitally filtered, constant time based,
2003 * event frequency meter. There are four routines:
2004 * fmeter_init() - initialize a frequency meter.
2005 * fmeter_markevent() - called each time the event happens.
2006 * fmeter_getrate() - returns the recent rate of such events.
2007 * fmeter_update() - internal routine used to update fmeter.
2009 * A common data structure is passed to each of these routines,
2010 * which is used to keep track of the state required to manage the
2011 * frequency meter and its digital filter.
2013 * The filter works on the number of events marked per unit time.
2014 * The filter is single-pole low-pass recursive (IIR). The time unit
2015 * is 1 second. Arithmetic is done using 32-bit integers scaled to
2016 * simulate 3 decimal digits of precision (multiplied by 1000).
2018 * With an FM_COEF of 933, and a time base of 1 second, the filter
2019 * has a half-life of 10 seconds, meaning that if the events quit
2020 * happening, then the rate returned from the fmeter_getrate()
2021 * will be cut in half each 10 seconds, until it converges to zero.
2023 * It is not worth doing a real infinitely recursive filter. If more
2024 * than FM_MAXTICKS ticks have elapsed since the last filter event,
2025 * just compute FM_MAXTICKS ticks worth, by which point the level
2028 * Limit the count of unprocessed events to FM_MAXCNT, so as to avoid
2029 * arithmetic overflow in the fmeter_update() routine.
2031 * Given the simple 32 bit integer arithmetic used, this meter works
2032 * best for reporting rates between one per millisecond (msec) and
2033 * one per 32 (approx) seconds. At constant rates faster than one
2034 * per msec it maxes out at values just under 1,000,000. At constant
2035 * rates between one per msec, and one per second it will stabilize
2036 * to a value N*1000, where N is the rate of events per second.
2037 * At constant rates between one per second and one per 32 seconds,
2038 * it will be choppy, moving up on the seconds that have an event,
2039 * and then decaying until the next event. At rates slower than
2040 * about one in 32 seconds, it decays all the way back to zero between
2044 #define FM_COEF 933 /* coefficient for half-life of 10 secs */
2045 #define FM_MAXTICKS ((u32)99) /* useless computing more ticks than this */
2046 #define FM_MAXCNT 1000000 /* limit cnt to avoid overflow */
2047 #define FM_SCALE 1000 /* faux fixed point scale */
2049 /* Initialize a frequency meter */
2050 static void fmeter_init(struct fmeter *fmp)
2055 spin_lock_init(&fmp->lock);
2058 /* Internal meter update - process cnt events and update value */
2059 static void fmeter_update(struct fmeter *fmp)
2064 now = ktime_get_seconds();
2065 ticks = now - fmp->time;
2070 ticks = min(FM_MAXTICKS, ticks);
2072 fmp->val = (FM_COEF * fmp->val) / FM_SCALE;
2075 fmp->val += ((FM_SCALE - FM_COEF) * fmp->cnt) / FM_SCALE;
2079 /* Process any previous ticks, then bump cnt by one (times scale). */
2080 static void fmeter_markevent(struct fmeter *fmp)
2082 spin_lock(&fmp->lock);
2084 fmp->cnt = min(FM_MAXCNT, fmp->cnt + FM_SCALE);
2085 spin_unlock(&fmp->lock);
2088 /* Process any previous ticks, then return current value. */
2089 static int fmeter_getrate(struct fmeter *fmp)
2093 spin_lock(&fmp->lock);
2096 spin_unlock(&fmp->lock);
2100 static struct cpuset *cpuset_attach_old_cs;
2102 /* Called by cgroups to determine if a cpuset is usable; cpuset_mutex held */
2103 static int cpuset_can_attach(struct cgroup_taskset *tset)
2105 struct cgroup_subsys_state *css;
2107 struct task_struct *task;
2110 /* used later by cpuset_attach() */
2111 cpuset_attach_old_cs = task_cs(cgroup_taskset_first(tset, &css));
2114 percpu_down_write(&cpuset_rwsem);
2116 /* allow moving tasks into an empty cpuset if on default hierarchy */
2118 if (!is_in_v2_mode() &&
2119 (cpumask_empty(cs->cpus_allowed) || nodes_empty(cs->mems_allowed)))
2122 cgroup_taskset_for_each(task, css, tset) {
2123 ret = task_can_attach(task, cs->cpus_allowed);
2126 ret = security_task_setscheduler(task);
2132 * Mark attach is in progress. This makes validate_change() fail
2133 * changes which zero cpus/mems_allowed.
2135 cs->attach_in_progress++;
2138 percpu_up_write(&cpuset_rwsem);
2142 static void cpuset_cancel_attach(struct cgroup_taskset *tset)
2144 struct cgroup_subsys_state *css;
2146 cgroup_taskset_first(tset, &css);
2148 percpu_down_write(&cpuset_rwsem);
2149 css_cs(css)->attach_in_progress--;
2150 percpu_up_write(&cpuset_rwsem);
2154 * Protected by cpuset_mutex. cpus_attach is used only by cpuset_attach()
2155 * but we can't allocate it dynamically there. Define it global and
2156 * allocate from cpuset_init().
2158 static cpumask_var_t cpus_attach;
2160 static void cpuset_attach(struct cgroup_taskset *tset)
2162 /* static buf protected by cpuset_mutex */
2163 static nodemask_t cpuset_attach_nodemask_to;
2164 struct task_struct *task;
2165 struct task_struct *leader;
2166 struct cgroup_subsys_state *css;
2168 struct cpuset *oldcs = cpuset_attach_old_cs;
2170 cgroup_taskset_first(tset, &css);
2173 percpu_down_write(&cpuset_rwsem);
2175 /* prepare for attach */
2176 if (cs == &top_cpuset)
2177 cpumask_copy(cpus_attach, cpu_possible_mask);
2179 guarantee_online_cpus(cs, cpus_attach);
2181 guarantee_online_mems(cs, &cpuset_attach_nodemask_to);
2183 cgroup_taskset_for_each(task, css, tset) {
2185 * can_attach beforehand should guarantee that this doesn't
2186 * fail. TODO: have a better way to handle failure here
2188 WARN_ON_ONCE(set_cpus_allowed_ptr(task, cpus_attach));
2190 cpuset_change_task_nodemask(task, &cpuset_attach_nodemask_to);
2191 cpuset_update_task_spread_flag(cs, task);
2195 * Change mm for all threadgroup leaders. This is expensive and may
2196 * sleep and should be moved outside migration path proper.
2198 cpuset_attach_nodemask_to = cs->effective_mems;
2199 cgroup_taskset_for_each_leader(leader, css, tset) {
2200 struct mm_struct *mm = get_task_mm(leader);
2203 mpol_rebind_mm(mm, &cpuset_attach_nodemask_to);
2206 * old_mems_allowed is the same with mems_allowed
2207 * here, except if this task is being moved
2208 * automatically due to hotplug. In that case
2209 * @mems_allowed has been updated and is empty, so
2210 * @old_mems_allowed is the right nodesets that we
2213 if (is_memory_migrate(cs))
2214 cpuset_migrate_mm(mm, &oldcs->old_mems_allowed,
2215 &cpuset_attach_nodemask_to);
2221 cs->old_mems_allowed = cpuset_attach_nodemask_to;
2223 cs->attach_in_progress--;
2224 if (!cs->attach_in_progress)
2225 wake_up(&cpuset_attach_wq);
2227 percpu_up_write(&cpuset_rwsem);
2230 /* The various types of files and directories in a cpuset file system */
2233 FILE_MEMORY_MIGRATE,
2236 FILE_EFFECTIVE_CPULIST,
2237 FILE_EFFECTIVE_MEMLIST,
2238 FILE_SUBPARTS_CPULIST,
2242 FILE_SCHED_LOAD_BALANCE,
2243 FILE_PARTITION_ROOT,
2244 FILE_SCHED_RELAX_DOMAIN_LEVEL,
2245 FILE_MEMORY_PRESSURE_ENABLED,
2246 FILE_MEMORY_PRESSURE,
2249 } cpuset_filetype_t;
2251 static int cpuset_write_u64(struct cgroup_subsys_state *css, struct cftype *cft,
2254 struct cpuset *cs = css_cs(css);
2255 cpuset_filetype_t type = cft->private;
2259 percpu_down_write(&cpuset_rwsem);
2260 if (!is_cpuset_online(cs)) {
2266 case FILE_CPU_EXCLUSIVE:
2267 retval = update_flag(CS_CPU_EXCLUSIVE, cs, val);
2269 case FILE_MEM_EXCLUSIVE:
2270 retval = update_flag(CS_MEM_EXCLUSIVE, cs, val);
2272 case FILE_MEM_HARDWALL:
2273 retval = update_flag(CS_MEM_HARDWALL, cs, val);
2275 case FILE_SCHED_LOAD_BALANCE:
2276 retval = update_flag(CS_SCHED_LOAD_BALANCE, cs, val);
2278 case FILE_MEMORY_MIGRATE:
2279 retval = update_flag(CS_MEMORY_MIGRATE, cs, val);
2281 case FILE_MEMORY_PRESSURE_ENABLED:
2282 cpuset_memory_pressure_enabled = !!val;
2284 case FILE_SPREAD_PAGE:
2285 retval = update_flag(CS_SPREAD_PAGE, cs, val);
2287 case FILE_SPREAD_SLAB:
2288 retval = update_flag(CS_SPREAD_SLAB, cs, val);
2295 percpu_up_write(&cpuset_rwsem);
2300 static int cpuset_write_s64(struct cgroup_subsys_state *css, struct cftype *cft,
2303 struct cpuset *cs = css_cs(css);
2304 cpuset_filetype_t type = cft->private;
2305 int retval = -ENODEV;
2308 percpu_down_write(&cpuset_rwsem);
2309 if (!is_cpuset_online(cs))
2313 case FILE_SCHED_RELAX_DOMAIN_LEVEL:
2314 retval = update_relax_domain_level(cs, val);
2321 percpu_up_write(&cpuset_rwsem);
2327 * Common handling for a write to a "cpus" or "mems" file.
2329 static ssize_t cpuset_write_resmask(struct kernfs_open_file *of,
2330 char *buf, size_t nbytes, loff_t off)
2332 struct cpuset *cs = css_cs(of_css(of));
2333 struct cpuset *trialcs;
2334 int retval = -ENODEV;
2336 buf = strstrip(buf);
2339 * CPU or memory hotunplug may leave @cs w/o any execution
2340 * resources, in which case the hotplug code asynchronously updates
2341 * configuration and transfers all tasks to the nearest ancestor
2342 * which can execute.
2344 * As writes to "cpus" or "mems" may restore @cs's execution
2345 * resources, wait for the previously scheduled operations before
2346 * proceeding, so that we don't end up keep removing tasks added
2347 * after execution capability is restored.
2349 * cpuset_hotplug_work calls back into cgroup core via
2350 * cgroup_transfer_tasks() and waiting for it from a cgroupfs
2351 * operation like this one can lead to a deadlock through kernfs
2352 * active_ref protection. Let's break the protection. Losing the
2353 * protection is okay as we check whether @cs is online after
2354 * grabbing cpuset_mutex anyway. This only happens on the legacy
2358 kernfs_break_active_protection(of->kn);
2359 flush_work(&cpuset_hotplug_work);
2362 percpu_down_write(&cpuset_rwsem);
2363 if (!is_cpuset_online(cs))
2366 trialcs = alloc_trial_cpuset(cs);
2372 switch (of_cft(of)->private) {
2374 retval = update_cpumask(cs, trialcs, buf);
2377 retval = update_nodemask(cs, trialcs, buf);
2384 free_cpuset(trialcs);
2386 percpu_up_write(&cpuset_rwsem);
2388 kernfs_unbreak_active_protection(of->kn);
2390 flush_workqueue(cpuset_migrate_mm_wq);
2391 return retval ?: nbytes;
2395 * These ascii lists should be read in a single call, by using a user
2396 * buffer large enough to hold the entire map. If read in smaller
2397 * chunks, there is no guarantee of atomicity. Since the display format
2398 * used, list of ranges of sequential numbers, is variable length,
2399 * and since these maps can change value dynamically, one could read
2400 * gibberish by doing partial reads while a list was changing.
2402 static int cpuset_common_seq_show(struct seq_file *sf, void *v)
2404 struct cpuset *cs = css_cs(seq_css(sf));
2405 cpuset_filetype_t type = seq_cft(sf)->private;
2408 spin_lock_irq(&callback_lock);
2412 seq_printf(sf, "%*pbl\n", cpumask_pr_args(cs->cpus_allowed));
2415 seq_printf(sf, "%*pbl\n", nodemask_pr_args(&cs->mems_allowed));
2417 case FILE_EFFECTIVE_CPULIST:
2418 seq_printf(sf, "%*pbl\n", cpumask_pr_args(cs->effective_cpus));
2420 case FILE_EFFECTIVE_MEMLIST:
2421 seq_printf(sf, "%*pbl\n", nodemask_pr_args(&cs->effective_mems));
2423 case FILE_SUBPARTS_CPULIST:
2424 seq_printf(sf, "%*pbl\n", cpumask_pr_args(cs->subparts_cpus));
2430 spin_unlock_irq(&callback_lock);
2434 static u64 cpuset_read_u64(struct cgroup_subsys_state *css, struct cftype *cft)
2436 struct cpuset *cs = css_cs(css);
2437 cpuset_filetype_t type = cft->private;
2439 case FILE_CPU_EXCLUSIVE:
2440 return is_cpu_exclusive(cs);
2441 case FILE_MEM_EXCLUSIVE:
2442 return is_mem_exclusive(cs);
2443 case FILE_MEM_HARDWALL:
2444 return is_mem_hardwall(cs);
2445 case FILE_SCHED_LOAD_BALANCE:
2446 return is_sched_load_balance(cs);
2447 case FILE_MEMORY_MIGRATE:
2448 return is_memory_migrate(cs);
2449 case FILE_MEMORY_PRESSURE_ENABLED:
2450 return cpuset_memory_pressure_enabled;
2451 case FILE_MEMORY_PRESSURE:
2452 return fmeter_getrate(&cs->fmeter);
2453 case FILE_SPREAD_PAGE:
2454 return is_spread_page(cs);
2455 case FILE_SPREAD_SLAB:
2456 return is_spread_slab(cs);
2461 /* Unreachable but makes gcc happy */
2465 static s64 cpuset_read_s64(struct cgroup_subsys_state *css, struct cftype *cft)
2467 struct cpuset *cs = css_cs(css);
2468 cpuset_filetype_t type = cft->private;
2470 case FILE_SCHED_RELAX_DOMAIN_LEVEL:
2471 return cs->relax_domain_level;
2476 /* Unrechable but makes gcc happy */
2480 static int sched_partition_show(struct seq_file *seq, void *v)
2482 struct cpuset *cs = css_cs(seq_css(seq));
2484 switch (cs->partition_root_state) {
2486 seq_puts(seq, "root\n");
2489 seq_puts(seq, "member\n");
2492 seq_puts(seq, "root invalid\n");
2498 static ssize_t sched_partition_write(struct kernfs_open_file *of, char *buf,
2499 size_t nbytes, loff_t off)
2501 struct cpuset *cs = css_cs(of_css(of));
2503 int retval = -ENODEV;
2505 buf = strstrip(buf);
2508 * Convert "root" to ENABLED, and convert "member" to DISABLED.
2510 if (!strcmp(buf, "root"))
2512 else if (!strcmp(buf, "member"))
2519 percpu_down_write(&cpuset_rwsem);
2520 if (!is_cpuset_online(cs))
2523 retval = update_prstate(cs, val);
2525 percpu_up_write(&cpuset_rwsem);
2528 return retval ?: nbytes;
2532 * for the common functions, 'private' gives the type of file
2535 static struct cftype legacy_files[] = {
2538 .seq_show = cpuset_common_seq_show,
2539 .write = cpuset_write_resmask,
2540 .max_write_len = (100U + 6 * NR_CPUS),
2541 .private = FILE_CPULIST,
2546 .seq_show = cpuset_common_seq_show,
2547 .write = cpuset_write_resmask,
2548 .max_write_len = (100U + 6 * MAX_NUMNODES),
2549 .private = FILE_MEMLIST,
2553 .name = "effective_cpus",
2554 .seq_show = cpuset_common_seq_show,
2555 .private = FILE_EFFECTIVE_CPULIST,
2559 .name = "effective_mems",
2560 .seq_show = cpuset_common_seq_show,
2561 .private = FILE_EFFECTIVE_MEMLIST,
2565 .name = "cpu_exclusive",
2566 .read_u64 = cpuset_read_u64,
2567 .write_u64 = cpuset_write_u64,
2568 .private = FILE_CPU_EXCLUSIVE,
2572 .name = "mem_exclusive",
2573 .read_u64 = cpuset_read_u64,
2574 .write_u64 = cpuset_write_u64,
2575 .private = FILE_MEM_EXCLUSIVE,
2579 .name = "mem_hardwall",
2580 .read_u64 = cpuset_read_u64,
2581 .write_u64 = cpuset_write_u64,
2582 .private = FILE_MEM_HARDWALL,
2586 .name = "sched_load_balance",
2587 .read_u64 = cpuset_read_u64,
2588 .write_u64 = cpuset_write_u64,
2589 .private = FILE_SCHED_LOAD_BALANCE,
2593 .name = "sched_relax_domain_level",
2594 .read_s64 = cpuset_read_s64,
2595 .write_s64 = cpuset_write_s64,
2596 .private = FILE_SCHED_RELAX_DOMAIN_LEVEL,
2600 .name = "memory_migrate",
2601 .read_u64 = cpuset_read_u64,
2602 .write_u64 = cpuset_write_u64,
2603 .private = FILE_MEMORY_MIGRATE,
2607 .name = "memory_pressure",
2608 .read_u64 = cpuset_read_u64,
2609 .private = FILE_MEMORY_PRESSURE,
2613 .name = "memory_spread_page",
2614 .read_u64 = cpuset_read_u64,
2615 .write_u64 = cpuset_write_u64,
2616 .private = FILE_SPREAD_PAGE,
2620 .name = "memory_spread_slab",
2621 .read_u64 = cpuset_read_u64,
2622 .write_u64 = cpuset_write_u64,
2623 .private = FILE_SPREAD_SLAB,
2627 .name = "memory_pressure_enabled",
2628 .flags = CFTYPE_ONLY_ON_ROOT,
2629 .read_u64 = cpuset_read_u64,
2630 .write_u64 = cpuset_write_u64,
2631 .private = FILE_MEMORY_PRESSURE_ENABLED,
2638 * This is currently a minimal set for the default hierarchy. It can be
2639 * expanded later on by migrating more features and control files from v1.
2641 static struct cftype dfl_files[] = {
2644 .seq_show = cpuset_common_seq_show,
2645 .write = cpuset_write_resmask,
2646 .max_write_len = (100U + 6 * NR_CPUS),
2647 .private = FILE_CPULIST,
2648 .flags = CFTYPE_NOT_ON_ROOT,
2653 .seq_show = cpuset_common_seq_show,
2654 .write = cpuset_write_resmask,
2655 .max_write_len = (100U + 6 * MAX_NUMNODES),
2656 .private = FILE_MEMLIST,
2657 .flags = CFTYPE_NOT_ON_ROOT,
2661 .name = "cpus.effective",
2662 .seq_show = cpuset_common_seq_show,
2663 .private = FILE_EFFECTIVE_CPULIST,
2667 .name = "mems.effective",
2668 .seq_show = cpuset_common_seq_show,
2669 .private = FILE_EFFECTIVE_MEMLIST,
2673 .name = "cpus.partition",
2674 .seq_show = sched_partition_show,
2675 .write = sched_partition_write,
2676 .private = FILE_PARTITION_ROOT,
2677 .flags = CFTYPE_NOT_ON_ROOT,
2681 .name = "cpus.subpartitions",
2682 .seq_show = cpuset_common_seq_show,
2683 .private = FILE_SUBPARTS_CPULIST,
2684 .flags = CFTYPE_DEBUG,
2692 * cpuset_css_alloc - allocate a cpuset css
2693 * cgrp: control group that the new cpuset will be part of
2696 static struct cgroup_subsys_state *
2697 cpuset_css_alloc(struct cgroup_subsys_state *parent_css)
2702 return &top_cpuset.css;
2704 cs = kzalloc(sizeof(*cs), GFP_KERNEL);
2706 return ERR_PTR(-ENOMEM);
2708 if (alloc_cpumasks(cs, NULL)) {
2710 return ERR_PTR(-ENOMEM);
2713 set_bit(CS_SCHED_LOAD_BALANCE, &cs->flags);
2714 nodes_clear(cs->mems_allowed);
2715 nodes_clear(cs->effective_mems);
2716 fmeter_init(&cs->fmeter);
2717 cs->relax_domain_level = -1;
2722 static int cpuset_css_online(struct cgroup_subsys_state *css)
2724 struct cpuset *cs = css_cs(css);
2725 struct cpuset *parent = parent_cs(cs);
2726 struct cpuset *tmp_cs;
2727 struct cgroup_subsys_state *pos_css;
2733 percpu_down_write(&cpuset_rwsem);
2735 set_bit(CS_ONLINE, &cs->flags);
2736 if (is_spread_page(parent))
2737 set_bit(CS_SPREAD_PAGE, &cs->flags);
2738 if (is_spread_slab(parent))
2739 set_bit(CS_SPREAD_SLAB, &cs->flags);
2743 spin_lock_irq(&callback_lock);
2744 if (is_in_v2_mode()) {
2745 cpumask_copy(cs->effective_cpus, parent->effective_cpus);
2746 cs->effective_mems = parent->effective_mems;
2747 cs->use_parent_ecpus = true;
2748 parent->child_ecpus_count++;
2750 spin_unlock_irq(&callback_lock);
2752 if (!test_bit(CGRP_CPUSET_CLONE_CHILDREN, &css->cgroup->flags))
2756 * Clone @parent's configuration if CGRP_CPUSET_CLONE_CHILDREN is
2757 * set. This flag handling is implemented in cgroup core for
2758 * histrical reasons - the flag may be specified during mount.
2760 * Currently, if any sibling cpusets have exclusive cpus or mem, we
2761 * refuse to clone the configuration - thereby refusing the task to
2762 * be entered, and as a result refusing the sys_unshare() or
2763 * clone() which initiated it. If this becomes a problem for some
2764 * users who wish to allow that scenario, then this could be
2765 * changed to grant parent->cpus_allowed-sibling_cpus_exclusive
2766 * (and likewise for mems) to the new cgroup.
2769 cpuset_for_each_child(tmp_cs, pos_css, parent) {
2770 if (is_mem_exclusive(tmp_cs) || is_cpu_exclusive(tmp_cs)) {
2777 spin_lock_irq(&callback_lock);
2778 cs->mems_allowed = parent->mems_allowed;
2779 cs->effective_mems = parent->mems_allowed;
2780 cpumask_copy(cs->cpus_allowed, parent->cpus_allowed);
2781 cpumask_copy(cs->effective_cpus, parent->cpus_allowed);
2782 spin_unlock_irq(&callback_lock);
2784 percpu_up_write(&cpuset_rwsem);
2790 * If the cpuset being removed has its flag 'sched_load_balance'
2791 * enabled, then simulate turning sched_load_balance off, which
2792 * will call rebuild_sched_domains_locked(). That is not needed
2793 * in the default hierarchy where only changes in partition
2794 * will cause repartitioning.
2796 * If the cpuset has the 'sched.partition' flag enabled, simulate
2797 * turning 'sched.partition" off.
2800 static void cpuset_css_offline(struct cgroup_subsys_state *css)
2802 struct cpuset *cs = css_cs(css);
2805 percpu_down_write(&cpuset_rwsem);
2807 if (is_partition_root(cs))
2808 update_prstate(cs, 0);
2810 if (!cgroup_subsys_on_dfl(cpuset_cgrp_subsys) &&
2811 is_sched_load_balance(cs))
2812 update_flag(CS_SCHED_LOAD_BALANCE, cs, 0);
2814 if (cs->use_parent_ecpus) {
2815 struct cpuset *parent = parent_cs(cs);
2817 cs->use_parent_ecpus = false;
2818 parent->child_ecpus_count--;
2822 clear_bit(CS_ONLINE, &cs->flags);
2824 percpu_up_write(&cpuset_rwsem);
2828 static void cpuset_css_free(struct cgroup_subsys_state *css)
2830 struct cpuset *cs = css_cs(css);
2835 static void cpuset_bind(struct cgroup_subsys_state *root_css)
2837 percpu_down_write(&cpuset_rwsem);
2838 spin_lock_irq(&callback_lock);
2840 if (is_in_v2_mode()) {
2841 cpumask_copy(top_cpuset.cpus_allowed, cpu_possible_mask);
2842 top_cpuset.mems_allowed = node_possible_map;
2844 cpumask_copy(top_cpuset.cpus_allowed,
2845 top_cpuset.effective_cpus);
2846 top_cpuset.mems_allowed = top_cpuset.effective_mems;
2849 spin_unlock_irq(&callback_lock);
2850 percpu_up_write(&cpuset_rwsem);
2854 * Make sure the new task conform to the current state of its parent,
2855 * which could have been changed by cpuset just after it inherits the
2856 * state from the parent and before it sits on the cgroup's task list.
2858 static void cpuset_fork(struct task_struct *task)
2860 if (task_css_is_root(task, cpuset_cgrp_id))
2863 set_cpus_allowed_ptr(task, current->cpus_ptr);
2864 task->mems_allowed = current->mems_allowed;
2867 struct cgroup_subsys cpuset_cgrp_subsys = {
2868 .css_alloc = cpuset_css_alloc,
2869 .css_online = cpuset_css_online,
2870 .css_offline = cpuset_css_offline,
2871 .css_free = cpuset_css_free,
2872 .can_attach = cpuset_can_attach,
2873 .cancel_attach = cpuset_cancel_attach,
2874 .attach = cpuset_attach,
2875 .post_attach = cpuset_post_attach,
2876 .bind = cpuset_bind,
2877 .fork = cpuset_fork,
2878 .legacy_cftypes = legacy_files,
2879 .dfl_cftypes = dfl_files,
2885 * cpuset_init - initialize cpusets at system boot
2887 * Description: Initialize top_cpuset
2890 int __init cpuset_init(void)
2892 BUG_ON(percpu_init_rwsem(&cpuset_rwsem));
2894 BUG_ON(!alloc_cpumask_var(&top_cpuset.cpus_allowed, GFP_KERNEL));
2895 BUG_ON(!alloc_cpumask_var(&top_cpuset.effective_cpus, GFP_KERNEL));
2896 BUG_ON(!zalloc_cpumask_var(&top_cpuset.subparts_cpus, GFP_KERNEL));
2898 cpumask_setall(top_cpuset.cpus_allowed);
2899 nodes_setall(top_cpuset.mems_allowed);
2900 cpumask_setall(top_cpuset.effective_cpus);
2901 nodes_setall(top_cpuset.effective_mems);
2903 fmeter_init(&top_cpuset.fmeter);
2904 set_bit(CS_SCHED_LOAD_BALANCE, &top_cpuset.flags);
2905 top_cpuset.relax_domain_level = -1;
2907 BUG_ON(!alloc_cpumask_var(&cpus_attach, GFP_KERNEL));
2913 * If CPU and/or memory hotplug handlers, below, unplug any CPUs
2914 * or memory nodes, we need to walk over the cpuset hierarchy,
2915 * removing that CPU or node from all cpusets. If this removes the
2916 * last CPU or node from a cpuset, then move the tasks in the empty
2917 * cpuset to its next-highest non-empty parent.
2919 static void remove_tasks_in_empty_cpuset(struct cpuset *cs)
2921 struct cpuset *parent;
2924 * Find its next-highest non-empty parent, (top cpuset
2925 * has online cpus, so can't be empty).
2927 parent = parent_cs(cs);
2928 while (cpumask_empty(parent->cpus_allowed) ||
2929 nodes_empty(parent->mems_allowed))
2930 parent = parent_cs(parent);
2932 if (cgroup_transfer_tasks(parent->css.cgroup, cs->css.cgroup)) {
2933 pr_err("cpuset: failed to transfer tasks out of empty cpuset ");
2934 pr_cont_cgroup_name(cs->css.cgroup);
2940 hotplug_update_tasks_legacy(struct cpuset *cs,
2941 struct cpumask *new_cpus, nodemask_t *new_mems,
2942 bool cpus_updated, bool mems_updated)
2946 spin_lock_irq(&callback_lock);
2947 cpumask_copy(cs->cpus_allowed, new_cpus);
2948 cpumask_copy(cs->effective_cpus, new_cpus);
2949 cs->mems_allowed = *new_mems;
2950 cs->effective_mems = *new_mems;
2951 spin_unlock_irq(&callback_lock);
2954 * Don't call update_tasks_cpumask() if the cpuset becomes empty,
2955 * as the tasks will be migratecd to an ancestor.
2957 if (cpus_updated && !cpumask_empty(cs->cpus_allowed))
2958 update_tasks_cpumask(cs);
2959 if (mems_updated && !nodes_empty(cs->mems_allowed))
2960 update_tasks_nodemask(cs);
2962 is_empty = cpumask_empty(cs->cpus_allowed) ||
2963 nodes_empty(cs->mems_allowed);
2965 percpu_up_write(&cpuset_rwsem);
2968 * Move tasks to the nearest ancestor with execution resources,
2969 * This is full cgroup operation which will also call back into
2970 * cpuset. Should be done outside any lock.
2973 remove_tasks_in_empty_cpuset(cs);
2975 percpu_down_write(&cpuset_rwsem);
2979 hotplug_update_tasks(struct cpuset *cs,
2980 struct cpumask *new_cpus, nodemask_t *new_mems,
2981 bool cpus_updated, bool mems_updated)
2983 if (cpumask_empty(new_cpus))
2984 cpumask_copy(new_cpus, parent_cs(cs)->effective_cpus);
2985 if (nodes_empty(*new_mems))
2986 *new_mems = parent_cs(cs)->effective_mems;
2988 spin_lock_irq(&callback_lock);
2989 cpumask_copy(cs->effective_cpus, new_cpus);
2990 cs->effective_mems = *new_mems;
2991 spin_unlock_irq(&callback_lock);
2994 update_tasks_cpumask(cs);
2996 update_tasks_nodemask(cs);
2999 static bool force_rebuild;
3001 void cpuset_force_rebuild(void)
3003 force_rebuild = true;
3007 * cpuset_hotplug_update_tasks - update tasks in a cpuset for hotunplug
3008 * @cs: cpuset in interest
3009 * @tmp: the tmpmasks structure pointer
3011 * Compare @cs's cpu and mem masks against top_cpuset and if some have gone
3012 * offline, update @cs accordingly. If @cs ends up with no CPU or memory,
3013 * all its tasks are moved to the nearest ancestor with both resources.
3015 static void cpuset_hotplug_update_tasks(struct cpuset *cs, struct tmpmasks *tmp)
3017 static cpumask_t new_cpus;
3018 static nodemask_t new_mems;
3021 struct cpuset *parent;
3023 wait_event(cpuset_attach_wq, cs->attach_in_progress == 0);
3025 percpu_down_write(&cpuset_rwsem);
3028 * We have raced with task attaching. We wait until attaching
3029 * is finished, so we won't attach a task to an empty cpuset.
3031 if (cs->attach_in_progress) {
3032 percpu_up_write(&cpuset_rwsem);
3036 parent = parent_cs(cs);
3037 compute_effective_cpumask(&new_cpus, cs, parent);
3038 nodes_and(new_mems, cs->mems_allowed, parent->effective_mems);
3040 if (cs->nr_subparts_cpus)
3042 * Make sure that CPUs allocated to child partitions
3043 * do not show up in effective_cpus.
3045 cpumask_andnot(&new_cpus, &new_cpus, cs->subparts_cpus);
3047 if (!tmp || !cs->partition_root_state)
3051 * In the unlikely event that a partition root has empty
3052 * effective_cpus or its parent becomes erroneous, we have to
3053 * transition it to the erroneous state.
3055 if (is_partition_root(cs) && (cpumask_empty(&new_cpus) ||
3056 (parent->partition_root_state == PRS_ERROR))) {
3057 if (cs->nr_subparts_cpus) {
3058 cs->nr_subparts_cpus = 0;
3059 cpumask_clear(cs->subparts_cpus);
3060 compute_effective_cpumask(&new_cpus, cs, parent);
3064 * If the effective_cpus is empty because the child
3065 * partitions take away all the CPUs, we can keep
3066 * the current partition and let the child partitions
3067 * fight for available CPUs.
3069 if ((parent->partition_root_state == PRS_ERROR) ||
3070 cpumask_empty(&new_cpus)) {
3071 update_parent_subparts_cpumask(cs, partcmd_disable,
3073 cs->partition_root_state = PRS_ERROR;
3075 cpuset_force_rebuild();
3079 * On the other hand, an erroneous partition root may be transitioned
3080 * back to a regular one or a partition root with no CPU allocated
3081 * from the parent may change to erroneous.
3083 if (is_partition_root(parent) &&
3084 ((cs->partition_root_state == PRS_ERROR) ||
3085 !cpumask_intersects(&new_cpus, parent->subparts_cpus)) &&
3086 update_parent_subparts_cpumask(cs, partcmd_update, NULL, tmp))
3087 cpuset_force_rebuild();
3090 cpus_updated = !cpumask_equal(&new_cpus, cs->effective_cpus);
3091 mems_updated = !nodes_equal(new_mems, cs->effective_mems);
3093 if (is_in_v2_mode())
3094 hotplug_update_tasks(cs, &new_cpus, &new_mems,
3095 cpus_updated, mems_updated);
3097 hotplug_update_tasks_legacy(cs, &new_cpus, &new_mems,
3098 cpus_updated, mems_updated);
3100 percpu_up_write(&cpuset_rwsem);
3104 * cpuset_hotplug_workfn - handle CPU/memory hotunplug for a cpuset
3106 * This function is called after either CPU or memory configuration has
3107 * changed and updates cpuset accordingly. The top_cpuset is always
3108 * synchronized to cpu_active_mask and N_MEMORY, which is necessary in
3109 * order to make cpusets transparent (of no affect) on systems that are
3110 * actively using CPU hotplug but making no active use of cpusets.
3112 * Non-root cpusets are only affected by offlining. If any CPUs or memory
3113 * nodes have been taken down, cpuset_hotplug_update_tasks() is invoked on
3116 * Note that CPU offlining during suspend is ignored. We don't modify
3117 * cpusets across suspend/resume cycles at all.
3119 static void cpuset_hotplug_workfn(struct work_struct *work)
3121 static cpumask_t new_cpus;
3122 static nodemask_t new_mems;
3123 bool cpus_updated, mems_updated;
3124 bool on_dfl = is_in_v2_mode();
3125 struct tmpmasks tmp, *ptmp = NULL;
3127 if (on_dfl && !alloc_cpumasks(NULL, &tmp))
3130 percpu_down_write(&cpuset_rwsem);
3132 /* fetch the available cpus/mems and find out which changed how */
3133 cpumask_copy(&new_cpus, cpu_active_mask);
3134 new_mems = node_states[N_MEMORY];
3137 * If subparts_cpus is populated, it is likely that the check below
3138 * will produce a false positive on cpus_updated when the cpu list
3139 * isn't changed. It is extra work, but it is better to be safe.
3141 cpus_updated = !cpumask_equal(top_cpuset.effective_cpus, &new_cpus);
3142 mems_updated = !nodes_equal(top_cpuset.effective_mems, new_mems);
3144 /* synchronize cpus_allowed to cpu_active_mask */
3146 spin_lock_irq(&callback_lock);
3148 cpumask_copy(top_cpuset.cpus_allowed, &new_cpus);
3150 * Make sure that CPUs allocated to child partitions
3151 * do not show up in effective_cpus. If no CPU is left,
3152 * we clear the subparts_cpus & let the child partitions
3153 * fight for the CPUs again.
3155 if (top_cpuset.nr_subparts_cpus) {
3156 if (cpumask_subset(&new_cpus,
3157 top_cpuset.subparts_cpus)) {
3158 top_cpuset.nr_subparts_cpus = 0;
3159 cpumask_clear(top_cpuset.subparts_cpus);
3161 cpumask_andnot(&new_cpus, &new_cpus,
3162 top_cpuset.subparts_cpus);
3165 cpumask_copy(top_cpuset.effective_cpus, &new_cpus);
3166 spin_unlock_irq(&callback_lock);
3167 /* we don't mess with cpumasks of tasks in top_cpuset */
3170 /* synchronize mems_allowed to N_MEMORY */
3172 spin_lock_irq(&callback_lock);
3174 top_cpuset.mems_allowed = new_mems;
3175 top_cpuset.effective_mems = new_mems;
3176 spin_unlock_irq(&callback_lock);
3177 update_tasks_nodemask(&top_cpuset);
3180 percpu_up_write(&cpuset_rwsem);
3182 /* if cpus or mems changed, we need to propagate to descendants */
3183 if (cpus_updated || mems_updated) {
3185 struct cgroup_subsys_state *pos_css;
3188 cpuset_for_each_descendant_pre(cs, pos_css, &top_cpuset) {
3189 if (cs == &top_cpuset || !css_tryget_online(&cs->css))
3193 cpuset_hotplug_update_tasks(cs, ptmp);
3201 /* rebuild sched domains if cpus_allowed has changed */
3202 if (cpus_updated || force_rebuild) {
3203 force_rebuild = false;
3204 rebuild_sched_domains();
3207 free_cpumasks(NULL, ptmp);
3210 void cpuset_update_active_cpus(void)
3213 * We're inside cpu hotplug critical region which usually nests
3214 * inside cgroup synchronization. Bounce actual hotplug processing
3215 * to a work item to avoid reverse locking order.
3217 schedule_work(&cpuset_hotplug_work);
3220 void cpuset_wait_for_hotplug(void)
3222 flush_work(&cpuset_hotplug_work);
3226 * Keep top_cpuset.mems_allowed tracking node_states[N_MEMORY].
3227 * Call this routine anytime after node_states[N_MEMORY] changes.
3228 * See cpuset_update_active_cpus() for CPU hotplug handling.
3230 static int cpuset_track_online_nodes(struct notifier_block *self,
3231 unsigned long action, void *arg)
3233 schedule_work(&cpuset_hotplug_work);
3237 static struct notifier_block cpuset_track_online_nodes_nb = {
3238 .notifier_call = cpuset_track_online_nodes,
3239 .priority = 10, /* ??! */
3243 * cpuset_init_smp - initialize cpus_allowed
3245 * Description: Finish top cpuset after cpu, node maps are initialized
3247 void __init cpuset_init_smp(void)
3249 cpumask_copy(top_cpuset.cpus_allowed, cpu_active_mask);
3250 top_cpuset.mems_allowed = node_states[N_MEMORY];
3251 top_cpuset.old_mems_allowed = top_cpuset.mems_allowed;
3253 cpumask_copy(top_cpuset.effective_cpus, cpu_active_mask);
3254 top_cpuset.effective_mems = node_states[N_MEMORY];
3256 register_hotmemory_notifier(&cpuset_track_online_nodes_nb);
3258 cpuset_migrate_mm_wq = alloc_ordered_workqueue("cpuset_migrate_mm", 0);
3259 BUG_ON(!cpuset_migrate_mm_wq);
3263 * cpuset_cpus_allowed - return cpus_allowed mask from a tasks cpuset.
3264 * @tsk: pointer to task_struct from which to obtain cpuset->cpus_allowed.
3265 * @pmask: pointer to struct cpumask variable to receive cpus_allowed set.
3267 * Description: Returns the cpumask_var_t cpus_allowed of the cpuset
3268 * attached to the specified @tsk. Guaranteed to return some non-empty
3269 * subset of cpu_online_mask, even if this means going outside the
3273 void cpuset_cpus_allowed(struct task_struct *tsk, struct cpumask *pmask)
3275 unsigned long flags;
3277 spin_lock_irqsave(&callback_lock, flags);
3279 guarantee_online_cpus(task_cs(tsk), pmask);
3281 spin_unlock_irqrestore(&callback_lock, flags);
3285 * cpuset_cpus_allowed_fallback - final fallback before complete catastrophe.
3286 * @tsk: pointer to task_struct with which the scheduler is struggling
3288 * Description: In the case that the scheduler cannot find an allowed cpu in
3289 * tsk->cpus_allowed, we fall back to task_cs(tsk)->cpus_allowed. In legacy
3290 * mode however, this value is the same as task_cs(tsk)->effective_cpus,
3291 * which will not contain a sane cpumask during cases such as cpu hotplugging.
3292 * This is the absolute last resort for the scheduler and it is only used if
3293 * _every_ other avenue has been traveled.
3296 void cpuset_cpus_allowed_fallback(struct task_struct *tsk)
3299 do_set_cpus_allowed(tsk, is_in_v2_mode() ?
3300 task_cs(tsk)->cpus_allowed : cpu_possible_mask);
3304 * We own tsk->cpus_allowed, nobody can change it under us.
3306 * But we used cs && cs->cpus_allowed lockless and thus can
3307 * race with cgroup_attach_task() or update_cpumask() and get
3308 * the wrong tsk->cpus_allowed. However, both cases imply the
3309 * subsequent cpuset_change_cpumask()->set_cpus_allowed_ptr()
3310 * which takes task_rq_lock().
3312 * If we are called after it dropped the lock we must see all
3313 * changes in tsk_cs()->cpus_allowed. Otherwise we can temporary
3314 * set any mask even if it is not right from task_cs() pov,
3315 * the pending set_cpus_allowed_ptr() will fix things.
3317 * select_fallback_rq() will fix things ups and set cpu_possible_mask
3322 void __init cpuset_init_current_mems_allowed(void)
3324 nodes_setall(current->mems_allowed);
3328 * cpuset_mems_allowed - return mems_allowed mask from a tasks cpuset.
3329 * @tsk: pointer to task_struct from which to obtain cpuset->mems_allowed.
3331 * Description: Returns the nodemask_t mems_allowed of the cpuset
3332 * attached to the specified @tsk. Guaranteed to return some non-empty
3333 * subset of node_states[N_MEMORY], even if this means going outside the
3337 nodemask_t cpuset_mems_allowed(struct task_struct *tsk)
3340 unsigned long flags;
3342 spin_lock_irqsave(&callback_lock, flags);
3344 guarantee_online_mems(task_cs(tsk), &mask);
3346 spin_unlock_irqrestore(&callback_lock, flags);
3352 * cpuset_nodemask_valid_mems_allowed - check nodemask vs. curremt mems_allowed
3353 * @nodemask: the nodemask to be checked
3355 * Are any of the nodes in the nodemask allowed in current->mems_allowed?
3357 int cpuset_nodemask_valid_mems_allowed(nodemask_t *nodemask)
3359 return nodes_intersects(*nodemask, current->mems_allowed);
3363 * nearest_hardwall_ancestor() - Returns the nearest mem_exclusive or
3364 * mem_hardwall ancestor to the specified cpuset. Call holding
3365 * callback_lock. If no ancestor is mem_exclusive or mem_hardwall
3366 * (an unusual configuration), then returns the root cpuset.
3368 static struct cpuset *nearest_hardwall_ancestor(struct cpuset *cs)
3370 while (!(is_mem_exclusive(cs) || is_mem_hardwall(cs)) && parent_cs(cs))
3376 * cpuset_node_allowed - Can we allocate on a memory node?
3377 * @node: is this an allowed node?
3378 * @gfp_mask: memory allocation flags
3380 * If we're in interrupt, yes, we can always allocate. If @node is set in
3381 * current's mems_allowed, yes. If it's not a __GFP_HARDWALL request and this
3382 * node is set in the nearest hardwalled cpuset ancestor to current's cpuset,
3383 * yes. If current has access to memory reserves as an oom victim, yes.
3386 * GFP_USER allocations are marked with the __GFP_HARDWALL bit,
3387 * and do not allow allocations outside the current tasks cpuset
3388 * unless the task has been OOM killed.
3389 * GFP_KERNEL allocations are not so marked, so can escape to the
3390 * nearest enclosing hardwalled ancestor cpuset.
3392 * Scanning up parent cpusets requires callback_lock. The
3393 * __alloc_pages() routine only calls here with __GFP_HARDWALL bit
3394 * _not_ set if it's a GFP_KERNEL allocation, and all nodes in the
3395 * current tasks mems_allowed came up empty on the first pass over
3396 * the zonelist. So only GFP_KERNEL allocations, if all nodes in the
3397 * cpuset are short of memory, might require taking the callback_lock.
3399 * The first call here from mm/page_alloc:get_page_from_freelist()
3400 * has __GFP_HARDWALL set in gfp_mask, enforcing hardwall cpusets,
3401 * so no allocation on a node outside the cpuset is allowed (unless
3402 * in interrupt, of course).
3404 * The second pass through get_page_from_freelist() doesn't even call
3405 * here for GFP_ATOMIC calls. For those calls, the __alloc_pages()
3406 * variable 'wait' is not set, and the bit ALLOC_CPUSET is not set
3407 * in alloc_flags. That logic and the checks below have the combined
3409 * in_interrupt - any node ok (current task context irrelevant)
3410 * GFP_ATOMIC - any node ok
3411 * tsk_is_oom_victim - any node ok
3412 * GFP_KERNEL - any node in enclosing hardwalled cpuset ok
3413 * GFP_USER - only nodes in current tasks mems allowed ok.
3415 bool __cpuset_node_allowed(int node, gfp_t gfp_mask)
3417 struct cpuset *cs; /* current cpuset ancestors */
3418 int allowed; /* is allocation in zone z allowed? */
3419 unsigned long flags;
3423 if (node_isset(node, current->mems_allowed))
3426 * Allow tasks that have access to memory reserves because they have
3427 * been OOM killed to get memory anywhere.
3429 if (unlikely(tsk_is_oom_victim(current)))
3431 if (gfp_mask & __GFP_HARDWALL) /* If hardwall request, stop here */
3434 if (current->flags & PF_EXITING) /* Let dying task have memory */
3437 /* Not hardwall and node outside mems_allowed: scan up cpusets */
3438 spin_lock_irqsave(&callback_lock, flags);
3441 cs = nearest_hardwall_ancestor(task_cs(current));
3442 allowed = node_isset(node, cs->mems_allowed);
3445 spin_unlock_irqrestore(&callback_lock, flags);
3450 * cpuset_mem_spread_node() - On which node to begin search for a file page
3451 * cpuset_slab_spread_node() - On which node to begin search for a slab page
3453 * If a task is marked PF_SPREAD_PAGE or PF_SPREAD_SLAB (as for
3454 * tasks in a cpuset with is_spread_page or is_spread_slab set),
3455 * and if the memory allocation used cpuset_mem_spread_node()
3456 * to determine on which node to start looking, as it will for
3457 * certain page cache or slab cache pages such as used for file
3458 * system buffers and inode caches, then instead of starting on the
3459 * local node to look for a free page, rather spread the starting
3460 * node around the tasks mems_allowed nodes.
3462 * We don't have to worry about the returned node being offline
3463 * because "it can't happen", and even if it did, it would be ok.
3465 * The routines calling guarantee_online_mems() are careful to
3466 * only set nodes in task->mems_allowed that are online. So it
3467 * should not be possible for the following code to return an
3468 * offline node. But if it did, that would be ok, as this routine
3469 * is not returning the node where the allocation must be, only
3470 * the node where the search should start. The zonelist passed to
3471 * __alloc_pages() will include all nodes. If the slab allocator
3472 * is passed an offline node, it will fall back to the local node.
3473 * See kmem_cache_alloc_node().
3476 static int cpuset_spread_node(int *rotor)
3478 return *rotor = next_node_in(*rotor, current->mems_allowed);
3481 int cpuset_mem_spread_node(void)
3483 if (current->cpuset_mem_spread_rotor == NUMA_NO_NODE)
3484 current->cpuset_mem_spread_rotor =
3485 node_random(¤t->mems_allowed);
3487 return cpuset_spread_node(¤t->cpuset_mem_spread_rotor);
3490 int cpuset_slab_spread_node(void)
3492 if (current->cpuset_slab_spread_rotor == NUMA_NO_NODE)
3493 current->cpuset_slab_spread_rotor =
3494 node_random(¤t->mems_allowed);
3496 return cpuset_spread_node(¤t->cpuset_slab_spread_rotor);
3499 EXPORT_SYMBOL_GPL(cpuset_mem_spread_node);
3502 * cpuset_mems_allowed_intersects - Does @tsk1's mems_allowed intersect @tsk2's?
3503 * @tsk1: pointer to task_struct of some task.
3504 * @tsk2: pointer to task_struct of some other task.
3506 * Description: Return true if @tsk1's mems_allowed intersects the
3507 * mems_allowed of @tsk2. Used by the OOM killer to determine if
3508 * one of the task's memory usage might impact the memory available
3512 int cpuset_mems_allowed_intersects(const struct task_struct *tsk1,
3513 const struct task_struct *tsk2)
3515 return nodes_intersects(tsk1->mems_allowed, tsk2->mems_allowed);
3519 * cpuset_print_current_mems_allowed - prints current's cpuset and mems_allowed
3521 * Description: Prints current's name, cpuset name, and cached copy of its
3522 * mems_allowed to the kernel log.
3524 void cpuset_print_current_mems_allowed(void)
3526 struct cgroup *cgrp;
3530 cgrp = task_cs(current)->css.cgroup;
3531 pr_cont(",cpuset=");
3532 pr_cont_cgroup_name(cgrp);
3533 pr_cont(",mems_allowed=%*pbl",
3534 nodemask_pr_args(¤t->mems_allowed));
3540 * Collection of memory_pressure is suppressed unless
3541 * this flag is enabled by writing "1" to the special
3542 * cpuset file 'memory_pressure_enabled' in the root cpuset.
3545 int cpuset_memory_pressure_enabled __read_mostly;
3548 * cpuset_memory_pressure_bump - keep stats of per-cpuset reclaims.
3550 * Keep a running average of the rate of synchronous (direct)
3551 * page reclaim efforts initiated by tasks in each cpuset.
3553 * This represents the rate at which some task in the cpuset
3554 * ran low on memory on all nodes it was allowed to use, and
3555 * had to enter the kernels page reclaim code in an effort to
3556 * create more free memory by tossing clean pages or swapping
3557 * or writing dirty pages.
3559 * Display to user space in the per-cpuset read-only file
3560 * "memory_pressure". Value displayed is an integer
3561 * representing the recent rate of entry into the synchronous
3562 * (direct) page reclaim by any task attached to the cpuset.
3565 void __cpuset_memory_pressure_bump(void)
3568 fmeter_markevent(&task_cs(current)->fmeter);
3572 #ifdef CONFIG_PROC_PID_CPUSET
3574 * proc_cpuset_show()
3575 * - Print tasks cpuset path into seq_file.
3576 * - Used for /proc/<pid>/cpuset.
3577 * - No need to task_lock(tsk) on this tsk->cpuset reference, as it
3578 * doesn't really matter if tsk->cpuset changes after we read it,
3579 * and we take cpuset_mutex, keeping cpuset_attach() from changing it
3582 int proc_cpuset_show(struct seq_file *m, struct pid_namespace *ns,
3583 struct pid *pid, struct task_struct *tsk)
3586 struct cgroup_subsys_state *css;
3590 buf = kmalloc(PATH_MAX, GFP_KERNEL);
3594 css = task_get_css(tsk, cpuset_cgrp_id);
3595 retval = cgroup_path_ns(css->cgroup, buf, PATH_MAX,
3596 current->nsproxy->cgroup_ns);
3598 if (retval >= PATH_MAX)
3599 retval = -ENAMETOOLONG;
3610 #endif /* CONFIG_PROC_PID_CPUSET */
3612 /* Display task mems_allowed in /proc/<pid>/status file. */
3613 void cpuset_task_status_allowed(struct seq_file *m, struct task_struct *task)
3615 seq_printf(m, "Mems_allowed:\t%*pb\n",
3616 nodemask_pr_args(&task->mems_allowed));
3617 seq_printf(m, "Mems_allowed_list:\t%*pbl\n",
3618 nodemask_pr_args(&task->mems_allowed));