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
18 * This file is subject to the terms and conditions of the GNU General Public
19 * License. See the file COPYING in the main directory of the Linux
20 * distribution for more details.
23 #include <linux/cpu.h>
24 #include <linux/cpumask.h>
25 #include <linux/cpuset.h>
26 #include <linux/err.h>
27 #include <linux/errno.h>
28 #include <linux/file.h>
30 #include <linux/init.h>
31 #include <linux/interrupt.h>
32 #include <linux/kernel.h>
33 #include <linux/kmod.h>
34 #include <linux/list.h>
35 #include <linux/mempolicy.h>
37 #include <linux/module.h>
38 #include <linux/mount.h>
39 #include <linux/namei.h>
40 #include <linux/pagemap.h>
41 #include <linux/proc_fs.h>
42 #include <linux/rcupdate.h>
43 #include <linux/sched.h>
44 #include <linux/seq_file.h>
45 #include <linux/security.h>
46 #include <linux/slab.h>
47 #include <linux/spinlock.h>
48 #include <linux/stat.h>
49 #include <linux/string.h>
50 #include <linux/time.h>
51 #include <linux/backing-dev.h>
52 #include <linux/sort.h>
54 #include <asm/uaccess.h>
55 #include <asm/atomic.h>
56 #include <linux/mutex.h>
57 #include <linux/kfifo.h>
60 * Tracks how many cpusets are currently defined in system.
61 * When there is only one cpuset (the root cpuset) we can
62 * short circuit some hooks.
64 int number_of_cpusets __read_mostly;
66 /* Retrieve the cpuset from a cgroup */
67 struct cgroup_subsys cpuset_subsys;
70 /* See "Frequency meter" comments, below. */
73 int cnt; /* unprocessed events count */
74 int val; /* most recent output value */
75 time_t time; /* clock (secs) when val computed */
76 spinlock_t lock; /* guards read or write of above */
80 struct cgroup_subsys_state css;
82 unsigned long flags; /* "unsigned long" so bitops work */
83 cpumask_t cpus_allowed; /* CPUs allowed to tasks in cpuset */
84 nodemask_t mems_allowed; /* Memory Nodes allowed to tasks */
86 struct cpuset *parent; /* my parent */
89 * Copy of global cpuset_mems_generation as of the most
90 * recent time this cpuset changed its mems_allowed.
94 struct fmeter fmeter; /* memory_pressure filter */
96 /* partition number for rebuild_sched_domains() */
100 /* Retrieve the cpuset for a cgroup */
101 static inline struct cpuset *cgroup_cs(struct cgroup *cont)
103 return container_of(cgroup_subsys_state(cont, cpuset_subsys_id),
107 /* Retrieve the cpuset for a task */
108 static inline struct cpuset *task_cs(struct task_struct *task)
110 return container_of(task_subsys_state(task, cpuset_subsys_id),
115 /* bits in struct cpuset flags field */
120 CS_SCHED_LOAD_BALANCE,
125 /* convenient tests for these bits */
126 static inline int is_cpu_exclusive(const struct cpuset *cs)
128 return test_bit(CS_CPU_EXCLUSIVE, &cs->flags);
131 static inline int is_mem_exclusive(const struct cpuset *cs)
133 return test_bit(CS_MEM_EXCLUSIVE, &cs->flags);
136 static inline int is_sched_load_balance(const struct cpuset *cs)
138 return test_bit(CS_SCHED_LOAD_BALANCE, &cs->flags);
141 static inline int is_memory_migrate(const struct cpuset *cs)
143 return test_bit(CS_MEMORY_MIGRATE, &cs->flags);
146 static inline int is_spread_page(const struct cpuset *cs)
148 return test_bit(CS_SPREAD_PAGE, &cs->flags);
151 static inline int is_spread_slab(const struct cpuset *cs)
153 return test_bit(CS_SPREAD_SLAB, &cs->flags);
157 * Increment this integer everytime any cpuset changes its
158 * mems_allowed value. Users of cpusets can track this generation
159 * number, and avoid having to lock and reload mems_allowed unless
160 * the cpuset they're using changes generation.
162 * A single, global generation is needed because attach_task() could
163 * reattach a task to a different cpuset, which must not have its
164 * generation numbers aliased with those of that tasks previous cpuset.
166 * Generations are needed for mems_allowed because one task cannot
167 * modify anothers memory placement. So we must enable every task,
168 * on every visit to __alloc_pages(), to efficiently check whether
169 * its current->cpuset->mems_allowed has changed, requiring an update
170 * of its current->mems_allowed.
172 * Since cpuset_mems_generation is guarded by manage_mutex,
173 * there is no need to mark it atomic.
175 static int cpuset_mems_generation;
177 static struct cpuset top_cpuset = {
178 .flags = ((1 << CS_CPU_EXCLUSIVE) | (1 << CS_MEM_EXCLUSIVE)),
179 .cpus_allowed = CPU_MASK_ALL,
180 .mems_allowed = NODE_MASK_ALL,
184 * We have two global cpuset mutexes below. They can nest.
185 * It is ok to first take manage_mutex, then nest callback_mutex. We also
186 * require taking task_lock() when dereferencing a tasks cpuset pointer.
187 * See "The task_lock() exception", at the end of this comment.
189 * A task must hold both mutexes to modify cpusets. If a task
190 * holds manage_mutex, then it blocks others wanting that mutex,
191 * ensuring that it is the only task able to also acquire callback_mutex
192 * and be able to modify cpusets. It can perform various checks on
193 * the cpuset structure first, knowing nothing will change. It can
194 * also allocate memory while just holding manage_mutex. While it is
195 * performing these checks, various callback routines can briefly
196 * acquire callback_mutex to query cpusets. Once it is ready to make
197 * the changes, it takes callback_mutex, blocking everyone else.
199 * Calls to the kernel memory allocator can not be made while holding
200 * callback_mutex, as that would risk double tripping on callback_mutex
201 * from one of the callbacks into the cpuset code from within
204 * If a task is only holding callback_mutex, then it has read-only
207 * The task_struct fields mems_allowed and mems_generation may only
208 * be accessed in the context of that task, so require no locks.
210 * Any task can increment and decrement the count field without lock.
211 * So in general, code holding manage_mutex or callback_mutex can't rely
212 * on the count field not changing. However, if the count goes to
213 * zero, then only attach_task(), which holds both mutexes, can
214 * increment it again. Because a count of zero means that no tasks
215 * are currently attached, therefore there is no way a task attached
216 * to that cpuset can fork (the other way to increment the count).
217 * So code holding manage_mutex or callback_mutex can safely assume that
218 * if the count is zero, it will stay zero. Similarly, if a task
219 * holds manage_mutex or callback_mutex on a cpuset with zero count, it
220 * knows that the cpuset won't be removed, as cpuset_rmdir() needs
221 * both of those mutexes.
223 * The cpuset_common_file_write handler for operations that modify
224 * the cpuset hierarchy holds manage_mutex across the entire operation,
225 * single threading all such cpuset modifications across the system.
227 * The cpuset_common_file_read() handlers only hold callback_mutex across
228 * small pieces of code, such as when reading out possibly multi-word
229 * cpumasks and nodemasks.
231 * The fork and exit callbacks cpuset_fork() and cpuset_exit(), don't
232 * (usually) take either mutex. These are the two most performance
233 * critical pieces of code here. The exception occurs on cpuset_exit(),
234 * when a task in a notify_on_release cpuset exits. Then manage_mutex
235 * is taken, and if the cpuset count is zero, a usermode call made
236 * to /sbin/cpuset_release_agent with the name of the cpuset (path
237 * relative to the root of cpuset file system) as the argument.
239 * A cpuset can only be deleted if both its 'count' of using tasks
240 * is zero, and its list of 'children' cpusets is empty. Since all
241 * tasks in the system use _some_ cpuset, and since there is always at
242 * least one task in the system (init), therefore, top_cpuset
243 * always has either children cpusets and/or using tasks. So we don't
244 * need a special hack to ensure that top_cpuset cannot be deleted.
246 * The above "Tale of Two Semaphores" would be complete, but for:
248 * The task_lock() exception
250 * The need for this exception arises from the action of attach_task(),
251 * which overwrites one tasks cpuset pointer with another. It does
252 * so using both mutexes, however there are several performance
253 * critical places that need to reference task->cpuset without the
254 * expense of grabbing a system global mutex. Therefore except as
255 * noted below, when dereferencing or, as in attach_task(), modifying
256 * a tasks cpuset pointer we use task_lock(), which acts on a spinlock
257 * (task->alloc_lock) already in the task_struct routinely used for
260 * P.S. One more locking exception. RCU is used to guard the
261 * update of a tasks cpuset pointer by attach_task() and the
262 * access of task->cpuset->mems_generation via that pointer in
263 * the routine cpuset_update_task_memory_state().
266 static DEFINE_MUTEX(callback_mutex);
268 /* This is ugly, but preserves the userspace API for existing cpuset
269 * users. If someone tries to mount the "cpuset" filesystem, we
270 * silently switch it to mount "cgroup" instead */
271 static int cpuset_get_sb(struct file_system_type *fs_type,
272 int flags, const char *unused_dev_name,
273 void *data, struct vfsmount *mnt)
275 struct file_system_type *cgroup_fs = get_fs_type("cgroup");
280 "release_agent=/sbin/cpuset_release_agent";
281 ret = cgroup_fs->get_sb(cgroup_fs, flags,
282 unused_dev_name, mountopts, mnt);
283 put_filesystem(cgroup_fs);
288 static struct file_system_type cpuset_fs_type = {
290 .get_sb = cpuset_get_sb,
294 * Return in *pmask the portion of a cpusets's cpus_allowed that
295 * are online. If none are online, walk up the cpuset hierarchy
296 * until we find one that does have some online cpus. If we get
297 * all the way to the top and still haven't found any online cpus,
298 * return cpu_online_map. Or if passed a NULL cs from an exit'ing
299 * task, return cpu_online_map.
301 * One way or another, we guarantee to return some non-empty subset
304 * Call with callback_mutex held.
307 static void guarantee_online_cpus(const struct cpuset *cs, cpumask_t *pmask)
309 while (cs && !cpus_intersects(cs->cpus_allowed, cpu_online_map))
312 cpus_and(*pmask, cs->cpus_allowed, cpu_online_map);
314 *pmask = cpu_online_map;
315 BUG_ON(!cpus_intersects(*pmask, cpu_online_map));
319 * Return in *pmask the portion of a cpusets's mems_allowed that
320 * are online, with memory. If none are online with memory, walk
321 * up the cpuset hierarchy until we find one that does have some
322 * online mems. If we get all the way to the top and still haven't
323 * found any online mems, return node_states[N_HIGH_MEMORY].
325 * One way or another, we guarantee to return some non-empty subset
326 * of node_states[N_HIGH_MEMORY].
328 * Call with callback_mutex held.
331 static void guarantee_online_mems(const struct cpuset *cs, nodemask_t *pmask)
333 while (cs && !nodes_intersects(cs->mems_allowed,
334 node_states[N_HIGH_MEMORY]))
337 nodes_and(*pmask, cs->mems_allowed,
338 node_states[N_HIGH_MEMORY]);
340 *pmask = node_states[N_HIGH_MEMORY];
341 BUG_ON(!nodes_intersects(*pmask, node_states[N_HIGH_MEMORY]));
345 * cpuset_update_task_memory_state - update task memory placement
347 * If the current tasks cpusets mems_allowed changed behind our
348 * backs, update current->mems_allowed, mems_generation and task NUMA
349 * mempolicy to the new value.
351 * Task mempolicy is updated by rebinding it relative to the
352 * current->cpuset if a task has its memory placement changed.
353 * Do not call this routine if in_interrupt().
355 * Call without callback_mutex or task_lock() held. May be
356 * called with or without manage_mutex held. Thanks in part to
357 * 'the_top_cpuset_hack', the tasks cpuset pointer will never
358 * be NULL. This routine also might acquire callback_mutex and
359 * current->mm->mmap_sem during call.
361 * Reading current->cpuset->mems_generation doesn't need task_lock
362 * to guard the current->cpuset derefence, because it is guarded
363 * from concurrent freeing of current->cpuset by attach_task(),
366 * The rcu_dereference() is technically probably not needed,
367 * as I don't actually mind if I see a new cpuset pointer but
368 * an old value of mems_generation. However this really only
369 * matters on alpha systems using cpusets heavily. If I dropped
370 * that rcu_dereference(), it would save them a memory barrier.
371 * For all other arch's, rcu_dereference is a no-op anyway, and for
372 * alpha systems not using cpusets, another planned optimization,
373 * avoiding the rcu critical section for tasks in the root cpuset
374 * which is statically allocated, so can't vanish, will make this
375 * irrelevant. Better to use RCU as intended, than to engage in
376 * some cute trick to save a memory barrier that is impossible to
377 * test, for alpha systems using cpusets heavily, which might not
380 * This routine is needed to update the per-task mems_allowed data,
381 * within the tasks context, when it is trying to allocate memory
382 * (in various mm/mempolicy.c routines) and notices that some other
383 * task has been modifying its cpuset.
386 void cpuset_update_task_memory_state(void)
388 int my_cpusets_mem_gen;
389 struct task_struct *tsk = current;
392 if (task_cs(tsk) == &top_cpuset) {
393 /* Don't need rcu for top_cpuset. It's never freed. */
394 my_cpusets_mem_gen = top_cpuset.mems_generation;
397 my_cpusets_mem_gen = task_cs(current)->mems_generation;
401 if (my_cpusets_mem_gen != tsk->cpuset_mems_generation) {
402 mutex_lock(&callback_mutex);
404 cs = task_cs(tsk); /* Maybe changed when task not locked */
405 guarantee_online_mems(cs, &tsk->mems_allowed);
406 tsk->cpuset_mems_generation = cs->mems_generation;
407 if (is_spread_page(cs))
408 tsk->flags |= PF_SPREAD_PAGE;
410 tsk->flags &= ~PF_SPREAD_PAGE;
411 if (is_spread_slab(cs))
412 tsk->flags |= PF_SPREAD_SLAB;
414 tsk->flags &= ~PF_SPREAD_SLAB;
416 mutex_unlock(&callback_mutex);
417 mpol_rebind_task(tsk, &tsk->mems_allowed);
422 * is_cpuset_subset(p, q) - Is cpuset p a subset of cpuset q?
424 * One cpuset is a subset of another if all its allowed CPUs and
425 * Memory Nodes are a subset of the other, and its exclusive flags
426 * are only set if the other's are set. Call holding manage_mutex.
429 static int is_cpuset_subset(const struct cpuset *p, const struct cpuset *q)
431 return cpus_subset(p->cpus_allowed, q->cpus_allowed) &&
432 nodes_subset(p->mems_allowed, q->mems_allowed) &&
433 is_cpu_exclusive(p) <= is_cpu_exclusive(q) &&
434 is_mem_exclusive(p) <= is_mem_exclusive(q);
438 * validate_change() - Used to validate that any proposed cpuset change
439 * follows the structural rules for cpusets.
441 * If we replaced the flag and mask values of the current cpuset
442 * (cur) with those values in the trial cpuset (trial), would
443 * our various subset and exclusive rules still be valid? Presumes
446 * 'cur' is the address of an actual, in-use cpuset. Operations
447 * such as list traversal that depend on the actual address of the
448 * cpuset in the list must use cur below, not trial.
450 * 'trial' is the address of bulk structure copy of cur, with
451 * perhaps one or more of the fields cpus_allowed, mems_allowed,
452 * or flags changed to new, trial values.
454 * Return 0 if valid, -errno if not.
457 static int validate_change(const struct cpuset *cur, const struct cpuset *trial)
460 struct cpuset *c, *par;
462 /* Each of our child cpusets must be a subset of us */
463 list_for_each_entry(cont, &cur->css.cgroup->children, sibling) {
464 if (!is_cpuset_subset(cgroup_cs(cont), trial))
468 /* Remaining checks don't apply to root cpuset */
469 if (cur == &top_cpuset)
474 /* We must be a subset of our parent cpuset */
475 if (!is_cpuset_subset(trial, par))
478 /* If either I or some sibling (!= me) is exclusive, we can't overlap */
479 list_for_each_entry(cont, &par->css.cgroup->children, sibling) {
481 if ((is_cpu_exclusive(trial) || is_cpu_exclusive(c)) &&
483 cpus_intersects(trial->cpus_allowed, c->cpus_allowed))
485 if ((is_mem_exclusive(trial) || is_mem_exclusive(c)) &&
487 nodes_intersects(trial->mems_allowed, c->mems_allowed))
495 * Helper routine for rebuild_sched_domains().
496 * Do cpusets a, b have overlapping cpus_allowed masks?
499 static int cpusets_overlap(struct cpuset *a, struct cpuset *b)
501 return cpus_intersects(a->cpus_allowed, b->cpus_allowed);
505 * rebuild_sched_domains()
507 * If the flag 'sched_load_balance' of any cpuset with non-empty
508 * 'cpus' changes, or if the 'cpus' allowed changes in any cpuset
509 * which has that flag enabled, or if any cpuset with a non-empty
510 * 'cpus' is removed, then call this routine to rebuild the
511 * scheduler's dynamic sched domains.
513 * This routine builds a partial partition of the systems CPUs
514 * (the set of non-overlappping cpumask_t's in the array 'part'
515 * below), and passes that partial partition to the kernel/sched.c
516 * partition_sched_domains() routine, which will rebuild the
517 * schedulers load balancing domains (sched domains) as specified
518 * by that partial partition. A 'partial partition' is a set of
519 * non-overlapping subsets whose union is a subset of that set.
521 * See "What is sched_load_balance" in Documentation/cpusets.txt
522 * for a background explanation of this.
524 * Does not return errors, on the theory that the callers of this
525 * routine would rather not worry about failures to rebuild sched
526 * domains when operating in the severe memory shortage situations
527 * that could cause allocation failures below.
529 * Call with cgroup_mutex held. May take callback_mutex during
530 * call due to the kfifo_alloc() and kmalloc() calls. May nest
531 * a call to the lock_cpu_hotplug()/unlock_cpu_hotplug() pair.
532 * Must not be called holding callback_mutex, because we must not
533 * call lock_cpu_hotplug() while holding callback_mutex. Elsewhere
534 * the kernel nests callback_mutex inside lock_cpu_hotplug() calls.
535 * So the reverse nesting would risk an ABBA deadlock.
537 * The three key local variables below are:
538 * q - a kfifo queue of cpuset pointers, used to implement a
539 * top-down scan of all cpusets. This scan loads a pointer
540 * to each cpuset marked is_sched_load_balance into the
541 * array 'csa'. For our purposes, rebuilding the schedulers
542 * sched domains, we can ignore !is_sched_load_balance cpusets.
543 * csa - (for CpuSet Array) Array of pointers to all the cpusets
544 * that need to be load balanced, for convenient iterative
545 * access by the subsequent code that finds the best partition,
546 * i.e the set of domains (subsets) of CPUs such that the
547 * cpus_allowed of every cpuset marked is_sched_load_balance
548 * is a subset of one of these domains, while there are as
549 * many such domains as possible, each as small as possible.
550 * doms - Conversion of 'csa' to an array of cpumasks, for passing to
551 * the kernel/sched.c routine partition_sched_domains() in a
552 * convenient format, that can be easily compared to the prior
553 * value to determine what partition elements (sched domains)
554 * were changed (added or removed.)
556 * Finding the best partition (set of domains):
557 * The triple nested loops below over i, j, k scan over the
558 * load balanced cpusets (using the array of cpuset pointers in
559 * csa[]) looking for pairs of cpusets that have overlapping
560 * cpus_allowed, but which don't have the same 'pn' partition
561 * number and gives them in the same partition number. It keeps
562 * looping on the 'restart' label until it can no longer find
565 * The union of the cpus_allowed masks from the set of
566 * all cpusets having the same 'pn' value then form the one
567 * element of the partition (one sched domain) to be passed to
568 * partition_sched_domains().
571 static void rebuild_sched_domains(void)
573 struct kfifo *q; /* queue of cpusets to be scanned */
574 struct cpuset *cp; /* scans q */
575 struct cpuset **csa; /* array of all cpuset ptrs */
576 int csn; /* how many cpuset ptrs in csa so far */
577 int i, j, k; /* indices for partition finding loops */
578 cpumask_t *doms; /* resulting partition; i.e. sched domains */
579 int ndoms; /* number of sched domains in result */
580 int nslot; /* next empty doms[] cpumask_t slot */
586 /* Special case for the 99% of systems with one, full, sched domain */
587 if (is_sched_load_balance(&top_cpuset)) {
589 doms = kmalloc(sizeof(cpumask_t), GFP_KERNEL);
592 *doms = top_cpuset.cpus_allowed;
596 q = kfifo_alloc(number_of_cpusets * sizeof(cp), GFP_KERNEL, NULL);
599 csa = kmalloc(number_of_cpusets * sizeof(cp), GFP_KERNEL);
605 __kfifo_put(q, (void *)&cp, sizeof(cp));
606 while (__kfifo_get(q, (void *)&cp, sizeof(cp))) {
608 struct cpuset *child; /* scans child cpusets of cp */
609 if (is_sched_load_balance(cp))
611 list_for_each_entry(cont, &cp->css.cgroup->children, sibling) {
612 child = cgroup_cs(cont);
613 __kfifo_put(q, (void *)&child, sizeof(cp));
617 for (i = 0; i < csn; i++)
622 /* Find the best partition (set of sched domains) */
623 for (i = 0; i < csn; i++) {
624 struct cpuset *a = csa[i];
627 for (j = 0; j < csn; j++) {
628 struct cpuset *b = csa[j];
631 if (apn != bpn && cpusets_overlap(a, b)) {
632 for (k = 0; k < csn; k++) {
633 struct cpuset *c = csa[k];
638 ndoms--; /* one less element */
644 /* Convert <csn, csa> to <ndoms, doms> */
645 doms = kmalloc(ndoms * sizeof(cpumask_t), GFP_KERNEL);
649 for (nslot = 0, i = 0; i < csn; i++) {
650 struct cpuset *a = csa[i];
654 cpumask_t *dp = doms + nslot;
656 if (nslot == ndoms) {
657 static int warnings = 10;
660 "rebuild_sched_domains confused:"
661 " nslot %d, ndoms %d, csn %d, i %d,"
663 nslot, ndoms, csn, i, apn);
670 for (j = i; j < csn; j++) {
671 struct cpuset *b = csa[j];
674 cpus_or(*dp, *dp, b->cpus_allowed);
681 BUG_ON(nslot != ndoms);
684 /* Have scheduler rebuild sched domains */
686 partition_sched_domains(ndoms, doms);
687 unlock_cpu_hotplug();
693 /* Don't kfree(doms) -- partition_sched_domains() does that. */
697 * Call with manage_mutex held. May take callback_mutex during call.
700 static int update_cpumask(struct cpuset *cs, char *buf)
702 struct cpuset trialcs;
704 int cpus_changed, is_load_balanced;
706 /* top_cpuset.cpus_allowed tracks cpu_online_map; it's read-only */
707 if (cs == &top_cpuset)
713 * We allow a cpuset's cpus_allowed to be empty; if it has attached
714 * tasks, we'll catch it later when we validate the change and return
717 if (!buf[0] || (buf[0] == '\n' && !buf[1])) {
718 cpus_clear(trialcs.cpus_allowed);
720 retval = cpulist_parse(buf, trialcs.cpus_allowed);
724 cpus_and(trialcs.cpus_allowed, trialcs.cpus_allowed, cpu_online_map);
725 /* cpus_allowed cannot be empty for a cpuset with attached tasks. */
726 if (cgroup_task_count(cs->css.cgroup) &&
727 cpus_empty(trialcs.cpus_allowed))
729 retval = validate_change(cs, &trialcs);
733 cpus_changed = !cpus_equal(cs->cpus_allowed, trialcs.cpus_allowed);
734 is_load_balanced = is_sched_load_balance(&trialcs);
736 mutex_lock(&callback_mutex);
737 cs->cpus_allowed = trialcs.cpus_allowed;
738 mutex_unlock(&callback_mutex);
740 if (cpus_changed && is_load_balanced)
741 rebuild_sched_domains();
749 * Migrate memory region from one set of nodes to another.
751 * Temporarilly set tasks mems_allowed to target nodes of migration,
752 * so that the migration code can allocate pages on these nodes.
754 * Call holding manage_mutex, so our current->cpuset won't change
755 * during this call, as manage_mutex holds off any attach_task()
756 * calls. Therefore we don't need to take task_lock around the
757 * call to guarantee_online_mems(), as we know no one is changing
760 * Hold callback_mutex around the two modifications of our tasks
761 * mems_allowed to synchronize with cpuset_mems_allowed().
763 * While the mm_struct we are migrating is typically from some
764 * other task, the task_struct mems_allowed that we are hacking
765 * is for our current task, which must allocate new pages for that
766 * migrating memory region.
768 * We call cpuset_update_task_memory_state() before hacking
769 * our tasks mems_allowed, so that we are assured of being in
770 * sync with our tasks cpuset, and in particular, callbacks to
771 * cpuset_update_task_memory_state() from nested page allocations
772 * won't see any mismatch of our cpuset and task mems_generation
773 * values, so won't overwrite our hacked tasks mems_allowed
777 static void cpuset_migrate_mm(struct mm_struct *mm, const nodemask_t *from,
778 const nodemask_t *to)
780 struct task_struct *tsk = current;
782 cpuset_update_task_memory_state();
784 mutex_lock(&callback_mutex);
785 tsk->mems_allowed = *to;
786 mutex_unlock(&callback_mutex);
788 do_migrate_pages(mm, from, to, MPOL_MF_MOVE_ALL);
790 mutex_lock(&callback_mutex);
791 guarantee_online_mems(task_cs(tsk),&tsk->mems_allowed);
792 mutex_unlock(&callback_mutex);
796 * Handle user request to change the 'mems' memory placement
797 * of a cpuset. Needs to validate the request, update the
798 * cpusets mems_allowed and mems_generation, and for each
799 * task in the cpuset, rebind any vma mempolicies and if
800 * the cpuset is marked 'memory_migrate', migrate the tasks
801 * pages to the new memory.
803 * Call with manage_mutex held. May take callback_mutex during call.
804 * Will take tasklist_lock, scan tasklist for tasks in cpuset cs,
805 * lock each such tasks mm->mmap_sem, scan its vma's and rebind
806 * their mempolicies to the cpusets new mems_allowed.
809 static void *cpuset_being_rebound;
811 static int update_nodemask(struct cpuset *cs, char *buf)
813 struct cpuset trialcs;
815 struct task_struct *p;
816 struct mm_struct **mmarray;
821 struct cgroup_iter it;
824 * top_cpuset.mems_allowed tracks node_stats[N_HIGH_MEMORY];
827 if (cs == &top_cpuset)
833 * We allow a cpuset's mems_allowed to be empty; if it has attached
834 * tasks, we'll catch it later when we validate the change and return
837 if (!buf[0] || (buf[0] == '\n' && !buf[1])) {
838 nodes_clear(trialcs.mems_allowed);
840 retval = nodelist_parse(buf, trialcs.mems_allowed);
843 if (!nodes_intersects(trialcs.mems_allowed,
844 node_states[N_HIGH_MEMORY])) {
846 * error if only memoryless nodes specified.
853 * Exclude memoryless nodes. We know that trialcs.mems_allowed
854 * contains at least one node with memory.
856 nodes_and(trialcs.mems_allowed, trialcs.mems_allowed,
857 node_states[N_HIGH_MEMORY]);
858 oldmem = cs->mems_allowed;
859 if (nodes_equal(oldmem, trialcs.mems_allowed)) {
860 retval = 0; /* Too easy - nothing to do */
863 /* mems_allowed cannot be empty for a cpuset with attached tasks. */
864 if (cgroup_task_count(cs->css.cgroup) &&
865 nodes_empty(trialcs.mems_allowed)) {
869 retval = validate_change(cs, &trialcs);
873 mutex_lock(&callback_mutex);
874 cs->mems_allowed = trialcs.mems_allowed;
875 cs->mems_generation = cpuset_mems_generation++;
876 mutex_unlock(&callback_mutex);
878 cpuset_being_rebound = cs; /* causes mpol_copy() rebind */
880 fudge = 10; /* spare mmarray[] slots */
881 fudge += cpus_weight(cs->cpus_allowed); /* imagine one fork-bomb/cpu */
885 * Allocate mmarray[] to hold mm reference for each task
886 * in cpuset cs. Can't kmalloc GFP_KERNEL while holding
887 * tasklist_lock. We could use GFP_ATOMIC, but with a
888 * few more lines of code, we can retry until we get a big
889 * enough mmarray[] w/o using GFP_ATOMIC.
892 ntasks = cgroup_task_count(cs->css.cgroup); /* guess */
894 mmarray = kmalloc(ntasks * sizeof(*mmarray), GFP_KERNEL);
897 read_lock(&tasklist_lock); /* block fork */
898 if (cgroup_task_count(cs->css.cgroup) <= ntasks)
899 break; /* got enough */
900 read_unlock(&tasklist_lock); /* try again */
906 /* Load up mmarray[] with mm reference for each task in cpuset. */
907 cgroup_iter_start(cs->css.cgroup, &it);
908 while ((p = cgroup_iter_next(cs->css.cgroup, &it))) {
909 struct mm_struct *mm;
913 "Cpuset mempolicy rebind incomplete.\n");
921 cgroup_iter_end(cs->css.cgroup, &it);
922 read_unlock(&tasklist_lock);
925 * Now that we've dropped the tasklist spinlock, we can
926 * rebind the vma mempolicies of each mm in mmarray[] to their
927 * new cpuset, and release that mm. The mpol_rebind_mm()
928 * call takes mmap_sem, which we couldn't take while holding
929 * tasklist_lock. Forks can happen again now - the mpol_copy()
930 * cpuset_being_rebound check will catch such forks, and rebind
931 * their vma mempolicies too. Because we still hold the global
932 * cpuset manage_mutex, we know that no other rebind effort will
933 * be contending for the global variable cpuset_being_rebound.
934 * It's ok if we rebind the same mm twice; mpol_rebind_mm()
935 * is idempotent. Also migrate pages in each mm to new nodes.
937 migrate = is_memory_migrate(cs);
938 for (i = 0; i < n; i++) {
939 struct mm_struct *mm = mmarray[i];
941 mpol_rebind_mm(mm, &cs->mems_allowed);
943 cpuset_migrate_mm(mm, &oldmem, &cs->mems_allowed);
947 /* We're done rebinding vma's to this cpusets new mems_allowed. */
949 cpuset_being_rebound = NULL;
955 int current_cpuset_is_being_rebound(void)
957 return task_cs(current) == cpuset_being_rebound;
961 * Call with manage_mutex held.
964 static int update_memory_pressure_enabled(struct cpuset *cs, char *buf)
966 if (simple_strtoul(buf, NULL, 10) != 0)
967 cpuset_memory_pressure_enabled = 1;
969 cpuset_memory_pressure_enabled = 0;
974 * update_flag - read a 0 or a 1 in a file and update associated flag
975 * bit: the bit to update (CS_CPU_EXCLUSIVE, CS_MEM_EXCLUSIVE,
976 * CS_SCHED_LOAD_BALANCE,
977 * CS_NOTIFY_ON_RELEASE, CS_MEMORY_MIGRATE,
978 * CS_SPREAD_PAGE, CS_SPREAD_SLAB)
979 * cs: the cpuset to update
980 * buf: the buffer where we read the 0 or 1
982 * Call with manage_mutex held.
985 static int update_flag(cpuset_flagbits_t bit, struct cpuset *cs, char *buf)
988 struct cpuset trialcs;
990 int cpus_nonempty, balance_flag_changed;
992 turning_on = (simple_strtoul(buf, NULL, 10) != 0);
996 set_bit(bit, &trialcs.flags);
998 clear_bit(bit, &trialcs.flags);
1000 err = validate_change(cs, &trialcs);
1004 cpus_nonempty = !cpus_empty(trialcs.cpus_allowed);
1005 balance_flag_changed = (is_sched_load_balance(cs) !=
1006 is_sched_load_balance(&trialcs));
1008 mutex_lock(&callback_mutex);
1009 cs->flags = trialcs.flags;
1010 mutex_unlock(&callback_mutex);
1012 if (cpus_nonempty && balance_flag_changed)
1013 rebuild_sched_domains();
1019 * Frequency meter - How fast is some event occurring?
1021 * These routines manage a digitally filtered, constant time based,
1022 * event frequency meter. There are four routines:
1023 * fmeter_init() - initialize a frequency meter.
1024 * fmeter_markevent() - called each time the event happens.
1025 * fmeter_getrate() - returns the recent rate of such events.
1026 * fmeter_update() - internal routine used to update fmeter.
1028 * A common data structure is passed to each of these routines,
1029 * which is used to keep track of the state required to manage the
1030 * frequency meter and its digital filter.
1032 * The filter works on the number of events marked per unit time.
1033 * The filter is single-pole low-pass recursive (IIR). The time unit
1034 * is 1 second. Arithmetic is done using 32-bit integers scaled to
1035 * simulate 3 decimal digits of precision (multiplied by 1000).
1037 * With an FM_COEF of 933, and a time base of 1 second, the filter
1038 * has a half-life of 10 seconds, meaning that if the events quit
1039 * happening, then the rate returned from the fmeter_getrate()
1040 * will be cut in half each 10 seconds, until it converges to zero.
1042 * It is not worth doing a real infinitely recursive filter. If more
1043 * than FM_MAXTICKS ticks have elapsed since the last filter event,
1044 * just compute FM_MAXTICKS ticks worth, by which point the level
1047 * Limit the count of unprocessed events to FM_MAXCNT, so as to avoid
1048 * arithmetic overflow in the fmeter_update() routine.
1050 * Given the simple 32 bit integer arithmetic used, this meter works
1051 * best for reporting rates between one per millisecond (msec) and
1052 * one per 32 (approx) seconds. At constant rates faster than one
1053 * per msec it maxes out at values just under 1,000,000. At constant
1054 * rates between one per msec, and one per second it will stabilize
1055 * to a value N*1000, where N is the rate of events per second.
1056 * At constant rates between one per second and one per 32 seconds,
1057 * it will be choppy, moving up on the seconds that have an event,
1058 * and then decaying until the next event. At rates slower than
1059 * about one in 32 seconds, it decays all the way back to zero between
1063 #define FM_COEF 933 /* coefficient for half-life of 10 secs */
1064 #define FM_MAXTICKS ((time_t)99) /* useless computing more ticks than this */
1065 #define FM_MAXCNT 1000000 /* limit cnt to avoid overflow */
1066 #define FM_SCALE 1000 /* faux fixed point scale */
1068 /* Initialize a frequency meter */
1069 static void fmeter_init(struct fmeter *fmp)
1074 spin_lock_init(&fmp->lock);
1077 /* Internal meter update - process cnt events and update value */
1078 static void fmeter_update(struct fmeter *fmp)
1080 time_t now = get_seconds();
1081 time_t ticks = now - fmp->time;
1086 ticks = min(FM_MAXTICKS, ticks);
1088 fmp->val = (FM_COEF * fmp->val) / FM_SCALE;
1091 fmp->val += ((FM_SCALE - FM_COEF) * fmp->cnt) / FM_SCALE;
1095 /* Process any previous ticks, then bump cnt by one (times scale). */
1096 static void fmeter_markevent(struct fmeter *fmp)
1098 spin_lock(&fmp->lock);
1100 fmp->cnt = min(FM_MAXCNT, fmp->cnt + FM_SCALE);
1101 spin_unlock(&fmp->lock);
1104 /* Process any previous ticks, then return current value. */
1105 static int fmeter_getrate(struct fmeter *fmp)
1109 spin_lock(&fmp->lock);
1112 spin_unlock(&fmp->lock);
1116 static int cpuset_can_attach(struct cgroup_subsys *ss,
1117 struct cgroup *cont, struct task_struct *tsk)
1119 struct cpuset *cs = cgroup_cs(cont);
1121 if (cpus_empty(cs->cpus_allowed) || nodes_empty(cs->mems_allowed))
1124 return security_task_setscheduler(tsk, 0, NULL);
1127 static void cpuset_attach(struct cgroup_subsys *ss,
1128 struct cgroup *cont, struct cgroup *oldcont,
1129 struct task_struct *tsk)
1132 nodemask_t from, to;
1133 struct mm_struct *mm;
1134 struct cpuset *cs = cgroup_cs(cont);
1135 struct cpuset *oldcs = cgroup_cs(oldcont);
1137 mutex_lock(&callback_mutex);
1138 guarantee_online_cpus(cs, &cpus);
1139 set_cpus_allowed(tsk, cpus);
1140 mutex_unlock(&callback_mutex);
1142 from = oldcs->mems_allowed;
1143 to = cs->mems_allowed;
1144 mm = get_task_mm(tsk);
1146 mpol_rebind_mm(mm, &to);
1147 if (is_memory_migrate(cs))
1148 cpuset_migrate_mm(mm, &from, &to);
1154 /* The various types of files and directories in a cpuset file system */
1157 FILE_MEMORY_MIGRATE,
1162 FILE_SCHED_LOAD_BALANCE,
1163 FILE_MEMORY_PRESSURE_ENABLED,
1164 FILE_MEMORY_PRESSURE,
1167 } cpuset_filetype_t;
1169 static ssize_t cpuset_common_file_write(struct cgroup *cont,
1172 const char __user *userbuf,
1173 size_t nbytes, loff_t *unused_ppos)
1175 struct cpuset *cs = cgroup_cs(cont);
1176 cpuset_filetype_t type = cft->private;
1180 /* Crude upper limit on largest legitimate cpulist user might write. */
1181 if (nbytes > 100U + 6 * max(NR_CPUS, MAX_NUMNODES))
1184 /* +1 for nul-terminator */
1185 if ((buffer = kmalloc(nbytes + 1, GFP_KERNEL)) == 0)
1188 if (copy_from_user(buffer, userbuf, nbytes)) {
1192 buffer[nbytes] = 0; /* nul-terminate */
1196 if (cgroup_is_removed(cont)) {
1203 retval = update_cpumask(cs, buffer);
1206 retval = update_nodemask(cs, buffer);
1208 case FILE_CPU_EXCLUSIVE:
1209 retval = update_flag(CS_CPU_EXCLUSIVE, cs, buffer);
1211 case FILE_MEM_EXCLUSIVE:
1212 retval = update_flag(CS_MEM_EXCLUSIVE, cs, buffer);
1214 case FILE_SCHED_LOAD_BALANCE:
1215 retval = update_flag(CS_SCHED_LOAD_BALANCE, cs, buffer);
1217 case FILE_MEMORY_MIGRATE:
1218 retval = update_flag(CS_MEMORY_MIGRATE, cs, buffer);
1220 case FILE_MEMORY_PRESSURE_ENABLED:
1221 retval = update_memory_pressure_enabled(cs, buffer);
1223 case FILE_MEMORY_PRESSURE:
1226 case FILE_SPREAD_PAGE:
1227 retval = update_flag(CS_SPREAD_PAGE, cs, buffer);
1228 cs->mems_generation = cpuset_mems_generation++;
1230 case FILE_SPREAD_SLAB:
1231 retval = update_flag(CS_SPREAD_SLAB, cs, buffer);
1232 cs->mems_generation = cpuset_mems_generation++;
1249 * These ascii lists should be read in a single call, by using a user
1250 * buffer large enough to hold the entire map. If read in smaller
1251 * chunks, there is no guarantee of atomicity. Since the display format
1252 * used, list of ranges of sequential numbers, is variable length,
1253 * and since these maps can change value dynamically, one could read
1254 * gibberish by doing partial reads while a list was changing.
1255 * A single large read to a buffer that crosses a page boundary is
1256 * ok, because the result being copied to user land is not recomputed
1257 * across a page fault.
1260 static int cpuset_sprintf_cpulist(char *page, struct cpuset *cs)
1264 mutex_lock(&callback_mutex);
1265 mask = cs->cpus_allowed;
1266 mutex_unlock(&callback_mutex);
1268 return cpulist_scnprintf(page, PAGE_SIZE, mask);
1271 static int cpuset_sprintf_memlist(char *page, struct cpuset *cs)
1275 mutex_lock(&callback_mutex);
1276 mask = cs->mems_allowed;
1277 mutex_unlock(&callback_mutex);
1279 return nodelist_scnprintf(page, PAGE_SIZE, mask);
1282 static ssize_t cpuset_common_file_read(struct cgroup *cont,
1286 size_t nbytes, loff_t *ppos)
1288 struct cpuset *cs = cgroup_cs(cont);
1289 cpuset_filetype_t type = cft->private;
1294 if (!(page = (char *)__get_free_page(GFP_TEMPORARY)))
1301 s += cpuset_sprintf_cpulist(s, cs);
1304 s += cpuset_sprintf_memlist(s, cs);
1306 case FILE_CPU_EXCLUSIVE:
1307 *s++ = is_cpu_exclusive(cs) ? '1' : '0';
1309 case FILE_MEM_EXCLUSIVE:
1310 *s++ = is_mem_exclusive(cs) ? '1' : '0';
1312 case FILE_SCHED_LOAD_BALANCE:
1313 *s++ = is_sched_load_balance(cs) ? '1' : '0';
1315 case FILE_MEMORY_MIGRATE:
1316 *s++ = is_memory_migrate(cs) ? '1' : '0';
1318 case FILE_MEMORY_PRESSURE_ENABLED:
1319 *s++ = cpuset_memory_pressure_enabled ? '1' : '0';
1321 case FILE_MEMORY_PRESSURE:
1322 s += sprintf(s, "%d", fmeter_getrate(&cs->fmeter));
1324 case FILE_SPREAD_PAGE:
1325 *s++ = is_spread_page(cs) ? '1' : '0';
1327 case FILE_SPREAD_SLAB:
1328 *s++ = is_spread_slab(cs) ? '1' : '0';
1336 retval = simple_read_from_buffer(buf, nbytes, ppos, page, s - page);
1338 free_page((unsigned long)page);
1347 * for the common functions, 'private' gives the type of file
1350 static struct cftype cft_cpus = {
1352 .read = cpuset_common_file_read,
1353 .write = cpuset_common_file_write,
1354 .private = FILE_CPULIST,
1357 static struct cftype cft_mems = {
1359 .read = cpuset_common_file_read,
1360 .write = cpuset_common_file_write,
1361 .private = FILE_MEMLIST,
1364 static struct cftype cft_cpu_exclusive = {
1365 .name = "cpu_exclusive",
1366 .read = cpuset_common_file_read,
1367 .write = cpuset_common_file_write,
1368 .private = FILE_CPU_EXCLUSIVE,
1371 static struct cftype cft_mem_exclusive = {
1372 .name = "mem_exclusive",
1373 .read = cpuset_common_file_read,
1374 .write = cpuset_common_file_write,
1375 .private = FILE_MEM_EXCLUSIVE,
1378 static struct cftype cft_sched_load_balance = {
1379 .name = "sched_load_balance",
1380 .read = cpuset_common_file_read,
1381 .write = cpuset_common_file_write,
1382 .private = FILE_SCHED_LOAD_BALANCE,
1385 static struct cftype cft_memory_migrate = {
1386 .name = "memory_migrate",
1387 .read = cpuset_common_file_read,
1388 .write = cpuset_common_file_write,
1389 .private = FILE_MEMORY_MIGRATE,
1392 static struct cftype cft_memory_pressure_enabled = {
1393 .name = "memory_pressure_enabled",
1394 .read = cpuset_common_file_read,
1395 .write = cpuset_common_file_write,
1396 .private = FILE_MEMORY_PRESSURE_ENABLED,
1399 static struct cftype cft_memory_pressure = {
1400 .name = "memory_pressure",
1401 .read = cpuset_common_file_read,
1402 .write = cpuset_common_file_write,
1403 .private = FILE_MEMORY_PRESSURE,
1406 static struct cftype cft_spread_page = {
1407 .name = "memory_spread_page",
1408 .read = cpuset_common_file_read,
1409 .write = cpuset_common_file_write,
1410 .private = FILE_SPREAD_PAGE,
1413 static struct cftype cft_spread_slab = {
1414 .name = "memory_spread_slab",
1415 .read = cpuset_common_file_read,
1416 .write = cpuset_common_file_write,
1417 .private = FILE_SPREAD_SLAB,
1420 static int cpuset_populate(struct cgroup_subsys *ss, struct cgroup *cont)
1424 if ((err = cgroup_add_file(cont, ss, &cft_cpus)) < 0)
1426 if ((err = cgroup_add_file(cont, ss, &cft_mems)) < 0)
1428 if ((err = cgroup_add_file(cont, ss, &cft_cpu_exclusive)) < 0)
1430 if ((err = cgroup_add_file(cont, ss, &cft_mem_exclusive)) < 0)
1432 if ((err = cgroup_add_file(cont, ss, &cft_memory_migrate)) < 0)
1434 if ((err = cgroup_add_file(cont, ss, &cft_sched_load_balance)) < 0)
1436 if ((err = cgroup_add_file(cont, ss, &cft_memory_pressure)) < 0)
1438 if ((err = cgroup_add_file(cont, ss, &cft_spread_page)) < 0)
1440 if ((err = cgroup_add_file(cont, ss, &cft_spread_slab)) < 0)
1442 /* memory_pressure_enabled is in root cpuset only */
1443 if (err == 0 && !cont->parent)
1444 err = cgroup_add_file(cont, ss,
1445 &cft_memory_pressure_enabled);
1450 * post_clone() is called at the end of cgroup_clone().
1451 * 'cgroup' was just created automatically as a result of
1452 * a cgroup_clone(), and the current task is about to
1453 * be moved into 'cgroup'.
1455 * Currently we refuse to set up the cgroup - thereby
1456 * refusing the task to be entered, and as a result refusing
1457 * the sys_unshare() or clone() which initiated it - if any
1458 * sibling cpusets have exclusive cpus or mem.
1460 * If this becomes a problem for some users who wish to
1461 * allow that scenario, then cpuset_post_clone() could be
1462 * changed to grant parent->cpus_allowed-sibling_cpus_exclusive
1463 * (and likewise for mems) to the new cgroup.
1465 static void cpuset_post_clone(struct cgroup_subsys *ss,
1466 struct cgroup *cgroup)
1468 struct cgroup *parent, *child;
1469 struct cpuset *cs, *parent_cs;
1471 parent = cgroup->parent;
1472 list_for_each_entry(child, &parent->children, sibling) {
1473 cs = cgroup_cs(child);
1474 if (is_mem_exclusive(cs) || is_cpu_exclusive(cs))
1477 cs = cgroup_cs(cgroup);
1478 parent_cs = cgroup_cs(parent);
1480 cs->mems_allowed = parent_cs->mems_allowed;
1481 cs->cpus_allowed = parent_cs->cpus_allowed;
1486 * cpuset_create - create a cpuset
1487 * parent: cpuset that will be parent of the new cpuset.
1488 * name: name of the new cpuset. Will be strcpy'ed.
1489 * mode: mode to set on new inode
1491 * Must be called with the mutex on the parent inode held
1494 static struct cgroup_subsys_state *cpuset_create(
1495 struct cgroup_subsys *ss,
1496 struct cgroup *cont)
1499 struct cpuset *parent;
1501 if (!cont->parent) {
1502 /* This is early initialization for the top cgroup */
1503 top_cpuset.mems_generation = cpuset_mems_generation++;
1504 return &top_cpuset.css;
1506 parent = cgroup_cs(cont->parent);
1507 cs = kmalloc(sizeof(*cs), GFP_KERNEL);
1509 return ERR_PTR(-ENOMEM);
1511 cpuset_update_task_memory_state();
1513 if (is_spread_page(parent))
1514 set_bit(CS_SPREAD_PAGE, &cs->flags);
1515 if (is_spread_slab(parent))
1516 set_bit(CS_SPREAD_SLAB, &cs->flags);
1517 set_bit(CS_SCHED_LOAD_BALANCE, &cs->flags);
1518 cs->cpus_allowed = CPU_MASK_NONE;
1519 cs->mems_allowed = NODE_MASK_NONE;
1520 cs->mems_generation = cpuset_mems_generation++;
1521 fmeter_init(&cs->fmeter);
1523 cs->parent = parent;
1524 number_of_cpusets++;
1529 * Locking note on the strange update_flag() call below:
1531 * If the cpuset being removed has its flag 'sched_load_balance'
1532 * enabled, then simulate turning sched_load_balance off, which
1533 * will call rebuild_sched_domains(). The lock_cpu_hotplug()
1534 * call in rebuild_sched_domains() must not be made while holding
1535 * callback_mutex. Elsewhere the kernel nests callback_mutex inside
1536 * lock_cpu_hotplug() calls. So the reverse nesting would risk an
1540 static void cpuset_destroy(struct cgroup_subsys *ss, struct cgroup *cont)
1542 struct cpuset *cs = cgroup_cs(cont);
1544 cpuset_update_task_memory_state();
1546 if (is_sched_load_balance(cs))
1547 update_flag(CS_SCHED_LOAD_BALANCE, cs, "0");
1549 number_of_cpusets--;
1553 struct cgroup_subsys cpuset_subsys = {
1555 .create = cpuset_create,
1556 .destroy = cpuset_destroy,
1557 .can_attach = cpuset_can_attach,
1558 .attach = cpuset_attach,
1559 .populate = cpuset_populate,
1560 .post_clone = cpuset_post_clone,
1561 .subsys_id = cpuset_subsys_id,
1566 * cpuset_init_early - just enough so that the calls to
1567 * cpuset_update_task_memory_state() in early init code
1571 int __init cpuset_init_early(void)
1573 top_cpuset.mems_generation = cpuset_mems_generation++;
1579 * cpuset_init - initialize cpusets at system boot
1581 * Description: Initialize top_cpuset and the cpuset internal file system,
1584 int __init cpuset_init(void)
1588 top_cpuset.cpus_allowed = CPU_MASK_ALL;
1589 top_cpuset.mems_allowed = NODE_MASK_ALL;
1591 fmeter_init(&top_cpuset.fmeter);
1592 top_cpuset.mems_generation = cpuset_mems_generation++;
1593 set_bit(CS_SCHED_LOAD_BALANCE, &top_cpuset.flags);
1595 err = register_filesystem(&cpuset_fs_type);
1599 number_of_cpusets = 1;
1604 * If common_cpu_mem_hotplug_unplug(), below, unplugs any CPUs
1605 * or memory nodes, we need to walk over the cpuset hierarchy,
1606 * removing that CPU or node from all cpusets. If this removes the
1607 * last CPU or node from a cpuset, then the guarantee_online_cpus()
1608 * or guarantee_online_mems() code will use that emptied cpusets
1609 * parent online CPUs or nodes. Cpusets that were already empty of
1610 * CPUs or nodes are left empty.
1612 * This routine is intentionally inefficient in a couple of regards.
1613 * It will check all cpusets in a subtree even if the top cpuset of
1614 * the subtree has no offline CPUs or nodes. It checks both CPUs and
1615 * nodes, even though the caller could have been coded to know that
1616 * only one of CPUs or nodes needed to be checked on a given call.
1617 * This was done to minimize text size rather than cpu cycles.
1619 * Call with both manage_mutex and callback_mutex held.
1621 * Recursive, on depth of cpuset subtree.
1624 static void guarantee_online_cpus_mems_in_subtree(const struct cpuset *cur)
1626 struct cgroup *cont;
1629 /* Each of our child cpusets mems must be online */
1630 list_for_each_entry(cont, &cur->css.cgroup->children, sibling) {
1631 c = cgroup_cs(cont);
1632 guarantee_online_cpus_mems_in_subtree(c);
1633 if (!cpus_empty(c->cpus_allowed))
1634 guarantee_online_cpus(c, &c->cpus_allowed);
1635 if (!nodes_empty(c->mems_allowed))
1636 guarantee_online_mems(c, &c->mems_allowed);
1641 * The cpus_allowed and mems_allowed nodemasks in the top_cpuset track
1642 * cpu_online_map and node_states[N_HIGH_MEMORY]. Force the top cpuset to
1643 * track what's online after any CPU or memory node hotplug or unplug
1646 * To ensure that we don't remove a CPU or node from the top cpuset
1647 * that is currently in use by a child cpuset (which would violate
1648 * the rule that cpusets must be subsets of their parent), we first
1649 * call the recursive routine guarantee_online_cpus_mems_in_subtree().
1651 * Since there are two callers of this routine, one for CPU hotplug
1652 * events and one for memory node hotplug events, we could have coded
1653 * two separate routines here. We code it as a single common routine
1654 * in order to minimize text size.
1657 static void common_cpu_mem_hotplug_unplug(void)
1660 mutex_lock(&callback_mutex);
1662 guarantee_online_cpus_mems_in_subtree(&top_cpuset);
1663 top_cpuset.cpus_allowed = cpu_online_map;
1664 top_cpuset.mems_allowed = node_states[N_HIGH_MEMORY];
1666 mutex_unlock(&callback_mutex);
1671 * The top_cpuset tracks what CPUs and Memory Nodes are online,
1672 * period. This is necessary in order to make cpusets transparent
1673 * (of no affect) on systems that are actively using CPU hotplug
1674 * but making no active use of cpusets.
1676 * This routine ensures that top_cpuset.cpus_allowed tracks
1677 * cpu_online_map on each CPU hotplug (cpuhp) event.
1680 static int cpuset_handle_cpuhp(struct notifier_block *unused_nb,
1681 unsigned long phase, void *unused_cpu)
1683 if (phase == CPU_DYING || phase == CPU_DYING_FROZEN)
1686 common_cpu_mem_hotplug_unplug();
1690 #ifdef CONFIG_MEMORY_HOTPLUG
1692 * Keep top_cpuset.mems_allowed tracking node_states[N_HIGH_MEMORY].
1693 * Call this routine anytime after you change
1694 * node_states[N_HIGH_MEMORY].
1695 * See also the previous routine cpuset_handle_cpuhp().
1698 void cpuset_track_online_nodes(void)
1700 common_cpu_mem_hotplug_unplug();
1705 * cpuset_init_smp - initialize cpus_allowed
1707 * Description: Finish top cpuset after cpu, node maps are initialized
1710 void __init cpuset_init_smp(void)
1712 top_cpuset.cpus_allowed = cpu_online_map;
1713 top_cpuset.mems_allowed = node_states[N_HIGH_MEMORY];
1715 hotcpu_notifier(cpuset_handle_cpuhp, 0);
1720 * cpuset_cpus_allowed - return cpus_allowed mask from a tasks cpuset.
1721 * @tsk: pointer to task_struct from which to obtain cpuset->cpus_allowed.
1723 * Description: Returns the cpumask_t cpus_allowed of the cpuset
1724 * attached to the specified @tsk. Guaranteed to return some non-empty
1725 * subset of cpu_online_map, even if this means going outside the
1729 cpumask_t cpuset_cpus_allowed(struct task_struct *tsk)
1733 mutex_lock(&callback_mutex);
1735 guarantee_online_cpus(task_cs(tsk), &mask);
1737 mutex_unlock(&callback_mutex);
1742 void cpuset_init_current_mems_allowed(void)
1744 current->mems_allowed = NODE_MASK_ALL;
1748 * cpuset_mems_allowed - return mems_allowed mask from a tasks cpuset.
1749 * @tsk: pointer to task_struct from which to obtain cpuset->mems_allowed.
1751 * Description: Returns the nodemask_t mems_allowed of the cpuset
1752 * attached to the specified @tsk. Guaranteed to return some non-empty
1753 * subset of node_states[N_HIGH_MEMORY], even if this means going outside the
1757 nodemask_t cpuset_mems_allowed(struct task_struct *tsk)
1761 mutex_lock(&callback_mutex);
1763 guarantee_online_mems(task_cs(tsk), &mask);
1765 mutex_unlock(&callback_mutex);
1771 * cpuset_zonelist_valid_mems_allowed - check zonelist vs. curremt mems_allowed
1772 * @zl: the zonelist to be checked
1774 * Are any of the nodes on zonelist zl allowed in current->mems_allowed?
1776 int cpuset_zonelist_valid_mems_allowed(struct zonelist *zl)
1780 for (i = 0; zl->zones[i]; i++) {
1781 int nid = zone_to_nid(zl->zones[i]);
1783 if (node_isset(nid, current->mems_allowed))
1790 * nearest_exclusive_ancestor() - Returns the nearest mem_exclusive
1791 * ancestor to the specified cpuset. Call holding callback_mutex.
1792 * If no ancestor is mem_exclusive (an unusual configuration), then
1793 * returns the root cpuset.
1795 static const struct cpuset *nearest_exclusive_ancestor(const struct cpuset *cs)
1797 while (!is_mem_exclusive(cs) && cs->parent)
1803 * cpuset_zone_allowed_softwall - Can we allocate on zone z's memory node?
1804 * @z: is this zone on an allowed node?
1805 * @gfp_mask: memory allocation flags
1807 * If we're in interrupt, yes, we can always allocate. If
1808 * __GFP_THISNODE is set, yes, we can always allocate. If zone
1809 * z's node is in our tasks mems_allowed, yes. If it's not a
1810 * __GFP_HARDWALL request and this zone's nodes is in the nearest
1811 * mem_exclusive cpuset ancestor to this tasks cpuset, yes.
1812 * If the task has been OOM killed and has access to memory reserves
1813 * as specified by the TIF_MEMDIE flag, yes.
1816 * If __GFP_HARDWALL is set, cpuset_zone_allowed_softwall()
1817 * reduces to cpuset_zone_allowed_hardwall(). Otherwise,
1818 * cpuset_zone_allowed_softwall() might sleep, and might allow a zone
1819 * from an enclosing cpuset.
1821 * cpuset_zone_allowed_hardwall() only handles the simpler case of
1822 * hardwall cpusets, and never sleeps.
1824 * The __GFP_THISNODE placement logic is really handled elsewhere,
1825 * by forcibly using a zonelist starting at a specified node, and by
1826 * (in get_page_from_freelist()) refusing to consider the zones for
1827 * any node on the zonelist except the first. By the time any such
1828 * calls get to this routine, we should just shut up and say 'yes'.
1830 * GFP_USER allocations are marked with the __GFP_HARDWALL bit,
1831 * and do not allow allocations outside the current tasks cpuset
1832 * unless the task has been OOM killed as is marked TIF_MEMDIE.
1833 * GFP_KERNEL allocations are not so marked, so can escape to the
1834 * nearest enclosing mem_exclusive ancestor cpuset.
1836 * Scanning up parent cpusets requires callback_mutex. The
1837 * __alloc_pages() routine only calls here with __GFP_HARDWALL bit
1838 * _not_ set if it's a GFP_KERNEL allocation, and all nodes in the
1839 * current tasks mems_allowed came up empty on the first pass over
1840 * the zonelist. So only GFP_KERNEL allocations, if all nodes in the
1841 * cpuset are short of memory, might require taking the callback_mutex
1844 * The first call here from mm/page_alloc:get_page_from_freelist()
1845 * has __GFP_HARDWALL set in gfp_mask, enforcing hardwall cpusets,
1846 * so no allocation on a node outside the cpuset is allowed (unless
1847 * in interrupt, of course).
1849 * The second pass through get_page_from_freelist() doesn't even call
1850 * here for GFP_ATOMIC calls. For those calls, the __alloc_pages()
1851 * variable 'wait' is not set, and the bit ALLOC_CPUSET is not set
1852 * in alloc_flags. That logic and the checks below have the combined
1854 * in_interrupt - any node ok (current task context irrelevant)
1855 * GFP_ATOMIC - any node ok
1856 * TIF_MEMDIE - any node ok
1857 * GFP_KERNEL - any node in enclosing mem_exclusive cpuset ok
1858 * GFP_USER - only nodes in current tasks mems allowed ok.
1861 * Don't call cpuset_zone_allowed_softwall if you can't sleep, unless you
1862 * pass in the __GFP_HARDWALL flag set in gfp_flag, which disables
1863 * the code that might scan up ancestor cpusets and sleep.
1866 int __cpuset_zone_allowed_softwall(struct zone *z, gfp_t gfp_mask)
1868 int node; /* node that zone z is on */
1869 const struct cpuset *cs; /* current cpuset ancestors */
1870 int allowed; /* is allocation in zone z allowed? */
1872 if (in_interrupt() || (gfp_mask & __GFP_THISNODE))
1874 node = zone_to_nid(z);
1875 might_sleep_if(!(gfp_mask & __GFP_HARDWALL));
1876 if (node_isset(node, current->mems_allowed))
1879 * Allow tasks that have access to memory reserves because they have
1880 * been OOM killed to get memory anywhere.
1882 if (unlikely(test_thread_flag(TIF_MEMDIE)))
1884 if (gfp_mask & __GFP_HARDWALL) /* If hardwall request, stop here */
1887 if (current->flags & PF_EXITING) /* Let dying task have memory */
1890 /* Not hardwall and node outside mems_allowed: scan up cpusets */
1891 mutex_lock(&callback_mutex);
1894 cs = nearest_exclusive_ancestor(task_cs(current));
1895 task_unlock(current);
1897 allowed = node_isset(node, cs->mems_allowed);
1898 mutex_unlock(&callback_mutex);
1903 * cpuset_zone_allowed_hardwall - Can we allocate on zone z's memory node?
1904 * @z: is this zone on an allowed node?
1905 * @gfp_mask: memory allocation flags
1907 * If we're in interrupt, yes, we can always allocate.
1908 * If __GFP_THISNODE is set, yes, we can always allocate. If zone
1909 * z's node is in our tasks mems_allowed, yes. If the task has been
1910 * OOM killed and has access to memory reserves as specified by the
1911 * TIF_MEMDIE flag, yes. Otherwise, no.
1913 * The __GFP_THISNODE placement logic is really handled elsewhere,
1914 * by forcibly using a zonelist starting at a specified node, and by
1915 * (in get_page_from_freelist()) refusing to consider the zones for
1916 * any node on the zonelist except the first. By the time any such
1917 * calls get to this routine, we should just shut up and say 'yes'.
1919 * Unlike the cpuset_zone_allowed_softwall() variant, above,
1920 * this variant requires that the zone be in the current tasks
1921 * mems_allowed or that we're in interrupt. It does not scan up the
1922 * cpuset hierarchy for the nearest enclosing mem_exclusive cpuset.
1926 int __cpuset_zone_allowed_hardwall(struct zone *z, gfp_t gfp_mask)
1928 int node; /* node that zone z is on */
1930 if (in_interrupt() || (gfp_mask & __GFP_THISNODE))
1932 node = zone_to_nid(z);
1933 if (node_isset(node, current->mems_allowed))
1936 * Allow tasks that have access to memory reserves because they have
1937 * been OOM killed to get memory anywhere.
1939 if (unlikely(test_thread_flag(TIF_MEMDIE)))
1945 * cpuset_lock - lock out any changes to cpuset structures
1947 * The out of memory (oom) code needs to mutex_lock cpusets
1948 * from being changed while it scans the tasklist looking for a
1949 * task in an overlapping cpuset. Expose callback_mutex via this
1950 * cpuset_lock() routine, so the oom code can lock it, before
1951 * locking the task list. The tasklist_lock is a spinlock, so
1952 * must be taken inside callback_mutex.
1955 void cpuset_lock(void)
1957 mutex_lock(&callback_mutex);
1961 * cpuset_unlock - release lock on cpuset changes
1963 * Undo the lock taken in a previous cpuset_lock() call.
1966 void cpuset_unlock(void)
1968 mutex_unlock(&callback_mutex);
1972 * cpuset_mem_spread_node() - On which node to begin search for a page
1974 * If a task is marked PF_SPREAD_PAGE or PF_SPREAD_SLAB (as for
1975 * tasks in a cpuset with is_spread_page or is_spread_slab set),
1976 * and if the memory allocation used cpuset_mem_spread_node()
1977 * to determine on which node to start looking, as it will for
1978 * certain page cache or slab cache pages such as used for file
1979 * system buffers and inode caches, then instead of starting on the
1980 * local node to look for a free page, rather spread the starting
1981 * node around the tasks mems_allowed nodes.
1983 * We don't have to worry about the returned node being offline
1984 * because "it can't happen", and even if it did, it would be ok.
1986 * The routines calling guarantee_online_mems() are careful to
1987 * only set nodes in task->mems_allowed that are online. So it
1988 * should not be possible for the following code to return an
1989 * offline node. But if it did, that would be ok, as this routine
1990 * is not returning the node where the allocation must be, only
1991 * the node where the search should start. The zonelist passed to
1992 * __alloc_pages() will include all nodes. If the slab allocator
1993 * is passed an offline node, it will fall back to the local node.
1994 * See kmem_cache_alloc_node().
1997 int cpuset_mem_spread_node(void)
2001 node = next_node(current->cpuset_mem_spread_rotor, current->mems_allowed);
2002 if (node == MAX_NUMNODES)
2003 node = first_node(current->mems_allowed);
2004 current->cpuset_mem_spread_rotor = node;
2007 EXPORT_SYMBOL_GPL(cpuset_mem_spread_node);
2010 * cpuset_mems_allowed_intersects - Does @tsk1's mems_allowed intersect @tsk2's?
2011 * @tsk1: pointer to task_struct of some task.
2012 * @tsk2: pointer to task_struct of some other task.
2014 * Description: Return true if @tsk1's mems_allowed intersects the
2015 * mems_allowed of @tsk2. Used by the OOM killer to determine if
2016 * one of the task's memory usage might impact the memory available
2020 int cpuset_mems_allowed_intersects(const struct task_struct *tsk1,
2021 const struct task_struct *tsk2)
2023 return nodes_intersects(tsk1->mems_allowed, tsk2->mems_allowed);
2027 * Collection of memory_pressure is suppressed unless
2028 * this flag is enabled by writing "1" to the special
2029 * cpuset file 'memory_pressure_enabled' in the root cpuset.
2032 int cpuset_memory_pressure_enabled __read_mostly;
2035 * cpuset_memory_pressure_bump - keep stats of per-cpuset reclaims.
2037 * Keep a running average of the rate of synchronous (direct)
2038 * page reclaim efforts initiated by tasks in each cpuset.
2040 * This represents the rate at which some task in the cpuset
2041 * ran low on memory on all nodes it was allowed to use, and
2042 * had to enter the kernels page reclaim code in an effort to
2043 * create more free memory by tossing clean pages or swapping
2044 * or writing dirty pages.
2046 * Display to user space in the per-cpuset read-only file
2047 * "memory_pressure". Value displayed is an integer
2048 * representing the recent rate of entry into the synchronous
2049 * (direct) page reclaim by any task attached to the cpuset.
2052 void __cpuset_memory_pressure_bump(void)
2055 fmeter_markevent(&task_cs(current)->fmeter);
2056 task_unlock(current);
2059 #ifdef CONFIG_PROC_PID_CPUSET
2061 * proc_cpuset_show()
2062 * - Print tasks cpuset path into seq_file.
2063 * - Used for /proc/<pid>/cpuset.
2064 * - No need to task_lock(tsk) on this tsk->cpuset reference, as it
2065 * doesn't really matter if tsk->cpuset changes after we read it,
2066 * and we take manage_mutex, keeping attach_task() from changing it
2067 * anyway. No need to check that tsk->cpuset != NULL, thanks to
2068 * the_top_cpuset_hack in cpuset_exit(), which sets an exiting tasks
2069 * cpuset to top_cpuset.
2071 static int proc_cpuset_show(struct seq_file *m, void *unused_v)
2074 struct task_struct *tsk;
2076 struct cgroup_subsys_state *css;
2080 buf = kmalloc(PAGE_SIZE, GFP_KERNEL);
2086 tsk = get_pid_task(pid, PIDTYPE_PID);
2092 css = task_subsys_state(tsk, cpuset_subsys_id);
2093 retval = cgroup_path(css->cgroup, buf, PAGE_SIZE);
2100 put_task_struct(tsk);
2107 static int cpuset_open(struct inode *inode, struct file *file)
2109 struct pid *pid = PROC_I(inode)->pid;
2110 return single_open(file, proc_cpuset_show, pid);
2113 const struct file_operations proc_cpuset_operations = {
2114 .open = cpuset_open,
2116 .llseek = seq_lseek,
2117 .release = single_release,
2119 #endif /* CONFIG_PROC_PID_CPUSET */
2121 /* Display task cpus_allowed, mems_allowed in /proc/<pid>/status file. */
2122 char *cpuset_task_status_allowed(struct task_struct *task, char *buffer)
2124 buffer += sprintf(buffer, "Cpus_allowed:\t");
2125 buffer += cpumask_scnprintf(buffer, PAGE_SIZE, task->cpus_allowed);
2126 buffer += sprintf(buffer, "\n");
2127 buffer += sprintf(buffer, "Mems_allowed:\t");
2128 buffer += nodemask_scnprintf(buffer, PAGE_SIZE, task->mems_allowed);
2129 buffer += sprintf(buffer, "\n");