4 * Processor and Memory placement constraints for sets of tasks.
6 * Copyright (C) 2003 BULL SA.
7 * Copyright (C) 2004-2007 Silicon Graphics, Inc.
8 * Copyright (C) 2006 Google, Inc
10 * Portions derived from Patrick Mochel's sysfs code.
11 * sysfs is Copyright (c) 2001-3 Patrick Mochel
13 * 2003-10-10 Written by Simon Derr.
14 * 2003-10-22 Updates by Stephen Hemminger.
15 * 2004 May-July Rework by Paul Jackson.
16 * 2006 Rework by Paul Menage to use generic cgroups
17 * 2008 Rework of the scheduler domains and CPU hotplug handling
20 * This file is subject to the terms and conditions of the GNU General Public
21 * License. See the file COPYING in the main directory of the Linux
22 * distribution for more details.
25 #include <linux/cpu.h>
26 #include <linux/cpumask.h>
27 #include <linux/cpuset.h>
28 #include <linux/err.h>
29 #include <linux/errno.h>
30 #include <linux/file.h>
32 #include <linux/init.h>
33 #include <linux/interrupt.h>
34 #include <linux/kernel.h>
35 #include <linux/kmod.h>
36 #include <linux/list.h>
37 #include <linux/mempolicy.h>
39 #include <linux/memory.h>
40 #include <linux/export.h>
41 #include <linux/mount.h>
42 #include <linux/fs_context.h>
43 #include <linux/namei.h>
44 #include <linux/pagemap.h>
45 #include <linux/proc_fs.h>
46 #include <linux/rcupdate.h>
47 #include <linux/sched.h>
48 #include <linux/sched/mm.h>
49 #include <linux/sched/task.h>
50 #include <linux/seq_file.h>
51 #include <linux/security.h>
52 #include <linux/slab.h>
53 #include <linux/spinlock.h>
54 #include <linux/stat.h>
55 #include <linux/string.h>
56 #include <linux/time.h>
57 #include <linux/time64.h>
58 #include <linux/backing-dev.h>
59 #include <linux/sort.h>
60 #include <linux/oom.h>
61 #include <linux/sched/isolation.h>
62 #include <linux/uaccess.h>
63 #include <linux/atomic.h>
64 #include <linux/mutex.h>
65 #include <linux/cgroup.h>
66 #include <linux/wait.h>
68 DEFINE_STATIC_KEY_FALSE(cpusets_pre_enable_key);
69 DEFINE_STATIC_KEY_FALSE(cpusets_enabled_key);
71 /* See "Frequency meter" comments, below. */
74 int cnt; /* unprocessed events count */
75 int val; /* most recent output value */
76 time64_t time; /* clock (secs) when val computed */
77 spinlock_t lock; /* guards read or write of above */
81 struct cgroup_subsys_state css;
83 unsigned long flags; /* "unsigned long" so bitops work */
86 * On default hierarchy:
88 * The user-configured masks can only be changed by writing to
89 * cpuset.cpus and cpuset.mems, and won't be limited by the
92 * The effective masks is the real masks that apply to the tasks
93 * in the cpuset. They may be changed if the configured masks are
94 * changed or hotplug happens.
96 * effective_mask == configured_mask & parent's effective_mask,
97 * and if it ends up empty, it will inherit the parent's mask.
100 * On legacy hierachy:
102 * The user-configured masks are always the same with effective masks.
105 /* user-configured CPUs and Memory Nodes allow to tasks */
106 cpumask_var_t cpus_allowed;
107 nodemask_t mems_allowed;
109 /* effective CPUs and Memory Nodes allow to tasks */
110 cpumask_var_t effective_cpus;
111 nodemask_t effective_mems;
114 * CPUs allocated to child sub-partitions (default hierarchy only)
115 * - CPUs granted by the parent = effective_cpus U subparts_cpus
116 * - effective_cpus and subparts_cpus are mutually exclusive.
118 * effective_cpus contains only onlined CPUs, but subparts_cpus
119 * may have offlined ones.
121 cpumask_var_t subparts_cpus;
124 * This is old Memory Nodes tasks took on.
126 * - top_cpuset.old_mems_allowed is initialized to mems_allowed.
127 * - A new cpuset's old_mems_allowed is initialized when some
128 * task is moved into it.
129 * - old_mems_allowed is used in cpuset_migrate_mm() when we change
130 * cpuset.mems_allowed and have tasks' nodemask updated, and
131 * then old_mems_allowed is updated to mems_allowed.
133 nodemask_t old_mems_allowed;
135 struct fmeter fmeter; /* memory_pressure filter */
138 * Tasks are being attached to this cpuset. Used to prevent
139 * zeroing cpus/mems_allowed between ->can_attach() and ->attach().
141 int attach_in_progress;
143 /* partition number for rebuild_sched_domains() */
146 /* for custom sched domain */
147 int relax_domain_level;
149 /* number of CPUs in subparts_cpus */
150 int nr_subparts_cpus;
152 /* partition root state */
153 int partition_root_state;
156 * Default hierarchy only:
157 * use_parent_ecpus - set if using parent's effective_cpus
158 * child_ecpus_count - # of children with use_parent_ecpus set
160 int use_parent_ecpus;
161 int child_ecpus_count;
165 * Partition root states:
167 * 0 - not a partition root
171 * -1 - invalid partition root
172 * None of the cpus in cpus_allowed can be put into the parent's
173 * subparts_cpus. In this case, the cpuset is not a real partition
174 * root anymore. However, the CPU_EXCLUSIVE bit will still be set
175 * and the cpuset can be restored back to a partition root if the
176 * parent cpuset can give more CPUs back to this child cpuset.
178 #define PRS_DISABLED 0
179 #define PRS_ENABLED 1
183 * Temporary cpumasks for working with partitions that are passed among
184 * functions to avoid memory allocation in inner functions.
187 cpumask_var_t addmask, delmask; /* For partition root */
188 cpumask_var_t new_cpus; /* For update_cpumasks_hier() */
191 static inline struct cpuset *css_cs(struct cgroup_subsys_state *css)
193 return css ? container_of(css, struct cpuset, css) : NULL;
196 /* Retrieve the cpuset for a task */
197 static inline struct cpuset *task_cs(struct task_struct *task)
199 return css_cs(task_css(task, cpuset_cgrp_id));
202 static inline struct cpuset *parent_cs(struct cpuset *cs)
204 return css_cs(cs->css.parent);
207 /* bits in struct cpuset flags field */
214 CS_SCHED_LOAD_BALANCE,
219 /* convenient tests for these bits */
220 static inline bool is_cpuset_online(struct cpuset *cs)
222 return test_bit(CS_ONLINE, &cs->flags) && !css_is_dying(&cs->css);
225 static inline int is_cpu_exclusive(const struct cpuset *cs)
227 return test_bit(CS_CPU_EXCLUSIVE, &cs->flags);
230 static inline int is_mem_exclusive(const struct cpuset *cs)
232 return test_bit(CS_MEM_EXCLUSIVE, &cs->flags);
235 static inline int is_mem_hardwall(const struct cpuset *cs)
237 return test_bit(CS_MEM_HARDWALL, &cs->flags);
240 static inline int is_sched_load_balance(const struct cpuset *cs)
242 return test_bit(CS_SCHED_LOAD_BALANCE, &cs->flags);
245 static inline int is_memory_migrate(const struct cpuset *cs)
247 return test_bit(CS_MEMORY_MIGRATE, &cs->flags);
250 static inline int is_spread_page(const struct cpuset *cs)
252 return test_bit(CS_SPREAD_PAGE, &cs->flags);
255 static inline int is_spread_slab(const struct cpuset *cs)
257 return test_bit(CS_SPREAD_SLAB, &cs->flags);
260 static inline int is_partition_root(const struct cpuset *cs)
262 return cs->partition_root_state > 0;
265 static struct cpuset top_cpuset = {
266 .flags = ((1 << CS_ONLINE) | (1 << CS_CPU_EXCLUSIVE) |
267 (1 << CS_MEM_EXCLUSIVE)),
268 .partition_root_state = PRS_ENABLED,
272 * cpuset_for_each_child - traverse online children of a cpuset
273 * @child_cs: loop cursor pointing to the current child
274 * @pos_css: used for iteration
275 * @parent_cs: target cpuset to walk children of
277 * Walk @child_cs through the online children of @parent_cs. Must be used
278 * with RCU read locked.
280 #define cpuset_for_each_child(child_cs, pos_css, parent_cs) \
281 css_for_each_child((pos_css), &(parent_cs)->css) \
282 if (is_cpuset_online(((child_cs) = css_cs((pos_css)))))
285 * cpuset_for_each_descendant_pre - pre-order walk of a cpuset's descendants
286 * @des_cs: loop cursor pointing to the current descendant
287 * @pos_css: used for iteration
288 * @root_cs: target cpuset to walk ancestor of
290 * Walk @des_cs through the online descendants of @root_cs. Must be used
291 * with RCU read locked. The caller may modify @pos_css by calling
292 * css_rightmost_descendant() to skip subtree. @root_cs is included in the
293 * iteration and the first node to be visited.
295 #define cpuset_for_each_descendant_pre(des_cs, pos_css, root_cs) \
296 css_for_each_descendant_pre((pos_css), &(root_cs)->css) \
297 if (is_cpuset_online(((des_cs) = css_cs((pos_css)))))
300 * There are two global locks guarding cpuset structures - cpuset_mutex and
301 * callback_lock. We also require taking task_lock() when dereferencing a
302 * task's cpuset pointer. See "The task_lock() exception", at the end of this
305 * A task must hold both locks to modify cpusets. If a task holds
306 * cpuset_mutex, then it blocks others wanting that mutex, ensuring that it
307 * is the only task able to also acquire callback_lock and be able to
308 * modify cpusets. It can perform various checks on the cpuset structure
309 * first, knowing nothing will change. It can also allocate memory while
310 * just holding cpuset_mutex. While it is performing these checks, various
311 * callback routines can briefly acquire callback_lock to query cpusets.
312 * Once it is ready to make the changes, it takes callback_lock, blocking
315 * Calls to the kernel memory allocator can not be made while holding
316 * callback_lock, as that would risk double tripping on callback_lock
317 * from one of the callbacks into the cpuset code from within
320 * If a task is only holding callback_lock, then it has read-only
323 * Now, the task_struct fields mems_allowed and mempolicy may be changed
324 * by other task, we use alloc_lock in the task_struct fields to protect
327 * The cpuset_common_file_read() handlers only hold callback_lock across
328 * small pieces of code, such as when reading out possibly multi-word
329 * cpumasks and nodemasks.
331 * Accessing a task's cpuset should be done in accordance with the
332 * guidelines for accessing subsystem state in kernel/cgroup.c
335 static DEFINE_MUTEX(cpuset_mutex);
336 static DEFINE_SPINLOCK(callback_lock);
338 static struct workqueue_struct *cpuset_migrate_mm_wq;
341 * CPU / memory hotplug is handled asynchronously.
343 static void cpuset_hotplug_workfn(struct work_struct *work);
344 static DECLARE_WORK(cpuset_hotplug_work, cpuset_hotplug_workfn);
346 static DECLARE_WAIT_QUEUE_HEAD(cpuset_attach_wq);
349 * Cgroup v2 behavior is used when on default hierarchy or the
350 * cgroup_v2_mode flag is set.
352 static inline bool is_in_v2_mode(void)
354 return cgroup_subsys_on_dfl(cpuset_cgrp_subsys) ||
355 (cpuset_cgrp_subsys.root->flags & CGRP_ROOT_CPUSET_V2_MODE);
359 * Return in pmask the portion of a cpusets's cpus_allowed that
360 * are online. If none are online, walk up the cpuset hierarchy
361 * until we find one that does have some online cpus.
363 * One way or another, we guarantee to return some non-empty subset
364 * of cpu_online_mask.
366 * Call with callback_lock or cpuset_mutex held.
368 static void guarantee_online_cpus(struct cpuset *cs, struct cpumask *pmask)
370 while (!cpumask_intersects(cs->effective_cpus, cpu_online_mask)) {
374 * The top cpuset doesn't have any online cpu as a
375 * consequence of a race between cpuset_hotplug_work
376 * and cpu hotplug notifier. But we know the top
377 * cpuset's effective_cpus is on its way to to be
378 * identical to cpu_online_mask.
380 cpumask_copy(pmask, cpu_online_mask);
384 cpumask_and(pmask, cs->effective_cpus, cpu_online_mask);
388 * Return in *pmask the portion of a cpusets's mems_allowed that
389 * are online, with memory. If none are online with memory, walk
390 * up the cpuset hierarchy until we find one that does have some
391 * online mems. The top cpuset always has some mems online.
393 * One way or another, we guarantee to return some non-empty subset
394 * of node_states[N_MEMORY].
396 * Call with callback_lock or cpuset_mutex held.
398 static void guarantee_online_mems(struct cpuset *cs, nodemask_t *pmask)
400 while (!nodes_intersects(cs->effective_mems, node_states[N_MEMORY]))
402 nodes_and(*pmask, cs->effective_mems, node_states[N_MEMORY]);
406 * update task's spread flag if cpuset's page/slab spread flag is set
408 * Call with callback_lock or cpuset_mutex held.
410 static void cpuset_update_task_spread_flag(struct cpuset *cs,
411 struct task_struct *tsk)
413 if (is_spread_page(cs))
414 task_set_spread_page(tsk);
416 task_clear_spread_page(tsk);
418 if (is_spread_slab(cs))
419 task_set_spread_slab(tsk);
421 task_clear_spread_slab(tsk);
425 * is_cpuset_subset(p, q) - Is cpuset p a subset of cpuset q?
427 * One cpuset is a subset of another if all its allowed CPUs and
428 * Memory Nodes are a subset of the other, and its exclusive flags
429 * are only set if the other's are set. Call holding cpuset_mutex.
432 static int is_cpuset_subset(const struct cpuset *p, const struct cpuset *q)
434 return cpumask_subset(p->cpus_allowed, q->cpus_allowed) &&
435 nodes_subset(p->mems_allowed, q->mems_allowed) &&
436 is_cpu_exclusive(p) <= is_cpu_exclusive(q) &&
437 is_mem_exclusive(p) <= is_mem_exclusive(q);
441 * alloc_cpumasks - allocate three cpumasks for cpuset
442 * @cs: the cpuset that have cpumasks to be allocated.
443 * @tmp: the tmpmasks structure pointer
444 * Return: 0 if successful, -ENOMEM otherwise.
446 * Only one of the two input arguments should be non-NULL.
448 static inline int alloc_cpumasks(struct cpuset *cs, struct tmpmasks *tmp)
450 cpumask_var_t *pmask1, *pmask2, *pmask3;
453 pmask1 = &cs->cpus_allowed;
454 pmask2 = &cs->effective_cpus;
455 pmask3 = &cs->subparts_cpus;
457 pmask1 = &tmp->new_cpus;
458 pmask2 = &tmp->addmask;
459 pmask3 = &tmp->delmask;
462 if (!zalloc_cpumask_var(pmask1, GFP_KERNEL))
465 if (!zalloc_cpumask_var(pmask2, GFP_KERNEL))
468 if (!zalloc_cpumask_var(pmask3, GFP_KERNEL))
474 free_cpumask_var(*pmask2);
476 free_cpumask_var(*pmask1);
481 * free_cpumasks - free cpumasks in a tmpmasks structure
482 * @cs: the cpuset that have cpumasks to be free.
483 * @tmp: the tmpmasks structure pointer
485 static inline void free_cpumasks(struct cpuset *cs, struct tmpmasks *tmp)
488 free_cpumask_var(cs->cpus_allowed);
489 free_cpumask_var(cs->effective_cpus);
490 free_cpumask_var(cs->subparts_cpus);
493 free_cpumask_var(tmp->new_cpus);
494 free_cpumask_var(tmp->addmask);
495 free_cpumask_var(tmp->delmask);
500 * alloc_trial_cpuset - allocate a trial cpuset
501 * @cs: the cpuset that the trial cpuset duplicates
503 static struct cpuset *alloc_trial_cpuset(struct cpuset *cs)
505 struct cpuset *trial;
507 trial = kmemdup(cs, sizeof(*cs), GFP_KERNEL);
511 if (alloc_cpumasks(trial, NULL)) {
516 cpumask_copy(trial->cpus_allowed, cs->cpus_allowed);
517 cpumask_copy(trial->effective_cpus, cs->effective_cpus);
522 * free_cpuset - free the cpuset
523 * @cs: the cpuset to be freed
525 static inline void free_cpuset(struct cpuset *cs)
527 free_cpumasks(cs, NULL);
532 * validate_change() - Used to validate that any proposed cpuset change
533 * follows the structural rules for cpusets.
535 * If we replaced the flag and mask values of the current cpuset
536 * (cur) with those values in the trial cpuset (trial), would
537 * our various subset and exclusive rules still be valid? Presumes
540 * 'cur' is the address of an actual, in-use cpuset. Operations
541 * such as list traversal that depend on the actual address of the
542 * cpuset in the list must use cur below, not trial.
544 * 'trial' is the address of bulk structure copy of cur, with
545 * perhaps one or more of the fields cpus_allowed, mems_allowed,
546 * or flags changed to new, trial values.
548 * Return 0 if valid, -errno if not.
551 static int validate_change(struct cpuset *cur, struct cpuset *trial)
553 struct cgroup_subsys_state *css;
554 struct cpuset *c, *par;
559 /* Each of our child cpusets must be a subset of us */
561 cpuset_for_each_child(c, css, cur)
562 if (!is_cpuset_subset(c, trial))
565 /* Remaining checks don't apply to root cpuset */
567 if (cur == &top_cpuset)
570 par = parent_cs(cur);
572 /* On legacy hiearchy, we must be a subset of our parent cpuset. */
574 if (!is_in_v2_mode() && !is_cpuset_subset(trial, par))
578 * If either I or some sibling (!= me) is exclusive, we can't
582 cpuset_for_each_child(c, css, par) {
583 if ((is_cpu_exclusive(trial) || is_cpu_exclusive(c)) &&
585 cpumask_intersects(trial->cpus_allowed, c->cpus_allowed))
587 if ((is_mem_exclusive(trial) || is_mem_exclusive(c)) &&
589 nodes_intersects(trial->mems_allowed, c->mems_allowed))
594 * Cpusets with tasks - existing or newly being attached - can't
595 * be changed to have empty cpus_allowed or mems_allowed.
598 if ((cgroup_is_populated(cur->css.cgroup) || cur->attach_in_progress)) {
599 if (!cpumask_empty(cur->cpus_allowed) &&
600 cpumask_empty(trial->cpus_allowed))
602 if (!nodes_empty(cur->mems_allowed) &&
603 nodes_empty(trial->mems_allowed))
608 * We can't shrink if we won't have enough room for SCHED_DEADLINE
612 if (is_cpu_exclusive(cur) &&
613 !cpuset_cpumask_can_shrink(cur->cpus_allowed,
614 trial->cpus_allowed))
625 * Helper routine for generate_sched_domains().
626 * Do cpusets a, b have overlapping effective cpus_allowed masks?
628 static int cpusets_overlap(struct cpuset *a, struct cpuset *b)
630 return cpumask_intersects(a->effective_cpus, b->effective_cpus);
634 update_domain_attr(struct sched_domain_attr *dattr, struct cpuset *c)
636 if (dattr->relax_domain_level < c->relax_domain_level)
637 dattr->relax_domain_level = c->relax_domain_level;
641 static void update_domain_attr_tree(struct sched_domain_attr *dattr,
642 struct cpuset *root_cs)
645 struct cgroup_subsys_state *pos_css;
648 cpuset_for_each_descendant_pre(cp, pos_css, root_cs) {
649 /* skip the whole subtree if @cp doesn't have any CPU */
650 if (cpumask_empty(cp->cpus_allowed)) {
651 pos_css = css_rightmost_descendant(pos_css);
655 if (is_sched_load_balance(cp))
656 update_domain_attr(dattr, cp);
661 /* Must be called with cpuset_mutex held. */
662 static inline int nr_cpusets(void)
664 /* jump label reference count + the top-level cpuset */
665 return static_key_count(&cpusets_enabled_key.key) + 1;
669 * generate_sched_domains()
671 * This function builds a partial partition of the systems CPUs
672 * A 'partial partition' is a set of non-overlapping subsets whose
673 * union is a subset of that set.
674 * The output of this function needs to be passed to kernel/sched/core.c
675 * partition_sched_domains() routine, which will rebuild the scheduler's
676 * load balancing domains (sched domains) as specified by that partial
679 * See "What is sched_load_balance" in Documentation/admin-guide/cgroup-v1/cpusets.rst
680 * for a background explanation of this.
682 * Does not return errors, on the theory that the callers of this
683 * routine would rather not worry about failures to rebuild sched
684 * domains when operating in the severe memory shortage situations
685 * that could cause allocation failures below.
687 * Must be called with cpuset_mutex held.
689 * The three key local variables below are:
690 * cp - cpuset pointer, used (together with pos_css) to perform a
691 * top-down scan of all cpusets. For our purposes, rebuilding
692 * the schedulers sched domains, we can ignore !is_sched_load_
694 * csa - (for CpuSet Array) Array of pointers to all the cpusets
695 * that need to be load balanced, for convenient iterative
696 * access by the subsequent code that finds the best partition,
697 * i.e the set of domains (subsets) of CPUs such that the
698 * cpus_allowed of every cpuset marked is_sched_load_balance
699 * is a subset of one of these domains, while there are as
700 * many such domains as possible, each as small as possible.
701 * doms - Conversion of 'csa' to an array of cpumasks, for passing to
702 * the kernel/sched/core.c routine partition_sched_domains() in a
703 * convenient format, that can be easily compared to the prior
704 * value to determine what partition elements (sched domains)
705 * were changed (added or removed.)
707 * Finding the best partition (set of domains):
708 * The triple nested loops below over i, j, k scan over the
709 * load balanced cpusets (using the array of cpuset pointers in
710 * csa[]) looking for pairs of cpusets that have overlapping
711 * cpus_allowed, but which don't have the same 'pn' partition
712 * number and gives them in the same partition number. It keeps
713 * looping on the 'restart' label until it can no longer find
716 * The union of the cpus_allowed masks from the set of
717 * all cpusets having the same 'pn' value then form the one
718 * element of the partition (one sched domain) to be passed to
719 * partition_sched_domains().
721 static int generate_sched_domains(cpumask_var_t **domains,
722 struct sched_domain_attr **attributes)
724 struct cpuset *cp; /* top-down scan of cpusets */
725 struct cpuset **csa; /* array of all cpuset ptrs */
726 int csn; /* how many cpuset ptrs in csa so far */
727 int i, j, k; /* indices for partition finding loops */
728 cpumask_var_t *doms; /* resulting partition; i.e. sched domains */
729 struct sched_domain_attr *dattr; /* attributes for custom domains */
730 int ndoms = 0; /* number of sched domains in result */
731 int nslot; /* next empty doms[] struct cpumask slot */
732 struct cgroup_subsys_state *pos_css;
733 bool root_load_balance = is_sched_load_balance(&top_cpuset);
739 /* Special case for the 99% of systems with one, full, sched domain */
740 if (root_load_balance && !top_cpuset.nr_subparts_cpus) {
742 doms = alloc_sched_domains(ndoms);
746 dattr = kmalloc(sizeof(struct sched_domain_attr), GFP_KERNEL);
748 *dattr = SD_ATTR_INIT;
749 update_domain_attr_tree(dattr, &top_cpuset);
751 cpumask_and(doms[0], top_cpuset.effective_cpus,
752 housekeeping_cpumask(HK_FLAG_DOMAIN));
757 csa = kmalloc_array(nr_cpusets(), sizeof(cp), GFP_KERNEL);
763 if (root_load_balance)
764 csa[csn++] = &top_cpuset;
765 cpuset_for_each_descendant_pre(cp, pos_css, &top_cpuset) {
766 if (cp == &top_cpuset)
769 * Continue traversing beyond @cp iff @cp has some CPUs and
770 * isn't load balancing. The former is obvious. The
771 * latter: All child cpusets contain a subset of the
772 * parent's cpus, so just skip them, and then we call
773 * update_domain_attr_tree() to calc relax_domain_level of
774 * the corresponding sched domain.
776 * If root is load-balancing, we can skip @cp if it
777 * is a subset of the root's effective_cpus.
779 if (!cpumask_empty(cp->cpus_allowed) &&
780 !(is_sched_load_balance(cp) &&
781 cpumask_intersects(cp->cpus_allowed,
782 housekeeping_cpumask(HK_FLAG_DOMAIN))))
785 if (root_load_balance &&
786 cpumask_subset(cp->cpus_allowed, top_cpuset.effective_cpus))
789 if (is_sched_load_balance(cp) &&
790 !cpumask_empty(cp->effective_cpus))
793 /* skip @cp's subtree if not a partition root */
794 if (!is_partition_root(cp))
795 pos_css = css_rightmost_descendant(pos_css);
799 for (i = 0; i < csn; i++)
804 /* Find the best partition (set of sched domains) */
805 for (i = 0; i < csn; i++) {
806 struct cpuset *a = csa[i];
809 for (j = 0; j < csn; j++) {
810 struct cpuset *b = csa[j];
813 if (apn != bpn && cpusets_overlap(a, b)) {
814 for (k = 0; k < csn; k++) {
815 struct cpuset *c = csa[k];
820 ndoms--; /* one less element */
827 * Now we know how many domains to create.
828 * Convert <csn, csa> to <ndoms, doms> and populate cpu masks.
830 doms = alloc_sched_domains(ndoms);
835 * The rest of the code, including the scheduler, can deal with
836 * dattr==NULL case. No need to abort if alloc fails.
838 dattr = kmalloc_array(ndoms, sizeof(struct sched_domain_attr),
841 for (nslot = 0, i = 0; i < csn; i++) {
842 struct cpuset *a = csa[i];
847 /* Skip completed partitions */
853 if (nslot == ndoms) {
854 static int warnings = 10;
856 pr_warn("rebuild_sched_domains confused: nslot %d, ndoms %d, csn %d, i %d, apn %d\n",
857 nslot, ndoms, csn, i, apn);
865 *(dattr + nslot) = SD_ATTR_INIT;
866 for (j = i; j < csn; j++) {
867 struct cpuset *b = csa[j];
870 cpumask_or(dp, dp, b->effective_cpus);
871 cpumask_and(dp, dp, housekeeping_cpumask(HK_FLAG_DOMAIN));
873 update_domain_attr_tree(dattr + nslot, b);
875 /* Done with this partition */
881 BUG_ON(nslot != ndoms);
887 * Fallback to the default domain if kmalloc() failed.
888 * See comments in partition_sched_domains().
899 * Rebuild scheduler domains.
901 * If the flag 'sched_load_balance' of any cpuset with non-empty
902 * 'cpus' changes, or if the 'cpus' allowed changes in any cpuset
903 * which has that flag enabled, or if any cpuset with a non-empty
904 * 'cpus' is removed, then call this routine to rebuild the
905 * scheduler's dynamic sched domains.
907 * Call with cpuset_mutex held. Takes get_online_cpus().
909 static void rebuild_sched_domains_locked(void)
911 struct sched_domain_attr *attr;
915 lockdep_assert_held(&cpuset_mutex);
919 * We have raced with CPU hotplug. Don't do anything to avoid
920 * passing doms with offlined cpu to partition_sched_domains().
921 * Anyways, hotplug work item will rebuild sched domains.
923 if (!top_cpuset.nr_subparts_cpus &&
924 !cpumask_equal(top_cpuset.effective_cpus, cpu_active_mask))
927 if (top_cpuset.nr_subparts_cpus &&
928 !cpumask_subset(top_cpuset.effective_cpus, cpu_active_mask))
931 /* Generate domain masks and attrs */
932 ndoms = generate_sched_domains(&doms, &attr);
934 /* Have scheduler rebuild the domains */
935 partition_sched_domains(ndoms, doms, attr);
939 #else /* !CONFIG_SMP */
940 static void rebuild_sched_domains_locked(void)
943 #endif /* CONFIG_SMP */
945 void rebuild_sched_domains(void)
947 mutex_lock(&cpuset_mutex);
948 rebuild_sched_domains_locked();
949 mutex_unlock(&cpuset_mutex);
953 * update_tasks_cpumask - Update the cpumasks of tasks in the cpuset.
954 * @cs: the cpuset in which each task's cpus_allowed mask needs to be changed
956 * Iterate through each task of @cs updating its cpus_allowed to the
957 * effective cpuset's. As this function is called with cpuset_mutex held,
958 * cpuset membership stays stable.
960 static void update_tasks_cpumask(struct cpuset *cs)
962 struct css_task_iter it;
963 struct task_struct *task;
965 css_task_iter_start(&cs->css, 0, &it);
966 while ((task = css_task_iter_next(&it)))
967 set_cpus_allowed_ptr(task, cs->effective_cpus);
968 css_task_iter_end(&it);
972 * compute_effective_cpumask - Compute the effective cpumask of the cpuset
973 * @new_cpus: the temp variable for the new effective_cpus mask
974 * @cs: the cpuset the need to recompute the new effective_cpus mask
975 * @parent: the parent cpuset
977 * If the parent has subpartition CPUs, include them in the list of
978 * allowable CPUs in computing the new effective_cpus mask. Since offlined
979 * CPUs are not removed from subparts_cpus, we have to use cpu_active_mask
982 static void compute_effective_cpumask(struct cpumask *new_cpus,
983 struct cpuset *cs, struct cpuset *parent)
985 if (parent->nr_subparts_cpus) {
986 cpumask_or(new_cpus, parent->effective_cpus,
987 parent->subparts_cpus);
988 cpumask_and(new_cpus, new_cpus, cs->cpus_allowed);
989 cpumask_and(new_cpus, new_cpus, cpu_active_mask);
991 cpumask_and(new_cpus, cs->cpus_allowed, parent->effective_cpus);
996 * Commands for update_parent_subparts_cpumask
999 partcmd_enable, /* Enable partition root */
1000 partcmd_disable, /* Disable partition root */
1001 partcmd_update, /* Update parent's subparts_cpus */
1005 * update_parent_subparts_cpumask - update subparts_cpus mask of parent cpuset
1006 * @cpuset: The cpuset that requests change in partition root state
1007 * @cmd: Partition root state change command
1008 * @newmask: Optional new cpumask for partcmd_update
1009 * @tmp: Temporary addmask and delmask
1010 * Return: 0, 1 or an error code
1012 * For partcmd_enable, the cpuset is being transformed from a non-partition
1013 * root to a partition root. The cpus_allowed mask of the given cpuset will
1014 * be put into parent's subparts_cpus and taken away from parent's
1015 * effective_cpus. The function will return 0 if all the CPUs listed in
1016 * cpus_allowed can be granted or an error code will be returned.
1018 * For partcmd_disable, the cpuset is being transofrmed from a partition
1019 * root back to a non-partition root. any CPUs in cpus_allowed that are in
1020 * parent's subparts_cpus will be taken away from that cpumask and put back
1021 * into parent's effective_cpus. 0 should always be returned.
1023 * For partcmd_update, if the optional newmask is specified, the cpu
1024 * list is to be changed from cpus_allowed to newmask. Otherwise,
1025 * cpus_allowed is assumed to remain the same. The cpuset should either
1026 * be a partition root or an invalid partition root. The partition root
1027 * state may change if newmask is NULL and none of the requested CPUs can
1028 * be granted by the parent. The function will return 1 if changes to
1029 * parent's subparts_cpus and effective_cpus happen or 0 otherwise.
1030 * Error code should only be returned when newmask is non-NULL.
1032 * The partcmd_enable and partcmd_disable commands are used by
1033 * update_prstate(). The partcmd_update command is used by
1034 * update_cpumasks_hier() with newmask NULL and update_cpumask() with
1037 * The checking is more strict when enabling partition root than the
1038 * other two commands.
1040 * Because of the implicit cpu exclusive nature of a partition root,
1041 * cpumask changes that violates the cpu exclusivity rule will not be
1042 * permitted when checked by validate_change(). The validate_change()
1043 * function will also prevent any changes to the cpu list if it is not
1044 * a superset of children's cpu lists.
1046 static int update_parent_subparts_cpumask(struct cpuset *cpuset, int cmd,
1047 struct cpumask *newmask,
1048 struct tmpmasks *tmp)
1050 struct cpuset *parent = parent_cs(cpuset);
1051 int adding; /* Moving cpus from effective_cpus to subparts_cpus */
1052 int deleting; /* Moving cpus from subparts_cpus to effective_cpus */
1053 bool part_error = false; /* Partition error? */
1055 lockdep_assert_held(&cpuset_mutex);
1058 * The parent must be a partition root.
1059 * The new cpumask, if present, or the current cpus_allowed must
1062 if (!is_partition_root(parent) ||
1063 (newmask && cpumask_empty(newmask)) ||
1064 (!newmask && cpumask_empty(cpuset->cpus_allowed)))
1068 * Enabling/disabling partition root is not allowed if there are
1071 if ((cmd != partcmd_update) && css_has_online_children(&cpuset->css))
1075 * Enabling partition root is not allowed if not all the CPUs
1076 * can be granted from parent's effective_cpus or at least one
1077 * CPU will be left after that.
1079 if ((cmd == partcmd_enable) &&
1080 (!cpumask_subset(cpuset->cpus_allowed, parent->effective_cpus) ||
1081 cpumask_equal(cpuset->cpus_allowed, parent->effective_cpus)))
1085 * A cpumask update cannot make parent's effective_cpus become empty.
1087 adding = deleting = false;
1088 if (cmd == partcmd_enable) {
1089 cpumask_copy(tmp->addmask, cpuset->cpus_allowed);
1091 } else if (cmd == partcmd_disable) {
1092 deleting = cpumask_and(tmp->delmask, cpuset->cpus_allowed,
1093 parent->subparts_cpus);
1094 } else if (newmask) {
1096 * partcmd_update with newmask:
1098 * delmask = cpus_allowed & ~newmask & parent->subparts_cpus
1099 * addmask = newmask & parent->effective_cpus
1100 * & ~parent->subparts_cpus
1102 cpumask_andnot(tmp->delmask, cpuset->cpus_allowed, newmask);
1103 deleting = cpumask_and(tmp->delmask, tmp->delmask,
1104 parent->subparts_cpus);
1106 cpumask_and(tmp->addmask, newmask, parent->effective_cpus);
1107 adding = cpumask_andnot(tmp->addmask, tmp->addmask,
1108 parent->subparts_cpus);
1110 * Return error if the new effective_cpus could become empty.
1113 cpumask_equal(parent->effective_cpus, tmp->addmask)) {
1117 * As some of the CPUs in subparts_cpus might have
1118 * been offlined, we need to compute the real delmask
1121 if (!cpumask_and(tmp->addmask, tmp->delmask,
1124 cpumask_copy(tmp->addmask, parent->effective_cpus);
1128 * partcmd_update w/o newmask:
1130 * addmask = cpus_allowed & parent->effectiveb_cpus
1132 * Note that parent's subparts_cpus may have been
1133 * pre-shrunk in case there is a change in the cpu list.
1134 * So no deletion is needed.
1136 adding = cpumask_and(tmp->addmask, cpuset->cpus_allowed,
1137 parent->effective_cpus);
1138 part_error = cpumask_equal(tmp->addmask,
1139 parent->effective_cpus);
1142 if (cmd == partcmd_update) {
1143 int prev_prs = cpuset->partition_root_state;
1146 * Check for possible transition between PRS_ENABLED
1149 switch (cpuset->partition_root_state) {
1152 cpuset->partition_root_state = PRS_ERROR;
1156 cpuset->partition_root_state = PRS_ENABLED;
1160 * Set part_error if previously in invalid state.
1162 part_error = (prev_prs == PRS_ERROR);
1165 if (!part_error && (cpuset->partition_root_state == PRS_ERROR))
1166 return 0; /* Nothing need to be done */
1168 if (cpuset->partition_root_state == PRS_ERROR) {
1170 * Remove all its cpus from parent's subparts_cpus.
1173 deleting = cpumask_and(tmp->delmask, cpuset->cpus_allowed,
1174 parent->subparts_cpus);
1177 if (!adding && !deleting)
1181 * Change the parent's subparts_cpus.
1182 * Newly added CPUs will be removed from effective_cpus and
1183 * newly deleted ones will be added back to effective_cpus.
1185 spin_lock_irq(&callback_lock);
1187 cpumask_or(parent->subparts_cpus,
1188 parent->subparts_cpus, tmp->addmask);
1189 cpumask_andnot(parent->effective_cpus,
1190 parent->effective_cpus, tmp->addmask);
1193 cpumask_andnot(parent->subparts_cpus,
1194 parent->subparts_cpus, tmp->delmask);
1196 * Some of the CPUs in subparts_cpus might have been offlined.
1198 cpumask_and(tmp->delmask, tmp->delmask, cpu_active_mask);
1199 cpumask_or(parent->effective_cpus,
1200 parent->effective_cpus, tmp->delmask);
1203 parent->nr_subparts_cpus = cpumask_weight(parent->subparts_cpus);
1204 spin_unlock_irq(&callback_lock);
1206 return cmd == partcmd_update;
1210 * update_cpumasks_hier - Update effective cpumasks and tasks in the subtree
1211 * @cs: the cpuset to consider
1212 * @tmp: temp variables for calculating effective_cpus & partition setup
1214 * When congifured cpumask is changed, the effective cpumasks of this cpuset
1215 * and all its descendants need to be updated.
1217 * On legacy hierachy, effective_cpus will be the same with cpu_allowed.
1219 * Called with cpuset_mutex held
1221 static void update_cpumasks_hier(struct cpuset *cs, struct tmpmasks *tmp)
1224 struct cgroup_subsys_state *pos_css;
1225 bool need_rebuild_sched_domains = false;
1228 cpuset_for_each_descendant_pre(cp, pos_css, cs) {
1229 struct cpuset *parent = parent_cs(cp);
1231 compute_effective_cpumask(tmp->new_cpus, cp, parent);
1234 * If it becomes empty, inherit the effective mask of the
1235 * parent, which is guaranteed to have some CPUs.
1237 if (is_in_v2_mode() && cpumask_empty(tmp->new_cpus)) {
1238 cpumask_copy(tmp->new_cpus, parent->effective_cpus);
1239 if (!cp->use_parent_ecpus) {
1240 cp->use_parent_ecpus = true;
1241 parent->child_ecpus_count++;
1243 } else if (cp->use_parent_ecpus) {
1244 cp->use_parent_ecpus = false;
1245 WARN_ON_ONCE(!parent->child_ecpus_count);
1246 parent->child_ecpus_count--;
1250 * Skip the whole subtree if the cpumask remains the same
1251 * and has no partition root state.
1253 if (!cp->partition_root_state &&
1254 cpumask_equal(tmp->new_cpus, cp->effective_cpus)) {
1255 pos_css = css_rightmost_descendant(pos_css);
1260 * update_parent_subparts_cpumask() should have been called
1261 * for cs already in update_cpumask(). We should also call
1262 * update_tasks_cpumask() again for tasks in the parent
1263 * cpuset if the parent's subparts_cpus changes.
1265 if ((cp != cs) && cp->partition_root_state) {
1266 switch (parent->partition_root_state) {
1269 * If parent is not a partition root or an
1270 * invalid partition root, clear the state
1271 * state and the CS_CPU_EXCLUSIVE flag.
1273 WARN_ON_ONCE(cp->partition_root_state
1275 cp->partition_root_state = 0;
1278 * clear_bit() is an atomic operation and
1279 * readers aren't interested in the state
1280 * of CS_CPU_EXCLUSIVE anyway. So we can
1281 * just update the flag without holding
1282 * the callback_lock.
1284 clear_bit(CS_CPU_EXCLUSIVE, &cp->flags);
1288 if (update_parent_subparts_cpumask(cp, partcmd_update, NULL, tmp))
1289 update_tasks_cpumask(parent);
1294 * When parent is invalid, it has to be too.
1296 cp->partition_root_state = PRS_ERROR;
1297 if (cp->nr_subparts_cpus) {
1298 cp->nr_subparts_cpus = 0;
1299 cpumask_clear(cp->subparts_cpus);
1305 if (!css_tryget_online(&cp->css))
1309 spin_lock_irq(&callback_lock);
1311 cpumask_copy(cp->effective_cpus, tmp->new_cpus);
1312 if (cp->nr_subparts_cpus &&
1313 (cp->partition_root_state != PRS_ENABLED)) {
1314 cp->nr_subparts_cpus = 0;
1315 cpumask_clear(cp->subparts_cpus);
1316 } else if (cp->nr_subparts_cpus) {
1318 * Make sure that effective_cpus & subparts_cpus
1319 * are mutually exclusive.
1321 * In the unlikely event that effective_cpus
1322 * becomes empty. we clear cp->nr_subparts_cpus and
1323 * let its child partition roots to compete for
1326 cpumask_andnot(cp->effective_cpus, cp->effective_cpus,
1328 if (cpumask_empty(cp->effective_cpus)) {
1329 cpumask_copy(cp->effective_cpus, tmp->new_cpus);
1330 cpumask_clear(cp->subparts_cpus);
1331 cp->nr_subparts_cpus = 0;
1332 } else if (!cpumask_subset(cp->subparts_cpus,
1334 cpumask_andnot(cp->subparts_cpus,
1335 cp->subparts_cpus, tmp->new_cpus);
1336 cp->nr_subparts_cpus
1337 = cpumask_weight(cp->subparts_cpus);
1340 spin_unlock_irq(&callback_lock);
1342 WARN_ON(!is_in_v2_mode() &&
1343 !cpumask_equal(cp->cpus_allowed, cp->effective_cpus));
1345 update_tasks_cpumask(cp);
1348 * On legacy hierarchy, if the effective cpumask of any non-
1349 * empty cpuset is changed, we need to rebuild sched domains.
1350 * On default hierarchy, the cpuset needs to be a partition
1353 if (!cpumask_empty(cp->cpus_allowed) &&
1354 is_sched_load_balance(cp) &&
1355 (!cgroup_subsys_on_dfl(cpuset_cgrp_subsys) ||
1356 is_partition_root(cp)))
1357 need_rebuild_sched_domains = true;
1364 if (need_rebuild_sched_domains)
1365 rebuild_sched_domains_locked();
1369 * update_sibling_cpumasks - Update siblings cpumasks
1370 * @parent: Parent cpuset
1371 * @cs: Current cpuset
1372 * @tmp: Temp variables
1374 static void update_sibling_cpumasks(struct cpuset *parent, struct cpuset *cs,
1375 struct tmpmasks *tmp)
1377 struct cpuset *sibling;
1378 struct cgroup_subsys_state *pos_css;
1381 * Check all its siblings and call update_cpumasks_hier()
1382 * if their use_parent_ecpus flag is set in order for them
1383 * to use the right effective_cpus value.
1386 cpuset_for_each_child(sibling, pos_css, parent) {
1389 if (!sibling->use_parent_ecpus)
1392 update_cpumasks_hier(sibling, tmp);
1398 * update_cpumask - update the cpus_allowed mask of a cpuset and all tasks in it
1399 * @cs: the cpuset to consider
1400 * @trialcs: trial cpuset
1401 * @buf: buffer of cpu numbers written to this cpuset
1403 static int update_cpumask(struct cpuset *cs, struct cpuset *trialcs,
1407 struct tmpmasks tmp;
1409 /* top_cpuset.cpus_allowed tracks cpu_online_mask; it's read-only */
1410 if (cs == &top_cpuset)
1414 * An empty cpus_allowed is ok only if the cpuset has no tasks.
1415 * Since cpulist_parse() fails on an empty mask, we special case
1416 * that parsing. The validate_change() call ensures that cpusets
1417 * with tasks have cpus.
1420 cpumask_clear(trialcs->cpus_allowed);
1422 retval = cpulist_parse(buf, trialcs->cpus_allowed);
1426 if (!cpumask_subset(trialcs->cpus_allowed,
1427 top_cpuset.cpus_allowed))
1431 /* Nothing to do if the cpus didn't change */
1432 if (cpumask_equal(cs->cpus_allowed, trialcs->cpus_allowed))
1435 retval = validate_change(cs, trialcs);
1439 #ifdef CONFIG_CPUMASK_OFFSTACK
1441 * Use the cpumasks in trialcs for tmpmasks when they are pointers
1442 * to allocated cpumasks.
1444 tmp.addmask = trialcs->subparts_cpus;
1445 tmp.delmask = trialcs->effective_cpus;
1446 tmp.new_cpus = trialcs->cpus_allowed;
1449 if (cs->partition_root_state) {
1450 /* Cpumask of a partition root cannot be empty */
1451 if (cpumask_empty(trialcs->cpus_allowed))
1453 if (update_parent_subparts_cpumask(cs, partcmd_update,
1454 trialcs->cpus_allowed, &tmp) < 0)
1458 spin_lock_irq(&callback_lock);
1459 cpumask_copy(cs->cpus_allowed, trialcs->cpus_allowed);
1462 * Make sure that subparts_cpus is a subset of cpus_allowed.
1464 if (cs->nr_subparts_cpus) {
1465 cpumask_andnot(cs->subparts_cpus, cs->subparts_cpus,
1467 cs->nr_subparts_cpus = cpumask_weight(cs->subparts_cpus);
1469 spin_unlock_irq(&callback_lock);
1471 update_cpumasks_hier(cs, &tmp);
1473 if (cs->partition_root_state) {
1474 struct cpuset *parent = parent_cs(cs);
1477 * For partition root, update the cpumasks of sibling
1478 * cpusets if they use parent's effective_cpus.
1480 if (parent->child_ecpus_count)
1481 update_sibling_cpumasks(parent, cs, &tmp);
1487 * Migrate memory region from one set of nodes to another. This is
1488 * performed asynchronously as it can be called from process migration path
1489 * holding locks involved in process management. All mm migrations are
1490 * performed in the queued order and can be waited for by flushing
1491 * cpuset_migrate_mm_wq.
1494 struct cpuset_migrate_mm_work {
1495 struct work_struct work;
1496 struct mm_struct *mm;
1501 static void cpuset_migrate_mm_workfn(struct work_struct *work)
1503 struct cpuset_migrate_mm_work *mwork =
1504 container_of(work, struct cpuset_migrate_mm_work, work);
1506 /* on a wq worker, no need to worry about %current's mems_allowed */
1507 do_migrate_pages(mwork->mm, &mwork->from, &mwork->to, MPOL_MF_MOVE_ALL);
1512 static void cpuset_migrate_mm(struct mm_struct *mm, const nodemask_t *from,
1513 const nodemask_t *to)
1515 struct cpuset_migrate_mm_work *mwork;
1517 mwork = kzalloc(sizeof(*mwork), GFP_KERNEL);
1520 mwork->from = *from;
1522 INIT_WORK(&mwork->work, cpuset_migrate_mm_workfn);
1523 queue_work(cpuset_migrate_mm_wq, &mwork->work);
1529 static void cpuset_post_attach(void)
1531 flush_workqueue(cpuset_migrate_mm_wq);
1535 * cpuset_change_task_nodemask - change task's mems_allowed and mempolicy
1536 * @tsk: the task to change
1537 * @newmems: new nodes that the task will be set
1539 * We use the mems_allowed_seq seqlock to safely update both tsk->mems_allowed
1540 * and rebind an eventual tasks' mempolicy. If the task is allocating in
1541 * parallel, it might temporarily see an empty intersection, which results in
1542 * a seqlock check and retry before OOM or allocation failure.
1544 static void cpuset_change_task_nodemask(struct task_struct *tsk,
1545 nodemask_t *newmems)
1549 local_irq_disable();
1550 write_seqcount_begin(&tsk->mems_allowed_seq);
1552 nodes_or(tsk->mems_allowed, tsk->mems_allowed, *newmems);
1553 mpol_rebind_task(tsk, newmems);
1554 tsk->mems_allowed = *newmems;
1556 write_seqcount_end(&tsk->mems_allowed_seq);
1562 static void *cpuset_being_rebound;
1565 * update_tasks_nodemask - Update the nodemasks of tasks in the cpuset.
1566 * @cs: the cpuset in which each task's mems_allowed mask needs to be changed
1568 * Iterate through each task of @cs updating its mems_allowed to the
1569 * effective cpuset's. As this function is called with cpuset_mutex held,
1570 * cpuset membership stays stable.
1572 static void update_tasks_nodemask(struct cpuset *cs)
1574 static nodemask_t newmems; /* protected by cpuset_mutex */
1575 struct css_task_iter it;
1576 struct task_struct *task;
1578 cpuset_being_rebound = cs; /* causes mpol_dup() rebind */
1580 guarantee_online_mems(cs, &newmems);
1583 * The mpol_rebind_mm() call takes mmap_sem, which we couldn't
1584 * take while holding tasklist_lock. Forks can happen - the
1585 * mpol_dup() cpuset_being_rebound check will catch such forks,
1586 * and rebind their vma mempolicies too. Because we still hold
1587 * the global cpuset_mutex, we know that no other rebind effort
1588 * will be contending for the global variable cpuset_being_rebound.
1589 * It's ok if we rebind the same mm twice; mpol_rebind_mm()
1590 * is idempotent. Also migrate pages in each mm to new nodes.
1592 css_task_iter_start(&cs->css, 0, &it);
1593 while ((task = css_task_iter_next(&it))) {
1594 struct mm_struct *mm;
1597 cpuset_change_task_nodemask(task, &newmems);
1599 mm = get_task_mm(task);
1603 migrate = is_memory_migrate(cs);
1605 mpol_rebind_mm(mm, &cs->mems_allowed);
1607 cpuset_migrate_mm(mm, &cs->old_mems_allowed, &newmems);
1611 css_task_iter_end(&it);
1614 * All the tasks' nodemasks have been updated, update
1615 * cs->old_mems_allowed.
1617 cs->old_mems_allowed = newmems;
1619 /* We're done rebinding vmas to this cpuset's new mems_allowed. */
1620 cpuset_being_rebound = NULL;
1624 * update_nodemasks_hier - Update effective nodemasks and tasks in the subtree
1625 * @cs: the cpuset to consider
1626 * @new_mems: a temp variable for calculating new effective_mems
1628 * When configured nodemask is changed, the effective nodemasks of this cpuset
1629 * and all its descendants need to be updated.
1631 * On legacy hiearchy, effective_mems will be the same with mems_allowed.
1633 * Called with cpuset_mutex held
1635 static void update_nodemasks_hier(struct cpuset *cs, nodemask_t *new_mems)
1638 struct cgroup_subsys_state *pos_css;
1641 cpuset_for_each_descendant_pre(cp, pos_css, cs) {
1642 struct cpuset *parent = parent_cs(cp);
1644 nodes_and(*new_mems, cp->mems_allowed, parent->effective_mems);
1647 * If it becomes empty, inherit the effective mask of the
1648 * parent, which is guaranteed to have some MEMs.
1650 if (is_in_v2_mode() && nodes_empty(*new_mems))
1651 *new_mems = parent->effective_mems;
1653 /* Skip the whole subtree if the nodemask remains the same. */
1654 if (nodes_equal(*new_mems, cp->effective_mems)) {
1655 pos_css = css_rightmost_descendant(pos_css);
1659 if (!css_tryget_online(&cp->css))
1663 spin_lock_irq(&callback_lock);
1664 cp->effective_mems = *new_mems;
1665 spin_unlock_irq(&callback_lock);
1667 WARN_ON(!is_in_v2_mode() &&
1668 !nodes_equal(cp->mems_allowed, cp->effective_mems));
1670 update_tasks_nodemask(cp);
1679 * Handle user request to change the 'mems' memory placement
1680 * of a cpuset. Needs to validate the request, update the
1681 * cpusets mems_allowed, and for each task in the cpuset,
1682 * update mems_allowed and rebind task's mempolicy and any vma
1683 * mempolicies and if the cpuset is marked 'memory_migrate',
1684 * migrate the tasks pages to the new memory.
1686 * Call with cpuset_mutex held. May take callback_lock during call.
1687 * Will take tasklist_lock, scan tasklist for tasks in cpuset cs,
1688 * lock each such tasks mm->mmap_sem, scan its vma's and rebind
1689 * their mempolicies to the cpusets new mems_allowed.
1691 static int update_nodemask(struct cpuset *cs, struct cpuset *trialcs,
1697 * top_cpuset.mems_allowed tracks node_stats[N_MEMORY];
1700 if (cs == &top_cpuset) {
1706 * An empty mems_allowed is ok iff there are no tasks in the cpuset.
1707 * Since nodelist_parse() fails on an empty mask, we special case
1708 * that parsing. The validate_change() call ensures that cpusets
1709 * with tasks have memory.
1712 nodes_clear(trialcs->mems_allowed);
1714 retval = nodelist_parse(buf, trialcs->mems_allowed);
1718 if (!nodes_subset(trialcs->mems_allowed,
1719 top_cpuset.mems_allowed)) {
1725 if (nodes_equal(cs->mems_allowed, trialcs->mems_allowed)) {
1726 retval = 0; /* Too easy - nothing to do */
1729 retval = validate_change(cs, trialcs);
1733 spin_lock_irq(&callback_lock);
1734 cs->mems_allowed = trialcs->mems_allowed;
1735 spin_unlock_irq(&callback_lock);
1737 /* use trialcs->mems_allowed as a temp variable */
1738 update_nodemasks_hier(cs, &trialcs->mems_allowed);
1743 bool current_cpuset_is_being_rebound(void)
1748 ret = task_cs(current) == cpuset_being_rebound;
1754 static int update_relax_domain_level(struct cpuset *cs, s64 val)
1757 if (val < -1 || val >= sched_domain_level_max)
1761 if (val != cs->relax_domain_level) {
1762 cs->relax_domain_level = val;
1763 if (!cpumask_empty(cs->cpus_allowed) &&
1764 is_sched_load_balance(cs))
1765 rebuild_sched_domains_locked();
1772 * update_tasks_flags - update the spread flags of tasks in the cpuset.
1773 * @cs: the cpuset in which each task's spread flags needs to be changed
1775 * Iterate through each task of @cs updating its spread flags. As this
1776 * function is called with cpuset_mutex held, cpuset membership stays
1779 static void update_tasks_flags(struct cpuset *cs)
1781 struct css_task_iter it;
1782 struct task_struct *task;
1784 css_task_iter_start(&cs->css, 0, &it);
1785 while ((task = css_task_iter_next(&it)))
1786 cpuset_update_task_spread_flag(cs, task);
1787 css_task_iter_end(&it);
1791 * update_flag - read a 0 or a 1 in a file and update associated flag
1792 * bit: the bit to update (see cpuset_flagbits_t)
1793 * cs: the cpuset to update
1794 * turning_on: whether the flag is being set or cleared
1796 * Call with cpuset_mutex held.
1799 static int update_flag(cpuset_flagbits_t bit, struct cpuset *cs,
1802 struct cpuset *trialcs;
1803 int balance_flag_changed;
1804 int spread_flag_changed;
1807 trialcs = alloc_trial_cpuset(cs);
1812 set_bit(bit, &trialcs->flags);
1814 clear_bit(bit, &trialcs->flags);
1816 err = validate_change(cs, trialcs);
1820 balance_flag_changed = (is_sched_load_balance(cs) !=
1821 is_sched_load_balance(trialcs));
1823 spread_flag_changed = ((is_spread_slab(cs) != is_spread_slab(trialcs))
1824 || (is_spread_page(cs) != is_spread_page(trialcs)));
1826 spin_lock_irq(&callback_lock);
1827 cs->flags = trialcs->flags;
1828 spin_unlock_irq(&callback_lock);
1830 if (!cpumask_empty(trialcs->cpus_allowed) && balance_flag_changed)
1831 rebuild_sched_domains_locked();
1833 if (spread_flag_changed)
1834 update_tasks_flags(cs);
1836 free_cpuset(trialcs);
1841 * update_prstate - update partititon_root_state
1842 * cs: the cpuset to update
1843 * val: 0 - disabled, 1 - enabled
1845 * Call with cpuset_mutex held.
1847 static int update_prstate(struct cpuset *cs, int val)
1850 struct cpuset *parent = parent_cs(cs);
1851 struct tmpmasks tmp;
1853 if ((val != 0) && (val != 1))
1855 if (val == cs->partition_root_state)
1859 * Cannot force a partial or invalid partition root to a full
1862 if (val && cs->partition_root_state)
1865 if (alloc_cpumasks(NULL, &tmp))
1869 if (!cs->partition_root_state) {
1871 * Turning on partition root requires setting the
1872 * CS_CPU_EXCLUSIVE bit implicitly as well and cpus_allowed
1875 if (cpumask_empty(cs->cpus_allowed))
1878 err = update_flag(CS_CPU_EXCLUSIVE, cs, 1);
1882 err = update_parent_subparts_cpumask(cs, partcmd_enable,
1885 update_flag(CS_CPU_EXCLUSIVE, cs, 0);
1888 cs->partition_root_state = PRS_ENABLED;
1891 * Turning off partition root will clear the
1892 * CS_CPU_EXCLUSIVE bit.
1894 if (cs->partition_root_state == PRS_ERROR) {
1895 cs->partition_root_state = 0;
1896 update_flag(CS_CPU_EXCLUSIVE, cs, 0);
1901 err = update_parent_subparts_cpumask(cs, partcmd_disable,
1906 cs->partition_root_state = 0;
1908 /* Turning off CS_CPU_EXCLUSIVE will not return error */
1909 update_flag(CS_CPU_EXCLUSIVE, cs, 0);
1913 * Update cpumask of parent's tasks except when it is the top
1914 * cpuset as some system daemons cannot be mapped to other CPUs.
1916 if (parent != &top_cpuset)
1917 update_tasks_cpumask(parent);
1919 if (parent->child_ecpus_count)
1920 update_sibling_cpumasks(parent, cs, &tmp);
1922 rebuild_sched_domains_locked();
1924 free_cpumasks(NULL, &tmp);
1929 * Frequency meter - How fast is some event occurring?
1931 * These routines manage a digitally filtered, constant time based,
1932 * event frequency meter. There are four routines:
1933 * fmeter_init() - initialize a frequency meter.
1934 * fmeter_markevent() - called each time the event happens.
1935 * fmeter_getrate() - returns the recent rate of such events.
1936 * fmeter_update() - internal routine used to update fmeter.
1938 * A common data structure is passed to each of these routines,
1939 * which is used to keep track of the state required to manage the
1940 * frequency meter and its digital filter.
1942 * The filter works on the number of events marked per unit time.
1943 * The filter is single-pole low-pass recursive (IIR). The time unit
1944 * is 1 second. Arithmetic is done using 32-bit integers scaled to
1945 * simulate 3 decimal digits of precision (multiplied by 1000).
1947 * With an FM_COEF of 933, and a time base of 1 second, the filter
1948 * has a half-life of 10 seconds, meaning that if the events quit
1949 * happening, then the rate returned from the fmeter_getrate()
1950 * will be cut in half each 10 seconds, until it converges to zero.
1952 * It is not worth doing a real infinitely recursive filter. If more
1953 * than FM_MAXTICKS ticks have elapsed since the last filter event,
1954 * just compute FM_MAXTICKS ticks worth, by which point the level
1957 * Limit the count of unprocessed events to FM_MAXCNT, so as to avoid
1958 * arithmetic overflow in the fmeter_update() routine.
1960 * Given the simple 32 bit integer arithmetic used, this meter works
1961 * best for reporting rates between one per millisecond (msec) and
1962 * one per 32 (approx) seconds. At constant rates faster than one
1963 * per msec it maxes out at values just under 1,000,000. At constant
1964 * rates between one per msec, and one per second it will stabilize
1965 * to a value N*1000, where N is the rate of events per second.
1966 * At constant rates between one per second and one per 32 seconds,
1967 * it will be choppy, moving up on the seconds that have an event,
1968 * and then decaying until the next event. At rates slower than
1969 * about one in 32 seconds, it decays all the way back to zero between
1973 #define FM_COEF 933 /* coefficient for half-life of 10 secs */
1974 #define FM_MAXTICKS ((u32)99) /* useless computing more ticks than this */
1975 #define FM_MAXCNT 1000000 /* limit cnt to avoid overflow */
1976 #define FM_SCALE 1000 /* faux fixed point scale */
1978 /* Initialize a frequency meter */
1979 static void fmeter_init(struct fmeter *fmp)
1984 spin_lock_init(&fmp->lock);
1987 /* Internal meter update - process cnt events and update value */
1988 static void fmeter_update(struct fmeter *fmp)
1993 now = ktime_get_seconds();
1994 ticks = now - fmp->time;
1999 ticks = min(FM_MAXTICKS, ticks);
2001 fmp->val = (FM_COEF * fmp->val) / FM_SCALE;
2004 fmp->val += ((FM_SCALE - FM_COEF) * fmp->cnt) / FM_SCALE;
2008 /* Process any previous ticks, then bump cnt by one (times scale). */
2009 static void fmeter_markevent(struct fmeter *fmp)
2011 spin_lock(&fmp->lock);
2013 fmp->cnt = min(FM_MAXCNT, fmp->cnt + FM_SCALE);
2014 spin_unlock(&fmp->lock);
2017 /* Process any previous ticks, then return current value. */
2018 static int fmeter_getrate(struct fmeter *fmp)
2022 spin_lock(&fmp->lock);
2025 spin_unlock(&fmp->lock);
2029 static struct cpuset *cpuset_attach_old_cs;
2031 /* Called by cgroups to determine if a cpuset is usable; cpuset_mutex held */
2032 static int cpuset_can_attach(struct cgroup_taskset *tset)
2034 struct cgroup_subsys_state *css;
2036 struct task_struct *task;
2039 /* used later by cpuset_attach() */
2040 cpuset_attach_old_cs = task_cs(cgroup_taskset_first(tset, &css));
2043 mutex_lock(&cpuset_mutex);
2045 /* allow moving tasks into an empty cpuset if on default hierarchy */
2047 if (!is_in_v2_mode() &&
2048 (cpumask_empty(cs->cpus_allowed) || nodes_empty(cs->mems_allowed)))
2051 cgroup_taskset_for_each(task, css, tset) {
2052 ret = task_can_attach(task, cs->cpus_allowed);
2055 ret = security_task_setscheduler(task);
2061 * Mark attach is in progress. This makes validate_change() fail
2062 * changes which zero cpus/mems_allowed.
2064 cs->attach_in_progress++;
2067 mutex_unlock(&cpuset_mutex);
2071 static void cpuset_cancel_attach(struct cgroup_taskset *tset)
2073 struct cgroup_subsys_state *css;
2075 cgroup_taskset_first(tset, &css);
2077 mutex_lock(&cpuset_mutex);
2078 css_cs(css)->attach_in_progress--;
2079 mutex_unlock(&cpuset_mutex);
2083 * Protected by cpuset_mutex. cpus_attach is used only by cpuset_attach()
2084 * but we can't allocate it dynamically there. Define it global and
2085 * allocate from cpuset_init().
2087 static cpumask_var_t cpus_attach;
2089 static void cpuset_attach(struct cgroup_taskset *tset)
2091 /* static buf protected by cpuset_mutex */
2092 static nodemask_t cpuset_attach_nodemask_to;
2093 struct task_struct *task;
2094 struct task_struct *leader;
2095 struct cgroup_subsys_state *css;
2097 struct cpuset *oldcs = cpuset_attach_old_cs;
2099 cgroup_taskset_first(tset, &css);
2102 mutex_lock(&cpuset_mutex);
2104 /* prepare for attach */
2105 if (cs == &top_cpuset)
2106 cpumask_copy(cpus_attach, cpu_possible_mask);
2108 guarantee_online_cpus(cs, cpus_attach);
2110 guarantee_online_mems(cs, &cpuset_attach_nodemask_to);
2112 cgroup_taskset_for_each(task, css, tset) {
2114 * can_attach beforehand should guarantee that this doesn't
2115 * fail. TODO: have a better way to handle failure here
2117 WARN_ON_ONCE(set_cpus_allowed_ptr(task, cpus_attach));
2119 cpuset_change_task_nodemask(task, &cpuset_attach_nodemask_to);
2120 cpuset_update_task_spread_flag(cs, task);
2124 * Change mm for all threadgroup leaders. This is expensive and may
2125 * sleep and should be moved outside migration path proper.
2127 cpuset_attach_nodemask_to = cs->effective_mems;
2128 cgroup_taskset_for_each_leader(leader, css, tset) {
2129 struct mm_struct *mm = get_task_mm(leader);
2132 mpol_rebind_mm(mm, &cpuset_attach_nodemask_to);
2135 * old_mems_allowed is the same with mems_allowed
2136 * here, except if this task is being moved
2137 * automatically due to hotplug. In that case
2138 * @mems_allowed has been updated and is empty, so
2139 * @old_mems_allowed is the right nodesets that we
2142 if (is_memory_migrate(cs))
2143 cpuset_migrate_mm(mm, &oldcs->old_mems_allowed,
2144 &cpuset_attach_nodemask_to);
2150 cs->old_mems_allowed = cpuset_attach_nodemask_to;
2152 cs->attach_in_progress--;
2153 if (!cs->attach_in_progress)
2154 wake_up(&cpuset_attach_wq);
2156 mutex_unlock(&cpuset_mutex);
2159 /* The various types of files and directories in a cpuset file system */
2162 FILE_MEMORY_MIGRATE,
2165 FILE_EFFECTIVE_CPULIST,
2166 FILE_EFFECTIVE_MEMLIST,
2167 FILE_SUBPARTS_CPULIST,
2171 FILE_SCHED_LOAD_BALANCE,
2172 FILE_PARTITION_ROOT,
2173 FILE_SCHED_RELAX_DOMAIN_LEVEL,
2174 FILE_MEMORY_PRESSURE_ENABLED,
2175 FILE_MEMORY_PRESSURE,
2178 } cpuset_filetype_t;
2180 static int cpuset_write_u64(struct cgroup_subsys_state *css, struct cftype *cft,
2183 struct cpuset *cs = css_cs(css);
2184 cpuset_filetype_t type = cft->private;
2187 mutex_lock(&cpuset_mutex);
2188 if (!is_cpuset_online(cs)) {
2194 case FILE_CPU_EXCLUSIVE:
2195 retval = update_flag(CS_CPU_EXCLUSIVE, cs, val);
2197 case FILE_MEM_EXCLUSIVE:
2198 retval = update_flag(CS_MEM_EXCLUSIVE, cs, val);
2200 case FILE_MEM_HARDWALL:
2201 retval = update_flag(CS_MEM_HARDWALL, cs, val);
2203 case FILE_SCHED_LOAD_BALANCE:
2204 retval = update_flag(CS_SCHED_LOAD_BALANCE, cs, val);
2206 case FILE_MEMORY_MIGRATE:
2207 retval = update_flag(CS_MEMORY_MIGRATE, cs, val);
2209 case FILE_MEMORY_PRESSURE_ENABLED:
2210 cpuset_memory_pressure_enabled = !!val;
2212 case FILE_SPREAD_PAGE:
2213 retval = update_flag(CS_SPREAD_PAGE, cs, val);
2215 case FILE_SPREAD_SLAB:
2216 retval = update_flag(CS_SPREAD_SLAB, cs, val);
2223 mutex_unlock(&cpuset_mutex);
2227 static int cpuset_write_s64(struct cgroup_subsys_state *css, struct cftype *cft,
2230 struct cpuset *cs = css_cs(css);
2231 cpuset_filetype_t type = cft->private;
2232 int retval = -ENODEV;
2234 mutex_lock(&cpuset_mutex);
2235 if (!is_cpuset_online(cs))
2239 case FILE_SCHED_RELAX_DOMAIN_LEVEL:
2240 retval = update_relax_domain_level(cs, val);
2247 mutex_unlock(&cpuset_mutex);
2252 * Common handling for a write to a "cpus" or "mems" file.
2254 static ssize_t cpuset_write_resmask(struct kernfs_open_file *of,
2255 char *buf, size_t nbytes, loff_t off)
2257 struct cpuset *cs = css_cs(of_css(of));
2258 struct cpuset *trialcs;
2259 int retval = -ENODEV;
2261 buf = strstrip(buf);
2264 * CPU or memory hotunplug may leave @cs w/o any execution
2265 * resources, in which case the hotplug code asynchronously updates
2266 * configuration and transfers all tasks to the nearest ancestor
2267 * which can execute.
2269 * As writes to "cpus" or "mems" may restore @cs's execution
2270 * resources, wait for the previously scheduled operations before
2271 * proceeding, so that we don't end up keep removing tasks added
2272 * after execution capability is restored.
2274 * cpuset_hotplug_work calls back into cgroup core via
2275 * cgroup_transfer_tasks() and waiting for it from a cgroupfs
2276 * operation like this one can lead to a deadlock through kernfs
2277 * active_ref protection. Let's break the protection. Losing the
2278 * protection is okay as we check whether @cs is online after
2279 * grabbing cpuset_mutex anyway. This only happens on the legacy
2283 kernfs_break_active_protection(of->kn);
2284 flush_work(&cpuset_hotplug_work);
2286 mutex_lock(&cpuset_mutex);
2287 if (!is_cpuset_online(cs))
2290 trialcs = alloc_trial_cpuset(cs);
2296 switch (of_cft(of)->private) {
2298 retval = update_cpumask(cs, trialcs, buf);
2301 retval = update_nodemask(cs, trialcs, buf);
2308 free_cpuset(trialcs);
2310 mutex_unlock(&cpuset_mutex);
2311 kernfs_unbreak_active_protection(of->kn);
2313 flush_workqueue(cpuset_migrate_mm_wq);
2314 return retval ?: nbytes;
2318 * These ascii lists should be read in a single call, by using a user
2319 * buffer large enough to hold the entire map. If read in smaller
2320 * chunks, there is no guarantee of atomicity. Since the display format
2321 * used, list of ranges of sequential numbers, is variable length,
2322 * and since these maps can change value dynamically, one could read
2323 * gibberish by doing partial reads while a list was changing.
2325 static int cpuset_common_seq_show(struct seq_file *sf, void *v)
2327 struct cpuset *cs = css_cs(seq_css(sf));
2328 cpuset_filetype_t type = seq_cft(sf)->private;
2331 spin_lock_irq(&callback_lock);
2335 seq_printf(sf, "%*pbl\n", cpumask_pr_args(cs->cpus_allowed));
2338 seq_printf(sf, "%*pbl\n", nodemask_pr_args(&cs->mems_allowed));
2340 case FILE_EFFECTIVE_CPULIST:
2341 seq_printf(sf, "%*pbl\n", cpumask_pr_args(cs->effective_cpus));
2343 case FILE_EFFECTIVE_MEMLIST:
2344 seq_printf(sf, "%*pbl\n", nodemask_pr_args(&cs->effective_mems));
2346 case FILE_SUBPARTS_CPULIST:
2347 seq_printf(sf, "%*pbl\n", cpumask_pr_args(cs->subparts_cpus));
2353 spin_unlock_irq(&callback_lock);
2357 static u64 cpuset_read_u64(struct cgroup_subsys_state *css, struct cftype *cft)
2359 struct cpuset *cs = css_cs(css);
2360 cpuset_filetype_t type = cft->private;
2362 case FILE_CPU_EXCLUSIVE:
2363 return is_cpu_exclusive(cs);
2364 case FILE_MEM_EXCLUSIVE:
2365 return is_mem_exclusive(cs);
2366 case FILE_MEM_HARDWALL:
2367 return is_mem_hardwall(cs);
2368 case FILE_SCHED_LOAD_BALANCE:
2369 return is_sched_load_balance(cs);
2370 case FILE_MEMORY_MIGRATE:
2371 return is_memory_migrate(cs);
2372 case FILE_MEMORY_PRESSURE_ENABLED:
2373 return cpuset_memory_pressure_enabled;
2374 case FILE_MEMORY_PRESSURE:
2375 return fmeter_getrate(&cs->fmeter);
2376 case FILE_SPREAD_PAGE:
2377 return is_spread_page(cs);
2378 case FILE_SPREAD_SLAB:
2379 return is_spread_slab(cs);
2384 /* Unreachable but makes gcc happy */
2388 static s64 cpuset_read_s64(struct cgroup_subsys_state *css, struct cftype *cft)
2390 struct cpuset *cs = css_cs(css);
2391 cpuset_filetype_t type = cft->private;
2393 case FILE_SCHED_RELAX_DOMAIN_LEVEL:
2394 return cs->relax_domain_level;
2399 /* Unrechable but makes gcc happy */
2403 static int sched_partition_show(struct seq_file *seq, void *v)
2405 struct cpuset *cs = css_cs(seq_css(seq));
2407 switch (cs->partition_root_state) {
2409 seq_puts(seq, "root\n");
2412 seq_puts(seq, "member\n");
2415 seq_puts(seq, "root invalid\n");
2421 static ssize_t sched_partition_write(struct kernfs_open_file *of, char *buf,
2422 size_t nbytes, loff_t off)
2424 struct cpuset *cs = css_cs(of_css(of));
2426 int retval = -ENODEV;
2428 buf = strstrip(buf);
2431 * Convert "root" to ENABLED, and convert "member" to DISABLED.
2433 if (!strcmp(buf, "root"))
2435 else if (!strcmp(buf, "member"))
2441 mutex_lock(&cpuset_mutex);
2442 if (!is_cpuset_online(cs))
2445 retval = update_prstate(cs, val);
2447 mutex_unlock(&cpuset_mutex);
2449 return retval ?: nbytes;
2453 * for the common functions, 'private' gives the type of file
2456 static struct cftype legacy_files[] = {
2459 .seq_show = cpuset_common_seq_show,
2460 .write = cpuset_write_resmask,
2461 .max_write_len = (100U + 6 * NR_CPUS),
2462 .private = FILE_CPULIST,
2467 .seq_show = cpuset_common_seq_show,
2468 .write = cpuset_write_resmask,
2469 .max_write_len = (100U + 6 * MAX_NUMNODES),
2470 .private = FILE_MEMLIST,
2474 .name = "effective_cpus",
2475 .seq_show = cpuset_common_seq_show,
2476 .private = FILE_EFFECTIVE_CPULIST,
2480 .name = "effective_mems",
2481 .seq_show = cpuset_common_seq_show,
2482 .private = FILE_EFFECTIVE_MEMLIST,
2486 .name = "cpu_exclusive",
2487 .read_u64 = cpuset_read_u64,
2488 .write_u64 = cpuset_write_u64,
2489 .private = FILE_CPU_EXCLUSIVE,
2493 .name = "mem_exclusive",
2494 .read_u64 = cpuset_read_u64,
2495 .write_u64 = cpuset_write_u64,
2496 .private = FILE_MEM_EXCLUSIVE,
2500 .name = "mem_hardwall",
2501 .read_u64 = cpuset_read_u64,
2502 .write_u64 = cpuset_write_u64,
2503 .private = FILE_MEM_HARDWALL,
2507 .name = "sched_load_balance",
2508 .read_u64 = cpuset_read_u64,
2509 .write_u64 = cpuset_write_u64,
2510 .private = FILE_SCHED_LOAD_BALANCE,
2514 .name = "sched_relax_domain_level",
2515 .read_s64 = cpuset_read_s64,
2516 .write_s64 = cpuset_write_s64,
2517 .private = FILE_SCHED_RELAX_DOMAIN_LEVEL,
2521 .name = "memory_migrate",
2522 .read_u64 = cpuset_read_u64,
2523 .write_u64 = cpuset_write_u64,
2524 .private = FILE_MEMORY_MIGRATE,
2528 .name = "memory_pressure",
2529 .read_u64 = cpuset_read_u64,
2530 .private = FILE_MEMORY_PRESSURE,
2534 .name = "memory_spread_page",
2535 .read_u64 = cpuset_read_u64,
2536 .write_u64 = cpuset_write_u64,
2537 .private = FILE_SPREAD_PAGE,
2541 .name = "memory_spread_slab",
2542 .read_u64 = cpuset_read_u64,
2543 .write_u64 = cpuset_write_u64,
2544 .private = FILE_SPREAD_SLAB,
2548 .name = "memory_pressure_enabled",
2549 .flags = CFTYPE_ONLY_ON_ROOT,
2550 .read_u64 = cpuset_read_u64,
2551 .write_u64 = cpuset_write_u64,
2552 .private = FILE_MEMORY_PRESSURE_ENABLED,
2559 * This is currently a minimal set for the default hierarchy. It can be
2560 * expanded later on by migrating more features and control files from v1.
2562 static struct cftype dfl_files[] = {
2565 .seq_show = cpuset_common_seq_show,
2566 .write = cpuset_write_resmask,
2567 .max_write_len = (100U + 6 * NR_CPUS),
2568 .private = FILE_CPULIST,
2569 .flags = CFTYPE_NOT_ON_ROOT,
2574 .seq_show = cpuset_common_seq_show,
2575 .write = cpuset_write_resmask,
2576 .max_write_len = (100U + 6 * MAX_NUMNODES),
2577 .private = FILE_MEMLIST,
2578 .flags = CFTYPE_NOT_ON_ROOT,
2582 .name = "cpus.effective",
2583 .seq_show = cpuset_common_seq_show,
2584 .private = FILE_EFFECTIVE_CPULIST,
2588 .name = "mems.effective",
2589 .seq_show = cpuset_common_seq_show,
2590 .private = FILE_EFFECTIVE_MEMLIST,
2594 .name = "cpus.partition",
2595 .seq_show = sched_partition_show,
2596 .write = sched_partition_write,
2597 .private = FILE_PARTITION_ROOT,
2598 .flags = CFTYPE_NOT_ON_ROOT,
2602 .name = "cpus.subpartitions",
2603 .seq_show = cpuset_common_seq_show,
2604 .private = FILE_SUBPARTS_CPULIST,
2605 .flags = CFTYPE_DEBUG,
2613 * cpuset_css_alloc - allocate a cpuset css
2614 * cgrp: control group that the new cpuset will be part of
2617 static struct cgroup_subsys_state *
2618 cpuset_css_alloc(struct cgroup_subsys_state *parent_css)
2623 return &top_cpuset.css;
2625 cs = kzalloc(sizeof(*cs), GFP_KERNEL);
2627 return ERR_PTR(-ENOMEM);
2629 if (alloc_cpumasks(cs, NULL)) {
2631 return ERR_PTR(-ENOMEM);
2634 set_bit(CS_SCHED_LOAD_BALANCE, &cs->flags);
2635 nodes_clear(cs->mems_allowed);
2636 nodes_clear(cs->effective_mems);
2637 fmeter_init(&cs->fmeter);
2638 cs->relax_domain_level = -1;
2643 static int cpuset_css_online(struct cgroup_subsys_state *css)
2645 struct cpuset *cs = css_cs(css);
2646 struct cpuset *parent = parent_cs(cs);
2647 struct cpuset *tmp_cs;
2648 struct cgroup_subsys_state *pos_css;
2653 mutex_lock(&cpuset_mutex);
2655 set_bit(CS_ONLINE, &cs->flags);
2656 if (is_spread_page(parent))
2657 set_bit(CS_SPREAD_PAGE, &cs->flags);
2658 if (is_spread_slab(parent))
2659 set_bit(CS_SPREAD_SLAB, &cs->flags);
2663 spin_lock_irq(&callback_lock);
2664 if (is_in_v2_mode()) {
2665 cpumask_copy(cs->effective_cpus, parent->effective_cpus);
2666 cs->effective_mems = parent->effective_mems;
2667 cs->use_parent_ecpus = true;
2668 parent->child_ecpus_count++;
2670 spin_unlock_irq(&callback_lock);
2672 if (!test_bit(CGRP_CPUSET_CLONE_CHILDREN, &css->cgroup->flags))
2676 * Clone @parent's configuration if CGRP_CPUSET_CLONE_CHILDREN is
2677 * set. This flag handling is implemented in cgroup core for
2678 * histrical reasons - the flag may be specified during mount.
2680 * Currently, if any sibling cpusets have exclusive cpus or mem, we
2681 * refuse to clone the configuration - thereby refusing the task to
2682 * be entered, and as a result refusing the sys_unshare() or
2683 * clone() which initiated it. If this becomes a problem for some
2684 * users who wish to allow that scenario, then this could be
2685 * changed to grant parent->cpus_allowed-sibling_cpus_exclusive
2686 * (and likewise for mems) to the new cgroup.
2689 cpuset_for_each_child(tmp_cs, pos_css, parent) {
2690 if (is_mem_exclusive(tmp_cs) || is_cpu_exclusive(tmp_cs)) {
2697 spin_lock_irq(&callback_lock);
2698 cs->mems_allowed = parent->mems_allowed;
2699 cs->effective_mems = parent->mems_allowed;
2700 cpumask_copy(cs->cpus_allowed, parent->cpus_allowed);
2701 cpumask_copy(cs->effective_cpus, parent->cpus_allowed);
2702 spin_unlock_irq(&callback_lock);
2704 mutex_unlock(&cpuset_mutex);
2709 * If the cpuset being removed has its flag 'sched_load_balance'
2710 * enabled, then simulate turning sched_load_balance off, which
2711 * will call rebuild_sched_domains_locked(). That is not needed
2712 * in the default hierarchy where only changes in partition
2713 * will cause repartitioning.
2715 * If the cpuset has the 'sched.partition' flag enabled, simulate
2716 * turning 'sched.partition" off.
2719 static void cpuset_css_offline(struct cgroup_subsys_state *css)
2721 struct cpuset *cs = css_cs(css);
2723 mutex_lock(&cpuset_mutex);
2725 if (is_partition_root(cs))
2726 update_prstate(cs, 0);
2728 if (!cgroup_subsys_on_dfl(cpuset_cgrp_subsys) &&
2729 is_sched_load_balance(cs))
2730 update_flag(CS_SCHED_LOAD_BALANCE, cs, 0);
2732 if (cs->use_parent_ecpus) {
2733 struct cpuset *parent = parent_cs(cs);
2735 cs->use_parent_ecpus = false;
2736 parent->child_ecpus_count--;
2740 clear_bit(CS_ONLINE, &cs->flags);
2742 mutex_unlock(&cpuset_mutex);
2745 static void cpuset_css_free(struct cgroup_subsys_state *css)
2747 struct cpuset *cs = css_cs(css);
2752 static void cpuset_bind(struct cgroup_subsys_state *root_css)
2754 mutex_lock(&cpuset_mutex);
2755 spin_lock_irq(&callback_lock);
2757 if (is_in_v2_mode()) {
2758 cpumask_copy(top_cpuset.cpus_allowed, cpu_possible_mask);
2759 top_cpuset.mems_allowed = node_possible_map;
2761 cpumask_copy(top_cpuset.cpus_allowed,
2762 top_cpuset.effective_cpus);
2763 top_cpuset.mems_allowed = top_cpuset.effective_mems;
2766 spin_unlock_irq(&callback_lock);
2767 mutex_unlock(&cpuset_mutex);
2771 * Make sure the new task conform to the current state of its parent,
2772 * which could have been changed by cpuset just after it inherits the
2773 * state from the parent and before it sits on the cgroup's task list.
2775 static void cpuset_fork(struct task_struct *task)
2777 if (task_css_is_root(task, cpuset_cgrp_id))
2780 set_cpus_allowed_ptr(task, current->cpus_ptr);
2781 task->mems_allowed = current->mems_allowed;
2784 struct cgroup_subsys cpuset_cgrp_subsys = {
2785 .css_alloc = cpuset_css_alloc,
2786 .css_online = cpuset_css_online,
2787 .css_offline = cpuset_css_offline,
2788 .css_free = cpuset_css_free,
2789 .can_attach = cpuset_can_attach,
2790 .cancel_attach = cpuset_cancel_attach,
2791 .attach = cpuset_attach,
2792 .post_attach = cpuset_post_attach,
2793 .bind = cpuset_bind,
2794 .fork = cpuset_fork,
2795 .legacy_cftypes = legacy_files,
2796 .dfl_cftypes = dfl_files,
2802 * cpuset_init - initialize cpusets at system boot
2804 * Description: Initialize top_cpuset
2807 int __init cpuset_init(void)
2809 BUG_ON(!alloc_cpumask_var(&top_cpuset.cpus_allowed, GFP_KERNEL));
2810 BUG_ON(!alloc_cpumask_var(&top_cpuset.effective_cpus, GFP_KERNEL));
2811 BUG_ON(!zalloc_cpumask_var(&top_cpuset.subparts_cpus, GFP_KERNEL));
2813 cpumask_setall(top_cpuset.cpus_allowed);
2814 nodes_setall(top_cpuset.mems_allowed);
2815 cpumask_setall(top_cpuset.effective_cpus);
2816 nodes_setall(top_cpuset.effective_mems);
2818 fmeter_init(&top_cpuset.fmeter);
2819 set_bit(CS_SCHED_LOAD_BALANCE, &top_cpuset.flags);
2820 top_cpuset.relax_domain_level = -1;
2822 BUG_ON(!alloc_cpumask_var(&cpus_attach, GFP_KERNEL));
2828 * If CPU and/or memory hotplug handlers, below, unplug any CPUs
2829 * or memory nodes, we need to walk over the cpuset hierarchy,
2830 * removing that CPU or node from all cpusets. If this removes the
2831 * last CPU or node from a cpuset, then move the tasks in the empty
2832 * cpuset to its next-highest non-empty parent.
2834 static void remove_tasks_in_empty_cpuset(struct cpuset *cs)
2836 struct cpuset *parent;
2839 * Find its next-highest non-empty parent, (top cpuset
2840 * has online cpus, so can't be empty).
2842 parent = parent_cs(cs);
2843 while (cpumask_empty(parent->cpus_allowed) ||
2844 nodes_empty(parent->mems_allowed))
2845 parent = parent_cs(parent);
2847 if (cgroup_transfer_tasks(parent->css.cgroup, cs->css.cgroup)) {
2848 pr_err("cpuset: failed to transfer tasks out of empty cpuset ");
2849 pr_cont_cgroup_name(cs->css.cgroup);
2855 hotplug_update_tasks_legacy(struct cpuset *cs,
2856 struct cpumask *new_cpus, nodemask_t *new_mems,
2857 bool cpus_updated, bool mems_updated)
2861 spin_lock_irq(&callback_lock);
2862 cpumask_copy(cs->cpus_allowed, new_cpus);
2863 cpumask_copy(cs->effective_cpus, new_cpus);
2864 cs->mems_allowed = *new_mems;
2865 cs->effective_mems = *new_mems;
2866 spin_unlock_irq(&callback_lock);
2869 * Don't call update_tasks_cpumask() if the cpuset becomes empty,
2870 * as the tasks will be migratecd to an ancestor.
2872 if (cpus_updated && !cpumask_empty(cs->cpus_allowed))
2873 update_tasks_cpumask(cs);
2874 if (mems_updated && !nodes_empty(cs->mems_allowed))
2875 update_tasks_nodemask(cs);
2877 is_empty = cpumask_empty(cs->cpus_allowed) ||
2878 nodes_empty(cs->mems_allowed);
2880 mutex_unlock(&cpuset_mutex);
2883 * Move tasks to the nearest ancestor with execution resources,
2884 * This is full cgroup operation which will also call back into
2885 * cpuset. Should be done outside any lock.
2888 remove_tasks_in_empty_cpuset(cs);
2890 mutex_lock(&cpuset_mutex);
2894 hotplug_update_tasks(struct cpuset *cs,
2895 struct cpumask *new_cpus, nodemask_t *new_mems,
2896 bool cpus_updated, bool mems_updated)
2898 if (cpumask_empty(new_cpus))
2899 cpumask_copy(new_cpus, parent_cs(cs)->effective_cpus);
2900 if (nodes_empty(*new_mems))
2901 *new_mems = parent_cs(cs)->effective_mems;
2903 spin_lock_irq(&callback_lock);
2904 cpumask_copy(cs->effective_cpus, new_cpus);
2905 cs->effective_mems = *new_mems;
2906 spin_unlock_irq(&callback_lock);
2909 update_tasks_cpumask(cs);
2911 update_tasks_nodemask(cs);
2914 static bool force_rebuild;
2916 void cpuset_force_rebuild(void)
2918 force_rebuild = true;
2922 * cpuset_hotplug_update_tasks - update tasks in a cpuset for hotunplug
2923 * @cs: cpuset in interest
2924 * @tmp: the tmpmasks structure pointer
2926 * Compare @cs's cpu and mem masks against top_cpuset and if some have gone
2927 * offline, update @cs accordingly. If @cs ends up with no CPU or memory,
2928 * all its tasks are moved to the nearest ancestor with both resources.
2930 static void cpuset_hotplug_update_tasks(struct cpuset *cs, struct tmpmasks *tmp)
2932 static cpumask_t new_cpus;
2933 static nodemask_t new_mems;
2936 struct cpuset *parent;
2938 wait_event(cpuset_attach_wq, cs->attach_in_progress == 0);
2940 mutex_lock(&cpuset_mutex);
2943 * We have raced with task attaching. We wait until attaching
2944 * is finished, so we won't attach a task to an empty cpuset.
2946 if (cs->attach_in_progress) {
2947 mutex_unlock(&cpuset_mutex);
2951 parent = parent_cs(cs);
2952 compute_effective_cpumask(&new_cpus, cs, parent);
2953 nodes_and(new_mems, cs->mems_allowed, parent->effective_mems);
2955 if (cs->nr_subparts_cpus)
2957 * Make sure that CPUs allocated to child partitions
2958 * do not show up in effective_cpus.
2960 cpumask_andnot(&new_cpus, &new_cpus, cs->subparts_cpus);
2962 if (!tmp || !cs->partition_root_state)
2966 * In the unlikely event that a partition root has empty
2967 * effective_cpus or its parent becomes erroneous, we have to
2968 * transition it to the erroneous state.
2970 if (is_partition_root(cs) && (cpumask_empty(&new_cpus) ||
2971 (parent->partition_root_state == PRS_ERROR))) {
2972 if (cs->nr_subparts_cpus) {
2973 cs->nr_subparts_cpus = 0;
2974 cpumask_clear(cs->subparts_cpus);
2975 compute_effective_cpumask(&new_cpus, cs, parent);
2979 * If the effective_cpus is empty because the child
2980 * partitions take away all the CPUs, we can keep
2981 * the current partition and let the child partitions
2982 * fight for available CPUs.
2984 if ((parent->partition_root_state == PRS_ERROR) ||
2985 cpumask_empty(&new_cpus)) {
2986 update_parent_subparts_cpumask(cs, partcmd_disable,
2988 cs->partition_root_state = PRS_ERROR;
2990 cpuset_force_rebuild();
2994 * On the other hand, an erroneous partition root may be transitioned
2995 * back to a regular one or a partition root with no CPU allocated
2996 * from the parent may change to erroneous.
2998 if (is_partition_root(parent) &&
2999 ((cs->partition_root_state == PRS_ERROR) ||
3000 !cpumask_intersects(&new_cpus, parent->subparts_cpus)) &&
3001 update_parent_subparts_cpumask(cs, partcmd_update, NULL, tmp))
3002 cpuset_force_rebuild();
3005 cpus_updated = !cpumask_equal(&new_cpus, cs->effective_cpus);
3006 mems_updated = !nodes_equal(new_mems, cs->effective_mems);
3008 if (is_in_v2_mode())
3009 hotplug_update_tasks(cs, &new_cpus, &new_mems,
3010 cpus_updated, mems_updated);
3012 hotplug_update_tasks_legacy(cs, &new_cpus, &new_mems,
3013 cpus_updated, mems_updated);
3015 mutex_unlock(&cpuset_mutex);
3019 * cpuset_hotplug_workfn - handle CPU/memory hotunplug for a cpuset
3021 * This function is called after either CPU or memory configuration has
3022 * changed and updates cpuset accordingly. The top_cpuset is always
3023 * synchronized to cpu_active_mask and N_MEMORY, which is necessary in
3024 * order to make cpusets transparent (of no affect) on systems that are
3025 * actively using CPU hotplug but making no active use of cpusets.
3027 * Non-root cpusets are only affected by offlining. If any CPUs or memory
3028 * nodes have been taken down, cpuset_hotplug_update_tasks() is invoked on
3031 * Note that CPU offlining during suspend is ignored. We don't modify
3032 * cpusets across suspend/resume cycles at all.
3034 static void cpuset_hotplug_workfn(struct work_struct *work)
3036 static cpumask_t new_cpus;
3037 static nodemask_t new_mems;
3038 bool cpus_updated, mems_updated;
3039 bool on_dfl = is_in_v2_mode();
3040 struct tmpmasks tmp, *ptmp = NULL;
3042 if (on_dfl && !alloc_cpumasks(NULL, &tmp))
3045 mutex_lock(&cpuset_mutex);
3047 /* fetch the available cpus/mems and find out which changed how */
3048 cpumask_copy(&new_cpus, cpu_active_mask);
3049 new_mems = node_states[N_MEMORY];
3052 * If subparts_cpus is populated, it is likely that the check below
3053 * will produce a false positive on cpus_updated when the cpu list
3054 * isn't changed. It is extra work, but it is better to be safe.
3056 cpus_updated = !cpumask_equal(top_cpuset.effective_cpus, &new_cpus);
3057 mems_updated = !nodes_equal(top_cpuset.effective_mems, new_mems);
3059 /* synchronize cpus_allowed to cpu_active_mask */
3061 spin_lock_irq(&callback_lock);
3063 cpumask_copy(top_cpuset.cpus_allowed, &new_cpus);
3065 * Make sure that CPUs allocated to child partitions
3066 * do not show up in effective_cpus. If no CPU is left,
3067 * we clear the subparts_cpus & let the child partitions
3068 * fight for the CPUs again.
3070 if (top_cpuset.nr_subparts_cpus) {
3071 if (cpumask_subset(&new_cpus,
3072 top_cpuset.subparts_cpus)) {
3073 top_cpuset.nr_subparts_cpus = 0;
3074 cpumask_clear(top_cpuset.subparts_cpus);
3076 cpumask_andnot(&new_cpus, &new_cpus,
3077 top_cpuset.subparts_cpus);
3080 cpumask_copy(top_cpuset.effective_cpus, &new_cpus);
3081 spin_unlock_irq(&callback_lock);
3082 /* we don't mess with cpumasks of tasks in top_cpuset */
3085 /* synchronize mems_allowed to N_MEMORY */
3087 spin_lock_irq(&callback_lock);
3089 top_cpuset.mems_allowed = new_mems;
3090 top_cpuset.effective_mems = new_mems;
3091 spin_unlock_irq(&callback_lock);
3092 update_tasks_nodemask(&top_cpuset);
3095 mutex_unlock(&cpuset_mutex);
3097 /* if cpus or mems changed, we need to propagate to descendants */
3098 if (cpus_updated || mems_updated) {
3100 struct cgroup_subsys_state *pos_css;
3103 cpuset_for_each_descendant_pre(cs, pos_css, &top_cpuset) {
3104 if (cs == &top_cpuset || !css_tryget_online(&cs->css))
3108 cpuset_hotplug_update_tasks(cs, ptmp);
3116 /* rebuild sched domains if cpus_allowed has changed */
3117 if (cpus_updated || force_rebuild) {
3118 force_rebuild = false;
3119 rebuild_sched_domains();
3122 free_cpumasks(NULL, ptmp);
3125 void cpuset_update_active_cpus(void)
3128 * We're inside cpu hotplug critical region which usually nests
3129 * inside cgroup synchronization. Bounce actual hotplug processing
3130 * to a work item to avoid reverse locking order.
3132 schedule_work(&cpuset_hotplug_work);
3135 void cpuset_wait_for_hotplug(void)
3137 flush_work(&cpuset_hotplug_work);
3141 * Keep top_cpuset.mems_allowed tracking node_states[N_MEMORY].
3142 * Call this routine anytime after node_states[N_MEMORY] changes.
3143 * See cpuset_update_active_cpus() for CPU hotplug handling.
3145 static int cpuset_track_online_nodes(struct notifier_block *self,
3146 unsigned long action, void *arg)
3148 schedule_work(&cpuset_hotplug_work);
3152 static struct notifier_block cpuset_track_online_nodes_nb = {
3153 .notifier_call = cpuset_track_online_nodes,
3154 .priority = 10, /* ??! */
3158 * cpuset_init_smp - initialize cpus_allowed
3160 * Description: Finish top cpuset after cpu, node maps are initialized
3162 void __init cpuset_init_smp(void)
3164 cpumask_copy(top_cpuset.cpus_allowed, cpu_active_mask);
3165 top_cpuset.mems_allowed = node_states[N_MEMORY];
3166 top_cpuset.old_mems_allowed = top_cpuset.mems_allowed;
3168 cpumask_copy(top_cpuset.effective_cpus, cpu_active_mask);
3169 top_cpuset.effective_mems = node_states[N_MEMORY];
3171 register_hotmemory_notifier(&cpuset_track_online_nodes_nb);
3173 cpuset_migrate_mm_wq = alloc_ordered_workqueue("cpuset_migrate_mm", 0);
3174 BUG_ON(!cpuset_migrate_mm_wq);
3178 * cpuset_cpus_allowed - return cpus_allowed mask from a tasks cpuset.
3179 * @tsk: pointer to task_struct from which to obtain cpuset->cpus_allowed.
3180 * @pmask: pointer to struct cpumask variable to receive cpus_allowed set.
3182 * Description: Returns the cpumask_var_t cpus_allowed of the cpuset
3183 * attached to the specified @tsk. Guaranteed to return some non-empty
3184 * subset of cpu_online_mask, even if this means going outside the
3188 void cpuset_cpus_allowed(struct task_struct *tsk, struct cpumask *pmask)
3190 unsigned long flags;
3192 spin_lock_irqsave(&callback_lock, flags);
3194 guarantee_online_cpus(task_cs(tsk), pmask);
3196 spin_unlock_irqrestore(&callback_lock, flags);
3200 * cpuset_cpus_allowed_fallback - final fallback before complete catastrophe.
3201 * @tsk: pointer to task_struct with which the scheduler is struggling
3203 * Description: In the case that the scheduler cannot find an allowed cpu in
3204 * tsk->cpus_allowed, we fall back to task_cs(tsk)->cpus_allowed. In legacy
3205 * mode however, this value is the same as task_cs(tsk)->effective_cpus,
3206 * which will not contain a sane cpumask during cases such as cpu hotplugging.
3207 * This is the absolute last resort for the scheduler and it is only used if
3208 * _every_ other avenue has been traveled.
3211 void cpuset_cpus_allowed_fallback(struct task_struct *tsk)
3214 do_set_cpus_allowed(tsk, is_in_v2_mode() ?
3215 task_cs(tsk)->cpus_allowed : cpu_possible_mask);
3219 * We own tsk->cpus_allowed, nobody can change it under us.
3221 * But we used cs && cs->cpus_allowed lockless and thus can
3222 * race with cgroup_attach_task() or update_cpumask() and get
3223 * the wrong tsk->cpus_allowed. However, both cases imply the
3224 * subsequent cpuset_change_cpumask()->set_cpus_allowed_ptr()
3225 * which takes task_rq_lock().
3227 * If we are called after it dropped the lock we must see all
3228 * changes in tsk_cs()->cpus_allowed. Otherwise we can temporary
3229 * set any mask even if it is not right from task_cs() pov,
3230 * the pending set_cpus_allowed_ptr() will fix things.
3232 * select_fallback_rq() will fix things ups and set cpu_possible_mask
3237 void __init cpuset_init_current_mems_allowed(void)
3239 nodes_setall(current->mems_allowed);
3243 * cpuset_mems_allowed - return mems_allowed mask from a tasks cpuset.
3244 * @tsk: pointer to task_struct from which to obtain cpuset->mems_allowed.
3246 * Description: Returns the nodemask_t mems_allowed of the cpuset
3247 * attached to the specified @tsk. Guaranteed to return some non-empty
3248 * subset of node_states[N_MEMORY], even if this means going outside the
3252 nodemask_t cpuset_mems_allowed(struct task_struct *tsk)
3255 unsigned long flags;
3257 spin_lock_irqsave(&callback_lock, flags);
3259 guarantee_online_mems(task_cs(tsk), &mask);
3261 spin_unlock_irqrestore(&callback_lock, flags);
3267 * cpuset_nodemask_valid_mems_allowed - check nodemask vs. curremt mems_allowed
3268 * @nodemask: the nodemask to be checked
3270 * Are any of the nodes in the nodemask allowed in current->mems_allowed?
3272 int cpuset_nodemask_valid_mems_allowed(nodemask_t *nodemask)
3274 return nodes_intersects(*nodemask, current->mems_allowed);
3278 * nearest_hardwall_ancestor() - Returns the nearest mem_exclusive or
3279 * mem_hardwall ancestor to the specified cpuset. Call holding
3280 * callback_lock. If no ancestor is mem_exclusive or mem_hardwall
3281 * (an unusual configuration), then returns the root cpuset.
3283 static struct cpuset *nearest_hardwall_ancestor(struct cpuset *cs)
3285 while (!(is_mem_exclusive(cs) || is_mem_hardwall(cs)) && parent_cs(cs))
3291 * cpuset_node_allowed - Can we allocate on a memory node?
3292 * @node: is this an allowed node?
3293 * @gfp_mask: memory allocation flags
3295 * If we're in interrupt, yes, we can always allocate. If @node is set in
3296 * current's mems_allowed, yes. If it's not a __GFP_HARDWALL request and this
3297 * node is set in the nearest hardwalled cpuset ancestor to current's cpuset,
3298 * yes. If current has access to memory reserves as an oom victim, yes.
3301 * GFP_USER allocations are marked with the __GFP_HARDWALL bit,
3302 * and do not allow allocations outside the current tasks cpuset
3303 * unless the task has been OOM killed.
3304 * GFP_KERNEL allocations are not so marked, so can escape to the
3305 * nearest enclosing hardwalled ancestor cpuset.
3307 * Scanning up parent cpusets requires callback_lock. The
3308 * __alloc_pages() routine only calls here with __GFP_HARDWALL bit
3309 * _not_ set if it's a GFP_KERNEL allocation, and all nodes in the
3310 * current tasks mems_allowed came up empty on the first pass over
3311 * the zonelist. So only GFP_KERNEL allocations, if all nodes in the
3312 * cpuset are short of memory, might require taking the callback_lock.
3314 * The first call here from mm/page_alloc:get_page_from_freelist()
3315 * has __GFP_HARDWALL set in gfp_mask, enforcing hardwall cpusets,
3316 * so no allocation on a node outside the cpuset is allowed (unless
3317 * in interrupt, of course).
3319 * The second pass through get_page_from_freelist() doesn't even call
3320 * here for GFP_ATOMIC calls. For those calls, the __alloc_pages()
3321 * variable 'wait' is not set, and the bit ALLOC_CPUSET is not set
3322 * in alloc_flags. That logic and the checks below have the combined
3324 * in_interrupt - any node ok (current task context irrelevant)
3325 * GFP_ATOMIC - any node ok
3326 * tsk_is_oom_victim - any node ok
3327 * GFP_KERNEL - any node in enclosing hardwalled cpuset ok
3328 * GFP_USER - only nodes in current tasks mems allowed ok.
3330 bool __cpuset_node_allowed(int node, gfp_t gfp_mask)
3332 struct cpuset *cs; /* current cpuset ancestors */
3333 int allowed; /* is allocation in zone z allowed? */
3334 unsigned long flags;
3338 if (node_isset(node, current->mems_allowed))
3341 * Allow tasks that have access to memory reserves because they have
3342 * been OOM killed to get memory anywhere.
3344 if (unlikely(tsk_is_oom_victim(current)))
3346 if (gfp_mask & __GFP_HARDWALL) /* If hardwall request, stop here */
3349 if (current->flags & PF_EXITING) /* Let dying task have memory */
3352 /* Not hardwall and node outside mems_allowed: scan up cpusets */
3353 spin_lock_irqsave(&callback_lock, flags);
3356 cs = nearest_hardwall_ancestor(task_cs(current));
3357 allowed = node_isset(node, cs->mems_allowed);
3360 spin_unlock_irqrestore(&callback_lock, flags);
3365 * cpuset_mem_spread_node() - On which node to begin search for a file page
3366 * cpuset_slab_spread_node() - On which node to begin search for a slab page
3368 * If a task is marked PF_SPREAD_PAGE or PF_SPREAD_SLAB (as for
3369 * tasks in a cpuset with is_spread_page or is_spread_slab set),
3370 * and if the memory allocation used cpuset_mem_spread_node()
3371 * to determine on which node to start looking, as it will for
3372 * certain page cache or slab cache pages such as used for file
3373 * system buffers and inode caches, then instead of starting on the
3374 * local node to look for a free page, rather spread the starting
3375 * node around the tasks mems_allowed nodes.
3377 * We don't have to worry about the returned node being offline
3378 * because "it can't happen", and even if it did, it would be ok.
3380 * The routines calling guarantee_online_mems() are careful to
3381 * only set nodes in task->mems_allowed that are online. So it
3382 * should not be possible for the following code to return an
3383 * offline node. But if it did, that would be ok, as this routine
3384 * is not returning the node where the allocation must be, only
3385 * the node where the search should start. The zonelist passed to
3386 * __alloc_pages() will include all nodes. If the slab allocator
3387 * is passed an offline node, it will fall back to the local node.
3388 * See kmem_cache_alloc_node().
3391 static int cpuset_spread_node(int *rotor)
3393 return *rotor = next_node_in(*rotor, current->mems_allowed);
3396 int cpuset_mem_spread_node(void)
3398 if (current->cpuset_mem_spread_rotor == NUMA_NO_NODE)
3399 current->cpuset_mem_spread_rotor =
3400 node_random(¤t->mems_allowed);
3402 return cpuset_spread_node(¤t->cpuset_mem_spread_rotor);
3405 int cpuset_slab_spread_node(void)
3407 if (current->cpuset_slab_spread_rotor == NUMA_NO_NODE)
3408 current->cpuset_slab_spread_rotor =
3409 node_random(¤t->mems_allowed);
3411 return cpuset_spread_node(¤t->cpuset_slab_spread_rotor);
3414 EXPORT_SYMBOL_GPL(cpuset_mem_spread_node);
3417 * cpuset_mems_allowed_intersects - Does @tsk1's mems_allowed intersect @tsk2's?
3418 * @tsk1: pointer to task_struct of some task.
3419 * @tsk2: pointer to task_struct of some other task.
3421 * Description: Return true if @tsk1's mems_allowed intersects the
3422 * mems_allowed of @tsk2. Used by the OOM killer to determine if
3423 * one of the task's memory usage might impact the memory available
3427 int cpuset_mems_allowed_intersects(const struct task_struct *tsk1,
3428 const struct task_struct *tsk2)
3430 return nodes_intersects(tsk1->mems_allowed, tsk2->mems_allowed);
3434 * cpuset_print_current_mems_allowed - prints current's cpuset and mems_allowed
3436 * Description: Prints current's name, cpuset name, and cached copy of its
3437 * mems_allowed to the kernel log.
3439 void cpuset_print_current_mems_allowed(void)
3441 struct cgroup *cgrp;
3445 cgrp = task_cs(current)->css.cgroup;
3446 pr_cont(",cpuset=");
3447 pr_cont_cgroup_name(cgrp);
3448 pr_cont(",mems_allowed=%*pbl",
3449 nodemask_pr_args(¤t->mems_allowed));
3455 * Collection of memory_pressure is suppressed unless
3456 * this flag is enabled by writing "1" to the special
3457 * cpuset file 'memory_pressure_enabled' in the root cpuset.
3460 int cpuset_memory_pressure_enabled __read_mostly;
3463 * cpuset_memory_pressure_bump - keep stats of per-cpuset reclaims.
3465 * Keep a running average of the rate of synchronous (direct)
3466 * page reclaim efforts initiated by tasks in each cpuset.
3468 * This represents the rate at which some task in the cpuset
3469 * ran low on memory on all nodes it was allowed to use, and
3470 * had to enter the kernels page reclaim code in an effort to
3471 * create more free memory by tossing clean pages or swapping
3472 * or writing dirty pages.
3474 * Display to user space in the per-cpuset read-only file
3475 * "memory_pressure". Value displayed is an integer
3476 * representing the recent rate of entry into the synchronous
3477 * (direct) page reclaim by any task attached to the cpuset.
3480 void __cpuset_memory_pressure_bump(void)
3483 fmeter_markevent(&task_cs(current)->fmeter);
3487 #ifdef CONFIG_PROC_PID_CPUSET
3489 * proc_cpuset_show()
3490 * - Print tasks cpuset path into seq_file.
3491 * - Used for /proc/<pid>/cpuset.
3492 * - No need to task_lock(tsk) on this tsk->cpuset reference, as it
3493 * doesn't really matter if tsk->cpuset changes after we read it,
3494 * and we take cpuset_mutex, keeping cpuset_attach() from changing it
3497 int proc_cpuset_show(struct seq_file *m, struct pid_namespace *ns,
3498 struct pid *pid, struct task_struct *tsk)
3501 struct cgroup_subsys_state *css;
3505 buf = kmalloc(PATH_MAX, GFP_KERNEL);
3509 css = task_get_css(tsk, cpuset_cgrp_id);
3510 retval = cgroup_path_ns(css->cgroup, buf, PATH_MAX,
3511 current->nsproxy->cgroup_ns);
3513 if (retval >= PATH_MAX)
3514 retval = -ENAMETOOLONG;
3525 #endif /* CONFIG_PROC_PID_CPUSET */
3527 /* Display task mems_allowed in /proc/<pid>/status file. */
3528 void cpuset_task_status_allowed(struct seq_file *m, struct task_struct *task)
3530 seq_printf(m, "Mems_allowed:\t%*pb\n",
3531 nodemask_pr_args(&task->mems_allowed));
3532 seq_printf(m, "Mems_allowed_list:\t%*pbl\n",
3533 nodemask_pr_args(&task->mems_allowed));