Merge branch 'for-linus' of git://git.kernel.dk/linux-2.6-block
[linux-2.6/s3c2410-cpufreq.git] / kernel / cpuset.c
bloba1b61f414228ea031cc40ea9ea17ab95e264ceaf
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 /* used for walking a cpuset heirarchy */
102 struct list_head stack_list;
105 /* Retrieve the cpuset for a cgroup */
106 static inline struct cpuset *cgroup_cs(struct cgroup *cont)
108 return container_of(cgroup_subsys_state(cont, cpuset_subsys_id),
109 struct cpuset, css);
112 /* Retrieve the cpuset for a task */
113 static inline struct cpuset *task_cs(struct task_struct *task)
115 return container_of(task_subsys_state(task, cpuset_subsys_id),
116 struct cpuset, css);
118 struct cpuset_hotplug_scanner {
119 struct cgroup_scanner scan;
120 struct cgroup *to;
123 /* bits in struct cpuset flags field */
124 typedef enum {
125 CS_CPU_EXCLUSIVE,
126 CS_MEM_EXCLUSIVE,
127 CS_MEMORY_MIGRATE,
128 CS_SCHED_LOAD_BALANCE,
129 CS_SPREAD_PAGE,
130 CS_SPREAD_SLAB,
131 } cpuset_flagbits_t;
133 /* convenient tests for these bits */
134 static inline int is_cpu_exclusive(const struct cpuset *cs)
136 return test_bit(CS_CPU_EXCLUSIVE, &cs->flags);
139 static inline int is_mem_exclusive(const struct cpuset *cs)
141 return test_bit(CS_MEM_EXCLUSIVE, &cs->flags);
144 static inline int is_sched_load_balance(const struct cpuset *cs)
146 return test_bit(CS_SCHED_LOAD_BALANCE, &cs->flags);
149 static inline int is_memory_migrate(const struct cpuset *cs)
151 return test_bit(CS_MEMORY_MIGRATE, &cs->flags);
154 static inline int is_spread_page(const struct cpuset *cs)
156 return test_bit(CS_SPREAD_PAGE, &cs->flags);
159 static inline int is_spread_slab(const struct cpuset *cs)
161 return test_bit(CS_SPREAD_SLAB, &cs->flags);
165 * Increment this integer everytime any cpuset changes its
166 * mems_allowed value. Users of cpusets can track this generation
167 * number, and avoid having to lock and reload mems_allowed unless
168 * the cpuset they're using changes generation.
170 * A single, global generation is needed because cpuset_attach_task() could
171 * reattach a task to a different cpuset, which must not have its
172 * generation numbers aliased with those of that tasks previous cpuset.
174 * Generations are needed for mems_allowed because one task cannot
175 * modify another's memory placement. So we must enable every task,
176 * on every visit to __alloc_pages(), to efficiently check whether
177 * its current->cpuset->mems_allowed has changed, requiring an update
178 * of its current->mems_allowed.
180 * Since writes to cpuset_mems_generation are guarded by the cgroup lock
181 * there is no need to mark it atomic.
183 static int cpuset_mems_generation;
185 static struct cpuset top_cpuset = {
186 .flags = ((1 << CS_CPU_EXCLUSIVE) | (1 << CS_MEM_EXCLUSIVE)),
187 .cpus_allowed = CPU_MASK_ALL,
188 .mems_allowed = NODE_MASK_ALL,
192 * There are two global mutexes guarding cpuset structures. The first
193 * is the main control groups cgroup_mutex, accessed via
194 * cgroup_lock()/cgroup_unlock(). The second is the cpuset-specific
195 * callback_mutex, below. They can nest. It is ok to first take
196 * cgroup_mutex, then nest callback_mutex. We also require taking
197 * task_lock() when dereferencing a task's cpuset pointer. See "The
198 * task_lock() exception", at the end of this comment.
200 * A task must hold both mutexes to modify cpusets. If a task
201 * holds cgroup_mutex, then it blocks others wanting that mutex,
202 * ensuring that it is the only task able to also acquire callback_mutex
203 * and be able to modify cpusets. It can perform various checks on
204 * the cpuset structure first, knowing nothing will change. It can
205 * also allocate memory while just holding cgroup_mutex. While it is
206 * performing these checks, various callback routines can briefly
207 * acquire callback_mutex to query cpusets. Once it is ready to make
208 * the changes, it takes callback_mutex, blocking everyone else.
210 * Calls to the kernel memory allocator can not be made while holding
211 * callback_mutex, as that would risk double tripping on callback_mutex
212 * from one of the callbacks into the cpuset code from within
213 * __alloc_pages().
215 * If a task is only holding callback_mutex, then it has read-only
216 * access to cpusets.
218 * The task_struct fields mems_allowed and mems_generation may only
219 * be accessed in the context of that task, so require no locks.
221 * The cpuset_common_file_write handler for operations that modify
222 * the cpuset hierarchy holds cgroup_mutex across the entire operation,
223 * single threading all such cpuset modifications across the system.
225 * The cpuset_common_file_read() handlers only hold callback_mutex across
226 * small pieces of code, such as when reading out possibly multi-word
227 * cpumasks and nodemasks.
229 * Accessing a task's cpuset should be done in accordance with the
230 * guidelines for accessing subsystem state in kernel/cgroup.c
233 static DEFINE_MUTEX(callback_mutex);
235 /* This is ugly, but preserves the userspace API for existing cpuset
236 * users. If someone tries to mount the "cpuset" filesystem, we
237 * silently switch it to mount "cgroup" instead */
238 static int cpuset_get_sb(struct file_system_type *fs_type,
239 int flags, const char *unused_dev_name,
240 void *data, struct vfsmount *mnt)
242 struct file_system_type *cgroup_fs = get_fs_type("cgroup");
243 int ret = -ENODEV;
244 if (cgroup_fs) {
245 char mountopts[] =
246 "cpuset,noprefix,"
247 "release_agent=/sbin/cpuset_release_agent";
248 ret = cgroup_fs->get_sb(cgroup_fs, flags,
249 unused_dev_name, mountopts, mnt);
250 put_filesystem(cgroup_fs);
252 return ret;
255 static struct file_system_type cpuset_fs_type = {
256 .name = "cpuset",
257 .get_sb = cpuset_get_sb,
261 * Return in *pmask the portion of a cpusets's cpus_allowed that
262 * are online. If none are online, walk up the cpuset hierarchy
263 * until we find one that does have some online cpus. If we get
264 * all the way to the top and still haven't found any online cpus,
265 * return cpu_online_map. Or if passed a NULL cs from an exit'ing
266 * task, return cpu_online_map.
268 * One way or another, we guarantee to return some non-empty subset
269 * of cpu_online_map.
271 * Call with callback_mutex held.
274 static void guarantee_online_cpus(const struct cpuset *cs, cpumask_t *pmask)
276 while (cs && !cpus_intersects(cs->cpus_allowed, cpu_online_map))
277 cs = cs->parent;
278 if (cs)
279 cpus_and(*pmask, cs->cpus_allowed, cpu_online_map);
280 else
281 *pmask = cpu_online_map;
282 BUG_ON(!cpus_intersects(*pmask, cpu_online_map));
286 * Return in *pmask the portion of a cpusets's mems_allowed that
287 * are online, with memory. If none are online with memory, walk
288 * up the cpuset hierarchy until we find one that does have some
289 * online mems. If we get all the way to the top and still haven't
290 * found any online mems, return node_states[N_HIGH_MEMORY].
292 * One way or another, we guarantee to return some non-empty subset
293 * of node_states[N_HIGH_MEMORY].
295 * Call with callback_mutex held.
298 static void guarantee_online_mems(const struct cpuset *cs, nodemask_t *pmask)
300 while (cs && !nodes_intersects(cs->mems_allowed,
301 node_states[N_HIGH_MEMORY]))
302 cs = cs->parent;
303 if (cs)
304 nodes_and(*pmask, cs->mems_allowed,
305 node_states[N_HIGH_MEMORY]);
306 else
307 *pmask = node_states[N_HIGH_MEMORY];
308 BUG_ON(!nodes_intersects(*pmask, node_states[N_HIGH_MEMORY]));
312 * cpuset_update_task_memory_state - update task memory placement
314 * If the current tasks cpusets mems_allowed changed behind our
315 * backs, update current->mems_allowed, mems_generation and task NUMA
316 * mempolicy to the new value.
318 * Task mempolicy is updated by rebinding it relative to the
319 * current->cpuset if a task has its memory placement changed.
320 * Do not call this routine if in_interrupt().
322 * Call without callback_mutex or task_lock() held. May be
323 * called with or without cgroup_mutex held. Thanks in part to
324 * 'the_top_cpuset_hack', the task's cpuset pointer will never
325 * be NULL. This routine also might acquire callback_mutex during
326 * call.
328 * Reading current->cpuset->mems_generation doesn't need task_lock
329 * to guard the current->cpuset derefence, because it is guarded
330 * from concurrent freeing of current->cpuset using RCU.
332 * The rcu_dereference() is technically probably not needed,
333 * as I don't actually mind if I see a new cpuset pointer but
334 * an old value of mems_generation. However this really only
335 * matters on alpha systems using cpusets heavily. If I dropped
336 * that rcu_dereference(), it would save them a memory barrier.
337 * For all other arch's, rcu_dereference is a no-op anyway, and for
338 * alpha systems not using cpusets, another planned optimization,
339 * avoiding the rcu critical section for tasks in the root cpuset
340 * which is statically allocated, so can't vanish, will make this
341 * irrelevant. Better to use RCU as intended, than to engage in
342 * some cute trick to save a memory barrier that is impossible to
343 * test, for alpha systems using cpusets heavily, which might not
344 * even exist.
346 * This routine is needed to update the per-task mems_allowed data,
347 * within the tasks context, when it is trying to allocate memory
348 * (in various mm/mempolicy.c routines) and notices that some other
349 * task has been modifying its cpuset.
352 void cpuset_update_task_memory_state(void)
354 int my_cpusets_mem_gen;
355 struct task_struct *tsk = current;
356 struct cpuset *cs;
358 if (task_cs(tsk) == &top_cpuset) {
359 /* Don't need rcu for top_cpuset. It's never freed. */
360 my_cpusets_mem_gen = top_cpuset.mems_generation;
361 } else {
362 rcu_read_lock();
363 my_cpusets_mem_gen = task_cs(current)->mems_generation;
364 rcu_read_unlock();
367 if (my_cpusets_mem_gen != tsk->cpuset_mems_generation) {
368 mutex_lock(&callback_mutex);
369 task_lock(tsk);
370 cs = task_cs(tsk); /* Maybe changed when task not locked */
371 guarantee_online_mems(cs, &tsk->mems_allowed);
372 tsk->cpuset_mems_generation = cs->mems_generation;
373 if (is_spread_page(cs))
374 tsk->flags |= PF_SPREAD_PAGE;
375 else
376 tsk->flags &= ~PF_SPREAD_PAGE;
377 if (is_spread_slab(cs))
378 tsk->flags |= PF_SPREAD_SLAB;
379 else
380 tsk->flags &= ~PF_SPREAD_SLAB;
381 task_unlock(tsk);
382 mutex_unlock(&callback_mutex);
383 mpol_rebind_task(tsk, &tsk->mems_allowed);
388 * is_cpuset_subset(p, q) - Is cpuset p a subset of cpuset q?
390 * One cpuset is a subset of another if all its allowed CPUs and
391 * Memory Nodes are a subset of the other, and its exclusive flags
392 * are only set if the other's are set. Call holding cgroup_mutex.
395 static int is_cpuset_subset(const struct cpuset *p, const struct cpuset *q)
397 return cpus_subset(p->cpus_allowed, q->cpus_allowed) &&
398 nodes_subset(p->mems_allowed, q->mems_allowed) &&
399 is_cpu_exclusive(p) <= is_cpu_exclusive(q) &&
400 is_mem_exclusive(p) <= is_mem_exclusive(q);
404 * validate_change() - Used to validate that any proposed cpuset change
405 * follows the structural rules for cpusets.
407 * If we replaced the flag and mask values of the current cpuset
408 * (cur) with those values in the trial cpuset (trial), would
409 * our various subset and exclusive rules still be valid? Presumes
410 * cgroup_mutex held.
412 * 'cur' is the address of an actual, in-use cpuset. Operations
413 * such as list traversal that depend on the actual address of the
414 * cpuset in the list must use cur below, not trial.
416 * 'trial' is the address of bulk structure copy of cur, with
417 * perhaps one or more of the fields cpus_allowed, mems_allowed,
418 * or flags changed to new, trial values.
420 * Return 0 if valid, -errno if not.
423 static int validate_change(const struct cpuset *cur, const struct cpuset *trial)
425 struct cgroup *cont;
426 struct cpuset *c, *par;
428 /* Each of our child cpusets must be a subset of us */
429 list_for_each_entry(cont, &cur->css.cgroup->children, sibling) {
430 if (!is_cpuset_subset(cgroup_cs(cont), trial))
431 return -EBUSY;
434 /* Remaining checks don't apply to root cpuset */
435 if (cur == &top_cpuset)
436 return 0;
438 par = cur->parent;
440 /* We must be a subset of our parent cpuset */
441 if (!is_cpuset_subset(trial, par))
442 return -EACCES;
445 * If either I or some sibling (!= me) is exclusive, we can't
446 * overlap
448 list_for_each_entry(cont, &par->css.cgroup->children, sibling) {
449 c = cgroup_cs(cont);
450 if ((is_cpu_exclusive(trial) || is_cpu_exclusive(c)) &&
451 c != cur &&
452 cpus_intersects(trial->cpus_allowed, c->cpus_allowed))
453 return -EINVAL;
454 if ((is_mem_exclusive(trial) || is_mem_exclusive(c)) &&
455 c != cur &&
456 nodes_intersects(trial->mems_allowed, c->mems_allowed))
457 return -EINVAL;
460 /* Cpusets with tasks can't have empty cpus_allowed or mems_allowed */
461 if (cgroup_task_count(cur->css.cgroup)) {
462 if (cpus_empty(trial->cpus_allowed) ||
463 nodes_empty(trial->mems_allowed)) {
464 return -ENOSPC;
468 return 0;
472 * Helper routine for rebuild_sched_domains().
473 * Do cpusets a, b have overlapping cpus_allowed masks?
476 static int cpusets_overlap(struct cpuset *a, struct cpuset *b)
478 return cpus_intersects(a->cpus_allowed, b->cpus_allowed);
482 * rebuild_sched_domains()
484 * If the flag 'sched_load_balance' of any cpuset with non-empty
485 * 'cpus' changes, or if the 'cpus' allowed changes in any cpuset
486 * which has that flag enabled, or if any cpuset with a non-empty
487 * 'cpus' is removed, then call this routine to rebuild the
488 * scheduler's dynamic sched domains.
490 * This routine builds a partial partition of the systems CPUs
491 * (the set of non-overlappping cpumask_t's in the array 'part'
492 * below), and passes that partial partition to the kernel/sched.c
493 * partition_sched_domains() routine, which will rebuild the
494 * schedulers load balancing domains (sched domains) as specified
495 * by that partial partition. A 'partial partition' is a set of
496 * non-overlapping subsets whose union is a subset of that set.
498 * See "What is sched_load_balance" in Documentation/cpusets.txt
499 * for a background explanation of this.
501 * Does not return errors, on the theory that the callers of this
502 * routine would rather not worry about failures to rebuild sched
503 * domains when operating in the severe memory shortage situations
504 * that could cause allocation failures below.
506 * Call with cgroup_mutex held. May take callback_mutex during
507 * call due to the kfifo_alloc() and kmalloc() calls. May nest
508 * a call to the get_online_cpus()/put_online_cpus() pair.
509 * Must not be called holding callback_mutex, because we must not
510 * call get_online_cpus() while holding callback_mutex. Elsewhere
511 * the kernel nests callback_mutex inside get_online_cpus() calls.
512 * So the reverse nesting would risk an ABBA deadlock.
514 * The three key local variables below are:
515 * q - a kfifo queue of cpuset pointers, used to implement a
516 * top-down scan of all cpusets. This scan loads a pointer
517 * to each cpuset marked is_sched_load_balance into the
518 * array 'csa'. For our purposes, rebuilding the schedulers
519 * sched domains, we can ignore !is_sched_load_balance cpusets.
520 * csa - (for CpuSet Array) Array of pointers to all the cpusets
521 * that need to be load balanced, for convenient iterative
522 * access by the subsequent code that finds the best partition,
523 * i.e the set of domains (subsets) of CPUs such that the
524 * cpus_allowed of every cpuset marked is_sched_load_balance
525 * is a subset of one of these domains, while there are as
526 * many such domains as possible, each as small as possible.
527 * doms - Conversion of 'csa' to an array of cpumasks, for passing to
528 * the kernel/sched.c routine partition_sched_domains() in a
529 * convenient format, that can be easily compared to the prior
530 * value to determine what partition elements (sched domains)
531 * were changed (added or removed.)
533 * Finding the best partition (set of domains):
534 * The triple nested loops below over i, j, k scan over the
535 * load balanced cpusets (using the array of cpuset pointers in
536 * csa[]) looking for pairs of cpusets that have overlapping
537 * cpus_allowed, but which don't have the same 'pn' partition
538 * number and gives them in the same partition number. It keeps
539 * looping on the 'restart' label until it can no longer find
540 * any such pairs.
542 * The union of the cpus_allowed masks from the set of
543 * all cpusets having the same 'pn' value then form the one
544 * element of the partition (one sched domain) to be passed to
545 * partition_sched_domains().
548 static void rebuild_sched_domains(void)
550 struct kfifo *q; /* queue of cpusets to be scanned */
551 struct cpuset *cp; /* scans q */
552 struct cpuset **csa; /* array of all cpuset ptrs */
553 int csn; /* how many cpuset ptrs in csa so far */
554 int i, j, k; /* indices for partition finding loops */
555 cpumask_t *doms; /* resulting partition; i.e. sched domains */
556 int ndoms; /* number of sched domains in result */
557 int nslot; /* next empty doms[] cpumask_t slot */
559 q = NULL;
560 csa = NULL;
561 doms = NULL;
563 /* Special case for the 99% of systems with one, full, sched domain */
564 if (is_sched_load_balance(&top_cpuset)) {
565 ndoms = 1;
566 doms = kmalloc(sizeof(cpumask_t), GFP_KERNEL);
567 if (!doms)
568 goto rebuild;
569 *doms = top_cpuset.cpus_allowed;
570 goto rebuild;
573 q = kfifo_alloc(number_of_cpusets * sizeof(cp), GFP_KERNEL, NULL);
574 if (IS_ERR(q))
575 goto done;
576 csa = kmalloc(number_of_cpusets * sizeof(cp), GFP_KERNEL);
577 if (!csa)
578 goto done;
579 csn = 0;
581 cp = &top_cpuset;
582 __kfifo_put(q, (void *)&cp, sizeof(cp));
583 while (__kfifo_get(q, (void *)&cp, sizeof(cp))) {
584 struct cgroup *cont;
585 struct cpuset *child; /* scans child cpusets of cp */
586 if (is_sched_load_balance(cp))
587 csa[csn++] = cp;
588 list_for_each_entry(cont, &cp->css.cgroup->children, sibling) {
589 child = cgroup_cs(cont);
590 __kfifo_put(q, (void *)&child, sizeof(cp));
594 for (i = 0; i < csn; i++)
595 csa[i]->pn = i;
596 ndoms = csn;
598 restart:
599 /* Find the best partition (set of sched domains) */
600 for (i = 0; i < csn; i++) {
601 struct cpuset *a = csa[i];
602 int apn = a->pn;
604 for (j = 0; j < csn; j++) {
605 struct cpuset *b = csa[j];
606 int bpn = b->pn;
608 if (apn != bpn && cpusets_overlap(a, b)) {
609 for (k = 0; k < csn; k++) {
610 struct cpuset *c = csa[k];
612 if (c->pn == bpn)
613 c->pn = apn;
615 ndoms--; /* one less element */
616 goto restart;
621 /* Convert <csn, csa> to <ndoms, doms> */
622 doms = kmalloc(ndoms * sizeof(cpumask_t), GFP_KERNEL);
623 if (!doms)
624 goto rebuild;
626 for (nslot = 0, i = 0; i < csn; i++) {
627 struct cpuset *a = csa[i];
628 int apn = a->pn;
630 if (apn >= 0) {
631 cpumask_t *dp = doms + nslot;
633 if (nslot == ndoms) {
634 static int warnings = 10;
635 if (warnings) {
636 printk(KERN_WARNING
637 "rebuild_sched_domains confused:"
638 " nslot %d, ndoms %d, csn %d, i %d,"
639 " apn %d\n",
640 nslot, ndoms, csn, i, apn);
641 warnings--;
643 continue;
646 cpus_clear(*dp);
647 for (j = i; j < csn; j++) {
648 struct cpuset *b = csa[j];
650 if (apn == b->pn) {
651 cpus_or(*dp, *dp, b->cpus_allowed);
652 b->pn = -1;
655 nslot++;
658 BUG_ON(nslot != ndoms);
660 rebuild:
661 /* Have scheduler rebuild sched domains */
662 get_online_cpus();
663 partition_sched_domains(ndoms, doms);
664 put_online_cpus();
666 done:
667 if (q && !IS_ERR(q))
668 kfifo_free(q);
669 kfree(csa);
670 /* Don't kfree(doms) -- partition_sched_domains() does that. */
673 static inline int started_after_time(struct task_struct *t1,
674 struct timespec *time,
675 struct task_struct *t2)
677 int start_diff = timespec_compare(&t1->start_time, time);
678 if (start_diff > 0) {
679 return 1;
680 } else if (start_diff < 0) {
681 return 0;
682 } else {
684 * Arbitrarily, if two processes started at the same
685 * time, we'll say that the lower pointer value
686 * started first. Note that t2 may have exited by now
687 * so this may not be a valid pointer any longer, but
688 * that's fine - it still serves to distinguish
689 * between two tasks started (effectively)
690 * simultaneously.
692 return t1 > t2;
696 static inline int started_after(void *p1, void *p2)
698 struct task_struct *t1 = p1;
699 struct task_struct *t2 = p2;
700 return started_after_time(t1, &t2->start_time, t2);
704 * cpuset_test_cpumask - test a task's cpus_allowed versus its cpuset's
705 * @tsk: task to test
706 * @scan: struct cgroup_scanner contained in its struct cpuset_hotplug_scanner
708 * Call with cgroup_mutex held. May take callback_mutex during call.
709 * Called for each task in a cgroup by cgroup_scan_tasks().
710 * Return nonzero if this tasks's cpus_allowed mask should be changed (in other
711 * words, if its mask is not equal to its cpuset's mask).
713 int cpuset_test_cpumask(struct task_struct *tsk, struct cgroup_scanner *scan)
715 return !cpus_equal(tsk->cpus_allowed,
716 (cgroup_cs(scan->cg))->cpus_allowed);
720 * cpuset_change_cpumask - make a task's cpus_allowed the same as its cpuset's
721 * @tsk: task to test
722 * @scan: struct cgroup_scanner containing the cgroup of the task
724 * Called by cgroup_scan_tasks() for each task in a cgroup whose
725 * cpus_allowed mask needs to be changed.
727 * We don't need to re-check for the cgroup/cpuset membership, since we're
728 * holding cgroup_lock() at this point.
730 void cpuset_change_cpumask(struct task_struct *tsk, struct cgroup_scanner *scan)
732 set_cpus_allowed(tsk, (cgroup_cs(scan->cg))->cpus_allowed);
736 * update_cpumask - update the cpus_allowed mask of a cpuset and all tasks in it
737 * @cs: the cpuset to consider
738 * @buf: buffer of cpu numbers written to this cpuset
740 static int update_cpumask(struct cpuset *cs, char *buf)
742 struct cpuset trialcs;
743 struct cgroup_scanner scan;
744 struct ptr_heap heap;
745 int retval;
746 int is_load_balanced;
748 /* top_cpuset.cpus_allowed tracks cpu_online_map; it's read-only */
749 if (cs == &top_cpuset)
750 return -EACCES;
752 trialcs = *cs;
755 * An empty cpus_allowed is ok only if the cpuset has no tasks.
756 * Since cpulist_parse() fails on an empty mask, we special case
757 * that parsing. The validate_change() call ensures that cpusets
758 * with tasks have cpus.
760 buf = strstrip(buf);
761 if (!*buf) {
762 cpus_clear(trialcs.cpus_allowed);
763 } else {
764 retval = cpulist_parse(buf, trialcs.cpus_allowed);
765 if (retval < 0)
766 return retval;
768 cpus_and(trialcs.cpus_allowed, trialcs.cpus_allowed, cpu_online_map);
769 retval = validate_change(cs, &trialcs);
770 if (retval < 0)
771 return retval;
773 /* Nothing to do if the cpus didn't change */
774 if (cpus_equal(cs->cpus_allowed, trialcs.cpus_allowed))
775 return 0;
777 retval = heap_init(&heap, PAGE_SIZE, GFP_KERNEL, &started_after);
778 if (retval)
779 return retval;
781 is_load_balanced = is_sched_load_balance(&trialcs);
783 mutex_lock(&callback_mutex);
784 cs->cpus_allowed = trialcs.cpus_allowed;
785 mutex_unlock(&callback_mutex);
788 * Scan tasks in the cpuset, and update the cpumasks of any
789 * that need an update.
791 scan.cg = cs->css.cgroup;
792 scan.test_task = cpuset_test_cpumask;
793 scan.process_task = cpuset_change_cpumask;
794 scan.heap = &heap;
795 cgroup_scan_tasks(&scan);
796 heap_free(&heap);
798 if (is_load_balanced)
799 rebuild_sched_domains();
800 return 0;
804 * cpuset_migrate_mm
806 * Migrate memory region from one set of nodes to another.
808 * Temporarilly set tasks mems_allowed to target nodes of migration,
809 * so that the migration code can allocate pages on these nodes.
811 * Call holding cgroup_mutex, so current's cpuset won't change
812 * during this call, as manage_mutex holds off any cpuset_attach()
813 * calls. Therefore we don't need to take task_lock around the
814 * call to guarantee_online_mems(), as we know no one is changing
815 * our task's cpuset.
817 * Hold callback_mutex around the two modifications of our tasks
818 * mems_allowed to synchronize with cpuset_mems_allowed().
820 * While the mm_struct we are migrating is typically from some
821 * other task, the task_struct mems_allowed that we are hacking
822 * is for our current task, which must allocate new pages for that
823 * migrating memory region.
825 * We call cpuset_update_task_memory_state() before hacking
826 * our tasks mems_allowed, so that we are assured of being in
827 * sync with our tasks cpuset, and in particular, callbacks to
828 * cpuset_update_task_memory_state() from nested page allocations
829 * won't see any mismatch of our cpuset and task mems_generation
830 * values, so won't overwrite our hacked tasks mems_allowed
831 * nodemask.
834 static void cpuset_migrate_mm(struct mm_struct *mm, const nodemask_t *from,
835 const nodemask_t *to)
837 struct task_struct *tsk = current;
839 cpuset_update_task_memory_state();
841 mutex_lock(&callback_mutex);
842 tsk->mems_allowed = *to;
843 mutex_unlock(&callback_mutex);
845 do_migrate_pages(mm, from, to, MPOL_MF_MOVE_ALL);
847 mutex_lock(&callback_mutex);
848 guarantee_online_mems(task_cs(tsk),&tsk->mems_allowed);
849 mutex_unlock(&callback_mutex);
853 * Handle user request to change the 'mems' memory placement
854 * of a cpuset. Needs to validate the request, update the
855 * cpusets mems_allowed and mems_generation, and for each
856 * task in the cpuset, rebind any vma mempolicies and if
857 * the cpuset is marked 'memory_migrate', migrate the tasks
858 * pages to the new memory.
860 * Call with cgroup_mutex held. May take callback_mutex during call.
861 * Will take tasklist_lock, scan tasklist for tasks in cpuset cs,
862 * lock each such tasks mm->mmap_sem, scan its vma's and rebind
863 * their mempolicies to the cpusets new mems_allowed.
866 static void *cpuset_being_rebound;
868 static int update_nodemask(struct cpuset *cs, char *buf)
870 struct cpuset trialcs;
871 nodemask_t oldmem;
872 struct task_struct *p;
873 struct mm_struct **mmarray;
874 int i, n, ntasks;
875 int migrate;
876 int fudge;
877 int retval;
878 struct cgroup_iter it;
881 * top_cpuset.mems_allowed tracks node_stats[N_HIGH_MEMORY];
882 * it's read-only
884 if (cs == &top_cpuset)
885 return -EACCES;
887 trialcs = *cs;
890 * An empty mems_allowed is ok iff there are no tasks in the cpuset.
891 * Since nodelist_parse() fails on an empty mask, we special case
892 * that parsing. The validate_change() call ensures that cpusets
893 * with tasks have memory.
895 buf = strstrip(buf);
896 if (!*buf) {
897 nodes_clear(trialcs.mems_allowed);
898 } else {
899 retval = nodelist_parse(buf, trialcs.mems_allowed);
900 if (retval < 0)
901 goto done;
903 nodes_and(trialcs.mems_allowed, trialcs.mems_allowed,
904 node_states[N_HIGH_MEMORY]);
905 oldmem = cs->mems_allowed;
906 if (nodes_equal(oldmem, trialcs.mems_allowed)) {
907 retval = 0; /* Too easy - nothing to do */
908 goto done;
910 retval = validate_change(cs, &trialcs);
911 if (retval < 0)
912 goto done;
914 mutex_lock(&callback_mutex);
915 cs->mems_allowed = trialcs.mems_allowed;
916 cs->mems_generation = cpuset_mems_generation++;
917 mutex_unlock(&callback_mutex);
919 cpuset_being_rebound = cs; /* causes mpol_copy() rebind */
921 fudge = 10; /* spare mmarray[] slots */
922 fudge += cpus_weight(cs->cpus_allowed); /* imagine one fork-bomb/cpu */
923 retval = -ENOMEM;
926 * Allocate mmarray[] to hold mm reference for each task
927 * in cpuset cs. Can't kmalloc GFP_KERNEL while holding
928 * tasklist_lock. We could use GFP_ATOMIC, but with a
929 * few more lines of code, we can retry until we get a big
930 * enough mmarray[] w/o using GFP_ATOMIC.
932 while (1) {
933 ntasks = cgroup_task_count(cs->css.cgroup); /* guess */
934 ntasks += fudge;
935 mmarray = kmalloc(ntasks * sizeof(*mmarray), GFP_KERNEL);
936 if (!mmarray)
937 goto done;
938 read_lock(&tasklist_lock); /* block fork */
939 if (cgroup_task_count(cs->css.cgroup) <= ntasks)
940 break; /* got enough */
941 read_unlock(&tasklist_lock); /* try again */
942 kfree(mmarray);
945 n = 0;
947 /* Load up mmarray[] with mm reference for each task in cpuset. */
948 cgroup_iter_start(cs->css.cgroup, &it);
949 while ((p = cgroup_iter_next(cs->css.cgroup, &it))) {
950 struct mm_struct *mm;
952 if (n >= ntasks) {
953 printk(KERN_WARNING
954 "Cpuset mempolicy rebind incomplete.\n");
955 break;
957 mm = get_task_mm(p);
958 if (!mm)
959 continue;
960 mmarray[n++] = mm;
962 cgroup_iter_end(cs->css.cgroup, &it);
963 read_unlock(&tasklist_lock);
966 * Now that we've dropped the tasklist spinlock, we can
967 * rebind the vma mempolicies of each mm in mmarray[] to their
968 * new cpuset, and release that mm. The mpol_rebind_mm()
969 * call takes mmap_sem, which we couldn't take while holding
970 * tasklist_lock. Forks can happen again now - the mpol_copy()
971 * cpuset_being_rebound check will catch such forks, and rebind
972 * their vma mempolicies too. Because we still hold the global
973 * cgroup_mutex, we know that no other rebind effort will
974 * be contending for the global variable cpuset_being_rebound.
975 * It's ok if we rebind the same mm twice; mpol_rebind_mm()
976 * is idempotent. Also migrate pages in each mm to new nodes.
978 migrate = is_memory_migrate(cs);
979 for (i = 0; i < n; i++) {
980 struct mm_struct *mm = mmarray[i];
982 mpol_rebind_mm(mm, &cs->mems_allowed);
983 if (migrate)
984 cpuset_migrate_mm(mm, &oldmem, &cs->mems_allowed);
985 mmput(mm);
988 /* We're done rebinding vmas to this cpuset's new mems_allowed. */
989 kfree(mmarray);
990 cpuset_being_rebound = NULL;
991 retval = 0;
992 done:
993 return retval;
996 int current_cpuset_is_being_rebound(void)
998 return task_cs(current) == cpuset_being_rebound;
1002 * Call with cgroup_mutex held.
1005 static int update_memory_pressure_enabled(struct cpuset *cs, char *buf)
1007 if (simple_strtoul(buf, NULL, 10) != 0)
1008 cpuset_memory_pressure_enabled = 1;
1009 else
1010 cpuset_memory_pressure_enabled = 0;
1011 return 0;
1015 * update_flag - read a 0 or a 1 in a file and update associated flag
1016 * bit: the bit to update (CS_CPU_EXCLUSIVE, CS_MEM_EXCLUSIVE,
1017 * CS_SCHED_LOAD_BALANCE,
1018 * CS_NOTIFY_ON_RELEASE, CS_MEMORY_MIGRATE,
1019 * CS_SPREAD_PAGE, CS_SPREAD_SLAB)
1020 * cs: the cpuset to update
1021 * buf: the buffer where we read the 0 or 1
1023 * Call with cgroup_mutex held.
1026 static int update_flag(cpuset_flagbits_t bit, struct cpuset *cs, char *buf)
1028 int turning_on;
1029 struct cpuset trialcs;
1030 int err;
1031 int cpus_nonempty, balance_flag_changed;
1033 turning_on = (simple_strtoul(buf, NULL, 10) != 0);
1035 trialcs = *cs;
1036 if (turning_on)
1037 set_bit(bit, &trialcs.flags);
1038 else
1039 clear_bit(bit, &trialcs.flags);
1041 err = validate_change(cs, &trialcs);
1042 if (err < 0)
1043 return err;
1045 cpus_nonempty = !cpus_empty(trialcs.cpus_allowed);
1046 balance_flag_changed = (is_sched_load_balance(cs) !=
1047 is_sched_load_balance(&trialcs));
1049 mutex_lock(&callback_mutex);
1050 cs->flags = trialcs.flags;
1051 mutex_unlock(&callback_mutex);
1053 if (cpus_nonempty && balance_flag_changed)
1054 rebuild_sched_domains();
1056 return 0;
1060 * Frequency meter - How fast is some event occurring?
1062 * These routines manage a digitally filtered, constant time based,
1063 * event frequency meter. There are four routines:
1064 * fmeter_init() - initialize a frequency meter.
1065 * fmeter_markevent() - called each time the event happens.
1066 * fmeter_getrate() - returns the recent rate of such events.
1067 * fmeter_update() - internal routine used to update fmeter.
1069 * A common data structure is passed to each of these routines,
1070 * which is used to keep track of the state required to manage the
1071 * frequency meter and its digital filter.
1073 * The filter works on the number of events marked per unit time.
1074 * The filter is single-pole low-pass recursive (IIR). The time unit
1075 * is 1 second. Arithmetic is done using 32-bit integers scaled to
1076 * simulate 3 decimal digits of precision (multiplied by 1000).
1078 * With an FM_COEF of 933, and a time base of 1 second, the filter
1079 * has a half-life of 10 seconds, meaning that if the events quit
1080 * happening, then the rate returned from the fmeter_getrate()
1081 * will be cut in half each 10 seconds, until it converges to zero.
1083 * It is not worth doing a real infinitely recursive filter. If more
1084 * than FM_MAXTICKS ticks have elapsed since the last filter event,
1085 * just compute FM_MAXTICKS ticks worth, by which point the level
1086 * will be stable.
1088 * Limit the count of unprocessed events to FM_MAXCNT, so as to avoid
1089 * arithmetic overflow in the fmeter_update() routine.
1091 * Given the simple 32 bit integer arithmetic used, this meter works
1092 * best for reporting rates between one per millisecond (msec) and
1093 * one per 32 (approx) seconds. At constant rates faster than one
1094 * per msec it maxes out at values just under 1,000,000. At constant
1095 * rates between one per msec, and one per second it will stabilize
1096 * to a value N*1000, where N is the rate of events per second.
1097 * At constant rates between one per second and one per 32 seconds,
1098 * it will be choppy, moving up on the seconds that have an event,
1099 * and then decaying until the next event. At rates slower than
1100 * about one in 32 seconds, it decays all the way back to zero between
1101 * each event.
1104 #define FM_COEF 933 /* coefficient for half-life of 10 secs */
1105 #define FM_MAXTICKS ((time_t)99) /* useless computing more ticks than this */
1106 #define FM_MAXCNT 1000000 /* limit cnt to avoid overflow */
1107 #define FM_SCALE 1000 /* faux fixed point scale */
1109 /* Initialize a frequency meter */
1110 static void fmeter_init(struct fmeter *fmp)
1112 fmp->cnt = 0;
1113 fmp->val = 0;
1114 fmp->time = 0;
1115 spin_lock_init(&fmp->lock);
1118 /* Internal meter update - process cnt events and update value */
1119 static void fmeter_update(struct fmeter *fmp)
1121 time_t now = get_seconds();
1122 time_t ticks = now - fmp->time;
1124 if (ticks == 0)
1125 return;
1127 ticks = min(FM_MAXTICKS, ticks);
1128 while (ticks-- > 0)
1129 fmp->val = (FM_COEF * fmp->val) / FM_SCALE;
1130 fmp->time = now;
1132 fmp->val += ((FM_SCALE - FM_COEF) * fmp->cnt) / FM_SCALE;
1133 fmp->cnt = 0;
1136 /* Process any previous ticks, then bump cnt by one (times scale). */
1137 static void fmeter_markevent(struct fmeter *fmp)
1139 spin_lock(&fmp->lock);
1140 fmeter_update(fmp);
1141 fmp->cnt = min(FM_MAXCNT, fmp->cnt + FM_SCALE);
1142 spin_unlock(&fmp->lock);
1145 /* Process any previous ticks, then return current value. */
1146 static int fmeter_getrate(struct fmeter *fmp)
1148 int val;
1150 spin_lock(&fmp->lock);
1151 fmeter_update(fmp);
1152 val = fmp->val;
1153 spin_unlock(&fmp->lock);
1154 return val;
1157 /* Called by cgroups to determine if a cpuset is usable; cgroup_mutex held */
1158 static int cpuset_can_attach(struct cgroup_subsys *ss,
1159 struct cgroup *cont, struct task_struct *tsk)
1161 struct cpuset *cs = cgroup_cs(cont);
1163 if (cpus_empty(cs->cpus_allowed) || nodes_empty(cs->mems_allowed))
1164 return -ENOSPC;
1166 return security_task_setscheduler(tsk, 0, NULL);
1169 static void cpuset_attach(struct cgroup_subsys *ss,
1170 struct cgroup *cont, struct cgroup *oldcont,
1171 struct task_struct *tsk)
1173 cpumask_t cpus;
1174 nodemask_t from, to;
1175 struct mm_struct *mm;
1176 struct cpuset *cs = cgroup_cs(cont);
1177 struct cpuset *oldcs = cgroup_cs(oldcont);
1179 mutex_lock(&callback_mutex);
1180 guarantee_online_cpus(cs, &cpus);
1181 set_cpus_allowed(tsk, cpus);
1182 mutex_unlock(&callback_mutex);
1184 from = oldcs->mems_allowed;
1185 to = cs->mems_allowed;
1186 mm = get_task_mm(tsk);
1187 if (mm) {
1188 mpol_rebind_mm(mm, &to);
1189 if (is_memory_migrate(cs))
1190 cpuset_migrate_mm(mm, &from, &to);
1191 mmput(mm);
1196 /* The various types of files and directories in a cpuset file system */
1198 typedef enum {
1199 FILE_MEMORY_MIGRATE,
1200 FILE_CPULIST,
1201 FILE_MEMLIST,
1202 FILE_CPU_EXCLUSIVE,
1203 FILE_MEM_EXCLUSIVE,
1204 FILE_SCHED_LOAD_BALANCE,
1205 FILE_MEMORY_PRESSURE_ENABLED,
1206 FILE_MEMORY_PRESSURE,
1207 FILE_SPREAD_PAGE,
1208 FILE_SPREAD_SLAB,
1209 } cpuset_filetype_t;
1211 static ssize_t cpuset_common_file_write(struct cgroup *cont,
1212 struct cftype *cft,
1213 struct file *file,
1214 const char __user *userbuf,
1215 size_t nbytes, loff_t *unused_ppos)
1217 struct cpuset *cs = cgroup_cs(cont);
1218 cpuset_filetype_t type = cft->private;
1219 char *buffer;
1220 int retval = 0;
1222 /* Crude upper limit on largest legitimate cpulist user might write. */
1223 if (nbytes > 100U + 6 * max(NR_CPUS, MAX_NUMNODES))
1224 return -E2BIG;
1226 /* +1 for nul-terminator */
1227 if ((buffer = kmalloc(nbytes + 1, GFP_KERNEL)) == 0)
1228 return -ENOMEM;
1230 if (copy_from_user(buffer, userbuf, nbytes)) {
1231 retval = -EFAULT;
1232 goto out1;
1234 buffer[nbytes] = 0; /* nul-terminate */
1236 cgroup_lock();
1238 if (cgroup_is_removed(cont)) {
1239 retval = -ENODEV;
1240 goto out2;
1243 switch (type) {
1244 case FILE_CPULIST:
1245 retval = update_cpumask(cs, buffer);
1246 break;
1247 case FILE_MEMLIST:
1248 retval = update_nodemask(cs, buffer);
1249 break;
1250 case FILE_CPU_EXCLUSIVE:
1251 retval = update_flag(CS_CPU_EXCLUSIVE, cs, buffer);
1252 break;
1253 case FILE_MEM_EXCLUSIVE:
1254 retval = update_flag(CS_MEM_EXCLUSIVE, cs, buffer);
1255 break;
1256 case FILE_SCHED_LOAD_BALANCE:
1257 retval = update_flag(CS_SCHED_LOAD_BALANCE, cs, buffer);
1258 break;
1259 case FILE_MEMORY_MIGRATE:
1260 retval = update_flag(CS_MEMORY_MIGRATE, cs, buffer);
1261 break;
1262 case FILE_MEMORY_PRESSURE_ENABLED:
1263 retval = update_memory_pressure_enabled(cs, buffer);
1264 break;
1265 case FILE_MEMORY_PRESSURE:
1266 retval = -EACCES;
1267 break;
1268 case FILE_SPREAD_PAGE:
1269 retval = update_flag(CS_SPREAD_PAGE, cs, buffer);
1270 cs->mems_generation = cpuset_mems_generation++;
1271 break;
1272 case FILE_SPREAD_SLAB:
1273 retval = update_flag(CS_SPREAD_SLAB, cs, buffer);
1274 cs->mems_generation = cpuset_mems_generation++;
1275 break;
1276 default:
1277 retval = -EINVAL;
1278 goto out2;
1281 if (retval == 0)
1282 retval = nbytes;
1283 out2:
1284 cgroup_unlock();
1285 out1:
1286 kfree(buffer);
1287 return retval;
1291 * These ascii lists should be read in a single call, by using a user
1292 * buffer large enough to hold the entire map. If read in smaller
1293 * chunks, there is no guarantee of atomicity. Since the display format
1294 * used, list of ranges of sequential numbers, is variable length,
1295 * and since these maps can change value dynamically, one could read
1296 * gibberish by doing partial reads while a list was changing.
1297 * A single large read to a buffer that crosses a page boundary is
1298 * ok, because the result being copied to user land is not recomputed
1299 * across a page fault.
1302 static int cpuset_sprintf_cpulist(char *page, struct cpuset *cs)
1304 cpumask_t mask;
1306 mutex_lock(&callback_mutex);
1307 mask = cs->cpus_allowed;
1308 mutex_unlock(&callback_mutex);
1310 return cpulist_scnprintf(page, PAGE_SIZE, mask);
1313 static int cpuset_sprintf_memlist(char *page, struct cpuset *cs)
1315 nodemask_t mask;
1317 mutex_lock(&callback_mutex);
1318 mask = cs->mems_allowed;
1319 mutex_unlock(&callback_mutex);
1321 return nodelist_scnprintf(page, PAGE_SIZE, mask);
1324 static ssize_t cpuset_common_file_read(struct cgroup *cont,
1325 struct cftype *cft,
1326 struct file *file,
1327 char __user *buf,
1328 size_t nbytes, loff_t *ppos)
1330 struct cpuset *cs = cgroup_cs(cont);
1331 cpuset_filetype_t type = cft->private;
1332 char *page;
1333 ssize_t retval = 0;
1334 char *s;
1336 if (!(page = (char *)__get_free_page(GFP_TEMPORARY)))
1337 return -ENOMEM;
1339 s = page;
1341 switch (type) {
1342 case FILE_CPULIST:
1343 s += cpuset_sprintf_cpulist(s, cs);
1344 break;
1345 case FILE_MEMLIST:
1346 s += cpuset_sprintf_memlist(s, cs);
1347 break;
1348 case FILE_CPU_EXCLUSIVE:
1349 *s++ = is_cpu_exclusive(cs) ? '1' : '0';
1350 break;
1351 case FILE_MEM_EXCLUSIVE:
1352 *s++ = is_mem_exclusive(cs) ? '1' : '0';
1353 break;
1354 case FILE_SCHED_LOAD_BALANCE:
1355 *s++ = is_sched_load_balance(cs) ? '1' : '0';
1356 break;
1357 case FILE_MEMORY_MIGRATE:
1358 *s++ = is_memory_migrate(cs) ? '1' : '0';
1359 break;
1360 case FILE_MEMORY_PRESSURE_ENABLED:
1361 *s++ = cpuset_memory_pressure_enabled ? '1' : '0';
1362 break;
1363 case FILE_MEMORY_PRESSURE:
1364 s += sprintf(s, "%d", fmeter_getrate(&cs->fmeter));
1365 break;
1366 case FILE_SPREAD_PAGE:
1367 *s++ = is_spread_page(cs) ? '1' : '0';
1368 break;
1369 case FILE_SPREAD_SLAB:
1370 *s++ = is_spread_slab(cs) ? '1' : '0';
1371 break;
1372 default:
1373 retval = -EINVAL;
1374 goto out;
1376 *s++ = '\n';
1378 retval = simple_read_from_buffer(buf, nbytes, ppos, page, s - page);
1379 out:
1380 free_page((unsigned long)page);
1381 return retval;
1389 * for the common functions, 'private' gives the type of file
1392 static struct cftype cft_cpus = {
1393 .name = "cpus",
1394 .read = cpuset_common_file_read,
1395 .write = cpuset_common_file_write,
1396 .private = FILE_CPULIST,
1399 static struct cftype cft_mems = {
1400 .name = "mems",
1401 .read = cpuset_common_file_read,
1402 .write = cpuset_common_file_write,
1403 .private = FILE_MEMLIST,
1406 static struct cftype cft_cpu_exclusive = {
1407 .name = "cpu_exclusive",
1408 .read = cpuset_common_file_read,
1409 .write = cpuset_common_file_write,
1410 .private = FILE_CPU_EXCLUSIVE,
1413 static struct cftype cft_mem_exclusive = {
1414 .name = "mem_exclusive",
1415 .read = cpuset_common_file_read,
1416 .write = cpuset_common_file_write,
1417 .private = FILE_MEM_EXCLUSIVE,
1420 static struct cftype cft_sched_load_balance = {
1421 .name = "sched_load_balance",
1422 .read = cpuset_common_file_read,
1423 .write = cpuset_common_file_write,
1424 .private = FILE_SCHED_LOAD_BALANCE,
1427 static struct cftype cft_memory_migrate = {
1428 .name = "memory_migrate",
1429 .read = cpuset_common_file_read,
1430 .write = cpuset_common_file_write,
1431 .private = FILE_MEMORY_MIGRATE,
1434 static struct cftype cft_memory_pressure_enabled = {
1435 .name = "memory_pressure_enabled",
1436 .read = cpuset_common_file_read,
1437 .write = cpuset_common_file_write,
1438 .private = FILE_MEMORY_PRESSURE_ENABLED,
1441 static struct cftype cft_memory_pressure = {
1442 .name = "memory_pressure",
1443 .read = cpuset_common_file_read,
1444 .write = cpuset_common_file_write,
1445 .private = FILE_MEMORY_PRESSURE,
1448 static struct cftype cft_spread_page = {
1449 .name = "memory_spread_page",
1450 .read = cpuset_common_file_read,
1451 .write = cpuset_common_file_write,
1452 .private = FILE_SPREAD_PAGE,
1455 static struct cftype cft_spread_slab = {
1456 .name = "memory_spread_slab",
1457 .read = cpuset_common_file_read,
1458 .write = cpuset_common_file_write,
1459 .private = FILE_SPREAD_SLAB,
1462 static int cpuset_populate(struct cgroup_subsys *ss, struct cgroup *cont)
1464 int err;
1466 if ((err = cgroup_add_file(cont, ss, &cft_cpus)) < 0)
1467 return err;
1468 if ((err = cgroup_add_file(cont, ss, &cft_mems)) < 0)
1469 return err;
1470 if ((err = cgroup_add_file(cont, ss, &cft_cpu_exclusive)) < 0)
1471 return err;
1472 if ((err = cgroup_add_file(cont, ss, &cft_mem_exclusive)) < 0)
1473 return err;
1474 if ((err = cgroup_add_file(cont, ss, &cft_memory_migrate)) < 0)
1475 return err;
1476 if ((err = cgroup_add_file(cont, ss, &cft_sched_load_balance)) < 0)
1477 return err;
1478 if ((err = cgroup_add_file(cont, ss, &cft_memory_pressure)) < 0)
1479 return err;
1480 if ((err = cgroup_add_file(cont, ss, &cft_spread_page)) < 0)
1481 return err;
1482 if ((err = cgroup_add_file(cont, ss, &cft_spread_slab)) < 0)
1483 return err;
1484 /* memory_pressure_enabled is in root cpuset only */
1485 if (err == 0 && !cont->parent)
1486 err = cgroup_add_file(cont, ss,
1487 &cft_memory_pressure_enabled);
1488 return 0;
1492 * post_clone() is called at the end of cgroup_clone().
1493 * 'cgroup' was just created automatically as a result of
1494 * a cgroup_clone(), and the current task is about to
1495 * be moved into 'cgroup'.
1497 * Currently we refuse to set up the cgroup - thereby
1498 * refusing the task to be entered, and as a result refusing
1499 * the sys_unshare() or clone() which initiated it - if any
1500 * sibling cpusets have exclusive cpus or mem.
1502 * If this becomes a problem for some users who wish to
1503 * allow that scenario, then cpuset_post_clone() could be
1504 * changed to grant parent->cpus_allowed-sibling_cpus_exclusive
1505 * (and likewise for mems) to the new cgroup. Called with cgroup_mutex
1506 * held.
1508 static void cpuset_post_clone(struct cgroup_subsys *ss,
1509 struct cgroup *cgroup)
1511 struct cgroup *parent, *child;
1512 struct cpuset *cs, *parent_cs;
1514 parent = cgroup->parent;
1515 list_for_each_entry(child, &parent->children, sibling) {
1516 cs = cgroup_cs(child);
1517 if (is_mem_exclusive(cs) || is_cpu_exclusive(cs))
1518 return;
1520 cs = cgroup_cs(cgroup);
1521 parent_cs = cgroup_cs(parent);
1523 cs->mems_allowed = parent_cs->mems_allowed;
1524 cs->cpus_allowed = parent_cs->cpus_allowed;
1525 return;
1529 * cpuset_create - create a cpuset
1530 * ss: cpuset cgroup subsystem
1531 * cont: control group that the new cpuset will be part of
1534 static struct cgroup_subsys_state *cpuset_create(
1535 struct cgroup_subsys *ss,
1536 struct cgroup *cont)
1538 struct cpuset *cs;
1539 struct cpuset *parent;
1541 if (!cont->parent) {
1542 /* This is early initialization for the top cgroup */
1543 top_cpuset.mems_generation = cpuset_mems_generation++;
1544 return &top_cpuset.css;
1546 parent = cgroup_cs(cont->parent);
1547 cs = kmalloc(sizeof(*cs), GFP_KERNEL);
1548 if (!cs)
1549 return ERR_PTR(-ENOMEM);
1551 cpuset_update_task_memory_state();
1552 cs->flags = 0;
1553 if (is_spread_page(parent))
1554 set_bit(CS_SPREAD_PAGE, &cs->flags);
1555 if (is_spread_slab(parent))
1556 set_bit(CS_SPREAD_SLAB, &cs->flags);
1557 set_bit(CS_SCHED_LOAD_BALANCE, &cs->flags);
1558 cs->cpus_allowed = CPU_MASK_NONE;
1559 cs->mems_allowed = NODE_MASK_NONE;
1560 cs->mems_generation = cpuset_mems_generation++;
1561 fmeter_init(&cs->fmeter);
1563 cs->parent = parent;
1564 number_of_cpusets++;
1565 return &cs->css ;
1569 * Locking note on the strange update_flag() call below:
1571 * If the cpuset being removed has its flag 'sched_load_balance'
1572 * enabled, then simulate turning sched_load_balance off, which
1573 * will call rebuild_sched_domains(). The get_online_cpus()
1574 * call in rebuild_sched_domains() must not be made while holding
1575 * callback_mutex. Elsewhere the kernel nests callback_mutex inside
1576 * get_online_cpus() calls. So the reverse nesting would risk an
1577 * ABBA deadlock.
1580 static void cpuset_destroy(struct cgroup_subsys *ss, struct cgroup *cont)
1582 struct cpuset *cs = cgroup_cs(cont);
1584 cpuset_update_task_memory_state();
1586 if (is_sched_load_balance(cs))
1587 update_flag(CS_SCHED_LOAD_BALANCE, cs, "0");
1589 number_of_cpusets--;
1590 kfree(cs);
1593 struct cgroup_subsys cpuset_subsys = {
1594 .name = "cpuset",
1595 .create = cpuset_create,
1596 .destroy = cpuset_destroy,
1597 .can_attach = cpuset_can_attach,
1598 .attach = cpuset_attach,
1599 .populate = cpuset_populate,
1600 .post_clone = cpuset_post_clone,
1601 .subsys_id = cpuset_subsys_id,
1602 .early_init = 1,
1606 * cpuset_init_early - just enough so that the calls to
1607 * cpuset_update_task_memory_state() in early init code
1608 * are harmless.
1611 int __init cpuset_init_early(void)
1613 top_cpuset.mems_generation = cpuset_mems_generation++;
1614 return 0;
1619 * cpuset_init - initialize cpusets at system boot
1621 * Description: Initialize top_cpuset and the cpuset internal file system,
1624 int __init cpuset_init(void)
1626 int err = 0;
1628 top_cpuset.cpus_allowed = CPU_MASK_ALL;
1629 top_cpuset.mems_allowed = NODE_MASK_ALL;
1631 fmeter_init(&top_cpuset.fmeter);
1632 top_cpuset.mems_generation = cpuset_mems_generation++;
1633 set_bit(CS_SCHED_LOAD_BALANCE, &top_cpuset.flags);
1635 err = register_filesystem(&cpuset_fs_type);
1636 if (err < 0)
1637 return err;
1639 number_of_cpusets = 1;
1640 return 0;
1644 * cpuset_do_move_task - move a given task to another cpuset
1645 * @tsk: pointer to task_struct the task to move
1646 * @scan: struct cgroup_scanner contained in its struct cpuset_hotplug_scanner
1648 * Called by cgroup_scan_tasks() for each task in a cgroup.
1649 * Return nonzero to stop the walk through the tasks.
1651 void cpuset_do_move_task(struct task_struct *tsk, struct cgroup_scanner *scan)
1653 struct cpuset_hotplug_scanner *chsp;
1655 chsp = container_of(scan, struct cpuset_hotplug_scanner, scan);
1656 cgroup_attach_task(chsp->to, tsk);
1660 * move_member_tasks_to_cpuset - move tasks from one cpuset to another
1661 * @from: cpuset in which the tasks currently reside
1662 * @to: cpuset to which the tasks will be moved
1664 * Called with cgroup_mutex held
1665 * callback_mutex must not be held, as cpuset_attach() will take it.
1667 * The cgroup_scan_tasks() function will scan all the tasks in a cgroup,
1668 * calling callback functions for each.
1670 static void move_member_tasks_to_cpuset(struct cpuset *from, struct cpuset *to)
1672 struct cpuset_hotplug_scanner scan;
1674 scan.scan.cg = from->css.cgroup;
1675 scan.scan.test_task = NULL; /* select all tasks in cgroup */
1676 scan.scan.process_task = cpuset_do_move_task;
1677 scan.scan.heap = NULL;
1678 scan.to = to->css.cgroup;
1680 if (cgroup_scan_tasks((struct cgroup_scanner *)&scan))
1681 printk(KERN_ERR "move_member_tasks_to_cpuset: "
1682 "cgroup_scan_tasks failed\n");
1686 * If common_cpu_mem_hotplug_unplug(), below, unplugs any CPUs
1687 * or memory nodes, we need to walk over the cpuset hierarchy,
1688 * removing that CPU or node from all cpusets. If this removes the
1689 * last CPU or node from a cpuset, then move the tasks in the empty
1690 * cpuset to its next-highest non-empty parent.
1692 * Called with cgroup_mutex held
1693 * callback_mutex must not be held, as cpuset_attach() will take it.
1695 static void remove_tasks_in_empty_cpuset(struct cpuset *cs)
1697 struct cpuset *parent;
1700 * The cgroup's css_sets list is in use if there are tasks
1701 * in the cpuset; the list is empty if there are none;
1702 * the cs->css.refcnt seems always 0.
1704 if (list_empty(&cs->css.cgroup->css_sets))
1705 return;
1708 * Find its next-highest non-empty parent, (top cpuset
1709 * has online cpus, so can't be empty).
1711 parent = cs->parent;
1712 while (cpus_empty(parent->cpus_allowed) ||
1713 nodes_empty(parent->mems_allowed))
1714 parent = parent->parent;
1716 move_member_tasks_to_cpuset(cs, parent);
1720 * Walk the specified cpuset subtree and look for empty cpusets.
1721 * The tasks of such cpuset must be moved to a parent cpuset.
1723 * Called with cgroup_mutex held. We take callback_mutex to modify
1724 * cpus_allowed and mems_allowed.
1726 * This walk processes the tree from top to bottom, completing one layer
1727 * before dropping down to the next. It always processes a node before
1728 * any of its children.
1730 * For now, since we lack memory hot unplug, we'll never see a cpuset
1731 * that has tasks along with an empty 'mems'. But if we did see such
1732 * a cpuset, we'd handle it just like we do if its 'cpus' was empty.
1734 static void scan_for_empty_cpusets(const struct cpuset *root)
1736 struct cpuset *cp; /* scans cpusets being updated */
1737 struct cpuset *child; /* scans child cpusets of cp */
1738 struct list_head queue;
1739 struct cgroup *cont;
1741 INIT_LIST_HEAD(&queue);
1743 list_add_tail((struct list_head *)&root->stack_list, &queue);
1745 while (!list_empty(&queue)) {
1746 cp = container_of(queue.next, struct cpuset, stack_list);
1747 list_del(queue.next);
1748 list_for_each_entry(cont, &cp->css.cgroup->children, sibling) {
1749 child = cgroup_cs(cont);
1750 list_add_tail(&child->stack_list, &queue);
1752 cont = cp->css.cgroup;
1754 /* Continue past cpusets with all cpus, mems online */
1755 if (cpus_subset(cp->cpus_allowed, cpu_online_map) &&
1756 nodes_subset(cp->mems_allowed, node_states[N_HIGH_MEMORY]))
1757 continue;
1759 /* Remove offline cpus and mems from this cpuset. */
1760 mutex_lock(&callback_mutex);
1761 cpus_and(cp->cpus_allowed, cp->cpus_allowed, cpu_online_map);
1762 nodes_and(cp->mems_allowed, cp->mems_allowed,
1763 node_states[N_HIGH_MEMORY]);
1764 mutex_unlock(&callback_mutex);
1766 /* Move tasks from the empty cpuset to a parent */
1767 if (cpus_empty(cp->cpus_allowed) ||
1768 nodes_empty(cp->mems_allowed))
1769 remove_tasks_in_empty_cpuset(cp);
1774 * The cpus_allowed and mems_allowed nodemasks in the top_cpuset track
1775 * cpu_online_map and node_states[N_HIGH_MEMORY]. Force the top cpuset to
1776 * track what's online after any CPU or memory node hotplug or unplug event.
1778 * Since there are two callers of this routine, one for CPU hotplug
1779 * events and one for memory node hotplug events, we could have coded
1780 * two separate routines here. We code it as a single common routine
1781 * in order to minimize text size.
1784 static void common_cpu_mem_hotplug_unplug(void)
1786 cgroup_lock();
1788 top_cpuset.cpus_allowed = cpu_online_map;
1789 top_cpuset.mems_allowed = node_states[N_HIGH_MEMORY];
1790 scan_for_empty_cpusets(&top_cpuset);
1792 cgroup_unlock();
1796 * The top_cpuset tracks what CPUs and Memory Nodes are online,
1797 * period. This is necessary in order to make cpusets transparent
1798 * (of no affect) on systems that are actively using CPU hotplug
1799 * but making no active use of cpusets.
1801 * This routine ensures that top_cpuset.cpus_allowed tracks
1802 * cpu_online_map on each CPU hotplug (cpuhp) event.
1805 static int cpuset_handle_cpuhp(struct notifier_block *unused_nb,
1806 unsigned long phase, void *unused_cpu)
1808 if (phase == CPU_DYING || phase == CPU_DYING_FROZEN)
1809 return NOTIFY_DONE;
1811 common_cpu_mem_hotplug_unplug();
1812 return 0;
1815 #ifdef CONFIG_MEMORY_HOTPLUG
1817 * Keep top_cpuset.mems_allowed tracking node_states[N_HIGH_MEMORY].
1818 * Call this routine anytime after you change
1819 * node_states[N_HIGH_MEMORY].
1820 * See also the previous routine cpuset_handle_cpuhp().
1823 void cpuset_track_online_nodes(void)
1825 common_cpu_mem_hotplug_unplug();
1827 #endif
1830 * cpuset_init_smp - initialize cpus_allowed
1832 * Description: Finish top cpuset after cpu, node maps are initialized
1835 void __init cpuset_init_smp(void)
1837 top_cpuset.cpus_allowed = cpu_online_map;
1838 top_cpuset.mems_allowed = node_states[N_HIGH_MEMORY];
1840 hotcpu_notifier(cpuset_handle_cpuhp, 0);
1845 * cpuset_cpus_allowed - return cpus_allowed mask from a tasks cpuset.
1846 * @tsk: pointer to task_struct from which to obtain cpuset->cpus_allowed.
1848 * Description: Returns the cpumask_t cpus_allowed of the cpuset
1849 * attached to the specified @tsk. Guaranteed to return some non-empty
1850 * subset of cpu_online_map, even if this means going outside the
1851 * tasks cpuset.
1854 cpumask_t cpuset_cpus_allowed(struct task_struct *tsk)
1856 cpumask_t mask;
1858 mutex_lock(&callback_mutex);
1859 mask = cpuset_cpus_allowed_locked(tsk);
1860 mutex_unlock(&callback_mutex);
1862 return mask;
1866 * cpuset_cpus_allowed_locked - return cpus_allowed mask from a tasks cpuset.
1867 * Must be called with callback_mutex held.
1869 cpumask_t cpuset_cpus_allowed_locked(struct task_struct *tsk)
1871 cpumask_t mask;
1873 task_lock(tsk);
1874 guarantee_online_cpus(task_cs(tsk), &mask);
1875 task_unlock(tsk);
1877 return mask;
1880 void cpuset_init_current_mems_allowed(void)
1882 current->mems_allowed = NODE_MASK_ALL;
1886 * cpuset_mems_allowed - return mems_allowed mask from a tasks cpuset.
1887 * @tsk: pointer to task_struct from which to obtain cpuset->mems_allowed.
1889 * Description: Returns the nodemask_t mems_allowed of the cpuset
1890 * attached to the specified @tsk. Guaranteed to return some non-empty
1891 * subset of node_states[N_HIGH_MEMORY], even if this means going outside the
1892 * tasks cpuset.
1895 nodemask_t cpuset_mems_allowed(struct task_struct *tsk)
1897 nodemask_t mask;
1899 mutex_lock(&callback_mutex);
1900 task_lock(tsk);
1901 guarantee_online_mems(task_cs(tsk), &mask);
1902 task_unlock(tsk);
1903 mutex_unlock(&callback_mutex);
1905 return mask;
1909 * cpuset_zonelist_valid_mems_allowed - check zonelist vs. curremt mems_allowed
1910 * @zl: the zonelist to be checked
1912 * Are any of the nodes on zonelist zl allowed in current->mems_allowed?
1914 int cpuset_zonelist_valid_mems_allowed(struct zonelist *zl)
1916 int i;
1918 for (i = 0; zl->zones[i]; i++) {
1919 int nid = zone_to_nid(zl->zones[i]);
1921 if (node_isset(nid, current->mems_allowed))
1922 return 1;
1924 return 0;
1928 * nearest_exclusive_ancestor() - Returns the nearest mem_exclusive
1929 * ancestor to the specified cpuset. Call holding callback_mutex.
1930 * If no ancestor is mem_exclusive (an unusual configuration), then
1931 * returns the root cpuset.
1933 static const struct cpuset *nearest_exclusive_ancestor(const struct cpuset *cs)
1935 while (!is_mem_exclusive(cs) && cs->parent)
1936 cs = cs->parent;
1937 return cs;
1941 * cpuset_zone_allowed_softwall - Can we allocate on zone z's memory node?
1942 * @z: is this zone on an allowed node?
1943 * @gfp_mask: memory allocation flags
1945 * If we're in interrupt, yes, we can always allocate. If
1946 * __GFP_THISNODE is set, yes, we can always allocate. If zone
1947 * z's node is in our tasks mems_allowed, yes. If it's not a
1948 * __GFP_HARDWALL request and this zone's nodes is in the nearest
1949 * mem_exclusive cpuset ancestor to this tasks cpuset, yes.
1950 * If the task has been OOM killed and has access to memory reserves
1951 * as specified by the TIF_MEMDIE flag, yes.
1952 * Otherwise, no.
1954 * If __GFP_HARDWALL is set, cpuset_zone_allowed_softwall()
1955 * reduces to cpuset_zone_allowed_hardwall(). Otherwise,
1956 * cpuset_zone_allowed_softwall() might sleep, and might allow a zone
1957 * from an enclosing cpuset.
1959 * cpuset_zone_allowed_hardwall() only handles the simpler case of
1960 * hardwall cpusets, and never sleeps.
1962 * The __GFP_THISNODE placement logic is really handled elsewhere,
1963 * by forcibly using a zonelist starting at a specified node, and by
1964 * (in get_page_from_freelist()) refusing to consider the zones for
1965 * any node on the zonelist except the first. By the time any such
1966 * calls get to this routine, we should just shut up and say 'yes'.
1968 * GFP_USER allocations are marked with the __GFP_HARDWALL bit,
1969 * and do not allow allocations outside the current tasks cpuset
1970 * unless the task has been OOM killed as is marked TIF_MEMDIE.
1971 * GFP_KERNEL allocations are not so marked, so can escape to the
1972 * nearest enclosing mem_exclusive ancestor cpuset.
1974 * Scanning up parent cpusets requires callback_mutex. The
1975 * __alloc_pages() routine only calls here with __GFP_HARDWALL bit
1976 * _not_ set if it's a GFP_KERNEL allocation, and all nodes in the
1977 * current tasks mems_allowed came up empty on the first pass over
1978 * the zonelist. So only GFP_KERNEL allocations, if all nodes in the
1979 * cpuset are short of memory, might require taking the callback_mutex
1980 * mutex.
1982 * The first call here from mm/page_alloc:get_page_from_freelist()
1983 * has __GFP_HARDWALL set in gfp_mask, enforcing hardwall cpusets,
1984 * so no allocation on a node outside the cpuset is allowed (unless
1985 * in interrupt, of course).
1987 * The second pass through get_page_from_freelist() doesn't even call
1988 * here for GFP_ATOMIC calls. For those calls, the __alloc_pages()
1989 * variable 'wait' is not set, and the bit ALLOC_CPUSET is not set
1990 * in alloc_flags. That logic and the checks below have the combined
1991 * affect that:
1992 * in_interrupt - any node ok (current task context irrelevant)
1993 * GFP_ATOMIC - any node ok
1994 * TIF_MEMDIE - any node ok
1995 * GFP_KERNEL - any node in enclosing mem_exclusive cpuset ok
1996 * GFP_USER - only nodes in current tasks mems allowed ok.
1998 * Rule:
1999 * Don't call cpuset_zone_allowed_softwall if you can't sleep, unless you
2000 * pass in the __GFP_HARDWALL flag set in gfp_flag, which disables
2001 * the code that might scan up ancestor cpusets and sleep.
2004 int __cpuset_zone_allowed_softwall(struct zone *z, gfp_t gfp_mask)
2006 int node; /* node that zone z is on */
2007 const struct cpuset *cs; /* current cpuset ancestors */
2008 int allowed; /* is allocation in zone z allowed? */
2010 if (in_interrupt() || (gfp_mask & __GFP_THISNODE))
2011 return 1;
2012 node = zone_to_nid(z);
2013 might_sleep_if(!(gfp_mask & __GFP_HARDWALL));
2014 if (node_isset(node, current->mems_allowed))
2015 return 1;
2017 * Allow tasks that have access to memory reserves because they have
2018 * been OOM killed to get memory anywhere.
2020 if (unlikely(test_thread_flag(TIF_MEMDIE)))
2021 return 1;
2022 if (gfp_mask & __GFP_HARDWALL) /* If hardwall request, stop here */
2023 return 0;
2025 if (current->flags & PF_EXITING) /* Let dying task have memory */
2026 return 1;
2028 /* Not hardwall and node outside mems_allowed: scan up cpusets */
2029 mutex_lock(&callback_mutex);
2031 task_lock(current);
2032 cs = nearest_exclusive_ancestor(task_cs(current));
2033 task_unlock(current);
2035 allowed = node_isset(node, cs->mems_allowed);
2036 mutex_unlock(&callback_mutex);
2037 return allowed;
2041 * cpuset_zone_allowed_hardwall - Can we allocate on zone z's memory node?
2042 * @z: is this zone on an allowed node?
2043 * @gfp_mask: memory allocation flags
2045 * If we're in interrupt, yes, we can always allocate.
2046 * If __GFP_THISNODE is set, yes, we can always allocate. If zone
2047 * z's node is in our tasks mems_allowed, yes. If the task has been
2048 * OOM killed and has access to memory reserves as specified by the
2049 * TIF_MEMDIE flag, yes. Otherwise, no.
2051 * The __GFP_THISNODE placement logic is really handled elsewhere,
2052 * by forcibly using a zonelist starting at a specified node, and by
2053 * (in get_page_from_freelist()) refusing to consider the zones for
2054 * any node on the zonelist except the first. By the time any such
2055 * calls get to this routine, we should just shut up and say 'yes'.
2057 * Unlike the cpuset_zone_allowed_softwall() variant, above,
2058 * this variant requires that the zone be in the current tasks
2059 * mems_allowed or that we're in interrupt. It does not scan up the
2060 * cpuset hierarchy for the nearest enclosing mem_exclusive cpuset.
2061 * It never sleeps.
2064 int __cpuset_zone_allowed_hardwall(struct zone *z, gfp_t gfp_mask)
2066 int node; /* node that zone z is on */
2068 if (in_interrupt() || (gfp_mask & __GFP_THISNODE))
2069 return 1;
2070 node = zone_to_nid(z);
2071 if (node_isset(node, current->mems_allowed))
2072 return 1;
2074 * Allow tasks that have access to memory reserves because they have
2075 * been OOM killed to get memory anywhere.
2077 if (unlikely(test_thread_flag(TIF_MEMDIE)))
2078 return 1;
2079 return 0;
2083 * cpuset_lock - lock out any changes to cpuset structures
2085 * The out of memory (oom) code needs to mutex_lock cpusets
2086 * from being changed while it scans the tasklist looking for a
2087 * task in an overlapping cpuset. Expose callback_mutex via this
2088 * cpuset_lock() routine, so the oom code can lock it, before
2089 * locking the task list. The tasklist_lock is a spinlock, so
2090 * must be taken inside callback_mutex.
2093 void cpuset_lock(void)
2095 mutex_lock(&callback_mutex);
2099 * cpuset_unlock - release lock on cpuset changes
2101 * Undo the lock taken in a previous cpuset_lock() call.
2104 void cpuset_unlock(void)
2106 mutex_unlock(&callback_mutex);
2110 * cpuset_mem_spread_node() - On which node to begin search for a page
2112 * If a task is marked PF_SPREAD_PAGE or PF_SPREAD_SLAB (as for
2113 * tasks in a cpuset with is_spread_page or is_spread_slab set),
2114 * and if the memory allocation used cpuset_mem_spread_node()
2115 * to determine on which node to start looking, as it will for
2116 * certain page cache or slab cache pages such as used for file
2117 * system buffers and inode caches, then instead of starting on the
2118 * local node to look for a free page, rather spread the starting
2119 * node around the tasks mems_allowed nodes.
2121 * We don't have to worry about the returned node being offline
2122 * because "it can't happen", and even if it did, it would be ok.
2124 * The routines calling guarantee_online_mems() are careful to
2125 * only set nodes in task->mems_allowed that are online. So it
2126 * should not be possible for the following code to return an
2127 * offline node. But if it did, that would be ok, as this routine
2128 * is not returning the node where the allocation must be, only
2129 * the node where the search should start. The zonelist passed to
2130 * __alloc_pages() will include all nodes. If the slab allocator
2131 * is passed an offline node, it will fall back to the local node.
2132 * See kmem_cache_alloc_node().
2135 int cpuset_mem_spread_node(void)
2137 int node;
2139 node = next_node(current->cpuset_mem_spread_rotor, current->mems_allowed);
2140 if (node == MAX_NUMNODES)
2141 node = first_node(current->mems_allowed);
2142 current->cpuset_mem_spread_rotor = node;
2143 return node;
2145 EXPORT_SYMBOL_GPL(cpuset_mem_spread_node);
2148 * cpuset_mems_allowed_intersects - Does @tsk1's mems_allowed intersect @tsk2's?
2149 * @tsk1: pointer to task_struct of some task.
2150 * @tsk2: pointer to task_struct of some other task.
2152 * Description: Return true if @tsk1's mems_allowed intersects the
2153 * mems_allowed of @tsk2. Used by the OOM killer to determine if
2154 * one of the task's memory usage might impact the memory available
2155 * to the other.
2158 int cpuset_mems_allowed_intersects(const struct task_struct *tsk1,
2159 const struct task_struct *tsk2)
2161 return nodes_intersects(tsk1->mems_allowed, tsk2->mems_allowed);
2165 * Collection of memory_pressure is suppressed unless
2166 * this flag is enabled by writing "1" to the special
2167 * cpuset file 'memory_pressure_enabled' in the root cpuset.
2170 int cpuset_memory_pressure_enabled __read_mostly;
2173 * cpuset_memory_pressure_bump - keep stats of per-cpuset reclaims.
2175 * Keep a running average of the rate of synchronous (direct)
2176 * page reclaim efforts initiated by tasks in each cpuset.
2178 * This represents the rate at which some task in the cpuset
2179 * ran low on memory on all nodes it was allowed to use, and
2180 * had to enter the kernels page reclaim code in an effort to
2181 * create more free memory by tossing clean pages or swapping
2182 * or writing dirty pages.
2184 * Display to user space in the per-cpuset read-only file
2185 * "memory_pressure". Value displayed is an integer
2186 * representing the recent rate of entry into the synchronous
2187 * (direct) page reclaim by any task attached to the cpuset.
2190 void __cpuset_memory_pressure_bump(void)
2192 task_lock(current);
2193 fmeter_markevent(&task_cs(current)->fmeter);
2194 task_unlock(current);
2197 #ifdef CONFIG_PROC_PID_CPUSET
2199 * proc_cpuset_show()
2200 * - Print tasks cpuset path into seq_file.
2201 * - Used for /proc/<pid>/cpuset.
2202 * - No need to task_lock(tsk) on this tsk->cpuset reference, as it
2203 * doesn't really matter if tsk->cpuset changes after we read it,
2204 * and we take cgroup_mutex, keeping cpuset_attach() from changing it
2205 * anyway.
2207 static int proc_cpuset_show(struct seq_file *m, void *unused_v)
2209 struct pid *pid;
2210 struct task_struct *tsk;
2211 char *buf;
2212 struct cgroup_subsys_state *css;
2213 int retval;
2215 retval = -ENOMEM;
2216 buf = kmalloc(PAGE_SIZE, GFP_KERNEL);
2217 if (!buf)
2218 goto out;
2220 retval = -ESRCH;
2221 pid = m->private;
2222 tsk = get_pid_task(pid, PIDTYPE_PID);
2223 if (!tsk)
2224 goto out_free;
2226 retval = -EINVAL;
2227 cgroup_lock();
2228 css = task_subsys_state(tsk, cpuset_subsys_id);
2229 retval = cgroup_path(css->cgroup, buf, PAGE_SIZE);
2230 if (retval < 0)
2231 goto out_unlock;
2232 seq_puts(m, buf);
2233 seq_putc(m, '\n');
2234 out_unlock:
2235 cgroup_unlock();
2236 put_task_struct(tsk);
2237 out_free:
2238 kfree(buf);
2239 out:
2240 return retval;
2243 static int cpuset_open(struct inode *inode, struct file *file)
2245 struct pid *pid = PROC_I(inode)->pid;
2246 return single_open(file, proc_cpuset_show, pid);
2249 const struct file_operations proc_cpuset_operations = {
2250 .open = cpuset_open,
2251 .read = seq_read,
2252 .llseek = seq_lseek,
2253 .release = single_release,
2255 #endif /* CONFIG_PROC_PID_CPUSET */
2257 /* Display task cpus_allowed, mems_allowed in /proc/<pid>/status file. */
2258 void cpuset_task_status_allowed(struct seq_file *m, struct task_struct *task)
2260 seq_printf(m, "Cpus_allowed:\t");
2261 m->count += cpumask_scnprintf(m->buf + m->count, m->size - m->count,
2262 task->cpus_allowed);
2263 seq_printf(m, "\n");
2264 seq_printf(m, "Mems_allowed:\t");
2265 m->count += nodemask_scnprintf(m->buf + m->count, m->size - m->count,
2266 task->mems_allowed);
2267 seq_printf(m, "\n");