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/init.h>
29 #include <linux/interrupt.h>
30 #include <linux/kernel.h>
31 #include <linux/mempolicy.h>
33 #include <linux/memory.h>
34 #include <linux/export.h>
35 #include <linux/rcupdate.h>
36 #include <linux/sched.h>
37 #include <linux/sched/deadline.h>
38 #include <linux/sched/mm.h>
39 #include <linux/sched/task.h>
40 #include <linux/security.h>
41 #include <linux/spinlock.h>
42 #include <linux/oom.h>
43 #include <linux/sched/isolation.h>
44 #include <linux/cgroup.h>
45 #include <linux/wait.h>
47 DEFINE_STATIC_KEY_FALSE(cpusets_pre_enable_key);
48 DEFINE_STATIC_KEY_FALSE(cpusets_enabled_key);
51 * There could be abnormal cpuset configurations for cpu or memory
52 * node binding, add this key to provide a quick low-cost judgment
55 DEFINE_STATIC_KEY_FALSE(cpusets_insane_config_key);
57 /* See "Frequency meter" comments, below. */
60 int cnt; /* unprocessed events count */
61 int val; /* most recent output value */
62 time64_t time; /* clock (secs) when val computed */
63 spinlock_t lock; /* guards read or write of above */
67 * Invalid partition error code
80 static const char * const perr_strings[] = {
81 [PERR_INVCPUS] = "Invalid cpu list in cpuset.cpus",
82 [PERR_INVPARENT] = "Parent is an invalid partition root",
83 [PERR_NOTPART] = "Parent is not a partition root",
84 [PERR_NOTEXCL] = "Cpu list in cpuset.cpus not exclusive",
85 [PERR_NOCPUS] = "Parent unable to distribute cpu downstream",
86 [PERR_HOTPLUG] = "No cpu available due to hotplug",
87 [PERR_CPUSEMPTY] = "cpuset.cpus is empty",
91 struct cgroup_subsys_state css;
93 unsigned long flags; /* "unsigned long" so bitops work */
96 * On default hierarchy:
98 * The user-configured masks can only be changed by writing to
99 * cpuset.cpus and cpuset.mems, and won't be limited by the
102 * The effective masks is the real masks that apply to the tasks
103 * in the cpuset. They may be changed if the configured masks are
104 * changed or hotplug happens.
106 * effective_mask == configured_mask & parent's effective_mask,
107 * and if it ends up empty, it will inherit the parent's mask.
110 * On legacy hierarchy:
112 * The user-configured masks are always the same with effective masks.
115 /* user-configured CPUs and Memory Nodes allow to tasks */
116 cpumask_var_t cpus_allowed;
117 nodemask_t mems_allowed;
119 /* effective CPUs and Memory Nodes allow to tasks */
120 cpumask_var_t effective_cpus;
121 nodemask_t effective_mems;
124 * CPUs allocated to child sub-partitions (default hierarchy only)
125 * - CPUs granted by the parent = effective_cpus U subparts_cpus
126 * - effective_cpus and subparts_cpus are mutually exclusive.
128 * effective_cpus contains only onlined CPUs, but subparts_cpus
129 * may have offlined ones.
131 cpumask_var_t subparts_cpus;
134 * This is old Memory Nodes tasks took on.
136 * - top_cpuset.old_mems_allowed is initialized to mems_allowed.
137 * - A new cpuset's old_mems_allowed is initialized when some
138 * task is moved into it.
139 * - old_mems_allowed is used in cpuset_migrate_mm() when we change
140 * cpuset.mems_allowed and have tasks' nodemask updated, and
141 * then old_mems_allowed is updated to mems_allowed.
143 nodemask_t old_mems_allowed;
145 struct fmeter fmeter; /* memory_pressure filter */
148 * Tasks are being attached to this cpuset. Used to prevent
149 * zeroing cpus/mems_allowed between ->can_attach() and ->attach().
151 int attach_in_progress;
153 /* partition number for rebuild_sched_domains() */
156 /* for custom sched domain */
157 int relax_domain_level;
159 /* number of CPUs in subparts_cpus */
160 int nr_subparts_cpus;
162 /* partition root state */
163 int partition_root_state;
166 * Default hierarchy only:
167 * use_parent_ecpus - set if using parent's effective_cpus
168 * child_ecpus_count - # of children with use_parent_ecpus set
170 int use_parent_ecpus;
171 int child_ecpus_count;
174 * number of SCHED_DEADLINE tasks attached to this cpuset, so that we
175 * know when to rebuild associated root domain bandwidth information.
177 int nr_deadline_tasks;
178 int nr_migrate_dl_tasks;
179 u64 sum_migrate_dl_bw;
181 /* Invalid partition error code, not lock protected */
182 enum prs_errcode prs_err;
184 /* Handle for cpuset.cpus.partition */
185 struct cgroup_file partition_file;
189 * Partition root states:
191 * 0 - member (not a partition root)
193 * 2 - partition root without load balancing (isolated)
194 * -1 - invalid partition root
195 * -2 - invalid isolated partition root
199 #define PRS_ISOLATED 2
200 #define PRS_INVALID_ROOT -1
201 #define PRS_INVALID_ISOLATED -2
203 static inline bool is_prs_invalid(int prs_state)
205 return prs_state < 0;
209 * Temporary cpumasks for working with partitions that are passed among
210 * functions to avoid memory allocation in inner functions.
213 cpumask_var_t addmask, delmask; /* For partition root */
214 cpumask_var_t new_cpus; /* For update_cpumasks_hier() */
217 static inline struct cpuset *css_cs(struct cgroup_subsys_state *css)
219 return css ? container_of(css, struct cpuset, css) : NULL;
222 /* Retrieve the cpuset for a task */
223 static inline struct cpuset *task_cs(struct task_struct *task)
225 return css_cs(task_css(task, cpuset_cgrp_id));
228 static inline struct cpuset *parent_cs(struct cpuset *cs)
230 return css_cs(cs->css.parent);
233 void inc_dl_tasks_cs(struct task_struct *p)
235 struct cpuset *cs = task_cs(p);
237 cs->nr_deadline_tasks++;
240 void dec_dl_tasks_cs(struct task_struct *p)
242 struct cpuset *cs = task_cs(p);
244 cs->nr_deadline_tasks--;
247 /* bits in struct cpuset flags field */
254 CS_SCHED_LOAD_BALANCE,
259 /* convenient tests for these bits */
260 static inline bool is_cpuset_online(struct cpuset *cs)
262 return test_bit(CS_ONLINE, &cs->flags) && !css_is_dying(&cs->css);
265 static inline int is_cpu_exclusive(const struct cpuset *cs)
267 return test_bit(CS_CPU_EXCLUSIVE, &cs->flags);
270 static inline int is_mem_exclusive(const struct cpuset *cs)
272 return test_bit(CS_MEM_EXCLUSIVE, &cs->flags);
275 static inline int is_mem_hardwall(const struct cpuset *cs)
277 return test_bit(CS_MEM_HARDWALL, &cs->flags);
280 static inline int is_sched_load_balance(const struct cpuset *cs)
282 return test_bit(CS_SCHED_LOAD_BALANCE, &cs->flags);
285 static inline int is_memory_migrate(const struct cpuset *cs)
287 return test_bit(CS_MEMORY_MIGRATE, &cs->flags);
290 static inline int is_spread_page(const struct cpuset *cs)
292 return test_bit(CS_SPREAD_PAGE, &cs->flags);
295 static inline int is_spread_slab(const struct cpuset *cs)
297 return test_bit(CS_SPREAD_SLAB, &cs->flags);
300 static inline int is_partition_valid(const struct cpuset *cs)
302 return cs->partition_root_state > 0;
305 static inline int is_partition_invalid(const struct cpuset *cs)
307 return cs->partition_root_state < 0;
311 * Callers should hold callback_lock to modify partition_root_state.
313 static inline void make_partition_invalid(struct cpuset *cs)
315 if (is_partition_valid(cs))
316 cs->partition_root_state = -cs->partition_root_state;
320 * Send notification event of whenever partition_root_state changes.
322 static inline void notify_partition_change(struct cpuset *cs, int old_prs)
324 if (old_prs == cs->partition_root_state)
326 cgroup_file_notify(&cs->partition_file);
328 /* Reset prs_err if not invalid */
329 if (is_partition_valid(cs))
330 WRITE_ONCE(cs->prs_err, PERR_NONE);
333 static struct cpuset top_cpuset = {
334 .flags = ((1 << CS_ONLINE) | (1 << CS_CPU_EXCLUSIVE) |
335 (1 << CS_MEM_EXCLUSIVE)),
336 .partition_root_state = PRS_ROOT,
340 * cpuset_for_each_child - traverse online children of a cpuset
341 * @child_cs: loop cursor pointing to the current child
342 * @pos_css: used for iteration
343 * @parent_cs: target cpuset to walk children of
345 * Walk @child_cs through the online children of @parent_cs. Must be used
346 * with RCU read locked.
348 #define cpuset_for_each_child(child_cs, pos_css, parent_cs) \
349 css_for_each_child((pos_css), &(parent_cs)->css) \
350 if (is_cpuset_online(((child_cs) = css_cs((pos_css)))))
353 * cpuset_for_each_descendant_pre - pre-order walk of a cpuset's descendants
354 * @des_cs: loop cursor pointing to the current descendant
355 * @pos_css: used for iteration
356 * @root_cs: target cpuset to walk ancestor of
358 * Walk @des_cs through the online descendants of @root_cs. Must be used
359 * with RCU read locked. The caller may modify @pos_css by calling
360 * css_rightmost_descendant() to skip subtree. @root_cs is included in the
361 * iteration and the first node to be visited.
363 #define cpuset_for_each_descendant_pre(des_cs, pos_css, root_cs) \
364 css_for_each_descendant_pre((pos_css), &(root_cs)->css) \
365 if (is_cpuset_online(((des_cs) = css_cs((pos_css)))))
368 * There are two global locks guarding cpuset structures - cpuset_mutex and
369 * callback_lock. We also require taking task_lock() when dereferencing a
370 * task's cpuset pointer. See "The task_lock() exception", at the end of this
371 * comment. The cpuset code uses only cpuset_mutex. Other kernel subsystems
372 * can use cpuset_lock()/cpuset_unlock() to prevent change to cpuset
373 * structures. Note that cpuset_mutex needs to be a mutex as it is used in
374 * paths that rely on priority inheritance (e.g. scheduler - on RT) for
377 * A task must hold both locks to modify cpusets. If a task holds
378 * cpuset_mutex, it blocks others, ensuring that it is the only task able to
379 * also acquire callback_lock and be able to modify cpusets. It can perform
380 * various checks on the cpuset structure first, knowing nothing will change.
381 * It can also allocate memory while just holding cpuset_mutex. While it is
382 * performing these checks, various callback routines can briefly acquire
383 * callback_lock to query cpusets. Once it is ready to make the changes, it
384 * takes callback_lock, blocking everyone else.
386 * Calls to the kernel memory allocator can not be made while holding
387 * callback_lock, as that would risk double tripping on callback_lock
388 * from one of the callbacks into the cpuset code from within
391 * If a task is only holding callback_lock, then it has read-only
394 * Now, the task_struct fields mems_allowed and mempolicy may be changed
395 * by other task, we use alloc_lock in the task_struct fields to protect
398 * The cpuset_common_file_read() handlers only hold callback_lock across
399 * small pieces of code, such as when reading out possibly multi-word
400 * cpumasks and nodemasks.
402 * Accessing a task's cpuset should be done in accordance with the
403 * guidelines for accessing subsystem state in kernel/cgroup.c
406 static DEFINE_MUTEX(cpuset_mutex);
408 void cpuset_lock(void)
410 mutex_lock(&cpuset_mutex);
413 void cpuset_unlock(void)
415 mutex_unlock(&cpuset_mutex);
418 static DEFINE_SPINLOCK(callback_lock);
420 static struct workqueue_struct *cpuset_migrate_mm_wq;
423 * CPU / memory hotplug is handled asynchronously.
425 static void cpuset_hotplug_workfn(struct work_struct *work);
426 static DECLARE_WORK(cpuset_hotplug_work, cpuset_hotplug_workfn);
428 static DECLARE_WAIT_QUEUE_HEAD(cpuset_attach_wq);
430 static inline void check_insane_mems_config(nodemask_t *nodes)
432 if (!cpusets_insane_config() &&
433 movable_only_nodes(nodes)) {
434 static_branch_enable(&cpusets_insane_config_key);
435 pr_info("Unsupported (movable nodes only) cpuset configuration detected (nmask=%*pbl)!\n"
436 "Cpuset allocations might fail even with a lot of memory available.\n",
437 nodemask_pr_args(nodes));
442 * Cgroup v2 behavior is used on the "cpus" and "mems" control files when
443 * on default hierarchy or when the cpuset_v2_mode flag is set by mounting
444 * the v1 cpuset cgroup filesystem with the "cpuset_v2_mode" mount option.
445 * With v2 behavior, "cpus" and "mems" are always what the users have
446 * requested and won't be changed by hotplug events. Only the effective
447 * cpus or mems will be affected.
449 static inline bool is_in_v2_mode(void)
451 return cgroup_subsys_on_dfl(cpuset_cgrp_subsys) ||
452 (cpuset_cgrp_subsys.root->flags & CGRP_ROOT_CPUSET_V2_MODE);
456 * partition_is_populated - check if partition has tasks
457 * @cs: partition root to be checked
458 * @excluded_child: a child cpuset to be excluded in task checking
459 * Return: true if there are tasks, false otherwise
461 * It is assumed that @cs is a valid partition root. @excluded_child should
462 * be non-NULL when this cpuset is going to become a partition itself.
464 static inline bool partition_is_populated(struct cpuset *cs,
465 struct cpuset *excluded_child)
467 struct cgroup_subsys_state *css;
468 struct cpuset *child;
470 if (cs->css.cgroup->nr_populated_csets)
472 if (!excluded_child && !cs->nr_subparts_cpus)
473 return cgroup_is_populated(cs->css.cgroup);
476 cpuset_for_each_child(child, css, cs) {
477 if (child == excluded_child)
479 if (is_partition_valid(child))
481 if (cgroup_is_populated(child->css.cgroup)) {
491 * Return in pmask the portion of a task's cpusets's cpus_allowed that
492 * are online and are capable of running the task. If none are found,
493 * walk up the cpuset hierarchy until we find one that does have some
496 * One way or another, we guarantee to return some non-empty subset
497 * of cpu_online_mask.
499 * Call with callback_lock or cpuset_mutex held.
501 static void guarantee_online_cpus(struct task_struct *tsk,
502 struct cpumask *pmask)
504 const struct cpumask *possible_mask = task_cpu_possible_mask(tsk);
507 if (WARN_ON(!cpumask_and(pmask, possible_mask, cpu_online_mask)))
508 cpumask_copy(pmask, cpu_online_mask);
513 while (!cpumask_intersects(cs->effective_cpus, pmask)) {
517 * The top cpuset doesn't have any online cpu as a
518 * consequence of a race between cpuset_hotplug_work
519 * and cpu hotplug notifier. But we know the top
520 * cpuset's effective_cpus is on its way to be
521 * identical to cpu_online_mask.
526 cpumask_and(pmask, pmask, cs->effective_cpus);
533 * Return in *pmask the portion of a cpusets's mems_allowed that
534 * are online, with memory. If none are online with memory, walk
535 * up the cpuset hierarchy until we find one that does have some
536 * online mems. The top cpuset always has some mems online.
538 * One way or another, we guarantee to return some non-empty subset
539 * of node_states[N_MEMORY].
541 * Call with callback_lock or cpuset_mutex held.
543 static void guarantee_online_mems(struct cpuset *cs, nodemask_t *pmask)
545 while (!nodes_intersects(cs->effective_mems, node_states[N_MEMORY]))
547 nodes_and(*pmask, cs->effective_mems, node_states[N_MEMORY]);
551 * update task's spread flag if cpuset's page/slab spread flag is set
553 * Call with callback_lock or cpuset_mutex held. The check can be skipped
554 * if on default hierarchy.
556 static void cpuset_update_task_spread_flags(struct cpuset *cs,
557 struct task_struct *tsk)
559 if (cgroup_subsys_on_dfl(cpuset_cgrp_subsys))
562 if (is_spread_page(cs))
563 task_set_spread_page(tsk);
565 task_clear_spread_page(tsk);
567 if (is_spread_slab(cs))
568 task_set_spread_slab(tsk);
570 task_clear_spread_slab(tsk);
574 * is_cpuset_subset(p, q) - Is cpuset p a subset of cpuset q?
576 * One cpuset is a subset of another if all its allowed CPUs and
577 * Memory Nodes are a subset of the other, and its exclusive flags
578 * are only set if the other's are set. Call holding cpuset_mutex.
581 static int is_cpuset_subset(const struct cpuset *p, const struct cpuset *q)
583 return cpumask_subset(p->cpus_allowed, q->cpus_allowed) &&
584 nodes_subset(p->mems_allowed, q->mems_allowed) &&
585 is_cpu_exclusive(p) <= is_cpu_exclusive(q) &&
586 is_mem_exclusive(p) <= is_mem_exclusive(q);
590 * alloc_cpumasks - allocate three cpumasks for cpuset
591 * @cs: the cpuset that have cpumasks to be allocated.
592 * @tmp: the tmpmasks structure pointer
593 * Return: 0 if successful, -ENOMEM otherwise.
595 * Only one of the two input arguments should be non-NULL.
597 static inline int alloc_cpumasks(struct cpuset *cs, struct tmpmasks *tmp)
599 cpumask_var_t *pmask1, *pmask2, *pmask3;
602 pmask1 = &cs->cpus_allowed;
603 pmask2 = &cs->effective_cpus;
604 pmask3 = &cs->subparts_cpus;
606 pmask1 = &tmp->new_cpus;
607 pmask2 = &tmp->addmask;
608 pmask3 = &tmp->delmask;
611 if (!zalloc_cpumask_var(pmask1, GFP_KERNEL))
614 if (!zalloc_cpumask_var(pmask2, GFP_KERNEL))
617 if (!zalloc_cpumask_var(pmask3, GFP_KERNEL))
623 free_cpumask_var(*pmask2);
625 free_cpumask_var(*pmask1);
630 * free_cpumasks - free cpumasks in a tmpmasks structure
631 * @cs: the cpuset that have cpumasks to be free.
632 * @tmp: the tmpmasks structure pointer
634 static inline void free_cpumasks(struct cpuset *cs, struct tmpmasks *tmp)
637 free_cpumask_var(cs->cpus_allowed);
638 free_cpumask_var(cs->effective_cpus);
639 free_cpumask_var(cs->subparts_cpus);
642 free_cpumask_var(tmp->new_cpus);
643 free_cpumask_var(tmp->addmask);
644 free_cpumask_var(tmp->delmask);
649 * alloc_trial_cpuset - allocate a trial cpuset
650 * @cs: the cpuset that the trial cpuset duplicates
652 static struct cpuset *alloc_trial_cpuset(struct cpuset *cs)
654 struct cpuset *trial;
656 trial = kmemdup(cs, sizeof(*cs), GFP_KERNEL);
660 if (alloc_cpumasks(trial, NULL)) {
665 cpumask_copy(trial->cpus_allowed, cs->cpus_allowed);
666 cpumask_copy(trial->effective_cpus, cs->effective_cpus);
671 * free_cpuset - free the cpuset
672 * @cs: the cpuset to be freed
674 static inline void free_cpuset(struct cpuset *cs)
676 free_cpumasks(cs, NULL);
681 * validate_change_legacy() - Validate conditions specific to legacy (v1)
684 static int validate_change_legacy(struct cpuset *cur, struct cpuset *trial)
686 struct cgroup_subsys_state *css;
687 struct cpuset *c, *par;
690 WARN_ON_ONCE(!rcu_read_lock_held());
692 /* Each of our child cpusets must be a subset of us */
694 cpuset_for_each_child(c, css, cur)
695 if (!is_cpuset_subset(c, trial))
698 /* On legacy hierarchy, we must be a subset of our parent cpuset. */
700 par = parent_cs(cur);
701 if (par && !is_cpuset_subset(trial, par))
710 * validate_change() - Used to validate that any proposed cpuset change
711 * follows the structural rules for cpusets.
713 * If we replaced the flag and mask values of the current cpuset
714 * (cur) with those values in the trial cpuset (trial), would
715 * our various subset and exclusive rules still be valid? Presumes
718 * 'cur' is the address of an actual, in-use cpuset. Operations
719 * such as list traversal that depend on the actual address of the
720 * cpuset in the list must use cur below, not trial.
722 * 'trial' is the address of bulk structure copy of cur, with
723 * perhaps one or more of the fields cpus_allowed, mems_allowed,
724 * or flags changed to new, trial values.
726 * Return 0 if valid, -errno if not.
729 static int validate_change(struct cpuset *cur, struct cpuset *trial)
731 struct cgroup_subsys_state *css;
732 struct cpuset *c, *par;
737 if (!is_in_v2_mode())
738 ret = validate_change_legacy(cur, trial);
742 /* Remaining checks don't apply to root cpuset */
743 if (cur == &top_cpuset)
746 par = parent_cs(cur);
749 * Cpusets with tasks - existing or newly being attached - can't
750 * be changed to have empty cpus_allowed or mems_allowed.
753 if ((cgroup_is_populated(cur->css.cgroup) || cur->attach_in_progress)) {
754 if (!cpumask_empty(cur->cpus_allowed) &&
755 cpumask_empty(trial->cpus_allowed))
757 if (!nodes_empty(cur->mems_allowed) &&
758 nodes_empty(trial->mems_allowed))
763 * We can't shrink if we won't have enough room for SCHED_DEADLINE
767 if (is_cpu_exclusive(cur) &&
768 !cpuset_cpumask_can_shrink(cur->cpus_allowed,
769 trial->cpus_allowed))
773 * If either I or some sibling (!= me) is exclusive, we can't
777 cpuset_for_each_child(c, css, par) {
778 if ((is_cpu_exclusive(trial) || is_cpu_exclusive(c)) &&
780 cpumask_intersects(trial->cpus_allowed, c->cpus_allowed))
782 if ((is_mem_exclusive(trial) || is_mem_exclusive(c)) &&
784 nodes_intersects(trial->mems_allowed, c->mems_allowed))
796 * Helper routine for generate_sched_domains().
797 * Do cpusets a, b have overlapping effective cpus_allowed masks?
799 static int cpusets_overlap(struct cpuset *a, struct cpuset *b)
801 return cpumask_intersects(a->effective_cpus, b->effective_cpus);
805 update_domain_attr(struct sched_domain_attr *dattr, struct cpuset *c)
807 if (dattr->relax_domain_level < c->relax_domain_level)
808 dattr->relax_domain_level = c->relax_domain_level;
812 static void update_domain_attr_tree(struct sched_domain_attr *dattr,
813 struct cpuset *root_cs)
816 struct cgroup_subsys_state *pos_css;
819 cpuset_for_each_descendant_pre(cp, pos_css, root_cs) {
820 /* skip the whole subtree if @cp doesn't have any CPU */
821 if (cpumask_empty(cp->cpus_allowed)) {
822 pos_css = css_rightmost_descendant(pos_css);
826 if (is_sched_load_balance(cp))
827 update_domain_attr(dattr, cp);
832 /* Must be called with cpuset_mutex held. */
833 static inline int nr_cpusets(void)
835 /* jump label reference count + the top-level cpuset */
836 return static_key_count(&cpusets_enabled_key.key) + 1;
840 * generate_sched_domains()
842 * This function builds a partial partition of the systems CPUs
843 * A 'partial partition' is a set of non-overlapping subsets whose
844 * union is a subset of that set.
845 * The output of this function needs to be passed to kernel/sched/core.c
846 * partition_sched_domains() routine, which will rebuild the scheduler's
847 * load balancing domains (sched domains) as specified by that partial
850 * See "What is sched_load_balance" in Documentation/admin-guide/cgroup-v1/cpusets.rst
851 * for a background explanation of this.
853 * Does not return errors, on the theory that the callers of this
854 * routine would rather not worry about failures to rebuild sched
855 * domains when operating in the severe memory shortage situations
856 * that could cause allocation failures below.
858 * Must be called with cpuset_mutex held.
860 * The three key local variables below are:
861 * cp - cpuset pointer, used (together with pos_css) to perform a
862 * top-down scan of all cpusets. For our purposes, rebuilding
863 * the schedulers sched domains, we can ignore !is_sched_load_
865 * csa - (for CpuSet Array) Array of pointers to all the cpusets
866 * that need to be load balanced, for convenient iterative
867 * access by the subsequent code that finds the best partition,
868 * i.e the set of domains (subsets) of CPUs such that the
869 * cpus_allowed of every cpuset marked is_sched_load_balance
870 * is a subset of one of these domains, while there are as
871 * many such domains as possible, each as small as possible.
872 * doms - Conversion of 'csa' to an array of cpumasks, for passing to
873 * the kernel/sched/core.c routine partition_sched_domains() in a
874 * convenient format, that can be easily compared to the prior
875 * value to determine what partition elements (sched domains)
876 * were changed (added or removed.)
878 * Finding the best partition (set of domains):
879 * The triple nested loops below over i, j, k scan over the
880 * load balanced cpusets (using the array of cpuset pointers in
881 * csa[]) looking for pairs of cpusets that have overlapping
882 * cpus_allowed, but which don't have the same 'pn' partition
883 * number and gives them in the same partition number. It keeps
884 * looping on the 'restart' label until it can no longer find
887 * The union of the cpus_allowed masks from the set of
888 * all cpusets having the same 'pn' value then form the one
889 * element of the partition (one sched domain) to be passed to
890 * partition_sched_domains().
892 static int generate_sched_domains(cpumask_var_t **domains,
893 struct sched_domain_attr **attributes)
895 struct cpuset *cp; /* top-down scan of cpusets */
896 struct cpuset **csa; /* array of all cpuset ptrs */
897 int csn; /* how many cpuset ptrs in csa so far */
898 int i, j, k; /* indices for partition finding loops */
899 cpumask_var_t *doms; /* resulting partition; i.e. sched domains */
900 struct sched_domain_attr *dattr; /* attributes for custom domains */
901 int ndoms = 0; /* number of sched domains in result */
902 int nslot; /* next empty doms[] struct cpumask slot */
903 struct cgroup_subsys_state *pos_css;
904 bool root_load_balance = is_sched_load_balance(&top_cpuset);
910 /* Special case for the 99% of systems with one, full, sched domain */
911 if (root_load_balance && !top_cpuset.nr_subparts_cpus) {
913 doms = alloc_sched_domains(ndoms);
917 dattr = kmalloc(sizeof(struct sched_domain_attr), GFP_KERNEL);
919 *dattr = SD_ATTR_INIT;
920 update_domain_attr_tree(dattr, &top_cpuset);
922 cpumask_and(doms[0], top_cpuset.effective_cpus,
923 housekeeping_cpumask(HK_TYPE_DOMAIN));
928 csa = kmalloc_array(nr_cpusets(), sizeof(cp), GFP_KERNEL);
934 if (root_load_balance)
935 csa[csn++] = &top_cpuset;
936 cpuset_for_each_descendant_pre(cp, pos_css, &top_cpuset) {
937 if (cp == &top_cpuset)
940 * Continue traversing beyond @cp iff @cp has some CPUs and
941 * isn't load balancing. The former is obvious. The
942 * latter: All child cpusets contain a subset of the
943 * parent's cpus, so just skip them, and then we call
944 * update_domain_attr_tree() to calc relax_domain_level of
945 * the corresponding sched domain.
947 * If root is load-balancing, we can skip @cp if it
948 * is a subset of the root's effective_cpus.
950 if (!cpumask_empty(cp->cpus_allowed) &&
951 !(is_sched_load_balance(cp) &&
952 cpumask_intersects(cp->cpus_allowed,
953 housekeeping_cpumask(HK_TYPE_DOMAIN))))
956 if (root_load_balance &&
957 cpumask_subset(cp->cpus_allowed, top_cpuset.effective_cpus))
960 if (is_sched_load_balance(cp) &&
961 !cpumask_empty(cp->effective_cpus))
964 /* skip @cp's subtree if not a partition root */
965 if (!is_partition_valid(cp))
966 pos_css = css_rightmost_descendant(pos_css);
970 for (i = 0; i < csn; i++)
975 /* Find the best partition (set of sched domains) */
976 for (i = 0; i < csn; i++) {
977 struct cpuset *a = csa[i];
980 for (j = 0; j < csn; j++) {
981 struct cpuset *b = csa[j];
984 if (apn != bpn && cpusets_overlap(a, b)) {
985 for (k = 0; k < csn; k++) {
986 struct cpuset *c = csa[k];
991 ndoms--; /* one less element */
998 * Now we know how many domains to create.
999 * Convert <csn, csa> to <ndoms, doms> and populate cpu masks.
1001 doms = alloc_sched_domains(ndoms);
1006 * The rest of the code, including the scheduler, can deal with
1007 * dattr==NULL case. No need to abort if alloc fails.
1009 dattr = kmalloc_array(ndoms, sizeof(struct sched_domain_attr),
1012 for (nslot = 0, i = 0; i < csn; i++) {
1013 struct cpuset *a = csa[i];
1018 /* Skip completed partitions */
1024 if (nslot == ndoms) {
1025 static int warnings = 10;
1027 pr_warn("rebuild_sched_domains confused: nslot %d, ndoms %d, csn %d, i %d, apn %d\n",
1028 nslot, ndoms, csn, i, apn);
1036 *(dattr + nslot) = SD_ATTR_INIT;
1037 for (j = i; j < csn; j++) {
1038 struct cpuset *b = csa[j];
1041 cpumask_or(dp, dp, b->effective_cpus);
1042 cpumask_and(dp, dp, housekeeping_cpumask(HK_TYPE_DOMAIN));
1044 update_domain_attr_tree(dattr + nslot, b);
1046 /* Done with this partition */
1052 BUG_ON(nslot != ndoms);
1058 * Fallback to the default domain if kmalloc() failed.
1059 * See comments in partition_sched_domains().
1065 *attributes = dattr;
1069 static void dl_update_tasks_root_domain(struct cpuset *cs)
1071 struct css_task_iter it;
1072 struct task_struct *task;
1074 if (cs->nr_deadline_tasks == 0)
1077 css_task_iter_start(&cs->css, 0, &it);
1079 while ((task = css_task_iter_next(&it)))
1080 dl_add_task_root_domain(task);
1082 css_task_iter_end(&it);
1085 static void dl_rebuild_rd_accounting(void)
1087 struct cpuset *cs = NULL;
1088 struct cgroup_subsys_state *pos_css;
1090 lockdep_assert_held(&cpuset_mutex);
1091 lockdep_assert_cpus_held();
1092 lockdep_assert_held(&sched_domains_mutex);
1097 * Clear default root domain DL accounting, it will be computed again
1098 * if a task belongs to it.
1100 dl_clear_root_domain(&def_root_domain);
1102 cpuset_for_each_descendant_pre(cs, pos_css, &top_cpuset) {
1104 if (cpumask_empty(cs->effective_cpus)) {
1105 pos_css = css_rightmost_descendant(pos_css);
1113 dl_update_tasks_root_domain(cs);
1122 partition_and_rebuild_sched_domains(int ndoms_new, cpumask_var_t doms_new[],
1123 struct sched_domain_attr *dattr_new)
1125 mutex_lock(&sched_domains_mutex);
1126 partition_sched_domains_locked(ndoms_new, doms_new, dattr_new);
1127 dl_rebuild_rd_accounting();
1128 mutex_unlock(&sched_domains_mutex);
1132 * Rebuild scheduler domains.
1134 * If the flag 'sched_load_balance' of any cpuset with non-empty
1135 * 'cpus' changes, or if the 'cpus' allowed changes in any cpuset
1136 * which has that flag enabled, or if any cpuset with a non-empty
1137 * 'cpus' is removed, then call this routine to rebuild the
1138 * scheduler's dynamic sched domains.
1140 * Call with cpuset_mutex held. Takes cpus_read_lock().
1142 static void rebuild_sched_domains_locked(void)
1144 struct cgroup_subsys_state *pos_css;
1145 struct sched_domain_attr *attr;
1146 cpumask_var_t *doms;
1150 lockdep_assert_cpus_held();
1151 lockdep_assert_held(&cpuset_mutex);
1154 * If we have raced with CPU hotplug, return early to avoid
1155 * passing doms with offlined cpu to partition_sched_domains().
1156 * Anyways, cpuset_hotplug_workfn() will rebuild sched domains.
1158 * With no CPUs in any subpartitions, top_cpuset's effective CPUs
1159 * should be the same as the active CPUs, so checking only top_cpuset
1160 * is enough to detect racing CPU offlines.
1162 if (!top_cpuset.nr_subparts_cpus &&
1163 !cpumask_equal(top_cpuset.effective_cpus, cpu_active_mask))
1167 * With subpartition CPUs, however, the effective CPUs of a partition
1168 * root should be only a subset of the active CPUs. Since a CPU in any
1169 * partition root could be offlined, all must be checked.
1171 if (top_cpuset.nr_subparts_cpus) {
1173 cpuset_for_each_descendant_pre(cs, pos_css, &top_cpuset) {
1174 if (!is_partition_valid(cs)) {
1175 pos_css = css_rightmost_descendant(pos_css);
1178 if (!cpumask_subset(cs->effective_cpus,
1187 /* Generate domain masks and attrs */
1188 ndoms = generate_sched_domains(&doms, &attr);
1190 /* Have scheduler rebuild the domains */
1191 partition_and_rebuild_sched_domains(ndoms, doms, attr);
1193 #else /* !CONFIG_SMP */
1194 static void rebuild_sched_domains_locked(void)
1197 #endif /* CONFIG_SMP */
1199 void rebuild_sched_domains(void)
1202 mutex_lock(&cpuset_mutex);
1203 rebuild_sched_domains_locked();
1204 mutex_unlock(&cpuset_mutex);
1209 * update_tasks_cpumask - Update the cpumasks of tasks in the cpuset.
1210 * @cs: the cpuset in which each task's cpus_allowed mask needs to be changed
1211 * @new_cpus: the temp variable for the new effective_cpus mask
1213 * Iterate through each task of @cs updating its cpus_allowed to the
1214 * effective cpuset's. As this function is called with cpuset_mutex held,
1215 * cpuset membership stays stable. For top_cpuset, task_cpu_possible_mask()
1216 * is used instead of effective_cpus to make sure all offline CPUs are also
1217 * included as hotplug code won't update cpumasks for tasks in top_cpuset.
1219 static void update_tasks_cpumask(struct cpuset *cs, struct cpumask *new_cpus)
1221 struct css_task_iter it;
1222 struct task_struct *task;
1223 bool top_cs = cs == &top_cpuset;
1225 css_task_iter_start(&cs->css, 0, &it);
1226 while ((task = css_task_iter_next(&it))) {
1227 const struct cpumask *possible_mask = task_cpu_possible_mask(task);
1231 * Percpu kthreads in top_cpuset are ignored
1233 if ((task->flags & PF_KTHREAD) && kthread_is_per_cpu(task))
1235 cpumask_andnot(new_cpus, possible_mask, cs->subparts_cpus);
1237 cpumask_and(new_cpus, possible_mask, cs->effective_cpus);
1239 set_cpus_allowed_ptr(task, new_cpus);
1241 css_task_iter_end(&it);
1245 * compute_effective_cpumask - Compute the effective cpumask of the cpuset
1246 * @new_cpus: the temp variable for the new effective_cpus mask
1247 * @cs: the cpuset the need to recompute the new effective_cpus mask
1248 * @parent: the parent cpuset
1250 * If the parent has subpartition CPUs, include them in the list of
1251 * allowable CPUs in computing the new effective_cpus mask. Since offlined
1252 * CPUs are not removed from subparts_cpus, we have to use cpu_active_mask
1253 * to mask those out.
1255 static void compute_effective_cpumask(struct cpumask *new_cpus,
1256 struct cpuset *cs, struct cpuset *parent)
1258 if (parent->nr_subparts_cpus) {
1259 cpumask_or(new_cpus, parent->effective_cpus,
1260 parent->subparts_cpus);
1261 cpumask_and(new_cpus, new_cpus, cs->cpus_allowed);
1262 cpumask_and(new_cpus, new_cpus, cpu_active_mask);
1264 cpumask_and(new_cpus, cs->cpus_allowed, parent->effective_cpus);
1269 * Commands for update_parent_subparts_cpumask
1272 partcmd_enable, /* Enable partition root */
1273 partcmd_disable, /* Disable partition root */
1274 partcmd_update, /* Update parent's subparts_cpus */
1275 partcmd_invalidate, /* Make partition invalid */
1278 static int update_flag(cpuset_flagbits_t bit, struct cpuset *cs,
1281 * update_parent_subparts_cpumask - update subparts_cpus mask of parent cpuset
1282 * @cs: The cpuset that requests change in partition root state
1283 * @cmd: Partition root state change command
1284 * @newmask: Optional new cpumask for partcmd_update
1285 * @tmp: Temporary addmask and delmask
1286 * Return: 0 or a partition root state error code
1288 * For partcmd_enable, the cpuset is being transformed from a non-partition
1289 * root to a partition root. The cpus_allowed mask of the given cpuset will
1290 * be put into parent's subparts_cpus and taken away from parent's
1291 * effective_cpus. The function will return 0 if all the CPUs listed in
1292 * cpus_allowed can be granted or an error code will be returned.
1294 * For partcmd_disable, the cpuset is being transformed from a partition
1295 * root back to a non-partition root. Any CPUs in cpus_allowed that are in
1296 * parent's subparts_cpus will be taken away from that cpumask and put back
1297 * into parent's effective_cpus. 0 will always be returned.
1299 * For partcmd_update, if the optional newmask is specified, the cpu list is
1300 * to be changed from cpus_allowed to newmask. Otherwise, cpus_allowed is
1301 * assumed to remain the same. The cpuset should either be a valid or invalid
1302 * partition root. The partition root state may change from valid to invalid
1303 * or vice versa. An error code will only be returned if transitioning from
1304 * invalid to valid violates the exclusivity rule.
1306 * For partcmd_invalidate, the current partition will be made invalid.
1308 * The partcmd_enable and partcmd_disable commands are used by
1309 * update_prstate(). An error code may be returned and the caller will check
1312 * The partcmd_update command is used by update_cpumasks_hier() with newmask
1313 * NULL and update_cpumask() with newmask set. The partcmd_invalidate is used
1314 * by update_cpumask() with NULL newmask. In both cases, the callers won't
1315 * check for error and so partition_root_state and prs_error will be updated
1318 static int update_parent_subparts_cpumask(struct cpuset *cs, int cmd,
1319 struct cpumask *newmask,
1320 struct tmpmasks *tmp)
1322 struct cpuset *parent = parent_cs(cs);
1323 int adding; /* Moving cpus from effective_cpus to subparts_cpus */
1324 int deleting; /* Moving cpus from subparts_cpus to effective_cpus */
1325 int old_prs, new_prs;
1326 int part_error = PERR_NONE; /* Partition error? */
1328 lockdep_assert_held(&cpuset_mutex);
1331 * The parent must be a partition root.
1332 * The new cpumask, if present, or the current cpus_allowed must
1335 if (!is_partition_valid(parent)) {
1336 return is_partition_invalid(parent)
1337 ? PERR_INVPARENT : PERR_NOTPART;
1339 if ((newmask && cpumask_empty(newmask)) ||
1340 (!newmask && cpumask_empty(cs->cpus_allowed)))
1341 return PERR_CPUSEMPTY;
1344 * new_prs will only be changed for the partcmd_update and
1345 * partcmd_invalidate commands.
1347 adding = deleting = false;
1348 old_prs = new_prs = cs->partition_root_state;
1349 if (cmd == partcmd_enable) {
1351 * Enabling partition root is not allowed if cpus_allowed
1352 * doesn't overlap parent's cpus_allowed.
1354 if (!cpumask_intersects(cs->cpus_allowed, parent->cpus_allowed))
1355 return PERR_INVCPUS;
1358 * A parent can be left with no CPU as long as there is no
1359 * task directly associated with the parent partition.
1361 if (cpumask_subset(parent->effective_cpus, cs->cpus_allowed) &&
1362 partition_is_populated(parent, cs))
1365 cpumask_copy(tmp->addmask, cs->cpus_allowed);
1367 } else if (cmd == partcmd_disable) {
1369 * Need to remove cpus from parent's subparts_cpus for valid
1372 deleting = !is_prs_invalid(old_prs) &&
1373 cpumask_and(tmp->delmask, cs->cpus_allowed,
1374 parent->subparts_cpus);
1375 } else if (cmd == partcmd_invalidate) {
1376 if (is_prs_invalid(old_prs))
1380 * Make the current partition invalid. It is assumed that
1381 * invalidation is caused by violating cpu exclusivity rule.
1383 deleting = cpumask_and(tmp->delmask, cs->cpus_allowed,
1384 parent->subparts_cpus);
1387 part_error = PERR_NOTEXCL;
1389 } else if (newmask) {
1391 * partcmd_update with newmask:
1393 * Compute add/delete mask to/from subparts_cpus
1395 * delmask = cpus_allowed & ~newmask & parent->subparts_cpus
1396 * addmask = newmask & parent->cpus_allowed
1397 * & ~parent->subparts_cpus
1399 cpumask_andnot(tmp->delmask, cs->cpus_allowed, newmask);
1400 deleting = cpumask_and(tmp->delmask, tmp->delmask,
1401 parent->subparts_cpus);
1403 cpumask_and(tmp->addmask, newmask, parent->cpus_allowed);
1404 adding = cpumask_andnot(tmp->addmask, tmp->addmask,
1405 parent->subparts_cpus);
1407 * Make partition invalid if parent's effective_cpus could
1408 * become empty and there are tasks in the parent.
1411 cpumask_subset(parent->effective_cpus, tmp->addmask) &&
1412 !cpumask_intersects(tmp->delmask, cpu_active_mask) &&
1413 partition_is_populated(parent, cs)) {
1414 part_error = PERR_NOCPUS;
1416 deleting = cpumask_and(tmp->delmask, cs->cpus_allowed,
1417 parent->subparts_cpus);
1421 * partcmd_update w/o newmask:
1423 * delmask = cpus_allowed & parent->subparts_cpus
1424 * addmask = cpus_allowed & parent->cpus_allowed
1425 * & ~parent->subparts_cpus
1427 * This gets invoked either due to a hotplug event or from
1428 * update_cpumasks_hier(). This can cause the state of a
1429 * partition root to transition from valid to invalid or vice
1430 * versa. So we still need to compute the addmask and delmask.
1432 * A partition error happens when:
1433 * 1) Cpuset is valid partition, but parent does not distribute
1435 * 2) Parent has tasks and all its effective CPUs will have
1436 * to be distributed out.
1438 cpumask_and(tmp->addmask, cs->cpus_allowed,
1439 parent->cpus_allowed);
1440 adding = cpumask_andnot(tmp->addmask, tmp->addmask,
1441 parent->subparts_cpus);
1443 if ((is_partition_valid(cs) && !parent->nr_subparts_cpus) ||
1445 cpumask_subset(parent->effective_cpus, tmp->addmask) &&
1446 partition_is_populated(parent, cs))) {
1447 part_error = PERR_NOCPUS;
1451 if (part_error && is_partition_valid(cs) &&
1452 parent->nr_subparts_cpus)
1453 deleting = cpumask_and(tmp->delmask, cs->cpus_allowed,
1454 parent->subparts_cpus);
1457 WRITE_ONCE(cs->prs_err, part_error);
1459 if (cmd == partcmd_update) {
1461 * Check for possible transition between valid and invalid
1464 switch (cs->partition_root_state) {
1470 case PRS_INVALID_ROOT:
1471 case PRS_INVALID_ISOLATED:
1478 if (!adding && !deleting && (new_prs == old_prs))
1482 * Transitioning between invalid to valid or vice versa may require
1483 * changing CS_CPU_EXCLUSIVE and CS_SCHED_LOAD_BALANCE.
1485 if (old_prs != new_prs) {
1486 if (is_prs_invalid(old_prs) && !is_cpu_exclusive(cs) &&
1487 (update_flag(CS_CPU_EXCLUSIVE, cs, 1) < 0))
1488 return PERR_NOTEXCL;
1489 if (is_prs_invalid(new_prs) && is_cpu_exclusive(cs))
1490 update_flag(CS_CPU_EXCLUSIVE, cs, 0);
1494 * Change the parent's subparts_cpus.
1495 * Newly added CPUs will be removed from effective_cpus and
1496 * newly deleted ones will be added back to effective_cpus.
1498 spin_lock_irq(&callback_lock);
1500 cpumask_or(parent->subparts_cpus,
1501 parent->subparts_cpus, tmp->addmask);
1502 cpumask_andnot(parent->effective_cpus,
1503 parent->effective_cpus, tmp->addmask);
1506 cpumask_andnot(parent->subparts_cpus,
1507 parent->subparts_cpus, tmp->delmask);
1509 * Some of the CPUs in subparts_cpus might have been offlined.
1511 cpumask_and(tmp->delmask, tmp->delmask, cpu_active_mask);
1512 cpumask_or(parent->effective_cpus,
1513 parent->effective_cpus, tmp->delmask);
1516 parent->nr_subparts_cpus = cpumask_weight(parent->subparts_cpus);
1518 if (old_prs != new_prs)
1519 cs->partition_root_state = new_prs;
1521 spin_unlock_irq(&callback_lock);
1523 if (adding || deleting)
1524 update_tasks_cpumask(parent, tmp->addmask);
1527 * Set or clear CS_SCHED_LOAD_BALANCE when partcmd_update, if necessary.
1528 * rebuild_sched_domains_locked() may be called.
1530 if (old_prs != new_prs) {
1531 if (old_prs == PRS_ISOLATED)
1532 update_flag(CS_SCHED_LOAD_BALANCE, cs, 1);
1533 else if (new_prs == PRS_ISOLATED)
1534 update_flag(CS_SCHED_LOAD_BALANCE, cs, 0);
1536 notify_partition_change(cs, old_prs);
1541 * update_cpumasks_hier - Update effective cpumasks and tasks in the subtree
1542 * @cs: the cpuset to consider
1543 * @tmp: temp variables for calculating effective_cpus & partition setup
1544 * @force: don't skip any descendant cpusets if set
1546 * When configured cpumask is changed, the effective cpumasks of this cpuset
1547 * and all its descendants need to be updated.
1549 * On legacy hierarchy, effective_cpus will be the same with cpu_allowed.
1551 * Called with cpuset_mutex held
1553 static void update_cpumasks_hier(struct cpuset *cs, struct tmpmasks *tmp,
1557 struct cgroup_subsys_state *pos_css;
1558 bool need_rebuild_sched_domains = false;
1559 int old_prs, new_prs;
1562 cpuset_for_each_descendant_pre(cp, pos_css, cs) {
1563 struct cpuset *parent = parent_cs(cp);
1564 bool update_parent = false;
1566 compute_effective_cpumask(tmp->new_cpus, cp, parent);
1569 * If it becomes empty, inherit the effective mask of the
1570 * parent, which is guaranteed to have some CPUs unless
1571 * it is a partition root that has explicitly distributed
1574 if (is_in_v2_mode() && cpumask_empty(tmp->new_cpus)) {
1575 if (is_partition_valid(cp) &&
1576 cpumask_equal(cp->cpus_allowed, cp->subparts_cpus))
1577 goto update_parent_subparts;
1579 cpumask_copy(tmp->new_cpus, parent->effective_cpus);
1580 if (!cp->use_parent_ecpus) {
1581 cp->use_parent_ecpus = true;
1582 parent->child_ecpus_count++;
1584 } else if (cp->use_parent_ecpus) {
1585 cp->use_parent_ecpus = false;
1586 WARN_ON_ONCE(!parent->child_ecpus_count);
1587 parent->child_ecpus_count--;
1591 * Skip the whole subtree if the cpumask remains the same
1592 * and has no partition root state and force flag not set.
1594 if (!cp->partition_root_state && !force &&
1595 cpumask_equal(tmp->new_cpus, cp->effective_cpus)) {
1596 pos_css = css_rightmost_descendant(pos_css);
1600 update_parent_subparts:
1602 * update_parent_subparts_cpumask() should have been called
1603 * for cs already in update_cpumask(). We should also call
1604 * update_tasks_cpumask() again for tasks in the parent
1605 * cpuset if the parent's subparts_cpus changes.
1607 old_prs = new_prs = cp->partition_root_state;
1608 if ((cp != cs) && old_prs) {
1609 switch (parent->partition_root_state) {
1612 update_parent = true;
1617 * When parent is not a partition root or is
1618 * invalid, child partition roots become
1621 if (is_partition_valid(cp))
1622 new_prs = -cp->partition_root_state;
1623 WRITE_ONCE(cp->prs_err,
1624 is_partition_invalid(parent)
1625 ? PERR_INVPARENT : PERR_NOTPART);
1630 if (!css_tryget_online(&cp->css))
1634 if (update_parent) {
1635 update_parent_subparts_cpumask(cp, partcmd_update, NULL,
1638 * The cpuset partition_root_state may become
1639 * invalid. Capture it.
1641 new_prs = cp->partition_root_state;
1644 spin_lock_irq(&callback_lock);
1646 if (cp->nr_subparts_cpus && !is_partition_valid(cp)) {
1648 * Put all active subparts_cpus back to effective_cpus.
1650 cpumask_or(tmp->new_cpus, tmp->new_cpus,
1652 cpumask_and(tmp->new_cpus, tmp->new_cpus,
1654 cp->nr_subparts_cpus = 0;
1655 cpumask_clear(cp->subparts_cpus);
1658 cpumask_copy(cp->effective_cpus, tmp->new_cpus);
1659 if (cp->nr_subparts_cpus) {
1661 * Make sure that effective_cpus & subparts_cpus
1662 * are mutually exclusive.
1664 cpumask_andnot(cp->effective_cpus, cp->effective_cpus,
1668 cp->partition_root_state = new_prs;
1669 spin_unlock_irq(&callback_lock);
1671 notify_partition_change(cp, old_prs);
1673 WARN_ON(!is_in_v2_mode() &&
1674 !cpumask_equal(cp->cpus_allowed, cp->effective_cpus));
1676 update_tasks_cpumask(cp, tmp->new_cpus);
1679 * On legacy hierarchy, if the effective cpumask of any non-
1680 * empty cpuset is changed, we need to rebuild sched domains.
1681 * On default hierarchy, the cpuset needs to be a partition
1684 if (!cpumask_empty(cp->cpus_allowed) &&
1685 is_sched_load_balance(cp) &&
1686 (!cgroup_subsys_on_dfl(cpuset_cgrp_subsys) ||
1687 is_partition_valid(cp)))
1688 need_rebuild_sched_domains = true;
1695 if (need_rebuild_sched_domains)
1696 rebuild_sched_domains_locked();
1700 * update_sibling_cpumasks - Update siblings cpumasks
1701 * @parent: Parent cpuset
1702 * @cs: Current cpuset
1703 * @tmp: Temp variables
1705 static void update_sibling_cpumasks(struct cpuset *parent, struct cpuset *cs,
1706 struct tmpmasks *tmp)
1708 struct cpuset *sibling;
1709 struct cgroup_subsys_state *pos_css;
1711 lockdep_assert_held(&cpuset_mutex);
1714 * Check all its siblings and call update_cpumasks_hier()
1715 * if their use_parent_ecpus flag is set in order for them
1716 * to use the right effective_cpus value.
1718 * The update_cpumasks_hier() function may sleep. So we have to
1719 * release the RCU read lock before calling it.
1722 cpuset_for_each_child(sibling, pos_css, parent) {
1725 if (!sibling->use_parent_ecpus)
1727 if (!css_tryget_online(&sibling->css))
1731 update_cpumasks_hier(sibling, tmp, false);
1733 css_put(&sibling->css);
1739 * update_cpumask - update the cpus_allowed mask of a cpuset and all tasks in it
1740 * @cs: the cpuset to consider
1741 * @trialcs: trial cpuset
1742 * @buf: buffer of cpu numbers written to this cpuset
1744 static int update_cpumask(struct cpuset *cs, struct cpuset *trialcs,
1748 struct tmpmasks tmp;
1749 bool invalidate = false;
1751 /* top_cpuset.cpus_allowed tracks cpu_online_mask; it's read-only */
1752 if (cs == &top_cpuset)
1756 * An empty cpus_allowed is ok only if the cpuset has no tasks.
1757 * Since cpulist_parse() fails on an empty mask, we special case
1758 * that parsing. The validate_change() call ensures that cpusets
1759 * with tasks have cpus.
1762 cpumask_clear(trialcs->cpus_allowed);
1764 retval = cpulist_parse(buf, trialcs->cpus_allowed);
1768 if (!cpumask_subset(trialcs->cpus_allowed,
1769 top_cpuset.cpus_allowed))
1773 /* Nothing to do if the cpus didn't change */
1774 if (cpumask_equal(cs->cpus_allowed, trialcs->cpus_allowed))
1777 #ifdef CONFIG_CPUMASK_OFFSTACK
1779 * Use the cpumasks in trialcs for tmpmasks when they are pointers
1780 * to allocated cpumasks.
1782 * Note that update_parent_subparts_cpumask() uses only addmask &
1783 * delmask, but not new_cpus.
1785 tmp.addmask = trialcs->subparts_cpus;
1786 tmp.delmask = trialcs->effective_cpus;
1787 tmp.new_cpus = NULL;
1790 retval = validate_change(cs, trialcs);
1792 if ((retval == -EINVAL) && cgroup_subsys_on_dfl(cpuset_cgrp_subsys)) {
1793 struct cpuset *cp, *parent;
1794 struct cgroup_subsys_state *css;
1797 * The -EINVAL error code indicates that partition sibling
1798 * CPU exclusivity rule has been violated. We still allow
1799 * the cpumask change to proceed while invalidating the
1800 * partition. However, any conflicting sibling partitions
1801 * have to be marked as invalid too.
1805 parent = parent_cs(cs);
1806 cpuset_for_each_child(cp, css, parent)
1807 if (is_partition_valid(cp) &&
1808 cpumask_intersects(trialcs->cpus_allowed, cp->cpus_allowed)) {
1810 update_parent_subparts_cpumask(cp, partcmd_invalidate, NULL, &tmp);
1819 if (cs->partition_root_state) {
1821 update_parent_subparts_cpumask(cs, partcmd_invalidate,
1824 update_parent_subparts_cpumask(cs, partcmd_update,
1825 trialcs->cpus_allowed, &tmp);
1828 compute_effective_cpumask(trialcs->effective_cpus, trialcs,
1830 spin_lock_irq(&callback_lock);
1831 cpumask_copy(cs->cpus_allowed, trialcs->cpus_allowed);
1834 * Make sure that subparts_cpus, if not empty, is a subset of
1835 * cpus_allowed. Clear subparts_cpus if partition not valid or
1836 * empty effective cpus with tasks.
1838 if (cs->nr_subparts_cpus) {
1839 if (!is_partition_valid(cs) ||
1840 (cpumask_subset(trialcs->effective_cpus, cs->subparts_cpus) &&
1841 partition_is_populated(cs, NULL))) {
1842 cs->nr_subparts_cpus = 0;
1843 cpumask_clear(cs->subparts_cpus);
1845 cpumask_and(cs->subparts_cpus, cs->subparts_cpus,
1847 cs->nr_subparts_cpus = cpumask_weight(cs->subparts_cpus);
1850 spin_unlock_irq(&callback_lock);
1852 #ifdef CONFIG_CPUMASK_OFFSTACK
1853 /* Now trialcs->cpus_allowed is available */
1854 tmp.new_cpus = trialcs->cpus_allowed;
1857 /* effective_cpus will be updated here */
1858 update_cpumasks_hier(cs, &tmp, false);
1860 if (cs->partition_root_state) {
1861 struct cpuset *parent = parent_cs(cs);
1864 * For partition root, update the cpumasks of sibling
1865 * cpusets if they use parent's effective_cpus.
1867 if (parent->child_ecpus_count)
1868 update_sibling_cpumasks(parent, cs, &tmp);
1874 * Migrate memory region from one set of nodes to another. This is
1875 * performed asynchronously as it can be called from process migration path
1876 * holding locks involved in process management. All mm migrations are
1877 * performed in the queued order and can be waited for by flushing
1878 * cpuset_migrate_mm_wq.
1881 struct cpuset_migrate_mm_work {
1882 struct work_struct work;
1883 struct mm_struct *mm;
1888 static void cpuset_migrate_mm_workfn(struct work_struct *work)
1890 struct cpuset_migrate_mm_work *mwork =
1891 container_of(work, struct cpuset_migrate_mm_work, work);
1893 /* on a wq worker, no need to worry about %current's mems_allowed */
1894 do_migrate_pages(mwork->mm, &mwork->from, &mwork->to, MPOL_MF_MOVE_ALL);
1899 static void cpuset_migrate_mm(struct mm_struct *mm, const nodemask_t *from,
1900 const nodemask_t *to)
1902 struct cpuset_migrate_mm_work *mwork;
1904 if (nodes_equal(*from, *to)) {
1909 mwork = kzalloc(sizeof(*mwork), GFP_KERNEL);
1912 mwork->from = *from;
1914 INIT_WORK(&mwork->work, cpuset_migrate_mm_workfn);
1915 queue_work(cpuset_migrate_mm_wq, &mwork->work);
1921 static void cpuset_post_attach(void)
1923 flush_workqueue(cpuset_migrate_mm_wq);
1927 * cpuset_change_task_nodemask - change task's mems_allowed and mempolicy
1928 * @tsk: the task to change
1929 * @newmems: new nodes that the task will be set
1931 * We use the mems_allowed_seq seqlock to safely update both tsk->mems_allowed
1932 * and rebind an eventual tasks' mempolicy. If the task is allocating in
1933 * parallel, it might temporarily see an empty intersection, which results in
1934 * a seqlock check and retry before OOM or allocation failure.
1936 static void cpuset_change_task_nodemask(struct task_struct *tsk,
1937 nodemask_t *newmems)
1941 local_irq_disable();
1942 write_seqcount_begin(&tsk->mems_allowed_seq);
1944 nodes_or(tsk->mems_allowed, tsk->mems_allowed, *newmems);
1945 mpol_rebind_task(tsk, newmems);
1946 tsk->mems_allowed = *newmems;
1948 write_seqcount_end(&tsk->mems_allowed_seq);
1954 static void *cpuset_being_rebound;
1957 * update_tasks_nodemask - Update the nodemasks of tasks in the cpuset.
1958 * @cs: the cpuset in which each task's mems_allowed mask needs to be changed
1960 * Iterate through each task of @cs updating its mems_allowed to the
1961 * effective cpuset's. As this function is called with cpuset_mutex held,
1962 * cpuset membership stays stable.
1964 static void update_tasks_nodemask(struct cpuset *cs)
1966 static nodemask_t newmems; /* protected by cpuset_mutex */
1967 struct css_task_iter it;
1968 struct task_struct *task;
1970 cpuset_being_rebound = cs; /* causes mpol_dup() rebind */
1972 guarantee_online_mems(cs, &newmems);
1975 * The mpol_rebind_mm() call takes mmap_lock, which we couldn't
1976 * take while holding tasklist_lock. Forks can happen - the
1977 * mpol_dup() cpuset_being_rebound check will catch such forks,
1978 * and rebind their vma mempolicies too. Because we still hold
1979 * the global cpuset_mutex, we know that no other rebind effort
1980 * will be contending for the global variable cpuset_being_rebound.
1981 * It's ok if we rebind the same mm twice; mpol_rebind_mm()
1982 * is idempotent. Also migrate pages in each mm to new nodes.
1984 css_task_iter_start(&cs->css, 0, &it);
1985 while ((task = css_task_iter_next(&it))) {
1986 struct mm_struct *mm;
1989 cpuset_change_task_nodemask(task, &newmems);
1991 mm = get_task_mm(task);
1995 migrate = is_memory_migrate(cs);
1997 mpol_rebind_mm(mm, &cs->mems_allowed);
1999 cpuset_migrate_mm(mm, &cs->old_mems_allowed, &newmems);
2003 css_task_iter_end(&it);
2006 * All the tasks' nodemasks have been updated, update
2007 * cs->old_mems_allowed.
2009 cs->old_mems_allowed = newmems;
2011 /* We're done rebinding vmas to this cpuset's new mems_allowed. */
2012 cpuset_being_rebound = NULL;
2016 * update_nodemasks_hier - Update effective nodemasks and tasks in the subtree
2017 * @cs: the cpuset to consider
2018 * @new_mems: a temp variable for calculating new effective_mems
2020 * When configured nodemask is changed, the effective nodemasks of this cpuset
2021 * and all its descendants need to be updated.
2023 * On legacy hierarchy, effective_mems will be the same with mems_allowed.
2025 * Called with cpuset_mutex held
2027 static void update_nodemasks_hier(struct cpuset *cs, nodemask_t *new_mems)
2030 struct cgroup_subsys_state *pos_css;
2033 cpuset_for_each_descendant_pre(cp, pos_css, cs) {
2034 struct cpuset *parent = parent_cs(cp);
2036 nodes_and(*new_mems, cp->mems_allowed, parent->effective_mems);
2039 * If it becomes empty, inherit the effective mask of the
2040 * parent, which is guaranteed to have some MEMs.
2042 if (is_in_v2_mode() && nodes_empty(*new_mems))
2043 *new_mems = parent->effective_mems;
2045 /* Skip the whole subtree if the nodemask remains the same. */
2046 if (nodes_equal(*new_mems, cp->effective_mems)) {
2047 pos_css = css_rightmost_descendant(pos_css);
2051 if (!css_tryget_online(&cp->css))
2055 spin_lock_irq(&callback_lock);
2056 cp->effective_mems = *new_mems;
2057 spin_unlock_irq(&callback_lock);
2059 WARN_ON(!is_in_v2_mode() &&
2060 !nodes_equal(cp->mems_allowed, cp->effective_mems));
2062 update_tasks_nodemask(cp);
2071 * Handle user request to change the 'mems' memory placement
2072 * of a cpuset. Needs to validate the request, update the
2073 * cpusets mems_allowed, and for each task in the cpuset,
2074 * update mems_allowed and rebind task's mempolicy and any vma
2075 * mempolicies and if the cpuset is marked 'memory_migrate',
2076 * migrate the tasks pages to the new memory.
2078 * Call with cpuset_mutex held. May take callback_lock during call.
2079 * Will take tasklist_lock, scan tasklist for tasks in cpuset cs,
2080 * lock each such tasks mm->mmap_lock, scan its vma's and rebind
2081 * their mempolicies to the cpusets new mems_allowed.
2083 static int update_nodemask(struct cpuset *cs, struct cpuset *trialcs,
2089 * top_cpuset.mems_allowed tracks node_stats[N_MEMORY];
2092 if (cs == &top_cpuset) {
2098 * An empty mems_allowed is ok iff there are no tasks in the cpuset.
2099 * Since nodelist_parse() fails on an empty mask, we special case
2100 * that parsing. The validate_change() call ensures that cpusets
2101 * with tasks have memory.
2104 nodes_clear(trialcs->mems_allowed);
2106 retval = nodelist_parse(buf, trialcs->mems_allowed);
2110 if (!nodes_subset(trialcs->mems_allowed,
2111 top_cpuset.mems_allowed)) {
2117 if (nodes_equal(cs->mems_allowed, trialcs->mems_allowed)) {
2118 retval = 0; /* Too easy - nothing to do */
2121 retval = validate_change(cs, trialcs);
2125 check_insane_mems_config(&trialcs->mems_allowed);
2127 spin_lock_irq(&callback_lock);
2128 cs->mems_allowed = trialcs->mems_allowed;
2129 spin_unlock_irq(&callback_lock);
2131 /* use trialcs->mems_allowed as a temp variable */
2132 update_nodemasks_hier(cs, &trialcs->mems_allowed);
2137 bool current_cpuset_is_being_rebound(void)
2142 ret = task_cs(current) == cpuset_being_rebound;
2148 static int update_relax_domain_level(struct cpuset *cs, s64 val)
2151 if (val < -1 || val >= sched_domain_level_max)
2155 if (val != cs->relax_domain_level) {
2156 cs->relax_domain_level = val;
2157 if (!cpumask_empty(cs->cpus_allowed) &&
2158 is_sched_load_balance(cs))
2159 rebuild_sched_domains_locked();
2166 * update_tasks_flags - update the spread flags of tasks in the cpuset.
2167 * @cs: the cpuset in which each task's spread flags needs to be changed
2169 * Iterate through each task of @cs updating its spread flags. As this
2170 * function is called with cpuset_mutex held, cpuset membership stays
2173 static void update_tasks_flags(struct cpuset *cs)
2175 struct css_task_iter it;
2176 struct task_struct *task;
2178 css_task_iter_start(&cs->css, 0, &it);
2179 while ((task = css_task_iter_next(&it)))
2180 cpuset_update_task_spread_flags(cs, task);
2181 css_task_iter_end(&it);
2185 * update_flag - read a 0 or a 1 in a file and update associated flag
2186 * bit: the bit to update (see cpuset_flagbits_t)
2187 * cs: the cpuset to update
2188 * turning_on: whether the flag is being set or cleared
2190 * Call with cpuset_mutex held.
2193 static int update_flag(cpuset_flagbits_t bit, struct cpuset *cs,
2196 struct cpuset *trialcs;
2197 int balance_flag_changed;
2198 int spread_flag_changed;
2201 trialcs = alloc_trial_cpuset(cs);
2206 set_bit(bit, &trialcs->flags);
2208 clear_bit(bit, &trialcs->flags);
2210 err = validate_change(cs, trialcs);
2214 balance_flag_changed = (is_sched_load_balance(cs) !=
2215 is_sched_load_balance(trialcs));
2217 spread_flag_changed = ((is_spread_slab(cs) != is_spread_slab(trialcs))
2218 || (is_spread_page(cs) != is_spread_page(trialcs)));
2220 spin_lock_irq(&callback_lock);
2221 cs->flags = trialcs->flags;
2222 spin_unlock_irq(&callback_lock);
2224 if (!cpumask_empty(trialcs->cpus_allowed) && balance_flag_changed)
2225 rebuild_sched_domains_locked();
2227 if (spread_flag_changed)
2228 update_tasks_flags(cs);
2230 free_cpuset(trialcs);
2235 * update_prstate - update partition_root_state
2236 * @cs: the cpuset to update
2237 * @new_prs: new partition root state
2238 * Return: 0 if successful, != 0 if error
2240 * Call with cpuset_mutex held.
2242 static int update_prstate(struct cpuset *cs, int new_prs)
2244 int err = PERR_NONE, old_prs = cs->partition_root_state;
2245 bool sched_domain_rebuilt = false;
2246 struct cpuset *parent = parent_cs(cs);
2247 struct tmpmasks tmpmask;
2249 if (old_prs == new_prs)
2253 * For a previously invalid partition root, leave it at being
2254 * invalid if new_prs is not "member".
2256 if (new_prs && is_prs_invalid(old_prs)) {
2257 cs->partition_root_state = -new_prs;
2261 if (alloc_cpumasks(NULL, &tmpmask))
2266 * Turning on partition root requires setting the
2267 * CS_CPU_EXCLUSIVE bit implicitly as well and cpus_allowed
2270 if (cpumask_empty(cs->cpus_allowed)) {
2271 err = PERR_CPUSEMPTY;
2275 err = update_flag(CS_CPU_EXCLUSIVE, cs, 1);
2281 err = update_parent_subparts_cpumask(cs, partcmd_enable,
2284 update_flag(CS_CPU_EXCLUSIVE, cs, 0);
2288 if (new_prs == PRS_ISOLATED) {
2290 * Disable the load balance flag should not return an
2291 * error unless the system is running out of memory.
2293 update_flag(CS_SCHED_LOAD_BALANCE, cs, 0);
2294 sched_domain_rebuilt = true;
2296 } else if (old_prs && new_prs) {
2298 * A change in load balance state only, no change in cpumasks.
2300 update_flag(CS_SCHED_LOAD_BALANCE, cs, (new_prs != PRS_ISOLATED));
2301 sched_domain_rebuilt = true;
2302 goto out; /* Sched domain is rebuilt in update_flag() */
2305 * Switching back to member is always allowed even if it
2306 * disables child partitions.
2308 update_parent_subparts_cpumask(cs, partcmd_disable, NULL,
2312 * If there are child partitions, they will all become invalid.
2314 if (unlikely(cs->nr_subparts_cpus)) {
2315 spin_lock_irq(&callback_lock);
2316 cs->nr_subparts_cpus = 0;
2317 cpumask_clear(cs->subparts_cpus);
2318 compute_effective_cpumask(cs->effective_cpus, cs, parent);
2319 spin_unlock_irq(&callback_lock);
2322 /* Turning off CS_CPU_EXCLUSIVE will not return error */
2323 update_flag(CS_CPU_EXCLUSIVE, cs, 0);
2325 if (!is_sched_load_balance(cs)) {
2326 /* Make sure load balance is on */
2327 update_flag(CS_SCHED_LOAD_BALANCE, cs, 1);
2328 sched_domain_rebuilt = true;
2332 update_tasks_cpumask(parent, tmpmask.new_cpus);
2334 if (parent->child_ecpus_count)
2335 update_sibling_cpumasks(parent, cs, &tmpmask);
2337 if (!sched_domain_rebuilt)
2338 rebuild_sched_domains_locked();
2341 * Make partition invalid if an error happen
2345 spin_lock_irq(&callback_lock);
2346 cs->partition_root_state = new_prs;
2347 WRITE_ONCE(cs->prs_err, err);
2348 spin_unlock_irq(&callback_lock);
2350 * Update child cpusets, if present.
2351 * Force update if switching back to member.
2353 if (!list_empty(&cs->css.children))
2354 update_cpumasks_hier(cs, &tmpmask, !new_prs);
2356 notify_partition_change(cs, old_prs);
2357 free_cpumasks(NULL, &tmpmask);
2362 * Frequency meter - How fast is some event occurring?
2364 * These routines manage a digitally filtered, constant time based,
2365 * event frequency meter. There are four routines:
2366 * fmeter_init() - initialize a frequency meter.
2367 * fmeter_markevent() - called each time the event happens.
2368 * fmeter_getrate() - returns the recent rate of such events.
2369 * fmeter_update() - internal routine used to update fmeter.
2371 * A common data structure is passed to each of these routines,
2372 * which is used to keep track of the state required to manage the
2373 * frequency meter and its digital filter.
2375 * The filter works on the number of events marked per unit time.
2376 * The filter is single-pole low-pass recursive (IIR). The time unit
2377 * is 1 second. Arithmetic is done using 32-bit integers scaled to
2378 * simulate 3 decimal digits of precision (multiplied by 1000).
2380 * With an FM_COEF of 933, and a time base of 1 second, the filter
2381 * has a half-life of 10 seconds, meaning that if the events quit
2382 * happening, then the rate returned from the fmeter_getrate()
2383 * will be cut in half each 10 seconds, until it converges to zero.
2385 * It is not worth doing a real infinitely recursive filter. If more
2386 * than FM_MAXTICKS ticks have elapsed since the last filter event,
2387 * just compute FM_MAXTICKS ticks worth, by which point the level
2390 * Limit the count of unprocessed events to FM_MAXCNT, so as to avoid
2391 * arithmetic overflow in the fmeter_update() routine.
2393 * Given the simple 32 bit integer arithmetic used, this meter works
2394 * best for reporting rates between one per millisecond (msec) and
2395 * one per 32 (approx) seconds. At constant rates faster than one
2396 * per msec it maxes out at values just under 1,000,000. At constant
2397 * rates between one per msec, and one per second it will stabilize
2398 * to a value N*1000, where N is the rate of events per second.
2399 * At constant rates between one per second and one per 32 seconds,
2400 * it will be choppy, moving up on the seconds that have an event,
2401 * and then decaying until the next event. At rates slower than
2402 * about one in 32 seconds, it decays all the way back to zero between
2406 #define FM_COEF 933 /* coefficient for half-life of 10 secs */
2407 #define FM_MAXTICKS ((u32)99) /* useless computing more ticks than this */
2408 #define FM_MAXCNT 1000000 /* limit cnt to avoid overflow */
2409 #define FM_SCALE 1000 /* faux fixed point scale */
2411 /* Initialize a frequency meter */
2412 static void fmeter_init(struct fmeter *fmp)
2417 spin_lock_init(&fmp->lock);
2420 /* Internal meter update - process cnt events and update value */
2421 static void fmeter_update(struct fmeter *fmp)
2426 now = ktime_get_seconds();
2427 ticks = now - fmp->time;
2432 ticks = min(FM_MAXTICKS, ticks);
2434 fmp->val = (FM_COEF * fmp->val) / FM_SCALE;
2437 fmp->val += ((FM_SCALE - FM_COEF) * fmp->cnt) / FM_SCALE;
2441 /* Process any previous ticks, then bump cnt by one (times scale). */
2442 static void fmeter_markevent(struct fmeter *fmp)
2444 spin_lock(&fmp->lock);
2446 fmp->cnt = min(FM_MAXCNT, fmp->cnt + FM_SCALE);
2447 spin_unlock(&fmp->lock);
2450 /* Process any previous ticks, then return current value. */
2451 static int fmeter_getrate(struct fmeter *fmp)
2455 spin_lock(&fmp->lock);
2458 spin_unlock(&fmp->lock);
2462 static struct cpuset *cpuset_attach_old_cs;
2465 * Check to see if a cpuset can accept a new task
2466 * For v1, cpus_allowed and mems_allowed can't be empty.
2467 * For v2, effective_cpus can't be empty.
2468 * Note that in v1, effective_cpus = cpus_allowed.
2470 static int cpuset_can_attach_check(struct cpuset *cs)
2472 if (cpumask_empty(cs->effective_cpus) ||
2473 (!is_in_v2_mode() && nodes_empty(cs->mems_allowed)))
2478 static void reset_migrate_dl_data(struct cpuset *cs)
2480 cs->nr_migrate_dl_tasks = 0;
2481 cs->sum_migrate_dl_bw = 0;
2484 /* Called by cgroups to determine if a cpuset is usable; cpuset_mutex held */
2485 static int cpuset_can_attach(struct cgroup_taskset *tset)
2487 struct cgroup_subsys_state *css;
2488 struct cpuset *cs, *oldcs;
2489 struct task_struct *task;
2492 /* used later by cpuset_attach() */
2493 cpuset_attach_old_cs = task_cs(cgroup_taskset_first(tset, &css));
2494 oldcs = cpuset_attach_old_cs;
2497 mutex_lock(&cpuset_mutex);
2499 /* Check to see if task is allowed in the cpuset */
2500 ret = cpuset_can_attach_check(cs);
2504 cgroup_taskset_for_each(task, css, tset) {
2505 ret = task_can_attach(task);
2508 ret = security_task_setscheduler(task);
2512 if (dl_task(task)) {
2513 cs->nr_migrate_dl_tasks++;
2514 cs->sum_migrate_dl_bw += task->dl.dl_bw;
2518 if (!cs->nr_migrate_dl_tasks)
2521 if (!cpumask_intersects(oldcs->effective_cpus, cs->effective_cpus)) {
2522 int cpu = cpumask_any_and(cpu_active_mask, cs->effective_cpus);
2524 if (unlikely(cpu >= nr_cpu_ids)) {
2525 reset_migrate_dl_data(cs);
2530 ret = dl_bw_alloc(cpu, cs->sum_migrate_dl_bw);
2532 reset_migrate_dl_data(cs);
2539 * Mark attach is in progress. This makes validate_change() fail
2540 * changes which zero cpus/mems_allowed.
2542 cs->attach_in_progress++;
2544 mutex_unlock(&cpuset_mutex);
2548 static void cpuset_cancel_attach(struct cgroup_taskset *tset)
2550 struct cgroup_subsys_state *css;
2553 cgroup_taskset_first(tset, &css);
2556 mutex_lock(&cpuset_mutex);
2557 cs->attach_in_progress--;
2558 if (!cs->attach_in_progress)
2559 wake_up(&cpuset_attach_wq);
2561 if (cs->nr_migrate_dl_tasks) {
2562 int cpu = cpumask_any(cs->effective_cpus);
2564 dl_bw_free(cpu, cs->sum_migrate_dl_bw);
2565 reset_migrate_dl_data(cs);
2568 mutex_unlock(&cpuset_mutex);
2572 * Protected by cpuset_mutex. cpus_attach is used only by cpuset_attach_task()
2573 * but we can't allocate it dynamically there. Define it global and
2574 * allocate from cpuset_init().
2576 static cpumask_var_t cpus_attach;
2577 static nodemask_t cpuset_attach_nodemask_to;
2579 static void cpuset_attach_task(struct cpuset *cs, struct task_struct *task)
2581 lockdep_assert_held(&cpuset_mutex);
2583 if (cs != &top_cpuset)
2584 guarantee_online_cpus(task, cpus_attach);
2586 cpumask_andnot(cpus_attach, task_cpu_possible_mask(task),
2589 * can_attach beforehand should guarantee that this doesn't
2590 * fail. TODO: have a better way to handle failure here
2592 WARN_ON_ONCE(set_cpus_allowed_ptr(task, cpus_attach));
2594 cpuset_change_task_nodemask(task, &cpuset_attach_nodemask_to);
2595 cpuset_update_task_spread_flags(cs, task);
2598 static void cpuset_attach(struct cgroup_taskset *tset)
2600 struct task_struct *task;
2601 struct task_struct *leader;
2602 struct cgroup_subsys_state *css;
2604 struct cpuset *oldcs = cpuset_attach_old_cs;
2605 bool cpus_updated, mems_updated;
2607 cgroup_taskset_first(tset, &css);
2610 lockdep_assert_cpus_held(); /* see cgroup_attach_lock() */
2611 mutex_lock(&cpuset_mutex);
2612 cpus_updated = !cpumask_equal(cs->effective_cpus,
2613 oldcs->effective_cpus);
2614 mems_updated = !nodes_equal(cs->effective_mems, oldcs->effective_mems);
2617 * In the default hierarchy, enabling cpuset in the child cgroups
2618 * will trigger a number of cpuset_attach() calls with no change
2619 * in effective cpus and mems. In that case, we can optimize out
2620 * by skipping the task iteration and update.
2622 if (cgroup_subsys_on_dfl(cpuset_cgrp_subsys) &&
2623 !cpus_updated && !mems_updated) {
2624 cpuset_attach_nodemask_to = cs->effective_mems;
2628 guarantee_online_mems(cs, &cpuset_attach_nodemask_to);
2630 cgroup_taskset_for_each(task, css, tset)
2631 cpuset_attach_task(cs, task);
2634 * Change mm for all threadgroup leaders. This is expensive and may
2635 * sleep and should be moved outside migration path proper. Skip it
2636 * if there is no change in effective_mems and CS_MEMORY_MIGRATE is
2639 cpuset_attach_nodemask_to = cs->effective_mems;
2640 if (!is_memory_migrate(cs) && !mems_updated)
2643 cgroup_taskset_for_each_leader(leader, css, tset) {
2644 struct mm_struct *mm = get_task_mm(leader);
2647 mpol_rebind_mm(mm, &cpuset_attach_nodemask_to);
2650 * old_mems_allowed is the same with mems_allowed
2651 * here, except if this task is being moved
2652 * automatically due to hotplug. In that case
2653 * @mems_allowed has been updated and is empty, so
2654 * @old_mems_allowed is the right nodesets that we
2657 if (is_memory_migrate(cs))
2658 cpuset_migrate_mm(mm, &oldcs->old_mems_allowed,
2659 &cpuset_attach_nodemask_to);
2666 cs->old_mems_allowed = cpuset_attach_nodemask_to;
2668 if (cs->nr_migrate_dl_tasks) {
2669 cs->nr_deadline_tasks += cs->nr_migrate_dl_tasks;
2670 oldcs->nr_deadline_tasks -= cs->nr_migrate_dl_tasks;
2671 reset_migrate_dl_data(cs);
2674 cs->attach_in_progress--;
2675 if (!cs->attach_in_progress)
2676 wake_up(&cpuset_attach_wq);
2678 mutex_unlock(&cpuset_mutex);
2681 /* The various types of files and directories in a cpuset file system */
2684 FILE_MEMORY_MIGRATE,
2687 FILE_EFFECTIVE_CPULIST,
2688 FILE_EFFECTIVE_MEMLIST,
2689 FILE_SUBPARTS_CPULIST,
2693 FILE_SCHED_LOAD_BALANCE,
2694 FILE_PARTITION_ROOT,
2695 FILE_SCHED_RELAX_DOMAIN_LEVEL,
2696 FILE_MEMORY_PRESSURE_ENABLED,
2697 FILE_MEMORY_PRESSURE,
2700 } cpuset_filetype_t;
2702 static int cpuset_write_u64(struct cgroup_subsys_state *css, struct cftype *cft,
2705 struct cpuset *cs = css_cs(css);
2706 cpuset_filetype_t type = cft->private;
2710 mutex_lock(&cpuset_mutex);
2711 if (!is_cpuset_online(cs)) {
2717 case FILE_CPU_EXCLUSIVE:
2718 retval = update_flag(CS_CPU_EXCLUSIVE, cs, val);
2720 case FILE_MEM_EXCLUSIVE:
2721 retval = update_flag(CS_MEM_EXCLUSIVE, cs, val);
2723 case FILE_MEM_HARDWALL:
2724 retval = update_flag(CS_MEM_HARDWALL, cs, val);
2726 case FILE_SCHED_LOAD_BALANCE:
2727 retval = update_flag(CS_SCHED_LOAD_BALANCE, cs, val);
2729 case FILE_MEMORY_MIGRATE:
2730 retval = update_flag(CS_MEMORY_MIGRATE, cs, val);
2732 case FILE_MEMORY_PRESSURE_ENABLED:
2733 cpuset_memory_pressure_enabled = !!val;
2735 case FILE_SPREAD_PAGE:
2736 retval = update_flag(CS_SPREAD_PAGE, cs, val);
2738 case FILE_SPREAD_SLAB:
2739 retval = update_flag(CS_SPREAD_SLAB, cs, val);
2746 mutex_unlock(&cpuset_mutex);
2751 static int cpuset_write_s64(struct cgroup_subsys_state *css, struct cftype *cft,
2754 struct cpuset *cs = css_cs(css);
2755 cpuset_filetype_t type = cft->private;
2756 int retval = -ENODEV;
2759 mutex_lock(&cpuset_mutex);
2760 if (!is_cpuset_online(cs))
2764 case FILE_SCHED_RELAX_DOMAIN_LEVEL:
2765 retval = update_relax_domain_level(cs, val);
2772 mutex_unlock(&cpuset_mutex);
2778 * Common handling for a write to a "cpus" or "mems" file.
2780 static ssize_t cpuset_write_resmask(struct kernfs_open_file *of,
2781 char *buf, size_t nbytes, loff_t off)
2783 struct cpuset *cs = css_cs(of_css(of));
2784 struct cpuset *trialcs;
2785 int retval = -ENODEV;
2787 buf = strstrip(buf);
2790 * CPU or memory hotunplug may leave @cs w/o any execution
2791 * resources, in which case the hotplug code asynchronously updates
2792 * configuration and transfers all tasks to the nearest ancestor
2793 * which can execute.
2795 * As writes to "cpus" or "mems" may restore @cs's execution
2796 * resources, wait for the previously scheduled operations before
2797 * proceeding, so that we don't end up keep removing tasks added
2798 * after execution capability is restored.
2800 * cpuset_hotplug_work calls back into cgroup core via
2801 * cgroup_transfer_tasks() and waiting for it from a cgroupfs
2802 * operation like this one can lead to a deadlock through kernfs
2803 * active_ref protection. Let's break the protection. Losing the
2804 * protection is okay as we check whether @cs is online after
2805 * grabbing cpuset_mutex anyway. This only happens on the legacy
2809 kernfs_break_active_protection(of->kn);
2810 flush_work(&cpuset_hotplug_work);
2813 mutex_lock(&cpuset_mutex);
2814 if (!is_cpuset_online(cs))
2817 trialcs = alloc_trial_cpuset(cs);
2823 switch (of_cft(of)->private) {
2825 retval = update_cpumask(cs, trialcs, buf);
2828 retval = update_nodemask(cs, trialcs, buf);
2835 free_cpuset(trialcs);
2837 mutex_unlock(&cpuset_mutex);
2839 kernfs_unbreak_active_protection(of->kn);
2841 flush_workqueue(cpuset_migrate_mm_wq);
2842 return retval ?: nbytes;
2846 * These ascii lists should be read in a single call, by using a user
2847 * buffer large enough to hold the entire map. If read in smaller
2848 * chunks, there is no guarantee of atomicity. Since the display format
2849 * used, list of ranges of sequential numbers, is variable length,
2850 * and since these maps can change value dynamically, one could read
2851 * gibberish by doing partial reads while a list was changing.
2853 static int cpuset_common_seq_show(struct seq_file *sf, void *v)
2855 struct cpuset *cs = css_cs(seq_css(sf));
2856 cpuset_filetype_t type = seq_cft(sf)->private;
2859 spin_lock_irq(&callback_lock);
2863 seq_printf(sf, "%*pbl\n", cpumask_pr_args(cs->cpus_allowed));
2866 seq_printf(sf, "%*pbl\n", nodemask_pr_args(&cs->mems_allowed));
2868 case FILE_EFFECTIVE_CPULIST:
2869 seq_printf(sf, "%*pbl\n", cpumask_pr_args(cs->effective_cpus));
2871 case FILE_EFFECTIVE_MEMLIST:
2872 seq_printf(sf, "%*pbl\n", nodemask_pr_args(&cs->effective_mems));
2874 case FILE_SUBPARTS_CPULIST:
2875 seq_printf(sf, "%*pbl\n", cpumask_pr_args(cs->subparts_cpus));
2881 spin_unlock_irq(&callback_lock);
2885 static u64 cpuset_read_u64(struct cgroup_subsys_state *css, struct cftype *cft)
2887 struct cpuset *cs = css_cs(css);
2888 cpuset_filetype_t type = cft->private;
2890 case FILE_CPU_EXCLUSIVE:
2891 return is_cpu_exclusive(cs);
2892 case FILE_MEM_EXCLUSIVE:
2893 return is_mem_exclusive(cs);
2894 case FILE_MEM_HARDWALL:
2895 return is_mem_hardwall(cs);
2896 case FILE_SCHED_LOAD_BALANCE:
2897 return is_sched_load_balance(cs);
2898 case FILE_MEMORY_MIGRATE:
2899 return is_memory_migrate(cs);
2900 case FILE_MEMORY_PRESSURE_ENABLED:
2901 return cpuset_memory_pressure_enabled;
2902 case FILE_MEMORY_PRESSURE:
2903 return fmeter_getrate(&cs->fmeter);
2904 case FILE_SPREAD_PAGE:
2905 return is_spread_page(cs);
2906 case FILE_SPREAD_SLAB:
2907 return is_spread_slab(cs);
2912 /* Unreachable but makes gcc happy */
2916 static s64 cpuset_read_s64(struct cgroup_subsys_state *css, struct cftype *cft)
2918 struct cpuset *cs = css_cs(css);
2919 cpuset_filetype_t type = cft->private;
2921 case FILE_SCHED_RELAX_DOMAIN_LEVEL:
2922 return cs->relax_domain_level;
2927 /* Unreachable but makes gcc happy */
2931 static int sched_partition_show(struct seq_file *seq, void *v)
2933 struct cpuset *cs = css_cs(seq_css(seq));
2934 const char *err, *type = NULL;
2936 switch (cs->partition_root_state) {
2938 seq_puts(seq, "root\n");
2941 seq_puts(seq, "isolated\n");
2944 seq_puts(seq, "member\n");
2946 case PRS_INVALID_ROOT:
2949 case PRS_INVALID_ISOLATED:
2952 err = perr_strings[READ_ONCE(cs->prs_err)];
2954 seq_printf(seq, "%s invalid (%s)\n", type, err);
2956 seq_printf(seq, "%s invalid\n", type);
2962 static ssize_t sched_partition_write(struct kernfs_open_file *of, char *buf,
2963 size_t nbytes, loff_t off)
2965 struct cpuset *cs = css_cs(of_css(of));
2967 int retval = -ENODEV;
2969 buf = strstrip(buf);
2972 * Convert "root" to ENABLED, and convert "member" to DISABLED.
2974 if (!strcmp(buf, "root"))
2976 else if (!strcmp(buf, "member"))
2978 else if (!strcmp(buf, "isolated"))
2985 mutex_lock(&cpuset_mutex);
2986 if (!is_cpuset_online(cs))
2989 retval = update_prstate(cs, val);
2991 mutex_unlock(&cpuset_mutex);
2994 return retval ?: nbytes;
2998 * for the common functions, 'private' gives the type of file
3001 static struct cftype legacy_files[] = {
3004 .seq_show = cpuset_common_seq_show,
3005 .write = cpuset_write_resmask,
3006 .max_write_len = (100U + 6 * NR_CPUS),
3007 .private = FILE_CPULIST,
3012 .seq_show = cpuset_common_seq_show,
3013 .write = cpuset_write_resmask,
3014 .max_write_len = (100U + 6 * MAX_NUMNODES),
3015 .private = FILE_MEMLIST,
3019 .name = "effective_cpus",
3020 .seq_show = cpuset_common_seq_show,
3021 .private = FILE_EFFECTIVE_CPULIST,
3025 .name = "effective_mems",
3026 .seq_show = cpuset_common_seq_show,
3027 .private = FILE_EFFECTIVE_MEMLIST,
3031 .name = "cpu_exclusive",
3032 .read_u64 = cpuset_read_u64,
3033 .write_u64 = cpuset_write_u64,
3034 .private = FILE_CPU_EXCLUSIVE,
3038 .name = "mem_exclusive",
3039 .read_u64 = cpuset_read_u64,
3040 .write_u64 = cpuset_write_u64,
3041 .private = FILE_MEM_EXCLUSIVE,
3045 .name = "mem_hardwall",
3046 .read_u64 = cpuset_read_u64,
3047 .write_u64 = cpuset_write_u64,
3048 .private = FILE_MEM_HARDWALL,
3052 .name = "sched_load_balance",
3053 .read_u64 = cpuset_read_u64,
3054 .write_u64 = cpuset_write_u64,
3055 .private = FILE_SCHED_LOAD_BALANCE,
3059 .name = "sched_relax_domain_level",
3060 .read_s64 = cpuset_read_s64,
3061 .write_s64 = cpuset_write_s64,
3062 .private = FILE_SCHED_RELAX_DOMAIN_LEVEL,
3066 .name = "memory_migrate",
3067 .read_u64 = cpuset_read_u64,
3068 .write_u64 = cpuset_write_u64,
3069 .private = FILE_MEMORY_MIGRATE,
3073 .name = "memory_pressure",
3074 .read_u64 = cpuset_read_u64,
3075 .private = FILE_MEMORY_PRESSURE,
3079 .name = "memory_spread_page",
3080 .read_u64 = cpuset_read_u64,
3081 .write_u64 = cpuset_write_u64,
3082 .private = FILE_SPREAD_PAGE,
3086 .name = "memory_spread_slab",
3087 .read_u64 = cpuset_read_u64,
3088 .write_u64 = cpuset_write_u64,
3089 .private = FILE_SPREAD_SLAB,
3093 .name = "memory_pressure_enabled",
3094 .flags = CFTYPE_ONLY_ON_ROOT,
3095 .read_u64 = cpuset_read_u64,
3096 .write_u64 = cpuset_write_u64,
3097 .private = FILE_MEMORY_PRESSURE_ENABLED,
3104 * This is currently a minimal set for the default hierarchy. It can be
3105 * expanded later on by migrating more features and control files from v1.
3107 static struct cftype dfl_files[] = {
3110 .seq_show = cpuset_common_seq_show,
3111 .write = cpuset_write_resmask,
3112 .max_write_len = (100U + 6 * NR_CPUS),
3113 .private = FILE_CPULIST,
3114 .flags = CFTYPE_NOT_ON_ROOT,
3119 .seq_show = cpuset_common_seq_show,
3120 .write = cpuset_write_resmask,
3121 .max_write_len = (100U + 6 * MAX_NUMNODES),
3122 .private = FILE_MEMLIST,
3123 .flags = CFTYPE_NOT_ON_ROOT,
3127 .name = "cpus.effective",
3128 .seq_show = cpuset_common_seq_show,
3129 .private = FILE_EFFECTIVE_CPULIST,
3133 .name = "mems.effective",
3134 .seq_show = cpuset_common_seq_show,
3135 .private = FILE_EFFECTIVE_MEMLIST,
3139 .name = "cpus.partition",
3140 .seq_show = sched_partition_show,
3141 .write = sched_partition_write,
3142 .private = FILE_PARTITION_ROOT,
3143 .flags = CFTYPE_NOT_ON_ROOT,
3144 .file_offset = offsetof(struct cpuset, partition_file),
3148 .name = "cpus.subpartitions",
3149 .seq_show = cpuset_common_seq_show,
3150 .private = FILE_SUBPARTS_CPULIST,
3151 .flags = CFTYPE_DEBUG,
3159 * cpuset_css_alloc - Allocate a cpuset css
3160 * @parent_css: Parent css of the control group that the new cpuset will be
3162 * Return: cpuset css on success, -ENOMEM on failure.
3164 * Allocate and initialize a new cpuset css, for non-NULL @parent_css, return
3165 * top cpuset css otherwise.
3167 static struct cgroup_subsys_state *
3168 cpuset_css_alloc(struct cgroup_subsys_state *parent_css)
3173 return &top_cpuset.css;
3175 cs = kzalloc(sizeof(*cs), GFP_KERNEL);
3177 return ERR_PTR(-ENOMEM);
3179 if (alloc_cpumasks(cs, NULL)) {
3181 return ERR_PTR(-ENOMEM);
3184 __set_bit(CS_SCHED_LOAD_BALANCE, &cs->flags);
3185 nodes_clear(cs->mems_allowed);
3186 nodes_clear(cs->effective_mems);
3187 fmeter_init(&cs->fmeter);
3188 cs->relax_domain_level = -1;
3190 /* Set CS_MEMORY_MIGRATE for default hierarchy */
3191 if (cgroup_subsys_on_dfl(cpuset_cgrp_subsys))
3192 __set_bit(CS_MEMORY_MIGRATE, &cs->flags);
3197 static int cpuset_css_online(struct cgroup_subsys_state *css)
3199 struct cpuset *cs = css_cs(css);
3200 struct cpuset *parent = parent_cs(cs);
3201 struct cpuset *tmp_cs;
3202 struct cgroup_subsys_state *pos_css;
3208 mutex_lock(&cpuset_mutex);
3210 set_bit(CS_ONLINE, &cs->flags);
3211 if (is_spread_page(parent))
3212 set_bit(CS_SPREAD_PAGE, &cs->flags);
3213 if (is_spread_slab(parent))
3214 set_bit(CS_SPREAD_SLAB, &cs->flags);
3218 spin_lock_irq(&callback_lock);
3219 if (is_in_v2_mode()) {
3220 cpumask_copy(cs->effective_cpus, parent->effective_cpus);
3221 cs->effective_mems = parent->effective_mems;
3222 cs->use_parent_ecpus = true;
3223 parent->child_ecpus_count++;
3225 spin_unlock_irq(&callback_lock);
3227 if (!test_bit(CGRP_CPUSET_CLONE_CHILDREN, &css->cgroup->flags))
3231 * Clone @parent's configuration if CGRP_CPUSET_CLONE_CHILDREN is
3232 * set. This flag handling is implemented in cgroup core for
3233 * historical reasons - the flag may be specified during mount.
3235 * Currently, if any sibling cpusets have exclusive cpus or mem, we
3236 * refuse to clone the configuration - thereby refusing the task to
3237 * be entered, and as a result refusing the sys_unshare() or
3238 * clone() which initiated it. If this becomes a problem for some
3239 * users who wish to allow that scenario, then this could be
3240 * changed to grant parent->cpus_allowed-sibling_cpus_exclusive
3241 * (and likewise for mems) to the new cgroup.
3244 cpuset_for_each_child(tmp_cs, pos_css, parent) {
3245 if (is_mem_exclusive(tmp_cs) || is_cpu_exclusive(tmp_cs)) {
3252 spin_lock_irq(&callback_lock);
3253 cs->mems_allowed = parent->mems_allowed;
3254 cs->effective_mems = parent->mems_allowed;
3255 cpumask_copy(cs->cpus_allowed, parent->cpus_allowed);
3256 cpumask_copy(cs->effective_cpus, parent->cpus_allowed);
3257 spin_unlock_irq(&callback_lock);
3259 mutex_unlock(&cpuset_mutex);
3265 * If the cpuset being removed has its flag 'sched_load_balance'
3266 * enabled, then simulate turning sched_load_balance off, which
3267 * will call rebuild_sched_domains_locked(). That is not needed
3268 * in the default hierarchy where only changes in partition
3269 * will cause repartitioning.
3271 * If the cpuset has the 'sched.partition' flag enabled, simulate
3272 * turning 'sched.partition" off.
3275 static void cpuset_css_offline(struct cgroup_subsys_state *css)
3277 struct cpuset *cs = css_cs(css);
3280 mutex_lock(&cpuset_mutex);
3282 if (is_partition_valid(cs))
3283 update_prstate(cs, 0);
3285 if (!cgroup_subsys_on_dfl(cpuset_cgrp_subsys) &&
3286 is_sched_load_balance(cs))
3287 update_flag(CS_SCHED_LOAD_BALANCE, cs, 0);
3289 if (cs->use_parent_ecpus) {
3290 struct cpuset *parent = parent_cs(cs);
3292 cs->use_parent_ecpus = false;
3293 parent->child_ecpus_count--;
3297 clear_bit(CS_ONLINE, &cs->flags);
3299 mutex_unlock(&cpuset_mutex);
3303 static void cpuset_css_free(struct cgroup_subsys_state *css)
3305 struct cpuset *cs = css_cs(css);
3310 static void cpuset_bind(struct cgroup_subsys_state *root_css)
3312 mutex_lock(&cpuset_mutex);
3313 spin_lock_irq(&callback_lock);
3315 if (is_in_v2_mode()) {
3316 cpumask_copy(top_cpuset.cpus_allowed, cpu_possible_mask);
3317 top_cpuset.mems_allowed = node_possible_map;
3319 cpumask_copy(top_cpuset.cpus_allowed,
3320 top_cpuset.effective_cpus);
3321 top_cpuset.mems_allowed = top_cpuset.effective_mems;
3324 spin_unlock_irq(&callback_lock);
3325 mutex_unlock(&cpuset_mutex);
3329 * In case the child is cloned into a cpuset different from its parent,
3330 * additional checks are done to see if the move is allowed.
3332 static int cpuset_can_fork(struct task_struct *task, struct css_set *cset)
3334 struct cpuset *cs = css_cs(cset->subsys[cpuset_cgrp_id]);
3339 same_cs = (cs == task_cs(current));
3345 lockdep_assert_held(&cgroup_mutex);
3346 mutex_lock(&cpuset_mutex);
3348 /* Check to see if task is allowed in the cpuset */
3349 ret = cpuset_can_attach_check(cs);
3353 ret = task_can_attach(task);
3357 ret = security_task_setscheduler(task);
3362 * Mark attach is in progress. This makes validate_change() fail
3363 * changes which zero cpus/mems_allowed.
3365 cs->attach_in_progress++;
3367 mutex_unlock(&cpuset_mutex);
3371 static void cpuset_cancel_fork(struct task_struct *task, struct css_set *cset)
3373 struct cpuset *cs = css_cs(cset->subsys[cpuset_cgrp_id]);
3377 same_cs = (cs == task_cs(current));
3383 mutex_lock(&cpuset_mutex);
3384 cs->attach_in_progress--;
3385 if (!cs->attach_in_progress)
3386 wake_up(&cpuset_attach_wq);
3387 mutex_unlock(&cpuset_mutex);
3391 * Make sure the new task conform to the current state of its parent,
3392 * which could have been changed by cpuset just after it inherits the
3393 * state from the parent and before it sits on the cgroup's task list.
3395 static void cpuset_fork(struct task_struct *task)
3402 same_cs = (cs == task_cs(current));
3406 if (cs == &top_cpuset)
3409 set_cpus_allowed_ptr(task, current->cpus_ptr);
3410 task->mems_allowed = current->mems_allowed;
3414 /* CLONE_INTO_CGROUP */
3415 mutex_lock(&cpuset_mutex);
3416 guarantee_online_mems(cs, &cpuset_attach_nodemask_to);
3417 cpuset_attach_task(cs, task);
3419 cs->attach_in_progress--;
3420 if (!cs->attach_in_progress)
3421 wake_up(&cpuset_attach_wq);
3423 mutex_unlock(&cpuset_mutex);
3426 struct cgroup_subsys cpuset_cgrp_subsys = {
3427 .css_alloc = cpuset_css_alloc,
3428 .css_online = cpuset_css_online,
3429 .css_offline = cpuset_css_offline,
3430 .css_free = cpuset_css_free,
3431 .can_attach = cpuset_can_attach,
3432 .cancel_attach = cpuset_cancel_attach,
3433 .attach = cpuset_attach,
3434 .post_attach = cpuset_post_attach,
3435 .bind = cpuset_bind,
3436 .can_fork = cpuset_can_fork,
3437 .cancel_fork = cpuset_cancel_fork,
3438 .fork = cpuset_fork,
3439 .legacy_cftypes = legacy_files,
3440 .dfl_cftypes = dfl_files,
3446 * cpuset_init - initialize cpusets at system boot
3448 * Description: Initialize top_cpuset
3451 int __init cpuset_init(void)
3453 BUG_ON(!alloc_cpumask_var(&top_cpuset.cpus_allowed, GFP_KERNEL));
3454 BUG_ON(!alloc_cpumask_var(&top_cpuset.effective_cpus, GFP_KERNEL));
3455 BUG_ON(!zalloc_cpumask_var(&top_cpuset.subparts_cpus, GFP_KERNEL));
3457 cpumask_setall(top_cpuset.cpus_allowed);
3458 nodes_setall(top_cpuset.mems_allowed);
3459 cpumask_setall(top_cpuset.effective_cpus);
3460 nodes_setall(top_cpuset.effective_mems);
3462 fmeter_init(&top_cpuset.fmeter);
3463 set_bit(CS_SCHED_LOAD_BALANCE, &top_cpuset.flags);
3464 top_cpuset.relax_domain_level = -1;
3466 BUG_ON(!alloc_cpumask_var(&cpus_attach, GFP_KERNEL));
3472 * If CPU and/or memory hotplug handlers, below, unplug any CPUs
3473 * or memory nodes, we need to walk over the cpuset hierarchy,
3474 * removing that CPU or node from all cpusets. If this removes the
3475 * last CPU or node from a cpuset, then move the tasks in the empty
3476 * cpuset to its next-highest non-empty parent.
3478 static void remove_tasks_in_empty_cpuset(struct cpuset *cs)
3480 struct cpuset *parent;
3483 * Find its next-highest non-empty parent, (top cpuset
3484 * has online cpus, so can't be empty).
3486 parent = parent_cs(cs);
3487 while (cpumask_empty(parent->cpus_allowed) ||
3488 nodes_empty(parent->mems_allowed))
3489 parent = parent_cs(parent);
3491 if (cgroup_transfer_tasks(parent->css.cgroup, cs->css.cgroup)) {
3492 pr_err("cpuset: failed to transfer tasks out of empty cpuset ");
3493 pr_cont_cgroup_name(cs->css.cgroup);
3499 hotplug_update_tasks_legacy(struct cpuset *cs,
3500 struct cpumask *new_cpus, nodemask_t *new_mems,
3501 bool cpus_updated, bool mems_updated)
3505 spin_lock_irq(&callback_lock);
3506 cpumask_copy(cs->cpus_allowed, new_cpus);
3507 cpumask_copy(cs->effective_cpus, new_cpus);
3508 cs->mems_allowed = *new_mems;
3509 cs->effective_mems = *new_mems;
3510 spin_unlock_irq(&callback_lock);
3513 * Don't call update_tasks_cpumask() if the cpuset becomes empty,
3514 * as the tasks will be migrated to an ancestor.
3516 if (cpus_updated && !cpumask_empty(cs->cpus_allowed))
3517 update_tasks_cpumask(cs, new_cpus);
3518 if (mems_updated && !nodes_empty(cs->mems_allowed))
3519 update_tasks_nodemask(cs);
3521 is_empty = cpumask_empty(cs->cpus_allowed) ||
3522 nodes_empty(cs->mems_allowed);
3524 mutex_unlock(&cpuset_mutex);
3527 * Move tasks to the nearest ancestor with execution resources,
3528 * This is full cgroup operation which will also call back into
3529 * cpuset. Should be done outside any lock.
3532 remove_tasks_in_empty_cpuset(cs);
3534 mutex_lock(&cpuset_mutex);
3538 hotplug_update_tasks(struct cpuset *cs,
3539 struct cpumask *new_cpus, nodemask_t *new_mems,
3540 bool cpus_updated, bool mems_updated)
3542 /* A partition root is allowed to have empty effective cpus */
3543 if (cpumask_empty(new_cpus) && !is_partition_valid(cs))
3544 cpumask_copy(new_cpus, parent_cs(cs)->effective_cpus);
3545 if (nodes_empty(*new_mems))
3546 *new_mems = parent_cs(cs)->effective_mems;
3548 spin_lock_irq(&callback_lock);
3549 cpumask_copy(cs->effective_cpus, new_cpus);
3550 cs->effective_mems = *new_mems;
3551 spin_unlock_irq(&callback_lock);
3554 update_tasks_cpumask(cs, new_cpus);
3556 update_tasks_nodemask(cs);
3559 static bool force_rebuild;
3561 void cpuset_force_rebuild(void)
3563 force_rebuild = true;
3567 * cpuset_hotplug_update_tasks - update tasks in a cpuset for hotunplug
3568 * @cs: cpuset in interest
3569 * @tmp: the tmpmasks structure pointer
3571 * Compare @cs's cpu and mem masks against top_cpuset and if some have gone
3572 * offline, update @cs accordingly. If @cs ends up with no CPU or memory,
3573 * all its tasks are moved to the nearest ancestor with both resources.
3575 static void cpuset_hotplug_update_tasks(struct cpuset *cs, struct tmpmasks *tmp)
3577 static cpumask_t new_cpus;
3578 static nodemask_t new_mems;
3581 struct cpuset *parent;
3583 wait_event(cpuset_attach_wq, cs->attach_in_progress == 0);
3585 mutex_lock(&cpuset_mutex);
3588 * We have raced with task attaching. We wait until attaching
3589 * is finished, so we won't attach a task to an empty cpuset.
3591 if (cs->attach_in_progress) {
3592 mutex_unlock(&cpuset_mutex);
3596 parent = parent_cs(cs);
3597 compute_effective_cpumask(&new_cpus, cs, parent);
3598 nodes_and(new_mems, cs->mems_allowed, parent->effective_mems);
3600 if (cs->nr_subparts_cpus)
3602 * Make sure that CPUs allocated to child partitions
3603 * do not show up in effective_cpus.
3605 cpumask_andnot(&new_cpus, &new_cpus, cs->subparts_cpus);
3607 if (!tmp || !cs->partition_root_state)
3611 * In the unlikely event that a partition root has empty
3612 * effective_cpus with tasks, we will have to invalidate child
3613 * partitions, if present, by setting nr_subparts_cpus to 0 to
3614 * reclaim their cpus.
3616 if (cs->nr_subparts_cpus && is_partition_valid(cs) &&
3617 cpumask_empty(&new_cpus) && partition_is_populated(cs, NULL)) {
3618 spin_lock_irq(&callback_lock);
3619 cs->nr_subparts_cpus = 0;
3620 cpumask_clear(cs->subparts_cpus);
3621 spin_unlock_irq(&callback_lock);
3622 compute_effective_cpumask(&new_cpus, cs, parent);
3626 * Force the partition to become invalid if either one of
3627 * the following conditions hold:
3628 * 1) empty effective cpus but not valid empty partition.
3629 * 2) parent is invalid or doesn't grant any cpus to child
3632 if (is_partition_valid(cs) && (!parent->nr_subparts_cpus ||
3633 (cpumask_empty(&new_cpus) && partition_is_populated(cs, NULL)))) {
3634 int old_prs, parent_prs;
3636 update_parent_subparts_cpumask(cs, partcmd_disable, NULL, tmp);
3637 if (cs->nr_subparts_cpus) {
3638 spin_lock_irq(&callback_lock);
3639 cs->nr_subparts_cpus = 0;
3640 cpumask_clear(cs->subparts_cpus);
3641 spin_unlock_irq(&callback_lock);
3642 compute_effective_cpumask(&new_cpus, cs, parent);
3645 old_prs = cs->partition_root_state;
3646 parent_prs = parent->partition_root_state;
3647 if (is_partition_valid(cs)) {
3648 spin_lock_irq(&callback_lock);
3649 make_partition_invalid(cs);
3650 spin_unlock_irq(&callback_lock);
3651 if (is_prs_invalid(parent_prs))
3652 WRITE_ONCE(cs->prs_err, PERR_INVPARENT);
3653 else if (!parent_prs)
3654 WRITE_ONCE(cs->prs_err, PERR_NOTPART);
3656 WRITE_ONCE(cs->prs_err, PERR_HOTPLUG);
3657 notify_partition_change(cs, old_prs);
3659 cpuset_force_rebuild();
3663 * On the other hand, an invalid partition root may be transitioned
3664 * back to a regular one.
3666 else if (is_partition_valid(parent) && is_partition_invalid(cs)) {
3667 update_parent_subparts_cpumask(cs, partcmd_update, NULL, tmp);
3668 if (is_partition_valid(cs))
3669 cpuset_force_rebuild();
3673 cpus_updated = !cpumask_equal(&new_cpus, cs->effective_cpus);
3674 mems_updated = !nodes_equal(new_mems, cs->effective_mems);
3675 if (!cpus_updated && !mems_updated)
3676 goto unlock; /* Hotplug doesn't affect this cpuset */
3679 check_insane_mems_config(&new_mems);
3681 if (is_in_v2_mode())
3682 hotplug_update_tasks(cs, &new_cpus, &new_mems,
3683 cpus_updated, mems_updated);
3685 hotplug_update_tasks_legacy(cs, &new_cpus, &new_mems,
3686 cpus_updated, mems_updated);
3689 mutex_unlock(&cpuset_mutex);
3693 * cpuset_hotplug_workfn - handle CPU/memory hotunplug for a cpuset
3695 * This function is called after either CPU or memory configuration has
3696 * changed and updates cpuset accordingly. The top_cpuset is always
3697 * synchronized to cpu_active_mask and N_MEMORY, which is necessary in
3698 * order to make cpusets transparent (of no affect) on systems that are
3699 * actively using CPU hotplug but making no active use of cpusets.
3701 * Non-root cpusets are only affected by offlining. If any CPUs or memory
3702 * nodes have been taken down, cpuset_hotplug_update_tasks() is invoked on
3705 * Note that CPU offlining during suspend is ignored. We don't modify
3706 * cpusets across suspend/resume cycles at all.
3708 static void cpuset_hotplug_workfn(struct work_struct *work)
3710 static cpumask_t new_cpus;
3711 static nodemask_t new_mems;
3712 bool cpus_updated, mems_updated;
3713 bool on_dfl = is_in_v2_mode();
3714 struct tmpmasks tmp, *ptmp = NULL;
3716 if (on_dfl && !alloc_cpumasks(NULL, &tmp))
3719 mutex_lock(&cpuset_mutex);
3721 /* fetch the available cpus/mems and find out which changed how */
3722 cpumask_copy(&new_cpus, cpu_active_mask);
3723 new_mems = node_states[N_MEMORY];
3726 * If subparts_cpus is populated, it is likely that the check below
3727 * will produce a false positive on cpus_updated when the cpu list
3728 * isn't changed. It is extra work, but it is better to be safe.
3730 cpus_updated = !cpumask_equal(top_cpuset.effective_cpus, &new_cpus);
3731 mems_updated = !nodes_equal(top_cpuset.effective_mems, new_mems);
3734 * In the rare case that hotplug removes all the cpus in subparts_cpus,
3735 * we assumed that cpus are updated.
3737 if (!cpus_updated && top_cpuset.nr_subparts_cpus)
3738 cpus_updated = true;
3740 /* synchronize cpus_allowed to cpu_active_mask */
3742 spin_lock_irq(&callback_lock);
3744 cpumask_copy(top_cpuset.cpus_allowed, &new_cpus);
3746 * Make sure that CPUs allocated to child partitions
3747 * do not show up in effective_cpus. If no CPU is left,
3748 * we clear the subparts_cpus & let the child partitions
3749 * fight for the CPUs again.
3751 if (top_cpuset.nr_subparts_cpus) {
3752 if (cpumask_subset(&new_cpus,
3753 top_cpuset.subparts_cpus)) {
3754 top_cpuset.nr_subparts_cpus = 0;
3755 cpumask_clear(top_cpuset.subparts_cpus);
3757 cpumask_andnot(&new_cpus, &new_cpus,
3758 top_cpuset.subparts_cpus);
3761 cpumask_copy(top_cpuset.effective_cpus, &new_cpus);
3762 spin_unlock_irq(&callback_lock);
3763 /* we don't mess with cpumasks of tasks in top_cpuset */
3766 /* synchronize mems_allowed to N_MEMORY */
3768 spin_lock_irq(&callback_lock);
3770 top_cpuset.mems_allowed = new_mems;
3771 top_cpuset.effective_mems = new_mems;
3772 spin_unlock_irq(&callback_lock);
3773 update_tasks_nodemask(&top_cpuset);
3776 mutex_unlock(&cpuset_mutex);
3778 /* if cpus or mems changed, we need to propagate to descendants */
3779 if (cpus_updated || mems_updated) {
3781 struct cgroup_subsys_state *pos_css;
3784 cpuset_for_each_descendant_pre(cs, pos_css, &top_cpuset) {
3785 if (cs == &top_cpuset || !css_tryget_online(&cs->css))
3789 cpuset_hotplug_update_tasks(cs, ptmp);
3797 /* rebuild sched domains if cpus_allowed has changed */
3798 if (cpus_updated || force_rebuild) {
3799 force_rebuild = false;
3800 rebuild_sched_domains();
3803 free_cpumasks(NULL, ptmp);
3806 void cpuset_update_active_cpus(void)
3809 * We're inside cpu hotplug critical region which usually nests
3810 * inside cgroup synchronization. Bounce actual hotplug processing
3811 * to a work item to avoid reverse locking order.
3813 schedule_work(&cpuset_hotplug_work);
3816 void cpuset_wait_for_hotplug(void)
3818 flush_work(&cpuset_hotplug_work);
3822 * Keep top_cpuset.mems_allowed tracking node_states[N_MEMORY].
3823 * Call this routine anytime after node_states[N_MEMORY] changes.
3824 * See cpuset_update_active_cpus() for CPU hotplug handling.
3826 static int cpuset_track_online_nodes(struct notifier_block *self,
3827 unsigned long action, void *arg)
3829 schedule_work(&cpuset_hotplug_work);
3834 * cpuset_init_smp - initialize cpus_allowed
3836 * Description: Finish top cpuset after cpu, node maps are initialized
3838 void __init cpuset_init_smp(void)
3841 * cpus_allowd/mems_allowed set to v2 values in the initial
3842 * cpuset_bind() call will be reset to v1 values in another
3843 * cpuset_bind() call when v1 cpuset is mounted.
3845 top_cpuset.old_mems_allowed = top_cpuset.mems_allowed;
3847 cpumask_copy(top_cpuset.effective_cpus, cpu_active_mask);
3848 top_cpuset.effective_mems = node_states[N_MEMORY];
3850 hotplug_memory_notifier(cpuset_track_online_nodes, CPUSET_CALLBACK_PRI);
3852 cpuset_migrate_mm_wq = alloc_ordered_workqueue("cpuset_migrate_mm", 0);
3853 BUG_ON(!cpuset_migrate_mm_wq);
3857 * cpuset_cpus_allowed - return cpus_allowed mask from a tasks cpuset.
3858 * @tsk: pointer to task_struct from which to obtain cpuset->cpus_allowed.
3859 * @pmask: pointer to struct cpumask variable to receive cpus_allowed set.
3861 * Description: Returns the cpumask_var_t cpus_allowed of the cpuset
3862 * attached to the specified @tsk. Guaranteed to return some non-empty
3863 * subset of cpu_online_mask, even if this means going outside the
3864 * tasks cpuset, except when the task is in the top cpuset.
3867 void cpuset_cpus_allowed(struct task_struct *tsk, struct cpumask *pmask)
3869 unsigned long flags;
3872 spin_lock_irqsave(&callback_lock, flags);
3876 if (cs != &top_cpuset)
3877 guarantee_online_cpus(tsk, pmask);
3879 * Tasks in the top cpuset won't get update to their cpumasks
3880 * when a hotplug online/offline event happens. So we include all
3881 * offline cpus in the allowed cpu list.
3883 if ((cs == &top_cpuset) || cpumask_empty(pmask)) {
3884 const struct cpumask *possible_mask = task_cpu_possible_mask(tsk);
3887 * We first exclude cpus allocated to partitions. If there is no
3888 * allowable online cpu left, we fall back to all possible cpus.
3890 cpumask_andnot(pmask, possible_mask, top_cpuset.subparts_cpus);
3891 if (!cpumask_intersects(pmask, cpu_online_mask))
3892 cpumask_copy(pmask, possible_mask);
3896 spin_unlock_irqrestore(&callback_lock, flags);
3900 * cpuset_cpus_allowed_fallback - final fallback before complete catastrophe.
3901 * @tsk: pointer to task_struct with which the scheduler is struggling
3903 * Description: In the case that the scheduler cannot find an allowed cpu in
3904 * tsk->cpus_allowed, we fall back to task_cs(tsk)->cpus_allowed. In legacy
3905 * mode however, this value is the same as task_cs(tsk)->effective_cpus,
3906 * which will not contain a sane cpumask during cases such as cpu hotplugging.
3907 * This is the absolute last resort for the scheduler and it is only used if
3908 * _every_ other avenue has been traveled.
3910 * Returns true if the affinity of @tsk was changed, false otherwise.
3913 bool cpuset_cpus_allowed_fallback(struct task_struct *tsk)
3915 const struct cpumask *possible_mask = task_cpu_possible_mask(tsk);
3916 const struct cpumask *cs_mask;
3917 bool changed = false;
3920 cs_mask = task_cs(tsk)->cpus_allowed;
3921 if (is_in_v2_mode() && cpumask_subset(cs_mask, possible_mask)) {
3922 do_set_cpus_allowed(tsk, cs_mask);
3928 * We own tsk->cpus_allowed, nobody can change it under us.
3930 * But we used cs && cs->cpus_allowed lockless and thus can
3931 * race with cgroup_attach_task() or update_cpumask() and get
3932 * the wrong tsk->cpus_allowed. However, both cases imply the
3933 * subsequent cpuset_change_cpumask()->set_cpus_allowed_ptr()
3934 * which takes task_rq_lock().
3936 * If we are called after it dropped the lock we must see all
3937 * changes in tsk_cs()->cpus_allowed. Otherwise we can temporary
3938 * set any mask even if it is not right from task_cs() pov,
3939 * the pending set_cpus_allowed_ptr() will fix things.
3941 * select_fallback_rq() will fix things ups and set cpu_possible_mask
3947 void __init cpuset_init_current_mems_allowed(void)
3949 nodes_setall(current->mems_allowed);
3953 * cpuset_mems_allowed - return mems_allowed mask from a tasks cpuset.
3954 * @tsk: pointer to task_struct from which to obtain cpuset->mems_allowed.
3956 * Description: Returns the nodemask_t mems_allowed of the cpuset
3957 * attached to the specified @tsk. Guaranteed to return some non-empty
3958 * subset of node_states[N_MEMORY], even if this means going outside the
3962 nodemask_t cpuset_mems_allowed(struct task_struct *tsk)
3965 unsigned long flags;
3967 spin_lock_irqsave(&callback_lock, flags);
3969 guarantee_online_mems(task_cs(tsk), &mask);
3971 spin_unlock_irqrestore(&callback_lock, flags);
3977 * cpuset_nodemask_valid_mems_allowed - check nodemask vs. current mems_allowed
3978 * @nodemask: the nodemask to be checked
3980 * Are any of the nodes in the nodemask allowed in current->mems_allowed?
3982 int cpuset_nodemask_valid_mems_allowed(nodemask_t *nodemask)
3984 return nodes_intersects(*nodemask, current->mems_allowed);
3988 * nearest_hardwall_ancestor() - Returns the nearest mem_exclusive or
3989 * mem_hardwall ancestor to the specified cpuset. Call holding
3990 * callback_lock. If no ancestor is mem_exclusive or mem_hardwall
3991 * (an unusual configuration), then returns the root cpuset.
3993 static struct cpuset *nearest_hardwall_ancestor(struct cpuset *cs)
3995 while (!(is_mem_exclusive(cs) || is_mem_hardwall(cs)) && parent_cs(cs))
4001 * cpuset_node_allowed - Can we allocate on a memory node?
4002 * @node: is this an allowed node?
4003 * @gfp_mask: memory allocation flags
4005 * If we're in interrupt, yes, we can always allocate. If @node is set in
4006 * current's mems_allowed, yes. If it's not a __GFP_HARDWALL request and this
4007 * node is set in the nearest hardwalled cpuset ancestor to current's cpuset,
4008 * yes. If current has access to memory reserves as an oom victim, yes.
4011 * GFP_USER allocations are marked with the __GFP_HARDWALL bit,
4012 * and do not allow allocations outside the current tasks cpuset
4013 * unless the task has been OOM killed.
4014 * GFP_KERNEL allocations are not so marked, so can escape to the
4015 * nearest enclosing hardwalled ancestor cpuset.
4017 * Scanning up parent cpusets requires callback_lock. The
4018 * __alloc_pages() routine only calls here with __GFP_HARDWALL bit
4019 * _not_ set if it's a GFP_KERNEL allocation, and all nodes in the
4020 * current tasks mems_allowed came up empty on the first pass over
4021 * the zonelist. So only GFP_KERNEL allocations, if all nodes in the
4022 * cpuset are short of memory, might require taking the callback_lock.
4024 * The first call here from mm/page_alloc:get_page_from_freelist()
4025 * has __GFP_HARDWALL set in gfp_mask, enforcing hardwall cpusets,
4026 * so no allocation on a node outside the cpuset is allowed (unless
4027 * in interrupt, of course).
4029 * The second pass through get_page_from_freelist() doesn't even call
4030 * here for GFP_ATOMIC calls. For those calls, the __alloc_pages()
4031 * variable 'wait' is not set, and the bit ALLOC_CPUSET is not set
4032 * in alloc_flags. That logic and the checks below have the combined
4034 * in_interrupt - any node ok (current task context irrelevant)
4035 * GFP_ATOMIC - any node ok
4036 * tsk_is_oom_victim - any node ok
4037 * GFP_KERNEL - any node in enclosing hardwalled cpuset ok
4038 * GFP_USER - only nodes in current tasks mems allowed ok.
4040 bool cpuset_node_allowed(int node, gfp_t gfp_mask)
4042 struct cpuset *cs; /* current cpuset ancestors */
4043 bool allowed; /* is allocation in zone z allowed? */
4044 unsigned long flags;
4048 if (node_isset(node, current->mems_allowed))
4051 * Allow tasks that have access to memory reserves because they have
4052 * been OOM killed to get memory anywhere.
4054 if (unlikely(tsk_is_oom_victim(current)))
4056 if (gfp_mask & __GFP_HARDWALL) /* If hardwall request, stop here */
4059 if (current->flags & PF_EXITING) /* Let dying task have memory */
4062 /* Not hardwall and node outside mems_allowed: scan up cpusets */
4063 spin_lock_irqsave(&callback_lock, flags);
4066 cs = nearest_hardwall_ancestor(task_cs(current));
4067 allowed = node_isset(node, cs->mems_allowed);
4070 spin_unlock_irqrestore(&callback_lock, flags);
4075 * cpuset_spread_node() - On which node to begin search for a page
4077 * If a task is marked PF_SPREAD_PAGE or PF_SPREAD_SLAB (as for
4078 * tasks in a cpuset with is_spread_page or is_spread_slab set),
4079 * and if the memory allocation used cpuset_mem_spread_node()
4080 * to determine on which node to start looking, as it will for
4081 * certain page cache or slab cache pages such as used for file
4082 * system buffers and inode caches, then instead of starting on the
4083 * local node to look for a free page, rather spread the starting
4084 * node around the tasks mems_allowed nodes.
4086 * We don't have to worry about the returned node being offline
4087 * because "it can't happen", and even if it did, it would be ok.
4089 * The routines calling guarantee_online_mems() are careful to
4090 * only set nodes in task->mems_allowed that are online. So it
4091 * should not be possible for the following code to return an
4092 * offline node. But if it did, that would be ok, as this routine
4093 * is not returning the node where the allocation must be, only
4094 * the node where the search should start. The zonelist passed to
4095 * __alloc_pages() will include all nodes. If the slab allocator
4096 * is passed an offline node, it will fall back to the local node.
4097 * See kmem_cache_alloc_node().
4099 static int cpuset_spread_node(int *rotor)
4101 return *rotor = next_node_in(*rotor, current->mems_allowed);
4105 * cpuset_mem_spread_node() - On which node to begin search for a file page
4107 int cpuset_mem_spread_node(void)
4109 if (current->cpuset_mem_spread_rotor == NUMA_NO_NODE)
4110 current->cpuset_mem_spread_rotor =
4111 node_random(¤t->mems_allowed);
4113 return cpuset_spread_node(¤t->cpuset_mem_spread_rotor);
4117 * cpuset_slab_spread_node() - On which node to begin search for a slab page
4119 int cpuset_slab_spread_node(void)
4121 if (current->cpuset_slab_spread_rotor == NUMA_NO_NODE)
4122 current->cpuset_slab_spread_rotor =
4123 node_random(¤t->mems_allowed);
4125 return cpuset_spread_node(¤t->cpuset_slab_spread_rotor);
4127 EXPORT_SYMBOL_GPL(cpuset_mem_spread_node);
4130 * cpuset_mems_allowed_intersects - Does @tsk1's mems_allowed intersect @tsk2's?
4131 * @tsk1: pointer to task_struct of some task.
4132 * @tsk2: pointer to task_struct of some other task.
4134 * Description: Return true if @tsk1's mems_allowed intersects the
4135 * mems_allowed of @tsk2. Used by the OOM killer to determine if
4136 * one of the task's memory usage might impact the memory available
4140 int cpuset_mems_allowed_intersects(const struct task_struct *tsk1,
4141 const struct task_struct *tsk2)
4143 return nodes_intersects(tsk1->mems_allowed, tsk2->mems_allowed);
4147 * cpuset_print_current_mems_allowed - prints current's cpuset and mems_allowed
4149 * Description: Prints current's name, cpuset name, and cached copy of its
4150 * mems_allowed to the kernel log.
4152 void cpuset_print_current_mems_allowed(void)
4154 struct cgroup *cgrp;
4158 cgrp = task_cs(current)->css.cgroup;
4159 pr_cont(",cpuset=");
4160 pr_cont_cgroup_name(cgrp);
4161 pr_cont(",mems_allowed=%*pbl",
4162 nodemask_pr_args(¤t->mems_allowed));
4168 * Collection of memory_pressure is suppressed unless
4169 * this flag is enabled by writing "1" to the special
4170 * cpuset file 'memory_pressure_enabled' in the root cpuset.
4173 int cpuset_memory_pressure_enabled __read_mostly;
4176 * __cpuset_memory_pressure_bump - keep stats of per-cpuset reclaims.
4178 * Keep a running average of the rate of synchronous (direct)
4179 * page reclaim efforts initiated by tasks in each cpuset.
4181 * This represents the rate at which some task in the cpuset
4182 * ran low on memory on all nodes it was allowed to use, and
4183 * had to enter the kernels page reclaim code in an effort to
4184 * create more free memory by tossing clean pages or swapping
4185 * or writing dirty pages.
4187 * Display to user space in the per-cpuset read-only file
4188 * "memory_pressure". Value displayed is an integer
4189 * representing the recent rate of entry into the synchronous
4190 * (direct) page reclaim by any task attached to the cpuset.
4193 void __cpuset_memory_pressure_bump(void)
4196 fmeter_markevent(&task_cs(current)->fmeter);
4200 #ifdef CONFIG_PROC_PID_CPUSET
4202 * proc_cpuset_show()
4203 * - Print tasks cpuset path into seq_file.
4204 * - Used for /proc/<pid>/cpuset.
4205 * - No need to task_lock(tsk) on this tsk->cpuset reference, as it
4206 * doesn't really matter if tsk->cpuset changes after we read it,
4207 * and we take cpuset_mutex, keeping cpuset_attach() from changing it
4210 int proc_cpuset_show(struct seq_file *m, struct pid_namespace *ns,
4211 struct pid *pid, struct task_struct *tsk)
4214 struct cgroup_subsys_state *css;
4218 buf = kmalloc(PATH_MAX, GFP_KERNEL);
4222 css = task_get_css(tsk, cpuset_cgrp_id);
4223 retval = cgroup_path_ns(css->cgroup, buf, PATH_MAX,
4224 current->nsproxy->cgroup_ns);
4226 if (retval >= PATH_MAX)
4227 retval = -ENAMETOOLONG;
4238 #endif /* CONFIG_PROC_PID_CPUSET */
4240 /* Display task mems_allowed in /proc/<pid>/status file. */
4241 void cpuset_task_status_allowed(struct seq_file *m, struct task_struct *task)
4243 seq_printf(m, "Mems_allowed:\t%*pb\n",
4244 nodemask_pr_args(&task->mems_allowed));
4245 seq_printf(m, "Mems_allowed_list:\t%*pbl\n",
4246 nodemask_pr_args(&task->mems_allowed));