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>
65 * Tracks how many cpusets are currently defined in system.
66 * When there is only one cpuset (the root cpuset) we can
67 * short circuit some hooks.
69 int number_of_cpusets __read_mostly;
71 /* Forward declare cgroup structures */
72 struct cgroup_subsys cpuset_subsys;
75 /* See "Frequency meter" comments, below. */
78 int cnt; /* unprocessed events count */
79 int val; /* most recent output value */
80 time_t time; /* clock (secs) when val computed */
81 spinlock_t lock; /* guards read or write of above */
85 struct cgroup_subsys_state css;
87 unsigned long flags; /* "unsigned long" so bitops work */
88 cpumask_var_t cpus_allowed; /* CPUs allowed to tasks in cpuset */
89 nodemask_t mems_allowed; /* Memory Nodes allowed to tasks */
92 * This is old Memory Nodes tasks took on.
94 * - top_cpuset.old_mems_allowed is initialized to mems_allowed.
95 * - A new cpuset's old_mems_allowed is initialized when some
96 * task is moved into it.
97 * - old_mems_allowed is used in cpuset_migrate_mm() when we change
98 * cpuset.mems_allowed and have tasks' nodemask updated, and
99 * then old_mems_allowed is updated to mems_allowed.
101 nodemask_t old_mems_allowed;
103 struct fmeter fmeter; /* memory_pressure filter */
106 * Tasks are being attached to this cpuset. Used to prevent
107 * zeroing cpus/mems_allowed between ->can_attach() and ->attach().
109 int attach_in_progress;
111 /* partition number for rebuild_sched_domains() */
114 /* for custom sched domain */
115 int relax_domain_level;
118 /* Retrieve the cpuset for a cgroup */
119 static inline struct cpuset *cgroup_cs(struct cgroup *cgrp)
121 return container_of(cgroup_subsys_state(cgrp, cpuset_subsys_id),
125 /* Retrieve the cpuset for a task */
126 static inline struct cpuset *task_cs(struct task_struct *task)
128 return container_of(task_subsys_state(task, cpuset_subsys_id),
132 static inline struct cpuset *parent_cs(const struct cpuset *cs)
134 struct cgroup *pcgrp = cs->css.cgroup->parent;
137 return cgroup_cs(pcgrp);
142 static inline bool task_has_mempolicy(struct task_struct *task)
144 return task->mempolicy;
147 static inline bool task_has_mempolicy(struct task_struct *task)
154 /* bits in struct cpuset flags field */
161 CS_SCHED_LOAD_BALANCE,
166 /* convenient tests for these bits */
167 static inline bool is_cpuset_online(const struct cpuset *cs)
169 return test_bit(CS_ONLINE, &cs->flags);
172 static inline int is_cpu_exclusive(const struct cpuset *cs)
174 return test_bit(CS_CPU_EXCLUSIVE, &cs->flags);
177 static inline int is_mem_exclusive(const struct cpuset *cs)
179 return test_bit(CS_MEM_EXCLUSIVE, &cs->flags);
182 static inline int is_mem_hardwall(const struct cpuset *cs)
184 return test_bit(CS_MEM_HARDWALL, &cs->flags);
187 static inline int is_sched_load_balance(const struct cpuset *cs)
189 return test_bit(CS_SCHED_LOAD_BALANCE, &cs->flags);
192 static inline int is_memory_migrate(const struct cpuset *cs)
194 return test_bit(CS_MEMORY_MIGRATE, &cs->flags);
197 static inline int is_spread_page(const struct cpuset *cs)
199 return test_bit(CS_SPREAD_PAGE, &cs->flags);
202 static inline int is_spread_slab(const struct cpuset *cs)
204 return test_bit(CS_SPREAD_SLAB, &cs->flags);
207 static struct cpuset top_cpuset = {
208 .flags = ((1 << CS_ONLINE) | (1 << CS_CPU_EXCLUSIVE) |
209 (1 << CS_MEM_EXCLUSIVE)),
213 * cpuset_for_each_child - traverse online children of a cpuset
214 * @child_cs: loop cursor pointing to the current child
215 * @pos_cgrp: used for iteration
216 * @parent_cs: target cpuset to walk children of
218 * Walk @child_cs through the online children of @parent_cs. Must be used
219 * with RCU read locked.
221 #define cpuset_for_each_child(child_cs, pos_cgrp, parent_cs) \
222 cgroup_for_each_child((pos_cgrp), (parent_cs)->css.cgroup) \
223 if (is_cpuset_online(((child_cs) = cgroup_cs((pos_cgrp)))))
226 * cpuset_for_each_descendant_pre - pre-order walk of a cpuset's descendants
227 * @des_cs: loop cursor pointing to the current descendant
228 * @pos_cgrp: used for iteration
229 * @root_cs: target cpuset to walk ancestor of
231 * Walk @des_cs through the online descendants of @root_cs. Must be used
232 * with RCU read locked. The caller may modify @pos_cgrp by calling
233 * cgroup_rightmost_descendant() to skip subtree.
235 #define cpuset_for_each_descendant_pre(des_cs, pos_cgrp, root_cs) \
236 cgroup_for_each_descendant_pre((pos_cgrp), (root_cs)->css.cgroup) \
237 if (is_cpuset_online(((des_cs) = cgroup_cs((pos_cgrp)))))
240 * There are two global mutexes guarding cpuset structures - cpuset_mutex
241 * and callback_mutex. The latter may nest inside the former. We also
242 * require taking task_lock() when dereferencing a task's cpuset pointer.
243 * See "The task_lock() exception", at the end of this comment.
245 * A task must hold both mutexes to modify cpusets. If a task holds
246 * cpuset_mutex, then it blocks others wanting that mutex, ensuring that it
247 * is the only task able to also acquire callback_mutex and be able to
248 * modify cpusets. It can perform various checks on the cpuset structure
249 * first, knowing nothing will change. It can also allocate memory while
250 * just holding cpuset_mutex. While it is performing these checks, various
251 * callback routines can briefly acquire callback_mutex to query cpusets.
252 * Once it is ready to make the changes, it takes callback_mutex, blocking
255 * Calls to the kernel memory allocator can not be made while holding
256 * callback_mutex, as that would risk double tripping on callback_mutex
257 * from one of the callbacks into the cpuset code from within
260 * If a task is only holding callback_mutex, then it has read-only
263 * Now, the task_struct fields mems_allowed and mempolicy may be changed
264 * by other task, we use alloc_lock in the task_struct fields to protect
267 * The cpuset_common_file_read() handlers only hold callback_mutex across
268 * small pieces of code, such as when reading out possibly multi-word
269 * cpumasks and nodemasks.
271 * Accessing a task's cpuset should be done in accordance with the
272 * guidelines for accessing subsystem state in kernel/cgroup.c
275 static DEFINE_MUTEX(cpuset_mutex);
276 static DEFINE_MUTEX(callback_mutex);
279 * CPU / memory hotplug is handled asynchronously.
281 static void cpuset_hotplug_workfn(struct work_struct *work);
282 static DECLARE_WORK(cpuset_hotplug_work, cpuset_hotplug_workfn);
284 static DECLARE_WAIT_QUEUE_HEAD(cpuset_attach_wq);
287 * This is ugly, but preserves the userspace API for existing cpuset
288 * users. If someone tries to mount the "cpuset" filesystem, we
289 * silently switch it to mount "cgroup" instead
291 static struct dentry *cpuset_mount(struct file_system_type *fs_type,
292 int flags, const char *unused_dev_name, void *data)
294 struct file_system_type *cgroup_fs = get_fs_type("cgroup");
295 struct dentry *ret = ERR_PTR(-ENODEV);
299 "release_agent=/sbin/cpuset_release_agent";
300 ret = cgroup_fs->mount(cgroup_fs, flags,
301 unused_dev_name, mountopts);
302 put_filesystem(cgroup_fs);
307 static struct file_system_type cpuset_fs_type = {
309 .mount = cpuset_mount,
313 * Return in pmask the portion of a cpusets's cpus_allowed that
314 * are online. If none are online, walk up the cpuset hierarchy
315 * until we find one that does have some online cpus. The top
316 * cpuset always has some cpus online.
318 * One way or another, we guarantee to return some non-empty subset
319 * of cpu_online_mask.
321 * Call with callback_mutex held.
323 static void guarantee_online_cpus(const struct cpuset *cs,
324 struct cpumask *pmask)
326 while (!cpumask_intersects(cs->cpus_allowed, cpu_online_mask))
328 cpumask_and(pmask, cs->cpus_allowed, cpu_online_mask);
332 * Return in *pmask the portion of a cpusets's mems_allowed that
333 * are online, with memory. If none are online with memory, walk
334 * up the cpuset hierarchy until we find one that does have some
335 * online mems. The top cpuset always has some mems online.
337 * One way or another, we guarantee to return some non-empty subset
338 * of node_states[N_MEMORY].
340 * Call with callback_mutex held.
342 static void guarantee_online_mems(const struct cpuset *cs, nodemask_t *pmask)
344 while (!nodes_intersects(cs->mems_allowed, node_states[N_MEMORY]))
346 nodes_and(*pmask, cs->mems_allowed, node_states[N_MEMORY]);
350 * update task's spread flag if cpuset's page/slab spread flag is set
352 * Called with callback_mutex/cpuset_mutex held
354 static void cpuset_update_task_spread_flag(struct cpuset *cs,
355 struct task_struct *tsk)
357 if (is_spread_page(cs))
358 tsk->flags |= PF_SPREAD_PAGE;
360 tsk->flags &= ~PF_SPREAD_PAGE;
361 if (is_spread_slab(cs))
362 tsk->flags |= PF_SPREAD_SLAB;
364 tsk->flags &= ~PF_SPREAD_SLAB;
368 * is_cpuset_subset(p, q) - Is cpuset p a subset of cpuset q?
370 * One cpuset is a subset of another if all its allowed CPUs and
371 * Memory Nodes are a subset of the other, and its exclusive flags
372 * are only set if the other's are set. Call holding cpuset_mutex.
375 static int is_cpuset_subset(const struct cpuset *p, const struct cpuset *q)
377 return cpumask_subset(p->cpus_allowed, q->cpus_allowed) &&
378 nodes_subset(p->mems_allowed, q->mems_allowed) &&
379 is_cpu_exclusive(p) <= is_cpu_exclusive(q) &&
380 is_mem_exclusive(p) <= is_mem_exclusive(q);
384 * alloc_trial_cpuset - allocate a trial cpuset
385 * @cs: the cpuset that the trial cpuset duplicates
387 static struct cpuset *alloc_trial_cpuset(const struct cpuset *cs)
389 struct cpuset *trial;
391 trial = kmemdup(cs, sizeof(*cs), GFP_KERNEL);
395 if (!alloc_cpumask_var(&trial->cpus_allowed, GFP_KERNEL)) {
399 cpumask_copy(trial->cpus_allowed, cs->cpus_allowed);
405 * free_trial_cpuset - free the trial cpuset
406 * @trial: the trial cpuset to be freed
408 static void free_trial_cpuset(struct cpuset *trial)
410 free_cpumask_var(trial->cpus_allowed);
415 * validate_change() - Used to validate that any proposed cpuset change
416 * follows the structural rules for cpusets.
418 * If we replaced the flag and mask values of the current cpuset
419 * (cur) with those values in the trial cpuset (trial), would
420 * our various subset and exclusive rules still be valid? Presumes
423 * 'cur' is the address of an actual, in-use cpuset. Operations
424 * such as list traversal that depend on the actual address of the
425 * cpuset in the list must use cur below, not trial.
427 * 'trial' is the address of bulk structure copy of cur, with
428 * perhaps one or more of the fields cpus_allowed, mems_allowed,
429 * or flags changed to new, trial values.
431 * Return 0 if valid, -errno if not.
434 static int validate_change(const struct cpuset *cur, const struct cpuset *trial)
437 struct cpuset *c, *par;
442 /* Each of our child cpusets must be a subset of us */
444 cpuset_for_each_child(c, cgrp, cur)
445 if (!is_cpuset_subset(c, trial))
448 /* Remaining checks don't apply to root cpuset */
450 if (cur == &top_cpuset)
453 par = parent_cs(cur);
455 /* We must be a subset of our parent cpuset */
457 if (!is_cpuset_subset(trial, par))
461 * If either I or some sibling (!= me) is exclusive, we can't
465 cpuset_for_each_child(c, cgrp, par) {
466 if ((is_cpu_exclusive(trial) || is_cpu_exclusive(c)) &&
468 cpumask_intersects(trial->cpus_allowed, c->cpus_allowed))
470 if ((is_mem_exclusive(trial) || is_mem_exclusive(c)) &&
472 nodes_intersects(trial->mems_allowed, c->mems_allowed))
477 * Cpusets with tasks - existing or newly being attached - can't
478 * be changed to have empty cpus_allowed or mems_allowed.
481 if ((cgroup_task_count(cur->css.cgroup) || cur->attach_in_progress)) {
482 if (!cpumask_empty(cur->cpus_allowed) &&
483 cpumask_empty(trial->cpus_allowed))
485 if (!nodes_empty(cur->mems_allowed) &&
486 nodes_empty(trial->mems_allowed))
498 * Helper routine for generate_sched_domains().
499 * Do cpusets a, b have overlapping cpus_allowed masks?
501 static int cpusets_overlap(struct cpuset *a, struct cpuset *b)
503 return cpumask_intersects(a->cpus_allowed, b->cpus_allowed);
507 update_domain_attr(struct sched_domain_attr *dattr, struct cpuset *c)
509 if (dattr->relax_domain_level < c->relax_domain_level)
510 dattr->relax_domain_level = c->relax_domain_level;
514 static void update_domain_attr_tree(struct sched_domain_attr *dattr,
515 struct cpuset *root_cs)
518 struct cgroup *pos_cgrp;
521 cpuset_for_each_descendant_pre(cp, pos_cgrp, root_cs) {
522 /* skip the whole subtree if @cp doesn't have any CPU */
523 if (cpumask_empty(cp->cpus_allowed)) {
524 pos_cgrp = cgroup_rightmost_descendant(pos_cgrp);
528 if (is_sched_load_balance(cp))
529 update_domain_attr(dattr, cp);
535 * generate_sched_domains()
537 * This function builds a partial partition of the systems CPUs
538 * A 'partial partition' is a set of non-overlapping subsets whose
539 * union is a subset of that set.
540 * The output of this function needs to be passed to kernel/sched/core.c
541 * partition_sched_domains() routine, which will rebuild the scheduler's
542 * load balancing domains (sched domains) as specified by that partial
545 * See "What is sched_load_balance" in Documentation/cgroups/cpusets.txt
546 * for a background explanation of this.
548 * Does not return errors, on the theory that the callers of this
549 * routine would rather not worry about failures to rebuild sched
550 * domains when operating in the severe memory shortage situations
551 * that could cause allocation failures below.
553 * Must be called with cpuset_mutex held.
555 * The three key local variables below are:
556 * q - a linked-list queue of cpuset pointers, used to implement a
557 * top-down scan of all cpusets. This scan loads a pointer
558 * to each cpuset marked is_sched_load_balance into the
559 * array 'csa'. For our purposes, rebuilding the schedulers
560 * sched domains, we can ignore !is_sched_load_balance cpusets.
561 * csa - (for CpuSet Array) Array of pointers to all the cpusets
562 * that need to be load balanced, for convenient iterative
563 * access by the subsequent code that finds the best partition,
564 * i.e the set of domains (subsets) of CPUs such that the
565 * cpus_allowed of every cpuset marked is_sched_load_balance
566 * is a subset of one of these domains, while there are as
567 * many such domains as possible, each as small as possible.
568 * doms - Conversion of 'csa' to an array of cpumasks, for passing to
569 * the kernel/sched/core.c routine partition_sched_domains() in a
570 * convenient format, that can be easily compared to the prior
571 * value to determine what partition elements (sched domains)
572 * were changed (added or removed.)
574 * Finding the best partition (set of domains):
575 * The triple nested loops below over i, j, k scan over the
576 * load balanced cpusets (using the array of cpuset pointers in
577 * csa[]) looking for pairs of cpusets that have overlapping
578 * cpus_allowed, but which don't have the same 'pn' partition
579 * number and gives them in the same partition number. It keeps
580 * looping on the 'restart' label until it can no longer find
583 * The union of the cpus_allowed masks from the set of
584 * all cpusets having the same 'pn' value then form the one
585 * element of the partition (one sched domain) to be passed to
586 * partition_sched_domains().
588 static int generate_sched_domains(cpumask_var_t **domains,
589 struct sched_domain_attr **attributes)
591 struct cpuset *cp; /* scans q */
592 struct cpuset **csa; /* array of all cpuset ptrs */
593 int csn; /* how many cpuset ptrs in csa so far */
594 int i, j, k; /* indices for partition finding loops */
595 cpumask_var_t *doms; /* resulting partition; i.e. sched domains */
596 struct sched_domain_attr *dattr; /* attributes for custom domains */
597 int ndoms = 0; /* number of sched domains in result */
598 int nslot; /* next empty doms[] struct cpumask slot */
599 struct cgroup *pos_cgrp;
605 /* Special case for the 99% of systems with one, full, sched domain */
606 if (is_sched_load_balance(&top_cpuset)) {
608 doms = alloc_sched_domains(ndoms);
612 dattr = kmalloc(sizeof(struct sched_domain_attr), GFP_KERNEL);
614 *dattr = SD_ATTR_INIT;
615 update_domain_attr_tree(dattr, &top_cpuset);
617 cpumask_copy(doms[0], top_cpuset.cpus_allowed);
622 csa = kmalloc(number_of_cpusets * sizeof(cp), GFP_KERNEL);
628 cpuset_for_each_descendant_pre(cp, pos_cgrp, &top_cpuset) {
630 * Continue traversing beyond @cp iff @cp has some CPUs and
631 * isn't load balancing. The former is obvious. The
632 * latter: All child cpusets contain a subset of the
633 * parent's cpus, so just skip them, and then we call
634 * update_domain_attr_tree() to calc relax_domain_level of
635 * the corresponding sched domain.
637 if (!cpumask_empty(cp->cpus_allowed) &&
638 !is_sched_load_balance(cp))
641 if (is_sched_load_balance(cp))
644 /* skip @cp's subtree */
645 pos_cgrp = cgroup_rightmost_descendant(pos_cgrp);
649 for (i = 0; i < csn; i++)
654 /* Find the best partition (set of sched domains) */
655 for (i = 0; i < csn; i++) {
656 struct cpuset *a = csa[i];
659 for (j = 0; j < csn; j++) {
660 struct cpuset *b = csa[j];
663 if (apn != bpn && cpusets_overlap(a, b)) {
664 for (k = 0; k < csn; k++) {
665 struct cpuset *c = csa[k];
670 ndoms--; /* one less element */
677 * Now we know how many domains to create.
678 * Convert <csn, csa> to <ndoms, doms> and populate cpu masks.
680 doms = alloc_sched_domains(ndoms);
685 * The rest of the code, including the scheduler, can deal with
686 * dattr==NULL case. No need to abort if alloc fails.
688 dattr = kmalloc(ndoms * sizeof(struct sched_domain_attr), GFP_KERNEL);
690 for (nslot = 0, i = 0; i < csn; i++) {
691 struct cpuset *a = csa[i];
696 /* Skip completed partitions */
702 if (nslot == ndoms) {
703 static int warnings = 10;
706 "rebuild_sched_domains confused:"
707 " nslot %d, ndoms %d, csn %d, i %d,"
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->cpus_allowed);
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.cpus_allowed, 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 * cpuset_change_cpumask - make a task's cpus_allowed the same as its cpuset's
840 * @scan: struct cgroup_scanner containing the cgroup of the task
842 * Called by cgroup_scan_tasks() for each task in a cgroup whose
843 * cpus_allowed mask needs to be changed.
845 * We don't need to re-check for the cgroup/cpuset membership, since we're
846 * holding cpuset_mutex at this point.
848 static void cpuset_change_cpumask(struct task_struct *tsk,
849 struct cgroup_scanner *scan)
851 struct cpuset *cpus_cs;
853 cpus_cs = effective_cpumask_cpuset(cgroup_cs(scan->cg));
854 set_cpus_allowed_ptr(tsk, cpus_cs->cpus_allowed);
858 * update_tasks_cpumask - Update the cpumasks of tasks in the cpuset.
859 * @cs: the cpuset in which each task's cpus_allowed mask needs to be changed
860 * @heap: if NULL, defer allocating heap memory to cgroup_scan_tasks()
862 * Called with cpuset_mutex held
864 * The cgroup_scan_tasks() function will scan all the tasks in a cgroup,
865 * calling callback functions for each.
867 * No return value. It's guaranteed that cgroup_scan_tasks() always returns 0
870 static void update_tasks_cpumask(struct cpuset *cs, struct ptr_heap *heap)
872 struct cgroup_scanner scan;
874 scan.cg = cs->css.cgroup;
875 scan.test_task = NULL;
876 scan.process_task = cpuset_change_cpumask;
878 cgroup_scan_tasks(&scan);
882 * update_tasks_cpumask_hier - Update the cpumasks of tasks in the hierarchy.
883 * @root_cs: the root cpuset of the hierarchy
884 * @update_root: update root cpuset or not?
885 * @heap: the heap used by cgroup_scan_tasks()
887 * This will update cpumasks of tasks in @root_cs and all other empty cpusets
888 * which take on cpumask of @root_cs.
890 * Called with cpuset_mutex held
892 static void update_tasks_cpumask_hier(struct cpuset *root_cs,
893 bool update_root, struct ptr_heap *heap)
896 struct cgroup *pos_cgrp;
899 update_tasks_cpumask(root_cs, heap);
902 cpuset_for_each_descendant_pre(cp, pos_cgrp, root_cs) {
903 /* skip the whole subtree if @cp have some CPU */
904 if (!cpumask_empty(cp->cpus_allowed)) {
905 pos_cgrp = cgroup_rightmost_descendant(pos_cgrp);
908 if (!css_tryget(&cp->css))
912 update_tasks_cpumask(cp, heap);
921 * update_cpumask - update the cpus_allowed mask of a cpuset and all tasks in it
922 * @cs: the cpuset to consider
923 * @buf: buffer of cpu numbers written to this cpuset
925 static int update_cpumask(struct cpuset *cs, struct cpuset *trialcs,
928 struct ptr_heap heap;
930 int is_load_balanced;
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, cpu_active_mask))
953 /* Nothing to do if the cpus didn't change */
954 if (cpumask_equal(cs->cpus_allowed, trialcs->cpus_allowed))
957 retval = validate_change(cs, trialcs);
961 retval = heap_init(&heap, PAGE_SIZE, GFP_KERNEL, NULL);
965 is_load_balanced = is_sched_load_balance(trialcs);
967 mutex_lock(&callback_mutex);
968 cpumask_copy(cs->cpus_allowed, trialcs->cpus_allowed);
969 mutex_unlock(&callback_mutex);
971 update_tasks_cpumask_hier(cs, true, &heap);
975 if (is_load_balanced)
976 rebuild_sched_domains_locked();
983 * Migrate memory region from one set of nodes to another.
985 * Temporarilly set tasks mems_allowed to target nodes of migration,
986 * so that the migration code can allocate pages on these nodes.
988 * Call holding cpuset_mutex, so current's cpuset won't change
989 * during this call, as manage_mutex holds off any cpuset_attach()
990 * calls. Therefore we don't need to take task_lock around the
991 * call to guarantee_online_mems(), as we know no one is changing
994 * While the mm_struct we are migrating is typically from some
995 * other task, the task_struct mems_allowed that we are hacking
996 * is for our current task, which must allocate new pages for that
997 * migrating memory region.
1000 static void cpuset_migrate_mm(struct mm_struct *mm, const nodemask_t *from,
1001 const nodemask_t *to)
1003 struct task_struct *tsk = current;
1004 struct cpuset *mems_cs;
1006 tsk->mems_allowed = *to;
1008 do_migrate_pages(mm, from, to, MPOL_MF_MOVE_ALL);
1010 mems_cs = effective_nodemask_cpuset(task_cs(tsk));
1011 guarantee_online_mems(mems_cs, &tsk->mems_allowed);
1015 * cpuset_change_task_nodemask - change task's mems_allowed and mempolicy
1016 * @tsk: the task to change
1017 * @newmems: new nodes that the task will be set
1019 * In order to avoid seeing no nodes if the old and new nodes are disjoint,
1020 * we structure updates as setting all new allowed nodes, then clearing newly
1023 static void cpuset_change_task_nodemask(struct task_struct *tsk,
1024 nodemask_t *newmems)
1029 * Allow tasks that have access to memory reserves because they have
1030 * been OOM killed to get memory anywhere.
1032 if (unlikely(test_thread_flag(TIF_MEMDIE)))
1034 if (current->flags & PF_EXITING) /* Let dying task have memory */
1039 * Determine if a loop is necessary if another thread is doing
1040 * get_mems_allowed(). If at least one node remains unchanged and
1041 * tsk does not have a mempolicy, then an empty nodemask will not be
1042 * possible when mems_allowed is larger than a word.
1044 need_loop = task_has_mempolicy(tsk) ||
1045 !nodes_intersects(*newmems, tsk->mems_allowed);
1048 write_seqcount_begin(&tsk->mems_allowed_seq);
1050 nodes_or(tsk->mems_allowed, tsk->mems_allowed, *newmems);
1051 mpol_rebind_task(tsk, newmems, MPOL_REBIND_STEP1);
1053 mpol_rebind_task(tsk, newmems, MPOL_REBIND_STEP2);
1054 tsk->mems_allowed = *newmems;
1057 write_seqcount_end(&tsk->mems_allowed_seq);
1063 * Update task's mems_allowed and rebind its mempolicy and vmas' mempolicy
1064 * of it to cpuset's new mems_allowed, and migrate pages to new nodes if
1065 * memory_migrate flag is set. Called with cpuset_mutex held.
1067 static void cpuset_change_nodemask(struct task_struct *p,
1068 struct cgroup_scanner *scan)
1070 struct cpuset *cs = cgroup_cs(scan->cg);
1071 struct mm_struct *mm;
1073 nodemask_t *newmems = scan->data;
1075 cpuset_change_task_nodemask(p, newmems);
1077 mm = get_task_mm(p);
1081 migrate = is_memory_migrate(cs);
1083 mpol_rebind_mm(mm, &cs->mems_allowed);
1085 cpuset_migrate_mm(mm, &cs->old_mems_allowed, newmems);
1089 static void *cpuset_being_rebound;
1092 * update_tasks_nodemask - Update the nodemasks of tasks in the cpuset.
1093 * @cs: the cpuset in which each task's mems_allowed mask needs to be changed
1094 * @heap: if NULL, defer allocating heap memory to cgroup_scan_tasks()
1096 * Called with cpuset_mutex held
1097 * No return value. It's guaranteed that cgroup_scan_tasks() always returns 0
1100 static void update_tasks_nodemask(struct cpuset *cs, struct ptr_heap *heap)
1102 static nodemask_t newmems; /* protected by cpuset_mutex */
1103 struct cgroup_scanner scan;
1104 struct cpuset *mems_cs = effective_nodemask_cpuset(cs);
1106 cpuset_being_rebound = cs; /* causes mpol_dup() rebind */
1108 guarantee_online_mems(mems_cs, &newmems);
1110 scan.cg = cs->css.cgroup;
1111 scan.test_task = NULL;
1112 scan.process_task = cpuset_change_nodemask;
1114 scan.data = &newmems;
1117 * The mpol_rebind_mm() call takes mmap_sem, which we couldn't
1118 * take while holding tasklist_lock. Forks can happen - the
1119 * mpol_dup() cpuset_being_rebound check will catch such forks,
1120 * and rebind their vma mempolicies too. Because we still hold
1121 * the global cpuset_mutex, we know that no other rebind effort
1122 * will be contending for the global variable cpuset_being_rebound.
1123 * It's ok if we rebind the same mm twice; mpol_rebind_mm()
1124 * is idempotent. Also migrate pages in each mm to new nodes.
1126 cgroup_scan_tasks(&scan);
1129 * All the tasks' nodemasks have been updated, update
1130 * cs->old_mems_allowed.
1132 cs->old_mems_allowed = newmems;
1134 /* We're done rebinding vmas to this cpuset's new mems_allowed. */
1135 cpuset_being_rebound = NULL;
1139 * update_tasks_nodemask_hier - Update the nodemasks of tasks in the hierarchy.
1140 * @cs: the root cpuset of the hierarchy
1141 * @update_root: update the root cpuset or not?
1142 * @heap: the heap used by cgroup_scan_tasks()
1144 * This will update nodemasks of tasks in @root_cs and all other empty cpusets
1145 * which take on nodemask of @root_cs.
1147 * Called with cpuset_mutex held
1149 static void update_tasks_nodemask_hier(struct cpuset *root_cs,
1150 bool update_root, struct ptr_heap *heap)
1153 struct cgroup *pos_cgrp;
1156 update_tasks_nodemask(root_cs, heap);
1159 cpuset_for_each_descendant_pre(cp, pos_cgrp, root_cs) {
1160 /* skip the whole subtree if @cp have some CPU */
1161 if (!nodes_empty(cp->mems_allowed)) {
1162 pos_cgrp = cgroup_rightmost_descendant(pos_cgrp);
1165 if (!css_tryget(&cp->css))
1169 update_tasks_nodemask(cp, heap);
1178 * Handle user request to change the 'mems' memory placement
1179 * of a cpuset. Needs to validate the request, update the
1180 * cpusets mems_allowed, and for each task in the cpuset,
1181 * update mems_allowed and rebind task's mempolicy and any vma
1182 * mempolicies and if the cpuset is marked 'memory_migrate',
1183 * migrate the tasks pages to the new memory.
1185 * Call with cpuset_mutex held. May take callback_mutex during call.
1186 * Will take tasklist_lock, scan tasklist for tasks in cpuset cs,
1187 * lock each such tasks mm->mmap_sem, scan its vma's and rebind
1188 * their mempolicies to the cpusets new mems_allowed.
1190 static int update_nodemask(struct cpuset *cs, struct cpuset *trialcs,
1194 struct ptr_heap heap;
1197 * top_cpuset.mems_allowed tracks node_stats[N_MEMORY];
1200 if (cs == &top_cpuset) {
1206 * An empty mems_allowed is ok iff there are no tasks in the cpuset.
1207 * Since nodelist_parse() fails on an empty mask, we special case
1208 * that parsing. The validate_change() call ensures that cpusets
1209 * with tasks have memory.
1212 nodes_clear(trialcs->mems_allowed);
1214 retval = nodelist_parse(buf, trialcs->mems_allowed);
1218 if (!nodes_subset(trialcs->mems_allowed,
1219 node_states[N_MEMORY])) {
1225 if (nodes_equal(cs->mems_allowed, trialcs->mems_allowed)) {
1226 retval = 0; /* Too easy - nothing to do */
1229 retval = validate_change(cs, trialcs);
1233 retval = heap_init(&heap, PAGE_SIZE, GFP_KERNEL, NULL);
1237 mutex_lock(&callback_mutex);
1238 cs->mems_allowed = trialcs->mems_allowed;
1239 mutex_unlock(&callback_mutex);
1241 update_tasks_nodemask_hier(cs, true, &heap);
1248 int current_cpuset_is_being_rebound(void)
1250 return task_cs(current) == cpuset_being_rebound;
1253 static int update_relax_domain_level(struct cpuset *cs, s64 val)
1256 if (val < -1 || val >= sched_domain_level_max)
1260 if (val != cs->relax_domain_level) {
1261 cs->relax_domain_level = val;
1262 if (!cpumask_empty(cs->cpus_allowed) &&
1263 is_sched_load_balance(cs))
1264 rebuild_sched_domains_locked();
1271 * cpuset_change_flag - make a task's spread flags the same as its cpuset's
1272 * @tsk: task to be updated
1273 * @scan: struct cgroup_scanner containing the cgroup of the task
1275 * Called by cgroup_scan_tasks() for each task in a cgroup.
1277 * We don't need to re-check for the cgroup/cpuset membership, since we're
1278 * holding cpuset_mutex at this point.
1280 static void cpuset_change_flag(struct task_struct *tsk,
1281 struct cgroup_scanner *scan)
1283 cpuset_update_task_spread_flag(cgroup_cs(scan->cg), tsk);
1287 * update_tasks_flags - update the spread flags of tasks in the cpuset.
1288 * @cs: the cpuset in which each task's spread flags needs to be changed
1289 * @heap: if NULL, defer allocating heap memory to cgroup_scan_tasks()
1291 * Called with cpuset_mutex held
1293 * The cgroup_scan_tasks() function will scan all the tasks in a cgroup,
1294 * calling callback functions for each.
1296 * No return value. It's guaranteed that cgroup_scan_tasks() always returns 0
1299 static void update_tasks_flags(struct cpuset *cs, struct ptr_heap *heap)
1301 struct cgroup_scanner scan;
1303 scan.cg = cs->css.cgroup;
1304 scan.test_task = NULL;
1305 scan.process_task = cpuset_change_flag;
1307 cgroup_scan_tasks(&scan);
1311 * update_flag - read a 0 or a 1 in a file and update associated flag
1312 * bit: the bit to update (see cpuset_flagbits_t)
1313 * cs: the cpuset to update
1314 * turning_on: whether the flag is being set or cleared
1316 * Call with cpuset_mutex held.
1319 static int update_flag(cpuset_flagbits_t bit, struct cpuset *cs,
1322 struct cpuset *trialcs;
1323 int balance_flag_changed;
1324 int spread_flag_changed;
1325 struct ptr_heap heap;
1328 trialcs = alloc_trial_cpuset(cs);
1333 set_bit(bit, &trialcs->flags);
1335 clear_bit(bit, &trialcs->flags);
1337 err = validate_change(cs, trialcs);
1341 err = heap_init(&heap, PAGE_SIZE, GFP_KERNEL, NULL);
1345 balance_flag_changed = (is_sched_load_balance(cs) !=
1346 is_sched_load_balance(trialcs));
1348 spread_flag_changed = ((is_spread_slab(cs) != is_spread_slab(trialcs))
1349 || (is_spread_page(cs) != is_spread_page(trialcs)));
1351 mutex_lock(&callback_mutex);
1352 cs->flags = trialcs->flags;
1353 mutex_unlock(&callback_mutex);
1355 if (!cpumask_empty(trialcs->cpus_allowed) && balance_flag_changed)
1356 rebuild_sched_domains_locked();
1358 if (spread_flag_changed)
1359 update_tasks_flags(cs, &heap);
1362 free_trial_cpuset(trialcs);
1367 * Frequency meter - How fast is some event occurring?
1369 * These routines manage a digitally filtered, constant time based,
1370 * event frequency meter. There are four routines:
1371 * fmeter_init() - initialize a frequency meter.
1372 * fmeter_markevent() - called each time the event happens.
1373 * fmeter_getrate() - returns the recent rate of such events.
1374 * fmeter_update() - internal routine used to update fmeter.
1376 * A common data structure is passed to each of these routines,
1377 * which is used to keep track of the state required to manage the
1378 * frequency meter and its digital filter.
1380 * The filter works on the number of events marked per unit time.
1381 * The filter is single-pole low-pass recursive (IIR). The time unit
1382 * is 1 second. Arithmetic is done using 32-bit integers scaled to
1383 * simulate 3 decimal digits of precision (multiplied by 1000).
1385 * With an FM_COEF of 933, and a time base of 1 second, the filter
1386 * has a half-life of 10 seconds, meaning that if the events quit
1387 * happening, then the rate returned from the fmeter_getrate()
1388 * will be cut in half each 10 seconds, until it converges to zero.
1390 * It is not worth doing a real infinitely recursive filter. If more
1391 * than FM_MAXTICKS ticks have elapsed since the last filter event,
1392 * just compute FM_MAXTICKS ticks worth, by which point the level
1395 * Limit the count of unprocessed events to FM_MAXCNT, so as to avoid
1396 * arithmetic overflow in the fmeter_update() routine.
1398 * Given the simple 32 bit integer arithmetic used, this meter works
1399 * best for reporting rates between one per millisecond (msec) and
1400 * one per 32 (approx) seconds. At constant rates faster than one
1401 * per msec it maxes out at values just under 1,000,000. At constant
1402 * rates between one per msec, and one per second it will stabilize
1403 * to a value N*1000, where N is the rate of events per second.
1404 * At constant rates between one per second and one per 32 seconds,
1405 * it will be choppy, moving up on the seconds that have an event,
1406 * and then decaying until the next event. At rates slower than
1407 * about one in 32 seconds, it decays all the way back to zero between
1411 #define FM_COEF 933 /* coefficient for half-life of 10 secs */
1412 #define FM_MAXTICKS ((time_t)99) /* useless computing more ticks than this */
1413 #define FM_MAXCNT 1000000 /* limit cnt to avoid overflow */
1414 #define FM_SCALE 1000 /* faux fixed point scale */
1416 /* Initialize a frequency meter */
1417 static void fmeter_init(struct fmeter *fmp)
1422 spin_lock_init(&fmp->lock);
1425 /* Internal meter update - process cnt events and update value */
1426 static void fmeter_update(struct fmeter *fmp)
1428 time_t now = get_seconds();
1429 time_t ticks = now - fmp->time;
1434 ticks = min(FM_MAXTICKS, ticks);
1436 fmp->val = (FM_COEF * fmp->val) / FM_SCALE;
1439 fmp->val += ((FM_SCALE - FM_COEF) * fmp->cnt) / FM_SCALE;
1443 /* Process any previous ticks, then bump cnt by one (times scale). */
1444 static void fmeter_markevent(struct fmeter *fmp)
1446 spin_lock(&fmp->lock);
1448 fmp->cnt = min(FM_MAXCNT, fmp->cnt + FM_SCALE);
1449 spin_unlock(&fmp->lock);
1452 /* Process any previous ticks, then return current value. */
1453 static int fmeter_getrate(struct fmeter *fmp)
1457 spin_lock(&fmp->lock);
1460 spin_unlock(&fmp->lock);
1464 /* Called by cgroups to determine if a cpuset is usable; cpuset_mutex held */
1465 static int cpuset_can_attach(struct cgroup *cgrp, struct cgroup_taskset *tset)
1467 struct cpuset *cs = cgroup_cs(cgrp);
1468 struct task_struct *task;
1471 mutex_lock(&cpuset_mutex);
1474 * We allow to move tasks into an empty cpuset if sane_behavior
1478 if (!cgroup_sane_behavior(cgrp) &&
1479 (cpumask_empty(cs->cpus_allowed) || nodes_empty(cs->mems_allowed)))
1482 cgroup_taskset_for_each(task, cgrp, tset) {
1484 * Kthreads which disallow setaffinity shouldn't be moved
1485 * to a new cpuset; we don't want to change their cpu
1486 * affinity and isolating such threads by their set of
1487 * allowed nodes is unnecessary. Thus, cpusets are not
1488 * applicable for such threads. This prevents checking for
1489 * success of set_cpus_allowed_ptr() on all attached tasks
1490 * before cpus_allowed may be changed.
1493 if (task->flags & PF_NO_SETAFFINITY)
1495 ret = security_task_setscheduler(task);
1501 * Mark attach is in progress. This makes validate_change() fail
1502 * changes which zero cpus/mems_allowed.
1504 cs->attach_in_progress++;
1507 mutex_unlock(&cpuset_mutex);
1511 static void cpuset_cancel_attach(struct cgroup *cgrp,
1512 struct cgroup_taskset *tset)
1514 mutex_lock(&cpuset_mutex);
1515 cgroup_cs(cgrp)->attach_in_progress--;
1516 mutex_unlock(&cpuset_mutex);
1520 * Protected by cpuset_mutex. cpus_attach is used only by cpuset_attach()
1521 * but we can't allocate it dynamically there. Define it global and
1522 * allocate from cpuset_init().
1524 static cpumask_var_t cpus_attach;
1526 static void cpuset_attach(struct cgroup *cgrp, struct cgroup_taskset *tset)
1528 /* static buf protected by cpuset_mutex */
1529 static nodemask_t cpuset_attach_nodemask_to;
1530 struct mm_struct *mm;
1531 struct task_struct *task;
1532 struct task_struct *leader = cgroup_taskset_first(tset);
1533 struct cgroup *oldcgrp = cgroup_taskset_cur_cgroup(tset);
1534 struct cpuset *cs = cgroup_cs(cgrp);
1535 struct cpuset *oldcs = cgroup_cs(oldcgrp);
1536 struct cpuset *cpus_cs = effective_cpumask_cpuset(cs);
1537 struct cpuset *mems_cs = effective_nodemask_cpuset(cs);
1539 mutex_lock(&cpuset_mutex);
1541 /* prepare for attach */
1542 if (cs == &top_cpuset)
1543 cpumask_copy(cpus_attach, cpu_possible_mask);
1545 guarantee_online_cpus(cpus_cs, cpus_attach);
1547 guarantee_online_mems(mems_cs, &cpuset_attach_nodemask_to);
1549 cgroup_taskset_for_each(task, cgrp, tset) {
1551 * can_attach beforehand should guarantee that this doesn't
1552 * fail. TODO: have a better way to handle failure here
1554 WARN_ON_ONCE(set_cpus_allowed_ptr(task, cpus_attach));
1556 cpuset_change_task_nodemask(task, &cpuset_attach_nodemask_to);
1557 cpuset_update_task_spread_flag(cs, task);
1561 * Change mm, possibly for multiple threads in a threadgroup. This is
1562 * expensive and may sleep.
1564 cpuset_attach_nodemask_to = cs->mems_allowed;
1565 mm = get_task_mm(leader);
1567 struct cpuset *mems_oldcs = effective_nodemask_cpuset(oldcs);
1569 mpol_rebind_mm(mm, &cpuset_attach_nodemask_to);
1572 * old_mems_allowed is the same with mems_allowed here, except
1573 * if this task is being moved automatically due to hotplug.
1574 * In that case @mems_allowed has been updated and is empty,
1575 * so @old_mems_allowed is the right nodesets that we migrate
1578 if (is_memory_migrate(cs)) {
1579 cpuset_migrate_mm(mm, &mems_oldcs->old_mems_allowed,
1580 &cpuset_attach_nodemask_to);
1585 cs->old_mems_allowed = cpuset_attach_nodemask_to;
1587 cs->attach_in_progress--;
1588 if (!cs->attach_in_progress)
1589 wake_up(&cpuset_attach_wq);
1591 mutex_unlock(&cpuset_mutex);
1594 /* The various types of files and directories in a cpuset file system */
1597 FILE_MEMORY_MIGRATE,
1603 FILE_SCHED_LOAD_BALANCE,
1604 FILE_SCHED_RELAX_DOMAIN_LEVEL,
1605 FILE_MEMORY_PRESSURE_ENABLED,
1606 FILE_MEMORY_PRESSURE,
1609 } cpuset_filetype_t;
1611 static int cpuset_write_u64(struct cgroup *cgrp, struct cftype *cft, u64 val)
1613 struct cpuset *cs = cgroup_cs(cgrp);
1614 cpuset_filetype_t type = cft->private;
1617 mutex_lock(&cpuset_mutex);
1618 if (!is_cpuset_online(cs)) {
1624 case FILE_CPU_EXCLUSIVE:
1625 retval = update_flag(CS_CPU_EXCLUSIVE, cs, val);
1627 case FILE_MEM_EXCLUSIVE:
1628 retval = update_flag(CS_MEM_EXCLUSIVE, cs, val);
1630 case FILE_MEM_HARDWALL:
1631 retval = update_flag(CS_MEM_HARDWALL, cs, val);
1633 case FILE_SCHED_LOAD_BALANCE:
1634 retval = update_flag(CS_SCHED_LOAD_BALANCE, cs, val);
1636 case FILE_MEMORY_MIGRATE:
1637 retval = update_flag(CS_MEMORY_MIGRATE, cs, val);
1639 case FILE_MEMORY_PRESSURE_ENABLED:
1640 cpuset_memory_pressure_enabled = !!val;
1642 case FILE_MEMORY_PRESSURE:
1645 case FILE_SPREAD_PAGE:
1646 retval = update_flag(CS_SPREAD_PAGE, cs, val);
1648 case FILE_SPREAD_SLAB:
1649 retval = update_flag(CS_SPREAD_SLAB, cs, val);
1656 mutex_unlock(&cpuset_mutex);
1660 static int cpuset_write_s64(struct cgroup *cgrp, struct cftype *cft, s64 val)
1662 struct cpuset *cs = cgroup_cs(cgrp);
1663 cpuset_filetype_t type = cft->private;
1664 int retval = -ENODEV;
1666 mutex_lock(&cpuset_mutex);
1667 if (!is_cpuset_online(cs))
1671 case FILE_SCHED_RELAX_DOMAIN_LEVEL:
1672 retval = update_relax_domain_level(cs, val);
1679 mutex_unlock(&cpuset_mutex);
1684 * Common handling for a write to a "cpus" or "mems" file.
1686 static int cpuset_write_resmask(struct cgroup *cgrp, struct cftype *cft,
1689 struct cpuset *cs = cgroup_cs(cgrp);
1690 struct cpuset *trialcs;
1691 int retval = -ENODEV;
1694 * CPU or memory hotunplug may leave @cs w/o any execution
1695 * resources, in which case the hotplug code asynchronously updates
1696 * configuration and transfers all tasks to the nearest ancestor
1697 * which can execute.
1699 * As writes to "cpus" or "mems" may restore @cs's execution
1700 * resources, wait for the previously scheduled operations before
1701 * proceeding, so that we don't end up keep removing tasks added
1702 * after execution capability is restored.
1704 flush_work(&cpuset_hotplug_work);
1706 mutex_lock(&cpuset_mutex);
1707 if (!is_cpuset_online(cs))
1710 trialcs = alloc_trial_cpuset(cs);
1716 switch (cft->private) {
1718 retval = update_cpumask(cs, trialcs, buf);
1721 retval = update_nodemask(cs, trialcs, buf);
1728 free_trial_cpuset(trialcs);
1730 mutex_unlock(&cpuset_mutex);
1735 * These ascii lists should be read in a single call, by using a user
1736 * buffer large enough to hold the entire map. If read in smaller
1737 * chunks, there is no guarantee of atomicity. Since the display format
1738 * used, list of ranges of sequential numbers, is variable length,
1739 * and since these maps can change value dynamically, one could read
1740 * gibberish by doing partial reads while a list was changing.
1741 * A single large read to a buffer that crosses a page boundary is
1742 * ok, because the result being copied to user land is not recomputed
1743 * across a page fault.
1746 static size_t cpuset_sprintf_cpulist(char *page, struct cpuset *cs)
1750 mutex_lock(&callback_mutex);
1751 count = cpulist_scnprintf(page, PAGE_SIZE, cs->cpus_allowed);
1752 mutex_unlock(&callback_mutex);
1757 static size_t cpuset_sprintf_memlist(char *page, struct cpuset *cs)
1761 mutex_lock(&callback_mutex);
1762 count = nodelist_scnprintf(page, PAGE_SIZE, cs->mems_allowed);
1763 mutex_unlock(&callback_mutex);
1768 static ssize_t cpuset_common_file_read(struct cgroup *cgrp,
1772 size_t nbytes, loff_t *ppos)
1774 struct cpuset *cs = cgroup_cs(cgrp);
1775 cpuset_filetype_t type = cft->private;
1780 if (!(page = (char *)__get_free_page(GFP_TEMPORARY)))
1787 s += cpuset_sprintf_cpulist(s, cs);
1790 s += cpuset_sprintf_memlist(s, cs);
1798 retval = simple_read_from_buffer(buf, nbytes, ppos, page, s - page);
1800 free_page((unsigned long)page);
1804 static u64 cpuset_read_u64(struct cgroup *cgrp, struct cftype *cft)
1806 struct cpuset *cs = cgroup_cs(cgrp);
1807 cpuset_filetype_t type = cft->private;
1809 case FILE_CPU_EXCLUSIVE:
1810 return is_cpu_exclusive(cs);
1811 case FILE_MEM_EXCLUSIVE:
1812 return is_mem_exclusive(cs);
1813 case FILE_MEM_HARDWALL:
1814 return is_mem_hardwall(cs);
1815 case FILE_SCHED_LOAD_BALANCE:
1816 return is_sched_load_balance(cs);
1817 case FILE_MEMORY_MIGRATE:
1818 return is_memory_migrate(cs);
1819 case FILE_MEMORY_PRESSURE_ENABLED:
1820 return cpuset_memory_pressure_enabled;
1821 case FILE_MEMORY_PRESSURE:
1822 return fmeter_getrate(&cs->fmeter);
1823 case FILE_SPREAD_PAGE:
1824 return is_spread_page(cs);
1825 case FILE_SPREAD_SLAB:
1826 return is_spread_slab(cs);
1831 /* Unreachable but makes gcc happy */
1835 static s64 cpuset_read_s64(struct cgroup *cgrp, struct cftype *cft)
1837 struct cpuset *cs = cgroup_cs(cgrp);
1838 cpuset_filetype_t type = cft->private;
1840 case FILE_SCHED_RELAX_DOMAIN_LEVEL:
1841 return cs->relax_domain_level;
1846 /* Unrechable but makes gcc happy */
1852 * for the common functions, 'private' gives the type of file
1855 static struct cftype files[] = {
1858 .read = cpuset_common_file_read,
1859 .write_string = cpuset_write_resmask,
1860 .max_write_len = (100U + 6 * NR_CPUS),
1861 .private = FILE_CPULIST,
1866 .read = cpuset_common_file_read,
1867 .write_string = cpuset_write_resmask,
1868 .max_write_len = (100U + 6 * MAX_NUMNODES),
1869 .private = FILE_MEMLIST,
1873 .name = "cpu_exclusive",
1874 .read_u64 = cpuset_read_u64,
1875 .write_u64 = cpuset_write_u64,
1876 .private = FILE_CPU_EXCLUSIVE,
1880 .name = "mem_exclusive",
1881 .read_u64 = cpuset_read_u64,
1882 .write_u64 = cpuset_write_u64,
1883 .private = FILE_MEM_EXCLUSIVE,
1887 .name = "mem_hardwall",
1888 .read_u64 = cpuset_read_u64,
1889 .write_u64 = cpuset_write_u64,
1890 .private = FILE_MEM_HARDWALL,
1894 .name = "sched_load_balance",
1895 .read_u64 = cpuset_read_u64,
1896 .write_u64 = cpuset_write_u64,
1897 .private = FILE_SCHED_LOAD_BALANCE,
1901 .name = "sched_relax_domain_level",
1902 .read_s64 = cpuset_read_s64,
1903 .write_s64 = cpuset_write_s64,
1904 .private = FILE_SCHED_RELAX_DOMAIN_LEVEL,
1908 .name = "memory_migrate",
1909 .read_u64 = cpuset_read_u64,
1910 .write_u64 = cpuset_write_u64,
1911 .private = FILE_MEMORY_MIGRATE,
1915 .name = "memory_pressure",
1916 .read_u64 = cpuset_read_u64,
1917 .write_u64 = cpuset_write_u64,
1918 .private = FILE_MEMORY_PRESSURE,
1923 .name = "memory_spread_page",
1924 .read_u64 = cpuset_read_u64,
1925 .write_u64 = cpuset_write_u64,
1926 .private = FILE_SPREAD_PAGE,
1930 .name = "memory_spread_slab",
1931 .read_u64 = cpuset_read_u64,
1932 .write_u64 = cpuset_write_u64,
1933 .private = FILE_SPREAD_SLAB,
1937 .name = "memory_pressure_enabled",
1938 .flags = CFTYPE_ONLY_ON_ROOT,
1939 .read_u64 = cpuset_read_u64,
1940 .write_u64 = cpuset_write_u64,
1941 .private = FILE_MEMORY_PRESSURE_ENABLED,
1948 * cpuset_css_alloc - allocate a cpuset css
1949 * cgrp: control group that the new cpuset will be part of
1952 static struct cgroup_subsys_state *cpuset_css_alloc(struct cgroup *cgrp)
1957 return &top_cpuset.css;
1959 cs = kzalloc(sizeof(*cs), GFP_KERNEL);
1961 return ERR_PTR(-ENOMEM);
1962 if (!alloc_cpumask_var(&cs->cpus_allowed, GFP_KERNEL)) {
1964 return ERR_PTR(-ENOMEM);
1967 set_bit(CS_SCHED_LOAD_BALANCE, &cs->flags);
1968 cpumask_clear(cs->cpus_allowed);
1969 nodes_clear(cs->mems_allowed);
1970 fmeter_init(&cs->fmeter);
1971 cs->relax_domain_level = -1;
1976 static int cpuset_css_online(struct cgroup *cgrp)
1978 struct cpuset *cs = cgroup_cs(cgrp);
1979 struct cpuset *parent = parent_cs(cs);
1980 struct cpuset *tmp_cs;
1981 struct cgroup *pos_cg;
1986 mutex_lock(&cpuset_mutex);
1988 set_bit(CS_ONLINE, &cs->flags);
1989 if (is_spread_page(parent))
1990 set_bit(CS_SPREAD_PAGE, &cs->flags);
1991 if (is_spread_slab(parent))
1992 set_bit(CS_SPREAD_SLAB, &cs->flags);
1994 number_of_cpusets++;
1996 if (!test_bit(CGRP_CPUSET_CLONE_CHILDREN, &cgrp->flags))
2000 * Clone @parent's configuration if CGRP_CPUSET_CLONE_CHILDREN is
2001 * set. This flag handling is implemented in cgroup core for
2002 * histrical reasons - the flag may be specified during mount.
2004 * Currently, if any sibling cpusets have exclusive cpus or mem, we
2005 * refuse to clone the configuration - thereby refusing the task to
2006 * be entered, and as a result refusing the sys_unshare() or
2007 * clone() which initiated it. If this becomes a problem for some
2008 * users who wish to allow that scenario, then this could be
2009 * changed to grant parent->cpus_allowed-sibling_cpus_exclusive
2010 * (and likewise for mems) to the new cgroup.
2013 cpuset_for_each_child(tmp_cs, pos_cg, parent) {
2014 if (is_mem_exclusive(tmp_cs) || is_cpu_exclusive(tmp_cs)) {
2021 mutex_lock(&callback_mutex);
2022 cs->mems_allowed = parent->mems_allowed;
2023 cpumask_copy(cs->cpus_allowed, parent->cpus_allowed);
2024 mutex_unlock(&callback_mutex);
2026 mutex_unlock(&cpuset_mutex);
2030 static void cpuset_css_offline(struct cgroup *cgrp)
2032 struct cpuset *cs = cgroup_cs(cgrp);
2034 mutex_lock(&cpuset_mutex);
2036 if (is_sched_load_balance(cs))
2037 update_flag(CS_SCHED_LOAD_BALANCE, cs, 0);
2039 number_of_cpusets--;
2040 clear_bit(CS_ONLINE, &cs->flags);
2042 mutex_unlock(&cpuset_mutex);
2046 * If the cpuset being removed has its flag 'sched_load_balance'
2047 * enabled, then simulate turning sched_load_balance off, which
2048 * will call rebuild_sched_domains_locked().
2051 static void cpuset_css_free(struct cgroup *cgrp)
2053 struct cpuset *cs = cgroup_cs(cgrp);
2055 free_cpumask_var(cs->cpus_allowed);
2059 struct cgroup_subsys cpuset_subsys = {
2061 .css_alloc = cpuset_css_alloc,
2062 .css_online = cpuset_css_online,
2063 .css_offline = cpuset_css_offline,
2064 .css_free = cpuset_css_free,
2065 .can_attach = cpuset_can_attach,
2066 .cancel_attach = cpuset_cancel_attach,
2067 .attach = cpuset_attach,
2068 .subsys_id = cpuset_subsys_id,
2069 .base_cftypes = files,
2074 * cpuset_init - initialize cpusets at system boot
2076 * Description: Initialize top_cpuset and the cpuset internal file system,
2079 int __init cpuset_init(void)
2083 if (!alloc_cpumask_var(&top_cpuset.cpus_allowed, GFP_KERNEL))
2086 cpumask_setall(top_cpuset.cpus_allowed);
2087 nodes_setall(top_cpuset.mems_allowed);
2089 fmeter_init(&top_cpuset.fmeter);
2090 set_bit(CS_SCHED_LOAD_BALANCE, &top_cpuset.flags);
2091 top_cpuset.relax_domain_level = -1;
2093 err = register_filesystem(&cpuset_fs_type);
2097 if (!alloc_cpumask_var(&cpus_attach, GFP_KERNEL))
2100 number_of_cpusets = 1;
2105 * If CPU and/or memory hotplug handlers, below, unplug any CPUs
2106 * or memory nodes, we need to walk over the cpuset hierarchy,
2107 * removing that CPU or node from all cpusets. If this removes the
2108 * last CPU or node from a cpuset, then move the tasks in the empty
2109 * cpuset to its next-highest non-empty parent.
2111 static void remove_tasks_in_empty_cpuset(struct cpuset *cs)
2113 struct cpuset *parent;
2116 * Find its next-highest non-empty parent, (top cpuset
2117 * has online cpus, so can't be empty).
2119 parent = parent_cs(cs);
2120 while (cpumask_empty(parent->cpus_allowed) ||
2121 nodes_empty(parent->mems_allowed))
2122 parent = parent_cs(parent);
2124 if (cgroup_transfer_tasks(parent->css.cgroup, cs->css.cgroup)) {
2126 printk(KERN_ERR "cpuset: failed to transfer tasks out of empty cpuset %s\n",
2127 cgroup_name(cs->css.cgroup));
2133 * cpuset_hotplug_update_tasks - update tasks in a cpuset for hotunplug
2134 * @cs: cpuset in interest
2136 * Compare @cs's cpu and mem masks against top_cpuset and if some have gone
2137 * offline, update @cs accordingly. If @cs ends up with no CPU or memory,
2138 * all its tasks are moved to the nearest ancestor with both resources.
2140 static void cpuset_hotplug_update_tasks(struct cpuset *cs)
2142 static cpumask_t off_cpus;
2143 static nodemask_t off_mems;
2145 bool sane = cgroup_sane_behavior(cs->css.cgroup);
2148 wait_event(cpuset_attach_wq, cs->attach_in_progress == 0);
2150 mutex_lock(&cpuset_mutex);
2153 * We have raced with task attaching. We wait until attaching
2154 * is finished, so we won't attach a task to an empty cpuset.
2156 if (cs->attach_in_progress) {
2157 mutex_unlock(&cpuset_mutex);
2161 cpumask_andnot(&off_cpus, cs->cpus_allowed, top_cpuset.cpus_allowed);
2162 nodes_andnot(off_mems, cs->mems_allowed, top_cpuset.mems_allowed);
2164 mutex_lock(&callback_mutex);
2165 cpumask_andnot(cs->cpus_allowed, cs->cpus_allowed, &off_cpus);
2166 mutex_unlock(&callback_mutex);
2169 * If sane_behavior flag is set, we need to update tasks' cpumask
2170 * for empty cpuset to take on ancestor's cpumask. Otherwise, don't
2171 * call update_tasks_cpumask() if the cpuset becomes empty, as
2172 * the tasks in it will be migrated to an ancestor.
2174 if ((sane && cpumask_empty(cs->cpus_allowed)) ||
2175 (!cpumask_empty(&off_cpus) && !cpumask_empty(cs->cpus_allowed)))
2176 update_tasks_cpumask(cs, NULL);
2178 mutex_lock(&callback_mutex);
2179 nodes_andnot(cs->mems_allowed, cs->mems_allowed, off_mems);
2180 mutex_unlock(&callback_mutex);
2183 * If sane_behavior flag is set, we need to update tasks' nodemask
2184 * for empty cpuset to take on ancestor's nodemask. Otherwise, don't
2185 * call update_tasks_nodemask() if the cpuset becomes empty, as
2186 * the tasks in it will be migratd to an ancestor.
2188 if ((sane && nodes_empty(cs->mems_allowed)) ||
2189 (!nodes_empty(off_mems) && !nodes_empty(cs->mems_allowed)))
2190 update_tasks_nodemask(cs, NULL);
2192 is_empty = cpumask_empty(cs->cpus_allowed) ||
2193 nodes_empty(cs->mems_allowed);
2195 mutex_unlock(&cpuset_mutex);
2198 * If sane_behavior flag is set, we'll keep tasks in empty cpusets.
2200 * Otherwise move tasks to the nearest ancestor with execution
2201 * resources. This is full cgroup operation which will
2202 * also call back into cpuset. Should be done outside any lock.
2204 if (!sane && is_empty)
2205 remove_tasks_in_empty_cpuset(cs);
2209 * cpuset_hotplug_workfn - handle CPU/memory hotunplug for a cpuset
2211 * This function is called after either CPU or memory configuration has
2212 * changed and updates cpuset accordingly. The top_cpuset is always
2213 * synchronized to cpu_active_mask and N_MEMORY, which is necessary in
2214 * order to make cpusets transparent (of no affect) on systems that are
2215 * actively using CPU hotplug but making no active use of cpusets.
2217 * Non-root cpusets are only affected by offlining. If any CPUs or memory
2218 * nodes have been taken down, cpuset_hotplug_update_tasks() is invoked on
2221 * Note that CPU offlining during suspend is ignored. We don't modify
2222 * cpusets across suspend/resume cycles at all.
2224 static void cpuset_hotplug_workfn(struct work_struct *work)
2226 static cpumask_t new_cpus;
2227 static nodemask_t new_mems;
2228 bool cpus_updated, mems_updated;
2230 mutex_lock(&cpuset_mutex);
2232 /* fetch the available cpus/mems and find out which changed how */
2233 cpumask_copy(&new_cpus, cpu_active_mask);
2234 new_mems = node_states[N_MEMORY];
2236 cpus_updated = !cpumask_equal(top_cpuset.cpus_allowed, &new_cpus);
2237 mems_updated = !nodes_equal(top_cpuset.mems_allowed, new_mems);
2239 /* synchronize cpus_allowed to cpu_active_mask */
2241 mutex_lock(&callback_mutex);
2242 cpumask_copy(top_cpuset.cpus_allowed, &new_cpus);
2243 mutex_unlock(&callback_mutex);
2244 /* we don't mess with cpumasks of tasks in top_cpuset */
2247 /* synchronize mems_allowed to N_MEMORY */
2249 mutex_lock(&callback_mutex);
2250 top_cpuset.mems_allowed = new_mems;
2251 mutex_unlock(&callback_mutex);
2252 update_tasks_nodemask(&top_cpuset, NULL);
2255 mutex_unlock(&cpuset_mutex);
2257 /* if cpus or mems changed, we need to propagate to descendants */
2258 if (cpus_updated || mems_updated) {
2260 struct cgroup *pos_cgrp;
2263 cpuset_for_each_descendant_pre(cs, pos_cgrp, &top_cpuset) {
2264 if (!css_tryget(&cs->css))
2268 cpuset_hotplug_update_tasks(cs);
2276 /* rebuild sched domains if cpus_allowed has changed */
2278 rebuild_sched_domains();
2281 void cpuset_update_active_cpus(bool cpu_online)
2284 * We're inside cpu hotplug critical region which usually nests
2285 * inside cgroup synchronization. Bounce actual hotplug processing
2286 * to a work item to avoid reverse locking order.
2288 * We still need to do partition_sched_domains() synchronously;
2289 * otherwise, the scheduler will get confused and put tasks to the
2290 * dead CPU. Fall back to the default single domain.
2291 * cpuset_hotplug_workfn() will rebuild it as necessary.
2293 partition_sched_domains(1, NULL, NULL);
2294 schedule_work(&cpuset_hotplug_work);
2298 * Keep top_cpuset.mems_allowed tracking node_states[N_MEMORY].
2299 * Call this routine anytime after node_states[N_MEMORY] changes.
2300 * See cpuset_update_active_cpus() for CPU hotplug handling.
2302 static int cpuset_track_online_nodes(struct notifier_block *self,
2303 unsigned long action, void *arg)
2305 schedule_work(&cpuset_hotplug_work);
2309 static struct notifier_block cpuset_track_online_nodes_nb = {
2310 .notifier_call = cpuset_track_online_nodes,
2311 .priority = 10, /* ??! */
2315 * cpuset_init_smp - initialize cpus_allowed
2317 * Description: Finish top cpuset after cpu, node maps are initialized
2319 void __init cpuset_init_smp(void)
2321 cpumask_copy(top_cpuset.cpus_allowed, cpu_active_mask);
2322 top_cpuset.mems_allowed = node_states[N_MEMORY];
2323 top_cpuset.old_mems_allowed = top_cpuset.mems_allowed;
2325 register_hotmemory_notifier(&cpuset_track_online_nodes_nb);
2329 * cpuset_cpus_allowed - return cpus_allowed mask from a tasks cpuset.
2330 * @tsk: pointer to task_struct from which to obtain cpuset->cpus_allowed.
2331 * @pmask: pointer to struct cpumask variable to receive cpus_allowed set.
2333 * Description: Returns the cpumask_var_t cpus_allowed of the cpuset
2334 * attached to the specified @tsk. Guaranteed to return some non-empty
2335 * subset of cpu_online_mask, even if this means going outside the
2339 void cpuset_cpus_allowed(struct task_struct *tsk, struct cpumask *pmask)
2341 struct cpuset *cpus_cs;
2343 mutex_lock(&callback_mutex);
2345 cpus_cs = effective_cpumask_cpuset(task_cs(tsk));
2346 guarantee_online_cpus(cpus_cs, pmask);
2348 mutex_unlock(&callback_mutex);
2351 void cpuset_cpus_allowed_fallback(struct task_struct *tsk)
2353 const struct cpuset *cpus_cs;
2356 cpus_cs = effective_cpumask_cpuset(task_cs(tsk));
2357 do_set_cpus_allowed(tsk, cpus_cs->cpus_allowed);
2361 * We own tsk->cpus_allowed, nobody can change it under us.
2363 * But we used cs && cs->cpus_allowed lockless and thus can
2364 * race with cgroup_attach_task() or update_cpumask() and get
2365 * the wrong tsk->cpus_allowed. However, both cases imply the
2366 * subsequent cpuset_change_cpumask()->set_cpus_allowed_ptr()
2367 * which takes task_rq_lock().
2369 * If we are called after it dropped the lock we must see all
2370 * changes in tsk_cs()->cpus_allowed. Otherwise we can temporary
2371 * set any mask even if it is not right from task_cs() pov,
2372 * the pending set_cpus_allowed_ptr() will fix things.
2374 * select_fallback_rq() will fix things ups and set cpu_possible_mask
2379 void cpuset_init_current_mems_allowed(void)
2381 nodes_setall(current->mems_allowed);
2385 * cpuset_mems_allowed - return mems_allowed mask from a tasks cpuset.
2386 * @tsk: pointer to task_struct from which to obtain cpuset->mems_allowed.
2388 * Description: Returns the nodemask_t mems_allowed of the cpuset
2389 * attached to the specified @tsk. Guaranteed to return some non-empty
2390 * subset of node_states[N_MEMORY], even if this means going outside the
2394 nodemask_t cpuset_mems_allowed(struct task_struct *tsk)
2396 struct cpuset *mems_cs;
2399 mutex_lock(&callback_mutex);
2401 mems_cs = effective_nodemask_cpuset(task_cs(tsk));
2402 guarantee_online_mems(mems_cs, &mask);
2404 mutex_unlock(&callback_mutex);
2410 * cpuset_nodemask_valid_mems_allowed - check nodemask vs. curremt mems_allowed
2411 * @nodemask: the nodemask to be checked
2413 * Are any of the nodes in the nodemask allowed in current->mems_allowed?
2415 int cpuset_nodemask_valid_mems_allowed(nodemask_t *nodemask)
2417 return nodes_intersects(*nodemask, current->mems_allowed);
2421 * nearest_hardwall_ancestor() - Returns the nearest mem_exclusive or
2422 * mem_hardwall ancestor to the specified cpuset. Call holding
2423 * callback_mutex. If no ancestor is mem_exclusive or mem_hardwall
2424 * (an unusual configuration), then returns the root cpuset.
2426 static const struct cpuset *nearest_hardwall_ancestor(const struct cpuset *cs)
2428 while (!(is_mem_exclusive(cs) || is_mem_hardwall(cs)) && parent_cs(cs))
2434 * cpuset_node_allowed_softwall - Can we allocate on a memory node?
2435 * @node: is this an allowed node?
2436 * @gfp_mask: memory allocation flags
2438 * If we're in interrupt, yes, we can always allocate. If __GFP_THISNODE is
2439 * set, yes, we can always allocate. If node is in our task's mems_allowed,
2440 * yes. If it's not a __GFP_HARDWALL request and this node is in the nearest
2441 * hardwalled cpuset ancestor to this task's cpuset, yes. If the task has been
2442 * OOM killed and has access to memory reserves as specified by the TIF_MEMDIE
2446 * If __GFP_HARDWALL is set, cpuset_node_allowed_softwall() reduces to
2447 * cpuset_node_allowed_hardwall(). Otherwise, cpuset_node_allowed_softwall()
2448 * might sleep, and might allow a node from an enclosing cpuset.
2450 * cpuset_node_allowed_hardwall() only handles the simpler case of hardwall
2451 * cpusets, and never sleeps.
2453 * The __GFP_THISNODE placement logic is really handled elsewhere,
2454 * by forcibly using a zonelist starting at a specified node, and by
2455 * (in get_page_from_freelist()) refusing to consider the zones for
2456 * any node on the zonelist except the first. By the time any such
2457 * calls get to this routine, we should just shut up and say 'yes'.
2459 * GFP_USER allocations are marked with the __GFP_HARDWALL bit,
2460 * and do not allow allocations outside the current tasks cpuset
2461 * unless the task has been OOM killed as is marked TIF_MEMDIE.
2462 * GFP_KERNEL allocations are not so marked, so can escape to the
2463 * nearest enclosing hardwalled ancestor cpuset.
2465 * Scanning up parent cpusets requires callback_mutex. The
2466 * __alloc_pages() routine only calls here with __GFP_HARDWALL bit
2467 * _not_ set if it's a GFP_KERNEL allocation, and all nodes in the
2468 * current tasks mems_allowed came up empty on the first pass over
2469 * the zonelist. So only GFP_KERNEL allocations, if all nodes in the
2470 * cpuset are short of memory, might require taking the callback_mutex
2473 * The first call here from mm/page_alloc:get_page_from_freelist()
2474 * has __GFP_HARDWALL set in gfp_mask, enforcing hardwall cpusets,
2475 * so no allocation on a node outside the cpuset is allowed (unless
2476 * in interrupt, of course).
2478 * The second pass through get_page_from_freelist() doesn't even call
2479 * here for GFP_ATOMIC calls. For those calls, the __alloc_pages()
2480 * variable 'wait' is not set, and the bit ALLOC_CPUSET is not set
2481 * in alloc_flags. That logic and the checks below have the combined
2483 * in_interrupt - any node ok (current task context irrelevant)
2484 * GFP_ATOMIC - any node ok
2485 * TIF_MEMDIE - any node ok
2486 * GFP_KERNEL - any node in enclosing hardwalled cpuset ok
2487 * GFP_USER - only nodes in current tasks mems allowed ok.
2490 * Don't call cpuset_node_allowed_softwall if you can't sleep, unless you
2491 * pass in the __GFP_HARDWALL flag set in gfp_flag, which disables
2492 * the code that might scan up ancestor cpusets and sleep.
2494 int __cpuset_node_allowed_softwall(int node, gfp_t gfp_mask)
2496 const struct cpuset *cs; /* current cpuset ancestors */
2497 int allowed; /* is allocation in zone z allowed? */
2499 if (in_interrupt() || (gfp_mask & __GFP_THISNODE))
2501 might_sleep_if(!(gfp_mask & __GFP_HARDWALL));
2502 if (node_isset(node, current->mems_allowed))
2505 * Allow tasks that have access to memory reserves because they have
2506 * been OOM killed to get memory anywhere.
2508 if (unlikely(test_thread_flag(TIF_MEMDIE)))
2510 if (gfp_mask & __GFP_HARDWALL) /* If hardwall request, stop here */
2513 if (current->flags & PF_EXITING) /* Let dying task have memory */
2516 /* Not hardwall and node outside mems_allowed: scan up cpusets */
2517 mutex_lock(&callback_mutex);
2520 cs = nearest_hardwall_ancestor(task_cs(current));
2521 task_unlock(current);
2523 allowed = node_isset(node, cs->mems_allowed);
2524 mutex_unlock(&callback_mutex);
2529 * cpuset_node_allowed_hardwall - Can we allocate on a memory node?
2530 * @node: is this an allowed node?
2531 * @gfp_mask: memory allocation flags
2533 * If we're in interrupt, yes, we can always allocate. If __GFP_THISNODE is
2534 * set, yes, we can always allocate. If node is in our task's mems_allowed,
2535 * yes. If the task has been OOM killed and has access to memory reserves as
2536 * specified by the TIF_MEMDIE flag, yes.
2539 * The __GFP_THISNODE placement logic is really handled elsewhere,
2540 * by forcibly using a zonelist starting at a specified node, and by
2541 * (in get_page_from_freelist()) refusing to consider the zones for
2542 * any node on the zonelist except the first. By the time any such
2543 * calls get to this routine, we should just shut up and say 'yes'.
2545 * Unlike the cpuset_node_allowed_softwall() variant, above,
2546 * this variant requires that the node be in the current task's
2547 * mems_allowed or that we're in interrupt. It does not scan up the
2548 * cpuset hierarchy for the nearest enclosing mem_exclusive cpuset.
2551 int __cpuset_node_allowed_hardwall(int node, gfp_t gfp_mask)
2553 if (in_interrupt() || (gfp_mask & __GFP_THISNODE))
2555 if (node_isset(node, current->mems_allowed))
2558 * Allow tasks that have access to memory reserves because they have
2559 * been OOM killed to get memory anywhere.
2561 if (unlikely(test_thread_flag(TIF_MEMDIE)))
2567 * cpuset_mem_spread_node() - On which node to begin search for a file page
2568 * cpuset_slab_spread_node() - On which node to begin search for a slab page
2570 * If a task is marked PF_SPREAD_PAGE or PF_SPREAD_SLAB (as for
2571 * tasks in a cpuset with is_spread_page or is_spread_slab set),
2572 * and if the memory allocation used cpuset_mem_spread_node()
2573 * to determine on which node to start looking, as it will for
2574 * certain page cache or slab cache pages such as used for file
2575 * system buffers and inode caches, then instead of starting on the
2576 * local node to look for a free page, rather spread the starting
2577 * node around the tasks mems_allowed nodes.
2579 * We don't have to worry about the returned node being offline
2580 * because "it can't happen", and even if it did, it would be ok.
2582 * The routines calling guarantee_online_mems() are careful to
2583 * only set nodes in task->mems_allowed that are online. So it
2584 * should not be possible for the following code to return an
2585 * offline node. But if it did, that would be ok, as this routine
2586 * is not returning the node where the allocation must be, only
2587 * the node where the search should start. The zonelist passed to
2588 * __alloc_pages() will include all nodes. If the slab allocator
2589 * is passed an offline node, it will fall back to the local node.
2590 * See kmem_cache_alloc_node().
2593 static int cpuset_spread_node(int *rotor)
2597 node = next_node(*rotor, current->mems_allowed);
2598 if (node == MAX_NUMNODES)
2599 node = first_node(current->mems_allowed);
2604 int cpuset_mem_spread_node(void)
2606 if (current->cpuset_mem_spread_rotor == NUMA_NO_NODE)
2607 current->cpuset_mem_spread_rotor =
2608 node_random(¤t->mems_allowed);
2610 return cpuset_spread_node(¤t->cpuset_mem_spread_rotor);
2613 int cpuset_slab_spread_node(void)
2615 if (current->cpuset_slab_spread_rotor == NUMA_NO_NODE)
2616 current->cpuset_slab_spread_rotor =
2617 node_random(¤t->mems_allowed);
2619 return cpuset_spread_node(¤t->cpuset_slab_spread_rotor);
2622 EXPORT_SYMBOL_GPL(cpuset_mem_spread_node);
2625 * cpuset_mems_allowed_intersects - Does @tsk1's mems_allowed intersect @tsk2's?
2626 * @tsk1: pointer to task_struct of some task.
2627 * @tsk2: pointer to task_struct of some other task.
2629 * Description: Return true if @tsk1's mems_allowed intersects the
2630 * mems_allowed of @tsk2. Used by the OOM killer to determine if
2631 * one of the task's memory usage might impact the memory available
2635 int cpuset_mems_allowed_intersects(const struct task_struct *tsk1,
2636 const struct task_struct *tsk2)
2638 return nodes_intersects(tsk1->mems_allowed, tsk2->mems_allowed);
2641 #define CPUSET_NODELIST_LEN (256)
2644 * cpuset_print_task_mems_allowed - prints task's cpuset and mems_allowed
2645 * @task: pointer to task_struct of some task.
2647 * Description: Prints @task's name, cpuset name, and cached copy of its
2648 * mems_allowed to the kernel log. Must hold task_lock(task) to allow
2649 * dereferencing task_cs(task).
2651 void cpuset_print_task_mems_allowed(struct task_struct *tsk)
2653 /* Statically allocated to prevent using excess stack. */
2654 static char cpuset_nodelist[CPUSET_NODELIST_LEN];
2655 static DEFINE_SPINLOCK(cpuset_buffer_lock);
2657 struct cgroup *cgrp = task_cs(tsk)->css.cgroup;
2660 spin_lock(&cpuset_buffer_lock);
2662 nodelist_scnprintf(cpuset_nodelist, CPUSET_NODELIST_LEN,
2664 printk(KERN_INFO "%s cpuset=%s mems_allowed=%s\n",
2665 tsk->comm, cgroup_name(cgrp), cpuset_nodelist);
2667 spin_unlock(&cpuset_buffer_lock);
2672 * Collection of memory_pressure is suppressed unless
2673 * this flag is enabled by writing "1" to the special
2674 * cpuset file 'memory_pressure_enabled' in the root cpuset.
2677 int cpuset_memory_pressure_enabled __read_mostly;
2680 * cpuset_memory_pressure_bump - keep stats of per-cpuset reclaims.
2682 * Keep a running average of the rate of synchronous (direct)
2683 * page reclaim efforts initiated by tasks in each cpuset.
2685 * This represents the rate at which some task in the cpuset
2686 * ran low on memory on all nodes it was allowed to use, and
2687 * had to enter the kernels page reclaim code in an effort to
2688 * create more free memory by tossing clean pages or swapping
2689 * or writing dirty pages.
2691 * Display to user space in the per-cpuset read-only file
2692 * "memory_pressure". Value displayed is an integer
2693 * representing the recent rate of entry into the synchronous
2694 * (direct) page reclaim by any task attached to the cpuset.
2697 void __cpuset_memory_pressure_bump(void)
2700 fmeter_markevent(&task_cs(current)->fmeter);
2701 task_unlock(current);
2704 #ifdef CONFIG_PROC_PID_CPUSET
2706 * proc_cpuset_show()
2707 * - Print tasks cpuset path into seq_file.
2708 * - Used for /proc/<pid>/cpuset.
2709 * - No need to task_lock(tsk) on this tsk->cpuset reference, as it
2710 * doesn't really matter if tsk->cpuset changes after we read it,
2711 * and we take cpuset_mutex, keeping cpuset_attach() from changing it
2714 int proc_cpuset_show(struct seq_file *m, void *unused_v)
2717 struct task_struct *tsk;
2719 struct cgroup_subsys_state *css;
2723 buf = kmalloc(PAGE_SIZE, GFP_KERNEL);
2729 tsk = get_pid_task(pid, PIDTYPE_PID);
2734 css = task_subsys_state(tsk, cpuset_subsys_id);
2735 retval = cgroup_path(css->cgroup, buf, PAGE_SIZE);
2742 put_task_struct(tsk);
2748 #endif /* CONFIG_PROC_PID_CPUSET */
2750 /* Display task mems_allowed in /proc/<pid>/status file. */
2751 void cpuset_task_status_allowed(struct seq_file *m, struct task_struct *task)
2753 seq_printf(m, "Mems_allowed:\t");
2754 seq_nodemask(m, &task->mems_allowed);
2755 seq_printf(m, "\n");
2756 seq_printf(m, "Mems_allowed_list:\t");
2757 seq_nodemask_list(m, &task->mems_allowed);
2758 seq_printf(m, "\n");