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 */
80 /* user-configured CPUs and Memory Nodes allow to tasks */
81 cpumask_var_t cpus_allowed;
82 nodemask_t mems_allowed;
84 /* effective CPUs and Memory Nodes allow to tasks */
85 cpumask_var_t effective_cpus;
86 nodemask_t effective_mems;
89 * This is old Memory Nodes tasks took on.
91 * - top_cpuset.old_mems_allowed is initialized to mems_allowed.
92 * - A new cpuset's old_mems_allowed is initialized when some
93 * task is moved into it.
94 * - old_mems_allowed is used in cpuset_migrate_mm() when we change
95 * cpuset.mems_allowed and have tasks' nodemask updated, and
96 * then old_mems_allowed is updated to mems_allowed.
98 nodemask_t old_mems_allowed;
100 struct fmeter fmeter; /* memory_pressure filter */
103 * Tasks are being attached to this cpuset. Used to prevent
104 * zeroing cpus/mems_allowed between ->can_attach() and ->attach().
106 int attach_in_progress;
108 /* partition number for rebuild_sched_domains() */
111 /* for custom sched domain */
112 int relax_domain_level;
115 static inline struct cpuset *css_cs(struct cgroup_subsys_state *css)
117 return css ? container_of(css, struct cpuset, css) : NULL;
120 /* Retrieve the cpuset for a task */
121 static inline struct cpuset *task_cs(struct task_struct *task)
123 return css_cs(task_css(task, cpuset_cgrp_id));
126 static inline struct cpuset *parent_cs(struct cpuset *cs)
128 return css_cs(cs->css.parent);
132 static inline bool task_has_mempolicy(struct task_struct *task)
134 return task->mempolicy;
137 static inline bool task_has_mempolicy(struct task_struct *task)
144 /* bits in struct cpuset flags field */
151 CS_SCHED_LOAD_BALANCE,
156 /* convenient tests for these bits */
157 static inline bool is_cpuset_online(const struct cpuset *cs)
159 return test_bit(CS_ONLINE, &cs->flags);
162 static inline int is_cpu_exclusive(const struct cpuset *cs)
164 return test_bit(CS_CPU_EXCLUSIVE, &cs->flags);
167 static inline int is_mem_exclusive(const struct cpuset *cs)
169 return test_bit(CS_MEM_EXCLUSIVE, &cs->flags);
172 static inline int is_mem_hardwall(const struct cpuset *cs)
174 return test_bit(CS_MEM_HARDWALL, &cs->flags);
177 static inline int is_sched_load_balance(const struct cpuset *cs)
179 return test_bit(CS_SCHED_LOAD_BALANCE, &cs->flags);
182 static inline int is_memory_migrate(const struct cpuset *cs)
184 return test_bit(CS_MEMORY_MIGRATE, &cs->flags);
187 static inline int is_spread_page(const struct cpuset *cs)
189 return test_bit(CS_SPREAD_PAGE, &cs->flags);
192 static inline int is_spread_slab(const struct cpuset *cs)
194 return test_bit(CS_SPREAD_SLAB, &cs->flags);
197 static struct cpuset top_cpuset = {
198 .flags = ((1 << CS_ONLINE) | (1 << CS_CPU_EXCLUSIVE) |
199 (1 << CS_MEM_EXCLUSIVE)),
203 * cpuset_for_each_child - traverse online children of a cpuset
204 * @child_cs: loop cursor pointing to the current child
205 * @pos_css: used for iteration
206 * @parent_cs: target cpuset to walk children of
208 * Walk @child_cs through the online children of @parent_cs. Must be used
209 * with RCU read locked.
211 #define cpuset_for_each_child(child_cs, pos_css, parent_cs) \
212 css_for_each_child((pos_css), &(parent_cs)->css) \
213 if (is_cpuset_online(((child_cs) = css_cs((pos_css)))))
216 * cpuset_for_each_descendant_pre - pre-order walk of a cpuset's descendants
217 * @des_cs: loop cursor pointing to the current descendant
218 * @pos_css: used for iteration
219 * @root_cs: target cpuset to walk ancestor of
221 * Walk @des_cs through the online descendants of @root_cs. Must be used
222 * with RCU read locked. The caller may modify @pos_css by calling
223 * css_rightmost_descendant() to skip subtree. @root_cs is included in the
224 * iteration and the first node to be visited.
226 #define cpuset_for_each_descendant_pre(des_cs, pos_css, root_cs) \
227 css_for_each_descendant_pre((pos_css), &(root_cs)->css) \
228 if (is_cpuset_online(((des_cs) = css_cs((pos_css)))))
231 * There are two global mutexes guarding cpuset structures - cpuset_mutex
232 * and callback_mutex. The latter may nest inside the former. We also
233 * require taking task_lock() when dereferencing a task's cpuset pointer.
234 * See "The task_lock() exception", at the end of this comment.
236 * A task must hold both mutexes to modify cpusets. If a task holds
237 * cpuset_mutex, then it blocks others wanting that mutex, ensuring that it
238 * is the only task able to also acquire callback_mutex and be able to
239 * modify cpusets. It can perform various checks on the cpuset structure
240 * first, knowing nothing will change. It can also allocate memory while
241 * just holding cpuset_mutex. While it is performing these checks, various
242 * callback routines can briefly acquire callback_mutex to query cpusets.
243 * Once it is ready to make the changes, it takes callback_mutex, blocking
246 * Calls to the kernel memory allocator can not be made while holding
247 * callback_mutex, as that would risk double tripping on callback_mutex
248 * from one of the callbacks into the cpuset code from within
251 * If a task is only holding callback_mutex, then it has read-only
254 * Now, the task_struct fields mems_allowed and mempolicy may be changed
255 * by other task, we use alloc_lock in the task_struct fields to protect
258 * The cpuset_common_file_read() handlers only hold callback_mutex across
259 * small pieces of code, such as when reading out possibly multi-word
260 * cpumasks and nodemasks.
262 * Accessing a task's cpuset should be done in accordance with the
263 * guidelines for accessing subsystem state in kernel/cgroup.c
266 static DEFINE_MUTEX(cpuset_mutex);
267 static DEFINE_MUTEX(callback_mutex);
270 * CPU / memory hotplug is handled asynchronously.
272 static void cpuset_hotplug_workfn(struct work_struct *work);
273 static DECLARE_WORK(cpuset_hotplug_work, cpuset_hotplug_workfn);
275 static DECLARE_WAIT_QUEUE_HEAD(cpuset_attach_wq);
278 * This is ugly, but preserves the userspace API for existing cpuset
279 * users. If someone tries to mount the "cpuset" filesystem, we
280 * silently switch it to mount "cgroup" instead
282 static struct dentry *cpuset_mount(struct file_system_type *fs_type,
283 int flags, const char *unused_dev_name, void *data)
285 struct file_system_type *cgroup_fs = get_fs_type("cgroup");
286 struct dentry *ret = ERR_PTR(-ENODEV);
290 "release_agent=/sbin/cpuset_release_agent";
291 ret = cgroup_fs->mount(cgroup_fs, flags,
292 unused_dev_name, mountopts);
293 put_filesystem(cgroup_fs);
298 static struct file_system_type cpuset_fs_type = {
300 .mount = cpuset_mount,
304 * Return in pmask the portion of a cpusets's cpus_allowed that
305 * are online. If none are online, walk up the cpuset hierarchy
306 * until we find one that does have some online cpus. The top
307 * cpuset always has some cpus online.
309 * One way or another, we guarantee to return some non-empty subset
310 * of cpu_online_mask.
312 * Call with callback_mutex held.
314 static void guarantee_online_cpus(struct cpuset *cs, struct cpumask *pmask)
316 while (!cpumask_intersects(cs->cpus_allowed, cpu_online_mask))
318 cpumask_and(pmask, cs->cpus_allowed, cpu_online_mask);
322 * Return in *pmask the portion of a cpusets's mems_allowed that
323 * are online, with memory. If none are online with memory, walk
324 * up the cpuset hierarchy until we find one that does have some
325 * online mems. The top cpuset always has some mems online.
327 * One way or another, we guarantee to return some non-empty subset
328 * of node_states[N_MEMORY].
330 * Call with callback_mutex held.
332 static void guarantee_online_mems(struct cpuset *cs, nodemask_t *pmask)
334 while (!nodes_intersects(cs->mems_allowed, node_states[N_MEMORY]))
336 nodes_and(*pmask, cs->mems_allowed, node_states[N_MEMORY]);
340 * update task's spread flag if cpuset's page/slab spread flag is set
342 * Called with callback_mutex/cpuset_mutex held
344 static void cpuset_update_task_spread_flag(struct cpuset *cs,
345 struct task_struct *tsk)
347 if (is_spread_page(cs))
348 tsk->flags |= PF_SPREAD_PAGE;
350 tsk->flags &= ~PF_SPREAD_PAGE;
351 if (is_spread_slab(cs))
352 tsk->flags |= PF_SPREAD_SLAB;
354 tsk->flags &= ~PF_SPREAD_SLAB;
358 * is_cpuset_subset(p, q) - Is cpuset p a subset of cpuset q?
360 * One cpuset is a subset of another if all its allowed CPUs and
361 * Memory Nodes are a subset of the other, and its exclusive flags
362 * are only set if the other's are set. Call holding cpuset_mutex.
365 static int is_cpuset_subset(const struct cpuset *p, const struct cpuset *q)
367 return cpumask_subset(p->cpus_allowed, q->cpus_allowed) &&
368 nodes_subset(p->mems_allowed, q->mems_allowed) &&
369 is_cpu_exclusive(p) <= is_cpu_exclusive(q) &&
370 is_mem_exclusive(p) <= is_mem_exclusive(q);
374 * alloc_trial_cpuset - allocate a trial cpuset
375 * @cs: the cpuset that the trial cpuset duplicates
377 static struct cpuset *alloc_trial_cpuset(struct cpuset *cs)
379 struct cpuset *trial;
381 trial = kmemdup(cs, sizeof(*cs), GFP_KERNEL);
385 if (!alloc_cpumask_var(&trial->cpus_allowed, GFP_KERNEL))
387 if (!alloc_cpumask_var(&trial->effective_cpus, GFP_KERNEL))
390 cpumask_copy(trial->cpus_allowed, cs->cpus_allowed);
391 cpumask_copy(trial->effective_cpus, cs->effective_cpus);
395 free_cpumask_var(trial->cpus_allowed);
402 * free_trial_cpuset - free the trial cpuset
403 * @trial: the trial cpuset to be freed
405 static void free_trial_cpuset(struct cpuset *trial)
407 free_cpumask_var(trial->effective_cpus);
408 free_cpumask_var(trial->cpus_allowed);
413 * validate_change() - Used to validate that any proposed cpuset change
414 * follows the structural rules for cpusets.
416 * If we replaced the flag and mask values of the current cpuset
417 * (cur) with those values in the trial cpuset (trial), would
418 * our various subset and exclusive rules still be valid? Presumes
421 * 'cur' is the address of an actual, in-use cpuset. Operations
422 * such as list traversal that depend on the actual address of the
423 * cpuset in the list must use cur below, not trial.
425 * 'trial' is the address of bulk structure copy of cur, with
426 * perhaps one or more of the fields cpus_allowed, mems_allowed,
427 * or flags changed to new, trial values.
429 * Return 0 if valid, -errno if not.
432 static int validate_change(struct cpuset *cur, struct cpuset *trial)
434 struct cgroup_subsys_state *css;
435 struct cpuset *c, *par;
440 /* Each of our child cpusets must be a subset of us */
442 cpuset_for_each_child(c, css, cur)
443 if (!is_cpuset_subset(c, trial))
446 /* Remaining checks don't apply to root cpuset */
448 if (cur == &top_cpuset)
451 par = parent_cs(cur);
453 /* We must be a subset of our parent cpuset */
455 if (!is_cpuset_subset(trial, par))
459 * If either I or some sibling (!= me) is exclusive, we can't
463 cpuset_for_each_child(c, css, par) {
464 if ((is_cpu_exclusive(trial) || is_cpu_exclusive(c)) &&
466 cpumask_intersects(trial->cpus_allowed, c->cpus_allowed))
468 if ((is_mem_exclusive(trial) || is_mem_exclusive(c)) &&
470 nodes_intersects(trial->mems_allowed, c->mems_allowed))
475 * Cpusets with tasks - existing or newly being attached - can't
476 * be changed to have empty cpus_allowed or mems_allowed.
479 if ((cgroup_has_tasks(cur->css.cgroup) || cur->attach_in_progress)) {
480 if (!cpumask_empty(cur->cpus_allowed) &&
481 cpumask_empty(trial->cpus_allowed))
483 if (!nodes_empty(cur->mems_allowed) &&
484 nodes_empty(trial->mems_allowed))
496 * Helper routine for generate_sched_domains().
497 * Do cpusets a, b have overlapping effective cpus_allowed masks?
499 static int cpusets_overlap(struct cpuset *a, struct cpuset *b)
501 return cpumask_intersects(a->effective_cpus, b->effective_cpus);
505 update_domain_attr(struct sched_domain_attr *dattr, struct cpuset *c)
507 if (dattr->relax_domain_level < c->relax_domain_level)
508 dattr->relax_domain_level = c->relax_domain_level;
512 static void update_domain_attr_tree(struct sched_domain_attr *dattr,
513 struct cpuset *root_cs)
516 struct cgroup_subsys_state *pos_css;
519 cpuset_for_each_descendant_pre(cp, pos_css, root_cs) {
523 /* skip the whole subtree if @cp doesn't have any CPU */
524 if (cpumask_empty(cp->cpus_allowed)) {
525 pos_css = css_rightmost_descendant(pos_css);
529 if (is_sched_load_balance(cp))
530 update_domain_attr(dattr, cp);
536 * generate_sched_domains()
538 * This function builds a partial partition of the systems CPUs
539 * A 'partial partition' is a set of non-overlapping subsets whose
540 * union is a subset of that set.
541 * The output of this function needs to be passed to kernel/sched/core.c
542 * partition_sched_domains() routine, which will rebuild the scheduler's
543 * load balancing domains (sched domains) as specified by that partial
546 * See "What is sched_load_balance" in Documentation/cgroups/cpusets.txt
547 * for a background explanation of this.
549 * Does not return errors, on the theory that the callers of this
550 * routine would rather not worry about failures to rebuild sched
551 * domains when operating in the severe memory shortage situations
552 * that could cause allocation failures below.
554 * Must be called with cpuset_mutex held.
556 * The three key local variables below are:
557 * q - a linked-list queue of cpuset pointers, used to implement a
558 * top-down scan of all cpusets. This scan loads a pointer
559 * to each cpuset marked is_sched_load_balance into the
560 * array 'csa'. For our purposes, rebuilding the schedulers
561 * sched domains, we can ignore !is_sched_load_balance cpusets.
562 * csa - (for CpuSet Array) Array of pointers to all the cpusets
563 * that need to be load balanced, for convenient iterative
564 * access by the subsequent code that finds the best partition,
565 * i.e the set of domains (subsets) of CPUs such that the
566 * cpus_allowed of every cpuset marked is_sched_load_balance
567 * is a subset of one of these domains, while there are as
568 * many such domains as possible, each as small as possible.
569 * doms - Conversion of 'csa' to an array of cpumasks, for passing to
570 * the kernel/sched/core.c routine partition_sched_domains() in a
571 * convenient format, that can be easily compared to the prior
572 * value to determine what partition elements (sched domains)
573 * were changed (added or removed.)
575 * Finding the best partition (set of domains):
576 * The triple nested loops below over i, j, k scan over the
577 * load balanced cpusets (using the array of cpuset pointers in
578 * csa[]) looking for pairs of cpusets that have overlapping
579 * cpus_allowed, but which don't have the same 'pn' partition
580 * number and gives them in the same partition number. It keeps
581 * looping on the 'restart' label until it can no longer find
584 * The union of the cpus_allowed masks from the set of
585 * all cpusets having the same 'pn' value then form the one
586 * element of the partition (one sched domain) to be passed to
587 * partition_sched_domains().
589 static int generate_sched_domains(cpumask_var_t **domains,
590 struct sched_domain_attr **attributes)
592 struct cpuset *cp; /* scans q */
593 struct cpuset **csa; /* array of all cpuset ptrs */
594 int csn; /* how many cpuset ptrs in csa so far */
595 int i, j, k; /* indices for partition finding loops */
596 cpumask_var_t *doms; /* resulting partition; i.e. sched domains */
597 struct sched_domain_attr *dattr; /* attributes for custom domains */
598 int ndoms = 0; /* number of sched domains in result */
599 int nslot; /* next empty doms[] struct cpumask slot */
600 struct cgroup_subsys_state *pos_css;
606 /* Special case for the 99% of systems with one, full, sched domain */
607 if (is_sched_load_balance(&top_cpuset)) {
609 doms = alloc_sched_domains(ndoms);
613 dattr = kmalloc(sizeof(struct sched_domain_attr), GFP_KERNEL);
615 *dattr = SD_ATTR_INIT;
616 update_domain_attr_tree(dattr, &top_cpuset);
618 cpumask_copy(doms[0], top_cpuset.effective_cpus);
623 csa = kmalloc(nr_cpusets() * sizeof(cp), GFP_KERNEL);
629 cpuset_for_each_descendant_pre(cp, pos_css, &top_cpuset) {
630 if (cp == &top_cpuset)
633 * Continue traversing beyond @cp iff @cp has some CPUs and
634 * isn't load balancing. The former is obvious. The
635 * latter: All child cpusets contain a subset of the
636 * parent's cpus, so just skip them, and then we call
637 * update_domain_attr_tree() to calc relax_domain_level of
638 * the corresponding sched domain.
640 if (!cpumask_empty(cp->cpus_allowed) &&
641 !is_sched_load_balance(cp))
644 if (is_sched_load_balance(cp))
647 /* skip @cp's subtree */
648 pos_css = css_rightmost_descendant(pos_css);
652 for (i = 0; i < csn; i++)
657 /* Find the best partition (set of sched domains) */
658 for (i = 0; i < csn; i++) {
659 struct cpuset *a = csa[i];
662 for (j = 0; j < csn; j++) {
663 struct cpuset *b = csa[j];
666 if (apn != bpn && cpusets_overlap(a, b)) {
667 for (k = 0; k < csn; k++) {
668 struct cpuset *c = csa[k];
673 ndoms--; /* one less element */
680 * Now we know how many domains to create.
681 * Convert <csn, csa> to <ndoms, doms> and populate cpu masks.
683 doms = alloc_sched_domains(ndoms);
688 * The rest of the code, including the scheduler, can deal with
689 * dattr==NULL case. No need to abort if alloc fails.
691 dattr = kmalloc(ndoms * sizeof(struct sched_domain_attr), GFP_KERNEL);
693 for (nslot = 0, i = 0; i < csn; i++) {
694 struct cpuset *a = csa[i];
699 /* Skip completed partitions */
705 if (nslot == ndoms) {
706 static int warnings = 10;
708 pr_warn("rebuild_sched_domains confused: nslot %d, ndoms %d, csn %d, i %d, apn %d\n",
709 nslot, ndoms, csn, i, apn);
717 *(dattr + nslot) = SD_ATTR_INIT;
718 for (j = i; j < csn; j++) {
719 struct cpuset *b = csa[j];
722 cpumask_or(dp, dp, b->effective_cpus);
724 update_domain_attr_tree(dattr + nslot, b);
726 /* Done with this partition */
732 BUG_ON(nslot != ndoms);
738 * Fallback to the default domain if kmalloc() failed.
739 * See comments in partition_sched_domains().
750 * Rebuild scheduler domains.
752 * If the flag 'sched_load_balance' of any cpuset with non-empty
753 * 'cpus' changes, or if the 'cpus' allowed changes in any cpuset
754 * which has that flag enabled, or if any cpuset with a non-empty
755 * 'cpus' is removed, then call this routine to rebuild the
756 * scheduler's dynamic sched domains.
758 * Call with cpuset_mutex held. Takes get_online_cpus().
760 static void rebuild_sched_domains_locked(void)
762 struct sched_domain_attr *attr;
766 lockdep_assert_held(&cpuset_mutex);
770 * We have raced with CPU hotplug. Don't do anything to avoid
771 * passing doms with offlined cpu to partition_sched_domains().
772 * Anyways, hotplug work item will rebuild sched domains.
774 if (!cpumask_equal(top_cpuset.effective_cpus, cpu_active_mask))
777 /* Generate domain masks and attrs */
778 ndoms = generate_sched_domains(&doms, &attr);
780 /* Have scheduler rebuild the domains */
781 partition_sched_domains(ndoms, doms, attr);
785 #else /* !CONFIG_SMP */
786 static void rebuild_sched_domains_locked(void)
789 #endif /* CONFIG_SMP */
791 void rebuild_sched_domains(void)
793 mutex_lock(&cpuset_mutex);
794 rebuild_sched_domains_locked();
795 mutex_unlock(&cpuset_mutex);
799 * effective_cpumask_cpuset - return nearest ancestor with non-empty cpus
800 * @cs: the cpuset in interest
802 * A cpuset's effective cpumask is the cpumask of the nearest ancestor
803 * with non-empty cpus. We use effective cpumask whenever:
804 * - we update tasks' cpus_allowed. (they take on the ancestor's cpumask
805 * if the cpuset they reside in has no cpus)
806 * - we want to retrieve task_cs(tsk)'s cpus_allowed.
808 * Called with cpuset_mutex held. cpuset_cpus_allowed_fallback() is an
809 * exception. See comments there.
811 static struct cpuset *effective_cpumask_cpuset(struct cpuset *cs)
813 while (cpumask_empty(cs->cpus_allowed))
819 * effective_nodemask_cpuset - return nearest ancestor with non-empty mems
820 * @cs: the cpuset in interest
822 * A cpuset's effective nodemask is the nodemask of the nearest ancestor
823 * with non-empty memss. We use effective nodemask whenever:
824 * - we update tasks' mems_allowed. (they take on the ancestor's nodemask
825 * if the cpuset they reside in has no mems)
826 * - we want to retrieve task_cs(tsk)'s mems_allowed.
828 * Called with cpuset_mutex held.
830 static struct cpuset *effective_nodemask_cpuset(struct cpuset *cs)
832 while (nodes_empty(cs->mems_allowed))
838 * update_tasks_cpumask - Update the cpumasks of tasks in the cpuset.
839 * @cs: the cpuset in which each task's cpus_allowed mask needs to be changed
841 * Iterate through each task of @cs updating its cpus_allowed to the
842 * effective cpuset's. As this function is called with cpuset_mutex held,
843 * cpuset membership stays stable.
845 static void update_tasks_cpumask(struct cpuset *cs)
847 struct cpuset *cpus_cs = effective_cpumask_cpuset(cs);
848 struct css_task_iter it;
849 struct task_struct *task;
851 css_task_iter_start(&cs->css, &it);
852 while ((task = css_task_iter_next(&it)))
853 set_cpus_allowed_ptr(task, cpus_cs->cpus_allowed);
854 css_task_iter_end(&it);
858 * update_cpumasks_hier - Update effective cpumasks and tasks in the subtree
859 * @cs: the cpuset to consider
860 * @new_cpus: temp variable for calculating new effective_cpus
862 * When congifured cpumask is changed, the effective cpumasks of this cpuset
863 * and all its descendants need to be updated.
865 * On legacy hierachy, effective_cpus will be the same with cpu_allowed.
867 * Called with cpuset_mutex held
869 static void update_cpumasks_hier(struct cpuset *cs, struct cpumask *new_cpus)
872 struct cgroup_subsys_state *pos_css;
873 bool need_rebuild_sched_domains = false;
876 cpuset_for_each_descendant_pre(cp, pos_css, cs) {
877 struct cpuset *parent = parent_cs(cp);
879 cpumask_and(new_cpus, cp->cpus_allowed, parent->effective_cpus);
882 * If it becomes empty, inherit the effective mask of the
883 * parent, which is guaranteed to have some CPUs.
885 if (cpumask_empty(new_cpus))
886 cpumask_copy(new_cpus, parent->effective_cpus);
888 /* Skip the whole subtree if the cpumask remains the same. */
889 if (cpumask_equal(new_cpus, cp->effective_cpus)) {
890 pos_css = css_rightmost_descendant(pos_css);
894 if (!css_tryget_online(&cp->css))
898 mutex_lock(&callback_mutex);
899 cpumask_copy(cp->effective_cpus, new_cpus);
900 mutex_unlock(&callback_mutex);
902 WARN_ON(!cgroup_on_dfl(cp->css.cgroup) &&
903 !cpumask_equal(cp->cpus_allowed, cp->effective_cpus));
905 update_tasks_cpumask(cp);
908 * If the effective cpumask of any non-empty cpuset is changed,
909 * we need to rebuild sched domains.
911 if (!cpumask_empty(cp->cpus_allowed) &&
912 is_sched_load_balance(cp))
913 need_rebuild_sched_domains = true;
920 if (need_rebuild_sched_domains)
921 rebuild_sched_domains_locked();
925 * update_cpumask - update the cpus_allowed mask of a cpuset and all tasks in it
926 * @cs: the cpuset to consider
927 * @trialcs: trial cpuset
928 * @buf: buffer of cpu numbers written to this cpuset
930 static int update_cpumask(struct cpuset *cs, struct cpuset *trialcs,
935 /* top_cpuset.cpus_allowed tracks cpu_online_mask; it's read-only */
936 if (cs == &top_cpuset)
940 * An empty cpus_allowed is ok only if the cpuset has no tasks.
941 * Since cpulist_parse() fails on an empty mask, we special case
942 * that parsing. The validate_change() call ensures that cpusets
943 * with tasks have cpus.
946 cpumask_clear(trialcs->cpus_allowed);
948 retval = cpulist_parse(buf, trialcs->cpus_allowed);
952 if (!cpumask_subset(trialcs->cpus_allowed, cpu_active_mask))
956 /* Nothing to do if the cpus didn't change */
957 if (cpumask_equal(cs->cpus_allowed, trialcs->cpus_allowed))
960 retval = validate_change(cs, trialcs);
964 mutex_lock(&callback_mutex);
965 cpumask_copy(cs->cpus_allowed, trialcs->cpus_allowed);
966 mutex_unlock(&callback_mutex);
968 /* use trialcs->cpus_allowed as a temp variable */
969 update_cpumasks_hier(cs, trialcs->cpus_allowed);
976 * Migrate memory region from one set of nodes to another.
978 * Temporarilly set tasks mems_allowed to target nodes of migration,
979 * so that the migration code can allocate pages on these nodes.
981 * While the mm_struct we are migrating is typically from some
982 * other task, the task_struct mems_allowed that we are hacking
983 * is for our current task, which must allocate new pages for that
984 * migrating memory region.
987 static void cpuset_migrate_mm(struct mm_struct *mm, const nodemask_t *from,
988 const nodemask_t *to)
990 struct task_struct *tsk = current;
991 struct cpuset *mems_cs;
993 tsk->mems_allowed = *to;
995 do_migrate_pages(mm, from, to, MPOL_MF_MOVE_ALL);
998 mems_cs = effective_nodemask_cpuset(task_cs(tsk));
999 guarantee_online_mems(mems_cs, &tsk->mems_allowed);
1004 * cpuset_change_task_nodemask - change task's mems_allowed and mempolicy
1005 * @tsk: the task to change
1006 * @newmems: new nodes that the task will be set
1008 * In order to avoid seeing no nodes if the old and new nodes are disjoint,
1009 * we structure updates as setting all new allowed nodes, then clearing newly
1012 static void cpuset_change_task_nodemask(struct task_struct *tsk,
1013 nodemask_t *newmems)
1018 * Allow tasks that have access to memory reserves because they have
1019 * been OOM killed to get memory anywhere.
1021 if (unlikely(test_thread_flag(TIF_MEMDIE)))
1023 if (current->flags & PF_EXITING) /* Let dying task have memory */
1028 * Determine if a loop is necessary if another thread is doing
1029 * read_mems_allowed_begin(). If at least one node remains unchanged and
1030 * tsk does not have a mempolicy, then an empty nodemask will not be
1031 * possible when mems_allowed is larger than a word.
1033 need_loop = task_has_mempolicy(tsk) ||
1034 !nodes_intersects(*newmems, tsk->mems_allowed);
1037 local_irq_disable();
1038 write_seqcount_begin(&tsk->mems_allowed_seq);
1041 nodes_or(tsk->mems_allowed, tsk->mems_allowed, *newmems);
1042 mpol_rebind_task(tsk, newmems, MPOL_REBIND_STEP1);
1044 mpol_rebind_task(tsk, newmems, MPOL_REBIND_STEP2);
1045 tsk->mems_allowed = *newmems;
1048 write_seqcount_end(&tsk->mems_allowed_seq);
1055 static void *cpuset_being_rebound;
1058 * update_tasks_nodemask - Update the nodemasks of tasks in the cpuset.
1059 * @cs: the cpuset in which each task's mems_allowed mask needs to be changed
1061 * Iterate through each task of @cs updating its mems_allowed to the
1062 * effective cpuset's. As this function is called with cpuset_mutex held,
1063 * cpuset membership stays stable.
1065 static void update_tasks_nodemask(struct cpuset *cs)
1067 static nodemask_t newmems; /* protected by cpuset_mutex */
1068 struct cpuset *mems_cs = effective_nodemask_cpuset(cs);
1069 struct css_task_iter it;
1070 struct task_struct *task;
1072 cpuset_being_rebound = cs; /* causes mpol_dup() rebind */
1074 guarantee_online_mems(mems_cs, &newmems);
1077 * The mpol_rebind_mm() call takes mmap_sem, which we couldn't
1078 * take while holding tasklist_lock. Forks can happen - the
1079 * mpol_dup() cpuset_being_rebound check will catch such forks,
1080 * and rebind their vma mempolicies too. Because we still hold
1081 * the global cpuset_mutex, we know that no other rebind effort
1082 * will be contending for the global variable cpuset_being_rebound.
1083 * It's ok if we rebind the same mm twice; mpol_rebind_mm()
1084 * is idempotent. Also migrate pages in each mm to new nodes.
1086 css_task_iter_start(&cs->css, &it);
1087 while ((task = css_task_iter_next(&it))) {
1088 struct mm_struct *mm;
1091 cpuset_change_task_nodemask(task, &newmems);
1093 mm = get_task_mm(task);
1097 migrate = is_memory_migrate(cs);
1099 mpol_rebind_mm(mm, &cs->mems_allowed);
1101 cpuset_migrate_mm(mm, &cs->old_mems_allowed, &newmems);
1104 css_task_iter_end(&it);
1107 * All the tasks' nodemasks have been updated, update
1108 * cs->old_mems_allowed.
1110 cs->old_mems_allowed = newmems;
1112 /* We're done rebinding vmas to this cpuset's new mems_allowed. */
1113 cpuset_being_rebound = NULL;
1117 * update_nodemasks_hier - Update effective nodemasks and tasks in the subtree
1118 * @cs: the cpuset to consider
1119 * @new_mems: a temp variable for calculating new effective_mems
1121 * When configured nodemask is changed, the effective nodemasks of this cpuset
1122 * and all its descendants need to be updated.
1124 * On legacy hiearchy, effective_mems will be the same with mems_allowed.
1126 * Called with cpuset_mutex held
1128 static void update_nodemasks_hier(struct cpuset *cs, nodemask_t *new_mems)
1131 struct cgroup_subsys_state *pos_css;
1134 cpuset_for_each_descendant_pre(cp, pos_css, cs) {
1135 struct cpuset *parent = parent_cs(cp);
1137 nodes_and(*new_mems, cp->mems_allowed, parent->effective_mems);
1140 * If it becomes empty, inherit the effective mask of the
1141 * parent, which is guaranteed to have some MEMs.
1143 if (nodes_empty(*new_mems))
1144 *new_mems = parent->effective_mems;
1146 /* Skip the whole subtree if the nodemask remains the same. */
1147 if (nodes_equal(*new_mems, cp->effective_mems)) {
1148 pos_css = css_rightmost_descendant(pos_css);
1152 if (!css_tryget_online(&cp->css))
1156 mutex_lock(&callback_mutex);
1157 cp->effective_mems = *new_mems;
1158 mutex_unlock(&callback_mutex);
1160 WARN_ON(!cgroup_on_dfl(cp->css.cgroup) &&
1161 nodes_equal(cp->mems_allowed, cp->effective_mems));
1163 update_tasks_nodemask(cp);
1172 * Handle user request to change the 'mems' memory placement
1173 * of a cpuset. Needs to validate the request, update the
1174 * cpusets mems_allowed, and for each task in the cpuset,
1175 * update mems_allowed and rebind task's mempolicy and any vma
1176 * mempolicies and if the cpuset is marked 'memory_migrate',
1177 * migrate the tasks pages to the new memory.
1179 * Call with cpuset_mutex held. May take callback_mutex during call.
1180 * Will take tasklist_lock, scan tasklist for tasks in cpuset cs,
1181 * lock each such tasks mm->mmap_sem, scan its vma's and rebind
1182 * their mempolicies to the cpusets new mems_allowed.
1184 static int update_nodemask(struct cpuset *cs, struct cpuset *trialcs,
1190 * top_cpuset.mems_allowed tracks node_stats[N_MEMORY];
1193 if (cs == &top_cpuset) {
1199 * An empty mems_allowed is ok iff there are no tasks in the cpuset.
1200 * Since nodelist_parse() fails on an empty mask, we special case
1201 * that parsing. The validate_change() call ensures that cpusets
1202 * with tasks have memory.
1205 nodes_clear(trialcs->mems_allowed);
1207 retval = nodelist_parse(buf, trialcs->mems_allowed);
1211 if (!nodes_subset(trialcs->mems_allowed,
1212 node_states[N_MEMORY])) {
1218 if (nodes_equal(cs->mems_allowed, trialcs->mems_allowed)) {
1219 retval = 0; /* Too easy - nothing to do */
1222 retval = validate_change(cs, trialcs);
1226 mutex_lock(&callback_mutex);
1227 cs->mems_allowed = trialcs->mems_allowed;
1228 mutex_unlock(&callback_mutex);
1230 /* use trialcs->mems_allowed as a temp variable */
1231 update_nodemasks_hier(cs, &cs->mems_allowed);
1236 int current_cpuset_is_being_rebound(void)
1238 return task_cs(current) == cpuset_being_rebound;
1241 static int update_relax_domain_level(struct cpuset *cs, s64 val)
1244 if (val < -1 || val >= sched_domain_level_max)
1248 if (val != cs->relax_domain_level) {
1249 cs->relax_domain_level = val;
1250 if (!cpumask_empty(cs->cpus_allowed) &&
1251 is_sched_load_balance(cs))
1252 rebuild_sched_domains_locked();
1259 * update_tasks_flags - update the spread flags of tasks in the cpuset.
1260 * @cs: the cpuset in which each task's spread flags needs to be changed
1262 * Iterate through each task of @cs updating its spread flags. As this
1263 * function is called with cpuset_mutex held, cpuset membership stays
1266 static void update_tasks_flags(struct cpuset *cs)
1268 struct css_task_iter it;
1269 struct task_struct *task;
1271 css_task_iter_start(&cs->css, &it);
1272 while ((task = css_task_iter_next(&it)))
1273 cpuset_update_task_spread_flag(cs, task);
1274 css_task_iter_end(&it);
1278 * update_flag - read a 0 or a 1 in a file and update associated flag
1279 * bit: the bit to update (see cpuset_flagbits_t)
1280 * cs: the cpuset to update
1281 * turning_on: whether the flag is being set or cleared
1283 * Call with cpuset_mutex held.
1286 static int update_flag(cpuset_flagbits_t bit, struct cpuset *cs,
1289 struct cpuset *trialcs;
1290 int balance_flag_changed;
1291 int spread_flag_changed;
1294 trialcs = alloc_trial_cpuset(cs);
1299 set_bit(bit, &trialcs->flags);
1301 clear_bit(bit, &trialcs->flags);
1303 err = validate_change(cs, trialcs);
1307 balance_flag_changed = (is_sched_load_balance(cs) !=
1308 is_sched_load_balance(trialcs));
1310 spread_flag_changed = ((is_spread_slab(cs) != is_spread_slab(trialcs))
1311 || (is_spread_page(cs) != is_spread_page(trialcs)));
1313 mutex_lock(&callback_mutex);
1314 cs->flags = trialcs->flags;
1315 mutex_unlock(&callback_mutex);
1317 if (!cpumask_empty(trialcs->cpus_allowed) && balance_flag_changed)
1318 rebuild_sched_domains_locked();
1320 if (spread_flag_changed)
1321 update_tasks_flags(cs);
1323 free_trial_cpuset(trialcs);
1328 * Frequency meter - How fast is some event occurring?
1330 * These routines manage a digitally filtered, constant time based,
1331 * event frequency meter. There are four routines:
1332 * fmeter_init() - initialize a frequency meter.
1333 * fmeter_markevent() - called each time the event happens.
1334 * fmeter_getrate() - returns the recent rate of such events.
1335 * fmeter_update() - internal routine used to update fmeter.
1337 * A common data structure is passed to each of these routines,
1338 * which is used to keep track of the state required to manage the
1339 * frequency meter and its digital filter.
1341 * The filter works on the number of events marked per unit time.
1342 * The filter is single-pole low-pass recursive (IIR). The time unit
1343 * is 1 second. Arithmetic is done using 32-bit integers scaled to
1344 * simulate 3 decimal digits of precision (multiplied by 1000).
1346 * With an FM_COEF of 933, and a time base of 1 second, the filter
1347 * has a half-life of 10 seconds, meaning that if the events quit
1348 * happening, then the rate returned from the fmeter_getrate()
1349 * will be cut in half each 10 seconds, until it converges to zero.
1351 * It is not worth doing a real infinitely recursive filter. If more
1352 * than FM_MAXTICKS ticks have elapsed since the last filter event,
1353 * just compute FM_MAXTICKS ticks worth, by which point the level
1356 * Limit the count of unprocessed events to FM_MAXCNT, so as to avoid
1357 * arithmetic overflow in the fmeter_update() routine.
1359 * Given the simple 32 bit integer arithmetic used, this meter works
1360 * best for reporting rates between one per millisecond (msec) and
1361 * one per 32 (approx) seconds. At constant rates faster than one
1362 * per msec it maxes out at values just under 1,000,000. At constant
1363 * rates between one per msec, and one per second it will stabilize
1364 * to a value N*1000, where N is the rate of events per second.
1365 * At constant rates between one per second and one per 32 seconds,
1366 * it will be choppy, moving up on the seconds that have an event,
1367 * and then decaying until the next event. At rates slower than
1368 * about one in 32 seconds, it decays all the way back to zero between
1372 #define FM_COEF 933 /* coefficient for half-life of 10 secs */
1373 #define FM_MAXTICKS ((time_t)99) /* useless computing more ticks than this */
1374 #define FM_MAXCNT 1000000 /* limit cnt to avoid overflow */
1375 #define FM_SCALE 1000 /* faux fixed point scale */
1377 /* Initialize a frequency meter */
1378 static void fmeter_init(struct fmeter *fmp)
1383 spin_lock_init(&fmp->lock);
1386 /* Internal meter update - process cnt events and update value */
1387 static void fmeter_update(struct fmeter *fmp)
1389 time_t now = get_seconds();
1390 time_t ticks = now - fmp->time;
1395 ticks = min(FM_MAXTICKS, ticks);
1397 fmp->val = (FM_COEF * fmp->val) / FM_SCALE;
1400 fmp->val += ((FM_SCALE - FM_COEF) * fmp->cnt) / FM_SCALE;
1404 /* Process any previous ticks, then bump cnt by one (times scale). */
1405 static void fmeter_markevent(struct fmeter *fmp)
1407 spin_lock(&fmp->lock);
1409 fmp->cnt = min(FM_MAXCNT, fmp->cnt + FM_SCALE);
1410 spin_unlock(&fmp->lock);
1413 /* Process any previous ticks, then return current value. */
1414 static int fmeter_getrate(struct fmeter *fmp)
1418 spin_lock(&fmp->lock);
1421 spin_unlock(&fmp->lock);
1425 static struct cpuset *cpuset_attach_old_cs;
1427 /* Called by cgroups to determine if a cpuset is usable; cpuset_mutex held */
1428 static int cpuset_can_attach(struct cgroup_subsys_state *css,
1429 struct cgroup_taskset *tset)
1431 struct cpuset *cs = css_cs(css);
1432 struct task_struct *task;
1435 /* used later by cpuset_attach() */
1436 cpuset_attach_old_cs = task_cs(cgroup_taskset_first(tset));
1438 mutex_lock(&cpuset_mutex);
1440 /* allow moving tasks into an empty cpuset if on default hierarchy */
1442 if (!cgroup_on_dfl(css->cgroup) &&
1443 (cpumask_empty(cs->cpus_allowed) || nodes_empty(cs->mems_allowed)))
1446 cgroup_taskset_for_each(task, tset) {
1448 * Kthreads which disallow setaffinity shouldn't be moved
1449 * to a new cpuset; we don't want to change their cpu
1450 * affinity and isolating such threads by their set of
1451 * allowed nodes is unnecessary. Thus, cpusets are not
1452 * applicable for such threads. This prevents checking for
1453 * success of set_cpus_allowed_ptr() on all attached tasks
1454 * before cpus_allowed may be changed.
1457 if (task->flags & PF_NO_SETAFFINITY)
1459 ret = security_task_setscheduler(task);
1465 * Mark attach is in progress. This makes validate_change() fail
1466 * changes which zero cpus/mems_allowed.
1468 cs->attach_in_progress++;
1471 mutex_unlock(&cpuset_mutex);
1475 static void cpuset_cancel_attach(struct cgroup_subsys_state *css,
1476 struct cgroup_taskset *tset)
1478 mutex_lock(&cpuset_mutex);
1479 css_cs(css)->attach_in_progress--;
1480 mutex_unlock(&cpuset_mutex);
1484 * Protected by cpuset_mutex. cpus_attach is used only by cpuset_attach()
1485 * but we can't allocate it dynamically there. Define it global and
1486 * allocate from cpuset_init().
1488 static cpumask_var_t cpus_attach;
1490 static void cpuset_attach(struct cgroup_subsys_state *css,
1491 struct cgroup_taskset *tset)
1493 /* static buf protected by cpuset_mutex */
1494 static nodemask_t cpuset_attach_nodemask_to;
1495 struct mm_struct *mm;
1496 struct task_struct *task;
1497 struct task_struct *leader = cgroup_taskset_first(tset);
1498 struct cpuset *cs = css_cs(css);
1499 struct cpuset *oldcs = cpuset_attach_old_cs;
1500 struct cpuset *cpus_cs = effective_cpumask_cpuset(cs);
1501 struct cpuset *mems_cs = effective_nodemask_cpuset(cs);
1503 mutex_lock(&cpuset_mutex);
1505 /* prepare for attach */
1506 if (cs == &top_cpuset)
1507 cpumask_copy(cpus_attach, cpu_possible_mask);
1509 guarantee_online_cpus(cpus_cs, cpus_attach);
1511 guarantee_online_mems(mems_cs, &cpuset_attach_nodemask_to);
1513 cgroup_taskset_for_each(task, tset) {
1515 * can_attach beforehand should guarantee that this doesn't
1516 * fail. TODO: have a better way to handle failure here
1518 WARN_ON_ONCE(set_cpus_allowed_ptr(task, cpus_attach));
1520 cpuset_change_task_nodemask(task, &cpuset_attach_nodemask_to);
1521 cpuset_update_task_spread_flag(cs, task);
1525 * Change mm, possibly for multiple threads in a threadgroup. This is
1526 * expensive and may sleep.
1528 cpuset_attach_nodemask_to = cs->mems_allowed;
1529 mm = get_task_mm(leader);
1531 struct cpuset *mems_oldcs = effective_nodemask_cpuset(oldcs);
1533 mpol_rebind_mm(mm, &cpuset_attach_nodemask_to);
1536 * old_mems_allowed is the same with mems_allowed here, except
1537 * if this task is being moved automatically due to hotplug.
1538 * In that case @mems_allowed has been updated and is empty,
1539 * so @old_mems_allowed is the right nodesets that we migrate
1542 if (is_memory_migrate(cs)) {
1543 cpuset_migrate_mm(mm, &mems_oldcs->old_mems_allowed,
1544 &cpuset_attach_nodemask_to);
1549 cs->old_mems_allowed = cpuset_attach_nodemask_to;
1551 cs->attach_in_progress--;
1552 if (!cs->attach_in_progress)
1553 wake_up(&cpuset_attach_wq);
1555 mutex_unlock(&cpuset_mutex);
1558 /* The various types of files and directories in a cpuset file system */
1561 FILE_MEMORY_MIGRATE,
1567 FILE_SCHED_LOAD_BALANCE,
1568 FILE_SCHED_RELAX_DOMAIN_LEVEL,
1569 FILE_MEMORY_PRESSURE_ENABLED,
1570 FILE_MEMORY_PRESSURE,
1573 } cpuset_filetype_t;
1575 static int cpuset_write_u64(struct cgroup_subsys_state *css, struct cftype *cft,
1578 struct cpuset *cs = css_cs(css);
1579 cpuset_filetype_t type = cft->private;
1582 mutex_lock(&cpuset_mutex);
1583 if (!is_cpuset_online(cs)) {
1589 case FILE_CPU_EXCLUSIVE:
1590 retval = update_flag(CS_CPU_EXCLUSIVE, cs, val);
1592 case FILE_MEM_EXCLUSIVE:
1593 retval = update_flag(CS_MEM_EXCLUSIVE, cs, val);
1595 case FILE_MEM_HARDWALL:
1596 retval = update_flag(CS_MEM_HARDWALL, cs, val);
1598 case FILE_SCHED_LOAD_BALANCE:
1599 retval = update_flag(CS_SCHED_LOAD_BALANCE, cs, val);
1601 case FILE_MEMORY_MIGRATE:
1602 retval = update_flag(CS_MEMORY_MIGRATE, cs, val);
1604 case FILE_MEMORY_PRESSURE_ENABLED:
1605 cpuset_memory_pressure_enabled = !!val;
1607 case FILE_MEMORY_PRESSURE:
1610 case FILE_SPREAD_PAGE:
1611 retval = update_flag(CS_SPREAD_PAGE, cs, val);
1613 case FILE_SPREAD_SLAB:
1614 retval = update_flag(CS_SPREAD_SLAB, cs, val);
1621 mutex_unlock(&cpuset_mutex);
1625 static int cpuset_write_s64(struct cgroup_subsys_state *css, struct cftype *cft,
1628 struct cpuset *cs = css_cs(css);
1629 cpuset_filetype_t type = cft->private;
1630 int retval = -ENODEV;
1632 mutex_lock(&cpuset_mutex);
1633 if (!is_cpuset_online(cs))
1637 case FILE_SCHED_RELAX_DOMAIN_LEVEL:
1638 retval = update_relax_domain_level(cs, val);
1645 mutex_unlock(&cpuset_mutex);
1650 * Common handling for a write to a "cpus" or "mems" file.
1652 static ssize_t cpuset_write_resmask(struct kernfs_open_file *of,
1653 char *buf, size_t nbytes, loff_t off)
1655 struct cpuset *cs = css_cs(of_css(of));
1656 struct cpuset *trialcs;
1657 int retval = -ENODEV;
1659 buf = strstrip(buf);
1662 * CPU or memory hotunplug may leave @cs w/o any execution
1663 * resources, in which case the hotplug code asynchronously updates
1664 * configuration and transfers all tasks to the nearest ancestor
1665 * which can execute.
1667 * As writes to "cpus" or "mems" may restore @cs's execution
1668 * resources, wait for the previously scheduled operations before
1669 * proceeding, so that we don't end up keep removing tasks added
1670 * after execution capability is restored.
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 return retval ?: nbytes;
1703 * These ascii lists should be read in a single call, by using a user
1704 * buffer large enough to hold the entire map. If read in smaller
1705 * chunks, there is no guarantee of atomicity. Since the display format
1706 * used, list of ranges of sequential numbers, is variable length,
1707 * and since these maps can change value dynamically, one could read
1708 * gibberish by doing partial reads while a list was changing.
1710 static int cpuset_common_seq_show(struct seq_file *sf, void *v)
1712 struct cpuset *cs = css_cs(seq_css(sf));
1713 cpuset_filetype_t type = seq_cft(sf)->private;
1718 count = seq_get_buf(sf, &buf);
1721 mutex_lock(&callback_mutex);
1725 s += cpulist_scnprintf(s, count, cs->cpus_allowed);
1728 s += nodelist_scnprintf(s, count, cs->mems_allowed);
1735 if (s < buf + count - 1) {
1737 seq_commit(sf, s - buf);
1742 mutex_unlock(&callback_mutex);
1746 static u64 cpuset_read_u64(struct cgroup_subsys_state *css, struct cftype *cft)
1748 struct cpuset *cs = css_cs(css);
1749 cpuset_filetype_t type = cft->private;
1751 case FILE_CPU_EXCLUSIVE:
1752 return is_cpu_exclusive(cs);
1753 case FILE_MEM_EXCLUSIVE:
1754 return is_mem_exclusive(cs);
1755 case FILE_MEM_HARDWALL:
1756 return is_mem_hardwall(cs);
1757 case FILE_SCHED_LOAD_BALANCE:
1758 return is_sched_load_balance(cs);
1759 case FILE_MEMORY_MIGRATE:
1760 return is_memory_migrate(cs);
1761 case FILE_MEMORY_PRESSURE_ENABLED:
1762 return cpuset_memory_pressure_enabled;
1763 case FILE_MEMORY_PRESSURE:
1764 return fmeter_getrate(&cs->fmeter);
1765 case FILE_SPREAD_PAGE:
1766 return is_spread_page(cs);
1767 case FILE_SPREAD_SLAB:
1768 return is_spread_slab(cs);
1773 /* Unreachable but makes gcc happy */
1777 static s64 cpuset_read_s64(struct cgroup_subsys_state *css, struct cftype *cft)
1779 struct cpuset *cs = css_cs(css);
1780 cpuset_filetype_t type = cft->private;
1782 case FILE_SCHED_RELAX_DOMAIN_LEVEL:
1783 return cs->relax_domain_level;
1788 /* Unrechable but makes gcc happy */
1794 * for the common functions, 'private' gives the type of file
1797 static struct cftype files[] = {
1800 .seq_show = cpuset_common_seq_show,
1801 .write = cpuset_write_resmask,
1802 .max_write_len = (100U + 6 * NR_CPUS),
1803 .private = FILE_CPULIST,
1808 .seq_show = cpuset_common_seq_show,
1809 .write = cpuset_write_resmask,
1810 .max_write_len = (100U + 6 * MAX_NUMNODES),
1811 .private = FILE_MEMLIST,
1815 .name = "cpu_exclusive",
1816 .read_u64 = cpuset_read_u64,
1817 .write_u64 = cpuset_write_u64,
1818 .private = FILE_CPU_EXCLUSIVE,
1822 .name = "mem_exclusive",
1823 .read_u64 = cpuset_read_u64,
1824 .write_u64 = cpuset_write_u64,
1825 .private = FILE_MEM_EXCLUSIVE,
1829 .name = "mem_hardwall",
1830 .read_u64 = cpuset_read_u64,
1831 .write_u64 = cpuset_write_u64,
1832 .private = FILE_MEM_HARDWALL,
1836 .name = "sched_load_balance",
1837 .read_u64 = cpuset_read_u64,
1838 .write_u64 = cpuset_write_u64,
1839 .private = FILE_SCHED_LOAD_BALANCE,
1843 .name = "sched_relax_domain_level",
1844 .read_s64 = cpuset_read_s64,
1845 .write_s64 = cpuset_write_s64,
1846 .private = FILE_SCHED_RELAX_DOMAIN_LEVEL,
1850 .name = "memory_migrate",
1851 .read_u64 = cpuset_read_u64,
1852 .write_u64 = cpuset_write_u64,
1853 .private = FILE_MEMORY_MIGRATE,
1857 .name = "memory_pressure",
1858 .read_u64 = cpuset_read_u64,
1859 .write_u64 = cpuset_write_u64,
1860 .private = FILE_MEMORY_PRESSURE,
1865 .name = "memory_spread_page",
1866 .read_u64 = cpuset_read_u64,
1867 .write_u64 = cpuset_write_u64,
1868 .private = FILE_SPREAD_PAGE,
1872 .name = "memory_spread_slab",
1873 .read_u64 = cpuset_read_u64,
1874 .write_u64 = cpuset_write_u64,
1875 .private = FILE_SPREAD_SLAB,
1879 .name = "memory_pressure_enabled",
1880 .flags = CFTYPE_ONLY_ON_ROOT,
1881 .read_u64 = cpuset_read_u64,
1882 .write_u64 = cpuset_write_u64,
1883 .private = FILE_MEMORY_PRESSURE_ENABLED,
1890 * cpuset_css_alloc - allocate a cpuset css
1891 * cgrp: control group that the new cpuset will be part of
1894 static struct cgroup_subsys_state *
1895 cpuset_css_alloc(struct cgroup_subsys_state *parent_css)
1900 return &top_cpuset.css;
1902 cs = kzalloc(sizeof(*cs), GFP_KERNEL);
1904 return ERR_PTR(-ENOMEM);
1905 if (!alloc_cpumask_var(&cs->cpus_allowed, GFP_KERNEL))
1907 if (!alloc_cpumask_var(&cs->effective_cpus, GFP_KERNEL))
1910 set_bit(CS_SCHED_LOAD_BALANCE, &cs->flags);
1911 cpumask_clear(cs->cpus_allowed);
1912 nodes_clear(cs->mems_allowed);
1913 cpumask_clear(cs->effective_cpus);
1914 nodes_clear(cs->effective_mems);
1915 fmeter_init(&cs->fmeter);
1916 cs->relax_domain_level = -1;
1921 free_cpumask_var(cs->cpus_allowed);
1924 return ERR_PTR(-ENOMEM);
1927 static int cpuset_css_online(struct cgroup_subsys_state *css)
1929 struct cpuset *cs = css_cs(css);
1930 struct cpuset *parent = parent_cs(cs);
1931 struct cpuset *tmp_cs;
1932 struct cgroup_subsys_state *pos_css;
1937 mutex_lock(&cpuset_mutex);
1939 set_bit(CS_ONLINE, &cs->flags);
1940 if (is_spread_page(parent))
1941 set_bit(CS_SPREAD_PAGE, &cs->flags);
1942 if (is_spread_slab(parent))
1943 set_bit(CS_SPREAD_SLAB, &cs->flags);
1947 mutex_lock(&callback_mutex);
1948 if (cgroup_on_dfl(cs->css.cgroup)) {
1949 cpumask_copy(cs->effective_cpus, parent->effective_cpus);
1950 cs->effective_mems = parent->effective_mems;
1952 mutex_unlock(&callback_mutex);
1954 if (!test_bit(CGRP_CPUSET_CLONE_CHILDREN, &css->cgroup->flags))
1958 * Clone @parent's configuration if CGRP_CPUSET_CLONE_CHILDREN is
1959 * set. This flag handling is implemented in cgroup core for
1960 * histrical reasons - the flag may be specified during mount.
1962 * Currently, if any sibling cpusets have exclusive cpus or mem, we
1963 * refuse to clone the configuration - thereby refusing the task to
1964 * be entered, and as a result refusing the sys_unshare() or
1965 * clone() which initiated it. If this becomes a problem for some
1966 * users who wish to allow that scenario, then this could be
1967 * changed to grant parent->cpus_allowed-sibling_cpus_exclusive
1968 * (and likewise for mems) to the new cgroup.
1971 cpuset_for_each_child(tmp_cs, pos_css, parent) {
1972 if (is_mem_exclusive(tmp_cs) || is_cpu_exclusive(tmp_cs)) {
1979 mutex_lock(&callback_mutex);
1980 cs->mems_allowed = parent->mems_allowed;
1981 cpumask_copy(cs->cpus_allowed, parent->cpus_allowed);
1982 mutex_unlock(&callback_mutex);
1984 mutex_unlock(&cpuset_mutex);
1989 * If the cpuset being removed has its flag 'sched_load_balance'
1990 * enabled, then simulate turning sched_load_balance off, which
1991 * will call rebuild_sched_domains_locked().
1994 static void cpuset_css_offline(struct cgroup_subsys_state *css)
1996 struct cpuset *cs = css_cs(css);
1998 mutex_lock(&cpuset_mutex);
2000 if (is_sched_load_balance(cs))
2001 update_flag(CS_SCHED_LOAD_BALANCE, cs, 0);
2004 clear_bit(CS_ONLINE, &cs->flags);
2006 mutex_unlock(&cpuset_mutex);
2009 static void cpuset_css_free(struct cgroup_subsys_state *css)
2011 struct cpuset *cs = css_cs(css);
2013 free_cpumask_var(cs->effective_cpus);
2014 free_cpumask_var(cs->cpus_allowed);
2018 static void cpuset_bind(struct cgroup_subsys_state *root_css)
2020 mutex_lock(&cpuset_mutex);
2021 mutex_lock(&callback_mutex);
2023 if (cgroup_on_dfl(root_css->cgroup)) {
2024 cpumask_copy(top_cpuset.cpus_allowed, cpu_possible_mask);
2025 top_cpuset.mems_allowed = node_possible_map;
2027 cpumask_copy(top_cpuset.cpus_allowed,
2028 top_cpuset.effective_cpus);
2029 top_cpuset.mems_allowed = top_cpuset.effective_mems;
2032 mutex_unlock(&callback_mutex);
2033 mutex_unlock(&cpuset_mutex);
2036 struct cgroup_subsys cpuset_cgrp_subsys = {
2037 .css_alloc = cpuset_css_alloc,
2038 .css_online = cpuset_css_online,
2039 .css_offline = cpuset_css_offline,
2040 .css_free = cpuset_css_free,
2041 .can_attach = cpuset_can_attach,
2042 .cancel_attach = cpuset_cancel_attach,
2043 .attach = cpuset_attach,
2044 .bind = cpuset_bind,
2045 .base_cftypes = files,
2050 * cpuset_init - initialize cpusets at system boot
2052 * Description: Initialize top_cpuset and the cpuset internal file system,
2055 int __init cpuset_init(void)
2059 if (!alloc_cpumask_var(&top_cpuset.cpus_allowed, GFP_KERNEL))
2061 if (!alloc_cpumask_var(&top_cpuset.effective_cpus, GFP_KERNEL))
2064 cpumask_setall(top_cpuset.cpus_allowed);
2065 nodes_setall(top_cpuset.mems_allowed);
2066 cpumask_setall(top_cpuset.effective_cpus);
2067 nodes_setall(top_cpuset.effective_mems);
2069 fmeter_init(&top_cpuset.fmeter);
2070 set_bit(CS_SCHED_LOAD_BALANCE, &top_cpuset.flags);
2071 top_cpuset.relax_domain_level = -1;
2073 err = register_filesystem(&cpuset_fs_type);
2077 if (!alloc_cpumask_var(&cpus_attach, GFP_KERNEL))
2084 * If CPU and/or memory hotplug handlers, below, unplug any CPUs
2085 * or memory nodes, we need to walk over the cpuset hierarchy,
2086 * removing that CPU or node from all cpusets. If this removes the
2087 * last CPU or node from a cpuset, then move the tasks in the empty
2088 * cpuset to its next-highest non-empty parent.
2090 static void remove_tasks_in_empty_cpuset(struct cpuset *cs)
2092 struct cpuset *parent;
2095 * Find its next-highest non-empty parent, (top cpuset
2096 * has online cpus, so can't be empty).
2098 parent = parent_cs(cs);
2099 while (cpumask_empty(parent->cpus_allowed) ||
2100 nodes_empty(parent->mems_allowed))
2101 parent = parent_cs(parent);
2103 if (cgroup_transfer_tasks(parent->css.cgroup, cs->css.cgroup)) {
2104 pr_err("cpuset: failed to transfer tasks out of empty cpuset ");
2105 pr_cont_cgroup_name(cs->css.cgroup);
2111 * cpuset_hotplug_update_tasks - update tasks in a cpuset for hotunplug
2112 * @cs: cpuset in interest
2114 * Compare @cs's cpu and mem masks against top_cpuset and if some have gone
2115 * offline, update @cs accordingly. If @cs ends up with no CPU or memory,
2116 * all its tasks are moved to the nearest ancestor with both resources.
2118 static void cpuset_hotplug_update_tasks(struct cpuset *cs)
2120 static cpumask_t off_cpus;
2121 static nodemask_t off_mems;
2123 bool on_dfl = cgroup_on_dfl(cs->css.cgroup);
2126 wait_event(cpuset_attach_wq, cs->attach_in_progress == 0);
2128 mutex_lock(&cpuset_mutex);
2131 * We have raced with task attaching. We wait until attaching
2132 * is finished, so we won't attach a task to an empty cpuset.
2134 if (cs->attach_in_progress) {
2135 mutex_unlock(&cpuset_mutex);
2139 cpumask_andnot(&off_cpus, cs->cpus_allowed, top_cpuset.cpus_allowed);
2140 nodes_andnot(off_mems, cs->mems_allowed, top_cpuset.mems_allowed);
2142 mutex_lock(&callback_mutex);
2143 cpumask_andnot(cs->cpus_allowed, cs->cpus_allowed, &off_cpus);
2145 /* Inherit the effective mask of the parent, if it becomes empty. */
2146 cpumask_andnot(cs->effective_cpus, cs->effective_cpus, &off_cpus);
2147 if (on_dfl && cpumask_empty(cs->effective_cpus))
2148 cpumask_copy(cs->effective_cpus, parent_cs(cs)->effective_cpus);
2149 mutex_unlock(&callback_mutex);
2152 * If on_dfl, we need to update tasks' cpumask for empty cpuset to
2153 * take on ancestor's cpumask. Otherwise, don't call
2154 * update_tasks_cpumask() if the cpuset becomes empty, as the tasks
2155 * in it will be migrated to an ancestor.
2157 if ((on_dfl && cpumask_empty(cs->cpus_allowed)) ||
2158 (!cpumask_empty(&off_cpus) && !cpumask_empty(cs->cpus_allowed)))
2159 update_tasks_cpumask(cs);
2161 mutex_lock(&callback_mutex);
2162 nodes_andnot(cs->mems_allowed, cs->mems_allowed, off_mems);
2164 /* Inherit the effective mask of the parent, if it becomes empty */
2165 nodes_andnot(cs->effective_mems, cs->effective_mems, off_mems);
2166 if (on_dfl && nodes_empty(cs->effective_mems))
2167 cs->effective_mems = parent_cs(cs)->effective_mems;
2168 mutex_unlock(&callback_mutex);
2171 * If on_dfl, we need to update tasks' nodemask for empty cpuset to
2172 * take on ancestor's nodemask. Otherwise, don't call
2173 * update_tasks_nodemask() if the cpuset becomes empty, as the
2174 * tasks in it will be migratd to an ancestor.
2176 if ((on_dfl && nodes_empty(cs->mems_allowed)) ||
2177 (!nodes_empty(off_mems) && !nodes_empty(cs->mems_allowed)))
2178 update_tasks_nodemask(cs);
2180 is_empty = cpumask_empty(cs->cpus_allowed) ||
2181 nodes_empty(cs->mems_allowed);
2183 mutex_unlock(&cpuset_mutex);
2186 * If on_dfl, we'll keep tasks in empty cpusets.
2188 * Otherwise move tasks to the nearest ancestor with execution
2189 * resources. This is full cgroup operation which will
2190 * also call back into cpuset. Should be done outside any lock.
2192 if (!on_dfl && is_empty)
2193 remove_tasks_in_empty_cpuset(cs);
2197 * cpuset_hotplug_workfn - handle CPU/memory hotunplug for a cpuset
2199 * This function is called after either CPU or memory configuration has
2200 * changed and updates cpuset accordingly. The top_cpuset is always
2201 * synchronized to cpu_active_mask and N_MEMORY, which is necessary in
2202 * order to make cpusets transparent (of no affect) on systems that are
2203 * actively using CPU hotplug but making no active use of cpusets.
2205 * Non-root cpusets are only affected by offlining. If any CPUs or memory
2206 * nodes have been taken down, cpuset_hotplug_update_tasks() is invoked on
2209 * Note that CPU offlining during suspend is ignored. We don't modify
2210 * cpusets across suspend/resume cycles at all.
2212 static void cpuset_hotplug_workfn(struct work_struct *work)
2214 static cpumask_t new_cpus;
2215 static nodemask_t new_mems;
2216 bool cpus_updated, mems_updated;
2218 mutex_lock(&cpuset_mutex);
2220 /* fetch the available cpus/mems and find out which changed how */
2221 cpumask_copy(&new_cpus, cpu_active_mask);
2222 new_mems = node_states[N_MEMORY];
2224 cpus_updated = !cpumask_equal(top_cpuset.cpus_allowed, &new_cpus);
2225 mems_updated = !nodes_equal(top_cpuset.mems_allowed, new_mems);
2227 /* synchronize cpus_allowed to cpu_active_mask */
2229 mutex_lock(&callback_mutex);
2230 cpumask_copy(top_cpuset.cpus_allowed, &new_cpus);
2231 cpumask_copy(top_cpuset.effective_cpus, &new_cpus);
2232 mutex_unlock(&callback_mutex);
2233 /* we don't mess with cpumasks of tasks in top_cpuset */
2236 /* synchronize mems_allowed to N_MEMORY */
2238 mutex_lock(&callback_mutex);
2239 top_cpuset.mems_allowed = new_mems;
2240 top_cpuset.effective_mems = new_mems;
2241 mutex_unlock(&callback_mutex);
2242 update_tasks_nodemask(&top_cpuset);
2245 mutex_unlock(&cpuset_mutex);
2247 /* if cpus or mems changed, we need to propagate to descendants */
2248 if (cpus_updated || mems_updated) {
2250 struct cgroup_subsys_state *pos_css;
2253 cpuset_for_each_descendant_pre(cs, pos_css, &top_cpuset) {
2254 if (cs == &top_cpuset || !css_tryget_online(&cs->css))
2258 cpuset_hotplug_update_tasks(cs);
2266 /* rebuild sched domains if cpus_allowed has changed */
2268 rebuild_sched_domains();
2271 void cpuset_update_active_cpus(bool cpu_online)
2274 * We're inside cpu hotplug critical region which usually nests
2275 * inside cgroup synchronization. Bounce actual hotplug processing
2276 * to a work item to avoid reverse locking order.
2278 * We still need to do partition_sched_domains() synchronously;
2279 * otherwise, the scheduler will get confused and put tasks to the
2280 * dead CPU. Fall back to the default single domain.
2281 * cpuset_hotplug_workfn() will rebuild it as necessary.
2283 partition_sched_domains(1, NULL, NULL);
2284 schedule_work(&cpuset_hotplug_work);
2288 * Keep top_cpuset.mems_allowed tracking node_states[N_MEMORY].
2289 * Call this routine anytime after node_states[N_MEMORY] changes.
2290 * See cpuset_update_active_cpus() for CPU hotplug handling.
2292 static int cpuset_track_online_nodes(struct notifier_block *self,
2293 unsigned long action, void *arg)
2295 schedule_work(&cpuset_hotplug_work);
2299 static struct notifier_block cpuset_track_online_nodes_nb = {
2300 .notifier_call = cpuset_track_online_nodes,
2301 .priority = 10, /* ??! */
2305 * cpuset_init_smp - initialize cpus_allowed
2307 * Description: Finish top cpuset after cpu, node maps are initialized
2309 void __init cpuset_init_smp(void)
2311 cpumask_copy(top_cpuset.cpus_allowed, cpu_active_mask);
2312 top_cpuset.mems_allowed = node_states[N_MEMORY];
2313 top_cpuset.old_mems_allowed = top_cpuset.mems_allowed;
2315 cpumask_copy(top_cpuset.effective_cpus, cpu_active_mask);
2316 top_cpuset.effective_mems = node_states[N_MEMORY];
2318 register_hotmemory_notifier(&cpuset_track_online_nodes_nb);
2322 * cpuset_cpus_allowed - return cpus_allowed mask from a tasks cpuset.
2323 * @tsk: pointer to task_struct from which to obtain cpuset->cpus_allowed.
2324 * @pmask: pointer to struct cpumask variable to receive cpus_allowed set.
2326 * Description: Returns the cpumask_var_t cpus_allowed of the cpuset
2327 * attached to the specified @tsk. Guaranteed to return some non-empty
2328 * subset of cpu_online_mask, even if this means going outside the
2332 void cpuset_cpus_allowed(struct task_struct *tsk, struct cpumask *pmask)
2334 struct cpuset *cpus_cs;
2336 mutex_lock(&callback_mutex);
2338 cpus_cs = effective_cpumask_cpuset(task_cs(tsk));
2339 guarantee_online_cpus(cpus_cs, pmask);
2341 mutex_unlock(&callback_mutex);
2344 void cpuset_cpus_allowed_fallback(struct task_struct *tsk)
2346 struct cpuset *cpus_cs;
2349 cpus_cs = effective_cpumask_cpuset(task_cs(tsk));
2350 do_set_cpus_allowed(tsk, cpus_cs->cpus_allowed);
2354 * We own tsk->cpus_allowed, nobody can change it under us.
2356 * But we used cs && cs->cpus_allowed lockless and thus can
2357 * race with cgroup_attach_task() or update_cpumask() and get
2358 * the wrong tsk->cpus_allowed. However, both cases imply the
2359 * subsequent cpuset_change_cpumask()->set_cpus_allowed_ptr()
2360 * which takes task_rq_lock().
2362 * If we are called after it dropped the lock we must see all
2363 * changes in tsk_cs()->cpus_allowed. Otherwise we can temporary
2364 * set any mask even if it is not right from task_cs() pov,
2365 * the pending set_cpus_allowed_ptr() will fix things.
2367 * select_fallback_rq() will fix things ups and set cpu_possible_mask
2372 void cpuset_init_current_mems_allowed(void)
2374 nodes_setall(current->mems_allowed);
2378 * cpuset_mems_allowed - return mems_allowed mask from a tasks cpuset.
2379 * @tsk: pointer to task_struct from which to obtain cpuset->mems_allowed.
2381 * Description: Returns the nodemask_t mems_allowed of the cpuset
2382 * attached to the specified @tsk. Guaranteed to return some non-empty
2383 * subset of node_states[N_MEMORY], even if this means going outside the
2387 nodemask_t cpuset_mems_allowed(struct task_struct *tsk)
2389 struct cpuset *mems_cs;
2392 mutex_lock(&callback_mutex);
2394 mems_cs = effective_nodemask_cpuset(task_cs(tsk));
2395 guarantee_online_mems(mems_cs, &mask);
2397 mutex_unlock(&callback_mutex);
2403 * cpuset_nodemask_valid_mems_allowed - check nodemask vs. curremt mems_allowed
2404 * @nodemask: the nodemask to be checked
2406 * Are any of the nodes in the nodemask allowed in current->mems_allowed?
2408 int cpuset_nodemask_valid_mems_allowed(nodemask_t *nodemask)
2410 return nodes_intersects(*nodemask, current->mems_allowed);
2414 * nearest_hardwall_ancestor() - Returns the nearest mem_exclusive or
2415 * mem_hardwall ancestor to the specified cpuset. Call holding
2416 * callback_mutex. If no ancestor is mem_exclusive or mem_hardwall
2417 * (an unusual configuration), then returns the root cpuset.
2419 static struct cpuset *nearest_hardwall_ancestor(struct cpuset *cs)
2421 while (!(is_mem_exclusive(cs) || is_mem_hardwall(cs)) && parent_cs(cs))
2427 * cpuset_node_allowed_softwall - Can we allocate on a memory node?
2428 * @node: is this an allowed node?
2429 * @gfp_mask: memory allocation flags
2431 * If we're in interrupt, yes, we can always allocate. If __GFP_THISNODE is
2432 * set, yes, we can always allocate. If node is in our task's mems_allowed,
2433 * yes. If it's not a __GFP_HARDWALL request and this node is in the nearest
2434 * hardwalled cpuset ancestor to this task's cpuset, yes. If the task has been
2435 * OOM killed and has access to memory reserves as specified by the TIF_MEMDIE
2439 * If __GFP_HARDWALL is set, cpuset_node_allowed_softwall() reduces to
2440 * cpuset_node_allowed_hardwall(). Otherwise, cpuset_node_allowed_softwall()
2441 * might sleep, and might allow a node from an enclosing cpuset.
2443 * cpuset_node_allowed_hardwall() only handles the simpler case of hardwall
2444 * cpusets, and never sleeps.
2446 * The __GFP_THISNODE placement logic is really handled elsewhere,
2447 * by forcibly using a zonelist starting at a specified node, and by
2448 * (in get_page_from_freelist()) refusing to consider the zones for
2449 * any node on the zonelist except the first. By the time any such
2450 * calls get to this routine, we should just shut up and say 'yes'.
2452 * GFP_USER allocations are marked with the __GFP_HARDWALL bit,
2453 * and do not allow allocations outside the current tasks cpuset
2454 * unless the task has been OOM killed as is marked TIF_MEMDIE.
2455 * GFP_KERNEL allocations are not so marked, so can escape to the
2456 * nearest enclosing hardwalled ancestor cpuset.
2458 * Scanning up parent cpusets requires callback_mutex. The
2459 * __alloc_pages() routine only calls here with __GFP_HARDWALL bit
2460 * _not_ set if it's a GFP_KERNEL allocation, and all nodes in the
2461 * current tasks mems_allowed came up empty on the first pass over
2462 * the zonelist. So only GFP_KERNEL allocations, if all nodes in the
2463 * cpuset are short of memory, might require taking the callback_mutex
2466 * The first call here from mm/page_alloc:get_page_from_freelist()
2467 * has __GFP_HARDWALL set in gfp_mask, enforcing hardwall cpusets,
2468 * so no allocation on a node outside the cpuset is allowed (unless
2469 * in interrupt, of course).
2471 * The second pass through get_page_from_freelist() doesn't even call
2472 * here for GFP_ATOMIC calls. For those calls, the __alloc_pages()
2473 * variable 'wait' is not set, and the bit ALLOC_CPUSET is not set
2474 * in alloc_flags. That logic and the checks below have the combined
2476 * in_interrupt - any node ok (current task context irrelevant)
2477 * GFP_ATOMIC - any node ok
2478 * TIF_MEMDIE - any node ok
2479 * GFP_KERNEL - any node in enclosing hardwalled cpuset ok
2480 * GFP_USER - only nodes in current tasks mems allowed ok.
2483 * Don't call cpuset_node_allowed_softwall if you can't sleep, unless you
2484 * pass in the __GFP_HARDWALL flag set in gfp_flag, which disables
2485 * the code that might scan up ancestor cpusets and sleep.
2487 int __cpuset_node_allowed_softwall(int node, gfp_t gfp_mask)
2489 struct cpuset *cs; /* current cpuset ancestors */
2490 int allowed; /* is allocation in zone z allowed? */
2492 if (in_interrupt() || (gfp_mask & __GFP_THISNODE))
2494 might_sleep_if(!(gfp_mask & __GFP_HARDWALL));
2495 if (node_isset(node, current->mems_allowed))
2498 * Allow tasks that have access to memory reserves because they have
2499 * been OOM killed to get memory anywhere.
2501 if (unlikely(test_thread_flag(TIF_MEMDIE)))
2503 if (gfp_mask & __GFP_HARDWALL) /* If hardwall request, stop here */
2506 if (current->flags & PF_EXITING) /* Let dying task have memory */
2509 /* Not hardwall and node outside mems_allowed: scan up cpusets */
2510 mutex_lock(&callback_mutex);
2513 cs = nearest_hardwall_ancestor(task_cs(current));
2514 allowed = node_isset(node, cs->mems_allowed);
2517 mutex_unlock(&callback_mutex);
2522 * cpuset_node_allowed_hardwall - Can we allocate on a memory node?
2523 * @node: is this an allowed node?
2524 * @gfp_mask: memory allocation flags
2526 * If we're in interrupt, yes, we can always allocate. If __GFP_THISNODE is
2527 * set, yes, we can always allocate. If node is in our task's mems_allowed,
2528 * yes. If the task has been OOM killed and has access to memory reserves as
2529 * specified by the TIF_MEMDIE flag, yes.
2532 * The __GFP_THISNODE placement logic is really handled elsewhere,
2533 * by forcibly using a zonelist starting at a specified node, and by
2534 * (in get_page_from_freelist()) refusing to consider the zones for
2535 * any node on the zonelist except the first. By the time any such
2536 * calls get to this routine, we should just shut up and say 'yes'.
2538 * Unlike the cpuset_node_allowed_softwall() variant, above,
2539 * this variant requires that the node be in the current task's
2540 * mems_allowed or that we're in interrupt. It does not scan up the
2541 * cpuset hierarchy for the nearest enclosing mem_exclusive cpuset.
2544 int __cpuset_node_allowed_hardwall(int node, gfp_t gfp_mask)
2546 if (in_interrupt() || (gfp_mask & __GFP_THISNODE))
2548 if (node_isset(node, current->mems_allowed))
2551 * Allow tasks that have access to memory reserves because they have
2552 * been OOM killed to get memory anywhere.
2554 if (unlikely(test_thread_flag(TIF_MEMDIE)))
2560 * cpuset_mem_spread_node() - On which node to begin search for a file page
2561 * cpuset_slab_spread_node() - On which node to begin search for a slab page
2563 * If a task is marked PF_SPREAD_PAGE or PF_SPREAD_SLAB (as for
2564 * tasks in a cpuset with is_spread_page or is_spread_slab set),
2565 * and if the memory allocation used cpuset_mem_spread_node()
2566 * to determine on which node to start looking, as it will for
2567 * certain page cache or slab cache pages such as used for file
2568 * system buffers and inode caches, then instead of starting on the
2569 * local node to look for a free page, rather spread the starting
2570 * node around the tasks mems_allowed nodes.
2572 * We don't have to worry about the returned node being offline
2573 * because "it can't happen", and even if it did, it would be ok.
2575 * The routines calling guarantee_online_mems() are careful to
2576 * only set nodes in task->mems_allowed that are online. So it
2577 * should not be possible for the following code to return an
2578 * offline node. But if it did, that would be ok, as this routine
2579 * is not returning the node where the allocation must be, only
2580 * the node where the search should start. The zonelist passed to
2581 * __alloc_pages() will include all nodes. If the slab allocator
2582 * is passed an offline node, it will fall back to the local node.
2583 * See kmem_cache_alloc_node().
2586 static int cpuset_spread_node(int *rotor)
2590 node = next_node(*rotor, current->mems_allowed);
2591 if (node == MAX_NUMNODES)
2592 node = first_node(current->mems_allowed);
2597 int cpuset_mem_spread_node(void)
2599 if (current->cpuset_mem_spread_rotor == NUMA_NO_NODE)
2600 current->cpuset_mem_spread_rotor =
2601 node_random(¤t->mems_allowed);
2603 return cpuset_spread_node(¤t->cpuset_mem_spread_rotor);
2606 int cpuset_slab_spread_node(void)
2608 if (current->cpuset_slab_spread_rotor == NUMA_NO_NODE)
2609 current->cpuset_slab_spread_rotor =
2610 node_random(¤t->mems_allowed);
2612 return cpuset_spread_node(¤t->cpuset_slab_spread_rotor);
2615 EXPORT_SYMBOL_GPL(cpuset_mem_spread_node);
2618 * cpuset_mems_allowed_intersects - Does @tsk1's mems_allowed intersect @tsk2's?
2619 * @tsk1: pointer to task_struct of some task.
2620 * @tsk2: pointer to task_struct of some other task.
2622 * Description: Return true if @tsk1's mems_allowed intersects the
2623 * mems_allowed of @tsk2. Used by the OOM killer to determine if
2624 * one of the task's memory usage might impact the memory available
2628 int cpuset_mems_allowed_intersects(const struct task_struct *tsk1,
2629 const struct task_struct *tsk2)
2631 return nodes_intersects(tsk1->mems_allowed, tsk2->mems_allowed);
2634 #define CPUSET_NODELIST_LEN (256)
2637 * cpuset_print_task_mems_allowed - prints task's cpuset and mems_allowed
2638 * @tsk: pointer to task_struct of some task.
2640 * Description: Prints @task's name, cpuset name, and cached copy of its
2641 * mems_allowed to the kernel log.
2643 void cpuset_print_task_mems_allowed(struct task_struct *tsk)
2645 /* Statically allocated to prevent using excess stack. */
2646 static char cpuset_nodelist[CPUSET_NODELIST_LEN];
2647 static DEFINE_SPINLOCK(cpuset_buffer_lock);
2648 struct cgroup *cgrp;
2650 spin_lock(&cpuset_buffer_lock);
2653 cgrp = task_cs(tsk)->css.cgroup;
2654 nodelist_scnprintf(cpuset_nodelist, CPUSET_NODELIST_LEN,
2656 pr_info("%s cpuset=", tsk->comm);
2657 pr_cont_cgroup_name(cgrp);
2658 pr_cont(" mems_allowed=%s\n", cpuset_nodelist);
2661 spin_unlock(&cpuset_buffer_lock);
2665 * Collection of memory_pressure is suppressed unless
2666 * this flag is enabled by writing "1" to the special
2667 * cpuset file 'memory_pressure_enabled' in the root cpuset.
2670 int cpuset_memory_pressure_enabled __read_mostly;
2673 * cpuset_memory_pressure_bump - keep stats of per-cpuset reclaims.
2675 * Keep a running average of the rate of synchronous (direct)
2676 * page reclaim efforts initiated by tasks in each cpuset.
2678 * This represents the rate at which some task in the cpuset
2679 * ran low on memory on all nodes it was allowed to use, and
2680 * had to enter the kernels page reclaim code in an effort to
2681 * create more free memory by tossing clean pages or swapping
2682 * or writing dirty pages.
2684 * Display to user space in the per-cpuset read-only file
2685 * "memory_pressure". Value displayed is an integer
2686 * representing the recent rate of entry into the synchronous
2687 * (direct) page reclaim by any task attached to the cpuset.
2690 void __cpuset_memory_pressure_bump(void)
2693 fmeter_markevent(&task_cs(current)->fmeter);
2697 #ifdef CONFIG_PROC_PID_CPUSET
2699 * proc_cpuset_show()
2700 * - Print tasks cpuset path into seq_file.
2701 * - Used for /proc/<pid>/cpuset.
2702 * - No need to task_lock(tsk) on this tsk->cpuset reference, as it
2703 * doesn't really matter if tsk->cpuset changes after we read it,
2704 * and we take cpuset_mutex, keeping cpuset_attach() from changing it
2707 int proc_cpuset_show(struct seq_file *m, void *unused_v)
2710 struct task_struct *tsk;
2712 struct cgroup_subsys_state *css;
2716 buf = kmalloc(PATH_MAX, GFP_KERNEL);
2722 tsk = get_pid_task(pid, PIDTYPE_PID);
2726 retval = -ENAMETOOLONG;
2728 css = task_css(tsk, cpuset_cgrp_id);
2729 p = cgroup_path(css->cgroup, buf, PATH_MAX);
2737 put_task_struct(tsk);
2743 #endif /* CONFIG_PROC_PID_CPUSET */
2745 /* Display task mems_allowed in /proc/<pid>/status file. */
2746 void cpuset_task_status_allowed(struct seq_file *m, struct task_struct *task)
2748 seq_puts(m, "Mems_allowed:\t");
2749 seq_nodemask(m, &task->mems_allowed);
2751 seq_puts(m, "Mems_allowed_list:\t");
2752 seq_nodemask_list(m, &task->mems_allowed);