Merge git://git.kernel.org/pub/scm/linux/kernel/git/mason/btrfs-unstable
[linux-2.6/mini2440.git] / kernel / cpuset.c
blob026faccca869e396d378693422e29e9c82f739ac
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/module.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 <asm/atomic.h>
59 #include <linux/mutex.h>
60 #include <linux/workqueue.h>
61 #include <linux/cgroup.h>
64 * Workqueue for cpuset related tasks.
66 * Using kevent workqueue may cause deadlock when memory_migrate
67 * is set. So we create a separate workqueue thread for cpuset.
69 static struct workqueue_struct *cpuset_wq;
72 * Tracks how many cpusets are currently defined in system.
73 * When there is only one cpuset (the root cpuset) we can
74 * short circuit some hooks.
76 int number_of_cpusets __read_mostly;
78 /* Forward declare cgroup structures */
79 struct cgroup_subsys cpuset_subsys;
80 struct cpuset;
82 /* See "Frequency meter" comments, below. */
84 struct fmeter {
85 int cnt; /* unprocessed events count */
86 int val; /* most recent output value */
87 time_t time; /* clock (secs) when val computed */
88 spinlock_t lock; /* guards read or write of above */
91 struct cpuset {
92 struct cgroup_subsys_state css;
94 unsigned long flags; /* "unsigned long" so bitops work */
95 cpumask_var_t cpus_allowed; /* CPUs allowed to tasks in cpuset */
96 nodemask_t mems_allowed; /* Memory Nodes allowed to tasks */
98 struct cpuset *parent; /* my parent */
101 * Copy of global cpuset_mems_generation as of the most
102 * recent time this cpuset changed its mems_allowed.
104 int mems_generation;
106 struct fmeter fmeter; /* memory_pressure filter */
108 /* partition number for rebuild_sched_domains() */
109 int pn;
111 /* for custom sched domain */
112 int relax_domain_level;
114 /* used for walking a cpuset heirarchy */
115 struct list_head stack_list;
118 /* Retrieve the cpuset for a cgroup */
119 static inline struct cpuset *cgroup_cs(struct cgroup *cont)
121 return container_of(cgroup_subsys_state(cont, 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 /* bits in struct cpuset flags field */
133 typedef enum {
134 CS_CPU_EXCLUSIVE,
135 CS_MEM_EXCLUSIVE,
136 CS_MEM_HARDWALL,
137 CS_MEMORY_MIGRATE,
138 CS_SCHED_LOAD_BALANCE,
139 CS_SPREAD_PAGE,
140 CS_SPREAD_SLAB,
141 } cpuset_flagbits_t;
143 /* convenient tests for these bits */
144 static inline int is_cpu_exclusive(const struct cpuset *cs)
146 return test_bit(CS_CPU_EXCLUSIVE, &cs->flags);
149 static inline int is_mem_exclusive(const struct cpuset *cs)
151 return test_bit(CS_MEM_EXCLUSIVE, &cs->flags);
154 static inline int is_mem_hardwall(const struct cpuset *cs)
156 return test_bit(CS_MEM_HARDWALL, &cs->flags);
159 static inline int is_sched_load_balance(const struct cpuset *cs)
161 return test_bit(CS_SCHED_LOAD_BALANCE, &cs->flags);
164 static inline int is_memory_migrate(const struct cpuset *cs)
166 return test_bit(CS_MEMORY_MIGRATE, &cs->flags);
169 static inline int is_spread_page(const struct cpuset *cs)
171 return test_bit(CS_SPREAD_PAGE, &cs->flags);
174 static inline int is_spread_slab(const struct cpuset *cs)
176 return test_bit(CS_SPREAD_SLAB, &cs->flags);
180 * Increment this integer everytime any cpuset changes its
181 * mems_allowed value. Users of cpusets can track this generation
182 * number, and avoid having to lock and reload mems_allowed unless
183 * the cpuset they're using changes generation.
185 * A single, global generation is needed because cpuset_attach_task() could
186 * reattach a task to a different cpuset, which must not have its
187 * generation numbers aliased with those of that tasks previous cpuset.
189 * Generations are needed for mems_allowed because one task cannot
190 * modify another's memory placement. So we must enable every task,
191 * on every visit to __alloc_pages(), to efficiently check whether
192 * its current->cpuset->mems_allowed has changed, requiring an update
193 * of its current->mems_allowed.
195 * Since writes to cpuset_mems_generation are guarded by the cgroup lock
196 * there is no need to mark it atomic.
198 static int cpuset_mems_generation;
200 static struct cpuset top_cpuset = {
201 .flags = ((1 << CS_CPU_EXCLUSIVE) | (1 << CS_MEM_EXCLUSIVE)),
205 * There are two global mutexes guarding cpuset structures. The first
206 * is the main control groups cgroup_mutex, accessed via
207 * cgroup_lock()/cgroup_unlock(). The second is the cpuset-specific
208 * callback_mutex, below. They can nest. It is ok to first take
209 * cgroup_mutex, then nest callback_mutex. We also require taking
210 * task_lock() when dereferencing a task's cpuset pointer. See "The
211 * task_lock() exception", at the end of this comment.
213 * A task must hold both mutexes to modify cpusets. If a task
214 * holds cgroup_mutex, then it blocks others wanting that mutex,
215 * ensuring that it is the only task able to also acquire callback_mutex
216 * and be able to modify cpusets. It can perform various checks on
217 * the cpuset structure first, knowing nothing will change. It can
218 * also allocate memory while just holding cgroup_mutex. While it is
219 * performing these checks, various callback routines can briefly
220 * acquire callback_mutex to query cpusets. Once it is ready to make
221 * the changes, it takes callback_mutex, blocking everyone else.
223 * Calls to the kernel memory allocator can not be made while holding
224 * callback_mutex, as that would risk double tripping on callback_mutex
225 * from one of the callbacks into the cpuset code from within
226 * __alloc_pages().
228 * If a task is only holding callback_mutex, then it has read-only
229 * access to cpusets.
231 * The task_struct fields mems_allowed and mems_generation may only
232 * be accessed in the context of that task, so require no locks.
234 * The cpuset_common_file_read() handlers only hold callback_mutex across
235 * small pieces of code, such as when reading out possibly multi-word
236 * cpumasks and nodemasks.
238 * Accessing a task's cpuset should be done in accordance with the
239 * guidelines for accessing subsystem state in kernel/cgroup.c
242 static DEFINE_MUTEX(callback_mutex);
245 * cpuset_buffer_lock protects both the cpuset_name and cpuset_nodelist
246 * buffers. They are statically allocated to prevent using excess stack
247 * when calling cpuset_print_task_mems_allowed().
249 #define CPUSET_NAME_LEN (128)
250 #define CPUSET_NODELIST_LEN (256)
251 static char cpuset_name[CPUSET_NAME_LEN];
252 static char cpuset_nodelist[CPUSET_NODELIST_LEN];
253 static DEFINE_SPINLOCK(cpuset_buffer_lock);
256 * This is ugly, but preserves the userspace API for existing cpuset
257 * users. If someone tries to mount the "cpuset" filesystem, we
258 * silently switch it to mount "cgroup" instead
260 static int cpuset_get_sb(struct file_system_type *fs_type,
261 int flags, const char *unused_dev_name,
262 void *data, struct vfsmount *mnt)
264 struct file_system_type *cgroup_fs = get_fs_type("cgroup");
265 int ret = -ENODEV;
266 if (cgroup_fs) {
267 char mountopts[] =
268 "cpuset,noprefix,"
269 "release_agent=/sbin/cpuset_release_agent";
270 ret = cgroup_fs->get_sb(cgroup_fs, flags,
271 unused_dev_name, mountopts, mnt);
272 put_filesystem(cgroup_fs);
274 return ret;
277 static struct file_system_type cpuset_fs_type = {
278 .name = "cpuset",
279 .get_sb = cpuset_get_sb,
283 * Return in pmask the portion of a cpusets's cpus_allowed that
284 * are online. If none are online, walk up the cpuset hierarchy
285 * until we find one that does have some online cpus. If we get
286 * all the way to the top and still haven't found any online cpus,
287 * return cpu_online_map. Or if passed a NULL cs from an exit'ing
288 * task, return cpu_online_map.
290 * One way or another, we guarantee to return some non-empty subset
291 * of cpu_online_map.
293 * Call with callback_mutex held.
296 static void guarantee_online_cpus(const struct cpuset *cs,
297 struct cpumask *pmask)
299 while (cs && !cpumask_intersects(cs->cpus_allowed, cpu_online_mask))
300 cs = cs->parent;
301 if (cs)
302 cpumask_and(pmask, cs->cpus_allowed, cpu_online_mask);
303 else
304 cpumask_copy(pmask, cpu_online_mask);
305 BUG_ON(!cpumask_intersects(pmask, cpu_online_mask));
309 * Return in *pmask the portion of a cpusets's mems_allowed that
310 * are online, with memory. If none are online with memory, walk
311 * up the cpuset hierarchy until we find one that does have some
312 * online mems. If we get all the way to the top and still haven't
313 * found any online mems, return node_states[N_HIGH_MEMORY].
315 * One way or another, we guarantee to return some non-empty subset
316 * of node_states[N_HIGH_MEMORY].
318 * Call with callback_mutex held.
321 static void guarantee_online_mems(const struct cpuset *cs, nodemask_t *pmask)
323 while (cs && !nodes_intersects(cs->mems_allowed,
324 node_states[N_HIGH_MEMORY]))
325 cs = cs->parent;
326 if (cs)
327 nodes_and(*pmask, cs->mems_allowed,
328 node_states[N_HIGH_MEMORY]);
329 else
330 *pmask = node_states[N_HIGH_MEMORY];
331 BUG_ON(!nodes_intersects(*pmask, node_states[N_HIGH_MEMORY]));
335 * cpuset_update_task_memory_state - update task memory placement
337 * If the current tasks cpusets mems_allowed changed behind our
338 * backs, update current->mems_allowed, mems_generation and task NUMA
339 * mempolicy to the new value.
341 * Task mempolicy is updated by rebinding it relative to the
342 * current->cpuset if a task has its memory placement changed.
343 * Do not call this routine if in_interrupt().
345 * Call without callback_mutex or task_lock() held. May be
346 * called with or without cgroup_mutex held. Thanks in part to
347 * 'the_top_cpuset_hack', the task's cpuset pointer will never
348 * be NULL. This routine also might acquire callback_mutex during
349 * call.
351 * Reading current->cpuset->mems_generation doesn't need task_lock
352 * to guard the current->cpuset derefence, because it is guarded
353 * from concurrent freeing of current->cpuset using RCU.
355 * The rcu_dereference() is technically probably not needed,
356 * as I don't actually mind if I see a new cpuset pointer but
357 * an old value of mems_generation. However this really only
358 * matters on alpha systems using cpusets heavily. If I dropped
359 * that rcu_dereference(), it would save them a memory barrier.
360 * For all other arch's, rcu_dereference is a no-op anyway, and for
361 * alpha systems not using cpusets, another planned optimization,
362 * avoiding the rcu critical section for tasks in the root cpuset
363 * which is statically allocated, so can't vanish, will make this
364 * irrelevant. Better to use RCU as intended, than to engage in
365 * some cute trick to save a memory barrier that is impossible to
366 * test, for alpha systems using cpusets heavily, which might not
367 * even exist.
369 * This routine is needed to update the per-task mems_allowed data,
370 * within the tasks context, when it is trying to allocate memory
371 * (in various mm/mempolicy.c routines) and notices that some other
372 * task has been modifying its cpuset.
375 void cpuset_update_task_memory_state(void)
377 int my_cpusets_mem_gen;
378 struct task_struct *tsk = current;
379 struct cpuset *cs;
381 rcu_read_lock();
382 my_cpusets_mem_gen = task_cs(tsk)->mems_generation;
383 rcu_read_unlock();
385 if (my_cpusets_mem_gen != tsk->cpuset_mems_generation) {
386 mutex_lock(&callback_mutex);
387 task_lock(tsk);
388 cs = task_cs(tsk); /* Maybe changed when task not locked */
389 guarantee_online_mems(cs, &tsk->mems_allowed);
390 tsk->cpuset_mems_generation = cs->mems_generation;
391 if (is_spread_page(cs))
392 tsk->flags |= PF_SPREAD_PAGE;
393 else
394 tsk->flags &= ~PF_SPREAD_PAGE;
395 if (is_spread_slab(cs))
396 tsk->flags |= PF_SPREAD_SLAB;
397 else
398 tsk->flags &= ~PF_SPREAD_SLAB;
399 task_unlock(tsk);
400 mutex_unlock(&callback_mutex);
401 mpol_rebind_task(tsk, &tsk->mems_allowed);
406 * is_cpuset_subset(p, q) - Is cpuset p a subset of cpuset q?
408 * One cpuset is a subset of another if all its allowed CPUs and
409 * Memory Nodes are a subset of the other, and its exclusive flags
410 * are only set if the other's are set. Call holding cgroup_mutex.
413 static int is_cpuset_subset(const struct cpuset *p, const struct cpuset *q)
415 return cpumask_subset(p->cpus_allowed, q->cpus_allowed) &&
416 nodes_subset(p->mems_allowed, q->mems_allowed) &&
417 is_cpu_exclusive(p) <= is_cpu_exclusive(q) &&
418 is_mem_exclusive(p) <= is_mem_exclusive(q);
422 * alloc_trial_cpuset - allocate a trial cpuset
423 * @cs: the cpuset that the trial cpuset duplicates
425 static struct cpuset *alloc_trial_cpuset(const struct cpuset *cs)
427 struct cpuset *trial;
429 trial = kmemdup(cs, sizeof(*cs), GFP_KERNEL);
430 if (!trial)
431 return NULL;
433 if (!alloc_cpumask_var(&trial->cpus_allowed, GFP_KERNEL)) {
434 kfree(trial);
435 return NULL;
437 cpumask_copy(trial->cpus_allowed, cs->cpus_allowed);
439 return trial;
443 * free_trial_cpuset - free the trial cpuset
444 * @trial: the trial cpuset to be freed
446 static void free_trial_cpuset(struct cpuset *trial)
448 free_cpumask_var(trial->cpus_allowed);
449 kfree(trial);
453 * validate_change() - Used to validate that any proposed cpuset change
454 * follows the structural rules for cpusets.
456 * If we replaced the flag and mask values of the current cpuset
457 * (cur) with those values in the trial cpuset (trial), would
458 * our various subset and exclusive rules still be valid? Presumes
459 * cgroup_mutex held.
461 * 'cur' is the address of an actual, in-use cpuset. Operations
462 * such as list traversal that depend on the actual address of the
463 * cpuset in the list must use cur below, not trial.
465 * 'trial' is the address of bulk structure copy of cur, with
466 * perhaps one or more of the fields cpus_allowed, mems_allowed,
467 * or flags changed to new, trial values.
469 * Return 0 if valid, -errno if not.
472 static int validate_change(const struct cpuset *cur, const struct cpuset *trial)
474 struct cgroup *cont;
475 struct cpuset *c, *par;
477 /* Each of our child cpusets must be a subset of us */
478 list_for_each_entry(cont, &cur->css.cgroup->children, sibling) {
479 if (!is_cpuset_subset(cgroup_cs(cont), trial))
480 return -EBUSY;
483 /* Remaining checks don't apply to root cpuset */
484 if (cur == &top_cpuset)
485 return 0;
487 par = cur->parent;
489 /* We must be a subset of our parent cpuset */
490 if (!is_cpuset_subset(trial, par))
491 return -EACCES;
494 * If either I or some sibling (!= me) is exclusive, we can't
495 * overlap
497 list_for_each_entry(cont, &par->css.cgroup->children, sibling) {
498 c = cgroup_cs(cont);
499 if ((is_cpu_exclusive(trial) || is_cpu_exclusive(c)) &&
500 c != cur &&
501 cpumask_intersects(trial->cpus_allowed, c->cpus_allowed))
502 return -EINVAL;
503 if ((is_mem_exclusive(trial) || is_mem_exclusive(c)) &&
504 c != cur &&
505 nodes_intersects(trial->mems_allowed, c->mems_allowed))
506 return -EINVAL;
509 /* Cpusets with tasks can't have empty cpus_allowed or mems_allowed */
510 if (cgroup_task_count(cur->css.cgroup)) {
511 if (cpumask_empty(trial->cpus_allowed) ||
512 nodes_empty(trial->mems_allowed)) {
513 return -ENOSPC;
517 return 0;
520 #ifdef CONFIG_SMP
522 * Helper routine for generate_sched_domains().
523 * Do cpusets a, b have overlapping cpus_allowed masks?
525 static int cpusets_overlap(struct cpuset *a, struct cpuset *b)
527 return cpumask_intersects(a->cpus_allowed, b->cpus_allowed);
530 static void
531 update_domain_attr(struct sched_domain_attr *dattr, struct cpuset *c)
533 if (dattr->relax_domain_level < c->relax_domain_level)
534 dattr->relax_domain_level = c->relax_domain_level;
535 return;
538 static void
539 update_domain_attr_tree(struct sched_domain_attr *dattr, struct cpuset *c)
541 LIST_HEAD(q);
543 list_add(&c->stack_list, &q);
544 while (!list_empty(&q)) {
545 struct cpuset *cp;
546 struct cgroup *cont;
547 struct cpuset *child;
549 cp = list_first_entry(&q, struct cpuset, stack_list);
550 list_del(q.next);
552 if (cpumask_empty(cp->cpus_allowed))
553 continue;
555 if (is_sched_load_balance(cp))
556 update_domain_attr(dattr, cp);
558 list_for_each_entry(cont, &cp->css.cgroup->children, sibling) {
559 child = cgroup_cs(cont);
560 list_add_tail(&child->stack_list, &q);
566 * generate_sched_domains()
568 * This function builds a partial partition of the systems CPUs
569 * A 'partial partition' is a set of non-overlapping subsets whose
570 * union is a subset of that set.
571 * The output of this function needs to be passed to kernel/sched.c
572 * partition_sched_domains() routine, which will rebuild the scheduler's
573 * load balancing domains (sched domains) as specified by that partial
574 * partition.
576 * See "What is sched_load_balance" in Documentation/cgroups/cpusets.txt
577 * for a background explanation of this.
579 * Does not return errors, on the theory that the callers of this
580 * routine would rather not worry about failures to rebuild sched
581 * domains when operating in the severe memory shortage situations
582 * that could cause allocation failures below.
584 * Must be called with cgroup_lock held.
586 * The three key local variables below are:
587 * q - a linked-list queue of cpuset pointers, used to implement a
588 * top-down scan of all cpusets. This scan loads a pointer
589 * to each cpuset marked is_sched_load_balance into the
590 * array 'csa'. For our purposes, rebuilding the schedulers
591 * sched domains, we can ignore !is_sched_load_balance cpusets.
592 * csa - (for CpuSet Array) Array of pointers to all the cpusets
593 * that need to be load balanced, for convenient iterative
594 * access by the subsequent code that finds the best partition,
595 * i.e the set of domains (subsets) of CPUs such that the
596 * cpus_allowed of every cpuset marked is_sched_load_balance
597 * is a subset of one of these domains, while there are as
598 * many such domains as possible, each as small as possible.
599 * doms - Conversion of 'csa' to an array of cpumasks, for passing to
600 * the kernel/sched.c routine partition_sched_domains() in a
601 * convenient format, that can be easily compared to the prior
602 * value to determine what partition elements (sched domains)
603 * were changed (added or removed.)
605 * Finding the best partition (set of domains):
606 * The triple nested loops below over i, j, k scan over the
607 * load balanced cpusets (using the array of cpuset pointers in
608 * csa[]) looking for pairs of cpusets that have overlapping
609 * cpus_allowed, but which don't have the same 'pn' partition
610 * number and gives them in the same partition number. It keeps
611 * looping on the 'restart' label until it can no longer find
612 * any such pairs.
614 * The union of the cpus_allowed masks from the set of
615 * all cpusets having the same 'pn' value then form the one
616 * element of the partition (one sched domain) to be passed to
617 * partition_sched_domains().
619 /* FIXME: see the FIXME in partition_sched_domains() */
620 static int generate_sched_domains(struct cpumask **domains,
621 struct sched_domain_attr **attributes)
623 LIST_HEAD(q); /* queue of cpusets to be scanned */
624 struct cpuset *cp; /* scans q */
625 struct cpuset **csa; /* array of all cpuset ptrs */
626 int csn; /* how many cpuset ptrs in csa so far */
627 int i, j, k; /* indices for partition finding loops */
628 struct cpumask *doms; /* resulting partition; i.e. sched domains */
629 struct sched_domain_attr *dattr; /* attributes for custom domains */
630 int ndoms = 0; /* number of sched domains in result */
631 int nslot; /* next empty doms[] struct cpumask slot */
633 doms = NULL;
634 dattr = NULL;
635 csa = NULL;
637 /* Special case for the 99% of systems with one, full, sched domain */
638 if (is_sched_load_balance(&top_cpuset)) {
639 doms = kmalloc(cpumask_size(), GFP_KERNEL);
640 if (!doms)
641 goto done;
643 dattr = kmalloc(sizeof(struct sched_domain_attr), GFP_KERNEL);
644 if (dattr) {
645 *dattr = SD_ATTR_INIT;
646 update_domain_attr_tree(dattr, &top_cpuset);
648 cpumask_copy(doms, top_cpuset.cpus_allowed);
650 ndoms = 1;
651 goto done;
654 csa = kmalloc(number_of_cpusets * sizeof(cp), GFP_KERNEL);
655 if (!csa)
656 goto done;
657 csn = 0;
659 list_add(&top_cpuset.stack_list, &q);
660 while (!list_empty(&q)) {
661 struct cgroup *cont;
662 struct cpuset *child; /* scans child cpusets of cp */
664 cp = list_first_entry(&q, struct cpuset, stack_list);
665 list_del(q.next);
667 if (cpumask_empty(cp->cpus_allowed))
668 continue;
671 * All child cpusets contain a subset of the parent's cpus, so
672 * just skip them, and then we call update_domain_attr_tree()
673 * to calc relax_domain_level of the corresponding sched
674 * domain.
676 if (is_sched_load_balance(cp)) {
677 csa[csn++] = cp;
678 continue;
681 list_for_each_entry(cont, &cp->css.cgroup->children, sibling) {
682 child = cgroup_cs(cont);
683 list_add_tail(&child->stack_list, &q);
687 for (i = 0; i < csn; i++)
688 csa[i]->pn = i;
689 ndoms = csn;
691 restart:
692 /* Find the best partition (set of sched domains) */
693 for (i = 0; i < csn; i++) {
694 struct cpuset *a = csa[i];
695 int apn = a->pn;
697 for (j = 0; j < csn; j++) {
698 struct cpuset *b = csa[j];
699 int bpn = b->pn;
701 if (apn != bpn && cpusets_overlap(a, b)) {
702 for (k = 0; k < csn; k++) {
703 struct cpuset *c = csa[k];
705 if (c->pn == bpn)
706 c->pn = apn;
708 ndoms--; /* one less element */
709 goto restart;
715 * Now we know how many domains to create.
716 * Convert <csn, csa> to <ndoms, doms> and populate cpu masks.
718 doms = kmalloc(ndoms * cpumask_size(), GFP_KERNEL);
719 if (!doms)
720 goto done;
723 * The rest of the code, including the scheduler, can deal with
724 * dattr==NULL case. No need to abort if alloc fails.
726 dattr = kmalloc(ndoms * sizeof(struct sched_domain_attr), GFP_KERNEL);
728 for (nslot = 0, i = 0; i < csn; i++) {
729 struct cpuset *a = csa[i];
730 struct cpumask *dp;
731 int apn = a->pn;
733 if (apn < 0) {
734 /* Skip completed partitions */
735 continue;
738 dp = doms + nslot;
740 if (nslot == ndoms) {
741 static int warnings = 10;
742 if (warnings) {
743 printk(KERN_WARNING
744 "rebuild_sched_domains confused:"
745 " nslot %d, ndoms %d, csn %d, i %d,"
746 " apn %d\n",
747 nslot, ndoms, csn, i, apn);
748 warnings--;
750 continue;
753 cpumask_clear(dp);
754 if (dattr)
755 *(dattr + nslot) = SD_ATTR_INIT;
756 for (j = i; j < csn; j++) {
757 struct cpuset *b = csa[j];
759 if (apn == b->pn) {
760 cpumask_or(dp, dp, b->cpus_allowed);
761 if (dattr)
762 update_domain_attr_tree(dattr + nslot, b);
764 /* Done with this partition */
765 b->pn = -1;
768 nslot++;
770 BUG_ON(nslot != ndoms);
772 done:
773 kfree(csa);
776 * Fallback to the default domain if kmalloc() failed.
777 * See comments in partition_sched_domains().
779 if (doms == NULL)
780 ndoms = 1;
782 *domains = doms;
783 *attributes = dattr;
784 return ndoms;
788 * Rebuild scheduler domains.
790 * Call with neither cgroup_mutex held nor within get_online_cpus().
791 * Takes both cgroup_mutex and get_online_cpus().
793 * Cannot be directly called from cpuset code handling changes
794 * to the cpuset pseudo-filesystem, because it cannot be called
795 * from code that already holds cgroup_mutex.
797 static void do_rebuild_sched_domains(struct work_struct *unused)
799 struct sched_domain_attr *attr;
800 struct cpumask *doms;
801 int ndoms;
803 get_online_cpus();
805 /* Generate domain masks and attrs */
806 cgroup_lock();
807 ndoms = generate_sched_domains(&doms, &attr);
808 cgroup_unlock();
810 /* Have scheduler rebuild the domains */
811 partition_sched_domains(ndoms, doms, attr);
813 put_online_cpus();
815 #else /* !CONFIG_SMP */
816 static void do_rebuild_sched_domains(struct work_struct *unused)
820 static int generate_sched_domains(struct cpumask **domains,
821 struct sched_domain_attr **attributes)
823 *domains = NULL;
824 return 1;
826 #endif /* CONFIG_SMP */
828 static DECLARE_WORK(rebuild_sched_domains_work, do_rebuild_sched_domains);
831 * Rebuild scheduler domains, asynchronously via workqueue.
833 * If the flag 'sched_load_balance' of any cpuset with non-empty
834 * 'cpus' changes, or if the 'cpus' allowed changes in any cpuset
835 * which has that flag enabled, or if any cpuset with a non-empty
836 * 'cpus' is removed, then call this routine to rebuild the
837 * scheduler's dynamic sched domains.
839 * The rebuild_sched_domains() and partition_sched_domains()
840 * routines must nest cgroup_lock() inside get_online_cpus(),
841 * but such cpuset changes as these must nest that locking the
842 * other way, holding cgroup_lock() for much of the code.
844 * So in order to avoid an ABBA deadlock, the cpuset code handling
845 * these user changes delegates the actual sched domain rebuilding
846 * to a separate workqueue thread, which ends up processing the
847 * above do_rebuild_sched_domains() function.
849 static void async_rebuild_sched_domains(void)
851 queue_work(cpuset_wq, &rebuild_sched_domains_work);
855 * Accomplishes the same scheduler domain rebuild as the above
856 * async_rebuild_sched_domains(), however it directly calls the
857 * rebuild routine synchronously rather than calling it via an
858 * asynchronous work thread.
860 * This can only be called from code that is not holding
861 * cgroup_mutex (not nested in a cgroup_lock() call.)
863 void rebuild_sched_domains(void)
865 do_rebuild_sched_domains(NULL);
869 * cpuset_test_cpumask - test a task's cpus_allowed versus its cpuset's
870 * @tsk: task to test
871 * @scan: struct cgroup_scanner contained in its struct cpuset_hotplug_scanner
873 * Call with cgroup_mutex held. May take callback_mutex during call.
874 * Called for each task in a cgroup by cgroup_scan_tasks().
875 * Return nonzero if this tasks's cpus_allowed mask should be changed (in other
876 * words, if its mask is not equal to its cpuset's mask).
878 static int cpuset_test_cpumask(struct task_struct *tsk,
879 struct cgroup_scanner *scan)
881 return !cpumask_equal(&tsk->cpus_allowed,
882 (cgroup_cs(scan->cg))->cpus_allowed);
886 * cpuset_change_cpumask - make a task's cpus_allowed the same as its cpuset's
887 * @tsk: task to test
888 * @scan: struct cgroup_scanner containing the cgroup of the task
890 * Called by cgroup_scan_tasks() for each task in a cgroup whose
891 * cpus_allowed mask needs to be changed.
893 * We don't need to re-check for the cgroup/cpuset membership, since we're
894 * holding cgroup_lock() at this point.
896 static void cpuset_change_cpumask(struct task_struct *tsk,
897 struct cgroup_scanner *scan)
899 set_cpus_allowed_ptr(tsk, ((cgroup_cs(scan->cg))->cpus_allowed));
903 * update_tasks_cpumask - Update the cpumasks of tasks in the cpuset.
904 * @cs: the cpuset in which each task's cpus_allowed mask needs to be changed
905 * @heap: if NULL, defer allocating heap memory to cgroup_scan_tasks()
907 * Called with cgroup_mutex held
909 * The cgroup_scan_tasks() function will scan all the tasks in a cgroup,
910 * calling callback functions for each.
912 * No return value. It's guaranteed that cgroup_scan_tasks() always returns 0
913 * if @heap != NULL.
915 static void update_tasks_cpumask(struct cpuset *cs, struct ptr_heap *heap)
917 struct cgroup_scanner scan;
919 scan.cg = cs->css.cgroup;
920 scan.test_task = cpuset_test_cpumask;
921 scan.process_task = cpuset_change_cpumask;
922 scan.heap = heap;
923 cgroup_scan_tasks(&scan);
927 * update_cpumask - update the cpus_allowed mask of a cpuset and all tasks in it
928 * @cs: the cpuset to consider
929 * @buf: buffer of cpu numbers written to this cpuset
931 static int update_cpumask(struct cpuset *cs, struct cpuset *trialcs,
932 const char *buf)
934 struct ptr_heap heap;
935 int retval;
936 int is_load_balanced;
938 /* top_cpuset.cpus_allowed tracks cpu_online_map; it's read-only */
939 if (cs == &top_cpuset)
940 return -EACCES;
943 * An empty cpus_allowed is ok only if the cpuset has no tasks.
944 * Since cpulist_parse() fails on an empty mask, we special case
945 * that parsing. The validate_change() call ensures that cpusets
946 * with tasks have cpus.
948 if (!*buf) {
949 cpumask_clear(trialcs->cpus_allowed);
950 } else {
951 retval = cpulist_parse(buf, trialcs->cpus_allowed);
952 if (retval < 0)
953 return retval;
955 if (!cpumask_subset(trialcs->cpus_allowed, cpu_online_mask))
956 return -EINVAL;
958 retval = validate_change(cs, trialcs);
959 if (retval < 0)
960 return retval;
962 /* Nothing to do if the cpus didn't change */
963 if (cpumask_equal(cs->cpus_allowed, trialcs->cpus_allowed))
964 return 0;
966 retval = heap_init(&heap, PAGE_SIZE, GFP_KERNEL, NULL);
967 if (retval)
968 return retval;
970 is_load_balanced = is_sched_load_balance(trialcs);
972 mutex_lock(&callback_mutex);
973 cpumask_copy(cs->cpus_allowed, trialcs->cpus_allowed);
974 mutex_unlock(&callback_mutex);
977 * Scan tasks in the cpuset, and update the cpumasks of any
978 * that need an update.
980 update_tasks_cpumask(cs, &heap);
982 heap_free(&heap);
984 if (is_load_balanced)
985 async_rebuild_sched_domains();
986 return 0;
990 * cpuset_migrate_mm
992 * Migrate memory region from one set of nodes to another.
994 * Temporarilly set tasks mems_allowed to target nodes of migration,
995 * so that the migration code can allocate pages on these nodes.
997 * Call holding cgroup_mutex, so current's cpuset won't change
998 * during this call, as manage_mutex holds off any cpuset_attach()
999 * calls. Therefore we don't need to take task_lock around the
1000 * call to guarantee_online_mems(), as we know no one is changing
1001 * our task's cpuset.
1003 * Hold callback_mutex around the two modifications of our tasks
1004 * mems_allowed to synchronize with cpuset_mems_allowed().
1006 * While the mm_struct we are migrating is typically from some
1007 * other task, the task_struct mems_allowed that we are hacking
1008 * is for our current task, which must allocate new pages for that
1009 * migrating memory region.
1011 * We call cpuset_update_task_memory_state() before hacking
1012 * our tasks mems_allowed, so that we are assured of being in
1013 * sync with our tasks cpuset, and in particular, callbacks to
1014 * cpuset_update_task_memory_state() from nested page allocations
1015 * won't see any mismatch of our cpuset and task mems_generation
1016 * values, so won't overwrite our hacked tasks mems_allowed
1017 * nodemask.
1020 static void cpuset_migrate_mm(struct mm_struct *mm, const nodemask_t *from,
1021 const nodemask_t *to)
1023 struct task_struct *tsk = current;
1025 cpuset_update_task_memory_state();
1027 mutex_lock(&callback_mutex);
1028 tsk->mems_allowed = *to;
1029 mutex_unlock(&callback_mutex);
1031 do_migrate_pages(mm, from, to, MPOL_MF_MOVE_ALL);
1033 mutex_lock(&callback_mutex);
1034 guarantee_online_mems(task_cs(tsk),&tsk->mems_allowed);
1035 mutex_unlock(&callback_mutex);
1039 * Rebind task's vmas to cpuset's new mems_allowed, and migrate pages to new
1040 * nodes if memory_migrate flag is set. Called with cgroup_mutex held.
1042 static void cpuset_change_nodemask(struct task_struct *p,
1043 struct cgroup_scanner *scan)
1045 struct mm_struct *mm;
1046 struct cpuset *cs;
1047 int migrate;
1048 const nodemask_t *oldmem = scan->data;
1050 mm = get_task_mm(p);
1051 if (!mm)
1052 return;
1054 cs = cgroup_cs(scan->cg);
1055 migrate = is_memory_migrate(cs);
1057 mpol_rebind_mm(mm, &cs->mems_allowed);
1058 if (migrate)
1059 cpuset_migrate_mm(mm, oldmem, &cs->mems_allowed);
1060 mmput(mm);
1063 static void *cpuset_being_rebound;
1066 * update_tasks_nodemask - Update the nodemasks of tasks in the cpuset.
1067 * @cs: the cpuset in which each task's mems_allowed mask needs to be changed
1068 * @oldmem: old mems_allowed of cpuset cs
1069 * @heap: if NULL, defer allocating heap memory to cgroup_scan_tasks()
1071 * Called with cgroup_mutex held
1072 * No return value. It's guaranteed that cgroup_scan_tasks() always returns 0
1073 * if @heap != NULL.
1075 static void update_tasks_nodemask(struct cpuset *cs, const nodemask_t *oldmem,
1076 struct ptr_heap *heap)
1078 struct cgroup_scanner scan;
1080 cpuset_being_rebound = cs; /* causes mpol_dup() rebind */
1082 scan.cg = cs->css.cgroup;
1083 scan.test_task = NULL;
1084 scan.process_task = cpuset_change_nodemask;
1085 scan.heap = heap;
1086 scan.data = (nodemask_t *)oldmem;
1089 * The mpol_rebind_mm() call takes mmap_sem, which we couldn't
1090 * take while holding tasklist_lock. Forks can happen - the
1091 * mpol_dup() cpuset_being_rebound check will catch such forks,
1092 * and rebind their vma mempolicies too. Because we still hold
1093 * the global cgroup_mutex, we know that no other rebind effort
1094 * will be contending for the global variable cpuset_being_rebound.
1095 * It's ok if we rebind the same mm twice; mpol_rebind_mm()
1096 * is idempotent. Also migrate pages in each mm to new nodes.
1098 cgroup_scan_tasks(&scan);
1100 /* We're done rebinding vmas to this cpuset's new mems_allowed. */
1101 cpuset_being_rebound = NULL;
1105 * Handle user request to change the 'mems' memory placement
1106 * of a cpuset. Needs to validate the request, update the
1107 * cpusets mems_allowed and mems_generation, and for each
1108 * task in the cpuset, rebind any vma mempolicies and if
1109 * the cpuset is marked 'memory_migrate', migrate the tasks
1110 * pages to the new memory.
1112 * Call with cgroup_mutex held. May take callback_mutex during call.
1113 * Will take tasklist_lock, scan tasklist for tasks in cpuset cs,
1114 * lock each such tasks mm->mmap_sem, scan its vma's and rebind
1115 * their mempolicies to the cpusets new mems_allowed.
1117 static int update_nodemask(struct cpuset *cs, struct cpuset *trialcs,
1118 const char *buf)
1120 nodemask_t oldmem;
1121 int retval;
1122 struct ptr_heap heap;
1125 * top_cpuset.mems_allowed tracks node_stats[N_HIGH_MEMORY];
1126 * it's read-only
1128 if (cs == &top_cpuset)
1129 return -EACCES;
1132 * An empty mems_allowed is ok iff there are no tasks in the cpuset.
1133 * Since nodelist_parse() fails on an empty mask, we special case
1134 * that parsing. The validate_change() call ensures that cpusets
1135 * with tasks have memory.
1137 if (!*buf) {
1138 nodes_clear(trialcs->mems_allowed);
1139 } else {
1140 retval = nodelist_parse(buf, trialcs->mems_allowed);
1141 if (retval < 0)
1142 goto done;
1144 if (!nodes_subset(trialcs->mems_allowed,
1145 node_states[N_HIGH_MEMORY]))
1146 return -EINVAL;
1148 oldmem = cs->mems_allowed;
1149 if (nodes_equal(oldmem, trialcs->mems_allowed)) {
1150 retval = 0; /* Too easy - nothing to do */
1151 goto done;
1153 retval = validate_change(cs, trialcs);
1154 if (retval < 0)
1155 goto done;
1157 retval = heap_init(&heap, PAGE_SIZE, GFP_KERNEL, NULL);
1158 if (retval < 0)
1159 goto done;
1161 mutex_lock(&callback_mutex);
1162 cs->mems_allowed = trialcs->mems_allowed;
1163 cs->mems_generation = cpuset_mems_generation++;
1164 mutex_unlock(&callback_mutex);
1166 update_tasks_nodemask(cs, &oldmem, &heap);
1168 heap_free(&heap);
1169 done:
1170 return retval;
1173 int current_cpuset_is_being_rebound(void)
1175 return task_cs(current) == cpuset_being_rebound;
1178 static int update_relax_domain_level(struct cpuset *cs, s64 val)
1180 #ifdef CONFIG_SMP
1181 if (val < -1 || val >= SD_LV_MAX)
1182 return -EINVAL;
1183 #endif
1185 if (val != cs->relax_domain_level) {
1186 cs->relax_domain_level = val;
1187 if (!cpumask_empty(cs->cpus_allowed) &&
1188 is_sched_load_balance(cs))
1189 async_rebuild_sched_domains();
1192 return 0;
1196 * update_flag - read a 0 or a 1 in a file and update associated flag
1197 * bit: the bit to update (see cpuset_flagbits_t)
1198 * cs: the cpuset to update
1199 * turning_on: whether the flag is being set or cleared
1201 * Call with cgroup_mutex held.
1204 static int update_flag(cpuset_flagbits_t bit, struct cpuset *cs,
1205 int turning_on)
1207 struct cpuset *trialcs;
1208 int err;
1209 int balance_flag_changed;
1211 trialcs = alloc_trial_cpuset(cs);
1212 if (!trialcs)
1213 return -ENOMEM;
1215 if (turning_on)
1216 set_bit(bit, &trialcs->flags);
1217 else
1218 clear_bit(bit, &trialcs->flags);
1220 err = validate_change(cs, trialcs);
1221 if (err < 0)
1222 goto out;
1224 balance_flag_changed = (is_sched_load_balance(cs) !=
1225 is_sched_load_balance(trialcs));
1227 mutex_lock(&callback_mutex);
1228 cs->flags = trialcs->flags;
1229 mutex_unlock(&callback_mutex);
1231 if (!cpumask_empty(trialcs->cpus_allowed) && balance_flag_changed)
1232 async_rebuild_sched_domains();
1234 out:
1235 free_trial_cpuset(trialcs);
1236 return err;
1240 * Frequency meter - How fast is some event occurring?
1242 * These routines manage a digitally filtered, constant time based,
1243 * event frequency meter. There are four routines:
1244 * fmeter_init() - initialize a frequency meter.
1245 * fmeter_markevent() - called each time the event happens.
1246 * fmeter_getrate() - returns the recent rate of such events.
1247 * fmeter_update() - internal routine used to update fmeter.
1249 * A common data structure is passed to each of these routines,
1250 * which is used to keep track of the state required to manage the
1251 * frequency meter and its digital filter.
1253 * The filter works on the number of events marked per unit time.
1254 * The filter is single-pole low-pass recursive (IIR). The time unit
1255 * is 1 second. Arithmetic is done using 32-bit integers scaled to
1256 * simulate 3 decimal digits of precision (multiplied by 1000).
1258 * With an FM_COEF of 933, and a time base of 1 second, the filter
1259 * has a half-life of 10 seconds, meaning that if the events quit
1260 * happening, then the rate returned from the fmeter_getrate()
1261 * will be cut in half each 10 seconds, until it converges to zero.
1263 * It is not worth doing a real infinitely recursive filter. If more
1264 * than FM_MAXTICKS ticks have elapsed since the last filter event,
1265 * just compute FM_MAXTICKS ticks worth, by which point the level
1266 * will be stable.
1268 * Limit the count of unprocessed events to FM_MAXCNT, so as to avoid
1269 * arithmetic overflow in the fmeter_update() routine.
1271 * Given the simple 32 bit integer arithmetic used, this meter works
1272 * best for reporting rates between one per millisecond (msec) and
1273 * one per 32 (approx) seconds. At constant rates faster than one
1274 * per msec it maxes out at values just under 1,000,000. At constant
1275 * rates between one per msec, and one per second it will stabilize
1276 * to a value N*1000, where N is the rate of events per second.
1277 * At constant rates between one per second and one per 32 seconds,
1278 * it will be choppy, moving up on the seconds that have an event,
1279 * and then decaying until the next event. At rates slower than
1280 * about one in 32 seconds, it decays all the way back to zero between
1281 * each event.
1284 #define FM_COEF 933 /* coefficient for half-life of 10 secs */
1285 #define FM_MAXTICKS ((time_t)99) /* useless computing more ticks than this */
1286 #define FM_MAXCNT 1000000 /* limit cnt to avoid overflow */
1287 #define FM_SCALE 1000 /* faux fixed point scale */
1289 /* Initialize a frequency meter */
1290 static void fmeter_init(struct fmeter *fmp)
1292 fmp->cnt = 0;
1293 fmp->val = 0;
1294 fmp->time = 0;
1295 spin_lock_init(&fmp->lock);
1298 /* Internal meter update - process cnt events and update value */
1299 static void fmeter_update(struct fmeter *fmp)
1301 time_t now = get_seconds();
1302 time_t ticks = now - fmp->time;
1304 if (ticks == 0)
1305 return;
1307 ticks = min(FM_MAXTICKS, ticks);
1308 while (ticks-- > 0)
1309 fmp->val = (FM_COEF * fmp->val) / FM_SCALE;
1310 fmp->time = now;
1312 fmp->val += ((FM_SCALE - FM_COEF) * fmp->cnt) / FM_SCALE;
1313 fmp->cnt = 0;
1316 /* Process any previous ticks, then bump cnt by one (times scale). */
1317 static void fmeter_markevent(struct fmeter *fmp)
1319 spin_lock(&fmp->lock);
1320 fmeter_update(fmp);
1321 fmp->cnt = min(FM_MAXCNT, fmp->cnt + FM_SCALE);
1322 spin_unlock(&fmp->lock);
1325 /* Process any previous ticks, then return current value. */
1326 static int fmeter_getrate(struct fmeter *fmp)
1328 int val;
1330 spin_lock(&fmp->lock);
1331 fmeter_update(fmp);
1332 val = fmp->val;
1333 spin_unlock(&fmp->lock);
1334 return val;
1337 /* Protected by cgroup_lock */
1338 static cpumask_var_t cpus_attach;
1340 /* Called by cgroups to determine if a cpuset is usable; cgroup_mutex held */
1341 static int cpuset_can_attach(struct cgroup_subsys *ss,
1342 struct cgroup *cont, struct task_struct *tsk)
1344 struct cpuset *cs = cgroup_cs(cont);
1346 if (cpumask_empty(cs->cpus_allowed) || nodes_empty(cs->mems_allowed))
1347 return -ENOSPC;
1350 * Kthreads bound to specific cpus cannot be moved to a new cpuset; we
1351 * cannot change their cpu affinity and isolating such threads by their
1352 * set of allowed nodes is unnecessary. Thus, cpusets are not
1353 * applicable for such threads. This prevents checking for success of
1354 * set_cpus_allowed_ptr() on all attached tasks before cpus_allowed may
1355 * be changed.
1357 if (tsk->flags & PF_THREAD_BOUND)
1358 return -EINVAL;
1360 return security_task_setscheduler(tsk, 0, NULL);
1363 static void cpuset_attach(struct cgroup_subsys *ss,
1364 struct cgroup *cont, struct cgroup *oldcont,
1365 struct task_struct *tsk)
1367 nodemask_t from, to;
1368 struct mm_struct *mm;
1369 struct cpuset *cs = cgroup_cs(cont);
1370 struct cpuset *oldcs = cgroup_cs(oldcont);
1371 int err;
1373 if (cs == &top_cpuset) {
1374 cpumask_copy(cpus_attach, cpu_possible_mask);
1375 } else {
1376 mutex_lock(&callback_mutex);
1377 guarantee_online_cpus(cs, cpus_attach);
1378 mutex_unlock(&callback_mutex);
1380 err = set_cpus_allowed_ptr(tsk, cpus_attach);
1381 if (err)
1382 return;
1384 from = oldcs->mems_allowed;
1385 to = cs->mems_allowed;
1386 mm = get_task_mm(tsk);
1387 if (mm) {
1388 mpol_rebind_mm(mm, &to);
1389 if (is_memory_migrate(cs))
1390 cpuset_migrate_mm(mm, &from, &to);
1391 mmput(mm);
1395 /* The various types of files and directories in a cpuset file system */
1397 typedef enum {
1398 FILE_MEMORY_MIGRATE,
1399 FILE_CPULIST,
1400 FILE_MEMLIST,
1401 FILE_CPU_EXCLUSIVE,
1402 FILE_MEM_EXCLUSIVE,
1403 FILE_MEM_HARDWALL,
1404 FILE_SCHED_LOAD_BALANCE,
1405 FILE_SCHED_RELAX_DOMAIN_LEVEL,
1406 FILE_MEMORY_PRESSURE_ENABLED,
1407 FILE_MEMORY_PRESSURE,
1408 FILE_SPREAD_PAGE,
1409 FILE_SPREAD_SLAB,
1410 } cpuset_filetype_t;
1412 static int cpuset_write_u64(struct cgroup *cgrp, struct cftype *cft, u64 val)
1414 int retval = 0;
1415 struct cpuset *cs = cgroup_cs(cgrp);
1416 cpuset_filetype_t type = cft->private;
1418 if (!cgroup_lock_live_group(cgrp))
1419 return -ENODEV;
1421 switch (type) {
1422 case FILE_CPU_EXCLUSIVE:
1423 retval = update_flag(CS_CPU_EXCLUSIVE, cs, val);
1424 break;
1425 case FILE_MEM_EXCLUSIVE:
1426 retval = update_flag(CS_MEM_EXCLUSIVE, cs, val);
1427 break;
1428 case FILE_MEM_HARDWALL:
1429 retval = update_flag(CS_MEM_HARDWALL, cs, val);
1430 break;
1431 case FILE_SCHED_LOAD_BALANCE:
1432 retval = update_flag(CS_SCHED_LOAD_BALANCE, cs, val);
1433 break;
1434 case FILE_MEMORY_MIGRATE:
1435 retval = update_flag(CS_MEMORY_MIGRATE, cs, val);
1436 break;
1437 case FILE_MEMORY_PRESSURE_ENABLED:
1438 cpuset_memory_pressure_enabled = !!val;
1439 break;
1440 case FILE_MEMORY_PRESSURE:
1441 retval = -EACCES;
1442 break;
1443 case FILE_SPREAD_PAGE:
1444 retval = update_flag(CS_SPREAD_PAGE, cs, val);
1445 cs->mems_generation = cpuset_mems_generation++;
1446 break;
1447 case FILE_SPREAD_SLAB:
1448 retval = update_flag(CS_SPREAD_SLAB, cs, val);
1449 cs->mems_generation = cpuset_mems_generation++;
1450 break;
1451 default:
1452 retval = -EINVAL;
1453 break;
1455 cgroup_unlock();
1456 return retval;
1459 static int cpuset_write_s64(struct cgroup *cgrp, struct cftype *cft, s64 val)
1461 int retval = 0;
1462 struct cpuset *cs = cgroup_cs(cgrp);
1463 cpuset_filetype_t type = cft->private;
1465 if (!cgroup_lock_live_group(cgrp))
1466 return -ENODEV;
1468 switch (type) {
1469 case FILE_SCHED_RELAX_DOMAIN_LEVEL:
1470 retval = update_relax_domain_level(cs, val);
1471 break;
1472 default:
1473 retval = -EINVAL;
1474 break;
1476 cgroup_unlock();
1477 return retval;
1481 * Common handling for a write to a "cpus" or "mems" file.
1483 static int cpuset_write_resmask(struct cgroup *cgrp, struct cftype *cft,
1484 const char *buf)
1486 int retval = 0;
1487 struct cpuset *cs = cgroup_cs(cgrp);
1488 struct cpuset *trialcs;
1490 if (!cgroup_lock_live_group(cgrp))
1491 return -ENODEV;
1493 trialcs = alloc_trial_cpuset(cs);
1494 if (!trialcs)
1495 return -ENOMEM;
1497 switch (cft->private) {
1498 case FILE_CPULIST:
1499 retval = update_cpumask(cs, trialcs, buf);
1500 break;
1501 case FILE_MEMLIST:
1502 retval = update_nodemask(cs, trialcs, buf);
1503 break;
1504 default:
1505 retval = -EINVAL;
1506 break;
1509 free_trial_cpuset(trialcs);
1510 cgroup_unlock();
1511 return retval;
1515 * These ascii lists should be read in a single call, by using a user
1516 * buffer large enough to hold the entire map. If read in smaller
1517 * chunks, there is no guarantee of atomicity. Since the display format
1518 * used, list of ranges of sequential numbers, is variable length,
1519 * and since these maps can change value dynamically, one could read
1520 * gibberish by doing partial reads while a list was changing.
1521 * A single large read to a buffer that crosses a page boundary is
1522 * ok, because the result being copied to user land is not recomputed
1523 * across a page fault.
1526 static int cpuset_sprintf_cpulist(char *page, struct cpuset *cs)
1528 int ret;
1530 mutex_lock(&callback_mutex);
1531 ret = cpulist_scnprintf(page, PAGE_SIZE, cs->cpus_allowed);
1532 mutex_unlock(&callback_mutex);
1534 return ret;
1537 static int cpuset_sprintf_memlist(char *page, struct cpuset *cs)
1539 nodemask_t mask;
1541 mutex_lock(&callback_mutex);
1542 mask = cs->mems_allowed;
1543 mutex_unlock(&callback_mutex);
1545 return nodelist_scnprintf(page, PAGE_SIZE, mask);
1548 static ssize_t cpuset_common_file_read(struct cgroup *cont,
1549 struct cftype *cft,
1550 struct file *file,
1551 char __user *buf,
1552 size_t nbytes, loff_t *ppos)
1554 struct cpuset *cs = cgroup_cs(cont);
1555 cpuset_filetype_t type = cft->private;
1556 char *page;
1557 ssize_t retval = 0;
1558 char *s;
1560 if (!(page = (char *)__get_free_page(GFP_TEMPORARY)))
1561 return -ENOMEM;
1563 s = page;
1565 switch (type) {
1566 case FILE_CPULIST:
1567 s += cpuset_sprintf_cpulist(s, cs);
1568 break;
1569 case FILE_MEMLIST:
1570 s += cpuset_sprintf_memlist(s, cs);
1571 break;
1572 default:
1573 retval = -EINVAL;
1574 goto out;
1576 *s++ = '\n';
1578 retval = simple_read_from_buffer(buf, nbytes, ppos, page, s - page);
1579 out:
1580 free_page((unsigned long)page);
1581 return retval;
1584 static u64 cpuset_read_u64(struct cgroup *cont, struct cftype *cft)
1586 struct cpuset *cs = cgroup_cs(cont);
1587 cpuset_filetype_t type = cft->private;
1588 switch (type) {
1589 case FILE_CPU_EXCLUSIVE:
1590 return is_cpu_exclusive(cs);
1591 case FILE_MEM_EXCLUSIVE:
1592 return is_mem_exclusive(cs);
1593 case FILE_MEM_HARDWALL:
1594 return is_mem_hardwall(cs);
1595 case FILE_SCHED_LOAD_BALANCE:
1596 return is_sched_load_balance(cs);
1597 case FILE_MEMORY_MIGRATE:
1598 return is_memory_migrate(cs);
1599 case FILE_MEMORY_PRESSURE_ENABLED:
1600 return cpuset_memory_pressure_enabled;
1601 case FILE_MEMORY_PRESSURE:
1602 return fmeter_getrate(&cs->fmeter);
1603 case FILE_SPREAD_PAGE:
1604 return is_spread_page(cs);
1605 case FILE_SPREAD_SLAB:
1606 return is_spread_slab(cs);
1607 default:
1608 BUG();
1611 /* Unreachable but makes gcc happy */
1612 return 0;
1615 static s64 cpuset_read_s64(struct cgroup *cont, struct cftype *cft)
1617 struct cpuset *cs = cgroup_cs(cont);
1618 cpuset_filetype_t type = cft->private;
1619 switch (type) {
1620 case FILE_SCHED_RELAX_DOMAIN_LEVEL:
1621 return cs->relax_domain_level;
1622 default:
1623 BUG();
1626 /* Unrechable but makes gcc happy */
1627 return 0;
1632 * for the common functions, 'private' gives the type of file
1635 static struct cftype files[] = {
1637 .name = "cpus",
1638 .read = cpuset_common_file_read,
1639 .write_string = cpuset_write_resmask,
1640 .max_write_len = (100U + 6 * NR_CPUS),
1641 .private = FILE_CPULIST,
1645 .name = "mems",
1646 .read = cpuset_common_file_read,
1647 .write_string = cpuset_write_resmask,
1648 .max_write_len = (100U + 6 * MAX_NUMNODES),
1649 .private = FILE_MEMLIST,
1653 .name = "cpu_exclusive",
1654 .read_u64 = cpuset_read_u64,
1655 .write_u64 = cpuset_write_u64,
1656 .private = FILE_CPU_EXCLUSIVE,
1660 .name = "mem_exclusive",
1661 .read_u64 = cpuset_read_u64,
1662 .write_u64 = cpuset_write_u64,
1663 .private = FILE_MEM_EXCLUSIVE,
1667 .name = "mem_hardwall",
1668 .read_u64 = cpuset_read_u64,
1669 .write_u64 = cpuset_write_u64,
1670 .private = FILE_MEM_HARDWALL,
1674 .name = "sched_load_balance",
1675 .read_u64 = cpuset_read_u64,
1676 .write_u64 = cpuset_write_u64,
1677 .private = FILE_SCHED_LOAD_BALANCE,
1681 .name = "sched_relax_domain_level",
1682 .read_s64 = cpuset_read_s64,
1683 .write_s64 = cpuset_write_s64,
1684 .private = FILE_SCHED_RELAX_DOMAIN_LEVEL,
1688 .name = "memory_migrate",
1689 .read_u64 = cpuset_read_u64,
1690 .write_u64 = cpuset_write_u64,
1691 .private = FILE_MEMORY_MIGRATE,
1695 .name = "memory_pressure",
1696 .read_u64 = cpuset_read_u64,
1697 .write_u64 = cpuset_write_u64,
1698 .private = FILE_MEMORY_PRESSURE,
1699 .mode = S_IRUGO,
1703 .name = "memory_spread_page",
1704 .read_u64 = cpuset_read_u64,
1705 .write_u64 = cpuset_write_u64,
1706 .private = FILE_SPREAD_PAGE,
1710 .name = "memory_spread_slab",
1711 .read_u64 = cpuset_read_u64,
1712 .write_u64 = cpuset_write_u64,
1713 .private = FILE_SPREAD_SLAB,
1717 static struct cftype cft_memory_pressure_enabled = {
1718 .name = "memory_pressure_enabled",
1719 .read_u64 = cpuset_read_u64,
1720 .write_u64 = cpuset_write_u64,
1721 .private = FILE_MEMORY_PRESSURE_ENABLED,
1724 static int cpuset_populate(struct cgroup_subsys *ss, struct cgroup *cont)
1726 int err;
1728 err = cgroup_add_files(cont, ss, files, ARRAY_SIZE(files));
1729 if (err)
1730 return err;
1731 /* memory_pressure_enabled is in root cpuset only */
1732 if (!cont->parent)
1733 err = cgroup_add_file(cont, ss,
1734 &cft_memory_pressure_enabled);
1735 return err;
1739 * post_clone() is called at the end of cgroup_clone().
1740 * 'cgroup' was just created automatically as a result of
1741 * a cgroup_clone(), and the current task is about to
1742 * be moved into 'cgroup'.
1744 * Currently we refuse to set up the cgroup - thereby
1745 * refusing the task to be entered, and as a result refusing
1746 * the sys_unshare() or clone() which initiated it - if any
1747 * sibling cpusets have exclusive cpus or mem.
1749 * If this becomes a problem for some users who wish to
1750 * allow that scenario, then cpuset_post_clone() could be
1751 * changed to grant parent->cpus_allowed-sibling_cpus_exclusive
1752 * (and likewise for mems) to the new cgroup. Called with cgroup_mutex
1753 * held.
1755 static void cpuset_post_clone(struct cgroup_subsys *ss,
1756 struct cgroup *cgroup)
1758 struct cgroup *parent, *child;
1759 struct cpuset *cs, *parent_cs;
1761 parent = cgroup->parent;
1762 list_for_each_entry(child, &parent->children, sibling) {
1763 cs = cgroup_cs(child);
1764 if (is_mem_exclusive(cs) || is_cpu_exclusive(cs))
1765 return;
1767 cs = cgroup_cs(cgroup);
1768 parent_cs = cgroup_cs(parent);
1770 cs->mems_allowed = parent_cs->mems_allowed;
1771 cpumask_copy(cs->cpus_allowed, parent_cs->cpus_allowed);
1772 return;
1776 * cpuset_create - create a cpuset
1777 * ss: cpuset cgroup subsystem
1778 * cont: control group that the new cpuset will be part of
1781 static struct cgroup_subsys_state *cpuset_create(
1782 struct cgroup_subsys *ss,
1783 struct cgroup *cont)
1785 struct cpuset *cs;
1786 struct cpuset *parent;
1788 if (!cont->parent) {
1789 /* This is early initialization for the top cgroup */
1790 top_cpuset.mems_generation = cpuset_mems_generation++;
1791 return &top_cpuset.css;
1793 parent = cgroup_cs(cont->parent);
1794 cs = kmalloc(sizeof(*cs), GFP_KERNEL);
1795 if (!cs)
1796 return ERR_PTR(-ENOMEM);
1797 if (!alloc_cpumask_var(&cs->cpus_allowed, GFP_KERNEL)) {
1798 kfree(cs);
1799 return ERR_PTR(-ENOMEM);
1802 cpuset_update_task_memory_state();
1803 cs->flags = 0;
1804 if (is_spread_page(parent))
1805 set_bit(CS_SPREAD_PAGE, &cs->flags);
1806 if (is_spread_slab(parent))
1807 set_bit(CS_SPREAD_SLAB, &cs->flags);
1808 set_bit(CS_SCHED_LOAD_BALANCE, &cs->flags);
1809 cpumask_clear(cs->cpus_allowed);
1810 nodes_clear(cs->mems_allowed);
1811 cs->mems_generation = cpuset_mems_generation++;
1812 fmeter_init(&cs->fmeter);
1813 cs->relax_domain_level = -1;
1815 cs->parent = parent;
1816 number_of_cpusets++;
1817 return &cs->css ;
1821 * If the cpuset being removed has its flag 'sched_load_balance'
1822 * enabled, then simulate turning sched_load_balance off, which
1823 * will call async_rebuild_sched_domains().
1826 static void cpuset_destroy(struct cgroup_subsys *ss, struct cgroup *cont)
1828 struct cpuset *cs = cgroup_cs(cont);
1830 cpuset_update_task_memory_state();
1832 if (is_sched_load_balance(cs))
1833 update_flag(CS_SCHED_LOAD_BALANCE, cs, 0);
1835 number_of_cpusets--;
1836 free_cpumask_var(cs->cpus_allowed);
1837 kfree(cs);
1840 struct cgroup_subsys cpuset_subsys = {
1841 .name = "cpuset",
1842 .create = cpuset_create,
1843 .destroy = cpuset_destroy,
1844 .can_attach = cpuset_can_attach,
1845 .attach = cpuset_attach,
1846 .populate = cpuset_populate,
1847 .post_clone = cpuset_post_clone,
1848 .subsys_id = cpuset_subsys_id,
1849 .early_init = 1,
1853 * cpuset_init_early - just enough so that the calls to
1854 * cpuset_update_task_memory_state() in early init code
1855 * are harmless.
1858 int __init cpuset_init_early(void)
1860 alloc_bootmem_cpumask_var(&top_cpuset.cpus_allowed);
1862 top_cpuset.mems_generation = cpuset_mems_generation++;
1863 return 0;
1868 * cpuset_init - initialize cpusets at system boot
1870 * Description: Initialize top_cpuset and the cpuset internal file system,
1873 int __init cpuset_init(void)
1875 int err = 0;
1877 cpumask_setall(top_cpuset.cpus_allowed);
1878 nodes_setall(top_cpuset.mems_allowed);
1880 fmeter_init(&top_cpuset.fmeter);
1881 top_cpuset.mems_generation = cpuset_mems_generation++;
1882 set_bit(CS_SCHED_LOAD_BALANCE, &top_cpuset.flags);
1883 top_cpuset.relax_domain_level = -1;
1885 err = register_filesystem(&cpuset_fs_type);
1886 if (err < 0)
1887 return err;
1889 if (!alloc_cpumask_var(&cpus_attach, GFP_KERNEL))
1890 BUG();
1892 number_of_cpusets = 1;
1893 return 0;
1897 * cpuset_do_move_task - move a given task to another cpuset
1898 * @tsk: pointer to task_struct the task to move
1899 * @scan: struct cgroup_scanner contained in its struct cpuset_hotplug_scanner
1901 * Called by cgroup_scan_tasks() for each task in a cgroup.
1902 * Return nonzero to stop the walk through the tasks.
1904 static void cpuset_do_move_task(struct task_struct *tsk,
1905 struct cgroup_scanner *scan)
1907 struct cgroup *new_cgroup = scan->data;
1909 cgroup_attach_task(new_cgroup, tsk);
1913 * move_member_tasks_to_cpuset - move tasks from one cpuset to another
1914 * @from: cpuset in which the tasks currently reside
1915 * @to: cpuset to which the tasks will be moved
1917 * Called with cgroup_mutex held
1918 * callback_mutex must not be held, as cpuset_attach() will take it.
1920 * The cgroup_scan_tasks() function will scan all the tasks in a cgroup,
1921 * calling callback functions for each.
1923 static void move_member_tasks_to_cpuset(struct cpuset *from, struct cpuset *to)
1925 struct cgroup_scanner scan;
1927 scan.cg = from->css.cgroup;
1928 scan.test_task = NULL; /* select all tasks in cgroup */
1929 scan.process_task = cpuset_do_move_task;
1930 scan.heap = NULL;
1931 scan.data = to->css.cgroup;
1933 if (cgroup_scan_tasks(&scan))
1934 printk(KERN_ERR "move_member_tasks_to_cpuset: "
1935 "cgroup_scan_tasks failed\n");
1939 * If CPU and/or memory hotplug handlers, below, unplug any CPUs
1940 * or memory nodes, we need to walk over the cpuset hierarchy,
1941 * removing that CPU or node from all cpusets. If this removes the
1942 * last CPU or node from a cpuset, then move the tasks in the empty
1943 * cpuset to its next-highest non-empty parent.
1945 * Called with cgroup_mutex held
1946 * callback_mutex must not be held, as cpuset_attach() will take it.
1948 static void remove_tasks_in_empty_cpuset(struct cpuset *cs)
1950 struct cpuset *parent;
1953 * The cgroup's css_sets list is in use if there are tasks
1954 * in the cpuset; the list is empty if there are none;
1955 * the cs->css.refcnt seems always 0.
1957 if (list_empty(&cs->css.cgroup->css_sets))
1958 return;
1961 * Find its next-highest non-empty parent, (top cpuset
1962 * has online cpus, so can't be empty).
1964 parent = cs->parent;
1965 while (cpumask_empty(parent->cpus_allowed) ||
1966 nodes_empty(parent->mems_allowed))
1967 parent = parent->parent;
1969 move_member_tasks_to_cpuset(cs, parent);
1973 * Walk the specified cpuset subtree and look for empty cpusets.
1974 * The tasks of such cpuset must be moved to a parent cpuset.
1976 * Called with cgroup_mutex held. We take callback_mutex to modify
1977 * cpus_allowed and mems_allowed.
1979 * This walk processes the tree from top to bottom, completing one layer
1980 * before dropping down to the next. It always processes a node before
1981 * any of its children.
1983 * For now, since we lack memory hot unplug, we'll never see a cpuset
1984 * that has tasks along with an empty 'mems'. But if we did see such
1985 * a cpuset, we'd handle it just like we do if its 'cpus' was empty.
1987 static void scan_for_empty_cpusets(struct cpuset *root)
1989 LIST_HEAD(queue);
1990 struct cpuset *cp; /* scans cpusets being updated */
1991 struct cpuset *child; /* scans child cpusets of cp */
1992 struct cgroup *cont;
1993 nodemask_t oldmems;
1995 list_add_tail((struct list_head *)&root->stack_list, &queue);
1997 while (!list_empty(&queue)) {
1998 cp = list_first_entry(&queue, struct cpuset, stack_list);
1999 list_del(queue.next);
2000 list_for_each_entry(cont, &cp->css.cgroup->children, sibling) {
2001 child = cgroup_cs(cont);
2002 list_add_tail(&child->stack_list, &queue);
2005 /* Continue past cpusets with all cpus, mems online */
2006 if (cpumask_subset(cp->cpus_allowed, cpu_online_mask) &&
2007 nodes_subset(cp->mems_allowed, node_states[N_HIGH_MEMORY]))
2008 continue;
2010 oldmems = cp->mems_allowed;
2012 /* Remove offline cpus and mems from this cpuset. */
2013 mutex_lock(&callback_mutex);
2014 cpumask_and(cp->cpus_allowed, cp->cpus_allowed,
2015 cpu_online_mask);
2016 nodes_and(cp->mems_allowed, cp->mems_allowed,
2017 node_states[N_HIGH_MEMORY]);
2018 mutex_unlock(&callback_mutex);
2020 /* Move tasks from the empty cpuset to a parent */
2021 if (cpumask_empty(cp->cpus_allowed) ||
2022 nodes_empty(cp->mems_allowed))
2023 remove_tasks_in_empty_cpuset(cp);
2024 else {
2025 update_tasks_cpumask(cp, NULL);
2026 update_tasks_nodemask(cp, &oldmems, NULL);
2032 * The top_cpuset tracks what CPUs and Memory Nodes are online,
2033 * period. This is necessary in order to make cpusets transparent
2034 * (of no affect) on systems that are actively using CPU hotplug
2035 * but making no active use of cpusets.
2037 * This routine ensures that top_cpuset.cpus_allowed tracks
2038 * cpu_online_map on each CPU hotplug (cpuhp) event.
2040 * Called within get_online_cpus(). Needs to call cgroup_lock()
2041 * before calling generate_sched_domains().
2043 static int cpuset_track_online_cpus(struct notifier_block *unused_nb,
2044 unsigned long phase, void *unused_cpu)
2046 struct sched_domain_attr *attr;
2047 struct cpumask *doms;
2048 int ndoms;
2050 switch (phase) {
2051 case CPU_ONLINE:
2052 case CPU_ONLINE_FROZEN:
2053 case CPU_DEAD:
2054 case CPU_DEAD_FROZEN:
2055 break;
2057 default:
2058 return NOTIFY_DONE;
2061 cgroup_lock();
2062 mutex_lock(&callback_mutex);
2063 cpumask_copy(top_cpuset.cpus_allowed, cpu_online_mask);
2064 mutex_unlock(&callback_mutex);
2065 scan_for_empty_cpusets(&top_cpuset);
2066 ndoms = generate_sched_domains(&doms, &attr);
2067 cgroup_unlock();
2069 /* Have scheduler rebuild the domains */
2070 partition_sched_domains(ndoms, doms, attr);
2072 return NOTIFY_OK;
2075 #ifdef CONFIG_MEMORY_HOTPLUG
2077 * Keep top_cpuset.mems_allowed tracking node_states[N_HIGH_MEMORY].
2078 * Call this routine anytime after node_states[N_HIGH_MEMORY] changes.
2079 * See also the previous routine cpuset_track_online_cpus().
2081 static int cpuset_track_online_nodes(struct notifier_block *self,
2082 unsigned long action, void *arg)
2084 cgroup_lock();
2085 switch (action) {
2086 case MEM_ONLINE:
2087 case MEM_OFFLINE:
2088 mutex_lock(&callback_mutex);
2089 top_cpuset.mems_allowed = node_states[N_HIGH_MEMORY];
2090 mutex_unlock(&callback_mutex);
2091 if (action == MEM_OFFLINE)
2092 scan_for_empty_cpusets(&top_cpuset);
2093 break;
2094 default:
2095 break;
2097 cgroup_unlock();
2098 return NOTIFY_OK;
2100 #endif
2103 * cpuset_init_smp - initialize cpus_allowed
2105 * Description: Finish top cpuset after cpu, node maps are initialized
2108 void __init cpuset_init_smp(void)
2110 cpumask_copy(top_cpuset.cpus_allowed, cpu_online_mask);
2111 top_cpuset.mems_allowed = node_states[N_HIGH_MEMORY];
2113 hotcpu_notifier(cpuset_track_online_cpus, 0);
2114 hotplug_memory_notifier(cpuset_track_online_nodes, 10);
2116 cpuset_wq = create_singlethread_workqueue("cpuset");
2117 BUG_ON(!cpuset_wq);
2121 * cpuset_cpus_allowed - return cpus_allowed mask from a tasks cpuset.
2122 * @tsk: pointer to task_struct from which to obtain cpuset->cpus_allowed.
2123 * @pmask: pointer to struct cpumask variable to receive cpus_allowed set.
2125 * Description: Returns the cpumask_var_t cpus_allowed of the cpuset
2126 * attached to the specified @tsk. Guaranteed to return some non-empty
2127 * subset of cpu_online_map, even if this means going outside the
2128 * tasks cpuset.
2131 void cpuset_cpus_allowed(struct task_struct *tsk, struct cpumask *pmask)
2133 mutex_lock(&callback_mutex);
2134 cpuset_cpus_allowed_locked(tsk, pmask);
2135 mutex_unlock(&callback_mutex);
2139 * cpuset_cpus_allowed_locked - return cpus_allowed mask from a tasks cpuset.
2140 * Must be called with callback_mutex held.
2142 void cpuset_cpus_allowed_locked(struct task_struct *tsk, struct cpumask *pmask)
2144 task_lock(tsk);
2145 guarantee_online_cpus(task_cs(tsk), pmask);
2146 task_unlock(tsk);
2149 void cpuset_init_current_mems_allowed(void)
2151 nodes_setall(current->mems_allowed);
2155 * cpuset_mems_allowed - return mems_allowed mask from a tasks cpuset.
2156 * @tsk: pointer to task_struct from which to obtain cpuset->mems_allowed.
2158 * Description: Returns the nodemask_t mems_allowed of the cpuset
2159 * attached to the specified @tsk. Guaranteed to return some non-empty
2160 * subset of node_states[N_HIGH_MEMORY], even if this means going outside the
2161 * tasks cpuset.
2164 nodemask_t cpuset_mems_allowed(struct task_struct *tsk)
2166 nodemask_t mask;
2168 mutex_lock(&callback_mutex);
2169 task_lock(tsk);
2170 guarantee_online_mems(task_cs(tsk), &mask);
2171 task_unlock(tsk);
2172 mutex_unlock(&callback_mutex);
2174 return mask;
2178 * cpuset_nodemask_valid_mems_allowed - check nodemask vs. curremt mems_allowed
2179 * @nodemask: the nodemask to be checked
2181 * Are any of the nodes in the nodemask allowed in current->mems_allowed?
2183 int cpuset_nodemask_valid_mems_allowed(nodemask_t *nodemask)
2185 return nodes_intersects(*nodemask, current->mems_allowed);
2189 * nearest_hardwall_ancestor() - Returns the nearest mem_exclusive or
2190 * mem_hardwall ancestor to the specified cpuset. Call holding
2191 * callback_mutex. If no ancestor is mem_exclusive or mem_hardwall
2192 * (an unusual configuration), then returns the root cpuset.
2194 static const struct cpuset *nearest_hardwall_ancestor(const struct cpuset *cs)
2196 while (!(is_mem_exclusive(cs) || is_mem_hardwall(cs)) && cs->parent)
2197 cs = cs->parent;
2198 return cs;
2202 * cpuset_node_allowed_softwall - Can we allocate on a memory node?
2203 * @node: is this an allowed node?
2204 * @gfp_mask: memory allocation flags
2206 * If we're in interrupt, yes, we can always allocate. If __GFP_THISNODE is
2207 * set, yes, we can always allocate. If node is in our task's mems_allowed,
2208 * yes. If it's not a __GFP_HARDWALL request and this node is in the nearest
2209 * hardwalled cpuset ancestor to this task's cpuset, yes. If the task has been
2210 * OOM killed and has access to memory reserves as specified by the TIF_MEMDIE
2211 * flag, yes.
2212 * Otherwise, no.
2214 * If __GFP_HARDWALL is set, cpuset_node_allowed_softwall() reduces to
2215 * cpuset_node_allowed_hardwall(). Otherwise, cpuset_node_allowed_softwall()
2216 * might sleep, and might allow a node from an enclosing cpuset.
2218 * cpuset_node_allowed_hardwall() only handles the simpler case of hardwall
2219 * cpusets, and never sleeps.
2221 * The __GFP_THISNODE placement logic is really handled elsewhere,
2222 * by forcibly using a zonelist starting at a specified node, and by
2223 * (in get_page_from_freelist()) refusing to consider the zones for
2224 * any node on the zonelist except the first. By the time any such
2225 * calls get to this routine, we should just shut up and say 'yes'.
2227 * GFP_USER allocations are marked with the __GFP_HARDWALL bit,
2228 * and do not allow allocations outside the current tasks cpuset
2229 * unless the task has been OOM killed as is marked TIF_MEMDIE.
2230 * GFP_KERNEL allocations are not so marked, so can escape to the
2231 * nearest enclosing hardwalled ancestor cpuset.
2233 * Scanning up parent cpusets requires callback_mutex. The
2234 * __alloc_pages() routine only calls here with __GFP_HARDWALL bit
2235 * _not_ set if it's a GFP_KERNEL allocation, and all nodes in the
2236 * current tasks mems_allowed came up empty on the first pass over
2237 * the zonelist. So only GFP_KERNEL allocations, if all nodes in the
2238 * cpuset are short of memory, might require taking the callback_mutex
2239 * mutex.
2241 * The first call here from mm/page_alloc:get_page_from_freelist()
2242 * has __GFP_HARDWALL set in gfp_mask, enforcing hardwall cpusets,
2243 * so no allocation on a node outside the cpuset is allowed (unless
2244 * in interrupt, of course).
2246 * The second pass through get_page_from_freelist() doesn't even call
2247 * here for GFP_ATOMIC calls. For those calls, the __alloc_pages()
2248 * variable 'wait' is not set, and the bit ALLOC_CPUSET is not set
2249 * in alloc_flags. That logic and the checks below have the combined
2250 * affect that:
2251 * in_interrupt - any node ok (current task context irrelevant)
2252 * GFP_ATOMIC - any node ok
2253 * TIF_MEMDIE - any node ok
2254 * GFP_KERNEL - any node in enclosing hardwalled cpuset ok
2255 * GFP_USER - only nodes in current tasks mems allowed ok.
2257 * Rule:
2258 * Don't call cpuset_node_allowed_softwall if you can't sleep, unless you
2259 * pass in the __GFP_HARDWALL flag set in gfp_flag, which disables
2260 * the code that might scan up ancestor cpusets and sleep.
2262 int __cpuset_node_allowed_softwall(int node, gfp_t gfp_mask)
2264 const struct cpuset *cs; /* current cpuset ancestors */
2265 int allowed; /* is allocation in zone z allowed? */
2267 if (in_interrupt() || (gfp_mask & __GFP_THISNODE))
2268 return 1;
2269 might_sleep_if(!(gfp_mask & __GFP_HARDWALL));
2270 if (node_isset(node, current->mems_allowed))
2271 return 1;
2273 * Allow tasks that have access to memory reserves because they have
2274 * been OOM killed to get memory anywhere.
2276 if (unlikely(test_thread_flag(TIF_MEMDIE)))
2277 return 1;
2278 if (gfp_mask & __GFP_HARDWALL) /* If hardwall request, stop here */
2279 return 0;
2281 if (current->flags & PF_EXITING) /* Let dying task have memory */
2282 return 1;
2284 /* Not hardwall and node outside mems_allowed: scan up cpusets */
2285 mutex_lock(&callback_mutex);
2287 task_lock(current);
2288 cs = nearest_hardwall_ancestor(task_cs(current));
2289 task_unlock(current);
2291 allowed = node_isset(node, cs->mems_allowed);
2292 mutex_unlock(&callback_mutex);
2293 return allowed;
2297 * cpuset_node_allowed_hardwall - Can we allocate on a memory node?
2298 * @node: is this an allowed node?
2299 * @gfp_mask: memory allocation flags
2301 * If we're in interrupt, yes, we can always allocate. If __GFP_THISNODE is
2302 * set, yes, we can always allocate. If node is in our task's mems_allowed,
2303 * yes. If the task has been OOM killed and has access to memory reserves as
2304 * specified by the TIF_MEMDIE flag, yes.
2305 * Otherwise, no.
2307 * The __GFP_THISNODE placement logic is really handled elsewhere,
2308 * by forcibly using a zonelist starting at a specified node, and by
2309 * (in get_page_from_freelist()) refusing to consider the zones for
2310 * any node on the zonelist except the first. By the time any such
2311 * calls get to this routine, we should just shut up and say 'yes'.
2313 * Unlike the cpuset_node_allowed_softwall() variant, above,
2314 * this variant requires that the node be in the current task's
2315 * mems_allowed or that we're in interrupt. It does not scan up the
2316 * cpuset hierarchy for the nearest enclosing mem_exclusive cpuset.
2317 * It never sleeps.
2319 int __cpuset_node_allowed_hardwall(int node, gfp_t gfp_mask)
2321 if (in_interrupt() || (gfp_mask & __GFP_THISNODE))
2322 return 1;
2323 if (node_isset(node, current->mems_allowed))
2324 return 1;
2326 * Allow tasks that have access to memory reserves because they have
2327 * been OOM killed to get memory anywhere.
2329 if (unlikely(test_thread_flag(TIF_MEMDIE)))
2330 return 1;
2331 return 0;
2335 * cpuset_lock - lock out any changes to cpuset structures
2337 * The out of memory (oom) code needs to mutex_lock cpusets
2338 * from being changed while it scans the tasklist looking for a
2339 * task in an overlapping cpuset. Expose callback_mutex via this
2340 * cpuset_lock() routine, so the oom code can lock it, before
2341 * locking the task list. The tasklist_lock is a spinlock, so
2342 * must be taken inside callback_mutex.
2345 void cpuset_lock(void)
2347 mutex_lock(&callback_mutex);
2351 * cpuset_unlock - release lock on cpuset changes
2353 * Undo the lock taken in a previous cpuset_lock() call.
2356 void cpuset_unlock(void)
2358 mutex_unlock(&callback_mutex);
2362 * cpuset_mem_spread_node() - On which node to begin search for a page
2364 * If a task is marked PF_SPREAD_PAGE or PF_SPREAD_SLAB (as for
2365 * tasks in a cpuset with is_spread_page or is_spread_slab set),
2366 * and if the memory allocation used cpuset_mem_spread_node()
2367 * to determine on which node to start looking, as it will for
2368 * certain page cache or slab cache pages such as used for file
2369 * system buffers and inode caches, then instead of starting on the
2370 * local node to look for a free page, rather spread the starting
2371 * node around the tasks mems_allowed nodes.
2373 * We don't have to worry about the returned node being offline
2374 * because "it can't happen", and even if it did, it would be ok.
2376 * The routines calling guarantee_online_mems() are careful to
2377 * only set nodes in task->mems_allowed that are online. So it
2378 * should not be possible for the following code to return an
2379 * offline node. But if it did, that would be ok, as this routine
2380 * is not returning the node where the allocation must be, only
2381 * the node where the search should start. The zonelist passed to
2382 * __alloc_pages() will include all nodes. If the slab allocator
2383 * is passed an offline node, it will fall back to the local node.
2384 * See kmem_cache_alloc_node().
2387 int cpuset_mem_spread_node(void)
2389 int node;
2391 node = next_node(current->cpuset_mem_spread_rotor, current->mems_allowed);
2392 if (node == MAX_NUMNODES)
2393 node = first_node(current->mems_allowed);
2394 current->cpuset_mem_spread_rotor = node;
2395 return node;
2397 EXPORT_SYMBOL_GPL(cpuset_mem_spread_node);
2400 * cpuset_mems_allowed_intersects - Does @tsk1's mems_allowed intersect @tsk2's?
2401 * @tsk1: pointer to task_struct of some task.
2402 * @tsk2: pointer to task_struct of some other task.
2404 * Description: Return true if @tsk1's mems_allowed intersects the
2405 * mems_allowed of @tsk2. Used by the OOM killer to determine if
2406 * one of the task's memory usage might impact the memory available
2407 * to the other.
2410 int cpuset_mems_allowed_intersects(const struct task_struct *tsk1,
2411 const struct task_struct *tsk2)
2413 return nodes_intersects(tsk1->mems_allowed, tsk2->mems_allowed);
2417 * cpuset_print_task_mems_allowed - prints task's cpuset and mems_allowed
2418 * @task: pointer to task_struct of some task.
2420 * Description: Prints @task's name, cpuset name, and cached copy of its
2421 * mems_allowed to the kernel log. Must hold task_lock(task) to allow
2422 * dereferencing task_cs(task).
2424 void cpuset_print_task_mems_allowed(struct task_struct *tsk)
2426 struct dentry *dentry;
2428 dentry = task_cs(tsk)->css.cgroup->dentry;
2429 spin_lock(&cpuset_buffer_lock);
2430 snprintf(cpuset_name, CPUSET_NAME_LEN,
2431 dentry ? (const char *)dentry->d_name.name : "/");
2432 nodelist_scnprintf(cpuset_nodelist, CPUSET_NODELIST_LEN,
2433 tsk->mems_allowed);
2434 printk(KERN_INFO "%s cpuset=%s mems_allowed=%s\n",
2435 tsk->comm, cpuset_name, cpuset_nodelist);
2436 spin_unlock(&cpuset_buffer_lock);
2440 * Collection of memory_pressure is suppressed unless
2441 * this flag is enabled by writing "1" to the special
2442 * cpuset file 'memory_pressure_enabled' in the root cpuset.
2445 int cpuset_memory_pressure_enabled __read_mostly;
2448 * cpuset_memory_pressure_bump - keep stats of per-cpuset reclaims.
2450 * Keep a running average of the rate of synchronous (direct)
2451 * page reclaim efforts initiated by tasks in each cpuset.
2453 * This represents the rate at which some task in the cpuset
2454 * ran low on memory on all nodes it was allowed to use, and
2455 * had to enter the kernels page reclaim code in an effort to
2456 * create more free memory by tossing clean pages or swapping
2457 * or writing dirty pages.
2459 * Display to user space in the per-cpuset read-only file
2460 * "memory_pressure". Value displayed is an integer
2461 * representing the recent rate of entry into the synchronous
2462 * (direct) page reclaim by any task attached to the cpuset.
2465 void __cpuset_memory_pressure_bump(void)
2467 task_lock(current);
2468 fmeter_markevent(&task_cs(current)->fmeter);
2469 task_unlock(current);
2472 #ifdef CONFIG_PROC_PID_CPUSET
2474 * proc_cpuset_show()
2475 * - Print tasks cpuset path into seq_file.
2476 * - Used for /proc/<pid>/cpuset.
2477 * - No need to task_lock(tsk) on this tsk->cpuset reference, as it
2478 * doesn't really matter if tsk->cpuset changes after we read it,
2479 * and we take cgroup_mutex, keeping cpuset_attach() from changing it
2480 * anyway.
2482 static int proc_cpuset_show(struct seq_file *m, void *unused_v)
2484 struct pid *pid;
2485 struct task_struct *tsk;
2486 char *buf;
2487 struct cgroup_subsys_state *css;
2488 int retval;
2490 retval = -ENOMEM;
2491 buf = kmalloc(PAGE_SIZE, GFP_KERNEL);
2492 if (!buf)
2493 goto out;
2495 retval = -ESRCH;
2496 pid = m->private;
2497 tsk = get_pid_task(pid, PIDTYPE_PID);
2498 if (!tsk)
2499 goto out_free;
2501 retval = -EINVAL;
2502 cgroup_lock();
2503 css = task_subsys_state(tsk, cpuset_subsys_id);
2504 retval = cgroup_path(css->cgroup, buf, PAGE_SIZE);
2505 if (retval < 0)
2506 goto out_unlock;
2507 seq_puts(m, buf);
2508 seq_putc(m, '\n');
2509 out_unlock:
2510 cgroup_unlock();
2511 put_task_struct(tsk);
2512 out_free:
2513 kfree(buf);
2514 out:
2515 return retval;
2518 static int cpuset_open(struct inode *inode, struct file *file)
2520 struct pid *pid = PROC_I(inode)->pid;
2521 return single_open(file, proc_cpuset_show, pid);
2524 const struct file_operations proc_cpuset_operations = {
2525 .open = cpuset_open,
2526 .read = seq_read,
2527 .llseek = seq_lseek,
2528 .release = single_release,
2530 #endif /* CONFIG_PROC_PID_CPUSET */
2532 /* Display task cpus_allowed, mems_allowed in /proc/<pid>/status file. */
2533 void cpuset_task_status_allowed(struct seq_file *m, struct task_struct *task)
2535 seq_printf(m, "Cpus_allowed:\t");
2536 seq_cpumask(m, &task->cpus_allowed);
2537 seq_printf(m, "\n");
2538 seq_printf(m, "Cpus_allowed_list:\t");
2539 seq_cpumask_list(m, &task->cpus_allowed);
2540 seq_printf(m, "\n");
2541 seq_printf(m, "Mems_allowed:\t");
2542 seq_nodemask(m, &task->mems_allowed);
2543 seq_printf(m, "\n");
2544 seq_printf(m, "Mems_allowed_list:\t");
2545 seq_nodemask_list(m, &task->mems_allowed);
2546 seq_printf(m, "\n");