Merge branch 'for-linus' of git://neil.brown.name/md
[linux-2.6/zen-sources.git] / kernel / cpuset.c
blob3e00526f52ec14d56eb89ee1ab5c1a40fbed1331
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/module.h>
40 #include <linux/mount.h>
41 #include <linux/namei.h>
42 #include <linux/pagemap.h>
43 #include <linux/proc_fs.h>
44 #include <linux/rcupdate.h>
45 #include <linux/sched.h>
46 #include <linux/seq_file.h>
47 #include <linux/security.h>
48 #include <linux/slab.h>
49 #include <linux/spinlock.h>
50 #include <linux/stat.h>
51 #include <linux/string.h>
52 #include <linux/time.h>
53 #include <linux/backing-dev.h>
54 #include <linux/sort.h>
56 #include <asm/uaccess.h>
57 #include <asm/atomic.h>
58 #include <linux/mutex.h>
59 #include <linux/workqueue.h>
60 #include <linux/cgroup.h>
63 * Tracks how many cpusets are currently defined in system.
64 * When there is only one cpuset (the root cpuset) we can
65 * short circuit some hooks.
67 int number_of_cpusets __read_mostly;
69 /* Forward declare cgroup structures */
70 struct cgroup_subsys cpuset_subsys;
71 struct cpuset;
73 /* See "Frequency meter" comments, below. */
75 struct fmeter {
76 int cnt; /* unprocessed events count */
77 int val; /* most recent output value */
78 time_t time; /* clock (secs) when val computed */
79 spinlock_t lock; /* guards read or write of above */
82 struct cpuset {
83 struct cgroup_subsys_state css;
85 unsigned long flags; /* "unsigned long" so bitops work */
86 cpumask_t cpus_allowed; /* CPUs allowed to tasks in cpuset */
87 nodemask_t mems_allowed; /* Memory Nodes allowed to tasks */
89 struct cpuset *parent; /* my parent */
92 * Copy of global cpuset_mems_generation as of the most
93 * recent time this cpuset changed its mems_allowed.
95 int mems_generation;
97 struct fmeter fmeter; /* memory_pressure filter */
99 /* partition number for rebuild_sched_domains() */
100 int pn;
102 /* for custom sched domain */
103 int relax_domain_level;
105 /* used for walking a cpuset heirarchy */
106 struct list_head stack_list;
109 /* Retrieve the cpuset for a cgroup */
110 static inline struct cpuset *cgroup_cs(struct cgroup *cont)
112 return container_of(cgroup_subsys_state(cont, cpuset_subsys_id),
113 struct cpuset, css);
116 /* Retrieve the cpuset for a task */
117 static inline struct cpuset *task_cs(struct task_struct *task)
119 return container_of(task_subsys_state(task, cpuset_subsys_id),
120 struct cpuset, css);
122 struct cpuset_hotplug_scanner {
123 struct cgroup_scanner scan;
124 struct cgroup *to;
127 /* bits in struct cpuset flags field */
128 typedef enum {
129 CS_CPU_EXCLUSIVE,
130 CS_MEM_EXCLUSIVE,
131 CS_MEM_HARDWALL,
132 CS_MEMORY_MIGRATE,
133 CS_SCHED_LOAD_BALANCE,
134 CS_SPREAD_PAGE,
135 CS_SPREAD_SLAB,
136 } cpuset_flagbits_t;
138 /* convenient tests for these bits */
139 static inline int is_cpu_exclusive(const struct cpuset *cs)
141 return test_bit(CS_CPU_EXCLUSIVE, &cs->flags);
144 static inline int is_mem_exclusive(const struct cpuset *cs)
146 return test_bit(CS_MEM_EXCLUSIVE, &cs->flags);
149 static inline int is_mem_hardwall(const struct cpuset *cs)
151 return test_bit(CS_MEM_HARDWALL, &cs->flags);
154 static inline int is_sched_load_balance(const struct cpuset *cs)
156 return test_bit(CS_SCHED_LOAD_BALANCE, &cs->flags);
159 static inline int is_memory_migrate(const struct cpuset *cs)
161 return test_bit(CS_MEMORY_MIGRATE, &cs->flags);
164 static inline int is_spread_page(const struct cpuset *cs)
166 return test_bit(CS_SPREAD_PAGE, &cs->flags);
169 static inline int is_spread_slab(const struct cpuset *cs)
171 return test_bit(CS_SPREAD_SLAB, &cs->flags);
175 * Increment this integer everytime any cpuset changes its
176 * mems_allowed value. Users of cpusets can track this generation
177 * number, and avoid having to lock and reload mems_allowed unless
178 * the cpuset they're using changes generation.
180 * A single, global generation is needed because cpuset_attach_task() could
181 * reattach a task to a different cpuset, which must not have its
182 * generation numbers aliased with those of that tasks previous cpuset.
184 * Generations are needed for mems_allowed because one task cannot
185 * modify another's memory placement. So we must enable every task,
186 * on every visit to __alloc_pages(), to efficiently check whether
187 * its current->cpuset->mems_allowed has changed, requiring an update
188 * of its current->mems_allowed.
190 * Since writes to cpuset_mems_generation are guarded by the cgroup lock
191 * there is no need to mark it atomic.
193 static int cpuset_mems_generation;
195 static struct cpuset top_cpuset = {
196 .flags = ((1 << CS_CPU_EXCLUSIVE) | (1 << CS_MEM_EXCLUSIVE)),
197 .cpus_allowed = CPU_MASK_ALL,
198 .mems_allowed = NODE_MASK_ALL,
202 * There are two global mutexes guarding cpuset structures. The first
203 * is the main control groups cgroup_mutex, accessed via
204 * cgroup_lock()/cgroup_unlock(). The second is the cpuset-specific
205 * callback_mutex, below. They can nest. It is ok to first take
206 * cgroup_mutex, then nest callback_mutex. We also require taking
207 * task_lock() when dereferencing a task's cpuset pointer. See "The
208 * task_lock() exception", at the end of this comment.
210 * A task must hold both mutexes to modify cpusets. If a task
211 * holds cgroup_mutex, then it blocks others wanting that mutex,
212 * ensuring that it is the only task able to also acquire callback_mutex
213 * and be able to modify cpusets. It can perform various checks on
214 * the cpuset structure first, knowing nothing will change. It can
215 * also allocate memory while just holding cgroup_mutex. While it is
216 * performing these checks, various callback routines can briefly
217 * acquire callback_mutex to query cpusets. Once it is ready to make
218 * the changes, it takes callback_mutex, blocking everyone else.
220 * Calls to the kernel memory allocator can not be made while holding
221 * callback_mutex, as that would risk double tripping on callback_mutex
222 * from one of the callbacks into the cpuset code from within
223 * __alloc_pages().
225 * If a task is only holding callback_mutex, then it has read-only
226 * access to cpusets.
228 * The task_struct fields mems_allowed and mems_generation may only
229 * be accessed in the context of that task, so require no locks.
231 * The cpuset_common_file_read() handlers only hold callback_mutex across
232 * small pieces of code, such as when reading out possibly multi-word
233 * cpumasks and nodemasks.
235 * Accessing a task's cpuset should be done in accordance with the
236 * guidelines for accessing subsystem state in kernel/cgroup.c
239 static DEFINE_MUTEX(callback_mutex);
242 * This is ugly, but preserves the userspace API for existing cpuset
243 * users. If someone tries to mount the "cpuset" filesystem, we
244 * silently switch it to mount "cgroup" instead
246 static int cpuset_get_sb(struct file_system_type *fs_type,
247 int flags, const char *unused_dev_name,
248 void *data, struct vfsmount *mnt)
250 struct file_system_type *cgroup_fs = get_fs_type("cgroup");
251 int ret = -ENODEV;
252 if (cgroup_fs) {
253 char mountopts[] =
254 "cpuset,noprefix,"
255 "release_agent=/sbin/cpuset_release_agent";
256 ret = cgroup_fs->get_sb(cgroup_fs, flags,
257 unused_dev_name, mountopts, mnt);
258 put_filesystem(cgroup_fs);
260 return ret;
263 static struct file_system_type cpuset_fs_type = {
264 .name = "cpuset",
265 .get_sb = cpuset_get_sb,
269 * Return in *pmask the portion of a cpusets's cpus_allowed that
270 * are online. If none are online, walk up the cpuset hierarchy
271 * until we find one that does have some online cpus. If we get
272 * all the way to the top and still haven't found any online cpus,
273 * return cpu_online_map. Or if passed a NULL cs from an exit'ing
274 * task, return cpu_online_map.
276 * One way or another, we guarantee to return some non-empty subset
277 * of cpu_online_map.
279 * Call with callback_mutex held.
282 static void guarantee_online_cpus(const struct cpuset *cs, cpumask_t *pmask)
284 while (cs && !cpus_intersects(cs->cpus_allowed, cpu_online_map))
285 cs = cs->parent;
286 if (cs)
287 cpus_and(*pmask, cs->cpus_allowed, cpu_online_map);
288 else
289 *pmask = cpu_online_map;
290 BUG_ON(!cpus_intersects(*pmask, cpu_online_map));
294 * Return in *pmask the portion of a cpusets's mems_allowed that
295 * are online, with memory. If none are online with memory, walk
296 * up the cpuset hierarchy until we find one that does have some
297 * online mems. If we get all the way to the top and still haven't
298 * found any online mems, return node_states[N_HIGH_MEMORY].
300 * One way or another, we guarantee to return some non-empty subset
301 * of node_states[N_HIGH_MEMORY].
303 * Call with callback_mutex held.
306 static void guarantee_online_mems(const struct cpuset *cs, nodemask_t *pmask)
308 while (cs && !nodes_intersects(cs->mems_allowed,
309 node_states[N_HIGH_MEMORY]))
310 cs = cs->parent;
311 if (cs)
312 nodes_and(*pmask, cs->mems_allowed,
313 node_states[N_HIGH_MEMORY]);
314 else
315 *pmask = node_states[N_HIGH_MEMORY];
316 BUG_ON(!nodes_intersects(*pmask, node_states[N_HIGH_MEMORY]));
320 * cpuset_update_task_memory_state - update task memory placement
322 * If the current tasks cpusets mems_allowed changed behind our
323 * backs, update current->mems_allowed, mems_generation and task NUMA
324 * mempolicy to the new value.
326 * Task mempolicy is updated by rebinding it relative to the
327 * current->cpuset if a task has its memory placement changed.
328 * Do not call this routine if in_interrupt().
330 * Call without callback_mutex or task_lock() held. May be
331 * called with or without cgroup_mutex held. Thanks in part to
332 * 'the_top_cpuset_hack', the task's cpuset pointer will never
333 * be NULL. This routine also might acquire callback_mutex during
334 * call.
336 * Reading current->cpuset->mems_generation doesn't need task_lock
337 * to guard the current->cpuset derefence, because it is guarded
338 * from concurrent freeing of current->cpuset using RCU.
340 * The rcu_dereference() is technically probably not needed,
341 * as I don't actually mind if I see a new cpuset pointer but
342 * an old value of mems_generation. However this really only
343 * matters on alpha systems using cpusets heavily. If I dropped
344 * that rcu_dereference(), it would save them a memory barrier.
345 * For all other arch's, rcu_dereference is a no-op anyway, and for
346 * alpha systems not using cpusets, another planned optimization,
347 * avoiding the rcu critical section for tasks in the root cpuset
348 * which is statically allocated, so can't vanish, will make this
349 * irrelevant. Better to use RCU as intended, than to engage in
350 * some cute trick to save a memory barrier that is impossible to
351 * test, for alpha systems using cpusets heavily, which might not
352 * even exist.
354 * This routine is needed to update the per-task mems_allowed data,
355 * within the tasks context, when it is trying to allocate memory
356 * (in various mm/mempolicy.c routines) and notices that some other
357 * task has been modifying its cpuset.
360 void cpuset_update_task_memory_state(void)
362 int my_cpusets_mem_gen;
363 struct task_struct *tsk = current;
364 struct cpuset *cs;
366 if (task_cs(tsk) == &top_cpuset) {
367 /* Don't need rcu for top_cpuset. It's never freed. */
368 my_cpusets_mem_gen = top_cpuset.mems_generation;
369 } else {
370 rcu_read_lock();
371 my_cpusets_mem_gen = task_cs(tsk)->mems_generation;
372 rcu_read_unlock();
375 if (my_cpusets_mem_gen != tsk->cpuset_mems_generation) {
376 mutex_lock(&callback_mutex);
377 task_lock(tsk);
378 cs = task_cs(tsk); /* Maybe changed when task not locked */
379 guarantee_online_mems(cs, &tsk->mems_allowed);
380 tsk->cpuset_mems_generation = cs->mems_generation;
381 if (is_spread_page(cs))
382 tsk->flags |= PF_SPREAD_PAGE;
383 else
384 tsk->flags &= ~PF_SPREAD_PAGE;
385 if (is_spread_slab(cs))
386 tsk->flags |= PF_SPREAD_SLAB;
387 else
388 tsk->flags &= ~PF_SPREAD_SLAB;
389 task_unlock(tsk);
390 mutex_unlock(&callback_mutex);
391 mpol_rebind_task(tsk, &tsk->mems_allowed);
396 * is_cpuset_subset(p, q) - Is cpuset p a subset of cpuset q?
398 * One cpuset is a subset of another if all its allowed CPUs and
399 * Memory Nodes are a subset of the other, and its exclusive flags
400 * are only set if the other's are set. Call holding cgroup_mutex.
403 static int is_cpuset_subset(const struct cpuset *p, const struct cpuset *q)
405 return cpus_subset(p->cpus_allowed, q->cpus_allowed) &&
406 nodes_subset(p->mems_allowed, q->mems_allowed) &&
407 is_cpu_exclusive(p) <= is_cpu_exclusive(q) &&
408 is_mem_exclusive(p) <= is_mem_exclusive(q);
412 * validate_change() - Used to validate that any proposed cpuset change
413 * follows the structural rules for cpusets.
415 * If we replaced the flag and mask values of the current cpuset
416 * (cur) with those values in the trial cpuset (trial), would
417 * our various subset and exclusive rules still be valid? Presumes
418 * cgroup_mutex held.
420 * 'cur' is the address of an actual, in-use cpuset. Operations
421 * such as list traversal that depend on the actual address of the
422 * cpuset in the list must use cur below, not trial.
424 * 'trial' is the address of bulk structure copy of cur, with
425 * perhaps one or more of the fields cpus_allowed, mems_allowed,
426 * or flags changed to new, trial values.
428 * Return 0 if valid, -errno if not.
431 static int validate_change(const struct cpuset *cur, const struct cpuset *trial)
433 struct cgroup *cont;
434 struct cpuset *c, *par;
436 /* Each of our child cpusets must be a subset of us */
437 list_for_each_entry(cont, &cur->css.cgroup->children, sibling) {
438 if (!is_cpuset_subset(cgroup_cs(cont), trial))
439 return -EBUSY;
442 /* Remaining checks don't apply to root cpuset */
443 if (cur == &top_cpuset)
444 return 0;
446 par = cur->parent;
448 /* We must be a subset of our parent cpuset */
449 if (!is_cpuset_subset(trial, par))
450 return -EACCES;
453 * If either I or some sibling (!= me) is exclusive, we can't
454 * overlap
456 list_for_each_entry(cont, &par->css.cgroup->children, sibling) {
457 c = cgroup_cs(cont);
458 if ((is_cpu_exclusive(trial) || is_cpu_exclusive(c)) &&
459 c != cur &&
460 cpus_intersects(trial->cpus_allowed, c->cpus_allowed))
461 return -EINVAL;
462 if ((is_mem_exclusive(trial) || is_mem_exclusive(c)) &&
463 c != cur &&
464 nodes_intersects(trial->mems_allowed, c->mems_allowed))
465 return -EINVAL;
468 /* Cpusets with tasks can't have empty cpus_allowed or mems_allowed */
469 if (cgroup_task_count(cur->css.cgroup)) {
470 if (cpus_empty(trial->cpus_allowed) ||
471 nodes_empty(trial->mems_allowed)) {
472 return -ENOSPC;
476 return 0;
480 * Helper routine for generate_sched_domains().
481 * Do cpusets a, b have overlapping cpus_allowed masks?
483 static int cpusets_overlap(struct cpuset *a, struct cpuset *b)
485 return cpus_intersects(a->cpus_allowed, b->cpus_allowed);
488 static void
489 update_domain_attr(struct sched_domain_attr *dattr, struct cpuset *c)
491 if (dattr->relax_domain_level < c->relax_domain_level)
492 dattr->relax_domain_level = c->relax_domain_level;
493 return;
496 static void
497 update_domain_attr_tree(struct sched_domain_attr *dattr, struct cpuset *c)
499 LIST_HEAD(q);
501 list_add(&c->stack_list, &q);
502 while (!list_empty(&q)) {
503 struct cpuset *cp;
504 struct cgroup *cont;
505 struct cpuset *child;
507 cp = list_first_entry(&q, struct cpuset, stack_list);
508 list_del(q.next);
510 if (cpus_empty(cp->cpus_allowed))
511 continue;
513 if (is_sched_load_balance(cp))
514 update_domain_attr(dattr, cp);
516 list_for_each_entry(cont, &cp->css.cgroup->children, sibling) {
517 child = cgroup_cs(cont);
518 list_add_tail(&child->stack_list, &q);
524 * generate_sched_domains()
526 * This function builds a partial partition of the systems CPUs
527 * A 'partial partition' is a set of non-overlapping subsets whose
528 * union is a subset of that set.
529 * The output of this function needs to be passed to kernel/sched.c
530 * partition_sched_domains() routine, which will rebuild the scheduler's
531 * load balancing domains (sched domains) as specified by that partial
532 * partition.
534 * See "What is sched_load_balance" in Documentation/cpusets.txt
535 * for a background explanation of this.
537 * Does not return errors, on the theory that the callers of this
538 * routine would rather not worry about failures to rebuild sched
539 * domains when operating in the severe memory shortage situations
540 * that could cause allocation failures below.
542 * Must be called with cgroup_lock held.
544 * The three key local variables below are:
545 * q - a linked-list queue of cpuset pointers, used to implement a
546 * top-down scan of all cpusets. This scan loads a pointer
547 * to each cpuset marked is_sched_load_balance into the
548 * array 'csa'. For our purposes, rebuilding the schedulers
549 * sched domains, we can ignore !is_sched_load_balance cpusets.
550 * csa - (for CpuSet Array) Array of pointers to all the cpusets
551 * that need to be load balanced, for convenient iterative
552 * access by the subsequent code that finds the best partition,
553 * i.e the set of domains (subsets) of CPUs such that the
554 * cpus_allowed of every cpuset marked is_sched_load_balance
555 * is a subset of one of these domains, while there are as
556 * many such domains as possible, each as small as possible.
557 * doms - Conversion of 'csa' to an array of cpumasks, for passing to
558 * the kernel/sched.c routine partition_sched_domains() in a
559 * convenient format, that can be easily compared to the prior
560 * value to determine what partition elements (sched domains)
561 * were changed (added or removed.)
563 * Finding the best partition (set of domains):
564 * The triple nested loops below over i, j, k scan over the
565 * load balanced cpusets (using the array of cpuset pointers in
566 * csa[]) looking for pairs of cpusets that have overlapping
567 * cpus_allowed, but which don't have the same 'pn' partition
568 * number and gives them in the same partition number. It keeps
569 * looping on the 'restart' label until it can no longer find
570 * any such pairs.
572 * The union of the cpus_allowed masks from the set of
573 * all cpusets having the same 'pn' value then form the one
574 * element of the partition (one sched domain) to be passed to
575 * partition_sched_domains().
577 static int generate_sched_domains(cpumask_t **domains,
578 struct sched_domain_attr **attributes)
580 LIST_HEAD(q); /* queue of cpusets to be scanned */
581 struct cpuset *cp; /* scans q */
582 struct cpuset **csa; /* array of all cpuset ptrs */
583 int csn; /* how many cpuset ptrs in csa so far */
584 int i, j, k; /* indices for partition finding loops */
585 cpumask_t *doms; /* resulting partition; i.e. sched domains */
586 struct sched_domain_attr *dattr; /* attributes for custom domains */
587 int ndoms; /* number of sched domains in result */
588 int nslot; /* next empty doms[] cpumask_t slot */
590 ndoms = 0;
591 doms = NULL;
592 dattr = NULL;
593 csa = NULL;
595 /* Special case for the 99% of systems with one, full, sched domain */
596 if (is_sched_load_balance(&top_cpuset)) {
597 doms = kmalloc(sizeof(cpumask_t), GFP_KERNEL);
598 if (!doms)
599 goto done;
601 dattr = kmalloc(sizeof(struct sched_domain_attr), GFP_KERNEL);
602 if (dattr) {
603 *dattr = SD_ATTR_INIT;
604 update_domain_attr_tree(dattr, &top_cpuset);
606 *doms = top_cpuset.cpus_allowed;
608 ndoms = 1;
609 goto done;
612 csa = kmalloc(number_of_cpusets * sizeof(cp), GFP_KERNEL);
613 if (!csa)
614 goto done;
615 csn = 0;
617 list_add(&top_cpuset.stack_list, &q);
618 while (!list_empty(&q)) {
619 struct cgroup *cont;
620 struct cpuset *child; /* scans child cpusets of cp */
622 cp = list_first_entry(&q, struct cpuset, stack_list);
623 list_del(q.next);
625 if (cpus_empty(cp->cpus_allowed))
626 continue;
629 * All child cpusets contain a subset of the parent's cpus, so
630 * just skip them, and then we call update_domain_attr_tree()
631 * to calc relax_domain_level of the corresponding sched
632 * domain.
634 if (is_sched_load_balance(cp)) {
635 csa[csn++] = cp;
636 continue;
639 list_for_each_entry(cont, &cp->css.cgroup->children, sibling) {
640 child = cgroup_cs(cont);
641 list_add_tail(&child->stack_list, &q);
645 for (i = 0; i < csn; i++)
646 csa[i]->pn = i;
647 ndoms = csn;
649 restart:
650 /* Find the best partition (set of sched domains) */
651 for (i = 0; i < csn; i++) {
652 struct cpuset *a = csa[i];
653 int apn = a->pn;
655 for (j = 0; j < csn; j++) {
656 struct cpuset *b = csa[j];
657 int bpn = b->pn;
659 if (apn != bpn && cpusets_overlap(a, b)) {
660 for (k = 0; k < csn; k++) {
661 struct cpuset *c = csa[k];
663 if (c->pn == bpn)
664 c->pn = apn;
666 ndoms--; /* one less element */
667 goto restart;
673 * Now we know how many domains to create.
674 * Convert <csn, csa> to <ndoms, doms> and populate cpu masks.
676 doms = kmalloc(ndoms * sizeof(cpumask_t), GFP_KERNEL);
677 if (!doms) {
678 ndoms = 0;
679 goto done;
683 * The rest of the code, including the scheduler, can deal with
684 * dattr==NULL case. No need to abort if alloc fails.
686 dattr = kmalloc(ndoms * sizeof(struct sched_domain_attr), GFP_KERNEL);
688 for (nslot = 0, i = 0; i < csn; i++) {
689 struct cpuset *a = csa[i];
690 cpumask_t *dp;
691 int apn = a->pn;
693 if (apn < 0) {
694 /* Skip completed partitions */
695 continue;
698 dp = doms + nslot;
700 if (nslot == ndoms) {
701 static int warnings = 10;
702 if (warnings) {
703 printk(KERN_WARNING
704 "rebuild_sched_domains confused:"
705 " nslot %d, ndoms %d, csn %d, i %d,"
706 " apn %d\n",
707 nslot, ndoms, csn, i, apn);
708 warnings--;
710 continue;
713 cpus_clear(*dp);
714 if (dattr)
715 *(dattr + nslot) = SD_ATTR_INIT;
716 for (j = i; j < csn; j++) {
717 struct cpuset *b = csa[j];
719 if (apn == b->pn) {
720 cpus_or(*dp, *dp, b->cpus_allowed);
721 if (dattr)
722 update_domain_attr_tree(dattr + nslot, b);
724 /* Done with this partition */
725 b->pn = -1;
728 nslot++;
730 BUG_ON(nslot != ndoms);
732 done:
733 kfree(csa);
735 *domains = doms;
736 *attributes = dattr;
737 return ndoms;
741 * Rebuild scheduler domains.
743 * Call with neither cgroup_mutex held nor within get_online_cpus().
744 * Takes both cgroup_mutex and get_online_cpus().
746 * Cannot be directly called from cpuset code handling changes
747 * to the cpuset pseudo-filesystem, because it cannot be called
748 * from code that already holds cgroup_mutex.
750 static void do_rebuild_sched_domains(struct work_struct *unused)
752 struct sched_domain_attr *attr;
753 cpumask_t *doms;
754 int ndoms;
756 get_online_cpus();
758 /* Generate domain masks and attrs */
759 cgroup_lock();
760 ndoms = generate_sched_domains(&doms, &attr);
761 cgroup_unlock();
763 /* Have scheduler rebuild the domains */
764 partition_sched_domains(ndoms, doms, attr);
766 put_online_cpus();
769 static DECLARE_WORK(rebuild_sched_domains_work, do_rebuild_sched_domains);
772 * Rebuild scheduler domains, asynchronously via workqueue.
774 * If the flag 'sched_load_balance' of any cpuset with non-empty
775 * 'cpus' changes, or if the 'cpus' allowed changes in any cpuset
776 * which has that flag enabled, or if any cpuset with a non-empty
777 * 'cpus' is removed, then call this routine to rebuild the
778 * scheduler's dynamic sched domains.
780 * The rebuild_sched_domains() and partition_sched_domains()
781 * routines must nest cgroup_lock() inside get_online_cpus(),
782 * but such cpuset changes as these must nest that locking the
783 * other way, holding cgroup_lock() for much of the code.
785 * So in order to avoid an ABBA deadlock, the cpuset code handling
786 * these user changes delegates the actual sched domain rebuilding
787 * to a separate workqueue thread, which ends up processing the
788 * above do_rebuild_sched_domains() function.
790 static void async_rebuild_sched_domains(void)
792 schedule_work(&rebuild_sched_domains_work);
796 * Accomplishes the same scheduler domain rebuild as the above
797 * async_rebuild_sched_domains(), however it directly calls the
798 * rebuild routine synchronously rather than calling it via an
799 * asynchronous work thread.
801 * This can only be called from code that is not holding
802 * cgroup_mutex (not nested in a cgroup_lock() call.)
804 void rebuild_sched_domains(void)
806 do_rebuild_sched_domains(NULL);
810 * cpuset_test_cpumask - test a task's cpus_allowed versus its cpuset's
811 * @tsk: task to test
812 * @scan: struct cgroup_scanner contained in its struct cpuset_hotplug_scanner
814 * Call with cgroup_mutex held. May take callback_mutex during call.
815 * Called for each task in a cgroup by cgroup_scan_tasks().
816 * Return nonzero if this tasks's cpus_allowed mask should be changed (in other
817 * words, if its mask is not equal to its cpuset's mask).
819 static int cpuset_test_cpumask(struct task_struct *tsk,
820 struct cgroup_scanner *scan)
822 return !cpus_equal(tsk->cpus_allowed,
823 (cgroup_cs(scan->cg))->cpus_allowed);
827 * cpuset_change_cpumask - make a task's cpus_allowed the same as its cpuset's
828 * @tsk: task to test
829 * @scan: struct cgroup_scanner containing the cgroup of the task
831 * Called by cgroup_scan_tasks() for each task in a cgroup whose
832 * cpus_allowed mask needs to be changed.
834 * We don't need to re-check for the cgroup/cpuset membership, since we're
835 * holding cgroup_lock() at this point.
837 static void cpuset_change_cpumask(struct task_struct *tsk,
838 struct cgroup_scanner *scan)
840 set_cpus_allowed_ptr(tsk, &((cgroup_cs(scan->cg))->cpus_allowed));
844 * update_tasks_cpumask - Update the cpumasks of tasks in the cpuset.
845 * @cs: the cpuset in which each task's cpus_allowed mask needs to be changed
846 * @heap: if NULL, defer allocating heap memory to cgroup_scan_tasks()
848 * Called with cgroup_mutex held
850 * The cgroup_scan_tasks() function will scan all the tasks in a cgroup,
851 * calling callback functions for each.
853 * No return value. It's guaranteed that cgroup_scan_tasks() always returns 0
854 * if @heap != NULL.
856 static void update_tasks_cpumask(struct cpuset *cs, struct ptr_heap *heap)
858 struct cgroup_scanner scan;
860 scan.cg = cs->css.cgroup;
861 scan.test_task = cpuset_test_cpumask;
862 scan.process_task = cpuset_change_cpumask;
863 scan.heap = heap;
864 cgroup_scan_tasks(&scan);
868 * update_cpumask - update the cpus_allowed mask of a cpuset and all tasks in it
869 * @cs: the cpuset to consider
870 * @buf: buffer of cpu numbers written to this cpuset
872 static int update_cpumask(struct cpuset *cs, const char *buf)
874 struct ptr_heap heap;
875 struct cpuset trialcs;
876 int retval;
877 int is_load_balanced;
879 /* top_cpuset.cpus_allowed tracks cpu_online_map; it's read-only */
880 if (cs == &top_cpuset)
881 return -EACCES;
883 trialcs = *cs;
886 * An empty cpus_allowed is ok only if the cpuset has no tasks.
887 * Since cpulist_parse() fails on an empty mask, we special case
888 * that parsing. The validate_change() call ensures that cpusets
889 * with tasks have cpus.
891 if (!*buf) {
892 cpus_clear(trialcs.cpus_allowed);
893 } else {
894 retval = cpulist_parse(buf, trialcs.cpus_allowed);
895 if (retval < 0)
896 return retval;
898 if (!cpus_subset(trialcs.cpus_allowed, cpu_online_map))
899 return -EINVAL;
901 retval = validate_change(cs, &trialcs);
902 if (retval < 0)
903 return retval;
905 /* Nothing to do if the cpus didn't change */
906 if (cpus_equal(cs->cpus_allowed, trialcs.cpus_allowed))
907 return 0;
909 retval = heap_init(&heap, PAGE_SIZE, GFP_KERNEL, NULL);
910 if (retval)
911 return retval;
913 is_load_balanced = is_sched_load_balance(&trialcs);
915 mutex_lock(&callback_mutex);
916 cs->cpus_allowed = trialcs.cpus_allowed;
917 mutex_unlock(&callback_mutex);
920 * Scan tasks in the cpuset, and update the cpumasks of any
921 * that need an update.
923 update_tasks_cpumask(cs, &heap);
925 heap_free(&heap);
927 if (is_load_balanced)
928 async_rebuild_sched_domains();
929 return 0;
933 * cpuset_migrate_mm
935 * Migrate memory region from one set of nodes to another.
937 * Temporarilly set tasks mems_allowed to target nodes of migration,
938 * so that the migration code can allocate pages on these nodes.
940 * Call holding cgroup_mutex, so current's cpuset won't change
941 * during this call, as manage_mutex holds off any cpuset_attach()
942 * calls. Therefore we don't need to take task_lock around the
943 * call to guarantee_online_mems(), as we know no one is changing
944 * our task's cpuset.
946 * Hold callback_mutex around the two modifications of our tasks
947 * mems_allowed to synchronize with cpuset_mems_allowed().
949 * While the mm_struct we are migrating is typically from some
950 * other task, the task_struct mems_allowed that we are hacking
951 * is for our current task, which must allocate new pages for that
952 * migrating memory region.
954 * We call cpuset_update_task_memory_state() before hacking
955 * our tasks mems_allowed, so that we are assured of being in
956 * sync with our tasks cpuset, and in particular, callbacks to
957 * cpuset_update_task_memory_state() from nested page allocations
958 * won't see any mismatch of our cpuset and task mems_generation
959 * values, so won't overwrite our hacked tasks mems_allowed
960 * nodemask.
963 static void cpuset_migrate_mm(struct mm_struct *mm, const nodemask_t *from,
964 const nodemask_t *to)
966 struct task_struct *tsk = current;
968 cpuset_update_task_memory_state();
970 mutex_lock(&callback_mutex);
971 tsk->mems_allowed = *to;
972 mutex_unlock(&callback_mutex);
974 do_migrate_pages(mm, from, to, MPOL_MF_MOVE_ALL);
976 mutex_lock(&callback_mutex);
977 guarantee_online_mems(task_cs(tsk),&tsk->mems_allowed);
978 mutex_unlock(&callback_mutex);
981 static void *cpuset_being_rebound;
984 * update_tasks_nodemask - Update the nodemasks of tasks in the cpuset.
985 * @cs: the cpuset in which each task's mems_allowed mask needs to be changed
986 * @oldmem: old mems_allowed of cpuset cs
988 * Called with cgroup_mutex held
989 * Return 0 if successful, -errno if not.
991 static int update_tasks_nodemask(struct cpuset *cs, const nodemask_t *oldmem)
993 struct task_struct *p;
994 struct mm_struct **mmarray;
995 int i, n, ntasks;
996 int migrate;
997 int fudge;
998 struct cgroup_iter it;
999 int retval;
1001 cpuset_being_rebound = cs; /* causes mpol_dup() rebind */
1003 fudge = 10; /* spare mmarray[] slots */
1004 fudge += cpus_weight(cs->cpus_allowed); /* imagine one fork-bomb/cpu */
1005 retval = -ENOMEM;
1008 * Allocate mmarray[] to hold mm reference for each task
1009 * in cpuset cs. Can't kmalloc GFP_KERNEL while holding
1010 * tasklist_lock. We could use GFP_ATOMIC, but with a
1011 * few more lines of code, we can retry until we get a big
1012 * enough mmarray[] w/o using GFP_ATOMIC.
1014 while (1) {
1015 ntasks = cgroup_task_count(cs->css.cgroup); /* guess */
1016 ntasks += fudge;
1017 mmarray = kmalloc(ntasks * sizeof(*mmarray), GFP_KERNEL);
1018 if (!mmarray)
1019 goto done;
1020 read_lock(&tasklist_lock); /* block fork */
1021 if (cgroup_task_count(cs->css.cgroup) <= ntasks)
1022 break; /* got enough */
1023 read_unlock(&tasklist_lock); /* try again */
1024 kfree(mmarray);
1027 n = 0;
1029 /* Load up mmarray[] with mm reference for each task in cpuset. */
1030 cgroup_iter_start(cs->css.cgroup, &it);
1031 while ((p = cgroup_iter_next(cs->css.cgroup, &it))) {
1032 struct mm_struct *mm;
1034 if (n >= ntasks) {
1035 printk(KERN_WARNING
1036 "Cpuset mempolicy rebind incomplete.\n");
1037 break;
1039 mm = get_task_mm(p);
1040 if (!mm)
1041 continue;
1042 mmarray[n++] = mm;
1044 cgroup_iter_end(cs->css.cgroup, &it);
1045 read_unlock(&tasklist_lock);
1048 * Now that we've dropped the tasklist spinlock, we can
1049 * rebind the vma mempolicies of each mm in mmarray[] to their
1050 * new cpuset, and release that mm. The mpol_rebind_mm()
1051 * call takes mmap_sem, which we couldn't take while holding
1052 * tasklist_lock. Forks can happen again now - the mpol_dup()
1053 * cpuset_being_rebound check will catch such forks, and rebind
1054 * their vma mempolicies too. Because we still hold the global
1055 * cgroup_mutex, we know that no other rebind effort will
1056 * be contending for the global variable cpuset_being_rebound.
1057 * It's ok if we rebind the same mm twice; mpol_rebind_mm()
1058 * is idempotent. Also migrate pages in each mm to new nodes.
1060 migrate = is_memory_migrate(cs);
1061 for (i = 0; i < n; i++) {
1062 struct mm_struct *mm = mmarray[i];
1064 mpol_rebind_mm(mm, &cs->mems_allowed);
1065 if (migrate)
1066 cpuset_migrate_mm(mm, oldmem, &cs->mems_allowed);
1067 mmput(mm);
1070 /* We're done rebinding vmas to this cpuset's new mems_allowed. */
1071 kfree(mmarray);
1072 cpuset_being_rebound = NULL;
1073 retval = 0;
1074 done:
1075 return retval;
1079 * Handle user request to change the 'mems' memory placement
1080 * of a cpuset. Needs to validate the request, update the
1081 * cpusets mems_allowed and mems_generation, and for each
1082 * task in the cpuset, rebind any vma mempolicies and if
1083 * the cpuset is marked 'memory_migrate', migrate the tasks
1084 * pages to the new memory.
1086 * Call with cgroup_mutex held. May take callback_mutex during call.
1087 * Will take tasklist_lock, scan tasklist for tasks in cpuset cs,
1088 * lock each such tasks mm->mmap_sem, scan its vma's and rebind
1089 * their mempolicies to the cpusets new mems_allowed.
1091 static int update_nodemask(struct cpuset *cs, const char *buf)
1093 struct cpuset trialcs;
1094 nodemask_t oldmem;
1095 int retval;
1098 * top_cpuset.mems_allowed tracks node_stats[N_HIGH_MEMORY];
1099 * it's read-only
1101 if (cs == &top_cpuset)
1102 return -EACCES;
1104 trialcs = *cs;
1107 * An empty mems_allowed is ok iff there are no tasks in the cpuset.
1108 * Since nodelist_parse() fails on an empty mask, we special case
1109 * that parsing. The validate_change() call ensures that cpusets
1110 * with tasks have memory.
1112 if (!*buf) {
1113 nodes_clear(trialcs.mems_allowed);
1114 } else {
1115 retval = nodelist_parse(buf, trialcs.mems_allowed);
1116 if (retval < 0)
1117 goto done;
1119 if (!nodes_subset(trialcs.mems_allowed,
1120 node_states[N_HIGH_MEMORY]))
1121 return -EINVAL;
1123 oldmem = cs->mems_allowed;
1124 if (nodes_equal(oldmem, trialcs.mems_allowed)) {
1125 retval = 0; /* Too easy - nothing to do */
1126 goto done;
1128 retval = validate_change(cs, &trialcs);
1129 if (retval < 0)
1130 goto done;
1132 mutex_lock(&callback_mutex);
1133 cs->mems_allowed = trialcs.mems_allowed;
1134 cs->mems_generation = cpuset_mems_generation++;
1135 mutex_unlock(&callback_mutex);
1137 retval = update_tasks_nodemask(cs, &oldmem);
1138 done:
1139 return retval;
1142 int current_cpuset_is_being_rebound(void)
1144 return task_cs(current) == cpuset_being_rebound;
1147 static int update_relax_domain_level(struct cpuset *cs, s64 val)
1149 if (val < -1 || val >= SD_LV_MAX)
1150 return -EINVAL;
1152 if (val != cs->relax_domain_level) {
1153 cs->relax_domain_level = val;
1154 if (!cpus_empty(cs->cpus_allowed) && is_sched_load_balance(cs))
1155 async_rebuild_sched_domains();
1158 return 0;
1162 * update_flag - read a 0 or a 1 in a file and update associated flag
1163 * bit: the bit to update (see cpuset_flagbits_t)
1164 * cs: the cpuset to update
1165 * turning_on: whether the flag is being set or cleared
1167 * Call with cgroup_mutex held.
1170 static int update_flag(cpuset_flagbits_t bit, struct cpuset *cs,
1171 int turning_on)
1173 struct cpuset trialcs;
1174 int err;
1175 int balance_flag_changed;
1177 trialcs = *cs;
1178 if (turning_on)
1179 set_bit(bit, &trialcs.flags);
1180 else
1181 clear_bit(bit, &trialcs.flags);
1183 err = validate_change(cs, &trialcs);
1184 if (err < 0)
1185 return err;
1187 balance_flag_changed = (is_sched_load_balance(cs) !=
1188 is_sched_load_balance(&trialcs));
1190 mutex_lock(&callback_mutex);
1191 cs->flags = trialcs.flags;
1192 mutex_unlock(&callback_mutex);
1194 if (!cpus_empty(trialcs.cpus_allowed) && balance_flag_changed)
1195 async_rebuild_sched_domains();
1197 return 0;
1201 * Frequency meter - How fast is some event occurring?
1203 * These routines manage a digitally filtered, constant time based,
1204 * event frequency meter. There are four routines:
1205 * fmeter_init() - initialize a frequency meter.
1206 * fmeter_markevent() - called each time the event happens.
1207 * fmeter_getrate() - returns the recent rate of such events.
1208 * fmeter_update() - internal routine used to update fmeter.
1210 * A common data structure is passed to each of these routines,
1211 * which is used to keep track of the state required to manage the
1212 * frequency meter and its digital filter.
1214 * The filter works on the number of events marked per unit time.
1215 * The filter is single-pole low-pass recursive (IIR). The time unit
1216 * is 1 second. Arithmetic is done using 32-bit integers scaled to
1217 * simulate 3 decimal digits of precision (multiplied by 1000).
1219 * With an FM_COEF of 933, and a time base of 1 second, the filter
1220 * has a half-life of 10 seconds, meaning that if the events quit
1221 * happening, then the rate returned from the fmeter_getrate()
1222 * will be cut in half each 10 seconds, until it converges to zero.
1224 * It is not worth doing a real infinitely recursive filter. If more
1225 * than FM_MAXTICKS ticks have elapsed since the last filter event,
1226 * just compute FM_MAXTICKS ticks worth, by which point the level
1227 * will be stable.
1229 * Limit the count of unprocessed events to FM_MAXCNT, so as to avoid
1230 * arithmetic overflow in the fmeter_update() routine.
1232 * Given the simple 32 bit integer arithmetic used, this meter works
1233 * best for reporting rates between one per millisecond (msec) and
1234 * one per 32 (approx) seconds. At constant rates faster than one
1235 * per msec it maxes out at values just under 1,000,000. At constant
1236 * rates between one per msec, and one per second it will stabilize
1237 * to a value N*1000, where N is the rate of events per second.
1238 * At constant rates between one per second and one per 32 seconds,
1239 * it will be choppy, moving up on the seconds that have an event,
1240 * and then decaying until the next event. At rates slower than
1241 * about one in 32 seconds, it decays all the way back to zero between
1242 * each event.
1245 #define FM_COEF 933 /* coefficient for half-life of 10 secs */
1246 #define FM_MAXTICKS ((time_t)99) /* useless computing more ticks than this */
1247 #define FM_MAXCNT 1000000 /* limit cnt to avoid overflow */
1248 #define FM_SCALE 1000 /* faux fixed point scale */
1250 /* Initialize a frequency meter */
1251 static void fmeter_init(struct fmeter *fmp)
1253 fmp->cnt = 0;
1254 fmp->val = 0;
1255 fmp->time = 0;
1256 spin_lock_init(&fmp->lock);
1259 /* Internal meter update - process cnt events and update value */
1260 static void fmeter_update(struct fmeter *fmp)
1262 time_t now = get_seconds();
1263 time_t ticks = now - fmp->time;
1265 if (ticks == 0)
1266 return;
1268 ticks = min(FM_MAXTICKS, ticks);
1269 while (ticks-- > 0)
1270 fmp->val = (FM_COEF * fmp->val) / FM_SCALE;
1271 fmp->time = now;
1273 fmp->val += ((FM_SCALE - FM_COEF) * fmp->cnt) / FM_SCALE;
1274 fmp->cnt = 0;
1277 /* Process any previous ticks, then bump cnt by one (times scale). */
1278 static void fmeter_markevent(struct fmeter *fmp)
1280 spin_lock(&fmp->lock);
1281 fmeter_update(fmp);
1282 fmp->cnt = min(FM_MAXCNT, fmp->cnt + FM_SCALE);
1283 spin_unlock(&fmp->lock);
1286 /* Process any previous ticks, then return current value. */
1287 static int fmeter_getrate(struct fmeter *fmp)
1289 int val;
1291 spin_lock(&fmp->lock);
1292 fmeter_update(fmp);
1293 val = fmp->val;
1294 spin_unlock(&fmp->lock);
1295 return val;
1298 /* Called by cgroups to determine if a cpuset is usable; cgroup_mutex held */
1299 static int cpuset_can_attach(struct cgroup_subsys *ss,
1300 struct cgroup *cont, struct task_struct *tsk)
1302 struct cpuset *cs = cgroup_cs(cont);
1304 if (cpus_empty(cs->cpus_allowed) || nodes_empty(cs->mems_allowed))
1305 return -ENOSPC;
1306 if (tsk->flags & PF_THREAD_BOUND) {
1307 cpumask_t mask;
1309 mutex_lock(&callback_mutex);
1310 mask = cs->cpus_allowed;
1311 mutex_unlock(&callback_mutex);
1312 if (!cpus_equal(tsk->cpus_allowed, mask))
1313 return -EINVAL;
1316 return security_task_setscheduler(tsk, 0, NULL);
1319 static void cpuset_attach(struct cgroup_subsys *ss,
1320 struct cgroup *cont, struct cgroup *oldcont,
1321 struct task_struct *tsk)
1323 cpumask_t cpus;
1324 nodemask_t from, to;
1325 struct mm_struct *mm;
1326 struct cpuset *cs = cgroup_cs(cont);
1327 struct cpuset *oldcs = cgroup_cs(oldcont);
1328 int err;
1330 mutex_lock(&callback_mutex);
1331 guarantee_online_cpus(cs, &cpus);
1332 err = set_cpus_allowed_ptr(tsk, &cpus);
1333 mutex_unlock(&callback_mutex);
1334 if (err)
1335 return;
1337 from = oldcs->mems_allowed;
1338 to = cs->mems_allowed;
1339 mm = get_task_mm(tsk);
1340 if (mm) {
1341 mpol_rebind_mm(mm, &to);
1342 if (is_memory_migrate(cs))
1343 cpuset_migrate_mm(mm, &from, &to);
1344 mmput(mm);
1349 /* The various types of files and directories in a cpuset file system */
1351 typedef enum {
1352 FILE_MEMORY_MIGRATE,
1353 FILE_CPULIST,
1354 FILE_MEMLIST,
1355 FILE_CPU_EXCLUSIVE,
1356 FILE_MEM_EXCLUSIVE,
1357 FILE_MEM_HARDWALL,
1358 FILE_SCHED_LOAD_BALANCE,
1359 FILE_SCHED_RELAX_DOMAIN_LEVEL,
1360 FILE_MEMORY_PRESSURE_ENABLED,
1361 FILE_MEMORY_PRESSURE,
1362 FILE_SPREAD_PAGE,
1363 FILE_SPREAD_SLAB,
1364 } cpuset_filetype_t;
1366 static int cpuset_write_u64(struct cgroup *cgrp, struct cftype *cft, u64 val)
1368 int retval = 0;
1369 struct cpuset *cs = cgroup_cs(cgrp);
1370 cpuset_filetype_t type = cft->private;
1372 if (!cgroup_lock_live_group(cgrp))
1373 return -ENODEV;
1375 switch (type) {
1376 case FILE_CPU_EXCLUSIVE:
1377 retval = update_flag(CS_CPU_EXCLUSIVE, cs, val);
1378 break;
1379 case FILE_MEM_EXCLUSIVE:
1380 retval = update_flag(CS_MEM_EXCLUSIVE, cs, val);
1381 break;
1382 case FILE_MEM_HARDWALL:
1383 retval = update_flag(CS_MEM_HARDWALL, cs, val);
1384 break;
1385 case FILE_SCHED_LOAD_BALANCE:
1386 retval = update_flag(CS_SCHED_LOAD_BALANCE, cs, val);
1387 break;
1388 case FILE_MEMORY_MIGRATE:
1389 retval = update_flag(CS_MEMORY_MIGRATE, cs, val);
1390 break;
1391 case FILE_MEMORY_PRESSURE_ENABLED:
1392 cpuset_memory_pressure_enabled = !!val;
1393 break;
1394 case FILE_MEMORY_PRESSURE:
1395 retval = -EACCES;
1396 break;
1397 case FILE_SPREAD_PAGE:
1398 retval = update_flag(CS_SPREAD_PAGE, cs, val);
1399 cs->mems_generation = cpuset_mems_generation++;
1400 break;
1401 case FILE_SPREAD_SLAB:
1402 retval = update_flag(CS_SPREAD_SLAB, cs, val);
1403 cs->mems_generation = cpuset_mems_generation++;
1404 break;
1405 default:
1406 retval = -EINVAL;
1407 break;
1409 cgroup_unlock();
1410 return retval;
1413 static int cpuset_write_s64(struct cgroup *cgrp, struct cftype *cft, s64 val)
1415 int retval = 0;
1416 struct cpuset *cs = cgroup_cs(cgrp);
1417 cpuset_filetype_t type = cft->private;
1419 if (!cgroup_lock_live_group(cgrp))
1420 return -ENODEV;
1422 switch (type) {
1423 case FILE_SCHED_RELAX_DOMAIN_LEVEL:
1424 retval = update_relax_domain_level(cs, val);
1425 break;
1426 default:
1427 retval = -EINVAL;
1428 break;
1430 cgroup_unlock();
1431 return retval;
1435 * Common handling for a write to a "cpus" or "mems" file.
1437 static int cpuset_write_resmask(struct cgroup *cgrp, struct cftype *cft,
1438 const char *buf)
1440 int retval = 0;
1442 if (!cgroup_lock_live_group(cgrp))
1443 return -ENODEV;
1445 switch (cft->private) {
1446 case FILE_CPULIST:
1447 retval = update_cpumask(cgroup_cs(cgrp), buf);
1448 break;
1449 case FILE_MEMLIST:
1450 retval = update_nodemask(cgroup_cs(cgrp), buf);
1451 break;
1452 default:
1453 retval = -EINVAL;
1454 break;
1456 cgroup_unlock();
1457 return retval;
1461 * These ascii lists should be read in a single call, by using a user
1462 * buffer large enough to hold the entire map. If read in smaller
1463 * chunks, there is no guarantee of atomicity. Since the display format
1464 * used, list of ranges of sequential numbers, is variable length,
1465 * and since these maps can change value dynamically, one could read
1466 * gibberish by doing partial reads while a list was changing.
1467 * A single large read to a buffer that crosses a page boundary is
1468 * ok, because the result being copied to user land is not recomputed
1469 * across a page fault.
1472 static int cpuset_sprintf_cpulist(char *page, struct cpuset *cs)
1474 cpumask_t mask;
1476 mutex_lock(&callback_mutex);
1477 mask = cs->cpus_allowed;
1478 mutex_unlock(&callback_mutex);
1480 return cpulist_scnprintf(page, PAGE_SIZE, mask);
1483 static int cpuset_sprintf_memlist(char *page, struct cpuset *cs)
1485 nodemask_t mask;
1487 mutex_lock(&callback_mutex);
1488 mask = cs->mems_allowed;
1489 mutex_unlock(&callback_mutex);
1491 return nodelist_scnprintf(page, PAGE_SIZE, mask);
1494 static ssize_t cpuset_common_file_read(struct cgroup *cont,
1495 struct cftype *cft,
1496 struct file *file,
1497 char __user *buf,
1498 size_t nbytes, loff_t *ppos)
1500 struct cpuset *cs = cgroup_cs(cont);
1501 cpuset_filetype_t type = cft->private;
1502 char *page;
1503 ssize_t retval = 0;
1504 char *s;
1506 if (!(page = (char *)__get_free_page(GFP_TEMPORARY)))
1507 return -ENOMEM;
1509 s = page;
1511 switch (type) {
1512 case FILE_CPULIST:
1513 s += cpuset_sprintf_cpulist(s, cs);
1514 break;
1515 case FILE_MEMLIST:
1516 s += cpuset_sprintf_memlist(s, cs);
1517 break;
1518 default:
1519 retval = -EINVAL;
1520 goto out;
1522 *s++ = '\n';
1524 retval = simple_read_from_buffer(buf, nbytes, ppos, page, s - page);
1525 out:
1526 free_page((unsigned long)page);
1527 return retval;
1530 static u64 cpuset_read_u64(struct cgroup *cont, struct cftype *cft)
1532 struct cpuset *cs = cgroup_cs(cont);
1533 cpuset_filetype_t type = cft->private;
1534 switch (type) {
1535 case FILE_CPU_EXCLUSIVE:
1536 return is_cpu_exclusive(cs);
1537 case FILE_MEM_EXCLUSIVE:
1538 return is_mem_exclusive(cs);
1539 case FILE_MEM_HARDWALL:
1540 return is_mem_hardwall(cs);
1541 case FILE_SCHED_LOAD_BALANCE:
1542 return is_sched_load_balance(cs);
1543 case FILE_MEMORY_MIGRATE:
1544 return is_memory_migrate(cs);
1545 case FILE_MEMORY_PRESSURE_ENABLED:
1546 return cpuset_memory_pressure_enabled;
1547 case FILE_MEMORY_PRESSURE:
1548 return fmeter_getrate(&cs->fmeter);
1549 case FILE_SPREAD_PAGE:
1550 return is_spread_page(cs);
1551 case FILE_SPREAD_SLAB:
1552 return is_spread_slab(cs);
1553 default:
1554 BUG();
1557 /* Unreachable but makes gcc happy */
1558 return 0;
1561 static s64 cpuset_read_s64(struct cgroup *cont, struct cftype *cft)
1563 struct cpuset *cs = cgroup_cs(cont);
1564 cpuset_filetype_t type = cft->private;
1565 switch (type) {
1566 case FILE_SCHED_RELAX_DOMAIN_LEVEL:
1567 return cs->relax_domain_level;
1568 default:
1569 BUG();
1572 /* Unrechable but makes gcc happy */
1573 return 0;
1578 * for the common functions, 'private' gives the type of file
1581 static struct cftype files[] = {
1583 .name = "cpus",
1584 .read = cpuset_common_file_read,
1585 .write_string = cpuset_write_resmask,
1586 .max_write_len = (100U + 6 * NR_CPUS),
1587 .private = FILE_CPULIST,
1591 .name = "mems",
1592 .read = cpuset_common_file_read,
1593 .write_string = cpuset_write_resmask,
1594 .max_write_len = (100U + 6 * MAX_NUMNODES),
1595 .private = FILE_MEMLIST,
1599 .name = "cpu_exclusive",
1600 .read_u64 = cpuset_read_u64,
1601 .write_u64 = cpuset_write_u64,
1602 .private = FILE_CPU_EXCLUSIVE,
1606 .name = "mem_exclusive",
1607 .read_u64 = cpuset_read_u64,
1608 .write_u64 = cpuset_write_u64,
1609 .private = FILE_MEM_EXCLUSIVE,
1613 .name = "mem_hardwall",
1614 .read_u64 = cpuset_read_u64,
1615 .write_u64 = cpuset_write_u64,
1616 .private = FILE_MEM_HARDWALL,
1620 .name = "sched_load_balance",
1621 .read_u64 = cpuset_read_u64,
1622 .write_u64 = cpuset_write_u64,
1623 .private = FILE_SCHED_LOAD_BALANCE,
1627 .name = "sched_relax_domain_level",
1628 .read_s64 = cpuset_read_s64,
1629 .write_s64 = cpuset_write_s64,
1630 .private = FILE_SCHED_RELAX_DOMAIN_LEVEL,
1634 .name = "memory_migrate",
1635 .read_u64 = cpuset_read_u64,
1636 .write_u64 = cpuset_write_u64,
1637 .private = FILE_MEMORY_MIGRATE,
1641 .name = "memory_pressure",
1642 .read_u64 = cpuset_read_u64,
1643 .write_u64 = cpuset_write_u64,
1644 .private = FILE_MEMORY_PRESSURE,
1648 .name = "memory_spread_page",
1649 .read_u64 = cpuset_read_u64,
1650 .write_u64 = cpuset_write_u64,
1651 .private = FILE_SPREAD_PAGE,
1655 .name = "memory_spread_slab",
1656 .read_u64 = cpuset_read_u64,
1657 .write_u64 = cpuset_write_u64,
1658 .private = FILE_SPREAD_SLAB,
1662 static struct cftype cft_memory_pressure_enabled = {
1663 .name = "memory_pressure_enabled",
1664 .read_u64 = cpuset_read_u64,
1665 .write_u64 = cpuset_write_u64,
1666 .private = FILE_MEMORY_PRESSURE_ENABLED,
1669 static int cpuset_populate(struct cgroup_subsys *ss, struct cgroup *cont)
1671 int err;
1673 err = cgroup_add_files(cont, ss, files, ARRAY_SIZE(files));
1674 if (err)
1675 return err;
1676 /* memory_pressure_enabled is in root cpuset only */
1677 if (!cont->parent)
1678 err = cgroup_add_file(cont, ss,
1679 &cft_memory_pressure_enabled);
1680 return err;
1684 * post_clone() is called at the end of cgroup_clone().
1685 * 'cgroup' was just created automatically as a result of
1686 * a cgroup_clone(), and the current task is about to
1687 * be moved into 'cgroup'.
1689 * Currently we refuse to set up the cgroup - thereby
1690 * refusing the task to be entered, and as a result refusing
1691 * the sys_unshare() or clone() which initiated it - if any
1692 * sibling cpusets have exclusive cpus or mem.
1694 * If this becomes a problem for some users who wish to
1695 * allow that scenario, then cpuset_post_clone() could be
1696 * changed to grant parent->cpus_allowed-sibling_cpus_exclusive
1697 * (and likewise for mems) to the new cgroup. Called with cgroup_mutex
1698 * held.
1700 static void cpuset_post_clone(struct cgroup_subsys *ss,
1701 struct cgroup *cgroup)
1703 struct cgroup *parent, *child;
1704 struct cpuset *cs, *parent_cs;
1706 parent = cgroup->parent;
1707 list_for_each_entry(child, &parent->children, sibling) {
1708 cs = cgroup_cs(child);
1709 if (is_mem_exclusive(cs) || is_cpu_exclusive(cs))
1710 return;
1712 cs = cgroup_cs(cgroup);
1713 parent_cs = cgroup_cs(parent);
1715 cs->mems_allowed = parent_cs->mems_allowed;
1716 cs->cpus_allowed = parent_cs->cpus_allowed;
1717 return;
1721 * cpuset_create - create a cpuset
1722 * ss: cpuset cgroup subsystem
1723 * cont: control group that the new cpuset will be part of
1726 static struct cgroup_subsys_state *cpuset_create(
1727 struct cgroup_subsys *ss,
1728 struct cgroup *cont)
1730 struct cpuset *cs;
1731 struct cpuset *parent;
1733 if (!cont->parent) {
1734 /* This is early initialization for the top cgroup */
1735 top_cpuset.mems_generation = cpuset_mems_generation++;
1736 return &top_cpuset.css;
1738 parent = cgroup_cs(cont->parent);
1739 cs = kmalloc(sizeof(*cs), GFP_KERNEL);
1740 if (!cs)
1741 return ERR_PTR(-ENOMEM);
1743 cpuset_update_task_memory_state();
1744 cs->flags = 0;
1745 if (is_spread_page(parent))
1746 set_bit(CS_SPREAD_PAGE, &cs->flags);
1747 if (is_spread_slab(parent))
1748 set_bit(CS_SPREAD_SLAB, &cs->flags);
1749 set_bit(CS_SCHED_LOAD_BALANCE, &cs->flags);
1750 cpus_clear(cs->cpus_allowed);
1751 nodes_clear(cs->mems_allowed);
1752 cs->mems_generation = cpuset_mems_generation++;
1753 fmeter_init(&cs->fmeter);
1754 cs->relax_domain_level = -1;
1756 cs->parent = parent;
1757 number_of_cpusets++;
1758 return &cs->css ;
1762 * If the cpuset being removed has its flag 'sched_load_balance'
1763 * enabled, then simulate turning sched_load_balance off, which
1764 * will call async_rebuild_sched_domains().
1767 static void cpuset_destroy(struct cgroup_subsys *ss, struct cgroup *cont)
1769 struct cpuset *cs = cgroup_cs(cont);
1771 cpuset_update_task_memory_state();
1773 if (is_sched_load_balance(cs))
1774 update_flag(CS_SCHED_LOAD_BALANCE, cs, 0);
1776 number_of_cpusets--;
1777 kfree(cs);
1780 struct cgroup_subsys cpuset_subsys = {
1781 .name = "cpuset",
1782 .create = cpuset_create,
1783 .destroy = cpuset_destroy,
1784 .can_attach = cpuset_can_attach,
1785 .attach = cpuset_attach,
1786 .populate = cpuset_populate,
1787 .post_clone = cpuset_post_clone,
1788 .subsys_id = cpuset_subsys_id,
1789 .early_init = 1,
1793 * cpuset_init_early - just enough so that the calls to
1794 * cpuset_update_task_memory_state() in early init code
1795 * are harmless.
1798 int __init cpuset_init_early(void)
1800 top_cpuset.mems_generation = cpuset_mems_generation++;
1801 return 0;
1806 * cpuset_init - initialize cpusets at system boot
1808 * Description: Initialize top_cpuset and the cpuset internal file system,
1811 int __init cpuset_init(void)
1813 int err = 0;
1815 cpus_setall(top_cpuset.cpus_allowed);
1816 nodes_setall(top_cpuset.mems_allowed);
1818 fmeter_init(&top_cpuset.fmeter);
1819 top_cpuset.mems_generation = cpuset_mems_generation++;
1820 set_bit(CS_SCHED_LOAD_BALANCE, &top_cpuset.flags);
1821 top_cpuset.relax_domain_level = -1;
1823 err = register_filesystem(&cpuset_fs_type);
1824 if (err < 0)
1825 return err;
1827 number_of_cpusets = 1;
1828 return 0;
1832 * cpuset_do_move_task - move a given task to another cpuset
1833 * @tsk: pointer to task_struct the task to move
1834 * @scan: struct cgroup_scanner contained in its struct cpuset_hotplug_scanner
1836 * Called by cgroup_scan_tasks() for each task in a cgroup.
1837 * Return nonzero to stop the walk through the tasks.
1839 static void cpuset_do_move_task(struct task_struct *tsk,
1840 struct cgroup_scanner *scan)
1842 struct cpuset_hotplug_scanner *chsp;
1844 chsp = container_of(scan, struct cpuset_hotplug_scanner, scan);
1845 cgroup_attach_task(chsp->to, tsk);
1849 * move_member_tasks_to_cpuset - move tasks from one cpuset to another
1850 * @from: cpuset in which the tasks currently reside
1851 * @to: cpuset to which the tasks will be moved
1853 * Called with cgroup_mutex held
1854 * callback_mutex must not be held, as cpuset_attach() will take it.
1856 * The cgroup_scan_tasks() function will scan all the tasks in a cgroup,
1857 * calling callback functions for each.
1859 static void move_member_tasks_to_cpuset(struct cpuset *from, struct cpuset *to)
1861 struct cpuset_hotplug_scanner scan;
1863 scan.scan.cg = from->css.cgroup;
1864 scan.scan.test_task = NULL; /* select all tasks in cgroup */
1865 scan.scan.process_task = cpuset_do_move_task;
1866 scan.scan.heap = NULL;
1867 scan.to = to->css.cgroup;
1869 if (cgroup_scan_tasks(&scan.scan))
1870 printk(KERN_ERR "move_member_tasks_to_cpuset: "
1871 "cgroup_scan_tasks failed\n");
1875 * If CPU and/or memory hotplug handlers, below, unplug any CPUs
1876 * or memory nodes, we need to walk over the cpuset hierarchy,
1877 * removing that CPU or node from all cpusets. If this removes the
1878 * last CPU or node from a cpuset, then move the tasks in the empty
1879 * cpuset to its next-highest non-empty parent.
1881 * Called with cgroup_mutex held
1882 * callback_mutex must not be held, as cpuset_attach() will take it.
1884 static void remove_tasks_in_empty_cpuset(struct cpuset *cs)
1886 struct cpuset *parent;
1889 * The cgroup's css_sets list is in use if there are tasks
1890 * in the cpuset; the list is empty if there are none;
1891 * the cs->css.refcnt seems always 0.
1893 if (list_empty(&cs->css.cgroup->css_sets))
1894 return;
1897 * Find its next-highest non-empty parent, (top cpuset
1898 * has online cpus, so can't be empty).
1900 parent = cs->parent;
1901 while (cpus_empty(parent->cpus_allowed) ||
1902 nodes_empty(parent->mems_allowed))
1903 parent = parent->parent;
1905 move_member_tasks_to_cpuset(cs, parent);
1909 * Walk the specified cpuset subtree and look for empty cpusets.
1910 * The tasks of such cpuset must be moved to a parent cpuset.
1912 * Called with cgroup_mutex held. We take callback_mutex to modify
1913 * cpus_allowed and mems_allowed.
1915 * This walk processes the tree from top to bottom, completing one layer
1916 * before dropping down to the next. It always processes a node before
1917 * any of its children.
1919 * For now, since we lack memory hot unplug, we'll never see a cpuset
1920 * that has tasks along with an empty 'mems'. But if we did see such
1921 * a cpuset, we'd handle it just like we do if its 'cpus' was empty.
1923 static void scan_for_empty_cpusets(struct cpuset *root)
1925 LIST_HEAD(queue);
1926 struct cpuset *cp; /* scans cpusets being updated */
1927 struct cpuset *child; /* scans child cpusets of cp */
1928 struct cgroup *cont;
1929 nodemask_t oldmems;
1931 list_add_tail((struct list_head *)&root->stack_list, &queue);
1933 while (!list_empty(&queue)) {
1934 cp = list_first_entry(&queue, struct cpuset, stack_list);
1935 list_del(queue.next);
1936 list_for_each_entry(cont, &cp->css.cgroup->children, sibling) {
1937 child = cgroup_cs(cont);
1938 list_add_tail(&child->stack_list, &queue);
1941 /* Continue past cpusets with all cpus, mems online */
1942 if (cpus_subset(cp->cpus_allowed, cpu_online_map) &&
1943 nodes_subset(cp->mems_allowed, node_states[N_HIGH_MEMORY]))
1944 continue;
1946 oldmems = cp->mems_allowed;
1948 /* Remove offline cpus and mems from this cpuset. */
1949 mutex_lock(&callback_mutex);
1950 cpus_and(cp->cpus_allowed, cp->cpus_allowed, cpu_online_map);
1951 nodes_and(cp->mems_allowed, cp->mems_allowed,
1952 node_states[N_HIGH_MEMORY]);
1953 mutex_unlock(&callback_mutex);
1955 /* Move tasks from the empty cpuset to a parent */
1956 if (cpus_empty(cp->cpus_allowed) ||
1957 nodes_empty(cp->mems_allowed))
1958 remove_tasks_in_empty_cpuset(cp);
1959 else {
1960 update_tasks_cpumask(cp, NULL);
1961 update_tasks_nodemask(cp, &oldmems);
1967 * The top_cpuset tracks what CPUs and Memory Nodes are online,
1968 * period. This is necessary in order to make cpusets transparent
1969 * (of no affect) on systems that are actively using CPU hotplug
1970 * but making no active use of cpusets.
1972 * This routine ensures that top_cpuset.cpus_allowed tracks
1973 * cpu_online_map on each CPU hotplug (cpuhp) event.
1975 * Called within get_online_cpus(). Needs to call cgroup_lock()
1976 * before calling generate_sched_domains().
1978 static int cpuset_track_online_cpus(struct notifier_block *unused_nb,
1979 unsigned long phase, void *unused_cpu)
1981 struct sched_domain_attr *attr;
1982 cpumask_t *doms;
1983 int ndoms;
1985 switch (phase) {
1986 case CPU_ONLINE:
1987 case CPU_ONLINE_FROZEN:
1988 case CPU_DEAD:
1989 case CPU_DEAD_FROZEN:
1990 break;
1992 default:
1993 return NOTIFY_DONE;
1996 cgroup_lock();
1997 top_cpuset.cpus_allowed = cpu_online_map;
1998 scan_for_empty_cpusets(&top_cpuset);
1999 ndoms = generate_sched_domains(&doms, &attr);
2000 cgroup_unlock();
2002 /* Have scheduler rebuild the domains */
2003 partition_sched_domains(ndoms, doms, attr);
2005 return NOTIFY_OK;
2008 #ifdef CONFIG_MEMORY_HOTPLUG
2010 * Keep top_cpuset.mems_allowed tracking node_states[N_HIGH_MEMORY].
2011 * Call this routine anytime after node_states[N_HIGH_MEMORY] changes.
2012 * See also the previous routine cpuset_track_online_cpus().
2014 void cpuset_track_online_nodes(void)
2016 cgroup_lock();
2017 top_cpuset.mems_allowed = node_states[N_HIGH_MEMORY];
2018 scan_for_empty_cpusets(&top_cpuset);
2019 cgroup_unlock();
2021 #endif
2024 * cpuset_init_smp - initialize cpus_allowed
2026 * Description: Finish top cpuset after cpu, node maps are initialized
2029 void __init cpuset_init_smp(void)
2031 top_cpuset.cpus_allowed = cpu_online_map;
2032 top_cpuset.mems_allowed = node_states[N_HIGH_MEMORY];
2034 hotcpu_notifier(cpuset_track_online_cpus, 0);
2038 * cpuset_cpus_allowed - return cpus_allowed mask from a tasks cpuset.
2039 * @tsk: pointer to task_struct from which to obtain cpuset->cpus_allowed.
2040 * @pmask: pointer to cpumask_t variable to receive cpus_allowed set.
2042 * Description: Returns the cpumask_t cpus_allowed of the cpuset
2043 * attached to the specified @tsk. Guaranteed to return some non-empty
2044 * subset of cpu_online_map, even if this means going outside the
2045 * tasks cpuset.
2048 void cpuset_cpus_allowed(struct task_struct *tsk, cpumask_t *pmask)
2050 mutex_lock(&callback_mutex);
2051 cpuset_cpus_allowed_locked(tsk, pmask);
2052 mutex_unlock(&callback_mutex);
2056 * cpuset_cpus_allowed_locked - return cpus_allowed mask from a tasks cpuset.
2057 * Must be called with callback_mutex held.
2059 void cpuset_cpus_allowed_locked(struct task_struct *tsk, cpumask_t *pmask)
2061 task_lock(tsk);
2062 guarantee_online_cpus(task_cs(tsk), pmask);
2063 task_unlock(tsk);
2066 void cpuset_init_current_mems_allowed(void)
2068 nodes_setall(current->mems_allowed);
2072 * cpuset_mems_allowed - return mems_allowed mask from a tasks cpuset.
2073 * @tsk: pointer to task_struct from which to obtain cpuset->mems_allowed.
2075 * Description: Returns the nodemask_t mems_allowed of the cpuset
2076 * attached to the specified @tsk. Guaranteed to return some non-empty
2077 * subset of node_states[N_HIGH_MEMORY], even if this means going outside the
2078 * tasks cpuset.
2081 nodemask_t cpuset_mems_allowed(struct task_struct *tsk)
2083 nodemask_t mask;
2085 mutex_lock(&callback_mutex);
2086 task_lock(tsk);
2087 guarantee_online_mems(task_cs(tsk), &mask);
2088 task_unlock(tsk);
2089 mutex_unlock(&callback_mutex);
2091 return mask;
2095 * cpuset_nodemask_valid_mems_allowed - check nodemask vs. curremt mems_allowed
2096 * @nodemask: the nodemask to be checked
2098 * Are any of the nodes in the nodemask allowed in current->mems_allowed?
2100 int cpuset_nodemask_valid_mems_allowed(nodemask_t *nodemask)
2102 return nodes_intersects(*nodemask, current->mems_allowed);
2106 * nearest_hardwall_ancestor() - Returns the nearest mem_exclusive or
2107 * mem_hardwall ancestor to the specified cpuset. Call holding
2108 * callback_mutex. If no ancestor is mem_exclusive or mem_hardwall
2109 * (an unusual configuration), then returns the root cpuset.
2111 static const struct cpuset *nearest_hardwall_ancestor(const struct cpuset *cs)
2113 while (!(is_mem_exclusive(cs) || is_mem_hardwall(cs)) && cs->parent)
2114 cs = cs->parent;
2115 return cs;
2119 * cpuset_zone_allowed_softwall - Can we allocate on zone z's memory node?
2120 * @z: is this zone on an allowed node?
2121 * @gfp_mask: memory allocation flags
2123 * If we're in interrupt, yes, we can always allocate. If
2124 * __GFP_THISNODE is set, yes, we can always allocate. If zone
2125 * z's node is in our tasks mems_allowed, yes. If it's not a
2126 * __GFP_HARDWALL request and this zone's nodes is in the nearest
2127 * hardwalled cpuset ancestor to this tasks cpuset, yes.
2128 * If the task has been OOM killed and has access to memory reserves
2129 * as specified by the TIF_MEMDIE flag, yes.
2130 * Otherwise, no.
2132 * If __GFP_HARDWALL is set, cpuset_zone_allowed_softwall()
2133 * reduces to cpuset_zone_allowed_hardwall(). Otherwise,
2134 * cpuset_zone_allowed_softwall() might sleep, and might allow a zone
2135 * from an enclosing cpuset.
2137 * cpuset_zone_allowed_hardwall() only handles the simpler case of
2138 * hardwall cpusets, and never sleeps.
2140 * The __GFP_THISNODE placement logic is really handled elsewhere,
2141 * by forcibly using a zonelist starting at a specified node, and by
2142 * (in get_page_from_freelist()) refusing to consider the zones for
2143 * any node on the zonelist except the first. By the time any such
2144 * calls get to this routine, we should just shut up and say 'yes'.
2146 * GFP_USER allocations are marked with the __GFP_HARDWALL bit,
2147 * and do not allow allocations outside the current tasks cpuset
2148 * unless the task has been OOM killed as is marked TIF_MEMDIE.
2149 * GFP_KERNEL allocations are not so marked, so can escape to the
2150 * nearest enclosing hardwalled ancestor cpuset.
2152 * Scanning up parent cpusets requires callback_mutex. The
2153 * __alloc_pages() routine only calls here with __GFP_HARDWALL bit
2154 * _not_ set if it's a GFP_KERNEL allocation, and all nodes in the
2155 * current tasks mems_allowed came up empty on the first pass over
2156 * the zonelist. So only GFP_KERNEL allocations, if all nodes in the
2157 * cpuset are short of memory, might require taking the callback_mutex
2158 * mutex.
2160 * The first call here from mm/page_alloc:get_page_from_freelist()
2161 * has __GFP_HARDWALL set in gfp_mask, enforcing hardwall cpusets,
2162 * so no allocation on a node outside the cpuset is allowed (unless
2163 * in interrupt, of course).
2165 * The second pass through get_page_from_freelist() doesn't even call
2166 * here for GFP_ATOMIC calls. For those calls, the __alloc_pages()
2167 * variable 'wait' is not set, and the bit ALLOC_CPUSET is not set
2168 * in alloc_flags. That logic and the checks below have the combined
2169 * affect that:
2170 * in_interrupt - any node ok (current task context irrelevant)
2171 * GFP_ATOMIC - any node ok
2172 * TIF_MEMDIE - any node ok
2173 * GFP_KERNEL - any node in enclosing hardwalled cpuset ok
2174 * GFP_USER - only nodes in current tasks mems allowed ok.
2176 * Rule:
2177 * Don't call cpuset_zone_allowed_softwall if you can't sleep, unless you
2178 * pass in the __GFP_HARDWALL flag set in gfp_flag, which disables
2179 * the code that might scan up ancestor cpusets and sleep.
2182 int __cpuset_zone_allowed_softwall(struct zone *z, gfp_t gfp_mask)
2184 int node; /* node that zone z is on */
2185 const struct cpuset *cs; /* current cpuset ancestors */
2186 int allowed; /* is allocation in zone z allowed? */
2188 if (in_interrupt() || (gfp_mask & __GFP_THISNODE))
2189 return 1;
2190 node = zone_to_nid(z);
2191 might_sleep_if(!(gfp_mask & __GFP_HARDWALL));
2192 if (node_isset(node, current->mems_allowed))
2193 return 1;
2195 * Allow tasks that have access to memory reserves because they have
2196 * been OOM killed to get memory anywhere.
2198 if (unlikely(test_thread_flag(TIF_MEMDIE)))
2199 return 1;
2200 if (gfp_mask & __GFP_HARDWALL) /* If hardwall request, stop here */
2201 return 0;
2203 if (current->flags & PF_EXITING) /* Let dying task have memory */
2204 return 1;
2206 /* Not hardwall and node outside mems_allowed: scan up cpusets */
2207 mutex_lock(&callback_mutex);
2209 task_lock(current);
2210 cs = nearest_hardwall_ancestor(task_cs(current));
2211 task_unlock(current);
2213 allowed = node_isset(node, cs->mems_allowed);
2214 mutex_unlock(&callback_mutex);
2215 return allowed;
2219 * cpuset_zone_allowed_hardwall - Can we allocate on zone z's memory node?
2220 * @z: is this zone on an allowed node?
2221 * @gfp_mask: memory allocation flags
2223 * If we're in interrupt, yes, we can always allocate.
2224 * If __GFP_THISNODE is set, yes, we can always allocate. If zone
2225 * z's node is in our tasks mems_allowed, yes. If the task has been
2226 * OOM killed and has access to memory reserves as specified by the
2227 * TIF_MEMDIE flag, yes. Otherwise, no.
2229 * The __GFP_THISNODE placement logic is really handled elsewhere,
2230 * by forcibly using a zonelist starting at a specified node, and by
2231 * (in get_page_from_freelist()) refusing to consider the zones for
2232 * any node on the zonelist except the first. By the time any such
2233 * calls get to this routine, we should just shut up and say 'yes'.
2235 * Unlike the cpuset_zone_allowed_softwall() variant, above,
2236 * this variant requires that the zone be in the current tasks
2237 * mems_allowed or that we're in interrupt. It does not scan up the
2238 * cpuset hierarchy for the nearest enclosing mem_exclusive cpuset.
2239 * It never sleeps.
2242 int __cpuset_zone_allowed_hardwall(struct zone *z, gfp_t gfp_mask)
2244 int node; /* node that zone z is on */
2246 if (in_interrupt() || (gfp_mask & __GFP_THISNODE))
2247 return 1;
2248 node = zone_to_nid(z);
2249 if (node_isset(node, current->mems_allowed))
2250 return 1;
2252 * Allow tasks that have access to memory reserves because they have
2253 * been OOM killed to get memory anywhere.
2255 if (unlikely(test_thread_flag(TIF_MEMDIE)))
2256 return 1;
2257 return 0;
2261 * cpuset_lock - lock out any changes to cpuset structures
2263 * The out of memory (oom) code needs to mutex_lock cpusets
2264 * from being changed while it scans the tasklist looking for a
2265 * task in an overlapping cpuset. Expose callback_mutex via this
2266 * cpuset_lock() routine, so the oom code can lock it, before
2267 * locking the task list. The tasklist_lock is a spinlock, so
2268 * must be taken inside callback_mutex.
2271 void cpuset_lock(void)
2273 mutex_lock(&callback_mutex);
2277 * cpuset_unlock - release lock on cpuset changes
2279 * Undo the lock taken in a previous cpuset_lock() call.
2282 void cpuset_unlock(void)
2284 mutex_unlock(&callback_mutex);
2288 * cpuset_mem_spread_node() - On which node to begin search for a page
2290 * If a task is marked PF_SPREAD_PAGE or PF_SPREAD_SLAB (as for
2291 * tasks in a cpuset with is_spread_page or is_spread_slab set),
2292 * and if the memory allocation used cpuset_mem_spread_node()
2293 * to determine on which node to start looking, as it will for
2294 * certain page cache or slab cache pages such as used for file
2295 * system buffers and inode caches, then instead of starting on the
2296 * local node to look for a free page, rather spread the starting
2297 * node around the tasks mems_allowed nodes.
2299 * We don't have to worry about the returned node being offline
2300 * because "it can't happen", and even if it did, it would be ok.
2302 * The routines calling guarantee_online_mems() are careful to
2303 * only set nodes in task->mems_allowed that are online. So it
2304 * should not be possible for the following code to return an
2305 * offline node. But if it did, that would be ok, as this routine
2306 * is not returning the node where the allocation must be, only
2307 * the node where the search should start. The zonelist passed to
2308 * __alloc_pages() will include all nodes. If the slab allocator
2309 * is passed an offline node, it will fall back to the local node.
2310 * See kmem_cache_alloc_node().
2313 int cpuset_mem_spread_node(void)
2315 int node;
2317 node = next_node(current->cpuset_mem_spread_rotor, current->mems_allowed);
2318 if (node == MAX_NUMNODES)
2319 node = first_node(current->mems_allowed);
2320 current->cpuset_mem_spread_rotor = node;
2321 return node;
2323 EXPORT_SYMBOL_GPL(cpuset_mem_spread_node);
2326 * cpuset_mems_allowed_intersects - Does @tsk1's mems_allowed intersect @tsk2's?
2327 * @tsk1: pointer to task_struct of some task.
2328 * @tsk2: pointer to task_struct of some other task.
2330 * Description: Return true if @tsk1's mems_allowed intersects the
2331 * mems_allowed of @tsk2. Used by the OOM killer to determine if
2332 * one of the task's memory usage might impact the memory available
2333 * to the other.
2336 int cpuset_mems_allowed_intersects(const struct task_struct *tsk1,
2337 const struct task_struct *tsk2)
2339 return nodes_intersects(tsk1->mems_allowed, tsk2->mems_allowed);
2343 * Collection of memory_pressure is suppressed unless
2344 * this flag is enabled by writing "1" to the special
2345 * cpuset file 'memory_pressure_enabled' in the root cpuset.
2348 int cpuset_memory_pressure_enabled __read_mostly;
2351 * cpuset_memory_pressure_bump - keep stats of per-cpuset reclaims.
2353 * Keep a running average of the rate of synchronous (direct)
2354 * page reclaim efforts initiated by tasks in each cpuset.
2356 * This represents the rate at which some task in the cpuset
2357 * ran low on memory on all nodes it was allowed to use, and
2358 * had to enter the kernels page reclaim code in an effort to
2359 * create more free memory by tossing clean pages or swapping
2360 * or writing dirty pages.
2362 * Display to user space in the per-cpuset read-only file
2363 * "memory_pressure". Value displayed is an integer
2364 * representing the recent rate of entry into the synchronous
2365 * (direct) page reclaim by any task attached to the cpuset.
2368 void __cpuset_memory_pressure_bump(void)
2370 task_lock(current);
2371 fmeter_markevent(&task_cs(current)->fmeter);
2372 task_unlock(current);
2375 #ifdef CONFIG_PROC_PID_CPUSET
2377 * proc_cpuset_show()
2378 * - Print tasks cpuset path into seq_file.
2379 * - Used for /proc/<pid>/cpuset.
2380 * - No need to task_lock(tsk) on this tsk->cpuset reference, as it
2381 * doesn't really matter if tsk->cpuset changes after we read it,
2382 * and we take cgroup_mutex, keeping cpuset_attach() from changing it
2383 * anyway.
2385 static int proc_cpuset_show(struct seq_file *m, void *unused_v)
2387 struct pid *pid;
2388 struct task_struct *tsk;
2389 char *buf;
2390 struct cgroup_subsys_state *css;
2391 int retval;
2393 retval = -ENOMEM;
2394 buf = kmalloc(PAGE_SIZE, GFP_KERNEL);
2395 if (!buf)
2396 goto out;
2398 retval = -ESRCH;
2399 pid = m->private;
2400 tsk = get_pid_task(pid, PIDTYPE_PID);
2401 if (!tsk)
2402 goto out_free;
2404 retval = -EINVAL;
2405 cgroup_lock();
2406 css = task_subsys_state(tsk, cpuset_subsys_id);
2407 retval = cgroup_path(css->cgroup, buf, PAGE_SIZE);
2408 if (retval < 0)
2409 goto out_unlock;
2410 seq_puts(m, buf);
2411 seq_putc(m, '\n');
2412 out_unlock:
2413 cgroup_unlock();
2414 put_task_struct(tsk);
2415 out_free:
2416 kfree(buf);
2417 out:
2418 return retval;
2421 static int cpuset_open(struct inode *inode, struct file *file)
2423 struct pid *pid = PROC_I(inode)->pid;
2424 return single_open(file, proc_cpuset_show, pid);
2427 const struct file_operations proc_cpuset_operations = {
2428 .open = cpuset_open,
2429 .read = seq_read,
2430 .llseek = seq_lseek,
2431 .release = single_release,
2433 #endif /* CONFIG_PROC_PID_CPUSET */
2435 /* Display task cpus_allowed, mems_allowed in /proc/<pid>/status file. */
2436 void cpuset_task_status_allowed(struct seq_file *m, struct task_struct *task)
2438 seq_printf(m, "Cpus_allowed:\t");
2439 seq_cpumask(m, &task->cpus_allowed);
2440 seq_printf(m, "\n");
2441 seq_printf(m, "Cpus_allowed_list:\t");
2442 seq_cpumask_list(m, &task->cpus_allowed);
2443 seq_printf(m, "\n");
2444 seq_printf(m, "Mems_allowed:\t");
2445 seq_nodemask(m, &task->mems_allowed);
2446 seq_printf(m, "\n");
2447 seq_printf(m, "Mems_allowed_list:\t");
2448 seq_nodemask_list(m, &task->mems_allowed);
2449 seq_printf(m, "\n");