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;
167 * Partition root states:
169 * 0 - not a partition root
173 * -1 - invalid partition root
174 * None of the cpus in cpus_allowed can be put into the parent's
175 * subparts_cpus. In this case, the cpuset is not a real partition
176 * root anymore. However, the CPU_EXCLUSIVE bit will still be set
177 * and the cpuset can be restored back to a partition root if the
178 * parent cpuset can give more CPUs back to this child cpuset.
180 #define PRS_DISABLED 0
181 #define PRS_ENABLED 1
185 * Temporary cpumasks for working with partitions that are passed among
186 * functions to avoid memory allocation in inner functions.
189 cpumask_var_t addmask, delmask; /* For partition root */
190 cpumask_var_t new_cpus; /* For update_cpumasks_hier() */
193 static inline struct cpuset *css_cs(struct cgroup_subsys_state *css)
195 return css ? container_of(css, struct cpuset, css) : NULL;
198 /* Retrieve the cpuset for a task */
199 static inline struct cpuset *task_cs(struct task_struct *task)
201 return css_cs(task_css(task, cpuset_cgrp_id));
204 static inline struct cpuset *parent_cs(struct cpuset *cs)
206 return css_cs(cs->css.parent);
209 /* bits in struct cpuset flags field */
216 CS_SCHED_LOAD_BALANCE,
221 /* convenient tests for these bits */
222 static inline bool is_cpuset_online(struct cpuset *cs)
224 return test_bit(CS_ONLINE, &cs->flags) && !css_is_dying(&cs->css);
227 static inline int is_cpu_exclusive(const struct cpuset *cs)
229 return test_bit(CS_CPU_EXCLUSIVE, &cs->flags);
232 static inline int is_mem_exclusive(const struct cpuset *cs)
234 return test_bit(CS_MEM_EXCLUSIVE, &cs->flags);
237 static inline int is_mem_hardwall(const struct cpuset *cs)
239 return test_bit(CS_MEM_HARDWALL, &cs->flags);
242 static inline int is_sched_load_balance(const struct cpuset *cs)
244 return test_bit(CS_SCHED_LOAD_BALANCE, &cs->flags);
247 static inline int is_memory_migrate(const struct cpuset *cs)
249 return test_bit(CS_MEMORY_MIGRATE, &cs->flags);
252 static inline int is_spread_page(const struct cpuset *cs)
254 return test_bit(CS_SPREAD_PAGE, &cs->flags);
257 static inline int is_spread_slab(const struct cpuset *cs)
259 return test_bit(CS_SPREAD_SLAB, &cs->flags);
262 static inline int is_partition_root(const struct cpuset *cs)
264 return cs->partition_root_state > 0;
267 static struct cpuset top_cpuset = {
268 .flags = ((1 << CS_ONLINE) | (1 << CS_CPU_EXCLUSIVE) |
269 (1 << CS_MEM_EXCLUSIVE)),
270 .partition_root_state = PRS_ENABLED,
274 * cpuset_for_each_child - traverse online children of a cpuset
275 * @child_cs: loop cursor pointing to the current child
276 * @pos_css: used for iteration
277 * @parent_cs: target cpuset to walk children of
279 * Walk @child_cs through the online children of @parent_cs. Must be used
280 * with RCU read locked.
282 #define cpuset_for_each_child(child_cs, pos_css, parent_cs) \
283 css_for_each_child((pos_css), &(parent_cs)->css) \
284 if (is_cpuset_online(((child_cs) = css_cs((pos_css)))))
287 * cpuset_for_each_descendant_pre - pre-order walk of a cpuset's descendants
288 * @des_cs: loop cursor pointing to the current descendant
289 * @pos_css: used for iteration
290 * @root_cs: target cpuset to walk ancestor of
292 * Walk @des_cs through the online descendants of @root_cs. Must be used
293 * with RCU read locked. The caller may modify @pos_css by calling
294 * css_rightmost_descendant() to skip subtree. @root_cs is included in the
295 * iteration and the first node to be visited.
297 #define cpuset_for_each_descendant_pre(des_cs, pos_css, root_cs) \
298 css_for_each_descendant_pre((pos_css), &(root_cs)->css) \
299 if (is_cpuset_online(((des_cs) = css_cs((pos_css)))))
302 * There are two global locks guarding cpuset structures - cpuset_mutex and
303 * callback_lock. We also require taking task_lock() when dereferencing a
304 * task's cpuset pointer. See "The task_lock() exception", at the end of this
307 * A task must hold both locks to modify cpusets. If a task holds
308 * cpuset_mutex, then it blocks others wanting that mutex, ensuring that it
309 * is the only task able to also acquire callback_lock and be able to
310 * modify cpusets. It can perform various checks on the cpuset structure
311 * first, knowing nothing will change. It can also allocate memory while
312 * just holding cpuset_mutex. While it is performing these checks, various
313 * callback routines can briefly acquire callback_lock to query cpusets.
314 * Once it is ready to make the changes, it takes callback_lock, blocking
317 * Calls to the kernel memory allocator can not be made while holding
318 * callback_lock, as that would risk double tripping on callback_lock
319 * from one of the callbacks into the cpuset code from within
322 * If a task is only holding callback_lock, then it has read-only
325 * Now, the task_struct fields mems_allowed and mempolicy may be changed
326 * by other task, we use alloc_lock in the task_struct fields to protect
329 * The cpuset_common_file_read() handlers only hold callback_lock across
330 * small pieces of code, such as when reading out possibly multi-word
331 * cpumasks and nodemasks.
333 * Accessing a task's cpuset should be done in accordance with the
334 * guidelines for accessing subsystem state in kernel/cgroup.c
337 DEFINE_STATIC_PERCPU_RWSEM(cpuset_rwsem);
339 void cpuset_read_lock(void)
341 percpu_down_read(&cpuset_rwsem);
344 void cpuset_read_unlock(void)
346 percpu_up_read(&cpuset_rwsem);
349 static DEFINE_SPINLOCK(callback_lock);
351 static struct workqueue_struct *cpuset_migrate_mm_wq;
354 * CPU / memory hotplug is handled asynchronously.
356 static void cpuset_hotplug_workfn(struct work_struct *work);
357 static DECLARE_WORK(cpuset_hotplug_work, cpuset_hotplug_workfn);
359 static DECLARE_WAIT_QUEUE_HEAD(cpuset_attach_wq);
362 * Cgroup v2 behavior is used when on default hierarchy or the
363 * cgroup_v2_mode flag is set.
365 static inline bool is_in_v2_mode(void)
367 return cgroup_subsys_on_dfl(cpuset_cgrp_subsys) ||
368 (cpuset_cgrp_subsys.root->flags & CGRP_ROOT_CPUSET_V2_MODE);
372 * Return in pmask the portion of a cpusets's cpus_allowed that
373 * are online. If none are online, walk up the cpuset hierarchy
374 * until we find one that does have some online cpus.
376 * One way or another, we guarantee to return some non-empty subset
377 * of cpu_online_mask.
379 * Call with callback_lock or cpuset_mutex held.
381 static void guarantee_online_cpus(struct cpuset *cs, struct cpumask *pmask)
383 while (!cpumask_intersects(cs->effective_cpus, cpu_online_mask)) {
387 * The top cpuset doesn't have any online cpu as a
388 * consequence of a race between cpuset_hotplug_work
389 * and cpu hotplug notifier. But we know the top
390 * cpuset's effective_cpus is on its way to to be
391 * identical to cpu_online_mask.
393 cpumask_copy(pmask, cpu_online_mask);
397 cpumask_and(pmask, cs->effective_cpus, cpu_online_mask);
401 * Return in *pmask the portion of a cpusets's mems_allowed that
402 * are online, with memory. If none are online with memory, walk
403 * up the cpuset hierarchy until we find one that does have some
404 * online mems. The top cpuset always has some mems online.
406 * One way or another, we guarantee to return some non-empty subset
407 * of node_states[N_MEMORY].
409 * Call with callback_lock or cpuset_mutex held.
411 static void guarantee_online_mems(struct cpuset *cs, nodemask_t *pmask)
413 while (!nodes_intersects(cs->effective_mems, node_states[N_MEMORY]))
415 nodes_and(*pmask, cs->effective_mems, node_states[N_MEMORY]);
419 * update task's spread flag if cpuset's page/slab spread flag is set
421 * Call with callback_lock or cpuset_mutex held.
423 static void cpuset_update_task_spread_flag(struct cpuset *cs,
424 struct task_struct *tsk)
426 if (is_spread_page(cs))
427 task_set_spread_page(tsk);
429 task_clear_spread_page(tsk);
431 if (is_spread_slab(cs))
432 task_set_spread_slab(tsk);
434 task_clear_spread_slab(tsk);
438 * is_cpuset_subset(p, q) - Is cpuset p a subset of cpuset q?
440 * One cpuset is a subset of another if all its allowed CPUs and
441 * Memory Nodes are a subset of the other, and its exclusive flags
442 * are only set if the other's are set. Call holding cpuset_mutex.
445 static int is_cpuset_subset(const struct cpuset *p, const struct cpuset *q)
447 return cpumask_subset(p->cpus_allowed, q->cpus_allowed) &&
448 nodes_subset(p->mems_allowed, q->mems_allowed) &&
449 is_cpu_exclusive(p) <= is_cpu_exclusive(q) &&
450 is_mem_exclusive(p) <= is_mem_exclusive(q);
454 * alloc_cpumasks - allocate three cpumasks for cpuset
455 * @cs: the cpuset that have cpumasks to be allocated.
456 * @tmp: the tmpmasks structure pointer
457 * Return: 0 if successful, -ENOMEM otherwise.
459 * Only one of the two input arguments should be non-NULL.
461 static inline int alloc_cpumasks(struct cpuset *cs, struct tmpmasks *tmp)
463 cpumask_var_t *pmask1, *pmask2, *pmask3;
466 pmask1 = &cs->cpus_allowed;
467 pmask2 = &cs->effective_cpus;
468 pmask3 = &cs->subparts_cpus;
470 pmask1 = &tmp->new_cpus;
471 pmask2 = &tmp->addmask;
472 pmask3 = &tmp->delmask;
475 if (!zalloc_cpumask_var(pmask1, GFP_KERNEL))
478 if (!zalloc_cpumask_var(pmask2, GFP_KERNEL))
481 if (!zalloc_cpumask_var(pmask3, GFP_KERNEL))
487 free_cpumask_var(*pmask2);
489 free_cpumask_var(*pmask1);
494 * free_cpumasks - free cpumasks in a tmpmasks structure
495 * @cs: the cpuset that have cpumasks to be free.
496 * @tmp: the tmpmasks structure pointer
498 static inline void free_cpumasks(struct cpuset *cs, struct tmpmasks *tmp)
501 free_cpumask_var(cs->cpus_allowed);
502 free_cpumask_var(cs->effective_cpus);
503 free_cpumask_var(cs->subparts_cpus);
506 free_cpumask_var(tmp->new_cpus);
507 free_cpumask_var(tmp->addmask);
508 free_cpumask_var(tmp->delmask);
513 * alloc_trial_cpuset - allocate a trial cpuset
514 * @cs: the cpuset that the trial cpuset duplicates
516 static struct cpuset *alloc_trial_cpuset(struct cpuset *cs)
518 struct cpuset *trial;
520 trial = kmemdup(cs, sizeof(*cs), GFP_KERNEL);
524 if (alloc_cpumasks(trial, NULL)) {
529 cpumask_copy(trial->cpus_allowed, cs->cpus_allowed);
530 cpumask_copy(trial->effective_cpus, cs->effective_cpus);
535 * free_cpuset - free the cpuset
536 * @cs: the cpuset to be freed
538 static inline void free_cpuset(struct cpuset *cs)
540 free_cpumasks(cs, NULL);
545 * validate_change() - Used to validate that any proposed cpuset change
546 * follows the structural rules for cpusets.
548 * If we replaced the flag and mask values of the current cpuset
549 * (cur) with those values in the trial cpuset (trial), would
550 * our various subset and exclusive rules still be valid? Presumes
553 * 'cur' is the address of an actual, in-use cpuset. Operations
554 * such as list traversal that depend on the actual address of the
555 * cpuset in the list must use cur below, not trial.
557 * 'trial' is the address of bulk structure copy of cur, with
558 * perhaps one or more of the fields cpus_allowed, mems_allowed,
559 * or flags changed to new, trial values.
561 * Return 0 if valid, -errno if not.
564 static int validate_change(struct cpuset *cur, struct cpuset *trial)
566 struct cgroup_subsys_state *css;
567 struct cpuset *c, *par;
572 /* Each of our child cpusets must be a subset of us */
574 cpuset_for_each_child(c, css, cur)
575 if (!is_cpuset_subset(c, trial))
578 /* Remaining checks don't apply to root cpuset */
580 if (cur == &top_cpuset)
583 par = parent_cs(cur);
585 /* On legacy hiearchy, we must be a subset of our parent cpuset. */
587 if (!is_in_v2_mode() && !is_cpuset_subset(trial, par))
591 * If either I or some sibling (!= me) is exclusive, we can't
595 cpuset_for_each_child(c, css, par) {
596 if ((is_cpu_exclusive(trial) || is_cpu_exclusive(c)) &&
598 cpumask_intersects(trial->cpus_allowed, c->cpus_allowed))
600 if ((is_mem_exclusive(trial) || is_mem_exclusive(c)) &&
602 nodes_intersects(trial->mems_allowed, c->mems_allowed))
607 * Cpusets with tasks - existing or newly being attached - can't
608 * be changed to have empty cpus_allowed or mems_allowed.
611 if ((cgroup_is_populated(cur->css.cgroup) || cur->attach_in_progress)) {
612 if (!cpumask_empty(cur->cpus_allowed) &&
613 cpumask_empty(trial->cpus_allowed))
615 if (!nodes_empty(cur->mems_allowed) &&
616 nodes_empty(trial->mems_allowed))
621 * We can't shrink if we won't have enough room for SCHED_DEADLINE
625 if (is_cpu_exclusive(cur) &&
626 !cpuset_cpumask_can_shrink(cur->cpus_allowed,
627 trial->cpus_allowed))
638 * Helper routine for generate_sched_domains().
639 * Do cpusets a, b have overlapping effective cpus_allowed masks?
641 static int cpusets_overlap(struct cpuset *a, struct cpuset *b)
643 return cpumask_intersects(a->effective_cpus, b->effective_cpus);
647 update_domain_attr(struct sched_domain_attr *dattr, struct cpuset *c)
649 if (dattr->relax_domain_level < c->relax_domain_level)
650 dattr->relax_domain_level = c->relax_domain_level;
654 static void update_domain_attr_tree(struct sched_domain_attr *dattr,
655 struct cpuset *root_cs)
658 struct cgroup_subsys_state *pos_css;
661 cpuset_for_each_descendant_pre(cp, pos_css, root_cs) {
662 /* skip the whole subtree if @cp doesn't have any CPU */
663 if (cpumask_empty(cp->cpus_allowed)) {
664 pos_css = css_rightmost_descendant(pos_css);
668 if (is_sched_load_balance(cp))
669 update_domain_attr(dattr, cp);
674 /* Must be called with cpuset_mutex held. */
675 static inline int nr_cpusets(void)
677 /* jump label reference count + the top-level cpuset */
678 return static_key_count(&cpusets_enabled_key.key) + 1;
682 * generate_sched_domains()
684 * This function builds a partial partition of the systems CPUs
685 * A 'partial partition' is a set of non-overlapping subsets whose
686 * union is a subset of that set.
687 * The output of this function needs to be passed to kernel/sched/core.c
688 * partition_sched_domains() routine, which will rebuild the scheduler's
689 * load balancing domains (sched domains) as specified by that partial
692 * See "What is sched_load_balance" in Documentation/admin-guide/cgroup-v1/cpusets.rst
693 * for a background explanation of this.
695 * Does not return errors, on the theory that the callers of this
696 * routine would rather not worry about failures to rebuild sched
697 * domains when operating in the severe memory shortage situations
698 * that could cause allocation failures below.
700 * Must be called with cpuset_mutex held.
702 * The three key local variables below are:
703 * cp - cpuset pointer, used (together with pos_css) to perform a
704 * top-down scan of all cpusets. For our purposes, rebuilding
705 * the schedulers sched domains, we can ignore !is_sched_load_
707 * csa - (for CpuSet Array) Array of pointers to all the cpusets
708 * that need to be load balanced, for convenient iterative
709 * access by the subsequent code that finds the best partition,
710 * i.e the set of domains (subsets) of CPUs such that the
711 * cpus_allowed of every cpuset marked is_sched_load_balance
712 * is a subset of one of these domains, while there are as
713 * many such domains as possible, each as small as possible.
714 * doms - Conversion of 'csa' to an array of cpumasks, for passing to
715 * the kernel/sched/core.c routine partition_sched_domains() in a
716 * convenient format, that can be easily compared to the prior
717 * value to determine what partition elements (sched domains)
718 * were changed (added or removed.)
720 * Finding the best partition (set of domains):
721 * The triple nested loops below over i, j, k scan over the
722 * load balanced cpusets (using the array of cpuset pointers in
723 * csa[]) looking for pairs of cpusets that have overlapping
724 * cpus_allowed, but which don't have the same 'pn' partition
725 * number and gives them in the same partition number. It keeps
726 * looping on the 'restart' label until it can no longer find
729 * The union of the cpus_allowed masks from the set of
730 * all cpusets having the same 'pn' value then form the one
731 * element of the partition (one sched domain) to be passed to
732 * partition_sched_domains().
734 static int generate_sched_domains(cpumask_var_t **domains,
735 struct sched_domain_attr **attributes)
737 struct cpuset *cp; /* top-down scan of cpusets */
738 struct cpuset **csa; /* array of all cpuset ptrs */
739 int csn; /* how many cpuset ptrs in csa so far */
740 int i, j, k; /* indices for partition finding loops */
741 cpumask_var_t *doms; /* resulting partition; i.e. sched domains */
742 struct sched_domain_attr *dattr; /* attributes for custom domains */
743 int ndoms = 0; /* number of sched domains in result */
744 int nslot; /* next empty doms[] struct cpumask slot */
745 struct cgroup_subsys_state *pos_css;
746 bool root_load_balance = is_sched_load_balance(&top_cpuset);
752 /* Special case for the 99% of systems with one, full, sched domain */
753 if (root_load_balance && !top_cpuset.nr_subparts_cpus) {
755 doms = alloc_sched_domains(ndoms);
759 dattr = kmalloc(sizeof(struct sched_domain_attr), GFP_KERNEL);
761 *dattr = SD_ATTR_INIT;
762 update_domain_attr_tree(dattr, &top_cpuset);
764 cpumask_and(doms[0], top_cpuset.effective_cpus,
765 housekeeping_cpumask(HK_FLAG_DOMAIN));
770 csa = kmalloc_array(nr_cpusets(), sizeof(cp), GFP_KERNEL);
776 if (root_load_balance)
777 csa[csn++] = &top_cpuset;
778 cpuset_for_each_descendant_pre(cp, pos_css, &top_cpuset) {
779 if (cp == &top_cpuset)
782 * Continue traversing beyond @cp iff @cp has some CPUs and
783 * isn't load balancing. The former is obvious. The
784 * latter: All child cpusets contain a subset of the
785 * parent's cpus, so just skip them, and then we call
786 * update_domain_attr_tree() to calc relax_domain_level of
787 * the corresponding sched domain.
789 * If root is load-balancing, we can skip @cp if it
790 * is a subset of the root's effective_cpus.
792 if (!cpumask_empty(cp->cpus_allowed) &&
793 !(is_sched_load_balance(cp) &&
794 cpumask_intersects(cp->cpus_allowed,
795 housekeeping_cpumask(HK_FLAG_DOMAIN))))
798 if (root_load_balance &&
799 cpumask_subset(cp->cpus_allowed, top_cpuset.effective_cpus))
802 if (is_sched_load_balance(cp) &&
803 !cpumask_empty(cp->effective_cpus))
806 /* skip @cp's subtree if not a partition root */
807 if (!is_partition_root(cp))
808 pos_css = css_rightmost_descendant(pos_css);
812 for (i = 0; i < csn; i++)
817 /* Find the best partition (set of sched domains) */
818 for (i = 0; i < csn; i++) {
819 struct cpuset *a = csa[i];
822 for (j = 0; j < csn; j++) {
823 struct cpuset *b = csa[j];
826 if (apn != bpn && cpusets_overlap(a, b)) {
827 for (k = 0; k < csn; k++) {
828 struct cpuset *c = csa[k];
833 ndoms--; /* one less element */
840 * Now we know how many domains to create.
841 * Convert <csn, csa> to <ndoms, doms> and populate cpu masks.
843 doms = alloc_sched_domains(ndoms);
848 * The rest of the code, including the scheduler, can deal with
849 * dattr==NULL case. No need to abort if alloc fails.
851 dattr = kmalloc_array(ndoms, sizeof(struct sched_domain_attr),
854 for (nslot = 0, i = 0; i < csn; i++) {
855 struct cpuset *a = csa[i];
860 /* Skip completed partitions */
866 if (nslot == ndoms) {
867 static int warnings = 10;
869 pr_warn("rebuild_sched_domains confused: nslot %d, ndoms %d, csn %d, i %d, apn %d\n",
870 nslot, ndoms, csn, i, apn);
878 *(dattr + nslot) = SD_ATTR_INIT;
879 for (j = i; j < csn; j++) {
880 struct cpuset *b = csa[j];
883 cpumask_or(dp, dp, b->effective_cpus);
884 cpumask_and(dp, dp, housekeeping_cpumask(HK_FLAG_DOMAIN));
886 update_domain_attr_tree(dattr + nslot, b);
888 /* Done with this partition */
894 BUG_ON(nslot != ndoms);
900 * Fallback to the default domain if kmalloc() failed.
901 * See comments in partition_sched_domains().
911 static void update_tasks_root_domain(struct cpuset *cs)
913 struct css_task_iter it;
914 struct task_struct *task;
916 css_task_iter_start(&cs->css, 0, &it);
918 while ((task = css_task_iter_next(&it)))
919 dl_add_task_root_domain(task);
921 css_task_iter_end(&it);
924 static void rebuild_root_domains(void)
926 struct cpuset *cs = NULL;
927 struct cgroup_subsys_state *pos_css;
929 percpu_rwsem_assert_held(&cpuset_rwsem);
930 lockdep_assert_cpus_held();
931 lockdep_assert_held(&sched_domains_mutex);
933 cgroup_enable_task_cg_lists();
938 * Clear default root domain DL accounting, it will be computed again
939 * if a task belongs to it.
941 dl_clear_root_domain(&def_root_domain);
943 cpuset_for_each_descendant_pre(cs, pos_css, &top_cpuset) {
945 if (cpumask_empty(cs->effective_cpus)) {
946 pos_css = css_rightmost_descendant(pos_css);
954 update_tasks_root_domain(cs);
963 partition_and_rebuild_sched_domains(int ndoms_new, cpumask_var_t doms_new[],
964 struct sched_domain_attr *dattr_new)
966 mutex_lock(&sched_domains_mutex);
967 partition_sched_domains_locked(ndoms_new, doms_new, dattr_new);
968 rebuild_root_domains();
969 mutex_unlock(&sched_domains_mutex);
973 * Rebuild scheduler domains.
975 * If the flag 'sched_load_balance' of any cpuset with non-empty
976 * 'cpus' changes, or if the 'cpus' allowed changes in any cpuset
977 * which has that flag enabled, or if any cpuset with a non-empty
978 * 'cpus' is removed, then call this routine to rebuild the
979 * scheduler's dynamic sched domains.
981 * Call with cpuset_mutex held. Takes get_online_cpus().
983 static void rebuild_sched_domains_locked(void)
985 struct cgroup_subsys_state *pos_css;
986 struct sched_domain_attr *attr;
991 lockdep_assert_cpus_held();
992 percpu_rwsem_assert_held(&cpuset_rwsem);
995 * If we have raced with CPU hotplug, return early to avoid
996 * passing doms with offlined cpu to partition_sched_domains().
997 * Anyways, cpuset_hotplug_workfn() will rebuild sched domains.
999 * With no CPUs in any subpartitions, top_cpuset's effective CPUs
1000 * should be the same as the active CPUs, so checking only top_cpuset
1001 * is enough to detect racing CPU offlines.
1003 if (!top_cpuset.nr_subparts_cpus &&
1004 !cpumask_equal(top_cpuset.effective_cpus, cpu_active_mask))
1008 * With subpartition CPUs, however, the effective CPUs of a partition
1009 * root should be only a subset of the active CPUs. Since a CPU in any
1010 * partition root could be offlined, all must be checked.
1012 if (top_cpuset.nr_subparts_cpus) {
1014 cpuset_for_each_descendant_pre(cs, pos_css, &top_cpuset) {
1015 if (!is_partition_root(cs)) {
1016 pos_css = css_rightmost_descendant(pos_css);
1019 if (!cpumask_subset(cs->effective_cpus,
1028 /* Generate domain masks and attrs */
1029 ndoms = generate_sched_domains(&doms, &attr);
1031 /* Have scheduler rebuild the domains */
1032 partition_and_rebuild_sched_domains(ndoms, doms, attr);
1034 #else /* !CONFIG_SMP */
1035 static void rebuild_sched_domains_locked(void)
1038 #endif /* CONFIG_SMP */
1040 void rebuild_sched_domains(void)
1043 percpu_down_write(&cpuset_rwsem);
1044 rebuild_sched_domains_locked();
1045 percpu_up_write(&cpuset_rwsem);
1050 * update_tasks_cpumask - Update the cpumasks of tasks in the cpuset.
1051 * @cs: the cpuset in which each task's cpus_allowed mask needs to be changed
1053 * Iterate through each task of @cs updating its cpus_allowed to the
1054 * effective cpuset's. As this function is called with cpuset_mutex held,
1055 * cpuset membership stays stable.
1057 static void update_tasks_cpumask(struct cpuset *cs)
1059 struct css_task_iter it;
1060 struct task_struct *task;
1061 bool top_cs = cs == &top_cpuset;
1063 css_task_iter_start(&cs->css, 0, &it);
1064 while ((task = css_task_iter_next(&it))) {
1066 * Percpu kthreads in top_cpuset are ignored
1068 if (top_cs && (task->flags & PF_KTHREAD) &&
1069 kthread_is_per_cpu(task))
1071 set_cpus_allowed_ptr(task, cs->effective_cpus);
1073 css_task_iter_end(&it);
1077 * compute_effective_cpumask - Compute the effective cpumask of the cpuset
1078 * @new_cpus: the temp variable for the new effective_cpus mask
1079 * @cs: the cpuset the need to recompute the new effective_cpus mask
1080 * @parent: the parent cpuset
1082 * If the parent has subpartition CPUs, include them in the list of
1083 * allowable CPUs in computing the new effective_cpus mask. Since offlined
1084 * CPUs are not removed from subparts_cpus, we have to use cpu_active_mask
1085 * to mask those out.
1087 static void compute_effective_cpumask(struct cpumask *new_cpus,
1088 struct cpuset *cs, struct cpuset *parent)
1090 if (parent->nr_subparts_cpus) {
1091 cpumask_or(new_cpus, parent->effective_cpus,
1092 parent->subparts_cpus);
1093 cpumask_and(new_cpus, new_cpus, cs->cpus_allowed);
1094 cpumask_and(new_cpus, new_cpus, cpu_active_mask);
1096 cpumask_and(new_cpus, cs->cpus_allowed, parent->effective_cpus);
1101 * Commands for update_parent_subparts_cpumask
1104 partcmd_enable, /* Enable partition root */
1105 partcmd_disable, /* Disable partition root */
1106 partcmd_update, /* Update parent's subparts_cpus */
1110 * update_parent_subparts_cpumask - update subparts_cpus mask of parent cpuset
1111 * @cpuset: The cpuset that requests change in partition root state
1112 * @cmd: Partition root state change command
1113 * @newmask: Optional new cpumask for partcmd_update
1114 * @tmp: Temporary addmask and delmask
1115 * Return: 0, 1 or an error code
1117 * For partcmd_enable, the cpuset is being transformed from a non-partition
1118 * root to a partition root. The cpus_allowed mask of the given cpuset will
1119 * be put into parent's subparts_cpus and taken away from parent's
1120 * effective_cpus. The function will return 0 if all the CPUs listed in
1121 * cpus_allowed can be granted or an error code will be returned.
1123 * For partcmd_disable, the cpuset is being transofrmed from a partition
1124 * root back to a non-partition root. any CPUs in cpus_allowed that are in
1125 * parent's subparts_cpus will be taken away from that cpumask and put back
1126 * into parent's effective_cpus. 0 should always be returned.
1128 * For partcmd_update, if the optional newmask is specified, the cpu
1129 * list is to be changed from cpus_allowed to newmask. Otherwise,
1130 * cpus_allowed is assumed to remain the same. The cpuset should either
1131 * be a partition root or an invalid partition root. The partition root
1132 * state may change if newmask is NULL and none of the requested CPUs can
1133 * be granted by the parent. The function will return 1 if changes to
1134 * parent's subparts_cpus and effective_cpus happen or 0 otherwise.
1135 * Error code should only be returned when newmask is non-NULL.
1137 * The partcmd_enable and partcmd_disable commands are used by
1138 * update_prstate(). The partcmd_update command is used by
1139 * update_cpumasks_hier() with newmask NULL and update_cpumask() with
1142 * The checking is more strict when enabling partition root than the
1143 * other two commands.
1145 * Because of the implicit cpu exclusive nature of a partition root,
1146 * cpumask changes that violates the cpu exclusivity rule will not be
1147 * permitted when checked by validate_change(). The validate_change()
1148 * function will also prevent any changes to the cpu list if it is not
1149 * a superset of children's cpu lists.
1151 static int update_parent_subparts_cpumask(struct cpuset *cpuset, int cmd,
1152 struct cpumask *newmask,
1153 struct tmpmasks *tmp)
1155 struct cpuset *parent = parent_cs(cpuset);
1156 int adding; /* Moving cpus from effective_cpus to subparts_cpus */
1157 int deleting; /* Moving cpus from subparts_cpus to effective_cpus */
1158 bool part_error = false; /* Partition error? */
1160 percpu_rwsem_assert_held(&cpuset_rwsem);
1163 * The parent must be a partition root.
1164 * The new cpumask, if present, or the current cpus_allowed must
1167 if (!is_partition_root(parent) ||
1168 (newmask && cpumask_empty(newmask)) ||
1169 (!newmask && cpumask_empty(cpuset->cpus_allowed)))
1173 * Enabling/disabling partition root is not allowed if there are
1176 if ((cmd != partcmd_update) && css_has_online_children(&cpuset->css))
1180 * Enabling partition root is not allowed if not all the CPUs
1181 * can be granted from parent's effective_cpus or at least one
1182 * CPU will be left after that.
1184 if ((cmd == partcmd_enable) &&
1185 (!cpumask_subset(cpuset->cpus_allowed, parent->effective_cpus) ||
1186 cpumask_equal(cpuset->cpus_allowed, parent->effective_cpus)))
1190 * A cpumask update cannot make parent's effective_cpus become empty.
1192 adding = deleting = false;
1193 if (cmd == partcmd_enable) {
1194 cpumask_copy(tmp->addmask, cpuset->cpus_allowed);
1196 } else if (cmd == partcmd_disable) {
1197 deleting = cpumask_and(tmp->delmask, cpuset->cpus_allowed,
1198 parent->subparts_cpus);
1199 } else if (newmask) {
1201 * partcmd_update with newmask:
1203 * delmask = cpus_allowed & ~newmask & parent->subparts_cpus
1204 * addmask = newmask & parent->effective_cpus
1205 * & ~parent->subparts_cpus
1207 cpumask_andnot(tmp->delmask, cpuset->cpus_allowed, newmask);
1208 deleting = cpumask_and(tmp->delmask, tmp->delmask,
1209 parent->subparts_cpus);
1211 cpumask_and(tmp->addmask, newmask, parent->effective_cpus);
1212 adding = cpumask_andnot(tmp->addmask, tmp->addmask,
1213 parent->subparts_cpus);
1215 * Return error if the new effective_cpus could become empty.
1218 cpumask_equal(parent->effective_cpus, tmp->addmask)) {
1222 * As some of the CPUs in subparts_cpus might have
1223 * been offlined, we need to compute the real delmask
1226 if (!cpumask_and(tmp->addmask, tmp->delmask,
1229 cpumask_copy(tmp->addmask, parent->effective_cpus);
1233 * partcmd_update w/o newmask:
1235 * addmask = cpus_allowed & parent->effectiveb_cpus
1237 * Note that parent's subparts_cpus may have been
1238 * pre-shrunk in case there is a change in the cpu list.
1239 * So no deletion is needed.
1241 adding = cpumask_and(tmp->addmask, cpuset->cpus_allowed,
1242 parent->effective_cpus);
1243 part_error = cpumask_equal(tmp->addmask,
1244 parent->effective_cpus);
1247 if (cmd == partcmd_update) {
1248 int prev_prs = cpuset->partition_root_state;
1251 * Check for possible transition between PRS_ENABLED
1254 switch (cpuset->partition_root_state) {
1257 cpuset->partition_root_state = PRS_ERROR;
1261 cpuset->partition_root_state = PRS_ENABLED;
1265 * Set part_error if previously in invalid state.
1267 part_error = (prev_prs == PRS_ERROR);
1270 if (!part_error && (cpuset->partition_root_state == PRS_ERROR))
1271 return 0; /* Nothing need to be done */
1273 if (cpuset->partition_root_state == PRS_ERROR) {
1275 * Remove all its cpus from parent's subparts_cpus.
1278 deleting = cpumask_and(tmp->delmask, cpuset->cpus_allowed,
1279 parent->subparts_cpus);
1282 if (!adding && !deleting)
1286 * Change the parent's subparts_cpus.
1287 * Newly added CPUs will be removed from effective_cpus and
1288 * newly deleted ones will be added back to effective_cpus.
1290 spin_lock_irq(&callback_lock);
1292 cpumask_or(parent->subparts_cpus,
1293 parent->subparts_cpus, tmp->addmask);
1294 cpumask_andnot(parent->effective_cpus,
1295 parent->effective_cpus, tmp->addmask);
1298 cpumask_andnot(parent->subparts_cpus,
1299 parent->subparts_cpus, tmp->delmask);
1301 * Some of the CPUs in subparts_cpus might have been offlined.
1303 cpumask_and(tmp->delmask, tmp->delmask, cpu_active_mask);
1304 cpumask_or(parent->effective_cpus,
1305 parent->effective_cpus, tmp->delmask);
1308 parent->nr_subparts_cpus = cpumask_weight(parent->subparts_cpus);
1309 spin_unlock_irq(&callback_lock);
1311 return cmd == partcmd_update;
1315 * update_cpumasks_hier - Update effective cpumasks and tasks in the subtree
1316 * @cs: the cpuset to consider
1317 * @tmp: temp variables for calculating effective_cpus & partition setup
1319 * When congifured cpumask is changed, the effective cpumasks of this cpuset
1320 * and all its descendants need to be updated.
1322 * On legacy hierachy, effective_cpus will be the same with cpu_allowed.
1324 * Called with cpuset_mutex held
1326 static void update_cpumasks_hier(struct cpuset *cs, struct tmpmasks *tmp)
1329 struct cgroup_subsys_state *pos_css;
1330 bool need_rebuild_sched_domains = false;
1333 cpuset_for_each_descendant_pre(cp, pos_css, cs) {
1334 struct cpuset *parent = parent_cs(cp);
1336 compute_effective_cpumask(tmp->new_cpus, cp, parent);
1339 * If it becomes empty, inherit the effective mask of the
1340 * parent, which is guaranteed to have some CPUs.
1342 if (is_in_v2_mode() && cpumask_empty(tmp->new_cpus)) {
1343 cpumask_copy(tmp->new_cpus, parent->effective_cpus);
1344 if (!cp->use_parent_ecpus) {
1345 cp->use_parent_ecpus = true;
1346 parent->child_ecpus_count++;
1348 } else if (cp->use_parent_ecpus) {
1349 cp->use_parent_ecpus = false;
1350 WARN_ON_ONCE(!parent->child_ecpus_count);
1351 parent->child_ecpus_count--;
1355 * Skip the whole subtree if the cpumask remains the same
1356 * and has no partition root state.
1358 if (!cp->partition_root_state &&
1359 cpumask_equal(tmp->new_cpus, cp->effective_cpus)) {
1360 pos_css = css_rightmost_descendant(pos_css);
1365 * update_parent_subparts_cpumask() should have been called
1366 * for cs already in update_cpumask(). We should also call
1367 * update_tasks_cpumask() again for tasks in the parent
1368 * cpuset if the parent's subparts_cpus changes.
1370 if ((cp != cs) && cp->partition_root_state) {
1371 switch (parent->partition_root_state) {
1374 * If parent is not a partition root or an
1375 * invalid partition root, clear the state
1376 * state and the CS_CPU_EXCLUSIVE flag.
1378 WARN_ON_ONCE(cp->partition_root_state
1380 cp->partition_root_state = 0;
1383 * clear_bit() is an atomic operation and
1384 * readers aren't interested in the state
1385 * of CS_CPU_EXCLUSIVE anyway. So we can
1386 * just update the flag without holding
1387 * the callback_lock.
1389 clear_bit(CS_CPU_EXCLUSIVE, &cp->flags);
1393 if (update_parent_subparts_cpumask(cp, partcmd_update, NULL, tmp))
1394 update_tasks_cpumask(parent);
1399 * When parent is invalid, it has to be too.
1401 cp->partition_root_state = PRS_ERROR;
1402 if (cp->nr_subparts_cpus) {
1403 cp->nr_subparts_cpus = 0;
1404 cpumask_clear(cp->subparts_cpus);
1410 if (!css_tryget_online(&cp->css))
1414 spin_lock_irq(&callback_lock);
1416 cpumask_copy(cp->effective_cpus, tmp->new_cpus);
1417 if (cp->nr_subparts_cpus &&
1418 (cp->partition_root_state != PRS_ENABLED)) {
1419 cp->nr_subparts_cpus = 0;
1420 cpumask_clear(cp->subparts_cpus);
1421 } else if (cp->nr_subparts_cpus) {
1423 * Make sure that effective_cpus & subparts_cpus
1424 * are mutually exclusive.
1426 * In the unlikely event that effective_cpus
1427 * becomes empty. we clear cp->nr_subparts_cpus and
1428 * let its child partition roots to compete for
1431 cpumask_andnot(cp->effective_cpus, cp->effective_cpus,
1433 if (cpumask_empty(cp->effective_cpus)) {
1434 cpumask_copy(cp->effective_cpus, tmp->new_cpus);
1435 cpumask_clear(cp->subparts_cpus);
1436 cp->nr_subparts_cpus = 0;
1437 } else if (!cpumask_subset(cp->subparts_cpus,
1439 cpumask_andnot(cp->subparts_cpus,
1440 cp->subparts_cpus, tmp->new_cpus);
1441 cp->nr_subparts_cpus
1442 = cpumask_weight(cp->subparts_cpus);
1445 spin_unlock_irq(&callback_lock);
1447 WARN_ON(!is_in_v2_mode() &&
1448 !cpumask_equal(cp->cpus_allowed, cp->effective_cpus));
1450 update_tasks_cpumask(cp);
1453 * On legacy hierarchy, if the effective cpumask of any non-
1454 * empty cpuset is changed, we need to rebuild sched domains.
1455 * On default hierarchy, the cpuset needs to be a partition
1458 if (!cpumask_empty(cp->cpus_allowed) &&
1459 is_sched_load_balance(cp) &&
1460 (!cgroup_subsys_on_dfl(cpuset_cgrp_subsys) ||
1461 is_partition_root(cp)))
1462 need_rebuild_sched_domains = true;
1469 if (need_rebuild_sched_domains)
1470 rebuild_sched_domains_locked();
1474 * update_sibling_cpumasks - Update siblings cpumasks
1475 * @parent: Parent cpuset
1476 * @cs: Current cpuset
1477 * @tmp: Temp variables
1479 static void update_sibling_cpumasks(struct cpuset *parent, struct cpuset *cs,
1480 struct tmpmasks *tmp)
1482 struct cpuset *sibling;
1483 struct cgroup_subsys_state *pos_css;
1485 percpu_rwsem_assert_held(&cpuset_rwsem);
1488 * Check all its siblings and call update_cpumasks_hier()
1489 * if their use_parent_ecpus flag is set in order for them
1490 * to use the right effective_cpus value.
1492 * The update_cpumasks_hier() function may sleep. So we have to
1493 * release the RCU read lock before calling it.
1496 cpuset_for_each_child(sibling, pos_css, parent) {
1499 if (!sibling->use_parent_ecpus)
1501 if (!css_tryget_online(&sibling->css))
1505 update_cpumasks_hier(sibling, tmp);
1507 css_put(&sibling->css);
1513 * update_cpumask - update the cpus_allowed mask of a cpuset and all tasks in it
1514 * @cs: the cpuset to consider
1515 * @trialcs: trial cpuset
1516 * @buf: buffer of cpu numbers written to this cpuset
1518 static int update_cpumask(struct cpuset *cs, struct cpuset *trialcs,
1522 struct tmpmasks tmp;
1524 /* top_cpuset.cpus_allowed tracks cpu_online_mask; it's read-only */
1525 if (cs == &top_cpuset)
1529 * An empty cpus_allowed is ok only if the cpuset has no tasks.
1530 * Since cpulist_parse() fails on an empty mask, we special case
1531 * that parsing. The validate_change() call ensures that cpusets
1532 * with tasks have cpus.
1535 cpumask_clear(trialcs->cpus_allowed);
1537 retval = cpulist_parse(buf, trialcs->cpus_allowed);
1541 if (!cpumask_subset(trialcs->cpus_allowed,
1542 top_cpuset.cpus_allowed))
1546 /* Nothing to do if the cpus didn't change */
1547 if (cpumask_equal(cs->cpus_allowed, trialcs->cpus_allowed))
1550 retval = validate_change(cs, trialcs);
1554 #ifdef CONFIG_CPUMASK_OFFSTACK
1556 * Use the cpumasks in trialcs for tmpmasks when they are pointers
1557 * to allocated cpumasks.
1559 tmp.addmask = trialcs->subparts_cpus;
1560 tmp.delmask = trialcs->effective_cpus;
1561 tmp.new_cpus = trialcs->cpus_allowed;
1564 if (cs->partition_root_state) {
1565 /* Cpumask of a partition root cannot be empty */
1566 if (cpumask_empty(trialcs->cpus_allowed))
1568 if (update_parent_subparts_cpumask(cs, partcmd_update,
1569 trialcs->cpus_allowed, &tmp) < 0)
1573 spin_lock_irq(&callback_lock);
1574 cpumask_copy(cs->cpus_allowed, trialcs->cpus_allowed);
1577 * Make sure that subparts_cpus is a subset of cpus_allowed.
1579 if (cs->nr_subparts_cpus) {
1580 cpumask_and(cs->subparts_cpus, cs->subparts_cpus, cs->cpus_allowed);
1581 cs->nr_subparts_cpus = cpumask_weight(cs->subparts_cpus);
1583 spin_unlock_irq(&callback_lock);
1585 update_cpumasks_hier(cs, &tmp);
1587 if (cs->partition_root_state) {
1588 struct cpuset *parent = parent_cs(cs);
1591 * For partition root, update the cpumasks of sibling
1592 * cpusets if they use parent's effective_cpus.
1594 if (parent->child_ecpus_count)
1595 update_sibling_cpumasks(parent, cs, &tmp);
1601 * Migrate memory region from one set of nodes to another. This is
1602 * performed asynchronously as it can be called from process migration path
1603 * holding locks involved in process management. All mm migrations are
1604 * performed in the queued order and can be waited for by flushing
1605 * cpuset_migrate_mm_wq.
1608 struct cpuset_migrate_mm_work {
1609 struct work_struct work;
1610 struct mm_struct *mm;
1615 static void cpuset_migrate_mm_workfn(struct work_struct *work)
1617 struct cpuset_migrate_mm_work *mwork =
1618 container_of(work, struct cpuset_migrate_mm_work, work);
1620 /* on a wq worker, no need to worry about %current's mems_allowed */
1621 do_migrate_pages(mwork->mm, &mwork->from, &mwork->to, MPOL_MF_MOVE_ALL);
1626 static void cpuset_migrate_mm(struct mm_struct *mm, const nodemask_t *from,
1627 const nodemask_t *to)
1629 struct cpuset_migrate_mm_work *mwork;
1631 mwork = kzalloc(sizeof(*mwork), GFP_KERNEL);
1634 mwork->from = *from;
1636 INIT_WORK(&mwork->work, cpuset_migrate_mm_workfn);
1637 queue_work(cpuset_migrate_mm_wq, &mwork->work);
1643 static void cpuset_post_attach(void)
1645 flush_workqueue(cpuset_migrate_mm_wq);
1649 * cpuset_change_task_nodemask - change task's mems_allowed and mempolicy
1650 * @tsk: the task to change
1651 * @newmems: new nodes that the task will be set
1653 * We use the mems_allowed_seq seqlock to safely update both tsk->mems_allowed
1654 * and rebind an eventual tasks' mempolicy. If the task is allocating in
1655 * parallel, it might temporarily see an empty intersection, which results in
1656 * a seqlock check and retry before OOM or allocation failure.
1658 static void cpuset_change_task_nodemask(struct task_struct *tsk,
1659 nodemask_t *newmems)
1663 local_irq_disable();
1664 write_seqcount_begin(&tsk->mems_allowed_seq);
1666 nodes_or(tsk->mems_allowed, tsk->mems_allowed, *newmems);
1667 mpol_rebind_task(tsk, newmems);
1668 tsk->mems_allowed = *newmems;
1670 write_seqcount_end(&tsk->mems_allowed_seq);
1676 static void *cpuset_being_rebound;
1679 * update_tasks_nodemask - Update the nodemasks of tasks in the cpuset.
1680 * @cs: the cpuset in which each task's mems_allowed mask needs to be changed
1682 * Iterate through each task of @cs updating its mems_allowed to the
1683 * effective cpuset's. As this function is called with cpuset_mutex held,
1684 * cpuset membership stays stable.
1686 static void update_tasks_nodemask(struct cpuset *cs)
1688 static nodemask_t newmems; /* protected by cpuset_mutex */
1689 struct css_task_iter it;
1690 struct task_struct *task;
1692 cpuset_being_rebound = cs; /* causes mpol_dup() rebind */
1694 guarantee_online_mems(cs, &newmems);
1697 * The mpol_rebind_mm() call takes mmap_sem, which we couldn't
1698 * take while holding tasklist_lock. Forks can happen - the
1699 * mpol_dup() cpuset_being_rebound check will catch such forks,
1700 * and rebind their vma mempolicies too. Because we still hold
1701 * the global cpuset_mutex, we know that no other rebind effort
1702 * will be contending for the global variable cpuset_being_rebound.
1703 * It's ok if we rebind the same mm twice; mpol_rebind_mm()
1704 * is idempotent. Also migrate pages in each mm to new nodes.
1706 css_task_iter_start(&cs->css, 0, &it);
1707 while ((task = css_task_iter_next(&it))) {
1708 struct mm_struct *mm;
1711 cpuset_change_task_nodemask(task, &newmems);
1713 mm = get_task_mm(task);
1717 migrate = is_memory_migrate(cs);
1719 mpol_rebind_mm(mm, &cs->mems_allowed);
1721 cpuset_migrate_mm(mm, &cs->old_mems_allowed, &newmems);
1725 css_task_iter_end(&it);
1728 * All the tasks' nodemasks have been updated, update
1729 * cs->old_mems_allowed.
1731 cs->old_mems_allowed = newmems;
1733 /* We're done rebinding vmas to this cpuset's new mems_allowed. */
1734 cpuset_being_rebound = NULL;
1738 * update_nodemasks_hier - Update effective nodemasks and tasks in the subtree
1739 * @cs: the cpuset to consider
1740 * @new_mems: a temp variable for calculating new effective_mems
1742 * When configured nodemask is changed, the effective nodemasks of this cpuset
1743 * and all its descendants need to be updated.
1745 * On legacy hiearchy, effective_mems will be the same with mems_allowed.
1747 * Called with cpuset_mutex held
1749 static void update_nodemasks_hier(struct cpuset *cs, nodemask_t *new_mems)
1752 struct cgroup_subsys_state *pos_css;
1755 cpuset_for_each_descendant_pre(cp, pos_css, cs) {
1756 struct cpuset *parent = parent_cs(cp);
1758 nodes_and(*new_mems, cp->mems_allowed, parent->effective_mems);
1761 * If it becomes empty, inherit the effective mask of the
1762 * parent, which is guaranteed to have some MEMs.
1764 if (is_in_v2_mode() && nodes_empty(*new_mems))
1765 *new_mems = parent->effective_mems;
1767 /* Skip the whole subtree if the nodemask remains the same. */
1768 if (nodes_equal(*new_mems, cp->effective_mems)) {
1769 pos_css = css_rightmost_descendant(pos_css);
1773 if (!css_tryget_online(&cp->css))
1777 spin_lock_irq(&callback_lock);
1778 cp->effective_mems = *new_mems;
1779 spin_unlock_irq(&callback_lock);
1781 WARN_ON(!is_in_v2_mode() &&
1782 !nodes_equal(cp->mems_allowed, cp->effective_mems));
1784 update_tasks_nodemask(cp);
1793 * Handle user request to change the 'mems' memory placement
1794 * of a cpuset. Needs to validate the request, update the
1795 * cpusets mems_allowed, and for each task in the cpuset,
1796 * update mems_allowed and rebind task's mempolicy and any vma
1797 * mempolicies and if the cpuset is marked 'memory_migrate',
1798 * migrate the tasks pages to the new memory.
1800 * Call with cpuset_mutex held. May take callback_lock during call.
1801 * Will take tasklist_lock, scan tasklist for tasks in cpuset cs,
1802 * lock each such tasks mm->mmap_sem, scan its vma's and rebind
1803 * their mempolicies to the cpusets new mems_allowed.
1805 static int update_nodemask(struct cpuset *cs, struct cpuset *trialcs,
1811 * top_cpuset.mems_allowed tracks node_stats[N_MEMORY];
1814 if (cs == &top_cpuset) {
1820 * An empty mems_allowed is ok iff there are no tasks in the cpuset.
1821 * Since nodelist_parse() fails on an empty mask, we special case
1822 * that parsing. The validate_change() call ensures that cpusets
1823 * with tasks have memory.
1826 nodes_clear(trialcs->mems_allowed);
1828 retval = nodelist_parse(buf, trialcs->mems_allowed);
1832 if (!nodes_subset(trialcs->mems_allowed,
1833 top_cpuset.mems_allowed)) {
1839 if (nodes_equal(cs->mems_allowed, trialcs->mems_allowed)) {
1840 retval = 0; /* Too easy - nothing to do */
1843 retval = validate_change(cs, trialcs);
1847 spin_lock_irq(&callback_lock);
1848 cs->mems_allowed = trialcs->mems_allowed;
1849 spin_unlock_irq(&callback_lock);
1851 /* use trialcs->mems_allowed as a temp variable */
1852 update_nodemasks_hier(cs, &trialcs->mems_allowed);
1857 bool current_cpuset_is_being_rebound(void)
1862 ret = task_cs(current) == cpuset_being_rebound;
1868 static int update_relax_domain_level(struct cpuset *cs, s64 val)
1871 if (val < -1 || val >= sched_domain_level_max)
1875 if (val != cs->relax_domain_level) {
1876 cs->relax_domain_level = val;
1877 if (!cpumask_empty(cs->cpus_allowed) &&
1878 is_sched_load_balance(cs))
1879 rebuild_sched_domains_locked();
1886 * update_tasks_flags - update the spread flags of tasks in the cpuset.
1887 * @cs: the cpuset in which each task's spread flags needs to be changed
1889 * Iterate through each task of @cs updating its spread flags. As this
1890 * function is called with cpuset_mutex held, cpuset membership stays
1893 static void update_tasks_flags(struct cpuset *cs)
1895 struct css_task_iter it;
1896 struct task_struct *task;
1898 css_task_iter_start(&cs->css, 0, &it);
1899 while ((task = css_task_iter_next(&it)))
1900 cpuset_update_task_spread_flag(cs, task);
1901 css_task_iter_end(&it);
1905 * update_flag - read a 0 or a 1 in a file and update associated flag
1906 * bit: the bit to update (see cpuset_flagbits_t)
1907 * cs: the cpuset to update
1908 * turning_on: whether the flag is being set or cleared
1910 * Call with cpuset_mutex held.
1913 static int update_flag(cpuset_flagbits_t bit, struct cpuset *cs,
1916 struct cpuset *trialcs;
1917 int balance_flag_changed;
1918 int spread_flag_changed;
1921 trialcs = alloc_trial_cpuset(cs);
1926 set_bit(bit, &trialcs->flags);
1928 clear_bit(bit, &trialcs->flags);
1930 err = validate_change(cs, trialcs);
1934 balance_flag_changed = (is_sched_load_balance(cs) !=
1935 is_sched_load_balance(trialcs));
1937 spread_flag_changed = ((is_spread_slab(cs) != is_spread_slab(trialcs))
1938 || (is_spread_page(cs) != is_spread_page(trialcs)));
1940 spin_lock_irq(&callback_lock);
1941 cs->flags = trialcs->flags;
1942 spin_unlock_irq(&callback_lock);
1944 if (!cpumask_empty(trialcs->cpus_allowed) && balance_flag_changed)
1945 rebuild_sched_domains_locked();
1947 if (spread_flag_changed)
1948 update_tasks_flags(cs);
1950 free_cpuset(trialcs);
1955 * update_prstate - update partititon_root_state
1956 * cs: the cpuset to update
1957 * val: 0 - disabled, 1 - enabled
1959 * Call with cpuset_mutex held.
1961 static int update_prstate(struct cpuset *cs, int val)
1964 struct cpuset *parent = parent_cs(cs);
1965 struct tmpmasks tmp;
1967 if ((val != 0) && (val != 1))
1969 if (val == cs->partition_root_state)
1973 * Cannot force a partial or invalid partition root to a full
1976 if (val && cs->partition_root_state)
1979 if (alloc_cpumasks(NULL, &tmp))
1983 if (!cs->partition_root_state) {
1985 * Turning on partition root requires setting the
1986 * CS_CPU_EXCLUSIVE bit implicitly as well and cpus_allowed
1989 if (cpumask_empty(cs->cpus_allowed))
1992 err = update_flag(CS_CPU_EXCLUSIVE, cs, 1);
1996 err = update_parent_subparts_cpumask(cs, partcmd_enable,
1999 update_flag(CS_CPU_EXCLUSIVE, cs, 0);
2002 cs->partition_root_state = PRS_ENABLED;
2005 * Turning off partition root will clear the
2006 * CS_CPU_EXCLUSIVE bit.
2008 if (cs->partition_root_state == PRS_ERROR) {
2009 cs->partition_root_state = 0;
2010 update_flag(CS_CPU_EXCLUSIVE, cs, 0);
2015 err = update_parent_subparts_cpumask(cs, partcmd_disable,
2020 cs->partition_root_state = 0;
2022 /* Turning off CS_CPU_EXCLUSIVE will not return error */
2023 update_flag(CS_CPU_EXCLUSIVE, cs, 0);
2026 update_tasks_cpumask(parent);
2028 if (parent->child_ecpus_count)
2029 update_sibling_cpumasks(parent, cs, &tmp);
2031 rebuild_sched_domains_locked();
2033 free_cpumasks(NULL, &tmp);
2038 * Frequency meter - How fast is some event occurring?
2040 * These routines manage a digitally filtered, constant time based,
2041 * event frequency meter. There are four routines:
2042 * fmeter_init() - initialize a frequency meter.
2043 * fmeter_markevent() - called each time the event happens.
2044 * fmeter_getrate() - returns the recent rate of such events.
2045 * fmeter_update() - internal routine used to update fmeter.
2047 * A common data structure is passed to each of these routines,
2048 * which is used to keep track of the state required to manage the
2049 * frequency meter and its digital filter.
2051 * The filter works on the number of events marked per unit time.
2052 * The filter is single-pole low-pass recursive (IIR). The time unit
2053 * is 1 second. Arithmetic is done using 32-bit integers scaled to
2054 * simulate 3 decimal digits of precision (multiplied by 1000).
2056 * With an FM_COEF of 933, and a time base of 1 second, the filter
2057 * has a half-life of 10 seconds, meaning that if the events quit
2058 * happening, then the rate returned from the fmeter_getrate()
2059 * will be cut in half each 10 seconds, until it converges to zero.
2061 * It is not worth doing a real infinitely recursive filter. If more
2062 * than FM_MAXTICKS ticks have elapsed since the last filter event,
2063 * just compute FM_MAXTICKS ticks worth, by which point the level
2066 * Limit the count of unprocessed events to FM_MAXCNT, so as to avoid
2067 * arithmetic overflow in the fmeter_update() routine.
2069 * Given the simple 32 bit integer arithmetic used, this meter works
2070 * best for reporting rates between one per millisecond (msec) and
2071 * one per 32 (approx) seconds. At constant rates faster than one
2072 * per msec it maxes out at values just under 1,000,000. At constant
2073 * rates between one per msec, and one per second it will stabilize
2074 * to a value N*1000, where N is the rate of events per second.
2075 * At constant rates between one per second and one per 32 seconds,
2076 * it will be choppy, moving up on the seconds that have an event,
2077 * and then decaying until the next event. At rates slower than
2078 * about one in 32 seconds, it decays all the way back to zero between
2082 #define FM_COEF 933 /* coefficient for half-life of 10 secs */
2083 #define FM_MAXTICKS ((u32)99) /* useless computing more ticks than this */
2084 #define FM_MAXCNT 1000000 /* limit cnt to avoid overflow */
2085 #define FM_SCALE 1000 /* faux fixed point scale */
2087 /* Initialize a frequency meter */
2088 static void fmeter_init(struct fmeter *fmp)
2093 spin_lock_init(&fmp->lock);
2096 /* Internal meter update - process cnt events and update value */
2097 static void fmeter_update(struct fmeter *fmp)
2102 now = ktime_get_seconds();
2103 ticks = now - fmp->time;
2108 ticks = min(FM_MAXTICKS, ticks);
2110 fmp->val = (FM_COEF * fmp->val) / FM_SCALE;
2113 fmp->val += ((FM_SCALE - FM_COEF) * fmp->cnt) / FM_SCALE;
2117 /* Process any previous ticks, then bump cnt by one (times scale). */
2118 static void fmeter_markevent(struct fmeter *fmp)
2120 spin_lock(&fmp->lock);
2122 fmp->cnt = min(FM_MAXCNT, fmp->cnt + FM_SCALE);
2123 spin_unlock(&fmp->lock);
2126 /* Process any previous ticks, then return current value. */
2127 static int fmeter_getrate(struct fmeter *fmp)
2131 spin_lock(&fmp->lock);
2134 spin_unlock(&fmp->lock);
2138 static struct cpuset *cpuset_attach_old_cs;
2140 /* Called by cgroups to determine if a cpuset is usable; cpuset_mutex held */
2141 static int cpuset_can_attach(struct cgroup_taskset *tset)
2143 struct cgroup_subsys_state *css;
2145 struct task_struct *task;
2148 /* used later by cpuset_attach() */
2149 cpuset_attach_old_cs = task_cs(cgroup_taskset_first(tset, &css));
2152 percpu_down_write(&cpuset_rwsem);
2154 /* allow moving tasks into an empty cpuset if on default hierarchy */
2156 if (!is_in_v2_mode() &&
2157 (cpumask_empty(cs->cpus_allowed) || nodes_empty(cs->mems_allowed)))
2160 cgroup_taskset_for_each(task, css, tset) {
2161 ret = task_can_attach(task, cs->cpus_allowed);
2164 ret = security_task_setscheduler(task);
2170 * Mark attach is in progress. This makes validate_change() fail
2171 * changes which zero cpus/mems_allowed.
2173 cs->attach_in_progress++;
2176 percpu_up_write(&cpuset_rwsem);
2180 static void cpuset_cancel_attach(struct cgroup_taskset *tset)
2182 struct cgroup_subsys_state *css;
2185 cgroup_taskset_first(tset, &css);
2188 percpu_down_write(&cpuset_rwsem);
2189 cs->attach_in_progress--;
2190 if (!cs->attach_in_progress)
2191 wake_up(&cpuset_attach_wq);
2192 percpu_up_write(&cpuset_rwsem);
2196 * Protected by cpuset_mutex. cpus_attach is used only by cpuset_attach()
2197 * but we can't allocate it dynamically there. Define it global and
2198 * allocate from cpuset_init().
2200 static cpumask_var_t cpus_attach;
2202 static void cpuset_attach(struct cgroup_taskset *tset)
2204 /* static buf protected by cpuset_mutex */
2205 static nodemask_t cpuset_attach_nodemask_to;
2206 struct task_struct *task;
2207 struct task_struct *leader;
2208 struct cgroup_subsys_state *css;
2210 struct cpuset *oldcs = cpuset_attach_old_cs;
2212 cgroup_taskset_first(tset, &css);
2215 lockdep_assert_cpus_held(); /* see cgroup_attach_lock() */
2216 percpu_down_write(&cpuset_rwsem);
2218 /* prepare for attach */
2219 if (cs == &top_cpuset)
2220 cpumask_copy(cpus_attach, cpu_possible_mask);
2222 guarantee_online_cpus(cs, cpus_attach);
2224 guarantee_online_mems(cs, &cpuset_attach_nodemask_to);
2226 cgroup_taskset_for_each(task, css, tset) {
2228 * can_attach beforehand should guarantee that this doesn't
2229 * fail. TODO: have a better way to handle failure here
2231 WARN_ON_ONCE(set_cpus_allowed_ptr(task, cpus_attach));
2233 cpuset_change_task_nodemask(task, &cpuset_attach_nodemask_to);
2234 cpuset_update_task_spread_flag(cs, task);
2238 * Change mm for all threadgroup leaders. This is expensive and may
2239 * sleep and should be moved outside migration path proper.
2241 cpuset_attach_nodemask_to = cs->effective_mems;
2242 cgroup_taskset_for_each_leader(leader, css, tset) {
2243 struct mm_struct *mm = get_task_mm(leader);
2246 mpol_rebind_mm(mm, &cpuset_attach_nodemask_to);
2249 * old_mems_allowed is the same with mems_allowed
2250 * here, except if this task is being moved
2251 * automatically due to hotplug. In that case
2252 * @mems_allowed has been updated and is empty, so
2253 * @old_mems_allowed is the right nodesets that we
2256 if (is_memory_migrate(cs))
2257 cpuset_migrate_mm(mm, &oldcs->old_mems_allowed,
2258 &cpuset_attach_nodemask_to);
2264 cs->old_mems_allowed = cpuset_attach_nodemask_to;
2266 cs->attach_in_progress--;
2267 if (!cs->attach_in_progress)
2268 wake_up(&cpuset_attach_wq);
2270 percpu_up_write(&cpuset_rwsem);
2273 /* The various types of files and directories in a cpuset file system */
2276 FILE_MEMORY_MIGRATE,
2279 FILE_EFFECTIVE_CPULIST,
2280 FILE_EFFECTIVE_MEMLIST,
2281 FILE_SUBPARTS_CPULIST,
2285 FILE_SCHED_LOAD_BALANCE,
2286 FILE_PARTITION_ROOT,
2287 FILE_SCHED_RELAX_DOMAIN_LEVEL,
2288 FILE_MEMORY_PRESSURE_ENABLED,
2289 FILE_MEMORY_PRESSURE,
2292 } cpuset_filetype_t;
2294 static int cpuset_write_u64(struct cgroup_subsys_state *css, struct cftype *cft,
2297 struct cpuset *cs = css_cs(css);
2298 cpuset_filetype_t type = cft->private;
2302 percpu_down_write(&cpuset_rwsem);
2303 if (!is_cpuset_online(cs)) {
2309 case FILE_CPU_EXCLUSIVE:
2310 retval = update_flag(CS_CPU_EXCLUSIVE, cs, val);
2312 case FILE_MEM_EXCLUSIVE:
2313 retval = update_flag(CS_MEM_EXCLUSIVE, cs, val);
2315 case FILE_MEM_HARDWALL:
2316 retval = update_flag(CS_MEM_HARDWALL, cs, val);
2318 case FILE_SCHED_LOAD_BALANCE:
2319 retval = update_flag(CS_SCHED_LOAD_BALANCE, cs, val);
2321 case FILE_MEMORY_MIGRATE:
2322 retval = update_flag(CS_MEMORY_MIGRATE, cs, val);
2324 case FILE_MEMORY_PRESSURE_ENABLED:
2325 cpuset_memory_pressure_enabled = !!val;
2327 case FILE_SPREAD_PAGE:
2328 retval = update_flag(CS_SPREAD_PAGE, cs, val);
2330 case FILE_SPREAD_SLAB:
2331 retval = update_flag(CS_SPREAD_SLAB, cs, val);
2338 percpu_up_write(&cpuset_rwsem);
2343 static int cpuset_write_s64(struct cgroup_subsys_state *css, struct cftype *cft,
2346 struct cpuset *cs = css_cs(css);
2347 cpuset_filetype_t type = cft->private;
2348 int retval = -ENODEV;
2351 percpu_down_write(&cpuset_rwsem);
2352 if (!is_cpuset_online(cs))
2356 case FILE_SCHED_RELAX_DOMAIN_LEVEL:
2357 retval = update_relax_domain_level(cs, val);
2364 percpu_up_write(&cpuset_rwsem);
2370 * Common handling for a write to a "cpus" or "mems" file.
2372 static ssize_t cpuset_write_resmask(struct kernfs_open_file *of,
2373 char *buf, size_t nbytes, loff_t off)
2375 struct cpuset *cs = css_cs(of_css(of));
2376 struct cpuset *trialcs;
2377 int retval = -ENODEV;
2379 buf = strstrip(buf);
2382 * CPU or memory hotunplug may leave @cs w/o any execution
2383 * resources, in which case the hotplug code asynchronously updates
2384 * configuration and transfers all tasks to the nearest ancestor
2385 * which can execute.
2387 * As writes to "cpus" or "mems" may restore @cs's execution
2388 * resources, wait for the previously scheduled operations before
2389 * proceeding, so that we don't end up keep removing tasks added
2390 * after execution capability is restored.
2392 * cpuset_hotplug_work calls back into cgroup core via
2393 * cgroup_transfer_tasks() and waiting for it from a cgroupfs
2394 * operation like this one can lead to a deadlock through kernfs
2395 * active_ref protection. Let's break the protection. Losing the
2396 * protection is okay as we check whether @cs is online after
2397 * grabbing cpuset_mutex anyway. This only happens on the legacy
2401 kernfs_break_active_protection(of->kn);
2402 flush_work(&cpuset_hotplug_work);
2405 percpu_down_write(&cpuset_rwsem);
2406 if (!is_cpuset_online(cs))
2409 trialcs = alloc_trial_cpuset(cs);
2415 switch (of_cft(of)->private) {
2417 retval = update_cpumask(cs, trialcs, buf);
2420 retval = update_nodemask(cs, trialcs, buf);
2427 free_cpuset(trialcs);
2429 percpu_up_write(&cpuset_rwsem);
2431 kernfs_unbreak_active_protection(of->kn);
2433 flush_workqueue(cpuset_migrate_mm_wq);
2434 return retval ?: nbytes;
2438 * These ascii lists should be read in a single call, by using a user
2439 * buffer large enough to hold the entire map. If read in smaller
2440 * chunks, there is no guarantee of atomicity. Since the display format
2441 * used, list of ranges of sequential numbers, is variable length,
2442 * and since these maps can change value dynamically, one could read
2443 * gibberish by doing partial reads while a list was changing.
2445 static int cpuset_common_seq_show(struct seq_file *sf, void *v)
2447 struct cpuset *cs = css_cs(seq_css(sf));
2448 cpuset_filetype_t type = seq_cft(sf)->private;
2451 spin_lock_irq(&callback_lock);
2455 seq_printf(sf, "%*pbl\n", cpumask_pr_args(cs->cpus_allowed));
2458 seq_printf(sf, "%*pbl\n", nodemask_pr_args(&cs->mems_allowed));
2460 case FILE_EFFECTIVE_CPULIST:
2461 seq_printf(sf, "%*pbl\n", cpumask_pr_args(cs->effective_cpus));
2463 case FILE_EFFECTIVE_MEMLIST:
2464 seq_printf(sf, "%*pbl\n", nodemask_pr_args(&cs->effective_mems));
2466 case FILE_SUBPARTS_CPULIST:
2467 seq_printf(sf, "%*pbl\n", cpumask_pr_args(cs->subparts_cpus));
2473 spin_unlock_irq(&callback_lock);
2477 static u64 cpuset_read_u64(struct cgroup_subsys_state *css, struct cftype *cft)
2479 struct cpuset *cs = css_cs(css);
2480 cpuset_filetype_t type = cft->private;
2482 case FILE_CPU_EXCLUSIVE:
2483 return is_cpu_exclusive(cs);
2484 case FILE_MEM_EXCLUSIVE:
2485 return is_mem_exclusive(cs);
2486 case FILE_MEM_HARDWALL:
2487 return is_mem_hardwall(cs);
2488 case FILE_SCHED_LOAD_BALANCE:
2489 return is_sched_load_balance(cs);
2490 case FILE_MEMORY_MIGRATE:
2491 return is_memory_migrate(cs);
2492 case FILE_MEMORY_PRESSURE_ENABLED:
2493 return cpuset_memory_pressure_enabled;
2494 case FILE_MEMORY_PRESSURE:
2495 return fmeter_getrate(&cs->fmeter);
2496 case FILE_SPREAD_PAGE:
2497 return is_spread_page(cs);
2498 case FILE_SPREAD_SLAB:
2499 return is_spread_slab(cs);
2504 /* Unreachable but makes gcc happy */
2508 static s64 cpuset_read_s64(struct cgroup_subsys_state *css, struct cftype *cft)
2510 struct cpuset *cs = css_cs(css);
2511 cpuset_filetype_t type = cft->private;
2513 case FILE_SCHED_RELAX_DOMAIN_LEVEL:
2514 return cs->relax_domain_level;
2519 /* Unrechable but makes gcc happy */
2523 static int sched_partition_show(struct seq_file *seq, void *v)
2525 struct cpuset *cs = css_cs(seq_css(seq));
2527 switch (cs->partition_root_state) {
2529 seq_puts(seq, "root\n");
2532 seq_puts(seq, "member\n");
2535 seq_puts(seq, "root invalid\n");
2541 static ssize_t sched_partition_write(struct kernfs_open_file *of, char *buf,
2542 size_t nbytes, loff_t off)
2544 struct cpuset *cs = css_cs(of_css(of));
2546 int retval = -ENODEV;
2548 buf = strstrip(buf);
2551 * Convert "root" to ENABLED, and convert "member" to DISABLED.
2553 if (!strcmp(buf, "root"))
2555 else if (!strcmp(buf, "member"))
2562 percpu_down_write(&cpuset_rwsem);
2563 if (!is_cpuset_online(cs))
2566 retval = update_prstate(cs, val);
2568 percpu_up_write(&cpuset_rwsem);
2571 return retval ?: nbytes;
2575 * for the common functions, 'private' gives the type of file
2578 static struct cftype legacy_files[] = {
2581 .seq_show = cpuset_common_seq_show,
2582 .write = cpuset_write_resmask,
2583 .max_write_len = (100U + 6 * NR_CPUS),
2584 .private = FILE_CPULIST,
2589 .seq_show = cpuset_common_seq_show,
2590 .write = cpuset_write_resmask,
2591 .max_write_len = (100U + 6 * MAX_NUMNODES),
2592 .private = FILE_MEMLIST,
2596 .name = "effective_cpus",
2597 .seq_show = cpuset_common_seq_show,
2598 .private = FILE_EFFECTIVE_CPULIST,
2602 .name = "effective_mems",
2603 .seq_show = cpuset_common_seq_show,
2604 .private = FILE_EFFECTIVE_MEMLIST,
2608 .name = "cpu_exclusive",
2609 .read_u64 = cpuset_read_u64,
2610 .write_u64 = cpuset_write_u64,
2611 .private = FILE_CPU_EXCLUSIVE,
2615 .name = "mem_exclusive",
2616 .read_u64 = cpuset_read_u64,
2617 .write_u64 = cpuset_write_u64,
2618 .private = FILE_MEM_EXCLUSIVE,
2622 .name = "mem_hardwall",
2623 .read_u64 = cpuset_read_u64,
2624 .write_u64 = cpuset_write_u64,
2625 .private = FILE_MEM_HARDWALL,
2629 .name = "sched_load_balance",
2630 .read_u64 = cpuset_read_u64,
2631 .write_u64 = cpuset_write_u64,
2632 .private = FILE_SCHED_LOAD_BALANCE,
2636 .name = "sched_relax_domain_level",
2637 .read_s64 = cpuset_read_s64,
2638 .write_s64 = cpuset_write_s64,
2639 .private = FILE_SCHED_RELAX_DOMAIN_LEVEL,
2643 .name = "memory_migrate",
2644 .read_u64 = cpuset_read_u64,
2645 .write_u64 = cpuset_write_u64,
2646 .private = FILE_MEMORY_MIGRATE,
2650 .name = "memory_pressure",
2651 .read_u64 = cpuset_read_u64,
2652 .private = FILE_MEMORY_PRESSURE,
2656 .name = "memory_spread_page",
2657 .read_u64 = cpuset_read_u64,
2658 .write_u64 = cpuset_write_u64,
2659 .private = FILE_SPREAD_PAGE,
2663 .name = "memory_spread_slab",
2664 .read_u64 = cpuset_read_u64,
2665 .write_u64 = cpuset_write_u64,
2666 .private = FILE_SPREAD_SLAB,
2670 .name = "memory_pressure_enabled",
2671 .flags = CFTYPE_ONLY_ON_ROOT,
2672 .read_u64 = cpuset_read_u64,
2673 .write_u64 = cpuset_write_u64,
2674 .private = FILE_MEMORY_PRESSURE_ENABLED,
2681 * This is currently a minimal set for the default hierarchy. It can be
2682 * expanded later on by migrating more features and control files from v1.
2684 static struct cftype dfl_files[] = {
2687 .seq_show = cpuset_common_seq_show,
2688 .write = cpuset_write_resmask,
2689 .max_write_len = (100U + 6 * NR_CPUS),
2690 .private = FILE_CPULIST,
2691 .flags = CFTYPE_NOT_ON_ROOT,
2696 .seq_show = cpuset_common_seq_show,
2697 .write = cpuset_write_resmask,
2698 .max_write_len = (100U + 6 * MAX_NUMNODES),
2699 .private = FILE_MEMLIST,
2700 .flags = CFTYPE_NOT_ON_ROOT,
2704 .name = "cpus.effective",
2705 .seq_show = cpuset_common_seq_show,
2706 .private = FILE_EFFECTIVE_CPULIST,
2710 .name = "mems.effective",
2711 .seq_show = cpuset_common_seq_show,
2712 .private = FILE_EFFECTIVE_MEMLIST,
2716 .name = "cpus.partition",
2717 .seq_show = sched_partition_show,
2718 .write = sched_partition_write,
2719 .private = FILE_PARTITION_ROOT,
2720 .flags = CFTYPE_NOT_ON_ROOT,
2724 .name = "cpus.subpartitions",
2725 .seq_show = cpuset_common_seq_show,
2726 .private = FILE_SUBPARTS_CPULIST,
2727 .flags = CFTYPE_DEBUG,
2735 * cpuset_css_alloc - allocate a cpuset css
2736 * cgrp: control group that the new cpuset will be part of
2739 static struct cgroup_subsys_state *
2740 cpuset_css_alloc(struct cgroup_subsys_state *parent_css)
2745 return &top_cpuset.css;
2747 cs = kzalloc(sizeof(*cs), GFP_KERNEL);
2749 return ERR_PTR(-ENOMEM);
2751 if (alloc_cpumasks(cs, NULL)) {
2753 return ERR_PTR(-ENOMEM);
2756 set_bit(CS_SCHED_LOAD_BALANCE, &cs->flags);
2757 nodes_clear(cs->mems_allowed);
2758 nodes_clear(cs->effective_mems);
2759 fmeter_init(&cs->fmeter);
2760 cs->relax_domain_level = -1;
2765 static int cpuset_css_online(struct cgroup_subsys_state *css)
2767 struct cpuset *cs = css_cs(css);
2768 struct cpuset *parent = parent_cs(cs);
2769 struct cpuset *tmp_cs;
2770 struct cgroup_subsys_state *pos_css;
2776 percpu_down_write(&cpuset_rwsem);
2778 set_bit(CS_ONLINE, &cs->flags);
2779 if (is_spread_page(parent))
2780 set_bit(CS_SPREAD_PAGE, &cs->flags);
2781 if (is_spread_slab(parent))
2782 set_bit(CS_SPREAD_SLAB, &cs->flags);
2786 spin_lock_irq(&callback_lock);
2787 if (is_in_v2_mode()) {
2788 cpumask_copy(cs->effective_cpus, parent->effective_cpus);
2789 cs->effective_mems = parent->effective_mems;
2790 cs->use_parent_ecpus = true;
2791 parent->child_ecpus_count++;
2793 spin_unlock_irq(&callback_lock);
2795 if (!test_bit(CGRP_CPUSET_CLONE_CHILDREN, &css->cgroup->flags))
2799 * Clone @parent's configuration if CGRP_CPUSET_CLONE_CHILDREN is
2800 * set. This flag handling is implemented in cgroup core for
2801 * histrical reasons - the flag may be specified during mount.
2803 * Currently, if any sibling cpusets have exclusive cpus or mem, we
2804 * refuse to clone the configuration - thereby refusing the task to
2805 * be entered, and as a result refusing the sys_unshare() or
2806 * clone() which initiated it. If this becomes a problem for some
2807 * users who wish to allow that scenario, then this could be
2808 * changed to grant parent->cpus_allowed-sibling_cpus_exclusive
2809 * (and likewise for mems) to the new cgroup.
2812 cpuset_for_each_child(tmp_cs, pos_css, parent) {
2813 if (is_mem_exclusive(tmp_cs) || is_cpu_exclusive(tmp_cs)) {
2820 spin_lock_irq(&callback_lock);
2821 cs->mems_allowed = parent->mems_allowed;
2822 cs->effective_mems = parent->mems_allowed;
2823 cpumask_copy(cs->cpus_allowed, parent->cpus_allowed);
2824 cpumask_copy(cs->effective_cpus, parent->cpus_allowed);
2825 spin_unlock_irq(&callback_lock);
2827 percpu_up_write(&cpuset_rwsem);
2833 * If the cpuset being removed has its flag 'sched_load_balance'
2834 * enabled, then simulate turning sched_load_balance off, which
2835 * will call rebuild_sched_domains_locked(). That is not needed
2836 * in the default hierarchy where only changes in partition
2837 * will cause repartitioning.
2839 * If the cpuset has the 'sched.partition' flag enabled, simulate
2840 * turning 'sched.partition" off.
2843 static void cpuset_css_offline(struct cgroup_subsys_state *css)
2845 struct cpuset *cs = css_cs(css);
2848 percpu_down_write(&cpuset_rwsem);
2850 if (is_partition_root(cs))
2851 update_prstate(cs, 0);
2853 if (!cgroup_subsys_on_dfl(cpuset_cgrp_subsys) &&
2854 is_sched_load_balance(cs))
2855 update_flag(CS_SCHED_LOAD_BALANCE, cs, 0);
2857 if (cs->use_parent_ecpus) {
2858 struct cpuset *parent = parent_cs(cs);
2860 cs->use_parent_ecpus = false;
2861 parent->child_ecpus_count--;
2865 clear_bit(CS_ONLINE, &cs->flags);
2867 percpu_up_write(&cpuset_rwsem);
2871 static void cpuset_css_free(struct cgroup_subsys_state *css)
2873 struct cpuset *cs = css_cs(css);
2878 static void cpuset_bind(struct cgroup_subsys_state *root_css)
2880 percpu_down_write(&cpuset_rwsem);
2881 spin_lock_irq(&callback_lock);
2883 if (is_in_v2_mode()) {
2884 cpumask_copy(top_cpuset.cpus_allowed, cpu_possible_mask);
2885 top_cpuset.mems_allowed = node_possible_map;
2887 cpumask_copy(top_cpuset.cpus_allowed,
2888 top_cpuset.effective_cpus);
2889 top_cpuset.mems_allowed = top_cpuset.effective_mems;
2892 spin_unlock_irq(&callback_lock);
2893 percpu_up_write(&cpuset_rwsem);
2897 * Make sure the new task conform to the current state of its parent,
2898 * which could have been changed by cpuset just after it inherits the
2899 * state from the parent and before it sits on the cgroup's task list.
2901 static void cpuset_fork(struct task_struct *task)
2903 if (task_css_is_root(task, cpuset_cgrp_id))
2906 set_cpus_allowed_ptr(task, current->cpus_ptr);
2907 task->mems_allowed = current->mems_allowed;
2910 struct cgroup_subsys cpuset_cgrp_subsys = {
2911 .css_alloc = cpuset_css_alloc,
2912 .css_online = cpuset_css_online,
2913 .css_offline = cpuset_css_offline,
2914 .css_free = cpuset_css_free,
2915 .can_attach = cpuset_can_attach,
2916 .cancel_attach = cpuset_cancel_attach,
2917 .attach = cpuset_attach,
2918 .post_attach = cpuset_post_attach,
2919 .bind = cpuset_bind,
2920 .fork = cpuset_fork,
2921 .legacy_cftypes = legacy_files,
2922 .dfl_cftypes = dfl_files,
2928 * cpuset_init - initialize cpusets at system boot
2930 * Description: Initialize top_cpuset
2933 int __init cpuset_init(void)
2935 BUG_ON(percpu_init_rwsem(&cpuset_rwsem));
2937 BUG_ON(!alloc_cpumask_var(&top_cpuset.cpus_allowed, GFP_KERNEL));
2938 BUG_ON(!alloc_cpumask_var(&top_cpuset.effective_cpus, GFP_KERNEL));
2939 BUG_ON(!zalloc_cpumask_var(&top_cpuset.subparts_cpus, GFP_KERNEL));
2941 cpumask_setall(top_cpuset.cpus_allowed);
2942 nodes_setall(top_cpuset.mems_allowed);
2943 cpumask_setall(top_cpuset.effective_cpus);
2944 nodes_setall(top_cpuset.effective_mems);
2946 fmeter_init(&top_cpuset.fmeter);
2947 set_bit(CS_SCHED_LOAD_BALANCE, &top_cpuset.flags);
2948 top_cpuset.relax_domain_level = -1;
2950 BUG_ON(!alloc_cpumask_var(&cpus_attach, GFP_KERNEL));
2956 * If CPU and/or memory hotplug handlers, below, unplug any CPUs
2957 * or memory nodes, we need to walk over the cpuset hierarchy,
2958 * removing that CPU or node from all cpusets. If this removes the
2959 * last CPU or node from a cpuset, then move the tasks in the empty
2960 * cpuset to its next-highest non-empty parent.
2962 static void remove_tasks_in_empty_cpuset(struct cpuset *cs)
2964 struct cpuset *parent;
2967 * Find its next-highest non-empty parent, (top cpuset
2968 * has online cpus, so can't be empty).
2970 parent = parent_cs(cs);
2971 while (cpumask_empty(parent->cpus_allowed) ||
2972 nodes_empty(parent->mems_allowed))
2973 parent = parent_cs(parent);
2975 if (cgroup_transfer_tasks(parent->css.cgroup, cs->css.cgroup)) {
2976 pr_err("cpuset: failed to transfer tasks out of empty cpuset ");
2977 pr_cont_cgroup_name(cs->css.cgroup);
2983 hotplug_update_tasks_legacy(struct cpuset *cs,
2984 struct cpumask *new_cpus, nodemask_t *new_mems,
2985 bool cpus_updated, bool mems_updated)
2989 spin_lock_irq(&callback_lock);
2990 cpumask_copy(cs->cpus_allowed, new_cpus);
2991 cpumask_copy(cs->effective_cpus, new_cpus);
2992 cs->mems_allowed = *new_mems;
2993 cs->effective_mems = *new_mems;
2994 spin_unlock_irq(&callback_lock);
2997 * Don't call update_tasks_cpumask() if the cpuset becomes empty,
2998 * as the tasks will be migratecd to an ancestor.
3000 if (cpus_updated && !cpumask_empty(cs->cpus_allowed))
3001 update_tasks_cpumask(cs);
3002 if (mems_updated && !nodes_empty(cs->mems_allowed))
3003 update_tasks_nodemask(cs);
3005 is_empty = cpumask_empty(cs->cpus_allowed) ||
3006 nodes_empty(cs->mems_allowed);
3008 percpu_up_write(&cpuset_rwsem);
3011 * Move tasks to the nearest ancestor with execution resources,
3012 * This is full cgroup operation which will also call back into
3013 * cpuset. Should be done outside any lock.
3016 remove_tasks_in_empty_cpuset(cs);
3018 percpu_down_write(&cpuset_rwsem);
3022 hotplug_update_tasks(struct cpuset *cs,
3023 struct cpumask *new_cpus, nodemask_t *new_mems,
3024 bool cpus_updated, bool mems_updated)
3026 if (cpumask_empty(new_cpus))
3027 cpumask_copy(new_cpus, parent_cs(cs)->effective_cpus);
3028 if (nodes_empty(*new_mems))
3029 *new_mems = parent_cs(cs)->effective_mems;
3031 spin_lock_irq(&callback_lock);
3032 cpumask_copy(cs->effective_cpus, new_cpus);
3033 cs->effective_mems = *new_mems;
3034 spin_unlock_irq(&callback_lock);
3037 update_tasks_cpumask(cs);
3039 update_tasks_nodemask(cs);
3042 static bool force_rebuild;
3044 void cpuset_force_rebuild(void)
3046 force_rebuild = true;
3050 * cpuset_hotplug_update_tasks - update tasks in a cpuset for hotunplug
3051 * @cs: cpuset in interest
3052 * @tmp: the tmpmasks structure pointer
3054 * Compare @cs's cpu and mem masks against top_cpuset and if some have gone
3055 * offline, update @cs accordingly. If @cs ends up with no CPU or memory,
3056 * all its tasks are moved to the nearest ancestor with both resources.
3058 static void cpuset_hotplug_update_tasks(struct cpuset *cs, struct tmpmasks *tmp)
3060 static cpumask_t new_cpus;
3061 static nodemask_t new_mems;
3064 struct cpuset *parent;
3066 wait_event(cpuset_attach_wq, cs->attach_in_progress == 0);
3068 percpu_down_write(&cpuset_rwsem);
3071 * We have raced with task attaching. We wait until attaching
3072 * is finished, so we won't attach a task to an empty cpuset.
3074 if (cs->attach_in_progress) {
3075 percpu_up_write(&cpuset_rwsem);
3079 parent = parent_cs(cs);
3080 compute_effective_cpumask(&new_cpus, cs, parent);
3081 nodes_and(new_mems, cs->mems_allowed, parent->effective_mems);
3083 if (cs->nr_subparts_cpus)
3085 * Make sure that CPUs allocated to child partitions
3086 * do not show up in effective_cpus.
3088 cpumask_andnot(&new_cpus, &new_cpus, cs->subparts_cpus);
3090 if (!tmp || !cs->partition_root_state)
3094 * In the unlikely event that a partition root has empty
3095 * effective_cpus or its parent becomes erroneous, we have to
3096 * transition it to the erroneous state.
3098 if (is_partition_root(cs) && (cpumask_empty(&new_cpus) ||
3099 (parent->partition_root_state == PRS_ERROR))) {
3100 if (cs->nr_subparts_cpus) {
3101 cs->nr_subparts_cpus = 0;
3102 cpumask_clear(cs->subparts_cpus);
3103 compute_effective_cpumask(&new_cpus, cs, parent);
3107 * If the effective_cpus is empty because the child
3108 * partitions take away all the CPUs, we can keep
3109 * the current partition and let the child partitions
3110 * fight for available CPUs.
3112 if ((parent->partition_root_state == PRS_ERROR) ||
3113 cpumask_empty(&new_cpus)) {
3114 update_parent_subparts_cpumask(cs, partcmd_disable,
3116 cs->partition_root_state = PRS_ERROR;
3118 cpuset_force_rebuild();
3122 * On the other hand, an erroneous partition root may be transitioned
3123 * back to a regular one or a partition root with no CPU allocated
3124 * from the parent may change to erroneous.
3126 if (is_partition_root(parent) &&
3127 ((cs->partition_root_state == PRS_ERROR) ||
3128 !cpumask_intersects(&new_cpus, parent->subparts_cpus)) &&
3129 update_parent_subparts_cpumask(cs, partcmd_update, NULL, tmp))
3130 cpuset_force_rebuild();
3133 cpus_updated = !cpumask_equal(&new_cpus, cs->effective_cpus);
3134 mems_updated = !nodes_equal(new_mems, cs->effective_mems);
3136 if (is_in_v2_mode())
3137 hotplug_update_tasks(cs, &new_cpus, &new_mems,
3138 cpus_updated, mems_updated);
3140 hotplug_update_tasks_legacy(cs, &new_cpus, &new_mems,
3141 cpus_updated, mems_updated);
3143 percpu_up_write(&cpuset_rwsem);
3147 * cpuset_hotplug_workfn - handle CPU/memory hotunplug for a cpuset
3149 * This function is called after either CPU or memory configuration has
3150 * changed and updates cpuset accordingly. The top_cpuset is always
3151 * synchronized to cpu_active_mask and N_MEMORY, which is necessary in
3152 * order to make cpusets transparent (of no affect) on systems that are
3153 * actively using CPU hotplug but making no active use of cpusets.
3155 * Non-root cpusets are only affected by offlining. If any CPUs or memory
3156 * nodes have been taken down, cpuset_hotplug_update_tasks() is invoked on
3159 * Note that CPU offlining during suspend is ignored. We don't modify
3160 * cpusets across suspend/resume cycles at all.
3162 static void cpuset_hotplug_workfn(struct work_struct *work)
3164 static cpumask_t new_cpus;
3165 static nodemask_t new_mems;
3166 bool cpus_updated, mems_updated;
3167 bool on_dfl = is_in_v2_mode();
3168 struct tmpmasks tmp, *ptmp = NULL;
3170 if (on_dfl && !alloc_cpumasks(NULL, &tmp))
3173 percpu_down_write(&cpuset_rwsem);
3175 /* fetch the available cpus/mems and find out which changed how */
3176 cpumask_copy(&new_cpus, cpu_active_mask);
3177 new_mems = node_states[N_MEMORY];
3180 * If subparts_cpus is populated, it is likely that the check below
3181 * will produce a false positive on cpus_updated when the cpu list
3182 * isn't changed. It is extra work, but it is better to be safe.
3184 cpus_updated = !cpumask_equal(top_cpuset.effective_cpus, &new_cpus);
3185 mems_updated = !nodes_equal(top_cpuset.effective_mems, new_mems);
3188 * In the rare case that hotplug removes all the cpus in subparts_cpus,
3189 * we assumed that cpus are updated.
3191 if (!cpus_updated && top_cpuset.nr_subparts_cpus)
3192 cpus_updated = true;
3194 /* synchronize cpus_allowed to cpu_active_mask */
3196 spin_lock_irq(&callback_lock);
3198 cpumask_copy(top_cpuset.cpus_allowed, &new_cpus);
3200 * Make sure that CPUs allocated to child partitions
3201 * do not show up in effective_cpus. If no CPU is left,
3202 * we clear the subparts_cpus & let the child partitions
3203 * fight for the CPUs again.
3205 if (top_cpuset.nr_subparts_cpus) {
3206 if (cpumask_subset(&new_cpus,
3207 top_cpuset.subparts_cpus)) {
3208 top_cpuset.nr_subparts_cpus = 0;
3209 cpumask_clear(top_cpuset.subparts_cpus);
3211 cpumask_andnot(&new_cpus, &new_cpus,
3212 top_cpuset.subparts_cpus);
3215 cpumask_copy(top_cpuset.effective_cpus, &new_cpus);
3216 spin_unlock_irq(&callback_lock);
3217 /* we don't mess with cpumasks of tasks in top_cpuset */
3220 /* synchronize mems_allowed to N_MEMORY */
3222 spin_lock_irq(&callback_lock);
3224 top_cpuset.mems_allowed = new_mems;
3225 top_cpuset.effective_mems = new_mems;
3226 spin_unlock_irq(&callback_lock);
3227 update_tasks_nodemask(&top_cpuset);
3230 percpu_up_write(&cpuset_rwsem);
3232 /* if cpus or mems changed, we need to propagate to descendants */
3233 if (cpus_updated || mems_updated) {
3235 struct cgroup_subsys_state *pos_css;
3238 cpuset_for_each_descendant_pre(cs, pos_css, &top_cpuset) {
3239 if (cs == &top_cpuset || !css_tryget_online(&cs->css))
3243 cpuset_hotplug_update_tasks(cs, ptmp);
3251 /* rebuild sched domains if cpus_allowed has changed */
3252 if (cpus_updated || force_rebuild) {
3253 force_rebuild = false;
3254 rebuild_sched_domains();
3257 free_cpumasks(NULL, ptmp);
3260 void cpuset_update_active_cpus(void)
3263 * We're inside cpu hotplug critical region which usually nests
3264 * inside cgroup synchronization. Bounce actual hotplug processing
3265 * to a work item to avoid reverse locking order.
3267 schedule_work(&cpuset_hotplug_work);
3270 void cpuset_wait_for_hotplug(void)
3272 flush_work(&cpuset_hotplug_work);
3276 * Keep top_cpuset.mems_allowed tracking node_states[N_MEMORY].
3277 * Call this routine anytime after node_states[N_MEMORY] changes.
3278 * See cpuset_update_active_cpus() for CPU hotplug handling.
3280 static int cpuset_track_online_nodes(struct notifier_block *self,
3281 unsigned long action, void *arg)
3283 schedule_work(&cpuset_hotplug_work);
3287 static struct notifier_block cpuset_track_online_nodes_nb = {
3288 .notifier_call = cpuset_track_online_nodes,
3289 .priority = 10, /* ??! */
3293 * cpuset_init_smp - initialize cpus_allowed
3295 * Description: Finish top cpuset after cpu, node maps are initialized
3297 void __init cpuset_init_smp(void)
3300 * cpus_allowd/mems_allowed set to v2 values in the initial
3301 * cpuset_bind() call will be reset to v1 values in another
3302 * cpuset_bind() call when v1 cpuset is mounted.
3304 top_cpuset.old_mems_allowed = top_cpuset.mems_allowed;
3306 cpumask_copy(top_cpuset.effective_cpus, cpu_active_mask);
3307 top_cpuset.effective_mems = node_states[N_MEMORY];
3309 register_hotmemory_notifier(&cpuset_track_online_nodes_nb);
3311 cpuset_migrate_mm_wq = alloc_ordered_workqueue("cpuset_migrate_mm", 0);
3312 BUG_ON(!cpuset_migrate_mm_wq);
3316 * cpuset_cpus_allowed - return cpus_allowed mask from a tasks cpuset.
3317 * @tsk: pointer to task_struct from which to obtain cpuset->cpus_allowed.
3318 * @pmask: pointer to struct cpumask variable to receive cpus_allowed set.
3320 * Description: Returns the cpumask_var_t cpus_allowed of the cpuset
3321 * attached to the specified @tsk. Guaranteed to return some non-empty
3322 * subset of cpu_online_mask, even if this means going outside the
3326 void cpuset_cpus_allowed(struct task_struct *tsk, struct cpumask *pmask)
3328 unsigned long flags;
3330 spin_lock_irqsave(&callback_lock, flags);
3332 guarantee_online_cpus(task_cs(tsk), pmask);
3334 spin_unlock_irqrestore(&callback_lock, flags);
3338 * cpuset_cpus_allowed_fallback - final fallback before complete catastrophe.
3339 * @tsk: pointer to task_struct with which the scheduler is struggling
3341 * Description: In the case that the scheduler cannot find an allowed cpu in
3342 * tsk->cpus_allowed, we fall back to task_cs(tsk)->cpus_allowed. In legacy
3343 * mode however, this value is the same as task_cs(tsk)->effective_cpus,
3344 * which will not contain a sane cpumask during cases such as cpu hotplugging.
3345 * This is the absolute last resort for the scheduler and it is only used if
3346 * _every_ other avenue has been traveled.
3349 void cpuset_cpus_allowed_fallback(struct task_struct *tsk)
3352 do_set_cpus_allowed(tsk, is_in_v2_mode() ?
3353 task_cs(tsk)->cpus_allowed : cpu_possible_mask);
3357 * We own tsk->cpus_allowed, nobody can change it under us.
3359 * But we used cs && cs->cpus_allowed lockless and thus can
3360 * race with cgroup_attach_task() or update_cpumask() and get
3361 * the wrong tsk->cpus_allowed. However, both cases imply the
3362 * subsequent cpuset_change_cpumask()->set_cpus_allowed_ptr()
3363 * which takes task_rq_lock().
3365 * If we are called after it dropped the lock we must see all
3366 * changes in tsk_cs()->cpus_allowed. Otherwise we can temporary
3367 * set any mask even if it is not right from task_cs() pov,
3368 * the pending set_cpus_allowed_ptr() will fix things.
3370 * select_fallback_rq() will fix things ups and set cpu_possible_mask
3375 void __init cpuset_init_current_mems_allowed(void)
3377 nodes_setall(current->mems_allowed);
3381 * cpuset_mems_allowed - return mems_allowed mask from a tasks cpuset.
3382 * @tsk: pointer to task_struct from which to obtain cpuset->mems_allowed.
3384 * Description: Returns the nodemask_t mems_allowed of the cpuset
3385 * attached to the specified @tsk. Guaranteed to return some non-empty
3386 * subset of node_states[N_MEMORY], even if this means going outside the
3390 nodemask_t cpuset_mems_allowed(struct task_struct *tsk)
3393 unsigned long flags;
3395 spin_lock_irqsave(&callback_lock, flags);
3397 guarantee_online_mems(task_cs(tsk), &mask);
3399 spin_unlock_irqrestore(&callback_lock, flags);
3405 * cpuset_nodemask_valid_mems_allowed - check nodemask vs. curremt mems_allowed
3406 * @nodemask: the nodemask to be checked
3408 * Are any of the nodes in the nodemask allowed in current->mems_allowed?
3410 int cpuset_nodemask_valid_mems_allowed(nodemask_t *nodemask)
3412 return nodes_intersects(*nodemask, current->mems_allowed);
3416 * nearest_hardwall_ancestor() - Returns the nearest mem_exclusive or
3417 * mem_hardwall ancestor to the specified cpuset. Call holding
3418 * callback_lock. If no ancestor is mem_exclusive or mem_hardwall
3419 * (an unusual configuration), then returns the root cpuset.
3421 static struct cpuset *nearest_hardwall_ancestor(struct cpuset *cs)
3423 while (!(is_mem_exclusive(cs) || is_mem_hardwall(cs)) && parent_cs(cs))
3429 * cpuset_node_allowed - Can we allocate on a memory node?
3430 * @node: is this an allowed node?
3431 * @gfp_mask: memory allocation flags
3433 * If we're in interrupt, yes, we can always allocate. If @node is set in
3434 * current's mems_allowed, yes. If it's not a __GFP_HARDWALL request and this
3435 * node is set in the nearest hardwalled cpuset ancestor to current's cpuset,
3436 * yes. If current has access to memory reserves as an oom victim, yes.
3439 * GFP_USER allocations are marked with the __GFP_HARDWALL bit,
3440 * and do not allow allocations outside the current tasks cpuset
3441 * unless the task has been OOM killed.
3442 * GFP_KERNEL allocations are not so marked, so can escape to the
3443 * nearest enclosing hardwalled ancestor cpuset.
3445 * Scanning up parent cpusets requires callback_lock. The
3446 * __alloc_pages() routine only calls here with __GFP_HARDWALL bit
3447 * _not_ set if it's a GFP_KERNEL allocation, and all nodes in the
3448 * current tasks mems_allowed came up empty on the first pass over
3449 * the zonelist. So only GFP_KERNEL allocations, if all nodes in the
3450 * cpuset are short of memory, might require taking the callback_lock.
3452 * The first call here from mm/page_alloc:get_page_from_freelist()
3453 * has __GFP_HARDWALL set in gfp_mask, enforcing hardwall cpusets,
3454 * so no allocation on a node outside the cpuset is allowed (unless
3455 * in interrupt, of course).
3457 * The second pass through get_page_from_freelist() doesn't even call
3458 * here for GFP_ATOMIC calls. For those calls, the __alloc_pages()
3459 * variable 'wait' is not set, and the bit ALLOC_CPUSET is not set
3460 * in alloc_flags. That logic and the checks below have the combined
3462 * in_interrupt - any node ok (current task context irrelevant)
3463 * GFP_ATOMIC - any node ok
3464 * tsk_is_oom_victim - any node ok
3465 * GFP_KERNEL - any node in enclosing hardwalled cpuset ok
3466 * GFP_USER - only nodes in current tasks mems allowed ok.
3468 bool __cpuset_node_allowed(int node, gfp_t gfp_mask)
3470 struct cpuset *cs; /* current cpuset ancestors */
3471 int allowed; /* is allocation in zone z allowed? */
3472 unsigned long flags;
3476 if (node_isset(node, current->mems_allowed))
3479 * Allow tasks that have access to memory reserves because they have
3480 * been OOM killed to get memory anywhere.
3482 if (unlikely(tsk_is_oom_victim(current)))
3484 if (gfp_mask & __GFP_HARDWALL) /* If hardwall request, stop here */
3487 if (current->flags & PF_EXITING) /* Let dying task have memory */
3490 /* Not hardwall and node outside mems_allowed: scan up cpusets */
3491 spin_lock_irqsave(&callback_lock, flags);
3494 cs = nearest_hardwall_ancestor(task_cs(current));
3495 allowed = node_isset(node, cs->mems_allowed);
3498 spin_unlock_irqrestore(&callback_lock, flags);
3503 * cpuset_mem_spread_node() - On which node to begin search for a file page
3504 * cpuset_slab_spread_node() - On which node to begin search for a slab page
3506 * If a task is marked PF_SPREAD_PAGE or PF_SPREAD_SLAB (as for
3507 * tasks in a cpuset with is_spread_page or is_spread_slab set),
3508 * and if the memory allocation used cpuset_mem_spread_node()
3509 * to determine on which node to start looking, as it will for
3510 * certain page cache or slab cache pages such as used for file
3511 * system buffers and inode caches, then instead of starting on the
3512 * local node to look for a free page, rather spread the starting
3513 * node around the tasks mems_allowed nodes.
3515 * We don't have to worry about the returned node being offline
3516 * because "it can't happen", and even if it did, it would be ok.
3518 * The routines calling guarantee_online_mems() are careful to
3519 * only set nodes in task->mems_allowed that are online. So it
3520 * should not be possible for the following code to return an
3521 * offline node. But if it did, that would be ok, as this routine
3522 * is not returning the node where the allocation must be, only
3523 * the node where the search should start. The zonelist passed to
3524 * __alloc_pages() will include all nodes. If the slab allocator
3525 * is passed an offline node, it will fall back to the local node.
3526 * See kmem_cache_alloc_node().
3529 static int cpuset_spread_node(int *rotor)
3531 return *rotor = next_node_in(*rotor, current->mems_allowed);
3534 int cpuset_mem_spread_node(void)
3536 if (current->cpuset_mem_spread_rotor == NUMA_NO_NODE)
3537 current->cpuset_mem_spread_rotor =
3538 node_random(¤t->mems_allowed);
3540 return cpuset_spread_node(¤t->cpuset_mem_spread_rotor);
3543 int cpuset_slab_spread_node(void)
3545 if (current->cpuset_slab_spread_rotor == NUMA_NO_NODE)
3546 current->cpuset_slab_spread_rotor =
3547 node_random(¤t->mems_allowed);
3549 return cpuset_spread_node(¤t->cpuset_slab_spread_rotor);
3552 EXPORT_SYMBOL_GPL(cpuset_mem_spread_node);
3555 * cpuset_mems_allowed_intersects - Does @tsk1's mems_allowed intersect @tsk2's?
3556 * @tsk1: pointer to task_struct of some task.
3557 * @tsk2: pointer to task_struct of some other task.
3559 * Description: Return true if @tsk1's mems_allowed intersects the
3560 * mems_allowed of @tsk2. Used by the OOM killer to determine if
3561 * one of the task's memory usage might impact the memory available
3565 int cpuset_mems_allowed_intersects(const struct task_struct *tsk1,
3566 const struct task_struct *tsk2)
3568 return nodes_intersects(tsk1->mems_allowed, tsk2->mems_allowed);
3572 * cpuset_print_current_mems_allowed - prints current's cpuset and mems_allowed
3574 * Description: Prints current's name, cpuset name, and cached copy of its
3575 * mems_allowed to the kernel log.
3577 void cpuset_print_current_mems_allowed(void)
3579 struct cgroup *cgrp;
3583 cgrp = task_cs(current)->css.cgroup;
3584 pr_cont(",cpuset=");
3585 pr_cont_cgroup_name(cgrp);
3586 pr_cont(",mems_allowed=%*pbl",
3587 nodemask_pr_args(¤t->mems_allowed));
3593 * Collection of memory_pressure is suppressed unless
3594 * this flag is enabled by writing "1" to the special
3595 * cpuset file 'memory_pressure_enabled' in the root cpuset.
3598 int cpuset_memory_pressure_enabled __read_mostly;
3601 * cpuset_memory_pressure_bump - keep stats of per-cpuset reclaims.
3603 * Keep a running average of the rate of synchronous (direct)
3604 * page reclaim efforts initiated by tasks in each cpuset.
3606 * This represents the rate at which some task in the cpuset
3607 * ran low on memory on all nodes it was allowed to use, and
3608 * had to enter the kernels page reclaim code in an effort to
3609 * create more free memory by tossing clean pages or swapping
3610 * or writing dirty pages.
3612 * Display to user space in the per-cpuset read-only file
3613 * "memory_pressure". Value displayed is an integer
3614 * representing the recent rate of entry into the synchronous
3615 * (direct) page reclaim by any task attached to the cpuset.
3618 void __cpuset_memory_pressure_bump(void)
3621 fmeter_markevent(&task_cs(current)->fmeter);
3625 #ifdef CONFIG_PROC_PID_CPUSET
3627 * proc_cpuset_show()
3628 * - Print tasks cpuset path into seq_file.
3629 * - Used for /proc/<pid>/cpuset.
3630 * - No need to task_lock(tsk) on this tsk->cpuset reference, as it
3631 * doesn't really matter if tsk->cpuset changes after we read it,
3632 * and we take cpuset_mutex, keeping cpuset_attach() from changing it
3635 int proc_cpuset_show(struct seq_file *m, struct pid_namespace *ns,
3636 struct pid *pid, struct task_struct *tsk)
3639 struct cgroup_subsys_state *css;
3643 buf = kmalloc(PATH_MAX, GFP_KERNEL);
3647 css = task_get_css(tsk, cpuset_cgrp_id);
3648 retval = cgroup_path_ns(css->cgroup, buf, PATH_MAX,
3649 current->nsproxy->cgroup_ns);
3651 if (retval >= PATH_MAX)
3652 retval = -ENAMETOOLONG;
3663 #endif /* CONFIG_PROC_PID_CPUSET */
3665 /* Display task mems_allowed in /proc/<pid>/status file. */
3666 void cpuset_task_status_allowed(struct seq_file *m, struct task_struct *task)
3668 seq_printf(m, "Mems_allowed:\t%*pb\n",
3669 nodemask_pr_args(&task->mems_allowed));
3670 seq_printf(m, "Mems_allowed_list:\t%*pbl\n",
3671 nodemask_pr_args(&task->mems_allowed));