Merge branch 'for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/rw/uml
[linux-2.6.git] / kernel / cpuset.c
blobe5657788feddfefaaed5f7ce3ce2ac26ca80a9c1
1 /*
2 * kernel/cpuset.c
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
18 * by Max Krasnyansky
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>
31 #include <linux/fs.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>
38 #include <linux/mm.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;
73 struct cpuset;
75 /* See "Frequency meter" comments, below. */
77 struct fmeter {
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 */
84 struct cpuset {
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() */
112 int pn;
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),
122 struct cpuset, css);
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),
129 struct cpuset, css);
132 static inline struct cpuset *parent_cs(const struct cpuset *cs)
134 struct cgroup *pcgrp = cs->css.cgroup->parent;
136 if (pcgrp)
137 return cgroup_cs(pcgrp);
138 return NULL;
141 #ifdef CONFIG_NUMA
142 static inline bool task_has_mempolicy(struct task_struct *task)
144 return task->mempolicy;
146 #else
147 static inline bool task_has_mempolicy(struct task_struct *task)
149 return false;
151 #endif
154 /* bits in struct cpuset flags field */
155 typedef enum {
156 CS_ONLINE,
157 CS_CPU_EXCLUSIVE,
158 CS_MEM_EXCLUSIVE,
159 CS_MEM_HARDWALL,
160 CS_MEMORY_MIGRATE,
161 CS_SCHED_LOAD_BALANCE,
162 CS_SPREAD_PAGE,
163 CS_SPREAD_SLAB,
164 } cpuset_flagbits_t;
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
253 * everyone else.
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
258 * __alloc_pages().
260 * If a task is only holding callback_mutex, then it has read-only
261 * access to cpusets.
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
265 * them.
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);
296 if (cgroup_fs) {
297 char mountopts[] =
298 "cpuset,noprefix,"
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);
304 return ret;
307 static struct file_system_type cpuset_fs_type = {
308 .name = "cpuset",
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))
327 cs = parent_cs(cs);
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]))
345 cs = parent_cs(cs);
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;
359 else
360 tsk->flags &= ~PF_SPREAD_PAGE;
361 if (is_spread_slab(cs))
362 tsk->flags |= PF_SPREAD_SLAB;
363 else
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);
392 if (!trial)
393 return NULL;
395 if (!alloc_cpumask_var(&trial->cpus_allowed, GFP_KERNEL)) {
396 kfree(trial);
397 return NULL;
399 cpumask_copy(trial->cpus_allowed, cs->cpus_allowed);
401 return trial;
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);
411 kfree(trial);
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
421 * cpuset_mutex held.
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)
436 struct cgroup *cgrp;
437 struct cpuset *c, *par;
438 int ret;
440 rcu_read_lock();
442 /* Each of our child cpusets must be a subset of us */
443 ret = -EBUSY;
444 cpuset_for_each_child(c, cgrp, cur)
445 if (!is_cpuset_subset(c, trial))
446 goto out;
448 /* Remaining checks don't apply to root cpuset */
449 ret = 0;
450 if (cur == &top_cpuset)
451 goto out;
453 par = parent_cs(cur);
455 /* We must be a subset of our parent cpuset */
456 ret = -EACCES;
457 if (!is_cpuset_subset(trial, par))
458 goto out;
461 * If either I or some sibling (!= me) is exclusive, we can't
462 * overlap
464 ret = -EINVAL;
465 cpuset_for_each_child(c, cgrp, par) {
466 if ((is_cpu_exclusive(trial) || is_cpu_exclusive(c)) &&
467 c != cur &&
468 cpumask_intersects(trial->cpus_allowed, c->cpus_allowed))
469 goto out;
470 if ((is_mem_exclusive(trial) || is_mem_exclusive(c)) &&
471 c != cur &&
472 nodes_intersects(trial->mems_allowed, c->mems_allowed))
473 goto out;
477 * Cpusets with tasks - existing or newly being attached - can't
478 * have empty cpus_allowed or mems_allowed.
480 ret = -ENOSPC;
481 if ((cgroup_task_count(cur->css.cgroup) || cur->attach_in_progress) &&
482 (cpumask_empty(trial->cpus_allowed) &&
483 nodes_empty(trial->mems_allowed)))
484 goto out;
486 ret = 0;
487 out:
488 rcu_read_unlock();
489 return ret;
492 #ifdef CONFIG_SMP
494 * Helper routine for generate_sched_domains().
495 * Do cpusets a, b have overlapping cpus_allowed masks?
497 static int cpusets_overlap(struct cpuset *a, struct cpuset *b)
499 return cpumask_intersects(a->cpus_allowed, b->cpus_allowed);
502 static void
503 update_domain_attr(struct sched_domain_attr *dattr, struct cpuset *c)
505 if (dattr->relax_domain_level < c->relax_domain_level)
506 dattr->relax_domain_level = c->relax_domain_level;
507 return;
510 static void update_domain_attr_tree(struct sched_domain_attr *dattr,
511 struct cpuset *root_cs)
513 struct cpuset *cp;
514 struct cgroup *pos_cgrp;
516 rcu_read_lock();
517 cpuset_for_each_descendant_pre(cp, pos_cgrp, root_cs) {
518 /* skip the whole subtree if @cp doesn't have any CPU */
519 if (cpumask_empty(cp->cpus_allowed)) {
520 pos_cgrp = cgroup_rightmost_descendant(pos_cgrp);
521 continue;
524 if (is_sched_load_balance(cp))
525 update_domain_attr(dattr, cp);
527 rcu_read_unlock();
531 * generate_sched_domains()
533 * This function builds a partial partition of the systems CPUs
534 * A 'partial partition' is a set of non-overlapping subsets whose
535 * union is a subset of that set.
536 * The output of this function needs to be passed to kernel/sched/core.c
537 * partition_sched_domains() routine, which will rebuild the scheduler's
538 * load balancing domains (sched domains) as specified by that partial
539 * partition.
541 * See "What is sched_load_balance" in Documentation/cgroups/cpusets.txt
542 * for a background explanation of this.
544 * Does not return errors, on the theory that the callers of this
545 * routine would rather not worry about failures to rebuild sched
546 * domains when operating in the severe memory shortage situations
547 * that could cause allocation failures below.
549 * Must be called with cpuset_mutex held.
551 * The three key local variables below are:
552 * q - a linked-list queue of cpuset pointers, used to implement a
553 * top-down scan of all cpusets. This scan loads a pointer
554 * to each cpuset marked is_sched_load_balance into the
555 * array 'csa'. For our purposes, rebuilding the schedulers
556 * sched domains, we can ignore !is_sched_load_balance cpusets.
557 * csa - (for CpuSet Array) Array of pointers to all the cpusets
558 * that need to be load balanced, for convenient iterative
559 * access by the subsequent code that finds the best partition,
560 * i.e the set of domains (subsets) of CPUs such that the
561 * cpus_allowed of every cpuset marked is_sched_load_balance
562 * is a subset of one of these domains, while there are as
563 * many such domains as possible, each as small as possible.
564 * doms - Conversion of 'csa' to an array of cpumasks, for passing to
565 * the kernel/sched/core.c routine partition_sched_domains() in a
566 * convenient format, that can be easily compared to the prior
567 * value to determine what partition elements (sched domains)
568 * were changed (added or removed.)
570 * Finding the best partition (set of domains):
571 * The triple nested loops below over i, j, k scan over the
572 * load balanced cpusets (using the array of cpuset pointers in
573 * csa[]) looking for pairs of cpusets that have overlapping
574 * cpus_allowed, but which don't have the same 'pn' partition
575 * number and gives them in the same partition number. It keeps
576 * looping on the 'restart' label until it can no longer find
577 * any such pairs.
579 * The union of the cpus_allowed masks from the set of
580 * all cpusets having the same 'pn' value then form the one
581 * element of the partition (one sched domain) to be passed to
582 * partition_sched_domains().
584 static int generate_sched_domains(cpumask_var_t **domains,
585 struct sched_domain_attr **attributes)
587 struct cpuset *cp; /* scans q */
588 struct cpuset **csa; /* array of all cpuset ptrs */
589 int csn; /* how many cpuset ptrs in csa so far */
590 int i, j, k; /* indices for partition finding loops */
591 cpumask_var_t *doms; /* resulting partition; i.e. sched domains */
592 struct sched_domain_attr *dattr; /* attributes for custom domains */
593 int ndoms = 0; /* number of sched domains in result */
594 int nslot; /* next empty doms[] struct cpumask slot */
595 struct cgroup *pos_cgrp;
597 doms = NULL;
598 dattr = NULL;
599 csa = NULL;
601 /* Special case for the 99% of systems with one, full, sched domain */
602 if (is_sched_load_balance(&top_cpuset)) {
603 ndoms = 1;
604 doms = alloc_sched_domains(ndoms);
605 if (!doms)
606 goto done;
608 dattr = kmalloc(sizeof(struct sched_domain_attr), GFP_KERNEL);
609 if (dattr) {
610 *dattr = SD_ATTR_INIT;
611 update_domain_attr_tree(dattr, &top_cpuset);
613 cpumask_copy(doms[0], top_cpuset.cpus_allowed);
615 goto done;
618 csa = kmalloc(number_of_cpusets * sizeof(cp), GFP_KERNEL);
619 if (!csa)
620 goto done;
621 csn = 0;
623 rcu_read_lock();
624 cpuset_for_each_descendant_pre(cp, pos_cgrp, &top_cpuset) {
626 * Continue traversing beyond @cp iff @cp has some CPUs and
627 * isn't load balancing. The former is obvious. The
628 * latter: All child cpusets contain a subset of the
629 * parent's cpus, so just skip them, and then we call
630 * update_domain_attr_tree() to calc relax_domain_level of
631 * the corresponding sched domain.
633 if (!cpumask_empty(cp->cpus_allowed) &&
634 !is_sched_load_balance(cp))
635 continue;
637 if (is_sched_load_balance(cp))
638 csa[csn++] = cp;
640 /* skip @cp's subtree */
641 pos_cgrp = cgroup_rightmost_descendant(pos_cgrp);
643 rcu_read_unlock();
645 for (i = 0; i < csn; i++)
646 csa[i]->pn = i;
647 ndoms = csn;
649 restart:
650 /* Find the best partition (set of sched domains) */
651 for (i = 0; i < csn; i++) {
652 struct cpuset *a = csa[i];
653 int apn = a->pn;
655 for (j = 0; j < csn; j++) {
656 struct cpuset *b = csa[j];
657 int bpn = b->pn;
659 if (apn != bpn && cpusets_overlap(a, b)) {
660 for (k = 0; k < csn; k++) {
661 struct cpuset *c = csa[k];
663 if (c->pn == bpn)
664 c->pn = apn;
666 ndoms--; /* one less element */
667 goto restart;
673 * Now we know how many domains to create.
674 * Convert <csn, csa> to <ndoms, doms> and populate cpu masks.
676 doms = alloc_sched_domains(ndoms);
677 if (!doms)
678 goto done;
681 * The rest of the code, including the scheduler, can deal with
682 * dattr==NULL case. No need to abort if alloc fails.
684 dattr = kmalloc(ndoms * sizeof(struct sched_domain_attr), GFP_KERNEL);
686 for (nslot = 0, i = 0; i < csn; i++) {
687 struct cpuset *a = csa[i];
688 struct cpumask *dp;
689 int apn = a->pn;
691 if (apn < 0) {
692 /* Skip completed partitions */
693 continue;
696 dp = doms[nslot];
698 if (nslot == ndoms) {
699 static int warnings = 10;
700 if (warnings) {
701 printk(KERN_WARNING
702 "rebuild_sched_domains confused:"
703 " nslot %d, ndoms %d, csn %d, i %d,"
704 " apn %d\n",
705 nslot, ndoms, csn, i, apn);
706 warnings--;
708 continue;
711 cpumask_clear(dp);
712 if (dattr)
713 *(dattr + nslot) = SD_ATTR_INIT;
714 for (j = i; j < csn; j++) {
715 struct cpuset *b = csa[j];
717 if (apn == b->pn) {
718 cpumask_or(dp, dp, b->cpus_allowed);
719 if (dattr)
720 update_domain_attr_tree(dattr + nslot, b);
722 /* Done with this partition */
723 b->pn = -1;
726 nslot++;
728 BUG_ON(nslot != ndoms);
730 done:
731 kfree(csa);
734 * Fallback to the default domain if kmalloc() failed.
735 * See comments in partition_sched_domains().
737 if (doms == NULL)
738 ndoms = 1;
740 *domains = doms;
741 *attributes = dattr;
742 return ndoms;
746 * Rebuild scheduler domains.
748 * If the flag 'sched_load_balance' of any cpuset with non-empty
749 * 'cpus' changes, or if the 'cpus' allowed changes in any cpuset
750 * which has that flag enabled, or if any cpuset with a non-empty
751 * 'cpus' is removed, then call this routine to rebuild the
752 * scheduler's dynamic sched domains.
754 * Call with cpuset_mutex held. Takes get_online_cpus().
756 static void rebuild_sched_domains_locked(void)
758 struct sched_domain_attr *attr;
759 cpumask_var_t *doms;
760 int ndoms;
762 lockdep_assert_held(&cpuset_mutex);
763 get_online_cpus();
766 * We have raced with CPU hotplug. Don't do anything to avoid
767 * passing doms with offlined cpu to partition_sched_domains().
768 * Anyways, hotplug work item will rebuild sched domains.
770 if (!cpumask_equal(top_cpuset.cpus_allowed, cpu_active_mask))
771 goto out;
773 /* Generate domain masks and attrs */
774 ndoms = generate_sched_domains(&doms, &attr);
776 /* Have scheduler rebuild the domains */
777 partition_sched_domains(ndoms, doms, attr);
778 out:
779 put_online_cpus();
781 #else /* !CONFIG_SMP */
782 static void rebuild_sched_domains_locked(void)
785 #endif /* CONFIG_SMP */
787 void rebuild_sched_domains(void)
789 mutex_lock(&cpuset_mutex);
790 rebuild_sched_domains_locked();
791 mutex_unlock(&cpuset_mutex);
795 * effective_cpumask_cpuset - return nearest ancestor with non-empty cpus
796 * @cs: the cpuset in interest
798 * A cpuset's effective cpumask is the cpumask of the nearest ancestor
799 * with non-empty cpus. We use effective cpumask whenever:
800 * - we update tasks' cpus_allowed. (they take on the ancestor's cpumask
801 * if the cpuset they reside in has no cpus)
802 * - we want to retrieve task_cs(tsk)'s cpus_allowed.
804 * Called with cpuset_mutex held. cpuset_cpus_allowed_fallback() is an
805 * exception. See comments there.
807 static struct cpuset *effective_cpumask_cpuset(struct cpuset *cs)
809 while (cpumask_empty(cs->cpus_allowed))
810 cs = parent_cs(cs);
811 return cs;
815 * effective_nodemask_cpuset - return nearest ancestor with non-empty mems
816 * @cs: the cpuset in interest
818 * A cpuset's effective nodemask is the nodemask of the nearest ancestor
819 * with non-empty memss. We use effective nodemask whenever:
820 * - we update tasks' mems_allowed. (they take on the ancestor's nodemask
821 * if the cpuset they reside in has no mems)
822 * - we want to retrieve task_cs(tsk)'s mems_allowed.
824 * Called with cpuset_mutex held.
826 static struct cpuset *effective_nodemask_cpuset(struct cpuset *cs)
828 while (nodes_empty(cs->mems_allowed))
829 cs = parent_cs(cs);
830 return cs;
834 * cpuset_change_cpumask - make a task's cpus_allowed the same as its cpuset's
835 * @tsk: task to test
836 * @scan: struct cgroup_scanner containing the cgroup of the task
838 * Called by cgroup_scan_tasks() for each task in a cgroup whose
839 * cpus_allowed mask needs to be changed.
841 * We don't need to re-check for the cgroup/cpuset membership, since we're
842 * holding cpuset_mutex at this point.
844 static void cpuset_change_cpumask(struct task_struct *tsk,
845 struct cgroup_scanner *scan)
847 struct cpuset *cpus_cs;
849 cpus_cs = effective_cpumask_cpuset(cgroup_cs(scan->cg));
850 set_cpus_allowed_ptr(tsk, cpus_cs->cpus_allowed);
854 * update_tasks_cpumask - Update the cpumasks of tasks in the cpuset.
855 * @cs: the cpuset in which each task's cpus_allowed mask needs to be changed
856 * @heap: if NULL, defer allocating heap memory to cgroup_scan_tasks()
858 * Called with cpuset_mutex held
860 * The cgroup_scan_tasks() function will scan all the tasks in a cgroup,
861 * calling callback functions for each.
863 * No return value. It's guaranteed that cgroup_scan_tasks() always returns 0
864 * if @heap != NULL.
866 static void update_tasks_cpumask(struct cpuset *cs, struct ptr_heap *heap)
868 struct cgroup_scanner scan;
870 scan.cg = cs->css.cgroup;
871 scan.test_task = NULL;
872 scan.process_task = cpuset_change_cpumask;
873 scan.heap = heap;
874 cgroup_scan_tasks(&scan);
878 * update_tasks_cpumask_hier - Update the cpumasks of tasks in the hierarchy.
879 * @root_cs: the root cpuset of the hierarchy
880 * @update_root: update root cpuset or not?
881 * @heap: the heap used by cgroup_scan_tasks()
883 * This will update cpumasks of tasks in @root_cs and all other empty cpusets
884 * which take on cpumask of @root_cs.
886 * Called with cpuset_mutex held
888 static void update_tasks_cpumask_hier(struct cpuset *root_cs,
889 bool update_root, struct ptr_heap *heap)
891 struct cpuset *cp;
892 struct cgroup *pos_cgrp;
894 if (update_root)
895 update_tasks_cpumask(root_cs, heap);
897 rcu_read_lock();
898 cpuset_for_each_descendant_pre(cp, pos_cgrp, root_cs) {
899 /* skip the whole subtree if @cp have some CPU */
900 if (!cpumask_empty(cp->cpus_allowed)) {
901 pos_cgrp = cgroup_rightmost_descendant(pos_cgrp);
902 continue;
904 if (!css_tryget(&cp->css))
905 continue;
906 rcu_read_unlock();
908 update_tasks_cpumask(cp, heap);
910 rcu_read_lock();
911 css_put(&cp->css);
913 rcu_read_unlock();
917 * update_cpumask - update the cpus_allowed mask of a cpuset and all tasks in it
918 * @cs: the cpuset to consider
919 * @buf: buffer of cpu numbers written to this cpuset
921 static int update_cpumask(struct cpuset *cs, struct cpuset *trialcs,
922 const char *buf)
924 struct ptr_heap heap;
925 int retval;
926 int is_load_balanced;
928 /* top_cpuset.cpus_allowed tracks cpu_online_mask; it's read-only */
929 if (cs == &top_cpuset)
930 return -EACCES;
933 * An empty cpus_allowed is ok only if the cpuset has no tasks.
934 * Since cpulist_parse() fails on an empty mask, we special case
935 * that parsing. The validate_change() call ensures that cpusets
936 * with tasks have cpus.
938 if (!*buf) {
939 cpumask_clear(trialcs->cpus_allowed);
940 } else {
941 retval = cpulist_parse(buf, trialcs->cpus_allowed);
942 if (retval < 0)
943 return retval;
945 if (!cpumask_subset(trialcs->cpus_allowed, cpu_active_mask))
946 return -EINVAL;
949 /* Nothing to do if the cpus didn't change */
950 if (cpumask_equal(cs->cpus_allowed, trialcs->cpus_allowed))
951 return 0;
953 retval = validate_change(cs, trialcs);
954 if (retval < 0)
955 return retval;
957 retval = heap_init(&heap, PAGE_SIZE, GFP_KERNEL, NULL);
958 if (retval)
959 return retval;
961 is_load_balanced = is_sched_load_balance(trialcs);
963 mutex_lock(&callback_mutex);
964 cpumask_copy(cs->cpus_allowed, trialcs->cpus_allowed);
965 mutex_unlock(&callback_mutex);
967 update_tasks_cpumask_hier(cs, true, &heap);
969 heap_free(&heap);
971 if (is_load_balanced)
972 rebuild_sched_domains_locked();
973 return 0;
977 * cpuset_migrate_mm
979 * Migrate memory region from one set of nodes to another.
981 * Temporarilly set tasks mems_allowed to target nodes of migration,
982 * so that the migration code can allocate pages on these nodes.
984 * Call holding cpuset_mutex, so current's cpuset won't change
985 * during this call, as manage_mutex holds off any cpuset_attach()
986 * calls. Therefore we don't need to take task_lock around the
987 * call to guarantee_online_mems(), as we know no one is changing
988 * our task's cpuset.
990 * While the mm_struct we are migrating is typically from some
991 * other task, the task_struct mems_allowed that we are hacking
992 * is for our current task, which must allocate new pages for that
993 * migrating memory region.
996 static void cpuset_migrate_mm(struct mm_struct *mm, const nodemask_t *from,
997 const nodemask_t *to)
999 struct task_struct *tsk = current;
1000 struct cpuset *mems_cs;
1002 tsk->mems_allowed = *to;
1004 do_migrate_pages(mm, from, to, MPOL_MF_MOVE_ALL);
1006 mems_cs = effective_nodemask_cpuset(task_cs(tsk));
1007 guarantee_online_mems(mems_cs, &tsk->mems_allowed);
1011 * cpuset_change_task_nodemask - change task's mems_allowed and mempolicy
1012 * @tsk: the task to change
1013 * @newmems: new nodes that the task will be set
1015 * In order to avoid seeing no nodes if the old and new nodes are disjoint,
1016 * we structure updates as setting all new allowed nodes, then clearing newly
1017 * disallowed ones.
1019 static void cpuset_change_task_nodemask(struct task_struct *tsk,
1020 nodemask_t *newmems)
1022 bool need_loop;
1025 * Allow tasks that have access to memory reserves because they have
1026 * been OOM killed to get memory anywhere.
1028 if (unlikely(test_thread_flag(TIF_MEMDIE)))
1029 return;
1030 if (current->flags & PF_EXITING) /* Let dying task have memory */
1031 return;
1033 task_lock(tsk);
1035 * Determine if a loop is necessary if another thread is doing
1036 * get_mems_allowed(). If at least one node remains unchanged and
1037 * tsk does not have a mempolicy, then an empty nodemask will not be
1038 * possible when mems_allowed is larger than a word.
1040 need_loop = task_has_mempolicy(tsk) ||
1041 !nodes_intersects(*newmems, tsk->mems_allowed);
1043 if (need_loop)
1044 write_seqcount_begin(&tsk->mems_allowed_seq);
1046 nodes_or(tsk->mems_allowed, tsk->mems_allowed, *newmems);
1047 mpol_rebind_task(tsk, newmems, MPOL_REBIND_STEP1);
1049 mpol_rebind_task(tsk, newmems, MPOL_REBIND_STEP2);
1050 tsk->mems_allowed = *newmems;
1052 if (need_loop)
1053 write_seqcount_end(&tsk->mems_allowed_seq);
1055 task_unlock(tsk);
1059 * Update task's mems_allowed and rebind its mempolicy and vmas' mempolicy
1060 * of it to cpuset's new mems_allowed, and migrate pages to new nodes if
1061 * memory_migrate flag is set. Called with cpuset_mutex held.
1063 static void cpuset_change_nodemask(struct task_struct *p,
1064 struct cgroup_scanner *scan)
1066 struct cpuset *cs = cgroup_cs(scan->cg);
1067 struct mm_struct *mm;
1068 int migrate;
1069 nodemask_t *newmems = scan->data;
1071 cpuset_change_task_nodemask(p, newmems);
1073 mm = get_task_mm(p);
1074 if (!mm)
1075 return;
1077 migrate = is_memory_migrate(cs);
1079 mpol_rebind_mm(mm, &cs->mems_allowed);
1080 if (migrate)
1081 cpuset_migrate_mm(mm, &cs->old_mems_allowed, newmems);
1082 mmput(mm);
1085 static void *cpuset_being_rebound;
1088 * update_tasks_nodemask - Update the nodemasks of tasks in the cpuset.
1089 * @cs: the cpuset in which each task's mems_allowed mask needs to be changed
1090 * @heap: if NULL, defer allocating heap memory to cgroup_scan_tasks()
1092 * Called with cpuset_mutex held
1093 * No return value. It's guaranteed that cgroup_scan_tasks() always returns 0
1094 * if @heap != NULL.
1096 static void update_tasks_nodemask(struct cpuset *cs, struct ptr_heap *heap)
1098 static nodemask_t newmems; /* protected by cpuset_mutex */
1099 struct cgroup_scanner scan;
1100 struct cpuset *mems_cs = effective_nodemask_cpuset(cs);
1102 cpuset_being_rebound = cs; /* causes mpol_dup() rebind */
1104 guarantee_online_mems(mems_cs, &newmems);
1106 scan.cg = cs->css.cgroup;
1107 scan.test_task = NULL;
1108 scan.process_task = cpuset_change_nodemask;
1109 scan.heap = heap;
1110 scan.data = &newmems;
1113 * The mpol_rebind_mm() call takes mmap_sem, which we couldn't
1114 * take while holding tasklist_lock. Forks can happen - the
1115 * mpol_dup() cpuset_being_rebound check will catch such forks,
1116 * and rebind their vma mempolicies too. Because we still hold
1117 * the global cpuset_mutex, we know that no other rebind effort
1118 * will be contending for the global variable cpuset_being_rebound.
1119 * It's ok if we rebind the same mm twice; mpol_rebind_mm()
1120 * is idempotent. Also migrate pages in each mm to new nodes.
1122 cgroup_scan_tasks(&scan);
1125 * All the tasks' nodemasks have been updated, update
1126 * cs->old_mems_allowed.
1128 cs->old_mems_allowed = newmems;
1130 /* We're done rebinding vmas to this cpuset's new mems_allowed. */
1131 cpuset_being_rebound = NULL;
1135 * update_tasks_nodemask_hier - Update the nodemasks of tasks in the hierarchy.
1136 * @cs: the root cpuset of the hierarchy
1137 * @update_root: update the root cpuset or not?
1138 * @heap: the heap used by cgroup_scan_tasks()
1140 * This will update nodemasks of tasks in @root_cs and all other empty cpusets
1141 * which take on nodemask of @root_cs.
1143 * Called with cpuset_mutex held
1145 static void update_tasks_nodemask_hier(struct cpuset *root_cs,
1146 bool update_root, struct ptr_heap *heap)
1148 struct cpuset *cp;
1149 struct cgroup *pos_cgrp;
1151 if (update_root)
1152 update_tasks_nodemask(root_cs, heap);
1154 rcu_read_lock();
1155 cpuset_for_each_descendant_pre(cp, pos_cgrp, root_cs) {
1156 /* skip the whole subtree if @cp have some CPU */
1157 if (!nodes_empty(cp->mems_allowed)) {
1158 pos_cgrp = cgroup_rightmost_descendant(pos_cgrp);
1159 continue;
1161 if (!css_tryget(&cp->css))
1162 continue;
1163 rcu_read_unlock();
1165 update_tasks_nodemask(cp, heap);
1167 rcu_read_lock();
1168 css_put(&cp->css);
1170 rcu_read_unlock();
1174 * Handle user request to change the 'mems' memory placement
1175 * of a cpuset. Needs to validate the request, update the
1176 * cpusets mems_allowed, and for each task in the cpuset,
1177 * update mems_allowed and rebind task's mempolicy and any vma
1178 * mempolicies and if the cpuset is marked 'memory_migrate',
1179 * migrate the tasks pages to the new memory.
1181 * Call with cpuset_mutex held. May take callback_mutex during call.
1182 * Will take tasklist_lock, scan tasklist for tasks in cpuset cs,
1183 * lock each such tasks mm->mmap_sem, scan its vma's and rebind
1184 * their mempolicies to the cpusets new mems_allowed.
1186 static int update_nodemask(struct cpuset *cs, struct cpuset *trialcs,
1187 const char *buf)
1189 int retval;
1190 struct ptr_heap heap;
1193 * top_cpuset.mems_allowed tracks node_stats[N_MEMORY];
1194 * it's read-only
1196 if (cs == &top_cpuset) {
1197 retval = -EACCES;
1198 goto done;
1202 * An empty mems_allowed is ok iff there are no tasks in the cpuset.
1203 * Since nodelist_parse() fails on an empty mask, we special case
1204 * that parsing. The validate_change() call ensures that cpusets
1205 * with tasks have memory.
1207 if (!*buf) {
1208 nodes_clear(trialcs->mems_allowed);
1209 } else {
1210 retval = nodelist_parse(buf, trialcs->mems_allowed);
1211 if (retval < 0)
1212 goto done;
1214 if (!nodes_subset(trialcs->mems_allowed,
1215 node_states[N_MEMORY])) {
1216 retval = -EINVAL;
1217 goto done;
1221 if (nodes_equal(cs->mems_allowed, trialcs->mems_allowed)) {
1222 retval = 0; /* Too easy - nothing to do */
1223 goto done;
1225 retval = validate_change(cs, trialcs);
1226 if (retval < 0)
1227 goto done;
1229 retval = heap_init(&heap, PAGE_SIZE, GFP_KERNEL, NULL);
1230 if (retval < 0)
1231 goto done;
1233 mutex_lock(&callback_mutex);
1234 cs->mems_allowed = trialcs->mems_allowed;
1235 mutex_unlock(&callback_mutex);
1237 update_tasks_nodemask_hier(cs, true, &heap);
1239 heap_free(&heap);
1240 done:
1241 return retval;
1244 int current_cpuset_is_being_rebound(void)
1246 return task_cs(current) == cpuset_being_rebound;
1249 static int update_relax_domain_level(struct cpuset *cs, s64 val)
1251 #ifdef CONFIG_SMP
1252 if (val < -1 || val >= sched_domain_level_max)
1253 return -EINVAL;
1254 #endif
1256 if (val != cs->relax_domain_level) {
1257 cs->relax_domain_level = val;
1258 if (!cpumask_empty(cs->cpus_allowed) &&
1259 is_sched_load_balance(cs))
1260 rebuild_sched_domains_locked();
1263 return 0;
1267 * cpuset_change_flag - make a task's spread flags the same as its cpuset's
1268 * @tsk: task to be updated
1269 * @scan: struct cgroup_scanner containing the cgroup of the task
1271 * Called by cgroup_scan_tasks() for each task in a cgroup.
1273 * We don't need to re-check for the cgroup/cpuset membership, since we're
1274 * holding cpuset_mutex at this point.
1276 static void cpuset_change_flag(struct task_struct *tsk,
1277 struct cgroup_scanner *scan)
1279 cpuset_update_task_spread_flag(cgroup_cs(scan->cg), tsk);
1283 * update_tasks_flags - update the spread flags of tasks in the cpuset.
1284 * @cs: the cpuset in which each task's spread flags needs to be changed
1285 * @heap: if NULL, defer allocating heap memory to cgroup_scan_tasks()
1287 * Called with cpuset_mutex held
1289 * The cgroup_scan_tasks() function will scan all the tasks in a cgroup,
1290 * calling callback functions for each.
1292 * No return value. It's guaranteed that cgroup_scan_tasks() always returns 0
1293 * if @heap != NULL.
1295 static void update_tasks_flags(struct cpuset *cs, struct ptr_heap *heap)
1297 struct cgroup_scanner scan;
1299 scan.cg = cs->css.cgroup;
1300 scan.test_task = NULL;
1301 scan.process_task = cpuset_change_flag;
1302 scan.heap = heap;
1303 cgroup_scan_tasks(&scan);
1307 * update_flag - read a 0 or a 1 in a file and update associated flag
1308 * bit: the bit to update (see cpuset_flagbits_t)
1309 * cs: the cpuset to update
1310 * turning_on: whether the flag is being set or cleared
1312 * Call with cpuset_mutex held.
1315 static int update_flag(cpuset_flagbits_t bit, struct cpuset *cs,
1316 int turning_on)
1318 struct cpuset *trialcs;
1319 int balance_flag_changed;
1320 int spread_flag_changed;
1321 struct ptr_heap heap;
1322 int err;
1324 trialcs = alloc_trial_cpuset(cs);
1325 if (!trialcs)
1326 return -ENOMEM;
1328 if (turning_on)
1329 set_bit(bit, &trialcs->flags);
1330 else
1331 clear_bit(bit, &trialcs->flags);
1333 err = validate_change(cs, trialcs);
1334 if (err < 0)
1335 goto out;
1337 err = heap_init(&heap, PAGE_SIZE, GFP_KERNEL, NULL);
1338 if (err < 0)
1339 goto out;
1341 balance_flag_changed = (is_sched_load_balance(cs) !=
1342 is_sched_load_balance(trialcs));
1344 spread_flag_changed = ((is_spread_slab(cs) != is_spread_slab(trialcs))
1345 || (is_spread_page(cs) != is_spread_page(trialcs)));
1347 mutex_lock(&callback_mutex);
1348 cs->flags = trialcs->flags;
1349 mutex_unlock(&callback_mutex);
1351 if (!cpumask_empty(trialcs->cpus_allowed) && balance_flag_changed)
1352 rebuild_sched_domains_locked();
1354 if (spread_flag_changed)
1355 update_tasks_flags(cs, &heap);
1356 heap_free(&heap);
1357 out:
1358 free_trial_cpuset(trialcs);
1359 return err;
1363 * Frequency meter - How fast is some event occurring?
1365 * These routines manage a digitally filtered, constant time based,
1366 * event frequency meter. There are four routines:
1367 * fmeter_init() - initialize a frequency meter.
1368 * fmeter_markevent() - called each time the event happens.
1369 * fmeter_getrate() - returns the recent rate of such events.
1370 * fmeter_update() - internal routine used to update fmeter.
1372 * A common data structure is passed to each of these routines,
1373 * which is used to keep track of the state required to manage the
1374 * frequency meter and its digital filter.
1376 * The filter works on the number of events marked per unit time.
1377 * The filter is single-pole low-pass recursive (IIR). The time unit
1378 * is 1 second. Arithmetic is done using 32-bit integers scaled to
1379 * simulate 3 decimal digits of precision (multiplied by 1000).
1381 * With an FM_COEF of 933, and a time base of 1 second, the filter
1382 * has a half-life of 10 seconds, meaning that if the events quit
1383 * happening, then the rate returned from the fmeter_getrate()
1384 * will be cut in half each 10 seconds, until it converges to zero.
1386 * It is not worth doing a real infinitely recursive filter. If more
1387 * than FM_MAXTICKS ticks have elapsed since the last filter event,
1388 * just compute FM_MAXTICKS ticks worth, by which point the level
1389 * will be stable.
1391 * Limit the count of unprocessed events to FM_MAXCNT, so as to avoid
1392 * arithmetic overflow in the fmeter_update() routine.
1394 * Given the simple 32 bit integer arithmetic used, this meter works
1395 * best for reporting rates between one per millisecond (msec) and
1396 * one per 32 (approx) seconds. At constant rates faster than one
1397 * per msec it maxes out at values just under 1,000,000. At constant
1398 * rates between one per msec, and one per second it will stabilize
1399 * to a value N*1000, where N is the rate of events per second.
1400 * At constant rates between one per second and one per 32 seconds,
1401 * it will be choppy, moving up on the seconds that have an event,
1402 * and then decaying until the next event. At rates slower than
1403 * about one in 32 seconds, it decays all the way back to zero between
1404 * each event.
1407 #define FM_COEF 933 /* coefficient for half-life of 10 secs */
1408 #define FM_MAXTICKS ((time_t)99) /* useless computing more ticks than this */
1409 #define FM_MAXCNT 1000000 /* limit cnt to avoid overflow */
1410 #define FM_SCALE 1000 /* faux fixed point scale */
1412 /* Initialize a frequency meter */
1413 static void fmeter_init(struct fmeter *fmp)
1415 fmp->cnt = 0;
1416 fmp->val = 0;
1417 fmp->time = 0;
1418 spin_lock_init(&fmp->lock);
1421 /* Internal meter update - process cnt events and update value */
1422 static void fmeter_update(struct fmeter *fmp)
1424 time_t now = get_seconds();
1425 time_t ticks = now - fmp->time;
1427 if (ticks == 0)
1428 return;
1430 ticks = min(FM_MAXTICKS, ticks);
1431 while (ticks-- > 0)
1432 fmp->val = (FM_COEF * fmp->val) / FM_SCALE;
1433 fmp->time = now;
1435 fmp->val += ((FM_SCALE - FM_COEF) * fmp->cnt) / FM_SCALE;
1436 fmp->cnt = 0;
1439 /* Process any previous ticks, then bump cnt by one (times scale). */
1440 static void fmeter_markevent(struct fmeter *fmp)
1442 spin_lock(&fmp->lock);
1443 fmeter_update(fmp);
1444 fmp->cnt = min(FM_MAXCNT, fmp->cnt + FM_SCALE);
1445 spin_unlock(&fmp->lock);
1448 /* Process any previous ticks, then return current value. */
1449 static int fmeter_getrate(struct fmeter *fmp)
1451 int val;
1453 spin_lock(&fmp->lock);
1454 fmeter_update(fmp);
1455 val = fmp->val;
1456 spin_unlock(&fmp->lock);
1457 return val;
1460 /* Called by cgroups to determine if a cpuset is usable; cpuset_mutex held */
1461 static int cpuset_can_attach(struct cgroup *cgrp, struct cgroup_taskset *tset)
1463 struct cpuset *cs = cgroup_cs(cgrp);
1464 struct task_struct *task;
1465 int ret;
1467 mutex_lock(&cpuset_mutex);
1470 * We allow to move tasks into an empty cpuset if sane_behavior
1471 * flag is set.
1473 ret = -ENOSPC;
1474 if (!cgroup_sane_behavior(cgrp) &&
1475 (cpumask_empty(cs->cpus_allowed) || nodes_empty(cs->mems_allowed)))
1476 goto out_unlock;
1478 cgroup_taskset_for_each(task, cgrp, tset) {
1480 * Kthreads which disallow setaffinity shouldn't be moved
1481 * to a new cpuset; we don't want to change their cpu
1482 * affinity and isolating such threads by their set of
1483 * allowed nodes is unnecessary. Thus, cpusets are not
1484 * applicable for such threads. This prevents checking for
1485 * success of set_cpus_allowed_ptr() on all attached tasks
1486 * before cpus_allowed may be changed.
1488 ret = -EINVAL;
1489 if (task->flags & PF_NO_SETAFFINITY)
1490 goto out_unlock;
1491 ret = security_task_setscheduler(task);
1492 if (ret)
1493 goto out_unlock;
1497 * Mark attach is in progress. This makes validate_change() fail
1498 * changes which zero cpus/mems_allowed.
1500 cs->attach_in_progress++;
1501 ret = 0;
1502 out_unlock:
1503 mutex_unlock(&cpuset_mutex);
1504 return ret;
1507 static void cpuset_cancel_attach(struct cgroup *cgrp,
1508 struct cgroup_taskset *tset)
1510 mutex_lock(&cpuset_mutex);
1511 cgroup_cs(cgrp)->attach_in_progress--;
1512 mutex_unlock(&cpuset_mutex);
1516 * Protected by cpuset_mutex. cpus_attach is used only by cpuset_attach()
1517 * but we can't allocate it dynamically there. Define it global and
1518 * allocate from cpuset_init().
1520 static cpumask_var_t cpus_attach;
1522 static void cpuset_attach(struct cgroup *cgrp, struct cgroup_taskset *tset)
1524 /* static buf protected by cpuset_mutex */
1525 static nodemask_t cpuset_attach_nodemask_to;
1526 struct mm_struct *mm;
1527 struct task_struct *task;
1528 struct task_struct *leader = cgroup_taskset_first(tset);
1529 struct cgroup *oldcgrp = cgroup_taskset_cur_cgroup(tset);
1530 struct cpuset *cs = cgroup_cs(cgrp);
1531 struct cpuset *oldcs = cgroup_cs(oldcgrp);
1532 struct cpuset *cpus_cs = effective_cpumask_cpuset(cs);
1533 struct cpuset *mems_cs = effective_nodemask_cpuset(cs);
1535 mutex_lock(&cpuset_mutex);
1537 /* prepare for attach */
1538 if (cs == &top_cpuset)
1539 cpumask_copy(cpus_attach, cpu_possible_mask);
1540 else
1541 guarantee_online_cpus(cpus_cs, cpus_attach);
1543 guarantee_online_mems(mems_cs, &cpuset_attach_nodemask_to);
1545 cgroup_taskset_for_each(task, cgrp, tset) {
1547 * can_attach beforehand should guarantee that this doesn't
1548 * fail. TODO: have a better way to handle failure here
1550 WARN_ON_ONCE(set_cpus_allowed_ptr(task, cpus_attach));
1552 cpuset_change_task_nodemask(task, &cpuset_attach_nodemask_to);
1553 cpuset_update_task_spread_flag(cs, task);
1557 * Change mm, possibly for multiple threads in a threadgroup. This is
1558 * expensive and may sleep.
1560 cpuset_attach_nodemask_to = cs->mems_allowed;
1561 mm = get_task_mm(leader);
1562 if (mm) {
1563 struct cpuset *mems_oldcs = effective_nodemask_cpuset(oldcs);
1565 mpol_rebind_mm(mm, &cpuset_attach_nodemask_to);
1568 * old_mems_allowed is the same with mems_allowed here, except
1569 * if this task is being moved automatically due to hotplug.
1570 * In that case @mems_allowed has been updated and is empty,
1571 * so @old_mems_allowed is the right nodesets that we migrate
1572 * mm from.
1574 if (is_memory_migrate(cs)) {
1575 cpuset_migrate_mm(mm, &mems_oldcs->old_mems_allowed,
1576 &cpuset_attach_nodemask_to);
1578 mmput(mm);
1581 cs->old_mems_allowed = cpuset_attach_nodemask_to;
1583 cs->attach_in_progress--;
1584 if (!cs->attach_in_progress)
1585 wake_up(&cpuset_attach_wq);
1587 mutex_unlock(&cpuset_mutex);
1590 /* The various types of files and directories in a cpuset file system */
1592 typedef enum {
1593 FILE_MEMORY_MIGRATE,
1594 FILE_CPULIST,
1595 FILE_MEMLIST,
1596 FILE_CPU_EXCLUSIVE,
1597 FILE_MEM_EXCLUSIVE,
1598 FILE_MEM_HARDWALL,
1599 FILE_SCHED_LOAD_BALANCE,
1600 FILE_SCHED_RELAX_DOMAIN_LEVEL,
1601 FILE_MEMORY_PRESSURE_ENABLED,
1602 FILE_MEMORY_PRESSURE,
1603 FILE_SPREAD_PAGE,
1604 FILE_SPREAD_SLAB,
1605 } cpuset_filetype_t;
1607 static int cpuset_write_u64(struct cgroup *cgrp, struct cftype *cft, u64 val)
1609 struct cpuset *cs = cgroup_cs(cgrp);
1610 cpuset_filetype_t type = cft->private;
1611 int retval = -ENODEV;
1613 mutex_lock(&cpuset_mutex);
1614 if (!is_cpuset_online(cs))
1615 goto out_unlock;
1617 switch (type) {
1618 case FILE_CPU_EXCLUSIVE:
1619 retval = update_flag(CS_CPU_EXCLUSIVE, cs, val);
1620 break;
1621 case FILE_MEM_EXCLUSIVE:
1622 retval = update_flag(CS_MEM_EXCLUSIVE, cs, val);
1623 break;
1624 case FILE_MEM_HARDWALL:
1625 retval = update_flag(CS_MEM_HARDWALL, cs, val);
1626 break;
1627 case FILE_SCHED_LOAD_BALANCE:
1628 retval = update_flag(CS_SCHED_LOAD_BALANCE, cs, val);
1629 break;
1630 case FILE_MEMORY_MIGRATE:
1631 retval = update_flag(CS_MEMORY_MIGRATE, cs, val);
1632 break;
1633 case FILE_MEMORY_PRESSURE_ENABLED:
1634 cpuset_memory_pressure_enabled = !!val;
1635 break;
1636 case FILE_MEMORY_PRESSURE:
1637 retval = -EACCES;
1638 break;
1639 case FILE_SPREAD_PAGE:
1640 retval = update_flag(CS_SPREAD_PAGE, cs, val);
1641 break;
1642 case FILE_SPREAD_SLAB:
1643 retval = update_flag(CS_SPREAD_SLAB, cs, val);
1644 break;
1645 default:
1646 retval = -EINVAL;
1647 break;
1649 out_unlock:
1650 mutex_unlock(&cpuset_mutex);
1651 return retval;
1654 static int cpuset_write_s64(struct cgroup *cgrp, struct cftype *cft, s64 val)
1656 struct cpuset *cs = cgroup_cs(cgrp);
1657 cpuset_filetype_t type = cft->private;
1658 int retval = -ENODEV;
1660 mutex_lock(&cpuset_mutex);
1661 if (!is_cpuset_online(cs))
1662 goto out_unlock;
1664 switch (type) {
1665 case FILE_SCHED_RELAX_DOMAIN_LEVEL:
1666 retval = update_relax_domain_level(cs, val);
1667 break;
1668 default:
1669 retval = -EINVAL;
1670 break;
1672 out_unlock:
1673 mutex_unlock(&cpuset_mutex);
1674 return retval;
1678 * Common handling for a write to a "cpus" or "mems" file.
1680 static int cpuset_write_resmask(struct cgroup *cgrp, struct cftype *cft,
1681 const char *buf)
1683 struct cpuset *cs = cgroup_cs(cgrp);
1684 struct cpuset *trialcs;
1685 int retval = -ENODEV;
1688 * CPU or memory hotunplug may leave @cs w/o any execution
1689 * resources, in which case the hotplug code asynchronously updates
1690 * configuration and transfers all tasks to the nearest ancestor
1691 * which can execute.
1693 * As writes to "cpus" or "mems" may restore @cs's execution
1694 * resources, wait for the previously scheduled operations before
1695 * proceeding, so that we don't end up keep removing tasks added
1696 * after execution capability is restored.
1698 flush_work(&cpuset_hotplug_work);
1700 mutex_lock(&cpuset_mutex);
1701 if (!is_cpuset_online(cs))
1702 goto out_unlock;
1704 trialcs = alloc_trial_cpuset(cs);
1705 if (!trialcs) {
1706 retval = -ENOMEM;
1707 goto out_unlock;
1710 switch (cft->private) {
1711 case FILE_CPULIST:
1712 retval = update_cpumask(cs, trialcs, buf);
1713 break;
1714 case FILE_MEMLIST:
1715 retval = update_nodemask(cs, trialcs, buf);
1716 break;
1717 default:
1718 retval = -EINVAL;
1719 break;
1722 free_trial_cpuset(trialcs);
1723 out_unlock:
1724 mutex_unlock(&cpuset_mutex);
1725 return retval;
1729 * These ascii lists should be read in a single call, by using a user
1730 * buffer large enough to hold the entire map. If read in smaller
1731 * chunks, there is no guarantee of atomicity. Since the display format
1732 * used, list of ranges of sequential numbers, is variable length,
1733 * and since these maps can change value dynamically, one could read
1734 * gibberish by doing partial reads while a list was changing.
1735 * A single large read to a buffer that crosses a page boundary is
1736 * ok, because the result being copied to user land is not recomputed
1737 * across a page fault.
1740 static size_t cpuset_sprintf_cpulist(char *page, struct cpuset *cs)
1742 size_t count;
1744 mutex_lock(&callback_mutex);
1745 count = cpulist_scnprintf(page, PAGE_SIZE, cs->cpus_allowed);
1746 mutex_unlock(&callback_mutex);
1748 return count;
1751 static size_t cpuset_sprintf_memlist(char *page, struct cpuset *cs)
1753 size_t count;
1755 mutex_lock(&callback_mutex);
1756 count = nodelist_scnprintf(page, PAGE_SIZE, cs->mems_allowed);
1757 mutex_unlock(&callback_mutex);
1759 return count;
1762 static ssize_t cpuset_common_file_read(struct cgroup *cgrp,
1763 struct cftype *cft,
1764 struct file *file,
1765 char __user *buf,
1766 size_t nbytes, loff_t *ppos)
1768 struct cpuset *cs = cgroup_cs(cgrp);
1769 cpuset_filetype_t type = cft->private;
1770 char *page;
1771 ssize_t retval = 0;
1772 char *s;
1774 if (!(page = (char *)__get_free_page(GFP_TEMPORARY)))
1775 return -ENOMEM;
1777 s = page;
1779 switch (type) {
1780 case FILE_CPULIST:
1781 s += cpuset_sprintf_cpulist(s, cs);
1782 break;
1783 case FILE_MEMLIST:
1784 s += cpuset_sprintf_memlist(s, cs);
1785 break;
1786 default:
1787 retval = -EINVAL;
1788 goto out;
1790 *s++ = '\n';
1792 retval = simple_read_from_buffer(buf, nbytes, ppos, page, s - page);
1793 out:
1794 free_page((unsigned long)page);
1795 return retval;
1798 static u64 cpuset_read_u64(struct cgroup *cgrp, struct cftype *cft)
1800 struct cpuset *cs = cgroup_cs(cgrp);
1801 cpuset_filetype_t type = cft->private;
1802 switch (type) {
1803 case FILE_CPU_EXCLUSIVE:
1804 return is_cpu_exclusive(cs);
1805 case FILE_MEM_EXCLUSIVE:
1806 return is_mem_exclusive(cs);
1807 case FILE_MEM_HARDWALL:
1808 return is_mem_hardwall(cs);
1809 case FILE_SCHED_LOAD_BALANCE:
1810 return is_sched_load_balance(cs);
1811 case FILE_MEMORY_MIGRATE:
1812 return is_memory_migrate(cs);
1813 case FILE_MEMORY_PRESSURE_ENABLED:
1814 return cpuset_memory_pressure_enabled;
1815 case FILE_MEMORY_PRESSURE:
1816 return fmeter_getrate(&cs->fmeter);
1817 case FILE_SPREAD_PAGE:
1818 return is_spread_page(cs);
1819 case FILE_SPREAD_SLAB:
1820 return is_spread_slab(cs);
1821 default:
1822 BUG();
1825 /* Unreachable but makes gcc happy */
1826 return 0;
1829 static s64 cpuset_read_s64(struct cgroup *cgrp, struct cftype *cft)
1831 struct cpuset *cs = cgroup_cs(cgrp);
1832 cpuset_filetype_t type = cft->private;
1833 switch (type) {
1834 case FILE_SCHED_RELAX_DOMAIN_LEVEL:
1835 return cs->relax_domain_level;
1836 default:
1837 BUG();
1840 /* Unrechable but makes gcc happy */
1841 return 0;
1846 * for the common functions, 'private' gives the type of file
1849 static struct cftype files[] = {
1851 .name = "cpus",
1852 .read = cpuset_common_file_read,
1853 .write_string = cpuset_write_resmask,
1854 .max_write_len = (100U + 6 * NR_CPUS),
1855 .private = FILE_CPULIST,
1859 .name = "mems",
1860 .read = cpuset_common_file_read,
1861 .write_string = cpuset_write_resmask,
1862 .max_write_len = (100U + 6 * MAX_NUMNODES),
1863 .private = FILE_MEMLIST,
1867 .name = "cpu_exclusive",
1868 .read_u64 = cpuset_read_u64,
1869 .write_u64 = cpuset_write_u64,
1870 .private = FILE_CPU_EXCLUSIVE,
1874 .name = "mem_exclusive",
1875 .read_u64 = cpuset_read_u64,
1876 .write_u64 = cpuset_write_u64,
1877 .private = FILE_MEM_EXCLUSIVE,
1881 .name = "mem_hardwall",
1882 .read_u64 = cpuset_read_u64,
1883 .write_u64 = cpuset_write_u64,
1884 .private = FILE_MEM_HARDWALL,
1888 .name = "sched_load_balance",
1889 .read_u64 = cpuset_read_u64,
1890 .write_u64 = cpuset_write_u64,
1891 .private = FILE_SCHED_LOAD_BALANCE,
1895 .name = "sched_relax_domain_level",
1896 .read_s64 = cpuset_read_s64,
1897 .write_s64 = cpuset_write_s64,
1898 .private = FILE_SCHED_RELAX_DOMAIN_LEVEL,
1902 .name = "memory_migrate",
1903 .read_u64 = cpuset_read_u64,
1904 .write_u64 = cpuset_write_u64,
1905 .private = FILE_MEMORY_MIGRATE,
1909 .name = "memory_pressure",
1910 .read_u64 = cpuset_read_u64,
1911 .write_u64 = cpuset_write_u64,
1912 .private = FILE_MEMORY_PRESSURE,
1913 .mode = S_IRUGO,
1917 .name = "memory_spread_page",
1918 .read_u64 = cpuset_read_u64,
1919 .write_u64 = cpuset_write_u64,
1920 .private = FILE_SPREAD_PAGE,
1924 .name = "memory_spread_slab",
1925 .read_u64 = cpuset_read_u64,
1926 .write_u64 = cpuset_write_u64,
1927 .private = FILE_SPREAD_SLAB,
1931 .name = "memory_pressure_enabled",
1932 .flags = CFTYPE_ONLY_ON_ROOT,
1933 .read_u64 = cpuset_read_u64,
1934 .write_u64 = cpuset_write_u64,
1935 .private = FILE_MEMORY_PRESSURE_ENABLED,
1938 { } /* terminate */
1942 * cpuset_css_alloc - allocate a cpuset css
1943 * cgrp: control group that the new cpuset will be part of
1946 static struct cgroup_subsys_state *cpuset_css_alloc(struct cgroup *cgrp)
1948 struct cpuset *cs;
1950 if (!cgrp->parent)
1951 return &top_cpuset.css;
1953 cs = kzalloc(sizeof(*cs), GFP_KERNEL);
1954 if (!cs)
1955 return ERR_PTR(-ENOMEM);
1956 if (!alloc_cpumask_var(&cs->cpus_allowed, GFP_KERNEL)) {
1957 kfree(cs);
1958 return ERR_PTR(-ENOMEM);
1961 set_bit(CS_SCHED_LOAD_BALANCE, &cs->flags);
1962 cpumask_clear(cs->cpus_allowed);
1963 nodes_clear(cs->mems_allowed);
1964 fmeter_init(&cs->fmeter);
1965 cs->relax_domain_level = -1;
1967 return &cs->css;
1970 static int cpuset_css_online(struct cgroup *cgrp)
1972 struct cpuset *cs = cgroup_cs(cgrp);
1973 struct cpuset *parent = parent_cs(cs);
1974 struct cpuset *tmp_cs;
1975 struct cgroup *pos_cg;
1977 if (!parent)
1978 return 0;
1980 mutex_lock(&cpuset_mutex);
1982 set_bit(CS_ONLINE, &cs->flags);
1983 if (is_spread_page(parent))
1984 set_bit(CS_SPREAD_PAGE, &cs->flags);
1985 if (is_spread_slab(parent))
1986 set_bit(CS_SPREAD_SLAB, &cs->flags);
1988 number_of_cpusets++;
1990 if (!test_bit(CGRP_CPUSET_CLONE_CHILDREN, &cgrp->flags))
1991 goto out_unlock;
1994 * Clone @parent's configuration if CGRP_CPUSET_CLONE_CHILDREN is
1995 * set. This flag handling is implemented in cgroup core for
1996 * histrical reasons - the flag may be specified during mount.
1998 * Currently, if any sibling cpusets have exclusive cpus or mem, we
1999 * refuse to clone the configuration - thereby refusing the task to
2000 * be entered, and as a result refusing the sys_unshare() or
2001 * clone() which initiated it. If this becomes a problem for some
2002 * users who wish to allow that scenario, then this could be
2003 * changed to grant parent->cpus_allowed-sibling_cpus_exclusive
2004 * (and likewise for mems) to the new cgroup.
2006 rcu_read_lock();
2007 cpuset_for_each_child(tmp_cs, pos_cg, parent) {
2008 if (is_mem_exclusive(tmp_cs) || is_cpu_exclusive(tmp_cs)) {
2009 rcu_read_unlock();
2010 goto out_unlock;
2013 rcu_read_unlock();
2015 mutex_lock(&callback_mutex);
2016 cs->mems_allowed = parent->mems_allowed;
2017 cpumask_copy(cs->cpus_allowed, parent->cpus_allowed);
2018 mutex_unlock(&callback_mutex);
2019 out_unlock:
2020 mutex_unlock(&cpuset_mutex);
2021 return 0;
2024 static void cpuset_css_offline(struct cgroup *cgrp)
2026 struct cpuset *cs = cgroup_cs(cgrp);
2028 mutex_lock(&cpuset_mutex);
2030 if (is_sched_load_balance(cs))
2031 update_flag(CS_SCHED_LOAD_BALANCE, cs, 0);
2033 number_of_cpusets--;
2034 clear_bit(CS_ONLINE, &cs->flags);
2036 mutex_unlock(&cpuset_mutex);
2040 * If the cpuset being removed has its flag 'sched_load_balance'
2041 * enabled, then simulate turning sched_load_balance off, which
2042 * will call rebuild_sched_domains_locked().
2045 static void cpuset_css_free(struct cgroup *cgrp)
2047 struct cpuset *cs = cgroup_cs(cgrp);
2049 free_cpumask_var(cs->cpus_allowed);
2050 kfree(cs);
2053 struct cgroup_subsys cpuset_subsys = {
2054 .name = "cpuset",
2055 .css_alloc = cpuset_css_alloc,
2056 .css_online = cpuset_css_online,
2057 .css_offline = cpuset_css_offline,
2058 .css_free = cpuset_css_free,
2059 .can_attach = cpuset_can_attach,
2060 .cancel_attach = cpuset_cancel_attach,
2061 .attach = cpuset_attach,
2062 .subsys_id = cpuset_subsys_id,
2063 .base_cftypes = files,
2064 .early_init = 1,
2068 * cpuset_init - initialize cpusets at system boot
2070 * Description: Initialize top_cpuset and the cpuset internal file system,
2073 int __init cpuset_init(void)
2075 int err = 0;
2077 if (!alloc_cpumask_var(&top_cpuset.cpus_allowed, GFP_KERNEL))
2078 BUG();
2080 cpumask_setall(top_cpuset.cpus_allowed);
2081 nodes_setall(top_cpuset.mems_allowed);
2083 fmeter_init(&top_cpuset.fmeter);
2084 set_bit(CS_SCHED_LOAD_BALANCE, &top_cpuset.flags);
2085 top_cpuset.relax_domain_level = -1;
2087 err = register_filesystem(&cpuset_fs_type);
2088 if (err < 0)
2089 return err;
2091 if (!alloc_cpumask_var(&cpus_attach, GFP_KERNEL))
2092 BUG();
2094 number_of_cpusets = 1;
2095 return 0;
2099 * If CPU and/or memory hotplug handlers, below, unplug any CPUs
2100 * or memory nodes, we need to walk over the cpuset hierarchy,
2101 * removing that CPU or node from all cpusets. If this removes the
2102 * last CPU or node from a cpuset, then move the tasks in the empty
2103 * cpuset to its next-highest non-empty parent.
2105 static void remove_tasks_in_empty_cpuset(struct cpuset *cs)
2107 struct cpuset *parent;
2110 * Find its next-highest non-empty parent, (top cpuset
2111 * has online cpus, so can't be empty).
2113 parent = parent_cs(cs);
2114 while (cpumask_empty(parent->cpus_allowed) ||
2115 nodes_empty(parent->mems_allowed))
2116 parent = parent_cs(parent);
2118 if (cgroup_transfer_tasks(parent->css.cgroup, cs->css.cgroup)) {
2119 rcu_read_lock();
2120 printk(KERN_ERR "cpuset: failed to transfer tasks out of empty cpuset %s\n",
2121 cgroup_name(cs->css.cgroup));
2122 rcu_read_unlock();
2127 * cpuset_hotplug_update_tasks - update tasks in a cpuset for hotunplug
2128 * @cs: cpuset in interest
2130 * Compare @cs's cpu and mem masks against top_cpuset and if some have gone
2131 * offline, update @cs accordingly. If @cs ends up with no CPU or memory,
2132 * all its tasks are moved to the nearest ancestor with both resources.
2134 static void cpuset_hotplug_update_tasks(struct cpuset *cs)
2136 static cpumask_t off_cpus;
2137 static nodemask_t off_mems;
2138 bool is_empty;
2139 bool sane = cgroup_sane_behavior(cs->css.cgroup);
2141 retry:
2142 wait_event(cpuset_attach_wq, cs->attach_in_progress == 0);
2144 mutex_lock(&cpuset_mutex);
2147 * We have raced with task attaching. We wait until attaching
2148 * is finished, so we won't attach a task to an empty cpuset.
2150 if (cs->attach_in_progress) {
2151 mutex_unlock(&cpuset_mutex);
2152 goto retry;
2155 cpumask_andnot(&off_cpus, cs->cpus_allowed, top_cpuset.cpus_allowed);
2156 nodes_andnot(off_mems, cs->mems_allowed, top_cpuset.mems_allowed);
2158 mutex_lock(&callback_mutex);
2159 cpumask_andnot(cs->cpus_allowed, cs->cpus_allowed, &off_cpus);
2160 mutex_unlock(&callback_mutex);
2163 * If sane_behavior flag is set, we need to update tasks' cpumask
2164 * for empty cpuset to take on ancestor's cpumask. Otherwise, don't
2165 * call update_tasks_cpumask() if the cpuset becomes empty, as
2166 * the tasks in it will be migrated to an ancestor.
2168 if ((sane && cpumask_empty(cs->cpus_allowed)) ||
2169 (!cpumask_empty(&off_cpus) && !cpumask_empty(cs->cpus_allowed)))
2170 update_tasks_cpumask(cs, NULL);
2172 mutex_lock(&callback_mutex);
2173 nodes_andnot(cs->mems_allowed, cs->mems_allowed, off_mems);
2174 mutex_unlock(&callback_mutex);
2177 * If sane_behavior flag is set, we need to update tasks' nodemask
2178 * for empty cpuset to take on ancestor's nodemask. Otherwise, don't
2179 * call update_tasks_nodemask() if the cpuset becomes empty, as
2180 * the tasks in it will be migratd to an ancestor.
2182 if ((sane && nodes_empty(cs->mems_allowed)) ||
2183 (!nodes_empty(off_mems) && !nodes_empty(cs->mems_allowed)))
2184 update_tasks_nodemask(cs, NULL);
2186 is_empty = cpumask_empty(cs->cpus_allowed) ||
2187 nodes_empty(cs->mems_allowed);
2189 mutex_unlock(&cpuset_mutex);
2192 * If sane_behavior flag is set, we'll keep tasks in empty cpusets.
2194 * Otherwise move tasks to the nearest ancestor with execution
2195 * resources. This is full cgroup operation which will
2196 * also call back into cpuset. Should be done outside any lock.
2198 if (!sane && is_empty)
2199 remove_tasks_in_empty_cpuset(cs);
2203 * cpuset_hotplug_workfn - handle CPU/memory hotunplug for a cpuset
2205 * This function is called after either CPU or memory configuration has
2206 * changed and updates cpuset accordingly. The top_cpuset is always
2207 * synchronized to cpu_active_mask and N_MEMORY, which is necessary in
2208 * order to make cpusets transparent (of no affect) on systems that are
2209 * actively using CPU hotplug but making no active use of cpusets.
2211 * Non-root cpusets are only affected by offlining. If any CPUs or memory
2212 * nodes have been taken down, cpuset_hotplug_update_tasks() is invoked on
2213 * all descendants.
2215 * Note that CPU offlining during suspend is ignored. We don't modify
2216 * cpusets across suspend/resume cycles at all.
2218 static void cpuset_hotplug_workfn(struct work_struct *work)
2220 static cpumask_t new_cpus;
2221 static nodemask_t new_mems;
2222 bool cpus_updated, mems_updated;
2224 mutex_lock(&cpuset_mutex);
2226 /* fetch the available cpus/mems and find out which changed how */
2227 cpumask_copy(&new_cpus, cpu_active_mask);
2228 new_mems = node_states[N_MEMORY];
2230 cpus_updated = !cpumask_equal(top_cpuset.cpus_allowed, &new_cpus);
2231 mems_updated = !nodes_equal(top_cpuset.mems_allowed, new_mems);
2233 /* synchronize cpus_allowed to cpu_active_mask */
2234 if (cpus_updated) {
2235 mutex_lock(&callback_mutex);
2236 cpumask_copy(top_cpuset.cpus_allowed, &new_cpus);
2237 mutex_unlock(&callback_mutex);
2238 /* we don't mess with cpumasks of tasks in top_cpuset */
2241 /* synchronize mems_allowed to N_MEMORY */
2242 if (mems_updated) {
2243 mutex_lock(&callback_mutex);
2244 top_cpuset.mems_allowed = new_mems;
2245 mutex_unlock(&callback_mutex);
2246 update_tasks_nodemask(&top_cpuset, NULL);
2249 mutex_unlock(&cpuset_mutex);
2251 /* if cpus or mems changed, we need to propagate to descendants */
2252 if (cpus_updated || mems_updated) {
2253 struct cpuset *cs;
2254 struct cgroup *pos_cgrp;
2256 rcu_read_lock();
2257 cpuset_for_each_descendant_pre(cs, pos_cgrp, &top_cpuset) {
2258 if (!css_tryget(&cs->css))
2259 continue;
2260 rcu_read_unlock();
2262 cpuset_hotplug_update_tasks(cs);
2264 rcu_read_lock();
2265 css_put(&cs->css);
2267 rcu_read_unlock();
2270 /* rebuild sched domains if cpus_allowed has changed */
2271 if (cpus_updated)
2272 rebuild_sched_domains();
2275 void cpuset_update_active_cpus(bool cpu_online)
2278 * We're inside cpu hotplug critical region which usually nests
2279 * inside cgroup synchronization. Bounce actual hotplug processing
2280 * to a work item to avoid reverse locking order.
2282 * We still need to do partition_sched_domains() synchronously;
2283 * otherwise, the scheduler will get confused and put tasks to the
2284 * dead CPU. Fall back to the default single domain.
2285 * cpuset_hotplug_workfn() will rebuild it as necessary.
2287 partition_sched_domains(1, NULL, NULL);
2288 schedule_work(&cpuset_hotplug_work);
2292 * Keep top_cpuset.mems_allowed tracking node_states[N_MEMORY].
2293 * Call this routine anytime after node_states[N_MEMORY] changes.
2294 * See cpuset_update_active_cpus() for CPU hotplug handling.
2296 static int cpuset_track_online_nodes(struct notifier_block *self,
2297 unsigned long action, void *arg)
2299 schedule_work(&cpuset_hotplug_work);
2300 return NOTIFY_OK;
2303 static struct notifier_block cpuset_track_online_nodes_nb = {
2304 .notifier_call = cpuset_track_online_nodes,
2305 .priority = 10, /* ??! */
2309 * cpuset_init_smp - initialize cpus_allowed
2311 * Description: Finish top cpuset after cpu, node maps are initialized
2313 void __init cpuset_init_smp(void)
2315 cpumask_copy(top_cpuset.cpus_allowed, cpu_active_mask);
2316 top_cpuset.mems_allowed = node_states[N_MEMORY];
2317 top_cpuset.old_mems_allowed = top_cpuset.mems_allowed;
2319 register_hotmemory_notifier(&cpuset_track_online_nodes_nb);
2323 * cpuset_cpus_allowed - return cpus_allowed mask from a tasks cpuset.
2324 * @tsk: pointer to task_struct from which to obtain cpuset->cpus_allowed.
2325 * @pmask: pointer to struct cpumask variable to receive cpus_allowed set.
2327 * Description: Returns the cpumask_var_t cpus_allowed of the cpuset
2328 * attached to the specified @tsk. Guaranteed to return some non-empty
2329 * subset of cpu_online_mask, even if this means going outside the
2330 * tasks cpuset.
2333 void cpuset_cpus_allowed(struct task_struct *tsk, struct cpumask *pmask)
2335 struct cpuset *cpus_cs;
2337 mutex_lock(&callback_mutex);
2338 task_lock(tsk);
2339 cpus_cs = effective_cpumask_cpuset(task_cs(tsk));
2340 guarantee_online_cpus(cpus_cs, pmask);
2341 task_unlock(tsk);
2342 mutex_unlock(&callback_mutex);
2345 void cpuset_cpus_allowed_fallback(struct task_struct *tsk)
2347 const struct cpuset *cpus_cs;
2349 rcu_read_lock();
2350 cpus_cs = effective_cpumask_cpuset(task_cs(tsk));
2351 do_set_cpus_allowed(tsk, cpus_cs->cpus_allowed);
2352 rcu_read_unlock();
2355 * We own tsk->cpus_allowed, nobody can change it under us.
2357 * But we used cs && cs->cpus_allowed lockless and thus can
2358 * race with cgroup_attach_task() or update_cpumask() and get
2359 * the wrong tsk->cpus_allowed. However, both cases imply the
2360 * subsequent cpuset_change_cpumask()->set_cpus_allowed_ptr()
2361 * which takes task_rq_lock().
2363 * If we are called after it dropped the lock we must see all
2364 * changes in tsk_cs()->cpus_allowed. Otherwise we can temporary
2365 * set any mask even if it is not right from task_cs() pov,
2366 * the pending set_cpus_allowed_ptr() will fix things.
2368 * select_fallback_rq() will fix things ups and set cpu_possible_mask
2369 * if required.
2373 void cpuset_init_current_mems_allowed(void)
2375 nodes_setall(current->mems_allowed);
2379 * cpuset_mems_allowed - return mems_allowed mask from a tasks cpuset.
2380 * @tsk: pointer to task_struct from which to obtain cpuset->mems_allowed.
2382 * Description: Returns the nodemask_t mems_allowed of the cpuset
2383 * attached to the specified @tsk. Guaranteed to return some non-empty
2384 * subset of node_states[N_MEMORY], even if this means going outside the
2385 * tasks cpuset.
2388 nodemask_t cpuset_mems_allowed(struct task_struct *tsk)
2390 struct cpuset *mems_cs;
2391 nodemask_t mask;
2393 mutex_lock(&callback_mutex);
2394 task_lock(tsk);
2395 mems_cs = effective_nodemask_cpuset(task_cs(tsk));
2396 guarantee_online_mems(mems_cs, &mask);
2397 task_unlock(tsk);
2398 mutex_unlock(&callback_mutex);
2400 return mask;
2404 * cpuset_nodemask_valid_mems_allowed - check nodemask vs. curremt mems_allowed
2405 * @nodemask: the nodemask to be checked
2407 * Are any of the nodes in the nodemask allowed in current->mems_allowed?
2409 int cpuset_nodemask_valid_mems_allowed(nodemask_t *nodemask)
2411 return nodes_intersects(*nodemask, current->mems_allowed);
2415 * nearest_hardwall_ancestor() - Returns the nearest mem_exclusive or
2416 * mem_hardwall ancestor to the specified cpuset. Call holding
2417 * callback_mutex. If no ancestor is mem_exclusive or mem_hardwall
2418 * (an unusual configuration), then returns the root cpuset.
2420 static const struct cpuset *nearest_hardwall_ancestor(const struct cpuset *cs)
2422 while (!(is_mem_exclusive(cs) || is_mem_hardwall(cs)) && parent_cs(cs))
2423 cs = parent_cs(cs);
2424 return cs;
2428 * cpuset_node_allowed_softwall - Can we allocate on a memory node?
2429 * @node: is this an allowed node?
2430 * @gfp_mask: memory allocation flags
2432 * If we're in interrupt, yes, we can always allocate. If __GFP_THISNODE is
2433 * set, yes, we can always allocate. If node is in our task's mems_allowed,
2434 * yes. If it's not a __GFP_HARDWALL request and this node is in the nearest
2435 * hardwalled cpuset ancestor to this task's cpuset, yes. If the task has been
2436 * OOM killed and has access to memory reserves as specified by the TIF_MEMDIE
2437 * flag, yes.
2438 * Otherwise, no.
2440 * If __GFP_HARDWALL is set, cpuset_node_allowed_softwall() reduces to
2441 * cpuset_node_allowed_hardwall(). Otherwise, cpuset_node_allowed_softwall()
2442 * might sleep, and might allow a node from an enclosing cpuset.
2444 * cpuset_node_allowed_hardwall() only handles the simpler case of hardwall
2445 * cpusets, and never sleeps.
2447 * The __GFP_THISNODE placement logic is really handled elsewhere,
2448 * by forcibly using a zonelist starting at a specified node, and by
2449 * (in get_page_from_freelist()) refusing to consider the zones for
2450 * any node on the zonelist except the first. By the time any such
2451 * calls get to this routine, we should just shut up and say 'yes'.
2453 * GFP_USER allocations are marked with the __GFP_HARDWALL bit,
2454 * and do not allow allocations outside the current tasks cpuset
2455 * unless the task has been OOM killed as is marked TIF_MEMDIE.
2456 * GFP_KERNEL allocations are not so marked, so can escape to the
2457 * nearest enclosing hardwalled ancestor cpuset.
2459 * Scanning up parent cpusets requires callback_mutex. The
2460 * __alloc_pages() routine only calls here with __GFP_HARDWALL bit
2461 * _not_ set if it's a GFP_KERNEL allocation, and all nodes in the
2462 * current tasks mems_allowed came up empty on the first pass over
2463 * the zonelist. So only GFP_KERNEL allocations, if all nodes in the
2464 * cpuset are short of memory, might require taking the callback_mutex
2465 * mutex.
2467 * The first call here from mm/page_alloc:get_page_from_freelist()
2468 * has __GFP_HARDWALL set in gfp_mask, enforcing hardwall cpusets,
2469 * so no allocation on a node outside the cpuset is allowed (unless
2470 * in interrupt, of course).
2472 * The second pass through get_page_from_freelist() doesn't even call
2473 * here for GFP_ATOMIC calls. For those calls, the __alloc_pages()
2474 * variable 'wait' is not set, and the bit ALLOC_CPUSET is not set
2475 * in alloc_flags. That logic and the checks below have the combined
2476 * affect that:
2477 * in_interrupt - any node ok (current task context irrelevant)
2478 * GFP_ATOMIC - any node ok
2479 * TIF_MEMDIE - any node ok
2480 * GFP_KERNEL - any node in enclosing hardwalled cpuset ok
2481 * GFP_USER - only nodes in current tasks mems allowed ok.
2483 * Rule:
2484 * Don't call cpuset_node_allowed_softwall if you can't sleep, unless you
2485 * pass in the __GFP_HARDWALL flag set in gfp_flag, which disables
2486 * the code that might scan up ancestor cpusets and sleep.
2488 int __cpuset_node_allowed_softwall(int node, gfp_t gfp_mask)
2490 const struct cpuset *cs; /* current cpuset ancestors */
2491 int allowed; /* is allocation in zone z allowed? */
2493 if (in_interrupt() || (gfp_mask & __GFP_THISNODE))
2494 return 1;
2495 might_sleep_if(!(gfp_mask & __GFP_HARDWALL));
2496 if (node_isset(node, current->mems_allowed))
2497 return 1;
2499 * Allow tasks that have access to memory reserves because they have
2500 * been OOM killed to get memory anywhere.
2502 if (unlikely(test_thread_flag(TIF_MEMDIE)))
2503 return 1;
2504 if (gfp_mask & __GFP_HARDWALL) /* If hardwall request, stop here */
2505 return 0;
2507 if (current->flags & PF_EXITING) /* Let dying task have memory */
2508 return 1;
2510 /* Not hardwall and node outside mems_allowed: scan up cpusets */
2511 mutex_lock(&callback_mutex);
2513 task_lock(current);
2514 cs = nearest_hardwall_ancestor(task_cs(current));
2515 task_unlock(current);
2517 allowed = node_isset(node, cs->mems_allowed);
2518 mutex_unlock(&callback_mutex);
2519 return allowed;
2523 * cpuset_node_allowed_hardwall - Can we allocate on a memory node?
2524 * @node: is this an allowed node?
2525 * @gfp_mask: memory allocation flags
2527 * If we're in interrupt, yes, we can always allocate. If __GFP_THISNODE is
2528 * set, yes, we can always allocate. If node is in our task's mems_allowed,
2529 * yes. If the task has been OOM killed and has access to memory reserves as
2530 * specified by the TIF_MEMDIE flag, yes.
2531 * Otherwise, no.
2533 * The __GFP_THISNODE placement logic is really handled elsewhere,
2534 * by forcibly using a zonelist starting at a specified node, and by
2535 * (in get_page_from_freelist()) refusing to consider the zones for
2536 * any node on the zonelist except the first. By the time any such
2537 * calls get to this routine, we should just shut up and say 'yes'.
2539 * Unlike the cpuset_node_allowed_softwall() variant, above,
2540 * this variant requires that the node be in the current task's
2541 * mems_allowed or that we're in interrupt. It does not scan up the
2542 * cpuset hierarchy for the nearest enclosing mem_exclusive cpuset.
2543 * It never sleeps.
2545 int __cpuset_node_allowed_hardwall(int node, gfp_t gfp_mask)
2547 if (in_interrupt() || (gfp_mask & __GFP_THISNODE))
2548 return 1;
2549 if (node_isset(node, current->mems_allowed))
2550 return 1;
2552 * Allow tasks that have access to memory reserves because they have
2553 * been OOM killed to get memory anywhere.
2555 if (unlikely(test_thread_flag(TIF_MEMDIE)))
2556 return 1;
2557 return 0;
2561 * cpuset_mem_spread_node() - On which node to begin search for a file page
2562 * cpuset_slab_spread_node() - On which node to begin search for a slab page
2564 * If a task is marked PF_SPREAD_PAGE or PF_SPREAD_SLAB (as for
2565 * tasks in a cpuset with is_spread_page or is_spread_slab set),
2566 * and if the memory allocation used cpuset_mem_spread_node()
2567 * to determine on which node to start looking, as it will for
2568 * certain page cache or slab cache pages such as used for file
2569 * system buffers and inode caches, then instead of starting on the
2570 * local node to look for a free page, rather spread the starting
2571 * node around the tasks mems_allowed nodes.
2573 * We don't have to worry about the returned node being offline
2574 * because "it can't happen", and even if it did, it would be ok.
2576 * The routines calling guarantee_online_mems() are careful to
2577 * only set nodes in task->mems_allowed that are online. So it
2578 * should not be possible for the following code to return an
2579 * offline node. But if it did, that would be ok, as this routine
2580 * is not returning the node where the allocation must be, only
2581 * the node where the search should start. The zonelist passed to
2582 * __alloc_pages() will include all nodes. If the slab allocator
2583 * is passed an offline node, it will fall back to the local node.
2584 * See kmem_cache_alloc_node().
2587 static int cpuset_spread_node(int *rotor)
2589 int node;
2591 node = next_node(*rotor, current->mems_allowed);
2592 if (node == MAX_NUMNODES)
2593 node = first_node(current->mems_allowed);
2594 *rotor = node;
2595 return node;
2598 int cpuset_mem_spread_node(void)
2600 if (current->cpuset_mem_spread_rotor == NUMA_NO_NODE)
2601 current->cpuset_mem_spread_rotor =
2602 node_random(&current->mems_allowed);
2604 return cpuset_spread_node(&current->cpuset_mem_spread_rotor);
2607 int cpuset_slab_spread_node(void)
2609 if (current->cpuset_slab_spread_rotor == NUMA_NO_NODE)
2610 current->cpuset_slab_spread_rotor =
2611 node_random(&current->mems_allowed);
2613 return cpuset_spread_node(&current->cpuset_slab_spread_rotor);
2616 EXPORT_SYMBOL_GPL(cpuset_mem_spread_node);
2619 * cpuset_mems_allowed_intersects - Does @tsk1's mems_allowed intersect @tsk2's?
2620 * @tsk1: pointer to task_struct of some task.
2621 * @tsk2: pointer to task_struct of some other task.
2623 * Description: Return true if @tsk1's mems_allowed intersects the
2624 * mems_allowed of @tsk2. Used by the OOM killer to determine if
2625 * one of the task's memory usage might impact the memory available
2626 * to the other.
2629 int cpuset_mems_allowed_intersects(const struct task_struct *tsk1,
2630 const struct task_struct *tsk2)
2632 return nodes_intersects(tsk1->mems_allowed, tsk2->mems_allowed);
2635 #define CPUSET_NODELIST_LEN (256)
2638 * cpuset_print_task_mems_allowed - prints task's cpuset and mems_allowed
2639 * @task: pointer to task_struct of some task.
2641 * Description: Prints @task's name, cpuset name, and cached copy of its
2642 * mems_allowed to the kernel log. Must hold task_lock(task) to allow
2643 * dereferencing task_cs(task).
2645 void cpuset_print_task_mems_allowed(struct task_struct *tsk)
2647 /* Statically allocated to prevent using excess stack. */
2648 static char cpuset_nodelist[CPUSET_NODELIST_LEN];
2649 static DEFINE_SPINLOCK(cpuset_buffer_lock);
2651 struct cgroup *cgrp = task_cs(tsk)->css.cgroup;
2653 rcu_read_lock();
2654 spin_lock(&cpuset_buffer_lock);
2656 nodelist_scnprintf(cpuset_nodelist, CPUSET_NODELIST_LEN,
2657 tsk->mems_allowed);
2658 printk(KERN_INFO "%s cpuset=%s mems_allowed=%s\n",
2659 tsk->comm, cgroup_name(cgrp), cpuset_nodelist);
2661 spin_unlock(&cpuset_buffer_lock);
2662 rcu_read_unlock();
2666 * Collection of memory_pressure is suppressed unless
2667 * this flag is enabled by writing "1" to the special
2668 * cpuset file 'memory_pressure_enabled' in the root cpuset.
2671 int cpuset_memory_pressure_enabled __read_mostly;
2674 * cpuset_memory_pressure_bump - keep stats of per-cpuset reclaims.
2676 * Keep a running average of the rate of synchronous (direct)
2677 * page reclaim efforts initiated by tasks in each cpuset.
2679 * This represents the rate at which some task in the cpuset
2680 * ran low on memory on all nodes it was allowed to use, and
2681 * had to enter the kernels page reclaim code in an effort to
2682 * create more free memory by tossing clean pages or swapping
2683 * or writing dirty pages.
2685 * Display to user space in the per-cpuset read-only file
2686 * "memory_pressure". Value displayed is an integer
2687 * representing the recent rate of entry into the synchronous
2688 * (direct) page reclaim by any task attached to the cpuset.
2691 void __cpuset_memory_pressure_bump(void)
2693 task_lock(current);
2694 fmeter_markevent(&task_cs(current)->fmeter);
2695 task_unlock(current);
2698 #ifdef CONFIG_PROC_PID_CPUSET
2700 * proc_cpuset_show()
2701 * - Print tasks cpuset path into seq_file.
2702 * - Used for /proc/<pid>/cpuset.
2703 * - No need to task_lock(tsk) on this tsk->cpuset reference, as it
2704 * doesn't really matter if tsk->cpuset changes after we read it,
2705 * and we take cpuset_mutex, keeping cpuset_attach() from changing it
2706 * anyway.
2708 int proc_cpuset_show(struct seq_file *m, void *unused_v)
2710 struct pid *pid;
2711 struct task_struct *tsk;
2712 char *buf;
2713 struct cgroup_subsys_state *css;
2714 int retval;
2716 retval = -ENOMEM;
2717 buf = kmalloc(PAGE_SIZE, GFP_KERNEL);
2718 if (!buf)
2719 goto out;
2721 retval = -ESRCH;
2722 pid = m->private;
2723 tsk = get_pid_task(pid, PIDTYPE_PID);
2724 if (!tsk)
2725 goto out_free;
2727 rcu_read_lock();
2728 css = task_subsys_state(tsk, cpuset_subsys_id);
2729 retval = cgroup_path(css->cgroup, buf, PAGE_SIZE);
2730 rcu_read_unlock();
2731 if (retval < 0)
2732 goto out_put_task;
2733 seq_puts(m, buf);
2734 seq_putc(m, '\n');
2735 out_put_task:
2736 put_task_struct(tsk);
2737 out_free:
2738 kfree(buf);
2739 out:
2740 return retval;
2742 #endif /* CONFIG_PROC_PID_CPUSET */
2744 /* Display task mems_allowed in /proc/<pid>/status file. */
2745 void cpuset_task_status_allowed(struct seq_file *m, struct task_struct *task)
2747 seq_printf(m, "Mems_allowed:\t");
2748 seq_nodemask(m, &task->mems_allowed);
2749 seq_printf(m, "\n");
2750 seq_printf(m, "Mems_allowed_list:\t");
2751 seq_nodemask_list(m, &task->mems_allowed);
2752 seq_printf(m, "\n");