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/mm.h>
49 #include <linux/sched/task.h>
50 #include <linux/seq_file.h>
51 #include <linux/security.h>
52 #include <linux/slab.h>
53 #include <linux/spinlock.h>
54 #include <linux/stat.h>
55 #include <linux/string.h>
56 #include <linux/time.h>
57 #include <linux/time64.h>
58 #include <linux/backing-dev.h>
59 #include <linux/sort.h>
60 #include <linux/oom.h>
61 #include <linux/sched/isolation.h>
62 #include <linux/uaccess.h>
63 #include <linux/atomic.h>
64 #include <linux/mutex.h>
65 #include <linux/cgroup.h>
66 #include <linux/wait.h>
68 DEFINE_STATIC_KEY_FALSE(cpusets_pre_enable_key);
69 DEFINE_STATIC_KEY_FALSE(cpusets_enabled_key);
71 /* See "Frequency meter" comments, below. */
74 int cnt; /* unprocessed events count */
75 int val; /* most recent output value */
76 time64_t time; /* clock (secs) when val computed */
77 spinlock_t lock; /* guards read or write of above */
81 struct cgroup_subsys_state css;
83 unsigned long flags; /* "unsigned long" so bitops work */
86 * On default hierarchy:
88 * The user-configured masks can only be changed by writing to
89 * cpuset.cpus and cpuset.mems, and won't be limited by the
92 * The effective masks is the real masks that apply to the tasks
93 * in the cpuset. They may be changed if the configured masks are
94 * changed or hotplug happens.
96 * effective_mask == configured_mask & parent's effective_mask,
97 * and if it ends up empty, it will inherit the parent's mask.
100 * On legacy hierachy:
102 * The user-configured masks are always the same with effective masks.
105 /* user-configured CPUs and Memory Nodes allow to tasks */
106 cpumask_var_t cpus_allowed;
107 nodemask_t mems_allowed;
109 /* effective CPUs and Memory Nodes allow to tasks */
110 cpumask_var_t effective_cpus;
111 nodemask_t effective_mems;
114 * CPUs allocated to child sub-partitions (default hierarchy only)
115 * - CPUs granted by the parent = effective_cpus U subparts_cpus
116 * - effective_cpus and subparts_cpus are mutually exclusive.
118 * effective_cpus contains only onlined CPUs, but subparts_cpus
119 * may have offlined ones.
121 cpumask_var_t subparts_cpus;
124 * This is old Memory Nodes tasks took on.
126 * - top_cpuset.old_mems_allowed is initialized to mems_allowed.
127 * - A new cpuset's old_mems_allowed is initialized when some
128 * task is moved into it.
129 * - old_mems_allowed is used in cpuset_migrate_mm() when we change
130 * cpuset.mems_allowed and have tasks' nodemask updated, and
131 * then old_mems_allowed is updated to mems_allowed.
133 nodemask_t old_mems_allowed;
135 struct fmeter fmeter; /* memory_pressure filter */
138 * Tasks are being attached to this cpuset. Used to prevent
139 * zeroing cpus/mems_allowed between ->can_attach() and ->attach().
141 int attach_in_progress;
143 /* partition number for rebuild_sched_domains() */
146 /* for custom sched domain */
147 int relax_domain_level;
149 /* number of CPUs in subparts_cpus */
150 int nr_subparts_cpus;
152 /* partition root state */
153 int partition_root_state;
156 * Default hierarchy only:
157 * use_parent_ecpus - set if using parent's effective_cpus
158 * child_ecpus_count - # of children with use_parent_ecpus set
160 int use_parent_ecpus;
161 int child_ecpus_count;
165 * Partition root states:
167 * 0 - not a partition root
171 * -1 - invalid partition root
172 * None of the cpus in cpus_allowed can be put into the parent's
173 * subparts_cpus. In this case, the cpuset is not a real partition
174 * root anymore. However, the CPU_EXCLUSIVE bit will still be set
175 * and the cpuset can be restored back to a partition root if the
176 * parent cpuset can give more CPUs back to this child cpuset.
178 #define PRS_DISABLED 0
179 #define PRS_ENABLED 1
183 * Temporary cpumasks for working with partitions that are passed among
184 * functions to avoid memory allocation in inner functions.
187 cpumask_var_t addmask, delmask; /* For partition root */
188 cpumask_var_t new_cpus; /* For update_cpumasks_hier() */
191 static inline struct cpuset *css_cs(struct cgroup_subsys_state *css)
193 return css ? container_of(css, struct cpuset, css) : NULL;
196 /* Retrieve the cpuset for a task */
197 static inline struct cpuset *task_cs(struct task_struct *task)
199 return css_cs(task_css(task, cpuset_cgrp_id));
202 static inline struct cpuset *parent_cs(struct cpuset *cs)
204 return css_cs(cs->css.parent);
207 /* bits in struct cpuset flags field */
214 CS_SCHED_LOAD_BALANCE,
219 /* convenient tests for these bits */
220 static inline bool is_cpuset_online(struct cpuset *cs)
222 return test_bit(CS_ONLINE, &cs->flags) && !css_is_dying(&cs->css);
225 static inline int is_cpu_exclusive(const struct cpuset *cs)
227 return test_bit(CS_CPU_EXCLUSIVE, &cs->flags);
230 static inline int is_mem_exclusive(const struct cpuset *cs)
232 return test_bit(CS_MEM_EXCLUSIVE, &cs->flags);
235 static inline int is_mem_hardwall(const struct cpuset *cs)
237 return test_bit(CS_MEM_HARDWALL, &cs->flags);
240 static inline int is_sched_load_balance(const struct cpuset *cs)
242 return test_bit(CS_SCHED_LOAD_BALANCE, &cs->flags);
245 static inline int is_memory_migrate(const struct cpuset *cs)
247 return test_bit(CS_MEMORY_MIGRATE, &cs->flags);
250 static inline int is_spread_page(const struct cpuset *cs)
252 return test_bit(CS_SPREAD_PAGE, &cs->flags);
255 static inline int is_spread_slab(const struct cpuset *cs)
257 return test_bit(CS_SPREAD_SLAB, &cs->flags);
260 static inline int is_partition_root(const struct cpuset *cs)
262 return cs->partition_root_state > 0;
265 static struct cpuset top_cpuset = {
266 .flags = ((1 << CS_ONLINE) | (1 << CS_CPU_EXCLUSIVE) |
267 (1 << CS_MEM_EXCLUSIVE)),
268 .partition_root_state = PRS_ENABLED,
272 * cpuset_for_each_child - traverse online children of a cpuset
273 * @child_cs: loop cursor pointing to the current child
274 * @pos_css: used for iteration
275 * @parent_cs: target cpuset to walk children of
277 * Walk @child_cs through the online children of @parent_cs. Must be used
278 * with RCU read locked.
280 #define cpuset_for_each_child(child_cs, pos_css, parent_cs) \
281 css_for_each_child((pos_css), &(parent_cs)->css) \
282 if (is_cpuset_online(((child_cs) = css_cs((pos_css)))))
285 * cpuset_for_each_descendant_pre - pre-order walk of a cpuset's descendants
286 * @des_cs: loop cursor pointing to the current descendant
287 * @pos_css: used for iteration
288 * @root_cs: target cpuset to walk ancestor of
290 * Walk @des_cs through the online descendants of @root_cs. Must be used
291 * with RCU read locked. The caller may modify @pos_css by calling
292 * css_rightmost_descendant() to skip subtree. @root_cs is included in the
293 * iteration and the first node to be visited.
295 #define cpuset_for_each_descendant_pre(des_cs, pos_css, root_cs) \
296 css_for_each_descendant_pre((pos_css), &(root_cs)->css) \
297 if (is_cpuset_online(((des_cs) = css_cs((pos_css)))))
300 * There are two global locks guarding cpuset structures - cpuset_mutex and
301 * callback_lock. We also require taking task_lock() when dereferencing a
302 * task's cpuset pointer. See "The task_lock() exception", at the end of this
305 * A task must hold both locks to modify cpusets. If a task holds
306 * cpuset_mutex, then it blocks others wanting that mutex, ensuring that it
307 * is the only task able to also acquire callback_lock and be able to
308 * modify cpusets. It can perform various checks on the cpuset structure
309 * first, knowing nothing will change. It can also allocate memory while
310 * just holding cpuset_mutex. While it is performing these checks, various
311 * callback routines can briefly acquire callback_lock to query cpusets.
312 * Once it is ready to make the changes, it takes callback_lock, blocking
315 * Calls to the kernel memory allocator can not be made while holding
316 * callback_lock, as that would risk double tripping on callback_lock
317 * from one of the callbacks into the cpuset code from within
320 * If a task is only holding callback_lock, then it has read-only
323 * Now, the task_struct fields mems_allowed and mempolicy may be changed
324 * by other task, we use alloc_lock in the task_struct fields to protect
327 * The cpuset_common_file_read() handlers only hold callback_lock across
328 * small pieces of code, such as when reading out possibly multi-word
329 * cpumasks and nodemasks.
331 * Accessing a task's cpuset should be done in accordance with the
332 * guidelines for accessing subsystem state in kernel/cgroup.c
335 static DEFINE_MUTEX(cpuset_mutex);
336 static DEFINE_SPINLOCK(callback_lock);
338 static struct workqueue_struct *cpuset_migrate_mm_wq;
341 * CPU / memory hotplug is handled asynchronously.
343 static void cpuset_hotplug_workfn(struct work_struct *work);
344 static DECLARE_WORK(cpuset_hotplug_work, cpuset_hotplug_workfn);
346 static DECLARE_WAIT_QUEUE_HEAD(cpuset_attach_wq);
349 * Cgroup v2 behavior is used when on default hierarchy or the
350 * cgroup_v2_mode flag is set.
352 static inline bool is_in_v2_mode(void)
354 return cgroup_subsys_on_dfl(cpuset_cgrp_subsys) ||
355 (cpuset_cgrp_subsys.root->flags & CGRP_ROOT_CPUSET_V2_MODE);
359 * Return in pmask the portion of a cpusets's cpus_allowed that
360 * are online. If none are online, walk up the cpuset hierarchy
361 * until we find one that does have some online cpus.
363 * One way or another, we guarantee to return some non-empty subset
364 * of cpu_online_mask.
366 * Call with callback_lock or cpuset_mutex held.
368 static void guarantee_online_cpus(struct cpuset *cs, struct cpumask *pmask)
370 while (!cpumask_intersects(cs->effective_cpus, cpu_online_mask)) {
374 * The top cpuset doesn't have any online cpu as a
375 * consequence of a race between cpuset_hotplug_work
376 * and cpu hotplug notifier. But we know the top
377 * cpuset's effective_cpus is on its way to to be
378 * identical to cpu_online_mask.
380 cpumask_copy(pmask, cpu_online_mask);
384 cpumask_and(pmask, cs->effective_cpus, cpu_online_mask);
388 * Return in *pmask the portion of a cpusets's mems_allowed that
389 * are online, with memory. If none are online with memory, walk
390 * up the cpuset hierarchy until we find one that does have some
391 * online mems. The top cpuset always has some mems online.
393 * One way or another, we guarantee to return some non-empty subset
394 * of node_states[N_MEMORY].
396 * Call with callback_lock or cpuset_mutex held.
398 static void guarantee_online_mems(struct cpuset *cs, nodemask_t *pmask)
400 while (!nodes_intersects(cs->effective_mems, node_states[N_MEMORY]))
402 nodes_and(*pmask, cs->effective_mems, node_states[N_MEMORY]);
406 * update task's spread flag if cpuset's page/slab spread flag is set
408 * Call with callback_lock or cpuset_mutex held.
410 static void cpuset_update_task_spread_flag(struct cpuset *cs,
411 struct task_struct *tsk)
413 if (is_spread_page(cs))
414 task_set_spread_page(tsk);
416 task_clear_spread_page(tsk);
418 if (is_spread_slab(cs))
419 task_set_spread_slab(tsk);
421 task_clear_spread_slab(tsk);
425 * is_cpuset_subset(p, q) - Is cpuset p a subset of cpuset q?
427 * One cpuset is a subset of another if all its allowed CPUs and
428 * Memory Nodes are a subset of the other, and its exclusive flags
429 * are only set if the other's are set. Call holding cpuset_mutex.
432 static int is_cpuset_subset(const struct cpuset *p, const struct cpuset *q)
434 return cpumask_subset(p->cpus_allowed, q->cpus_allowed) &&
435 nodes_subset(p->mems_allowed, q->mems_allowed) &&
436 is_cpu_exclusive(p) <= is_cpu_exclusive(q) &&
437 is_mem_exclusive(p) <= is_mem_exclusive(q);
441 * alloc_cpumasks - allocate three cpumasks for cpuset
442 * @cs: the cpuset that have cpumasks to be allocated.
443 * @tmp: the tmpmasks structure pointer
444 * Return: 0 if successful, -ENOMEM otherwise.
446 * Only one of the two input arguments should be non-NULL.
448 static inline int alloc_cpumasks(struct cpuset *cs, struct tmpmasks *tmp)
450 cpumask_var_t *pmask1, *pmask2, *pmask3;
453 pmask1 = &cs->cpus_allowed;
454 pmask2 = &cs->effective_cpus;
455 pmask3 = &cs->subparts_cpus;
457 pmask1 = &tmp->new_cpus;
458 pmask2 = &tmp->addmask;
459 pmask3 = &tmp->delmask;
462 if (!zalloc_cpumask_var(pmask1, GFP_KERNEL))
465 if (!zalloc_cpumask_var(pmask2, GFP_KERNEL))
468 if (!zalloc_cpumask_var(pmask3, GFP_KERNEL))
474 free_cpumask_var(*pmask2);
476 free_cpumask_var(*pmask1);
481 * free_cpumasks - free cpumasks in a tmpmasks structure
482 * @cs: the cpuset that have cpumasks to be free.
483 * @tmp: the tmpmasks structure pointer
485 static inline void free_cpumasks(struct cpuset *cs, struct tmpmasks *tmp)
488 free_cpumask_var(cs->cpus_allowed);
489 free_cpumask_var(cs->effective_cpus);
490 free_cpumask_var(cs->subparts_cpus);
493 free_cpumask_var(tmp->new_cpus);
494 free_cpumask_var(tmp->addmask);
495 free_cpumask_var(tmp->delmask);
500 * alloc_trial_cpuset - allocate a trial cpuset
501 * @cs: the cpuset that the trial cpuset duplicates
503 static struct cpuset *alloc_trial_cpuset(struct cpuset *cs)
505 struct cpuset *trial;
507 trial = kmemdup(cs, sizeof(*cs), GFP_KERNEL);
511 if (alloc_cpumasks(trial, NULL)) {
516 cpumask_copy(trial->cpus_allowed, cs->cpus_allowed);
517 cpumask_copy(trial->effective_cpus, cs->effective_cpus);
522 * free_cpuset - free the cpuset
523 * @cs: the cpuset to be freed
525 static inline void free_cpuset(struct cpuset *cs)
527 free_cpumasks(cs, NULL);
532 * validate_change() - Used to validate that any proposed cpuset change
533 * follows the structural rules for cpusets.
535 * If we replaced the flag and mask values of the current cpuset
536 * (cur) with those values in the trial cpuset (trial), would
537 * our various subset and exclusive rules still be valid? Presumes
540 * 'cur' is the address of an actual, in-use cpuset. Operations
541 * such as list traversal that depend on the actual address of the
542 * cpuset in the list must use cur below, not trial.
544 * 'trial' is the address of bulk structure copy of cur, with
545 * perhaps one or more of the fields cpus_allowed, mems_allowed,
546 * or flags changed to new, trial values.
548 * Return 0 if valid, -errno if not.
551 static int validate_change(struct cpuset *cur, struct cpuset *trial)
553 struct cgroup_subsys_state *css;
554 struct cpuset *c, *par;
559 /* Each of our child cpusets must be a subset of us */
561 cpuset_for_each_child(c, css, cur)
562 if (!is_cpuset_subset(c, trial))
565 /* Remaining checks don't apply to root cpuset */
567 if (cur == &top_cpuset)
570 par = parent_cs(cur);
572 /* On legacy hiearchy, we must be a subset of our parent cpuset. */
574 if (!is_in_v2_mode() && !is_cpuset_subset(trial, par))
578 * If either I or some sibling (!= me) is exclusive, we can't
582 cpuset_for_each_child(c, css, par) {
583 if ((is_cpu_exclusive(trial) || is_cpu_exclusive(c)) &&
585 cpumask_intersects(trial->cpus_allowed, c->cpus_allowed))
587 if ((is_mem_exclusive(trial) || is_mem_exclusive(c)) &&
589 nodes_intersects(trial->mems_allowed, c->mems_allowed))
594 * Cpusets with tasks - existing or newly being attached - can't
595 * be changed to have empty cpus_allowed or mems_allowed.
598 if ((cgroup_is_populated(cur->css.cgroup) || cur->attach_in_progress)) {
599 if (!cpumask_empty(cur->cpus_allowed) &&
600 cpumask_empty(trial->cpus_allowed))
602 if (!nodes_empty(cur->mems_allowed) &&
603 nodes_empty(trial->mems_allowed))
608 * We can't shrink if we won't have enough room for SCHED_DEADLINE
612 if (is_cpu_exclusive(cur) &&
613 !cpuset_cpumask_can_shrink(cur->cpus_allowed,
614 trial->cpus_allowed))
625 * Helper routine for generate_sched_domains().
626 * Do cpusets a, b have overlapping effective cpus_allowed masks?
628 static int cpusets_overlap(struct cpuset *a, struct cpuset *b)
630 return cpumask_intersects(a->effective_cpus, b->effective_cpus);
634 update_domain_attr(struct sched_domain_attr *dattr, struct cpuset *c)
636 if (dattr->relax_domain_level < c->relax_domain_level)
637 dattr->relax_domain_level = c->relax_domain_level;
641 static void update_domain_attr_tree(struct sched_domain_attr *dattr,
642 struct cpuset *root_cs)
645 struct cgroup_subsys_state *pos_css;
648 cpuset_for_each_descendant_pre(cp, pos_css, root_cs) {
649 /* skip the whole subtree if @cp doesn't have any CPU */
650 if (cpumask_empty(cp->cpus_allowed)) {
651 pos_css = css_rightmost_descendant(pos_css);
655 if (is_sched_load_balance(cp))
656 update_domain_attr(dattr, cp);
661 /* Must be called with cpuset_mutex held. */
662 static inline int nr_cpusets(void)
664 /* jump label reference count + the top-level cpuset */
665 return static_key_count(&cpusets_enabled_key.key) + 1;
669 * generate_sched_domains()
671 * This function builds a partial partition of the systems CPUs
672 * A 'partial partition' is a set of non-overlapping subsets whose
673 * union is a subset of that set.
674 * The output of this function needs to be passed to kernel/sched/core.c
675 * partition_sched_domains() routine, which will rebuild the scheduler's
676 * load balancing domains (sched domains) as specified by that partial
679 * See "What is sched_load_balance" in Documentation/admin-guide/cgroup-v1/cpusets.rst
680 * for a background explanation of this.
682 * Does not return errors, on the theory that the callers of this
683 * routine would rather not worry about failures to rebuild sched
684 * domains when operating in the severe memory shortage situations
685 * that could cause allocation failures below.
687 * Must be called with cpuset_mutex held.
689 * The three key local variables below are:
690 * cp - cpuset pointer, used (together with pos_css) to perform a
691 * top-down scan of all cpusets. For our purposes, rebuilding
692 * the schedulers sched domains, we can ignore !is_sched_load_
694 * csa - (for CpuSet Array) Array of pointers to all the cpusets
695 * that need to be load balanced, for convenient iterative
696 * access by the subsequent code that finds the best partition,
697 * i.e the set of domains (subsets) of CPUs such that the
698 * cpus_allowed of every cpuset marked is_sched_load_balance
699 * is a subset of one of these domains, while there are as
700 * many such domains as possible, each as small as possible.
701 * doms - Conversion of 'csa' to an array of cpumasks, for passing to
702 * the kernel/sched/core.c routine partition_sched_domains() in a
703 * convenient format, that can be easily compared to the prior
704 * value to determine what partition elements (sched domains)
705 * were changed (added or removed.)
707 * Finding the best partition (set of domains):
708 * The triple nested loops below over i, j, k scan over the
709 * load balanced cpusets (using the array of cpuset pointers in
710 * csa[]) looking for pairs of cpusets that have overlapping
711 * cpus_allowed, but which don't have the same 'pn' partition
712 * number and gives them in the same partition number. It keeps
713 * looping on the 'restart' label until it can no longer find
716 * The union of the cpus_allowed masks from the set of
717 * all cpusets having the same 'pn' value then form the one
718 * element of the partition (one sched domain) to be passed to
719 * partition_sched_domains().
721 static int generate_sched_domains(cpumask_var_t **domains,
722 struct sched_domain_attr **attributes)
724 struct cpuset *cp; /* top-down scan of cpusets */
725 struct cpuset **csa; /* array of all cpuset ptrs */
726 int csn; /* how many cpuset ptrs in csa so far */
727 int i, j, k; /* indices for partition finding loops */
728 cpumask_var_t *doms; /* resulting partition; i.e. sched domains */
729 struct sched_domain_attr *dattr; /* attributes for custom domains */
730 int ndoms = 0; /* number of sched domains in result */
731 int nslot; /* next empty doms[] struct cpumask slot */
732 struct cgroup_subsys_state *pos_css;
733 bool root_load_balance = is_sched_load_balance(&top_cpuset);
739 /* Special case for the 99% of systems with one, full, sched domain */
740 if (root_load_balance && !top_cpuset.nr_subparts_cpus) {
742 doms = alloc_sched_domains(ndoms);
746 dattr = kmalloc(sizeof(struct sched_domain_attr), GFP_KERNEL);
748 *dattr = SD_ATTR_INIT;
749 update_domain_attr_tree(dattr, &top_cpuset);
751 cpumask_and(doms[0], top_cpuset.effective_cpus,
752 housekeeping_cpumask(HK_FLAG_DOMAIN));
757 csa = kmalloc_array(nr_cpusets(), sizeof(cp), GFP_KERNEL);
763 if (root_load_balance)
764 csa[csn++] = &top_cpuset;
765 cpuset_for_each_descendant_pre(cp, pos_css, &top_cpuset) {
766 if (cp == &top_cpuset)
769 * Continue traversing beyond @cp iff @cp has some CPUs and
770 * isn't load balancing. The former is obvious. The
771 * latter: All child cpusets contain a subset of the
772 * parent's cpus, so just skip them, and then we call
773 * update_domain_attr_tree() to calc relax_domain_level of
774 * the corresponding sched domain.
776 * If root is load-balancing, we can skip @cp if it
777 * is a subset of the root's effective_cpus.
779 if (!cpumask_empty(cp->cpus_allowed) &&
780 !(is_sched_load_balance(cp) &&
781 cpumask_intersects(cp->cpus_allowed,
782 housekeeping_cpumask(HK_FLAG_DOMAIN))))
785 if (root_load_balance &&
786 cpumask_subset(cp->cpus_allowed, top_cpuset.effective_cpus))
789 if (is_sched_load_balance(cp))
792 /* skip @cp's subtree if not a partition root */
793 if (!is_partition_root(cp))
794 pos_css = css_rightmost_descendant(pos_css);
798 for (i = 0; i < csn; i++)
803 /* Find the best partition (set of sched domains) */
804 for (i = 0; i < csn; i++) {
805 struct cpuset *a = csa[i];
808 for (j = 0; j < csn; j++) {
809 struct cpuset *b = csa[j];
812 if (apn != bpn && cpusets_overlap(a, b)) {
813 for (k = 0; k < csn; k++) {
814 struct cpuset *c = csa[k];
819 ndoms--; /* one less element */
826 * Now we know how many domains to create.
827 * Convert <csn, csa> to <ndoms, doms> and populate cpu masks.
829 doms = alloc_sched_domains(ndoms);
834 * The rest of the code, including the scheduler, can deal with
835 * dattr==NULL case. No need to abort if alloc fails.
837 dattr = kmalloc_array(ndoms, sizeof(struct sched_domain_attr),
840 for (nslot = 0, i = 0; i < csn; i++) {
841 struct cpuset *a = csa[i];
846 /* Skip completed partitions */
852 if (nslot == ndoms) {
853 static int warnings = 10;
855 pr_warn("rebuild_sched_domains confused: nslot %d, ndoms %d, csn %d, i %d, apn %d\n",
856 nslot, ndoms, csn, i, apn);
864 *(dattr + nslot) = SD_ATTR_INIT;
865 for (j = i; j < csn; j++) {
866 struct cpuset *b = csa[j];
869 cpumask_or(dp, dp, b->effective_cpus);
870 cpumask_and(dp, dp, housekeeping_cpumask(HK_FLAG_DOMAIN));
872 update_domain_attr_tree(dattr + nslot, b);
874 /* Done with this partition */
880 BUG_ON(nslot != ndoms);
886 * Fallback to the default domain if kmalloc() failed.
887 * See comments in partition_sched_domains().
898 * Rebuild scheduler domains.
900 * If the flag 'sched_load_balance' of any cpuset with non-empty
901 * 'cpus' changes, or if the 'cpus' allowed changes in any cpuset
902 * which has that flag enabled, or if any cpuset with a non-empty
903 * 'cpus' is removed, then call this routine to rebuild the
904 * scheduler's dynamic sched domains.
906 * Call with cpuset_mutex held. Takes get_online_cpus().
908 static void rebuild_sched_domains_locked(void)
910 struct sched_domain_attr *attr;
914 lockdep_assert_held(&cpuset_mutex);
918 * We have raced with CPU hotplug. Don't do anything to avoid
919 * passing doms with offlined cpu to partition_sched_domains().
920 * Anyways, hotplug work item will rebuild sched domains.
922 if (!top_cpuset.nr_subparts_cpus &&
923 !cpumask_equal(top_cpuset.effective_cpus, cpu_active_mask))
926 if (top_cpuset.nr_subparts_cpus &&
927 !cpumask_subset(top_cpuset.effective_cpus, cpu_active_mask))
930 /* Generate domain masks and attrs */
931 ndoms = generate_sched_domains(&doms, &attr);
933 /* Have scheduler rebuild the domains */
934 partition_sched_domains(ndoms, doms, attr);
938 #else /* !CONFIG_SMP */
939 static void rebuild_sched_domains_locked(void)
942 #endif /* CONFIG_SMP */
944 void rebuild_sched_domains(void)
946 mutex_lock(&cpuset_mutex);
947 rebuild_sched_domains_locked();
948 mutex_unlock(&cpuset_mutex);
952 * update_tasks_cpumask - Update the cpumasks of tasks in the cpuset.
953 * @cs: the cpuset in which each task's cpus_allowed mask needs to be changed
955 * Iterate through each task of @cs updating its cpus_allowed to the
956 * effective cpuset's. As this function is called with cpuset_mutex held,
957 * cpuset membership stays stable.
959 static void update_tasks_cpumask(struct cpuset *cs)
961 struct css_task_iter it;
962 struct task_struct *task;
964 css_task_iter_start(&cs->css, 0, &it);
965 while ((task = css_task_iter_next(&it)))
966 set_cpus_allowed_ptr(task, cs->effective_cpus);
967 css_task_iter_end(&it);
971 * compute_effective_cpumask - Compute the effective cpumask of the cpuset
972 * @new_cpus: the temp variable for the new effective_cpus mask
973 * @cs: the cpuset the need to recompute the new effective_cpus mask
974 * @parent: the parent cpuset
976 * If the parent has subpartition CPUs, include them in the list of
977 * allowable CPUs in computing the new effective_cpus mask. Since offlined
978 * CPUs are not removed from subparts_cpus, we have to use cpu_active_mask
981 static void compute_effective_cpumask(struct cpumask *new_cpus,
982 struct cpuset *cs, struct cpuset *parent)
984 if (parent->nr_subparts_cpus) {
985 cpumask_or(new_cpus, parent->effective_cpus,
986 parent->subparts_cpus);
987 cpumask_and(new_cpus, new_cpus, cs->cpus_allowed);
988 cpumask_and(new_cpus, new_cpus, cpu_active_mask);
990 cpumask_and(new_cpus, cs->cpus_allowed, parent->effective_cpus);
995 * Commands for update_parent_subparts_cpumask
998 partcmd_enable, /* Enable partition root */
999 partcmd_disable, /* Disable partition root */
1000 partcmd_update, /* Update parent's subparts_cpus */
1004 * update_parent_subparts_cpumask - update subparts_cpus mask of parent cpuset
1005 * @cpuset: The cpuset that requests change in partition root state
1006 * @cmd: Partition root state change command
1007 * @newmask: Optional new cpumask for partcmd_update
1008 * @tmp: Temporary addmask and delmask
1009 * Return: 0, 1 or an error code
1011 * For partcmd_enable, the cpuset is being transformed from a non-partition
1012 * root to a partition root. The cpus_allowed mask of the given cpuset will
1013 * be put into parent's subparts_cpus and taken away from parent's
1014 * effective_cpus. The function will return 0 if all the CPUs listed in
1015 * cpus_allowed can be granted or an error code will be returned.
1017 * For partcmd_disable, the cpuset is being transofrmed from a partition
1018 * root back to a non-partition root. any CPUs in cpus_allowed that are in
1019 * parent's subparts_cpus will be taken away from that cpumask and put back
1020 * into parent's effective_cpus. 0 should always be returned.
1022 * For partcmd_update, if the optional newmask is specified, the cpu
1023 * list is to be changed from cpus_allowed to newmask. Otherwise,
1024 * cpus_allowed is assumed to remain the same. The cpuset should either
1025 * be a partition root or an invalid partition root. The partition root
1026 * state may change if newmask is NULL and none of the requested CPUs can
1027 * be granted by the parent. The function will return 1 if changes to
1028 * parent's subparts_cpus and effective_cpus happen or 0 otherwise.
1029 * Error code should only be returned when newmask is non-NULL.
1031 * The partcmd_enable and partcmd_disable commands are used by
1032 * update_prstate(). The partcmd_update command is used by
1033 * update_cpumasks_hier() with newmask NULL and update_cpumask() with
1036 * The checking is more strict when enabling partition root than the
1037 * other two commands.
1039 * Because of the implicit cpu exclusive nature of a partition root,
1040 * cpumask changes that violates the cpu exclusivity rule will not be
1041 * permitted when checked by validate_change(). The validate_change()
1042 * function will also prevent any changes to the cpu list if it is not
1043 * a superset of children's cpu lists.
1045 static int update_parent_subparts_cpumask(struct cpuset *cpuset, int cmd,
1046 struct cpumask *newmask,
1047 struct tmpmasks *tmp)
1049 struct cpuset *parent = parent_cs(cpuset);
1050 int adding; /* Moving cpus from effective_cpus to subparts_cpus */
1051 int deleting; /* Moving cpus from subparts_cpus to effective_cpus */
1052 bool part_error = false; /* Partition error? */
1054 lockdep_assert_held(&cpuset_mutex);
1057 * The parent must be a partition root.
1058 * The new cpumask, if present, or the current cpus_allowed must
1061 if (!is_partition_root(parent) ||
1062 (newmask && cpumask_empty(newmask)) ||
1063 (!newmask && cpumask_empty(cpuset->cpus_allowed)))
1067 * Enabling/disabling partition root is not allowed if there are
1070 if ((cmd != partcmd_update) && css_has_online_children(&cpuset->css))
1074 * Enabling partition root is not allowed if not all the CPUs
1075 * can be granted from parent's effective_cpus or at least one
1076 * CPU will be left after that.
1078 if ((cmd == partcmd_enable) &&
1079 (!cpumask_subset(cpuset->cpus_allowed, parent->effective_cpus) ||
1080 cpumask_equal(cpuset->cpus_allowed, parent->effective_cpus)))
1084 * A cpumask update cannot make parent's effective_cpus become empty.
1086 adding = deleting = false;
1087 if (cmd == partcmd_enable) {
1088 cpumask_copy(tmp->addmask, cpuset->cpus_allowed);
1090 } else if (cmd == partcmd_disable) {
1091 deleting = cpumask_and(tmp->delmask, cpuset->cpus_allowed,
1092 parent->subparts_cpus);
1093 } else if (newmask) {
1095 * partcmd_update with newmask:
1097 * delmask = cpus_allowed & ~newmask & parent->subparts_cpus
1098 * addmask = newmask & parent->effective_cpus
1099 * & ~parent->subparts_cpus
1101 cpumask_andnot(tmp->delmask, cpuset->cpus_allowed, newmask);
1102 deleting = cpumask_and(tmp->delmask, tmp->delmask,
1103 parent->subparts_cpus);
1105 cpumask_and(tmp->addmask, newmask, parent->effective_cpus);
1106 adding = cpumask_andnot(tmp->addmask, tmp->addmask,
1107 parent->subparts_cpus);
1109 * Return error if the new effective_cpus could become empty.
1112 cpumask_equal(parent->effective_cpus, tmp->addmask)) {
1116 * As some of the CPUs in subparts_cpus might have
1117 * been offlined, we need to compute the real delmask
1120 if (!cpumask_and(tmp->addmask, tmp->delmask,
1123 cpumask_copy(tmp->addmask, parent->effective_cpus);
1127 * partcmd_update w/o newmask:
1129 * addmask = cpus_allowed & parent->effectiveb_cpus
1131 * Note that parent's subparts_cpus may have been
1132 * pre-shrunk in case there is a change in the cpu list.
1133 * So no deletion is needed.
1135 adding = cpumask_and(tmp->addmask, cpuset->cpus_allowed,
1136 parent->effective_cpus);
1137 part_error = cpumask_equal(tmp->addmask,
1138 parent->effective_cpus);
1141 if (cmd == partcmd_update) {
1142 int prev_prs = cpuset->partition_root_state;
1145 * Check for possible transition between PRS_ENABLED
1148 switch (cpuset->partition_root_state) {
1151 cpuset->partition_root_state = PRS_ERROR;
1155 cpuset->partition_root_state = PRS_ENABLED;
1159 * Set part_error if previously in invalid state.
1161 part_error = (prev_prs == PRS_ERROR);
1164 if (!part_error && (cpuset->partition_root_state == PRS_ERROR))
1165 return 0; /* Nothing need to be done */
1167 if (cpuset->partition_root_state == PRS_ERROR) {
1169 * Remove all its cpus from parent's subparts_cpus.
1172 deleting = cpumask_and(tmp->delmask, cpuset->cpus_allowed,
1173 parent->subparts_cpus);
1176 if (!adding && !deleting)
1180 * Change the parent's subparts_cpus.
1181 * Newly added CPUs will be removed from effective_cpus and
1182 * newly deleted ones will be added back to effective_cpus.
1184 spin_lock_irq(&callback_lock);
1186 cpumask_or(parent->subparts_cpus,
1187 parent->subparts_cpus, tmp->addmask);
1188 cpumask_andnot(parent->effective_cpus,
1189 parent->effective_cpus, tmp->addmask);
1192 cpumask_andnot(parent->subparts_cpus,
1193 parent->subparts_cpus, tmp->delmask);
1195 * Some of the CPUs in subparts_cpus might have been offlined.
1197 cpumask_and(tmp->delmask, tmp->delmask, cpu_active_mask);
1198 cpumask_or(parent->effective_cpus,
1199 parent->effective_cpus, tmp->delmask);
1202 parent->nr_subparts_cpus = cpumask_weight(parent->subparts_cpus);
1203 spin_unlock_irq(&callback_lock);
1205 return cmd == partcmd_update;
1209 * update_cpumasks_hier - Update effective cpumasks and tasks in the subtree
1210 * @cs: the cpuset to consider
1211 * @tmp: temp variables for calculating effective_cpus & partition setup
1213 * When congifured cpumask is changed, the effective cpumasks of this cpuset
1214 * and all its descendants need to be updated.
1216 * On legacy hierachy, effective_cpus will be the same with cpu_allowed.
1218 * Called with cpuset_mutex held
1220 static void update_cpumasks_hier(struct cpuset *cs, struct tmpmasks *tmp)
1223 struct cgroup_subsys_state *pos_css;
1224 bool need_rebuild_sched_domains = false;
1227 cpuset_for_each_descendant_pre(cp, pos_css, cs) {
1228 struct cpuset *parent = parent_cs(cp);
1230 compute_effective_cpumask(tmp->new_cpus, cp, parent);
1233 * If it becomes empty, inherit the effective mask of the
1234 * parent, which is guaranteed to have some CPUs.
1236 if (is_in_v2_mode() && cpumask_empty(tmp->new_cpus)) {
1237 cpumask_copy(tmp->new_cpus, parent->effective_cpus);
1238 if (!cp->use_parent_ecpus) {
1239 cp->use_parent_ecpus = true;
1240 parent->child_ecpus_count++;
1242 } else if (cp->use_parent_ecpus) {
1243 cp->use_parent_ecpus = false;
1244 WARN_ON_ONCE(!parent->child_ecpus_count);
1245 parent->child_ecpus_count--;
1249 * Skip the whole subtree if the cpumask remains the same
1250 * and has no partition root state.
1252 if (!cp->partition_root_state &&
1253 cpumask_equal(tmp->new_cpus, cp->effective_cpus)) {
1254 pos_css = css_rightmost_descendant(pos_css);
1259 * update_parent_subparts_cpumask() should have been called
1260 * for cs already in update_cpumask(). We should also call
1261 * update_tasks_cpumask() again for tasks in the parent
1262 * cpuset if the parent's subparts_cpus changes.
1264 if ((cp != cs) && cp->partition_root_state) {
1265 switch (parent->partition_root_state) {
1268 * If parent is not a partition root or an
1269 * invalid partition root, clear the state
1270 * state and the CS_CPU_EXCLUSIVE flag.
1272 WARN_ON_ONCE(cp->partition_root_state
1274 cp->partition_root_state = 0;
1277 * clear_bit() is an atomic operation and
1278 * readers aren't interested in the state
1279 * of CS_CPU_EXCLUSIVE anyway. So we can
1280 * just update the flag without holding
1281 * the callback_lock.
1283 clear_bit(CS_CPU_EXCLUSIVE, &cp->flags);
1287 if (update_parent_subparts_cpumask(cp, partcmd_update, NULL, tmp))
1288 update_tasks_cpumask(parent);
1293 * When parent is invalid, it has to be too.
1295 cp->partition_root_state = PRS_ERROR;
1296 if (cp->nr_subparts_cpus) {
1297 cp->nr_subparts_cpus = 0;
1298 cpumask_clear(cp->subparts_cpus);
1304 if (!css_tryget_online(&cp->css))
1308 spin_lock_irq(&callback_lock);
1310 cpumask_copy(cp->effective_cpus, tmp->new_cpus);
1311 if (cp->nr_subparts_cpus &&
1312 (cp->partition_root_state != PRS_ENABLED)) {
1313 cp->nr_subparts_cpus = 0;
1314 cpumask_clear(cp->subparts_cpus);
1315 } else if (cp->nr_subparts_cpus) {
1317 * Make sure that effective_cpus & subparts_cpus
1318 * are mutually exclusive.
1320 * In the unlikely event that effective_cpus
1321 * becomes empty. we clear cp->nr_subparts_cpus and
1322 * let its child partition roots to compete for
1325 cpumask_andnot(cp->effective_cpus, cp->effective_cpus,
1327 if (cpumask_empty(cp->effective_cpus)) {
1328 cpumask_copy(cp->effective_cpus, tmp->new_cpus);
1329 cpumask_clear(cp->subparts_cpus);
1330 cp->nr_subparts_cpus = 0;
1331 } else if (!cpumask_subset(cp->subparts_cpus,
1333 cpumask_andnot(cp->subparts_cpus,
1334 cp->subparts_cpus, tmp->new_cpus);
1335 cp->nr_subparts_cpus
1336 = cpumask_weight(cp->subparts_cpus);
1339 spin_unlock_irq(&callback_lock);
1341 WARN_ON(!is_in_v2_mode() &&
1342 !cpumask_equal(cp->cpus_allowed, cp->effective_cpus));
1344 update_tasks_cpumask(cp);
1347 * On legacy hierarchy, if the effective cpumask of any non-
1348 * empty cpuset is changed, we need to rebuild sched domains.
1349 * On default hierarchy, the cpuset needs to be a partition
1352 if (!cpumask_empty(cp->cpus_allowed) &&
1353 is_sched_load_balance(cp) &&
1354 (!cgroup_subsys_on_dfl(cpuset_cgrp_subsys) ||
1355 is_partition_root(cp)))
1356 need_rebuild_sched_domains = true;
1363 if (need_rebuild_sched_domains)
1364 rebuild_sched_domains_locked();
1368 * update_sibling_cpumasks - Update siblings cpumasks
1369 * @parent: Parent cpuset
1370 * @cs: Current cpuset
1371 * @tmp: Temp variables
1373 static void update_sibling_cpumasks(struct cpuset *parent, struct cpuset *cs,
1374 struct tmpmasks *tmp)
1376 struct cpuset *sibling;
1377 struct cgroup_subsys_state *pos_css;
1380 * Check all its siblings and call update_cpumasks_hier()
1381 * if their use_parent_ecpus flag is set in order for them
1382 * to use the right effective_cpus value.
1385 cpuset_for_each_child(sibling, pos_css, parent) {
1388 if (!sibling->use_parent_ecpus)
1391 update_cpumasks_hier(sibling, tmp);
1397 * update_cpumask - update the cpus_allowed mask of a cpuset and all tasks in it
1398 * @cs: the cpuset to consider
1399 * @trialcs: trial cpuset
1400 * @buf: buffer of cpu numbers written to this cpuset
1402 static int update_cpumask(struct cpuset *cs, struct cpuset *trialcs,
1406 struct tmpmasks tmp;
1408 /* top_cpuset.cpus_allowed tracks cpu_online_mask; it's read-only */
1409 if (cs == &top_cpuset)
1413 * An empty cpus_allowed is ok only if the cpuset has no tasks.
1414 * Since cpulist_parse() fails on an empty mask, we special case
1415 * that parsing. The validate_change() call ensures that cpusets
1416 * with tasks have cpus.
1419 cpumask_clear(trialcs->cpus_allowed);
1421 retval = cpulist_parse(buf, trialcs->cpus_allowed);
1425 if (!cpumask_subset(trialcs->cpus_allowed,
1426 top_cpuset.cpus_allowed))
1430 /* Nothing to do if the cpus didn't change */
1431 if (cpumask_equal(cs->cpus_allowed, trialcs->cpus_allowed))
1434 retval = validate_change(cs, trialcs);
1438 #ifdef CONFIG_CPUMASK_OFFSTACK
1440 * Use the cpumasks in trialcs for tmpmasks when they are pointers
1441 * to allocated cpumasks.
1443 tmp.addmask = trialcs->subparts_cpus;
1444 tmp.delmask = trialcs->effective_cpus;
1445 tmp.new_cpus = trialcs->cpus_allowed;
1448 if (cs->partition_root_state) {
1449 /* Cpumask of a partition root cannot be empty */
1450 if (cpumask_empty(trialcs->cpus_allowed))
1452 if (update_parent_subparts_cpumask(cs, partcmd_update,
1453 trialcs->cpus_allowed, &tmp) < 0)
1457 spin_lock_irq(&callback_lock);
1458 cpumask_copy(cs->cpus_allowed, trialcs->cpus_allowed);
1461 * Make sure that subparts_cpus is a subset of cpus_allowed.
1463 if (cs->nr_subparts_cpus) {
1464 cpumask_andnot(cs->subparts_cpus, cs->subparts_cpus,
1466 cs->nr_subparts_cpus = cpumask_weight(cs->subparts_cpus);
1468 spin_unlock_irq(&callback_lock);
1470 update_cpumasks_hier(cs, &tmp);
1472 if (cs->partition_root_state) {
1473 struct cpuset *parent = parent_cs(cs);
1476 * For partition root, update the cpumasks of sibling
1477 * cpusets if they use parent's effective_cpus.
1479 if (parent->child_ecpus_count)
1480 update_sibling_cpumasks(parent, cs, &tmp);
1486 * Migrate memory region from one set of nodes to another. This is
1487 * performed asynchronously as it can be called from process migration path
1488 * holding locks involved in process management. All mm migrations are
1489 * performed in the queued order and can be waited for by flushing
1490 * cpuset_migrate_mm_wq.
1493 struct cpuset_migrate_mm_work {
1494 struct work_struct work;
1495 struct mm_struct *mm;
1500 static void cpuset_migrate_mm_workfn(struct work_struct *work)
1502 struct cpuset_migrate_mm_work *mwork =
1503 container_of(work, struct cpuset_migrate_mm_work, work);
1505 /* on a wq worker, no need to worry about %current's mems_allowed */
1506 do_migrate_pages(mwork->mm, &mwork->from, &mwork->to, MPOL_MF_MOVE_ALL);
1511 static void cpuset_migrate_mm(struct mm_struct *mm, const nodemask_t *from,
1512 const nodemask_t *to)
1514 struct cpuset_migrate_mm_work *mwork;
1516 mwork = kzalloc(sizeof(*mwork), GFP_KERNEL);
1519 mwork->from = *from;
1521 INIT_WORK(&mwork->work, cpuset_migrate_mm_workfn);
1522 queue_work(cpuset_migrate_mm_wq, &mwork->work);
1528 static void cpuset_post_attach(void)
1530 flush_workqueue(cpuset_migrate_mm_wq);
1534 * cpuset_change_task_nodemask - change task's mems_allowed and mempolicy
1535 * @tsk: the task to change
1536 * @newmems: new nodes that the task will be set
1538 * We use the mems_allowed_seq seqlock to safely update both tsk->mems_allowed
1539 * and rebind an eventual tasks' mempolicy. If the task is allocating in
1540 * parallel, it might temporarily see an empty intersection, which results in
1541 * a seqlock check and retry before OOM or allocation failure.
1543 static void cpuset_change_task_nodemask(struct task_struct *tsk,
1544 nodemask_t *newmems)
1548 local_irq_disable();
1549 write_seqcount_begin(&tsk->mems_allowed_seq);
1551 nodes_or(tsk->mems_allowed, tsk->mems_allowed, *newmems);
1552 mpol_rebind_task(tsk, newmems);
1553 tsk->mems_allowed = *newmems;
1555 write_seqcount_end(&tsk->mems_allowed_seq);
1561 static void *cpuset_being_rebound;
1564 * update_tasks_nodemask - Update the nodemasks of tasks in the cpuset.
1565 * @cs: the cpuset in which each task's mems_allowed mask needs to be changed
1567 * Iterate through each task of @cs updating its mems_allowed to the
1568 * effective cpuset's. As this function is called with cpuset_mutex held,
1569 * cpuset membership stays stable.
1571 static void update_tasks_nodemask(struct cpuset *cs)
1573 static nodemask_t newmems; /* protected by cpuset_mutex */
1574 struct css_task_iter it;
1575 struct task_struct *task;
1577 cpuset_being_rebound = cs; /* causes mpol_dup() rebind */
1579 guarantee_online_mems(cs, &newmems);
1582 * The mpol_rebind_mm() call takes mmap_sem, which we couldn't
1583 * take while holding tasklist_lock. Forks can happen - the
1584 * mpol_dup() cpuset_being_rebound check will catch such forks,
1585 * and rebind their vma mempolicies too. Because we still hold
1586 * the global cpuset_mutex, we know that no other rebind effort
1587 * will be contending for the global variable cpuset_being_rebound.
1588 * It's ok if we rebind the same mm twice; mpol_rebind_mm()
1589 * is idempotent. Also migrate pages in each mm to new nodes.
1591 css_task_iter_start(&cs->css, 0, &it);
1592 while ((task = css_task_iter_next(&it))) {
1593 struct mm_struct *mm;
1596 cpuset_change_task_nodemask(task, &newmems);
1598 mm = get_task_mm(task);
1602 migrate = is_memory_migrate(cs);
1604 mpol_rebind_mm(mm, &cs->mems_allowed);
1606 cpuset_migrate_mm(mm, &cs->old_mems_allowed, &newmems);
1610 css_task_iter_end(&it);
1613 * All the tasks' nodemasks have been updated, update
1614 * cs->old_mems_allowed.
1616 cs->old_mems_allowed = newmems;
1618 /* We're done rebinding vmas to this cpuset's new mems_allowed. */
1619 cpuset_being_rebound = NULL;
1623 * update_nodemasks_hier - Update effective nodemasks and tasks in the subtree
1624 * @cs: the cpuset to consider
1625 * @new_mems: a temp variable for calculating new effective_mems
1627 * When configured nodemask is changed, the effective nodemasks of this cpuset
1628 * and all its descendants need to be updated.
1630 * On legacy hiearchy, effective_mems will be the same with mems_allowed.
1632 * Called with cpuset_mutex held
1634 static void update_nodemasks_hier(struct cpuset *cs, nodemask_t *new_mems)
1637 struct cgroup_subsys_state *pos_css;
1640 cpuset_for_each_descendant_pre(cp, pos_css, cs) {
1641 struct cpuset *parent = parent_cs(cp);
1643 nodes_and(*new_mems, cp->mems_allowed, parent->effective_mems);
1646 * If it becomes empty, inherit the effective mask of the
1647 * parent, which is guaranteed to have some MEMs.
1649 if (is_in_v2_mode() && nodes_empty(*new_mems))
1650 *new_mems = parent->effective_mems;
1652 /* Skip the whole subtree if the nodemask remains the same. */
1653 if (nodes_equal(*new_mems, cp->effective_mems)) {
1654 pos_css = css_rightmost_descendant(pos_css);
1658 if (!css_tryget_online(&cp->css))
1662 spin_lock_irq(&callback_lock);
1663 cp->effective_mems = *new_mems;
1664 spin_unlock_irq(&callback_lock);
1666 WARN_ON(!is_in_v2_mode() &&
1667 !nodes_equal(cp->mems_allowed, cp->effective_mems));
1669 update_tasks_nodemask(cp);
1678 * Handle user request to change the 'mems' memory placement
1679 * of a cpuset. Needs to validate the request, update the
1680 * cpusets mems_allowed, and for each task in the cpuset,
1681 * update mems_allowed and rebind task's mempolicy and any vma
1682 * mempolicies and if the cpuset is marked 'memory_migrate',
1683 * migrate the tasks pages to the new memory.
1685 * Call with cpuset_mutex held. May take callback_lock during call.
1686 * Will take tasklist_lock, scan tasklist for tasks in cpuset cs,
1687 * lock each such tasks mm->mmap_sem, scan its vma's and rebind
1688 * their mempolicies to the cpusets new mems_allowed.
1690 static int update_nodemask(struct cpuset *cs, struct cpuset *trialcs,
1696 * top_cpuset.mems_allowed tracks node_stats[N_MEMORY];
1699 if (cs == &top_cpuset) {
1705 * An empty mems_allowed is ok iff there are no tasks in the cpuset.
1706 * Since nodelist_parse() fails on an empty mask, we special case
1707 * that parsing. The validate_change() call ensures that cpusets
1708 * with tasks have memory.
1711 nodes_clear(trialcs->mems_allowed);
1713 retval = nodelist_parse(buf, trialcs->mems_allowed);
1717 if (!nodes_subset(trialcs->mems_allowed,
1718 top_cpuset.mems_allowed)) {
1724 if (nodes_equal(cs->mems_allowed, trialcs->mems_allowed)) {
1725 retval = 0; /* Too easy - nothing to do */
1728 retval = validate_change(cs, trialcs);
1732 spin_lock_irq(&callback_lock);
1733 cs->mems_allowed = trialcs->mems_allowed;
1734 spin_unlock_irq(&callback_lock);
1736 /* use trialcs->mems_allowed as a temp variable */
1737 update_nodemasks_hier(cs, &trialcs->mems_allowed);
1742 bool current_cpuset_is_being_rebound(void)
1747 ret = task_cs(current) == cpuset_being_rebound;
1753 static int update_relax_domain_level(struct cpuset *cs, s64 val)
1756 if (val < -1 || val >= sched_domain_level_max)
1760 if (val != cs->relax_domain_level) {
1761 cs->relax_domain_level = val;
1762 if (!cpumask_empty(cs->cpus_allowed) &&
1763 is_sched_load_balance(cs))
1764 rebuild_sched_domains_locked();
1771 * update_tasks_flags - update the spread flags of tasks in the cpuset.
1772 * @cs: the cpuset in which each task's spread flags needs to be changed
1774 * Iterate through each task of @cs updating its spread flags. As this
1775 * function is called with cpuset_mutex held, cpuset membership stays
1778 static void update_tasks_flags(struct cpuset *cs)
1780 struct css_task_iter it;
1781 struct task_struct *task;
1783 css_task_iter_start(&cs->css, 0, &it);
1784 while ((task = css_task_iter_next(&it)))
1785 cpuset_update_task_spread_flag(cs, task);
1786 css_task_iter_end(&it);
1790 * update_flag - read a 0 or a 1 in a file and update associated flag
1791 * bit: the bit to update (see cpuset_flagbits_t)
1792 * cs: the cpuset to update
1793 * turning_on: whether the flag is being set or cleared
1795 * Call with cpuset_mutex held.
1798 static int update_flag(cpuset_flagbits_t bit, struct cpuset *cs,
1801 struct cpuset *trialcs;
1802 int balance_flag_changed;
1803 int spread_flag_changed;
1806 trialcs = alloc_trial_cpuset(cs);
1811 set_bit(bit, &trialcs->flags);
1813 clear_bit(bit, &trialcs->flags);
1815 err = validate_change(cs, trialcs);
1819 balance_flag_changed = (is_sched_load_balance(cs) !=
1820 is_sched_load_balance(trialcs));
1822 spread_flag_changed = ((is_spread_slab(cs) != is_spread_slab(trialcs))
1823 || (is_spread_page(cs) != is_spread_page(trialcs)));
1825 spin_lock_irq(&callback_lock);
1826 cs->flags = trialcs->flags;
1827 spin_unlock_irq(&callback_lock);
1829 if (!cpumask_empty(trialcs->cpus_allowed) && balance_flag_changed)
1830 rebuild_sched_domains_locked();
1832 if (spread_flag_changed)
1833 update_tasks_flags(cs);
1835 free_cpuset(trialcs);
1840 * update_prstate - update partititon_root_state
1841 * cs: the cpuset to update
1842 * val: 0 - disabled, 1 - enabled
1844 * Call with cpuset_mutex held.
1846 static int update_prstate(struct cpuset *cs, int val)
1849 struct cpuset *parent = parent_cs(cs);
1850 struct tmpmasks tmp;
1852 if ((val != 0) && (val != 1))
1854 if (val == cs->partition_root_state)
1858 * Cannot force a partial or invalid partition root to a full
1861 if (val && cs->partition_root_state)
1864 if (alloc_cpumasks(NULL, &tmp))
1868 if (!cs->partition_root_state) {
1870 * Turning on partition root requires setting the
1871 * CS_CPU_EXCLUSIVE bit implicitly as well and cpus_allowed
1874 if (cpumask_empty(cs->cpus_allowed))
1877 err = update_flag(CS_CPU_EXCLUSIVE, cs, 1);
1881 err = update_parent_subparts_cpumask(cs, partcmd_enable,
1884 update_flag(CS_CPU_EXCLUSIVE, cs, 0);
1887 cs->partition_root_state = PRS_ENABLED;
1890 * Turning off partition root will clear the
1891 * CS_CPU_EXCLUSIVE bit.
1893 if (cs->partition_root_state == PRS_ERROR) {
1894 cs->partition_root_state = 0;
1895 update_flag(CS_CPU_EXCLUSIVE, cs, 0);
1900 err = update_parent_subparts_cpumask(cs, partcmd_disable,
1905 cs->partition_root_state = 0;
1907 /* Turning off CS_CPU_EXCLUSIVE will not return error */
1908 update_flag(CS_CPU_EXCLUSIVE, cs, 0);
1912 * Update cpumask of parent's tasks except when it is the top
1913 * cpuset as some system daemons cannot be mapped to other CPUs.
1915 if (parent != &top_cpuset)
1916 update_tasks_cpumask(parent);
1918 if (parent->child_ecpus_count)
1919 update_sibling_cpumasks(parent, cs, &tmp);
1921 rebuild_sched_domains_locked();
1923 free_cpumasks(NULL, &tmp);
1928 * Frequency meter - How fast is some event occurring?
1930 * These routines manage a digitally filtered, constant time based,
1931 * event frequency meter. There are four routines:
1932 * fmeter_init() - initialize a frequency meter.
1933 * fmeter_markevent() - called each time the event happens.
1934 * fmeter_getrate() - returns the recent rate of such events.
1935 * fmeter_update() - internal routine used to update fmeter.
1937 * A common data structure is passed to each of these routines,
1938 * which is used to keep track of the state required to manage the
1939 * frequency meter and its digital filter.
1941 * The filter works on the number of events marked per unit time.
1942 * The filter is single-pole low-pass recursive (IIR). The time unit
1943 * is 1 second. Arithmetic is done using 32-bit integers scaled to
1944 * simulate 3 decimal digits of precision (multiplied by 1000).
1946 * With an FM_COEF of 933, and a time base of 1 second, the filter
1947 * has a half-life of 10 seconds, meaning that if the events quit
1948 * happening, then the rate returned from the fmeter_getrate()
1949 * will be cut in half each 10 seconds, until it converges to zero.
1951 * It is not worth doing a real infinitely recursive filter. If more
1952 * than FM_MAXTICKS ticks have elapsed since the last filter event,
1953 * just compute FM_MAXTICKS ticks worth, by which point the level
1956 * Limit the count of unprocessed events to FM_MAXCNT, so as to avoid
1957 * arithmetic overflow in the fmeter_update() routine.
1959 * Given the simple 32 bit integer arithmetic used, this meter works
1960 * best for reporting rates between one per millisecond (msec) and
1961 * one per 32 (approx) seconds. At constant rates faster than one
1962 * per msec it maxes out at values just under 1,000,000. At constant
1963 * rates between one per msec, and one per second it will stabilize
1964 * to a value N*1000, where N is the rate of events per second.
1965 * At constant rates between one per second and one per 32 seconds,
1966 * it will be choppy, moving up on the seconds that have an event,
1967 * and then decaying until the next event. At rates slower than
1968 * about one in 32 seconds, it decays all the way back to zero between
1972 #define FM_COEF 933 /* coefficient for half-life of 10 secs */
1973 #define FM_MAXTICKS ((u32)99) /* useless computing more ticks than this */
1974 #define FM_MAXCNT 1000000 /* limit cnt to avoid overflow */
1975 #define FM_SCALE 1000 /* faux fixed point scale */
1977 /* Initialize a frequency meter */
1978 static void fmeter_init(struct fmeter *fmp)
1983 spin_lock_init(&fmp->lock);
1986 /* Internal meter update - process cnt events and update value */
1987 static void fmeter_update(struct fmeter *fmp)
1992 now = ktime_get_seconds();
1993 ticks = now - fmp->time;
1998 ticks = min(FM_MAXTICKS, ticks);
2000 fmp->val = (FM_COEF * fmp->val) / FM_SCALE;
2003 fmp->val += ((FM_SCALE - FM_COEF) * fmp->cnt) / FM_SCALE;
2007 /* Process any previous ticks, then bump cnt by one (times scale). */
2008 static void fmeter_markevent(struct fmeter *fmp)
2010 spin_lock(&fmp->lock);
2012 fmp->cnt = min(FM_MAXCNT, fmp->cnt + FM_SCALE);
2013 spin_unlock(&fmp->lock);
2016 /* Process any previous ticks, then return current value. */
2017 static int fmeter_getrate(struct fmeter *fmp)
2021 spin_lock(&fmp->lock);
2024 spin_unlock(&fmp->lock);
2028 static struct cpuset *cpuset_attach_old_cs;
2030 /* Called by cgroups to determine if a cpuset is usable; cpuset_mutex held */
2031 static int cpuset_can_attach(struct cgroup_taskset *tset)
2033 struct cgroup_subsys_state *css;
2035 struct task_struct *task;
2038 /* used later by cpuset_attach() */
2039 cpuset_attach_old_cs = task_cs(cgroup_taskset_first(tset, &css));
2042 mutex_lock(&cpuset_mutex);
2044 /* allow moving tasks into an empty cpuset if on default hierarchy */
2046 if (!is_in_v2_mode() &&
2047 (cpumask_empty(cs->cpus_allowed) || nodes_empty(cs->mems_allowed)))
2050 cgroup_taskset_for_each(task, css, tset) {
2051 ret = task_can_attach(task, cs->cpus_allowed);
2054 ret = security_task_setscheduler(task);
2060 * Mark attach is in progress. This makes validate_change() fail
2061 * changes which zero cpus/mems_allowed.
2063 cs->attach_in_progress++;
2066 mutex_unlock(&cpuset_mutex);
2070 static void cpuset_cancel_attach(struct cgroup_taskset *tset)
2072 struct cgroup_subsys_state *css;
2074 cgroup_taskset_first(tset, &css);
2076 mutex_lock(&cpuset_mutex);
2077 css_cs(css)->attach_in_progress--;
2078 mutex_unlock(&cpuset_mutex);
2082 * Protected by cpuset_mutex. cpus_attach is used only by cpuset_attach()
2083 * but we can't allocate it dynamically there. Define it global and
2084 * allocate from cpuset_init().
2086 static cpumask_var_t cpus_attach;
2088 static void cpuset_attach(struct cgroup_taskset *tset)
2090 /* static buf protected by cpuset_mutex */
2091 static nodemask_t cpuset_attach_nodemask_to;
2092 struct task_struct *task;
2093 struct task_struct *leader;
2094 struct cgroup_subsys_state *css;
2096 struct cpuset *oldcs = cpuset_attach_old_cs;
2098 cgroup_taskset_first(tset, &css);
2101 mutex_lock(&cpuset_mutex);
2103 /* prepare for attach */
2104 if (cs == &top_cpuset)
2105 cpumask_copy(cpus_attach, cpu_possible_mask);
2107 guarantee_online_cpus(cs, cpus_attach);
2109 guarantee_online_mems(cs, &cpuset_attach_nodemask_to);
2111 cgroup_taskset_for_each(task, css, tset) {
2113 * can_attach beforehand should guarantee that this doesn't
2114 * fail. TODO: have a better way to handle failure here
2116 WARN_ON_ONCE(set_cpus_allowed_ptr(task, cpus_attach));
2118 cpuset_change_task_nodemask(task, &cpuset_attach_nodemask_to);
2119 cpuset_update_task_spread_flag(cs, task);
2123 * Change mm for all threadgroup leaders. This is expensive and may
2124 * sleep and should be moved outside migration path proper.
2126 cpuset_attach_nodemask_to = cs->effective_mems;
2127 cgroup_taskset_for_each_leader(leader, css, tset) {
2128 struct mm_struct *mm = get_task_mm(leader);
2131 mpol_rebind_mm(mm, &cpuset_attach_nodemask_to);
2134 * old_mems_allowed is the same with mems_allowed
2135 * here, except if this task is being moved
2136 * automatically due to hotplug. In that case
2137 * @mems_allowed has been updated and is empty, so
2138 * @old_mems_allowed is the right nodesets that we
2141 if (is_memory_migrate(cs))
2142 cpuset_migrate_mm(mm, &oldcs->old_mems_allowed,
2143 &cpuset_attach_nodemask_to);
2149 cs->old_mems_allowed = cpuset_attach_nodemask_to;
2151 cs->attach_in_progress--;
2152 if (!cs->attach_in_progress)
2153 wake_up(&cpuset_attach_wq);
2155 mutex_unlock(&cpuset_mutex);
2158 /* The various types of files and directories in a cpuset file system */
2161 FILE_MEMORY_MIGRATE,
2164 FILE_EFFECTIVE_CPULIST,
2165 FILE_EFFECTIVE_MEMLIST,
2166 FILE_SUBPARTS_CPULIST,
2170 FILE_SCHED_LOAD_BALANCE,
2171 FILE_PARTITION_ROOT,
2172 FILE_SCHED_RELAX_DOMAIN_LEVEL,
2173 FILE_MEMORY_PRESSURE_ENABLED,
2174 FILE_MEMORY_PRESSURE,
2177 } cpuset_filetype_t;
2179 static int cpuset_write_u64(struct cgroup_subsys_state *css, struct cftype *cft,
2182 struct cpuset *cs = css_cs(css);
2183 cpuset_filetype_t type = cft->private;
2186 mutex_lock(&cpuset_mutex);
2187 if (!is_cpuset_online(cs)) {
2193 case FILE_CPU_EXCLUSIVE:
2194 retval = update_flag(CS_CPU_EXCLUSIVE, cs, val);
2196 case FILE_MEM_EXCLUSIVE:
2197 retval = update_flag(CS_MEM_EXCLUSIVE, cs, val);
2199 case FILE_MEM_HARDWALL:
2200 retval = update_flag(CS_MEM_HARDWALL, cs, val);
2202 case FILE_SCHED_LOAD_BALANCE:
2203 retval = update_flag(CS_SCHED_LOAD_BALANCE, cs, val);
2205 case FILE_MEMORY_MIGRATE:
2206 retval = update_flag(CS_MEMORY_MIGRATE, cs, val);
2208 case FILE_MEMORY_PRESSURE_ENABLED:
2209 cpuset_memory_pressure_enabled = !!val;
2211 case FILE_SPREAD_PAGE:
2212 retval = update_flag(CS_SPREAD_PAGE, cs, val);
2214 case FILE_SPREAD_SLAB:
2215 retval = update_flag(CS_SPREAD_SLAB, cs, val);
2222 mutex_unlock(&cpuset_mutex);
2226 static int cpuset_write_s64(struct cgroup_subsys_state *css, struct cftype *cft,
2229 struct cpuset *cs = css_cs(css);
2230 cpuset_filetype_t type = cft->private;
2231 int retval = -ENODEV;
2233 mutex_lock(&cpuset_mutex);
2234 if (!is_cpuset_online(cs))
2238 case FILE_SCHED_RELAX_DOMAIN_LEVEL:
2239 retval = update_relax_domain_level(cs, val);
2246 mutex_unlock(&cpuset_mutex);
2251 * Common handling for a write to a "cpus" or "mems" file.
2253 static ssize_t cpuset_write_resmask(struct kernfs_open_file *of,
2254 char *buf, size_t nbytes, loff_t off)
2256 struct cpuset *cs = css_cs(of_css(of));
2257 struct cpuset *trialcs;
2258 int retval = -ENODEV;
2260 buf = strstrip(buf);
2263 * CPU or memory hotunplug may leave @cs w/o any execution
2264 * resources, in which case the hotplug code asynchronously updates
2265 * configuration and transfers all tasks to the nearest ancestor
2266 * which can execute.
2268 * As writes to "cpus" or "mems" may restore @cs's execution
2269 * resources, wait for the previously scheduled operations before
2270 * proceeding, so that we don't end up keep removing tasks added
2271 * after execution capability is restored.
2273 * cpuset_hotplug_work calls back into cgroup core via
2274 * cgroup_transfer_tasks() and waiting for it from a cgroupfs
2275 * operation like this one can lead to a deadlock through kernfs
2276 * active_ref protection. Let's break the protection. Losing the
2277 * protection is okay as we check whether @cs is online after
2278 * grabbing cpuset_mutex anyway. This only happens on the legacy
2282 kernfs_break_active_protection(of->kn);
2283 flush_work(&cpuset_hotplug_work);
2285 mutex_lock(&cpuset_mutex);
2286 if (!is_cpuset_online(cs))
2289 trialcs = alloc_trial_cpuset(cs);
2295 switch (of_cft(of)->private) {
2297 retval = update_cpumask(cs, trialcs, buf);
2300 retval = update_nodemask(cs, trialcs, buf);
2307 free_cpuset(trialcs);
2309 mutex_unlock(&cpuset_mutex);
2310 kernfs_unbreak_active_protection(of->kn);
2312 flush_workqueue(cpuset_migrate_mm_wq);
2313 return retval ?: nbytes;
2317 * These ascii lists should be read in a single call, by using a user
2318 * buffer large enough to hold the entire map. If read in smaller
2319 * chunks, there is no guarantee of atomicity. Since the display format
2320 * used, list of ranges of sequential numbers, is variable length,
2321 * and since these maps can change value dynamically, one could read
2322 * gibberish by doing partial reads while a list was changing.
2324 static int cpuset_common_seq_show(struct seq_file *sf, void *v)
2326 struct cpuset *cs = css_cs(seq_css(sf));
2327 cpuset_filetype_t type = seq_cft(sf)->private;
2330 spin_lock_irq(&callback_lock);
2334 seq_printf(sf, "%*pbl\n", cpumask_pr_args(cs->cpus_allowed));
2337 seq_printf(sf, "%*pbl\n", nodemask_pr_args(&cs->mems_allowed));
2339 case FILE_EFFECTIVE_CPULIST:
2340 seq_printf(sf, "%*pbl\n", cpumask_pr_args(cs->effective_cpus));
2342 case FILE_EFFECTIVE_MEMLIST:
2343 seq_printf(sf, "%*pbl\n", nodemask_pr_args(&cs->effective_mems));
2345 case FILE_SUBPARTS_CPULIST:
2346 seq_printf(sf, "%*pbl\n", cpumask_pr_args(cs->subparts_cpus));
2352 spin_unlock_irq(&callback_lock);
2356 static u64 cpuset_read_u64(struct cgroup_subsys_state *css, struct cftype *cft)
2358 struct cpuset *cs = css_cs(css);
2359 cpuset_filetype_t type = cft->private;
2361 case FILE_CPU_EXCLUSIVE:
2362 return is_cpu_exclusive(cs);
2363 case FILE_MEM_EXCLUSIVE:
2364 return is_mem_exclusive(cs);
2365 case FILE_MEM_HARDWALL:
2366 return is_mem_hardwall(cs);
2367 case FILE_SCHED_LOAD_BALANCE:
2368 return is_sched_load_balance(cs);
2369 case FILE_MEMORY_MIGRATE:
2370 return is_memory_migrate(cs);
2371 case FILE_MEMORY_PRESSURE_ENABLED:
2372 return cpuset_memory_pressure_enabled;
2373 case FILE_MEMORY_PRESSURE:
2374 return fmeter_getrate(&cs->fmeter);
2375 case FILE_SPREAD_PAGE:
2376 return is_spread_page(cs);
2377 case FILE_SPREAD_SLAB:
2378 return is_spread_slab(cs);
2383 /* Unreachable but makes gcc happy */
2387 static s64 cpuset_read_s64(struct cgroup_subsys_state *css, struct cftype *cft)
2389 struct cpuset *cs = css_cs(css);
2390 cpuset_filetype_t type = cft->private;
2392 case FILE_SCHED_RELAX_DOMAIN_LEVEL:
2393 return cs->relax_domain_level;
2398 /* Unrechable but makes gcc happy */
2402 static int sched_partition_show(struct seq_file *seq, void *v)
2404 struct cpuset *cs = css_cs(seq_css(seq));
2406 switch (cs->partition_root_state) {
2408 seq_puts(seq, "root\n");
2411 seq_puts(seq, "member\n");
2414 seq_puts(seq, "root invalid\n");
2420 static ssize_t sched_partition_write(struct kernfs_open_file *of, char *buf,
2421 size_t nbytes, loff_t off)
2423 struct cpuset *cs = css_cs(of_css(of));
2425 int retval = -ENODEV;
2427 buf = strstrip(buf);
2430 * Convert "root" to ENABLED, and convert "member" to DISABLED.
2432 if (!strcmp(buf, "root"))
2434 else if (!strcmp(buf, "member"))
2440 mutex_lock(&cpuset_mutex);
2441 if (!is_cpuset_online(cs))
2444 retval = update_prstate(cs, val);
2446 mutex_unlock(&cpuset_mutex);
2448 return retval ?: nbytes;
2452 * for the common functions, 'private' gives the type of file
2455 static struct cftype legacy_files[] = {
2458 .seq_show = cpuset_common_seq_show,
2459 .write = cpuset_write_resmask,
2460 .max_write_len = (100U + 6 * NR_CPUS),
2461 .private = FILE_CPULIST,
2466 .seq_show = cpuset_common_seq_show,
2467 .write = cpuset_write_resmask,
2468 .max_write_len = (100U + 6 * MAX_NUMNODES),
2469 .private = FILE_MEMLIST,
2473 .name = "effective_cpus",
2474 .seq_show = cpuset_common_seq_show,
2475 .private = FILE_EFFECTIVE_CPULIST,
2479 .name = "effective_mems",
2480 .seq_show = cpuset_common_seq_show,
2481 .private = FILE_EFFECTIVE_MEMLIST,
2485 .name = "cpu_exclusive",
2486 .read_u64 = cpuset_read_u64,
2487 .write_u64 = cpuset_write_u64,
2488 .private = FILE_CPU_EXCLUSIVE,
2492 .name = "mem_exclusive",
2493 .read_u64 = cpuset_read_u64,
2494 .write_u64 = cpuset_write_u64,
2495 .private = FILE_MEM_EXCLUSIVE,
2499 .name = "mem_hardwall",
2500 .read_u64 = cpuset_read_u64,
2501 .write_u64 = cpuset_write_u64,
2502 .private = FILE_MEM_HARDWALL,
2506 .name = "sched_load_balance",
2507 .read_u64 = cpuset_read_u64,
2508 .write_u64 = cpuset_write_u64,
2509 .private = FILE_SCHED_LOAD_BALANCE,
2513 .name = "sched_relax_domain_level",
2514 .read_s64 = cpuset_read_s64,
2515 .write_s64 = cpuset_write_s64,
2516 .private = FILE_SCHED_RELAX_DOMAIN_LEVEL,
2520 .name = "memory_migrate",
2521 .read_u64 = cpuset_read_u64,
2522 .write_u64 = cpuset_write_u64,
2523 .private = FILE_MEMORY_MIGRATE,
2527 .name = "memory_pressure",
2528 .read_u64 = cpuset_read_u64,
2529 .private = FILE_MEMORY_PRESSURE,
2533 .name = "memory_spread_page",
2534 .read_u64 = cpuset_read_u64,
2535 .write_u64 = cpuset_write_u64,
2536 .private = FILE_SPREAD_PAGE,
2540 .name = "memory_spread_slab",
2541 .read_u64 = cpuset_read_u64,
2542 .write_u64 = cpuset_write_u64,
2543 .private = FILE_SPREAD_SLAB,
2547 .name = "memory_pressure_enabled",
2548 .flags = CFTYPE_ONLY_ON_ROOT,
2549 .read_u64 = cpuset_read_u64,
2550 .write_u64 = cpuset_write_u64,
2551 .private = FILE_MEMORY_PRESSURE_ENABLED,
2558 * This is currently a minimal set for the default hierarchy. It can be
2559 * expanded later on by migrating more features and control files from v1.
2561 static struct cftype dfl_files[] = {
2564 .seq_show = cpuset_common_seq_show,
2565 .write = cpuset_write_resmask,
2566 .max_write_len = (100U + 6 * NR_CPUS),
2567 .private = FILE_CPULIST,
2568 .flags = CFTYPE_NOT_ON_ROOT,
2573 .seq_show = cpuset_common_seq_show,
2574 .write = cpuset_write_resmask,
2575 .max_write_len = (100U + 6 * MAX_NUMNODES),
2576 .private = FILE_MEMLIST,
2577 .flags = CFTYPE_NOT_ON_ROOT,
2581 .name = "cpus.effective",
2582 .seq_show = cpuset_common_seq_show,
2583 .private = FILE_EFFECTIVE_CPULIST,
2587 .name = "mems.effective",
2588 .seq_show = cpuset_common_seq_show,
2589 .private = FILE_EFFECTIVE_MEMLIST,
2593 .name = "cpus.partition",
2594 .seq_show = sched_partition_show,
2595 .write = sched_partition_write,
2596 .private = FILE_PARTITION_ROOT,
2597 .flags = CFTYPE_NOT_ON_ROOT,
2601 .name = "cpus.subpartitions",
2602 .seq_show = cpuset_common_seq_show,
2603 .private = FILE_SUBPARTS_CPULIST,
2604 .flags = CFTYPE_DEBUG,
2612 * cpuset_css_alloc - allocate a cpuset css
2613 * cgrp: control group that the new cpuset will be part of
2616 static struct cgroup_subsys_state *
2617 cpuset_css_alloc(struct cgroup_subsys_state *parent_css)
2622 return &top_cpuset.css;
2624 cs = kzalloc(sizeof(*cs), GFP_KERNEL);
2626 return ERR_PTR(-ENOMEM);
2628 if (alloc_cpumasks(cs, NULL)) {
2630 return ERR_PTR(-ENOMEM);
2633 set_bit(CS_SCHED_LOAD_BALANCE, &cs->flags);
2634 nodes_clear(cs->mems_allowed);
2635 nodes_clear(cs->effective_mems);
2636 fmeter_init(&cs->fmeter);
2637 cs->relax_domain_level = -1;
2642 static int cpuset_css_online(struct cgroup_subsys_state *css)
2644 struct cpuset *cs = css_cs(css);
2645 struct cpuset *parent = parent_cs(cs);
2646 struct cpuset *tmp_cs;
2647 struct cgroup_subsys_state *pos_css;
2652 mutex_lock(&cpuset_mutex);
2654 set_bit(CS_ONLINE, &cs->flags);
2655 if (is_spread_page(parent))
2656 set_bit(CS_SPREAD_PAGE, &cs->flags);
2657 if (is_spread_slab(parent))
2658 set_bit(CS_SPREAD_SLAB, &cs->flags);
2662 spin_lock_irq(&callback_lock);
2663 if (is_in_v2_mode()) {
2664 cpumask_copy(cs->effective_cpus, parent->effective_cpus);
2665 cs->effective_mems = parent->effective_mems;
2666 cs->use_parent_ecpus = true;
2667 parent->child_ecpus_count++;
2669 spin_unlock_irq(&callback_lock);
2671 if (!test_bit(CGRP_CPUSET_CLONE_CHILDREN, &css->cgroup->flags))
2675 * Clone @parent's configuration if CGRP_CPUSET_CLONE_CHILDREN is
2676 * set. This flag handling is implemented in cgroup core for
2677 * histrical reasons - the flag may be specified during mount.
2679 * Currently, if any sibling cpusets have exclusive cpus or mem, we
2680 * refuse to clone the configuration - thereby refusing the task to
2681 * be entered, and as a result refusing the sys_unshare() or
2682 * clone() which initiated it. If this becomes a problem for some
2683 * users who wish to allow that scenario, then this could be
2684 * changed to grant parent->cpus_allowed-sibling_cpus_exclusive
2685 * (and likewise for mems) to the new cgroup.
2688 cpuset_for_each_child(tmp_cs, pos_css, parent) {
2689 if (is_mem_exclusive(tmp_cs) || is_cpu_exclusive(tmp_cs)) {
2696 spin_lock_irq(&callback_lock);
2697 cs->mems_allowed = parent->mems_allowed;
2698 cs->effective_mems = parent->mems_allowed;
2699 cpumask_copy(cs->cpus_allowed, parent->cpus_allowed);
2700 cpumask_copy(cs->effective_cpus, parent->cpus_allowed);
2701 spin_unlock_irq(&callback_lock);
2703 mutex_unlock(&cpuset_mutex);
2708 * If the cpuset being removed has its flag 'sched_load_balance'
2709 * enabled, then simulate turning sched_load_balance off, which
2710 * will call rebuild_sched_domains_locked(). That is not needed
2711 * in the default hierarchy where only changes in partition
2712 * will cause repartitioning.
2714 * If the cpuset has the 'sched.partition' flag enabled, simulate
2715 * turning 'sched.partition" off.
2718 static void cpuset_css_offline(struct cgroup_subsys_state *css)
2720 struct cpuset *cs = css_cs(css);
2722 mutex_lock(&cpuset_mutex);
2724 if (is_partition_root(cs))
2725 update_prstate(cs, 0);
2727 if (!cgroup_subsys_on_dfl(cpuset_cgrp_subsys) &&
2728 is_sched_load_balance(cs))
2729 update_flag(CS_SCHED_LOAD_BALANCE, cs, 0);
2731 if (cs->use_parent_ecpus) {
2732 struct cpuset *parent = parent_cs(cs);
2734 cs->use_parent_ecpus = false;
2735 parent->child_ecpus_count--;
2739 clear_bit(CS_ONLINE, &cs->flags);
2741 mutex_unlock(&cpuset_mutex);
2744 static void cpuset_css_free(struct cgroup_subsys_state *css)
2746 struct cpuset *cs = css_cs(css);
2751 static void cpuset_bind(struct cgroup_subsys_state *root_css)
2753 mutex_lock(&cpuset_mutex);
2754 spin_lock_irq(&callback_lock);
2756 if (is_in_v2_mode()) {
2757 cpumask_copy(top_cpuset.cpus_allowed, cpu_possible_mask);
2758 top_cpuset.mems_allowed = node_possible_map;
2760 cpumask_copy(top_cpuset.cpus_allowed,
2761 top_cpuset.effective_cpus);
2762 top_cpuset.mems_allowed = top_cpuset.effective_mems;
2765 spin_unlock_irq(&callback_lock);
2766 mutex_unlock(&cpuset_mutex);
2770 * Make sure the new task conform to the current state of its parent,
2771 * which could have been changed by cpuset just after it inherits the
2772 * state from the parent and before it sits on the cgroup's task list.
2774 static void cpuset_fork(struct task_struct *task)
2776 if (task_css_is_root(task, cpuset_cgrp_id))
2779 set_cpus_allowed_ptr(task, current->cpus_ptr);
2780 task->mems_allowed = current->mems_allowed;
2783 struct cgroup_subsys cpuset_cgrp_subsys = {
2784 .css_alloc = cpuset_css_alloc,
2785 .css_online = cpuset_css_online,
2786 .css_offline = cpuset_css_offline,
2787 .css_free = cpuset_css_free,
2788 .can_attach = cpuset_can_attach,
2789 .cancel_attach = cpuset_cancel_attach,
2790 .attach = cpuset_attach,
2791 .post_attach = cpuset_post_attach,
2792 .bind = cpuset_bind,
2793 .fork = cpuset_fork,
2794 .legacy_cftypes = legacy_files,
2795 .dfl_cftypes = dfl_files,
2801 * cpuset_init - initialize cpusets at system boot
2803 * Description: Initialize top_cpuset
2806 int __init cpuset_init(void)
2808 BUG_ON(!alloc_cpumask_var(&top_cpuset.cpus_allowed, GFP_KERNEL));
2809 BUG_ON(!alloc_cpumask_var(&top_cpuset.effective_cpus, GFP_KERNEL));
2810 BUG_ON(!zalloc_cpumask_var(&top_cpuset.subparts_cpus, GFP_KERNEL));
2812 cpumask_setall(top_cpuset.cpus_allowed);
2813 nodes_setall(top_cpuset.mems_allowed);
2814 cpumask_setall(top_cpuset.effective_cpus);
2815 nodes_setall(top_cpuset.effective_mems);
2817 fmeter_init(&top_cpuset.fmeter);
2818 set_bit(CS_SCHED_LOAD_BALANCE, &top_cpuset.flags);
2819 top_cpuset.relax_domain_level = -1;
2821 BUG_ON(!alloc_cpumask_var(&cpus_attach, GFP_KERNEL));
2827 * If CPU and/or memory hotplug handlers, below, unplug any CPUs
2828 * or memory nodes, we need to walk over the cpuset hierarchy,
2829 * removing that CPU or node from all cpusets. If this removes the
2830 * last CPU or node from a cpuset, then move the tasks in the empty
2831 * cpuset to its next-highest non-empty parent.
2833 static void remove_tasks_in_empty_cpuset(struct cpuset *cs)
2835 struct cpuset *parent;
2838 * Find its next-highest non-empty parent, (top cpuset
2839 * has online cpus, so can't be empty).
2841 parent = parent_cs(cs);
2842 while (cpumask_empty(parent->cpus_allowed) ||
2843 nodes_empty(parent->mems_allowed))
2844 parent = parent_cs(parent);
2846 if (cgroup_transfer_tasks(parent->css.cgroup, cs->css.cgroup)) {
2847 pr_err("cpuset: failed to transfer tasks out of empty cpuset ");
2848 pr_cont_cgroup_name(cs->css.cgroup);
2854 hotplug_update_tasks_legacy(struct cpuset *cs,
2855 struct cpumask *new_cpus, nodemask_t *new_mems,
2856 bool cpus_updated, bool mems_updated)
2860 spin_lock_irq(&callback_lock);
2861 cpumask_copy(cs->cpus_allowed, new_cpus);
2862 cpumask_copy(cs->effective_cpus, new_cpus);
2863 cs->mems_allowed = *new_mems;
2864 cs->effective_mems = *new_mems;
2865 spin_unlock_irq(&callback_lock);
2868 * Don't call update_tasks_cpumask() if the cpuset becomes empty,
2869 * as the tasks will be migratecd to an ancestor.
2871 if (cpus_updated && !cpumask_empty(cs->cpus_allowed))
2872 update_tasks_cpumask(cs);
2873 if (mems_updated && !nodes_empty(cs->mems_allowed))
2874 update_tasks_nodemask(cs);
2876 is_empty = cpumask_empty(cs->cpus_allowed) ||
2877 nodes_empty(cs->mems_allowed);
2879 mutex_unlock(&cpuset_mutex);
2882 * Move tasks to the nearest ancestor with execution resources,
2883 * This is full cgroup operation which will also call back into
2884 * cpuset. Should be done outside any lock.
2887 remove_tasks_in_empty_cpuset(cs);
2889 mutex_lock(&cpuset_mutex);
2893 hotplug_update_tasks(struct cpuset *cs,
2894 struct cpumask *new_cpus, nodemask_t *new_mems,
2895 bool cpus_updated, bool mems_updated)
2897 if (cpumask_empty(new_cpus))
2898 cpumask_copy(new_cpus, parent_cs(cs)->effective_cpus);
2899 if (nodes_empty(*new_mems))
2900 *new_mems = parent_cs(cs)->effective_mems;
2902 spin_lock_irq(&callback_lock);
2903 cpumask_copy(cs->effective_cpus, new_cpus);
2904 cs->effective_mems = *new_mems;
2905 spin_unlock_irq(&callback_lock);
2908 update_tasks_cpumask(cs);
2910 update_tasks_nodemask(cs);
2913 static bool force_rebuild;
2915 void cpuset_force_rebuild(void)
2917 force_rebuild = true;
2921 * cpuset_hotplug_update_tasks - update tasks in a cpuset for hotunplug
2922 * @cs: cpuset in interest
2923 * @tmp: the tmpmasks structure pointer
2925 * Compare @cs's cpu and mem masks against top_cpuset and if some have gone
2926 * offline, update @cs accordingly. If @cs ends up with no CPU or memory,
2927 * all its tasks are moved to the nearest ancestor with both resources.
2929 static void cpuset_hotplug_update_tasks(struct cpuset *cs, struct tmpmasks *tmp)
2931 static cpumask_t new_cpus;
2932 static nodemask_t new_mems;
2935 struct cpuset *parent;
2937 wait_event(cpuset_attach_wq, cs->attach_in_progress == 0);
2939 mutex_lock(&cpuset_mutex);
2942 * We have raced with task attaching. We wait until attaching
2943 * is finished, so we won't attach a task to an empty cpuset.
2945 if (cs->attach_in_progress) {
2946 mutex_unlock(&cpuset_mutex);
2950 parent = parent_cs(cs);
2951 compute_effective_cpumask(&new_cpus, cs, parent);
2952 nodes_and(new_mems, cs->mems_allowed, parent->effective_mems);
2954 if (cs->nr_subparts_cpus)
2956 * Make sure that CPUs allocated to child partitions
2957 * do not show up in effective_cpus.
2959 cpumask_andnot(&new_cpus, &new_cpus, cs->subparts_cpus);
2961 if (!tmp || !cs->partition_root_state)
2965 * In the unlikely event that a partition root has empty
2966 * effective_cpus or its parent becomes erroneous, we have to
2967 * transition it to the erroneous state.
2969 if (is_partition_root(cs) && (cpumask_empty(&new_cpus) ||
2970 (parent->partition_root_state == PRS_ERROR))) {
2971 if (cs->nr_subparts_cpus) {
2972 cs->nr_subparts_cpus = 0;
2973 cpumask_clear(cs->subparts_cpus);
2974 compute_effective_cpumask(&new_cpus, cs, parent);
2978 * If the effective_cpus is empty because the child
2979 * partitions take away all the CPUs, we can keep
2980 * the current partition and let the child partitions
2981 * fight for available CPUs.
2983 if ((parent->partition_root_state == PRS_ERROR) ||
2984 cpumask_empty(&new_cpus)) {
2985 update_parent_subparts_cpumask(cs, partcmd_disable,
2987 cs->partition_root_state = PRS_ERROR;
2989 cpuset_force_rebuild();
2993 * On the other hand, an erroneous partition root may be transitioned
2994 * back to a regular one or a partition root with no CPU allocated
2995 * from the parent may change to erroneous.
2997 if (is_partition_root(parent) &&
2998 ((cs->partition_root_state == PRS_ERROR) ||
2999 !cpumask_intersects(&new_cpus, parent->subparts_cpus)) &&
3000 update_parent_subparts_cpumask(cs, partcmd_update, NULL, tmp))
3001 cpuset_force_rebuild();
3004 cpus_updated = !cpumask_equal(&new_cpus, cs->effective_cpus);
3005 mems_updated = !nodes_equal(new_mems, cs->effective_mems);
3007 if (is_in_v2_mode())
3008 hotplug_update_tasks(cs, &new_cpus, &new_mems,
3009 cpus_updated, mems_updated);
3011 hotplug_update_tasks_legacy(cs, &new_cpus, &new_mems,
3012 cpus_updated, mems_updated);
3014 mutex_unlock(&cpuset_mutex);
3018 * cpuset_hotplug_workfn - handle CPU/memory hotunplug for a cpuset
3020 * This function is called after either CPU or memory configuration has
3021 * changed and updates cpuset accordingly. The top_cpuset is always
3022 * synchronized to cpu_active_mask and N_MEMORY, which is necessary in
3023 * order to make cpusets transparent (of no affect) on systems that are
3024 * actively using CPU hotplug but making no active use of cpusets.
3026 * Non-root cpusets are only affected by offlining. If any CPUs or memory
3027 * nodes have been taken down, cpuset_hotplug_update_tasks() is invoked on
3030 * Note that CPU offlining during suspend is ignored. We don't modify
3031 * cpusets across suspend/resume cycles at all.
3033 static void cpuset_hotplug_workfn(struct work_struct *work)
3035 static cpumask_t new_cpus;
3036 static nodemask_t new_mems;
3037 bool cpus_updated, mems_updated;
3038 bool on_dfl = is_in_v2_mode();
3039 struct tmpmasks tmp, *ptmp = NULL;
3041 if (on_dfl && !alloc_cpumasks(NULL, &tmp))
3044 mutex_lock(&cpuset_mutex);
3046 /* fetch the available cpus/mems and find out which changed how */
3047 cpumask_copy(&new_cpus, cpu_active_mask);
3048 new_mems = node_states[N_MEMORY];
3051 * If subparts_cpus is populated, it is likely that the check below
3052 * will produce a false positive on cpus_updated when the cpu list
3053 * isn't changed. It is extra work, but it is better to be safe.
3055 cpus_updated = !cpumask_equal(top_cpuset.effective_cpus, &new_cpus);
3056 mems_updated = !nodes_equal(top_cpuset.effective_mems, new_mems);
3058 /* synchronize cpus_allowed to cpu_active_mask */
3060 spin_lock_irq(&callback_lock);
3062 cpumask_copy(top_cpuset.cpus_allowed, &new_cpus);
3064 * Make sure that CPUs allocated to child partitions
3065 * do not show up in effective_cpus. If no CPU is left,
3066 * we clear the subparts_cpus & let the child partitions
3067 * fight for the CPUs again.
3069 if (top_cpuset.nr_subparts_cpus) {
3070 if (cpumask_subset(&new_cpus,
3071 top_cpuset.subparts_cpus)) {
3072 top_cpuset.nr_subparts_cpus = 0;
3073 cpumask_clear(top_cpuset.subparts_cpus);
3075 cpumask_andnot(&new_cpus, &new_cpus,
3076 top_cpuset.subparts_cpus);
3079 cpumask_copy(top_cpuset.effective_cpus, &new_cpus);
3080 spin_unlock_irq(&callback_lock);
3081 /* we don't mess with cpumasks of tasks in top_cpuset */
3084 /* synchronize mems_allowed to N_MEMORY */
3086 spin_lock_irq(&callback_lock);
3088 top_cpuset.mems_allowed = new_mems;
3089 top_cpuset.effective_mems = new_mems;
3090 spin_unlock_irq(&callback_lock);
3091 update_tasks_nodemask(&top_cpuset);
3094 mutex_unlock(&cpuset_mutex);
3096 /* if cpus or mems changed, we need to propagate to descendants */
3097 if (cpus_updated || mems_updated) {
3099 struct cgroup_subsys_state *pos_css;
3102 cpuset_for_each_descendant_pre(cs, pos_css, &top_cpuset) {
3103 if (cs == &top_cpuset || !css_tryget_online(&cs->css))
3107 cpuset_hotplug_update_tasks(cs, ptmp);
3115 /* rebuild sched domains if cpus_allowed has changed */
3116 if (cpus_updated || force_rebuild) {
3117 force_rebuild = false;
3118 rebuild_sched_domains();
3121 free_cpumasks(NULL, ptmp);
3124 void cpuset_update_active_cpus(void)
3127 * We're inside cpu hotplug critical region which usually nests
3128 * inside cgroup synchronization. Bounce actual hotplug processing
3129 * to a work item to avoid reverse locking order.
3131 schedule_work(&cpuset_hotplug_work);
3134 void cpuset_wait_for_hotplug(void)
3136 flush_work(&cpuset_hotplug_work);
3140 * Keep top_cpuset.mems_allowed tracking node_states[N_MEMORY].
3141 * Call this routine anytime after node_states[N_MEMORY] changes.
3142 * See cpuset_update_active_cpus() for CPU hotplug handling.
3144 static int cpuset_track_online_nodes(struct notifier_block *self,
3145 unsigned long action, void *arg)
3147 schedule_work(&cpuset_hotplug_work);
3151 static struct notifier_block cpuset_track_online_nodes_nb = {
3152 .notifier_call = cpuset_track_online_nodes,
3153 .priority = 10, /* ??! */
3157 * cpuset_init_smp - initialize cpus_allowed
3159 * Description: Finish top cpuset after cpu, node maps are initialized
3161 void __init cpuset_init_smp(void)
3163 cpumask_copy(top_cpuset.cpus_allowed, cpu_active_mask);
3164 top_cpuset.mems_allowed = node_states[N_MEMORY];
3165 top_cpuset.old_mems_allowed = top_cpuset.mems_allowed;
3167 cpumask_copy(top_cpuset.effective_cpus, cpu_active_mask);
3168 top_cpuset.effective_mems = node_states[N_MEMORY];
3170 register_hotmemory_notifier(&cpuset_track_online_nodes_nb);
3172 cpuset_migrate_mm_wq = alloc_ordered_workqueue("cpuset_migrate_mm", 0);
3173 BUG_ON(!cpuset_migrate_mm_wq);
3177 * cpuset_cpus_allowed - return cpus_allowed mask from a tasks cpuset.
3178 * @tsk: pointer to task_struct from which to obtain cpuset->cpus_allowed.
3179 * @pmask: pointer to struct cpumask variable to receive cpus_allowed set.
3181 * Description: Returns the cpumask_var_t cpus_allowed of the cpuset
3182 * attached to the specified @tsk. Guaranteed to return some non-empty
3183 * subset of cpu_online_mask, even if this means going outside the
3187 void cpuset_cpus_allowed(struct task_struct *tsk, struct cpumask *pmask)
3189 unsigned long flags;
3191 spin_lock_irqsave(&callback_lock, flags);
3193 guarantee_online_cpus(task_cs(tsk), pmask);
3195 spin_unlock_irqrestore(&callback_lock, flags);
3199 * cpuset_cpus_allowed_fallback - final fallback before complete catastrophe.
3200 * @tsk: pointer to task_struct with which the scheduler is struggling
3202 * Description: In the case that the scheduler cannot find an allowed cpu in
3203 * tsk->cpus_allowed, we fall back to task_cs(tsk)->cpus_allowed. In legacy
3204 * mode however, this value is the same as task_cs(tsk)->effective_cpus,
3205 * which will not contain a sane cpumask during cases such as cpu hotplugging.
3206 * This is the absolute last resort for the scheduler and it is only used if
3207 * _every_ other avenue has been traveled.
3210 void cpuset_cpus_allowed_fallback(struct task_struct *tsk)
3213 do_set_cpus_allowed(tsk, is_in_v2_mode() ?
3214 task_cs(tsk)->cpus_allowed : cpu_possible_mask);
3218 * We own tsk->cpus_allowed, nobody can change it under us.
3220 * But we used cs && cs->cpus_allowed lockless and thus can
3221 * race with cgroup_attach_task() or update_cpumask() and get
3222 * the wrong tsk->cpus_allowed. However, both cases imply the
3223 * subsequent cpuset_change_cpumask()->set_cpus_allowed_ptr()
3224 * which takes task_rq_lock().
3226 * If we are called after it dropped the lock we must see all
3227 * changes in tsk_cs()->cpus_allowed. Otherwise we can temporary
3228 * set any mask even if it is not right from task_cs() pov,
3229 * the pending set_cpus_allowed_ptr() will fix things.
3231 * select_fallback_rq() will fix things ups and set cpu_possible_mask
3236 void __init cpuset_init_current_mems_allowed(void)
3238 nodes_setall(current->mems_allowed);
3242 * cpuset_mems_allowed - return mems_allowed mask from a tasks cpuset.
3243 * @tsk: pointer to task_struct from which to obtain cpuset->mems_allowed.
3245 * Description: Returns the nodemask_t mems_allowed of the cpuset
3246 * attached to the specified @tsk. Guaranteed to return some non-empty
3247 * subset of node_states[N_MEMORY], even if this means going outside the
3251 nodemask_t cpuset_mems_allowed(struct task_struct *tsk)
3254 unsigned long flags;
3256 spin_lock_irqsave(&callback_lock, flags);
3258 guarantee_online_mems(task_cs(tsk), &mask);
3260 spin_unlock_irqrestore(&callback_lock, flags);
3266 * cpuset_nodemask_valid_mems_allowed - check nodemask vs. curremt mems_allowed
3267 * @nodemask: the nodemask to be checked
3269 * Are any of the nodes in the nodemask allowed in current->mems_allowed?
3271 int cpuset_nodemask_valid_mems_allowed(nodemask_t *nodemask)
3273 return nodes_intersects(*nodemask, current->mems_allowed);
3277 * nearest_hardwall_ancestor() - Returns the nearest mem_exclusive or
3278 * mem_hardwall ancestor to the specified cpuset. Call holding
3279 * callback_lock. If no ancestor is mem_exclusive or mem_hardwall
3280 * (an unusual configuration), then returns the root cpuset.
3282 static struct cpuset *nearest_hardwall_ancestor(struct cpuset *cs)
3284 while (!(is_mem_exclusive(cs) || is_mem_hardwall(cs)) && parent_cs(cs))
3290 * cpuset_node_allowed - Can we allocate on a memory node?
3291 * @node: is this an allowed node?
3292 * @gfp_mask: memory allocation flags
3294 * If we're in interrupt, yes, we can always allocate. If @node is set in
3295 * current's mems_allowed, yes. If it's not a __GFP_HARDWALL request and this
3296 * node is set in the nearest hardwalled cpuset ancestor to current's cpuset,
3297 * yes. If current has access to memory reserves as an oom victim, yes.
3300 * GFP_USER allocations are marked with the __GFP_HARDWALL bit,
3301 * and do not allow allocations outside the current tasks cpuset
3302 * unless the task has been OOM killed.
3303 * GFP_KERNEL allocations are not so marked, so can escape to the
3304 * nearest enclosing hardwalled ancestor cpuset.
3306 * Scanning up parent cpusets requires callback_lock. The
3307 * __alloc_pages() routine only calls here with __GFP_HARDWALL bit
3308 * _not_ set if it's a GFP_KERNEL allocation, and all nodes in the
3309 * current tasks mems_allowed came up empty on the first pass over
3310 * the zonelist. So only GFP_KERNEL allocations, if all nodes in the
3311 * cpuset are short of memory, might require taking the callback_lock.
3313 * The first call here from mm/page_alloc:get_page_from_freelist()
3314 * has __GFP_HARDWALL set in gfp_mask, enforcing hardwall cpusets,
3315 * so no allocation on a node outside the cpuset is allowed (unless
3316 * in interrupt, of course).
3318 * The second pass through get_page_from_freelist() doesn't even call
3319 * here for GFP_ATOMIC calls. For those calls, the __alloc_pages()
3320 * variable 'wait' is not set, and the bit ALLOC_CPUSET is not set
3321 * in alloc_flags. That logic and the checks below have the combined
3323 * in_interrupt - any node ok (current task context irrelevant)
3324 * GFP_ATOMIC - any node ok
3325 * tsk_is_oom_victim - any node ok
3326 * GFP_KERNEL - any node in enclosing hardwalled cpuset ok
3327 * GFP_USER - only nodes in current tasks mems allowed ok.
3329 bool __cpuset_node_allowed(int node, gfp_t gfp_mask)
3331 struct cpuset *cs; /* current cpuset ancestors */
3332 int allowed; /* is allocation in zone z allowed? */
3333 unsigned long flags;
3337 if (node_isset(node, current->mems_allowed))
3340 * Allow tasks that have access to memory reserves because they have
3341 * been OOM killed to get memory anywhere.
3343 if (unlikely(tsk_is_oom_victim(current)))
3345 if (gfp_mask & __GFP_HARDWALL) /* If hardwall request, stop here */
3348 if (current->flags & PF_EXITING) /* Let dying task have memory */
3351 /* Not hardwall and node outside mems_allowed: scan up cpusets */
3352 spin_lock_irqsave(&callback_lock, flags);
3355 cs = nearest_hardwall_ancestor(task_cs(current));
3356 allowed = node_isset(node, cs->mems_allowed);
3359 spin_unlock_irqrestore(&callback_lock, flags);
3364 * cpuset_mem_spread_node() - On which node to begin search for a file page
3365 * cpuset_slab_spread_node() - On which node to begin search for a slab page
3367 * If a task is marked PF_SPREAD_PAGE or PF_SPREAD_SLAB (as for
3368 * tasks in a cpuset with is_spread_page or is_spread_slab set),
3369 * and if the memory allocation used cpuset_mem_spread_node()
3370 * to determine on which node to start looking, as it will for
3371 * certain page cache or slab cache pages such as used for file
3372 * system buffers and inode caches, then instead of starting on the
3373 * local node to look for a free page, rather spread the starting
3374 * node around the tasks mems_allowed nodes.
3376 * We don't have to worry about the returned node being offline
3377 * because "it can't happen", and even if it did, it would be ok.
3379 * The routines calling guarantee_online_mems() are careful to
3380 * only set nodes in task->mems_allowed that are online. So it
3381 * should not be possible for the following code to return an
3382 * offline node. But if it did, that would be ok, as this routine
3383 * is not returning the node where the allocation must be, only
3384 * the node where the search should start. The zonelist passed to
3385 * __alloc_pages() will include all nodes. If the slab allocator
3386 * is passed an offline node, it will fall back to the local node.
3387 * See kmem_cache_alloc_node().
3390 static int cpuset_spread_node(int *rotor)
3392 return *rotor = next_node_in(*rotor, current->mems_allowed);
3395 int cpuset_mem_spread_node(void)
3397 if (current->cpuset_mem_spread_rotor == NUMA_NO_NODE)
3398 current->cpuset_mem_spread_rotor =
3399 node_random(¤t->mems_allowed);
3401 return cpuset_spread_node(¤t->cpuset_mem_spread_rotor);
3404 int cpuset_slab_spread_node(void)
3406 if (current->cpuset_slab_spread_rotor == NUMA_NO_NODE)
3407 current->cpuset_slab_spread_rotor =
3408 node_random(¤t->mems_allowed);
3410 return cpuset_spread_node(¤t->cpuset_slab_spread_rotor);
3413 EXPORT_SYMBOL_GPL(cpuset_mem_spread_node);
3416 * cpuset_mems_allowed_intersects - Does @tsk1's mems_allowed intersect @tsk2's?
3417 * @tsk1: pointer to task_struct of some task.
3418 * @tsk2: pointer to task_struct of some other task.
3420 * Description: Return true if @tsk1's mems_allowed intersects the
3421 * mems_allowed of @tsk2. Used by the OOM killer to determine if
3422 * one of the task's memory usage might impact the memory available
3426 int cpuset_mems_allowed_intersects(const struct task_struct *tsk1,
3427 const struct task_struct *tsk2)
3429 return nodes_intersects(tsk1->mems_allowed, tsk2->mems_allowed);
3433 * cpuset_print_current_mems_allowed - prints current's cpuset and mems_allowed
3435 * Description: Prints current's name, cpuset name, and cached copy of its
3436 * mems_allowed to the kernel log.
3438 void cpuset_print_current_mems_allowed(void)
3440 struct cgroup *cgrp;
3444 cgrp = task_cs(current)->css.cgroup;
3445 pr_cont(",cpuset=");
3446 pr_cont_cgroup_name(cgrp);
3447 pr_cont(",mems_allowed=%*pbl",
3448 nodemask_pr_args(¤t->mems_allowed));
3454 * Collection of memory_pressure is suppressed unless
3455 * this flag is enabled by writing "1" to the special
3456 * cpuset file 'memory_pressure_enabled' in the root cpuset.
3459 int cpuset_memory_pressure_enabled __read_mostly;
3462 * cpuset_memory_pressure_bump - keep stats of per-cpuset reclaims.
3464 * Keep a running average of the rate of synchronous (direct)
3465 * page reclaim efforts initiated by tasks in each cpuset.
3467 * This represents the rate at which some task in the cpuset
3468 * ran low on memory on all nodes it was allowed to use, and
3469 * had to enter the kernels page reclaim code in an effort to
3470 * create more free memory by tossing clean pages or swapping
3471 * or writing dirty pages.
3473 * Display to user space in the per-cpuset read-only file
3474 * "memory_pressure". Value displayed is an integer
3475 * representing the recent rate of entry into the synchronous
3476 * (direct) page reclaim by any task attached to the cpuset.
3479 void __cpuset_memory_pressure_bump(void)
3482 fmeter_markevent(&task_cs(current)->fmeter);
3486 #ifdef CONFIG_PROC_PID_CPUSET
3488 * proc_cpuset_show()
3489 * - Print tasks cpuset path into seq_file.
3490 * - Used for /proc/<pid>/cpuset.
3491 * - No need to task_lock(tsk) on this tsk->cpuset reference, as it
3492 * doesn't really matter if tsk->cpuset changes after we read it,
3493 * and we take cpuset_mutex, keeping cpuset_attach() from changing it
3496 int proc_cpuset_show(struct seq_file *m, struct pid_namespace *ns,
3497 struct pid *pid, struct task_struct *tsk)
3500 struct cgroup_subsys_state *css;
3504 buf = kmalloc(PATH_MAX, GFP_KERNEL);
3508 css = task_get_css(tsk, cpuset_cgrp_id);
3509 retval = cgroup_path_ns(css->cgroup, buf, PATH_MAX,
3510 current->nsproxy->cgroup_ns);
3512 if (retval >= PATH_MAX)
3513 retval = -ENAMETOOLONG;
3524 #endif /* CONFIG_PROC_PID_CPUSET */
3526 /* Display task mems_allowed in /proc/<pid>/status file. */
3527 void cpuset_task_status_allowed(struct seq_file *m, struct task_struct *task)
3529 seq_printf(m, "Mems_allowed:\t%*pb\n",
3530 nodemask_pr_args(&task->mems_allowed));
3531 seq_printf(m, "Mems_allowed_list:\t%*pbl\n",
3532 nodemask_pr_args(&task->mems_allowed));