KVM: Allow reading aliases with mmu_lock
[linux-2.6/linux-acpi-2.6/ibm-acpi-2.6.git] / kernel / cpuset.c
blob91cf85b36dd577c2a231498575fcea1805ba44b7
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
18 * This file is subject to the terms and conditions of the GNU General Public
19 * License. See the file COPYING in the main directory of the Linux
20 * distribution for more details.
23 #include <linux/cpu.h>
24 #include <linux/cpumask.h>
25 #include <linux/cpuset.h>
26 #include <linux/err.h>
27 #include <linux/errno.h>
28 #include <linux/file.h>
29 #include <linux/fs.h>
30 #include <linux/init.h>
31 #include <linux/interrupt.h>
32 #include <linux/kernel.h>
33 #include <linux/kmod.h>
34 #include <linux/list.h>
35 #include <linux/mempolicy.h>
36 #include <linux/mm.h>
37 #include <linux/module.h>
38 #include <linux/mount.h>
39 #include <linux/namei.h>
40 #include <linux/pagemap.h>
41 #include <linux/proc_fs.h>
42 #include <linux/rcupdate.h>
43 #include <linux/sched.h>
44 #include <linux/seq_file.h>
45 #include <linux/security.h>
46 #include <linux/slab.h>
47 #include <linux/spinlock.h>
48 #include <linux/stat.h>
49 #include <linux/string.h>
50 #include <linux/time.h>
51 #include <linux/backing-dev.h>
52 #include <linux/sort.h>
54 #include <asm/uaccess.h>
55 #include <asm/atomic.h>
56 #include <linux/mutex.h>
57 #include <linux/kfifo.h>
58 #include <linux/workqueue.h>
59 #include <linux/cgroup.h>
62 * Tracks how many cpusets are currently defined in system.
63 * When there is only one cpuset (the root cpuset) we can
64 * short circuit some hooks.
66 int number_of_cpusets __read_mostly;
68 /* Forward declare cgroup structures */
69 struct cgroup_subsys cpuset_subsys;
70 struct cpuset;
72 /* See "Frequency meter" comments, below. */
74 struct fmeter {
75 int cnt; /* unprocessed events count */
76 int val; /* most recent output value */
77 time_t time; /* clock (secs) when val computed */
78 spinlock_t lock; /* guards read or write of above */
81 struct cpuset {
82 struct cgroup_subsys_state css;
84 unsigned long flags; /* "unsigned long" so bitops work */
85 cpumask_t cpus_allowed; /* CPUs allowed to tasks in cpuset */
86 nodemask_t mems_allowed; /* Memory Nodes allowed to tasks */
88 struct cpuset *parent; /* my parent */
91 * Copy of global cpuset_mems_generation as of the most
92 * recent time this cpuset changed its mems_allowed.
94 int mems_generation;
96 struct fmeter fmeter; /* memory_pressure filter */
98 /* partition number for rebuild_sched_domains() */
99 int pn;
101 /* for custom sched domain */
102 int relax_domain_level;
104 /* used for walking a cpuset heirarchy */
105 struct list_head stack_list;
108 /* Retrieve the cpuset for a cgroup */
109 static inline struct cpuset *cgroup_cs(struct cgroup *cont)
111 return container_of(cgroup_subsys_state(cont, cpuset_subsys_id),
112 struct cpuset, css);
115 /* Retrieve the cpuset for a task */
116 static inline struct cpuset *task_cs(struct task_struct *task)
118 return container_of(task_subsys_state(task, cpuset_subsys_id),
119 struct cpuset, css);
121 struct cpuset_hotplug_scanner {
122 struct cgroup_scanner scan;
123 struct cgroup *to;
126 /* bits in struct cpuset flags field */
127 typedef enum {
128 CS_CPU_EXCLUSIVE,
129 CS_MEM_EXCLUSIVE,
130 CS_MEM_HARDWALL,
131 CS_MEMORY_MIGRATE,
132 CS_SCHED_LOAD_BALANCE,
133 CS_SPREAD_PAGE,
134 CS_SPREAD_SLAB,
135 } cpuset_flagbits_t;
137 /* convenient tests for these bits */
138 static inline int is_cpu_exclusive(const struct cpuset *cs)
140 return test_bit(CS_CPU_EXCLUSIVE, &cs->flags);
143 static inline int is_mem_exclusive(const struct cpuset *cs)
145 return test_bit(CS_MEM_EXCLUSIVE, &cs->flags);
148 static inline int is_mem_hardwall(const struct cpuset *cs)
150 return test_bit(CS_MEM_HARDWALL, &cs->flags);
153 static inline int is_sched_load_balance(const struct cpuset *cs)
155 return test_bit(CS_SCHED_LOAD_BALANCE, &cs->flags);
158 static inline int is_memory_migrate(const struct cpuset *cs)
160 return test_bit(CS_MEMORY_MIGRATE, &cs->flags);
163 static inline int is_spread_page(const struct cpuset *cs)
165 return test_bit(CS_SPREAD_PAGE, &cs->flags);
168 static inline int is_spread_slab(const struct cpuset *cs)
170 return test_bit(CS_SPREAD_SLAB, &cs->flags);
174 * Increment this integer everytime any cpuset changes its
175 * mems_allowed value. Users of cpusets can track this generation
176 * number, and avoid having to lock and reload mems_allowed unless
177 * the cpuset they're using changes generation.
179 * A single, global generation is needed because cpuset_attach_task() could
180 * reattach a task to a different cpuset, which must not have its
181 * generation numbers aliased with those of that tasks previous cpuset.
183 * Generations are needed for mems_allowed because one task cannot
184 * modify another's memory placement. So we must enable every task,
185 * on every visit to __alloc_pages(), to efficiently check whether
186 * its current->cpuset->mems_allowed has changed, requiring an update
187 * of its current->mems_allowed.
189 * Since writes to cpuset_mems_generation are guarded by the cgroup lock
190 * there is no need to mark it atomic.
192 static int cpuset_mems_generation;
194 static struct cpuset top_cpuset = {
195 .flags = ((1 << CS_CPU_EXCLUSIVE) | (1 << CS_MEM_EXCLUSIVE)),
196 .cpus_allowed = CPU_MASK_ALL,
197 .mems_allowed = NODE_MASK_ALL,
201 * There are two global mutexes guarding cpuset structures. The first
202 * is the main control groups cgroup_mutex, accessed via
203 * cgroup_lock()/cgroup_unlock(). The second is the cpuset-specific
204 * callback_mutex, below. They can nest. It is ok to first take
205 * cgroup_mutex, then nest callback_mutex. We also require taking
206 * task_lock() when dereferencing a task's cpuset pointer. See "The
207 * task_lock() exception", at the end of this comment.
209 * A task must hold both mutexes to modify cpusets. If a task
210 * holds cgroup_mutex, then it blocks others wanting that mutex,
211 * ensuring that it is the only task able to also acquire callback_mutex
212 * and be able to modify cpusets. It can perform various checks on
213 * the cpuset structure first, knowing nothing will change. It can
214 * also allocate memory while just holding cgroup_mutex. While it is
215 * performing these checks, various callback routines can briefly
216 * acquire callback_mutex to query cpusets. Once it is ready to make
217 * the changes, it takes callback_mutex, blocking everyone else.
219 * Calls to the kernel memory allocator can not be made while holding
220 * callback_mutex, as that would risk double tripping on callback_mutex
221 * from one of the callbacks into the cpuset code from within
222 * __alloc_pages().
224 * If a task is only holding callback_mutex, then it has read-only
225 * access to cpusets.
227 * The task_struct fields mems_allowed and mems_generation may only
228 * be accessed in the context of that task, so require no locks.
230 * The cpuset_common_file_read() handlers only hold callback_mutex across
231 * small pieces of code, such as when reading out possibly multi-word
232 * cpumasks and nodemasks.
234 * Accessing a task's cpuset should be done in accordance with the
235 * guidelines for accessing subsystem state in kernel/cgroup.c
238 static DEFINE_MUTEX(callback_mutex);
240 /* This is ugly, but preserves the userspace API for existing cpuset
241 * users. If someone tries to mount the "cpuset" filesystem, we
242 * silently switch it to mount "cgroup" instead */
243 static int cpuset_get_sb(struct file_system_type *fs_type,
244 int flags, const char *unused_dev_name,
245 void *data, struct vfsmount *mnt)
247 struct file_system_type *cgroup_fs = get_fs_type("cgroup");
248 int ret = -ENODEV;
249 if (cgroup_fs) {
250 char mountopts[] =
251 "cpuset,noprefix,"
252 "release_agent=/sbin/cpuset_release_agent";
253 ret = cgroup_fs->get_sb(cgroup_fs, flags,
254 unused_dev_name, mountopts, mnt);
255 put_filesystem(cgroup_fs);
257 return ret;
260 static struct file_system_type cpuset_fs_type = {
261 .name = "cpuset",
262 .get_sb = cpuset_get_sb,
266 * Return in *pmask the portion of a cpusets's cpus_allowed that
267 * are online. If none are online, walk up the cpuset hierarchy
268 * until we find one that does have some online cpus. If we get
269 * all the way to the top and still haven't found any online cpus,
270 * return cpu_online_map. Or if passed a NULL cs from an exit'ing
271 * task, return cpu_online_map.
273 * One way or another, we guarantee to return some non-empty subset
274 * of cpu_online_map.
276 * Call with callback_mutex held.
279 static void guarantee_online_cpus(const struct cpuset *cs, cpumask_t *pmask)
281 while (cs && !cpus_intersects(cs->cpus_allowed, cpu_online_map))
282 cs = cs->parent;
283 if (cs)
284 cpus_and(*pmask, cs->cpus_allowed, cpu_online_map);
285 else
286 *pmask = cpu_online_map;
287 BUG_ON(!cpus_intersects(*pmask, cpu_online_map));
291 * Return in *pmask the portion of a cpusets's mems_allowed that
292 * are online, with memory. If none are online with memory, walk
293 * up the cpuset hierarchy until we find one that does have some
294 * online mems. If we get all the way to the top and still haven't
295 * found any online mems, return node_states[N_HIGH_MEMORY].
297 * One way or another, we guarantee to return some non-empty subset
298 * of node_states[N_HIGH_MEMORY].
300 * Call with callback_mutex held.
303 static void guarantee_online_mems(const struct cpuset *cs, nodemask_t *pmask)
305 while (cs && !nodes_intersects(cs->mems_allowed,
306 node_states[N_HIGH_MEMORY]))
307 cs = cs->parent;
308 if (cs)
309 nodes_and(*pmask, cs->mems_allowed,
310 node_states[N_HIGH_MEMORY]);
311 else
312 *pmask = node_states[N_HIGH_MEMORY];
313 BUG_ON(!nodes_intersects(*pmask, node_states[N_HIGH_MEMORY]));
317 * cpuset_update_task_memory_state - update task memory placement
319 * If the current tasks cpusets mems_allowed changed behind our
320 * backs, update current->mems_allowed, mems_generation and task NUMA
321 * mempolicy to the new value.
323 * Task mempolicy is updated by rebinding it relative to the
324 * current->cpuset if a task has its memory placement changed.
325 * Do not call this routine if in_interrupt().
327 * Call without callback_mutex or task_lock() held. May be
328 * called with or without cgroup_mutex held. Thanks in part to
329 * 'the_top_cpuset_hack', the task's cpuset pointer will never
330 * be NULL. This routine also might acquire callback_mutex during
331 * call.
333 * Reading current->cpuset->mems_generation doesn't need task_lock
334 * to guard the current->cpuset derefence, because it is guarded
335 * from concurrent freeing of current->cpuset using RCU.
337 * The rcu_dereference() is technically probably not needed,
338 * as I don't actually mind if I see a new cpuset pointer but
339 * an old value of mems_generation. However this really only
340 * matters on alpha systems using cpusets heavily. If I dropped
341 * that rcu_dereference(), it would save them a memory barrier.
342 * For all other arch's, rcu_dereference is a no-op anyway, and for
343 * alpha systems not using cpusets, another planned optimization,
344 * avoiding the rcu critical section for tasks in the root cpuset
345 * which is statically allocated, so can't vanish, will make this
346 * irrelevant. Better to use RCU as intended, than to engage in
347 * some cute trick to save a memory barrier that is impossible to
348 * test, for alpha systems using cpusets heavily, which might not
349 * even exist.
351 * This routine is needed to update the per-task mems_allowed data,
352 * within the tasks context, when it is trying to allocate memory
353 * (in various mm/mempolicy.c routines) and notices that some other
354 * task has been modifying its cpuset.
357 void cpuset_update_task_memory_state(void)
359 int my_cpusets_mem_gen;
360 struct task_struct *tsk = current;
361 struct cpuset *cs;
363 if (task_cs(tsk) == &top_cpuset) {
364 /* Don't need rcu for top_cpuset. It's never freed. */
365 my_cpusets_mem_gen = top_cpuset.mems_generation;
366 } else {
367 rcu_read_lock();
368 my_cpusets_mem_gen = task_cs(tsk)->mems_generation;
369 rcu_read_unlock();
372 if (my_cpusets_mem_gen != tsk->cpuset_mems_generation) {
373 mutex_lock(&callback_mutex);
374 task_lock(tsk);
375 cs = task_cs(tsk); /* Maybe changed when task not locked */
376 guarantee_online_mems(cs, &tsk->mems_allowed);
377 tsk->cpuset_mems_generation = cs->mems_generation;
378 if (is_spread_page(cs))
379 tsk->flags |= PF_SPREAD_PAGE;
380 else
381 tsk->flags &= ~PF_SPREAD_PAGE;
382 if (is_spread_slab(cs))
383 tsk->flags |= PF_SPREAD_SLAB;
384 else
385 tsk->flags &= ~PF_SPREAD_SLAB;
386 task_unlock(tsk);
387 mutex_unlock(&callback_mutex);
388 mpol_rebind_task(tsk, &tsk->mems_allowed);
393 * is_cpuset_subset(p, q) - Is cpuset p a subset of cpuset q?
395 * One cpuset is a subset of another if all its allowed CPUs and
396 * Memory Nodes are a subset of the other, and its exclusive flags
397 * are only set if the other's are set. Call holding cgroup_mutex.
400 static int is_cpuset_subset(const struct cpuset *p, const struct cpuset *q)
402 return cpus_subset(p->cpus_allowed, q->cpus_allowed) &&
403 nodes_subset(p->mems_allowed, q->mems_allowed) &&
404 is_cpu_exclusive(p) <= is_cpu_exclusive(q) &&
405 is_mem_exclusive(p) <= is_mem_exclusive(q);
409 * validate_change() - Used to validate that any proposed cpuset change
410 * follows the structural rules for cpusets.
412 * If we replaced the flag and mask values of the current cpuset
413 * (cur) with those values in the trial cpuset (trial), would
414 * our various subset and exclusive rules still be valid? Presumes
415 * cgroup_mutex held.
417 * 'cur' is the address of an actual, in-use cpuset. Operations
418 * such as list traversal that depend on the actual address of the
419 * cpuset in the list must use cur below, not trial.
421 * 'trial' is the address of bulk structure copy of cur, with
422 * perhaps one or more of the fields cpus_allowed, mems_allowed,
423 * or flags changed to new, trial values.
425 * Return 0 if valid, -errno if not.
428 static int validate_change(const struct cpuset *cur, const struct cpuset *trial)
430 struct cgroup *cont;
431 struct cpuset *c, *par;
433 /* Each of our child cpusets must be a subset of us */
434 list_for_each_entry(cont, &cur->css.cgroup->children, sibling) {
435 if (!is_cpuset_subset(cgroup_cs(cont), trial))
436 return -EBUSY;
439 /* Remaining checks don't apply to root cpuset */
440 if (cur == &top_cpuset)
441 return 0;
443 par = cur->parent;
445 /* We must be a subset of our parent cpuset */
446 if (!is_cpuset_subset(trial, par))
447 return -EACCES;
450 * If either I or some sibling (!= me) is exclusive, we can't
451 * overlap
453 list_for_each_entry(cont, &par->css.cgroup->children, sibling) {
454 c = cgroup_cs(cont);
455 if ((is_cpu_exclusive(trial) || is_cpu_exclusive(c)) &&
456 c != cur &&
457 cpus_intersects(trial->cpus_allowed, c->cpus_allowed))
458 return -EINVAL;
459 if ((is_mem_exclusive(trial) || is_mem_exclusive(c)) &&
460 c != cur &&
461 nodes_intersects(trial->mems_allowed, c->mems_allowed))
462 return -EINVAL;
465 /* Cpusets with tasks can't have empty cpus_allowed or mems_allowed */
466 if (cgroup_task_count(cur->css.cgroup)) {
467 if (cpus_empty(trial->cpus_allowed) ||
468 nodes_empty(trial->mems_allowed)) {
469 return -ENOSPC;
473 return 0;
477 * Helper routine for rebuild_sched_domains().
478 * Do cpusets a, b have overlapping cpus_allowed masks?
481 static int cpusets_overlap(struct cpuset *a, struct cpuset *b)
483 return cpus_intersects(a->cpus_allowed, b->cpus_allowed);
486 static void
487 update_domain_attr(struct sched_domain_attr *dattr, struct cpuset *c)
489 if (!dattr)
490 return;
491 if (dattr->relax_domain_level < c->relax_domain_level)
492 dattr->relax_domain_level = c->relax_domain_level;
493 return;
497 * rebuild_sched_domains()
499 * This routine will be called to rebuild the scheduler's dynamic
500 * sched domains:
501 * - if the flag 'sched_load_balance' of any cpuset with non-empty
502 * 'cpus' changes,
503 * - or if the 'cpus' allowed changes in any cpuset which has that
504 * flag enabled,
505 * - or if the 'sched_relax_domain_level' of any cpuset which has
506 * that flag enabled and with non-empty 'cpus' changes,
507 * - or if any cpuset with non-empty 'cpus' is removed,
508 * - or if a cpu gets offlined.
510 * This routine builds a partial partition of the systems CPUs
511 * (the set of non-overlappping cpumask_t's in the array 'part'
512 * below), and passes that partial partition to the kernel/sched.c
513 * partition_sched_domains() routine, which will rebuild the
514 * schedulers load balancing domains (sched domains) as specified
515 * by that partial partition. A 'partial partition' is a set of
516 * non-overlapping subsets whose union is a subset of that set.
518 * See "What is sched_load_balance" in Documentation/cpusets.txt
519 * for a background explanation of this.
521 * Does not return errors, on the theory that the callers of this
522 * routine would rather not worry about failures to rebuild sched
523 * domains when operating in the severe memory shortage situations
524 * that could cause allocation failures below.
526 * Call with cgroup_mutex held. May take callback_mutex during
527 * call due to the kfifo_alloc() and kmalloc() calls. May nest
528 * a call to the get_online_cpus()/put_online_cpus() pair.
529 * Must not be called holding callback_mutex, because we must not
530 * call get_online_cpus() while holding callback_mutex. Elsewhere
531 * the kernel nests callback_mutex inside get_online_cpus() calls.
532 * So the reverse nesting would risk an ABBA deadlock.
534 * The three key local variables below are:
535 * q - a kfifo queue of cpuset pointers, used to implement a
536 * top-down scan of all cpusets. This scan loads a pointer
537 * to each cpuset marked is_sched_load_balance into the
538 * array 'csa'. For our purposes, rebuilding the schedulers
539 * sched domains, we can ignore !is_sched_load_balance cpusets.
540 * csa - (for CpuSet Array) Array of pointers to all the cpusets
541 * that need to be load balanced, for convenient iterative
542 * access by the subsequent code that finds the best partition,
543 * i.e the set of domains (subsets) of CPUs such that the
544 * cpus_allowed of every cpuset marked is_sched_load_balance
545 * is a subset of one of these domains, while there are as
546 * many such domains as possible, each as small as possible.
547 * doms - Conversion of 'csa' to an array of cpumasks, for passing to
548 * the kernel/sched.c routine partition_sched_domains() in a
549 * convenient format, that can be easily compared to the prior
550 * value to determine what partition elements (sched domains)
551 * were changed (added or removed.)
553 * Finding the best partition (set of domains):
554 * The triple nested loops below over i, j, k scan over the
555 * load balanced cpusets (using the array of cpuset pointers in
556 * csa[]) looking for pairs of cpusets that have overlapping
557 * cpus_allowed, but which don't have the same 'pn' partition
558 * number and gives them in the same partition number. It keeps
559 * looping on the 'restart' label until it can no longer find
560 * any such pairs.
562 * The union of the cpus_allowed masks from the set of
563 * all cpusets having the same 'pn' value then form the one
564 * element of the partition (one sched domain) to be passed to
565 * partition_sched_domains().
568 void rebuild_sched_domains(void)
570 struct kfifo *q; /* queue of cpusets to be scanned */
571 struct cpuset *cp; /* scans q */
572 struct cpuset **csa; /* array of all cpuset ptrs */
573 int csn; /* how many cpuset ptrs in csa so far */
574 int i, j, k; /* indices for partition finding loops */
575 cpumask_t *doms; /* resulting partition; i.e. sched domains */
576 struct sched_domain_attr *dattr; /* attributes for custom domains */
577 int ndoms; /* number of sched domains in result */
578 int nslot; /* next empty doms[] cpumask_t slot */
580 q = NULL;
581 csa = NULL;
582 doms = NULL;
583 dattr = NULL;
585 /* Special case for the 99% of systems with one, full, sched domain */
586 if (is_sched_load_balance(&top_cpuset)) {
587 ndoms = 1;
588 doms = kmalloc(sizeof(cpumask_t), GFP_KERNEL);
589 if (!doms)
590 goto rebuild;
591 dattr = kmalloc(sizeof(struct sched_domain_attr), GFP_KERNEL);
592 if (dattr) {
593 *dattr = SD_ATTR_INIT;
594 update_domain_attr(dattr, &top_cpuset);
596 *doms = top_cpuset.cpus_allowed;
597 goto rebuild;
600 q = kfifo_alloc(number_of_cpusets * sizeof(cp), GFP_KERNEL, NULL);
601 if (IS_ERR(q))
602 goto done;
603 csa = kmalloc(number_of_cpusets * sizeof(cp), GFP_KERNEL);
604 if (!csa)
605 goto done;
606 csn = 0;
608 cp = &top_cpuset;
609 __kfifo_put(q, (void *)&cp, sizeof(cp));
610 while (__kfifo_get(q, (void *)&cp, sizeof(cp))) {
611 struct cgroup *cont;
612 struct cpuset *child; /* scans child cpusets of cp */
614 if (cpus_empty(cp->cpus_allowed))
615 continue;
617 if (is_sched_load_balance(cp))
618 csa[csn++] = cp;
620 list_for_each_entry(cont, &cp->css.cgroup->children, sibling) {
621 child = cgroup_cs(cont);
622 __kfifo_put(q, (void *)&child, sizeof(cp));
626 for (i = 0; i < csn; i++)
627 csa[i]->pn = i;
628 ndoms = csn;
630 restart:
631 /* Find the best partition (set of sched domains) */
632 for (i = 0; i < csn; i++) {
633 struct cpuset *a = csa[i];
634 int apn = a->pn;
636 for (j = 0; j < csn; j++) {
637 struct cpuset *b = csa[j];
638 int bpn = b->pn;
640 if (apn != bpn && cpusets_overlap(a, b)) {
641 for (k = 0; k < csn; k++) {
642 struct cpuset *c = csa[k];
644 if (c->pn == bpn)
645 c->pn = apn;
647 ndoms--; /* one less element */
648 goto restart;
653 /* Convert <csn, csa> to <ndoms, doms> */
654 doms = kmalloc(ndoms * sizeof(cpumask_t), GFP_KERNEL);
655 if (!doms)
656 goto rebuild;
657 dattr = kmalloc(ndoms * sizeof(struct sched_domain_attr), GFP_KERNEL);
659 for (nslot = 0, i = 0; i < csn; i++) {
660 struct cpuset *a = csa[i];
661 int apn = a->pn;
663 if (apn >= 0) {
664 cpumask_t *dp = doms + nslot;
666 if (nslot == ndoms) {
667 static int warnings = 10;
668 if (warnings) {
669 printk(KERN_WARNING
670 "rebuild_sched_domains confused:"
671 " nslot %d, ndoms %d, csn %d, i %d,"
672 " apn %d\n",
673 nslot, ndoms, csn, i, apn);
674 warnings--;
676 continue;
679 cpus_clear(*dp);
680 if (dattr)
681 *(dattr + nslot) = SD_ATTR_INIT;
682 for (j = i; j < csn; j++) {
683 struct cpuset *b = csa[j];
685 if (apn == b->pn) {
686 cpus_or(*dp, *dp, b->cpus_allowed);
687 b->pn = -1;
688 if (dattr)
689 update_domain_attr(dattr
690 + nslot, b);
693 nslot++;
696 BUG_ON(nslot != ndoms);
698 rebuild:
699 /* Have scheduler rebuild sched domains */
700 get_online_cpus();
701 partition_sched_domains(ndoms, doms, dattr);
702 put_online_cpus();
704 done:
705 if (q && !IS_ERR(q))
706 kfifo_free(q);
707 kfree(csa);
708 /* Don't kfree(doms) -- partition_sched_domains() does that. */
709 /* Don't kfree(dattr) -- partition_sched_domains() does that. */
713 * cpuset_test_cpumask - test a task's cpus_allowed versus its cpuset's
714 * @tsk: task to test
715 * @scan: struct cgroup_scanner contained in its struct cpuset_hotplug_scanner
717 * Call with cgroup_mutex held. May take callback_mutex during call.
718 * Called for each task in a cgroup by cgroup_scan_tasks().
719 * Return nonzero if this tasks's cpus_allowed mask should be changed (in other
720 * words, if its mask is not equal to its cpuset's mask).
722 static int cpuset_test_cpumask(struct task_struct *tsk,
723 struct cgroup_scanner *scan)
725 return !cpus_equal(tsk->cpus_allowed,
726 (cgroup_cs(scan->cg))->cpus_allowed);
730 * cpuset_change_cpumask - make a task's cpus_allowed the same as its cpuset's
731 * @tsk: task to test
732 * @scan: struct cgroup_scanner containing the cgroup of the task
734 * Called by cgroup_scan_tasks() for each task in a cgroup whose
735 * cpus_allowed mask needs to be changed.
737 * We don't need to re-check for the cgroup/cpuset membership, since we're
738 * holding cgroup_lock() at this point.
740 static void cpuset_change_cpumask(struct task_struct *tsk,
741 struct cgroup_scanner *scan)
743 set_cpus_allowed_ptr(tsk, &((cgroup_cs(scan->cg))->cpus_allowed));
747 * update_tasks_cpumask - Update the cpumasks of tasks in the cpuset.
748 * @cs: the cpuset in which each task's cpus_allowed mask needs to be changed
750 * Called with cgroup_mutex held
752 * The cgroup_scan_tasks() function will scan all the tasks in a cgroup,
753 * calling callback functions for each.
755 * Return 0 if successful, -errno if not.
757 static int update_tasks_cpumask(struct cpuset *cs)
759 struct cgroup_scanner scan;
760 struct ptr_heap heap;
761 int retval;
764 * cgroup_scan_tasks() will initialize heap->gt for us.
765 * heap_init() is still needed here for we should not change
766 * cs->cpus_allowed when heap_init() fails.
768 retval = heap_init(&heap, PAGE_SIZE, GFP_KERNEL, NULL);
769 if (retval)
770 return retval;
772 scan.cg = cs->css.cgroup;
773 scan.test_task = cpuset_test_cpumask;
774 scan.process_task = cpuset_change_cpumask;
775 scan.heap = &heap;
776 retval = cgroup_scan_tasks(&scan);
778 heap_free(&heap);
779 return retval;
783 * update_cpumask - update the cpus_allowed mask of a cpuset and all tasks in it
784 * @cs: the cpuset to consider
785 * @buf: buffer of cpu numbers written to this cpuset
787 static int update_cpumask(struct cpuset *cs, const char *buf)
789 struct cpuset trialcs;
790 int retval;
791 int is_load_balanced;
793 /* top_cpuset.cpus_allowed tracks cpu_online_map; it's read-only */
794 if (cs == &top_cpuset)
795 return -EACCES;
797 trialcs = *cs;
800 * An empty cpus_allowed is ok only if the cpuset has no tasks.
801 * Since cpulist_parse() fails on an empty mask, we special case
802 * that parsing. The validate_change() call ensures that cpusets
803 * with tasks have cpus.
805 if (!*buf) {
806 cpus_clear(trialcs.cpus_allowed);
807 } else {
808 retval = cpulist_parse(buf, trialcs.cpus_allowed);
809 if (retval < 0)
810 return retval;
812 if (!cpus_subset(trialcs.cpus_allowed, cpu_online_map))
813 return -EINVAL;
815 retval = validate_change(cs, &trialcs);
816 if (retval < 0)
817 return retval;
819 /* Nothing to do if the cpus didn't change */
820 if (cpus_equal(cs->cpus_allowed, trialcs.cpus_allowed))
821 return 0;
823 is_load_balanced = is_sched_load_balance(&trialcs);
825 mutex_lock(&callback_mutex);
826 cs->cpus_allowed = trialcs.cpus_allowed;
827 mutex_unlock(&callback_mutex);
830 * Scan tasks in the cpuset, and update the cpumasks of any
831 * that need an update.
833 retval = update_tasks_cpumask(cs);
834 if (retval < 0)
835 return retval;
837 if (is_load_balanced)
838 rebuild_sched_domains();
839 return 0;
843 * cpuset_migrate_mm
845 * Migrate memory region from one set of nodes to another.
847 * Temporarilly set tasks mems_allowed to target nodes of migration,
848 * so that the migration code can allocate pages on these nodes.
850 * Call holding cgroup_mutex, so current's cpuset won't change
851 * during this call, as manage_mutex holds off any cpuset_attach()
852 * calls. Therefore we don't need to take task_lock around the
853 * call to guarantee_online_mems(), as we know no one is changing
854 * our task's cpuset.
856 * Hold callback_mutex around the two modifications of our tasks
857 * mems_allowed to synchronize with cpuset_mems_allowed().
859 * While the mm_struct we are migrating is typically from some
860 * other task, the task_struct mems_allowed that we are hacking
861 * is for our current task, which must allocate new pages for that
862 * migrating memory region.
864 * We call cpuset_update_task_memory_state() before hacking
865 * our tasks mems_allowed, so that we are assured of being in
866 * sync with our tasks cpuset, and in particular, callbacks to
867 * cpuset_update_task_memory_state() from nested page allocations
868 * won't see any mismatch of our cpuset and task mems_generation
869 * values, so won't overwrite our hacked tasks mems_allowed
870 * nodemask.
873 static void cpuset_migrate_mm(struct mm_struct *mm, const nodemask_t *from,
874 const nodemask_t *to)
876 struct task_struct *tsk = current;
878 cpuset_update_task_memory_state();
880 mutex_lock(&callback_mutex);
881 tsk->mems_allowed = *to;
882 mutex_unlock(&callback_mutex);
884 do_migrate_pages(mm, from, to, MPOL_MF_MOVE_ALL);
886 mutex_lock(&callback_mutex);
887 guarantee_online_mems(task_cs(tsk),&tsk->mems_allowed);
888 mutex_unlock(&callback_mutex);
891 static void *cpuset_being_rebound;
894 * update_tasks_nodemask - Update the nodemasks of tasks in the cpuset.
895 * @cs: the cpuset in which each task's mems_allowed mask needs to be changed
896 * @oldmem: old mems_allowed of cpuset cs
898 * Called with cgroup_mutex held
899 * Return 0 if successful, -errno if not.
901 static int update_tasks_nodemask(struct cpuset *cs, const nodemask_t *oldmem)
903 struct task_struct *p;
904 struct mm_struct **mmarray;
905 int i, n, ntasks;
906 int migrate;
907 int fudge;
908 struct cgroup_iter it;
909 int retval;
911 cpuset_being_rebound = cs; /* causes mpol_dup() rebind */
913 fudge = 10; /* spare mmarray[] slots */
914 fudge += cpus_weight(cs->cpus_allowed); /* imagine one fork-bomb/cpu */
915 retval = -ENOMEM;
918 * Allocate mmarray[] to hold mm reference for each task
919 * in cpuset cs. Can't kmalloc GFP_KERNEL while holding
920 * tasklist_lock. We could use GFP_ATOMIC, but with a
921 * few more lines of code, we can retry until we get a big
922 * enough mmarray[] w/o using GFP_ATOMIC.
924 while (1) {
925 ntasks = cgroup_task_count(cs->css.cgroup); /* guess */
926 ntasks += fudge;
927 mmarray = kmalloc(ntasks * sizeof(*mmarray), GFP_KERNEL);
928 if (!mmarray)
929 goto done;
930 read_lock(&tasklist_lock); /* block fork */
931 if (cgroup_task_count(cs->css.cgroup) <= ntasks)
932 break; /* got enough */
933 read_unlock(&tasklist_lock); /* try again */
934 kfree(mmarray);
937 n = 0;
939 /* Load up mmarray[] with mm reference for each task in cpuset. */
940 cgroup_iter_start(cs->css.cgroup, &it);
941 while ((p = cgroup_iter_next(cs->css.cgroup, &it))) {
942 struct mm_struct *mm;
944 if (n >= ntasks) {
945 printk(KERN_WARNING
946 "Cpuset mempolicy rebind incomplete.\n");
947 break;
949 mm = get_task_mm(p);
950 if (!mm)
951 continue;
952 mmarray[n++] = mm;
954 cgroup_iter_end(cs->css.cgroup, &it);
955 read_unlock(&tasklist_lock);
958 * Now that we've dropped the tasklist spinlock, we can
959 * rebind the vma mempolicies of each mm in mmarray[] to their
960 * new cpuset, and release that mm. The mpol_rebind_mm()
961 * call takes mmap_sem, which we couldn't take while holding
962 * tasklist_lock. Forks can happen again now - the mpol_dup()
963 * cpuset_being_rebound check will catch such forks, and rebind
964 * their vma mempolicies too. Because we still hold the global
965 * cgroup_mutex, we know that no other rebind effort will
966 * be contending for the global variable cpuset_being_rebound.
967 * It's ok if we rebind the same mm twice; mpol_rebind_mm()
968 * is idempotent. Also migrate pages in each mm to new nodes.
970 migrate = is_memory_migrate(cs);
971 for (i = 0; i < n; i++) {
972 struct mm_struct *mm = mmarray[i];
974 mpol_rebind_mm(mm, &cs->mems_allowed);
975 if (migrate)
976 cpuset_migrate_mm(mm, oldmem, &cs->mems_allowed);
977 mmput(mm);
980 /* We're done rebinding vmas to this cpuset's new mems_allowed. */
981 kfree(mmarray);
982 cpuset_being_rebound = NULL;
983 retval = 0;
984 done:
985 return retval;
989 * Handle user request to change the 'mems' memory placement
990 * of a cpuset. Needs to validate the request, update the
991 * cpusets mems_allowed and mems_generation, and for each
992 * task in the cpuset, rebind any vma mempolicies and if
993 * the cpuset is marked 'memory_migrate', migrate the tasks
994 * pages to the new memory.
996 * Call with cgroup_mutex held. May take callback_mutex during call.
997 * Will take tasklist_lock, scan tasklist for tasks in cpuset cs,
998 * lock each such tasks mm->mmap_sem, scan its vma's and rebind
999 * their mempolicies to the cpusets new mems_allowed.
1001 static int update_nodemask(struct cpuset *cs, const char *buf)
1003 struct cpuset trialcs;
1004 nodemask_t oldmem;
1005 int retval;
1008 * top_cpuset.mems_allowed tracks node_stats[N_HIGH_MEMORY];
1009 * it's read-only
1011 if (cs == &top_cpuset)
1012 return -EACCES;
1014 trialcs = *cs;
1017 * An empty mems_allowed is ok iff there are no tasks in the cpuset.
1018 * Since nodelist_parse() fails on an empty mask, we special case
1019 * that parsing. The validate_change() call ensures that cpusets
1020 * with tasks have memory.
1022 if (!*buf) {
1023 nodes_clear(trialcs.mems_allowed);
1024 } else {
1025 retval = nodelist_parse(buf, trialcs.mems_allowed);
1026 if (retval < 0)
1027 goto done;
1029 if (!nodes_subset(trialcs.mems_allowed,
1030 node_states[N_HIGH_MEMORY]))
1031 return -EINVAL;
1033 oldmem = cs->mems_allowed;
1034 if (nodes_equal(oldmem, trialcs.mems_allowed)) {
1035 retval = 0; /* Too easy - nothing to do */
1036 goto done;
1038 retval = validate_change(cs, &trialcs);
1039 if (retval < 0)
1040 goto done;
1042 mutex_lock(&callback_mutex);
1043 cs->mems_allowed = trialcs.mems_allowed;
1044 cs->mems_generation = cpuset_mems_generation++;
1045 mutex_unlock(&callback_mutex);
1047 retval = update_tasks_nodemask(cs, &oldmem);
1048 done:
1049 return retval;
1052 int current_cpuset_is_being_rebound(void)
1054 return task_cs(current) == cpuset_being_rebound;
1057 static int update_relax_domain_level(struct cpuset *cs, s64 val)
1059 if (val < -1 || val >= SD_LV_MAX)
1060 return -EINVAL;
1062 if (val != cs->relax_domain_level) {
1063 cs->relax_domain_level = val;
1064 if (!cpus_empty(cs->cpus_allowed) && is_sched_load_balance(cs))
1065 rebuild_sched_domains();
1068 return 0;
1072 * update_flag - read a 0 or a 1 in a file and update associated flag
1073 * bit: the bit to update (see cpuset_flagbits_t)
1074 * cs: the cpuset to update
1075 * turning_on: whether the flag is being set or cleared
1077 * Call with cgroup_mutex held.
1080 static int update_flag(cpuset_flagbits_t bit, struct cpuset *cs,
1081 int turning_on)
1083 struct cpuset trialcs;
1084 int err;
1085 int cpus_nonempty, balance_flag_changed;
1087 trialcs = *cs;
1088 if (turning_on)
1089 set_bit(bit, &trialcs.flags);
1090 else
1091 clear_bit(bit, &trialcs.flags);
1093 err = validate_change(cs, &trialcs);
1094 if (err < 0)
1095 return err;
1097 cpus_nonempty = !cpus_empty(trialcs.cpus_allowed);
1098 balance_flag_changed = (is_sched_load_balance(cs) !=
1099 is_sched_load_balance(&trialcs));
1101 mutex_lock(&callback_mutex);
1102 cs->flags = trialcs.flags;
1103 mutex_unlock(&callback_mutex);
1105 if (cpus_nonempty && balance_flag_changed)
1106 rebuild_sched_domains();
1108 return 0;
1112 * Frequency meter - How fast is some event occurring?
1114 * These routines manage a digitally filtered, constant time based,
1115 * event frequency meter. There are four routines:
1116 * fmeter_init() - initialize a frequency meter.
1117 * fmeter_markevent() - called each time the event happens.
1118 * fmeter_getrate() - returns the recent rate of such events.
1119 * fmeter_update() - internal routine used to update fmeter.
1121 * A common data structure is passed to each of these routines,
1122 * which is used to keep track of the state required to manage the
1123 * frequency meter and its digital filter.
1125 * The filter works on the number of events marked per unit time.
1126 * The filter is single-pole low-pass recursive (IIR). The time unit
1127 * is 1 second. Arithmetic is done using 32-bit integers scaled to
1128 * simulate 3 decimal digits of precision (multiplied by 1000).
1130 * With an FM_COEF of 933, and a time base of 1 second, the filter
1131 * has a half-life of 10 seconds, meaning that if the events quit
1132 * happening, then the rate returned from the fmeter_getrate()
1133 * will be cut in half each 10 seconds, until it converges to zero.
1135 * It is not worth doing a real infinitely recursive filter. If more
1136 * than FM_MAXTICKS ticks have elapsed since the last filter event,
1137 * just compute FM_MAXTICKS ticks worth, by which point the level
1138 * will be stable.
1140 * Limit the count of unprocessed events to FM_MAXCNT, so as to avoid
1141 * arithmetic overflow in the fmeter_update() routine.
1143 * Given the simple 32 bit integer arithmetic used, this meter works
1144 * best for reporting rates between one per millisecond (msec) and
1145 * one per 32 (approx) seconds. At constant rates faster than one
1146 * per msec it maxes out at values just under 1,000,000. At constant
1147 * rates between one per msec, and one per second it will stabilize
1148 * to a value N*1000, where N is the rate of events per second.
1149 * At constant rates between one per second and one per 32 seconds,
1150 * it will be choppy, moving up on the seconds that have an event,
1151 * and then decaying until the next event. At rates slower than
1152 * about one in 32 seconds, it decays all the way back to zero between
1153 * each event.
1156 #define FM_COEF 933 /* coefficient for half-life of 10 secs */
1157 #define FM_MAXTICKS ((time_t)99) /* useless computing more ticks than this */
1158 #define FM_MAXCNT 1000000 /* limit cnt to avoid overflow */
1159 #define FM_SCALE 1000 /* faux fixed point scale */
1161 /* Initialize a frequency meter */
1162 static void fmeter_init(struct fmeter *fmp)
1164 fmp->cnt = 0;
1165 fmp->val = 0;
1166 fmp->time = 0;
1167 spin_lock_init(&fmp->lock);
1170 /* Internal meter update - process cnt events and update value */
1171 static void fmeter_update(struct fmeter *fmp)
1173 time_t now = get_seconds();
1174 time_t ticks = now - fmp->time;
1176 if (ticks == 0)
1177 return;
1179 ticks = min(FM_MAXTICKS, ticks);
1180 while (ticks-- > 0)
1181 fmp->val = (FM_COEF * fmp->val) / FM_SCALE;
1182 fmp->time = now;
1184 fmp->val += ((FM_SCALE - FM_COEF) * fmp->cnt) / FM_SCALE;
1185 fmp->cnt = 0;
1188 /* Process any previous ticks, then bump cnt by one (times scale). */
1189 static void fmeter_markevent(struct fmeter *fmp)
1191 spin_lock(&fmp->lock);
1192 fmeter_update(fmp);
1193 fmp->cnt = min(FM_MAXCNT, fmp->cnt + FM_SCALE);
1194 spin_unlock(&fmp->lock);
1197 /* Process any previous ticks, then return current value. */
1198 static int fmeter_getrate(struct fmeter *fmp)
1200 int val;
1202 spin_lock(&fmp->lock);
1203 fmeter_update(fmp);
1204 val = fmp->val;
1205 spin_unlock(&fmp->lock);
1206 return val;
1209 /* Called by cgroups to determine if a cpuset is usable; cgroup_mutex held */
1210 static int cpuset_can_attach(struct cgroup_subsys *ss,
1211 struct cgroup *cont, struct task_struct *tsk)
1213 struct cpuset *cs = cgroup_cs(cont);
1215 if (cpus_empty(cs->cpus_allowed) || nodes_empty(cs->mems_allowed))
1216 return -ENOSPC;
1217 if (tsk->flags & PF_THREAD_BOUND) {
1218 cpumask_t mask;
1220 mutex_lock(&callback_mutex);
1221 mask = cs->cpus_allowed;
1222 mutex_unlock(&callback_mutex);
1223 if (!cpus_equal(tsk->cpus_allowed, mask))
1224 return -EINVAL;
1227 return security_task_setscheduler(tsk, 0, NULL);
1230 static void cpuset_attach(struct cgroup_subsys *ss,
1231 struct cgroup *cont, struct cgroup *oldcont,
1232 struct task_struct *tsk)
1234 cpumask_t cpus;
1235 nodemask_t from, to;
1236 struct mm_struct *mm;
1237 struct cpuset *cs = cgroup_cs(cont);
1238 struct cpuset *oldcs = cgroup_cs(oldcont);
1239 int err;
1241 mutex_lock(&callback_mutex);
1242 guarantee_online_cpus(cs, &cpus);
1243 err = set_cpus_allowed_ptr(tsk, &cpus);
1244 mutex_unlock(&callback_mutex);
1245 if (err)
1246 return;
1248 from = oldcs->mems_allowed;
1249 to = cs->mems_allowed;
1250 mm = get_task_mm(tsk);
1251 if (mm) {
1252 mpol_rebind_mm(mm, &to);
1253 if (is_memory_migrate(cs))
1254 cpuset_migrate_mm(mm, &from, &to);
1255 mmput(mm);
1260 /* The various types of files and directories in a cpuset file system */
1262 typedef enum {
1263 FILE_MEMORY_MIGRATE,
1264 FILE_CPULIST,
1265 FILE_MEMLIST,
1266 FILE_CPU_EXCLUSIVE,
1267 FILE_MEM_EXCLUSIVE,
1268 FILE_MEM_HARDWALL,
1269 FILE_SCHED_LOAD_BALANCE,
1270 FILE_SCHED_RELAX_DOMAIN_LEVEL,
1271 FILE_MEMORY_PRESSURE_ENABLED,
1272 FILE_MEMORY_PRESSURE,
1273 FILE_SPREAD_PAGE,
1274 FILE_SPREAD_SLAB,
1275 } cpuset_filetype_t;
1277 static int cpuset_write_u64(struct cgroup *cgrp, struct cftype *cft, u64 val)
1279 int retval = 0;
1280 struct cpuset *cs = cgroup_cs(cgrp);
1281 cpuset_filetype_t type = cft->private;
1283 if (!cgroup_lock_live_group(cgrp))
1284 return -ENODEV;
1286 switch (type) {
1287 case FILE_CPU_EXCLUSIVE:
1288 retval = update_flag(CS_CPU_EXCLUSIVE, cs, val);
1289 break;
1290 case FILE_MEM_EXCLUSIVE:
1291 retval = update_flag(CS_MEM_EXCLUSIVE, cs, val);
1292 break;
1293 case FILE_MEM_HARDWALL:
1294 retval = update_flag(CS_MEM_HARDWALL, cs, val);
1295 break;
1296 case FILE_SCHED_LOAD_BALANCE:
1297 retval = update_flag(CS_SCHED_LOAD_BALANCE, cs, val);
1298 break;
1299 case FILE_MEMORY_MIGRATE:
1300 retval = update_flag(CS_MEMORY_MIGRATE, cs, val);
1301 break;
1302 case FILE_MEMORY_PRESSURE_ENABLED:
1303 cpuset_memory_pressure_enabled = !!val;
1304 break;
1305 case FILE_MEMORY_PRESSURE:
1306 retval = -EACCES;
1307 break;
1308 case FILE_SPREAD_PAGE:
1309 retval = update_flag(CS_SPREAD_PAGE, cs, val);
1310 cs->mems_generation = cpuset_mems_generation++;
1311 break;
1312 case FILE_SPREAD_SLAB:
1313 retval = update_flag(CS_SPREAD_SLAB, cs, val);
1314 cs->mems_generation = cpuset_mems_generation++;
1315 break;
1316 default:
1317 retval = -EINVAL;
1318 break;
1320 cgroup_unlock();
1321 return retval;
1324 static int cpuset_write_s64(struct cgroup *cgrp, struct cftype *cft, s64 val)
1326 int retval = 0;
1327 struct cpuset *cs = cgroup_cs(cgrp);
1328 cpuset_filetype_t type = cft->private;
1330 if (!cgroup_lock_live_group(cgrp))
1331 return -ENODEV;
1333 switch (type) {
1334 case FILE_SCHED_RELAX_DOMAIN_LEVEL:
1335 retval = update_relax_domain_level(cs, val);
1336 break;
1337 default:
1338 retval = -EINVAL;
1339 break;
1341 cgroup_unlock();
1342 return retval;
1346 * Common handling for a write to a "cpus" or "mems" file.
1348 static int cpuset_write_resmask(struct cgroup *cgrp, struct cftype *cft,
1349 const char *buf)
1351 int retval = 0;
1353 if (!cgroup_lock_live_group(cgrp))
1354 return -ENODEV;
1356 switch (cft->private) {
1357 case FILE_CPULIST:
1358 retval = update_cpumask(cgroup_cs(cgrp), buf);
1359 break;
1360 case FILE_MEMLIST:
1361 retval = update_nodemask(cgroup_cs(cgrp), buf);
1362 break;
1363 default:
1364 retval = -EINVAL;
1365 break;
1367 cgroup_unlock();
1368 return retval;
1372 * These ascii lists should be read in a single call, by using a user
1373 * buffer large enough to hold the entire map. If read in smaller
1374 * chunks, there is no guarantee of atomicity. Since the display format
1375 * used, list of ranges of sequential numbers, is variable length,
1376 * and since these maps can change value dynamically, one could read
1377 * gibberish by doing partial reads while a list was changing.
1378 * A single large read to a buffer that crosses a page boundary is
1379 * ok, because the result being copied to user land is not recomputed
1380 * across a page fault.
1383 static int cpuset_sprintf_cpulist(char *page, struct cpuset *cs)
1385 cpumask_t mask;
1387 mutex_lock(&callback_mutex);
1388 mask = cs->cpus_allowed;
1389 mutex_unlock(&callback_mutex);
1391 return cpulist_scnprintf(page, PAGE_SIZE, mask);
1394 static int cpuset_sprintf_memlist(char *page, struct cpuset *cs)
1396 nodemask_t mask;
1398 mutex_lock(&callback_mutex);
1399 mask = cs->mems_allowed;
1400 mutex_unlock(&callback_mutex);
1402 return nodelist_scnprintf(page, PAGE_SIZE, mask);
1405 static ssize_t cpuset_common_file_read(struct cgroup *cont,
1406 struct cftype *cft,
1407 struct file *file,
1408 char __user *buf,
1409 size_t nbytes, loff_t *ppos)
1411 struct cpuset *cs = cgroup_cs(cont);
1412 cpuset_filetype_t type = cft->private;
1413 char *page;
1414 ssize_t retval = 0;
1415 char *s;
1417 if (!(page = (char *)__get_free_page(GFP_TEMPORARY)))
1418 return -ENOMEM;
1420 s = page;
1422 switch (type) {
1423 case FILE_CPULIST:
1424 s += cpuset_sprintf_cpulist(s, cs);
1425 break;
1426 case FILE_MEMLIST:
1427 s += cpuset_sprintf_memlist(s, cs);
1428 break;
1429 default:
1430 retval = -EINVAL;
1431 goto out;
1433 *s++ = '\n';
1435 retval = simple_read_from_buffer(buf, nbytes, ppos, page, s - page);
1436 out:
1437 free_page((unsigned long)page);
1438 return retval;
1441 static u64 cpuset_read_u64(struct cgroup *cont, struct cftype *cft)
1443 struct cpuset *cs = cgroup_cs(cont);
1444 cpuset_filetype_t type = cft->private;
1445 switch (type) {
1446 case FILE_CPU_EXCLUSIVE:
1447 return is_cpu_exclusive(cs);
1448 case FILE_MEM_EXCLUSIVE:
1449 return is_mem_exclusive(cs);
1450 case FILE_MEM_HARDWALL:
1451 return is_mem_hardwall(cs);
1452 case FILE_SCHED_LOAD_BALANCE:
1453 return is_sched_load_balance(cs);
1454 case FILE_MEMORY_MIGRATE:
1455 return is_memory_migrate(cs);
1456 case FILE_MEMORY_PRESSURE_ENABLED:
1457 return cpuset_memory_pressure_enabled;
1458 case FILE_MEMORY_PRESSURE:
1459 return fmeter_getrate(&cs->fmeter);
1460 case FILE_SPREAD_PAGE:
1461 return is_spread_page(cs);
1462 case FILE_SPREAD_SLAB:
1463 return is_spread_slab(cs);
1464 default:
1465 BUG();
1469 static s64 cpuset_read_s64(struct cgroup *cont, struct cftype *cft)
1471 struct cpuset *cs = cgroup_cs(cont);
1472 cpuset_filetype_t type = cft->private;
1473 switch (type) {
1474 case FILE_SCHED_RELAX_DOMAIN_LEVEL:
1475 return cs->relax_domain_level;
1476 default:
1477 BUG();
1483 * for the common functions, 'private' gives the type of file
1486 static struct cftype files[] = {
1488 .name = "cpus",
1489 .read = cpuset_common_file_read,
1490 .write_string = cpuset_write_resmask,
1491 .max_write_len = (100U + 6 * NR_CPUS),
1492 .private = FILE_CPULIST,
1496 .name = "mems",
1497 .read = cpuset_common_file_read,
1498 .write_string = cpuset_write_resmask,
1499 .max_write_len = (100U + 6 * MAX_NUMNODES),
1500 .private = FILE_MEMLIST,
1504 .name = "cpu_exclusive",
1505 .read_u64 = cpuset_read_u64,
1506 .write_u64 = cpuset_write_u64,
1507 .private = FILE_CPU_EXCLUSIVE,
1511 .name = "mem_exclusive",
1512 .read_u64 = cpuset_read_u64,
1513 .write_u64 = cpuset_write_u64,
1514 .private = FILE_MEM_EXCLUSIVE,
1518 .name = "mem_hardwall",
1519 .read_u64 = cpuset_read_u64,
1520 .write_u64 = cpuset_write_u64,
1521 .private = FILE_MEM_HARDWALL,
1525 .name = "sched_load_balance",
1526 .read_u64 = cpuset_read_u64,
1527 .write_u64 = cpuset_write_u64,
1528 .private = FILE_SCHED_LOAD_BALANCE,
1532 .name = "sched_relax_domain_level",
1533 .read_s64 = cpuset_read_s64,
1534 .write_s64 = cpuset_write_s64,
1535 .private = FILE_SCHED_RELAX_DOMAIN_LEVEL,
1539 .name = "memory_migrate",
1540 .read_u64 = cpuset_read_u64,
1541 .write_u64 = cpuset_write_u64,
1542 .private = FILE_MEMORY_MIGRATE,
1546 .name = "memory_pressure",
1547 .read_u64 = cpuset_read_u64,
1548 .write_u64 = cpuset_write_u64,
1549 .private = FILE_MEMORY_PRESSURE,
1553 .name = "memory_spread_page",
1554 .read_u64 = cpuset_read_u64,
1555 .write_u64 = cpuset_write_u64,
1556 .private = FILE_SPREAD_PAGE,
1560 .name = "memory_spread_slab",
1561 .read_u64 = cpuset_read_u64,
1562 .write_u64 = cpuset_write_u64,
1563 .private = FILE_SPREAD_SLAB,
1567 static struct cftype cft_memory_pressure_enabled = {
1568 .name = "memory_pressure_enabled",
1569 .read_u64 = cpuset_read_u64,
1570 .write_u64 = cpuset_write_u64,
1571 .private = FILE_MEMORY_PRESSURE_ENABLED,
1574 static int cpuset_populate(struct cgroup_subsys *ss, struct cgroup *cont)
1576 int err;
1578 err = cgroup_add_files(cont, ss, files, ARRAY_SIZE(files));
1579 if (err)
1580 return err;
1581 /* memory_pressure_enabled is in root cpuset only */
1582 if (!cont->parent)
1583 err = cgroup_add_file(cont, ss,
1584 &cft_memory_pressure_enabled);
1585 return err;
1589 * post_clone() is called at the end of cgroup_clone().
1590 * 'cgroup' was just created automatically as a result of
1591 * a cgroup_clone(), and the current task is about to
1592 * be moved into 'cgroup'.
1594 * Currently we refuse to set up the cgroup - thereby
1595 * refusing the task to be entered, and as a result refusing
1596 * the sys_unshare() or clone() which initiated it - if any
1597 * sibling cpusets have exclusive cpus or mem.
1599 * If this becomes a problem for some users who wish to
1600 * allow that scenario, then cpuset_post_clone() could be
1601 * changed to grant parent->cpus_allowed-sibling_cpus_exclusive
1602 * (and likewise for mems) to the new cgroup. Called with cgroup_mutex
1603 * held.
1605 static void cpuset_post_clone(struct cgroup_subsys *ss,
1606 struct cgroup *cgroup)
1608 struct cgroup *parent, *child;
1609 struct cpuset *cs, *parent_cs;
1611 parent = cgroup->parent;
1612 list_for_each_entry(child, &parent->children, sibling) {
1613 cs = cgroup_cs(child);
1614 if (is_mem_exclusive(cs) || is_cpu_exclusive(cs))
1615 return;
1617 cs = cgroup_cs(cgroup);
1618 parent_cs = cgroup_cs(parent);
1620 cs->mems_allowed = parent_cs->mems_allowed;
1621 cs->cpus_allowed = parent_cs->cpus_allowed;
1622 return;
1626 * cpuset_create - create a cpuset
1627 * ss: cpuset cgroup subsystem
1628 * cont: control group that the new cpuset will be part of
1631 static struct cgroup_subsys_state *cpuset_create(
1632 struct cgroup_subsys *ss,
1633 struct cgroup *cont)
1635 struct cpuset *cs;
1636 struct cpuset *parent;
1638 if (!cont->parent) {
1639 /* This is early initialization for the top cgroup */
1640 top_cpuset.mems_generation = cpuset_mems_generation++;
1641 return &top_cpuset.css;
1643 parent = cgroup_cs(cont->parent);
1644 cs = kmalloc(sizeof(*cs), GFP_KERNEL);
1645 if (!cs)
1646 return ERR_PTR(-ENOMEM);
1648 cpuset_update_task_memory_state();
1649 cs->flags = 0;
1650 if (is_spread_page(parent))
1651 set_bit(CS_SPREAD_PAGE, &cs->flags);
1652 if (is_spread_slab(parent))
1653 set_bit(CS_SPREAD_SLAB, &cs->flags);
1654 set_bit(CS_SCHED_LOAD_BALANCE, &cs->flags);
1655 cpus_clear(cs->cpus_allowed);
1656 nodes_clear(cs->mems_allowed);
1657 cs->mems_generation = cpuset_mems_generation++;
1658 fmeter_init(&cs->fmeter);
1659 cs->relax_domain_level = -1;
1661 cs->parent = parent;
1662 number_of_cpusets++;
1663 return &cs->css ;
1667 * Locking note on the strange update_flag() call below:
1669 * If the cpuset being removed has its flag 'sched_load_balance'
1670 * enabled, then simulate turning sched_load_balance off, which
1671 * will call rebuild_sched_domains(). The get_online_cpus()
1672 * call in rebuild_sched_domains() must not be made while holding
1673 * callback_mutex. Elsewhere the kernel nests callback_mutex inside
1674 * get_online_cpus() calls. So the reverse nesting would risk an
1675 * ABBA deadlock.
1678 static void cpuset_destroy(struct cgroup_subsys *ss, struct cgroup *cont)
1680 struct cpuset *cs = cgroup_cs(cont);
1682 cpuset_update_task_memory_state();
1684 if (is_sched_load_balance(cs))
1685 update_flag(CS_SCHED_LOAD_BALANCE, cs, 0);
1687 number_of_cpusets--;
1688 kfree(cs);
1691 struct cgroup_subsys cpuset_subsys = {
1692 .name = "cpuset",
1693 .create = cpuset_create,
1694 .destroy = cpuset_destroy,
1695 .can_attach = cpuset_can_attach,
1696 .attach = cpuset_attach,
1697 .populate = cpuset_populate,
1698 .post_clone = cpuset_post_clone,
1699 .subsys_id = cpuset_subsys_id,
1700 .early_init = 1,
1704 * cpuset_init_early - just enough so that the calls to
1705 * cpuset_update_task_memory_state() in early init code
1706 * are harmless.
1709 int __init cpuset_init_early(void)
1711 top_cpuset.mems_generation = cpuset_mems_generation++;
1712 return 0;
1717 * cpuset_init - initialize cpusets at system boot
1719 * Description: Initialize top_cpuset and the cpuset internal file system,
1722 int __init cpuset_init(void)
1724 int err = 0;
1726 cpus_setall(top_cpuset.cpus_allowed);
1727 nodes_setall(top_cpuset.mems_allowed);
1729 fmeter_init(&top_cpuset.fmeter);
1730 top_cpuset.mems_generation = cpuset_mems_generation++;
1731 set_bit(CS_SCHED_LOAD_BALANCE, &top_cpuset.flags);
1732 top_cpuset.relax_domain_level = -1;
1734 err = register_filesystem(&cpuset_fs_type);
1735 if (err < 0)
1736 return err;
1738 number_of_cpusets = 1;
1739 return 0;
1743 * cpuset_do_move_task - move a given task to another cpuset
1744 * @tsk: pointer to task_struct the task to move
1745 * @scan: struct cgroup_scanner contained in its struct cpuset_hotplug_scanner
1747 * Called by cgroup_scan_tasks() for each task in a cgroup.
1748 * Return nonzero to stop the walk through the tasks.
1750 static void cpuset_do_move_task(struct task_struct *tsk,
1751 struct cgroup_scanner *scan)
1753 struct cpuset_hotplug_scanner *chsp;
1755 chsp = container_of(scan, struct cpuset_hotplug_scanner, scan);
1756 cgroup_attach_task(chsp->to, tsk);
1760 * move_member_tasks_to_cpuset - move tasks from one cpuset to another
1761 * @from: cpuset in which the tasks currently reside
1762 * @to: cpuset to which the tasks will be moved
1764 * Called with cgroup_mutex held
1765 * callback_mutex must not be held, as cpuset_attach() will take it.
1767 * The cgroup_scan_tasks() function will scan all the tasks in a cgroup,
1768 * calling callback functions for each.
1770 static void move_member_tasks_to_cpuset(struct cpuset *from, struct cpuset *to)
1772 struct cpuset_hotplug_scanner scan;
1774 scan.scan.cg = from->css.cgroup;
1775 scan.scan.test_task = NULL; /* select all tasks in cgroup */
1776 scan.scan.process_task = cpuset_do_move_task;
1777 scan.scan.heap = NULL;
1778 scan.to = to->css.cgroup;
1780 if (cgroup_scan_tasks(&scan.scan))
1781 printk(KERN_ERR "move_member_tasks_to_cpuset: "
1782 "cgroup_scan_tasks failed\n");
1786 * If common_cpu_mem_hotplug_unplug(), below, unplugs any CPUs
1787 * or memory nodes, we need to walk over the cpuset hierarchy,
1788 * removing that CPU or node from all cpusets. If this removes the
1789 * last CPU or node from a cpuset, then move the tasks in the empty
1790 * cpuset to its next-highest non-empty parent.
1792 * Called with cgroup_mutex held
1793 * callback_mutex must not be held, as cpuset_attach() will take it.
1795 static void remove_tasks_in_empty_cpuset(struct cpuset *cs)
1797 struct cpuset *parent;
1800 * The cgroup's css_sets list is in use if there are tasks
1801 * in the cpuset; the list is empty if there are none;
1802 * the cs->css.refcnt seems always 0.
1804 if (list_empty(&cs->css.cgroup->css_sets))
1805 return;
1808 * Find its next-highest non-empty parent, (top cpuset
1809 * has online cpus, so can't be empty).
1811 parent = cs->parent;
1812 while (cpus_empty(parent->cpus_allowed) ||
1813 nodes_empty(parent->mems_allowed))
1814 parent = parent->parent;
1816 move_member_tasks_to_cpuset(cs, parent);
1820 * Walk the specified cpuset subtree and look for empty cpusets.
1821 * The tasks of such cpuset must be moved to a parent cpuset.
1823 * Called with cgroup_mutex held. We take callback_mutex to modify
1824 * cpus_allowed and mems_allowed.
1826 * This walk processes the tree from top to bottom, completing one layer
1827 * before dropping down to the next. It always processes a node before
1828 * any of its children.
1830 * For now, since we lack memory hot unplug, we'll never see a cpuset
1831 * that has tasks along with an empty 'mems'. But if we did see such
1832 * a cpuset, we'd handle it just like we do if its 'cpus' was empty.
1834 static void scan_for_empty_cpusets(const struct cpuset *root)
1836 struct cpuset *cp; /* scans cpusets being updated */
1837 struct cpuset *child; /* scans child cpusets of cp */
1838 struct list_head queue;
1839 struct cgroup *cont;
1840 nodemask_t oldmems;
1842 INIT_LIST_HEAD(&queue);
1844 list_add_tail((struct list_head *)&root->stack_list, &queue);
1846 while (!list_empty(&queue)) {
1847 cp = container_of(queue.next, struct cpuset, stack_list);
1848 list_del(queue.next);
1849 list_for_each_entry(cont, &cp->css.cgroup->children, sibling) {
1850 child = cgroup_cs(cont);
1851 list_add_tail(&child->stack_list, &queue);
1853 cont = cp->css.cgroup;
1855 /* Continue past cpusets with all cpus, mems online */
1856 if (cpus_subset(cp->cpus_allowed, cpu_online_map) &&
1857 nodes_subset(cp->mems_allowed, node_states[N_HIGH_MEMORY]))
1858 continue;
1860 oldmems = cp->mems_allowed;
1862 /* Remove offline cpus and mems from this cpuset. */
1863 mutex_lock(&callback_mutex);
1864 cpus_and(cp->cpus_allowed, cp->cpus_allowed, cpu_online_map);
1865 nodes_and(cp->mems_allowed, cp->mems_allowed,
1866 node_states[N_HIGH_MEMORY]);
1867 mutex_unlock(&callback_mutex);
1869 /* Move tasks from the empty cpuset to a parent */
1870 if (cpus_empty(cp->cpus_allowed) ||
1871 nodes_empty(cp->mems_allowed))
1872 remove_tasks_in_empty_cpuset(cp);
1873 else {
1874 update_tasks_cpumask(cp);
1875 update_tasks_nodemask(cp, &oldmems);
1881 * The cpus_allowed and mems_allowed nodemasks in the top_cpuset track
1882 * cpu_online_map and node_states[N_HIGH_MEMORY]. Force the top cpuset to
1883 * track what's online after any CPU or memory node hotplug or unplug event.
1885 * Since there are two callers of this routine, one for CPU hotplug
1886 * events and one for memory node hotplug events, we could have coded
1887 * two separate routines here. We code it as a single common routine
1888 * in order to minimize text size.
1891 static void common_cpu_mem_hotplug_unplug(int rebuild_sd)
1893 cgroup_lock();
1895 top_cpuset.cpus_allowed = cpu_online_map;
1896 top_cpuset.mems_allowed = node_states[N_HIGH_MEMORY];
1897 scan_for_empty_cpusets(&top_cpuset);
1900 * Scheduler destroys domains on hotplug events.
1901 * Rebuild them based on the current settings.
1903 if (rebuild_sd)
1904 rebuild_sched_domains();
1906 cgroup_unlock();
1910 * The top_cpuset tracks what CPUs and Memory Nodes are online,
1911 * period. This is necessary in order to make cpusets transparent
1912 * (of no affect) on systems that are actively using CPU hotplug
1913 * but making no active use of cpusets.
1915 * This routine ensures that top_cpuset.cpus_allowed tracks
1916 * cpu_online_map on each CPU hotplug (cpuhp) event.
1919 static int cpuset_handle_cpuhp(struct notifier_block *unused_nb,
1920 unsigned long phase, void *unused_cpu)
1922 switch (phase) {
1923 case CPU_UP_CANCELED:
1924 case CPU_UP_CANCELED_FROZEN:
1925 case CPU_DOWN_FAILED:
1926 case CPU_DOWN_FAILED_FROZEN:
1927 case CPU_ONLINE:
1928 case CPU_ONLINE_FROZEN:
1929 case CPU_DEAD:
1930 case CPU_DEAD_FROZEN:
1931 common_cpu_mem_hotplug_unplug(1);
1932 break;
1933 default:
1934 return NOTIFY_DONE;
1937 return NOTIFY_OK;
1940 #ifdef CONFIG_MEMORY_HOTPLUG
1942 * Keep top_cpuset.mems_allowed tracking node_states[N_HIGH_MEMORY].
1943 * Call this routine anytime after you change
1944 * node_states[N_HIGH_MEMORY].
1945 * See also the previous routine cpuset_handle_cpuhp().
1948 void cpuset_track_online_nodes(void)
1950 common_cpu_mem_hotplug_unplug(0);
1952 #endif
1955 * cpuset_init_smp - initialize cpus_allowed
1957 * Description: Finish top cpuset after cpu, node maps are initialized
1960 void __init cpuset_init_smp(void)
1962 top_cpuset.cpus_allowed = cpu_online_map;
1963 top_cpuset.mems_allowed = node_states[N_HIGH_MEMORY];
1965 hotcpu_notifier(cpuset_handle_cpuhp, 0);
1969 * cpuset_cpus_allowed - return cpus_allowed mask from a tasks cpuset.
1970 * @tsk: pointer to task_struct from which to obtain cpuset->cpus_allowed.
1971 * @pmask: pointer to cpumask_t variable to receive cpus_allowed set.
1973 * Description: Returns the cpumask_t cpus_allowed of the cpuset
1974 * attached to the specified @tsk. Guaranteed to return some non-empty
1975 * subset of cpu_online_map, even if this means going outside the
1976 * tasks cpuset.
1979 void cpuset_cpus_allowed(struct task_struct *tsk, cpumask_t *pmask)
1981 mutex_lock(&callback_mutex);
1982 cpuset_cpus_allowed_locked(tsk, pmask);
1983 mutex_unlock(&callback_mutex);
1987 * cpuset_cpus_allowed_locked - return cpus_allowed mask from a tasks cpuset.
1988 * Must be called with callback_mutex held.
1990 void cpuset_cpus_allowed_locked(struct task_struct *tsk, cpumask_t *pmask)
1992 task_lock(tsk);
1993 guarantee_online_cpus(task_cs(tsk), pmask);
1994 task_unlock(tsk);
1997 void cpuset_init_current_mems_allowed(void)
1999 nodes_setall(current->mems_allowed);
2003 * cpuset_mems_allowed - return mems_allowed mask from a tasks cpuset.
2004 * @tsk: pointer to task_struct from which to obtain cpuset->mems_allowed.
2006 * Description: Returns the nodemask_t mems_allowed of the cpuset
2007 * attached to the specified @tsk. Guaranteed to return some non-empty
2008 * subset of node_states[N_HIGH_MEMORY], even if this means going outside the
2009 * tasks cpuset.
2012 nodemask_t cpuset_mems_allowed(struct task_struct *tsk)
2014 nodemask_t mask;
2016 mutex_lock(&callback_mutex);
2017 task_lock(tsk);
2018 guarantee_online_mems(task_cs(tsk), &mask);
2019 task_unlock(tsk);
2020 mutex_unlock(&callback_mutex);
2022 return mask;
2026 * cpuset_nodemask_valid_mems_allowed - check nodemask vs. curremt mems_allowed
2027 * @nodemask: the nodemask to be checked
2029 * Are any of the nodes in the nodemask allowed in current->mems_allowed?
2031 int cpuset_nodemask_valid_mems_allowed(nodemask_t *nodemask)
2033 return nodes_intersects(*nodemask, current->mems_allowed);
2037 * nearest_hardwall_ancestor() - Returns the nearest mem_exclusive or
2038 * mem_hardwall ancestor to the specified cpuset. Call holding
2039 * callback_mutex. If no ancestor is mem_exclusive or mem_hardwall
2040 * (an unusual configuration), then returns the root cpuset.
2042 static const struct cpuset *nearest_hardwall_ancestor(const struct cpuset *cs)
2044 while (!(is_mem_exclusive(cs) || is_mem_hardwall(cs)) && cs->parent)
2045 cs = cs->parent;
2046 return cs;
2050 * cpuset_zone_allowed_softwall - Can we allocate on zone z's memory node?
2051 * @z: is this zone on an allowed node?
2052 * @gfp_mask: memory allocation flags
2054 * If we're in interrupt, yes, we can always allocate. If
2055 * __GFP_THISNODE is set, yes, we can always allocate. If zone
2056 * z's node is in our tasks mems_allowed, yes. If it's not a
2057 * __GFP_HARDWALL request and this zone's nodes is in the nearest
2058 * hardwalled cpuset ancestor to this tasks cpuset, yes.
2059 * If the task has been OOM killed and has access to memory reserves
2060 * as specified by the TIF_MEMDIE flag, yes.
2061 * Otherwise, no.
2063 * If __GFP_HARDWALL is set, cpuset_zone_allowed_softwall()
2064 * reduces to cpuset_zone_allowed_hardwall(). Otherwise,
2065 * cpuset_zone_allowed_softwall() might sleep, and might allow a zone
2066 * from an enclosing cpuset.
2068 * cpuset_zone_allowed_hardwall() only handles the simpler case of
2069 * hardwall cpusets, and never sleeps.
2071 * The __GFP_THISNODE placement logic is really handled elsewhere,
2072 * by forcibly using a zonelist starting at a specified node, and by
2073 * (in get_page_from_freelist()) refusing to consider the zones for
2074 * any node on the zonelist except the first. By the time any such
2075 * calls get to this routine, we should just shut up and say 'yes'.
2077 * GFP_USER allocations are marked with the __GFP_HARDWALL bit,
2078 * and do not allow allocations outside the current tasks cpuset
2079 * unless the task has been OOM killed as is marked TIF_MEMDIE.
2080 * GFP_KERNEL allocations are not so marked, so can escape to the
2081 * nearest enclosing hardwalled ancestor cpuset.
2083 * Scanning up parent cpusets requires callback_mutex. The
2084 * __alloc_pages() routine only calls here with __GFP_HARDWALL bit
2085 * _not_ set if it's a GFP_KERNEL allocation, and all nodes in the
2086 * current tasks mems_allowed came up empty on the first pass over
2087 * the zonelist. So only GFP_KERNEL allocations, if all nodes in the
2088 * cpuset are short of memory, might require taking the callback_mutex
2089 * mutex.
2091 * The first call here from mm/page_alloc:get_page_from_freelist()
2092 * has __GFP_HARDWALL set in gfp_mask, enforcing hardwall cpusets,
2093 * so no allocation on a node outside the cpuset is allowed (unless
2094 * in interrupt, of course).
2096 * The second pass through get_page_from_freelist() doesn't even call
2097 * here for GFP_ATOMIC calls. For those calls, the __alloc_pages()
2098 * variable 'wait' is not set, and the bit ALLOC_CPUSET is not set
2099 * in alloc_flags. That logic and the checks below have the combined
2100 * affect that:
2101 * in_interrupt - any node ok (current task context irrelevant)
2102 * GFP_ATOMIC - any node ok
2103 * TIF_MEMDIE - any node ok
2104 * GFP_KERNEL - any node in enclosing hardwalled cpuset ok
2105 * GFP_USER - only nodes in current tasks mems allowed ok.
2107 * Rule:
2108 * Don't call cpuset_zone_allowed_softwall if you can't sleep, unless you
2109 * pass in the __GFP_HARDWALL flag set in gfp_flag, which disables
2110 * the code that might scan up ancestor cpusets and sleep.
2113 int __cpuset_zone_allowed_softwall(struct zone *z, gfp_t gfp_mask)
2115 int node; /* node that zone z is on */
2116 const struct cpuset *cs; /* current cpuset ancestors */
2117 int allowed; /* is allocation in zone z allowed? */
2119 if (in_interrupt() || (gfp_mask & __GFP_THISNODE))
2120 return 1;
2121 node = zone_to_nid(z);
2122 might_sleep_if(!(gfp_mask & __GFP_HARDWALL));
2123 if (node_isset(node, current->mems_allowed))
2124 return 1;
2126 * Allow tasks that have access to memory reserves because they have
2127 * been OOM killed to get memory anywhere.
2129 if (unlikely(test_thread_flag(TIF_MEMDIE)))
2130 return 1;
2131 if (gfp_mask & __GFP_HARDWALL) /* If hardwall request, stop here */
2132 return 0;
2134 if (current->flags & PF_EXITING) /* Let dying task have memory */
2135 return 1;
2137 /* Not hardwall and node outside mems_allowed: scan up cpusets */
2138 mutex_lock(&callback_mutex);
2140 task_lock(current);
2141 cs = nearest_hardwall_ancestor(task_cs(current));
2142 task_unlock(current);
2144 allowed = node_isset(node, cs->mems_allowed);
2145 mutex_unlock(&callback_mutex);
2146 return allowed;
2150 * cpuset_zone_allowed_hardwall - Can we allocate on zone z's memory node?
2151 * @z: is this zone on an allowed node?
2152 * @gfp_mask: memory allocation flags
2154 * If we're in interrupt, yes, we can always allocate.
2155 * If __GFP_THISNODE is set, yes, we can always allocate. If zone
2156 * z's node is in our tasks mems_allowed, yes. If the task has been
2157 * OOM killed and has access to memory reserves as specified by the
2158 * TIF_MEMDIE flag, yes. Otherwise, no.
2160 * The __GFP_THISNODE placement logic is really handled elsewhere,
2161 * by forcibly using a zonelist starting at a specified node, and by
2162 * (in get_page_from_freelist()) refusing to consider the zones for
2163 * any node on the zonelist except the first. By the time any such
2164 * calls get to this routine, we should just shut up and say 'yes'.
2166 * Unlike the cpuset_zone_allowed_softwall() variant, above,
2167 * this variant requires that the zone be in the current tasks
2168 * mems_allowed or that we're in interrupt. It does not scan up the
2169 * cpuset hierarchy for the nearest enclosing mem_exclusive cpuset.
2170 * It never sleeps.
2173 int __cpuset_zone_allowed_hardwall(struct zone *z, gfp_t gfp_mask)
2175 int node; /* node that zone z is on */
2177 if (in_interrupt() || (gfp_mask & __GFP_THISNODE))
2178 return 1;
2179 node = zone_to_nid(z);
2180 if (node_isset(node, current->mems_allowed))
2181 return 1;
2183 * Allow tasks that have access to memory reserves because they have
2184 * been OOM killed to get memory anywhere.
2186 if (unlikely(test_thread_flag(TIF_MEMDIE)))
2187 return 1;
2188 return 0;
2192 * cpuset_lock - lock out any changes to cpuset structures
2194 * The out of memory (oom) code needs to mutex_lock cpusets
2195 * from being changed while it scans the tasklist looking for a
2196 * task in an overlapping cpuset. Expose callback_mutex via this
2197 * cpuset_lock() routine, so the oom code can lock it, before
2198 * locking the task list. The tasklist_lock is a spinlock, so
2199 * must be taken inside callback_mutex.
2202 void cpuset_lock(void)
2204 mutex_lock(&callback_mutex);
2208 * cpuset_unlock - release lock on cpuset changes
2210 * Undo the lock taken in a previous cpuset_lock() call.
2213 void cpuset_unlock(void)
2215 mutex_unlock(&callback_mutex);
2219 * cpuset_mem_spread_node() - On which node to begin search for a page
2221 * If a task is marked PF_SPREAD_PAGE or PF_SPREAD_SLAB (as for
2222 * tasks in a cpuset with is_spread_page or is_spread_slab set),
2223 * and if the memory allocation used cpuset_mem_spread_node()
2224 * to determine on which node to start looking, as it will for
2225 * certain page cache or slab cache pages such as used for file
2226 * system buffers and inode caches, then instead of starting on the
2227 * local node to look for a free page, rather spread the starting
2228 * node around the tasks mems_allowed nodes.
2230 * We don't have to worry about the returned node being offline
2231 * because "it can't happen", and even if it did, it would be ok.
2233 * The routines calling guarantee_online_mems() are careful to
2234 * only set nodes in task->mems_allowed that are online. So it
2235 * should not be possible for the following code to return an
2236 * offline node. But if it did, that would be ok, as this routine
2237 * is not returning the node where the allocation must be, only
2238 * the node where the search should start. The zonelist passed to
2239 * __alloc_pages() will include all nodes. If the slab allocator
2240 * is passed an offline node, it will fall back to the local node.
2241 * See kmem_cache_alloc_node().
2244 int cpuset_mem_spread_node(void)
2246 int node;
2248 node = next_node(current->cpuset_mem_spread_rotor, current->mems_allowed);
2249 if (node == MAX_NUMNODES)
2250 node = first_node(current->mems_allowed);
2251 current->cpuset_mem_spread_rotor = node;
2252 return node;
2254 EXPORT_SYMBOL_GPL(cpuset_mem_spread_node);
2257 * cpuset_mems_allowed_intersects - Does @tsk1's mems_allowed intersect @tsk2's?
2258 * @tsk1: pointer to task_struct of some task.
2259 * @tsk2: pointer to task_struct of some other task.
2261 * Description: Return true if @tsk1's mems_allowed intersects the
2262 * mems_allowed of @tsk2. Used by the OOM killer to determine if
2263 * one of the task's memory usage might impact the memory available
2264 * to the other.
2267 int cpuset_mems_allowed_intersects(const struct task_struct *tsk1,
2268 const struct task_struct *tsk2)
2270 return nodes_intersects(tsk1->mems_allowed, tsk2->mems_allowed);
2274 * Collection of memory_pressure is suppressed unless
2275 * this flag is enabled by writing "1" to the special
2276 * cpuset file 'memory_pressure_enabled' in the root cpuset.
2279 int cpuset_memory_pressure_enabled __read_mostly;
2282 * cpuset_memory_pressure_bump - keep stats of per-cpuset reclaims.
2284 * Keep a running average of the rate of synchronous (direct)
2285 * page reclaim efforts initiated by tasks in each cpuset.
2287 * This represents the rate at which some task in the cpuset
2288 * ran low on memory on all nodes it was allowed to use, and
2289 * had to enter the kernels page reclaim code in an effort to
2290 * create more free memory by tossing clean pages or swapping
2291 * or writing dirty pages.
2293 * Display to user space in the per-cpuset read-only file
2294 * "memory_pressure". Value displayed is an integer
2295 * representing the recent rate of entry into the synchronous
2296 * (direct) page reclaim by any task attached to the cpuset.
2299 void __cpuset_memory_pressure_bump(void)
2301 task_lock(current);
2302 fmeter_markevent(&task_cs(current)->fmeter);
2303 task_unlock(current);
2306 #ifdef CONFIG_PROC_PID_CPUSET
2308 * proc_cpuset_show()
2309 * - Print tasks cpuset path into seq_file.
2310 * - Used for /proc/<pid>/cpuset.
2311 * - No need to task_lock(tsk) on this tsk->cpuset reference, as it
2312 * doesn't really matter if tsk->cpuset changes after we read it,
2313 * and we take cgroup_mutex, keeping cpuset_attach() from changing it
2314 * anyway.
2316 static int proc_cpuset_show(struct seq_file *m, void *unused_v)
2318 struct pid *pid;
2319 struct task_struct *tsk;
2320 char *buf;
2321 struct cgroup_subsys_state *css;
2322 int retval;
2324 retval = -ENOMEM;
2325 buf = kmalloc(PAGE_SIZE, GFP_KERNEL);
2326 if (!buf)
2327 goto out;
2329 retval = -ESRCH;
2330 pid = m->private;
2331 tsk = get_pid_task(pid, PIDTYPE_PID);
2332 if (!tsk)
2333 goto out_free;
2335 retval = -EINVAL;
2336 cgroup_lock();
2337 css = task_subsys_state(tsk, cpuset_subsys_id);
2338 retval = cgroup_path(css->cgroup, buf, PAGE_SIZE);
2339 if (retval < 0)
2340 goto out_unlock;
2341 seq_puts(m, buf);
2342 seq_putc(m, '\n');
2343 out_unlock:
2344 cgroup_unlock();
2345 put_task_struct(tsk);
2346 out_free:
2347 kfree(buf);
2348 out:
2349 return retval;
2352 static int cpuset_open(struct inode *inode, struct file *file)
2354 struct pid *pid = PROC_I(inode)->pid;
2355 return single_open(file, proc_cpuset_show, pid);
2358 const struct file_operations proc_cpuset_operations = {
2359 .open = cpuset_open,
2360 .read = seq_read,
2361 .llseek = seq_lseek,
2362 .release = single_release,
2364 #endif /* CONFIG_PROC_PID_CPUSET */
2366 /* Display task cpus_allowed, mems_allowed in /proc/<pid>/status file. */
2367 void cpuset_task_status_allowed(struct seq_file *m, struct task_struct *task)
2369 seq_printf(m, "Cpus_allowed:\t");
2370 m->count += cpumask_scnprintf(m->buf + m->count, m->size - m->count,
2371 task->cpus_allowed);
2372 seq_printf(m, "\n");
2373 seq_printf(m, "Cpus_allowed_list:\t");
2374 m->count += cpulist_scnprintf(m->buf + m->count, m->size - m->count,
2375 task->cpus_allowed);
2376 seq_printf(m, "\n");
2377 seq_printf(m, "Mems_allowed:\t");
2378 m->count += nodemask_scnprintf(m->buf + m->count, m->size - m->count,
2379 task->mems_allowed);
2380 seq_printf(m, "\n");
2381 seq_printf(m, "Mems_allowed_list:\t");
2382 m->count += nodelist_scnprintf(m->buf + m->count, m->size - m->count,
2383 task->mems_allowed);
2384 seq_printf(m, "\n");