4 * Processor and Memory placement constraints for sets of tasks.
6 * Copyright (C) 2003 BULL SA.
7 * Copyright (C) 2004-2007 Silicon Graphics, Inc.
8 * Copyright (C) 2006 Google, Inc
10 * Portions derived from Patrick Mochel's sysfs code.
11 * sysfs is Copyright (c) 2001-3 Patrick Mochel
13 * 2003-10-10 Written by Simon Derr.
14 * 2003-10-22 Updates by Stephen Hemminger.
15 * 2004 May-July Rework by Paul Jackson.
16 * 2006 Rework by Paul Menage to use generic cgroups
17 * 2008 Rework of the scheduler domains and CPU hotplug handling
20 * This file is subject to the terms and conditions of the GNU General Public
21 * License. See the file COPYING in the main directory of the Linux
22 * distribution for more details.
25 #include <linux/cpu.h>
26 #include <linux/cpumask.h>
27 #include <linux/cpuset.h>
28 #include <linux/err.h>
29 #include <linux/errno.h>
30 #include <linux/file.h>
32 #include <linux/init.h>
33 #include <linux/interrupt.h>
34 #include <linux/kernel.h>
35 #include <linux/kmod.h>
36 #include <linux/list.h>
37 #include <linux/mempolicy.h>
39 #include <linux/memory.h>
40 #include <linux/export.h>
41 #include <linux/mount.h>
42 #include <linux/namei.h>
43 #include <linux/pagemap.h>
44 #include <linux/proc_fs.h>
45 #include <linux/rcupdate.h>
46 #include <linux/sched.h>
47 #include <linux/seq_file.h>
48 #include <linux/security.h>
49 #include <linux/slab.h>
50 #include <linux/spinlock.h>
51 #include <linux/stat.h>
52 #include <linux/string.h>
53 #include <linux/time.h>
54 #include <linux/backing-dev.h>
55 #include <linux/sort.h>
57 #include <asm/uaccess.h>
58 #include <linux/atomic.h>
59 #include <linux/mutex.h>
60 #include <linux/workqueue.h>
61 #include <linux/cgroup.h>
62 #include <linux/wait.h>
64 struct static_key cpusets_enabled_key __read_mostly = STATIC_KEY_INIT_FALSE;
66 /* See "Frequency meter" comments, below. */
69 int cnt; /* unprocessed events count */
70 int val; /* most recent output value */
71 time_t time; /* clock (secs) when val computed */
72 spinlock_t lock; /* guards read or write of above */
76 struct cgroup_subsys_state css;
78 unsigned long flags; /* "unsigned long" so bitops work */
81 * On default hierarchy:
83 * The user-configured masks can only be changed by writing to
84 * cpuset.cpus and cpuset.mems, and won't be limited by the
87 * The effective masks is the real masks that apply to the tasks
88 * in the cpuset. They may be changed if the configured masks are
89 * changed or hotplug happens.
91 * effective_mask == configured_mask & parent's effective_mask,
92 * and if it ends up empty, it will inherit the parent's mask.
97 * The user-configured masks are always the same with effective masks.
100 /* user-configured CPUs and Memory Nodes allow to tasks */
101 cpumask_var_t cpus_allowed;
102 nodemask_t mems_allowed;
104 /* effective CPUs and Memory Nodes allow to tasks */
105 cpumask_var_t effective_cpus;
106 nodemask_t effective_mems;
109 * This is old Memory Nodes tasks took on.
111 * - top_cpuset.old_mems_allowed is initialized to mems_allowed.
112 * - A new cpuset's old_mems_allowed is initialized when some
113 * task is moved into it.
114 * - old_mems_allowed is used in cpuset_migrate_mm() when we change
115 * cpuset.mems_allowed and have tasks' nodemask updated, and
116 * then old_mems_allowed is updated to mems_allowed.
118 nodemask_t old_mems_allowed;
120 struct fmeter fmeter; /* memory_pressure filter */
123 * Tasks are being attached to this cpuset. Used to prevent
124 * zeroing cpus/mems_allowed between ->can_attach() and ->attach().
126 int attach_in_progress;
128 /* partition number for rebuild_sched_domains() */
131 /* for custom sched domain */
132 int relax_domain_level;
135 static inline struct cpuset *css_cs(struct cgroup_subsys_state *css)
137 return css ? container_of(css, struct cpuset, css) : NULL;
140 /* Retrieve the cpuset for a task */
141 static inline struct cpuset *task_cs(struct task_struct *task)
143 return css_cs(task_css(task, cpuset_cgrp_id));
146 static inline struct cpuset *parent_cs(struct cpuset *cs)
148 return css_cs(cs->css.parent);
152 static inline bool task_has_mempolicy(struct task_struct *task)
154 return task->mempolicy;
157 static inline bool task_has_mempolicy(struct task_struct *task)
164 /* bits in struct cpuset flags field */
171 CS_SCHED_LOAD_BALANCE,
176 /* convenient tests for these bits */
177 static inline bool is_cpuset_online(const struct cpuset *cs)
179 return test_bit(CS_ONLINE, &cs->flags);
182 static inline int is_cpu_exclusive(const struct cpuset *cs)
184 return test_bit(CS_CPU_EXCLUSIVE, &cs->flags);
187 static inline int is_mem_exclusive(const struct cpuset *cs)
189 return test_bit(CS_MEM_EXCLUSIVE, &cs->flags);
192 static inline int is_mem_hardwall(const struct cpuset *cs)
194 return test_bit(CS_MEM_HARDWALL, &cs->flags);
197 static inline int is_sched_load_balance(const struct cpuset *cs)
199 return test_bit(CS_SCHED_LOAD_BALANCE, &cs->flags);
202 static inline int is_memory_migrate(const struct cpuset *cs)
204 return test_bit(CS_MEMORY_MIGRATE, &cs->flags);
207 static inline int is_spread_page(const struct cpuset *cs)
209 return test_bit(CS_SPREAD_PAGE, &cs->flags);
212 static inline int is_spread_slab(const struct cpuset *cs)
214 return test_bit(CS_SPREAD_SLAB, &cs->flags);
217 static struct cpuset top_cpuset = {
218 .flags = ((1 << CS_ONLINE) | (1 << CS_CPU_EXCLUSIVE) |
219 (1 << CS_MEM_EXCLUSIVE)),
223 * cpuset_for_each_child - traverse online children of a cpuset
224 * @child_cs: loop cursor pointing to the current child
225 * @pos_css: used for iteration
226 * @parent_cs: target cpuset to walk children of
228 * Walk @child_cs through the online children of @parent_cs. Must be used
229 * with RCU read locked.
231 #define cpuset_for_each_child(child_cs, pos_css, parent_cs) \
232 css_for_each_child((pos_css), &(parent_cs)->css) \
233 if (is_cpuset_online(((child_cs) = css_cs((pos_css)))))
236 * cpuset_for_each_descendant_pre - pre-order walk of a cpuset's descendants
237 * @des_cs: loop cursor pointing to the current descendant
238 * @pos_css: used for iteration
239 * @root_cs: target cpuset to walk ancestor of
241 * Walk @des_cs through the online descendants of @root_cs. Must be used
242 * with RCU read locked. The caller may modify @pos_css by calling
243 * css_rightmost_descendant() to skip subtree. @root_cs is included in the
244 * iteration and the first node to be visited.
246 #define cpuset_for_each_descendant_pre(des_cs, pos_css, root_cs) \
247 css_for_each_descendant_pre((pos_css), &(root_cs)->css) \
248 if (is_cpuset_online(((des_cs) = css_cs((pos_css)))))
251 * There are two global locks guarding cpuset structures - cpuset_mutex and
252 * callback_lock. We also require taking task_lock() when dereferencing a
253 * task's cpuset pointer. See "The task_lock() exception", at the end of this
256 * A task must hold both locks to modify cpusets. If a task holds
257 * cpuset_mutex, then it blocks others wanting that mutex, ensuring that it
258 * is the only task able to also acquire callback_lock and be able to
259 * modify cpusets. It can perform various checks on the cpuset structure
260 * first, knowing nothing will change. It can also allocate memory while
261 * just holding cpuset_mutex. While it is performing these checks, various
262 * callback routines can briefly acquire callback_lock to query cpusets.
263 * Once it is ready to make the changes, it takes callback_lock, blocking
266 * Calls to the kernel memory allocator can not be made while holding
267 * callback_lock, as that would risk double tripping on callback_lock
268 * from one of the callbacks into the cpuset code from within
271 * If a task is only holding callback_lock, then it has read-only
274 * Now, the task_struct fields mems_allowed and mempolicy may be changed
275 * by other task, we use alloc_lock in the task_struct fields to protect
278 * The cpuset_common_file_read() handlers only hold callback_lock across
279 * small pieces of code, such as when reading out possibly multi-word
280 * cpumasks and nodemasks.
282 * Accessing a task's cpuset should be done in accordance with the
283 * guidelines for accessing subsystem state in kernel/cgroup.c
286 static DEFINE_MUTEX(cpuset_mutex);
287 static DEFINE_SPINLOCK(callback_lock);
290 * CPU / memory hotplug is handled asynchronously.
292 static void cpuset_hotplug_workfn(struct work_struct *work);
293 static DECLARE_WORK(cpuset_hotplug_work, cpuset_hotplug_workfn);
295 static DECLARE_WAIT_QUEUE_HEAD(cpuset_attach_wq);
298 * This is ugly, but preserves the userspace API for existing cpuset
299 * users. If someone tries to mount the "cpuset" filesystem, we
300 * silently switch it to mount "cgroup" instead
302 static struct dentry *cpuset_mount(struct file_system_type *fs_type,
303 int flags, const char *unused_dev_name, void *data)
305 struct file_system_type *cgroup_fs = get_fs_type("cgroup");
306 struct dentry *ret = ERR_PTR(-ENODEV);
310 "release_agent=/sbin/cpuset_release_agent";
311 ret = cgroup_fs->mount(cgroup_fs, flags,
312 unused_dev_name, mountopts);
313 put_filesystem(cgroup_fs);
318 static struct file_system_type cpuset_fs_type = {
320 .mount = cpuset_mount,
324 * Return in pmask the portion of a cpusets's cpus_allowed that
325 * are online. If none are online, walk up the cpuset hierarchy
326 * until we find one that does have some online cpus. The top
327 * cpuset always has some cpus online.
329 * One way or another, we guarantee to return some non-empty subset
330 * of cpu_online_mask.
332 * Call with callback_lock or cpuset_mutex held.
334 static void guarantee_online_cpus(struct cpuset *cs, struct cpumask *pmask)
336 while (!cpumask_intersects(cs->effective_cpus, cpu_online_mask))
338 cpumask_and(pmask, cs->effective_cpus, cpu_online_mask);
342 * Return in *pmask the portion of a cpusets's mems_allowed that
343 * are online, with memory. If none are online with memory, walk
344 * up the cpuset hierarchy until we find one that does have some
345 * online mems. The top cpuset always has some mems online.
347 * One way or another, we guarantee to return some non-empty subset
348 * of node_states[N_MEMORY].
350 * Call with callback_lock or cpuset_mutex held.
352 static void guarantee_online_mems(struct cpuset *cs, nodemask_t *pmask)
354 while (!nodes_intersects(cs->effective_mems, node_states[N_MEMORY]))
356 nodes_and(*pmask, cs->effective_mems, node_states[N_MEMORY]);
360 * update task's spread flag if cpuset's page/slab spread flag is set
362 * Call with callback_lock or cpuset_mutex held.
364 static void cpuset_update_task_spread_flag(struct cpuset *cs,
365 struct task_struct *tsk)
367 if (is_spread_page(cs))
368 task_set_spread_page(tsk);
370 task_clear_spread_page(tsk);
372 if (is_spread_slab(cs))
373 task_set_spread_slab(tsk);
375 task_clear_spread_slab(tsk);
379 * is_cpuset_subset(p, q) - Is cpuset p a subset of cpuset q?
381 * One cpuset is a subset of another if all its allowed CPUs and
382 * Memory Nodes are a subset of the other, and its exclusive flags
383 * are only set if the other's are set. Call holding cpuset_mutex.
386 static int is_cpuset_subset(const struct cpuset *p, const struct cpuset *q)
388 return cpumask_subset(p->cpus_allowed, q->cpus_allowed) &&
389 nodes_subset(p->mems_allowed, q->mems_allowed) &&
390 is_cpu_exclusive(p) <= is_cpu_exclusive(q) &&
391 is_mem_exclusive(p) <= is_mem_exclusive(q);
395 * alloc_trial_cpuset - allocate a trial cpuset
396 * @cs: the cpuset that the trial cpuset duplicates
398 static struct cpuset *alloc_trial_cpuset(struct cpuset *cs)
400 struct cpuset *trial;
402 trial = kmemdup(cs, sizeof(*cs), GFP_KERNEL);
406 if (!alloc_cpumask_var(&trial->cpus_allowed, GFP_KERNEL))
408 if (!alloc_cpumask_var(&trial->effective_cpus, GFP_KERNEL))
411 cpumask_copy(trial->cpus_allowed, cs->cpus_allowed);
412 cpumask_copy(trial->effective_cpus, cs->effective_cpus);
416 free_cpumask_var(trial->cpus_allowed);
423 * free_trial_cpuset - free the trial cpuset
424 * @trial: the trial cpuset to be freed
426 static void free_trial_cpuset(struct cpuset *trial)
428 free_cpumask_var(trial->effective_cpus);
429 free_cpumask_var(trial->cpus_allowed);
434 * validate_change() - Used to validate that any proposed cpuset change
435 * follows the structural rules for cpusets.
437 * If we replaced the flag and mask values of the current cpuset
438 * (cur) with those values in the trial cpuset (trial), would
439 * our various subset and exclusive rules still be valid? Presumes
442 * 'cur' is the address of an actual, in-use cpuset. Operations
443 * such as list traversal that depend on the actual address of the
444 * cpuset in the list must use cur below, not trial.
446 * 'trial' is the address of bulk structure copy of cur, with
447 * perhaps one or more of the fields cpus_allowed, mems_allowed,
448 * or flags changed to new, trial values.
450 * Return 0 if valid, -errno if not.
453 static int validate_change(struct cpuset *cur, struct cpuset *trial)
455 struct cgroup_subsys_state *css;
456 struct cpuset *c, *par;
461 /* Each of our child cpusets must be a subset of us */
463 cpuset_for_each_child(c, css, cur)
464 if (!is_cpuset_subset(c, trial))
467 /* Remaining checks don't apply to root cpuset */
469 if (cur == &top_cpuset)
472 par = parent_cs(cur);
474 /* On legacy hiearchy, we must be a subset of our parent cpuset. */
476 if (!cgroup_on_dfl(cur->css.cgroup) && !is_cpuset_subset(trial, par))
480 * If either I or some sibling (!= me) is exclusive, we can't
484 cpuset_for_each_child(c, css, par) {
485 if ((is_cpu_exclusive(trial) || is_cpu_exclusive(c)) &&
487 cpumask_intersects(trial->cpus_allowed, c->cpus_allowed))
489 if ((is_mem_exclusive(trial) || is_mem_exclusive(c)) &&
491 nodes_intersects(trial->mems_allowed, c->mems_allowed))
496 * Cpusets with tasks - existing or newly being attached - can't
497 * be changed to have empty cpus_allowed or mems_allowed.
500 if ((cgroup_has_tasks(cur->css.cgroup) || cur->attach_in_progress)) {
501 if (!cpumask_empty(cur->cpus_allowed) &&
502 cpumask_empty(trial->cpus_allowed))
504 if (!nodes_empty(cur->mems_allowed) &&
505 nodes_empty(trial->mems_allowed))
510 * We can't shrink if we won't have enough room for SCHED_DEADLINE
514 if (is_cpu_exclusive(cur) &&
515 !cpuset_cpumask_can_shrink(cur->cpus_allowed,
516 trial->cpus_allowed))
527 * Helper routine for generate_sched_domains().
528 * Do cpusets a, b have overlapping effective cpus_allowed masks?
530 static int cpusets_overlap(struct cpuset *a, struct cpuset *b)
532 return cpumask_intersects(a->effective_cpus, b->effective_cpus);
536 update_domain_attr(struct sched_domain_attr *dattr, struct cpuset *c)
538 if (dattr->relax_domain_level < c->relax_domain_level)
539 dattr->relax_domain_level = c->relax_domain_level;
543 static void update_domain_attr_tree(struct sched_domain_attr *dattr,
544 struct cpuset *root_cs)
547 struct cgroup_subsys_state *pos_css;
550 cpuset_for_each_descendant_pre(cp, pos_css, root_cs) {
551 /* skip the whole subtree if @cp doesn't have any CPU */
552 if (cpumask_empty(cp->cpus_allowed)) {
553 pos_css = css_rightmost_descendant(pos_css);
557 if (is_sched_load_balance(cp))
558 update_domain_attr(dattr, cp);
564 * generate_sched_domains()
566 * This function builds a partial partition of the systems CPUs
567 * A 'partial partition' is a set of non-overlapping subsets whose
568 * union is a subset of that set.
569 * The output of this function needs to be passed to kernel/sched/core.c
570 * partition_sched_domains() routine, which will rebuild the scheduler's
571 * load balancing domains (sched domains) as specified by that partial
574 * See "What is sched_load_balance" in Documentation/cgroups/cpusets.txt
575 * for a background explanation of this.
577 * Does not return errors, on the theory that the callers of this
578 * routine would rather not worry about failures to rebuild sched
579 * domains when operating in the severe memory shortage situations
580 * that could cause allocation failures below.
582 * Must be called with cpuset_mutex held.
584 * The three key local variables below are:
585 * q - a linked-list queue of cpuset pointers, used to implement a
586 * top-down scan of all cpusets. This scan loads a pointer
587 * to each cpuset marked is_sched_load_balance into the
588 * array 'csa'. For our purposes, rebuilding the schedulers
589 * sched domains, we can ignore !is_sched_load_balance cpusets.
590 * csa - (for CpuSet Array) Array of pointers to all the cpusets
591 * that need to be load balanced, for convenient iterative
592 * access by the subsequent code that finds the best partition,
593 * i.e the set of domains (subsets) of CPUs such that the
594 * cpus_allowed of every cpuset marked is_sched_load_balance
595 * is a subset of one of these domains, while there are as
596 * many such domains as possible, each as small as possible.
597 * doms - Conversion of 'csa' to an array of cpumasks, for passing to
598 * the kernel/sched/core.c routine partition_sched_domains() in a
599 * convenient format, that can be easily compared to the prior
600 * value to determine what partition elements (sched domains)
601 * were changed (added or removed.)
603 * Finding the best partition (set of domains):
604 * The triple nested loops below over i, j, k scan over the
605 * load balanced cpusets (using the array of cpuset pointers in
606 * csa[]) looking for pairs of cpusets that have overlapping
607 * cpus_allowed, but which don't have the same 'pn' partition
608 * number and gives them in the same partition number. It keeps
609 * looping on the 'restart' label until it can no longer find
612 * The union of the cpus_allowed masks from the set of
613 * all cpusets having the same 'pn' value then form the one
614 * element of the partition (one sched domain) to be passed to
615 * partition_sched_domains().
617 static int generate_sched_domains(cpumask_var_t **domains,
618 struct sched_domain_attr **attributes)
620 struct cpuset *cp; /* scans q */
621 struct cpuset **csa; /* array of all cpuset ptrs */
622 int csn; /* how many cpuset ptrs in csa so far */
623 int i, j, k; /* indices for partition finding loops */
624 cpumask_var_t *doms; /* resulting partition; i.e. sched domains */
625 cpumask_var_t non_isolated_cpus; /* load balanced CPUs */
626 struct sched_domain_attr *dattr; /* attributes for custom domains */
627 int ndoms = 0; /* number of sched domains in result */
628 int nslot; /* next empty doms[] struct cpumask slot */
629 struct cgroup_subsys_state *pos_css;
635 if (!alloc_cpumask_var(&non_isolated_cpus, GFP_KERNEL))
637 cpumask_andnot(non_isolated_cpus, cpu_possible_mask, cpu_isolated_map);
639 /* Special case for the 99% of systems with one, full, sched domain */
640 if (is_sched_load_balance(&top_cpuset)) {
642 doms = alloc_sched_domains(ndoms);
646 dattr = kmalloc(sizeof(struct sched_domain_attr), GFP_KERNEL);
648 *dattr = SD_ATTR_INIT;
649 update_domain_attr_tree(dattr, &top_cpuset);
651 cpumask_and(doms[0], top_cpuset.effective_cpus,
657 csa = kmalloc(nr_cpusets() * sizeof(cp), GFP_KERNEL);
663 cpuset_for_each_descendant_pre(cp, pos_css, &top_cpuset) {
664 if (cp == &top_cpuset)
667 * Continue traversing beyond @cp iff @cp has some CPUs and
668 * isn't load balancing. The former is obvious. The
669 * latter: All child cpusets contain a subset of the
670 * parent's cpus, so just skip them, and then we call
671 * update_domain_attr_tree() to calc relax_domain_level of
672 * the corresponding sched domain.
674 if (!cpumask_empty(cp->cpus_allowed) &&
675 !(is_sched_load_balance(cp) &&
676 cpumask_intersects(cp->cpus_allowed, non_isolated_cpus)))
679 if (is_sched_load_balance(cp))
682 /* skip @cp's subtree */
683 pos_css = css_rightmost_descendant(pos_css);
687 for (i = 0; i < csn; i++)
692 /* Find the best partition (set of sched domains) */
693 for (i = 0; i < csn; i++) {
694 struct cpuset *a = csa[i];
697 for (j = 0; j < csn; j++) {
698 struct cpuset *b = csa[j];
701 if (apn != bpn && cpusets_overlap(a, b)) {
702 for (k = 0; k < csn; k++) {
703 struct cpuset *c = csa[k];
708 ndoms--; /* one less element */
715 * Now we know how many domains to create.
716 * Convert <csn, csa> to <ndoms, doms> and populate cpu masks.
718 doms = alloc_sched_domains(ndoms);
723 * The rest of the code, including the scheduler, can deal with
724 * dattr==NULL case. No need to abort if alloc fails.
726 dattr = kmalloc(ndoms * sizeof(struct sched_domain_attr), GFP_KERNEL);
728 for (nslot = 0, i = 0; i < csn; i++) {
729 struct cpuset *a = csa[i];
734 /* Skip completed partitions */
740 if (nslot == ndoms) {
741 static int warnings = 10;
743 pr_warn("rebuild_sched_domains confused: nslot %d, ndoms %d, csn %d, i %d, apn %d\n",
744 nslot, ndoms, csn, i, apn);
752 *(dattr + nslot) = SD_ATTR_INIT;
753 for (j = i; j < csn; j++) {
754 struct cpuset *b = csa[j];
757 cpumask_or(dp, dp, b->effective_cpus);
758 cpumask_and(dp, dp, non_isolated_cpus);
760 update_domain_attr_tree(dattr + nslot, b);
762 /* Done with this partition */
768 BUG_ON(nslot != ndoms);
771 free_cpumask_var(non_isolated_cpus);
775 * Fallback to the default domain if kmalloc() failed.
776 * See comments in partition_sched_domains().
787 * Rebuild scheduler domains.
789 * If the flag 'sched_load_balance' of any cpuset with non-empty
790 * 'cpus' changes, or if the 'cpus' allowed changes in any cpuset
791 * which has that flag enabled, or if any cpuset with a non-empty
792 * 'cpus' is removed, then call this routine to rebuild the
793 * scheduler's dynamic sched domains.
795 * Call with cpuset_mutex held. Takes get_online_cpus().
797 static void rebuild_sched_domains_locked(void)
799 struct sched_domain_attr *attr;
803 lockdep_assert_held(&cpuset_mutex);
807 * We have raced with CPU hotplug. Don't do anything to avoid
808 * passing doms with offlined cpu to partition_sched_domains().
809 * Anyways, hotplug work item will rebuild sched domains.
811 if (!cpumask_equal(top_cpuset.effective_cpus, cpu_active_mask))
814 /* Generate domain masks and attrs */
815 ndoms = generate_sched_domains(&doms, &attr);
817 /* Have scheduler rebuild the domains */
818 partition_sched_domains(ndoms, doms, attr);
822 #else /* !CONFIG_SMP */
823 static void rebuild_sched_domains_locked(void)
826 #endif /* CONFIG_SMP */
828 void rebuild_sched_domains(void)
830 mutex_lock(&cpuset_mutex);
831 rebuild_sched_domains_locked();
832 mutex_unlock(&cpuset_mutex);
836 * update_tasks_cpumask - Update the cpumasks of tasks in the cpuset.
837 * @cs: the cpuset in which each task's cpus_allowed mask needs to be changed
839 * Iterate through each task of @cs updating its cpus_allowed to the
840 * effective cpuset's. As this function is called with cpuset_mutex held,
841 * cpuset membership stays stable.
843 static void update_tasks_cpumask(struct cpuset *cs)
845 struct css_task_iter it;
846 struct task_struct *task;
848 css_task_iter_start(&cs->css, &it);
849 while ((task = css_task_iter_next(&it)))
850 set_cpus_allowed_ptr(task, cs->effective_cpus);
851 css_task_iter_end(&it);
855 * update_cpumasks_hier - Update effective cpumasks and tasks in the subtree
856 * @cs: the cpuset to consider
857 * @new_cpus: temp variable for calculating new effective_cpus
859 * When congifured cpumask is changed, the effective cpumasks of this cpuset
860 * and all its descendants need to be updated.
862 * On legacy hierachy, effective_cpus will be the same with cpu_allowed.
864 * Called with cpuset_mutex held
866 static void update_cpumasks_hier(struct cpuset *cs, struct cpumask *new_cpus)
869 struct cgroup_subsys_state *pos_css;
870 bool need_rebuild_sched_domains = false;
873 cpuset_for_each_descendant_pre(cp, pos_css, cs) {
874 struct cpuset *parent = parent_cs(cp);
876 cpumask_and(new_cpus, cp->cpus_allowed, parent->effective_cpus);
879 * If it becomes empty, inherit the effective mask of the
880 * parent, which is guaranteed to have some CPUs.
882 if (cgroup_on_dfl(cp->css.cgroup) && cpumask_empty(new_cpus))
883 cpumask_copy(new_cpus, parent->effective_cpus);
885 /* Skip the whole subtree if the cpumask remains the same. */
886 if (cpumask_equal(new_cpus, cp->effective_cpus)) {
887 pos_css = css_rightmost_descendant(pos_css);
891 if (!css_tryget_online(&cp->css))
895 spin_lock_irq(&callback_lock);
896 cpumask_copy(cp->effective_cpus, new_cpus);
897 spin_unlock_irq(&callback_lock);
899 WARN_ON(!cgroup_on_dfl(cp->css.cgroup) &&
900 !cpumask_equal(cp->cpus_allowed, cp->effective_cpus));
902 update_tasks_cpumask(cp);
905 * If the effective cpumask of any non-empty cpuset is changed,
906 * we need to rebuild sched domains.
908 if (!cpumask_empty(cp->cpus_allowed) &&
909 is_sched_load_balance(cp))
910 need_rebuild_sched_domains = true;
917 if (need_rebuild_sched_domains)
918 rebuild_sched_domains_locked();
922 * update_cpumask - update the cpus_allowed mask of a cpuset and all tasks in it
923 * @cs: the cpuset to consider
924 * @trialcs: trial cpuset
925 * @buf: buffer of cpu numbers written to this cpuset
927 static int update_cpumask(struct cpuset *cs, struct cpuset *trialcs,
932 /* top_cpuset.cpus_allowed tracks cpu_online_mask; it's read-only */
933 if (cs == &top_cpuset)
937 * An empty cpus_allowed is ok only if the cpuset has no tasks.
938 * Since cpulist_parse() fails on an empty mask, we special case
939 * that parsing. The validate_change() call ensures that cpusets
940 * with tasks have cpus.
943 cpumask_clear(trialcs->cpus_allowed);
945 retval = cpulist_parse(buf, trialcs->cpus_allowed);
949 if (!cpumask_subset(trialcs->cpus_allowed,
950 top_cpuset.cpus_allowed))
954 /* Nothing to do if the cpus didn't change */
955 if (cpumask_equal(cs->cpus_allowed, trialcs->cpus_allowed))
958 retval = validate_change(cs, trialcs);
962 spin_lock_irq(&callback_lock);
963 cpumask_copy(cs->cpus_allowed, trialcs->cpus_allowed);
964 spin_unlock_irq(&callback_lock);
966 /* use trialcs->cpus_allowed as a temp variable */
967 update_cpumasks_hier(cs, trialcs->cpus_allowed);
974 * Migrate memory region from one set of nodes to another.
976 * Temporarilly set tasks mems_allowed to target nodes of migration,
977 * so that the migration code can allocate pages on these nodes.
979 * While the mm_struct we are migrating is typically from some
980 * other task, the task_struct mems_allowed that we are hacking
981 * is for our current task, which must allocate new pages for that
982 * migrating memory region.
985 static void cpuset_migrate_mm(struct mm_struct *mm, const nodemask_t *from,
986 const nodemask_t *to)
988 struct task_struct *tsk = current;
990 tsk->mems_allowed = *to;
992 do_migrate_pages(mm, from, to, MPOL_MF_MOVE_ALL);
995 guarantee_online_mems(task_cs(tsk), &tsk->mems_allowed);
1000 * cpuset_change_task_nodemask - change task's mems_allowed and mempolicy
1001 * @tsk: the task to change
1002 * @newmems: new nodes that the task will be set
1004 * In order to avoid seeing no nodes if the old and new nodes are disjoint,
1005 * we structure updates as setting all new allowed nodes, then clearing newly
1008 static void cpuset_change_task_nodemask(struct task_struct *tsk,
1009 nodemask_t *newmems)
1014 * Allow tasks that have access to memory reserves because they have
1015 * been OOM killed to get memory anywhere.
1017 if (unlikely(test_thread_flag(TIF_MEMDIE)))
1019 if (current->flags & PF_EXITING) /* Let dying task have memory */
1024 * Determine if a loop is necessary if another thread is doing
1025 * read_mems_allowed_begin(). If at least one node remains unchanged and
1026 * tsk does not have a mempolicy, then an empty nodemask will not be
1027 * possible when mems_allowed is larger than a word.
1029 need_loop = task_has_mempolicy(tsk) ||
1030 !nodes_intersects(*newmems, tsk->mems_allowed);
1033 local_irq_disable();
1034 write_seqcount_begin(&tsk->mems_allowed_seq);
1037 nodes_or(tsk->mems_allowed, tsk->mems_allowed, *newmems);
1038 mpol_rebind_task(tsk, newmems, MPOL_REBIND_STEP1);
1040 mpol_rebind_task(tsk, newmems, MPOL_REBIND_STEP2);
1041 tsk->mems_allowed = *newmems;
1044 write_seqcount_end(&tsk->mems_allowed_seq);
1051 static void *cpuset_being_rebound;
1054 * update_tasks_nodemask - Update the nodemasks of tasks in the cpuset.
1055 * @cs: the cpuset in which each task's mems_allowed mask needs to be changed
1057 * Iterate through each task of @cs updating its mems_allowed to the
1058 * effective cpuset's. As this function is called with cpuset_mutex held,
1059 * cpuset membership stays stable.
1061 static void update_tasks_nodemask(struct cpuset *cs)
1063 static nodemask_t newmems; /* protected by cpuset_mutex */
1064 struct css_task_iter it;
1065 struct task_struct *task;
1067 cpuset_being_rebound = cs; /* causes mpol_dup() rebind */
1069 guarantee_online_mems(cs, &newmems);
1072 * The mpol_rebind_mm() call takes mmap_sem, which we couldn't
1073 * take while holding tasklist_lock. Forks can happen - the
1074 * mpol_dup() cpuset_being_rebound check will catch such forks,
1075 * and rebind their vma mempolicies too. Because we still hold
1076 * the global cpuset_mutex, we know that no other rebind effort
1077 * will be contending for the global variable cpuset_being_rebound.
1078 * It's ok if we rebind the same mm twice; mpol_rebind_mm()
1079 * is idempotent. Also migrate pages in each mm to new nodes.
1081 css_task_iter_start(&cs->css, &it);
1082 while ((task = css_task_iter_next(&it))) {
1083 struct mm_struct *mm;
1086 cpuset_change_task_nodemask(task, &newmems);
1088 mm = get_task_mm(task);
1092 migrate = is_memory_migrate(cs);
1094 mpol_rebind_mm(mm, &cs->mems_allowed);
1096 cpuset_migrate_mm(mm, &cs->old_mems_allowed, &newmems);
1099 css_task_iter_end(&it);
1102 * All the tasks' nodemasks have been updated, update
1103 * cs->old_mems_allowed.
1105 cs->old_mems_allowed = newmems;
1107 /* We're done rebinding vmas to this cpuset's new mems_allowed. */
1108 cpuset_being_rebound = NULL;
1112 * update_nodemasks_hier - Update effective nodemasks and tasks in the subtree
1113 * @cs: the cpuset to consider
1114 * @new_mems: a temp variable for calculating new effective_mems
1116 * When configured nodemask is changed, the effective nodemasks of this cpuset
1117 * and all its descendants need to be updated.
1119 * On legacy hiearchy, effective_mems will be the same with mems_allowed.
1121 * Called with cpuset_mutex held
1123 static void update_nodemasks_hier(struct cpuset *cs, nodemask_t *new_mems)
1126 struct cgroup_subsys_state *pos_css;
1129 cpuset_for_each_descendant_pre(cp, pos_css, cs) {
1130 struct cpuset *parent = parent_cs(cp);
1132 nodes_and(*new_mems, cp->mems_allowed, parent->effective_mems);
1135 * If it becomes empty, inherit the effective mask of the
1136 * parent, which is guaranteed to have some MEMs.
1138 if (cgroup_on_dfl(cp->css.cgroup) && nodes_empty(*new_mems))
1139 *new_mems = parent->effective_mems;
1141 /* Skip the whole subtree if the nodemask remains the same. */
1142 if (nodes_equal(*new_mems, cp->effective_mems)) {
1143 pos_css = css_rightmost_descendant(pos_css);
1147 if (!css_tryget_online(&cp->css))
1151 spin_lock_irq(&callback_lock);
1152 cp->effective_mems = *new_mems;
1153 spin_unlock_irq(&callback_lock);
1155 WARN_ON(!cgroup_on_dfl(cp->css.cgroup) &&
1156 !nodes_equal(cp->mems_allowed, cp->effective_mems));
1158 update_tasks_nodemask(cp);
1167 * Handle user request to change the 'mems' memory placement
1168 * of a cpuset. Needs to validate the request, update the
1169 * cpusets mems_allowed, and for each task in the cpuset,
1170 * update mems_allowed and rebind task's mempolicy and any vma
1171 * mempolicies and if the cpuset is marked 'memory_migrate',
1172 * migrate the tasks pages to the new memory.
1174 * Call with cpuset_mutex held. May take callback_lock during call.
1175 * Will take tasklist_lock, scan tasklist for tasks in cpuset cs,
1176 * lock each such tasks mm->mmap_sem, scan its vma's and rebind
1177 * their mempolicies to the cpusets new mems_allowed.
1179 static int update_nodemask(struct cpuset *cs, struct cpuset *trialcs,
1185 * top_cpuset.mems_allowed tracks node_stats[N_MEMORY];
1188 if (cs == &top_cpuset) {
1194 * An empty mems_allowed is ok iff there are no tasks in the cpuset.
1195 * Since nodelist_parse() fails on an empty mask, we special case
1196 * that parsing. The validate_change() call ensures that cpusets
1197 * with tasks have memory.
1200 nodes_clear(trialcs->mems_allowed);
1202 retval = nodelist_parse(buf, trialcs->mems_allowed);
1206 if (!nodes_subset(trialcs->mems_allowed,
1207 top_cpuset.mems_allowed)) {
1213 if (nodes_equal(cs->mems_allowed, trialcs->mems_allowed)) {
1214 retval = 0; /* Too easy - nothing to do */
1217 retval = validate_change(cs, trialcs);
1221 spin_lock_irq(&callback_lock);
1222 cs->mems_allowed = trialcs->mems_allowed;
1223 spin_unlock_irq(&callback_lock);
1225 /* use trialcs->mems_allowed as a temp variable */
1226 update_nodemasks_hier(cs, &cs->mems_allowed);
1231 int current_cpuset_is_being_rebound(void)
1236 ret = task_cs(current) == cpuset_being_rebound;
1242 static int update_relax_domain_level(struct cpuset *cs, s64 val)
1245 if (val < -1 || val >= sched_domain_level_max)
1249 if (val != cs->relax_domain_level) {
1250 cs->relax_domain_level = val;
1251 if (!cpumask_empty(cs->cpus_allowed) &&
1252 is_sched_load_balance(cs))
1253 rebuild_sched_domains_locked();
1260 * update_tasks_flags - update the spread flags of tasks in the cpuset.
1261 * @cs: the cpuset in which each task's spread flags needs to be changed
1263 * Iterate through each task of @cs updating its spread flags. As this
1264 * function is called with cpuset_mutex held, cpuset membership stays
1267 static void update_tasks_flags(struct cpuset *cs)
1269 struct css_task_iter it;
1270 struct task_struct *task;
1272 css_task_iter_start(&cs->css, &it);
1273 while ((task = css_task_iter_next(&it)))
1274 cpuset_update_task_spread_flag(cs, task);
1275 css_task_iter_end(&it);
1279 * update_flag - read a 0 or a 1 in a file and update associated flag
1280 * bit: the bit to update (see cpuset_flagbits_t)
1281 * cs: the cpuset to update
1282 * turning_on: whether the flag is being set or cleared
1284 * Call with cpuset_mutex held.
1287 static int update_flag(cpuset_flagbits_t bit, struct cpuset *cs,
1290 struct cpuset *trialcs;
1291 int balance_flag_changed;
1292 int spread_flag_changed;
1295 trialcs = alloc_trial_cpuset(cs);
1300 set_bit(bit, &trialcs->flags);
1302 clear_bit(bit, &trialcs->flags);
1304 err = validate_change(cs, trialcs);
1308 balance_flag_changed = (is_sched_load_balance(cs) !=
1309 is_sched_load_balance(trialcs));
1311 spread_flag_changed = ((is_spread_slab(cs) != is_spread_slab(trialcs))
1312 || (is_spread_page(cs) != is_spread_page(trialcs)));
1314 spin_lock_irq(&callback_lock);
1315 cs->flags = trialcs->flags;
1316 spin_unlock_irq(&callback_lock);
1318 if (!cpumask_empty(trialcs->cpus_allowed) && balance_flag_changed)
1319 rebuild_sched_domains_locked();
1321 if (spread_flag_changed)
1322 update_tasks_flags(cs);
1324 free_trial_cpuset(trialcs);
1329 * Frequency meter - How fast is some event occurring?
1331 * These routines manage a digitally filtered, constant time based,
1332 * event frequency meter. There are four routines:
1333 * fmeter_init() - initialize a frequency meter.
1334 * fmeter_markevent() - called each time the event happens.
1335 * fmeter_getrate() - returns the recent rate of such events.
1336 * fmeter_update() - internal routine used to update fmeter.
1338 * A common data structure is passed to each of these routines,
1339 * which is used to keep track of the state required to manage the
1340 * frequency meter and its digital filter.
1342 * The filter works on the number of events marked per unit time.
1343 * The filter is single-pole low-pass recursive (IIR). The time unit
1344 * is 1 second. Arithmetic is done using 32-bit integers scaled to
1345 * simulate 3 decimal digits of precision (multiplied by 1000).
1347 * With an FM_COEF of 933, and a time base of 1 second, the filter
1348 * has a half-life of 10 seconds, meaning that if the events quit
1349 * happening, then the rate returned from the fmeter_getrate()
1350 * will be cut in half each 10 seconds, until it converges to zero.
1352 * It is not worth doing a real infinitely recursive filter. If more
1353 * than FM_MAXTICKS ticks have elapsed since the last filter event,
1354 * just compute FM_MAXTICKS ticks worth, by which point the level
1357 * Limit the count of unprocessed events to FM_MAXCNT, so as to avoid
1358 * arithmetic overflow in the fmeter_update() routine.
1360 * Given the simple 32 bit integer arithmetic used, this meter works
1361 * best for reporting rates between one per millisecond (msec) and
1362 * one per 32 (approx) seconds. At constant rates faster than one
1363 * per msec it maxes out at values just under 1,000,000. At constant
1364 * rates between one per msec, and one per second it will stabilize
1365 * to a value N*1000, where N is the rate of events per second.
1366 * At constant rates between one per second and one per 32 seconds,
1367 * it will be choppy, moving up on the seconds that have an event,
1368 * and then decaying until the next event. At rates slower than
1369 * about one in 32 seconds, it decays all the way back to zero between
1373 #define FM_COEF 933 /* coefficient for half-life of 10 secs */
1374 #define FM_MAXTICKS ((time_t)99) /* useless computing more ticks than this */
1375 #define FM_MAXCNT 1000000 /* limit cnt to avoid overflow */
1376 #define FM_SCALE 1000 /* faux fixed point scale */
1378 /* Initialize a frequency meter */
1379 static void fmeter_init(struct fmeter *fmp)
1384 spin_lock_init(&fmp->lock);
1387 /* Internal meter update - process cnt events and update value */
1388 static void fmeter_update(struct fmeter *fmp)
1390 time_t now = get_seconds();
1391 time_t ticks = now - fmp->time;
1396 ticks = min(FM_MAXTICKS, ticks);
1398 fmp->val = (FM_COEF * fmp->val) / FM_SCALE;
1401 fmp->val += ((FM_SCALE - FM_COEF) * fmp->cnt) / FM_SCALE;
1405 /* Process any previous ticks, then bump cnt by one (times scale). */
1406 static void fmeter_markevent(struct fmeter *fmp)
1408 spin_lock(&fmp->lock);
1410 fmp->cnt = min(FM_MAXCNT, fmp->cnt + FM_SCALE);
1411 spin_unlock(&fmp->lock);
1414 /* Process any previous ticks, then return current value. */
1415 static int fmeter_getrate(struct fmeter *fmp)
1419 spin_lock(&fmp->lock);
1422 spin_unlock(&fmp->lock);
1426 static struct cpuset *cpuset_attach_old_cs;
1428 /* Called by cgroups to determine if a cpuset is usable; cpuset_mutex held */
1429 static int cpuset_can_attach(struct cgroup_subsys_state *css,
1430 struct cgroup_taskset *tset)
1432 struct cpuset *cs = css_cs(css);
1433 struct task_struct *task;
1436 /* used later by cpuset_attach() */
1437 cpuset_attach_old_cs = task_cs(cgroup_taskset_first(tset));
1439 mutex_lock(&cpuset_mutex);
1441 /* allow moving tasks into an empty cpuset if on default hierarchy */
1443 if (!cgroup_on_dfl(css->cgroup) &&
1444 (cpumask_empty(cs->cpus_allowed) || nodes_empty(cs->mems_allowed)))
1447 cgroup_taskset_for_each(task, tset) {
1448 ret = task_can_attach(task, cs->cpus_allowed);
1451 ret = security_task_setscheduler(task);
1457 * Mark attach is in progress. This makes validate_change() fail
1458 * changes which zero cpus/mems_allowed.
1460 cs->attach_in_progress++;
1463 mutex_unlock(&cpuset_mutex);
1467 static void cpuset_cancel_attach(struct cgroup_subsys_state *css,
1468 struct cgroup_taskset *tset)
1470 mutex_lock(&cpuset_mutex);
1471 css_cs(css)->attach_in_progress--;
1472 mutex_unlock(&cpuset_mutex);
1476 * Protected by cpuset_mutex. cpus_attach is used only by cpuset_attach()
1477 * but we can't allocate it dynamically there. Define it global and
1478 * allocate from cpuset_init().
1480 static cpumask_var_t cpus_attach;
1482 static void cpuset_attach(struct cgroup_subsys_state *css,
1483 struct cgroup_taskset *tset)
1485 /* static buf protected by cpuset_mutex */
1486 static nodemask_t cpuset_attach_nodemask_to;
1487 struct mm_struct *mm;
1488 struct task_struct *task;
1489 struct task_struct *leader = cgroup_taskset_first(tset);
1490 struct cpuset *cs = css_cs(css);
1491 struct cpuset *oldcs = cpuset_attach_old_cs;
1493 mutex_lock(&cpuset_mutex);
1495 /* prepare for attach */
1496 if (cs == &top_cpuset)
1497 cpumask_copy(cpus_attach, cpu_possible_mask);
1499 guarantee_online_cpus(cs, cpus_attach);
1501 guarantee_online_mems(cs, &cpuset_attach_nodemask_to);
1503 cgroup_taskset_for_each(task, tset) {
1505 * can_attach beforehand should guarantee that this doesn't
1506 * fail. TODO: have a better way to handle failure here
1508 WARN_ON_ONCE(set_cpus_allowed_ptr(task, cpus_attach));
1510 cpuset_change_task_nodemask(task, &cpuset_attach_nodemask_to);
1511 cpuset_update_task_spread_flag(cs, task);
1515 * Change mm, possibly for multiple threads in a threadgroup. This is
1516 * expensive and may sleep.
1518 cpuset_attach_nodemask_to = cs->effective_mems;
1519 mm = get_task_mm(leader);
1521 mpol_rebind_mm(mm, &cpuset_attach_nodemask_to);
1524 * old_mems_allowed is the same with mems_allowed here, except
1525 * if this task is being moved automatically due to hotplug.
1526 * In that case @mems_allowed has been updated and is empty,
1527 * so @old_mems_allowed is the right nodesets that we migrate
1530 if (is_memory_migrate(cs)) {
1531 cpuset_migrate_mm(mm, &oldcs->old_mems_allowed,
1532 &cpuset_attach_nodemask_to);
1537 cs->old_mems_allowed = cpuset_attach_nodemask_to;
1539 cs->attach_in_progress--;
1540 if (!cs->attach_in_progress)
1541 wake_up(&cpuset_attach_wq);
1543 mutex_unlock(&cpuset_mutex);
1546 /* The various types of files and directories in a cpuset file system */
1549 FILE_MEMORY_MIGRATE,
1552 FILE_EFFECTIVE_CPULIST,
1553 FILE_EFFECTIVE_MEMLIST,
1557 FILE_SCHED_LOAD_BALANCE,
1558 FILE_SCHED_RELAX_DOMAIN_LEVEL,
1559 FILE_MEMORY_PRESSURE_ENABLED,
1560 FILE_MEMORY_PRESSURE,
1563 } cpuset_filetype_t;
1565 static int cpuset_write_u64(struct cgroup_subsys_state *css, struct cftype *cft,
1568 struct cpuset *cs = css_cs(css);
1569 cpuset_filetype_t type = cft->private;
1572 mutex_lock(&cpuset_mutex);
1573 if (!is_cpuset_online(cs)) {
1579 case FILE_CPU_EXCLUSIVE:
1580 retval = update_flag(CS_CPU_EXCLUSIVE, cs, val);
1582 case FILE_MEM_EXCLUSIVE:
1583 retval = update_flag(CS_MEM_EXCLUSIVE, cs, val);
1585 case FILE_MEM_HARDWALL:
1586 retval = update_flag(CS_MEM_HARDWALL, cs, val);
1588 case FILE_SCHED_LOAD_BALANCE:
1589 retval = update_flag(CS_SCHED_LOAD_BALANCE, cs, val);
1591 case FILE_MEMORY_MIGRATE:
1592 retval = update_flag(CS_MEMORY_MIGRATE, cs, val);
1594 case FILE_MEMORY_PRESSURE_ENABLED:
1595 cpuset_memory_pressure_enabled = !!val;
1597 case FILE_MEMORY_PRESSURE:
1600 case FILE_SPREAD_PAGE:
1601 retval = update_flag(CS_SPREAD_PAGE, cs, val);
1603 case FILE_SPREAD_SLAB:
1604 retval = update_flag(CS_SPREAD_SLAB, cs, val);
1611 mutex_unlock(&cpuset_mutex);
1615 static int cpuset_write_s64(struct cgroup_subsys_state *css, struct cftype *cft,
1618 struct cpuset *cs = css_cs(css);
1619 cpuset_filetype_t type = cft->private;
1620 int retval = -ENODEV;
1622 mutex_lock(&cpuset_mutex);
1623 if (!is_cpuset_online(cs))
1627 case FILE_SCHED_RELAX_DOMAIN_LEVEL:
1628 retval = update_relax_domain_level(cs, val);
1635 mutex_unlock(&cpuset_mutex);
1640 * Common handling for a write to a "cpus" or "mems" file.
1642 static ssize_t cpuset_write_resmask(struct kernfs_open_file *of,
1643 char *buf, size_t nbytes, loff_t off)
1645 struct cpuset *cs = css_cs(of_css(of));
1646 struct cpuset *trialcs;
1647 int retval = -ENODEV;
1649 buf = strstrip(buf);
1652 * CPU or memory hotunplug may leave @cs w/o any execution
1653 * resources, in which case the hotplug code asynchronously updates
1654 * configuration and transfers all tasks to the nearest ancestor
1655 * which can execute.
1657 * As writes to "cpus" or "mems" may restore @cs's execution
1658 * resources, wait for the previously scheduled operations before
1659 * proceeding, so that we don't end up keep removing tasks added
1660 * after execution capability is restored.
1662 * cpuset_hotplug_work calls back into cgroup core via
1663 * cgroup_transfer_tasks() and waiting for it from a cgroupfs
1664 * operation like this one can lead to a deadlock through kernfs
1665 * active_ref protection. Let's break the protection. Losing the
1666 * protection is okay as we check whether @cs is online after
1667 * grabbing cpuset_mutex anyway. This only happens on the legacy
1671 kernfs_break_active_protection(of->kn);
1672 flush_work(&cpuset_hotplug_work);
1674 mutex_lock(&cpuset_mutex);
1675 if (!is_cpuset_online(cs))
1678 trialcs = alloc_trial_cpuset(cs);
1684 switch (of_cft(of)->private) {
1686 retval = update_cpumask(cs, trialcs, buf);
1689 retval = update_nodemask(cs, trialcs, buf);
1696 free_trial_cpuset(trialcs);
1698 mutex_unlock(&cpuset_mutex);
1699 kernfs_unbreak_active_protection(of->kn);
1701 return retval ?: nbytes;
1705 * These ascii lists should be read in a single call, by using a user
1706 * buffer large enough to hold the entire map. If read in smaller
1707 * chunks, there is no guarantee of atomicity. Since the display format
1708 * used, list of ranges of sequential numbers, is variable length,
1709 * and since these maps can change value dynamically, one could read
1710 * gibberish by doing partial reads while a list was changing.
1712 static int cpuset_common_seq_show(struct seq_file *sf, void *v)
1714 struct cpuset *cs = css_cs(seq_css(sf));
1715 cpuset_filetype_t type = seq_cft(sf)->private;
1718 spin_lock_irq(&callback_lock);
1722 seq_printf(sf, "%*pbl\n", cpumask_pr_args(cs->cpus_allowed));
1725 seq_printf(sf, "%*pbl\n", nodemask_pr_args(&cs->mems_allowed));
1727 case FILE_EFFECTIVE_CPULIST:
1728 seq_printf(sf, "%*pbl\n", cpumask_pr_args(cs->effective_cpus));
1730 case FILE_EFFECTIVE_MEMLIST:
1731 seq_printf(sf, "%*pbl\n", nodemask_pr_args(&cs->effective_mems));
1737 spin_unlock_irq(&callback_lock);
1741 static u64 cpuset_read_u64(struct cgroup_subsys_state *css, struct cftype *cft)
1743 struct cpuset *cs = css_cs(css);
1744 cpuset_filetype_t type = cft->private;
1746 case FILE_CPU_EXCLUSIVE:
1747 return is_cpu_exclusive(cs);
1748 case FILE_MEM_EXCLUSIVE:
1749 return is_mem_exclusive(cs);
1750 case FILE_MEM_HARDWALL:
1751 return is_mem_hardwall(cs);
1752 case FILE_SCHED_LOAD_BALANCE:
1753 return is_sched_load_balance(cs);
1754 case FILE_MEMORY_MIGRATE:
1755 return is_memory_migrate(cs);
1756 case FILE_MEMORY_PRESSURE_ENABLED:
1757 return cpuset_memory_pressure_enabled;
1758 case FILE_MEMORY_PRESSURE:
1759 return fmeter_getrate(&cs->fmeter);
1760 case FILE_SPREAD_PAGE:
1761 return is_spread_page(cs);
1762 case FILE_SPREAD_SLAB:
1763 return is_spread_slab(cs);
1768 /* Unreachable but makes gcc happy */
1772 static s64 cpuset_read_s64(struct cgroup_subsys_state *css, struct cftype *cft)
1774 struct cpuset *cs = css_cs(css);
1775 cpuset_filetype_t type = cft->private;
1777 case FILE_SCHED_RELAX_DOMAIN_LEVEL:
1778 return cs->relax_domain_level;
1783 /* Unrechable but makes gcc happy */
1789 * for the common functions, 'private' gives the type of file
1792 static struct cftype files[] = {
1795 .seq_show = cpuset_common_seq_show,
1796 .write = cpuset_write_resmask,
1797 .max_write_len = (100U + 6 * NR_CPUS),
1798 .private = FILE_CPULIST,
1803 .seq_show = cpuset_common_seq_show,
1804 .write = cpuset_write_resmask,
1805 .max_write_len = (100U + 6 * MAX_NUMNODES),
1806 .private = FILE_MEMLIST,
1810 .name = "effective_cpus",
1811 .seq_show = cpuset_common_seq_show,
1812 .private = FILE_EFFECTIVE_CPULIST,
1816 .name = "effective_mems",
1817 .seq_show = cpuset_common_seq_show,
1818 .private = FILE_EFFECTIVE_MEMLIST,
1822 .name = "cpu_exclusive",
1823 .read_u64 = cpuset_read_u64,
1824 .write_u64 = cpuset_write_u64,
1825 .private = FILE_CPU_EXCLUSIVE,
1829 .name = "mem_exclusive",
1830 .read_u64 = cpuset_read_u64,
1831 .write_u64 = cpuset_write_u64,
1832 .private = FILE_MEM_EXCLUSIVE,
1836 .name = "mem_hardwall",
1837 .read_u64 = cpuset_read_u64,
1838 .write_u64 = cpuset_write_u64,
1839 .private = FILE_MEM_HARDWALL,
1843 .name = "sched_load_balance",
1844 .read_u64 = cpuset_read_u64,
1845 .write_u64 = cpuset_write_u64,
1846 .private = FILE_SCHED_LOAD_BALANCE,
1850 .name = "sched_relax_domain_level",
1851 .read_s64 = cpuset_read_s64,
1852 .write_s64 = cpuset_write_s64,
1853 .private = FILE_SCHED_RELAX_DOMAIN_LEVEL,
1857 .name = "memory_migrate",
1858 .read_u64 = cpuset_read_u64,
1859 .write_u64 = cpuset_write_u64,
1860 .private = FILE_MEMORY_MIGRATE,
1864 .name = "memory_pressure",
1865 .read_u64 = cpuset_read_u64,
1866 .write_u64 = cpuset_write_u64,
1867 .private = FILE_MEMORY_PRESSURE,
1872 .name = "memory_spread_page",
1873 .read_u64 = cpuset_read_u64,
1874 .write_u64 = cpuset_write_u64,
1875 .private = FILE_SPREAD_PAGE,
1879 .name = "memory_spread_slab",
1880 .read_u64 = cpuset_read_u64,
1881 .write_u64 = cpuset_write_u64,
1882 .private = FILE_SPREAD_SLAB,
1886 .name = "memory_pressure_enabled",
1887 .flags = CFTYPE_ONLY_ON_ROOT,
1888 .read_u64 = cpuset_read_u64,
1889 .write_u64 = cpuset_write_u64,
1890 .private = FILE_MEMORY_PRESSURE_ENABLED,
1897 * cpuset_css_alloc - allocate a cpuset css
1898 * cgrp: control group that the new cpuset will be part of
1901 static struct cgroup_subsys_state *
1902 cpuset_css_alloc(struct cgroup_subsys_state *parent_css)
1907 return &top_cpuset.css;
1909 cs = kzalloc(sizeof(*cs), GFP_KERNEL);
1911 return ERR_PTR(-ENOMEM);
1912 if (!alloc_cpumask_var(&cs->cpus_allowed, GFP_KERNEL))
1914 if (!alloc_cpumask_var(&cs->effective_cpus, GFP_KERNEL))
1917 set_bit(CS_SCHED_LOAD_BALANCE, &cs->flags);
1918 cpumask_clear(cs->cpus_allowed);
1919 nodes_clear(cs->mems_allowed);
1920 cpumask_clear(cs->effective_cpus);
1921 nodes_clear(cs->effective_mems);
1922 fmeter_init(&cs->fmeter);
1923 cs->relax_domain_level = -1;
1928 free_cpumask_var(cs->cpus_allowed);
1931 return ERR_PTR(-ENOMEM);
1934 static int cpuset_css_online(struct cgroup_subsys_state *css)
1936 struct cpuset *cs = css_cs(css);
1937 struct cpuset *parent = parent_cs(cs);
1938 struct cpuset *tmp_cs;
1939 struct cgroup_subsys_state *pos_css;
1944 mutex_lock(&cpuset_mutex);
1946 set_bit(CS_ONLINE, &cs->flags);
1947 if (is_spread_page(parent))
1948 set_bit(CS_SPREAD_PAGE, &cs->flags);
1949 if (is_spread_slab(parent))
1950 set_bit(CS_SPREAD_SLAB, &cs->flags);
1954 spin_lock_irq(&callback_lock);
1955 if (cgroup_on_dfl(cs->css.cgroup)) {
1956 cpumask_copy(cs->effective_cpus, parent->effective_cpus);
1957 cs->effective_mems = parent->effective_mems;
1959 spin_unlock_irq(&callback_lock);
1961 if (!test_bit(CGRP_CPUSET_CLONE_CHILDREN, &css->cgroup->flags))
1965 * Clone @parent's configuration if CGRP_CPUSET_CLONE_CHILDREN is
1966 * set. This flag handling is implemented in cgroup core for
1967 * histrical reasons - the flag may be specified during mount.
1969 * Currently, if any sibling cpusets have exclusive cpus or mem, we
1970 * refuse to clone the configuration - thereby refusing the task to
1971 * be entered, and as a result refusing the sys_unshare() or
1972 * clone() which initiated it. If this becomes a problem for some
1973 * users who wish to allow that scenario, then this could be
1974 * changed to grant parent->cpus_allowed-sibling_cpus_exclusive
1975 * (and likewise for mems) to the new cgroup.
1978 cpuset_for_each_child(tmp_cs, pos_css, parent) {
1979 if (is_mem_exclusive(tmp_cs) || is_cpu_exclusive(tmp_cs)) {
1986 spin_lock_irq(&callback_lock);
1987 cs->mems_allowed = parent->mems_allowed;
1988 cs->effective_mems = parent->mems_allowed;
1989 cpumask_copy(cs->cpus_allowed, parent->cpus_allowed);
1990 cpumask_copy(cs->effective_cpus, parent->cpus_allowed);
1991 spin_unlock_irq(&callback_lock);
1993 mutex_unlock(&cpuset_mutex);
1998 * If the cpuset being removed has its flag 'sched_load_balance'
1999 * enabled, then simulate turning sched_load_balance off, which
2000 * will call rebuild_sched_domains_locked().
2003 static void cpuset_css_offline(struct cgroup_subsys_state *css)
2005 struct cpuset *cs = css_cs(css);
2007 mutex_lock(&cpuset_mutex);
2009 if (is_sched_load_balance(cs))
2010 update_flag(CS_SCHED_LOAD_BALANCE, cs, 0);
2013 clear_bit(CS_ONLINE, &cs->flags);
2015 mutex_unlock(&cpuset_mutex);
2018 static void cpuset_css_free(struct cgroup_subsys_state *css)
2020 struct cpuset *cs = css_cs(css);
2022 free_cpumask_var(cs->effective_cpus);
2023 free_cpumask_var(cs->cpus_allowed);
2027 static void cpuset_bind(struct cgroup_subsys_state *root_css)
2029 mutex_lock(&cpuset_mutex);
2030 spin_lock_irq(&callback_lock);
2032 if (cgroup_on_dfl(root_css->cgroup)) {
2033 cpumask_copy(top_cpuset.cpus_allowed, cpu_possible_mask);
2034 top_cpuset.mems_allowed = node_possible_map;
2036 cpumask_copy(top_cpuset.cpus_allowed,
2037 top_cpuset.effective_cpus);
2038 top_cpuset.mems_allowed = top_cpuset.effective_mems;
2041 spin_unlock_irq(&callback_lock);
2042 mutex_unlock(&cpuset_mutex);
2045 struct cgroup_subsys cpuset_cgrp_subsys = {
2046 .css_alloc = cpuset_css_alloc,
2047 .css_online = cpuset_css_online,
2048 .css_offline = cpuset_css_offline,
2049 .css_free = cpuset_css_free,
2050 .can_attach = cpuset_can_attach,
2051 .cancel_attach = cpuset_cancel_attach,
2052 .attach = cpuset_attach,
2053 .bind = cpuset_bind,
2054 .legacy_cftypes = files,
2059 * cpuset_init - initialize cpusets at system boot
2061 * Description: Initialize top_cpuset and the cpuset internal file system,
2064 int __init cpuset_init(void)
2068 if (!alloc_cpumask_var(&top_cpuset.cpus_allowed, GFP_KERNEL))
2070 if (!alloc_cpumask_var(&top_cpuset.effective_cpus, GFP_KERNEL))
2073 cpumask_setall(top_cpuset.cpus_allowed);
2074 nodes_setall(top_cpuset.mems_allowed);
2075 cpumask_setall(top_cpuset.effective_cpus);
2076 nodes_setall(top_cpuset.effective_mems);
2078 fmeter_init(&top_cpuset.fmeter);
2079 set_bit(CS_SCHED_LOAD_BALANCE, &top_cpuset.flags);
2080 top_cpuset.relax_domain_level = -1;
2082 err = register_filesystem(&cpuset_fs_type);
2086 if (!alloc_cpumask_var(&cpus_attach, GFP_KERNEL))
2093 * If CPU and/or memory hotplug handlers, below, unplug any CPUs
2094 * or memory nodes, we need to walk over the cpuset hierarchy,
2095 * removing that CPU or node from all cpusets. If this removes the
2096 * last CPU or node from a cpuset, then move the tasks in the empty
2097 * cpuset to its next-highest non-empty parent.
2099 static void remove_tasks_in_empty_cpuset(struct cpuset *cs)
2101 struct cpuset *parent;
2104 * Find its next-highest non-empty parent, (top cpuset
2105 * has online cpus, so can't be empty).
2107 parent = parent_cs(cs);
2108 while (cpumask_empty(parent->cpus_allowed) ||
2109 nodes_empty(parent->mems_allowed))
2110 parent = parent_cs(parent);
2112 if (cgroup_transfer_tasks(parent->css.cgroup, cs->css.cgroup)) {
2113 pr_err("cpuset: failed to transfer tasks out of empty cpuset ");
2114 pr_cont_cgroup_name(cs->css.cgroup);
2120 hotplug_update_tasks_legacy(struct cpuset *cs,
2121 struct cpumask *new_cpus, nodemask_t *new_mems,
2122 bool cpus_updated, bool mems_updated)
2126 spin_lock_irq(&callback_lock);
2127 cpumask_copy(cs->cpus_allowed, new_cpus);
2128 cpumask_copy(cs->effective_cpus, new_cpus);
2129 cs->mems_allowed = *new_mems;
2130 cs->effective_mems = *new_mems;
2131 spin_unlock_irq(&callback_lock);
2134 * Don't call update_tasks_cpumask() if the cpuset becomes empty,
2135 * as the tasks will be migratecd to an ancestor.
2137 if (cpus_updated && !cpumask_empty(cs->cpus_allowed))
2138 update_tasks_cpumask(cs);
2139 if (mems_updated && !nodes_empty(cs->mems_allowed))
2140 update_tasks_nodemask(cs);
2142 is_empty = cpumask_empty(cs->cpus_allowed) ||
2143 nodes_empty(cs->mems_allowed);
2145 mutex_unlock(&cpuset_mutex);
2148 * Move tasks to the nearest ancestor with execution resources,
2149 * This is full cgroup operation which will also call back into
2150 * cpuset. Should be done outside any lock.
2153 remove_tasks_in_empty_cpuset(cs);
2155 mutex_lock(&cpuset_mutex);
2159 hotplug_update_tasks(struct cpuset *cs,
2160 struct cpumask *new_cpus, nodemask_t *new_mems,
2161 bool cpus_updated, bool mems_updated)
2163 if (cpumask_empty(new_cpus))
2164 cpumask_copy(new_cpus, parent_cs(cs)->effective_cpus);
2165 if (nodes_empty(*new_mems))
2166 *new_mems = parent_cs(cs)->effective_mems;
2168 spin_lock_irq(&callback_lock);
2169 cpumask_copy(cs->effective_cpus, new_cpus);
2170 cs->effective_mems = *new_mems;
2171 spin_unlock_irq(&callback_lock);
2174 update_tasks_cpumask(cs);
2176 update_tasks_nodemask(cs);
2180 * cpuset_hotplug_update_tasks - update tasks in a cpuset for hotunplug
2181 * @cs: cpuset in interest
2183 * Compare @cs's cpu and mem masks against top_cpuset and if some have gone
2184 * offline, update @cs accordingly. If @cs ends up with no CPU or memory,
2185 * all its tasks are moved to the nearest ancestor with both resources.
2187 static void cpuset_hotplug_update_tasks(struct cpuset *cs)
2189 static cpumask_t new_cpus;
2190 static nodemask_t new_mems;
2194 wait_event(cpuset_attach_wq, cs->attach_in_progress == 0);
2196 mutex_lock(&cpuset_mutex);
2199 * We have raced with task attaching. We wait until attaching
2200 * is finished, so we won't attach a task to an empty cpuset.
2202 if (cs->attach_in_progress) {
2203 mutex_unlock(&cpuset_mutex);
2207 cpumask_and(&new_cpus, cs->cpus_allowed, parent_cs(cs)->effective_cpus);
2208 nodes_and(new_mems, cs->mems_allowed, parent_cs(cs)->effective_mems);
2210 cpus_updated = !cpumask_equal(&new_cpus, cs->effective_cpus);
2211 mems_updated = !nodes_equal(new_mems, cs->effective_mems);
2213 if (cgroup_on_dfl(cs->css.cgroup))
2214 hotplug_update_tasks(cs, &new_cpus, &new_mems,
2215 cpus_updated, mems_updated);
2217 hotplug_update_tasks_legacy(cs, &new_cpus, &new_mems,
2218 cpus_updated, mems_updated);
2220 mutex_unlock(&cpuset_mutex);
2224 * cpuset_hotplug_workfn - handle CPU/memory hotunplug for a cpuset
2226 * This function is called after either CPU or memory configuration has
2227 * changed and updates cpuset accordingly. The top_cpuset is always
2228 * synchronized to cpu_active_mask and N_MEMORY, which is necessary in
2229 * order to make cpusets transparent (of no affect) on systems that are
2230 * actively using CPU hotplug but making no active use of cpusets.
2232 * Non-root cpusets are only affected by offlining. If any CPUs or memory
2233 * nodes have been taken down, cpuset_hotplug_update_tasks() is invoked on
2236 * Note that CPU offlining during suspend is ignored. We don't modify
2237 * cpusets across suspend/resume cycles at all.
2239 static void cpuset_hotplug_workfn(struct work_struct *work)
2241 static cpumask_t new_cpus;
2242 static nodemask_t new_mems;
2243 bool cpus_updated, mems_updated;
2244 bool on_dfl = cgroup_on_dfl(top_cpuset.css.cgroup);
2246 mutex_lock(&cpuset_mutex);
2248 /* fetch the available cpus/mems and find out which changed how */
2249 cpumask_copy(&new_cpus, cpu_active_mask);
2250 new_mems = node_states[N_MEMORY];
2252 cpus_updated = !cpumask_equal(top_cpuset.effective_cpus, &new_cpus);
2253 mems_updated = !nodes_equal(top_cpuset.effective_mems, new_mems);
2255 /* synchronize cpus_allowed to cpu_active_mask */
2257 spin_lock_irq(&callback_lock);
2259 cpumask_copy(top_cpuset.cpus_allowed, &new_cpus);
2260 cpumask_copy(top_cpuset.effective_cpus, &new_cpus);
2261 spin_unlock_irq(&callback_lock);
2262 /* we don't mess with cpumasks of tasks in top_cpuset */
2265 /* synchronize mems_allowed to N_MEMORY */
2267 spin_lock_irq(&callback_lock);
2269 top_cpuset.mems_allowed = new_mems;
2270 top_cpuset.effective_mems = new_mems;
2271 spin_unlock_irq(&callback_lock);
2272 update_tasks_nodemask(&top_cpuset);
2275 mutex_unlock(&cpuset_mutex);
2277 /* if cpus or mems changed, we need to propagate to descendants */
2278 if (cpus_updated || mems_updated) {
2280 struct cgroup_subsys_state *pos_css;
2283 cpuset_for_each_descendant_pre(cs, pos_css, &top_cpuset) {
2284 if (cs == &top_cpuset || !css_tryget_online(&cs->css))
2288 cpuset_hotplug_update_tasks(cs);
2296 /* rebuild sched domains if cpus_allowed has changed */
2298 rebuild_sched_domains();
2301 void cpuset_update_active_cpus(bool cpu_online)
2304 * We're inside cpu hotplug critical region which usually nests
2305 * inside cgroup synchronization. Bounce actual hotplug processing
2306 * to a work item to avoid reverse locking order.
2308 * We still need to do partition_sched_domains() synchronously;
2309 * otherwise, the scheduler will get confused and put tasks to the
2310 * dead CPU. Fall back to the default single domain.
2311 * cpuset_hotplug_workfn() will rebuild it as necessary.
2313 partition_sched_domains(1, NULL, NULL);
2314 schedule_work(&cpuset_hotplug_work);
2318 * Keep top_cpuset.mems_allowed tracking node_states[N_MEMORY].
2319 * Call this routine anytime after node_states[N_MEMORY] changes.
2320 * See cpuset_update_active_cpus() for CPU hotplug handling.
2322 static int cpuset_track_online_nodes(struct notifier_block *self,
2323 unsigned long action, void *arg)
2325 schedule_work(&cpuset_hotplug_work);
2329 static struct notifier_block cpuset_track_online_nodes_nb = {
2330 .notifier_call = cpuset_track_online_nodes,
2331 .priority = 10, /* ??! */
2335 * cpuset_init_smp - initialize cpus_allowed
2337 * Description: Finish top cpuset after cpu, node maps are initialized
2339 void __init cpuset_init_smp(void)
2341 cpumask_copy(top_cpuset.cpus_allowed, cpu_active_mask);
2342 top_cpuset.mems_allowed = node_states[N_MEMORY];
2343 top_cpuset.old_mems_allowed = top_cpuset.mems_allowed;
2345 cpumask_copy(top_cpuset.effective_cpus, cpu_active_mask);
2346 top_cpuset.effective_mems = node_states[N_MEMORY];
2348 register_hotmemory_notifier(&cpuset_track_online_nodes_nb);
2352 * cpuset_cpus_allowed - return cpus_allowed mask from a tasks cpuset.
2353 * @tsk: pointer to task_struct from which to obtain cpuset->cpus_allowed.
2354 * @pmask: pointer to struct cpumask variable to receive cpus_allowed set.
2356 * Description: Returns the cpumask_var_t cpus_allowed of the cpuset
2357 * attached to the specified @tsk. Guaranteed to return some non-empty
2358 * subset of cpu_online_mask, even if this means going outside the
2362 void cpuset_cpus_allowed(struct task_struct *tsk, struct cpumask *pmask)
2364 unsigned long flags;
2366 spin_lock_irqsave(&callback_lock, flags);
2368 guarantee_online_cpus(task_cs(tsk), pmask);
2370 spin_unlock_irqrestore(&callback_lock, flags);
2373 void cpuset_cpus_allowed_fallback(struct task_struct *tsk)
2376 do_set_cpus_allowed(tsk, task_cs(tsk)->effective_cpus);
2380 * We own tsk->cpus_allowed, nobody can change it under us.
2382 * But we used cs && cs->cpus_allowed lockless and thus can
2383 * race with cgroup_attach_task() or update_cpumask() and get
2384 * the wrong tsk->cpus_allowed. However, both cases imply the
2385 * subsequent cpuset_change_cpumask()->set_cpus_allowed_ptr()
2386 * which takes task_rq_lock().
2388 * If we are called after it dropped the lock we must see all
2389 * changes in tsk_cs()->cpus_allowed. Otherwise we can temporary
2390 * set any mask even if it is not right from task_cs() pov,
2391 * the pending set_cpus_allowed_ptr() will fix things.
2393 * select_fallback_rq() will fix things ups and set cpu_possible_mask
2398 void __init cpuset_init_current_mems_allowed(void)
2400 nodes_setall(current->mems_allowed);
2404 * cpuset_mems_allowed - return mems_allowed mask from a tasks cpuset.
2405 * @tsk: pointer to task_struct from which to obtain cpuset->mems_allowed.
2407 * Description: Returns the nodemask_t mems_allowed of the cpuset
2408 * attached to the specified @tsk. Guaranteed to return some non-empty
2409 * subset of node_states[N_MEMORY], even if this means going outside the
2413 nodemask_t cpuset_mems_allowed(struct task_struct *tsk)
2416 unsigned long flags;
2418 spin_lock_irqsave(&callback_lock, flags);
2420 guarantee_online_mems(task_cs(tsk), &mask);
2422 spin_unlock_irqrestore(&callback_lock, flags);
2428 * cpuset_nodemask_valid_mems_allowed - check nodemask vs. curremt mems_allowed
2429 * @nodemask: the nodemask to be checked
2431 * Are any of the nodes in the nodemask allowed in current->mems_allowed?
2433 int cpuset_nodemask_valid_mems_allowed(nodemask_t *nodemask)
2435 return nodes_intersects(*nodemask, current->mems_allowed);
2439 * nearest_hardwall_ancestor() - Returns the nearest mem_exclusive or
2440 * mem_hardwall ancestor to the specified cpuset. Call holding
2441 * callback_lock. If no ancestor is mem_exclusive or mem_hardwall
2442 * (an unusual configuration), then returns the root cpuset.
2444 static struct cpuset *nearest_hardwall_ancestor(struct cpuset *cs)
2446 while (!(is_mem_exclusive(cs) || is_mem_hardwall(cs)) && parent_cs(cs))
2452 * cpuset_node_allowed - Can we allocate on a memory node?
2453 * @node: is this an allowed node?
2454 * @gfp_mask: memory allocation flags
2456 * If we're in interrupt, yes, we can always allocate. If @node is set in
2457 * current's mems_allowed, yes. If it's not a __GFP_HARDWALL request and this
2458 * node is set in the nearest hardwalled cpuset ancestor to current's cpuset,
2459 * yes. If current has access to memory reserves due to TIF_MEMDIE, yes.
2462 * GFP_USER allocations are marked with the __GFP_HARDWALL bit,
2463 * and do not allow allocations outside the current tasks cpuset
2464 * unless the task has been OOM killed as is marked TIF_MEMDIE.
2465 * GFP_KERNEL allocations are not so marked, so can escape to the
2466 * nearest enclosing hardwalled ancestor cpuset.
2468 * Scanning up parent cpusets requires callback_lock. The
2469 * __alloc_pages() routine only calls here with __GFP_HARDWALL bit
2470 * _not_ set if it's a GFP_KERNEL allocation, and all nodes in the
2471 * current tasks mems_allowed came up empty on the first pass over
2472 * the zonelist. So only GFP_KERNEL allocations, if all nodes in the
2473 * cpuset are short of memory, might require taking the callback_lock.
2475 * The first call here from mm/page_alloc:get_page_from_freelist()
2476 * has __GFP_HARDWALL set in gfp_mask, enforcing hardwall cpusets,
2477 * so no allocation on a node outside the cpuset is allowed (unless
2478 * in interrupt, of course).
2480 * The second pass through get_page_from_freelist() doesn't even call
2481 * here for GFP_ATOMIC calls. For those calls, the __alloc_pages()
2482 * variable 'wait' is not set, and the bit ALLOC_CPUSET is not set
2483 * in alloc_flags. That logic and the checks below have the combined
2485 * in_interrupt - any node ok (current task context irrelevant)
2486 * GFP_ATOMIC - any node ok
2487 * TIF_MEMDIE - any node ok
2488 * GFP_KERNEL - any node in enclosing hardwalled cpuset ok
2489 * GFP_USER - only nodes in current tasks mems allowed ok.
2491 int __cpuset_node_allowed(int node, gfp_t gfp_mask)
2493 struct cpuset *cs; /* current cpuset ancestors */
2494 int allowed; /* is allocation in zone z allowed? */
2495 unsigned long flags;
2499 if (node_isset(node, current->mems_allowed))
2502 * Allow tasks that have access to memory reserves because they have
2503 * been OOM killed to get memory anywhere.
2505 if (unlikely(test_thread_flag(TIF_MEMDIE)))
2507 if (gfp_mask & __GFP_HARDWALL) /* If hardwall request, stop here */
2510 if (current->flags & PF_EXITING) /* Let dying task have memory */
2513 /* Not hardwall and node outside mems_allowed: scan up cpusets */
2514 spin_lock_irqsave(&callback_lock, flags);
2517 cs = nearest_hardwall_ancestor(task_cs(current));
2518 allowed = node_isset(node, cs->mems_allowed);
2521 spin_unlock_irqrestore(&callback_lock, flags);
2526 * cpuset_mem_spread_node() - On which node to begin search for a file page
2527 * cpuset_slab_spread_node() - On which node to begin search for a slab page
2529 * If a task is marked PF_SPREAD_PAGE or PF_SPREAD_SLAB (as for
2530 * tasks in a cpuset with is_spread_page or is_spread_slab set),
2531 * and if the memory allocation used cpuset_mem_spread_node()
2532 * to determine on which node to start looking, as it will for
2533 * certain page cache or slab cache pages such as used for file
2534 * system buffers and inode caches, then instead of starting on the
2535 * local node to look for a free page, rather spread the starting
2536 * node around the tasks mems_allowed nodes.
2538 * We don't have to worry about the returned node being offline
2539 * because "it can't happen", and even if it did, it would be ok.
2541 * The routines calling guarantee_online_mems() are careful to
2542 * only set nodes in task->mems_allowed that are online. So it
2543 * should not be possible for the following code to return an
2544 * offline node. But if it did, that would be ok, as this routine
2545 * is not returning the node where the allocation must be, only
2546 * the node where the search should start. The zonelist passed to
2547 * __alloc_pages() will include all nodes. If the slab allocator
2548 * is passed an offline node, it will fall back to the local node.
2549 * See kmem_cache_alloc_node().
2552 static int cpuset_spread_node(int *rotor)
2556 node = next_node(*rotor, current->mems_allowed);
2557 if (node == MAX_NUMNODES)
2558 node = first_node(current->mems_allowed);
2563 int cpuset_mem_spread_node(void)
2565 if (current->cpuset_mem_spread_rotor == NUMA_NO_NODE)
2566 current->cpuset_mem_spread_rotor =
2567 node_random(¤t->mems_allowed);
2569 return cpuset_spread_node(¤t->cpuset_mem_spread_rotor);
2572 int cpuset_slab_spread_node(void)
2574 if (current->cpuset_slab_spread_rotor == NUMA_NO_NODE)
2575 current->cpuset_slab_spread_rotor =
2576 node_random(¤t->mems_allowed);
2578 return cpuset_spread_node(¤t->cpuset_slab_spread_rotor);
2581 EXPORT_SYMBOL_GPL(cpuset_mem_spread_node);
2584 * cpuset_mems_allowed_intersects - Does @tsk1's mems_allowed intersect @tsk2's?
2585 * @tsk1: pointer to task_struct of some task.
2586 * @tsk2: pointer to task_struct of some other task.
2588 * Description: Return true if @tsk1's mems_allowed intersects the
2589 * mems_allowed of @tsk2. Used by the OOM killer to determine if
2590 * one of the task's memory usage might impact the memory available
2594 int cpuset_mems_allowed_intersects(const struct task_struct *tsk1,
2595 const struct task_struct *tsk2)
2597 return nodes_intersects(tsk1->mems_allowed, tsk2->mems_allowed);
2601 * cpuset_print_task_mems_allowed - prints task's cpuset and mems_allowed
2602 * @tsk: pointer to task_struct of some task.
2604 * Description: Prints @task's name, cpuset name, and cached copy of its
2605 * mems_allowed to the kernel log.
2607 void cpuset_print_task_mems_allowed(struct task_struct *tsk)
2609 struct cgroup *cgrp;
2613 cgrp = task_cs(tsk)->css.cgroup;
2614 pr_info("%s cpuset=", tsk->comm);
2615 pr_cont_cgroup_name(cgrp);
2616 pr_cont(" mems_allowed=%*pbl\n", nodemask_pr_args(&tsk->mems_allowed));
2622 * Collection of memory_pressure is suppressed unless
2623 * this flag is enabled by writing "1" to the special
2624 * cpuset file 'memory_pressure_enabled' in the root cpuset.
2627 int cpuset_memory_pressure_enabled __read_mostly;
2630 * cpuset_memory_pressure_bump - keep stats of per-cpuset reclaims.
2632 * Keep a running average of the rate of synchronous (direct)
2633 * page reclaim efforts initiated by tasks in each cpuset.
2635 * This represents the rate at which some task in the cpuset
2636 * ran low on memory on all nodes it was allowed to use, and
2637 * had to enter the kernels page reclaim code in an effort to
2638 * create more free memory by tossing clean pages or swapping
2639 * or writing dirty pages.
2641 * Display to user space in the per-cpuset read-only file
2642 * "memory_pressure". Value displayed is an integer
2643 * representing the recent rate of entry into the synchronous
2644 * (direct) page reclaim by any task attached to the cpuset.
2647 void __cpuset_memory_pressure_bump(void)
2650 fmeter_markevent(&task_cs(current)->fmeter);
2654 #ifdef CONFIG_PROC_PID_CPUSET
2656 * proc_cpuset_show()
2657 * - Print tasks cpuset path into seq_file.
2658 * - Used for /proc/<pid>/cpuset.
2659 * - No need to task_lock(tsk) on this tsk->cpuset reference, as it
2660 * doesn't really matter if tsk->cpuset changes after we read it,
2661 * and we take cpuset_mutex, keeping cpuset_attach() from changing it
2664 int proc_cpuset_show(struct seq_file *m, struct pid_namespace *ns,
2665 struct pid *pid, struct task_struct *tsk)
2668 struct cgroup_subsys_state *css;
2672 buf = kmalloc(PATH_MAX, GFP_KERNEL);
2676 retval = -ENAMETOOLONG;
2678 css = task_css(tsk, cpuset_cgrp_id);
2679 p = cgroup_path(css->cgroup, buf, PATH_MAX);
2691 #endif /* CONFIG_PROC_PID_CPUSET */
2693 /* Display task mems_allowed in /proc/<pid>/status file. */
2694 void cpuset_task_status_allowed(struct seq_file *m, struct task_struct *task)
2696 seq_printf(m, "Mems_allowed:\t%*pb\n",
2697 nodemask_pr_args(&task->mems_allowed));
2698 seq_printf(m, "Mems_allowed_list:\t%*pbl\n",
2699 nodemask_pr_args(&task->mems_allowed));