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 * effective_cpus contains only onlined CPUs, but subparts_cpus
118 * may have offlined ones.
120 cpumask_var_t subparts_cpus;
123 * This is old Memory Nodes tasks took on.
125 * - top_cpuset.old_mems_allowed is initialized to mems_allowed.
126 * - A new cpuset's old_mems_allowed is initialized when some
127 * task is moved into it.
128 * - old_mems_allowed is used in cpuset_migrate_mm() when we change
129 * cpuset.mems_allowed and have tasks' nodemask updated, and
130 * then old_mems_allowed is updated to mems_allowed.
132 nodemask_t old_mems_allowed;
134 struct fmeter fmeter; /* memory_pressure filter */
137 * Tasks are being attached to this cpuset. Used to prevent
138 * zeroing cpus/mems_allowed between ->can_attach() and ->attach().
140 int attach_in_progress;
142 /* partition number for rebuild_sched_domains() */
145 /* for custom sched domain */
146 int relax_domain_level;
148 /* number of CPUs in subparts_cpus */
149 int nr_subparts_cpus;
151 /* partition root state */
152 int partition_root_state;
155 * Default hierarchy only:
156 * use_parent_ecpus - set if using parent's effective_cpus
157 * child_ecpus_count - # of children with use_parent_ecpus set
159 int use_parent_ecpus;
160 int child_ecpus_count;
164 * Partition root states:
166 * 0 - not a partition root
170 * -1 - invalid partition root
171 * None of the cpus in cpus_allowed can be put into the parent's
172 * subparts_cpus. In this case, the cpuset is not a real partition
173 * root anymore. However, the CPU_EXCLUSIVE bit will still be set
174 * and the cpuset can be restored back to a partition root if the
175 * parent cpuset can give more CPUs back to this child cpuset.
177 #define PRS_DISABLED 0
178 #define PRS_ENABLED 1
182 * Temporary cpumasks for working with partitions that are passed among
183 * functions to avoid memory allocation in inner functions.
186 cpumask_var_t addmask, delmask; /* For partition root */
187 cpumask_var_t new_cpus; /* For update_cpumasks_hier() */
190 static inline struct cpuset *css_cs(struct cgroup_subsys_state *css)
192 return css ? container_of(css, struct cpuset, css) : NULL;
195 /* Retrieve the cpuset for a task */
196 static inline struct cpuset *task_cs(struct task_struct *task)
198 return css_cs(task_css(task, cpuset_cgrp_id));
201 static inline struct cpuset *parent_cs(struct cpuset *cs)
203 return css_cs(cs->css.parent);
207 static inline bool task_has_mempolicy(struct task_struct *task)
209 return task->mempolicy;
212 static inline bool task_has_mempolicy(struct task_struct *task)
219 /* bits in struct cpuset flags field */
226 CS_SCHED_LOAD_BALANCE,
231 /* convenient tests for these bits */
232 static inline bool is_cpuset_online(struct cpuset *cs)
234 return test_bit(CS_ONLINE, &cs->flags) && !css_is_dying(&cs->css);
237 static inline int is_cpu_exclusive(const struct cpuset *cs)
239 return test_bit(CS_CPU_EXCLUSIVE, &cs->flags);
242 static inline int is_mem_exclusive(const struct cpuset *cs)
244 return test_bit(CS_MEM_EXCLUSIVE, &cs->flags);
247 static inline int is_mem_hardwall(const struct cpuset *cs)
249 return test_bit(CS_MEM_HARDWALL, &cs->flags);
252 static inline int is_sched_load_balance(const struct cpuset *cs)
254 return test_bit(CS_SCHED_LOAD_BALANCE, &cs->flags);
257 static inline int is_memory_migrate(const struct cpuset *cs)
259 return test_bit(CS_MEMORY_MIGRATE, &cs->flags);
262 static inline int is_spread_page(const struct cpuset *cs)
264 return test_bit(CS_SPREAD_PAGE, &cs->flags);
267 static inline int is_spread_slab(const struct cpuset *cs)
269 return test_bit(CS_SPREAD_SLAB, &cs->flags);
272 static inline int is_partition_root(const struct cpuset *cs)
274 return cs->partition_root_state > 0;
277 static struct cpuset top_cpuset = {
278 .flags = ((1 << CS_ONLINE) | (1 << CS_CPU_EXCLUSIVE) |
279 (1 << CS_MEM_EXCLUSIVE)),
280 .partition_root_state = PRS_ENABLED,
284 * cpuset_for_each_child - traverse online children of a cpuset
285 * @child_cs: loop cursor pointing to the current child
286 * @pos_css: used for iteration
287 * @parent_cs: target cpuset to walk children of
289 * Walk @child_cs through the online children of @parent_cs. Must be used
290 * with RCU read locked.
292 #define cpuset_for_each_child(child_cs, pos_css, parent_cs) \
293 css_for_each_child((pos_css), &(parent_cs)->css) \
294 if (is_cpuset_online(((child_cs) = css_cs((pos_css)))))
297 * cpuset_for_each_descendant_pre - pre-order walk of a cpuset's descendants
298 * @des_cs: loop cursor pointing to the current descendant
299 * @pos_css: used for iteration
300 * @root_cs: target cpuset to walk ancestor of
302 * Walk @des_cs through the online descendants of @root_cs. Must be used
303 * with RCU read locked. The caller may modify @pos_css by calling
304 * css_rightmost_descendant() to skip subtree. @root_cs is included in the
305 * iteration and the first node to be visited.
307 #define cpuset_for_each_descendant_pre(des_cs, pos_css, root_cs) \
308 css_for_each_descendant_pre((pos_css), &(root_cs)->css) \
309 if (is_cpuset_online(((des_cs) = css_cs((pos_css)))))
312 * There are two global locks guarding cpuset structures - cpuset_mutex and
313 * callback_lock. We also require taking task_lock() when dereferencing a
314 * task's cpuset pointer. See "The task_lock() exception", at the end of this
317 * A task must hold both locks to modify cpusets. If a task holds
318 * cpuset_mutex, then it blocks others wanting that mutex, ensuring that it
319 * is the only task able to also acquire callback_lock and be able to
320 * modify cpusets. It can perform various checks on the cpuset structure
321 * first, knowing nothing will change. It can also allocate memory while
322 * just holding cpuset_mutex. While it is performing these checks, various
323 * callback routines can briefly acquire callback_lock to query cpusets.
324 * Once it is ready to make the changes, it takes callback_lock, blocking
327 * Calls to the kernel memory allocator can not be made while holding
328 * callback_lock, as that would risk double tripping on callback_lock
329 * from one of the callbacks into the cpuset code from within
332 * If a task is only holding callback_lock, then it has read-only
335 * Now, the task_struct fields mems_allowed and mempolicy may be changed
336 * by other task, we use alloc_lock in the task_struct fields to protect
339 * The cpuset_common_file_read() handlers only hold callback_lock across
340 * small pieces of code, such as when reading out possibly multi-word
341 * cpumasks and nodemasks.
343 * Accessing a task's cpuset should be done in accordance with the
344 * guidelines for accessing subsystem state in kernel/cgroup.c
347 static DEFINE_MUTEX(cpuset_mutex);
348 static DEFINE_SPINLOCK(callback_lock);
350 static struct workqueue_struct *cpuset_migrate_mm_wq;
353 * CPU / memory hotplug is handled asynchronously.
355 static void cpuset_hotplug_workfn(struct work_struct *work);
356 static DECLARE_WORK(cpuset_hotplug_work, cpuset_hotplug_workfn);
358 static DECLARE_WAIT_QUEUE_HEAD(cpuset_attach_wq);
361 * Cgroup v2 behavior is used when on default hierarchy or the
362 * cgroup_v2_mode flag is set.
364 static inline bool is_in_v2_mode(void)
366 return cgroup_subsys_on_dfl(cpuset_cgrp_subsys) ||
367 (cpuset_cgrp_subsys.root->flags & CGRP_ROOT_CPUSET_V2_MODE);
371 * This is ugly, but preserves the userspace API for existing cpuset
372 * users. If someone tries to mount the "cpuset" filesystem, we
373 * silently switch it to mount "cgroup" instead
375 static struct dentry *cpuset_mount(struct file_system_type *fs_type,
376 int flags, const char *unused_dev_name, void *data)
378 struct file_system_type *cgroup_fs = get_fs_type("cgroup");
379 struct dentry *ret = ERR_PTR(-ENODEV);
383 "release_agent=/sbin/cpuset_release_agent";
384 ret = cgroup_fs->mount(cgroup_fs, flags,
385 unused_dev_name, mountopts);
386 put_filesystem(cgroup_fs);
391 static struct file_system_type cpuset_fs_type = {
393 .mount = cpuset_mount,
397 * Return in pmask the portion of a cpusets's cpus_allowed that
398 * are online. If none are online, walk up the cpuset hierarchy
399 * until we find one that does have some online cpus.
401 * One way or another, we guarantee to return some non-empty subset
402 * of cpu_online_mask.
404 * Call with callback_lock or cpuset_mutex held.
406 static void guarantee_online_cpus(struct cpuset *cs, struct cpumask *pmask)
408 while (!cpumask_intersects(cs->effective_cpus, cpu_online_mask)) {
412 * The top cpuset doesn't have any online cpu as a
413 * consequence of a race between cpuset_hotplug_work
414 * and cpu hotplug notifier. But we know the top
415 * cpuset's effective_cpus is on its way to to be
416 * identical to cpu_online_mask.
418 cpumask_copy(pmask, cpu_online_mask);
422 cpumask_and(pmask, cs->effective_cpus, cpu_online_mask);
426 * Return in *pmask the portion of a cpusets's mems_allowed that
427 * are online, with memory. If none are online with memory, walk
428 * up the cpuset hierarchy until we find one that does have some
429 * online mems. The top cpuset always has some mems online.
431 * One way or another, we guarantee to return some non-empty subset
432 * of node_states[N_MEMORY].
434 * Call with callback_lock or cpuset_mutex held.
436 static void guarantee_online_mems(struct cpuset *cs, nodemask_t *pmask)
438 while (!nodes_intersects(cs->effective_mems, node_states[N_MEMORY]))
440 nodes_and(*pmask, cs->effective_mems, node_states[N_MEMORY]);
444 * update task's spread flag if cpuset's page/slab spread flag is set
446 * Call with callback_lock or cpuset_mutex held.
448 static void cpuset_update_task_spread_flag(struct cpuset *cs,
449 struct task_struct *tsk)
451 if (is_spread_page(cs))
452 task_set_spread_page(tsk);
454 task_clear_spread_page(tsk);
456 if (is_spread_slab(cs))
457 task_set_spread_slab(tsk);
459 task_clear_spread_slab(tsk);
463 * is_cpuset_subset(p, q) - Is cpuset p a subset of cpuset q?
465 * One cpuset is a subset of another if all its allowed CPUs and
466 * Memory Nodes are a subset of the other, and its exclusive flags
467 * are only set if the other's are set. Call holding cpuset_mutex.
470 static int is_cpuset_subset(const struct cpuset *p, const struct cpuset *q)
472 return cpumask_subset(p->cpus_allowed, q->cpus_allowed) &&
473 nodes_subset(p->mems_allowed, q->mems_allowed) &&
474 is_cpu_exclusive(p) <= is_cpu_exclusive(q) &&
475 is_mem_exclusive(p) <= is_mem_exclusive(q);
479 * alloc_cpumasks - allocate three cpumasks for cpuset
480 * @cs: the cpuset that have cpumasks to be allocated.
481 * @tmp: the tmpmasks structure pointer
482 * Return: 0 if successful, -ENOMEM otherwise.
484 * Only one of the two input arguments should be non-NULL.
486 static inline int alloc_cpumasks(struct cpuset *cs, struct tmpmasks *tmp)
488 cpumask_var_t *pmask1, *pmask2, *pmask3;
491 pmask1 = &cs->cpus_allowed;
492 pmask2 = &cs->effective_cpus;
493 pmask3 = &cs->subparts_cpus;
495 pmask1 = &tmp->new_cpus;
496 pmask2 = &tmp->addmask;
497 pmask3 = &tmp->delmask;
500 if (!zalloc_cpumask_var(pmask1, GFP_KERNEL))
503 if (!zalloc_cpumask_var(pmask2, GFP_KERNEL))
506 if (!zalloc_cpumask_var(pmask3, GFP_KERNEL))
512 free_cpumask_var(*pmask2);
514 free_cpumask_var(*pmask1);
519 * free_cpumasks - free cpumasks in a tmpmasks structure
520 * @cs: the cpuset that have cpumasks to be free.
521 * @tmp: the tmpmasks structure pointer
523 static inline void free_cpumasks(struct cpuset *cs, struct tmpmasks *tmp)
526 free_cpumask_var(cs->cpus_allowed);
527 free_cpumask_var(cs->effective_cpus);
528 free_cpumask_var(cs->subparts_cpus);
531 free_cpumask_var(tmp->new_cpus);
532 free_cpumask_var(tmp->addmask);
533 free_cpumask_var(tmp->delmask);
538 * alloc_trial_cpuset - allocate a trial cpuset
539 * @cs: the cpuset that the trial cpuset duplicates
541 static struct cpuset *alloc_trial_cpuset(struct cpuset *cs)
543 struct cpuset *trial;
545 trial = kmemdup(cs, sizeof(*cs), GFP_KERNEL);
549 if (alloc_cpumasks(trial, NULL)) {
554 cpumask_copy(trial->cpus_allowed, cs->cpus_allowed);
555 cpumask_copy(trial->effective_cpus, cs->effective_cpus);
560 * free_cpuset - free the cpuset
561 * @cs: the cpuset to be freed
563 static inline void free_cpuset(struct cpuset *cs)
565 free_cpumasks(cs, NULL);
570 * validate_change() - Used to validate that any proposed cpuset change
571 * follows the structural rules for cpusets.
573 * If we replaced the flag and mask values of the current cpuset
574 * (cur) with those values in the trial cpuset (trial), would
575 * our various subset and exclusive rules still be valid? Presumes
578 * 'cur' is the address of an actual, in-use cpuset. Operations
579 * such as list traversal that depend on the actual address of the
580 * cpuset in the list must use cur below, not trial.
582 * 'trial' is the address of bulk structure copy of cur, with
583 * perhaps one or more of the fields cpus_allowed, mems_allowed,
584 * or flags changed to new, trial values.
586 * Return 0 if valid, -errno if not.
589 static int validate_change(struct cpuset *cur, struct cpuset *trial)
591 struct cgroup_subsys_state *css;
592 struct cpuset *c, *par;
597 /* Each of our child cpusets must be a subset of us */
599 cpuset_for_each_child(c, css, cur)
600 if (!is_cpuset_subset(c, trial))
603 /* Remaining checks don't apply to root cpuset */
605 if (cur == &top_cpuset)
608 par = parent_cs(cur);
610 /* On legacy hiearchy, we must be a subset of our parent cpuset. */
612 if (!is_in_v2_mode() && !is_cpuset_subset(trial, par))
616 * If either I or some sibling (!= me) is exclusive, we can't
620 cpuset_for_each_child(c, css, par) {
621 if ((is_cpu_exclusive(trial) || is_cpu_exclusive(c)) &&
623 cpumask_intersects(trial->cpus_allowed, c->cpus_allowed))
625 if ((is_mem_exclusive(trial) || is_mem_exclusive(c)) &&
627 nodes_intersects(trial->mems_allowed, c->mems_allowed))
632 * Cpusets with tasks - existing or newly being attached - can't
633 * be changed to have empty cpus_allowed or mems_allowed.
636 if ((cgroup_is_populated(cur->css.cgroup) || cur->attach_in_progress)) {
637 if (!cpumask_empty(cur->cpus_allowed) &&
638 cpumask_empty(trial->cpus_allowed))
640 if (!nodes_empty(cur->mems_allowed) &&
641 nodes_empty(trial->mems_allowed))
646 * We can't shrink if we won't have enough room for SCHED_DEADLINE
650 if (is_cpu_exclusive(cur) &&
651 !cpuset_cpumask_can_shrink(cur->cpus_allowed,
652 trial->cpus_allowed))
663 * Helper routine for generate_sched_domains().
664 * Do cpusets a, b have overlapping effective cpus_allowed masks?
666 static int cpusets_overlap(struct cpuset *a, struct cpuset *b)
668 return cpumask_intersects(a->effective_cpus, b->effective_cpus);
672 update_domain_attr(struct sched_domain_attr *dattr, struct cpuset *c)
674 if (dattr->relax_domain_level < c->relax_domain_level)
675 dattr->relax_domain_level = c->relax_domain_level;
679 static void update_domain_attr_tree(struct sched_domain_attr *dattr,
680 struct cpuset *root_cs)
683 struct cgroup_subsys_state *pos_css;
686 cpuset_for_each_descendant_pre(cp, pos_css, root_cs) {
687 /* skip the whole subtree if @cp doesn't have any CPU */
688 if (cpumask_empty(cp->cpus_allowed)) {
689 pos_css = css_rightmost_descendant(pos_css);
693 if (is_sched_load_balance(cp))
694 update_domain_attr(dattr, cp);
699 /* Must be called with cpuset_mutex held. */
700 static inline int nr_cpusets(void)
702 /* jump label reference count + the top-level cpuset */
703 return static_key_count(&cpusets_enabled_key.key) + 1;
707 * generate_sched_domains()
709 * This function builds a partial partition of the systems CPUs
710 * A 'partial partition' is a set of non-overlapping subsets whose
711 * union is a subset of that set.
712 * The output of this function needs to be passed to kernel/sched/core.c
713 * partition_sched_domains() routine, which will rebuild the scheduler's
714 * load balancing domains (sched domains) as specified by that partial
717 * See "What is sched_load_balance" in Documentation/cgroup-v1/cpusets.txt
718 * for a background explanation of this.
720 * Does not return errors, on the theory that the callers of this
721 * routine would rather not worry about failures to rebuild sched
722 * domains when operating in the severe memory shortage situations
723 * that could cause allocation failures below.
725 * Must be called with cpuset_mutex held.
727 * The three key local variables below are:
728 * q - a linked-list queue of cpuset pointers, used to implement a
729 * top-down scan of all cpusets. This scan loads a pointer
730 * to each cpuset marked is_sched_load_balance into the
731 * array 'csa'. For our purposes, rebuilding the schedulers
732 * sched domains, we can ignore !is_sched_load_balance cpusets.
733 * csa - (for CpuSet Array) Array of pointers to all the cpusets
734 * that need to be load balanced, for convenient iterative
735 * access by the subsequent code that finds the best partition,
736 * i.e the set of domains (subsets) of CPUs such that the
737 * cpus_allowed of every cpuset marked is_sched_load_balance
738 * is a subset of one of these domains, while there are as
739 * many such domains as possible, each as small as possible.
740 * doms - Conversion of 'csa' to an array of cpumasks, for passing to
741 * the kernel/sched/core.c routine partition_sched_domains() in a
742 * convenient format, that can be easily compared to the prior
743 * value to determine what partition elements (sched domains)
744 * were changed (added or removed.)
746 * Finding the best partition (set of domains):
747 * The triple nested loops below over i, j, k scan over the
748 * load balanced cpusets (using the array of cpuset pointers in
749 * csa[]) looking for pairs of cpusets that have overlapping
750 * cpus_allowed, but which don't have the same 'pn' partition
751 * number and gives them in the same partition number. It keeps
752 * looping on the 'restart' label until it can no longer find
755 * The union of the cpus_allowed masks from the set of
756 * all cpusets having the same 'pn' value then form the one
757 * element of the partition (one sched domain) to be passed to
758 * partition_sched_domains().
760 static int generate_sched_domains(cpumask_var_t **domains,
761 struct sched_domain_attr **attributes)
763 struct cpuset *cp; /* scans q */
764 struct cpuset **csa; /* array of all cpuset ptrs */
765 int csn; /* how many cpuset ptrs in csa so far */
766 int i, j, k; /* indices for partition finding loops */
767 cpumask_var_t *doms; /* resulting partition; i.e. sched domains */
768 struct sched_domain_attr *dattr; /* attributes for custom domains */
769 int ndoms = 0; /* number of sched domains in result */
770 int nslot; /* next empty doms[] struct cpumask slot */
771 struct cgroup_subsys_state *pos_css;
772 bool root_load_balance = is_sched_load_balance(&top_cpuset);
778 /* Special case for the 99% of systems with one, full, sched domain */
779 if (root_load_balance && !top_cpuset.nr_subparts_cpus) {
781 doms = alloc_sched_domains(ndoms);
785 dattr = kmalloc(sizeof(struct sched_domain_attr), GFP_KERNEL);
787 *dattr = SD_ATTR_INIT;
788 update_domain_attr_tree(dattr, &top_cpuset);
790 cpumask_and(doms[0], top_cpuset.effective_cpus,
791 housekeeping_cpumask(HK_FLAG_DOMAIN));
796 csa = kmalloc_array(nr_cpusets(), sizeof(cp), GFP_KERNEL);
802 if (root_load_balance)
803 csa[csn++] = &top_cpuset;
804 cpuset_for_each_descendant_pre(cp, pos_css, &top_cpuset) {
805 if (cp == &top_cpuset)
808 * Continue traversing beyond @cp iff @cp has some CPUs and
809 * isn't load balancing. The former is obvious. The
810 * latter: All child cpusets contain a subset of the
811 * parent's cpus, so just skip them, and then we call
812 * update_domain_attr_tree() to calc relax_domain_level of
813 * the corresponding sched domain.
815 * If root is load-balancing, we can skip @cp if it
816 * is a subset of the root's effective_cpus.
818 if (!cpumask_empty(cp->cpus_allowed) &&
819 !(is_sched_load_balance(cp) &&
820 cpumask_intersects(cp->cpus_allowed,
821 housekeeping_cpumask(HK_FLAG_DOMAIN))))
824 if (root_load_balance &&
825 cpumask_subset(cp->cpus_allowed, top_cpuset.effective_cpus))
828 if (is_sched_load_balance(cp))
831 /* skip @cp's subtree if not a partition root */
832 if (!is_partition_root(cp))
833 pos_css = css_rightmost_descendant(pos_css);
837 for (i = 0; i < csn; i++)
842 /* Find the best partition (set of sched domains) */
843 for (i = 0; i < csn; i++) {
844 struct cpuset *a = csa[i];
847 for (j = 0; j < csn; j++) {
848 struct cpuset *b = csa[j];
851 if (apn != bpn && cpusets_overlap(a, b)) {
852 for (k = 0; k < csn; k++) {
853 struct cpuset *c = csa[k];
858 ndoms--; /* one less element */
865 * Now we know how many domains to create.
866 * Convert <csn, csa> to <ndoms, doms> and populate cpu masks.
868 doms = alloc_sched_domains(ndoms);
873 * The rest of the code, including the scheduler, can deal with
874 * dattr==NULL case. No need to abort if alloc fails.
876 dattr = kmalloc_array(ndoms, sizeof(struct sched_domain_attr),
879 for (nslot = 0, i = 0; i < csn; i++) {
880 struct cpuset *a = csa[i];
885 /* Skip completed partitions */
891 if (nslot == ndoms) {
892 static int warnings = 10;
894 pr_warn("rebuild_sched_domains confused: nslot %d, ndoms %d, csn %d, i %d, apn %d\n",
895 nslot, ndoms, csn, i, apn);
903 *(dattr + nslot) = SD_ATTR_INIT;
904 for (j = i; j < csn; j++) {
905 struct cpuset *b = csa[j];
908 cpumask_or(dp, dp, b->effective_cpus);
909 cpumask_and(dp, dp, housekeeping_cpumask(HK_FLAG_DOMAIN));
911 update_domain_attr_tree(dattr + nslot, b);
913 /* Done with this partition */
919 BUG_ON(nslot != ndoms);
925 * Fallback to the default domain if kmalloc() failed.
926 * See comments in partition_sched_domains().
937 * Rebuild scheduler domains.
939 * If the flag 'sched_load_balance' of any cpuset with non-empty
940 * 'cpus' changes, or if the 'cpus' allowed changes in any cpuset
941 * which has that flag enabled, or if any cpuset with a non-empty
942 * 'cpus' is removed, then call this routine to rebuild the
943 * scheduler's dynamic sched domains.
945 * Call with cpuset_mutex held. Takes get_online_cpus().
947 static void rebuild_sched_domains_locked(void)
949 struct sched_domain_attr *attr;
953 lockdep_assert_held(&cpuset_mutex);
957 * We have raced with CPU hotplug. Don't do anything to avoid
958 * passing doms with offlined cpu to partition_sched_domains().
959 * Anyways, hotplug work item will rebuild sched domains.
961 if (!top_cpuset.nr_subparts_cpus &&
962 !cpumask_equal(top_cpuset.effective_cpus, cpu_active_mask))
965 if (top_cpuset.nr_subparts_cpus &&
966 !cpumask_subset(top_cpuset.effective_cpus, cpu_active_mask))
969 /* Generate domain masks and attrs */
970 ndoms = generate_sched_domains(&doms, &attr);
972 /* Have scheduler rebuild the domains */
973 partition_sched_domains(ndoms, doms, attr);
977 #else /* !CONFIG_SMP */
978 static void rebuild_sched_domains_locked(void)
981 #endif /* CONFIG_SMP */
983 void rebuild_sched_domains(void)
985 mutex_lock(&cpuset_mutex);
986 rebuild_sched_domains_locked();
987 mutex_unlock(&cpuset_mutex);
991 * update_tasks_cpumask - Update the cpumasks of tasks in the cpuset.
992 * @cs: the cpuset in which each task's cpus_allowed mask needs to be changed
994 * Iterate through each task of @cs updating its cpus_allowed to the
995 * effective cpuset's. As this function is called with cpuset_mutex held,
996 * cpuset membership stays stable.
998 static void update_tasks_cpumask(struct cpuset *cs)
1000 struct css_task_iter it;
1001 struct task_struct *task;
1003 css_task_iter_start(&cs->css, 0, &it);
1004 while ((task = css_task_iter_next(&it)))
1005 set_cpus_allowed_ptr(task, cs->effective_cpus);
1006 css_task_iter_end(&it);
1010 * compute_effective_cpumask - Compute the effective cpumask of the cpuset
1011 * @new_cpus: the temp variable for the new effective_cpus mask
1012 * @cs: the cpuset the need to recompute the new effective_cpus mask
1013 * @parent: the parent cpuset
1015 * If the parent has subpartition CPUs, include them in the list of
1016 * allowable CPUs in computing the new effective_cpus mask. Since offlined
1017 * CPUs are not removed from subparts_cpus, we have to use cpu_active_mask
1018 * to mask those out.
1020 static void compute_effective_cpumask(struct cpumask *new_cpus,
1021 struct cpuset *cs, struct cpuset *parent)
1023 if (parent->nr_subparts_cpus) {
1024 cpumask_or(new_cpus, parent->effective_cpus,
1025 parent->subparts_cpus);
1026 cpumask_and(new_cpus, new_cpus, cs->cpus_allowed);
1027 cpumask_and(new_cpus, new_cpus, cpu_active_mask);
1029 cpumask_and(new_cpus, cs->cpus_allowed, parent->effective_cpus);
1034 * Commands for update_parent_subparts_cpumask
1037 partcmd_enable, /* Enable partition root */
1038 partcmd_disable, /* Disable partition root */
1039 partcmd_update, /* Update parent's subparts_cpus */
1043 * update_parent_subparts_cpumask - update subparts_cpus mask of parent cpuset
1044 * @cpuset: The cpuset that requests change in partition root state
1045 * @cmd: Partition root state change command
1046 * @newmask: Optional new cpumask for partcmd_update
1047 * @tmp: Temporary addmask and delmask
1048 * Return: 0, 1 or an error code
1050 * For partcmd_enable, the cpuset is being transformed from a non-partition
1051 * root to a partition root. The cpus_allowed mask of the given cpuset will
1052 * be put into parent's subparts_cpus and taken away from parent's
1053 * effective_cpus. The function will return 0 if all the CPUs listed in
1054 * cpus_allowed can be granted or an error code will be returned.
1056 * For partcmd_disable, the cpuset is being transofrmed from a partition
1057 * root back to a non-partition root. any CPUs in cpus_allowed that are in
1058 * parent's subparts_cpus will be taken away from that cpumask and put back
1059 * into parent's effective_cpus. 0 should always be returned.
1061 * For partcmd_update, if the optional newmask is specified, the cpu
1062 * list is to be changed from cpus_allowed to newmask. Otherwise,
1063 * cpus_allowed is assumed to remain the same. The cpuset should either
1064 * be a partition root or an invalid partition root. The partition root
1065 * state may change if newmask is NULL and none of the requested CPUs can
1066 * be granted by the parent. The function will return 1 if changes to
1067 * parent's subparts_cpus and effective_cpus happen or 0 otherwise.
1068 * Error code should only be returned when newmask is non-NULL.
1070 * The partcmd_enable and partcmd_disable commands are used by
1071 * update_prstate(). The partcmd_update command is used by
1072 * update_cpumasks_hier() with newmask NULL and update_cpumask() with
1075 * The checking is more strict when enabling partition root than the
1076 * other two commands.
1078 * Because of the implicit cpu exclusive nature of a partition root,
1079 * cpumask changes that violates the cpu exclusivity rule will not be
1080 * permitted when checked by validate_change(). The validate_change()
1081 * function will also prevent any changes to the cpu list if it is not
1082 * a superset of children's cpu lists.
1084 static int update_parent_subparts_cpumask(struct cpuset *cpuset, int cmd,
1085 struct cpumask *newmask,
1086 struct tmpmasks *tmp)
1088 struct cpuset *parent = parent_cs(cpuset);
1089 int adding; /* Moving cpus from effective_cpus to subparts_cpus */
1090 int deleting; /* Moving cpus from subparts_cpus to effective_cpus */
1091 bool part_error = false; /* Partition error? */
1093 lockdep_assert_held(&cpuset_mutex);
1096 * The parent must be a partition root.
1097 * The new cpumask, if present, or the current cpus_allowed must
1100 if (!is_partition_root(parent) ||
1101 (newmask && cpumask_empty(newmask)) ||
1102 (!newmask && cpumask_empty(cpuset->cpus_allowed)))
1106 * Enabling/disabling partition root is not allowed if there are
1109 if ((cmd != partcmd_update) && css_has_online_children(&cpuset->css))
1113 * Enabling partition root is not allowed if not all the CPUs
1114 * can be granted from parent's effective_cpus or at least one
1115 * CPU will be left after that.
1117 if ((cmd == partcmd_enable) &&
1118 (!cpumask_subset(cpuset->cpus_allowed, parent->effective_cpus) ||
1119 cpumask_equal(cpuset->cpus_allowed, parent->effective_cpus)))
1123 * A cpumask update cannot make parent's effective_cpus become empty.
1125 adding = deleting = false;
1126 if (cmd == partcmd_enable) {
1127 cpumask_copy(tmp->addmask, cpuset->cpus_allowed);
1129 } else if (cmd == partcmd_disable) {
1130 deleting = cpumask_and(tmp->delmask, cpuset->cpus_allowed,
1131 parent->subparts_cpus);
1132 } else if (newmask) {
1134 * partcmd_update with newmask:
1136 * delmask = cpus_allowed & ~newmask & parent->subparts_cpus
1137 * addmask = newmask & parent->effective_cpus
1138 * & ~parent->subparts_cpus
1140 cpumask_andnot(tmp->delmask, cpuset->cpus_allowed, newmask);
1141 deleting = cpumask_and(tmp->delmask, tmp->delmask,
1142 parent->subparts_cpus);
1144 cpumask_and(tmp->addmask, newmask, parent->effective_cpus);
1145 adding = cpumask_andnot(tmp->addmask, tmp->addmask,
1146 parent->subparts_cpus);
1148 * Return error if the new effective_cpus could become empty.
1151 cpumask_equal(parent->effective_cpus, tmp->addmask)) {
1155 * As some of the CPUs in subparts_cpus might have
1156 * been offlined, we need to compute the real delmask
1159 if (!cpumask_and(tmp->addmask, tmp->delmask,
1162 cpumask_copy(tmp->addmask, parent->effective_cpus);
1166 * partcmd_update w/o newmask:
1168 * addmask = cpus_allowed & parent->effectiveb_cpus
1170 * Note that parent's subparts_cpus may have been
1171 * pre-shrunk in case there is a change in the cpu list.
1172 * So no deletion is needed.
1174 adding = cpumask_and(tmp->addmask, cpuset->cpus_allowed,
1175 parent->effective_cpus);
1176 part_error = cpumask_equal(tmp->addmask,
1177 parent->effective_cpus);
1180 if (cmd == partcmd_update) {
1181 int prev_prs = cpuset->partition_root_state;
1184 * Check for possible transition between PRS_ENABLED
1187 switch (cpuset->partition_root_state) {
1190 cpuset->partition_root_state = PRS_ERROR;
1194 cpuset->partition_root_state = PRS_ENABLED;
1198 * Set part_error if previously in invalid state.
1200 part_error = (prev_prs == PRS_ERROR);
1203 if (!part_error && (cpuset->partition_root_state == PRS_ERROR))
1204 return 0; /* Nothing need to be done */
1206 if (cpuset->partition_root_state == PRS_ERROR) {
1208 * Remove all its cpus from parent's subparts_cpus.
1211 deleting = cpumask_and(tmp->delmask, cpuset->cpus_allowed,
1212 parent->subparts_cpus);
1215 if (!adding && !deleting)
1219 * Change the parent's subparts_cpus.
1220 * Newly added CPUs will be removed from effective_cpus and
1221 * newly deleted ones will be added back to effective_cpus.
1223 spin_lock_irq(&callback_lock);
1225 cpumask_or(parent->subparts_cpus,
1226 parent->subparts_cpus, tmp->addmask);
1227 cpumask_andnot(parent->effective_cpus,
1228 parent->effective_cpus, tmp->addmask);
1231 cpumask_andnot(parent->subparts_cpus,
1232 parent->subparts_cpus, tmp->delmask);
1234 * Some of the CPUs in subparts_cpus might have been offlined.
1236 cpumask_and(tmp->delmask, tmp->delmask, cpu_active_mask);
1237 cpumask_or(parent->effective_cpus,
1238 parent->effective_cpus, tmp->delmask);
1241 parent->nr_subparts_cpus = cpumask_weight(parent->subparts_cpus);
1242 spin_unlock_irq(&callback_lock);
1244 return cmd == partcmd_update;
1248 * update_cpumasks_hier - Update effective cpumasks and tasks in the subtree
1249 * @cs: the cpuset to consider
1250 * @tmp: temp variables for calculating effective_cpus & partition setup
1252 * When congifured cpumask is changed, the effective cpumasks of this cpuset
1253 * and all its descendants need to be updated.
1255 * On legacy hierachy, effective_cpus will be the same with cpu_allowed.
1257 * Called with cpuset_mutex held
1259 static void update_cpumasks_hier(struct cpuset *cs, struct tmpmasks *tmp)
1262 struct cgroup_subsys_state *pos_css;
1263 bool need_rebuild_sched_domains = false;
1266 cpuset_for_each_descendant_pre(cp, pos_css, cs) {
1267 struct cpuset *parent = parent_cs(cp);
1269 compute_effective_cpumask(tmp->new_cpus, cp, parent);
1272 * If it becomes empty, inherit the effective mask of the
1273 * parent, which is guaranteed to have some CPUs.
1275 if (is_in_v2_mode() && cpumask_empty(tmp->new_cpus)) {
1276 cpumask_copy(tmp->new_cpus, parent->effective_cpus);
1277 if (!cp->use_parent_ecpus) {
1278 cp->use_parent_ecpus = true;
1279 parent->child_ecpus_count++;
1281 } else if (cp->use_parent_ecpus) {
1282 cp->use_parent_ecpus = false;
1283 WARN_ON_ONCE(!parent->child_ecpus_count);
1284 parent->child_ecpus_count--;
1288 * Skip the whole subtree if the cpumask remains the same
1289 * and has no partition root state.
1291 if (!cp->partition_root_state &&
1292 cpumask_equal(tmp->new_cpus, cp->effective_cpus)) {
1293 pos_css = css_rightmost_descendant(pos_css);
1298 * update_parent_subparts_cpumask() should have been called
1299 * for cs already in update_cpumask(). We should also call
1300 * update_tasks_cpumask() again for tasks in the parent
1301 * cpuset if the parent's subparts_cpus changes.
1303 if ((cp != cs) && cp->partition_root_state) {
1304 switch (parent->partition_root_state) {
1307 * If parent is not a partition root or an
1308 * invalid partition root, clear the state
1309 * state and the CS_CPU_EXCLUSIVE flag.
1311 WARN_ON_ONCE(cp->partition_root_state
1313 cp->partition_root_state = 0;
1316 * clear_bit() is an atomic operation and
1317 * readers aren't interested in the state
1318 * of CS_CPU_EXCLUSIVE anyway. So we can
1319 * just update the flag without holding
1320 * the callback_lock.
1322 clear_bit(CS_CPU_EXCLUSIVE, &cp->flags);
1326 if (update_parent_subparts_cpumask(cp, partcmd_update, NULL, tmp))
1327 update_tasks_cpumask(parent);
1332 * When parent is invalid, it has to be too.
1334 cp->partition_root_state = PRS_ERROR;
1335 if (cp->nr_subparts_cpus) {
1336 cp->nr_subparts_cpus = 0;
1337 cpumask_clear(cp->subparts_cpus);
1343 if (!css_tryget_online(&cp->css))
1347 spin_lock_irq(&callback_lock);
1349 cpumask_copy(cp->effective_cpus, tmp->new_cpus);
1350 if (cp->nr_subparts_cpus &&
1351 (cp->partition_root_state != PRS_ENABLED)) {
1352 cp->nr_subparts_cpus = 0;
1353 cpumask_clear(cp->subparts_cpus);
1354 } else if (cp->nr_subparts_cpus) {
1356 * Make sure that effective_cpus & subparts_cpus
1357 * are mutually exclusive.
1359 * In the unlikely event that effective_cpus
1360 * becomes empty. we clear cp->nr_subparts_cpus and
1361 * let its child partition roots to compete for
1364 cpumask_andnot(cp->effective_cpus, cp->effective_cpus,
1366 if (cpumask_empty(cp->effective_cpus)) {
1367 cpumask_copy(cp->effective_cpus, tmp->new_cpus);
1368 cpumask_clear(cp->subparts_cpus);
1369 cp->nr_subparts_cpus = 0;
1370 } else if (!cpumask_subset(cp->subparts_cpus,
1372 cpumask_andnot(cp->subparts_cpus,
1373 cp->subparts_cpus, tmp->new_cpus);
1374 cp->nr_subparts_cpus
1375 = cpumask_weight(cp->subparts_cpus);
1378 spin_unlock_irq(&callback_lock);
1380 WARN_ON(!is_in_v2_mode() &&
1381 !cpumask_equal(cp->cpus_allowed, cp->effective_cpus));
1383 update_tasks_cpumask(cp);
1386 * On legacy hierarchy, if the effective cpumask of any non-
1387 * empty cpuset is changed, we need to rebuild sched domains.
1388 * On default hierarchy, the cpuset needs to be a partition
1391 if (!cpumask_empty(cp->cpus_allowed) &&
1392 is_sched_load_balance(cp) &&
1393 (!cgroup_subsys_on_dfl(cpuset_cgrp_subsys) ||
1394 is_partition_root(cp)))
1395 need_rebuild_sched_domains = true;
1402 if (need_rebuild_sched_domains)
1403 rebuild_sched_domains_locked();
1407 * update_sibling_cpumasks - Update siblings cpumasks
1408 * @parent: Parent cpuset
1409 * @cs: Current cpuset
1410 * @tmp: Temp variables
1412 static void update_sibling_cpumasks(struct cpuset *parent, struct cpuset *cs,
1413 struct tmpmasks *tmp)
1415 struct cpuset *sibling;
1416 struct cgroup_subsys_state *pos_css;
1419 * Check all its siblings and call update_cpumasks_hier()
1420 * if their use_parent_ecpus flag is set in order for them
1421 * to use the right effective_cpus value.
1424 cpuset_for_each_child(sibling, pos_css, parent) {
1427 if (!sibling->use_parent_ecpus)
1430 update_cpumasks_hier(sibling, tmp);
1436 * update_cpumask - update the cpus_allowed mask of a cpuset and all tasks in it
1437 * @cs: the cpuset to consider
1438 * @trialcs: trial cpuset
1439 * @buf: buffer of cpu numbers written to this cpuset
1441 static int update_cpumask(struct cpuset *cs, struct cpuset *trialcs,
1445 struct tmpmasks tmp;
1447 /* top_cpuset.cpus_allowed tracks cpu_online_mask; it's read-only */
1448 if (cs == &top_cpuset)
1452 * An empty cpus_allowed is ok only if the cpuset has no tasks.
1453 * Since cpulist_parse() fails on an empty mask, we special case
1454 * that parsing. The validate_change() call ensures that cpusets
1455 * with tasks have cpus.
1458 cpumask_clear(trialcs->cpus_allowed);
1460 retval = cpulist_parse(buf, trialcs->cpus_allowed);
1464 if (!cpumask_subset(trialcs->cpus_allowed,
1465 top_cpuset.cpus_allowed))
1469 /* Nothing to do if the cpus didn't change */
1470 if (cpumask_equal(cs->cpus_allowed, trialcs->cpus_allowed))
1473 retval = validate_change(cs, trialcs);
1477 #ifdef CONFIG_CPUMASK_OFFSTACK
1479 * Use the cpumasks in trialcs for tmpmasks when they are pointers
1480 * to allocated cpumasks.
1482 tmp.addmask = trialcs->subparts_cpus;
1483 tmp.delmask = trialcs->effective_cpus;
1484 tmp.new_cpus = trialcs->cpus_allowed;
1487 if (cs->partition_root_state) {
1488 /* Cpumask of a partition root cannot be empty */
1489 if (cpumask_empty(trialcs->cpus_allowed))
1491 if (update_parent_subparts_cpumask(cs, partcmd_update,
1492 trialcs->cpus_allowed, &tmp) < 0)
1496 spin_lock_irq(&callback_lock);
1497 cpumask_copy(cs->cpus_allowed, trialcs->cpus_allowed);
1500 * Make sure that subparts_cpus is a subset of cpus_allowed.
1502 if (cs->nr_subparts_cpus) {
1503 cpumask_andnot(cs->subparts_cpus, cs->subparts_cpus,
1505 cs->nr_subparts_cpus = cpumask_weight(cs->subparts_cpus);
1507 spin_unlock_irq(&callback_lock);
1509 update_cpumasks_hier(cs, &tmp);
1511 if (cs->partition_root_state) {
1512 struct cpuset *parent = parent_cs(cs);
1515 * For partition root, update the cpumasks of sibling
1516 * cpusets if they use parent's effective_cpus.
1518 if (parent->child_ecpus_count)
1519 update_sibling_cpumasks(parent, cs, &tmp);
1525 * Migrate memory region from one set of nodes to another. This is
1526 * performed asynchronously as it can be called from process migration path
1527 * holding locks involved in process management. All mm migrations are
1528 * performed in the queued order and can be waited for by flushing
1529 * cpuset_migrate_mm_wq.
1532 struct cpuset_migrate_mm_work {
1533 struct work_struct work;
1534 struct mm_struct *mm;
1539 static void cpuset_migrate_mm_workfn(struct work_struct *work)
1541 struct cpuset_migrate_mm_work *mwork =
1542 container_of(work, struct cpuset_migrate_mm_work, work);
1544 /* on a wq worker, no need to worry about %current's mems_allowed */
1545 do_migrate_pages(mwork->mm, &mwork->from, &mwork->to, MPOL_MF_MOVE_ALL);
1550 static void cpuset_migrate_mm(struct mm_struct *mm, const nodemask_t *from,
1551 const nodemask_t *to)
1553 struct cpuset_migrate_mm_work *mwork;
1555 mwork = kzalloc(sizeof(*mwork), GFP_KERNEL);
1558 mwork->from = *from;
1560 INIT_WORK(&mwork->work, cpuset_migrate_mm_workfn);
1561 queue_work(cpuset_migrate_mm_wq, &mwork->work);
1567 static void cpuset_post_attach(void)
1569 flush_workqueue(cpuset_migrate_mm_wq);
1573 * cpuset_change_task_nodemask - change task's mems_allowed and mempolicy
1574 * @tsk: the task to change
1575 * @newmems: new nodes that the task will be set
1577 * We use the mems_allowed_seq seqlock to safely update both tsk->mems_allowed
1578 * and rebind an eventual tasks' mempolicy. If the task is allocating in
1579 * parallel, it might temporarily see an empty intersection, which results in
1580 * a seqlock check and retry before OOM or allocation failure.
1582 static void cpuset_change_task_nodemask(struct task_struct *tsk,
1583 nodemask_t *newmems)
1587 local_irq_disable();
1588 write_seqcount_begin(&tsk->mems_allowed_seq);
1590 nodes_or(tsk->mems_allowed, tsk->mems_allowed, *newmems);
1591 mpol_rebind_task(tsk, newmems);
1592 tsk->mems_allowed = *newmems;
1594 write_seqcount_end(&tsk->mems_allowed_seq);
1600 static void *cpuset_being_rebound;
1603 * update_tasks_nodemask - Update the nodemasks of tasks in the cpuset.
1604 * @cs: the cpuset in which each task's mems_allowed mask needs to be changed
1606 * Iterate through each task of @cs updating its mems_allowed to the
1607 * effective cpuset's. As this function is called with cpuset_mutex held,
1608 * cpuset membership stays stable.
1610 static void update_tasks_nodemask(struct cpuset *cs)
1612 static nodemask_t newmems; /* protected by cpuset_mutex */
1613 struct css_task_iter it;
1614 struct task_struct *task;
1616 cpuset_being_rebound = cs; /* causes mpol_dup() rebind */
1618 guarantee_online_mems(cs, &newmems);
1621 * The mpol_rebind_mm() call takes mmap_sem, which we couldn't
1622 * take while holding tasklist_lock. Forks can happen - the
1623 * mpol_dup() cpuset_being_rebound check will catch such forks,
1624 * and rebind their vma mempolicies too. Because we still hold
1625 * the global cpuset_mutex, we know that no other rebind effort
1626 * will be contending for the global variable cpuset_being_rebound.
1627 * It's ok if we rebind the same mm twice; mpol_rebind_mm()
1628 * is idempotent. Also migrate pages in each mm to new nodes.
1630 css_task_iter_start(&cs->css, 0, &it);
1631 while ((task = css_task_iter_next(&it))) {
1632 struct mm_struct *mm;
1635 cpuset_change_task_nodemask(task, &newmems);
1637 mm = get_task_mm(task);
1641 migrate = is_memory_migrate(cs);
1643 mpol_rebind_mm(mm, &cs->mems_allowed);
1645 cpuset_migrate_mm(mm, &cs->old_mems_allowed, &newmems);
1649 css_task_iter_end(&it);
1652 * All the tasks' nodemasks have been updated, update
1653 * cs->old_mems_allowed.
1655 cs->old_mems_allowed = newmems;
1657 /* We're done rebinding vmas to this cpuset's new mems_allowed. */
1658 cpuset_being_rebound = NULL;
1662 * update_nodemasks_hier - Update effective nodemasks and tasks in the subtree
1663 * @cs: the cpuset to consider
1664 * @new_mems: a temp variable for calculating new effective_mems
1666 * When configured nodemask is changed, the effective nodemasks of this cpuset
1667 * and all its descendants need to be updated.
1669 * On legacy hiearchy, effective_mems will be the same with mems_allowed.
1671 * Called with cpuset_mutex held
1673 static void update_nodemasks_hier(struct cpuset *cs, nodemask_t *new_mems)
1676 struct cgroup_subsys_state *pos_css;
1679 cpuset_for_each_descendant_pre(cp, pos_css, cs) {
1680 struct cpuset *parent = parent_cs(cp);
1682 nodes_and(*new_mems, cp->mems_allowed, parent->effective_mems);
1685 * If it becomes empty, inherit the effective mask of the
1686 * parent, which is guaranteed to have some MEMs.
1688 if (is_in_v2_mode() && nodes_empty(*new_mems))
1689 *new_mems = parent->effective_mems;
1691 /* Skip the whole subtree if the nodemask remains the same. */
1692 if (nodes_equal(*new_mems, cp->effective_mems)) {
1693 pos_css = css_rightmost_descendant(pos_css);
1697 if (!css_tryget_online(&cp->css))
1701 spin_lock_irq(&callback_lock);
1702 cp->effective_mems = *new_mems;
1703 spin_unlock_irq(&callback_lock);
1705 WARN_ON(!is_in_v2_mode() &&
1706 !nodes_equal(cp->mems_allowed, cp->effective_mems));
1708 update_tasks_nodemask(cp);
1717 * Handle user request to change the 'mems' memory placement
1718 * of a cpuset. Needs to validate the request, update the
1719 * cpusets mems_allowed, and for each task in the cpuset,
1720 * update mems_allowed and rebind task's mempolicy and any vma
1721 * mempolicies and if the cpuset is marked 'memory_migrate',
1722 * migrate the tasks pages to the new memory.
1724 * Call with cpuset_mutex held. May take callback_lock during call.
1725 * Will take tasklist_lock, scan tasklist for tasks in cpuset cs,
1726 * lock each such tasks mm->mmap_sem, scan its vma's and rebind
1727 * their mempolicies to the cpusets new mems_allowed.
1729 static int update_nodemask(struct cpuset *cs, struct cpuset *trialcs,
1735 * top_cpuset.mems_allowed tracks node_stats[N_MEMORY];
1738 if (cs == &top_cpuset) {
1744 * An empty mems_allowed is ok iff there are no tasks in the cpuset.
1745 * Since nodelist_parse() fails on an empty mask, we special case
1746 * that parsing. The validate_change() call ensures that cpusets
1747 * with tasks have memory.
1750 nodes_clear(trialcs->mems_allowed);
1752 retval = nodelist_parse(buf, trialcs->mems_allowed);
1756 if (!nodes_subset(trialcs->mems_allowed,
1757 top_cpuset.mems_allowed)) {
1763 if (nodes_equal(cs->mems_allowed, trialcs->mems_allowed)) {
1764 retval = 0; /* Too easy - nothing to do */
1767 retval = validate_change(cs, trialcs);
1771 spin_lock_irq(&callback_lock);
1772 cs->mems_allowed = trialcs->mems_allowed;
1773 spin_unlock_irq(&callback_lock);
1775 /* use trialcs->mems_allowed as a temp variable */
1776 update_nodemasks_hier(cs, &trialcs->mems_allowed);
1781 bool current_cpuset_is_being_rebound(void)
1786 ret = task_cs(current) == cpuset_being_rebound;
1792 static int update_relax_domain_level(struct cpuset *cs, s64 val)
1795 if (val < -1 || val >= sched_domain_level_max)
1799 if (val != cs->relax_domain_level) {
1800 cs->relax_domain_level = val;
1801 if (!cpumask_empty(cs->cpus_allowed) &&
1802 is_sched_load_balance(cs))
1803 rebuild_sched_domains_locked();
1810 * update_tasks_flags - update the spread flags of tasks in the cpuset.
1811 * @cs: the cpuset in which each task's spread flags needs to be changed
1813 * Iterate through each task of @cs updating its spread flags. As this
1814 * function is called with cpuset_mutex held, cpuset membership stays
1817 static void update_tasks_flags(struct cpuset *cs)
1819 struct css_task_iter it;
1820 struct task_struct *task;
1822 css_task_iter_start(&cs->css, 0, &it);
1823 while ((task = css_task_iter_next(&it)))
1824 cpuset_update_task_spread_flag(cs, task);
1825 css_task_iter_end(&it);
1829 * update_flag - read a 0 or a 1 in a file and update associated flag
1830 * bit: the bit to update (see cpuset_flagbits_t)
1831 * cs: the cpuset to update
1832 * turning_on: whether the flag is being set or cleared
1834 * Call with cpuset_mutex held.
1837 static int update_flag(cpuset_flagbits_t bit, struct cpuset *cs,
1840 struct cpuset *trialcs;
1841 int balance_flag_changed;
1842 int spread_flag_changed;
1845 trialcs = alloc_trial_cpuset(cs);
1850 set_bit(bit, &trialcs->flags);
1852 clear_bit(bit, &trialcs->flags);
1854 err = validate_change(cs, trialcs);
1858 balance_flag_changed = (is_sched_load_balance(cs) !=
1859 is_sched_load_balance(trialcs));
1861 spread_flag_changed = ((is_spread_slab(cs) != is_spread_slab(trialcs))
1862 || (is_spread_page(cs) != is_spread_page(trialcs)));
1864 spin_lock_irq(&callback_lock);
1865 cs->flags = trialcs->flags;
1866 spin_unlock_irq(&callback_lock);
1868 if (!cpumask_empty(trialcs->cpus_allowed) && balance_flag_changed)
1869 rebuild_sched_domains_locked();
1871 if (spread_flag_changed)
1872 update_tasks_flags(cs);
1874 free_cpuset(trialcs);
1879 * update_prstate - update partititon_root_state
1880 * cs: the cpuset to update
1881 * val: 0 - disabled, 1 - enabled
1883 * Call with cpuset_mutex held.
1885 static int update_prstate(struct cpuset *cs, int val)
1888 struct cpuset *parent = parent_cs(cs);
1889 struct tmpmasks tmp;
1891 if ((val != 0) && (val != 1))
1893 if (val == cs->partition_root_state)
1897 * Cannot force a partial or invalid partition root to a full
1900 if (val && cs->partition_root_state)
1903 if (alloc_cpumasks(NULL, &tmp))
1907 if (!cs->partition_root_state) {
1909 * Turning on partition root requires setting the
1910 * CS_CPU_EXCLUSIVE bit implicitly as well and cpus_allowed
1913 if (cpumask_empty(cs->cpus_allowed))
1916 err = update_flag(CS_CPU_EXCLUSIVE, cs, 1);
1920 err = update_parent_subparts_cpumask(cs, partcmd_enable,
1923 update_flag(CS_CPU_EXCLUSIVE, cs, 0);
1926 cs->partition_root_state = PRS_ENABLED;
1929 * Turning off partition root will clear the
1930 * CS_CPU_EXCLUSIVE bit.
1932 if (cs->partition_root_state == PRS_ERROR) {
1933 cs->partition_root_state = 0;
1934 update_flag(CS_CPU_EXCLUSIVE, cs, 0);
1939 err = update_parent_subparts_cpumask(cs, partcmd_disable,
1944 cs->partition_root_state = 0;
1946 /* Turning off CS_CPU_EXCLUSIVE will not return error */
1947 update_flag(CS_CPU_EXCLUSIVE, cs, 0);
1951 * Update cpumask of parent's tasks except when it is the top
1952 * cpuset as some system daemons cannot be mapped to other CPUs.
1954 if (parent != &top_cpuset)
1955 update_tasks_cpumask(parent);
1957 if (parent->child_ecpus_count)
1958 update_sibling_cpumasks(parent, cs, &tmp);
1960 rebuild_sched_domains_locked();
1962 free_cpumasks(NULL, &tmp);
1967 * Frequency meter - How fast is some event occurring?
1969 * These routines manage a digitally filtered, constant time based,
1970 * event frequency meter. There are four routines:
1971 * fmeter_init() - initialize a frequency meter.
1972 * fmeter_markevent() - called each time the event happens.
1973 * fmeter_getrate() - returns the recent rate of such events.
1974 * fmeter_update() - internal routine used to update fmeter.
1976 * A common data structure is passed to each of these routines,
1977 * which is used to keep track of the state required to manage the
1978 * frequency meter and its digital filter.
1980 * The filter works on the number of events marked per unit time.
1981 * The filter is single-pole low-pass recursive (IIR). The time unit
1982 * is 1 second. Arithmetic is done using 32-bit integers scaled to
1983 * simulate 3 decimal digits of precision (multiplied by 1000).
1985 * With an FM_COEF of 933, and a time base of 1 second, the filter
1986 * has a half-life of 10 seconds, meaning that if the events quit
1987 * happening, then the rate returned from the fmeter_getrate()
1988 * will be cut in half each 10 seconds, until it converges to zero.
1990 * It is not worth doing a real infinitely recursive filter. If more
1991 * than FM_MAXTICKS ticks have elapsed since the last filter event,
1992 * just compute FM_MAXTICKS ticks worth, by which point the level
1995 * Limit the count of unprocessed events to FM_MAXCNT, so as to avoid
1996 * arithmetic overflow in the fmeter_update() routine.
1998 * Given the simple 32 bit integer arithmetic used, this meter works
1999 * best for reporting rates between one per millisecond (msec) and
2000 * one per 32 (approx) seconds. At constant rates faster than one
2001 * per msec it maxes out at values just under 1,000,000. At constant
2002 * rates between one per msec, and one per second it will stabilize
2003 * to a value N*1000, where N is the rate of events per second.
2004 * At constant rates between one per second and one per 32 seconds,
2005 * it will be choppy, moving up on the seconds that have an event,
2006 * and then decaying until the next event. At rates slower than
2007 * about one in 32 seconds, it decays all the way back to zero between
2011 #define FM_COEF 933 /* coefficient for half-life of 10 secs */
2012 #define FM_MAXTICKS ((u32)99) /* useless computing more ticks than this */
2013 #define FM_MAXCNT 1000000 /* limit cnt to avoid overflow */
2014 #define FM_SCALE 1000 /* faux fixed point scale */
2016 /* Initialize a frequency meter */
2017 static void fmeter_init(struct fmeter *fmp)
2022 spin_lock_init(&fmp->lock);
2025 /* Internal meter update - process cnt events and update value */
2026 static void fmeter_update(struct fmeter *fmp)
2031 now = ktime_get_seconds();
2032 ticks = now - fmp->time;
2037 ticks = min(FM_MAXTICKS, ticks);
2039 fmp->val = (FM_COEF * fmp->val) / FM_SCALE;
2042 fmp->val += ((FM_SCALE - FM_COEF) * fmp->cnt) / FM_SCALE;
2046 /* Process any previous ticks, then bump cnt by one (times scale). */
2047 static void fmeter_markevent(struct fmeter *fmp)
2049 spin_lock(&fmp->lock);
2051 fmp->cnt = min(FM_MAXCNT, fmp->cnt + FM_SCALE);
2052 spin_unlock(&fmp->lock);
2055 /* Process any previous ticks, then return current value. */
2056 static int fmeter_getrate(struct fmeter *fmp)
2060 spin_lock(&fmp->lock);
2063 spin_unlock(&fmp->lock);
2067 static struct cpuset *cpuset_attach_old_cs;
2069 /* Called by cgroups to determine if a cpuset is usable; cpuset_mutex held */
2070 static int cpuset_can_attach(struct cgroup_taskset *tset)
2072 struct cgroup_subsys_state *css;
2074 struct task_struct *task;
2077 /* used later by cpuset_attach() */
2078 cpuset_attach_old_cs = task_cs(cgroup_taskset_first(tset, &css));
2081 mutex_lock(&cpuset_mutex);
2083 /* allow moving tasks into an empty cpuset if on default hierarchy */
2085 if (!is_in_v2_mode() &&
2086 (cpumask_empty(cs->cpus_allowed) || nodes_empty(cs->mems_allowed)))
2089 cgroup_taskset_for_each(task, css, tset) {
2090 ret = task_can_attach(task, cs->cpus_allowed);
2093 ret = security_task_setscheduler(task);
2099 * Mark attach is in progress. This makes validate_change() fail
2100 * changes which zero cpus/mems_allowed.
2102 cs->attach_in_progress++;
2105 mutex_unlock(&cpuset_mutex);
2109 static void cpuset_cancel_attach(struct cgroup_taskset *tset)
2111 struct cgroup_subsys_state *css;
2113 cgroup_taskset_first(tset, &css);
2115 mutex_lock(&cpuset_mutex);
2116 css_cs(css)->attach_in_progress--;
2117 mutex_unlock(&cpuset_mutex);
2121 * Protected by cpuset_mutex. cpus_attach is used only by cpuset_attach()
2122 * but we can't allocate it dynamically there. Define it global and
2123 * allocate from cpuset_init().
2125 static cpumask_var_t cpus_attach;
2127 static void cpuset_attach(struct cgroup_taskset *tset)
2129 /* static buf protected by cpuset_mutex */
2130 static nodemask_t cpuset_attach_nodemask_to;
2131 struct task_struct *task;
2132 struct task_struct *leader;
2133 struct cgroup_subsys_state *css;
2135 struct cpuset *oldcs = cpuset_attach_old_cs;
2137 cgroup_taskset_first(tset, &css);
2140 mutex_lock(&cpuset_mutex);
2142 /* prepare for attach */
2143 if (cs == &top_cpuset)
2144 cpumask_copy(cpus_attach, cpu_possible_mask);
2146 guarantee_online_cpus(cs, cpus_attach);
2148 guarantee_online_mems(cs, &cpuset_attach_nodemask_to);
2150 cgroup_taskset_for_each(task, css, tset) {
2152 * can_attach beforehand should guarantee that this doesn't
2153 * fail. TODO: have a better way to handle failure here
2155 WARN_ON_ONCE(set_cpus_allowed_ptr(task, cpus_attach));
2157 cpuset_change_task_nodemask(task, &cpuset_attach_nodemask_to);
2158 cpuset_update_task_spread_flag(cs, task);
2162 * Change mm for all threadgroup leaders. This is expensive and may
2163 * sleep and should be moved outside migration path proper.
2165 cpuset_attach_nodemask_to = cs->effective_mems;
2166 cgroup_taskset_for_each_leader(leader, css, tset) {
2167 struct mm_struct *mm = get_task_mm(leader);
2170 mpol_rebind_mm(mm, &cpuset_attach_nodemask_to);
2173 * old_mems_allowed is the same with mems_allowed
2174 * here, except if this task is being moved
2175 * automatically due to hotplug. In that case
2176 * @mems_allowed has been updated and is empty, so
2177 * @old_mems_allowed is the right nodesets that we
2180 if (is_memory_migrate(cs))
2181 cpuset_migrate_mm(mm, &oldcs->old_mems_allowed,
2182 &cpuset_attach_nodemask_to);
2188 cs->old_mems_allowed = cpuset_attach_nodemask_to;
2190 cs->attach_in_progress--;
2191 if (!cs->attach_in_progress)
2192 wake_up(&cpuset_attach_wq);
2194 mutex_unlock(&cpuset_mutex);
2197 /* The various types of files and directories in a cpuset file system */
2200 FILE_MEMORY_MIGRATE,
2203 FILE_EFFECTIVE_CPULIST,
2204 FILE_EFFECTIVE_MEMLIST,
2205 FILE_SUBPARTS_CPULIST,
2209 FILE_SCHED_LOAD_BALANCE,
2210 FILE_PARTITION_ROOT,
2211 FILE_SCHED_RELAX_DOMAIN_LEVEL,
2212 FILE_MEMORY_PRESSURE_ENABLED,
2213 FILE_MEMORY_PRESSURE,
2216 } cpuset_filetype_t;
2218 static int cpuset_write_u64(struct cgroup_subsys_state *css, struct cftype *cft,
2221 struct cpuset *cs = css_cs(css);
2222 cpuset_filetype_t type = cft->private;
2225 mutex_lock(&cpuset_mutex);
2226 if (!is_cpuset_online(cs)) {
2232 case FILE_CPU_EXCLUSIVE:
2233 retval = update_flag(CS_CPU_EXCLUSIVE, cs, val);
2235 case FILE_MEM_EXCLUSIVE:
2236 retval = update_flag(CS_MEM_EXCLUSIVE, cs, val);
2238 case FILE_MEM_HARDWALL:
2239 retval = update_flag(CS_MEM_HARDWALL, cs, val);
2241 case FILE_SCHED_LOAD_BALANCE:
2242 retval = update_flag(CS_SCHED_LOAD_BALANCE, cs, val);
2244 case FILE_MEMORY_MIGRATE:
2245 retval = update_flag(CS_MEMORY_MIGRATE, cs, val);
2247 case FILE_MEMORY_PRESSURE_ENABLED:
2248 cpuset_memory_pressure_enabled = !!val;
2250 case FILE_SPREAD_PAGE:
2251 retval = update_flag(CS_SPREAD_PAGE, cs, val);
2253 case FILE_SPREAD_SLAB:
2254 retval = update_flag(CS_SPREAD_SLAB, cs, val);
2261 mutex_unlock(&cpuset_mutex);
2265 static int cpuset_write_s64(struct cgroup_subsys_state *css, struct cftype *cft,
2268 struct cpuset *cs = css_cs(css);
2269 cpuset_filetype_t type = cft->private;
2270 int retval = -ENODEV;
2272 mutex_lock(&cpuset_mutex);
2273 if (!is_cpuset_online(cs))
2277 case FILE_SCHED_RELAX_DOMAIN_LEVEL:
2278 retval = update_relax_domain_level(cs, val);
2285 mutex_unlock(&cpuset_mutex);
2290 * Common handling for a write to a "cpus" or "mems" file.
2292 static ssize_t cpuset_write_resmask(struct kernfs_open_file *of,
2293 char *buf, size_t nbytes, loff_t off)
2295 struct cpuset *cs = css_cs(of_css(of));
2296 struct cpuset *trialcs;
2297 int retval = -ENODEV;
2299 buf = strstrip(buf);
2302 * CPU or memory hotunplug may leave @cs w/o any execution
2303 * resources, in which case the hotplug code asynchronously updates
2304 * configuration and transfers all tasks to the nearest ancestor
2305 * which can execute.
2307 * As writes to "cpus" or "mems" may restore @cs's execution
2308 * resources, wait for the previously scheduled operations before
2309 * proceeding, so that we don't end up keep removing tasks added
2310 * after execution capability is restored.
2312 * cpuset_hotplug_work calls back into cgroup core via
2313 * cgroup_transfer_tasks() and waiting for it from a cgroupfs
2314 * operation like this one can lead to a deadlock through kernfs
2315 * active_ref protection. Let's break the protection. Losing the
2316 * protection is okay as we check whether @cs is online after
2317 * grabbing cpuset_mutex anyway. This only happens on the legacy
2321 kernfs_break_active_protection(of->kn);
2322 flush_work(&cpuset_hotplug_work);
2324 mutex_lock(&cpuset_mutex);
2325 if (!is_cpuset_online(cs))
2328 trialcs = alloc_trial_cpuset(cs);
2334 switch (of_cft(of)->private) {
2336 retval = update_cpumask(cs, trialcs, buf);
2339 retval = update_nodemask(cs, trialcs, buf);
2346 free_cpuset(trialcs);
2348 mutex_unlock(&cpuset_mutex);
2349 kernfs_unbreak_active_protection(of->kn);
2351 flush_workqueue(cpuset_migrate_mm_wq);
2352 return retval ?: nbytes;
2356 * These ascii lists should be read in a single call, by using a user
2357 * buffer large enough to hold the entire map. If read in smaller
2358 * chunks, there is no guarantee of atomicity. Since the display format
2359 * used, list of ranges of sequential numbers, is variable length,
2360 * and since these maps can change value dynamically, one could read
2361 * gibberish by doing partial reads while a list was changing.
2363 static int cpuset_common_seq_show(struct seq_file *sf, void *v)
2365 struct cpuset *cs = css_cs(seq_css(sf));
2366 cpuset_filetype_t type = seq_cft(sf)->private;
2369 spin_lock_irq(&callback_lock);
2373 seq_printf(sf, "%*pbl\n", cpumask_pr_args(cs->cpus_allowed));
2376 seq_printf(sf, "%*pbl\n", nodemask_pr_args(&cs->mems_allowed));
2378 case FILE_EFFECTIVE_CPULIST:
2379 seq_printf(sf, "%*pbl\n", cpumask_pr_args(cs->effective_cpus));
2381 case FILE_EFFECTIVE_MEMLIST:
2382 seq_printf(sf, "%*pbl\n", nodemask_pr_args(&cs->effective_mems));
2384 case FILE_SUBPARTS_CPULIST:
2385 seq_printf(sf, "%*pbl\n", cpumask_pr_args(cs->subparts_cpus));
2391 spin_unlock_irq(&callback_lock);
2395 static u64 cpuset_read_u64(struct cgroup_subsys_state *css, struct cftype *cft)
2397 struct cpuset *cs = css_cs(css);
2398 cpuset_filetype_t type = cft->private;
2400 case FILE_CPU_EXCLUSIVE:
2401 return is_cpu_exclusive(cs);
2402 case FILE_MEM_EXCLUSIVE:
2403 return is_mem_exclusive(cs);
2404 case FILE_MEM_HARDWALL:
2405 return is_mem_hardwall(cs);
2406 case FILE_SCHED_LOAD_BALANCE:
2407 return is_sched_load_balance(cs);
2408 case FILE_MEMORY_MIGRATE:
2409 return is_memory_migrate(cs);
2410 case FILE_MEMORY_PRESSURE_ENABLED:
2411 return cpuset_memory_pressure_enabled;
2412 case FILE_MEMORY_PRESSURE:
2413 return fmeter_getrate(&cs->fmeter);
2414 case FILE_SPREAD_PAGE:
2415 return is_spread_page(cs);
2416 case FILE_SPREAD_SLAB:
2417 return is_spread_slab(cs);
2422 /* Unreachable but makes gcc happy */
2426 static s64 cpuset_read_s64(struct cgroup_subsys_state *css, struct cftype *cft)
2428 struct cpuset *cs = css_cs(css);
2429 cpuset_filetype_t type = cft->private;
2431 case FILE_SCHED_RELAX_DOMAIN_LEVEL:
2432 return cs->relax_domain_level;
2437 /* Unrechable but makes gcc happy */
2441 static int sched_partition_show(struct seq_file *seq, void *v)
2443 struct cpuset *cs = css_cs(seq_css(seq));
2445 switch (cs->partition_root_state) {
2447 seq_puts(seq, "root\n");
2450 seq_puts(seq, "member\n");
2453 seq_puts(seq, "root invalid\n");
2459 static ssize_t sched_partition_write(struct kernfs_open_file *of, char *buf,
2460 size_t nbytes, loff_t off)
2462 struct cpuset *cs = css_cs(of_css(of));
2464 int retval = -ENODEV;
2466 buf = strstrip(buf);
2469 * Convert "root" to ENABLED, and convert "member" to DISABLED.
2471 if (!strcmp(buf, "root"))
2473 else if (!strcmp(buf, "member"))
2479 mutex_lock(&cpuset_mutex);
2480 if (!is_cpuset_online(cs))
2483 retval = update_prstate(cs, val);
2485 mutex_unlock(&cpuset_mutex);
2487 return retval ?: nbytes;
2491 * for the common functions, 'private' gives the type of file
2494 static struct cftype legacy_files[] = {
2497 .seq_show = cpuset_common_seq_show,
2498 .write = cpuset_write_resmask,
2499 .max_write_len = (100U + 6 * NR_CPUS),
2500 .private = FILE_CPULIST,
2505 .seq_show = cpuset_common_seq_show,
2506 .write = cpuset_write_resmask,
2507 .max_write_len = (100U + 6 * MAX_NUMNODES),
2508 .private = FILE_MEMLIST,
2512 .name = "effective_cpus",
2513 .seq_show = cpuset_common_seq_show,
2514 .private = FILE_EFFECTIVE_CPULIST,
2518 .name = "effective_mems",
2519 .seq_show = cpuset_common_seq_show,
2520 .private = FILE_EFFECTIVE_MEMLIST,
2524 .name = "cpu_exclusive",
2525 .read_u64 = cpuset_read_u64,
2526 .write_u64 = cpuset_write_u64,
2527 .private = FILE_CPU_EXCLUSIVE,
2531 .name = "mem_exclusive",
2532 .read_u64 = cpuset_read_u64,
2533 .write_u64 = cpuset_write_u64,
2534 .private = FILE_MEM_EXCLUSIVE,
2538 .name = "mem_hardwall",
2539 .read_u64 = cpuset_read_u64,
2540 .write_u64 = cpuset_write_u64,
2541 .private = FILE_MEM_HARDWALL,
2545 .name = "sched_load_balance",
2546 .read_u64 = cpuset_read_u64,
2547 .write_u64 = cpuset_write_u64,
2548 .private = FILE_SCHED_LOAD_BALANCE,
2552 .name = "sched_relax_domain_level",
2553 .read_s64 = cpuset_read_s64,
2554 .write_s64 = cpuset_write_s64,
2555 .private = FILE_SCHED_RELAX_DOMAIN_LEVEL,
2559 .name = "memory_migrate",
2560 .read_u64 = cpuset_read_u64,
2561 .write_u64 = cpuset_write_u64,
2562 .private = FILE_MEMORY_MIGRATE,
2566 .name = "memory_pressure",
2567 .read_u64 = cpuset_read_u64,
2568 .private = FILE_MEMORY_PRESSURE,
2572 .name = "memory_spread_page",
2573 .read_u64 = cpuset_read_u64,
2574 .write_u64 = cpuset_write_u64,
2575 .private = FILE_SPREAD_PAGE,
2579 .name = "memory_spread_slab",
2580 .read_u64 = cpuset_read_u64,
2581 .write_u64 = cpuset_write_u64,
2582 .private = FILE_SPREAD_SLAB,
2586 .name = "memory_pressure_enabled",
2587 .flags = CFTYPE_ONLY_ON_ROOT,
2588 .read_u64 = cpuset_read_u64,
2589 .write_u64 = cpuset_write_u64,
2590 .private = FILE_MEMORY_PRESSURE_ENABLED,
2597 * This is currently a minimal set for the default hierarchy. It can be
2598 * expanded later on by migrating more features and control files from v1.
2600 static struct cftype dfl_files[] = {
2603 .seq_show = cpuset_common_seq_show,
2604 .write = cpuset_write_resmask,
2605 .max_write_len = (100U + 6 * NR_CPUS),
2606 .private = FILE_CPULIST,
2607 .flags = CFTYPE_NOT_ON_ROOT,
2612 .seq_show = cpuset_common_seq_show,
2613 .write = cpuset_write_resmask,
2614 .max_write_len = (100U + 6 * MAX_NUMNODES),
2615 .private = FILE_MEMLIST,
2616 .flags = CFTYPE_NOT_ON_ROOT,
2620 .name = "cpus.effective",
2621 .seq_show = cpuset_common_seq_show,
2622 .private = FILE_EFFECTIVE_CPULIST,
2626 .name = "mems.effective",
2627 .seq_show = cpuset_common_seq_show,
2628 .private = FILE_EFFECTIVE_MEMLIST,
2632 .name = "cpus.partition",
2633 .seq_show = sched_partition_show,
2634 .write = sched_partition_write,
2635 .private = FILE_PARTITION_ROOT,
2636 .flags = CFTYPE_NOT_ON_ROOT,
2640 .name = "cpus.subpartitions",
2641 .seq_show = cpuset_common_seq_show,
2642 .private = FILE_SUBPARTS_CPULIST,
2643 .flags = CFTYPE_DEBUG,
2651 * cpuset_css_alloc - allocate a cpuset css
2652 * cgrp: control group that the new cpuset will be part of
2655 static struct cgroup_subsys_state *
2656 cpuset_css_alloc(struct cgroup_subsys_state *parent_css)
2661 return &top_cpuset.css;
2663 cs = kzalloc(sizeof(*cs), GFP_KERNEL);
2665 return ERR_PTR(-ENOMEM);
2667 if (alloc_cpumasks(cs, NULL)) {
2669 return ERR_PTR(-ENOMEM);
2672 set_bit(CS_SCHED_LOAD_BALANCE, &cs->flags);
2673 nodes_clear(cs->mems_allowed);
2674 nodes_clear(cs->effective_mems);
2675 fmeter_init(&cs->fmeter);
2676 cs->relax_domain_level = -1;
2681 static int cpuset_css_online(struct cgroup_subsys_state *css)
2683 struct cpuset *cs = css_cs(css);
2684 struct cpuset *parent = parent_cs(cs);
2685 struct cpuset *tmp_cs;
2686 struct cgroup_subsys_state *pos_css;
2691 mutex_lock(&cpuset_mutex);
2693 set_bit(CS_ONLINE, &cs->flags);
2694 if (is_spread_page(parent))
2695 set_bit(CS_SPREAD_PAGE, &cs->flags);
2696 if (is_spread_slab(parent))
2697 set_bit(CS_SPREAD_SLAB, &cs->flags);
2701 spin_lock_irq(&callback_lock);
2702 if (is_in_v2_mode()) {
2703 cpumask_copy(cs->effective_cpus, parent->effective_cpus);
2704 cs->effective_mems = parent->effective_mems;
2705 cs->use_parent_ecpus = true;
2706 parent->child_ecpus_count++;
2708 spin_unlock_irq(&callback_lock);
2710 if (!test_bit(CGRP_CPUSET_CLONE_CHILDREN, &css->cgroup->flags))
2714 * Clone @parent's configuration if CGRP_CPUSET_CLONE_CHILDREN is
2715 * set. This flag handling is implemented in cgroup core for
2716 * histrical reasons - the flag may be specified during mount.
2718 * Currently, if any sibling cpusets have exclusive cpus or mem, we
2719 * refuse to clone the configuration - thereby refusing the task to
2720 * be entered, and as a result refusing the sys_unshare() or
2721 * clone() which initiated it. If this becomes a problem for some
2722 * users who wish to allow that scenario, then this could be
2723 * changed to grant parent->cpus_allowed-sibling_cpus_exclusive
2724 * (and likewise for mems) to the new cgroup.
2727 cpuset_for_each_child(tmp_cs, pos_css, parent) {
2728 if (is_mem_exclusive(tmp_cs) || is_cpu_exclusive(tmp_cs)) {
2735 spin_lock_irq(&callback_lock);
2736 cs->mems_allowed = parent->mems_allowed;
2737 cs->effective_mems = parent->mems_allowed;
2738 cpumask_copy(cs->cpus_allowed, parent->cpus_allowed);
2739 cpumask_copy(cs->effective_cpus, parent->cpus_allowed);
2740 spin_unlock_irq(&callback_lock);
2742 mutex_unlock(&cpuset_mutex);
2747 * If the cpuset being removed has its flag 'sched_load_balance'
2748 * enabled, then simulate turning sched_load_balance off, which
2749 * will call rebuild_sched_domains_locked(). That is not needed
2750 * in the default hierarchy where only changes in partition
2751 * will cause repartitioning.
2753 * If the cpuset has the 'sched.partition' flag enabled, simulate
2754 * turning 'sched.partition" off.
2757 static void cpuset_css_offline(struct cgroup_subsys_state *css)
2759 struct cpuset *cs = css_cs(css);
2761 mutex_lock(&cpuset_mutex);
2763 if (is_partition_root(cs))
2764 update_prstate(cs, 0);
2766 if (!cgroup_subsys_on_dfl(cpuset_cgrp_subsys) &&
2767 is_sched_load_balance(cs))
2768 update_flag(CS_SCHED_LOAD_BALANCE, cs, 0);
2770 if (cs->use_parent_ecpus) {
2771 struct cpuset *parent = parent_cs(cs);
2773 cs->use_parent_ecpus = false;
2774 parent->child_ecpus_count--;
2778 clear_bit(CS_ONLINE, &cs->flags);
2780 mutex_unlock(&cpuset_mutex);
2783 static void cpuset_css_free(struct cgroup_subsys_state *css)
2785 struct cpuset *cs = css_cs(css);
2790 static void cpuset_bind(struct cgroup_subsys_state *root_css)
2792 mutex_lock(&cpuset_mutex);
2793 spin_lock_irq(&callback_lock);
2795 if (is_in_v2_mode()) {
2796 cpumask_copy(top_cpuset.cpus_allowed, cpu_possible_mask);
2797 top_cpuset.mems_allowed = node_possible_map;
2799 cpumask_copy(top_cpuset.cpus_allowed,
2800 top_cpuset.effective_cpus);
2801 top_cpuset.mems_allowed = top_cpuset.effective_mems;
2804 spin_unlock_irq(&callback_lock);
2805 mutex_unlock(&cpuset_mutex);
2809 * Make sure the new task conform to the current state of its parent,
2810 * which could have been changed by cpuset just after it inherits the
2811 * state from the parent and before it sits on the cgroup's task list.
2813 static void cpuset_fork(struct task_struct *task)
2815 if (task_css_is_root(task, cpuset_cgrp_id))
2818 set_cpus_allowed_ptr(task, ¤t->cpus_allowed);
2819 task->mems_allowed = current->mems_allowed;
2822 struct cgroup_subsys cpuset_cgrp_subsys = {
2823 .css_alloc = cpuset_css_alloc,
2824 .css_online = cpuset_css_online,
2825 .css_offline = cpuset_css_offline,
2826 .css_free = cpuset_css_free,
2827 .can_attach = cpuset_can_attach,
2828 .cancel_attach = cpuset_cancel_attach,
2829 .attach = cpuset_attach,
2830 .post_attach = cpuset_post_attach,
2831 .bind = cpuset_bind,
2832 .fork = cpuset_fork,
2833 .legacy_cftypes = legacy_files,
2834 .dfl_cftypes = dfl_files,
2840 * cpuset_init - initialize cpusets at system boot
2842 * Description: Initialize top_cpuset and the cpuset internal file system,
2845 int __init cpuset_init(void)
2849 BUG_ON(!alloc_cpumask_var(&top_cpuset.cpus_allowed, GFP_KERNEL));
2850 BUG_ON(!alloc_cpumask_var(&top_cpuset.effective_cpus, GFP_KERNEL));
2851 BUG_ON(!zalloc_cpumask_var(&top_cpuset.subparts_cpus, GFP_KERNEL));
2853 cpumask_setall(top_cpuset.cpus_allowed);
2854 nodes_setall(top_cpuset.mems_allowed);
2855 cpumask_setall(top_cpuset.effective_cpus);
2856 nodes_setall(top_cpuset.effective_mems);
2858 fmeter_init(&top_cpuset.fmeter);
2859 set_bit(CS_SCHED_LOAD_BALANCE, &top_cpuset.flags);
2860 top_cpuset.relax_domain_level = -1;
2862 err = register_filesystem(&cpuset_fs_type);
2866 BUG_ON(!alloc_cpumask_var(&cpus_attach, GFP_KERNEL));
2872 * If CPU and/or memory hotplug handlers, below, unplug any CPUs
2873 * or memory nodes, we need to walk over the cpuset hierarchy,
2874 * removing that CPU or node from all cpusets. If this removes the
2875 * last CPU or node from a cpuset, then move the tasks in the empty
2876 * cpuset to its next-highest non-empty parent.
2878 static void remove_tasks_in_empty_cpuset(struct cpuset *cs)
2880 struct cpuset *parent;
2883 * Find its next-highest non-empty parent, (top cpuset
2884 * has online cpus, so can't be empty).
2886 parent = parent_cs(cs);
2887 while (cpumask_empty(parent->cpus_allowed) ||
2888 nodes_empty(parent->mems_allowed))
2889 parent = parent_cs(parent);
2891 if (cgroup_transfer_tasks(parent->css.cgroup, cs->css.cgroup)) {
2892 pr_err("cpuset: failed to transfer tasks out of empty cpuset ");
2893 pr_cont_cgroup_name(cs->css.cgroup);
2899 hotplug_update_tasks_legacy(struct cpuset *cs,
2900 struct cpumask *new_cpus, nodemask_t *new_mems,
2901 bool cpus_updated, bool mems_updated)
2905 spin_lock_irq(&callback_lock);
2906 cpumask_copy(cs->cpus_allowed, new_cpus);
2907 cpumask_copy(cs->effective_cpus, new_cpus);
2908 cs->mems_allowed = *new_mems;
2909 cs->effective_mems = *new_mems;
2910 spin_unlock_irq(&callback_lock);
2913 * Don't call update_tasks_cpumask() if the cpuset becomes empty,
2914 * as the tasks will be migratecd to an ancestor.
2916 if (cpus_updated && !cpumask_empty(cs->cpus_allowed))
2917 update_tasks_cpumask(cs);
2918 if (mems_updated && !nodes_empty(cs->mems_allowed))
2919 update_tasks_nodemask(cs);
2921 is_empty = cpumask_empty(cs->cpus_allowed) ||
2922 nodes_empty(cs->mems_allowed);
2924 mutex_unlock(&cpuset_mutex);
2927 * Move tasks to the nearest ancestor with execution resources,
2928 * This is full cgroup operation which will also call back into
2929 * cpuset. Should be done outside any lock.
2932 remove_tasks_in_empty_cpuset(cs);
2934 mutex_lock(&cpuset_mutex);
2938 hotplug_update_tasks(struct cpuset *cs,
2939 struct cpumask *new_cpus, nodemask_t *new_mems,
2940 bool cpus_updated, bool mems_updated)
2942 if (cpumask_empty(new_cpus))
2943 cpumask_copy(new_cpus, parent_cs(cs)->effective_cpus);
2944 if (nodes_empty(*new_mems))
2945 *new_mems = parent_cs(cs)->effective_mems;
2947 spin_lock_irq(&callback_lock);
2948 cpumask_copy(cs->effective_cpus, new_cpus);
2949 cs->effective_mems = *new_mems;
2950 spin_unlock_irq(&callback_lock);
2953 update_tasks_cpumask(cs);
2955 update_tasks_nodemask(cs);
2958 static bool force_rebuild;
2960 void cpuset_force_rebuild(void)
2962 force_rebuild = true;
2966 * cpuset_hotplug_update_tasks - update tasks in a cpuset for hotunplug
2967 * @cs: cpuset in interest
2968 * @tmp: the tmpmasks structure pointer
2970 * Compare @cs's cpu and mem masks against top_cpuset and if some have gone
2971 * offline, update @cs accordingly. If @cs ends up with no CPU or memory,
2972 * all its tasks are moved to the nearest ancestor with both resources.
2974 static void cpuset_hotplug_update_tasks(struct cpuset *cs, struct tmpmasks *tmp)
2976 static cpumask_t new_cpus;
2977 static nodemask_t new_mems;
2980 struct cpuset *parent;
2982 wait_event(cpuset_attach_wq, cs->attach_in_progress == 0);
2984 mutex_lock(&cpuset_mutex);
2987 * We have raced with task attaching. We wait until attaching
2988 * is finished, so we won't attach a task to an empty cpuset.
2990 if (cs->attach_in_progress) {
2991 mutex_unlock(&cpuset_mutex);
2995 parent = parent_cs(cs);
2996 compute_effective_cpumask(&new_cpus, cs, parent);
2997 nodes_and(new_mems, cs->mems_allowed, parent->effective_mems);
2999 if (cs->nr_subparts_cpus)
3001 * Make sure that CPUs allocated to child partitions
3002 * do not show up in effective_cpus.
3004 cpumask_andnot(&new_cpus, &new_cpus, cs->subparts_cpus);
3006 if (!tmp || !cs->partition_root_state)
3010 * In the unlikely event that a partition root has empty
3011 * effective_cpus or its parent becomes erroneous, we have to
3012 * transition it to the erroneous state.
3014 if (is_partition_root(cs) && (cpumask_empty(&new_cpus) ||
3015 (parent->partition_root_state == PRS_ERROR))) {
3016 if (cs->nr_subparts_cpus) {
3017 cs->nr_subparts_cpus = 0;
3018 cpumask_clear(cs->subparts_cpus);
3019 compute_effective_cpumask(&new_cpus, cs, parent);
3023 * If the effective_cpus is empty because the child
3024 * partitions take away all the CPUs, we can keep
3025 * the current partition and let the child partitions
3026 * fight for available CPUs.
3028 if ((parent->partition_root_state == PRS_ERROR) ||
3029 cpumask_empty(&new_cpus)) {
3030 update_parent_subparts_cpumask(cs, partcmd_disable,
3032 cs->partition_root_state = PRS_ERROR;
3034 cpuset_force_rebuild();
3038 * On the other hand, an erroneous partition root may be transitioned
3039 * back to a regular one or a partition root with no CPU allocated
3040 * from the parent may change to erroneous.
3042 if (is_partition_root(parent) &&
3043 ((cs->partition_root_state == PRS_ERROR) ||
3044 !cpumask_intersects(&new_cpus, parent->subparts_cpus)) &&
3045 update_parent_subparts_cpumask(cs, partcmd_update, NULL, tmp))
3046 cpuset_force_rebuild();
3049 cpus_updated = !cpumask_equal(&new_cpus, cs->effective_cpus);
3050 mems_updated = !nodes_equal(new_mems, cs->effective_mems);
3052 if (is_in_v2_mode())
3053 hotplug_update_tasks(cs, &new_cpus, &new_mems,
3054 cpus_updated, mems_updated);
3056 hotplug_update_tasks_legacy(cs, &new_cpus, &new_mems,
3057 cpus_updated, mems_updated);
3059 mutex_unlock(&cpuset_mutex);
3063 * cpuset_hotplug_workfn - handle CPU/memory hotunplug for a cpuset
3065 * This function is called after either CPU or memory configuration has
3066 * changed and updates cpuset accordingly. The top_cpuset is always
3067 * synchronized to cpu_active_mask and N_MEMORY, which is necessary in
3068 * order to make cpusets transparent (of no affect) on systems that are
3069 * actively using CPU hotplug but making no active use of cpusets.
3071 * Non-root cpusets are only affected by offlining. If any CPUs or memory
3072 * nodes have been taken down, cpuset_hotplug_update_tasks() is invoked on
3075 * Note that CPU offlining during suspend is ignored. We don't modify
3076 * cpusets across suspend/resume cycles at all.
3078 static void cpuset_hotplug_workfn(struct work_struct *work)
3080 static cpumask_t new_cpus;
3081 static nodemask_t new_mems;
3082 bool cpus_updated, mems_updated;
3083 bool on_dfl = is_in_v2_mode();
3084 struct tmpmasks tmp, *ptmp = NULL;
3086 if (on_dfl && !alloc_cpumasks(NULL, &tmp))
3089 mutex_lock(&cpuset_mutex);
3091 /* fetch the available cpus/mems and find out which changed how */
3092 cpumask_copy(&new_cpus, cpu_active_mask);
3093 new_mems = node_states[N_MEMORY];
3096 * If subparts_cpus is populated, it is likely that the check below
3097 * will produce a false positive on cpus_updated when the cpu list
3098 * isn't changed. It is extra work, but it is better to be safe.
3100 cpus_updated = !cpumask_equal(top_cpuset.effective_cpus, &new_cpus);
3101 mems_updated = !nodes_equal(top_cpuset.effective_mems, new_mems);
3103 /* synchronize cpus_allowed to cpu_active_mask */
3105 spin_lock_irq(&callback_lock);
3107 cpumask_copy(top_cpuset.cpus_allowed, &new_cpus);
3109 * Make sure that CPUs allocated to child partitions
3110 * do not show up in effective_cpus. If no CPU is left,
3111 * we clear the subparts_cpus & let the child partitions
3112 * fight for the CPUs again.
3114 if (top_cpuset.nr_subparts_cpus) {
3115 if (cpumask_subset(&new_cpus,
3116 top_cpuset.subparts_cpus)) {
3117 top_cpuset.nr_subparts_cpus = 0;
3118 cpumask_clear(top_cpuset.subparts_cpus);
3120 cpumask_andnot(&new_cpus, &new_cpus,
3121 top_cpuset.subparts_cpus);
3124 cpumask_copy(top_cpuset.effective_cpus, &new_cpus);
3125 spin_unlock_irq(&callback_lock);
3126 /* we don't mess with cpumasks of tasks in top_cpuset */
3129 /* synchronize mems_allowed to N_MEMORY */
3131 spin_lock_irq(&callback_lock);
3133 top_cpuset.mems_allowed = new_mems;
3134 top_cpuset.effective_mems = new_mems;
3135 spin_unlock_irq(&callback_lock);
3136 update_tasks_nodemask(&top_cpuset);
3139 mutex_unlock(&cpuset_mutex);
3141 /* if cpus or mems changed, we need to propagate to descendants */
3142 if (cpus_updated || mems_updated) {
3144 struct cgroup_subsys_state *pos_css;
3147 cpuset_for_each_descendant_pre(cs, pos_css, &top_cpuset) {
3148 if (cs == &top_cpuset || !css_tryget_online(&cs->css))
3152 cpuset_hotplug_update_tasks(cs, ptmp);
3160 /* rebuild sched domains if cpus_allowed has changed */
3161 if (cpus_updated || force_rebuild) {
3162 force_rebuild = false;
3163 rebuild_sched_domains();
3166 free_cpumasks(NULL, ptmp);
3169 void cpuset_update_active_cpus(void)
3172 * We're inside cpu hotplug critical region which usually nests
3173 * inside cgroup synchronization. Bounce actual hotplug processing
3174 * to a work item to avoid reverse locking order.
3176 schedule_work(&cpuset_hotplug_work);
3179 void cpuset_wait_for_hotplug(void)
3181 flush_work(&cpuset_hotplug_work);
3185 * Keep top_cpuset.mems_allowed tracking node_states[N_MEMORY].
3186 * Call this routine anytime after node_states[N_MEMORY] changes.
3187 * See cpuset_update_active_cpus() for CPU hotplug handling.
3189 static int cpuset_track_online_nodes(struct notifier_block *self,
3190 unsigned long action, void *arg)
3192 schedule_work(&cpuset_hotplug_work);
3196 static struct notifier_block cpuset_track_online_nodes_nb = {
3197 .notifier_call = cpuset_track_online_nodes,
3198 .priority = 10, /* ??! */
3202 * cpuset_init_smp - initialize cpus_allowed
3204 * Description: Finish top cpuset after cpu, node maps are initialized
3206 void __init cpuset_init_smp(void)
3208 cpumask_copy(top_cpuset.cpus_allowed, cpu_active_mask);
3209 top_cpuset.mems_allowed = node_states[N_MEMORY];
3210 top_cpuset.old_mems_allowed = top_cpuset.mems_allowed;
3212 cpumask_copy(top_cpuset.effective_cpus, cpu_active_mask);
3213 top_cpuset.effective_mems = node_states[N_MEMORY];
3215 register_hotmemory_notifier(&cpuset_track_online_nodes_nb);
3217 cpuset_migrate_mm_wq = alloc_ordered_workqueue("cpuset_migrate_mm", 0);
3218 BUG_ON(!cpuset_migrate_mm_wq);
3222 * cpuset_cpus_allowed - return cpus_allowed mask from a tasks cpuset.
3223 * @tsk: pointer to task_struct from which to obtain cpuset->cpus_allowed.
3224 * @pmask: pointer to struct cpumask variable to receive cpus_allowed set.
3226 * Description: Returns the cpumask_var_t cpus_allowed of the cpuset
3227 * attached to the specified @tsk. Guaranteed to return some non-empty
3228 * subset of cpu_online_mask, even if this means going outside the
3232 void cpuset_cpus_allowed(struct task_struct *tsk, struct cpumask *pmask)
3234 unsigned long flags;
3236 spin_lock_irqsave(&callback_lock, flags);
3238 guarantee_online_cpus(task_cs(tsk), pmask);
3240 spin_unlock_irqrestore(&callback_lock, flags);
3243 void cpuset_cpus_allowed_fallback(struct task_struct *tsk)
3246 do_set_cpus_allowed(tsk, task_cs(tsk)->effective_cpus);
3250 * We own tsk->cpus_allowed, nobody can change it under us.
3252 * But we used cs && cs->cpus_allowed lockless and thus can
3253 * race with cgroup_attach_task() or update_cpumask() and get
3254 * the wrong tsk->cpus_allowed. However, both cases imply the
3255 * subsequent cpuset_change_cpumask()->set_cpus_allowed_ptr()
3256 * which takes task_rq_lock().
3258 * If we are called after it dropped the lock we must see all
3259 * changes in tsk_cs()->cpus_allowed. Otherwise we can temporary
3260 * set any mask even if it is not right from task_cs() pov,
3261 * the pending set_cpus_allowed_ptr() will fix things.
3263 * select_fallback_rq() will fix things ups and set cpu_possible_mask
3268 void __init cpuset_init_current_mems_allowed(void)
3270 nodes_setall(current->mems_allowed);
3274 * cpuset_mems_allowed - return mems_allowed mask from a tasks cpuset.
3275 * @tsk: pointer to task_struct from which to obtain cpuset->mems_allowed.
3277 * Description: Returns the nodemask_t mems_allowed of the cpuset
3278 * attached to the specified @tsk. Guaranteed to return some non-empty
3279 * subset of node_states[N_MEMORY], even if this means going outside the
3283 nodemask_t cpuset_mems_allowed(struct task_struct *tsk)
3286 unsigned long flags;
3288 spin_lock_irqsave(&callback_lock, flags);
3290 guarantee_online_mems(task_cs(tsk), &mask);
3292 spin_unlock_irqrestore(&callback_lock, flags);
3298 * cpuset_nodemask_valid_mems_allowed - check nodemask vs. curremt mems_allowed
3299 * @nodemask: the nodemask to be checked
3301 * Are any of the nodes in the nodemask allowed in current->mems_allowed?
3303 int cpuset_nodemask_valid_mems_allowed(nodemask_t *nodemask)
3305 return nodes_intersects(*nodemask, current->mems_allowed);
3309 * nearest_hardwall_ancestor() - Returns the nearest mem_exclusive or
3310 * mem_hardwall ancestor to the specified cpuset. Call holding
3311 * callback_lock. If no ancestor is mem_exclusive or mem_hardwall
3312 * (an unusual configuration), then returns the root cpuset.
3314 static struct cpuset *nearest_hardwall_ancestor(struct cpuset *cs)
3316 while (!(is_mem_exclusive(cs) || is_mem_hardwall(cs)) && parent_cs(cs))
3322 * cpuset_node_allowed - Can we allocate on a memory node?
3323 * @node: is this an allowed node?
3324 * @gfp_mask: memory allocation flags
3326 * If we're in interrupt, yes, we can always allocate. If @node is set in
3327 * current's mems_allowed, yes. If it's not a __GFP_HARDWALL request and this
3328 * node is set in the nearest hardwalled cpuset ancestor to current's cpuset,
3329 * yes. If current has access to memory reserves as an oom victim, yes.
3332 * GFP_USER allocations are marked with the __GFP_HARDWALL bit,
3333 * and do not allow allocations outside the current tasks cpuset
3334 * unless the task has been OOM killed.
3335 * GFP_KERNEL allocations are not so marked, so can escape to the
3336 * nearest enclosing hardwalled ancestor cpuset.
3338 * Scanning up parent cpusets requires callback_lock. The
3339 * __alloc_pages() routine only calls here with __GFP_HARDWALL bit
3340 * _not_ set if it's a GFP_KERNEL allocation, and all nodes in the
3341 * current tasks mems_allowed came up empty on the first pass over
3342 * the zonelist. So only GFP_KERNEL allocations, if all nodes in the
3343 * cpuset are short of memory, might require taking the callback_lock.
3345 * The first call here from mm/page_alloc:get_page_from_freelist()
3346 * has __GFP_HARDWALL set in gfp_mask, enforcing hardwall cpusets,
3347 * so no allocation on a node outside the cpuset is allowed (unless
3348 * in interrupt, of course).
3350 * The second pass through get_page_from_freelist() doesn't even call
3351 * here for GFP_ATOMIC calls. For those calls, the __alloc_pages()
3352 * variable 'wait' is not set, and the bit ALLOC_CPUSET is not set
3353 * in alloc_flags. That logic and the checks below have the combined
3355 * in_interrupt - any node ok (current task context irrelevant)
3356 * GFP_ATOMIC - any node ok
3357 * tsk_is_oom_victim - any node ok
3358 * GFP_KERNEL - any node in enclosing hardwalled cpuset ok
3359 * GFP_USER - only nodes in current tasks mems allowed ok.
3361 bool __cpuset_node_allowed(int node, gfp_t gfp_mask)
3363 struct cpuset *cs; /* current cpuset ancestors */
3364 int allowed; /* is allocation in zone z allowed? */
3365 unsigned long flags;
3369 if (node_isset(node, current->mems_allowed))
3372 * Allow tasks that have access to memory reserves because they have
3373 * been OOM killed to get memory anywhere.
3375 if (unlikely(tsk_is_oom_victim(current)))
3377 if (gfp_mask & __GFP_HARDWALL) /* If hardwall request, stop here */
3380 if (current->flags & PF_EXITING) /* Let dying task have memory */
3383 /* Not hardwall and node outside mems_allowed: scan up cpusets */
3384 spin_lock_irqsave(&callback_lock, flags);
3387 cs = nearest_hardwall_ancestor(task_cs(current));
3388 allowed = node_isset(node, cs->mems_allowed);
3391 spin_unlock_irqrestore(&callback_lock, flags);
3396 * cpuset_mem_spread_node() - On which node to begin search for a file page
3397 * cpuset_slab_spread_node() - On which node to begin search for a slab page
3399 * If a task is marked PF_SPREAD_PAGE or PF_SPREAD_SLAB (as for
3400 * tasks in a cpuset with is_spread_page or is_spread_slab set),
3401 * and if the memory allocation used cpuset_mem_spread_node()
3402 * to determine on which node to start looking, as it will for
3403 * certain page cache or slab cache pages such as used for file
3404 * system buffers and inode caches, then instead of starting on the
3405 * local node to look for a free page, rather spread the starting
3406 * node around the tasks mems_allowed nodes.
3408 * We don't have to worry about the returned node being offline
3409 * because "it can't happen", and even if it did, it would be ok.
3411 * The routines calling guarantee_online_mems() are careful to
3412 * only set nodes in task->mems_allowed that are online. So it
3413 * should not be possible for the following code to return an
3414 * offline node. But if it did, that would be ok, as this routine
3415 * is not returning the node where the allocation must be, only
3416 * the node where the search should start. The zonelist passed to
3417 * __alloc_pages() will include all nodes. If the slab allocator
3418 * is passed an offline node, it will fall back to the local node.
3419 * See kmem_cache_alloc_node().
3422 static int cpuset_spread_node(int *rotor)
3424 return *rotor = next_node_in(*rotor, current->mems_allowed);
3427 int cpuset_mem_spread_node(void)
3429 if (current->cpuset_mem_spread_rotor == NUMA_NO_NODE)
3430 current->cpuset_mem_spread_rotor =
3431 node_random(¤t->mems_allowed);
3433 return cpuset_spread_node(¤t->cpuset_mem_spread_rotor);
3436 int cpuset_slab_spread_node(void)
3438 if (current->cpuset_slab_spread_rotor == NUMA_NO_NODE)
3439 current->cpuset_slab_spread_rotor =
3440 node_random(¤t->mems_allowed);
3442 return cpuset_spread_node(¤t->cpuset_slab_spread_rotor);
3445 EXPORT_SYMBOL_GPL(cpuset_mem_spread_node);
3448 * cpuset_mems_allowed_intersects - Does @tsk1's mems_allowed intersect @tsk2's?
3449 * @tsk1: pointer to task_struct of some task.
3450 * @tsk2: pointer to task_struct of some other task.
3452 * Description: Return true if @tsk1's mems_allowed intersects the
3453 * mems_allowed of @tsk2. Used by the OOM killer to determine if
3454 * one of the task's memory usage might impact the memory available
3458 int cpuset_mems_allowed_intersects(const struct task_struct *tsk1,
3459 const struct task_struct *tsk2)
3461 return nodes_intersects(tsk1->mems_allowed, tsk2->mems_allowed);
3465 * cpuset_print_current_mems_allowed - prints current's cpuset and mems_allowed
3467 * Description: Prints current's name, cpuset name, and cached copy of its
3468 * mems_allowed to the kernel log.
3470 void cpuset_print_current_mems_allowed(void)
3472 struct cgroup *cgrp;
3476 cgrp = task_cs(current)->css.cgroup;
3477 pr_cont(",cpuset=");
3478 pr_cont_cgroup_name(cgrp);
3479 pr_cont(",mems_allowed=%*pbl",
3480 nodemask_pr_args(¤t->mems_allowed));
3486 * Collection of memory_pressure is suppressed unless
3487 * this flag is enabled by writing "1" to the special
3488 * cpuset file 'memory_pressure_enabled' in the root cpuset.
3491 int cpuset_memory_pressure_enabled __read_mostly;
3494 * cpuset_memory_pressure_bump - keep stats of per-cpuset reclaims.
3496 * Keep a running average of the rate of synchronous (direct)
3497 * page reclaim efforts initiated by tasks in each cpuset.
3499 * This represents the rate at which some task in the cpuset
3500 * ran low on memory on all nodes it was allowed to use, and
3501 * had to enter the kernels page reclaim code in an effort to
3502 * create more free memory by tossing clean pages or swapping
3503 * or writing dirty pages.
3505 * Display to user space in the per-cpuset read-only file
3506 * "memory_pressure". Value displayed is an integer
3507 * representing the recent rate of entry into the synchronous
3508 * (direct) page reclaim by any task attached to the cpuset.
3511 void __cpuset_memory_pressure_bump(void)
3514 fmeter_markevent(&task_cs(current)->fmeter);
3518 #ifdef CONFIG_PROC_PID_CPUSET
3520 * proc_cpuset_show()
3521 * - Print tasks cpuset path into seq_file.
3522 * - Used for /proc/<pid>/cpuset.
3523 * - No need to task_lock(tsk) on this tsk->cpuset reference, as it
3524 * doesn't really matter if tsk->cpuset changes after we read it,
3525 * and we take cpuset_mutex, keeping cpuset_attach() from changing it
3528 int proc_cpuset_show(struct seq_file *m, struct pid_namespace *ns,
3529 struct pid *pid, struct task_struct *tsk)
3532 struct cgroup_subsys_state *css;
3536 buf = kmalloc(PATH_MAX, GFP_KERNEL);
3540 css = task_get_css(tsk, cpuset_cgrp_id);
3541 retval = cgroup_path_ns(css->cgroup, buf, PATH_MAX,
3542 current->nsproxy->cgroup_ns);
3544 if (retval >= PATH_MAX)
3545 retval = -ENAMETOOLONG;
3556 #endif /* CONFIG_PROC_PID_CPUSET */
3558 /* Display task mems_allowed in /proc/<pid>/status file. */
3559 void cpuset_task_status_allowed(struct seq_file *m, struct task_struct *task)
3561 seq_printf(m, "Mems_allowed:\t%*pb\n",
3562 nodemask_pr_args(&task->mems_allowed));
3563 seq_printf(m, "Mems_allowed_list:\t%*pbl\n",
3564 nodemask_pr_args(&task->mems_allowed));