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/namei.h>
43 #include <linux/pagemap.h>
44 #include <linux/proc_fs.h>
45 #include <linux/rcupdate.h>
46 #include <linux/sched.h>
47 #include <linux/sched/mm.h>
48 #include <linux/sched/task.h>
49 #include <linux/seq_file.h>
50 #include <linux/security.h>
51 #include <linux/slab.h>
52 #include <linux/spinlock.h>
53 #include <linux/stat.h>
54 #include <linux/string.h>
55 #include <linux/time.h>
56 #include <linux/time64.h>
57 #include <linux/backing-dev.h>
58 #include <linux/sort.h>
59 #include <linux/oom.h>
60 #include <linux/sched/isolation.h>
61 #include <linux/uaccess.h>
62 #include <linux/atomic.h>
63 #include <linux/mutex.h>
64 #include <linux/cgroup.h>
65 #include <linux/wait.h>
67 DEFINE_STATIC_KEY_FALSE(cpusets_pre_enable_key);
68 DEFINE_STATIC_KEY_FALSE(cpusets_enabled_key);
70 /* See "Frequency meter" comments, below. */
73 int cnt; /* unprocessed events count */
74 int val; /* most recent output value */
75 time64_t time; /* clock (secs) when val computed */
76 spinlock_t lock; /* guards read or write of above */
80 struct cgroup_subsys_state css;
82 unsigned long flags; /* "unsigned long" so bitops work */
85 * On default hierarchy:
87 * The user-configured masks can only be changed by writing to
88 * cpuset.cpus and cpuset.mems, and won't be limited by the
91 * The effective masks is the real masks that apply to the tasks
92 * in the cpuset. They may be changed if the configured masks are
93 * changed or hotplug happens.
95 * effective_mask == configured_mask & parent's effective_mask,
96 * and if it ends up empty, it will inherit the parent's mask.
101 * The user-configured masks are always the same with effective masks.
104 /* user-configured CPUs and Memory Nodes allow to tasks */
105 cpumask_var_t cpus_allowed;
106 nodemask_t mems_allowed;
108 /* effective CPUs and Memory Nodes allow to tasks */
109 cpumask_var_t effective_cpus;
110 nodemask_t effective_mems;
113 * CPUs allocated to child sub-partitions (default hierarchy only)
114 * - CPUs granted by the parent = effective_cpus U subparts_cpus
115 * - effective_cpus and subparts_cpus are mutually exclusive.
117 cpumask_var_t subparts_cpus;
120 * This is old Memory Nodes tasks took on.
122 * - top_cpuset.old_mems_allowed is initialized to mems_allowed.
123 * - A new cpuset's old_mems_allowed is initialized when some
124 * task is moved into it.
125 * - old_mems_allowed is used in cpuset_migrate_mm() when we change
126 * cpuset.mems_allowed and have tasks' nodemask updated, and
127 * then old_mems_allowed is updated to mems_allowed.
129 nodemask_t old_mems_allowed;
131 struct fmeter fmeter; /* memory_pressure filter */
134 * Tasks are being attached to this cpuset. Used to prevent
135 * zeroing cpus/mems_allowed between ->can_attach() and ->attach().
137 int attach_in_progress;
139 /* partition number for rebuild_sched_domains() */
142 /* for custom sched domain */
143 int relax_domain_level;
145 /* number of CPUs in subparts_cpus */
146 int nr_subparts_cpus;
148 /* partition root state */
149 int partition_root_state;
153 * Partition root states:
155 * 0 - not a partition root
158 #define PRS_DISABLED 0
159 #define PRS_ENABLED 1
162 * Temporary cpumasks for working with partitions that are passed among
163 * functions to avoid memory allocation in inner functions.
166 cpumask_var_t addmask, delmask; /* For partition root */
167 cpumask_var_t new_cpus; /* For update_cpumasks_hier() */
170 static inline struct cpuset *css_cs(struct cgroup_subsys_state *css)
172 return css ? container_of(css, struct cpuset, css) : NULL;
175 /* Retrieve the cpuset for a task */
176 static inline struct cpuset *task_cs(struct task_struct *task)
178 return css_cs(task_css(task, cpuset_cgrp_id));
181 static inline struct cpuset *parent_cs(struct cpuset *cs)
183 return css_cs(cs->css.parent);
187 static inline bool task_has_mempolicy(struct task_struct *task)
189 return task->mempolicy;
192 static inline bool task_has_mempolicy(struct task_struct *task)
199 /* bits in struct cpuset flags field */
206 CS_SCHED_LOAD_BALANCE,
211 /* convenient tests for these bits */
212 static inline bool is_cpuset_online(struct cpuset *cs)
214 return test_bit(CS_ONLINE, &cs->flags) && !css_is_dying(&cs->css);
217 static inline int is_cpu_exclusive(const struct cpuset *cs)
219 return test_bit(CS_CPU_EXCLUSIVE, &cs->flags);
222 static inline int is_mem_exclusive(const struct cpuset *cs)
224 return test_bit(CS_MEM_EXCLUSIVE, &cs->flags);
227 static inline int is_mem_hardwall(const struct cpuset *cs)
229 return test_bit(CS_MEM_HARDWALL, &cs->flags);
232 static inline int is_sched_load_balance(const struct cpuset *cs)
234 return test_bit(CS_SCHED_LOAD_BALANCE, &cs->flags);
237 static inline int is_memory_migrate(const struct cpuset *cs)
239 return test_bit(CS_MEMORY_MIGRATE, &cs->flags);
242 static inline int is_spread_page(const struct cpuset *cs)
244 return test_bit(CS_SPREAD_PAGE, &cs->flags);
247 static inline int is_spread_slab(const struct cpuset *cs)
249 return test_bit(CS_SPREAD_SLAB, &cs->flags);
252 static inline int is_partition_root(const struct cpuset *cs)
254 return cs->partition_root_state;
257 static struct cpuset top_cpuset = {
258 .flags = ((1 << CS_ONLINE) | (1 << CS_CPU_EXCLUSIVE) |
259 (1 << CS_MEM_EXCLUSIVE)),
260 .partition_root_state = PRS_ENABLED,
264 * cpuset_for_each_child - traverse online children of a cpuset
265 * @child_cs: loop cursor pointing to the current child
266 * @pos_css: used for iteration
267 * @parent_cs: target cpuset to walk children of
269 * Walk @child_cs through the online children of @parent_cs. Must be used
270 * with RCU read locked.
272 #define cpuset_for_each_child(child_cs, pos_css, parent_cs) \
273 css_for_each_child((pos_css), &(parent_cs)->css) \
274 if (is_cpuset_online(((child_cs) = css_cs((pos_css)))))
277 * cpuset_for_each_descendant_pre - pre-order walk of a cpuset's descendants
278 * @des_cs: loop cursor pointing to the current descendant
279 * @pos_css: used for iteration
280 * @root_cs: target cpuset to walk ancestor of
282 * Walk @des_cs through the online descendants of @root_cs. Must be used
283 * with RCU read locked. The caller may modify @pos_css by calling
284 * css_rightmost_descendant() to skip subtree. @root_cs is included in the
285 * iteration and the first node to be visited.
287 #define cpuset_for_each_descendant_pre(des_cs, pos_css, root_cs) \
288 css_for_each_descendant_pre((pos_css), &(root_cs)->css) \
289 if (is_cpuset_online(((des_cs) = css_cs((pos_css)))))
292 * There are two global locks guarding cpuset structures - cpuset_mutex and
293 * callback_lock. We also require taking task_lock() when dereferencing a
294 * task's cpuset pointer. See "The task_lock() exception", at the end of this
297 * A task must hold both locks to modify cpusets. If a task holds
298 * cpuset_mutex, then it blocks others wanting that mutex, ensuring that it
299 * is the only task able to also acquire callback_lock and be able to
300 * modify cpusets. It can perform various checks on the cpuset structure
301 * first, knowing nothing will change. It can also allocate memory while
302 * just holding cpuset_mutex. While it is performing these checks, various
303 * callback routines can briefly acquire callback_lock to query cpusets.
304 * Once it is ready to make the changes, it takes callback_lock, blocking
307 * Calls to the kernel memory allocator can not be made while holding
308 * callback_lock, as that would risk double tripping on callback_lock
309 * from one of the callbacks into the cpuset code from within
312 * If a task is only holding callback_lock, then it has read-only
315 * Now, the task_struct fields mems_allowed and mempolicy may be changed
316 * by other task, we use alloc_lock in the task_struct fields to protect
319 * The cpuset_common_file_read() handlers only hold callback_lock across
320 * small pieces of code, such as when reading out possibly multi-word
321 * cpumasks and nodemasks.
323 * Accessing a task's cpuset should be done in accordance with the
324 * guidelines for accessing subsystem state in kernel/cgroup.c
327 static DEFINE_MUTEX(cpuset_mutex);
328 static DEFINE_SPINLOCK(callback_lock);
330 static struct workqueue_struct *cpuset_migrate_mm_wq;
333 * CPU / memory hotplug is handled asynchronously.
335 static void cpuset_hotplug_workfn(struct work_struct *work);
336 static DECLARE_WORK(cpuset_hotplug_work, cpuset_hotplug_workfn);
338 static DECLARE_WAIT_QUEUE_HEAD(cpuset_attach_wq);
341 * Cgroup v2 behavior is used when on default hierarchy or the
342 * cgroup_v2_mode flag is set.
344 static inline bool is_in_v2_mode(void)
346 return cgroup_subsys_on_dfl(cpuset_cgrp_subsys) ||
347 (cpuset_cgrp_subsys.root->flags & CGRP_ROOT_CPUSET_V2_MODE);
351 * This is ugly, but preserves the userspace API for existing cpuset
352 * users. If someone tries to mount the "cpuset" filesystem, we
353 * silently switch it to mount "cgroup" instead
355 static struct dentry *cpuset_mount(struct file_system_type *fs_type,
356 int flags, const char *unused_dev_name, void *data)
358 struct file_system_type *cgroup_fs = get_fs_type("cgroup");
359 struct dentry *ret = ERR_PTR(-ENODEV);
363 "release_agent=/sbin/cpuset_release_agent";
364 ret = cgroup_fs->mount(cgroup_fs, flags,
365 unused_dev_name, mountopts);
366 put_filesystem(cgroup_fs);
371 static struct file_system_type cpuset_fs_type = {
373 .mount = cpuset_mount,
377 * Return in pmask the portion of a cpusets's cpus_allowed that
378 * are online. If none are online, walk up the cpuset hierarchy
379 * until we find one that does have some online cpus.
381 * One way or another, we guarantee to return some non-empty subset
382 * of cpu_online_mask.
384 * Call with callback_lock or cpuset_mutex held.
386 static void guarantee_online_cpus(struct cpuset *cs, struct cpumask *pmask)
388 while (!cpumask_intersects(cs->effective_cpus, cpu_online_mask)) {
392 * The top cpuset doesn't have any online cpu as a
393 * consequence of a race between cpuset_hotplug_work
394 * and cpu hotplug notifier. But we know the top
395 * cpuset's effective_cpus is on its way to to be
396 * identical to cpu_online_mask.
398 cpumask_copy(pmask, cpu_online_mask);
402 cpumask_and(pmask, cs->effective_cpus, cpu_online_mask);
406 * Return in *pmask the portion of a cpusets's mems_allowed that
407 * are online, with memory. If none are online with memory, walk
408 * up the cpuset hierarchy until we find one that does have some
409 * online mems. The top cpuset always has some mems online.
411 * One way or another, we guarantee to return some non-empty subset
412 * of node_states[N_MEMORY].
414 * Call with callback_lock or cpuset_mutex held.
416 static void guarantee_online_mems(struct cpuset *cs, nodemask_t *pmask)
418 while (!nodes_intersects(cs->effective_mems, node_states[N_MEMORY]))
420 nodes_and(*pmask, cs->effective_mems, node_states[N_MEMORY]);
424 * update task's spread flag if cpuset's page/slab spread flag is set
426 * Call with callback_lock or cpuset_mutex held.
428 static void cpuset_update_task_spread_flag(struct cpuset *cs,
429 struct task_struct *tsk)
431 if (is_spread_page(cs))
432 task_set_spread_page(tsk);
434 task_clear_spread_page(tsk);
436 if (is_spread_slab(cs))
437 task_set_spread_slab(tsk);
439 task_clear_spread_slab(tsk);
443 * is_cpuset_subset(p, q) - Is cpuset p a subset of cpuset q?
445 * One cpuset is a subset of another if all its allowed CPUs and
446 * Memory Nodes are a subset of the other, and its exclusive flags
447 * are only set if the other's are set. Call holding cpuset_mutex.
450 static int is_cpuset_subset(const struct cpuset *p, const struct cpuset *q)
452 return cpumask_subset(p->cpus_allowed, q->cpus_allowed) &&
453 nodes_subset(p->mems_allowed, q->mems_allowed) &&
454 is_cpu_exclusive(p) <= is_cpu_exclusive(q) &&
455 is_mem_exclusive(p) <= is_mem_exclusive(q);
459 * alloc_cpumasks - allocate three cpumasks for cpuset
460 * @cs: the cpuset that have cpumasks to be allocated.
461 * @tmp: the tmpmasks structure pointer
462 * Return: 0 if successful, -ENOMEM otherwise.
464 * Only one of the two input arguments should be non-NULL.
466 static inline int alloc_cpumasks(struct cpuset *cs, struct tmpmasks *tmp)
468 cpumask_var_t *pmask1, *pmask2, *pmask3;
471 pmask1 = &cs->cpus_allowed;
472 pmask2 = &cs->effective_cpus;
473 pmask3 = &cs->subparts_cpus;
475 pmask1 = &tmp->new_cpus;
476 pmask2 = &tmp->addmask;
477 pmask3 = &tmp->delmask;
480 if (!zalloc_cpumask_var(pmask1, GFP_KERNEL))
483 if (!zalloc_cpumask_var(pmask2, GFP_KERNEL))
486 if (!zalloc_cpumask_var(pmask3, GFP_KERNEL))
492 free_cpumask_var(*pmask2);
494 free_cpumask_var(*pmask1);
499 * free_cpumasks - free cpumasks in a tmpmasks structure
500 * @cs: the cpuset that have cpumasks to be free.
501 * @tmp: the tmpmasks structure pointer
503 static inline void free_cpumasks(struct cpuset *cs, struct tmpmasks *tmp)
506 free_cpumask_var(cs->cpus_allowed);
507 free_cpumask_var(cs->effective_cpus);
508 free_cpumask_var(cs->subparts_cpus);
511 free_cpumask_var(tmp->new_cpus);
512 free_cpumask_var(tmp->addmask);
513 free_cpumask_var(tmp->delmask);
518 * alloc_trial_cpuset - allocate a trial cpuset
519 * @cs: the cpuset that the trial cpuset duplicates
521 static struct cpuset *alloc_trial_cpuset(struct cpuset *cs)
523 struct cpuset *trial;
525 trial = kmemdup(cs, sizeof(*cs), GFP_KERNEL);
529 if (alloc_cpumasks(trial, NULL)) {
534 cpumask_copy(trial->cpus_allowed, cs->cpus_allowed);
535 cpumask_copy(trial->effective_cpus, cs->effective_cpus);
540 * free_cpuset - free the cpuset
541 * @cs: the cpuset to be freed
543 static inline void free_cpuset(struct cpuset *cs)
545 free_cpumasks(cs, NULL);
550 * validate_change() - Used to validate that any proposed cpuset change
551 * follows the structural rules for cpusets.
553 * If we replaced the flag and mask values of the current cpuset
554 * (cur) with those values in the trial cpuset (trial), would
555 * our various subset and exclusive rules still be valid? Presumes
558 * 'cur' is the address of an actual, in-use cpuset. Operations
559 * such as list traversal that depend on the actual address of the
560 * cpuset in the list must use cur below, not trial.
562 * 'trial' is the address of bulk structure copy of cur, with
563 * perhaps one or more of the fields cpus_allowed, mems_allowed,
564 * or flags changed to new, trial values.
566 * Return 0 if valid, -errno if not.
569 static int validate_change(struct cpuset *cur, struct cpuset *trial)
571 struct cgroup_subsys_state *css;
572 struct cpuset *c, *par;
577 /* Each of our child cpusets must be a subset of us */
579 cpuset_for_each_child(c, css, cur)
580 if (!is_cpuset_subset(c, trial))
583 /* Remaining checks don't apply to root cpuset */
585 if (cur == &top_cpuset)
588 par = parent_cs(cur);
590 /* On legacy hiearchy, we must be a subset of our parent cpuset. */
592 if (!is_in_v2_mode() && !is_cpuset_subset(trial, par))
596 * If either I or some sibling (!= me) is exclusive, we can't
600 cpuset_for_each_child(c, css, par) {
601 if ((is_cpu_exclusive(trial) || is_cpu_exclusive(c)) &&
603 cpumask_intersects(trial->cpus_allowed, c->cpus_allowed))
605 if ((is_mem_exclusive(trial) || is_mem_exclusive(c)) &&
607 nodes_intersects(trial->mems_allowed, c->mems_allowed))
612 * Cpusets with tasks - existing or newly being attached - can't
613 * be changed to have empty cpus_allowed or mems_allowed.
616 if ((cgroup_is_populated(cur->css.cgroup) || cur->attach_in_progress)) {
617 if (!cpumask_empty(cur->cpus_allowed) &&
618 cpumask_empty(trial->cpus_allowed))
620 if (!nodes_empty(cur->mems_allowed) &&
621 nodes_empty(trial->mems_allowed))
626 * We can't shrink if we won't have enough room for SCHED_DEADLINE
630 if (is_cpu_exclusive(cur) &&
631 !cpuset_cpumask_can_shrink(cur->cpus_allowed,
632 trial->cpus_allowed))
643 * Helper routine for generate_sched_domains().
644 * Do cpusets a, b have overlapping effective cpus_allowed masks?
646 static int cpusets_overlap(struct cpuset *a, struct cpuset *b)
648 return cpumask_intersects(a->effective_cpus, b->effective_cpus);
652 update_domain_attr(struct sched_domain_attr *dattr, struct cpuset *c)
654 if (dattr->relax_domain_level < c->relax_domain_level)
655 dattr->relax_domain_level = c->relax_domain_level;
659 static void update_domain_attr_tree(struct sched_domain_attr *dattr,
660 struct cpuset *root_cs)
663 struct cgroup_subsys_state *pos_css;
666 cpuset_for_each_descendant_pre(cp, pos_css, root_cs) {
667 /* skip the whole subtree if @cp doesn't have any CPU */
668 if (cpumask_empty(cp->cpus_allowed)) {
669 pos_css = css_rightmost_descendant(pos_css);
673 if (is_sched_load_balance(cp))
674 update_domain_attr(dattr, cp);
679 /* Must be called with cpuset_mutex held. */
680 static inline int nr_cpusets(void)
682 /* jump label reference count + the top-level cpuset */
683 return static_key_count(&cpusets_enabled_key.key) + 1;
687 * generate_sched_domains()
689 * This function builds a partial partition of the systems CPUs
690 * A 'partial partition' is a set of non-overlapping subsets whose
691 * union is a subset of that set.
692 * The output of this function needs to be passed to kernel/sched/core.c
693 * partition_sched_domains() routine, which will rebuild the scheduler's
694 * load balancing domains (sched domains) as specified by that partial
697 * See "What is sched_load_balance" in Documentation/cgroup-v1/cpusets.txt
698 * for a background explanation of this.
700 * Does not return errors, on the theory that the callers of this
701 * routine would rather not worry about failures to rebuild sched
702 * domains when operating in the severe memory shortage situations
703 * that could cause allocation failures below.
705 * Must be called with cpuset_mutex held.
707 * The three key local variables below are:
708 * q - a linked-list queue of cpuset pointers, used to implement a
709 * top-down scan of all cpusets. This scan loads a pointer
710 * to each cpuset marked is_sched_load_balance into the
711 * array 'csa'. For our purposes, rebuilding the schedulers
712 * sched domains, we can ignore !is_sched_load_balance cpusets.
713 * csa - (for CpuSet Array) Array of pointers to all the cpusets
714 * that need to be load balanced, for convenient iterative
715 * access by the subsequent code that finds the best partition,
716 * i.e the set of domains (subsets) of CPUs such that the
717 * cpus_allowed of every cpuset marked is_sched_load_balance
718 * is a subset of one of these domains, while there are as
719 * many such domains as possible, each as small as possible.
720 * doms - Conversion of 'csa' to an array of cpumasks, for passing to
721 * the kernel/sched/core.c routine partition_sched_domains() in a
722 * convenient format, that can be easily compared to the prior
723 * value to determine what partition elements (sched domains)
724 * were changed (added or removed.)
726 * Finding the best partition (set of domains):
727 * The triple nested loops below over i, j, k scan over the
728 * load balanced cpusets (using the array of cpuset pointers in
729 * csa[]) looking for pairs of cpusets that have overlapping
730 * cpus_allowed, but which don't have the same 'pn' partition
731 * number and gives them in the same partition number. It keeps
732 * looping on the 'restart' label until it can no longer find
735 * The union of the cpus_allowed masks from the set of
736 * all cpusets having the same 'pn' value then form the one
737 * element of the partition (one sched domain) to be passed to
738 * partition_sched_domains().
740 static int generate_sched_domains(cpumask_var_t **domains,
741 struct sched_domain_attr **attributes)
743 struct cpuset *cp; /* scans q */
744 struct cpuset **csa; /* array of all cpuset ptrs */
745 int csn; /* how many cpuset ptrs in csa so far */
746 int i, j, k; /* indices for partition finding loops */
747 cpumask_var_t *doms; /* resulting partition; i.e. sched domains */
748 struct sched_domain_attr *dattr; /* attributes for custom domains */
749 int ndoms = 0; /* number of sched domains in result */
750 int nslot; /* next empty doms[] struct cpumask slot */
751 struct cgroup_subsys_state *pos_css;
757 /* Special case for the 99% of systems with one, full, sched domain */
758 if (is_sched_load_balance(&top_cpuset)) {
760 doms = alloc_sched_domains(ndoms);
764 dattr = kmalloc(sizeof(struct sched_domain_attr), GFP_KERNEL);
766 *dattr = SD_ATTR_INIT;
767 update_domain_attr_tree(dattr, &top_cpuset);
769 cpumask_and(doms[0], top_cpuset.effective_cpus,
770 housekeeping_cpumask(HK_FLAG_DOMAIN));
775 csa = kmalloc_array(nr_cpusets(), sizeof(cp), GFP_KERNEL);
781 cpuset_for_each_descendant_pre(cp, pos_css, &top_cpuset) {
782 if (cp == &top_cpuset)
785 * Continue traversing beyond @cp iff @cp has some CPUs and
786 * isn't load balancing. The former is obvious. The
787 * latter: All child cpusets contain a subset of the
788 * parent's cpus, so just skip them, and then we call
789 * update_domain_attr_tree() to calc relax_domain_level of
790 * the corresponding sched domain.
792 if (!cpumask_empty(cp->cpus_allowed) &&
793 !(is_sched_load_balance(cp) &&
794 cpumask_intersects(cp->cpus_allowed,
795 housekeeping_cpumask(HK_FLAG_DOMAIN))))
798 if (is_sched_load_balance(cp))
801 /* skip @cp's subtree */
802 pos_css = css_rightmost_descendant(pos_css);
806 for (i = 0; i < csn; i++)
811 /* Find the best partition (set of sched domains) */
812 for (i = 0; i < csn; i++) {
813 struct cpuset *a = csa[i];
816 for (j = 0; j < csn; j++) {
817 struct cpuset *b = csa[j];
820 if (apn != bpn && cpusets_overlap(a, b)) {
821 for (k = 0; k < csn; k++) {
822 struct cpuset *c = csa[k];
827 ndoms--; /* one less element */
834 * Now we know how many domains to create.
835 * Convert <csn, csa> to <ndoms, doms> and populate cpu masks.
837 doms = alloc_sched_domains(ndoms);
842 * The rest of the code, including the scheduler, can deal with
843 * dattr==NULL case. No need to abort if alloc fails.
845 dattr = kmalloc_array(ndoms, sizeof(struct sched_domain_attr),
848 for (nslot = 0, i = 0; i < csn; i++) {
849 struct cpuset *a = csa[i];
854 /* Skip completed partitions */
860 if (nslot == ndoms) {
861 static int warnings = 10;
863 pr_warn("rebuild_sched_domains confused: nslot %d, ndoms %d, csn %d, i %d, apn %d\n",
864 nslot, ndoms, csn, i, apn);
872 *(dattr + nslot) = SD_ATTR_INIT;
873 for (j = i; j < csn; j++) {
874 struct cpuset *b = csa[j];
877 cpumask_or(dp, dp, b->effective_cpus);
878 cpumask_and(dp, dp, housekeeping_cpumask(HK_FLAG_DOMAIN));
880 update_domain_attr_tree(dattr + nslot, b);
882 /* Done with this partition */
888 BUG_ON(nslot != ndoms);
894 * Fallback to the default domain if kmalloc() failed.
895 * See comments in partition_sched_domains().
906 * Rebuild scheduler domains.
908 * If the flag 'sched_load_balance' of any cpuset with non-empty
909 * 'cpus' changes, or if the 'cpus' allowed changes in any cpuset
910 * which has that flag enabled, or if any cpuset with a non-empty
911 * 'cpus' is removed, then call this routine to rebuild the
912 * scheduler's dynamic sched domains.
914 * Call with cpuset_mutex held. Takes get_online_cpus().
916 static void rebuild_sched_domains_locked(void)
918 struct sched_domain_attr *attr;
922 lockdep_assert_held(&cpuset_mutex);
926 * We have raced with CPU hotplug. Don't do anything to avoid
927 * passing doms with offlined cpu to partition_sched_domains().
928 * Anyways, hotplug work item will rebuild sched domains.
930 if (!cpumask_equal(top_cpuset.effective_cpus, cpu_active_mask))
933 /* Generate domain masks and attrs */
934 ndoms = generate_sched_domains(&doms, &attr);
936 /* Have scheduler rebuild the domains */
937 partition_sched_domains(ndoms, doms, attr);
941 #else /* !CONFIG_SMP */
942 static void rebuild_sched_domains_locked(void)
945 #endif /* CONFIG_SMP */
947 void rebuild_sched_domains(void)
949 mutex_lock(&cpuset_mutex);
950 rebuild_sched_domains_locked();
951 mutex_unlock(&cpuset_mutex);
955 * update_tasks_cpumask - Update the cpumasks of tasks in the cpuset.
956 * @cs: the cpuset in which each task's cpus_allowed mask needs to be changed
958 * Iterate through each task of @cs updating its cpus_allowed to the
959 * effective cpuset's. As this function is called with cpuset_mutex held,
960 * cpuset membership stays stable.
962 static void update_tasks_cpumask(struct cpuset *cs)
964 struct css_task_iter it;
965 struct task_struct *task;
967 css_task_iter_start(&cs->css, 0, &it);
968 while ((task = css_task_iter_next(&it)))
969 set_cpus_allowed_ptr(task, cs->effective_cpus);
970 css_task_iter_end(&it);
974 * compute_effective_cpumask - Compute the effective cpumask of the cpuset
975 * @new_cpus: the temp variable for the new effective_cpus mask
976 * @cs: the cpuset the need to recompute the new effective_cpus mask
977 * @parent: the parent cpuset
979 * If the parent has subpartition CPUs, include them in the list of
980 * allowable CPUs in computing the new effective_cpus 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);
990 cpumask_and(new_cpus, cs->cpus_allowed, parent->effective_cpus);
995 * Commands for update_parent_subparts_cpumask
998 partcmd_enable, /* Enable partition root */
999 partcmd_disable, /* Disable partition root */
1000 partcmd_update, /* Update parent's subparts_cpus */
1004 * update_parent_subparts_cpumask - update subparts_cpus mask of parent cpuset
1005 * @cpuset: The cpuset that requests change in partition root state
1006 * @cmd: Partition root state change command
1007 * @newmask: Optional new cpumask for partcmd_update
1008 * @tmp: Temporary addmask and delmask
1009 * Return: 0, 1 or an error code
1011 * For partcmd_enable, the cpuset is being transformed from a non-partition
1012 * root to a partition root. The cpus_allowed mask of the given cpuset will
1013 * be put into parent's subparts_cpus and taken away from parent's
1014 * effective_cpus. The function will return 0 if all the CPUs listed in
1015 * cpus_allowed can be granted or an error code will be returned.
1017 * For partcmd_disable, the cpuset is being transofrmed from a partition
1018 * root back to a non-partition root. any CPUs in cpus_allowed that are in
1019 * parent's subparts_cpus will be taken away from that cpumask and put back
1020 * into parent's effective_cpus. 0 should always be returned.
1022 * For partcmd_update, if the optional newmask is specified, the cpu
1023 * list is to be changed from cpus_allowed to newmask. Otherwise,
1024 * cpus_allowed is assumed to remain the same. The function will return
1025 * 1 if changes to parent's subparts_cpus and effective_cpus happen or 0
1026 * otherwise. In case of error, an error code will be returned.
1028 * The partcmd_enable and partcmd_disable commands are used by
1029 * update_prstate(). The partcmd_update command is used by
1030 * update_cpumasks_hier() with newmask NULL and update_cpumask() with
1033 * The checking is more strict when enabling partition root than the
1034 * other two commands.
1036 * Because of the implicit cpu exclusive nature of a partition root,
1037 * cpumask changes that violates the cpu exclusivity rule will not be
1038 * permitted when checked by validate_change(). The validate_change()
1039 * function will also prevent any changes to the cpu list if it is not
1040 * a superset of children's cpu lists.
1042 static int update_parent_subparts_cpumask(struct cpuset *cpuset, int cmd,
1043 struct cpumask *newmask,
1044 struct tmpmasks *tmp)
1046 struct cpuset *parent = parent_cs(cpuset);
1047 int adding; /* Moving cpus from effective_cpus to subparts_cpus */
1048 int deleting; /* Moving cpus from subparts_cpus to effective_cpus */
1050 lockdep_assert_held(&cpuset_mutex);
1053 * The parent must be a partition root.
1054 * The new cpumask, if present, or the current cpus_allowed must
1057 if (!is_partition_root(parent) ||
1058 (newmask && cpumask_empty(newmask)) ||
1059 (!newmask && cpumask_empty(cpuset->cpus_allowed)))
1063 * Enabling/disabling partition root is not allowed if there are
1066 if ((cmd != partcmd_update) && css_has_online_children(&cpuset->css))
1070 * Enabling partition root is not allowed if not all the CPUs
1071 * can be granted from parent's effective_cpus or at least one
1072 * CPU will be left after that.
1074 if ((cmd == partcmd_enable) &&
1075 (!cpumask_subset(cpuset->cpus_allowed, parent->effective_cpus) ||
1076 cpumask_equal(cpuset->cpus_allowed, parent->effective_cpus)))
1080 * A cpumask update cannot make parent's effective_cpus become empty.
1082 adding = deleting = false;
1083 if (cmd == partcmd_enable) {
1084 cpumask_copy(tmp->addmask, cpuset->cpus_allowed);
1086 } else if (cmd == partcmd_disable) {
1087 deleting = cpumask_and(tmp->delmask, cpuset->cpus_allowed,
1088 parent->subparts_cpus);
1089 } else if (newmask) {
1091 * partcmd_update with newmask:
1093 * delmask = cpus_allowed & ~newmask & parent->subparts_cpus
1094 * addmask = newmask & parent->effective_cpus
1095 * & ~parent->subparts_cpus
1097 cpumask_andnot(tmp->delmask, cpuset->cpus_allowed, newmask);
1098 deleting = cpumask_and(tmp->delmask, tmp->delmask,
1099 parent->subparts_cpus);
1101 cpumask_and(tmp->addmask, newmask, parent->effective_cpus);
1102 adding = cpumask_andnot(tmp->addmask, tmp->addmask,
1103 parent->subparts_cpus);
1105 * Return error if the new effective_cpus could become empty.
1107 if (adding && !deleting &&
1108 cpumask_equal(parent->effective_cpus, tmp->addmask))
1112 * partcmd_update w/o newmask:
1114 * addmask = cpus_allowed & parent->effectiveb_cpus
1116 * Note that parent's subparts_cpus may have been
1117 * pre-shrunk in case the CPUs granted to the parent
1118 * by the grandparent changes. So no deletion is needed.
1120 adding = cpumask_and(tmp->addmask, cpuset->cpus_allowed,
1121 parent->effective_cpus);
1122 if (cpumask_equal(tmp->addmask, parent->effective_cpus))
1126 if (!adding && !deleting)
1130 * Change the parent's subparts_cpus.
1131 * Newly added CPUs will be removed from effective_cpus and
1132 * newly deleted ones will be added back to effective_cpus.
1134 spin_lock_irq(&callback_lock);
1136 cpumask_or(parent->subparts_cpus,
1137 parent->subparts_cpus, tmp->addmask);
1138 cpumask_andnot(parent->effective_cpus,
1139 parent->effective_cpus, tmp->addmask);
1142 cpumask_andnot(parent->subparts_cpus,
1143 parent->subparts_cpus, tmp->delmask);
1144 cpumask_or(parent->effective_cpus,
1145 parent->effective_cpus, tmp->delmask);
1148 parent->nr_subparts_cpus = cpumask_weight(parent->subparts_cpus);
1149 spin_unlock_irq(&callback_lock);
1151 return cmd == partcmd_update;
1155 * update_cpumasks_hier - Update effective cpumasks and tasks in the subtree
1156 * @cs: the cpuset to consider
1157 * @tmp: temp variables for calculating effective_cpus & partition setup
1159 * When congifured cpumask is changed, the effective cpumasks of this cpuset
1160 * and all its descendants need to be updated.
1162 * On legacy hierachy, effective_cpus will be the same with cpu_allowed.
1164 * Called with cpuset_mutex held
1166 static void update_cpumasks_hier(struct cpuset *cs, struct tmpmasks *tmp)
1169 struct cgroup_subsys_state *pos_css;
1170 bool need_rebuild_sched_domains = false;
1173 cpuset_for_each_descendant_pre(cp, pos_css, cs) {
1174 struct cpuset *parent = parent_cs(cp);
1177 compute_effective_cpumask(tmp->new_cpus, cp, parent);
1178 cs_empty = cpumask_empty(tmp->new_cpus);
1181 * A partition root cannot have empty effective_cpus
1183 WARN_ON_ONCE(cs_empty && is_partition_root(cp));
1186 * If it becomes empty, inherit the effective mask of the
1187 * parent, which is guaranteed to have some CPUs.
1189 if (is_in_v2_mode() && cs_empty)
1190 cpumask_copy(tmp->new_cpus, parent->effective_cpus);
1193 * Skip the whole subtree if the cpumask remains the same
1194 * and has no partition root state.
1196 if (!is_partition_root(cp) &&
1197 cpumask_equal(tmp->new_cpus, cp->effective_cpus)) {
1198 pos_css = css_rightmost_descendant(pos_css);
1203 * update_parent_subparts_cpumask() should have been called
1204 * for cs already in update_cpumask(). We should also call
1205 * update_tasks_cpumask() again for tasks in the parent
1206 * cpuset if the parent's subparts_cpus changes.
1208 if ((cp != cs) && cp->partition_root_state &&
1209 update_parent_subparts_cpumask(cp, partcmd_update,
1211 if (parent != &top_cpuset)
1212 update_tasks_cpumask(parent);
1215 if (!css_tryget_online(&cp->css))
1219 spin_lock_irq(&callback_lock);
1221 cpumask_copy(cp->effective_cpus, tmp->new_cpus);
1222 if (cp->nr_subparts_cpus) {
1224 * Make sure that effective_cpus & subparts_cpus
1225 * are mutually exclusive.
1227 cpumask_andnot(cp->effective_cpus, cp->effective_cpus,
1230 spin_unlock_irq(&callback_lock);
1232 WARN_ON(!is_in_v2_mode() &&
1233 !cpumask_equal(cp->cpus_allowed, cp->effective_cpus));
1235 update_tasks_cpumask(cp);
1238 * If the effective cpumask of any non-empty cpuset is changed,
1239 * we need to rebuild sched domains.
1241 if (!cpumask_empty(cp->cpus_allowed) &&
1242 is_sched_load_balance(cp))
1243 need_rebuild_sched_domains = true;
1250 if (need_rebuild_sched_domains)
1251 rebuild_sched_domains_locked();
1255 * update_cpumask - update the cpus_allowed mask of a cpuset and all tasks in it
1256 * @cs: the cpuset to consider
1257 * @trialcs: trial cpuset
1258 * @buf: buffer of cpu numbers written to this cpuset
1260 static int update_cpumask(struct cpuset *cs, struct cpuset *trialcs,
1264 struct tmpmasks tmp;
1266 /* top_cpuset.cpus_allowed tracks cpu_online_mask; it's read-only */
1267 if (cs == &top_cpuset)
1271 * An empty cpus_allowed is ok only if the cpuset has no tasks.
1272 * Since cpulist_parse() fails on an empty mask, we special case
1273 * that parsing. The validate_change() call ensures that cpusets
1274 * with tasks have cpus.
1277 cpumask_clear(trialcs->cpus_allowed);
1279 retval = cpulist_parse(buf, trialcs->cpus_allowed);
1283 if (!cpumask_subset(trialcs->cpus_allowed,
1284 top_cpuset.cpus_allowed))
1288 /* Nothing to do if the cpus didn't change */
1289 if (cpumask_equal(cs->cpus_allowed, trialcs->cpus_allowed))
1292 retval = validate_change(cs, trialcs);
1296 #ifdef CONFIG_CPUMASK_OFFSTACK
1298 * Use the cpumasks in trialcs for tmpmasks when they are pointers
1299 * to allocated cpumasks.
1301 tmp.addmask = trialcs->subparts_cpus;
1302 tmp.delmask = trialcs->effective_cpus;
1303 tmp.new_cpus = trialcs->cpus_allowed;
1306 if (cs->partition_root_state) {
1307 /* Cpumask of a partition root cannot be empty */
1308 if (cpumask_empty(trialcs->cpus_allowed))
1310 if (update_parent_subparts_cpumask(cs, partcmd_update,
1311 trialcs->cpus_allowed, &tmp) < 0)
1315 spin_lock_irq(&callback_lock);
1316 cpumask_copy(cs->cpus_allowed, trialcs->cpus_allowed);
1319 * Make sure that subparts_cpus is a subset of cpus_allowed.
1321 if (cs->nr_subparts_cpus) {
1322 cpumask_andnot(cs->subparts_cpus, cs->subparts_cpus,
1324 cs->nr_subparts_cpus = cpumask_weight(cs->subparts_cpus);
1326 spin_unlock_irq(&callback_lock);
1328 update_cpumasks_hier(cs, &tmp);
1333 * Migrate memory region from one set of nodes to another. This is
1334 * performed asynchronously as it can be called from process migration path
1335 * holding locks involved in process management. All mm migrations are
1336 * performed in the queued order and can be waited for by flushing
1337 * cpuset_migrate_mm_wq.
1340 struct cpuset_migrate_mm_work {
1341 struct work_struct work;
1342 struct mm_struct *mm;
1347 static void cpuset_migrate_mm_workfn(struct work_struct *work)
1349 struct cpuset_migrate_mm_work *mwork =
1350 container_of(work, struct cpuset_migrate_mm_work, work);
1352 /* on a wq worker, no need to worry about %current's mems_allowed */
1353 do_migrate_pages(mwork->mm, &mwork->from, &mwork->to, MPOL_MF_MOVE_ALL);
1358 static void cpuset_migrate_mm(struct mm_struct *mm, const nodemask_t *from,
1359 const nodemask_t *to)
1361 struct cpuset_migrate_mm_work *mwork;
1363 mwork = kzalloc(sizeof(*mwork), GFP_KERNEL);
1366 mwork->from = *from;
1368 INIT_WORK(&mwork->work, cpuset_migrate_mm_workfn);
1369 queue_work(cpuset_migrate_mm_wq, &mwork->work);
1375 static void cpuset_post_attach(void)
1377 flush_workqueue(cpuset_migrate_mm_wq);
1381 * cpuset_change_task_nodemask - change task's mems_allowed and mempolicy
1382 * @tsk: the task to change
1383 * @newmems: new nodes that the task will be set
1385 * We use the mems_allowed_seq seqlock to safely update both tsk->mems_allowed
1386 * and rebind an eventual tasks' mempolicy. If the task is allocating in
1387 * parallel, it might temporarily see an empty intersection, which results in
1388 * a seqlock check and retry before OOM or allocation failure.
1390 static void cpuset_change_task_nodemask(struct task_struct *tsk,
1391 nodemask_t *newmems)
1395 local_irq_disable();
1396 write_seqcount_begin(&tsk->mems_allowed_seq);
1398 nodes_or(tsk->mems_allowed, tsk->mems_allowed, *newmems);
1399 mpol_rebind_task(tsk, newmems);
1400 tsk->mems_allowed = *newmems;
1402 write_seqcount_end(&tsk->mems_allowed_seq);
1408 static void *cpuset_being_rebound;
1411 * update_tasks_nodemask - Update the nodemasks of tasks in the cpuset.
1412 * @cs: the cpuset in which each task's mems_allowed mask needs to be changed
1414 * Iterate through each task of @cs updating its mems_allowed to the
1415 * effective cpuset's. As this function is called with cpuset_mutex held,
1416 * cpuset membership stays stable.
1418 static void update_tasks_nodemask(struct cpuset *cs)
1420 static nodemask_t newmems; /* protected by cpuset_mutex */
1421 struct css_task_iter it;
1422 struct task_struct *task;
1424 cpuset_being_rebound = cs; /* causes mpol_dup() rebind */
1426 guarantee_online_mems(cs, &newmems);
1429 * The mpol_rebind_mm() call takes mmap_sem, which we couldn't
1430 * take while holding tasklist_lock. Forks can happen - the
1431 * mpol_dup() cpuset_being_rebound check will catch such forks,
1432 * and rebind their vma mempolicies too. Because we still hold
1433 * the global cpuset_mutex, we know that no other rebind effort
1434 * will be contending for the global variable cpuset_being_rebound.
1435 * It's ok if we rebind the same mm twice; mpol_rebind_mm()
1436 * is idempotent. Also migrate pages in each mm to new nodes.
1438 css_task_iter_start(&cs->css, 0, &it);
1439 while ((task = css_task_iter_next(&it))) {
1440 struct mm_struct *mm;
1443 cpuset_change_task_nodemask(task, &newmems);
1445 mm = get_task_mm(task);
1449 migrate = is_memory_migrate(cs);
1451 mpol_rebind_mm(mm, &cs->mems_allowed);
1453 cpuset_migrate_mm(mm, &cs->old_mems_allowed, &newmems);
1457 css_task_iter_end(&it);
1460 * All the tasks' nodemasks have been updated, update
1461 * cs->old_mems_allowed.
1463 cs->old_mems_allowed = newmems;
1465 /* We're done rebinding vmas to this cpuset's new mems_allowed. */
1466 cpuset_being_rebound = NULL;
1470 * update_nodemasks_hier - Update effective nodemasks and tasks in the subtree
1471 * @cs: the cpuset to consider
1472 * @new_mems: a temp variable for calculating new effective_mems
1474 * When configured nodemask is changed, the effective nodemasks of this cpuset
1475 * and all its descendants need to be updated.
1477 * On legacy hiearchy, effective_mems will be the same with mems_allowed.
1479 * Called with cpuset_mutex held
1481 static void update_nodemasks_hier(struct cpuset *cs, nodemask_t *new_mems)
1484 struct cgroup_subsys_state *pos_css;
1487 cpuset_for_each_descendant_pre(cp, pos_css, cs) {
1488 struct cpuset *parent = parent_cs(cp);
1490 nodes_and(*new_mems, cp->mems_allowed, parent->effective_mems);
1493 * If it becomes empty, inherit the effective mask of the
1494 * parent, which is guaranteed to have some MEMs.
1496 if (is_in_v2_mode() && nodes_empty(*new_mems))
1497 *new_mems = parent->effective_mems;
1499 /* Skip the whole subtree if the nodemask remains the same. */
1500 if (nodes_equal(*new_mems, cp->effective_mems)) {
1501 pos_css = css_rightmost_descendant(pos_css);
1505 if (!css_tryget_online(&cp->css))
1509 spin_lock_irq(&callback_lock);
1510 cp->effective_mems = *new_mems;
1511 spin_unlock_irq(&callback_lock);
1513 WARN_ON(!is_in_v2_mode() &&
1514 !nodes_equal(cp->mems_allowed, cp->effective_mems));
1516 update_tasks_nodemask(cp);
1525 * Handle user request to change the 'mems' memory placement
1526 * of a cpuset. Needs to validate the request, update the
1527 * cpusets mems_allowed, and for each task in the cpuset,
1528 * update mems_allowed and rebind task's mempolicy and any vma
1529 * mempolicies and if the cpuset is marked 'memory_migrate',
1530 * migrate the tasks pages to the new memory.
1532 * Call with cpuset_mutex held. May take callback_lock during call.
1533 * Will take tasklist_lock, scan tasklist for tasks in cpuset cs,
1534 * lock each such tasks mm->mmap_sem, scan its vma's and rebind
1535 * their mempolicies to the cpusets new mems_allowed.
1537 static int update_nodemask(struct cpuset *cs, struct cpuset *trialcs,
1543 * top_cpuset.mems_allowed tracks node_stats[N_MEMORY];
1546 if (cs == &top_cpuset) {
1552 * An empty mems_allowed is ok iff there are no tasks in the cpuset.
1553 * Since nodelist_parse() fails on an empty mask, we special case
1554 * that parsing. The validate_change() call ensures that cpusets
1555 * with tasks have memory.
1558 nodes_clear(trialcs->mems_allowed);
1560 retval = nodelist_parse(buf, trialcs->mems_allowed);
1564 if (!nodes_subset(trialcs->mems_allowed,
1565 top_cpuset.mems_allowed)) {
1571 if (nodes_equal(cs->mems_allowed, trialcs->mems_allowed)) {
1572 retval = 0; /* Too easy - nothing to do */
1575 retval = validate_change(cs, trialcs);
1579 spin_lock_irq(&callback_lock);
1580 cs->mems_allowed = trialcs->mems_allowed;
1581 spin_unlock_irq(&callback_lock);
1583 /* use trialcs->mems_allowed as a temp variable */
1584 update_nodemasks_hier(cs, &trialcs->mems_allowed);
1589 bool current_cpuset_is_being_rebound(void)
1594 ret = task_cs(current) == cpuset_being_rebound;
1600 static int update_relax_domain_level(struct cpuset *cs, s64 val)
1603 if (val < -1 || val >= sched_domain_level_max)
1607 if (val != cs->relax_domain_level) {
1608 cs->relax_domain_level = val;
1609 if (!cpumask_empty(cs->cpus_allowed) &&
1610 is_sched_load_balance(cs))
1611 rebuild_sched_domains_locked();
1618 * update_tasks_flags - update the spread flags of tasks in the cpuset.
1619 * @cs: the cpuset in which each task's spread flags needs to be changed
1621 * Iterate through each task of @cs updating its spread flags. As this
1622 * function is called with cpuset_mutex held, cpuset membership stays
1625 static void update_tasks_flags(struct cpuset *cs)
1627 struct css_task_iter it;
1628 struct task_struct *task;
1630 css_task_iter_start(&cs->css, 0, &it);
1631 while ((task = css_task_iter_next(&it)))
1632 cpuset_update_task_spread_flag(cs, task);
1633 css_task_iter_end(&it);
1637 * update_flag - read a 0 or a 1 in a file and update associated flag
1638 * bit: the bit to update (see cpuset_flagbits_t)
1639 * cs: the cpuset to update
1640 * turning_on: whether the flag is being set or cleared
1642 * Call with cpuset_mutex held.
1645 static int update_flag(cpuset_flagbits_t bit, struct cpuset *cs,
1648 struct cpuset *trialcs;
1649 int balance_flag_changed;
1650 int spread_flag_changed;
1653 trialcs = alloc_trial_cpuset(cs);
1658 set_bit(bit, &trialcs->flags);
1660 clear_bit(bit, &trialcs->flags);
1662 err = validate_change(cs, trialcs);
1666 balance_flag_changed = (is_sched_load_balance(cs) !=
1667 is_sched_load_balance(trialcs));
1669 spread_flag_changed = ((is_spread_slab(cs) != is_spread_slab(trialcs))
1670 || (is_spread_page(cs) != is_spread_page(trialcs)));
1672 spin_lock_irq(&callback_lock);
1673 cs->flags = trialcs->flags;
1674 spin_unlock_irq(&callback_lock);
1676 if (!cpumask_empty(trialcs->cpus_allowed) && balance_flag_changed)
1677 rebuild_sched_domains_locked();
1679 if (spread_flag_changed)
1680 update_tasks_flags(cs);
1682 free_cpuset(trialcs);
1687 * update_prstate - update partititon_root_state
1688 * cs: the cpuset to update
1689 * val: 0 - disabled, 1 - enabled
1691 * Call with cpuset_mutex held.
1693 static int update_prstate(struct cpuset *cs, int val)
1696 struct cpuset *parent = parent_cs(cs);
1697 struct tmpmasks tmp;
1699 if ((val != 0) && (val != 1))
1701 if (val == cs->partition_root_state)
1705 * Cannot force a partial or erroneous partition root to a full
1708 if (val && cs->partition_root_state)
1711 if (alloc_cpumasks(NULL, &tmp))
1715 if (!cs->partition_root_state) {
1717 * Turning on partition root requires setting the
1718 * CS_CPU_EXCLUSIVE bit implicitly as well and cpus_allowed
1721 if (cpumask_empty(cs->cpus_allowed))
1724 err = update_flag(CS_CPU_EXCLUSIVE, cs, 1);
1728 err = update_parent_subparts_cpumask(cs, partcmd_enable,
1731 update_flag(CS_CPU_EXCLUSIVE, cs, 0);
1734 cs->partition_root_state = PRS_ENABLED;
1736 err = update_parent_subparts_cpumask(cs, partcmd_disable,
1741 cs->partition_root_state = 0;
1743 /* Turning off CS_CPU_EXCLUSIVE will not return error */
1744 update_flag(CS_CPU_EXCLUSIVE, cs, 0);
1748 * Update cpumask of parent's tasks except when it is the top
1749 * cpuset as some system daemons cannot be mapped to other CPUs.
1751 if (parent != &top_cpuset)
1752 update_tasks_cpumask(parent);
1754 rebuild_sched_domains_locked();
1756 free_cpumasks(NULL, &tmp);
1761 * Frequency meter - How fast is some event occurring?
1763 * These routines manage a digitally filtered, constant time based,
1764 * event frequency meter. There are four routines:
1765 * fmeter_init() - initialize a frequency meter.
1766 * fmeter_markevent() - called each time the event happens.
1767 * fmeter_getrate() - returns the recent rate of such events.
1768 * fmeter_update() - internal routine used to update fmeter.
1770 * A common data structure is passed to each of these routines,
1771 * which is used to keep track of the state required to manage the
1772 * frequency meter and its digital filter.
1774 * The filter works on the number of events marked per unit time.
1775 * The filter is single-pole low-pass recursive (IIR). The time unit
1776 * is 1 second. Arithmetic is done using 32-bit integers scaled to
1777 * simulate 3 decimal digits of precision (multiplied by 1000).
1779 * With an FM_COEF of 933, and a time base of 1 second, the filter
1780 * has a half-life of 10 seconds, meaning that if the events quit
1781 * happening, then the rate returned from the fmeter_getrate()
1782 * will be cut in half each 10 seconds, until it converges to zero.
1784 * It is not worth doing a real infinitely recursive filter. If more
1785 * than FM_MAXTICKS ticks have elapsed since the last filter event,
1786 * just compute FM_MAXTICKS ticks worth, by which point the level
1789 * Limit the count of unprocessed events to FM_MAXCNT, so as to avoid
1790 * arithmetic overflow in the fmeter_update() routine.
1792 * Given the simple 32 bit integer arithmetic used, this meter works
1793 * best for reporting rates between one per millisecond (msec) and
1794 * one per 32 (approx) seconds. At constant rates faster than one
1795 * per msec it maxes out at values just under 1,000,000. At constant
1796 * rates between one per msec, and one per second it will stabilize
1797 * to a value N*1000, where N is the rate of events per second.
1798 * At constant rates between one per second and one per 32 seconds,
1799 * it will be choppy, moving up on the seconds that have an event,
1800 * and then decaying until the next event. At rates slower than
1801 * about one in 32 seconds, it decays all the way back to zero between
1805 #define FM_COEF 933 /* coefficient for half-life of 10 secs */
1806 #define FM_MAXTICKS ((u32)99) /* useless computing more ticks than this */
1807 #define FM_MAXCNT 1000000 /* limit cnt to avoid overflow */
1808 #define FM_SCALE 1000 /* faux fixed point scale */
1810 /* Initialize a frequency meter */
1811 static void fmeter_init(struct fmeter *fmp)
1816 spin_lock_init(&fmp->lock);
1819 /* Internal meter update - process cnt events and update value */
1820 static void fmeter_update(struct fmeter *fmp)
1825 now = ktime_get_seconds();
1826 ticks = now - fmp->time;
1831 ticks = min(FM_MAXTICKS, ticks);
1833 fmp->val = (FM_COEF * fmp->val) / FM_SCALE;
1836 fmp->val += ((FM_SCALE - FM_COEF) * fmp->cnt) / FM_SCALE;
1840 /* Process any previous ticks, then bump cnt by one (times scale). */
1841 static void fmeter_markevent(struct fmeter *fmp)
1843 spin_lock(&fmp->lock);
1845 fmp->cnt = min(FM_MAXCNT, fmp->cnt + FM_SCALE);
1846 spin_unlock(&fmp->lock);
1849 /* Process any previous ticks, then return current value. */
1850 static int fmeter_getrate(struct fmeter *fmp)
1854 spin_lock(&fmp->lock);
1857 spin_unlock(&fmp->lock);
1861 static struct cpuset *cpuset_attach_old_cs;
1863 /* Called by cgroups to determine if a cpuset is usable; cpuset_mutex held */
1864 static int cpuset_can_attach(struct cgroup_taskset *tset)
1866 struct cgroup_subsys_state *css;
1868 struct task_struct *task;
1871 /* used later by cpuset_attach() */
1872 cpuset_attach_old_cs = task_cs(cgroup_taskset_first(tset, &css));
1875 mutex_lock(&cpuset_mutex);
1877 /* allow moving tasks into an empty cpuset if on default hierarchy */
1879 if (!is_in_v2_mode() &&
1880 (cpumask_empty(cs->cpus_allowed) || nodes_empty(cs->mems_allowed)))
1883 cgroup_taskset_for_each(task, css, tset) {
1884 ret = task_can_attach(task, cs->cpus_allowed);
1887 ret = security_task_setscheduler(task);
1893 * Mark attach is in progress. This makes validate_change() fail
1894 * changes which zero cpus/mems_allowed.
1896 cs->attach_in_progress++;
1899 mutex_unlock(&cpuset_mutex);
1903 static void cpuset_cancel_attach(struct cgroup_taskset *tset)
1905 struct cgroup_subsys_state *css;
1908 cgroup_taskset_first(tset, &css);
1911 mutex_lock(&cpuset_mutex);
1912 css_cs(css)->attach_in_progress--;
1913 mutex_unlock(&cpuset_mutex);
1917 * Protected by cpuset_mutex. cpus_attach is used only by cpuset_attach()
1918 * but we can't allocate it dynamically there. Define it global and
1919 * allocate from cpuset_init().
1921 static cpumask_var_t cpus_attach;
1923 static void cpuset_attach(struct cgroup_taskset *tset)
1925 /* static buf protected by cpuset_mutex */
1926 static nodemask_t cpuset_attach_nodemask_to;
1927 struct task_struct *task;
1928 struct task_struct *leader;
1929 struct cgroup_subsys_state *css;
1931 struct cpuset *oldcs = cpuset_attach_old_cs;
1933 cgroup_taskset_first(tset, &css);
1936 mutex_lock(&cpuset_mutex);
1938 /* prepare for attach */
1939 if (cs == &top_cpuset)
1940 cpumask_copy(cpus_attach, cpu_possible_mask);
1942 guarantee_online_cpus(cs, cpus_attach);
1944 guarantee_online_mems(cs, &cpuset_attach_nodemask_to);
1946 cgroup_taskset_for_each(task, css, tset) {
1948 * can_attach beforehand should guarantee that this doesn't
1949 * fail. TODO: have a better way to handle failure here
1951 WARN_ON_ONCE(set_cpus_allowed_ptr(task, cpus_attach));
1953 cpuset_change_task_nodemask(task, &cpuset_attach_nodemask_to);
1954 cpuset_update_task_spread_flag(cs, task);
1958 * Change mm for all threadgroup leaders. This is expensive and may
1959 * sleep and should be moved outside migration path proper.
1961 cpuset_attach_nodemask_to = cs->effective_mems;
1962 cgroup_taskset_for_each_leader(leader, css, tset) {
1963 struct mm_struct *mm = get_task_mm(leader);
1966 mpol_rebind_mm(mm, &cpuset_attach_nodemask_to);
1969 * old_mems_allowed is the same with mems_allowed
1970 * here, except if this task is being moved
1971 * automatically due to hotplug. In that case
1972 * @mems_allowed has been updated and is empty, so
1973 * @old_mems_allowed is the right nodesets that we
1976 if (is_memory_migrate(cs))
1977 cpuset_migrate_mm(mm, &oldcs->old_mems_allowed,
1978 &cpuset_attach_nodemask_to);
1984 cs->old_mems_allowed = cpuset_attach_nodemask_to;
1986 cs->attach_in_progress--;
1987 if (!cs->attach_in_progress)
1988 wake_up(&cpuset_attach_wq);
1990 mutex_unlock(&cpuset_mutex);
1993 /* The various types of files and directories in a cpuset file system */
1996 FILE_MEMORY_MIGRATE,
1999 FILE_EFFECTIVE_CPULIST,
2000 FILE_EFFECTIVE_MEMLIST,
2004 FILE_SCHED_LOAD_BALANCE,
2005 FILE_PARTITION_ROOT,
2006 FILE_SCHED_RELAX_DOMAIN_LEVEL,
2007 FILE_MEMORY_PRESSURE_ENABLED,
2008 FILE_MEMORY_PRESSURE,
2011 } cpuset_filetype_t;
2013 static int cpuset_write_u64(struct cgroup_subsys_state *css, struct cftype *cft,
2016 struct cpuset *cs = css_cs(css);
2017 cpuset_filetype_t type = cft->private;
2020 mutex_lock(&cpuset_mutex);
2021 if (!is_cpuset_online(cs)) {
2027 case FILE_CPU_EXCLUSIVE:
2028 retval = update_flag(CS_CPU_EXCLUSIVE, cs, val);
2030 case FILE_MEM_EXCLUSIVE:
2031 retval = update_flag(CS_MEM_EXCLUSIVE, cs, val);
2033 case FILE_MEM_HARDWALL:
2034 retval = update_flag(CS_MEM_HARDWALL, cs, val);
2036 case FILE_SCHED_LOAD_BALANCE:
2037 retval = update_flag(CS_SCHED_LOAD_BALANCE, cs, val);
2039 case FILE_MEMORY_MIGRATE:
2040 retval = update_flag(CS_MEMORY_MIGRATE, cs, val);
2042 case FILE_MEMORY_PRESSURE_ENABLED:
2043 cpuset_memory_pressure_enabled = !!val;
2045 case FILE_SPREAD_PAGE:
2046 retval = update_flag(CS_SPREAD_PAGE, cs, val);
2048 case FILE_SPREAD_SLAB:
2049 retval = update_flag(CS_SPREAD_SLAB, cs, val);
2056 mutex_unlock(&cpuset_mutex);
2060 static int cpuset_write_s64(struct cgroup_subsys_state *css, struct cftype *cft,
2063 struct cpuset *cs = css_cs(css);
2064 cpuset_filetype_t type = cft->private;
2065 int retval = -ENODEV;
2067 mutex_lock(&cpuset_mutex);
2068 if (!is_cpuset_online(cs))
2072 case FILE_SCHED_RELAX_DOMAIN_LEVEL:
2073 retval = update_relax_domain_level(cs, val);
2075 case FILE_PARTITION_ROOT:
2076 retval = update_prstate(cs, val);
2083 mutex_unlock(&cpuset_mutex);
2088 * Common handling for a write to a "cpus" or "mems" file.
2090 static ssize_t cpuset_write_resmask(struct kernfs_open_file *of,
2091 char *buf, size_t nbytes, loff_t off)
2093 struct cpuset *cs = css_cs(of_css(of));
2094 struct cpuset *trialcs;
2095 int retval = -ENODEV;
2097 buf = strstrip(buf);
2100 * CPU or memory hotunplug may leave @cs w/o any execution
2101 * resources, in which case the hotplug code asynchronously updates
2102 * configuration and transfers all tasks to the nearest ancestor
2103 * which can execute.
2105 * As writes to "cpus" or "mems" may restore @cs's execution
2106 * resources, wait for the previously scheduled operations before
2107 * proceeding, so that we don't end up keep removing tasks added
2108 * after execution capability is restored.
2110 * cpuset_hotplug_work calls back into cgroup core via
2111 * cgroup_transfer_tasks() and waiting for it from a cgroupfs
2112 * operation like this one can lead to a deadlock through kernfs
2113 * active_ref protection. Let's break the protection. Losing the
2114 * protection is okay as we check whether @cs is online after
2115 * grabbing cpuset_mutex anyway. This only happens on the legacy
2119 kernfs_break_active_protection(of->kn);
2120 flush_work(&cpuset_hotplug_work);
2122 mutex_lock(&cpuset_mutex);
2123 if (!is_cpuset_online(cs))
2126 trialcs = alloc_trial_cpuset(cs);
2132 switch (of_cft(of)->private) {
2134 retval = update_cpumask(cs, trialcs, buf);
2137 retval = update_nodemask(cs, trialcs, buf);
2144 free_cpuset(trialcs);
2146 mutex_unlock(&cpuset_mutex);
2147 kernfs_unbreak_active_protection(of->kn);
2149 flush_workqueue(cpuset_migrate_mm_wq);
2150 return retval ?: nbytes;
2154 * These ascii lists should be read in a single call, by using a user
2155 * buffer large enough to hold the entire map. If read in smaller
2156 * chunks, there is no guarantee of atomicity. Since the display format
2157 * used, list of ranges of sequential numbers, is variable length,
2158 * and since these maps can change value dynamically, one could read
2159 * gibberish by doing partial reads while a list was changing.
2161 static int cpuset_common_seq_show(struct seq_file *sf, void *v)
2163 struct cpuset *cs = css_cs(seq_css(sf));
2164 cpuset_filetype_t type = seq_cft(sf)->private;
2167 spin_lock_irq(&callback_lock);
2171 seq_printf(sf, "%*pbl\n", cpumask_pr_args(cs->cpus_allowed));
2174 seq_printf(sf, "%*pbl\n", nodemask_pr_args(&cs->mems_allowed));
2176 case FILE_EFFECTIVE_CPULIST:
2177 seq_printf(sf, "%*pbl\n", cpumask_pr_args(cs->effective_cpus));
2179 case FILE_EFFECTIVE_MEMLIST:
2180 seq_printf(sf, "%*pbl\n", nodemask_pr_args(&cs->effective_mems));
2186 spin_unlock_irq(&callback_lock);
2190 static u64 cpuset_read_u64(struct cgroup_subsys_state *css, struct cftype *cft)
2192 struct cpuset *cs = css_cs(css);
2193 cpuset_filetype_t type = cft->private;
2195 case FILE_CPU_EXCLUSIVE:
2196 return is_cpu_exclusive(cs);
2197 case FILE_MEM_EXCLUSIVE:
2198 return is_mem_exclusive(cs);
2199 case FILE_MEM_HARDWALL:
2200 return is_mem_hardwall(cs);
2201 case FILE_SCHED_LOAD_BALANCE:
2202 return is_sched_load_balance(cs);
2203 case FILE_MEMORY_MIGRATE:
2204 return is_memory_migrate(cs);
2205 case FILE_MEMORY_PRESSURE_ENABLED:
2206 return cpuset_memory_pressure_enabled;
2207 case FILE_MEMORY_PRESSURE:
2208 return fmeter_getrate(&cs->fmeter);
2209 case FILE_SPREAD_PAGE:
2210 return is_spread_page(cs);
2211 case FILE_SPREAD_SLAB:
2212 return is_spread_slab(cs);
2217 /* Unreachable but makes gcc happy */
2221 static s64 cpuset_read_s64(struct cgroup_subsys_state *css, struct cftype *cft)
2223 struct cpuset *cs = css_cs(css);
2224 cpuset_filetype_t type = cft->private;
2226 case FILE_SCHED_RELAX_DOMAIN_LEVEL:
2227 return cs->relax_domain_level;
2228 case FILE_PARTITION_ROOT:
2229 return cs->partition_root_state;
2234 /* Unrechable but makes gcc happy */
2239 * for the common functions, 'private' gives the type of file
2242 static struct cftype legacy_files[] = {
2245 .seq_show = cpuset_common_seq_show,
2246 .write = cpuset_write_resmask,
2247 .max_write_len = (100U + 6 * NR_CPUS),
2248 .private = FILE_CPULIST,
2253 .seq_show = cpuset_common_seq_show,
2254 .write = cpuset_write_resmask,
2255 .max_write_len = (100U + 6 * MAX_NUMNODES),
2256 .private = FILE_MEMLIST,
2260 .name = "effective_cpus",
2261 .seq_show = cpuset_common_seq_show,
2262 .private = FILE_EFFECTIVE_CPULIST,
2266 .name = "effective_mems",
2267 .seq_show = cpuset_common_seq_show,
2268 .private = FILE_EFFECTIVE_MEMLIST,
2272 .name = "cpu_exclusive",
2273 .read_u64 = cpuset_read_u64,
2274 .write_u64 = cpuset_write_u64,
2275 .private = FILE_CPU_EXCLUSIVE,
2279 .name = "mem_exclusive",
2280 .read_u64 = cpuset_read_u64,
2281 .write_u64 = cpuset_write_u64,
2282 .private = FILE_MEM_EXCLUSIVE,
2286 .name = "mem_hardwall",
2287 .read_u64 = cpuset_read_u64,
2288 .write_u64 = cpuset_write_u64,
2289 .private = FILE_MEM_HARDWALL,
2293 .name = "sched_load_balance",
2294 .read_u64 = cpuset_read_u64,
2295 .write_u64 = cpuset_write_u64,
2296 .private = FILE_SCHED_LOAD_BALANCE,
2300 .name = "sched_relax_domain_level",
2301 .read_s64 = cpuset_read_s64,
2302 .write_s64 = cpuset_write_s64,
2303 .private = FILE_SCHED_RELAX_DOMAIN_LEVEL,
2307 .name = "memory_migrate",
2308 .read_u64 = cpuset_read_u64,
2309 .write_u64 = cpuset_write_u64,
2310 .private = FILE_MEMORY_MIGRATE,
2314 .name = "memory_pressure",
2315 .read_u64 = cpuset_read_u64,
2316 .private = FILE_MEMORY_PRESSURE,
2320 .name = "memory_spread_page",
2321 .read_u64 = cpuset_read_u64,
2322 .write_u64 = cpuset_write_u64,
2323 .private = FILE_SPREAD_PAGE,
2327 .name = "memory_spread_slab",
2328 .read_u64 = cpuset_read_u64,
2329 .write_u64 = cpuset_write_u64,
2330 .private = FILE_SPREAD_SLAB,
2334 .name = "memory_pressure_enabled",
2335 .flags = CFTYPE_ONLY_ON_ROOT,
2336 .read_u64 = cpuset_read_u64,
2337 .write_u64 = cpuset_write_u64,
2338 .private = FILE_MEMORY_PRESSURE_ENABLED,
2345 * This is currently a minimal set for the default hierarchy. It can be
2346 * expanded later on by migrating more features and control files from v1.
2348 static struct cftype dfl_files[] = {
2351 .seq_show = cpuset_common_seq_show,
2352 .write = cpuset_write_resmask,
2353 .max_write_len = (100U + 6 * NR_CPUS),
2354 .private = FILE_CPULIST,
2355 .flags = CFTYPE_NOT_ON_ROOT,
2360 .seq_show = cpuset_common_seq_show,
2361 .write = cpuset_write_resmask,
2362 .max_write_len = (100U + 6 * MAX_NUMNODES),
2363 .private = FILE_MEMLIST,
2364 .flags = CFTYPE_NOT_ON_ROOT,
2368 .name = "cpus.effective",
2369 .seq_show = cpuset_common_seq_show,
2370 .private = FILE_EFFECTIVE_CPULIST,
2371 .flags = CFTYPE_NOT_ON_ROOT,
2375 .name = "mems.effective",
2376 .seq_show = cpuset_common_seq_show,
2377 .private = FILE_EFFECTIVE_MEMLIST,
2378 .flags = CFTYPE_NOT_ON_ROOT,
2382 .name = "sched.partition",
2383 .read_s64 = cpuset_read_s64,
2384 .write_s64 = cpuset_write_s64,
2385 .private = FILE_PARTITION_ROOT,
2386 .flags = CFTYPE_NOT_ON_ROOT,
2394 * cpuset_css_alloc - allocate a cpuset css
2395 * cgrp: control group that the new cpuset will be part of
2398 static struct cgroup_subsys_state *
2399 cpuset_css_alloc(struct cgroup_subsys_state *parent_css)
2404 return &top_cpuset.css;
2406 cs = kzalloc(sizeof(*cs), GFP_KERNEL);
2408 return ERR_PTR(-ENOMEM);
2410 if (alloc_cpumasks(cs, NULL)) {
2412 return ERR_PTR(-ENOMEM);
2415 set_bit(CS_SCHED_LOAD_BALANCE, &cs->flags);
2416 nodes_clear(cs->mems_allowed);
2417 nodes_clear(cs->effective_mems);
2418 fmeter_init(&cs->fmeter);
2419 cs->relax_domain_level = -1;
2424 static int cpuset_css_online(struct cgroup_subsys_state *css)
2426 struct cpuset *cs = css_cs(css);
2427 struct cpuset *parent = parent_cs(cs);
2428 struct cpuset *tmp_cs;
2429 struct cgroup_subsys_state *pos_css;
2434 mutex_lock(&cpuset_mutex);
2436 set_bit(CS_ONLINE, &cs->flags);
2437 if (is_spread_page(parent))
2438 set_bit(CS_SPREAD_PAGE, &cs->flags);
2439 if (is_spread_slab(parent))
2440 set_bit(CS_SPREAD_SLAB, &cs->flags);
2444 spin_lock_irq(&callback_lock);
2445 if (is_in_v2_mode()) {
2446 cpumask_copy(cs->effective_cpus, parent->effective_cpus);
2447 cs->effective_mems = parent->effective_mems;
2449 spin_unlock_irq(&callback_lock);
2451 if (!test_bit(CGRP_CPUSET_CLONE_CHILDREN, &css->cgroup->flags))
2455 * Clone @parent's configuration if CGRP_CPUSET_CLONE_CHILDREN is
2456 * set. This flag handling is implemented in cgroup core for
2457 * histrical reasons - the flag may be specified during mount.
2459 * Currently, if any sibling cpusets have exclusive cpus or mem, we
2460 * refuse to clone the configuration - thereby refusing the task to
2461 * be entered, and as a result refusing the sys_unshare() or
2462 * clone() which initiated it. If this becomes a problem for some
2463 * users who wish to allow that scenario, then this could be
2464 * changed to grant parent->cpus_allowed-sibling_cpus_exclusive
2465 * (and likewise for mems) to the new cgroup.
2468 cpuset_for_each_child(tmp_cs, pos_css, parent) {
2469 if (is_mem_exclusive(tmp_cs) || is_cpu_exclusive(tmp_cs)) {
2476 spin_lock_irq(&callback_lock);
2477 cs->mems_allowed = parent->mems_allowed;
2478 cs->effective_mems = parent->mems_allowed;
2479 cpumask_copy(cs->cpus_allowed, parent->cpus_allowed);
2480 cpumask_copy(cs->effective_cpus, parent->cpus_allowed);
2481 spin_unlock_irq(&callback_lock);
2483 mutex_unlock(&cpuset_mutex);
2488 * If the cpuset being removed has its flag 'sched_load_balance'
2489 * enabled, then simulate turning sched_load_balance off, which
2490 * will call rebuild_sched_domains_locked(). That is not needed
2491 * in the default hierarchy where only changes in partition
2492 * will cause repartitioning.
2494 * If the cpuset has the 'sched.partition' flag enabled, simulate
2495 * turning 'sched.partition" off.
2498 static void cpuset_css_offline(struct cgroup_subsys_state *css)
2500 struct cpuset *cs = css_cs(css);
2502 mutex_lock(&cpuset_mutex);
2504 if (is_partition_root(cs))
2505 update_prstate(cs, 0);
2507 if (!cgroup_subsys_on_dfl(cpuset_cgrp_subsys) &&
2508 is_sched_load_balance(cs))
2509 update_flag(CS_SCHED_LOAD_BALANCE, cs, 0);
2512 clear_bit(CS_ONLINE, &cs->flags);
2514 mutex_unlock(&cpuset_mutex);
2517 static void cpuset_css_free(struct cgroup_subsys_state *css)
2519 struct cpuset *cs = css_cs(css);
2524 static void cpuset_bind(struct cgroup_subsys_state *root_css)
2526 mutex_lock(&cpuset_mutex);
2527 spin_lock_irq(&callback_lock);
2529 if (is_in_v2_mode()) {
2530 cpumask_copy(top_cpuset.cpus_allowed, cpu_possible_mask);
2531 top_cpuset.mems_allowed = node_possible_map;
2533 cpumask_copy(top_cpuset.cpus_allowed,
2534 top_cpuset.effective_cpus);
2535 top_cpuset.mems_allowed = top_cpuset.effective_mems;
2538 spin_unlock_irq(&callback_lock);
2539 mutex_unlock(&cpuset_mutex);
2543 * Make sure the new task conform to the current state of its parent,
2544 * which could have been changed by cpuset just after it inherits the
2545 * state from the parent and before it sits on the cgroup's task list.
2547 static void cpuset_fork(struct task_struct *task)
2549 if (task_css_is_root(task, cpuset_cgrp_id))
2552 set_cpus_allowed_ptr(task, ¤t->cpus_allowed);
2553 task->mems_allowed = current->mems_allowed;
2556 struct cgroup_subsys cpuset_cgrp_subsys = {
2557 .css_alloc = cpuset_css_alloc,
2558 .css_online = cpuset_css_online,
2559 .css_offline = cpuset_css_offline,
2560 .css_free = cpuset_css_free,
2561 .can_attach = cpuset_can_attach,
2562 .cancel_attach = cpuset_cancel_attach,
2563 .attach = cpuset_attach,
2564 .post_attach = cpuset_post_attach,
2565 .bind = cpuset_bind,
2566 .fork = cpuset_fork,
2567 .legacy_cftypes = legacy_files,
2568 .dfl_cftypes = dfl_files,
2574 * cpuset_init - initialize cpusets at system boot
2576 * Description: Initialize top_cpuset and the cpuset internal file system,
2579 int __init cpuset_init(void)
2583 BUG_ON(!alloc_cpumask_var(&top_cpuset.cpus_allowed, GFP_KERNEL));
2584 BUG_ON(!alloc_cpumask_var(&top_cpuset.effective_cpus, GFP_KERNEL));
2585 BUG_ON(!zalloc_cpumask_var(&top_cpuset.subparts_cpus, GFP_KERNEL));
2587 cpumask_setall(top_cpuset.cpus_allowed);
2588 nodes_setall(top_cpuset.mems_allowed);
2589 cpumask_setall(top_cpuset.effective_cpus);
2590 nodes_setall(top_cpuset.effective_mems);
2592 fmeter_init(&top_cpuset.fmeter);
2593 set_bit(CS_SCHED_LOAD_BALANCE, &top_cpuset.flags);
2594 top_cpuset.relax_domain_level = -1;
2596 err = register_filesystem(&cpuset_fs_type);
2600 BUG_ON(!alloc_cpumask_var(&cpus_attach, GFP_KERNEL));
2606 * If CPU and/or memory hotplug handlers, below, unplug any CPUs
2607 * or memory nodes, we need to walk over the cpuset hierarchy,
2608 * removing that CPU or node from all cpusets. If this removes the
2609 * last CPU or node from a cpuset, then move the tasks in the empty
2610 * cpuset to its next-highest non-empty parent.
2612 static void remove_tasks_in_empty_cpuset(struct cpuset *cs)
2614 struct cpuset *parent;
2617 * Find its next-highest non-empty parent, (top cpuset
2618 * has online cpus, so can't be empty).
2620 parent = parent_cs(cs);
2621 while (cpumask_empty(parent->cpus_allowed) ||
2622 nodes_empty(parent->mems_allowed))
2623 parent = parent_cs(parent);
2625 if (cgroup_transfer_tasks(parent->css.cgroup, cs->css.cgroup)) {
2626 pr_err("cpuset: failed to transfer tasks out of empty cpuset ");
2627 pr_cont_cgroup_name(cs->css.cgroup);
2633 hotplug_update_tasks_legacy(struct cpuset *cs,
2634 struct cpumask *new_cpus, nodemask_t *new_mems,
2635 bool cpus_updated, bool mems_updated)
2639 spin_lock_irq(&callback_lock);
2640 cpumask_copy(cs->cpus_allowed, new_cpus);
2641 cpumask_copy(cs->effective_cpus, new_cpus);
2642 cs->mems_allowed = *new_mems;
2643 cs->effective_mems = *new_mems;
2644 spin_unlock_irq(&callback_lock);
2647 * Don't call update_tasks_cpumask() if the cpuset becomes empty,
2648 * as the tasks will be migratecd to an ancestor.
2650 if (cpus_updated && !cpumask_empty(cs->cpus_allowed))
2651 update_tasks_cpumask(cs);
2652 if (mems_updated && !nodes_empty(cs->mems_allowed))
2653 update_tasks_nodemask(cs);
2655 is_empty = cpumask_empty(cs->cpus_allowed) ||
2656 nodes_empty(cs->mems_allowed);
2658 mutex_unlock(&cpuset_mutex);
2661 * Move tasks to the nearest ancestor with execution resources,
2662 * This is full cgroup operation which will also call back into
2663 * cpuset. Should be done outside any lock.
2666 remove_tasks_in_empty_cpuset(cs);
2668 mutex_lock(&cpuset_mutex);
2672 hotplug_update_tasks(struct cpuset *cs,
2673 struct cpumask *new_cpus, nodemask_t *new_mems,
2674 bool cpus_updated, bool mems_updated)
2676 if (cpumask_empty(new_cpus))
2677 cpumask_copy(new_cpus, parent_cs(cs)->effective_cpus);
2678 if (nodes_empty(*new_mems))
2679 *new_mems = parent_cs(cs)->effective_mems;
2681 spin_lock_irq(&callback_lock);
2682 cpumask_copy(cs->effective_cpus, new_cpus);
2683 cs->effective_mems = *new_mems;
2684 spin_unlock_irq(&callback_lock);
2687 update_tasks_cpumask(cs);
2689 update_tasks_nodemask(cs);
2693 * cpuset_hotplug_update_tasks - update tasks in a cpuset for hotunplug
2694 * @cs: cpuset in interest
2696 * Compare @cs's cpu and mem masks against top_cpuset and if some have gone
2697 * offline, update @cs accordingly. If @cs ends up with no CPU or memory,
2698 * all its tasks are moved to the nearest ancestor with both resources.
2700 static void cpuset_hotplug_update_tasks(struct cpuset *cs)
2702 static cpumask_t new_cpus;
2703 static nodemask_t new_mems;
2707 wait_event(cpuset_attach_wq, cs->attach_in_progress == 0);
2709 mutex_lock(&cpuset_mutex);
2712 * We have raced with task attaching. We wait until attaching
2713 * is finished, so we won't attach a task to an empty cpuset.
2715 if (cs->attach_in_progress) {
2716 mutex_unlock(&cpuset_mutex);
2720 cpumask_and(&new_cpus, cs->cpus_allowed, parent_cs(cs)->effective_cpus);
2721 nodes_and(new_mems, cs->mems_allowed, parent_cs(cs)->effective_mems);
2723 cpus_updated = !cpumask_equal(&new_cpus, cs->effective_cpus);
2724 mems_updated = !nodes_equal(new_mems, cs->effective_mems);
2726 if (is_in_v2_mode())
2727 hotplug_update_tasks(cs, &new_cpus, &new_mems,
2728 cpus_updated, mems_updated);
2730 hotplug_update_tasks_legacy(cs, &new_cpus, &new_mems,
2731 cpus_updated, mems_updated);
2733 mutex_unlock(&cpuset_mutex);
2736 static bool force_rebuild;
2738 void cpuset_force_rebuild(void)
2740 force_rebuild = true;
2744 * cpuset_hotplug_workfn - handle CPU/memory hotunplug for a cpuset
2746 * This function is called after either CPU or memory configuration has
2747 * changed and updates cpuset accordingly. The top_cpuset is always
2748 * synchronized to cpu_active_mask and N_MEMORY, which is necessary in
2749 * order to make cpusets transparent (of no affect) on systems that are
2750 * actively using CPU hotplug but making no active use of cpusets.
2752 * Non-root cpusets are only affected by offlining. If any CPUs or memory
2753 * nodes have been taken down, cpuset_hotplug_update_tasks() is invoked on
2756 * Note that CPU offlining during suspend is ignored. We don't modify
2757 * cpusets across suspend/resume cycles at all.
2759 static void cpuset_hotplug_workfn(struct work_struct *work)
2761 static cpumask_t new_cpus;
2762 static nodemask_t new_mems;
2763 bool cpus_updated, mems_updated;
2764 bool on_dfl = is_in_v2_mode();
2766 mutex_lock(&cpuset_mutex);
2768 /* fetch the available cpus/mems and find out which changed how */
2769 cpumask_copy(&new_cpus, cpu_active_mask);
2770 new_mems = node_states[N_MEMORY];
2772 cpus_updated = !cpumask_equal(top_cpuset.effective_cpus, &new_cpus);
2773 mems_updated = !nodes_equal(top_cpuset.effective_mems, new_mems);
2775 /* synchronize cpus_allowed to cpu_active_mask */
2777 spin_lock_irq(&callback_lock);
2779 cpumask_copy(top_cpuset.cpus_allowed, &new_cpus);
2780 cpumask_copy(top_cpuset.effective_cpus, &new_cpus);
2781 spin_unlock_irq(&callback_lock);
2782 /* we don't mess with cpumasks of tasks in top_cpuset */
2785 /* synchronize mems_allowed to N_MEMORY */
2787 spin_lock_irq(&callback_lock);
2789 top_cpuset.mems_allowed = new_mems;
2790 top_cpuset.effective_mems = new_mems;
2791 spin_unlock_irq(&callback_lock);
2792 update_tasks_nodemask(&top_cpuset);
2795 mutex_unlock(&cpuset_mutex);
2797 /* if cpus or mems changed, we need to propagate to descendants */
2798 if (cpus_updated || mems_updated) {
2800 struct cgroup_subsys_state *pos_css;
2803 cpuset_for_each_descendant_pre(cs, pos_css, &top_cpuset) {
2804 if (cs == &top_cpuset || !css_tryget_online(&cs->css))
2808 cpuset_hotplug_update_tasks(cs);
2816 /* rebuild sched domains if cpus_allowed has changed */
2817 if (cpus_updated || force_rebuild) {
2818 force_rebuild = false;
2819 rebuild_sched_domains();
2823 void cpuset_update_active_cpus(void)
2826 * We're inside cpu hotplug critical region which usually nests
2827 * inside cgroup synchronization. Bounce actual hotplug processing
2828 * to a work item to avoid reverse locking order.
2830 schedule_work(&cpuset_hotplug_work);
2833 void cpuset_wait_for_hotplug(void)
2835 flush_work(&cpuset_hotplug_work);
2839 * Keep top_cpuset.mems_allowed tracking node_states[N_MEMORY].
2840 * Call this routine anytime after node_states[N_MEMORY] changes.
2841 * See cpuset_update_active_cpus() for CPU hotplug handling.
2843 static int cpuset_track_online_nodes(struct notifier_block *self,
2844 unsigned long action, void *arg)
2846 schedule_work(&cpuset_hotplug_work);
2850 static struct notifier_block cpuset_track_online_nodes_nb = {
2851 .notifier_call = cpuset_track_online_nodes,
2852 .priority = 10, /* ??! */
2856 * cpuset_init_smp - initialize cpus_allowed
2858 * Description: Finish top cpuset after cpu, node maps are initialized
2860 void __init cpuset_init_smp(void)
2862 cpumask_copy(top_cpuset.cpus_allowed, cpu_active_mask);
2863 top_cpuset.mems_allowed = node_states[N_MEMORY];
2864 top_cpuset.old_mems_allowed = top_cpuset.mems_allowed;
2866 cpumask_copy(top_cpuset.effective_cpus, cpu_active_mask);
2867 top_cpuset.effective_mems = node_states[N_MEMORY];
2869 register_hotmemory_notifier(&cpuset_track_online_nodes_nb);
2871 cpuset_migrate_mm_wq = alloc_ordered_workqueue("cpuset_migrate_mm", 0);
2872 BUG_ON(!cpuset_migrate_mm_wq);
2876 * cpuset_cpus_allowed - return cpus_allowed mask from a tasks cpuset.
2877 * @tsk: pointer to task_struct from which to obtain cpuset->cpus_allowed.
2878 * @pmask: pointer to struct cpumask variable to receive cpus_allowed set.
2880 * Description: Returns the cpumask_var_t cpus_allowed of the cpuset
2881 * attached to the specified @tsk. Guaranteed to return some non-empty
2882 * subset of cpu_online_mask, even if this means going outside the
2886 void cpuset_cpus_allowed(struct task_struct *tsk, struct cpumask *pmask)
2888 unsigned long flags;
2890 spin_lock_irqsave(&callback_lock, flags);
2892 guarantee_online_cpus(task_cs(tsk), pmask);
2894 spin_unlock_irqrestore(&callback_lock, flags);
2897 void cpuset_cpus_allowed_fallback(struct task_struct *tsk)
2900 do_set_cpus_allowed(tsk, task_cs(tsk)->effective_cpus);
2904 * We own tsk->cpus_allowed, nobody can change it under us.
2906 * But we used cs && cs->cpus_allowed lockless and thus can
2907 * race with cgroup_attach_task() or update_cpumask() and get
2908 * the wrong tsk->cpus_allowed. However, both cases imply the
2909 * subsequent cpuset_change_cpumask()->set_cpus_allowed_ptr()
2910 * which takes task_rq_lock().
2912 * If we are called after it dropped the lock we must see all
2913 * changes in tsk_cs()->cpus_allowed. Otherwise we can temporary
2914 * set any mask even if it is not right from task_cs() pov,
2915 * the pending set_cpus_allowed_ptr() will fix things.
2917 * select_fallback_rq() will fix things ups and set cpu_possible_mask
2922 void __init cpuset_init_current_mems_allowed(void)
2924 nodes_setall(current->mems_allowed);
2928 * cpuset_mems_allowed - return mems_allowed mask from a tasks cpuset.
2929 * @tsk: pointer to task_struct from which to obtain cpuset->mems_allowed.
2931 * Description: Returns the nodemask_t mems_allowed of the cpuset
2932 * attached to the specified @tsk. Guaranteed to return some non-empty
2933 * subset of node_states[N_MEMORY], even if this means going outside the
2937 nodemask_t cpuset_mems_allowed(struct task_struct *tsk)
2940 unsigned long flags;
2942 spin_lock_irqsave(&callback_lock, flags);
2944 guarantee_online_mems(task_cs(tsk), &mask);
2946 spin_unlock_irqrestore(&callback_lock, flags);
2952 * cpuset_nodemask_valid_mems_allowed - check nodemask vs. curremt mems_allowed
2953 * @nodemask: the nodemask to be checked
2955 * Are any of the nodes in the nodemask allowed in current->mems_allowed?
2957 int cpuset_nodemask_valid_mems_allowed(nodemask_t *nodemask)
2959 return nodes_intersects(*nodemask, current->mems_allowed);
2963 * nearest_hardwall_ancestor() - Returns the nearest mem_exclusive or
2964 * mem_hardwall ancestor to the specified cpuset. Call holding
2965 * callback_lock. If no ancestor is mem_exclusive or mem_hardwall
2966 * (an unusual configuration), then returns the root cpuset.
2968 static struct cpuset *nearest_hardwall_ancestor(struct cpuset *cs)
2970 while (!(is_mem_exclusive(cs) || is_mem_hardwall(cs)) && parent_cs(cs))
2976 * cpuset_node_allowed - Can we allocate on a memory node?
2977 * @node: is this an allowed node?
2978 * @gfp_mask: memory allocation flags
2980 * If we're in interrupt, yes, we can always allocate. If @node is set in
2981 * current's mems_allowed, yes. If it's not a __GFP_HARDWALL request and this
2982 * node is set in the nearest hardwalled cpuset ancestor to current's cpuset,
2983 * yes. If current has access to memory reserves as an oom victim, yes.
2986 * GFP_USER allocations are marked with the __GFP_HARDWALL bit,
2987 * and do not allow allocations outside the current tasks cpuset
2988 * unless the task has been OOM killed.
2989 * GFP_KERNEL allocations are not so marked, so can escape to the
2990 * nearest enclosing hardwalled ancestor cpuset.
2992 * Scanning up parent cpusets requires callback_lock. The
2993 * __alloc_pages() routine only calls here with __GFP_HARDWALL bit
2994 * _not_ set if it's a GFP_KERNEL allocation, and all nodes in the
2995 * current tasks mems_allowed came up empty on the first pass over
2996 * the zonelist. So only GFP_KERNEL allocations, if all nodes in the
2997 * cpuset are short of memory, might require taking the callback_lock.
2999 * The first call here from mm/page_alloc:get_page_from_freelist()
3000 * has __GFP_HARDWALL set in gfp_mask, enforcing hardwall cpusets,
3001 * so no allocation on a node outside the cpuset is allowed (unless
3002 * in interrupt, of course).
3004 * The second pass through get_page_from_freelist() doesn't even call
3005 * here for GFP_ATOMIC calls. For those calls, the __alloc_pages()
3006 * variable 'wait' is not set, and the bit ALLOC_CPUSET is not set
3007 * in alloc_flags. That logic and the checks below have the combined
3009 * in_interrupt - any node ok (current task context irrelevant)
3010 * GFP_ATOMIC - any node ok
3011 * tsk_is_oom_victim - any node ok
3012 * GFP_KERNEL - any node in enclosing hardwalled cpuset ok
3013 * GFP_USER - only nodes in current tasks mems allowed ok.
3015 bool __cpuset_node_allowed(int node, gfp_t gfp_mask)
3017 struct cpuset *cs; /* current cpuset ancestors */
3018 int allowed; /* is allocation in zone z allowed? */
3019 unsigned long flags;
3023 if (node_isset(node, current->mems_allowed))
3026 * Allow tasks that have access to memory reserves because they have
3027 * been OOM killed to get memory anywhere.
3029 if (unlikely(tsk_is_oom_victim(current)))
3031 if (gfp_mask & __GFP_HARDWALL) /* If hardwall request, stop here */
3034 if (current->flags & PF_EXITING) /* Let dying task have memory */
3037 /* Not hardwall and node outside mems_allowed: scan up cpusets */
3038 spin_lock_irqsave(&callback_lock, flags);
3041 cs = nearest_hardwall_ancestor(task_cs(current));
3042 allowed = node_isset(node, cs->mems_allowed);
3045 spin_unlock_irqrestore(&callback_lock, flags);
3050 * cpuset_mem_spread_node() - On which node to begin search for a file page
3051 * cpuset_slab_spread_node() - On which node to begin search for a slab page
3053 * If a task is marked PF_SPREAD_PAGE or PF_SPREAD_SLAB (as for
3054 * tasks in a cpuset with is_spread_page or is_spread_slab set),
3055 * and if the memory allocation used cpuset_mem_spread_node()
3056 * to determine on which node to start looking, as it will for
3057 * certain page cache or slab cache pages such as used for file
3058 * system buffers and inode caches, then instead of starting on the
3059 * local node to look for a free page, rather spread the starting
3060 * node around the tasks mems_allowed nodes.
3062 * We don't have to worry about the returned node being offline
3063 * because "it can't happen", and even if it did, it would be ok.
3065 * The routines calling guarantee_online_mems() are careful to
3066 * only set nodes in task->mems_allowed that are online. So it
3067 * should not be possible for the following code to return an
3068 * offline node. But if it did, that would be ok, as this routine
3069 * is not returning the node where the allocation must be, only
3070 * the node where the search should start. The zonelist passed to
3071 * __alloc_pages() will include all nodes. If the slab allocator
3072 * is passed an offline node, it will fall back to the local node.
3073 * See kmem_cache_alloc_node().
3076 static int cpuset_spread_node(int *rotor)
3078 return *rotor = next_node_in(*rotor, current->mems_allowed);
3081 int cpuset_mem_spread_node(void)
3083 if (current->cpuset_mem_spread_rotor == NUMA_NO_NODE)
3084 current->cpuset_mem_spread_rotor =
3085 node_random(¤t->mems_allowed);
3087 return cpuset_spread_node(¤t->cpuset_mem_spread_rotor);
3090 int cpuset_slab_spread_node(void)
3092 if (current->cpuset_slab_spread_rotor == NUMA_NO_NODE)
3093 current->cpuset_slab_spread_rotor =
3094 node_random(¤t->mems_allowed);
3096 return cpuset_spread_node(¤t->cpuset_slab_spread_rotor);
3099 EXPORT_SYMBOL_GPL(cpuset_mem_spread_node);
3102 * cpuset_mems_allowed_intersects - Does @tsk1's mems_allowed intersect @tsk2's?
3103 * @tsk1: pointer to task_struct of some task.
3104 * @tsk2: pointer to task_struct of some other task.
3106 * Description: Return true if @tsk1's mems_allowed intersects the
3107 * mems_allowed of @tsk2. Used by the OOM killer to determine if
3108 * one of the task's memory usage might impact the memory available
3112 int cpuset_mems_allowed_intersects(const struct task_struct *tsk1,
3113 const struct task_struct *tsk2)
3115 return nodes_intersects(tsk1->mems_allowed, tsk2->mems_allowed);
3119 * cpuset_print_current_mems_allowed - prints current's cpuset and mems_allowed
3121 * Description: Prints current's name, cpuset name, and cached copy of its
3122 * mems_allowed to the kernel log.
3124 void cpuset_print_current_mems_allowed(void)
3126 struct cgroup *cgrp;
3130 cgrp = task_cs(current)->css.cgroup;
3131 pr_info("%s cpuset=", current->comm);
3132 pr_cont_cgroup_name(cgrp);
3133 pr_cont(" mems_allowed=%*pbl\n",
3134 nodemask_pr_args(¤t->mems_allowed));
3140 * Collection of memory_pressure is suppressed unless
3141 * this flag is enabled by writing "1" to the special
3142 * cpuset file 'memory_pressure_enabled' in the root cpuset.
3145 int cpuset_memory_pressure_enabled __read_mostly;
3148 * cpuset_memory_pressure_bump - keep stats of per-cpuset reclaims.
3150 * Keep a running average of the rate of synchronous (direct)
3151 * page reclaim efforts initiated by tasks in each cpuset.
3153 * This represents the rate at which some task in the cpuset
3154 * ran low on memory on all nodes it was allowed to use, and
3155 * had to enter the kernels page reclaim code in an effort to
3156 * create more free memory by tossing clean pages or swapping
3157 * or writing dirty pages.
3159 * Display to user space in the per-cpuset read-only file
3160 * "memory_pressure". Value displayed is an integer
3161 * representing the recent rate of entry into the synchronous
3162 * (direct) page reclaim by any task attached to the cpuset.
3165 void __cpuset_memory_pressure_bump(void)
3168 fmeter_markevent(&task_cs(current)->fmeter);
3172 #ifdef CONFIG_PROC_PID_CPUSET
3174 * proc_cpuset_show()
3175 * - Print tasks cpuset path into seq_file.
3176 * - Used for /proc/<pid>/cpuset.
3177 * - No need to task_lock(tsk) on this tsk->cpuset reference, as it
3178 * doesn't really matter if tsk->cpuset changes after we read it,
3179 * and we take cpuset_mutex, keeping cpuset_attach() from changing it
3182 int proc_cpuset_show(struct seq_file *m, struct pid_namespace *ns,
3183 struct pid *pid, struct task_struct *tsk)
3186 struct cgroup_subsys_state *css;
3190 buf = kmalloc(PATH_MAX, GFP_KERNEL);
3194 css = task_get_css(tsk, cpuset_cgrp_id);
3195 retval = cgroup_path_ns(css->cgroup, buf, PATH_MAX,
3196 current->nsproxy->cgroup_ns);
3198 if (retval >= PATH_MAX)
3199 retval = -ENAMETOOLONG;
3210 #endif /* CONFIG_PROC_PID_CPUSET */
3212 /* Display task mems_allowed in /proc/<pid>/status file. */
3213 void cpuset_task_status_allowed(struct seq_file *m, struct task_struct *task)
3215 seq_printf(m, "Mems_allowed:\t%*pb\n",
3216 nodemask_pr_args(&task->mems_allowed));
3217 seq_printf(m, "Mems_allowed_list:\t%*pbl\n",
3218 nodemask_pr_args(&task->mems_allowed));