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 */
91 struct fmeter fmeter; /* memory_pressure filter */
94 * Tasks are being attached to this cpuset. Used to prevent
95 * zeroing cpus/mems_allowed between ->can_attach() and ->attach().
97 int attach_in_progress;
99 /* partition number for rebuild_sched_domains() */
102 /* for custom sched domain */
103 int relax_domain_level;
106 /* Retrieve the cpuset for a cgroup */
107 static inline struct cpuset *cgroup_cs(struct cgroup *cont)
109 return container_of(cgroup_subsys_state(cont, cpuset_subsys_id),
113 /* Retrieve the cpuset for a task */
114 static inline struct cpuset *task_cs(struct task_struct *task)
116 return container_of(task_subsys_state(task, cpuset_subsys_id),
120 static inline struct cpuset *parent_cs(const struct cpuset *cs)
122 struct cgroup *pcgrp = cs->css.cgroup->parent;
125 return cgroup_cs(pcgrp);
130 static inline bool task_has_mempolicy(struct task_struct *task)
132 return task->mempolicy;
135 static inline bool task_has_mempolicy(struct task_struct *task)
142 /* bits in struct cpuset flags field */
149 CS_SCHED_LOAD_BALANCE,
154 /* convenient tests for these bits */
155 static inline bool is_cpuset_online(const struct cpuset *cs)
157 return test_bit(CS_ONLINE, &cs->flags);
160 static inline int is_cpu_exclusive(const struct cpuset *cs)
162 return test_bit(CS_CPU_EXCLUSIVE, &cs->flags);
165 static inline int is_mem_exclusive(const struct cpuset *cs)
167 return test_bit(CS_MEM_EXCLUSIVE, &cs->flags);
170 static inline int is_mem_hardwall(const struct cpuset *cs)
172 return test_bit(CS_MEM_HARDWALL, &cs->flags);
175 static inline int is_sched_load_balance(const struct cpuset *cs)
177 return test_bit(CS_SCHED_LOAD_BALANCE, &cs->flags);
180 static inline int is_memory_migrate(const struct cpuset *cs)
182 return test_bit(CS_MEMORY_MIGRATE, &cs->flags);
185 static inline int is_spread_page(const struct cpuset *cs)
187 return test_bit(CS_SPREAD_PAGE, &cs->flags);
190 static inline int is_spread_slab(const struct cpuset *cs)
192 return test_bit(CS_SPREAD_SLAB, &cs->flags);
195 static struct cpuset top_cpuset = {
196 .flags = ((1 << CS_ONLINE) | (1 << CS_CPU_EXCLUSIVE) |
197 (1 << CS_MEM_EXCLUSIVE)),
201 * cpuset_for_each_child - traverse online children of a cpuset
202 * @child_cs: loop cursor pointing to the current child
203 * @pos_cgrp: used for iteration
204 * @parent_cs: target cpuset to walk children of
206 * Walk @child_cs through the online children of @parent_cs. Must be used
207 * with RCU read locked.
209 #define cpuset_for_each_child(child_cs, pos_cgrp, parent_cs) \
210 cgroup_for_each_child((pos_cgrp), (parent_cs)->css.cgroup) \
211 if (is_cpuset_online(((child_cs) = cgroup_cs((pos_cgrp)))))
214 * cpuset_for_each_descendant_pre - pre-order walk of a cpuset's descendants
215 * @des_cs: loop cursor pointing to the current descendant
216 * @pos_cgrp: used for iteration
217 * @root_cs: target cpuset to walk ancestor of
219 * Walk @des_cs through the online descendants of @root_cs. Must be used
220 * with RCU read locked. The caller may modify @pos_cgrp by calling
221 * cgroup_rightmost_descendant() to skip subtree.
223 #define cpuset_for_each_descendant_pre(des_cs, pos_cgrp, root_cs) \
224 cgroup_for_each_descendant_pre((pos_cgrp), (root_cs)->css.cgroup) \
225 if (is_cpuset_online(((des_cs) = cgroup_cs((pos_cgrp)))))
228 * There are two global mutexes guarding cpuset structures - cpuset_mutex
229 * and callback_mutex. The latter may nest inside the former. We also
230 * require taking task_lock() when dereferencing a task's cpuset pointer.
231 * See "The task_lock() exception", at the end of this comment.
233 * A task must hold both mutexes to modify cpusets. If a task holds
234 * cpuset_mutex, then it blocks others wanting that mutex, ensuring that it
235 * is the only task able to also acquire callback_mutex and be able to
236 * modify cpusets. It can perform various checks on the cpuset structure
237 * first, knowing nothing will change. It can also allocate memory while
238 * just holding cpuset_mutex. While it is performing these checks, various
239 * callback routines can briefly acquire callback_mutex to query cpusets.
240 * Once it is ready to make the changes, it takes callback_mutex, blocking
243 * Calls to the kernel memory allocator can not be made while holding
244 * callback_mutex, as that would risk double tripping on callback_mutex
245 * from one of the callbacks into the cpuset code from within
248 * If a task is only holding callback_mutex, then it has read-only
251 * Now, the task_struct fields mems_allowed and mempolicy may be changed
252 * by other task, we use alloc_lock in the task_struct fields to protect
255 * The cpuset_common_file_read() handlers only hold callback_mutex across
256 * small pieces of code, such as when reading out possibly multi-word
257 * cpumasks and nodemasks.
259 * Accessing a task's cpuset should be done in accordance with the
260 * guidelines for accessing subsystem state in kernel/cgroup.c
263 static DEFINE_MUTEX(cpuset_mutex);
264 static DEFINE_MUTEX(callback_mutex);
267 * CPU / memory hotplug is handled asynchronously.
269 static void cpuset_hotplug_workfn(struct work_struct *work);
270 static DECLARE_WORK(cpuset_hotplug_work, cpuset_hotplug_workfn);
272 static DECLARE_WAIT_QUEUE_HEAD(cpuset_attach_wq);
275 * This is ugly, but preserves the userspace API for existing cpuset
276 * users. If someone tries to mount the "cpuset" filesystem, we
277 * silently switch it to mount "cgroup" instead
279 static struct dentry *cpuset_mount(struct file_system_type *fs_type,
280 int flags, const char *unused_dev_name, void *data)
282 struct file_system_type *cgroup_fs = get_fs_type("cgroup");
283 struct dentry *ret = ERR_PTR(-ENODEV);
287 "release_agent=/sbin/cpuset_release_agent";
288 ret = cgroup_fs->mount(cgroup_fs, flags,
289 unused_dev_name, mountopts);
290 put_filesystem(cgroup_fs);
295 static struct file_system_type cpuset_fs_type = {
297 .mount = cpuset_mount,
301 * Return in pmask the portion of a cpusets's cpus_allowed that
302 * are online. If none are online, walk up the cpuset hierarchy
303 * until we find one that does have some online cpus. The top
304 * cpuset always has some cpus online.
306 * One way or another, we guarantee to return some non-empty subset
307 * of cpu_online_mask.
309 * Call with callback_mutex held.
311 static void guarantee_online_cpus(const struct cpuset *cs,
312 struct cpumask *pmask)
314 while (!cpumask_intersects(cs->cpus_allowed, cpu_online_mask))
316 cpumask_and(pmask, cs->cpus_allowed, cpu_online_mask);
320 * Return in *pmask the portion of a cpusets's mems_allowed that
321 * are online, with memory. If none are online with memory, walk
322 * up the cpuset hierarchy until we find one that does have some
323 * online mems. The top cpuset always has some mems online.
325 * One way or another, we guarantee to return some non-empty subset
326 * of node_states[N_MEMORY].
328 * Call with callback_mutex held.
330 static void guarantee_online_mems(const struct cpuset *cs, nodemask_t *pmask)
332 while (!nodes_intersects(cs->mems_allowed, node_states[N_MEMORY]))
334 nodes_and(*pmask, cs->mems_allowed, node_states[N_MEMORY]);
338 * update task's spread flag if cpuset's page/slab spread flag is set
340 * Called with callback_mutex/cpuset_mutex held
342 static void cpuset_update_task_spread_flag(struct cpuset *cs,
343 struct task_struct *tsk)
345 if (is_spread_page(cs))
346 tsk->flags |= PF_SPREAD_PAGE;
348 tsk->flags &= ~PF_SPREAD_PAGE;
349 if (is_spread_slab(cs))
350 tsk->flags |= PF_SPREAD_SLAB;
352 tsk->flags &= ~PF_SPREAD_SLAB;
356 * is_cpuset_subset(p, q) - Is cpuset p a subset of cpuset q?
358 * One cpuset is a subset of another if all its allowed CPUs and
359 * Memory Nodes are a subset of the other, and its exclusive flags
360 * are only set if the other's are set. Call holding cpuset_mutex.
363 static int is_cpuset_subset(const struct cpuset *p, const struct cpuset *q)
365 return cpumask_subset(p->cpus_allowed, q->cpus_allowed) &&
366 nodes_subset(p->mems_allowed, q->mems_allowed) &&
367 is_cpu_exclusive(p) <= is_cpu_exclusive(q) &&
368 is_mem_exclusive(p) <= is_mem_exclusive(q);
372 * alloc_trial_cpuset - allocate a trial cpuset
373 * @cs: the cpuset that the trial cpuset duplicates
375 static struct cpuset *alloc_trial_cpuset(const struct cpuset *cs)
377 struct cpuset *trial;
379 trial = kmemdup(cs, sizeof(*cs), GFP_KERNEL);
383 if (!alloc_cpumask_var(&trial->cpus_allowed, GFP_KERNEL)) {
387 cpumask_copy(trial->cpus_allowed, cs->cpus_allowed);
393 * free_trial_cpuset - free the trial cpuset
394 * @trial: the trial cpuset to be freed
396 static void free_trial_cpuset(struct cpuset *trial)
398 free_cpumask_var(trial->cpus_allowed);
403 * validate_change() - Used to validate that any proposed cpuset change
404 * follows the structural rules for cpusets.
406 * If we replaced the flag and mask values of the current cpuset
407 * (cur) with those values in the trial cpuset (trial), would
408 * our various subset and exclusive rules still be valid? Presumes
411 * 'cur' is the address of an actual, in-use cpuset. Operations
412 * such as list traversal that depend on the actual address of the
413 * cpuset in the list must use cur below, not trial.
415 * 'trial' is the address of bulk structure copy of cur, with
416 * perhaps one or more of the fields cpus_allowed, mems_allowed,
417 * or flags changed to new, trial values.
419 * Return 0 if valid, -errno if not.
422 static int validate_change(const struct cpuset *cur, const struct cpuset *trial)
425 struct cpuset *c, *par;
430 /* Each of our child cpusets must be a subset of us */
432 cpuset_for_each_child(c, cont, cur)
433 if (!is_cpuset_subset(c, trial))
436 /* Remaining checks don't apply to root cpuset */
438 if (cur == &top_cpuset)
441 par = parent_cs(cur);
443 /* We must be a subset of our parent cpuset */
445 if (!is_cpuset_subset(trial, par))
449 * If either I or some sibling (!= me) is exclusive, we can't
453 cpuset_for_each_child(c, cont, par) {
454 if ((is_cpu_exclusive(trial) || is_cpu_exclusive(c)) &&
456 cpumask_intersects(trial->cpus_allowed, c->cpus_allowed))
458 if ((is_mem_exclusive(trial) || is_mem_exclusive(c)) &&
460 nodes_intersects(trial->mems_allowed, c->mems_allowed))
465 * Cpusets with tasks - existing or newly being attached - can't
466 * have empty cpus_allowed or mems_allowed.
469 if ((cgroup_task_count(cur->css.cgroup) || cur->attach_in_progress) &&
470 (cpumask_empty(trial->cpus_allowed) ||
471 nodes_empty(trial->mems_allowed)))
482 * Helper routine for generate_sched_domains().
483 * Do cpusets a, b have overlapping cpus_allowed masks?
485 static int cpusets_overlap(struct cpuset *a, struct cpuset *b)
487 return cpumask_intersects(a->cpus_allowed, b->cpus_allowed);
491 update_domain_attr(struct sched_domain_attr *dattr, struct cpuset *c)
493 if (dattr->relax_domain_level < c->relax_domain_level)
494 dattr->relax_domain_level = c->relax_domain_level;
498 static void update_domain_attr_tree(struct sched_domain_attr *dattr,
499 struct cpuset *root_cs)
502 struct cgroup *pos_cgrp;
505 cpuset_for_each_descendant_pre(cp, pos_cgrp, root_cs) {
506 /* skip the whole subtree if @cp doesn't have any CPU */
507 if (cpumask_empty(cp->cpus_allowed)) {
508 pos_cgrp = cgroup_rightmost_descendant(pos_cgrp);
512 if (is_sched_load_balance(cp))
513 update_domain_attr(dattr, cp);
519 * generate_sched_domains()
521 * This function builds a partial partition of the systems CPUs
522 * A 'partial partition' is a set of non-overlapping subsets whose
523 * union is a subset of that set.
524 * The output of this function needs to be passed to kernel/sched.c
525 * partition_sched_domains() routine, which will rebuild the scheduler's
526 * load balancing domains (sched domains) as specified by that partial
529 * See "What is sched_load_balance" in Documentation/cgroups/cpusets.txt
530 * for a background explanation of this.
532 * Does not return errors, on the theory that the callers of this
533 * routine would rather not worry about failures to rebuild sched
534 * domains when operating in the severe memory shortage situations
535 * that could cause allocation failures below.
537 * Must be called with cpuset_mutex held.
539 * The three key local variables below are:
540 * q - a linked-list queue of cpuset pointers, used to implement a
541 * top-down scan of all cpusets. This scan loads a pointer
542 * to each cpuset marked is_sched_load_balance into the
543 * array 'csa'. For our purposes, rebuilding the schedulers
544 * sched domains, we can ignore !is_sched_load_balance cpusets.
545 * csa - (for CpuSet Array) Array of pointers to all the cpusets
546 * that need to be load balanced, for convenient iterative
547 * access by the subsequent code that finds the best partition,
548 * i.e the set of domains (subsets) of CPUs such that the
549 * cpus_allowed of every cpuset marked is_sched_load_balance
550 * is a subset of one of these domains, while there are as
551 * many such domains as possible, each as small as possible.
552 * doms - Conversion of 'csa' to an array of cpumasks, for passing to
553 * the kernel/sched.c routine partition_sched_domains() in a
554 * convenient format, that can be easily compared to the prior
555 * value to determine what partition elements (sched domains)
556 * were changed (added or removed.)
558 * Finding the best partition (set of domains):
559 * The triple nested loops below over i, j, k scan over the
560 * load balanced cpusets (using the array of cpuset pointers in
561 * csa[]) looking for pairs of cpusets that have overlapping
562 * cpus_allowed, but which don't have the same 'pn' partition
563 * number and gives them in the same partition number. It keeps
564 * looping on the 'restart' label until it can no longer find
567 * The union of the cpus_allowed masks from the set of
568 * all cpusets having the same 'pn' value then form the one
569 * element of the partition (one sched domain) to be passed to
570 * partition_sched_domains().
572 static int generate_sched_domains(cpumask_var_t **domains,
573 struct sched_domain_attr **attributes)
575 struct cpuset *cp; /* scans q */
576 struct cpuset **csa; /* array of all cpuset ptrs */
577 int csn; /* how many cpuset ptrs in csa so far */
578 int i, j, k; /* indices for partition finding loops */
579 cpumask_var_t *doms; /* resulting partition; i.e. sched domains */
580 struct sched_domain_attr *dattr; /* attributes for custom domains */
581 int ndoms = 0; /* number of sched domains in result */
582 int nslot; /* next empty doms[] struct cpumask slot */
583 struct cgroup *pos_cgrp;
589 /* Special case for the 99% of systems with one, full, sched domain */
590 if (is_sched_load_balance(&top_cpuset)) {
592 doms = alloc_sched_domains(ndoms);
596 dattr = kmalloc(sizeof(struct sched_domain_attr), GFP_KERNEL);
598 *dattr = SD_ATTR_INIT;
599 update_domain_attr_tree(dattr, &top_cpuset);
601 cpumask_copy(doms[0], top_cpuset.cpus_allowed);
606 csa = kmalloc(number_of_cpusets * sizeof(cp), GFP_KERNEL);
612 cpuset_for_each_descendant_pre(cp, pos_cgrp, &top_cpuset) {
614 * Continue traversing beyond @cp iff @cp has some CPUs and
615 * isn't load balancing. The former is obvious. The
616 * latter: All child cpusets contain a subset of the
617 * parent's cpus, so just skip them, and then we call
618 * update_domain_attr_tree() to calc relax_domain_level of
619 * the corresponding sched domain.
621 if (!cpumask_empty(cp->cpus_allowed) &&
622 !is_sched_load_balance(cp))
625 if (is_sched_load_balance(cp))
628 /* skip @cp's subtree */
629 pos_cgrp = cgroup_rightmost_descendant(pos_cgrp);
633 for (i = 0; i < csn; i++)
638 /* Find the best partition (set of sched domains) */
639 for (i = 0; i < csn; i++) {
640 struct cpuset *a = csa[i];
643 for (j = 0; j < csn; j++) {
644 struct cpuset *b = csa[j];
647 if (apn != bpn && cpusets_overlap(a, b)) {
648 for (k = 0; k < csn; k++) {
649 struct cpuset *c = csa[k];
654 ndoms--; /* one less element */
661 * Now we know how many domains to create.
662 * Convert <csn, csa> to <ndoms, doms> and populate cpu masks.
664 doms = alloc_sched_domains(ndoms);
669 * The rest of the code, including the scheduler, can deal with
670 * dattr==NULL case. No need to abort if alloc fails.
672 dattr = kmalloc(ndoms * sizeof(struct sched_domain_attr), GFP_KERNEL);
674 for (nslot = 0, i = 0; i < csn; i++) {
675 struct cpuset *a = csa[i];
680 /* Skip completed partitions */
686 if (nslot == ndoms) {
687 static int warnings = 10;
690 "rebuild_sched_domains confused:"
691 " nslot %d, ndoms %d, csn %d, i %d,"
693 nslot, ndoms, csn, i, apn);
701 *(dattr + nslot) = SD_ATTR_INIT;
702 for (j = i; j < csn; j++) {
703 struct cpuset *b = csa[j];
706 cpumask_or(dp, dp, b->cpus_allowed);
708 update_domain_attr_tree(dattr + nslot, b);
710 /* Done with this partition */
716 BUG_ON(nslot != ndoms);
722 * Fallback to the default domain if kmalloc() failed.
723 * See comments in partition_sched_domains().
734 * Rebuild scheduler domains.
736 * If the flag 'sched_load_balance' of any cpuset with non-empty
737 * 'cpus' changes, or if the 'cpus' allowed changes in any cpuset
738 * which has that flag enabled, or if any cpuset with a non-empty
739 * 'cpus' is removed, then call this routine to rebuild the
740 * scheduler's dynamic sched domains.
742 * Call with cpuset_mutex held. Takes get_online_cpus().
744 static void rebuild_sched_domains_locked(void)
746 struct sched_domain_attr *attr;
750 lockdep_assert_held(&cpuset_mutex);
754 * We have raced with CPU hotplug. Don't do anything to avoid
755 * passing doms with offlined cpu to partition_sched_domains().
756 * Anyways, hotplug work item will rebuild sched domains.
758 if (!cpumask_equal(top_cpuset.cpus_allowed, cpu_active_mask))
761 /* Generate domain masks and attrs */
762 ndoms = generate_sched_domains(&doms, &attr);
764 /* Have scheduler rebuild the domains */
765 partition_sched_domains(ndoms, doms, attr);
769 #else /* !CONFIG_SMP */
770 static void rebuild_sched_domains_locked(void)
773 #endif /* CONFIG_SMP */
775 void rebuild_sched_domains(void)
777 mutex_lock(&cpuset_mutex);
778 rebuild_sched_domains_locked();
779 mutex_unlock(&cpuset_mutex);
783 * cpuset_change_cpumask - make a task's cpus_allowed the same as its cpuset's
785 * @scan: struct cgroup_scanner containing the cgroup of the task
787 * Called by cgroup_scan_tasks() for each task in a cgroup whose
788 * cpus_allowed mask needs to be changed.
790 * We don't need to re-check for the cgroup/cpuset membership, since we're
791 * holding cpuset_mutex at this point.
793 static void cpuset_change_cpumask(struct task_struct *tsk,
794 struct cgroup_scanner *scan)
796 set_cpus_allowed_ptr(tsk, ((cgroup_cs(scan->cg))->cpus_allowed));
800 * update_tasks_cpumask - Update the cpumasks of tasks in the cpuset.
801 * @cs: the cpuset in which each task's cpus_allowed mask needs to be changed
802 * @heap: if NULL, defer allocating heap memory to cgroup_scan_tasks()
804 * Called with cpuset_mutex held
806 * The cgroup_scan_tasks() function will scan all the tasks in a cgroup,
807 * calling callback functions for each.
809 * No return value. It's guaranteed that cgroup_scan_tasks() always returns 0
812 static void update_tasks_cpumask(struct cpuset *cs, struct ptr_heap *heap)
814 struct cgroup_scanner scan;
816 scan.cg = cs->css.cgroup;
817 scan.test_task = NULL;
818 scan.process_task = cpuset_change_cpumask;
820 cgroup_scan_tasks(&scan);
824 * update_cpumask - update the cpus_allowed mask of a cpuset and all tasks in it
825 * @cs: the cpuset to consider
826 * @buf: buffer of cpu numbers written to this cpuset
828 static int update_cpumask(struct cpuset *cs, struct cpuset *trialcs,
831 struct ptr_heap heap;
833 int is_load_balanced;
835 /* top_cpuset.cpus_allowed tracks cpu_online_mask; it's read-only */
836 if (cs == &top_cpuset)
840 * An empty cpus_allowed is ok only if the cpuset has no tasks.
841 * Since cpulist_parse() fails on an empty mask, we special case
842 * that parsing. The validate_change() call ensures that cpusets
843 * with tasks have cpus.
846 cpumask_clear(trialcs->cpus_allowed);
848 retval = cpulist_parse(buf, trialcs->cpus_allowed);
852 if (!cpumask_subset(trialcs->cpus_allowed, cpu_active_mask))
856 /* Nothing to do if the cpus didn't change */
857 if (cpumask_equal(cs->cpus_allowed, trialcs->cpus_allowed))
860 retval = validate_change(cs, trialcs);
864 retval = heap_init(&heap, PAGE_SIZE, GFP_KERNEL, NULL);
868 is_load_balanced = is_sched_load_balance(trialcs);
870 mutex_lock(&callback_mutex);
871 cpumask_copy(cs->cpus_allowed, trialcs->cpus_allowed);
872 mutex_unlock(&callback_mutex);
875 * Scan tasks in the cpuset, and update the cpumasks of any
876 * that need an update.
878 update_tasks_cpumask(cs, &heap);
882 if (is_load_balanced)
883 rebuild_sched_domains_locked();
890 * Migrate memory region from one set of nodes to another.
892 * Temporarilly set tasks mems_allowed to target nodes of migration,
893 * so that the migration code can allocate pages on these nodes.
895 * Call holding cpuset_mutex, so current's cpuset won't change
896 * during this call, as manage_mutex holds off any cpuset_attach()
897 * calls. Therefore we don't need to take task_lock around the
898 * call to guarantee_online_mems(), as we know no one is changing
901 * While the mm_struct we are migrating is typically from some
902 * other task, the task_struct mems_allowed that we are hacking
903 * is for our current task, which must allocate new pages for that
904 * migrating memory region.
907 static void cpuset_migrate_mm(struct mm_struct *mm, const nodemask_t *from,
908 const nodemask_t *to)
910 struct task_struct *tsk = current;
912 tsk->mems_allowed = *to;
914 do_migrate_pages(mm, from, to, MPOL_MF_MOVE_ALL);
916 guarantee_online_mems(task_cs(tsk),&tsk->mems_allowed);
920 * cpuset_change_task_nodemask - change task's mems_allowed and mempolicy
921 * @tsk: the task to change
922 * @newmems: new nodes that the task will be set
924 * In order to avoid seeing no nodes if the old and new nodes are disjoint,
925 * we structure updates as setting all new allowed nodes, then clearing newly
928 static void cpuset_change_task_nodemask(struct task_struct *tsk,
934 * Allow tasks that have access to memory reserves because they have
935 * been OOM killed to get memory anywhere.
937 if (unlikely(test_thread_flag(TIF_MEMDIE)))
939 if (current->flags & PF_EXITING) /* Let dying task have memory */
944 * Determine if a loop is necessary if another thread is doing
945 * get_mems_allowed(). If at least one node remains unchanged and
946 * tsk does not have a mempolicy, then an empty nodemask will not be
947 * possible when mems_allowed is larger than a word.
949 need_loop = task_has_mempolicy(tsk) ||
950 !nodes_intersects(*newmems, tsk->mems_allowed);
953 write_seqcount_begin(&tsk->mems_allowed_seq);
955 nodes_or(tsk->mems_allowed, tsk->mems_allowed, *newmems);
956 mpol_rebind_task(tsk, newmems, MPOL_REBIND_STEP1);
958 mpol_rebind_task(tsk, newmems, MPOL_REBIND_STEP2);
959 tsk->mems_allowed = *newmems;
962 write_seqcount_end(&tsk->mems_allowed_seq);
968 * Update task's mems_allowed and rebind its mempolicy and vmas' mempolicy
969 * of it to cpuset's new mems_allowed, and migrate pages to new nodes if
970 * memory_migrate flag is set. Called with cpuset_mutex held.
972 static void cpuset_change_nodemask(struct task_struct *p,
973 struct cgroup_scanner *scan)
975 struct mm_struct *mm;
978 const nodemask_t *oldmem = scan->data;
979 static nodemask_t newmems; /* protected by cpuset_mutex */
981 cs = cgroup_cs(scan->cg);
982 guarantee_online_mems(cs, &newmems);
984 cpuset_change_task_nodemask(p, &newmems);
990 migrate = is_memory_migrate(cs);
992 mpol_rebind_mm(mm, &cs->mems_allowed);
994 cpuset_migrate_mm(mm, oldmem, &cs->mems_allowed);
998 static void *cpuset_being_rebound;
1001 * update_tasks_nodemask - Update the nodemasks of tasks in the cpuset.
1002 * @cs: the cpuset in which each task's mems_allowed mask needs to be changed
1003 * @oldmem: old mems_allowed of cpuset cs
1004 * @heap: if NULL, defer allocating heap memory to cgroup_scan_tasks()
1006 * Called with cpuset_mutex held
1007 * No return value. It's guaranteed that cgroup_scan_tasks() always returns 0
1010 static void update_tasks_nodemask(struct cpuset *cs, const nodemask_t *oldmem,
1011 struct ptr_heap *heap)
1013 struct cgroup_scanner scan;
1015 cpuset_being_rebound = cs; /* causes mpol_dup() rebind */
1017 scan.cg = cs->css.cgroup;
1018 scan.test_task = NULL;
1019 scan.process_task = cpuset_change_nodemask;
1021 scan.data = (nodemask_t *)oldmem;
1024 * The mpol_rebind_mm() call takes mmap_sem, which we couldn't
1025 * take while holding tasklist_lock. Forks can happen - the
1026 * mpol_dup() cpuset_being_rebound check will catch such forks,
1027 * and rebind their vma mempolicies too. Because we still hold
1028 * the global cpuset_mutex, we know that no other rebind effort
1029 * will be contending for the global variable cpuset_being_rebound.
1030 * It's ok if we rebind the same mm twice; mpol_rebind_mm()
1031 * is idempotent. Also migrate pages in each mm to new nodes.
1033 cgroup_scan_tasks(&scan);
1035 /* We're done rebinding vmas to this cpuset's new mems_allowed. */
1036 cpuset_being_rebound = NULL;
1040 * Handle user request to change the 'mems' memory placement
1041 * of a cpuset. Needs to validate the request, update the
1042 * cpusets mems_allowed, and for each task in the cpuset,
1043 * update mems_allowed and rebind task's mempolicy and any vma
1044 * mempolicies and if the cpuset is marked 'memory_migrate',
1045 * migrate the tasks pages to the new memory.
1047 * Call with cpuset_mutex held. May take callback_mutex during call.
1048 * Will take tasklist_lock, scan tasklist for tasks in cpuset cs,
1049 * lock each such tasks mm->mmap_sem, scan its vma's and rebind
1050 * their mempolicies to the cpusets new mems_allowed.
1052 static int update_nodemask(struct cpuset *cs, struct cpuset *trialcs,
1055 NODEMASK_ALLOC(nodemask_t, oldmem, GFP_KERNEL);
1057 struct ptr_heap heap;
1063 * top_cpuset.mems_allowed tracks node_stats[N_MEMORY];
1066 if (cs == &top_cpuset) {
1072 * An empty mems_allowed is ok iff there are no tasks in the cpuset.
1073 * Since nodelist_parse() fails on an empty mask, we special case
1074 * that parsing. The validate_change() call ensures that cpusets
1075 * with tasks have memory.
1078 nodes_clear(trialcs->mems_allowed);
1080 retval = nodelist_parse(buf, trialcs->mems_allowed);
1084 if (!nodes_subset(trialcs->mems_allowed,
1085 node_states[N_MEMORY])) {
1090 *oldmem = cs->mems_allowed;
1091 if (nodes_equal(*oldmem, trialcs->mems_allowed)) {
1092 retval = 0; /* Too easy - nothing to do */
1095 retval = validate_change(cs, trialcs);
1099 retval = heap_init(&heap, PAGE_SIZE, GFP_KERNEL, NULL);
1103 mutex_lock(&callback_mutex);
1104 cs->mems_allowed = trialcs->mems_allowed;
1105 mutex_unlock(&callback_mutex);
1107 update_tasks_nodemask(cs, oldmem, &heap);
1111 NODEMASK_FREE(oldmem);
1115 int current_cpuset_is_being_rebound(void)
1117 return task_cs(current) == cpuset_being_rebound;
1120 static int update_relax_domain_level(struct cpuset *cs, s64 val)
1123 if (val < -1 || val >= sched_domain_level_max)
1127 if (val != cs->relax_domain_level) {
1128 cs->relax_domain_level = val;
1129 if (!cpumask_empty(cs->cpus_allowed) &&
1130 is_sched_load_balance(cs))
1131 rebuild_sched_domains_locked();
1138 * cpuset_change_flag - make a task's spread flags the same as its cpuset's
1139 * @tsk: task to be updated
1140 * @scan: struct cgroup_scanner containing the cgroup of the task
1142 * Called by cgroup_scan_tasks() for each task in a cgroup.
1144 * We don't need to re-check for the cgroup/cpuset membership, since we're
1145 * holding cpuset_mutex at this point.
1147 static void cpuset_change_flag(struct task_struct *tsk,
1148 struct cgroup_scanner *scan)
1150 cpuset_update_task_spread_flag(cgroup_cs(scan->cg), tsk);
1154 * update_tasks_flags - update the spread flags of tasks in the cpuset.
1155 * @cs: the cpuset in which each task's spread flags needs to be changed
1156 * @heap: if NULL, defer allocating heap memory to cgroup_scan_tasks()
1158 * Called with cpuset_mutex held
1160 * The cgroup_scan_tasks() function will scan all the tasks in a cgroup,
1161 * calling callback functions for each.
1163 * No return value. It's guaranteed that cgroup_scan_tasks() always returns 0
1166 static void update_tasks_flags(struct cpuset *cs, struct ptr_heap *heap)
1168 struct cgroup_scanner scan;
1170 scan.cg = cs->css.cgroup;
1171 scan.test_task = NULL;
1172 scan.process_task = cpuset_change_flag;
1174 cgroup_scan_tasks(&scan);
1178 * update_flag - read a 0 or a 1 in a file and update associated flag
1179 * bit: the bit to update (see cpuset_flagbits_t)
1180 * cs: the cpuset to update
1181 * turning_on: whether the flag is being set or cleared
1183 * Call with cpuset_mutex held.
1186 static int update_flag(cpuset_flagbits_t bit, struct cpuset *cs,
1189 struct cpuset *trialcs;
1190 int balance_flag_changed;
1191 int spread_flag_changed;
1192 struct ptr_heap heap;
1195 trialcs = alloc_trial_cpuset(cs);
1200 set_bit(bit, &trialcs->flags);
1202 clear_bit(bit, &trialcs->flags);
1204 err = validate_change(cs, trialcs);
1208 err = heap_init(&heap, PAGE_SIZE, GFP_KERNEL, NULL);
1212 balance_flag_changed = (is_sched_load_balance(cs) !=
1213 is_sched_load_balance(trialcs));
1215 spread_flag_changed = ((is_spread_slab(cs) != is_spread_slab(trialcs))
1216 || (is_spread_page(cs) != is_spread_page(trialcs)));
1218 mutex_lock(&callback_mutex);
1219 cs->flags = trialcs->flags;
1220 mutex_unlock(&callback_mutex);
1222 if (!cpumask_empty(trialcs->cpus_allowed) && balance_flag_changed)
1223 rebuild_sched_domains_locked();
1225 if (spread_flag_changed)
1226 update_tasks_flags(cs, &heap);
1229 free_trial_cpuset(trialcs);
1234 * Frequency meter - How fast is some event occurring?
1236 * These routines manage a digitally filtered, constant time based,
1237 * event frequency meter. There are four routines:
1238 * fmeter_init() - initialize a frequency meter.
1239 * fmeter_markevent() - called each time the event happens.
1240 * fmeter_getrate() - returns the recent rate of such events.
1241 * fmeter_update() - internal routine used to update fmeter.
1243 * A common data structure is passed to each of these routines,
1244 * which is used to keep track of the state required to manage the
1245 * frequency meter and its digital filter.
1247 * The filter works on the number of events marked per unit time.
1248 * The filter is single-pole low-pass recursive (IIR). The time unit
1249 * is 1 second. Arithmetic is done using 32-bit integers scaled to
1250 * simulate 3 decimal digits of precision (multiplied by 1000).
1252 * With an FM_COEF of 933, and a time base of 1 second, the filter
1253 * has a half-life of 10 seconds, meaning that if the events quit
1254 * happening, then the rate returned from the fmeter_getrate()
1255 * will be cut in half each 10 seconds, until it converges to zero.
1257 * It is not worth doing a real infinitely recursive filter. If more
1258 * than FM_MAXTICKS ticks have elapsed since the last filter event,
1259 * just compute FM_MAXTICKS ticks worth, by which point the level
1262 * Limit the count of unprocessed events to FM_MAXCNT, so as to avoid
1263 * arithmetic overflow in the fmeter_update() routine.
1265 * Given the simple 32 bit integer arithmetic used, this meter works
1266 * best for reporting rates between one per millisecond (msec) and
1267 * one per 32 (approx) seconds. At constant rates faster than one
1268 * per msec it maxes out at values just under 1,000,000. At constant
1269 * rates between one per msec, and one per second it will stabilize
1270 * to a value N*1000, where N is the rate of events per second.
1271 * At constant rates between one per second and one per 32 seconds,
1272 * it will be choppy, moving up on the seconds that have an event,
1273 * and then decaying until the next event. At rates slower than
1274 * about one in 32 seconds, it decays all the way back to zero between
1278 #define FM_COEF 933 /* coefficient for half-life of 10 secs */
1279 #define FM_MAXTICKS ((time_t)99) /* useless computing more ticks than this */
1280 #define FM_MAXCNT 1000000 /* limit cnt to avoid overflow */
1281 #define FM_SCALE 1000 /* faux fixed point scale */
1283 /* Initialize a frequency meter */
1284 static void fmeter_init(struct fmeter *fmp)
1289 spin_lock_init(&fmp->lock);
1292 /* Internal meter update - process cnt events and update value */
1293 static void fmeter_update(struct fmeter *fmp)
1295 time_t now = get_seconds();
1296 time_t ticks = now - fmp->time;
1301 ticks = min(FM_MAXTICKS, ticks);
1303 fmp->val = (FM_COEF * fmp->val) / FM_SCALE;
1306 fmp->val += ((FM_SCALE - FM_COEF) * fmp->cnt) / FM_SCALE;
1310 /* Process any previous ticks, then bump cnt by one (times scale). */
1311 static void fmeter_markevent(struct fmeter *fmp)
1313 spin_lock(&fmp->lock);
1315 fmp->cnt = min(FM_MAXCNT, fmp->cnt + FM_SCALE);
1316 spin_unlock(&fmp->lock);
1319 /* Process any previous ticks, then return current value. */
1320 static int fmeter_getrate(struct fmeter *fmp)
1324 spin_lock(&fmp->lock);
1327 spin_unlock(&fmp->lock);
1331 /* Called by cgroups to determine if a cpuset is usable; cpuset_mutex held */
1332 static int cpuset_can_attach(struct cgroup *cgrp, struct cgroup_taskset *tset)
1334 struct cpuset *cs = cgroup_cs(cgrp);
1335 struct task_struct *task;
1338 mutex_lock(&cpuset_mutex);
1341 if (cpumask_empty(cs->cpus_allowed) || nodes_empty(cs->mems_allowed))
1344 cgroup_taskset_for_each(task, cgrp, tset) {
1346 * Kthreads which disallow setaffinity shouldn't be moved
1347 * to a new cpuset; we don't want to change their cpu
1348 * affinity and isolating such threads by their set of
1349 * allowed nodes is unnecessary. Thus, cpusets are not
1350 * applicable for such threads. This prevents checking for
1351 * success of set_cpus_allowed_ptr() on all attached tasks
1352 * before cpus_allowed may be changed.
1355 if (task->flags & PF_NO_SETAFFINITY)
1357 ret = security_task_setscheduler(task);
1363 * Mark attach is in progress. This makes validate_change() fail
1364 * changes which zero cpus/mems_allowed.
1366 cs->attach_in_progress++;
1369 mutex_unlock(&cpuset_mutex);
1373 static void cpuset_cancel_attach(struct cgroup *cgrp,
1374 struct cgroup_taskset *tset)
1376 mutex_lock(&cpuset_mutex);
1377 cgroup_cs(cgrp)->attach_in_progress--;
1378 mutex_unlock(&cpuset_mutex);
1382 * Protected by cpuset_mutex. cpus_attach is used only by cpuset_attach()
1383 * but we can't allocate it dynamically there. Define it global and
1384 * allocate from cpuset_init().
1386 static cpumask_var_t cpus_attach;
1388 static void cpuset_attach(struct cgroup *cgrp, struct cgroup_taskset *tset)
1390 /* static buf protected by cpuset_mutex */
1391 static nodemask_t cpuset_attach_nodemask_to;
1392 struct mm_struct *mm;
1393 struct task_struct *task;
1394 struct task_struct *leader = cgroup_taskset_first(tset);
1395 struct cgroup *oldcgrp = cgroup_taskset_cur_cgroup(tset);
1396 struct cpuset *cs = cgroup_cs(cgrp);
1397 struct cpuset *oldcs = cgroup_cs(oldcgrp);
1399 mutex_lock(&cpuset_mutex);
1401 /* prepare for attach */
1402 if (cs == &top_cpuset)
1403 cpumask_copy(cpus_attach, cpu_possible_mask);
1405 guarantee_online_cpus(cs, cpus_attach);
1407 guarantee_online_mems(cs, &cpuset_attach_nodemask_to);
1409 cgroup_taskset_for_each(task, cgrp, tset) {
1411 * can_attach beforehand should guarantee that this doesn't
1412 * fail. TODO: have a better way to handle failure here
1414 WARN_ON_ONCE(set_cpus_allowed_ptr(task, cpus_attach));
1416 cpuset_change_task_nodemask(task, &cpuset_attach_nodemask_to);
1417 cpuset_update_task_spread_flag(cs, task);
1421 * Change mm, possibly for multiple threads in a threadgroup. This is
1422 * expensive and may sleep.
1424 cpuset_attach_nodemask_to = cs->mems_allowed;
1425 mm = get_task_mm(leader);
1427 mpol_rebind_mm(mm, &cpuset_attach_nodemask_to);
1428 if (is_memory_migrate(cs))
1429 cpuset_migrate_mm(mm, &oldcs->mems_allowed,
1430 &cpuset_attach_nodemask_to);
1434 cs->attach_in_progress--;
1435 if (!cs->attach_in_progress)
1436 wake_up(&cpuset_attach_wq);
1438 mutex_unlock(&cpuset_mutex);
1441 /* The various types of files and directories in a cpuset file system */
1444 FILE_MEMORY_MIGRATE,
1450 FILE_SCHED_LOAD_BALANCE,
1451 FILE_SCHED_RELAX_DOMAIN_LEVEL,
1452 FILE_MEMORY_PRESSURE_ENABLED,
1453 FILE_MEMORY_PRESSURE,
1456 } cpuset_filetype_t;
1458 static int cpuset_write_u64(struct cgroup *cgrp, struct cftype *cft, u64 val)
1460 struct cpuset *cs = cgroup_cs(cgrp);
1461 cpuset_filetype_t type = cft->private;
1462 int retval = -ENODEV;
1464 mutex_lock(&cpuset_mutex);
1465 if (!is_cpuset_online(cs))
1469 case FILE_CPU_EXCLUSIVE:
1470 retval = update_flag(CS_CPU_EXCLUSIVE, cs, val);
1472 case FILE_MEM_EXCLUSIVE:
1473 retval = update_flag(CS_MEM_EXCLUSIVE, cs, val);
1475 case FILE_MEM_HARDWALL:
1476 retval = update_flag(CS_MEM_HARDWALL, cs, val);
1478 case FILE_SCHED_LOAD_BALANCE:
1479 retval = update_flag(CS_SCHED_LOAD_BALANCE, cs, val);
1481 case FILE_MEMORY_MIGRATE:
1482 retval = update_flag(CS_MEMORY_MIGRATE, cs, val);
1484 case FILE_MEMORY_PRESSURE_ENABLED:
1485 cpuset_memory_pressure_enabled = !!val;
1487 case FILE_MEMORY_PRESSURE:
1490 case FILE_SPREAD_PAGE:
1491 retval = update_flag(CS_SPREAD_PAGE, cs, val);
1493 case FILE_SPREAD_SLAB:
1494 retval = update_flag(CS_SPREAD_SLAB, cs, val);
1501 mutex_unlock(&cpuset_mutex);
1505 static int cpuset_write_s64(struct cgroup *cgrp, struct cftype *cft, s64 val)
1507 struct cpuset *cs = cgroup_cs(cgrp);
1508 cpuset_filetype_t type = cft->private;
1509 int retval = -ENODEV;
1511 mutex_lock(&cpuset_mutex);
1512 if (!is_cpuset_online(cs))
1516 case FILE_SCHED_RELAX_DOMAIN_LEVEL:
1517 retval = update_relax_domain_level(cs, val);
1524 mutex_unlock(&cpuset_mutex);
1529 * Common handling for a write to a "cpus" or "mems" file.
1531 static int cpuset_write_resmask(struct cgroup *cgrp, struct cftype *cft,
1534 struct cpuset *cs = cgroup_cs(cgrp);
1535 struct cpuset *trialcs;
1536 int retval = -ENODEV;
1539 * CPU or memory hotunplug may leave @cs w/o any execution
1540 * resources, in which case the hotplug code asynchronously updates
1541 * configuration and transfers all tasks to the nearest ancestor
1542 * which can execute.
1544 * As writes to "cpus" or "mems" may restore @cs's execution
1545 * resources, wait for the previously scheduled operations before
1546 * proceeding, so that we don't end up keep removing tasks added
1547 * after execution capability is restored.
1549 flush_work(&cpuset_hotplug_work);
1551 mutex_lock(&cpuset_mutex);
1552 if (!is_cpuset_online(cs))
1555 trialcs = alloc_trial_cpuset(cs);
1561 switch (cft->private) {
1563 retval = update_cpumask(cs, trialcs, buf);
1566 retval = update_nodemask(cs, trialcs, buf);
1573 free_trial_cpuset(trialcs);
1575 mutex_unlock(&cpuset_mutex);
1580 * These ascii lists should be read in a single call, by using a user
1581 * buffer large enough to hold the entire map. If read in smaller
1582 * chunks, there is no guarantee of atomicity. Since the display format
1583 * used, list of ranges of sequential numbers, is variable length,
1584 * and since these maps can change value dynamically, one could read
1585 * gibberish by doing partial reads while a list was changing.
1586 * A single large read to a buffer that crosses a page boundary is
1587 * ok, because the result being copied to user land is not recomputed
1588 * across a page fault.
1591 static size_t cpuset_sprintf_cpulist(char *page, struct cpuset *cs)
1595 mutex_lock(&callback_mutex);
1596 count = cpulist_scnprintf(page, PAGE_SIZE, cs->cpus_allowed);
1597 mutex_unlock(&callback_mutex);
1602 static size_t cpuset_sprintf_memlist(char *page, struct cpuset *cs)
1606 mutex_lock(&callback_mutex);
1607 count = nodelist_scnprintf(page, PAGE_SIZE, cs->mems_allowed);
1608 mutex_unlock(&callback_mutex);
1613 static ssize_t cpuset_common_file_read(struct cgroup *cont,
1617 size_t nbytes, loff_t *ppos)
1619 struct cpuset *cs = cgroup_cs(cont);
1620 cpuset_filetype_t type = cft->private;
1625 if (!(page = (char *)__get_free_page(GFP_TEMPORARY)))
1632 s += cpuset_sprintf_cpulist(s, cs);
1635 s += cpuset_sprintf_memlist(s, cs);
1643 retval = simple_read_from_buffer(buf, nbytes, ppos, page, s - page);
1645 free_page((unsigned long)page);
1649 static u64 cpuset_read_u64(struct cgroup *cont, struct cftype *cft)
1651 struct cpuset *cs = cgroup_cs(cont);
1652 cpuset_filetype_t type = cft->private;
1654 case FILE_CPU_EXCLUSIVE:
1655 return is_cpu_exclusive(cs);
1656 case FILE_MEM_EXCLUSIVE:
1657 return is_mem_exclusive(cs);
1658 case FILE_MEM_HARDWALL:
1659 return is_mem_hardwall(cs);
1660 case FILE_SCHED_LOAD_BALANCE:
1661 return is_sched_load_balance(cs);
1662 case FILE_MEMORY_MIGRATE:
1663 return is_memory_migrate(cs);
1664 case FILE_MEMORY_PRESSURE_ENABLED:
1665 return cpuset_memory_pressure_enabled;
1666 case FILE_MEMORY_PRESSURE:
1667 return fmeter_getrate(&cs->fmeter);
1668 case FILE_SPREAD_PAGE:
1669 return is_spread_page(cs);
1670 case FILE_SPREAD_SLAB:
1671 return is_spread_slab(cs);
1676 /* Unreachable but makes gcc happy */
1680 static s64 cpuset_read_s64(struct cgroup *cont, struct cftype *cft)
1682 struct cpuset *cs = cgroup_cs(cont);
1683 cpuset_filetype_t type = cft->private;
1685 case FILE_SCHED_RELAX_DOMAIN_LEVEL:
1686 return cs->relax_domain_level;
1691 /* Unrechable but makes gcc happy */
1697 * for the common functions, 'private' gives the type of file
1700 static struct cftype files[] = {
1703 .read = cpuset_common_file_read,
1704 .write_string = cpuset_write_resmask,
1705 .max_write_len = (100U + 6 * NR_CPUS),
1706 .private = FILE_CPULIST,
1711 .read = cpuset_common_file_read,
1712 .write_string = cpuset_write_resmask,
1713 .max_write_len = (100U + 6 * MAX_NUMNODES),
1714 .private = FILE_MEMLIST,
1718 .name = "cpu_exclusive",
1719 .read_u64 = cpuset_read_u64,
1720 .write_u64 = cpuset_write_u64,
1721 .private = FILE_CPU_EXCLUSIVE,
1725 .name = "mem_exclusive",
1726 .read_u64 = cpuset_read_u64,
1727 .write_u64 = cpuset_write_u64,
1728 .private = FILE_MEM_EXCLUSIVE,
1732 .name = "mem_hardwall",
1733 .read_u64 = cpuset_read_u64,
1734 .write_u64 = cpuset_write_u64,
1735 .private = FILE_MEM_HARDWALL,
1739 .name = "sched_load_balance",
1740 .read_u64 = cpuset_read_u64,
1741 .write_u64 = cpuset_write_u64,
1742 .private = FILE_SCHED_LOAD_BALANCE,
1746 .name = "sched_relax_domain_level",
1747 .read_s64 = cpuset_read_s64,
1748 .write_s64 = cpuset_write_s64,
1749 .private = FILE_SCHED_RELAX_DOMAIN_LEVEL,
1753 .name = "memory_migrate",
1754 .read_u64 = cpuset_read_u64,
1755 .write_u64 = cpuset_write_u64,
1756 .private = FILE_MEMORY_MIGRATE,
1760 .name = "memory_pressure",
1761 .read_u64 = cpuset_read_u64,
1762 .write_u64 = cpuset_write_u64,
1763 .private = FILE_MEMORY_PRESSURE,
1768 .name = "memory_spread_page",
1769 .read_u64 = cpuset_read_u64,
1770 .write_u64 = cpuset_write_u64,
1771 .private = FILE_SPREAD_PAGE,
1775 .name = "memory_spread_slab",
1776 .read_u64 = cpuset_read_u64,
1777 .write_u64 = cpuset_write_u64,
1778 .private = FILE_SPREAD_SLAB,
1782 .name = "memory_pressure_enabled",
1783 .flags = CFTYPE_ONLY_ON_ROOT,
1784 .read_u64 = cpuset_read_u64,
1785 .write_u64 = cpuset_write_u64,
1786 .private = FILE_MEMORY_PRESSURE_ENABLED,
1793 * cpuset_css_alloc - allocate a cpuset css
1794 * cont: control group that the new cpuset will be part of
1797 static struct cgroup_subsys_state *cpuset_css_alloc(struct cgroup *cont)
1802 return &top_cpuset.css;
1804 cs = kzalloc(sizeof(*cs), GFP_KERNEL);
1806 return ERR_PTR(-ENOMEM);
1807 if (!alloc_cpumask_var(&cs->cpus_allowed, GFP_KERNEL)) {
1809 return ERR_PTR(-ENOMEM);
1812 set_bit(CS_SCHED_LOAD_BALANCE, &cs->flags);
1813 cpumask_clear(cs->cpus_allowed);
1814 nodes_clear(cs->mems_allowed);
1815 fmeter_init(&cs->fmeter);
1816 cs->relax_domain_level = -1;
1821 static int cpuset_css_online(struct cgroup *cgrp)
1823 struct cpuset *cs = cgroup_cs(cgrp);
1824 struct cpuset *parent = parent_cs(cs);
1825 struct cpuset *tmp_cs;
1826 struct cgroup *pos_cg;
1831 mutex_lock(&cpuset_mutex);
1833 set_bit(CS_ONLINE, &cs->flags);
1834 if (is_spread_page(parent))
1835 set_bit(CS_SPREAD_PAGE, &cs->flags);
1836 if (is_spread_slab(parent))
1837 set_bit(CS_SPREAD_SLAB, &cs->flags);
1839 number_of_cpusets++;
1841 if (!test_bit(CGRP_CPUSET_CLONE_CHILDREN, &cgrp->flags))
1845 * Clone @parent's configuration if CGRP_CPUSET_CLONE_CHILDREN is
1846 * set. This flag handling is implemented in cgroup core for
1847 * histrical reasons - the flag may be specified during mount.
1849 * Currently, if any sibling cpusets have exclusive cpus or mem, we
1850 * refuse to clone the configuration - thereby refusing the task to
1851 * be entered, and as a result refusing the sys_unshare() or
1852 * clone() which initiated it. If this becomes a problem for some
1853 * users who wish to allow that scenario, then this could be
1854 * changed to grant parent->cpus_allowed-sibling_cpus_exclusive
1855 * (and likewise for mems) to the new cgroup.
1858 cpuset_for_each_child(tmp_cs, pos_cg, parent) {
1859 if (is_mem_exclusive(tmp_cs) || is_cpu_exclusive(tmp_cs)) {
1866 mutex_lock(&callback_mutex);
1867 cs->mems_allowed = parent->mems_allowed;
1868 cpumask_copy(cs->cpus_allowed, parent->cpus_allowed);
1869 mutex_unlock(&callback_mutex);
1871 mutex_unlock(&cpuset_mutex);
1875 static void cpuset_css_offline(struct cgroup *cgrp)
1877 struct cpuset *cs = cgroup_cs(cgrp);
1879 mutex_lock(&cpuset_mutex);
1881 if (is_sched_load_balance(cs))
1882 update_flag(CS_SCHED_LOAD_BALANCE, cs, 0);
1884 number_of_cpusets--;
1885 clear_bit(CS_ONLINE, &cs->flags);
1887 mutex_unlock(&cpuset_mutex);
1891 * If the cpuset being removed has its flag 'sched_load_balance'
1892 * enabled, then simulate turning sched_load_balance off, which
1893 * will call rebuild_sched_domains_locked().
1896 static void cpuset_css_free(struct cgroup *cont)
1898 struct cpuset *cs = cgroup_cs(cont);
1900 free_cpumask_var(cs->cpus_allowed);
1904 struct cgroup_subsys cpuset_subsys = {
1906 .css_alloc = cpuset_css_alloc,
1907 .css_online = cpuset_css_online,
1908 .css_offline = cpuset_css_offline,
1909 .css_free = cpuset_css_free,
1910 .can_attach = cpuset_can_attach,
1911 .cancel_attach = cpuset_cancel_attach,
1912 .attach = cpuset_attach,
1913 .subsys_id = cpuset_subsys_id,
1914 .base_cftypes = files,
1919 * cpuset_init - initialize cpusets at system boot
1921 * Description: Initialize top_cpuset and the cpuset internal file system,
1924 int __init cpuset_init(void)
1928 if (!alloc_cpumask_var(&top_cpuset.cpus_allowed, GFP_KERNEL))
1931 cpumask_setall(top_cpuset.cpus_allowed);
1932 nodes_setall(top_cpuset.mems_allowed);
1934 fmeter_init(&top_cpuset.fmeter);
1935 set_bit(CS_SCHED_LOAD_BALANCE, &top_cpuset.flags);
1936 top_cpuset.relax_domain_level = -1;
1938 err = register_filesystem(&cpuset_fs_type);
1942 if (!alloc_cpumask_var(&cpus_attach, GFP_KERNEL))
1945 number_of_cpusets = 1;
1950 * If CPU and/or memory hotplug handlers, below, unplug any CPUs
1951 * or memory nodes, we need to walk over the cpuset hierarchy,
1952 * removing that CPU or node from all cpusets. If this removes the
1953 * last CPU or node from a cpuset, then move the tasks in the empty
1954 * cpuset to its next-highest non-empty parent.
1956 static void remove_tasks_in_empty_cpuset(struct cpuset *cs)
1958 struct cpuset *parent;
1961 * Find its next-highest non-empty parent, (top cpuset
1962 * has online cpus, so can't be empty).
1964 parent = parent_cs(cs);
1965 while (cpumask_empty(parent->cpus_allowed) ||
1966 nodes_empty(parent->mems_allowed))
1967 parent = parent_cs(parent);
1969 if (cgroup_transfer_tasks(parent->css.cgroup, cs->css.cgroup)) {
1971 printk(KERN_ERR "cpuset: failed to transfer tasks out of empty cpuset %s\n",
1972 cgroup_name(cs->css.cgroup));
1978 * cpuset_hotplug_update_tasks - update tasks in a cpuset for hotunplug
1979 * @cs: cpuset in interest
1981 * Compare @cs's cpu and mem masks against top_cpuset and if some have gone
1982 * offline, update @cs accordingly. If @cs ends up with no CPU or memory,
1983 * all its tasks are moved to the nearest ancestor with both resources.
1985 static void cpuset_hotplug_update_tasks(struct cpuset *cs)
1987 static cpumask_t off_cpus;
1988 static nodemask_t off_mems, tmp_mems;
1992 wait_event(cpuset_attach_wq, cs->attach_in_progress == 0);
1994 mutex_lock(&cpuset_mutex);
1997 * We have raced with task attaching. We wait until attaching
1998 * is finished, so we won't attach a task to an empty cpuset.
2000 if (cs->attach_in_progress) {
2001 mutex_unlock(&cpuset_mutex);
2005 cpumask_andnot(&off_cpus, cs->cpus_allowed, top_cpuset.cpus_allowed);
2006 nodes_andnot(off_mems, cs->mems_allowed, top_cpuset.mems_allowed);
2008 /* remove offline cpus from @cs */
2009 if (!cpumask_empty(&off_cpus)) {
2010 mutex_lock(&callback_mutex);
2011 cpumask_andnot(cs->cpus_allowed, cs->cpus_allowed, &off_cpus);
2012 mutex_unlock(&callback_mutex);
2013 update_tasks_cpumask(cs, NULL);
2016 /* remove offline mems from @cs */
2017 if (!nodes_empty(off_mems)) {
2018 tmp_mems = cs->mems_allowed;
2019 mutex_lock(&callback_mutex);
2020 nodes_andnot(cs->mems_allowed, cs->mems_allowed, off_mems);
2021 mutex_unlock(&callback_mutex);
2022 update_tasks_nodemask(cs, &tmp_mems, NULL);
2025 is_empty = cpumask_empty(cs->cpus_allowed) ||
2026 nodes_empty(cs->mems_allowed);
2028 mutex_unlock(&cpuset_mutex);
2031 * If @cs became empty, move tasks to the nearest ancestor with
2032 * execution resources. This is full cgroup operation which will
2033 * also call back into cpuset. Should be done outside any lock.
2036 remove_tasks_in_empty_cpuset(cs);
2040 * cpuset_hotplug_workfn - handle CPU/memory hotunplug for a cpuset
2042 * This function is called after either CPU or memory configuration has
2043 * changed and updates cpuset accordingly. The top_cpuset is always
2044 * synchronized to cpu_active_mask and N_MEMORY, which is necessary in
2045 * order to make cpusets transparent (of no affect) on systems that are
2046 * actively using CPU hotplug but making no active use of cpusets.
2048 * Non-root cpusets are only affected by offlining. If any CPUs or memory
2049 * nodes have been taken down, cpuset_hotplug_update_tasks() is invoked on
2052 * Note that CPU offlining during suspend is ignored. We don't modify
2053 * cpusets across suspend/resume cycles at all.
2055 static void cpuset_hotplug_workfn(struct work_struct *work)
2057 static cpumask_t new_cpus, tmp_cpus;
2058 static nodemask_t new_mems, tmp_mems;
2059 bool cpus_updated, mems_updated;
2060 bool cpus_offlined, mems_offlined;
2062 mutex_lock(&cpuset_mutex);
2064 /* fetch the available cpus/mems and find out which changed how */
2065 cpumask_copy(&new_cpus, cpu_active_mask);
2066 new_mems = node_states[N_MEMORY];
2068 cpus_updated = !cpumask_equal(top_cpuset.cpus_allowed, &new_cpus);
2069 cpus_offlined = cpumask_andnot(&tmp_cpus, top_cpuset.cpus_allowed,
2072 mems_updated = !nodes_equal(top_cpuset.mems_allowed, new_mems);
2073 nodes_andnot(tmp_mems, top_cpuset.mems_allowed, new_mems);
2074 mems_offlined = !nodes_empty(tmp_mems);
2076 /* synchronize cpus_allowed to cpu_active_mask */
2078 mutex_lock(&callback_mutex);
2079 cpumask_copy(top_cpuset.cpus_allowed, &new_cpus);
2080 mutex_unlock(&callback_mutex);
2081 /* we don't mess with cpumasks of tasks in top_cpuset */
2084 /* synchronize mems_allowed to N_MEMORY */
2086 tmp_mems = top_cpuset.mems_allowed;
2087 mutex_lock(&callback_mutex);
2088 top_cpuset.mems_allowed = new_mems;
2089 mutex_unlock(&callback_mutex);
2090 update_tasks_nodemask(&top_cpuset, &tmp_mems, NULL);
2093 mutex_unlock(&cpuset_mutex);
2095 /* if cpus or mems went down, we need to propagate to descendants */
2096 if (cpus_offlined || mems_offlined) {
2098 struct cgroup *pos_cgrp;
2101 cpuset_for_each_descendant_pre(cs, pos_cgrp, &top_cpuset) {
2102 if (!css_tryget(&cs->css))
2106 cpuset_hotplug_update_tasks(cs);
2114 /* rebuild sched domains if cpus_allowed has changed */
2116 rebuild_sched_domains();
2119 void cpuset_update_active_cpus(bool cpu_online)
2122 * We're inside cpu hotplug critical region which usually nests
2123 * inside cgroup synchronization. Bounce actual hotplug processing
2124 * to a work item to avoid reverse locking order.
2126 * We still need to do partition_sched_domains() synchronously;
2127 * otherwise, the scheduler will get confused and put tasks to the
2128 * dead CPU. Fall back to the default single domain.
2129 * cpuset_hotplug_workfn() will rebuild it as necessary.
2131 partition_sched_domains(1, NULL, NULL);
2132 schedule_work(&cpuset_hotplug_work);
2136 * Keep top_cpuset.mems_allowed tracking node_states[N_MEMORY].
2137 * Call this routine anytime after node_states[N_MEMORY] changes.
2138 * See cpuset_update_active_cpus() for CPU hotplug handling.
2140 static int cpuset_track_online_nodes(struct notifier_block *self,
2141 unsigned long action, void *arg)
2143 schedule_work(&cpuset_hotplug_work);
2147 static struct notifier_block cpuset_track_online_nodes_nb = {
2148 .notifier_call = cpuset_track_online_nodes,
2149 .priority = 10, /* ??! */
2153 * cpuset_init_smp - initialize cpus_allowed
2155 * Description: Finish top cpuset after cpu, node maps are initialized
2157 void __init cpuset_init_smp(void)
2159 cpumask_copy(top_cpuset.cpus_allowed, cpu_active_mask);
2160 top_cpuset.mems_allowed = node_states[N_MEMORY];
2162 register_hotmemory_notifier(&cpuset_track_online_nodes_nb);
2166 * cpuset_cpus_allowed - return cpus_allowed mask from a tasks cpuset.
2167 * @tsk: pointer to task_struct from which to obtain cpuset->cpus_allowed.
2168 * @pmask: pointer to struct cpumask variable to receive cpus_allowed set.
2170 * Description: Returns the cpumask_var_t cpus_allowed of the cpuset
2171 * attached to the specified @tsk. Guaranteed to return some non-empty
2172 * subset of cpu_online_mask, even if this means going outside the
2176 void cpuset_cpus_allowed(struct task_struct *tsk, struct cpumask *pmask)
2178 mutex_lock(&callback_mutex);
2180 guarantee_online_cpus(task_cs(tsk), pmask);
2182 mutex_unlock(&callback_mutex);
2185 void cpuset_cpus_allowed_fallback(struct task_struct *tsk)
2187 const struct cpuset *cs;
2191 do_set_cpus_allowed(tsk, cs->cpus_allowed);
2195 * We own tsk->cpus_allowed, nobody can change it under us.
2197 * But we used cs && cs->cpus_allowed lockless and thus can
2198 * race with cgroup_attach_task() or update_cpumask() and get
2199 * the wrong tsk->cpus_allowed. However, both cases imply the
2200 * subsequent cpuset_change_cpumask()->set_cpus_allowed_ptr()
2201 * which takes task_rq_lock().
2203 * If we are called after it dropped the lock we must see all
2204 * changes in tsk_cs()->cpus_allowed. Otherwise we can temporary
2205 * set any mask even if it is not right from task_cs() pov,
2206 * the pending set_cpus_allowed_ptr() will fix things.
2208 * select_fallback_rq() will fix things ups and set cpu_possible_mask
2213 void cpuset_init_current_mems_allowed(void)
2215 nodes_setall(current->mems_allowed);
2219 * cpuset_mems_allowed - return mems_allowed mask from a tasks cpuset.
2220 * @tsk: pointer to task_struct from which to obtain cpuset->mems_allowed.
2222 * Description: Returns the nodemask_t mems_allowed of the cpuset
2223 * attached to the specified @tsk. Guaranteed to return some non-empty
2224 * subset of node_states[N_MEMORY], even if this means going outside the
2228 nodemask_t cpuset_mems_allowed(struct task_struct *tsk)
2232 mutex_lock(&callback_mutex);
2234 guarantee_online_mems(task_cs(tsk), &mask);
2236 mutex_unlock(&callback_mutex);
2242 * cpuset_nodemask_valid_mems_allowed - check nodemask vs. curremt mems_allowed
2243 * @nodemask: the nodemask to be checked
2245 * Are any of the nodes in the nodemask allowed in current->mems_allowed?
2247 int cpuset_nodemask_valid_mems_allowed(nodemask_t *nodemask)
2249 return nodes_intersects(*nodemask, current->mems_allowed);
2253 * nearest_hardwall_ancestor() - Returns the nearest mem_exclusive or
2254 * mem_hardwall ancestor to the specified cpuset. Call holding
2255 * callback_mutex. If no ancestor is mem_exclusive or mem_hardwall
2256 * (an unusual configuration), then returns the root cpuset.
2258 static const struct cpuset *nearest_hardwall_ancestor(const struct cpuset *cs)
2260 while (!(is_mem_exclusive(cs) || is_mem_hardwall(cs)) && parent_cs(cs))
2266 * cpuset_node_allowed_softwall - Can we allocate on a memory node?
2267 * @node: is this an allowed node?
2268 * @gfp_mask: memory allocation flags
2270 * If we're in interrupt, yes, we can always allocate. If __GFP_THISNODE is
2271 * set, yes, we can always allocate. If node is in our task's mems_allowed,
2272 * yes. If it's not a __GFP_HARDWALL request and this node is in the nearest
2273 * hardwalled cpuset ancestor to this task's cpuset, yes. If the task has been
2274 * OOM killed and has access to memory reserves as specified by the TIF_MEMDIE
2278 * If __GFP_HARDWALL is set, cpuset_node_allowed_softwall() reduces to
2279 * cpuset_node_allowed_hardwall(). Otherwise, cpuset_node_allowed_softwall()
2280 * might sleep, and might allow a node from an enclosing cpuset.
2282 * cpuset_node_allowed_hardwall() only handles the simpler case of hardwall
2283 * cpusets, and never sleeps.
2285 * The __GFP_THISNODE placement logic is really handled elsewhere,
2286 * by forcibly using a zonelist starting at a specified node, and by
2287 * (in get_page_from_freelist()) refusing to consider the zones for
2288 * any node on the zonelist except the first. By the time any such
2289 * calls get to this routine, we should just shut up and say 'yes'.
2291 * GFP_USER allocations are marked with the __GFP_HARDWALL bit,
2292 * and do not allow allocations outside the current tasks cpuset
2293 * unless the task has been OOM killed as is marked TIF_MEMDIE.
2294 * GFP_KERNEL allocations are not so marked, so can escape to the
2295 * nearest enclosing hardwalled ancestor cpuset.
2297 * Scanning up parent cpusets requires callback_mutex. The
2298 * __alloc_pages() routine only calls here with __GFP_HARDWALL bit
2299 * _not_ set if it's a GFP_KERNEL allocation, and all nodes in the
2300 * current tasks mems_allowed came up empty on the first pass over
2301 * the zonelist. So only GFP_KERNEL allocations, if all nodes in the
2302 * cpuset are short of memory, might require taking the callback_mutex
2305 * The first call here from mm/page_alloc:get_page_from_freelist()
2306 * has __GFP_HARDWALL set in gfp_mask, enforcing hardwall cpusets,
2307 * so no allocation on a node outside the cpuset is allowed (unless
2308 * in interrupt, of course).
2310 * The second pass through get_page_from_freelist() doesn't even call
2311 * here for GFP_ATOMIC calls. For those calls, the __alloc_pages()
2312 * variable 'wait' is not set, and the bit ALLOC_CPUSET is not set
2313 * in alloc_flags. That logic and the checks below have the combined
2315 * in_interrupt - any node ok (current task context irrelevant)
2316 * GFP_ATOMIC - any node ok
2317 * TIF_MEMDIE - any node ok
2318 * GFP_KERNEL - any node in enclosing hardwalled cpuset ok
2319 * GFP_USER - only nodes in current tasks mems allowed ok.
2322 * Don't call cpuset_node_allowed_softwall if you can't sleep, unless you
2323 * pass in the __GFP_HARDWALL flag set in gfp_flag, which disables
2324 * the code that might scan up ancestor cpusets and sleep.
2326 int __cpuset_node_allowed_softwall(int node, gfp_t gfp_mask)
2328 const struct cpuset *cs; /* current cpuset ancestors */
2329 int allowed; /* is allocation in zone z allowed? */
2331 if (in_interrupt() || (gfp_mask & __GFP_THISNODE))
2333 might_sleep_if(!(gfp_mask & __GFP_HARDWALL));
2334 if (node_isset(node, current->mems_allowed))
2337 * Allow tasks that have access to memory reserves because they have
2338 * been OOM killed to get memory anywhere.
2340 if (unlikely(test_thread_flag(TIF_MEMDIE)))
2342 if (gfp_mask & __GFP_HARDWALL) /* If hardwall request, stop here */
2345 if (current->flags & PF_EXITING) /* Let dying task have memory */
2348 /* Not hardwall and node outside mems_allowed: scan up cpusets */
2349 mutex_lock(&callback_mutex);
2352 cs = nearest_hardwall_ancestor(task_cs(current));
2353 task_unlock(current);
2355 allowed = node_isset(node, cs->mems_allowed);
2356 mutex_unlock(&callback_mutex);
2361 * cpuset_node_allowed_hardwall - Can we allocate on a memory node?
2362 * @node: is this an allowed node?
2363 * @gfp_mask: memory allocation flags
2365 * If we're in interrupt, yes, we can always allocate. If __GFP_THISNODE is
2366 * set, yes, we can always allocate. If node is in our task's mems_allowed,
2367 * yes. If the task has been OOM killed and has access to memory reserves as
2368 * specified by the TIF_MEMDIE flag, yes.
2371 * The __GFP_THISNODE placement logic is really handled elsewhere,
2372 * by forcibly using a zonelist starting at a specified node, and by
2373 * (in get_page_from_freelist()) refusing to consider the zones for
2374 * any node on the zonelist except the first. By the time any such
2375 * calls get to this routine, we should just shut up and say 'yes'.
2377 * Unlike the cpuset_node_allowed_softwall() variant, above,
2378 * this variant requires that the node be in the current task's
2379 * mems_allowed or that we're in interrupt. It does not scan up the
2380 * cpuset hierarchy for the nearest enclosing mem_exclusive cpuset.
2383 int __cpuset_node_allowed_hardwall(int node, gfp_t gfp_mask)
2385 if (in_interrupt() || (gfp_mask & __GFP_THISNODE))
2387 if (node_isset(node, current->mems_allowed))
2390 * Allow tasks that have access to memory reserves because they have
2391 * been OOM killed to get memory anywhere.
2393 if (unlikely(test_thread_flag(TIF_MEMDIE)))
2399 * cpuset_mem_spread_node() - On which node to begin search for a file page
2400 * cpuset_slab_spread_node() - On which node to begin search for a slab page
2402 * If a task is marked PF_SPREAD_PAGE or PF_SPREAD_SLAB (as for
2403 * tasks in a cpuset with is_spread_page or is_spread_slab set),
2404 * and if the memory allocation used cpuset_mem_spread_node()
2405 * to determine on which node to start looking, as it will for
2406 * certain page cache or slab cache pages such as used for file
2407 * system buffers and inode caches, then instead of starting on the
2408 * local node to look for a free page, rather spread the starting
2409 * node around the tasks mems_allowed nodes.
2411 * We don't have to worry about the returned node being offline
2412 * because "it can't happen", and even if it did, it would be ok.
2414 * The routines calling guarantee_online_mems() are careful to
2415 * only set nodes in task->mems_allowed that are online. So it
2416 * should not be possible for the following code to return an
2417 * offline node. But if it did, that would be ok, as this routine
2418 * is not returning the node where the allocation must be, only
2419 * the node where the search should start. The zonelist passed to
2420 * __alloc_pages() will include all nodes. If the slab allocator
2421 * is passed an offline node, it will fall back to the local node.
2422 * See kmem_cache_alloc_node().
2425 static int cpuset_spread_node(int *rotor)
2429 node = next_node(*rotor, current->mems_allowed);
2430 if (node == MAX_NUMNODES)
2431 node = first_node(current->mems_allowed);
2436 int cpuset_mem_spread_node(void)
2438 if (current->cpuset_mem_spread_rotor == NUMA_NO_NODE)
2439 current->cpuset_mem_spread_rotor =
2440 node_random(¤t->mems_allowed);
2442 return cpuset_spread_node(¤t->cpuset_mem_spread_rotor);
2445 int cpuset_slab_spread_node(void)
2447 if (current->cpuset_slab_spread_rotor == NUMA_NO_NODE)
2448 current->cpuset_slab_spread_rotor =
2449 node_random(¤t->mems_allowed);
2451 return cpuset_spread_node(¤t->cpuset_slab_spread_rotor);
2454 EXPORT_SYMBOL_GPL(cpuset_mem_spread_node);
2457 * cpuset_mems_allowed_intersects - Does @tsk1's mems_allowed intersect @tsk2's?
2458 * @tsk1: pointer to task_struct of some task.
2459 * @tsk2: pointer to task_struct of some other task.
2461 * Description: Return true if @tsk1's mems_allowed intersects the
2462 * mems_allowed of @tsk2. Used by the OOM killer to determine if
2463 * one of the task's memory usage might impact the memory available
2467 int cpuset_mems_allowed_intersects(const struct task_struct *tsk1,
2468 const struct task_struct *tsk2)
2470 return nodes_intersects(tsk1->mems_allowed, tsk2->mems_allowed);
2473 #define CPUSET_NODELIST_LEN (256)
2476 * cpuset_print_task_mems_allowed - prints task's cpuset and mems_allowed
2477 * @task: pointer to task_struct of some task.
2479 * Description: Prints @task's name, cpuset name, and cached copy of its
2480 * mems_allowed to the kernel log. Must hold task_lock(task) to allow
2481 * dereferencing task_cs(task).
2483 void cpuset_print_task_mems_allowed(struct task_struct *tsk)
2485 /* Statically allocated to prevent using excess stack. */
2486 static char cpuset_nodelist[CPUSET_NODELIST_LEN];
2487 static DEFINE_SPINLOCK(cpuset_buffer_lock);
2489 struct cgroup *cgrp = task_cs(tsk)->css.cgroup;
2492 spin_lock(&cpuset_buffer_lock);
2494 nodelist_scnprintf(cpuset_nodelist, CPUSET_NODELIST_LEN,
2496 printk(KERN_INFO "%s cpuset=%s mems_allowed=%s\n",
2497 tsk->comm, cgroup_name(cgrp), cpuset_nodelist);
2499 spin_unlock(&cpuset_buffer_lock);
2504 * Collection of memory_pressure is suppressed unless
2505 * this flag is enabled by writing "1" to the special
2506 * cpuset file 'memory_pressure_enabled' in the root cpuset.
2509 int cpuset_memory_pressure_enabled __read_mostly;
2512 * cpuset_memory_pressure_bump - keep stats of per-cpuset reclaims.
2514 * Keep a running average of the rate of synchronous (direct)
2515 * page reclaim efforts initiated by tasks in each cpuset.
2517 * This represents the rate at which some task in the cpuset
2518 * ran low on memory on all nodes it was allowed to use, and
2519 * had to enter the kernels page reclaim code in an effort to
2520 * create more free memory by tossing clean pages or swapping
2521 * or writing dirty pages.
2523 * Display to user space in the per-cpuset read-only file
2524 * "memory_pressure". Value displayed is an integer
2525 * representing the recent rate of entry into the synchronous
2526 * (direct) page reclaim by any task attached to the cpuset.
2529 void __cpuset_memory_pressure_bump(void)
2532 fmeter_markevent(&task_cs(current)->fmeter);
2533 task_unlock(current);
2536 #ifdef CONFIG_PROC_PID_CPUSET
2538 * proc_cpuset_show()
2539 * - Print tasks cpuset path into seq_file.
2540 * - Used for /proc/<pid>/cpuset.
2541 * - No need to task_lock(tsk) on this tsk->cpuset reference, as it
2542 * doesn't really matter if tsk->cpuset changes after we read it,
2543 * and we take cpuset_mutex, keeping cpuset_attach() from changing it
2546 int proc_cpuset_show(struct seq_file *m, void *unused_v)
2549 struct task_struct *tsk;
2551 struct cgroup_subsys_state *css;
2555 buf = kmalloc(PAGE_SIZE, GFP_KERNEL);
2561 tsk = get_pid_task(pid, PIDTYPE_PID);
2566 css = task_subsys_state(tsk, cpuset_subsys_id);
2567 retval = cgroup_path(css->cgroup, buf, PAGE_SIZE);
2574 put_task_struct(tsk);
2580 #endif /* CONFIG_PROC_PID_CPUSET */
2582 /* Display task mems_allowed in /proc/<pid>/status file. */
2583 void cpuset_task_status_allowed(struct seq_file *m, struct task_struct *task)
2585 seq_printf(m, "Mems_allowed:\t");
2586 seq_nodemask(m, &task->mems_allowed);
2587 seq_printf(m, "\n");
2588 seq_printf(m, "Mems_allowed_list:\t");
2589 seq_nodemask_list(m, &task->mems_allowed);
2590 seq_printf(m, "\n");