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/kthread.h>
37 #include <linux/list.h>
38 #include <linux/mempolicy.h>
40 #include <linux/memory.h>
41 #include <linux/export.h>
42 #include <linux/mount.h>
43 #include <linux/fs_context.h>
44 #include <linux/namei.h>
45 #include <linux/pagemap.h>
46 #include <linux/proc_fs.h>
47 #include <linux/rcupdate.h>
48 #include <linux/sched.h>
49 #include <linux/sched/deadline.h>
50 #include <linux/sched/mm.h>
51 #include <linux/sched/task.h>
52 #include <linux/seq_file.h>
53 #include <linux/security.h>
54 #include <linux/slab.h>
55 #include <linux/spinlock.h>
56 #include <linux/stat.h>
57 #include <linux/string.h>
58 #include <linux/time.h>
59 #include <linux/time64.h>
60 #include <linux/backing-dev.h>
61 #include <linux/sort.h>
62 #include <linux/oom.h>
63 #include <linux/sched/isolation.h>
64 #include <linux/uaccess.h>
65 #include <linux/atomic.h>
66 #include <linux/mutex.h>
67 #include <linux/cgroup.h>
68 #include <linux/wait.h>
70 DEFINE_STATIC_KEY_FALSE(cpusets_pre_enable_key);
71 DEFINE_STATIC_KEY_FALSE(cpusets_enabled_key);
73 /* See "Frequency meter" comments, below. */
76 int cnt; /* unprocessed events count */
77 int val; /* most recent output value */
78 time64_t time; /* clock (secs) when val computed */
79 spinlock_t lock; /* guards read or write of above */
83 struct cgroup_subsys_state css;
85 unsigned long flags; /* "unsigned long" so bitops work */
88 * On default hierarchy:
90 * The user-configured masks can only be changed by writing to
91 * cpuset.cpus and cpuset.mems, and won't be limited by the
94 * The effective masks is the real masks that apply to the tasks
95 * in the cpuset. They may be changed if the configured masks are
96 * changed or hotplug happens.
98 * effective_mask == configured_mask & parent's effective_mask,
99 * and if it ends up empty, it will inherit the parent's mask.
102 * On legacy hierachy:
104 * The user-configured masks are always the same with effective masks.
107 /* user-configured CPUs and Memory Nodes allow to tasks */
108 cpumask_var_t cpus_allowed;
109 nodemask_t mems_allowed;
111 /* effective CPUs and Memory Nodes allow to tasks */
112 cpumask_var_t effective_cpus;
113 nodemask_t effective_mems;
116 * CPUs allocated to child sub-partitions (default hierarchy only)
117 * - CPUs granted by the parent = effective_cpus U subparts_cpus
118 * - effective_cpus and subparts_cpus are mutually exclusive.
120 * effective_cpus contains only onlined CPUs, but subparts_cpus
121 * may have offlined ones.
123 cpumask_var_t subparts_cpus;
126 * This is old Memory Nodes tasks took on.
128 * - top_cpuset.old_mems_allowed is initialized to mems_allowed.
129 * - A new cpuset's old_mems_allowed is initialized when some
130 * task is moved into it.
131 * - old_mems_allowed is used in cpuset_migrate_mm() when we change
132 * cpuset.mems_allowed and have tasks' nodemask updated, and
133 * then old_mems_allowed is updated to mems_allowed.
135 nodemask_t old_mems_allowed;
137 struct fmeter fmeter; /* memory_pressure filter */
140 * Tasks are being attached to this cpuset. Used to prevent
141 * zeroing cpus/mems_allowed between ->can_attach() and ->attach().
143 int attach_in_progress;
145 /* partition number for rebuild_sched_domains() */
148 /* for custom sched domain */
149 int relax_domain_level;
151 /* number of CPUs in subparts_cpus */
152 int nr_subparts_cpus;
154 /* partition root state */
155 int partition_root_state;
158 * Default hierarchy only:
159 * use_parent_ecpus - set if using parent's effective_cpus
160 * child_ecpus_count - # of children with use_parent_ecpus set
162 int use_parent_ecpus;
163 int child_ecpus_count;
166 * number of SCHED_DEADLINE tasks attached to this cpuset, so that we
167 * know when to rebuild associated root domain bandwidth information.
169 int nr_deadline_tasks;
170 int nr_migrate_dl_tasks;
171 u64 sum_migrate_dl_bw;
175 * Partition root states:
177 * 0 - not a partition root
181 * -1 - invalid partition root
182 * None of the cpus in cpus_allowed can be put into the parent's
183 * subparts_cpus. In this case, the cpuset is not a real partition
184 * root anymore. However, the CPU_EXCLUSIVE bit will still be set
185 * and the cpuset can be restored back to a partition root if the
186 * parent cpuset can give more CPUs back to this child cpuset.
188 #define PRS_DISABLED 0
189 #define PRS_ENABLED 1
193 * Temporary cpumasks for working with partitions that are passed among
194 * functions to avoid memory allocation in inner functions.
197 cpumask_var_t addmask, delmask; /* For partition root */
198 cpumask_var_t new_cpus; /* For update_cpumasks_hier() */
201 static inline struct cpuset *css_cs(struct cgroup_subsys_state *css)
203 return css ? container_of(css, struct cpuset, css) : NULL;
206 /* Retrieve the cpuset for a task */
207 static inline struct cpuset *task_cs(struct task_struct *task)
209 return css_cs(task_css(task, cpuset_cgrp_id));
212 static inline struct cpuset *parent_cs(struct cpuset *cs)
214 return css_cs(cs->css.parent);
217 void inc_dl_tasks_cs(struct task_struct *p)
219 struct cpuset *cs = task_cs(p);
221 cs->nr_deadline_tasks++;
224 void dec_dl_tasks_cs(struct task_struct *p)
226 struct cpuset *cs = task_cs(p);
228 cs->nr_deadline_tasks--;
231 /* bits in struct cpuset flags field */
238 CS_SCHED_LOAD_BALANCE,
243 /* convenient tests for these bits */
244 static inline bool is_cpuset_online(struct cpuset *cs)
246 return test_bit(CS_ONLINE, &cs->flags) && !css_is_dying(&cs->css);
249 static inline int is_cpu_exclusive(const struct cpuset *cs)
251 return test_bit(CS_CPU_EXCLUSIVE, &cs->flags);
254 static inline int is_mem_exclusive(const struct cpuset *cs)
256 return test_bit(CS_MEM_EXCLUSIVE, &cs->flags);
259 static inline int is_mem_hardwall(const struct cpuset *cs)
261 return test_bit(CS_MEM_HARDWALL, &cs->flags);
264 static inline int is_sched_load_balance(const struct cpuset *cs)
266 return test_bit(CS_SCHED_LOAD_BALANCE, &cs->flags);
269 static inline int is_memory_migrate(const struct cpuset *cs)
271 return test_bit(CS_MEMORY_MIGRATE, &cs->flags);
274 static inline int is_spread_page(const struct cpuset *cs)
276 return test_bit(CS_SPREAD_PAGE, &cs->flags);
279 static inline int is_spread_slab(const struct cpuset *cs)
281 return test_bit(CS_SPREAD_SLAB, &cs->flags);
284 static inline int is_partition_root(const struct cpuset *cs)
286 return cs->partition_root_state > 0;
289 static struct cpuset top_cpuset = {
290 .flags = ((1 << CS_ONLINE) | (1 << CS_CPU_EXCLUSIVE) |
291 (1 << CS_MEM_EXCLUSIVE)),
292 .partition_root_state = PRS_ENABLED,
296 * cpuset_for_each_child - traverse online children of a cpuset
297 * @child_cs: loop cursor pointing to the current child
298 * @pos_css: used for iteration
299 * @parent_cs: target cpuset to walk children of
301 * Walk @child_cs through the online children of @parent_cs. Must be used
302 * with RCU read locked.
304 #define cpuset_for_each_child(child_cs, pos_css, parent_cs) \
305 css_for_each_child((pos_css), &(parent_cs)->css) \
306 if (is_cpuset_online(((child_cs) = css_cs((pos_css)))))
309 * cpuset_for_each_descendant_pre - pre-order walk of a cpuset's descendants
310 * @des_cs: loop cursor pointing to the current descendant
311 * @pos_css: used for iteration
312 * @root_cs: target cpuset to walk ancestor of
314 * Walk @des_cs through the online descendants of @root_cs. Must be used
315 * with RCU read locked. The caller may modify @pos_css by calling
316 * css_rightmost_descendant() to skip subtree. @root_cs is included in the
317 * iteration and the first node to be visited.
319 #define cpuset_for_each_descendant_pre(des_cs, pos_css, root_cs) \
320 css_for_each_descendant_pre((pos_css), &(root_cs)->css) \
321 if (is_cpuset_online(((des_cs) = css_cs((pos_css)))))
324 * There are two global locks guarding cpuset structures - cpuset_mutex and
325 * callback_lock. We also require taking task_lock() when dereferencing a
326 * task's cpuset pointer. See "The task_lock() exception", at the end of this
329 * A task must hold both locks to modify cpusets. If a task holds
330 * cpuset_mutex, then it blocks others wanting that mutex, ensuring that it
331 * is the only task able to also acquire callback_lock and be able to
332 * modify cpusets. It can perform various checks on the cpuset structure
333 * first, knowing nothing will change. It can also allocate memory while
334 * just holding cpuset_mutex. While it is performing these checks, various
335 * callback routines can briefly acquire callback_lock to query cpusets.
336 * Once it is ready to make the changes, it takes callback_lock, blocking
339 * Calls to the kernel memory allocator can not be made while holding
340 * callback_lock, as that would risk double tripping on callback_lock
341 * from one of the callbacks into the cpuset code from within
344 * If a task is only holding callback_lock, then it has read-only
347 * Now, the task_struct fields mems_allowed and mempolicy may be changed
348 * by other task, we use alloc_lock in the task_struct fields to protect
351 * The cpuset_common_file_read() handlers only hold callback_lock across
352 * small pieces of code, such as when reading out possibly multi-word
353 * cpumasks and nodemasks.
355 * Accessing a task's cpuset should be done in accordance with the
356 * guidelines for accessing subsystem state in kernel/cgroup.c
359 static DEFINE_MUTEX(cpuset_mutex);
361 void cpuset_lock(void)
363 mutex_lock(&cpuset_mutex);
366 void cpuset_unlock(void)
368 mutex_unlock(&cpuset_mutex);
371 static DEFINE_SPINLOCK(callback_lock);
373 static struct workqueue_struct *cpuset_migrate_mm_wq;
376 * CPU / memory hotplug is handled asynchronously.
378 static void cpuset_hotplug_workfn(struct work_struct *work);
379 static DECLARE_WORK(cpuset_hotplug_work, cpuset_hotplug_workfn);
381 static DECLARE_WAIT_QUEUE_HEAD(cpuset_attach_wq);
384 * Cgroup v2 behavior is used on the "cpus" and "mems" control files when
385 * on default hierarchy or when the cpuset_v2_mode flag is set by mounting
386 * the v1 cpuset cgroup filesystem with the "cpuset_v2_mode" mount option.
387 * With v2 behavior, "cpus" and "mems" are always what the users have
388 * requested and won't be changed by hotplug events. Only the effective
389 * cpus or mems will be affected.
391 static inline bool is_in_v2_mode(void)
393 return cgroup_subsys_on_dfl(cpuset_cgrp_subsys) ||
394 (cpuset_cgrp_subsys.root->flags & CGRP_ROOT_CPUSET_V2_MODE);
398 * Return in pmask the portion of a cpusets's cpus_allowed that
399 * are online. If none are online, walk up the cpuset hierarchy
400 * until we find one that does have some online cpus.
402 * One way or another, we guarantee to return some non-empty subset
403 * of cpu_online_mask.
405 * Call with callback_lock or cpuset_mutex held.
407 static void guarantee_online_cpus(struct cpuset *cs, struct cpumask *pmask)
409 while (!cpumask_intersects(cs->effective_cpus, cpu_online_mask)) {
413 * The top cpuset doesn't have any online cpu as a
414 * consequence of a race between cpuset_hotplug_work
415 * and cpu hotplug notifier. But we know the top
416 * cpuset's effective_cpus is on its way to be
417 * identical to cpu_online_mask.
419 cpumask_copy(pmask, cpu_online_mask);
423 cpumask_and(pmask, cs->effective_cpus, cpu_online_mask);
427 * Return in *pmask the portion of a cpusets's mems_allowed that
428 * are online, with memory. If none are online with memory, walk
429 * up the cpuset hierarchy until we find one that does have some
430 * online mems. The top cpuset always has some mems online.
432 * One way or another, we guarantee to return some non-empty subset
433 * of node_states[N_MEMORY].
435 * Call with callback_lock or cpuset_mutex held.
437 static void guarantee_online_mems(struct cpuset *cs, nodemask_t *pmask)
439 while (!nodes_intersects(cs->effective_mems, node_states[N_MEMORY]))
441 nodes_and(*pmask, cs->effective_mems, node_states[N_MEMORY]);
445 * update task's spread flag if cpuset's page/slab spread flag is set
447 * Call with callback_lock or cpuset_mutex held.
449 static void cpuset_update_task_spread_flag(struct cpuset *cs,
450 struct task_struct *tsk)
452 if (is_spread_page(cs))
453 task_set_spread_page(tsk);
455 task_clear_spread_page(tsk);
457 if (is_spread_slab(cs))
458 task_set_spread_slab(tsk);
460 task_clear_spread_slab(tsk);
464 * is_cpuset_subset(p, q) - Is cpuset p a subset of cpuset q?
466 * One cpuset is a subset of another if all its allowed CPUs and
467 * Memory Nodes are a subset of the other, and its exclusive flags
468 * are only set if the other's are set. Call holding cpuset_mutex.
471 static int is_cpuset_subset(const struct cpuset *p, const struct cpuset *q)
473 return cpumask_subset(p->cpus_allowed, q->cpus_allowed) &&
474 nodes_subset(p->mems_allowed, q->mems_allowed) &&
475 is_cpu_exclusive(p) <= is_cpu_exclusive(q) &&
476 is_mem_exclusive(p) <= is_mem_exclusive(q);
480 * alloc_cpumasks - allocate three cpumasks for cpuset
481 * @cs: the cpuset that have cpumasks to be allocated.
482 * @tmp: the tmpmasks structure pointer
483 * Return: 0 if successful, -ENOMEM otherwise.
485 * Only one of the two input arguments should be non-NULL.
487 static inline int alloc_cpumasks(struct cpuset *cs, struct tmpmasks *tmp)
489 cpumask_var_t *pmask1, *pmask2, *pmask3;
492 pmask1 = &cs->cpus_allowed;
493 pmask2 = &cs->effective_cpus;
494 pmask3 = &cs->subparts_cpus;
496 pmask1 = &tmp->new_cpus;
497 pmask2 = &tmp->addmask;
498 pmask3 = &tmp->delmask;
501 if (!zalloc_cpumask_var(pmask1, GFP_KERNEL))
504 if (!zalloc_cpumask_var(pmask2, GFP_KERNEL))
507 if (!zalloc_cpumask_var(pmask3, GFP_KERNEL))
513 free_cpumask_var(*pmask2);
515 free_cpumask_var(*pmask1);
520 * free_cpumasks - free cpumasks in a tmpmasks structure
521 * @cs: the cpuset that have cpumasks to be free.
522 * @tmp: the tmpmasks structure pointer
524 static inline void free_cpumasks(struct cpuset *cs, struct tmpmasks *tmp)
527 free_cpumask_var(cs->cpus_allowed);
528 free_cpumask_var(cs->effective_cpus);
529 free_cpumask_var(cs->subparts_cpus);
532 free_cpumask_var(tmp->new_cpus);
533 free_cpumask_var(tmp->addmask);
534 free_cpumask_var(tmp->delmask);
539 * alloc_trial_cpuset - allocate a trial cpuset
540 * @cs: the cpuset that the trial cpuset duplicates
542 static struct cpuset *alloc_trial_cpuset(struct cpuset *cs)
544 struct cpuset *trial;
546 trial = kmemdup(cs, sizeof(*cs), GFP_KERNEL);
550 if (alloc_cpumasks(trial, NULL)) {
555 cpumask_copy(trial->cpus_allowed, cs->cpus_allowed);
556 cpumask_copy(trial->effective_cpus, cs->effective_cpus);
561 * free_cpuset - free the cpuset
562 * @cs: the cpuset to be freed
564 static inline void free_cpuset(struct cpuset *cs)
566 free_cpumasks(cs, NULL);
571 * validate_change() - Used to validate that any proposed cpuset change
572 * follows the structural rules for cpusets.
574 * If we replaced the flag and mask values of the current cpuset
575 * (cur) with those values in the trial cpuset (trial), would
576 * our various subset and exclusive rules still be valid? Presumes
579 * 'cur' is the address of an actual, in-use cpuset. Operations
580 * such as list traversal that depend on the actual address of the
581 * cpuset in the list must use cur below, not trial.
583 * 'trial' is the address of bulk structure copy of cur, with
584 * perhaps one or more of the fields cpus_allowed, mems_allowed,
585 * or flags changed to new, trial values.
587 * Return 0 if valid, -errno if not.
590 static int validate_change(struct cpuset *cur, struct cpuset *trial)
592 struct cgroup_subsys_state *css;
593 struct cpuset *c, *par;
598 /* Each of our child cpusets must be a subset of us */
600 cpuset_for_each_child(c, css, cur)
601 if (!is_cpuset_subset(c, trial))
604 /* Remaining checks don't apply to root cpuset */
606 if (cur == &top_cpuset)
609 par = parent_cs(cur);
611 /* On legacy hiearchy, we must be a subset of our parent cpuset. */
613 if (!is_in_v2_mode() && !is_cpuset_subset(trial, par))
617 * If either I or some sibling (!= me) is exclusive, we can't
621 cpuset_for_each_child(c, css, par) {
622 if ((is_cpu_exclusive(trial) || is_cpu_exclusive(c)) &&
624 cpumask_intersects(trial->cpus_allowed, c->cpus_allowed))
626 if ((is_mem_exclusive(trial) || is_mem_exclusive(c)) &&
628 nodes_intersects(trial->mems_allowed, c->mems_allowed))
633 * Cpusets with tasks - existing or newly being attached - can't
634 * be changed to have empty cpus_allowed or mems_allowed.
637 if ((cgroup_is_populated(cur->css.cgroup) || cur->attach_in_progress)) {
638 if (!cpumask_empty(cur->cpus_allowed) &&
639 cpumask_empty(trial->cpus_allowed))
641 if (!nodes_empty(cur->mems_allowed) &&
642 nodes_empty(trial->mems_allowed))
647 * We can't shrink if we won't have enough room for SCHED_DEADLINE
651 if (is_cpu_exclusive(cur) &&
652 !cpuset_cpumask_can_shrink(cur->cpus_allowed,
653 trial->cpus_allowed))
664 * Helper routine for generate_sched_domains().
665 * Do cpusets a, b have overlapping effective cpus_allowed masks?
667 static int cpusets_overlap(struct cpuset *a, struct cpuset *b)
669 return cpumask_intersects(a->effective_cpus, b->effective_cpus);
673 update_domain_attr(struct sched_domain_attr *dattr, struct cpuset *c)
675 if (dattr->relax_domain_level < c->relax_domain_level)
676 dattr->relax_domain_level = c->relax_domain_level;
680 static void update_domain_attr_tree(struct sched_domain_attr *dattr,
681 struct cpuset *root_cs)
684 struct cgroup_subsys_state *pos_css;
687 cpuset_for_each_descendant_pre(cp, pos_css, root_cs) {
688 /* skip the whole subtree if @cp doesn't have any CPU */
689 if (cpumask_empty(cp->cpus_allowed)) {
690 pos_css = css_rightmost_descendant(pos_css);
694 if (is_sched_load_balance(cp))
695 update_domain_attr(dattr, cp);
700 /* Must be called with cpuset_mutex held. */
701 static inline int nr_cpusets(void)
703 /* jump label reference count + the top-level cpuset */
704 return static_key_count(&cpusets_enabled_key.key) + 1;
708 * generate_sched_domains()
710 * This function builds a partial partition of the systems CPUs
711 * A 'partial partition' is a set of non-overlapping subsets whose
712 * union is a subset of that set.
713 * The output of this function needs to be passed to kernel/sched/core.c
714 * partition_sched_domains() routine, which will rebuild the scheduler's
715 * load balancing domains (sched domains) as specified by that partial
718 * See "What is sched_load_balance" in Documentation/admin-guide/cgroup-v1/cpusets.rst
719 * for a background explanation of this.
721 * Does not return errors, on the theory that the callers of this
722 * routine would rather not worry about failures to rebuild sched
723 * domains when operating in the severe memory shortage situations
724 * that could cause allocation failures below.
726 * Must be called with cpuset_mutex held.
728 * The three key local variables below are:
729 * cp - cpuset pointer, used (together with pos_css) to perform a
730 * top-down scan of all cpusets. For our purposes, rebuilding
731 * the schedulers sched domains, we can ignore !is_sched_load_
733 * csa - (for CpuSet Array) Array of pointers to all the cpusets
734 * that need to be load balanced, for convenient iterative
735 * access by the subsequent code that finds the best partition,
736 * i.e the set of domains (subsets) of CPUs such that the
737 * cpus_allowed of every cpuset marked is_sched_load_balance
738 * is a subset of one of these domains, while there are as
739 * many such domains as possible, each as small as possible.
740 * doms - Conversion of 'csa' to an array of cpumasks, for passing to
741 * the kernel/sched/core.c routine partition_sched_domains() in a
742 * convenient format, that can be easily compared to the prior
743 * value to determine what partition elements (sched domains)
744 * were changed (added or removed.)
746 * Finding the best partition (set of domains):
747 * The triple nested loops below over i, j, k scan over the
748 * load balanced cpusets (using the array of cpuset pointers in
749 * csa[]) looking for pairs of cpusets that have overlapping
750 * cpus_allowed, but which don't have the same 'pn' partition
751 * number and gives them in the same partition number. It keeps
752 * looping on the 'restart' label until it can no longer find
755 * The union of the cpus_allowed masks from the set of
756 * all cpusets having the same 'pn' value then form the one
757 * element of the partition (one sched domain) to be passed to
758 * partition_sched_domains().
760 static int generate_sched_domains(cpumask_var_t **domains,
761 struct sched_domain_attr **attributes)
763 struct cpuset *cp; /* top-down scan of cpusets */
764 struct cpuset **csa; /* array of all cpuset ptrs */
765 int csn; /* how many cpuset ptrs in csa so far */
766 int i, j, k; /* indices for partition finding loops */
767 cpumask_var_t *doms; /* resulting partition; i.e. sched domains */
768 struct sched_domain_attr *dattr; /* attributes for custom domains */
769 int ndoms = 0; /* number of sched domains in result */
770 int nslot; /* next empty doms[] struct cpumask slot */
771 struct cgroup_subsys_state *pos_css;
772 bool root_load_balance = is_sched_load_balance(&top_cpuset);
778 /* Special case for the 99% of systems with one, full, sched domain */
779 if (root_load_balance && !top_cpuset.nr_subparts_cpus) {
781 doms = alloc_sched_domains(ndoms);
785 dattr = kmalloc(sizeof(struct sched_domain_attr), GFP_KERNEL);
787 *dattr = SD_ATTR_INIT;
788 update_domain_attr_tree(dattr, &top_cpuset);
790 cpumask_and(doms[0], top_cpuset.effective_cpus,
791 housekeeping_cpumask(HK_FLAG_DOMAIN));
796 csa = kmalloc_array(nr_cpusets(), sizeof(cp), GFP_KERNEL);
802 if (root_load_balance)
803 csa[csn++] = &top_cpuset;
804 cpuset_for_each_descendant_pre(cp, pos_css, &top_cpuset) {
805 if (cp == &top_cpuset)
808 * Continue traversing beyond @cp iff @cp has some CPUs and
809 * isn't load balancing. The former is obvious. The
810 * latter: All child cpusets contain a subset of the
811 * parent's cpus, so just skip them, and then we call
812 * update_domain_attr_tree() to calc relax_domain_level of
813 * the corresponding sched domain.
815 * If root is load-balancing, we can skip @cp if it
816 * is a subset of the root's effective_cpus.
818 if (!cpumask_empty(cp->cpus_allowed) &&
819 !(is_sched_load_balance(cp) &&
820 cpumask_intersects(cp->cpus_allowed,
821 housekeeping_cpumask(HK_FLAG_DOMAIN))))
824 if (root_load_balance &&
825 cpumask_subset(cp->cpus_allowed, top_cpuset.effective_cpus))
828 if (is_sched_load_balance(cp) &&
829 !cpumask_empty(cp->effective_cpus))
832 /* skip @cp's subtree if not a partition root */
833 if (!is_partition_root(cp))
834 pos_css = css_rightmost_descendant(pos_css);
838 for (i = 0; i < csn; i++)
843 /* Find the best partition (set of sched domains) */
844 for (i = 0; i < csn; i++) {
845 struct cpuset *a = csa[i];
848 for (j = 0; j < csn; j++) {
849 struct cpuset *b = csa[j];
852 if (apn != bpn && cpusets_overlap(a, b)) {
853 for (k = 0; k < csn; k++) {
854 struct cpuset *c = csa[k];
859 ndoms--; /* one less element */
866 * Now we know how many domains to create.
867 * Convert <csn, csa> to <ndoms, doms> and populate cpu masks.
869 doms = alloc_sched_domains(ndoms);
874 * The rest of the code, including the scheduler, can deal with
875 * dattr==NULL case. No need to abort if alloc fails.
877 dattr = kmalloc_array(ndoms, sizeof(struct sched_domain_attr),
880 for (nslot = 0, i = 0; i < csn; i++) {
881 struct cpuset *a = csa[i];
886 /* Skip completed partitions */
892 if (nslot == ndoms) {
893 static int warnings = 10;
895 pr_warn("rebuild_sched_domains confused: nslot %d, ndoms %d, csn %d, i %d, apn %d\n",
896 nslot, ndoms, csn, i, apn);
904 *(dattr + nslot) = SD_ATTR_INIT;
905 for (j = i; j < csn; j++) {
906 struct cpuset *b = csa[j];
909 cpumask_or(dp, dp, b->effective_cpus);
910 cpumask_and(dp, dp, housekeeping_cpumask(HK_FLAG_DOMAIN));
912 update_domain_attr_tree(dattr + nslot, b);
914 /* Done with this partition */
920 BUG_ON(nslot != ndoms);
926 * Fallback to the default domain if kmalloc() failed.
927 * See comments in partition_sched_domains().
937 static void dl_update_tasks_root_domain(struct cpuset *cs)
939 struct css_task_iter it;
940 struct task_struct *task;
942 if (cs->nr_deadline_tasks == 0)
945 css_task_iter_start(&cs->css, 0, &it);
947 while ((task = css_task_iter_next(&it)))
948 dl_add_task_root_domain(task);
950 css_task_iter_end(&it);
953 static void dl_rebuild_rd_accounting(void)
955 struct cpuset *cs = NULL;
956 struct cgroup_subsys_state *pos_css;
958 lockdep_assert_held(&cpuset_mutex);
959 lockdep_assert_cpus_held();
960 lockdep_assert_held(&sched_domains_mutex);
965 * Clear default root domain DL accounting, it will be computed again
966 * if a task belongs to it.
968 dl_clear_root_domain(&def_root_domain);
970 cpuset_for_each_descendant_pre(cs, pos_css, &top_cpuset) {
972 if (cpumask_empty(cs->effective_cpus)) {
973 pos_css = css_rightmost_descendant(pos_css);
981 dl_update_tasks_root_domain(cs);
990 partition_and_rebuild_sched_domains(int ndoms_new, cpumask_var_t doms_new[],
991 struct sched_domain_attr *dattr_new)
993 mutex_lock(&sched_domains_mutex);
994 partition_sched_domains_locked(ndoms_new, doms_new, dattr_new);
995 dl_rebuild_rd_accounting();
996 mutex_unlock(&sched_domains_mutex);
1000 * Rebuild scheduler domains.
1002 * If the flag 'sched_load_balance' of any cpuset with non-empty
1003 * 'cpus' changes, or if the 'cpus' allowed changes in any cpuset
1004 * which has that flag enabled, or if any cpuset with a non-empty
1005 * 'cpus' is removed, then call this routine to rebuild the
1006 * scheduler's dynamic sched domains.
1008 * Call with cpuset_mutex held. Takes get_online_cpus().
1010 static void rebuild_sched_domains_locked(void)
1012 struct cgroup_subsys_state *pos_css;
1013 struct sched_domain_attr *attr;
1014 cpumask_var_t *doms;
1018 lockdep_assert_cpus_held();
1019 lockdep_assert_held(&cpuset_mutex);
1022 * If we have raced with CPU hotplug, return early to avoid
1023 * passing doms with offlined cpu to partition_sched_domains().
1024 * Anyways, cpuset_hotplug_workfn() will rebuild sched domains.
1026 * With no CPUs in any subpartitions, top_cpuset's effective CPUs
1027 * should be the same as the active CPUs, so checking only top_cpuset
1028 * is enough to detect racing CPU offlines.
1030 if (!top_cpuset.nr_subparts_cpus &&
1031 !cpumask_equal(top_cpuset.effective_cpus, cpu_active_mask))
1035 * With subpartition CPUs, however, the effective CPUs of a partition
1036 * root should be only a subset of the active CPUs. Since a CPU in any
1037 * partition root could be offlined, all must be checked.
1039 if (top_cpuset.nr_subparts_cpus) {
1041 cpuset_for_each_descendant_pre(cs, pos_css, &top_cpuset) {
1042 if (!is_partition_root(cs)) {
1043 pos_css = css_rightmost_descendant(pos_css);
1046 if (!cpumask_subset(cs->effective_cpus,
1055 /* Generate domain masks and attrs */
1056 ndoms = generate_sched_domains(&doms, &attr);
1058 /* Have scheduler rebuild the domains */
1059 partition_and_rebuild_sched_domains(ndoms, doms, attr);
1061 #else /* !CONFIG_SMP */
1062 static void rebuild_sched_domains_locked(void)
1065 #endif /* CONFIG_SMP */
1067 void rebuild_sched_domains(void)
1070 mutex_lock(&cpuset_mutex);
1071 rebuild_sched_domains_locked();
1072 mutex_unlock(&cpuset_mutex);
1077 * update_tasks_cpumask - Update the cpumasks of tasks in the cpuset.
1078 * @cs: the cpuset in which each task's cpus_allowed mask needs to be changed
1080 * Iterate through each task of @cs updating its cpus_allowed to the
1081 * effective cpuset's. As this function is called with cpuset_mutex held,
1082 * cpuset membership stays stable.
1084 static void update_tasks_cpumask(struct cpuset *cs)
1086 struct css_task_iter it;
1087 struct task_struct *task;
1088 bool top_cs = cs == &top_cpuset;
1090 css_task_iter_start(&cs->css, 0, &it);
1091 while ((task = css_task_iter_next(&it))) {
1093 * Percpu kthreads in top_cpuset are ignored
1095 if (top_cs && (task->flags & PF_KTHREAD) &&
1096 kthread_is_per_cpu(task))
1098 set_cpus_allowed_ptr(task, cs->effective_cpus);
1100 css_task_iter_end(&it);
1104 * compute_effective_cpumask - Compute the effective cpumask of the cpuset
1105 * @new_cpus: the temp variable for the new effective_cpus mask
1106 * @cs: the cpuset the need to recompute the new effective_cpus mask
1107 * @parent: the parent cpuset
1109 * If the parent has subpartition CPUs, include them in the list of
1110 * allowable CPUs in computing the new effective_cpus mask. Since offlined
1111 * CPUs are not removed from subparts_cpus, we have to use cpu_active_mask
1112 * to mask those out.
1114 static void compute_effective_cpumask(struct cpumask *new_cpus,
1115 struct cpuset *cs, struct cpuset *parent)
1117 if (parent->nr_subparts_cpus) {
1118 cpumask_or(new_cpus, parent->effective_cpus,
1119 parent->subparts_cpus);
1120 cpumask_and(new_cpus, new_cpus, cs->cpus_allowed);
1121 cpumask_and(new_cpus, new_cpus, cpu_active_mask);
1123 cpumask_and(new_cpus, cs->cpus_allowed, parent->effective_cpus);
1128 * Commands for update_parent_subparts_cpumask
1131 partcmd_enable, /* Enable partition root */
1132 partcmd_disable, /* Disable partition root */
1133 partcmd_update, /* Update parent's subparts_cpus */
1137 * update_parent_subparts_cpumask - update subparts_cpus mask of parent cpuset
1138 * @cpuset: The cpuset that requests change in partition root state
1139 * @cmd: Partition root state change command
1140 * @newmask: Optional new cpumask for partcmd_update
1141 * @tmp: Temporary addmask and delmask
1142 * Return: 0, 1 or an error code
1144 * For partcmd_enable, the cpuset is being transformed from a non-partition
1145 * root to a partition root. The cpus_allowed mask of the given cpuset will
1146 * be put into parent's subparts_cpus and taken away from parent's
1147 * effective_cpus. The function will return 0 if all the CPUs listed in
1148 * cpus_allowed can be granted or an error code will be returned.
1150 * For partcmd_disable, the cpuset is being transofrmed from a partition
1151 * root back to a non-partition root. Any CPUs in cpus_allowed that are in
1152 * parent's subparts_cpus will be taken away from that cpumask and put back
1153 * into parent's effective_cpus. 0 should always be returned.
1155 * For partcmd_update, if the optional newmask is specified, the cpu
1156 * list is to be changed from cpus_allowed to newmask. Otherwise,
1157 * cpus_allowed is assumed to remain the same. The cpuset should either
1158 * be a partition root or an invalid partition root. The partition root
1159 * state may change if newmask is NULL and none of the requested CPUs can
1160 * be granted by the parent. The function will return 1 if changes to
1161 * parent's subparts_cpus and effective_cpus happen or 0 otherwise.
1162 * Error code should only be returned when newmask is non-NULL.
1164 * The partcmd_enable and partcmd_disable commands are used by
1165 * update_prstate(). The partcmd_update command is used by
1166 * update_cpumasks_hier() with newmask NULL and update_cpumask() with
1169 * The checking is more strict when enabling partition root than the
1170 * other two commands.
1172 * Because of the implicit cpu exclusive nature of a partition root,
1173 * cpumask changes that violates the cpu exclusivity rule will not be
1174 * permitted when checked by validate_change(). The validate_change()
1175 * function will also prevent any changes to the cpu list if it is not
1176 * a superset of children's cpu lists.
1178 static int update_parent_subparts_cpumask(struct cpuset *cpuset, int cmd,
1179 struct cpumask *newmask,
1180 struct tmpmasks *tmp)
1182 struct cpuset *parent = parent_cs(cpuset);
1183 int adding; /* Moving cpus from effective_cpus to subparts_cpus */
1184 int deleting; /* Moving cpus from subparts_cpus to effective_cpus */
1186 bool part_error = false; /* Partition error? */
1188 lockdep_assert_held(&cpuset_mutex);
1191 * The parent must be a partition root.
1192 * The new cpumask, if present, or the current cpus_allowed must
1195 if (!is_partition_root(parent) ||
1196 (newmask && cpumask_empty(newmask)) ||
1197 (!newmask && cpumask_empty(cpuset->cpus_allowed)))
1201 * Enabling/disabling partition root is not allowed if there are
1204 if ((cmd != partcmd_update) && css_has_online_children(&cpuset->css))
1208 * Enabling partition root is not allowed if not all the CPUs
1209 * can be granted from parent's effective_cpus or at least one
1210 * CPU will be left after that.
1212 if ((cmd == partcmd_enable) &&
1213 (!cpumask_subset(cpuset->cpus_allowed, parent->effective_cpus) ||
1214 cpumask_equal(cpuset->cpus_allowed, parent->effective_cpus)))
1218 * A cpumask update cannot make parent's effective_cpus become empty.
1220 adding = deleting = false;
1221 new_prs = cpuset->partition_root_state;
1222 if (cmd == partcmd_enable) {
1223 cpumask_copy(tmp->addmask, cpuset->cpus_allowed);
1225 } else if (cmd == partcmd_disable) {
1226 deleting = cpumask_and(tmp->delmask, cpuset->cpus_allowed,
1227 parent->subparts_cpus);
1228 } else if (newmask) {
1230 * partcmd_update with newmask:
1232 * delmask = cpus_allowed & ~newmask & parent->subparts_cpus
1233 * addmask = newmask & parent->effective_cpus
1234 * & ~parent->subparts_cpus
1236 cpumask_andnot(tmp->delmask, cpuset->cpus_allowed, newmask);
1237 deleting = cpumask_and(tmp->delmask, tmp->delmask,
1238 parent->subparts_cpus);
1240 cpumask_and(tmp->addmask, newmask, parent->effective_cpus);
1241 adding = cpumask_andnot(tmp->addmask, tmp->addmask,
1242 parent->subparts_cpus);
1244 * Return error if the new effective_cpus could become empty.
1247 cpumask_equal(parent->effective_cpus, tmp->addmask)) {
1251 * As some of the CPUs in subparts_cpus might have
1252 * been offlined, we need to compute the real delmask
1255 if (!cpumask_and(tmp->addmask, tmp->delmask,
1258 cpumask_copy(tmp->addmask, parent->effective_cpus);
1262 * partcmd_update w/o newmask:
1264 * addmask = cpus_allowed & parent->effective_cpus
1266 * Note that parent's subparts_cpus may have been
1267 * pre-shrunk in case there is a change in the cpu list.
1268 * So no deletion is needed.
1270 adding = cpumask_and(tmp->addmask, cpuset->cpus_allowed,
1271 parent->effective_cpus);
1272 part_error = cpumask_equal(tmp->addmask,
1273 parent->effective_cpus);
1276 if (cmd == partcmd_update) {
1277 int prev_prs = cpuset->partition_root_state;
1280 * Check for possible transition between PRS_ENABLED
1283 switch (cpuset->partition_root_state) {
1286 new_prs = PRS_ERROR;
1290 new_prs = PRS_ENABLED;
1294 * Set part_error if previously in invalid state.
1296 part_error = (prev_prs == PRS_ERROR);
1299 if (!part_error && (new_prs == PRS_ERROR))
1300 return 0; /* Nothing need to be done */
1302 if (new_prs == PRS_ERROR) {
1304 * Remove all its cpus from parent's subparts_cpus.
1307 deleting = cpumask_and(tmp->delmask, cpuset->cpus_allowed,
1308 parent->subparts_cpus);
1311 if (!adding && !deleting && (new_prs == cpuset->partition_root_state))
1315 * Change the parent's subparts_cpus.
1316 * Newly added CPUs will be removed from effective_cpus and
1317 * newly deleted ones will be added back to effective_cpus.
1319 spin_lock_irq(&callback_lock);
1321 cpumask_or(parent->subparts_cpus,
1322 parent->subparts_cpus, tmp->addmask);
1323 cpumask_andnot(parent->effective_cpus,
1324 parent->effective_cpus, tmp->addmask);
1327 cpumask_andnot(parent->subparts_cpus,
1328 parent->subparts_cpus, tmp->delmask);
1330 * Some of the CPUs in subparts_cpus might have been offlined.
1332 cpumask_and(tmp->delmask, tmp->delmask, cpu_active_mask);
1333 cpumask_or(parent->effective_cpus,
1334 parent->effective_cpus, tmp->delmask);
1337 parent->nr_subparts_cpus = cpumask_weight(parent->subparts_cpus);
1339 if (cpuset->partition_root_state != new_prs)
1340 cpuset->partition_root_state = new_prs;
1341 spin_unlock_irq(&callback_lock);
1343 return cmd == partcmd_update;
1347 * update_cpumasks_hier - Update effective cpumasks and tasks in the subtree
1348 * @cs: the cpuset to consider
1349 * @tmp: temp variables for calculating effective_cpus & partition setup
1351 * When congifured cpumask is changed, the effective cpumasks of this cpuset
1352 * and all its descendants need to be updated.
1354 * On legacy hierachy, effective_cpus will be the same with cpu_allowed.
1356 * Called with cpuset_mutex held
1358 static void update_cpumasks_hier(struct cpuset *cs, struct tmpmasks *tmp)
1361 struct cgroup_subsys_state *pos_css;
1362 bool need_rebuild_sched_domains = false;
1366 cpuset_for_each_descendant_pre(cp, pos_css, cs) {
1367 struct cpuset *parent = parent_cs(cp);
1369 compute_effective_cpumask(tmp->new_cpus, cp, parent);
1372 * If it becomes empty, inherit the effective mask of the
1373 * parent, which is guaranteed to have some CPUs.
1375 if (is_in_v2_mode() && cpumask_empty(tmp->new_cpus)) {
1376 cpumask_copy(tmp->new_cpus, parent->effective_cpus);
1377 if (!cp->use_parent_ecpus) {
1378 cp->use_parent_ecpus = true;
1379 parent->child_ecpus_count++;
1381 } else if (cp->use_parent_ecpus) {
1382 cp->use_parent_ecpus = false;
1383 WARN_ON_ONCE(!parent->child_ecpus_count);
1384 parent->child_ecpus_count--;
1388 * Skip the whole subtree if the cpumask remains the same
1389 * and has no partition root state.
1391 if (!cp->partition_root_state &&
1392 cpumask_equal(tmp->new_cpus, cp->effective_cpus)) {
1393 pos_css = css_rightmost_descendant(pos_css);
1398 * update_parent_subparts_cpumask() should have been called
1399 * for cs already in update_cpumask(). We should also call
1400 * update_tasks_cpumask() again for tasks in the parent
1401 * cpuset if the parent's subparts_cpus changes.
1403 new_prs = cp->partition_root_state;
1404 if ((cp != cs) && new_prs) {
1405 switch (parent->partition_root_state) {
1408 * If parent is not a partition root or an
1409 * invalid partition root, clear its state
1410 * and its CS_CPU_EXCLUSIVE flag.
1412 WARN_ON_ONCE(cp->partition_root_state
1414 new_prs = PRS_DISABLED;
1417 * clear_bit() is an atomic operation and
1418 * readers aren't interested in the state
1419 * of CS_CPU_EXCLUSIVE anyway. So we can
1420 * just update the flag without holding
1421 * the callback_lock.
1423 clear_bit(CS_CPU_EXCLUSIVE, &cp->flags);
1427 if (update_parent_subparts_cpumask(cp, partcmd_update, NULL, tmp))
1428 update_tasks_cpumask(parent);
1433 * When parent is invalid, it has to be too.
1435 new_prs = PRS_ERROR;
1440 if (!css_tryget_online(&cp->css))
1444 spin_lock_irq(&callback_lock);
1446 cpumask_copy(cp->effective_cpus, tmp->new_cpus);
1447 if (cp->nr_subparts_cpus && (new_prs != PRS_ENABLED)) {
1448 cp->nr_subparts_cpus = 0;
1449 cpumask_clear(cp->subparts_cpus);
1450 } else if (cp->nr_subparts_cpus) {
1452 * Make sure that effective_cpus & subparts_cpus
1453 * are mutually exclusive.
1455 * In the unlikely event that effective_cpus
1456 * becomes empty. we clear cp->nr_subparts_cpus and
1457 * let its child partition roots to compete for
1460 cpumask_andnot(cp->effective_cpus, cp->effective_cpus,
1462 if (cpumask_empty(cp->effective_cpus)) {
1463 cpumask_copy(cp->effective_cpus, tmp->new_cpus);
1464 cpumask_clear(cp->subparts_cpus);
1465 cp->nr_subparts_cpus = 0;
1466 } else if (!cpumask_subset(cp->subparts_cpus,
1468 cpumask_andnot(cp->subparts_cpus,
1469 cp->subparts_cpus, tmp->new_cpus);
1470 cp->nr_subparts_cpus
1471 = cpumask_weight(cp->subparts_cpus);
1475 if (new_prs != cp->partition_root_state)
1476 cp->partition_root_state = new_prs;
1478 spin_unlock_irq(&callback_lock);
1480 WARN_ON(!is_in_v2_mode() &&
1481 !cpumask_equal(cp->cpus_allowed, cp->effective_cpus));
1483 update_tasks_cpumask(cp);
1486 * On legacy hierarchy, if the effective cpumask of any non-
1487 * empty cpuset is changed, we need to rebuild sched domains.
1488 * On default hierarchy, the cpuset needs to be a partition
1491 if (!cpumask_empty(cp->cpus_allowed) &&
1492 is_sched_load_balance(cp) &&
1493 (!cgroup_subsys_on_dfl(cpuset_cgrp_subsys) ||
1494 is_partition_root(cp)))
1495 need_rebuild_sched_domains = true;
1502 if (need_rebuild_sched_domains)
1503 rebuild_sched_domains_locked();
1507 * update_sibling_cpumasks - Update siblings cpumasks
1508 * @parent: Parent cpuset
1509 * @cs: Current cpuset
1510 * @tmp: Temp variables
1512 static void update_sibling_cpumasks(struct cpuset *parent, struct cpuset *cs,
1513 struct tmpmasks *tmp)
1515 struct cpuset *sibling;
1516 struct cgroup_subsys_state *pos_css;
1518 lockdep_assert_held(&cpuset_mutex);
1521 * Check all its siblings and call update_cpumasks_hier()
1522 * if their use_parent_ecpus flag is set in order for them
1523 * to use the right effective_cpus value.
1525 * The update_cpumasks_hier() function may sleep. So we have to
1526 * release the RCU read lock before calling it.
1529 cpuset_for_each_child(sibling, pos_css, parent) {
1532 if (!sibling->use_parent_ecpus)
1534 if (!css_tryget_online(&sibling->css))
1538 update_cpumasks_hier(sibling, tmp);
1540 css_put(&sibling->css);
1546 * update_cpumask - update the cpus_allowed mask of a cpuset and all tasks in it
1547 * @cs: the cpuset to consider
1548 * @trialcs: trial cpuset
1549 * @buf: buffer of cpu numbers written to this cpuset
1551 static int update_cpumask(struct cpuset *cs, struct cpuset *trialcs,
1555 struct tmpmasks tmp;
1557 /* top_cpuset.cpus_allowed tracks cpu_online_mask; it's read-only */
1558 if (cs == &top_cpuset)
1562 * An empty cpus_allowed is ok only if the cpuset has no tasks.
1563 * Since cpulist_parse() fails on an empty mask, we special case
1564 * that parsing. The validate_change() call ensures that cpusets
1565 * with tasks have cpus.
1568 cpumask_clear(trialcs->cpus_allowed);
1570 retval = cpulist_parse(buf, trialcs->cpus_allowed);
1574 if (!cpumask_subset(trialcs->cpus_allowed,
1575 top_cpuset.cpus_allowed))
1579 /* Nothing to do if the cpus didn't change */
1580 if (cpumask_equal(cs->cpus_allowed, trialcs->cpus_allowed))
1583 retval = validate_change(cs, trialcs);
1587 #ifdef CONFIG_CPUMASK_OFFSTACK
1589 * Use the cpumasks in trialcs for tmpmasks when they are pointers
1590 * to allocated cpumasks.
1592 tmp.addmask = trialcs->subparts_cpus;
1593 tmp.delmask = trialcs->effective_cpus;
1594 tmp.new_cpus = trialcs->cpus_allowed;
1597 if (cs->partition_root_state) {
1598 /* Cpumask of a partition root cannot be empty */
1599 if (cpumask_empty(trialcs->cpus_allowed))
1601 if (update_parent_subparts_cpumask(cs, partcmd_update,
1602 trialcs->cpus_allowed, &tmp) < 0)
1606 spin_lock_irq(&callback_lock);
1607 cpumask_copy(cs->cpus_allowed, trialcs->cpus_allowed);
1610 * Make sure that subparts_cpus is a subset of cpus_allowed.
1612 if (cs->nr_subparts_cpus) {
1613 cpumask_and(cs->subparts_cpus, cs->subparts_cpus, cs->cpus_allowed);
1614 cs->nr_subparts_cpus = cpumask_weight(cs->subparts_cpus);
1616 spin_unlock_irq(&callback_lock);
1618 update_cpumasks_hier(cs, &tmp);
1620 if (cs->partition_root_state) {
1621 struct cpuset *parent = parent_cs(cs);
1624 * For partition root, update the cpumasks of sibling
1625 * cpusets if they use parent's effective_cpus.
1627 if (parent->child_ecpus_count)
1628 update_sibling_cpumasks(parent, cs, &tmp);
1634 * Migrate memory region from one set of nodes to another. This is
1635 * performed asynchronously as it can be called from process migration path
1636 * holding locks involved in process management. All mm migrations are
1637 * performed in the queued order and can be waited for by flushing
1638 * cpuset_migrate_mm_wq.
1641 struct cpuset_migrate_mm_work {
1642 struct work_struct work;
1643 struct mm_struct *mm;
1648 static void cpuset_migrate_mm_workfn(struct work_struct *work)
1650 struct cpuset_migrate_mm_work *mwork =
1651 container_of(work, struct cpuset_migrate_mm_work, work);
1653 /* on a wq worker, no need to worry about %current's mems_allowed */
1654 do_migrate_pages(mwork->mm, &mwork->from, &mwork->to, MPOL_MF_MOVE_ALL);
1659 static void cpuset_migrate_mm(struct mm_struct *mm, const nodemask_t *from,
1660 const nodemask_t *to)
1662 struct cpuset_migrate_mm_work *mwork;
1664 mwork = kzalloc(sizeof(*mwork), GFP_KERNEL);
1667 mwork->from = *from;
1669 INIT_WORK(&mwork->work, cpuset_migrate_mm_workfn);
1670 queue_work(cpuset_migrate_mm_wq, &mwork->work);
1676 static void cpuset_post_attach(void)
1678 flush_workqueue(cpuset_migrate_mm_wq);
1682 * cpuset_change_task_nodemask - change task's mems_allowed and mempolicy
1683 * @tsk: the task to change
1684 * @newmems: new nodes that the task will be set
1686 * We use the mems_allowed_seq seqlock to safely update both tsk->mems_allowed
1687 * and rebind an eventual tasks' mempolicy. If the task is allocating in
1688 * parallel, it might temporarily see an empty intersection, which results in
1689 * a seqlock check and retry before OOM or allocation failure.
1691 static void cpuset_change_task_nodemask(struct task_struct *tsk,
1692 nodemask_t *newmems)
1696 local_irq_disable();
1697 write_seqcount_begin(&tsk->mems_allowed_seq);
1699 nodes_or(tsk->mems_allowed, tsk->mems_allowed, *newmems);
1700 mpol_rebind_task(tsk, newmems);
1701 tsk->mems_allowed = *newmems;
1703 write_seqcount_end(&tsk->mems_allowed_seq);
1709 static void *cpuset_being_rebound;
1712 * update_tasks_nodemask - Update the nodemasks of tasks in the cpuset.
1713 * @cs: the cpuset in which each task's mems_allowed mask needs to be changed
1715 * Iterate through each task of @cs updating its mems_allowed to the
1716 * effective cpuset's. As this function is called with cpuset_mutex held,
1717 * cpuset membership stays stable.
1719 static void update_tasks_nodemask(struct cpuset *cs)
1721 static nodemask_t newmems; /* protected by cpuset_mutex */
1722 struct css_task_iter it;
1723 struct task_struct *task;
1725 cpuset_being_rebound = cs; /* causes mpol_dup() rebind */
1727 guarantee_online_mems(cs, &newmems);
1730 * The mpol_rebind_mm() call takes mmap_lock, which we couldn't
1731 * take while holding tasklist_lock. Forks can happen - the
1732 * mpol_dup() cpuset_being_rebound check will catch such forks,
1733 * and rebind their vma mempolicies too. Because we still hold
1734 * the global cpuset_mutex, we know that no other rebind effort
1735 * will be contending for the global variable cpuset_being_rebound.
1736 * It's ok if we rebind the same mm twice; mpol_rebind_mm()
1737 * is idempotent. Also migrate pages in each mm to new nodes.
1739 css_task_iter_start(&cs->css, 0, &it);
1740 while ((task = css_task_iter_next(&it))) {
1741 struct mm_struct *mm;
1744 cpuset_change_task_nodemask(task, &newmems);
1746 mm = get_task_mm(task);
1750 migrate = is_memory_migrate(cs);
1752 mpol_rebind_mm(mm, &cs->mems_allowed);
1754 cpuset_migrate_mm(mm, &cs->old_mems_allowed, &newmems);
1758 css_task_iter_end(&it);
1761 * All the tasks' nodemasks have been updated, update
1762 * cs->old_mems_allowed.
1764 cs->old_mems_allowed = newmems;
1766 /* We're done rebinding vmas to this cpuset's new mems_allowed. */
1767 cpuset_being_rebound = NULL;
1771 * update_nodemasks_hier - Update effective nodemasks and tasks in the subtree
1772 * @cs: the cpuset to consider
1773 * @new_mems: a temp variable for calculating new effective_mems
1775 * When configured nodemask is changed, the effective nodemasks of this cpuset
1776 * and all its descendants need to be updated.
1778 * On legacy hiearchy, effective_mems will be the same with mems_allowed.
1780 * Called with cpuset_mutex held
1782 static void update_nodemasks_hier(struct cpuset *cs, nodemask_t *new_mems)
1785 struct cgroup_subsys_state *pos_css;
1788 cpuset_for_each_descendant_pre(cp, pos_css, cs) {
1789 struct cpuset *parent = parent_cs(cp);
1791 nodes_and(*new_mems, cp->mems_allowed, parent->effective_mems);
1794 * If it becomes empty, inherit the effective mask of the
1795 * parent, which is guaranteed to have some MEMs.
1797 if (is_in_v2_mode() && nodes_empty(*new_mems))
1798 *new_mems = parent->effective_mems;
1800 /* Skip the whole subtree if the nodemask remains the same. */
1801 if (nodes_equal(*new_mems, cp->effective_mems)) {
1802 pos_css = css_rightmost_descendant(pos_css);
1806 if (!css_tryget_online(&cp->css))
1810 spin_lock_irq(&callback_lock);
1811 cp->effective_mems = *new_mems;
1812 spin_unlock_irq(&callback_lock);
1814 WARN_ON(!is_in_v2_mode() &&
1815 !nodes_equal(cp->mems_allowed, cp->effective_mems));
1817 update_tasks_nodemask(cp);
1826 * Handle user request to change the 'mems' memory placement
1827 * of a cpuset. Needs to validate the request, update the
1828 * cpusets mems_allowed, and for each task in the cpuset,
1829 * update mems_allowed and rebind task's mempolicy and any vma
1830 * mempolicies and if the cpuset is marked 'memory_migrate',
1831 * migrate the tasks pages to the new memory.
1833 * Call with cpuset_mutex held. May take callback_lock during call.
1834 * Will take tasklist_lock, scan tasklist for tasks in cpuset cs,
1835 * lock each such tasks mm->mmap_lock, scan its vma's and rebind
1836 * their mempolicies to the cpusets new mems_allowed.
1838 static int update_nodemask(struct cpuset *cs, struct cpuset *trialcs,
1844 * top_cpuset.mems_allowed tracks node_stats[N_MEMORY];
1847 if (cs == &top_cpuset) {
1853 * An empty mems_allowed is ok iff there are no tasks in the cpuset.
1854 * Since nodelist_parse() fails on an empty mask, we special case
1855 * that parsing. The validate_change() call ensures that cpusets
1856 * with tasks have memory.
1859 nodes_clear(trialcs->mems_allowed);
1861 retval = nodelist_parse(buf, trialcs->mems_allowed);
1865 if (!nodes_subset(trialcs->mems_allowed,
1866 top_cpuset.mems_allowed)) {
1872 if (nodes_equal(cs->mems_allowed, trialcs->mems_allowed)) {
1873 retval = 0; /* Too easy - nothing to do */
1876 retval = validate_change(cs, trialcs);
1880 spin_lock_irq(&callback_lock);
1881 cs->mems_allowed = trialcs->mems_allowed;
1882 spin_unlock_irq(&callback_lock);
1884 /* use trialcs->mems_allowed as a temp variable */
1885 update_nodemasks_hier(cs, &trialcs->mems_allowed);
1890 bool current_cpuset_is_being_rebound(void)
1895 ret = task_cs(current) == cpuset_being_rebound;
1901 static int update_relax_domain_level(struct cpuset *cs, s64 val)
1904 if (val < -1 || val >= sched_domain_level_max)
1908 if (val != cs->relax_domain_level) {
1909 cs->relax_domain_level = val;
1910 if (!cpumask_empty(cs->cpus_allowed) &&
1911 is_sched_load_balance(cs))
1912 rebuild_sched_domains_locked();
1919 * update_tasks_flags - update the spread flags of tasks in the cpuset.
1920 * @cs: the cpuset in which each task's spread flags needs to be changed
1922 * Iterate through each task of @cs updating its spread flags. As this
1923 * function is called with cpuset_mutex held, cpuset membership stays
1926 static void update_tasks_flags(struct cpuset *cs)
1928 struct css_task_iter it;
1929 struct task_struct *task;
1931 css_task_iter_start(&cs->css, 0, &it);
1932 while ((task = css_task_iter_next(&it)))
1933 cpuset_update_task_spread_flag(cs, task);
1934 css_task_iter_end(&it);
1938 * update_flag - read a 0 or a 1 in a file and update associated flag
1939 * bit: the bit to update (see cpuset_flagbits_t)
1940 * cs: the cpuset to update
1941 * turning_on: whether the flag is being set or cleared
1943 * Call with cpuset_mutex held.
1946 static int update_flag(cpuset_flagbits_t bit, struct cpuset *cs,
1949 struct cpuset *trialcs;
1950 int balance_flag_changed;
1951 int spread_flag_changed;
1954 trialcs = alloc_trial_cpuset(cs);
1959 set_bit(bit, &trialcs->flags);
1961 clear_bit(bit, &trialcs->flags);
1963 err = validate_change(cs, trialcs);
1967 balance_flag_changed = (is_sched_load_balance(cs) !=
1968 is_sched_load_balance(trialcs));
1970 spread_flag_changed = ((is_spread_slab(cs) != is_spread_slab(trialcs))
1971 || (is_spread_page(cs) != is_spread_page(trialcs)));
1973 spin_lock_irq(&callback_lock);
1974 cs->flags = trialcs->flags;
1975 spin_unlock_irq(&callback_lock);
1977 if (!cpumask_empty(trialcs->cpus_allowed) && balance_flag_changed)
1978 rebuild_sched_domains_locked();
1980 if (spread_flag_changed)
1981 update_tasks_flags(cs);
1983 free_cpuset(trialcs);
1988 * update_prstate - update partititon_root_state
1989 * cs: the cpuset to update
1990 * new_prs: new partition root state
1992 * Call with cpuset_mutex held.
1994 static int update_prstate(struct cpuset *cs, int new_prs)
1996 int err, old_prs = cs->partition_root_state;
1997 struct cpuset *parent = parent_cs(cs);
1998 struct tmpmasks tmpmask;
2000 if (old_prs == new_prs)
2004 * Cannot force a partial or invalid partition root to a full
2007 if (new_prs && (old_prs == PRS_ERROR))
2010 if (alloc_cpumasks(NULL, &tmpmask))
2016 * Turning on partition root requires setting the
2017 * CS_CPU_EXCLUSIVE bit implicitly as well and cpus_allowed
2020 if (cpumask_empty(cs->cpus_allowed))
2023 err = update_flag(CS_CPU_EXCLUSIVE, cs, 1);
2027 err = update_parent_subparts_cpumask(cs, partcmd_enable,
2030 update_flag(CS_CPU_EXCLUSIVE, cs, 0);
2035 * Turning off partition root will clear the
2036 * CS_CPU_EXCLUSIVE bit.
2038 if (old_prs == PRS_ERROR) {
2039 update_flag(CS_CPU_EXCLUSIVE, cs, 0);
2044 err = update_parent_subparts_cpumask(cs, partcmd_disable,
2049 /* Turning off CS_CPU_EXCLUSIVE will not return error */
2050 update_flag(CS_CPU_EXCLUSIVE, cs, 0);
2053 update_tasks_cpumask(parent);
2055 if (parent->child_ecpus_count)
2056 update_sibling_cpumasks(parent, cs, &tmpmask);
2058 rebuild_sched_domains_locked();
2061 spin_lock_irq(&callback_lock);
2062 cs->partition_root_state = new_prs;
2063 spin_unlock_irq(&callback_lock);
2066 free_cpumasks(NULL, &tmpmask);
2071 * Frequency meter - How fast is some event occurring?
2073 * These routines manage a digitally filtered, constant time based,
2074 * event frequency meter. There are four routines:
2075 * fmeter_init() - initialize a frequency meter.
2076 * fmeter_markevent() - called each time the event happens.
2077 * fmeter_getrate() - returns the recent rate of such events.
2078 * fmeter_update() - internal routine used to update fmeter.
2080 * A common data structure is passed to each of these routines,
2081 * which is used to keep track of the state required to manage the
2082 * frequency meter and its digital filter.
2084 * The filter works on the number of events marked per unit time.
2085 * The filter is single-pole low-pass recursive (IIR). The time unit
2086 * is 1 second. Arithmetic is done using 32-bit integers scaled to
2087 * simulate 3 decimal digits of precision (multiplied by 1000).
2089 * With an FM_COEF of 933, and a time base of 1 second, the filter
2090 * has a half-life of 10 seconds, meaning that if the events quit
2091 * happening, then the rate returned from the fmeter_getrate()
2092 * will be cut in half each 10 seconds, until it converges to zero.
2094 * It is not worth doing a real infinitely recursive filter. If more
2095 * than FM_MAXTICKS ticks have elapsed since the last filter event,
2096 * just compute FM_MAXTICKS ticks worth, by which point the level
2099 * Limit the count of unprocessed events to FM_MAXCNT, so as to avoid
2100 * arithmetic overflow in the fmeter_update() routine.
2102 * Given the simple 32 bit integer arithmetic used, this meter works
2103 * best for reporting rates between one per millisecond (msec) and
2104 * one per 32 (approx) seconds. At constant rates faster than one
2105 * per msec it maxes out at values just under 1,000,000. At constant
2106 * rates between one per msec, and one per second it will stabilize
2107 * to a value N*1000, where N is the rate of events per second.
2108 * At constant rates between one per second and one per 32 seconds,
2109 * it will be choppy, moving up on the seconds that have an event,
2110 * and then decaying until the next event. At rates slower than
2111 * about one in 32 seconds, it decays all the way back to zero between
2115 #define FM_COEF 933 /* coefficient for half-life of 10 secs */
2116 #define FM_MAXTICKS ((u32)99) /* useless computing more ticks than this */
2117 #define FM_MAXCNT 1000000 /* limit cnt to avoid overflow */
2118 #define FM_SCALE 1000 /* faux fixed point scale */
2120 /* Initialize a frequency meter */
2121 static void fmeter_init(struct fmeter *fmp)
2126 spin_lock_init(&fmp->lock);
2129 /* Internal meter update - process cnt events and update value */
2130 static void fmeter_update(struct fmeter *fmp)
2135 now = ktime_get_seconds();
2136 ticks = now - fmp->time;
2141 ticks = min(FM_MAXTICKS, ticks);
2143 fmp->val = (FM_COEF * fmp->val) / FM_SCALE;
2146 fmp->val += ((FM_SCALE - FM_COEF) * fmp->cnt) / FM_SCALE;
2150 /* Process any previous ticks, then bump cnt by one (times scale). */
2151 static void fmeter_markevent(struct fmeter *fmp)
2153 spin_lock(&fmp->lock);
2155 fmp->cnt = min(FM_MAXCNT, fmp->cnt + FM_SCALE);
2156 spin_unlock(&fmp->lock);
2159 /* Process any previous ticks, then return current value. */
2160 static int fmeter_getrate(struct fmeter *fmp)
2164 spin_lock(&fmp->lock);
2167 spin_unlock(&fmp->lock);
2171 static struct cpuset *cpuset_attach_old_cs;
2173 static void reset_migrate_dl_data(struct cpuset *cs)
2175 cs->nr_migrate_dl_tasks = 0;
2176 cs->sum_migrate_dl_bw = 0;
2179 /* Called by cgroups to determine if a cpuset is usable; cpuset_mutex held */
2180 static int cpuset_can_attach(struct cgroup_taskset *tset)
2182 struct cgroup_subsys_state *css;
2183 struct cpuset *cs, *oldcs;
2184 struct task_struct *task;
2187 /* used later by cpuset_attach() */
2188 cpuset_attach_old_cs = task_cs(cgroup_taskset_first(tset, &css));
2189 oldcs = cpuset_attach_old_cs;
2192 mutex_lock(&cpuset_mutex);
2194 /* allow moving tasks into an empty cpuset if on default hierarchy */
2196 if (!is_in_v2_mode() &&
2197 (cpumask_empty(cs->cpus_allowed) || nodes_empty(cs->mems_allowed)))
2200 cgroup_taskset_for_each(task, css, tset) {
2201 ret = task_can_attach(task);
2204 ret = security_task_setscheduler(task);
2208 if (dl_task(task)) {
2209 cs->nr_migrate_dl_tasks++;
2210 cs->sum_migrate_dl_bw += task->dl.dl_bw;
2214 if (!cs->nr_migrate_dl_tasks)
2217 if (!cpumask_intersects(oldcs->effective_cpus, cs->effective_cpus)) {
2218 int cpu = cpumask_any_and(cpu_active_mask, cs->effective_cpus);
2220 if (unlikely(cpu >= nr_cpu_ids)) {
2221 reset_migrate_dl_data(cs);
2226 ret = dl_bw_alloc(cpu, cs->sum_migrate_dl_bw);
2228 reset_migrate_dl_data(cs);
2235 * Mark attach is in progress. This makes validate_change() fail
2236 * changes which zero cpus/mems_allowed.
2238 cs->attach_in_progress++;
2241 mutex_unlock(&cpuset_mutex);
2245 static void cpuset_cancel_attach(struct cgroup_taskset *tset)
2247 struct cgroup_subsys_state *css;
2250 cgroup_taskset_first(tset, &css);
2253 mutex_lock(&cpuset_mutex);
2254 cs->attach_in_progress--;
2255 if (!cs->attach_in_progress)
2256 wake_up(&cpuset_attach_wq);
2258 if (cs->nr_migrate_dl_tasks) {
2259 int cpu = cpumask_any(cs->effective_cpus);
2261 dl_bw_free(cpu, cs->sum_migrate_dl_bw);
2262 reset_migrate_dl_data(cs);
2265 mutex_unlock(&cpuset_mutex);
2269 * Protected by cpuset_mutex. cpus_attach is used only by cpuset_attach()
2270 * but we can't allocate it dynamically there. Define it global and
2271 * allocate from cpuset_init().
2273 static cpumask_var_t cpus_attach;
2275 static void cpuset_attach(struct cgroup_taskset *tset)
2277 /* static buf protected by cpuset_mutex */
2278 static nodemask_t cpuset_attach_nodemask_to;
2279 struct task_struct *task;
2280 struct task_struct *leader;
2281 struct cgroup_subsys_state *css;
2283 struct cpuset *oldcs = cpuset_attach_old_cs;
2285 cgroup_taskset_first(tset, &css);
2288 lockdep_assert_cpus_held(); /* see cgroup_attach_lock() */
2289 mutex_lock(&cpuset_mutex);
2291 /* prepare for attach */
2292 if (cs == &top_cpuset)
2293 cpumask_copy(cpus_attach, cpu_possible_mask);
2295 guarantee_online_cpus(cs, cpus_attach);
2297 guarantee_online_mems(cs, &cpuset_attach_nodemask_to);
2299 cgroup_taskset_for_each(task, css, tset) {
2301 * can_attach beforehand should guarantee that this doesn't
2302 * fail. TODO: have a better way to handle failure here
2304 WARN_ON_ONCE(set_cpus_allowed_ptr(task, cpus_attach));
2306 cpuset_change_task_nodemask(task, &cpuset_attach_nodemask_to);
2307 cpuset_update_task_spread_flag(cs, task);
2311 * Change mm for all threadgroup leaders. This is expensive and may
2312 * sleep and should be moved outside migration path proper.
2314 cpuset_attach_nodemask_to = cs->effective_mems;
2315 cgroup_taskset_for_each_leader(leader, css, tset) {
2316 struct mm_struct *mm = get_task_mm(leader);
2319 mpol_rebind_mm(mm, &cpuset_attach_nodemask_to);
2322 * old_mems_allowed is the same with mems_allowed
2323 * here, except if this task is being moved
2324 * automatically due to hotplug. In that case
2325 * @mems_allowed has been updated and is empty, so
2326 * @old_mems_allowed is the right nodesets that we
2329 if (is_memory_migrate(cs))
2330 cpuset_migrate_mm(mm, &oldcs->old_mems_allowed,
2331 &cpuset_attach_nodemask_to);
2337 cs->old_mems_allowed = cpuset_attach_nodemask_to;
2339 if (cs->nr_migrate_dl_tasks) {
2340 cs->nr_deadline_tasks += cs->nr_migrate_dl_tasks;
2341 oldcs->nr_deadline_tasks -= cs->nr_migrate_dl_tasks;
2342 reset_migrate_dl_data(cs);
2345 cs->attach_in_progress--;
2346 if (!cs->attach_in_progress)
2347 wake_up(&cpuset_attach_wq);
2349 mutex_unlock(&cpuset_mutex);
2352 /* The various types of files and directories in a cpuset file system */
2355 FILE_MEMORY_MIGRATE,
2358 FILE_EFFECTIVE_CPULIST,
2359 FILE_EFFECTIVE_MEMLIST,
2360 FILE_SUBPARTS_CPULIST,
2364 FILE_SCHED_LOAD_BALANCE,
2365 FILE_PARTITION_ROOT,
2366 FILE_SCHED_RELAX_DOMAIN_LEVEL,
2367 FILE_MEMORY_PRESSURE_ENABLED,
2368 FILE_MEMORY_PRESSURE,
2371 } cpuset_filetype_t;
2373 static int cpuset_write_u64(struct cgroup_subsys_state *css, struct cftype *cft,
2376 struct cpuset *cs = css_cs(css);
2377 cpuset_filetype_t type = cft->private;
2381 mutex_lock(&cpuset_mutex);
2382 if (!is_cpuset_online(cs)) {
2388 case FILE_CPU_EXCLUSIVE:
2389 retval = update_flag(CS_CPU_EXCLUSIVE, cs, val);
2391 case FILE_MEM_EXCLUSIVE:
2392 retval = update_flag(CS_MEM_EXCLUSIVE, cs, val);
2394 case FILE_MEM_HARDWALL:
2395 retval = update_flag(CS_MEM_HARDWALL, cs, val);
2397 case FILE_SCHED_LOAD_BALANCE:
2398 retval = update_flag(CS_SCHED_LOAD_BALANCE, cs, val);
2400 case FILE_MEMORY_MIGRATE:
2401 retval = update_flag(CS_MEMORY_MIGRATE, cs, val);
2403 case FILE_MEMORY_PRESSURE_ENABLED:
2404 cpuset_memory_pressure_enabled = !!val;
2406 case FILE_SPREAD_PAGE:
2407 retval = update_flag(CS_SPREAD_PAGE, cs, val);
2409 case FILE_SPREAD_SLAB:
2410 retval = update_flag(CS_SPREAD_SLAB, cs, val);
2417 mutex_unlock(&cpuset_mutex);
2422 static int cpuset_write_s64(struct cgroup_subsys_state *css, struct cftype *cft,
2425 struct cpuset *cs = css_cs(css);
2426 cpuset_filetype_t type = cft->private;
2427 int retval = -ENODEV;
2430 mutex_lock(&cpuset_mutex);
2431 if (!is_cpuset_online(cs))
2435 case FILE_SCHED_RELAX_DOMAIN_LEVEL:
2436 retval = update_relax_domain_level(cs, val);
2443 mutex_unlock(&cpuset_mutex);
2449 * Common handling for a write to a "cpus" or "mems" file.
2451 static ssize_t cpuset_write_resmask(struct kernfs_open_file *of,
2452 char *buf, size_t nbytes, loff_t off)
2454 struct cpuset *cs = css_cs(of_css(of));
2455 struct cpuset *trialcs;
2456 int retval = -ENODEV;
2458 buf = strstrip(buf);
2461 * CPU or memory hotunplug may leave @cs w/o any execution
2462 * resources, in which case the hotplug code asynchronously updates
2463 * configuration and transfers all tasks to the nearest ancestor
2464 * which can execute.
2466 * As writes to "cpus" or "mems" may restore @cs's execution
2467 * resources, wait for the previously scheduled operations before
2468 * proceeding, so that we don't end up keep removing tasks added
2469 * after execution capability is restored.
2471 * cpuset_hotplug_work calls back into cgroup core via
2472 * cgroup_transfer_tasks() and waiting for it from a cgroupfs
2473 * operation like this one can lead to a deadlock through kernfs
2474 * active_ref protection. Let's break the protection. Losing the
2475 * protection is okay as we check whether @cs is online after
2476 * grabbing cpuset_mutex anyway. This only happens on the legacy
2480 kernfs_break_active_protection(of->kn);
2481 flush_work(&cpuset_hotplug_work);
2484 mutex_lock(&cpuset_mutex);
2485 if (!is_cpuset_online(cs))
2488 trialcs = alloc_trial_cpuset(cs);
2494 switch (of_cft(of)->private) {
2496 retval = update_cpumask(cs, trialcs, buf);
2499 retval = update_nodemask(cs, trialcs, buf);
2506 free_cpuset(trialcs);
2508 mutex_unlock(&cpuset_mutex);
2510 kernfs_unbreak_active_protection(of->kn);
2512 flush_workqueue(cpuset_migrate_mm_wq);
2513 return retval ?: nbytes;
2517 * These ascii lists should be read in a single call, by using a user
2518 * buffer large enough to hold the entire map. If read in smaller
2519 * chunks, there is no guarantee of atomicity. Since the display format
2520 * used, list of ranges of sequential numbers, is variable length,
2521 * and since these maps can change value dynamically, one could read
2522 * gibberish by doing partial reads while a list was changing.
2524 static int cpuset_common_seq_show(struct seq_file *sf, void *v)
2526 struct cpuset *cs = css_cs(seq_css(sf));
2527 cpuset_filetype_t type = seq_cft(sf)->private;
2530 spin_lock_irq(&callback_lock);
2534 seq_printf(sf, "%*pbl\n", cpumask_pr_args(cs->cpus_allowed));
2537 seq_printf(sf, "%*pbl\n", nodemask_pr_args(&cs->mems_allowed));
2539 case FILE_EFFECTIVE_CPULIST:
2540 seq_printf(sf, "%*pbl\n", cpumask_pr_args(cs->effective_cpus));
2542 case FILE_EFFECTIVE_MEMLIST:
2543 seq_printf(sf, "%*pbl\n", nodemask_pr_args(&cs->effective_mems));
2545 case FILE_SUBPARTS_CPULIST:
2546 seq_printf(sf, "%*pbl\n", cpumask_pr_args(cs->subparts_cpus));
2552 spin_unlock_irq(&callback_lock);
2556 static u64 cpuset_read_u64(struct cgroup_subsys_state *css, struct cftype *cft)
2558 struct cpuset *cs = css_cs(css);
2559 cpuset_filetype_t type = cft->private;
2561 case FILE_CPU_EXCLUSIVE:
2562 return is_cpu_exclusive(cs);
2563 case FILE_MEM_EXCLUSIVE:
2564 return is_mem_exclusive(cs);
2565 case FILE_MEM_HARDWALL:
2566 return is_mem_hardwall(cs);
2567 case FILE_SCHED_LOAD_BALANCE:
2568 return is_sched_load_balance(cs);
2569 case FILE_MEMORY_MIGRATE:
2570 return is_memory_migrate(cs);
2571 case FILE_MEMORY_PRESSURE_ENABLED:
2572 return cpuset_memory_pressure_enabled;
2573 case FILE_MEMORY_PRESSURE:
2574 return fmeter_getrate(&cs->fmeter);
2575 case FILE_SPREAD_PAGE:
2576 return is_spread_page(cs);
2577 case FILE_SPREAD_SLAB:
2578 return is_spread_slab(cs);
2583 /* Unreachable but makes gcc happy */
2587 static s64 cpuset_read_s64(struct cgroup_subsys_state *css, struct cftype *cft)
2589 struct cpuset *cs = css_cs(css);
2590 cpuset_filetype_t type = cft->private;
2592 case FILE_SCHED_RELAX_DOMAIN_LEVEL:
2593 return cs->relax_domain_level;
2598 /* Unrechable but makes gcc happy */
2602 static int sched_partition_show(struct seq_file *seq, void *v)
2604 struct cpuset *cs = css_cs(seq_css(seq));
2606 switch (cs->partition_root_state) {
2608 seq_puts(seq, "root\n");
2611 seq_puts(seq, "member\n");
2614 seq_puts(seq, "root invalid\n");
2620 static ssize_t sched_partition_write(struct kernfs_open_file *of, char *buf,
2621 size_t nbytes, loff_t off)
2623 struct cpuset *cs = css_cs(of_css(of));
2625 int retval = -ENODEV;
2627 buf = strstrip(buf);
2630 * Convert "root" to ENABLED, and convert "member" to DISABLED.
2632 if (!strcmp(buf, "root"))
2634 else if (!strcmp(buf, "member"))
2641 mutex_lock(&cpuset_mutex);
2642 if (!is_cpuset_online(cs))
2645 retval = update_prstate(cs, val);
2647 mutex_unlock(&cpuset_mutex);
2650 return retval ?: nbytes;
2654 * for the common functions, 'private' gives the type of file
2657 static struct cftype legacy_files[] = {
2660 .seq_show = cpuset_common_seq_show,
2661 .write = cpuset_write_resmask,
2662 .max_write_len = (100U + 6 * NR_CPUS),
2663 .private = FILE_CPULIST,
2668 .seq_show = cpuset_common_seq_show,
2669 .write = cpuset_write_resmask,
2670 .max_write_len = (100U + 6 * MAX_NUMNODES),
2671 .private = FILE_MEMLIST,
2675 .name = "effective_cpus",
2676 .seq_show = cpuset_common_seq_show,
2677 .private = FILE_EFFECTIVE_CPULIST,
2681 .name = "effective_mems",
2682 .seq_show = cpuset_common_seq_show,
2683 .private = FILE_EFFECTIVE_MEMLIST,
2687 .name = "cpu_exclusive",
2688 .read_u64 = cpuset_read_u64,
2689 .write_u64 = cpuset_write_u64,
2690 .private = FILE_CPU_EXCLUSIVE,
2694 .name = "mem_exclusive",
2695 .read_u64 = cpuset_read_u64,
2696 .write_u64 = cpuset_write_u64,
2697 .private = FILE_MEM_EXCLUSIVE,
2701 .name = "mem_hardwall",
2702 .read_u64 = cpuset_read_u64,
2703 .write_u64 = cpuset_write_u64,
2704 .private = FILE_MEM_HARDWALL,
2708 .name = "sched_load_balance",
2709 .read_u64 = cpuset_read_u64,
2710 .write_u64 = cpuset_write_u64,
2711 .private = FILE_SCHED_LOAD_BALANCE,
2715 .name = "sched_relax_domain_level",
2716 .read_s64 = cpuset_read_s64,
2717 .write_s64 = cpuset_write_s64,
2718 .private = FILE_SCHED_RELAX_DOMAIN_LEVEL,
2722 .name = "memory_migrate",
2723 .read_u64 = cpuset_read_u64,
2724 .write_u64 = cpuset_write_u64,
2725 .private = FILE_MEMORY_MIGRATE,
2729 .name = "memory_pressure",
2730 .read_u64 = cpuset_read_u64,
2731 .private = FILE_MEMORY_PRESSURE,
2735 .name = "memory_spread_page",
2736 .read_u64 = cpuset_read_u64,
2737 .write_u64 = cpuset_write_u64,
2738 .private = FILE_SPREAD_PAGE,
2742 .name = "memory_spread_slab",
2743 .read_u64 = cpuset_read_u64,
2744 .write_u64 = cpuset_write_u64,
2745 .private = FILE_SPREAD_SLAB,
2749 .name = "memory_pressure_enabled",
2750 .flags = CFTYPE_ONLY_ON_ROOT,
2751 .read_u64 = cpuset_read_u64,
2752 .write_u64 = cpuset_write_u64,
2753 .private = FILE_MEMORY_PRESSURE_ENABLED,
2760 * This is currently a minimal set for the default hierarchy. It can be
2761 * expanded later on by migrating more features and control files from v1.
2763 static struct cftype dfl_files[] = {
2766 .seq_show = cpuset_common_seq_show,
2767 .write = cpuset_write_resmask,
2768 .max_write_len = (100U + 6 * NR_CPUS),
2769 .private = FILE_CPULIST,
2770 .flags = CFTYPE_NOT_ON_ROOT,
2775 .seq_show = cpuset_common_seq_show,
2776 .write = cpuset_write_resmask,
2777 .max_write_len = (100U + 6 * MAX_NUMNODES),
2778 .private = FILE_MEMLIST,
2779 .flags = CFTYPE_NOT_ON_ROOT,
2783 .name = "cpus.effective",
2784 .seq_show = cpuset_common_seq_show,
2785 .private = FILE_EFFECTIVE_CPULIST,
2789 .name = "mems.effective",
2790 .seq_show = cpuset_common_seq_show,
2791 .private = FILE_EFFECTIVE_MEMLIST,
2795 .name = "cpus.partition",
2796 .seq_show = sched_partition_show,
2797 .write = sched_partition_write,
2798 .private = FILE_PARTITION_ROOT,
2799 .flags = CFTYPE_NOT_ON_ROOT,
2803 .name = "cpus.subpartitions",
2804 .seq_show = cpuset_common_seq_show,
2805 .private = FILE_SUBPARTS_CPULIST,
2806 .flags = CFTYPE_DEBUG,
2814 * cpuset_css_alloc - allocate a cpuset css
2815 * cgrp: control group that the new cpuset will be part of
2818 static struct cgroup_subsys_state *
2819 cpuset_css_alloc(struct cgroup_subsys_state *parent_css)
2824 return &top_cpuset.css;
2826 cs = kzalloc(sizeof(*cs), GFP_KERNEL);
2828 return ERR_PTR(-ENOMEM);
2830 if (alloc_cpumasks(cs, NULL)) {
2832 return ERR_PTR(-ENOMEM);
2835 set_bit(CS_SCHED_LOAD_BALANCE, &cs->flags);
2836 nodes_clear(cs->mems_allowed);
2837 nodes_clear(cs->effective_mems);
2838 fmeter_init(&cs->fmeter);
2839 cs->relax_domain_level = -1;
2844 static int cpuset_css_online(struct cgroup_subsys_state *css)
2846 struct cpuset *cs = css_cs(css);
2847 struct cpuset *parent = parent_cs(cs);
2848 struct cpuset *tmp_cs;
2849 struct cgroup_subsys_state *pos_css;
2855 mutex_lock(&cpuset_mutex);
2857 set_bit(CS_ONLINE, &cs->flags);
2858 if (is_spread_page(parent))
2859 set_bit(CS_SPREAD_PAGE, &cs->flags);
2860 if (is_spread_slab(parent))
2861 set_bit(CS_SPREAD_SLAB, &cs->flags);
2865 spin_lock_irq(&callback_lock);
2866 if (is_in_v2_mode()) {
2867 cpumask_copy(cs->effective_cpus, parent->effective_cpus);
2868 cs->effective_mems = parent->effective_mems;
2869 cs->use_parent_ecpus = true;
2870 parent->child_ecpus_count++;
2872 spin_unlock_irq(&callback_lock);
2874 if (!test_bit(CGRP_CPUSET_CLONE_CHILDREN, &css->cgroup->flags))
2878 * Clone @parent's configuration if CGRP_CPUSET_CLONE_CHILDREN is
2879 * set. This flag handling is implemented in cgroup core for
2880 * histrical reasons - the flag may be specified during mount.
2882 * Currently, if any sibling cpusets have exclusive cpus or mem, we
2883 * refuse to clone the configuration - thereby refusing the task to
2884 * be entered, and as a result refusing the sys_unshare() or
2885 * clone() which initiated it. If this becomes a problem for some
2886 * users who wish to allow that scenario, then this could be
2887 * changed to grant parent->cpus_allowed-sibling_cpus_exclusive
2888 * (and likewise for mems) to the new cgroup.
2891 cpuset_for_each_child(tmp_cs, pos_css, parent) {
2892 if (is_mem_exclusive(tmp_cs) || is_cpu_exclusive(tmp_cs)) {
2899 spin_lock_irq(&callback_lock);
2900 cs->mems_allowed = parent->mems_allowed;
2901 cs->effective_mems = parent->mems_allowed;
2902 cpumask_copy(cs->cpus_allowed, parent->cpus_allowed);
2903 cpumask_copy(cs->effective_cpus, parent->cpus_allowed);
2904 spin_unlock_irq(&callback_lock);
2906 mutex_unlock(&cpuset_mutex);
2912 * If the cpuset being removed has its flag 'sched_load_balance'
2913 * enabled, then simulate turning sched_load_balance off, which
2914 * will call rebuild_sched_domains_locked(). That is not needed
2915 * in the default hierarchy where only changes in partition
2916 * will cause repartitioning.
2918 * If the cpuset has the 'sched.partition' flag enabled, simulate
2919 * turning 'sched.partition" off.
2922 static void cpuset_css_offline(struct cgroup_subsys_state *css)
2924 struct cpuset *cs = css_cs(css);
2927 mutex_lock(&cpuset_mutex);
2929 if (is_partition_root(cs))
2930 update_prstate(cs, 0);
2932 if (!cgroup_subsys_on_dfl(cpuset_cgrp_subsys) &&
2933 is_sched_load_balance(cs))
2934 update_flag(CS_SCHED_LOAD_BALANCE, cs, 0);
2936 if (cs->use_parent_ecpus) {
2937 struct cpuset *parent = parent_cs(cs);
2939 cs->use_parent_ecpus = false;
2940 parent->child_ecpus_count--;
2944 clear_bit(CS_ONLINE, &cs->flags);
2946 mutex_unlock(&cpuset_mutex);
2950 static void cpuset_css_free(struct cgroup_subsys_state *css)
2952 struct cpuset *cs = css_cs(css);
2957 static void cpuset_bind(struct cgroup_subsys_state *root_css)
2959 mutex_lock(&cpuset_mutex);
2960 spin_lock_irq(&callback_lock);
2962 if (is_in_v2_mode()) {
2963 cpumask_copy(top_cpuset.cpus_allowed, cpu_possible_mask);
2964 top_cpuset.mems_allowed = node_possible_map;
2966 cpumask_copy(top_cpuset.cpus_allowed,
2967 top_cpuset.effective_cpus);
2968 top_cpuset.mems_allowed = top_cpuset.effective_mems;
2971 spin_unlock_irq(&callback_lock);
2972 mutex_unlock(&cpuset_mutex);
2976 * Make sure the new task conform to the current state of its parent,
2977 * which could have been changed by cpuset just after it inherits the
2978 * state from the parent and before it sits on the cgroup's task list.
2980 static void cpuset_fork(struct task_struct *task)
2982 if (task_css_is_root(task, cpuset_cgrp_id))
2985 set_cpus_allowed_ptr(task, current->cpus_ptr);
2986 task->mems_allowed = current->mems_allowed;
2989 struct cgroup_subsys cpuset_cgrp_subsys = {
2990 .css_alloc = cpuset_css_alloc,
2991 .css_online = cpuset_css_online,
2992 .css_offline = cpuset_css_offline,
2993 .css_free = cpuset_css_free,
2994 .can_attach = cpuset_can_attach,
2995 .cancel_attach = cpuset_cancel_attach,
2996 .attach = cpuset_attach,
2997 .post_attach = cpuset_post_attach,
2998 .bind = cpuset_bind,
2999 .fork = cpuset_fork,
3000 .legacy_cftypes = legacy_files,
3001 .dfl_cftypes = dfl_files,
3007 * cpuset_init - initialize cpusets at system boot
3009 * Description: Initialize top_cpuset
3012 int __init cpuset_init(void)
3014 BUG_ON(!alloc_cpumask_var(&top_cpuset.cpus_allowed, GFP_KERNEL));
3015 BUG_ON(!alloc_cpumask_var(&top_cpuset.effective_cpus, GFP_KERNEL));
3016 BUG_ON(!zalloc_cpumask_var(&top_cpuset.subparts_cpus, GFP_KERNEL));
3018 cpumask_setall(top_cpuset.cpus_allowed);
3019 nodes_setall(top_cpuset.mems_allowed);
3020 cpumask_setall(top_cpuset.effective_cpus);
3021 nodes_setall(top_cpuset.effective_mems);
3023 fmeter_init(&top_cpuset.fmeter);
3024 set_bit(CS_SCHED_LOAD_BALANCE, &top_cpuset.flags);
3025 top_cpuset.relax_domain_level = -1;
3027 BUG_ON(!alloc_cpumask_var(&cpus_attach, GFP_KERNEL));
3033 * If CPU and/or memory hotplug handlers, below, unplug any CPUs
3034 * or memory nodes, we need to walk over the cpuset hierarchy,
3035 * removing that CPU or node from all cpusets. If this removes the
3036 * last CPU or node from a cpuset, then move the tasks in the empty
3037 * cpuset to its next-highest non-empty parent.
3039 static void remove_tasks_in_empty_cpuset(struct cpuset *cs)
3041 struct cpuset *parent;
3044 * Find its next-highest non-empty parent, (top cpuset
3045 * has online cpus, so can't be empty).
3047 parent = parent_cs(cs);
3048 while (cpumask_empty(parent->cpus_allowed) ||
3049 nodes_empty(parent->mems_allowed))
3050 parent = parent_cs(parent);
3052 if (cgroup_transfer_tasks(parent->css.cgroup, cs->css.cgroup)) {
3053 pr_err("cpuset: failed to transfer tasks out of empty cpuset ");
3054 pr_cont_cgroup_name(cs->css.cgroup);
3060 hotplug_update_tasks_legacy(struct cpuset *cs,
3061 struct cpumask *new_cpus, nodemask_t *new_mems,
3062 bool cpus_updated, bool mems_updated)
3066 spin_lock_irq(&callback_lock);
3067 cpumask_copy(cs->cpus_allowed, new_cpus);
3068 cpumask_copy(cs->effective_cpus, new_cpus);
3069 cs->mems_allowed = *new_mems;
3070 cs->effective_mems = *new_mems;
3071 spin_unlock_irq(&callback_lock);
3074 * Don't call update_tasks_cpumask() if the cpuset becomes empty,
3075 * as the tasks will be migratecd to an ancestor.
3077 if (cpus_updated && !cpumask_empty(cs->cpus_allowed))
3078 update_tasks_cpumask(cs);
3079 if (mems_updated && !nodes_empty(cs->mems_allowed))
3080 update_tasks_nodemask(cs);
3082 is_empty = cpumask_empty(cs->cpus_allowed) ||
3083 nodes_empty(cs->mems_allowed);
3085 mutex_unlock(&cpuset_mutex);
3088 * Move tasks to the nearest ancestor with execution resources,
3089 * This is full cgroup operation which will also call back into
3090 * cpuset. Should be done outside any lock.
3093 remove_tasks_in_empty_cpuset(cs);
3095 mutex_lock(&cpuset_mutex);
3099 hotplug_update_tasks(struct cpuset *cs,
3100 struct cpumask *new_cpus, nodemask_t *new_mems,
3101 bool cpus_updated, bool mems_updated)
3103 if (cpumask_empty(new_cpus))
3104 cpumask_copy(new_cpus, parent_cs(cs)->effective_cpus);
3105 if (nodes_empty(*new_mems))
3106 *new_mems = parent_cs(cs)->effective_mems;
3108 spin_lock_irq(&callback_lock);
3109 cpumask_copy(cs->effective_cpus, new_cpus);
3110 cs->effective_mems = *new_mems;
3111 spin_unlock_irq(&callback_lock);
3114 update_tasks_cpumask(cs);
3116 update_tasks_nodemask(cs);
3119 static bool force_rebuild;
3121 void cpuset_force_rebuild(void)
3123 force_rebuild = true;
3127 * cpuset_hotplug_update_tasks - update tasks in a cpuset for hotunplug
3128 * @cs: cpuset in interest
3129 * @tmp: the tmpmasks structure pointer
3131 * Compare @cs's cpu and mem masks against top_cpuset and if some have gone
3132 * offline, update @cs accordingly. If @cs ends up with no CPU or memory,
3133 * all its tasks are moved to the nearest ancestor with both resources.
3135 static void cpuset_hotplug_update_tasks(struct cpuset *cs, struct tmpmasks *tmp)
3137 static cpumask_t new_cpus;
3138 static nodemask_t new_mems;
3141 struct cpuset *parent;
3143 wait_event(cpuset_attach_wq, cs->attach_in_progress == 0);
3145 mutex_lock(&cpuset_mutex);
3148 * We have raced with task attaching. We wait until attaching
3149 * is finished, so we won't attach a task to an empty cpuset.
3151 if (cs->attach_in_progress) {
3152 mutex_unlock(&cpuset_mutex);
3156 parent = parent_cs(cs);
3157 compute_effective_cpumask(&new_cpus, cs, parent);
3158 nodes_and(new_mems, cs->mems_allowed, parent->effective_mems);
3160 if (cs->nr_subparts_cpus)
3162 * Make sure that CPUs allocated to child partitions
3163 * do not show up in effective_cpus.
3165 cpumask_andnot(&new_cpus, &new_cpus, cs->subparts_cpus);
3167 if (!tmp || !cs->partition_root_state)
3171 * In the unlikely event that a partition root has empty
3172 * effective_cpus or its parent becomes erroneous, we have to
3173 * transition it to the erroneous state.
3175 if (is_partition_root(cs) && (cpumask_empty(&new_cpus) ||
3176 (parent->partition_root_state == PRS_ERROR))) {
3177 if (cs->nr_subparts_cpus) {
3178 spin_lock_irq(&callback_lock);
3179 cs->nr_subparts_cpus = 0;
3180 cpumask_clear(cs->subparts_cpus);
3181 spin_unlock_irq(&callback_lock);
3182 compute_effective_cpumask(&new_cpus, cs, parent);
3186 * If the effective_cpus is empty because the child
3187 * partitions take away all the CPUs, we can keep
3188 * the current partition and let the child partitions
3189 * fight for available CPUs.
3191 if ((parent->partition_root_state == PRS_ERROR) ||
3192 cpumask_empty(&new_cpus)) {
3193 update_parent_subparts_cpumask(cs, partcmd_disable,
3195 spin_lock_irq(&callback_lock);
3196 cs->partition_root_state = PRS_ERROR;
3197 spin_unlock_irq(&callback_lock);
3199 cpuset_force_rebuild();
3203 * On the other hand, an erroneous partition root may be transitioned
3204 * back to a regular one or a partition root with no CPU allocated
3205 * from the parent may change to erroneous.
3207 if (is_partition_root(parent) &&
3208 ((cs->partition_root_state == PRS_ERROR) ||
3209 !cpumask_intersects(&new_cpus, parent->subparts_cpus)) &&
3210 update_parent_subparts_cpumask(cs, partcmd_update, NULL, tmp))
3211 cpuset_force_rebuild();
3214 cpus_updated = !cpumask_equal(&new_cpus, cs->effective_cpus);
3215 mems_updated = !nodes_equal(new_mems, cs->effective_mems);
3217 if (is_in_v2_mode())
3218 hotplug_update_tasks(cs, &new_cpus, &new_mems,
3219 cpus_updated, mems_updated);
3221 hotplug_update_tasks_legacy(cs, &new_cpus, &new_mems,
3222 cpus_updated, mems_updated);
3224 mutex_unlock(&cpuset_mutex);
3228 * cpuset_hotplug_workfn - handle CPU/memory hotunplug for a cpuset
3230 * This function is called after either CPU or memory configuration has
3231 * changed and updates cpuset accordingly. The top_cpuset is always
3232 * synchronized to cpu_active_mask and N_MEMORY, which is necessary in
3233 * order to make cpusets transparent (of no affect) on systems that are
3234 * actively using CPU hotplug but making no active use of cpusets.
3236 * Non-root cpusets are only affected by offlining. If any CPUs or memory
3237 * nodes have been taken down, cpuset_hotplug_update_tasks() is invoked on
3240 * Note that CPU offlining during suspend is ignored. We don't modify
3241 * cpusets across suspend/resume cycles at all.
3243 static void cpuset_hotplug_workfn(struct work_struct *work)
3245 static cpumask_t new_cpus;
3246 static nodemask_t new_mems;
3247 bool cpus_updated, mems_updated;
3248 bool on_dfl = is_in_v2_mode();
3249 struct tmpmasks tmp, *ptmp = NULL;
3251 if (on_dfl && !alloc_cpumasks(NULL, &tmp))
3254 mutex_lock(&cpuset_mutex);
3256 /* fetch the available cpus/mems and find out which changed how */
3257 cpumask_copy(&new_cpus, cpu_active_mask);
3258 new_mems = node_states[N_MEMORY];
3261 * If subparts_cpus is populated, it is likely that the check below
3262 * will produce a false positive on cpus_updated when the cpu list
3263 * isn't changed. It is extra work, but it is better to be safe.
3265 cpus_updated = !cpumask_equal(top_cpuset.effective_cpus, &new_cpus);
3266 mems_updated = !nodes_equal(top_cpuset.effective_mems, new_mems);
3269 * In the rare case that hotplug removes all the cpus in subparts_cpus,
3270 * we assumed that cpus are updated.
3272 if (!cpus_updated && top_cpuset.nr_subparts_cpus)
3273 cpus_updated = true;
3275 /* synchronize cpus_allowed to cpu_active_mask */
3277 spin_lock_irq(&callback_lock);
3279 cpumask_copy(top_cpuset.cpus_allowed, &new_cpus);
3281 * Make sure that CPUs allocated to child partitions
3282 * do not show up in effective_cpus. If no CPU is left,
3283 * we clear the subparts_cpus & let the child partitions
3284 * fight for the CPUs again.
3286 if (top_cpuset.nr_subparts_cpus) {
3287 if (cpumask_subset(&new_cpus,
3288 top_cpuset.subparts_cpus)) {
3289 top_cpuset.nr_subparts_cpus = 0;
3290 cpumask_clear(top_cpuset.subparts_cpus);
3292 cpumask_andnot(&new_cpus, &new_cpus,
3293 top_cpuset.subparts_cpus);
3296 cpumask_copy(top_cpuset.effective_cpus, &new_cpus);
3297 spin_unlock_irq(&callback_lock);
3298 /* we don't mess with cpumasks of tasks in top_cpuset */
3301 /* synchronize mems_allowed to N_MEMORY */
3303 spin_lock_irq(&callback_lock);
3305 top_cpuset.mems_allowed = new_mems;
3306 top_cpuset.effective_mems = new_mems;
3307 spin_unlock_irq(&callback_lock);
3308 update_tasks_nodemask(&top_cpuset);
3311 mutex_unlock(&cpuset_mutex);
3313 /* if cpus or mems changed, we need to propagate to descendants */
3314 if (cpus_updated || mems_updated) {
3316 struct cgroup_subsys_state *pos_css;
3319 cpuset_for_each_descendant_pre(cs, pos_css, &top_cpuset) {
3320 if (cs == &top_cpuset || !css_tryget_online(&cs->css))
3324 cpuset_hotplug_update_tasks(cs, ptmp);
3332 /* rebuild sched domains if cpus_allowed has changed */
3333 if (cpus_updated || force_rebuild) {
3334 force_rebuild = false;
3335 rebuild_sched_domains();
3338 free_cpumasks(NULL, ptmp);
3341 void cpuset_update_active_cpus(void)
3344 * We're inside cpu hotplug critical region which usually nests
3345 * inside cgroup synchronization. Bounce actual hotplug processing
3346 * to a work item to avoid reverse locking order.
3348 schedule_work(&cpuset_hotplug_work);
3351 void cpuset_wait_for_hotplug(void)
3353 flush_work(&cpuset_hotplug_work);
3357 * Keep top_cpuset.mems_allowed tracking node_states[N_MEMORY].
3358 * Call this routine anytime after node_states[N_MEMORY] changes.
3359 * See cpuset_update_active_cpus() for CPU hotplug handling.
3361 static int cpuset_track_online_nodes(struct notifier_block *self,
3362 unsigned long action, void *arg)
3364 schedule_work(&cpuset_hotplug_work);
3368 static struct notifier_block cpuset_track_online_nodes_nb = {
3369 .notifier_call = cpuset_track_online_nodes,
3370 .priority = 10, /* ??! */
3374 * cpuset_init_smp - initialize cpus_allowed
3376 * Description: Finish top cpuset after cpu, node maps are initialized
3378 void __init cpuset_init_smp(void)
3381 * cpus_allowd/mems_allowed set to v2 values in the initial
3382 * cpuset_bind() call will be reset to v1 values in another
3383 * cpuset_bind() call when v1 cpuset is mounted.
3385 top_cpuset.old_mems_allowed = top_cpuset.mems_allowed;
3387 cpumask_copy(top_cpuset.effective_cpus, cpu_active_mask);
3388 top_cpuset.effective_mems = node_states[N_MEMORY];
3390 register_hotmemory_notifier(&cpuset_track_online_nodes_nb);
3392 cpuset_migrate_mm_wq = alloc_ordered_workqueue("cpuset_migrate_mm", 0);
3393 BUG_ON(!cpuset_migrate_mm_wq);
3397 * cpuset_cpus_allowed - return cpus_allowed mask from a tasks cpuset.
3398 * @tsk: pointer to task_struct from which to obtain cpuset->cpus_allowed.
3399 * @pmask: pointer to struct cpumask variable to receive cpus_allowed set.
3401 * Description: Returns the cpumask_var_t cpus_allowed of the cpuset
3402 * attached to the specified @tsk. Guaranteed to return some non-empty
3403 * subset of cpu_online_mask, even if this means going outside the
3407 void cpuset_cpus_allowed(struct task_struct *tsk, struct cpumask *pmask)
3409 unsigned long flags;
3411 spin_lock_irqsave(&callback_lock, flags);
3413 guarantee_online_cpus(task_cs(tsk), pmask);
3415 spin_unlock_irqrestore(&callback_lock, flags);
3419 * cpuset_cpus_allowed_fallback - final fallback before complete catastrophe.
3420 * @tsk: pointer to task_struct with which the scheduler is struggling
3422 * Description: In the case that the scheduler cannot find an allowed cpu in
3423 * tsk->cpus_allowed, we fall back to task_cs(tsk)->cpus_allowed. In legacy
3424 * mode however, this value is the same as task_cs(tsk)->effective_cpus,
3425 * which will not contain a sane cpumask during cases such as cpu hotplugging.
3426 * This is the absolute last resort for the scheduler and it is only used if
3427 * _every_ other avenue has been traveled.
3430 void cpuset_cpus_allowed_fallback(struct task_struct *tsk)
3433 do_set_cpus_allowed(tsk, is_in_v2_mode() ?
3434 task_cs(tsk)->cpus_allowed : cpu_possible_mask);
3438 * We own tsk->cpus_allowed, nobody can change it under us.
3440 * But we used cs && cs->cpus_allowed lockless and thus can
3441 * race with cgroup_attach_task() or update_cpumask() and get
3442 * the wrong tsk->cpus_allowed. However, both cases imply the
3443 * subsequent cpuset_change_cpumask()->set_cpus_allowed_ptr()
3444 * which takes task_rq_lock().
3446 * If we are called after it dropped the lock we must see all
3447 * changes in tsk_cs()->cpus_allowed. Otherwise we can temporary
3448 * set any mask even if it is not right from task_cs() pov,
3449 * the pending set_cpus_allowed_ptr() will fix things.
3451 * select_fallback_rq() will fix things ups and set cpu_possible_mask
3456 void __init cpuset_init_current_mems_allowed(void)
3458 nodes_setall(current->mems_allowed);
3462 * cpuset_mems_allowed - return mems_allowed mask from a tasks cpuset.
3463 * @tsk: pointer to task_struct from which to obtain cpuset->mems_allowed.
3465 * Description: Returns the nodemask_t mems_allowed of the cpuset
3466 * attached to the specified @tsk. Guaranteed to return some non-empty
3467 * subset of node_states[N_MEMORY], even if this means going outside the
3471 nodemask_t cpuset_mems_allowed(struct task_struct *tsk)
3474 unsigned long flags;
3476 spin_lock_irqsave(&callback_lock, flags);
3478 guarantee_online_mems(task_cs(tsk), &mask);
3480 spin_unlock_irqrestore(&callback_lock, flags);
3486 * cpuset_nodemask_valid_mems_allowed - check nodemask vs. curremt mems_allowed
3487 * @nodemask: the nodemask to be checked
3489 * Are any of the nodes in the nodemask allowed in current->mems_allowed?
3491 int cpuset_nodemask_valid_mems_allowed(nodemask_t *nodemask)
3493 return nodes_intersects(*nodemask, current->mems_allowed);
3497 * nearest_hardwall_ancestor() - Returns the nearest mem_exclusive or
3498 * mem_hardwall ancestor to the specified cpuset. Call holding
3499 * callback_lock. If no ancestor is mem_exclusive or mem_hardwall
3500 * (an unusual configuration), then returns the root cpuset.
3502 static struct cpuset *nearest_hardwall_ancestor(struct cpuset *cs)
3504 while (!(is_mem_exclusive(cs) || is_mem_hardwall(cs)) && parent_cs(cs))
3510 * cpuset_node_allowed - Can we allocate on a memory node?
3511 * @node: is this an allowed node?
3512 * @gfp_mask: memory allocation flags
3514 * If we're in interrupt, yes, we can always allocate. If @node is set in
3515 * current's mems_allowed, yes. If it's not a __GFP_HARDWALL request and this
3516 * node is set in the nearest hardwalled cpuset ancestor to current's cpuset,
3517 * yes. If current has access to memory reserves as an oom victim, yes.
3520 * GFP_USER allocations are marked with the __GFP_HARDWALL bit,
3521 * and do not allow allocations outside the current tasks cpuset
3522 * unless the task has been OOM killed.
3523 * GFP_KERNEL allocations are not so marked, so can escape to the
3524 * nearest enclosing hardwalled ancestor cpuset.
3526 * Scanning up parent cpusets requires callback_lock. The
3527 * __alloc_pages() routine only calls here with __GFP_HARDWALL bit
3528 * _not_ set if it's a GFP_KERNEL allocation, and all nodes in the
3529 * current tasks mems_allowed came up empty on the first pass over
3530 * the zonelist. So only GFP_KERNEL allocations, if all nodes in the
3531 * cpuset are short of memory, might require taking the callback_lock.
3533 * The first call here from mm/page_alloc:get_page_from_freelist()
3534 * has __GFP_HARDWALL set in gfp_mask, enforcing hardwall cpusets,
3535 * so no allocation on a node outside the cpuset is allowed (unless
3536 * in interrupt, of course).
3538 * The second pass through get_page_from_freelist() doesn't even call
3539 * here for GFP_ATOMIC calls. For those calls, the __alloc_pages()
3540 * variable 'wait' is not set, and the bit ALLOC_CPUSET is not set
3541 * in alloc_flags. That logic and the checks below have the combined
3543 * in_interrupt - any node ok (current task context irrelevant)
3544 * GFP_ATOMIC - any node ok
3545 * tsk_is_oom_victim - any node ok
3546 * GFP_KERNEL - any node in enclosing hardwalled cpuset ok
3547 * GFP_USER - only nodes in current tasks mems allowed ok.
3549 bool __cpuset_node_allowed(int node, gfp_t gfp_mask)
3551 struct cpuset *cs; /* current cpuset ancestors */
3552 int allowed; /* is allocation in zone z allowed? */
3553 unsigned long flags;
3557 if (node_isset(node, current->mems_allowed))
3560 * Allow tasks that have access to memory reserves because they have
3561 * been OOM killed to get memory anywhere.
3563 if (unlikely(tsk_is_oom_victim(current)))
3565 if (gfp_mask & __GFP_HARDWALL) /* If hardwall request, stop here */
3568 if (current->flags & PF_EXITING) /* Let dying task have memory */
3571 /* Not hardwall and node outside mems_allowed: scan up cpusets */
3572 spin_lock_irqsave(&callback_lock, flags);
3575 cs = nearest_hardwall_ancestor(task_cs(current));
3576 allowed = node_isset(node, cs->mems_allowed);
3579 spin_unlock_irqrestore(&callback_lock, flags);
3584 * cpuset_mem_spread_node() - On which node to begin search for a file page
3585 * cpuset_slab_spread_node() - On which node to begin search for a slab page
3587 * If a task is marked PF_SPREAD_PAGE or PF_SPREAD_SLAB (as for
3588 * tasks in a cpuset with is_spread_page or is_spread_slab set),
3589 * and if the memory allocation used cpuset_mem_spread_node()
3590 * to determine on which node to start looking, as it will for
3591 * certain page cache or slab cache pages such as used for file
3592 * system buffers and inode caches, then instead of starting on the
3593 * local node to look for a free page, rather spread the starting
3594 * node around the tasks mems_allowed nodes.
3596 * We don't have to worry about the returned node being offline
3597 * because "it can't happen", and even if it did, it would be ok.
3599 * The routines calling guarantee_online_mems() are careful to
3600 * only set nodes in task->mems_allowed that are online. So it
3601 * should not be possible for the following code to return an
3602 * offline node. But if it did, that would be ok, as this routine
3603 * is not returning the node where the allocation must be, only
3604 * the node where the search should start. The zonelist passed to
3605 * __alloc_pages() will include all nodes. If the slab allocator
3606 * is passed an offline node, it will fall back to the local node.
3607 * See kmem_cache_alloc_node().
3610 static int cpuset_spread_node(int *rotor)
3612 return *rotor = next_node_in(*rotor, current->mems_allowed);
3615 int cpuset_mem_spread_node(void)
3617 if (current->cpuset_mem_spread_rotor == NUMA_NO_NODE)
3618 current->cpuset_mem_spread_rotor =
3619 node_random(¤t->mems_allowed);
3621 return cpuset_spread_node(¤t->cpuset_mem_spread_rotor);
3624 int cpuset_slab_spread_node(void)
3626 if (current->cpuset_slab_spread_rotor == NUMA_NO_NODE)
3627 current->cpuset_slab_spread_rotor =
3628 node_random(¤t->mems_allowed);
3630 return cpuset_spread_node(¤t->cpuset_slab_spread_rotor);
3633 EXPORT_SYMBOL_GPL(cpuset_mem_spread_node);
3636 * cpuset_mems_allowed_intersects - Does @tsk1's mems_allowed intersect @tsk2's?
3637 * @tsk1: pointer to task_struct of some task.
3638 * @tsk2: pointer to task_struct of some other task.
3640 * Description: Return true if @tsk1's mems_allowed intersects the
3641 * mems_allowed of @tsk2. Used by the OOM killer to determine if
3642 * one of the task's memory usage might impact the memory available
3646 int cpuset_mems_allowed_intersects(const struct task_struct *tsk1,
3647 const struct task_struct *tsk2)
3649 return nodes_intersects(tsk1->mems_allowed, tsk2->mems_allowed);
3653 * cpuset_print_current_mems_allowed - prints current's cpuset and mems_allowed
3655 * Description: Prints current's name, cpuset name, and cached copy of its
3656 * mems_allowed to the kernel log.
3658 void cpuset_print_current_mems_allowed(void)
3660 struct cgroup *cgrp;
3664 cgrp = task_cs(current)->css.cgroup;
3665 pr_cont(",cpuset=");
3666 pr_cont_cgroup_name(cgrp);
3667 pr_cont(",mems_allowed=%*pbl",
3668 nodemask_pr_args(¤t->mems_allowed));
3674 * Collection of memory_pressure is suppressed unless
3675 * this flag is enabled by writing "1" to the special
3676 * cpuset file 'memory_pressure_enabled' in the root cpuset.
3679 int cpuset_memory_pressure_enabled __read_mostly;
3682 * cpuset_memory_pressure_bump - keep stats of per-cpuset reclaims.
3684 * Keep a running average of the rate of synchronous (direct)
3685 * page reclaim efforts initiated by tasks in each cpuset.
3687 * This represents the rate at which some task in the cpuset
3688 * ran low on memory on all nodes it was allowed to use, and
3689 * had to enter the kernels page reclaim code in an effort to
3690 * create more free memory by tossing clean pages or swapping
3691 * or writing dirty pages.
3693 * Display to user space in the per-cpuset read-only file
3694 * "memory_pressure". Value displayed is an integer
3695 * representing the recent rate of entry into the synchronous
3696 * (direct) page reclaim by any task attached to the cpuset.
3699 void __cpuset_memory_pressure_bump(void)
3702 fmeter_markevent(&task_cs(current)->fmeter);
3706 #ifdef CONFIG_PROC_PID_CPUSET
3708 * proc_cpuset_show()
3709 * - Print tasks cpuset path into seq_file.
3710 * - Used for /proc/<pid>/cpuset.
3711 * - No need to task_lock(tsk) on this tsk->cpuset reference, as it
3712 * doesn't really matter if tsk->cpuset changes after we read it,
3713 * and we take cpuset_mutex, keeping cpuset_attach() from changing it
3716 int proc_cpuset_show(struct seq_file *m, struct pid_namespace *ns,
3717 struct pid *pid, struct task_struct *tsk)
3720 struct cgroup_subsys_state *css;
3724 buf = kmalloc(PATH_MAX, GFP_KERNEL);
3728 css = task_get_css(tsk, cpuset_cgrp_id);
3729 retval = cgroup_path_ns(css->cgroup, buf, PATH_MAX,
3730 current->nsproxy->cgroup_ns);
3732 if (retval >= PATH_MAX)
3733 retval = -ENAMETOOLONG;
3744 #endif /* CONFIG_PROC_PID_CPUSET */
3746 /* Display task mems_allowed in /proc/<pid>/status file. */
3747 void cpuset_task_status_allowed(struct seq_file *m, struct task_struct *task)
3749 seq_printf(m, "Mems_allowed:\t%*pb\n",
3750 nodemask_pr_args(&task->mems_allowed));
3751 seq_printf(m, "Mems_allowed_list:\t%*pbl\n",
3752 nodemask_pr_args(&task->mems_allowed));