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 * This is ugly, but preserves the userspace API for existing cpuset
360 * users. If someone tries to mount the "cpuset" filesystem, we
361 * silently switch it to mount "cgroup" instead
363 static int cpuset_get_tree(struct fs_context *fc)
365 struct file_system_type *cgroup_fs;
366 struct fs_context *new_fc;
369 cgroup_fs = get_fs_type("cgroup");
373 new_fc = fs_context_for_mount(cgroup_fs, fc->sb_flags);
374 if (IS_ERR(new_fc)) {
375 ret = PTR_ERR(new_fc);
377 static const char agent_path[] = "/sbin/cpuset_release_agent";
378 ret = vfs_parse_fs_string(new_fc, "cpuset", NULL, 0);
380 ret = vfs_parse_fs_string(new_fc, "noprefix", NULL, 0);
382 ret = vfs_parse_fs_string(new_fc, "release_agent",
383 agent_path, sizeof(agent_path) - 1);
385 ret = vfs_get_tree(new_fc);
386 if (!ret) { /* steal the result */
387 fc->root = new_fc->root;
390 put_fs_context(new_fc);
392 put_filesystem(cgroup_fs);
396 static const struct fs_context_operations cpuset_fs_context_ops = {
397 .get_tree = cpuset_get_tree,
400 static int cpuset_init_fs_context(struct fs_context *fc)
402 fc->ops = &cpuset_fs_context_ops;
406 static struct file_system_type cpuset_fs_type = {
408 .init_fs_context = cpuset_init_fs_context,
412 * Return in pmask the portion of a cpusets's cpus_allowed that
413 * are online. If none are online, walk up the cpuset hierarchy
414 * until we find one that does have some online cpus.
416 * One way or another, we guarantee to return some non-empty subset
417 * of cpu_online_mask.
419 * Call with callback_lock or cpuset_mutex held.
421 static void guarantee_online_cpus(struct cpuset *cs, struct cpumask *pmask)
423 while (!cpumask_intersects(cs->effective_cpus, cpu_online_mask)) {
427 * The top cpuset doesn't have any online cpu as a
428 * consequence of a race between cpuset_hotplug_work
429 * and cpu hotplug notifier. But we know the top
430 * cpuset's effective_cpus is on its way to to be
431 * identical to cpu_online_mask.
433 cpumask_copy(pmask, cpu_online_mask);
437 cpumask_and(pmask, cs->effective_cpus, cpu_online_mask);
441 * Return in *pmask the portion of a cpusets's mems_allowed that
442 * are online, with memory. If none are online with memory, walk
443 * up the cpuset hierarchy until we find one that does have some
444 * online mems. The top cpuset always has some mems online.
446 * One way or another, we guarantee to return some non-empty subset
447 * of node_states[N_MEMORY].
449 * Call with callback_lock or cpuset_mutex held.
451 static void guarantee_online_mems(struct cpuset *cs, nodemask_t *pmask)
453 while (!nodes_intersects(cs->effective_mems, node_states[N_MEMORY]))
455 nodes_and(*pmask, cs->effective_mems, node_states[N_MEMORY]);
459 * update task's spread flag if cpuset's page/slab spread flag is set
461 * Call with callback_lock or cpuset_mutex held.
463 static void cpuset_update_task_spread_flag(struct cpuset *cs,
464 struct task_struct *tsk)
466 if (is_spread_page(cs))
467 task_set_spread_page(tsk);
469 task_clear_spread_page(tsk);
471 if (is_spread_slab(cs))
472 task_set_spread_slab(tsk);
474 task_clear_spread_slab(tsk);
478 * is_cpuset_subset(p, q) - Is cpuset p a subset of cpuset q?
480 * One cpuset is a subset of another if all its allowed CPUs and
481 * Memory Nodes are a subset of the other, and its exclusive flags
482 * are only set if the other's are set. Call holding cpuset_mutex.
485 static int is_cpuset_subset(const struct cpuset *p, const struct cpuset *q)
487 return cpumask_subset(p->cpus_allowed, q->cpus_allowed) &&
488 nodes_subset(p->mems_allowed, q->mems_allowed) &&
489 is_cpu_exclusive(p) <= is_cpu_exclusive(q) &&
490 is_mem_exclusive(p) <= is_mem_exclusive(q);
494 * alloc_cpumasks - allocate three cpumasks for cpuset
495 * @cs: the cpuset that have cpumasks to be allocated.
496 * @tmp: the tmpmasks structure pointer
497 * Return: 0 if successful, -ENOMEM otherwise.
499 * Only one of the two input arguments should be non-NULL.
501 static inline int alloc_cpumasks(struct cpuset *cs, struct tmpmasks *tmp)
503 cpumask_var_t *pmask1, *pmask2, *pmask3;
506 pmask1 = &cs->cpus_allowed;
507 pmask2 = &cs->effective_cpus;
508 pmask3 = &cs->subparts_cpus;
510 pmask1 = &tmp->new_cpus;
511 pmask2 = &tmp->addmask;
512 pmask3 = &tmp->delmask;
515 if (!zalloc_cpumask_var(pmask1, GFP_KERNEL))
518 if (!zalloc_cpumask_var(pmask2, GFP_KERNEL))
521 if (!zalloc_cpumask_var(pmask3, GFP_KERNEL))
527 free_cpumask_var(*pmask2);
529 free_cpumask_var(*pmask1);
534 * free_cpumasks - free cpumasks in a tmpmasks structure
535 * @cs: the cpuset that have cpumasks to be free.
536 * @tmp: the tmpmasks structure pointer
538 static inline void free_cpumasks(struct cpuset *cs, struct tmpmasks *tmp)
541 free_cpumask_var(cs->cpus_allowed);
542 free_cpumask_var(cs->effective_cpus);
543 free_cpumask_var(cs->subparts_cpus);
546 free_cpumask_var(tmp->new_cpus);
547 free_cpumask_var(tmp->addmask);
548 free_cpumask_var(tmp->delmask);
553 * alloc_trial_cpuset - allocate a trial cpuset
554 * @cs: the cpuset that the trial cpuset duplicates
556 static struct cpuset *alloc_trial_cpuset(struct cpuset *cs)
558 struct cpuset *trial;
560 trial = kmemdup(cs, sizeof(*cs), GFP_KERNEL);
564 if (alloc_cpumasks(trial, NULL)) {
569 cpumask_copy(trial->cpus_allowed, cs->cpus_allowed);
570 cpumask_copy(trial->effective_cpus, cs->effective_cpus);
575 * free_cpuset - free the cpuset
576 * @cs: the cpuset to be freed
578 static inline void free_cpuset(struct cpuset *cs)
580 free_cpumasks(cs, NULL);
585 * validate_change() - Used to validate that any proposed cpuset change
586 * follows the structural rules for cpusets.
588 * If we replaced the flag and mask values of the current cpuset
589 * (cur) with those values in the trial cpuset (trial), would
590 * our various subset and exclusive rules still be valid? Presumes
593 * 'cur' is the address of an actual, in-use cpuset. Operations
594 * such as list traversal that depend on the actual address of the
595 * cpuset in the list must use cur below, not trial.
597 * 'trial' is the address of bulk structure copy of cur, with
598 * perhaps one or more of the fields cpus_allowed, mems_allowed,
599 * or flags changed to new, trial values.
601 * Return 0 if valid, -errno if not.
604 static int validate_change(struct cpuset *cur, struct cpuset *trial)
606 struct cgroup_subsys_state *css;
607 struct cpuset *c, *par;
612 /* Each of our child cpusets must be a subset of us */
614 cpuset_for_each_child(c, css, cur)
615 if (!is_cpuset_subset(c, trial))
618 /* Remaining checks don't apply to root cpuset */
620 if (cur == &top_cpuset)
623 par = parent_cs(cur);
625 /* On legacy hiearchy, we must be a subset of our parent cpuset. */
627 if (!is_in_v2_mode() && !is_cpuset_subset(trial, par))
631 * If either I or some sibling (!= me) is exclusive, we can't
635 cpuset_for_each_child(c, css, par) {
636 if ((is_cpu_exclusive(trial) || is_cpu_exclusive(c)) &&
638 cpumask_intersects(trial->cpus_allowed, c->cpus_allowed))
640 if ((is_mem_exclusive(trial) || is_mem_exclusive(c)) &&
642 nodes_intersects(trial->mems_allowed, c->mems_allowed))
647 * Cpusets with tasks - existing or newly being attached - can't
648 * be changed to have empty cpus_allowed or mems_allowed.
651 if ((cgroup_is_populated(cur->css.cgroup) || cur->attach_in_progress)) {
652 if (!cpumask_empty(cur->cpus_allowed) &&
653 cpumask_empty(trial->cpus_allowed))
655 if (!nodes_empty(cur->mems_allowed) &&
656 nodes_empty(trial->mems_allowed))
661 * We can't shrink if we won't have enough room for SCHED_DEADLINE
665 if (is_cpu_exclusive(cur) &&
666 !cpuset_cpumask_can_shrink(cur->cpus_allowed,
667 trial->cpus_allowed))
678 * Helper routine for generate_sched_domains().
679 * Do cpusets a, b have overlapping effective cpus_allowed masks?
681 static int cpusets_overlap(struct cpuset *a, struct cpuset *b)
683 return cpumask_intersects(a->effective_cpus, b->effective_cpus);
687 update_domain_attr(struct sched_domain_attr *dattr, struct cpuset *c)
689 if (dattr->relax_domain_level < c->relax_domain_level)
690 dattr->relax_domain_level = c->relax_domain_level;
694 static void update_domain_attr_tree(struct sched_domain_attr *dattr,
695 struct cpuset *root_cs)
698 struct cgroup_subsys_state *pos_css;
701 cpuset_for_each_descendant_pre(cp, pos_css, root_cs) {
702 /* skip the whole subtree if @cp doesn't have any CPU */
703 if (cpumask_empty(cp->cpus_allowed)) {
704 pos_css = css_rightmost_descendant(pos_css);
708 if (is_sched_load_balance(cp))
709 update_domain_attr(dattr, cp);
714 /* Must be called with cpuset_mutex held. */
715 static inline int nr_cpusets(void)
717 /* jump label reference count + the top-level cpuset */
718 return static_key_count(&cpusets_enabled_key.key) + 1;
722 * generate_sched_domains()
724 * This function builds a partial partition of the systems CPUs
725 * A 'partial partition' is a set of non-overlapping subsets whose
726 * union is a subset of that set.
727 * The output of this function needs to be passed to kernel/sched/core.c
728 * partition_sched_domains() routine, which will rebuild the scheduler's
729 * load balancing domains (sched domains) as specified by that partial
732 * See "What is sched_load_balance" in Documentation/cgroup-v1/cpusets.txt
733 * for a background explanation of this.
735 * Does not return errors, on the theory that the callers of this
736 * routine would rather not worry about failures to rebuild sched
737 * domains when operating in the severe memory shortage situations
738 * that could cause allocation failures below.
740 * Must be called with cpuset_mutex held.
742 * The three key local variables below are:
743 * q - a linked-list queue of cpuset pointers, used to implement a
744 * top-down scan of all cpusets. This scan loads a pointer
745 * to each cpuset marked is_sched_load_balance into the
746 * array 'csa'. For our purposes, rebuilding the schedulers
747 * sched domains, we can ignore !is_sched_load_balance cpusets.
748 * csa - (for CpuSet Array) Array of pointers to all the cpusets
749 * that need to be load balanced, for convenient iterative
750 * access by the subsequent code that finds the best partition,
751 * i.e the set of domains (subsets) of CPUs such that the
752 * cpus_allowed of every cpuset marked is_sched_load_balance
753 * is a subset of one of these domains, while there are as
754 * many such domains as possible, each as small as possible.
755 * doms - Conversion of 'csa' to an array of cpumasks, for passing to
756 * the kernel/sched/core.c routine partition_sched_domains() in a
757 * convenient format, that can be easily compared to the prior
758 * value to determine what partition elements (sched domains)
759 * were changed (added or removed.)
761 * Finding the best partition (set of domains):
762 * The triple nested loops below over i, j, k scan over the
763 * load balanced cpusets (using the array of cpuset pointers in
764 * csa[]) looking for pairs of cpusets that have overlapping
765 * cpus_allowed, but which don't have the same 'pn' partition
766 * number and gives them in the same partition number. It keeps
767 * looping on the 'restart' label until it can no longer find
770 * The union of the cpus_allowed masks from the set of
771 * all cpusets having the same 'pn' value then form the one
772 * element of the partition (one sched domain) to be passed to
773 * partition_sched_domains().
775 static int generate_sched_domains(cpumask_var_t **domains,
776 struct sched_domain_attr **attributes)
778 struct cpuset *cp; /* scans q */
779 struct cpuset **csa; /* array of all cpuset ptrs */
780 int csn; /* how many cpuset ptrs in csa so far */
781 int i, j, k; /* indices for partition finding loops */
782 cpumask_var_t *doms; /* resulting partition; i.e. sched domains */
783 struct sched_domain_attr *dattr; /* attributes for custom domains */
784 int ndoms = 0; /* number of sched domains in result */
785 int nslot; /* next empty doms[] struct cpumask slot */
786 struct cgroup_subsys_state *pos_css;
787 bool root_load_balance = is_sched_load_balance(&top_cpuset);
793 /* Special case for the 99% of systems with one, full, sched domain */
794 if (root_load_balance && !top_cpuset.nr_subparts_cpus) {
796 doms = alloc_sched_domains(ndoms);
800 dattr = kmalloc(sizeof(struct sched_domain_attr), GFP_KERNEL);
802 *dattr = SD_ATTR_INIT;
803 update_domain_attr_tree(dattr, &top_cpuset);
805 cpumask_and(doms[0], top_cpuset.effective_cpus,
806 housekeeping_cpumask(HK_FLAG_DOMAIN));
811 csa = kmalloc_array(nr_cpusets(), sizeof(cp), GFP_KERNEL);
817 if (root_load_balance)
818 csa[csn++] = &top_cpuset;
819 cpuset_for_each_descendant_pre(cp, pos_css, &top_cpuset) {
820 if (cp == &top_cpuset)
823 * Continue traversing beyond @cp iff @cp has some CPUs and
824 * isn't load balancing. The former is obvious. The
825 * latter: All child cpusets contain a subset of the
826 * parent's cpus, so just skip them, and then we call
827 * update_domain_attr_tree() to calc relax_domain_level of
828 * the corresponding sched domain.
830 * If root is load-balancing, we can skip @cp if it
831 * is a subset of the root's effective_cpus.
833 if (!cpumask_empty(cp->cpus_allowed) &&
834 !(is_sched_load_balance(cp) &&
835 cpumask_intersects(cp->cpus_allowed,
836 housekeeping_cpumask(HK_FLAG_DOMAIN))))
839 if (root_load_balance &&
840 cpumask_subset(cp->cpus_allowed, top_cpuset.effective_cpus))
843 if (is_sched_load_balance(cp))
846 /* skip @cp's subtree if not a partition root */
847 if (!is_partition_root(cp))
848 pos_css = css_rightmost_descendant(pos_css);
852 for (i = 0; i < csn; i++)
857 /* Find the best partition (set of sched domains) */
858 for (i = 0; i < csn; i++) {
859 struct cpuset *a = csa[i];
862 for (j = 0; j < csn; j++) {
863 struct cpuset *b = csa[j];
866 if (apn != bpn && cpusets_overlap(a, b)) {
867 for (k = 0; k < csn; k++) {
868 struct cpuset *c = csa[k];
873 ndoms--; /* one less element */
880 * Now we know how many domains to create.
881 * Convert <csn, csa> to <ndoms, doms> and populate cpu masks.
883 doms = alloc_sched_domains(ndoms);
888 * The rest of the code, including the scheduler, can deal with
889 * dattr==NULL case. No need to abort if alloc fails.
891 dattr = kmalloc_array(ndoms, sizeof(struct sched_domain_attr),
894 for (nslot = 0, i = 0; i < csn; i++) {
895 struct cpuset *a = csa[i];
900 /* Skip completed partitions */
906 if (nslot == ndoms) {
907 static int warnings = 10;
909 pr_warn("rebuild_sched_domains confused: nslot %d, ndoms %d, csn %d, i %d, apn %d\n",
910 nslot, ndoms, csn, i, apn);
918 *(dattr + nslot) = SD_ATTR_INIT;
919 for (j = i; j < csn; j++) {
920 struct cpuset *b = csa[j];
923 cpumask_or(dp, dp, b->effective_cpus);
924 cpumask_and(dp, dp, housekeeping_cpumask(HK_FLAG_DOMAIN));
926 update_domain_attr_tree(dattr + nslot, b);
928 /* Done with this partition */
934 BUG_ON(nslot != ndoms);
940 * Fallback to the default domain if kmalloc() failed.
941 * See comments in partition_sched_domains().
952 * Rebuild scheduler domains.
954 * If the flag 'sched_load_balance' of any cpuset with non-empty
955 * 'cpus' changes, or if the 'cpus' allowed changes in any cpuset
956 * which has that flag enabled, or if any cpuset with a non-empty
957 * 'cpus' is removed, then call this routine to rebuild the
958 * scheduler's dynamic sched domains.
960 * Call with cpuset_mutex held. Takes get_online_cpus().
962 static void rebuild_sched_domains_locked(void)
964 struct sched_domain_attr *attr;
968 lockdep_assert_held(&cpuset_mutex);
972 * We have raced with CPU hotplug. Don't do anything to avoid
973 * passing doms with offlined cpu to partition_sched_domains().
974 * Anyways, hotplug work item will rebuild sched domains.
976 if (!top_cpuset.nr_subparts_cpus &&
977 !cpumask_equal(top_cpuset.effective_cpus, cpu_active_mask))
980 if (top_cpuset.nr_subparts_cpus &&
981 !cpumask_subset(top_cpuset.effective_cpus, cpu_active_mask))
984 /* Generate domain masks and attrs */
985 ndoms = generate_sched_domains(&doms, &attr);
987 /* Have scheduler rebuild the domains */
988 partition_sched_domains(ndoms, doms, attr);
992 #else /* !CONFIG_SMP */
993 static void rebuild_sched_domains_locked(void)
996 #endif /* CONFIG_SMP */
998 void rebuild_sched_domains(void)
1000 mutex_lock(&cpuset_mutex);
1001 rebuild_sched_domains_locked();
1002 mutex_unlock(&cpuset_mutex);
1006 * update_tasks_cpumask - Update the cpumasks of tasks in the cpuset.
1007 * @cs: the cpuset in which each task's cpus_allowed mask needs to be changed
1009 * Iterate through each task of @cs updating its cpus_allowed to the
1010 * effective cpuset's. As this function is called with cpuset_mutex held,
1011 * cpuset membership stays stable.
1013 static void update_tasks_cpumask(struct cpuset *cs)
1015 struct css_task_iter it;
1016 struct task_struct *task;
1018 css_task_iter_start(&cs->css, 0, &it);
1019 while ((task = css_task_iter_next(&it)))
1020 set_cpus_allowed_ptr(task, cs->effective_cpus);
1021 css_task_iter_end(&it);
1025 * compute_effective_cpumask - Compute the effective cpumask of the cpuset
1026 * @new_cpus: the temp variable for the new effective_cpus mask
1027 * @cs: the cpuset the need to recompute the new effective_cpus mask
1028 * @parent: the parent cpuset
1030 * If the parent has subpartition CPUs, include them in the list of
1031 * allowable CPUs in computing the new effective_cpus mask. Since offlined
1032 * CPUs are not removed from subparts_cpus, we have to use cpu_active_mask
1033 * to mask those out.
1035 static void compute_effective_cpumask(struct cpumask *new_cpus,
1036 struct cpuset *cs, struct cpuset *parent)
1038 if (parent->nr_subparts_cpus) {
1039 cpumask_or(new_cpus, parent->effective_cpus,
1040 parent->subparts_cpus);
1041 cpumask_and(new_cpus, new_cpus, cs->cpus_allowed);
1042 cpumask_and(new_cpus, new_cpus, cpu_active_mask);
1044 cpumask_and(new_cpus, cs->cpus_allowed, parent->effective_cpus);
1049 * Commands for update_parent_subparts_cpumask
1052 partcmd_enable, /* Enable partition root */
1053 partcmd_disable, /* Disable partition root */
1054 partcmd_update, /* Update parent's subparts_cpus */
1058 * update_parent_subparts_cpumask - update subparts_cpus mask of parent cpuset
1059 * @cpuset: The cpuset that requests change in partition root state
1060 * @cmd: Partition root state change command
1061 * @newmask: Optional new cpumask for partcmd_update
1062 * @tmp: Temporary addmask and delmask
1063 * Return: 0, 1 or an error code
1065 * For partcmd_enable, the cpuset is being transformed from a non-partition
1066 * root to a partition root. The cpus_allowed mask of the given cpuset will
1067 * be put into parent's subparts_cpus and taken away from parent's
1068 * effective_cpus. The function will return 0 if all the CPUs listed in
1069 * cpus_allowed can be granted or an error code will be returned.
1071 * For partcmd_disable, the cpuset is being transofrmed from a partition
1072 * root back to a non-partition root. any CPUs in cpus_allowed that are in
1073 * parent's subparts_cpus will be taken away from that cpumask and put back
1074 * into parent's effective_cpus. 0 should always be returned.
1076 * For partcmd_update, if the optional newmask is specified, the cpu
1077 * list is to be changed from cpus_allowed to newmask. Otherwise,
1078 * cpus_allowed is assumed to remain the same. The cpuset should either
1079 * be a partition root or an invalid partition root. The partition root
1080 * state may change if newmask is NULL and none of the requested CPUs can
1081 * be granted by the parent. The function will return 1 if changes to
1082 * parent's subparts_cpus and effective_cpus happen or 0 otherwise.
1083 * Error code should only be returned when newmask is non-NULL.
1085 * The partcmd_enable and partcmd_disable commands are used by
1086 * update_prstate(). The partcmd_update command is used by
1087 * update_cpumasks_hier() with newmask NULL and update_cpumask() with
1090 * The checking is more strict when enabling partition root than the
1091 * other two commands.
1093 * Because of the implicit cpu exclusive nature of a partition root,
1094 * cpumask changes that violates the cpu exclusivity rule will not be
1095 * permitted when checked by validate_change(). The validate_change()
1096 * function will also prevent any changes to the cpu list if it is not
1097 * a superset of children's cpu lists.
1099 static int update_parent_subparts_cpumask(struct cpuset *cpuset, int cmd,
1100 struct cpumask *newmask,
1101 struct tmpmasks *tmp)
1103 struct cpuset *parent = parent_cs(cpuset);
1104 int adding; /* Moving cpus from effective_cpus to subparts_cpus */
1105 int deleting; /* Moving cpus from subparts_cpus to effective_cpus */
1106 bool part_error = false; /* Partition error? */
1108 lockdep_assert_held(&cpuset_mutex);
1111 * The parent must be a partition root.
1112 * The new cpumask, if present, or the current cpus_allowed must
1115 if (!is_partition_root(parent) ||
1116 (newmask && cpumask_empty(newmask)) ||
1117 (!newmask && cpumask_empty(cpuset->cpus_allowed)))
1121 * Enabling/disabling partition root is not allowed if there are
1124 if ((cmd != partcmd_update) && css_has_online_children(&cpuset->css))
1128 * Enabling partition root is not allowed if not all the CPUs
1129 * can be granted from parent's effective_cpus or at least one
1130 * CPU will be left after that.
1132 if ((cmd == partcmd_enable) &&
1133 (!cpumask_subset(cpuset->cpus_allowed, parent->effective_cpus) ||
1134 cpumask_equal(cpuset->cpus_allowed, parent->effective_cpus)))
1138 * A cpumask update cannot make parent's effective_cpus become empty.
1140 adding = deleting = false;
1141 if (cmd == partcmd_enable) {
1142 cpumask_copy(tmp->addmask, cpuset->cpus_allowed);
1144 } else if (cmd == partcmd_disable) {
1145 deleting = cpumask_and(tmp->delmask, cpuset->cpus_allowed,
1146 parent->subparts_cpus);
1147 } else if (newmask) {
1149 * partcmd_update with newmask:
1151 * delmask = cpus_allowed & ~newmask & parent->subparts_cpus
1152 * addmask = newmask & parent->effective_cpus
1153 * & ~parent->subparts_cpus
1155 cpumask_andnot(tmp->delmask, cpuset->cpus_allowed, newmask);
1156 deleting = cpumask_and(tmp->delmask, tmp->delmask,
1157 parent->subparts_cpus);
1159 cpumask_and(tmp->addmask, newmask, parent->effective_cpus);
1160 adding = cpumask_andnot(tmp->addmask, tmp->addmask,
1161 parent->subparts_cpus);
1163 * Return error if the new effective_cpus could become empty.
1166 cpumask_equal(parent->effective_cpus, tmp->addmask)) {
1170 * As some of the CPUs in subparts_cpus might have
1171 * been offlined, we need to compute the real delmask
1174 if (!cpumask_and(tmp->addmask, tmp->delmask,
1177 cpumask_copy(tmp->addmask, parent->effective_cpus);
1181 * partcmd_update w/o newmask:
1183 * addmask = cpus_allowed & parent->effectiveb_cpus
1185 * Note that parent's subparts_cpus may have been
1186 * pre-shrunk in case there is a change in the cpu list.
1187 * So no deletion is needed.
1189 adding = cpumask_and(tmp->addmask, cpuset->cpus_allowed,
1190 parent->effective_cpus);
1191 part_error = cpumask_equal(tmp->addmask,
1192 parent->effective_cpus);
1195 if (cmd == partcmd_update) {
1196 int prev_prs = cpuset->partition_root_state;
1199 * Check for possible transition between PRS_ENABLED
1202 switch (cpuset->partition_root_state) {
1205 cpuset->partition_root_state = PRS_ERROR;
1209 cpuset->partition_root_state = PRS_ENABLED;
1213 * Set part_error if previously in invalid state.
1215 part_error = (prev_prs == PRS_ERROR);
1218 if (!part_error && (cpuset->partition_root_state == PRS_ERROR))
1219 return 0; /* Nothing need to be done */
1221 if (cpuset->partition_root_state == PRS_ERROR) {
1223 * Remove all its cpus from parent's subparts_cpus.
1226 deleting = cpumask_and(tmp->delmask, cpuset->cpus_allowed,
1227 parent->subparts_cpus);
1230 if (!adding && !deleting)
1234 * Change the parent's subparts_cpus.
1235 * Newly added CPUs will be removed from effective_cpus and
1236 * newly deleted ones will be added back to effective_cpus.
1238 spin_lock_irq(&callback_lock);
1240 cpumask_or(parent->subparts_cpus,
1241 parent->subparts_cpus, tmp->addmask);
1242 cpumask_andnot(parent->effective_cpus,
1243 parent->effective_cpus, tmp->addmask);
1246 cpumask_andnot(parent->subparts_cpus,
1247 parent->subparts_cpus, tmp->delmask);
1249 * Some of the CPUs in subparts_cpus might have been offlined.
1251 cpumask_and(tmp->delmask, tmp->delmask, cpu_active_mask);
1252 cpumask_or(parent->effective_cpus,
1253 parent->effective_cpus, tmp->delmask);
1256 parent->nr_subparts_cpus = cpumask_weight(parent->subparts_cpus);
1257 spin_unlock_irq(&callback_lock);
1259 return cmd == partcmd_update;
1263 * update_cpumasks_hier - Update effective cpumasks and tasks in the subtree
1264 * @cs: the cpuset to consider
1265 * @tmp: temp variables for calculating effective_cpus & partition setup
1267 * When congifured cpumask is changed, the effective cpumasks of this cpuset
1268 * and all its descendants need to be updated.
1270 * On legacy hierachy, effective_cpus will be the same with cpu_allowed.
1272 * Called with cpuset_mutex held
1274 static void update_cpumasks_hier(struct cpuset *cs, struct tmpmasks *tmp)
1277 struct cgroup_subsys_state *pos_css;
1278 bool need_rebuild_sched_domains = false;
1281 cpuset_for_each_descendant_pre(cp, pos_css, cs) {
1282 struct cpuset *parent = parent_cs(cp);
1284 compute_effective_cpumask(tmp->new_cpus, cp, parent);
1287 * If it becomes empty, inherit the effective mask of the
1288 * parent, which is guaranteed to have some CPUs.
1290 if (is_in_v2_mode() && cpumask_empty(tmp->new_cpus)) {
1291 cpumask_copy(tmp->new_cpus, parent->effective_cpus);
1292 if (!cp->use_parent_ecpus) {
1293 cp->use_parent_ecpus = true;
1294 parent->child_ecpus_count++;
1296 } else if (cp->use_parent_ecpus) {
1297 cp->use_parent_ecpus = false;
1298 WARN_ON_ONCE(!parent->child_ecpus_count);
1299 parent->child_ecpus_count--;
1303 * Skip the whole subtree if the cpumask remains the same
1304 * and has no partition root state.
1306 if (!cp->partition_root_state &&
1307 cpumask_equal(tmp->new_cpus, cp->effective_cpus)) {
1308 pos_css = css_rightmost_descendant(pos_css);
1313 * update_parent_subparts_cpumask() should have been called
1314 * for cs already in update_cpumask(). We should also call
1315 * update_tasks_cpumask() again for tasks in the parent
1316 * cpuset if the parent's subparts_cpus changes.
1318 if ((cp != cs) && cp->partition_root_state) {
1319 switch (parent->partition_root_state) {
1322 * If parent is not a partition root or an
1323 * invalid partition root, clear the state
1324 * state and the CS_CPU_EXCLUSIVE flag.
1326 WARN_ON_ONCE(cp->partition_root_state
1328 cp->partition_root_state = 0;
1331 * clear_bit() is an atomic operation and
1332 * readers aren't interested in the state
1333 * of CS_CPU_EXCLUSIVE anyway. So we can
1334 * just update the flag without holding
1335 * the callback_lock.
1337 clear_bit(CS_CPU_EXCLUSIVE, &cp->flags);
1341 if (update_parent_subparts_cpumask(cp, partcmd_update, NULL, tmp))
1342 update_tasks_cpumask(parent);
1347 * When parent is invalid, it has to be too.
1349 cp->partition_root_state = PRS_ERROR;
1350 if (cp->nr_subparts_cpus) {
1351 cp->nr_subparts_cpus = 0;
1352 cpumask_clear(cp->subparts_cpus);
1358 if (!css_tryget_online(&cp->css))
1362 spin_lock_irq(&callback_lock);
1364 cpumask_copy(cp->effective_cpus, tmp->new_cpus);
1365 if (cp->nr_subparts_cpus &&
1366 (cp->partition_root_state != PRS_ENABLED)) {
1367 cp->nr_subparts_cpus = 0;
1368 cpumask_clear(cp->subparts_cpus);
1369 } else if (cp->nr_subparts_cpus) {
1371 * Make sure that effective_cpus & subparts_cpus
1372 * are mutually exclusive.
1374 * In the unlikely event that effective_cpus
1375 * becomes empty. we clear cp->nr_subparts_cpus and
1376 * let its child partition roots to compete for
1379 cpumask_andnot(cp->effective_cpus, cp->effective_cpus,
1381 if (cpumask_empty(cp->effective_cpus)) {
1382 cpumask_copy(cp->effective_cpus, tmp->new_cpus);
1383 cpumask_clear(cp->subparts_cpus);
1384 cp->nr_subparts_cpus = 0;
1385 } else if (!cpumask_subset(cp->subparts_cpus,
1387 cpumask_andnot(cp->subparts_cpus,
1388 cp->subparts_cpus, tmp->new_cpus);
1389 cp->nr_subparts_cpus
1390 = cpumask_weight(cp->subparts_cpus);
1393 spin_unlock_irq(&callback_lock);
1395 WARN_ON(!is_in_v2_mode() &&
1396 !cpumask_equal(cp->cpus_allowed, cp->effective_cpus));
1398 update_tasks_cpumask(cp);
1401 * On legacy hierarchy, if the effective cpumask of any non-
1402 * empty cpuset is changed, we need to rebuild sched domains.
1403 * On default hierarchy, the cpuset needs to be a partition
1406 if (!cpumask_empty(cp->cpus_allowed) &&
1407 is_sched_load_balance(cp) &&
1408 (!cgroup_subsys_on_dfl(cpuset_cgrp_subsys) ||
1409 is_partition_root(cp)))
1410 need_rebuild_sched_domains = true;
1417 if (need_rebuild_sched_domains)
1418 rebuild_sched_domains_locked();
1422 * update_sibling_cpumasks - Update siblings cpumasks
1423 * @parent: Parent cpuset
1424 * @cs: Current cpuset
1425 * @tmp: Temp variables
1427 static void update_sibling_cpumasks(struct cpuset *parent, struct cpuset *cs,
1428 struct tmpmasks *tmp)
1430 struct cpuset *sibling;
1431 struct cgroup_subsys_state *pos_css;
1434 * Check all its siblings and call update_cpumasks_hier()
1435 * if their use_parent_ecpus flag is set in order for them
1436 * to use the right effective_cpus value.
1439 cpuset_for_each_child(sibling, pos_css, parent) {
1442 if (!sibling->use_parent_ecpus)
1445 update_cpumasks_hier(sibling, tmp);
1451 * update_cpumask - update the cpus_allowed mask of a cpuset and all tasks in it
1452 * @cs: the cpuset to consider
1453 * @trialcs: trial cpuset
1454 * @buf: buffer of cpu numbers written to this cpuset
1456 static int update_cpumask(struct cpuset *cs, struct cpuset *trialcs,
1460 struct tmpmasks tmp;
1462 /* top_cpuset.cpus_allowed tracks cpu_online_mask; it's read-only */
1463 if (cs == &top_cpuset)
1467 * An empty cpus_allowed is ok only if the cpuset has no tasks.
1468 * Since cpulist_parse() fails on an empty mask, we special case
1469 * that parsing. The validate_change() call ensures that cpusets
1470 * with tasks have cpus.
1473 cpumask_clear(trialcs->cpus_allowed);
1475 retval = cpulist_parse(buf, trialcs->cpus_allowed);
1479 if (!cpumask_subset(trialcs->cpus_allowed,
1480 top_cpuset.cpus_allowed))
1484 /* Nothing to do if the cpus didn't change */
1485 if (cpumask_equal(cs->cpus_allowed, trialcs->cpus_allowed))
1488 retval = validate_change(cs, trialcs);
1492 #ifdef CONFIG_CPUMASK_OFFSTACK
1494 * Use the cpumasks in trialcs for tmpmasks when they are pointers
1495 * to allocated cpumasks.
1497 tmp.addmask = trialcs->subparts_cpus;
1498 tmp.delmask = trialcs->effective_cpus;
1499 tmp.new_cpus = trialcs->cpus_allowed;
1502 if (cs->partition_root_state) {
1503 /* Cpumask of a partition root cannot be empty */
1504 if (cpumask_empty(trialcs->cpus_allowed))
1506 if (update_parent_subparts_cpumask(cs, partcmd_update,
1507 trialcs->cpus_allowed, &tmp) < 0)
1511 spin_lock_irq(&callback_lock);
1512 cpumask_copy(cs->cpus_allowed, trialcs->cpus_allowed);
1515 * Make sure that subparts_cpus is a subset of cpus_allowed.
1517 if (cs->nr_subparts_cpus) {
1518 cpumask_andnot(cs->subparts_cpus, cs->subparts_cpus,
1520 cs->nr_subparts_cpus = cpumask_weight(cs->subparts_cpus);
1522 spin_unlock_irq(&callback_lock);
1524 update_cpumasks_hier(cs, &tmp);
1526 if (cs->partition_root_state) {
1527 struct cpuset *parent = parent_cs(cs);
1530 * For partition root, update the cpumasks of sibling
1531 * cpusets if they use parent's effective_cpus.
1533 if (parent->child_ecpus_count)
1534 update_sibling_cpumasks(parent, cs, &tmp);
1540 * Migrate memory region from one set of nodes to another. This is
1541 * performed asynchronously as it can be called from process migration path
1542 * holding locks involved in process management. All mm migrations are
1543 * performed in the queued order and can be waited for by flushing
1544 * cpuset_migrate_mm_wq.
1547 struct cpuset_migrate_mm_work {
1548 struct work_struct work;
1549 struct mm_struct *mm;
1554 static void cpuset_migrate_mm_workfn(struct work_struct *work)
1556 struct cpuset_migrate_mm_work *mwork =
1557 container_of(work, struct cpuset_migrate_mm_work, work);
1559 /* on a wq worker, no need to worry about %current's mems_allowed */
1560 do_migrate_pages(mwork->mm, &mwork->from, &mwork->to, MPOL_MF_MOVE_ALL);
1565 static void cpuset_migrate_mm(struct mm_struct *mm, const nodemask_t *from,
1566 const nodemask_t *to)
1568 struct cpuset_migrate_mm_work *mwork;
1570 mwork = kzalloc(sizeof(*mwork), GFP_KERNEL);
1573 mwork->from = *from;
1575 INIT_WORK(&mwork->work, cpuset_migrate_mm_workfn);
1576 queue_work(cpuset_migrate_mm_wq, &mwork->work);
1582 static void cpuset_post_attach(void)
1584 flush_workqueue(cpuset_migrate_mm_wq);
1588 * cpuset_change_task_nodemask - change task's mems_allowed and mempolicy
1589 * @tsk: the task to change
1590 * @newmems: new nodes that the task will be set
1592 * We use the mems_allowed_seq seqlock to safely update both tsk->mems_allowed
1593 * and rebind an eventual tasks' mempolicy. If the task is allocating in
1594 * parallel, it might temporarily see an empty intersection, which results in
1595 * a seqlock check and retry before OOM or allocation failure.
1597 static void cpuset_change_task_nodemask(struct task_struct *tsk,
1598 nodemask_t *newmems)
1602 local_irq_disable();
1603 write_seqcount_begin(&tsk->mems_allowed_seq);
1605 nodes_or(tsk->mems_allowed, tsk->mems_allowed, *newmems);
1606 mpol_rebind_task(tsk, newmems);
1607 tsk->mems_allowed = *newmems;
1609 write_seqcount_end(&tsk->mems_allowed_seq);
1615 static void *cpuset_being_rebound;
1618 * update_tasks_nodemask - Update the nodemasks of tasks in the cpuset.
1619 * @cs: the cpuset in which each task's mems_allowed mask needs to be changed
1621 * Iterate through each task of @cs updating its mems_allowed to the
1622 * effective cpuset's. As this function is called with cpuset_mutex held,
1623 * cpuset membership stays stable.
1625 static void update_tasks_nodemask(struct cpuset *cs)
1627 static nodemask_t newmems; /* protected by cpuset_mutex */
1628 struct css_task_iter it;
1629 struct task_struct *task;
1631 cpuset_being_rebound = cs; /* causes mpol_dup() rebind */
1633 guarantee_online_mems(cs, &newmems);
1636 * The mpol_rebind_mm() call takes mmap_sem, which we couldn't
1637 * take while holding tasklist_lock. Forks can happen - the
1638 * mpol_dup() cpuset_being_rebound check will catch such forks,
1639 * and rebind their vma mempolicies too. Because we still hold
1640 * the global cpuset_mutex, we know that no other rebind effort
1641 * will be contending for the global variable cpuset_being_rebound.
1642 * It's ok if we rebind the same mm twice; mpol_rebind_mm()
1643 * is idempotent. Also migrate pages in each mm to new nodes.
1645 css_task_iter_start(&cs->css, 0, &it);
1646 while ((task = css_task_iter_next(&it))) {
1647 struct mm_struct *mm;
1650 cpuset_change_task_nodemask(task, &newmems);
1652 mm = get_task_mm(task);
1656 migrate = is_memory_migrate(cs);
1658 mpol_rebind_mm(mm, &cs->mems_allowed);
1660 cpuset_migrate_mm(mm, &cs->old_mems_allowed, &newmems);
1664 css_task_iter_end(&it);
1667 * All the tasks' nodemasks have been updated, update
1668 * cs->old_mems_allowed.
1670 cs->old_mems_allowed = newmems;
1672 /* We're done rebinding vmas to this cpuset's new mems_allowed. */
1673 cpuset_being_rebound = NULL;
1677 * update_nodemasks_hier - Update effective nodemasks and tasks in the subtree
1678 * @cs: the cpuset to consider
1679 * @new_mems: a temp variable for calculating new effective_mems
1681 * When configured nodemask is changed, the effective nodemasks of this cpuset
1682 * and all its descendants need to be updated.
1684 * On legacy hiearchy, effective_mems will be the same with mems_allowed.
1686 * Called with cpuset_mutex held
1688 static void update_nodemasks_hier(struct cpuset *cs, nodemask_t *new_mems)
1691 struct cgroup_subsys_state *pos_css;
1694 cpuset_for_each_descendant_pre(cp, pos_css, cs) {
1695 struct cpuset *parent = parent_cs(cp);
1697 nodes_and(*new_mems, cp->mems_allowed, parent->effective_mems);
1700 * If it becomes empty, inherit the effective mask of the
1701 * parent, which is guaranteed to have some MEMs.
1703 if (is_in_v2_mode() && nodes_empty(*new_mems))
1704 *new_mems = parent->effective_mems;
1706 /* Skip the whole subtree if the nodemask remains the same. */
1707 if (nodes_equal(*new_mems, cp->effective_mems)) {
1708 pos_css = css_rightmost_descendant(pos_css);
1712 if (!css_tryget_online(&cp->css))
1716 spin_lock_irq(&callback_lock);
1717 cp->effective_mems = *new_mems;
1718 spin_unlock_irq(&callback_lock);
1720 WARN_ON(!is_in_v2_mode() &&
1721 !nodes_equal(cp->mems_allowed, cp->effective_mems));
1723 update_tasks_nodemask(cp);
1732 * Handle user request to change the 'mems' memory placement
1733 * of a cpuset. Needs to validate the request, update the
1734 * cpusets mems_allowed, and for each task in the cpuset,
1735 * update mems_allowed and rebind task's mempolicy and any vma
1736 * mempolicies and if the cpuset is marked 'memory_migrate',
1737 * migrate the tasks pages to the new memory.
1739 * Call with cpuset_mutex held. May take callback_lock during call.
1740 * Will take tasklist_lock, scan tasklist for tasks in cpuset cs,
1741 * lock each such tasks mm->mmap_sem, scan its vma's and rebind
1742 * their mempolicies to the cpusets new mems_allowed.
1744 static int update_nodemask(struct cpuset *cs, struct cpuset *trialcs,
1750 * top_cpuset.mems_allowed tracks node_stats[N_MEMORY];
1753 if (cs == &top_cpuset) {
1759 * An empty mems_allowed is ok iff there are no tasks in the cpuset.
1760 * Since nodelist_parse() fails on an empty mask, we special case
1761 * that parsing. The validate_change() call ensures that cpusets
1762 * with tasks have memory.
1765 nodes_clear(trialcs->mems_allowed);
1767 retval = nodelist_parse(buf, trialcs->mems_allowed);
1771 if (!nodes_subset(trialcs->mems_allowed,
1772 top_cpuset.mems_allowed)) {
1778 if (nodes_equal(cs->mems_allowed, trialcs->mems_allowed)) {
1779 retval = 0; /* Too easy - nothing to do */
1782 retval = validate_change(cs, trialcs);
1786 spin_lock_irq(&callback_lock);
1787 cs->mems_allowed = trialcs->mems_allowed;
1788 spin_unlock_irq(&callback_lock);
1790 /* use trialcs->mems_allowed as a temp variable */
1791 update_nodemasks_hier(cs, &trialcs->mems_allowed);
1796 bool current_cpuset_is_being_rebound(void)
1801 ret = task_cs(current) == cpuset_being_rebound;
1807 static int update_relax_domain_level(struct cpuset *cs, s64 val)
1810 if (val < -1 || val >= sched_domain_level_max)
1814 if (val != cs->relax_domain_level) {
1815 cs->relax_domain_level = val;
1816 if (!cpumask_empty(cs->cpus_allowed) &&
1817 is_sched_load_balance(cs))
1818 rebuild_sched_domains_locked();
1825 * update_tasks_flags - update the spread flags of tasks in the cpuset.
1826 * @cs: the cpuset in which each task's spread flags needs to be changed
1828 * Iterate through each task of @cs updating its spread flags. As this
1829 * function is called with cpuset_mutex held, cpuset membership stays
1832 static void update_tasks_flags(struct cpuset *cs)
1834 struct css_task_iter it;
1835 struct task_struct *task;
1837 css_task_iter_start(&cs->css, 0, &it);
1838 while ((task = css_task_iter_next(&it)))
1839 cpuset_update_task_spread_flag(cs, task);
1840 css_task_iter_end(&it);
1844 * update_flag - read a 0 or a 1 in a file and update associated flag
1845 * bit: the bit to update (see cpuset_flagbits_t)
1846 * cs: the cpuset to update
1847 * turning_on: whether the flag is being set or cleared
1849 * Call with cpuset_mutex held.
1852 static int update_flag(cpuset_flagbits_t bit, struct cpuset *cs,
1855 struct cpuset *trialcs;
1856 int balance_flag_changed;
1857 int spread_flag_changed;
1860 trialcs = alloc_trial_cpuset(cs);
1865 set_bit(bit, &trialcs->flags);
1867 clear_bit(bit, &trialcs->flags);
1869 err = validate_change(cs, trialcs);
1873 balance_flag_changed = (is_sched_load_balance(cs) !=
1874 is_sched_load_balance(trialcs));
1876 spread_flag_changed = ((is_spread_slab(cs) != is_spread_slab(trialcs))
1877 || (is_spread_page(cs) != is_spread_page(trialcs)));
1879 spin_lock_irq(&callback_lock);
1880 cs->flags = trialcs->flags;
1881 spin_unlock_irq(&callback_lock);
1883 if (!cpumask_empty(trialcs->cpus_allowed) && balance_flag_changed)
1884 rebuild_sched_domains_locked();
1886 if (spread_flag_changed)
1887 update_tasks_flags(cs);
1889 free_cpuset(trialcs);
1894 * update_prstate - update partititon_root_state
1895 * cs: the cpuset to update
1896 * val: 0 - disabled, 1 - enabled
1898 * Call with cpuset_mutex held.
1900 static int update_prstate(struct cpuset *cs, int val)
1903 struct cpuset *parent = parent_cs(cs);
1904 struct tmpmasks tmp;
1906 if ((val != 0) && (val != 1))
1908 if (val == cs->partition_root_state)
1912 * Cannot force a partial or invalid partition root to a full
1915 if (val && cs->partition_root_state)
1918 if (alloc_cpumasks(NULL, &tmp))
1922 if (!cs->partition_root_state) {
1924 * Turning on partition root requires setting the
1925 * CS_CPU_EXCLUSIVE bit implicitly as well and cpus_allowed
1928 if (cpumask_empty(cs->cpus_allowed))
1931 err = update_flag(CS_CPU_EXCLUSIVE, cs, 1);
1935 err = update_parent_subparts_cpumask(cs, partcmd_enable,
1938 update_flag(CS_CPU_EXCLUSIVE, cs, 0);
1941 cs->partition_root_state = PRS_ENABLED;
1944 * Turning off partition root will clear the
1945 * CS_CPU_EXCLUSIVE bit.
1947 if (cs->partition_root_state == PRS_ERROR) {
1948 cs->partition_root_state = 0;
1949 update_flag(CS_CPU_EXCLUSIVE, cs, 0);
1954 err = update_parent_subparts_cpumask(cs, partcmd_disable,
1959 cs->partition_root_state = 0;
1961 /* Turning off CS_CPU_EXCLUSIVE will not return error */
1962 update_flag(CS_CPU_EXCLUSIVE, cs, 0);
1966 * Update cpumask of parent's tasks except when it is the top
1967 * cpuset as some system daemons cannot be mapped to other CPUs.
1969 if (parent != &top_cpuset)
1970 update_tasks_cpumask(parent);
1972 if (parent->child_ecpus_count)
1973 update_sibling_cpumasks(parent, cs, &tmp);
1975 rebuild_sched_domains_locked();
1977 free_cpumasks(NULL, &tmp);
1982 * Frequency meter - How fast is some event occurring?
1984 * These routines manage a digitally filtered, constant time based,
1985 * event frequency meter. There are four routines:
1986 * fmeter_init() - initialize a frequency meter.
1987 * fmeter_markevent() - called each time the event happens.
1988 * fmeter_getrate() - returns the recent rate of such events.
1989 * fmeter_update() - internal routine used to update fmeter.
1991 * A common data structure is passed to each of these routines,
1992 * which is used to keep track of the state required to manage the
1993 * frequency meter and its digital filter.
1995 * The filter works on the number of events marked per unit time.
1996 * The filter is single-pole low-pass recursive (IIR). The time unit
1997 * is 1 second. Arithmetic is done using 32-bit integers scaled to
1998 * simulate 3 decimal digits of precision (multiplied by 1000).
2000 * With an FM_COEF of 933, and a time base of 1 second, the filter
2001 * has a half-life of 10 seconds, meaning that if the events quit
2002 * happening, then the rate returned from the fmeter_getrate()
2003 * will be cut in half each 10 seconds, until it converges to zero.
2005 * It is not worth doing a real infinitely recursive filter. If more
2006 * than FM_MAXTICKS ticks have elapsed since the last filter event,
2007 * just compute FM_MAXTICKS ticks worth, by which point the level
2010 * Limit the count of unprocessed events to FM_MAXCNT, so as to avoid
2011 * arithmetic overflow in the fmeter_update() routine.
2013 * Given the simple 32 bit integer arithmetic used, this meter works
2014 * best for reporting rates between one per millisecond (msec) and
2015 * one per 32 (approx) seconds. At constant rates faster than one
2016 * per msec it maxes out at values just under 1,000,000. At constant
2017 * rates between one per msec, and one per second it will stabilize
2018 * to a value N*1000, where N is the rate of events per second.
2019 * At constant rates between one per second and one per 32 seconds,
2020 * it will be choppy, moving up on the seconds that have an event,
2021 * and then decaying until the next event. At rates slower than
2022 * about one in 32 seconds, it decays all the way back to zero between
2026 #define FM_COEF 933 /* coefficient for half-life of 10 secs */
2027 #define FM_MAXTICKS ((u32)99) /* useless computing more ticks than this */
2028 #define FM_MAXCNT 1000000 /* limit cnt to avoid overflow */
2029 #define FM_SCALE 1000 /* faux fixed point scale */
2031 /* Initialize a frequency meter */
2032 static void fmeter_init(struct fmeter *fmp)
2037 spin_lock_init(&fmp->lock);
2040 /* Internal meter update - process cnt events and update value */
2041 static void fmeter_update(struct fmeter *fmp)
2046 now = ktime_get_seconds();
2047 ticks = now - fmp->time;
2052 ticks = min(FM_MAXTICKS, ticks);
2054 fmp->val = (FM_COEF * fmp->val) / FM_SCALE;
2057 fmp->val += ((FM_SCALE - FM_COEF) * fmp->cnt) / FM_SCALE;
2061 /* Process any previous ticks, then bump cnt by one (times scale). */
2062 static void fmeter_markevent(struct fmeter *fmp)
2064 spin_lock(&fmp->lock);
2066 fmp->cnt = min(FM_MAXCNT, fmp->cnt + FM_SCALE);
2067 spin_unlock(&fmp->lock);
2070 /* Process any previous ticks, then return current value. */
2071 static int fmeter_getrate(struct fmeter *fmp)
2075 spin_lock(&fmp->lock);
2078 spin_unlock(&fmp->lock);
2082 static struct cpuset *cpuset_attach_old_cs;
2084 /* Called by cgroups to determine if a cpuset is usable; cpuset_mutex held */
2085 static int cpuset_can_attach(struct cgroup_taskset *tset)
2087 struct cgroup_subsys_state *css;
2089 struct task_struct *task;
2092 /* used later by cpuset_attach() */
2093 cpuset_attach_old_cs = task_cs(cgroup_taskset_first(tset, &css));
2096 mutex_lock(&cpuset_mutex);
2098 /* allow moving tasks into an empty cpuset if on default hierarchy */
2100 if (!is_in_v2_mode() &&
2101 (cpumask_empty(cs->cpus_allowed) || nodes_empty(cs->mems_allowed)))
2104 cgroup_taskset_for_each(task, css, tset) {
2105 ret = task_can_attach(task, cs->cpus_allowed);
2108 ret = security_task_setscheduler(task);
2114 * Mark attach is in progress. This makes validate_change() fail
2115 * changes which zero cpus/mems_allowed.
2117 cs->attach_in_progress++;
2120 mutex_unlock(&cpuset_mutex);
2124 static void cpuset_cancel_attach(struct cgroup_taskset *tset)
2126 struct cgroup_subsys_state *css;
2128 cgroup_taskset_first(tset, &css);
2130 mutex_lock(&cpuset_mutex);
2131 css_cs(css)->attach_in_progress--;
2132 mutex_unlock(&cpuset_mutex);
2136 * Protected by cpuset_mutex. cpus_attach is used only by cpuset_attach()
2137 * but we can't allocate it dynamically there. Define it global and
2138 * allocate from cpuset_init().
2140 static cpumask_var_t cpus_attach;
2142 static void cpuset_attach(struct cgroup_taskset *tset)
2144 /* static buf protected by cpuset_mutex */
2145 static nodemask_t cpuset_attach_nodemask_to;
2146 struct task_struct *task;
2147 struct task_struct *leader;
2148 struct cgroup_subsys_state *css;
2150 struct cpuset *oldcs = cpuset_attach_old_cs;
2152 cgroup_taskset_first(tset, &css);
2155 mutex_lock(&cpuset_mutex);
2157 /* prepare for attach */
2158 if (cs == &top_cpuset)
2159 cpumask_copy(cpus_attach, cpu_possible_mask);
2161 guarantee_online_cpus(cs, cpus_attach);
2163 guarantee_online_mems(cs, &cpuset_attach_nodemask_to);
2165 cgroup_taskset_for_each(task, css, tset) {
2167 * can_attach beforehand should guarantee that this doesn't
2168 * fail. TODO: have a better way to handle failure here
2170 WARN_ON_ONCE(set_cpus_allowed_ptr(task, cpus_attach));
2172 cpuset_change_task_nodemask(task, &cpuset_attach_nodemask_to);
2173 cpuset_update_task_spread_flag(cs, task);
2177 * Change mm for all threadgroup leaders. This is expensive and may
2178 * sleep and should be moved outside migration path proper.
2180 cpuset_attach_nodemask_to = cs->effective_mems;
2181 cgroup_taskset_for_each_leader(leader, css, tset) {
2182 struct mm_struct *mm = get_task_mm(leader);
2185 mpol_rebind_mm(mm, &cpuset_attach_nodemask_to);
2188 * old_mems_allowed is the same with mems_allowed
2189 * here, except if this task is being moved
2190 * automatically due to hotplug. In that case
2191 * @mems_allowed has been updated and is empty, so
2192 * @old_mems_allowed is the right nodesets that we
2195 if (is_memory_migrate(cs))
2196 cpuset_migrate_mm(mm, &oldcs->old_mems_allowed,
2197 &cpuset_attach_nodemask_to);
2203 cs->old_mems_allowed = cpuset_attach_nodemask_to;
2205 cs->attach_in_progress--;
2206 if (!cs->attach_in_progress)
2207 wake_up(&cpuset_attach_wq);
2209 mutex_unlock(&cpuset_mutex);
2212 /* The various types of files and directories in a cpuset file system */
2215 FILE_MEMORY_MIGRATE,
2218 FILE_EFFECTIVE_CPULIST,
2219 FILE_EFFECTIVE_MEMLIST,
2220 FILE_SUBPARTS_CPULIST,
2224 FILE_SCHED_LOAD_BALANCE,
2225 FILE_PARTITION_ROOT,
2226 FILE_SCHED_RELAX_DOMAIN_LEVEL,
2227 FILE_MEMORY_PRESSURE_ENABLED,
2228 FILE_MEMORY_PRESSURE,
2231 } cpuset_filetype_t;
2233 static int cpuset_write_u64(struct cgroup_subsys_state *css, struct cftype *cft,
2236 struct cpuset *cs = css_cs(css);
2237 cpuset_filetype_t type = cft->private;
2240 mutex_lock(&cpuset_mutex);
2241 if (!is_cpuset_online(cs)) {
2247 case FILE_CPU_EXCLUSIVE:
2248 retval = update_flag(CS_CPU_EXCLUSIVE, cs, val);
2250 case FILE_MEM_EXCLUSIVE:
2251 retval = update_flag(CS_MEM_EXCLUSIVE, cs, val);
2253 case FILE_MEM_HARDWALL:
2254 retval = update_flag(CS_MEM_HARDWALL, cs, val);
2256 case FILE_SCHED_LOAD_BALANCE:
2257 retval = update_flag(CS_SCHED_LOAD_BALANCE, cs, val);
2259 case FILE_MEMORY_MIGRATE:
2260 retval = update_flag(CS_MEMORY_MIGRATE, cs, val);
2262 case FILE_MEMORY_PRESSURE_ENABLED:
2263 cpuset_memory_pressure_enabled = !!val;
2265 case FILE_SPREAD_PAGE:
2266 retval = update_flag(CS_SPREAD_PAGE, cs, val);
2268 case FILE_SPREAD_SLAB:
2269 retval = update_flag(CS_SPREAD_SLAB, cs, val);
2276 mutex_unlock(&cpuset_mutex);
2280 static int cpuset_write_s64(struct cgroup_subsys_state *css, struct cftype *cft,
2283 struct cpuset *cs = css_cs(css);
2284 cpuset_filetype_t type = cft->private;
2285 int retval = -ENODEV;
2287 mutex_lock(&cpuset_mutex);
2288 if (!is_cpuset_online(cs))
2292 case FILE_SCHED_RELAX_DOMAIN_LEVEL:
2293 retval = update_relax_domain_level(cs, val);
2300 mutex_unlock(&cpuset_mutex);
2305 * Common handling for a write to a "cpus" or "mems" file.
2307 static ssize_t cpuset_write_resmask(struct kernfs_open_file *of,
2308 char *buf, size_t nbytes, loff_t off)
2310 struct cpuset *cs = css_cs(of_css(of));
2311 struct cpuset *trialcs;
2312 int retval = -ENODEV;
2314 buf = strstrip(buf);
2317 * CPU or memory hotunplug may leave @cs w/o any execution
2318 * resources, in which case the hotplug code asynchronously updates
2319 * configuration and transfers all tasks to the nearest ancestor
2320 * which can execute.
2322 * As writes to "cpus" or "mems" may restore @cs's execution
2323 * resources, wait for the previously scheduled operations before
2324 * proceeding, so that we don't end up keep removing tasks added
2325 * after execution capability is restored.
2327 * cpuset_hotplug_work calls back into cgroup core via
2328 * cgroup_transfer_tasks() and waiting for it from a cgroupfs
2329 * operation like this one can lead to a deadlock through kernfs
2330 * active_ref protection. Let's break the protection. Losing the
2331 * protection is okay as we check whether @cs is online after
2332 * grabbing cpuset_mutex anyway. This only happens on the legacy
2336 kernfs_break_active_protection(of->kn);
2337 flush_work(&cpuset_hotplug_work);
2339 mutex_lock(&cpuset_mutex);
2340 if (!is_cpuset_online(cs))
2343 trialcs = alloc_trial_cpuset(cs);
2349 switch (of_cft(of)->private) {
2351 retval = update_cpumask(cs, trialcs, buf);
2354 retval = update_nodemask(cs, trialcs, buf);
2361 free_cpuset(trialcs);
2363 mutex_unlock(&cpuset_mutex);
2364 kernfs_unbreak_active_protection(of->kn);
2366 flush_workqueue(cpuset_migrate_mm_wq);
2367 return retval ?: nbytes;
2371 * These ascii lists should be read in a single call, by using a user
2372 * buffer large enough to hold the entire map. If read in smaller
2373 * chunks, there is no guarantee of atomicity. Since the display format
2374 * used, list of ranges of sequential numbers, is variable length,
2375 * and since these maps can change value dynamically, one could read
2376 * gibberish by doing partial reads while a list was changing.
2378 static int cpuset_common_seq_show(struct seq_file *sf, void *v)
2380 struct cpuset *cs = css_cs(seq_css(sf));
2381 cpuset_filetype_t type = seq_cft(sf)->private;
2384 spin_lock_irq(&callback_lock);
2388 seq_printf(sf, "%*pbl\n", cpumask_pr_args(cs->cpus_allowed));
2391 seq_printf(sf, "%*pbl\n", nodemask_pr_args(&cs->mems_allowed));
2393 case FILE_EFFECTIVE_CPULIST:
2394 seq_printf(sf, "%*pbl\n", cpumask_pr_args(cs->effective_cpus));
2396 case FILE_EFFECTIVE_MEMLIST:
2397 seq_printf(sf, "%*pbl\n", nodemask_pr_args(&cs->effective_mems));
2399 case FILE_SUBPARTS_CPULIST:
2400 seq_printf(sf, "%*pbl\n", cpumask_pr_args(cs->subparts_cpus));
2406 spin_unlock_irq(&callback_lock);
2410 static u64 cpuset_read_u64(struct cgroup_subsys_state *css, struct cftype *cft)
2412 struct cpuset *cs = css_cs(css);
2413 cpuset_filetype_t type = cft->private;
2415 case FILE_CPU_EXCLUSIVE:
2416 return is_cpu_exclusive(cs);
2417 case FILE_MEM_EXCLUSIVE:
2418 return is_mem_exclusive(cs);
2419 case FILE_MEM_HARDWALL:
2420 return is_mem_hardwall(cs);
2421 case FILE_SCHED_LOAD_BALANCE:
2422 return is_sched_load_balance(cs);
2423 case FILE_MEMORY_MIGRATE:
2424 return is_memory_migrate(cs);
2425 case FILE_MEMORY_PRESSURE_ENABLED:
2426 return cpuset_memory_pressure_enabled;
2427 case FILE_MEMORY_PRESSURE:
2428 return fmeter_getrate(&cs->fmeter);
2429 case FILE_SPREAD_PAGE:
2430 return is_spread_page(cs);
2431 case FILE_SPREAD_SLAB:
2432 return is_spread_slab(cs);
2437 /* Unreachable but makes gcc happy */
2441 static s64 cpuset_read_s64(struct cgroup_subsys_state *css, struct cftype *cft)
2443 struct cpuset *cs = css_cs(css);
2444 cpuset_filetype_t type = cft->private;
2446 case FILE_SCHED_RELAX_DOMAIN_LEVEL:
2447 return cs->relax_domain_level;
2452 /* Unrechable but makes gcc happy */
2456 static int sched_partition_show(struct seq_file *seq, void *v)
2458 struct cpuset *cs = css_cs(seq_css(seq));
2460 switch (cs->partition_root_state) {
2462 seq_puts(seq, "root\n");
2465 seq_puts(seq, "member\n");
2468 seq_puts(seq, "root invalid\n");
2474 static ssize_t sched_partition_write(struct kernfs_open_file *of, char *buf,
2475 size_t nbytes, loff_t off)
2477 struct cpuset *cs = css_cs(of_css(of));
2479 int retval = -ENODEV;
2481 buf = strstrip(buf);
2484 * Convert "root" to ENABLED, and convert "member" to DISABLED.
2486 if (!strcmp(buf, "root"))
2488 else if (!strcmp(buf, "member"))
2494 mutex_lock(&cpuset_mutex);
2495 if (!is_cpuset_online(cs))
2498 retval = update_prstate(cs, val);
2500 mutex_unlock(&cpuset_mutex);
2502 return retval ?: nbytes;
2506 * for the common functions, 'private' gives the type of file
2509 static struct cftype legacy_files[] = {
2512 .seq_show = cpuset_common_seq_show,
2513 .write = cpuset_write_resmask,
2514 .max_write_len = (100U + 6 * NR_CPUS),
2515 .private = FILE_CPULIST,
2520 .seq_show = cpuset_common_seq_show,
2521 .write = cpuset_write_resmask,
2522 .max_write_len = (100U + 6 * MAX_NUMNODES),
2523 .private = FILE_MEMLIST,
2527 .name = "effective_cpus",
2528 .seq_show = cpuset_common_seq_show,
2529 .private = FILE_EFFECTIVE_CPULIST,
2533 .name = "effective_mems",
2534 .seq_show = cpuset_common_seq_show,
2535 .private = FILE_EFFECTIVE_MEMLIST,
2539 .name = "cpu_exclusive",
2540 .read_u64 = cpuset_read_u64,
2541 .write_u64 = cpuset_write_u64,
2542 .private = FILE_CPU_EXCLUSIVE,
2546 .name = "mem_exclusive",
2547 .read_u64 = cpuset_read_u64,
2548 .write_u64 = cpuset_write_u64,
2549 .private = FILE_MEM_EXCLUSIVE,
2553 .name = "mem_hardwall",
2554 .read_u64 = cpuset_read_u64,
2555 .write_u64 = cpuset_write_u64,
2556 .private = FILE_MEM_HARDWALL,
2560 .name = "sched_load_balance",
2561 .read_u64 = cpuset_read_u64,
2562 .write_u64 = cpuset_write_u64,
2563 .private = FILE_SCHED_LOAD_BALANCE,
2567 .name = "sched_relax_domain_level",
2568 .read_s64 = cpuset_read_s64,
2569 .write_s64 = cpuset_write_s64,
2570 .private = FILE_SCHED_RELAX_DOMAIN_LEVEL,
2574 .name = "memory_migrate",
2575 .read_u64 = cpuset_read_u64,
2576 .write_u64 = cpuset_write_u64,
2577 .private = FILE_MEMORY_MIGRATE,
2581 .name = "memory_pressure",
2582 .read_u64 = cpuset_read_u64,
2583 .private = FILE_MEMORY_PRESSURE,
2587 .name = "memory_spread_page",
2588 .read_u64 = cpuset_read_u64,
2589 .write_u64 = cpuset_write_u64,
2590 .private = FILE_SPREAD_PAGE,
2594 .name = "memory_spread_slab",
2595 .read_u64 = cpuset_read_u64,
2596 .write_u64 = cpuset_write_u64,
2597 .private = FILE_SPREAD_SLAB,
2601 .name = "memory_pressure_enabled",
2602 .flags = CFTYPE_ONLY_ON_ROOT,
2603 .read_u64 = cpuset_read_u64,
2604 .write_u64 = cpuset_write_u64,
2605 .private = FILE_MEMORY_PRESSURE_ENABLED,
2612 * This is currently a minimal set for the default hierarchy. It can be
2613 * expanded later on by migrating more features and control files from v1.
2615 static struct cftype dfl_files[] = {
2618 .seq_show = cpuset_common_seq_show,
2619 .write = cpuset_write_resmask,
2620 .max_write_len = (100U + 6 * NR_CPUS),
2621 .private = FILE_CPULIST,
2622 .flags = CFTYPE_NOT_ON_ROOT,
2627 .seq_show = cpuset_common_seq_show,
2628 .write = cpuset_write_resmask,
2629 .max_write_len = (100U + 6 * MAX_NUMNODES),
2630 .private = FILE_MEMLIST,
2631 .flags = CFTYPE_NOT_ON_ROOT,
2635 .name = "cpus.effective",
2636 .seq_show = cpuset_common_seq_show,
2637 .private = FILE_EFFECTIVE_CPULIST,
2641 .name = "mems.effective",
2642 .seq_show = cpuset_common_seq_show,
2643 .private = FILE_EFFECTIVE_MEMLIST,
2647 .name = "cpus.partition",
2648 .seq_show = sched_partition_show,
2649 .write = sched_partition_write,
2650 .private = FILE_PARTITION_ROOT,
2651 .flags = CFTYPE_NOT_ON_ROOT,
2655 .name = "cpus.subpartitions",
2656 .seq_show = cpuset_common_seq_show,
2657 .private = FILE_SUBPARTS_CPULIST,
2658 .flags = CFTYPE_DEBUG,
2666 * cpuset_css_alloc - allocate a cpuset css
2667 * cgrp: control group that the new cpuset will be part of
2670 static struct cgroup_subsys_state *
2671 cpuset_css_alloc(struct cgroup_subsys_state *parent_css)
2676 return &top_cpuset.css;
2678 cs = kzalloc(sizeof(*cs), GFP_KERNEL);
2680 return ERR_PTR(-ENOMEM);
2682 if (alloc_cpumasks(cs, NULL)) {
2684 return ERR_PTR(-ENOMEM);
2687 set_bit(CS_SCHED_LOAD_BALANCE, &cs->flags);
2688 nodes_clear(cs->mems_allowed);
2689 nodes_clear(cs->effective_mems);
2690 fmeter_init(&cs->fmeter);
2691 cs->relax_domain_level = -1;
2696 static int cpuset_css_online(struct cgroup_subsys_state *css)
2698 struct cpuset *cs = css_cs(css);
2699 struct cpuset *parent = parent_cs(cs);
2700 struct cpuset *tmp_cs;
2701 struct cgroup_subsys_state *pos_css;
2706 mutex_lock(&cpuset_mutex);
2708 set_bit(CS_ONLINE, &cs->flags);
2709 if (is_spread_page(parent))
2710 set_bit(CS_SPREAD_PAGE, &cs->flags);
2711 if (is_spread_slab(parent))
2712 set_bit(CS_SPREAD_SLAB, &cs->flags);
2716 spin_lock_irq(&callback_lock);
2717 if (is_in_v2_mode()) {
2718 cpumask_copy(cs->effective_cpus, parent->effective_cpus);
2719 cs->effective_mems = parent->effective_mems;
2720 cs->use_parent_ecpus = true;
2721 parent->child_ecpus_count++;
2723 spin_unlock_irq(&callback_lock);
2725 if (!test_bit(CGRP_CPUSET_CLONE_CHILDREN, &css->cgroup->flags))
2729 * Clone @parent's configuration if CGRP_CPUSET_CLONE_CHILDREN is
2730 * set. This flag handling is implemented in cgroup core for
2731 * histrical reasons - the flag may be specified during mount.
2733 * Currently, if any sibling cpusets have exclusive cpus or mem, we
2734 * refuse to clone the configuration - thereby refusing the task to
2735 * be entered, and as a result refusing the sys_unshare() or
2736 * clone() which initiated it. If this becomes a problem for some
2737 * users who wish to allow that scenario, then this could be
2738 * changed to grant parent->cpus_allowed-sibling_cpus_exclusive
2739 * (and likewise for mems) to the new cgroup.
2742 cpuset_for_each_child(tmp_cs, pos_css, parent) {
2743 if (is_mem_exclusive(tmp_cs) || is_cpu_exclusive(tmp_cs)) {
2750 spin_lock_irq(&callback_lock);
2751 cs->mems_allowed = parent->mems_allowed;
2752 cs->effective_mems = parent->mems_allowed;
2753 cpumask_copy(cs->cpus_allowed, parent->cpus_allowed);
2754 cpumask_copy(cs->effective_cpus, parent->cpus_allowed);
2755 spin_unlock_irq(&callback_lock);
2757 mutex_unlock(&cpuset_mutex);
2762 * If the cpuset being removed has its flag 'sched_load_balance'
2763 * enabled, then simulate turning sched_load_balance off, which
2764 * will call rebuild_sched_domains_locked(). That is not needed
2765 * in the default hierarchy where only changes in partition
2766 * will cause repartitioning.
2768 * If the cpuset has the 'sched.partition' flag enabled, simulate
2769 * turning 'sched.partition" off.
2772 static void cpuset_css_offline(struct cgroup_subsys_state *css)
2774 struct cpuset *cs = css_cs(css);
2776 mutex_lock(&cpuset_mutex);
2778 if (is_partition_root(cs))
2779 update_prstate(cs, 0);
2781 if (!cgroup_subsys_on_dfl(cpuset_cgrp_subsys) &&
2782 is_sched_load_balance(cs))
2783 update_flag(CS_SCHED_LOAD_BALANCE, cs, 0);
2785 if (cs->use_parent_ecpus) {
2786 struct cpuset *parent = parent_cs(cs);
2788 cs->use_parent_ecpus = false;
2789 parent->child_ecpus_count--;
2793 clear_bit(CS_ONLINE, &cs->flags);
2795 mutex_unlock(&cpuset_mutex);
2798 static void cpuset_css_free(struct cgroup_subsys_state *css)
2800 struct cpuset *cs = css_cs(css);
2805 static void cpuset_bind(struct cgroup_subsys_state *root_css)
2807 mutex_lock(&cpuset_mutex);
2808 spin_lock_irq(&callback_lock);
2810 if (is_in_v2_mode()) {
2811 cpumask_copy(top_cpuset.cpus_allowed, cpu_possible_mask);
2812 top_cpuset.mems_allowed = node_possible_map;
2814 cpumask_copy(top_cpuset.cpus_allowed,
2815 top_cpuset.effective_cpus);
2816 top_cpuset.mems_allowed = top_cpuset.effective_mems;
2819 spin_unlock_irq(&callback_lock);
2820 mutex_unlock(&cpuset_mutex);
2824 * Make sure the new task conform to the current state of its parent,
2825 * which could have been changed by cpuset just after it inherits the
2826 * state from the parent and before it sits on the cgroup's task list.
2828 static void cpuset_fork(struct task_struct *task)
2830 if (task_css_is_root(task, cpuset_cgrp_id))
2833 set_cpus_allowed_ptr(task, ¤t->cpus_allowed);
2834 task->mems_allowed = current->mems_allowed;
2837 struct cgroup_subsys cpuset_cgrp_subsys = {
2838 .css_alloc = cpuset_css_alloc,
2839 .css_online = cpuset_css_online,
2840 .css_offline = cpuset_css_offline,
2841 .css_free = cpuset_css_free,
2842 .can_attach = cpuset_can_attach,
2843 .cancel_attach = cpuset_cancel_attach,
2844 .attach = cpuset_attach,
2845 .post_attach = cpuset_post_attach,
2846 .bind = cpuset_bind,
2847 .fork = cpuset_fork,
2848 .legacy_cftypes = legacy_files,
2849 .dfl_cftypes = dfl_files,
2855 * cpuset_init - initialize cpusets at system boot
2857 * Description: Initialize top_cpuset and the cpuset internal file system,
2860 int __init cpuset_init(void)
2864 BUG_ON(!alloc_cpumask_var(&top_cpuset.cpus_allowed, GFP_KERNEL));
2865 BUG_ON(!alloc_cpumask_var(&top_cpuset.effective_cpus, GFP_KERNEL));
2866 BUG_ON(!zalloc_cpumask_var(&top_cpuset.subparts_cpus, GFP_KERNEL));
2868 cpumask_setall(top_cpuset.cpus_allowed);
2869 nodes_setall(top_cpuset.mems_allowed);
2870 cpumask_setall(top_cpuset.effective_cpus);
2871 nodes_setall(top_cpuset.effective_mems);
2873 fmeter_init(&top_cpuset.fmeter);
2874 set_bit(CS_SCHED_LOAD_BALANCE, &top_cpuset.flags);
2875 top_cpuset.relax_domain_level = -1;
2877 err = register_filesystem(&cpuset_fs_type);
2881 BUG_ON(!alloc_cpumask_var(&cpus_attach, GFP_KERNEL));
2887 * If CPU and/or memory hotplug handlers, below, unplug any CPUs
2888 * or memory nodes, we need to walk over the cpuset hierarchy,
2889 * removing that CPU or node from all cpusets. If this removes the
2890 * last CPU or node from a cpuset, then move the tasks in the empty
2891 * cpuset to its next-highest non-empty parent.
2893 static void remove_tasks_in_empty_cpuset(struct cpuset *cs)
2895 struct cpuset *parent;
2898 * Find its next-highest non-empty parent, (top cpuset
2899 * has online cpus, so can't be empty).
2901 parent = parent_cs(cs);
2902 while (cpumask_empty(parent->cpus_allowed) ||
2903 nodes_empty(parent->mems_allowed))
2904 parent = parent_cs(parent);
2906 if (cgroup_transfer_tasks(parent->css.cgroup, cs->css.cgroup)) {
2907 pr_err("cpuset: failed to transfer tasks out of empty cpuset ");
2908 pr_cont_cgroup_name(cs->css.cgroup);
2914 hotplug_update_tasks_legacy(struct cpuset *cs,
2915 struct cpumask *new_cpus, nodemask_t *new_mems,
2916 bool cpus_updated, bool mems_updated)
2920 spin_lock_irq(&callback_lock);
2921 cpumask_copy(cs->cpus_allowed, new_cpus);
2922 cpumask_copy(cs->effective_cpus, new_cpus);
2923 cs->mems_allowed = *new_mems;
2924 cs->effective_mems = *new_mems;
2925 spin_unlock_irq(&callback_lock);
2928 * Don't call update_tasks_cpumask() if the cpuset becomes empty,
2929 * as the tasks will be migratecd to an ancestor.
2931 if (cpus_updated && !cpumask_empty(cs->cpus_allowed))
2932 update_tasks_cpumask(cs);
2933 if (mems_updated && !nodes_empty(cs->mems_allowed))
2934 update_tasks_nodemask(cs);
2936 is_empty = cpumask_empty(cs->cpus_allowed) ||
2937 nodes_empty(cs->mems_allowed);
2939 mutex_unlock(&cpuset_mutex);
2942 * Move tasks to the nearest ancestor with execution resources,
2943 * This is full cgroup operation which will also call back into
2944 * cpuset. Should be done outside any lock.
2947 remove_tasks_in_empty_cpuset(cs);
2949 mutex_lock(&cpuset_mutex);
2953 hotplug_update_tasks(struct cpuset *cs,
2954 struct cpumask *new_cpus, nodemask_t *new_mems,
2955 bool cpus_updated, bool mems_updated)
2957 if (cpumask_empty(new_cpus))
2958 cpumask_copy(new_cpus, parent_cs(cs)->effective_cpus);
2959 if (nodes_empty(*new_mems))
2960 *new_mems = parent_cs(cs)->effective_mems;
2962 spin_lock_irq(&callback_lock);
2963 cpumask_copy(cs->effective_cpus, new_cpus);
2964 cs->effective_mems = *new_mems;
2965 spin_unlock_irq(&callback_lock);
2968 update_tasks_cpumask(cs);
2970 update_tasks_nodemask(cs);
2973 static bool force_rebuild;
2975 void cpuset_force_rebuild(void)
2977 force_rebuild = true;
2981 * cpuset_hotplug_update_tasks - update tasks in a cpuset for hotunplug
2982 * @cs: cpuset in interest
2983 * @tmp: the tmpmasks structure pointer
2985 * Compare @cs's cpu and mem masks against top_cpuset and if some have gone
2986 * offline, update @cs accordingly. If @cs ends up with no CPU or memory,
2987 * all its tasks are moved to the nearest ancestor with both resources.
2989 static void cpuset_hotplug_update_tasks(struct cpuset *cs, struct tmpmasks *tmp)
2991 static cpumask_t new_cpus;
2992 static nodemask_t new_mems;
2995 struct cpuset *parent;
2997 wait_event(cpuset_attach_wq, cs->attach_in_progress == 0);
2999 mutex_lock(&cpuset_mutex);
3002 * We have raced with task attaching. We wait until attaching
3003 * is finished, so we won't attach a task to an empty cpuset.
3005 if (cs->attach_in_progress) {
3006 mutex_unlock(&cpuset_mutex);
3010 parent = parent_cs(cs);
3011 compute_effective_cpumask(&new_cpus, cs, parent);
3012 nodes_and(new_mems, cs->mems_allowed, parent->effective_mems);
3014 if (cs->nr_subparts_cpus)
3016 * Make sure that CPUs allocated to child partitions
3017 * do not show up in effective_cpus.
3019 cpumask_andnot(&new_cpus, &new_cpus, cs->subparts_cpus);
3021 if (!tmp || !cs->partition_root_state)
3025 * In the unlikely event that a partition root has empty
3026 * effective_cpus or its parent becomes erroneous, we have to
3027 * transition it to the erroneous state.
3029 if (is_partition_root(cs) && (cpumask_empty(&new_cpus) ||
3030 (parent->partition_root_state == PRS_ERROR))) {
3031 if (cs->nr_subparts_cpus) {
3032 cs->nr_subparts_cpus = 0;
3033 cpumask_clear(cs->subparts_cpus);
3034 compute_effective_cpumask(&new_cpus, cs, parent);
3038 * If the effective_cpus is empty because the child
3039 * partitions take away all the CPUs, we can keep
3040 * the current partition and let the child partitions
3041 * fight for available CPUs.
3043 if ((parent->partition_root_state == PRS_ERROR) ||
3044 cpumask_empty(&new_cpus)) {
3045 update_parent_subparts_cpumask(cs, partcmd_disable,
3047 cs->partition_root_state = PRS_ERROR;
3049 cpuset_force_rebuild();
3053 * On the other hand, an erroneous partition root may be transitioned
3054 * back to a regular one or a partition root with no CPU allocated
3055 * from the parent may change to erroneous.
3057 if (is_partition_root(parent) &&
3058 ((cs->partition_root_state == PRS_ERROR) ||
3059 !cpumask_intersects(&new_cpus, parent->subparts_cpus)) &&
3060 update_parent_subparts_cpumask(cs, partcmd_update, NULL, tmp))
3061 cpuset_force_rebuild();
3064 cpus_updated = !cpumask_equal(&new_cpus, cs->effective_cpus);
3065 mems_updated = !nodes_equal(new_mems, cs->effective_mems);
3067 if (is_in_v2_mode())
3068 hotplug_update_tasks(cs, &new_cpus, &new_mems,
3069 cpus_updated, mems_updated);
3071 hotplug_update_tasks_legacy(cs, &new_cpus, &new_mems,
3072 cpus_updated, mems_updated);
3074 mutex_unlock(&cpuset_mutex);
3078 * cpuset_hotplug_workfn - handle CPU/memory hotunplug for a cpuset
3080 * This function is called after either CPU or memory configuration has
3081 * changed and updates cpuset accordingly. The top_cpuset is always
3082 * synchronized to cpu_active_mask and N_MEMORY, which is necessary in
3083 * order to make cpusets transparent (of no affect) on systems that are
3084 * actively using CPU hotplug but making no active use of cpusets.
3086 * Non-root cpusets are only affected by offlining. If any CPUs or memory
3087 * nodes have been taken down, cpuset_hotplug_update_tasks() is invoked on
3090 * Note that CPU offlining during suspend is ignored. We don't modify
3091 * cpusets across suspend/resume cycles at all.
3093 static void cpuset_hotplug_workfn(struct work_struct *work)
3095 static cpumask_t new_cpus;
3096 static nodemask_t new_mems;
3097 bool cpus_updated, mems_updated;
3098 bool on_dfl = is_in_v2_mode();
3099 struct tmpmasks tmp, *ptmp = NULL;
3101 if (on_dfl && !alloc_cpumasks(NULL, &tmp))
3104 mutex_lock(&cpuset_mutex);
3106 /* fetch the available cpus/mems and find out which changed how */
3107 cpumask_copy(&new_cpus, cpu_active_mask);
3108 new_mems = node_states[N_MEMORY];
3111 * If subparts_cpus is populated, it is likely that the check below
3112 * will produce a false positive on cpus_updated when the cpu list
3113 * isn't changed. It is extra work, but it is better to be safe.
3115 cpus_updated = !cpumask_equal(top_cpuset.effective_cpus, &new_cpus);
3116 mems_updated = !nodes_equal(top_cpuset.effective_mems, new_mems);
3118 /* synchronize cpus_allowed to cpu_active_mask */
3120 spin_lock_irq(&callback_lock);
3122 cpumask_copy(top_cpuset.cpus_allowed, &new_cpus);
3124 * Make sure that CPUs allocated to child partitions
3125 * do not show up in effective_cpus. If no CPU is left,
3126 * we clear the subparts_cpus & let the child partitions
3127 * fight for the CPUs again.
3129 if (top_cpuset.nr_subparts_cpus) {
3130 if (cpumask_subset(&new_cpus,
3131 top_cpuset.subparts_cpus)) {
3132 top_cpuset.nr_subparts_cpus = 0;
3133 cpumask_clear(top_cpuset.subparts_cpus);
3135 cpumask_andnot(&new_cpus, &new_cpus,
3136 top_cpuset.subparts_cpus);
3139 cpumask_copy(top_cpuset.effective_cpus, &new_cpus);
3140 spin_unlock_irq(&callback_lock);
3141 /* we don't mess with cpumasks of tasks in top_cpuset */
3144 /* synchronize mems_allowed to N_MEMORY */
3146 spin_lock_irq(&callback_lock);
3148 top_cpuset.mems_allowed = new_mems;
3149 top_cpuset.effective_mems = new_mems;
3150 spin_unlock_irq(&callback_lock);
3151 update_tasks_nodemask(&top_cpuset);
3154 mutex_unlock(&cpuset_mutex);
3156 /* if cpus or mems changed, we need to propagate to descendants */
3157 if (cpus_updated || mems_updated) {
3159 struct cgroup_subsys_state *pos_css;
3162 cpuset_for_each_descendant_pre(cs, pos_css, &top_cpuset) {
3163 if (cs == &top_cpuset || !css_tryget_online(&cs->css))
3167 cpuset_hotplug_update_tasks(cs, ptmp);
3175 /* rebuild sched domains if cpus_allowed has changed */
3176 if (cpus_updated || force_rebuild) {
3177 force_rebuild = false;
3178 rebuild_sched_domains();
3181 free_cpumasks(NULL, ptmp);
3184 void cpuset_update_active_cpus(void)
3187 * We're inside cpu hotplug critical region which usually nests
3188 * inside cgroup synchronization. Bounce actual hotplug processing
3189 * to a work item to avoid reverse locking order.
3191 schedule_work(&cpuset_hotplug_work);
3194 void cpuset_wait_for_hotplug(void)
3196 flush_work(&cpuset_hotplug_work);
3200 * Keep top_cpuset.mems_allowed tracking node_states[N_MEMORY].
3201 * Call this routine anytime after node_states[N_MEMORY] changes.
3202 * See cpuset_update_active_cpus() for CPU hotplug handling.
3204 static int cpuset_track_online_nodes(struct notifier_block *self,
3205 unsigned long action, void *arg)
3207 schedule_work(&cpuset_hotplug_work);
3211 static struct notifier_block cpuset_track_online_nodes_nb = {
3212 .notifier_call = cpuset_track_online_nodes,
3213 .priority = 10, /* ??! */
3217 * cpuset_init_smp - initialize cpus_allowed
3219 * Description: Finish top cpuset after cpu, node maps are initialized
3221 void __init cpuset_init_smp(void)
3223 cpumask_copy(top_cpuset.cpus_allowed, cpu_active_mask);
3224 top_cpuset.mems_allowed = node_states[N_MEMORY];
3225 top_cpuset.old_mems_allowed = top_cpuset.mems_allowed;
3227 cpumask_copy(top_cpuset.effective_cpus, cpu_active_mask);
3228 top_cpuset.effective_mems = node_states[N_MEMORY];
3230 register_hotmemory_notifier(&cpuset_track_online_nodes_nb);
3232 cpuset_migrate_mm_wq = alloc_ordered_workqueue("cpuset_migrate_mm", 0);
3233 BUG_ON(!cpuset_migrate_mm_wq);
3237 * cpuset_cpus_allowed - return cpus_allowed mask from a tasks cpuset.
3238 * @tsk: pointer to task_struct from which to obtain cpuset->cpus_allowed.
3239 * @pmask: pointer to struct cpumask variable to receive cpus_allowed set.
3241 * Description: Returns the cpumask_var_t cpus_allowed of the cpuset
3242 * attached to the specified @tsk. Guaranteed to return some non-empty
3243 * subset of cpu_online_mask, even if this means going outside the
3247 void cpuset_cpus_allowed(struct task_struct *tsk, struct cpumask *pmask)
3249 unsigned long flags;
3251 spin_lock_irqsave(&callback_lock, flags);
3253 guarantee_online_cpus(task_cs(tsk), pmask);
3255 spin_unlock_irqrestore(&callback_lock, flags);
3258 void cpuset_cpus_allowed_fallback(struct task_struct *tsk)
3261 do_set_cpus_allowed(tsk, task_cs(tsk)->effective_cpus);
3265 * We own tsk->cpus_allowed, nobody can change it under us.
3267 * But we used cs && cs->cpus_allowed lockless and thus can
3268 * race with cgroup_attach_task() or update_cpumask() and get
3269 * the wrong tsk->cpus_allowed. However, both cases imply the
3270 * subsequent cpuset_change_cpumask()->set_cpus_allowed_ptr()
3271 * which takes task_rq_lock().
3273 * If we are called after it dropped the lock we must see all
3274 * changes in tsk_cs()->cpus_allowed. Otherwise we can temporary
3275 * set any mask even if it is not right from task_cs() pov,
3276 * the pending set_cpus_allowed_ptr() will fix things.
3278 * select_fallback_rq() will fix things ups and set cpu_possible_mask
3283 void __init cpuset_init_current_mems_allowed(void)
3285 nodes_setall(current->mems_allowed);
3289 * cpuset_mems_allowed - return mems_allowed mask from a tasks cpuset.
3290 * @tsk: pointer to task_struct from which to obtain cpuset->mems_allowed.
3292 * Description: Returns the nodemask_t mems_allowed of the cpuset
3293 * attached to the specified @tsk. Guaranteed to return some non-empty
3294 * subset of node_states[N_MEMORY], even if this means going outside the
3298 nodemask_t cpuset_mems_allowed(struct task_struct *tsk)
3301 unsigned long flags;
3303 spin_lock_irqsave(&callback_lock, flags);
3305 guarantee_online_mems(task_cs(tsk), &mask);
3307 spin_unlock_irqrestore(&callback_lock, flags);
3313 * cpuset_nodemask_valid_mems_allowed - check nodemask vs. curremt mems_allowed
3314 * @nodemask: the nodemask to be checked
3316 * Are any of the nodes in the nodemask allowed in current->mems_allowed?
3318 int cpuset_nodemask_valid_mems_allowed(nodemask_t *nodemask)
3320 return nodes_intersects(*nodemask, current->mems_allowed);
3324 * nearest_hardwall_ancestor() - Returns the nearest mem_exclusive or
3325 * mem_hardwall ancestor to the specified cpuset. Call holding
3326 * callback_lock. If no ancestor is mem_exclusive or mem_hardwall
3327 * (an unusual configuration), then returns the root cpuset.
3329 static struct cpuset *nearest_hardwall_ancestor(struct cpuset *cs)
3331 while (!(is_mem_exclusive(cs) || is_mem_hardwall(cs)) && parent_cs(cs))
3337 * cpuset_node_allowed - Can we allocate on a memory node?
3338 * @node: is this an allowed node?
3339 * @gfp_mask: memory allocation flags
3341 * If we're in interrupt, yes, we can always allocate. If @node is set in
3342 * current's mems_allowed, yes. If it's not a __GFP_HARDWALL request and this
3343 * node is set in the nearest hardwalled cpuset ancestor to current's cpuset,
3344 * yes. If current has access to memory reserves as an oom victim, yes.
3347 * GFP_USER allocations are marked with the __GFP_HARDWALL bit,
3348 * and do not allow allocations outside the current tasks cpuset
3349 * unless the task has been OOM killed.
3350 * GFP_KERNEL allocations are not so marked, so can escape to the
3351 * nearest enclosing hardwalled ancestor cpuset.
3353 * Scanning up parent cpusets requires callback_lock. The
3354 * __alloc_pages() routine only calls here with __GFP_HARDWALL bit
3355 * _not_ set if it's a GFP_KERNEL allocation, and all nodes in the
3356 * current tasks mems_allowed came up empty on the first pass over
3357 * the zonelist. So only GFP_KERNEL allocations, if all nodes in the
3358 * cpuset are short of memory, might require taking the callback_lock.
3360 * The first call here from mm/page_alloc:get_page_from_freelist()
3361 * has __GFP_HARDWALL set in gfp_mask, enforcing hardwall cpusets,
3362 * so no allocation on a node outside the cpuset is allowed (unless
3363 * in interrupt, of course).
3365 * The second pass through get_page_from_freelist() doesn't even call
3366 * here for GFP_ATOMIC calls. For those calls, the __alloc_pages()
3367 * variable 'wait' is not set, and the bit ALLOC_CPUSET is not set
3368 * in alloc_flags. That logic and the checks below have the combined
3370 * in_interrupt - any node ok (current task context irrelevant)
3371 * GFP_ATOMIC - any node ok
3372 * tsk_is_oom_victim - any node ok
3373 * GFP_KERNEL - any node in enclosing hardwalled cpuset ok
3374 * GFP_USER - only nodes in current tasks mems allowed ok.
3376 bool __cpuset_node_allowed(int node, gfp_t gfp_mask)
3378 struct cpuset *cs; /* current cpuset ancestors */
3379 int allowed; /* is allocation in zone z allowed? */
3380 unsigned long flags;
3384 if (node_isset(node, current->mems_allowed))
3387 * Allow tasks that have access to memory reserves because they have
3388 * been OOM killed to get memory anywhere.
3390 if (unlikely(tsk_is_oom_victim(current)))
3392 if (gfp_mask & __GFP_HARDWALL) /* If hardwall request, stop here */
3395 if (current->flags & PF_EXITING) /* Let dying task have memory */
3398 /* Not hardwall and node outside mems_allowed: scan up cpusets */
3399 spin_lock_irqsave(&callback_lock, flags);
3402 cs = nearest_hardwall_ancestor(task_cs(current));
3403 allowed = node_isset(node, cs->mems_allowed);
3406 spin_unlock_irqrestore(&callback_lock, flags);
3411 * cpuset_mem_spread_node() - On which node to begin search for a file page
3412 * cpuset_slab_spread_node() - On which node to begin search for a slab page
3414 * If a task is marked PF_SPREAD_PAGE or PF_SPREAD_SLAB (as for
3415 * tasks in a cpuset with is_spread_page or is_spread_slab set),
3416 * and if the memory allocation used cpuset_mem_spread_node()
3417 * to determine on which node to start looking, as it will for
3418 * certain page cache or slab cache pages such as used for file
3419 * system buffers and inode caches, then instead of starting on the
3420 * local node to look for a free page, rather spread the starting
3421 * node around the tasks mems_allowed nodes.
3423 * We don't have to worry about the returned node being offline
3424 * because "it can't happen", and even if it did, it would be ok.
3426 * The routines calling guarantee_online_mems() are careful to
3427 * only set nodes in task->mems_allowed that are online. So it
3428 * should not be possible for the following code to return an
3429 * offline node. But if it did, that would be ok, as this routine
3430 * is not returning the node where the allocation must be, only
3431 * the node where the search should start. The zonelist passed to
3432 * __alloc_pages() will include all nodes. If the slab allocator
3433 * is passed an offline node, it will fall back to the local node.
3434 * See kmem_cache_alloc_node().
3437 static int cpuset_spread_node(int *rotor)
3439 return *rotor = next_node_in(*rotor, current->mems_allowed);
3442 int cpuset_mem_spread_node(void)
3444 if (current->cpuset_mem_spread_rotor == NUMA_NO_NODE)
3445 current->cpuset_mem_spread_rotor =
3446 node_random(¤t->mems_allowed);
3448 return cpuset_spread_node(¤t->cpuset_mem_spread_rotor);
3451 int cpuset_slab_spread_node(void)
3453 if (current->cpuset_slab_spread_rotor == NUMA_NO_NODE)
3454 current->cpuset_slab_spread_rotor =
3455 node_random(¤t->mems_allowed);
3457 return cpuset_spread_node(¤t->cpuset_slab_spread_rotor);
3460 EXPORT_SYMBOL_GPL(cpuset_mem_spread_node);
3463 * cpuset_mems_allowed_intersects - Does @tsk1's mems_allowed intersect @tsk2's?
3464 * @tsk1: pointer to task_struct of some task.
3465 * @tsk2: pointer to task_struct of some other task.
3467 * Description: Return true if @tsk1's mems_allowed intersects the
3468 * mems_allowed of @tsk2. Used by the OOM killer to determine if
3469 * one of the task's memory usage might impact the memory available
3473 int cpuset_mems_allowed_intersects(const struct task_struct *tsk1,
3474 const struct task_struct *tsk2)
3476 return nodes_intersects(tsk1->mems_allowed, tsk2->mems_allowed);
3480 * cpuset_print_current_mems_allowed - prints current's cpuset and mems_allowed
3482 * Description: Prints current's name, cpuset name, and cached copy of its
3483 * mems_allowed to the kernel log.
3485 void cpuset_print_current_mems_allowed(void)
3487 struct cgroup *cgrp;
3491 cgrp = task_cs(current)->css.cgroup;
3492 pr_cont(",cpuset=");
3493 pr_cont_cgroup_name(cgrp);
3494 pr_cont(",mems_allowed=%*pbl",
3495 nodemask_pr_args(¤t->mems_allowed));
3501 * Collection of memory_pressure is suppressed unless
3502 * this flag is enabled by writing "1" to the special
3503 * cpuset file 'memory_pressure_enabled' in the root cpuset.
3506 int cpuset_memory_pressure_enabled __read_mostly;
3509 * cpuset_memory_pressure_bump - keep stats of per-cpuset reclaims.
3511 * Keep a running average of the rate of synchronous (direct)
3512 * page reclaim efforts initiated by tasks in each cpuset.
3514 * This represents the rate at which some task in the cpuset
3515 * ran low on memory on all nodes it was allowed to use, and
3516 * had to enter the kernels page reclaim code in an effort to
3517 * create more free memory by tossing clean pages or swapping
3518 * or writing dirty pages.
3520 * Display to user space in the per-cpuset read-only file
3521 * "memory_pressure". Value displayed is an integer
3522 * representing the recent rate of entry into the synchronous
3523 * (direct) page reclaim by any task attached to the cpuset.
3526 void __cpuset_memory_pressure_bump(void)
3529 fmeter_markevent(&task_cs(current)->fmeter);
3533 #ifdef CONFIG_PROC_PID_CPUSET
3535 * proc_cpuset_show()
3536 * - Print tasks cpuset path into seq_file.
3537 * - Used for /proc/<pid>/cpuset.
3538 * - No need to task_lock(tsk) on this tsk->cpuset reference, as it
3539 * doesn't really matter if tsk->cpuset changes after we read it,
3540 * and we take cpuset_mutex, keeping cpuset_attach() from changing it
3543 int proc_cpuset_show(struct seq_file *m, struct pid_namespace *ns,
3544 struct pid *pid, struct task_struct *tsk)
3547 struct cgroup_subsys_state *css;
3551 buf = kmalloc(PATH_MAX, GFP_KERNEL);
3555 css = task_get_css(tsk, cpuset_cgrp_id);
3556 retval = cgroup_path_ns(css->cgroup, buf, PATH_MAX,
3557 current->nsproxy->cgroup_ns);
3559 if (retval >= PATH_MAX)
3560 retval = -ENAMETOOLONG;
3571 #endif /* CONFIG_PROC_PID_CPUSET */
3573 /* Display task mems_allowed in /proc/<pid>/status file. */
3574 void cpuset_task_status_allowed(struct seq_file *m, struct task_struct *task)
3576 seq_printf(m, "Mems_allowed:\t%*pb\n",
3577 nodemask_pr_args(&task->mems_allowed));
3578 seq_printf(m, "Mems_allowed_list:\t%*pbl\n",
3579 nodemask_pr_args(&task->mems_allowed));