4 * Processor and Memory placement constraints for sets of tasks.
6 * Copyright (C) 2003 BULL SA.
7 * Copyright (C) 2004-2007 Silicon Graphics, Inc.
8 * Copyright (C) 2006 Google, Inc
10 * Portions derived from Patrick Mochel's sysfs code.
11 * sysfs is Copyright (c) 2001-3 Patrick Mochel
13 * 2003-10-10 Written by Simon Derr.
14 * 2003-10-22 Updates by Stephen Hemminger.
15 * 2004 May-July Rework by Paul Jackson.
16 * 2006 Rework by Paul Menage to use generic cgroups
17 * 2008 Rework of the scheduler domains and CPU hotplug handling
20 * This file is subject to the terms and conditions of the GNU General Public
21 * License. See the file COPYING in the main directory of the Linux
22 * distribution for more details.
25 #include <linux/cpu.h>
26 #include <linux/cpumask.h>
27 #include <linux/cpuset.h>
28 #include <linux/err.h>
29 #include <linux/errno.h>
30 #include <linux/file.h>
32 #include <linux/init.h>
33 #include <linux/interrupt.h>
34 #include <linux/kernel.h>
35 #include <linux/kmod.h>
36 #include <linux/list.h>
37 #include <linux/mempolicy.h>
39 #include <linux/memory.h>
40 #include <linux/module.h>
41 #include <linux/mount.h>
42 #include <linux/namei.h>
43 #include <linux/pagemap.h>
44 #include <linux/proc_fs.h>
45 #include <linux/rcupdate.h>
46 #include <linux/sched.h>
47 #include <linux/seq_file.h>
48 #include <linux/security.h>
49 #include <linux/slab.h>
50 #include <linux/spinlock.h>
51 #include <linux/stat.h>
52 #include <linux/string.h>
53 #include <linux/time.h>
54 #include <linux/backing-dev.h>
55 #include <linux/sort.h>
57 #include <asm/uaccess.h>
58 #include <asm/atomic.h>
59 #include <linux/mutex.h>
60 #include <linux/workqueue.h>
61 #include <linux/cgroup.h>
64 * Tracks how many cpusets are currently defined in system.
65 * When there is only one cpuset (the root cpuset) we can
66 * short circuit some hooks.
68 int number_of_cpusets __read_mostly;
70 /* Forward declare cgroup structures */
71 struct cgroup_subsys cpuset_subsys;
74 /* See "Frequency meter" comments, below. */
77 int cnt; /* unprocessed events count */
78 int val; /* most recent output value */
79 time_t time; /* clock (secs) when val computed */
80 spinlock_t lock; /* guards read or write of above */
84 struct cgroup_subsys_state css;
86 unsigned long flags; /* "unsigned long" so bitops work */
87 cpumask_t cpus_allowed; /* CPUs allowed to tasks in cpuset */
88 nodemask_t mems_allowed; /* Memory Nodes allowed to tasks */
90 struct cpuset *parent; /* my parent */
93 * Copy of global cpuset_mems_generation as of the most
94 * recent time this cpuset changed its mems_allowed.
98 struct fmeter fmeter; /* memory_pressure filter */
100 /* partition number for rebuild_sched_domains() */
103 /* for custom sched domain */
104 int relax_domain_level;
106 /* used for walking a cpuset heirarchy */
107 struct list_head stack_list;
110 /* Retrieve the cpuset for a cgroup */
111 static inline struct cpuset *cgroup_cs(struct cgroup *cont)
113 return container_of(cgroup_subsys_state(cont, cpuset_subsys_id),
117 /* Retrieve the cpuset for a task */
118 static inline struct cpuset *task_cs(struct task_struct *task)
120 return container_of(task_subsys_state(task, cpuset_subsys_id),
123 struct cpuset_hotplug_scanner {
124 struct cgroup_scanner scan;
128 /* bits in struct cpuset flags field */
134 CS_SCHED_LOAD_BALANCE,
139 /* convenient tests for these bits */
140 static inline int is_cpu_exclusive(const struct cpuset *cs)
142 return test_bit(CS_CPU_EXCLUSIVE, &cs->flags);
145 static inline int is_mem_exclusive(const struct cpuset *cs)
147 return test_bit(CS_MEM_EXCLUSIVE, &cs->flags);
150 static inline int is_mem_hardwall(const struct cpuset *cs)
152 return test_bit(CS_MEM_HARDWALL, &cs->flags);
155 static inline int is_sched_load_balance(const struct cpuset *cs)
157 return test_bit(CS_SCHED_LOAD_BALANCE, &cs->flags);
160 static inline int is_memory_migrate(const struct cpuset *cs)
162 return test_bit(CS_MEMORY_MIGRATE, &cs->flags);
165 static inline int is_spread_page(const struct cpuset *cs)
167 return test_bit(CS_SPREAD_PAGE, &cs->flags);
170 static inline int is_spread_slab(const struct cpuset *cs)
172 return test_bit(CS_SPREAD_SLAB, &cs->flags);
176 * Increment this integer everytime any cpuset changes its
177 * mems_allowed value. Users of cpusets can track this generation
178 * number, and avoid having to lock and reload mems_allowed unless
179 * the cpuset they're using changes generation.
181 * A single, global generation is needed because cpuset_attach_task() could
182 * reattach a task to a different cpuset, which must not have its
183 * generation numbers aliased with those of that tasks previous cpuset.
185 * Generations are needed for mems_allowed because one task cannot
186 * modify another's memory placement. So we must enable every task,
187 * on every visit to __alloc_pages(), to efficiently check whether
188 * its current->cpuset->mems_allowed has changed, requiring an update
189 * of its current->mems_allowed.
191 * Since writes to cpuset_mems_generation are guarded by the cgroup lock
192 * there is no need to mark it atomic.
194 static int cpuset_mems_generation;
196 static struct cpuset top_cpuset = {
197 .flags = ((1 << CS_CPU_EXCLUSIVE) | (1 << CS_MEM_EXCLUSIVE)),
198 .cpus_allowed = CPU_MASK_ALL,
199 .mems_allowed = NODE_MASK_ALL,
203 * There are two global mutexes guarding cpuset structures. The first
204 * is the main control groups cgroup_mutex, accessed via
205 * cgroup_lock()/cgroup_unlock(). The second is the cpuset-specific
206 * callback_mutex, below. They can nest. It is ok to first take
207 * cgroup_mutex, then nest callback_mutex. We also require taking
208 * task_lock() when dereferencing a task's cpuset pointer. See "The
209 * task_lock() exception", at the end of this comment.
211 * A task must hold both mutexes to modify cpusets. If a task
212 * holds cgroup_mutex, then it blocks others wanting that mutex,
213 * ensuring that it is the only task able to also acquire callback_mutex
214 * and be able to modify cpusets. It can perform various checks on
215 * the cpuset structure first, knowing nothing will change. It can
216 * also allocate memory while just holding cgroup_mutex. While it is
217 * performing these checks, various callback routines can briefly
218 * acquire callback_mutex to query cpusets. Once it is ready to make
219 * the changes, it takes callback_mutex, blocking everyone else.
221 * Calls to the kernel memory allocator can not be made while holding
222 * callback_mutex, as that would risk double tripping on callback_mutex
223 * from one of the callbacks into the cpuset code from within
226 * If a task is only holding callback_mutex, then it has read-only
229 * The task_struct fields mems_allowed and mems_generation may only
230 * be accessed in the context of that task, so require no locks.
232 * The cpuset_common_file_read() handlers only hold callback_mutex across
233 * small pieces of code, such as when reading out possibly multi-word
234 * cpumasks and nodemasks.
236 * Accessing a task's cpuset should be done in accordance with the
237 * guidelines for accessing subsystem state in kernel/cgroup.c
240 static DEFINE_MUTEX(callback_mutex);
243 * cpuset_buffer_lock protects both the cpuset_name and cpuset_nodelist
244 * buffers. They are statically allocated to prevent using excess stack
245 * when calling cpuset_print_task_mems_allowed().
247 #define CPUSET_NAME_LEN (128)
248 #define CPUSET_NODELIST_LEN (256)
249 static char cpuset_name[CPUSET_NAME_LEN];
250 static char cpuset_nodelist[CPUSET_NODELIST_LEN];
251 static DEFINE_SPINLOCK(cpuset_buffer_lock);
254 * This is ugly, but preserves the userspace API for existing cpuset
255 * users. If someone tries to mount the "cpuset" filesystem, we
256 * silently switch it to mount "cgroup" instead
258 static int cpuset_get_sb(struct file_system_type *fs_type,
259 int flags, const char *unused_dev_name,
260 void *data, struct vfsmount *mnt)
262 struct file_system_type *cgroup_fs = get_fs_type("cgroup");
267 "release_agent=/sbin/cpuset_release_agent";
268 ret = cgroup_fs->get_sb(cgroup_fs, flags,
269 unused_dev_name, mountopts, mnt);
270 put_filesystem(cgroup_fs);
275 static struct file_system_type cpuset_fs_type = {
277 .get_sb = cpuset_get_sb,
281 * Return in *pmask the portion of a cpusets's cpus_allowed that
282 * are online. If none are online, walk up the cpuset hierarchy
283 * until we find one that does have some online cpus. If we get
284 * all the way to the top and still haven't found any online cpus,
285 * return cpu_online_map. Or if passed a NULL cs from an exit'ing
286 * task, return cpu_online_map.
288 * One way or another, we guarantee to return some non-empty subset
291 * Call with callback_mutex held.
294 static void guarantee_online_cpus(const struct cpuset *cs, cpumask_t *pmask)
296 while (cs && !cpus_intersects(cs->cpus_allowed, cpu_online_map))
299 cpus_and(*pmask, cs->cpus_allowed, cpu_online_map);
301 *pmask = cpu_online_map;
302 BUG_ON(!cpus_intersects(*pmask, cpu_online_map));
306 * Return in *pmask the portion of a cpusets's mems_allowed that
307 * are online, with memory. If none are online with memory, walk
308 * up the cpuset hierarchy until we find one that does have some
309 * online mems. If we get all the way to the top and still haven't
310 * found any online mems, return node_states[N_HIGH_MEMORY].
312 * One way or another, we guarantee to return some non-empty subset
313 * of node_states[N_HIGH_MEMORY].
315 * Call with callback_mutex held.
318 static void guarantee_online_mems(const struct cpuset *cs, nodemask_t *pmask)
320 while (cs && !nodes_intersects(cs->mems_allowed,
321 node_states[N_HIGH_MEMORY]))
324 nodes_and(*pmask, cs->mems_allowed,
325 node_states[N_HIGH_MEMORY]);
327 *pmask = node_states[N_HIGH_MEMORY];
328 BUG_ON(!nodes_intersects(*pmask, node_states[N_HIGH_MEMORY]));
332 * cpuset_update_task_memory_state - update task memory placement
334 * If the current tasks cpusets mems_allowed changed behind our
335 * backs, update current->mems_allowed, mems_generation and task NUMA
336 * mempolicy to the new value.
338 * Task mempolicy is updated by rebinding it relative to the
339 * current->cpuset if a task has its memory placement changed.
340 * Do not call this routine if in_interrupt().
342 * Call without callback_mutex or task_lock() held. May be
343 * called with or without cgroup_mutex held. Thanks in part to
344 * 'the_top_cpuset_hack', the task's cpuset pointer will never
345 * be NULL. This routine also might acquire callback_mutex during
348 * Reading current->cpuset->mems_generation doesn't need task_lock
349 * to guard the current->cpuset derefence, because it is guarded
350 * from concurrent freeing of current->cpuset using RCU.
352 * The rcu_dereference() is technically probably not needed,
353 * as I don't actually mind if I see a new cpuset pointer but
354 * an old value of mems_generation. However this really only
355 * matters on alpha systems using cpusets heavily. If I dropped
356 * that rcu_dereference(), it would save them a memory barrier.
357 * For all other arch's, rcu_dereference is a no-op anyway, and for
358 * alpha systems not using cpusets, another planned optimization,
359 * avoiding the rcu critical section for tasks in the root cpuset
360 * which is statically allocated, so can't vanish, will make this
361 * irrelevant. Better to use RCU as intended, than to engage in
362 * some cute trick to save a memory barrier that is impossible to
363 * test, for alpha systems using cpusets heavily, which might not
366 * This routine is needed to update the per-task mems_allowed data,
367 * within the tasks context, when it is trying to allocate memory
368 * (in various mm/mempolicy.c routines) and notices that some other
369 * task has been modifying its cpuset.
372 void cpuset_update_task_memory_state(void)
374 int my_cpusets_mem_gen;
375 struct task_struct *tsk = current;
379 my_cpusets_mem_gen = task_cs(tsk)->mems_generation;
382 if (my_cpusets_mem_gen != tsk->cpuset_mems_generation) {
383 mutex_lock(&callback_mutex);
385 cs = task_cs(tsk); /* Maybe changed when task not locked */
386 guarantee_online_mems(cs, &tsk->mems_allowed);
387 tsk->cpuset_mems_generation = cs->mems_generation;
388 if (is_spread_page(cs))
389 tsk->flags |= PF_SPREAD_PAGE;
391 tsk->flags &= ~PF_SPREAD_PAGE;
392 if (is_spread_slab(cs))
393 tsk->flags |= PF_SPREAD_SLAB;
395 tsk->flags &= ~PF_SPREAD_SLAB;
397 mutex_unlock(&callback_mutex);
398 mpol_rebind_task(tsk, &tsk->mems_allowed);
403 * is_cpuset_subset(p, q) - Is cpuset p a subset of cpuset q?
405 * One cpuset is a subset of another if all its allowed CPUs and
406 * Memory Nodes are a subset of the other, and its exclusive flags
407 * are only set if the other's are set. Call holding cgroup_mutex.
410 static int is_cpuset_subset(const struct cpuset *p, const struct cpuset *q)
412 return cpus_subset(p->cpus_allowed, q->cpus_allowed) &&
413 nodes_subset(p->mems_allowed, q->mems_allowed) &&
414 is_cpu_exclusive(p) <= is_cpu_exclusive(q) &&
415 is_mem_exclusive(p) <= is_mem_exclusive(q);
419 * validate_change() - Used to validate that any proposed cpuset change
420 * follows the structural rules for cpusets.
422 * If we replaced the flag and mask values of the current cpuset
423 * (cur) with those values in the trial cpuset (trial), would
424 * our various subset and exclusive rules still be valid? Presumes
427 * 'cur' is the address of an actual, in-use cpuset. Operations
428 * such as list traversal that depend on the actual address of the
429 * cpuset in the list must use cur below, not trial.
431 * 'trial' is the address of bulk structure copy of cur, with
432 * perhaps one or more of the fields cpus_allowed, mems_allowed,
433 * or flags changed to new, trial values.
435 * Return 0 if valid, -errno if not.
438 static int validate_change(const struct cpuset *cur, const struct cpuset *trial)
441 struct cpuset *c, *par;
443 /* Each of our child cpusets must be a subset of us */
444 list_for_each_entry(cont, &cur->css.cgroup->children, sibling) {
445 if (!is_cpuset_subset(cgroup_cs(cont), trial))
449 /* Remaining checks don't apply to root cpuset */
450 if (cur == &top_cpuset)
455 /* We must be a subset of our parent cpuset */
456 if (!is_cpuset_subset(trial, par))
460 * If either I or some sibling (!= me) is exclusive, we can't
463 list_for_each_entry(cont, &par->css.cgroup->children, sibling) {
465 if ((is_cpu_exclusive(trial) || is_cpu_exclusive(c)) &&
467 cpus_intersects(trial->cpus_allowed, c->cpus_allowed))
469 if ((is_mem_exclusive(trial) || is_mem_exclusive(c)) &&
471 nodes_intersects(trial->mems_allowed, c->mems_allowed))
475 /* Cpusets with tasks can't have empty cpus_allowed or mems_allowed */
476 if (cgroup_task_count(cur->css.cgroup)) {
477 if (cpus_empty(trial->cpus_allowed) ||
478 nodes_empty(trial->mems_allowed)) {
487 * Helper routine for generate_sched_domains().
488 * Do cpusets a, b have overlapping cpus_allowed masks?
490 static int cpusets_overlap(struct cpuset *a, struct cpuset *b)
492 return cpus_intersects(a->cpus_allowed, b->cpus_allowed);
496 update_domain_attr(struct sched_domain_attr *dattr, struct cpuset *c)
498 if (dattr->relax_domain_level < c->relax_domain_level)
499 dattr->relax_domain_level = c->relax_domain_level;
504 update_domain_attr_tree(struct sched_domain_attr *dattr, struct cpuset *c)
508 list_add(&c->stack_list, &q);
509 while (!list_empty(&q)) {
512 struct cpuset *child;
514 cp = list_first_entry(&q, struct cpuset, stack_list);
517 if (cpus_empty(cp->cpus_allowed))
520 if (is_sched_load_balance(cp))
521 update_domain_attr(dattr, cp);
523 list_for_each_entry(cont, &cp->css.cgroup->children, sibling) {
524 child = cgroup_cs(cont);
525 list_add_tail(&child->stack_list, &q);
531 * generate_sched_domains()
533 * This function builds a partial partition of the systems CPUs
534 * A 'partial partition' is a set of non-overlapping subsets whose
535 * union is a subset of that set.
536 * The output of this function needs to be passed to kernel/sched.c
537 * partition_sched_domains() routine, which will rebuild the scheduler's
538 * load balancing domains (sched domains) as specified by that partial
541 * See "What is sched_load_balance" in Documentation/cpusets.txt
542 * for a background explanation of this.
544 * Does not return errors, on the theory that the callers of this
545 * routine would rather not worry about failures to rebuild sched
546 * domains when operating in the severe memory shortage situations
547 * that could cause allocation failures below.
549 * Must be called with cgroup_lock held.
551 * The three key local variables below are:
552 * q - a linked-list queue of cpuset pointers, used to implement a
553 * top-down scan of all cpusets. This scan loads a pointer
554 * to each cpuset marked is_sched_load_balance into the
555 * array 'csa'. For our purposes, rebuilding the schedulers
556 * sched domains, we can ignore !is_sched_load_balance cpusets.
557 * csa - (for CpuSet Array) Array of pointers to all the cpusets
558 * that need to be load balanced, for convenient iterative
559 * access by the subsequent code that finds the best partition,
560 * i.e the set of domains (subsets) of CPUs such that the
561 * cpus_allowed of every cpuset marked is_sched_load_balance
562 * is a subset of one of these domains, while there are as
563 * many such domains as possible, each as small as possible.
564 * doms - Conversion of 'csa' to an array of cpumasks, for passing to
565 * the kernel/sched.c routine partition_sched_domains() in a
566 * convenient format, that can be easily compared to the prior
567 * value to determine what partition elements (sched domains)
568 * were changed (added or removed.)
570 * Finding the best partition (set of domains):
571 * The triple nested loops below over i, j, k scan over the
572 * load balanced cpusets (using the array of cpuset pointers in
573 * csa[]) looking for pairs of cpusets that have overlapping
574 * cpus_allowed, but which don't have the same 'pn' partition
575 * number and gives them in the same partition number. It keeps
576 * looping on the 'restart' label until it can no longer find
579 * The union of the cpus_allowed masks from the set of
580 * all cpusets having the same 'pn' value then form the one
581 * element of the partition (one sched domain) to be passed to
582 * partition_sched_domains().
584 static int generate_sched_domains(cpumask_t **domains,
585 struct sched_domain_attr **attributes)
587 LIST_HEAD(q); /* queue of cpusets to be scanned */
588 struct cpuset *cp; /* scans q */
589 struct cpuset **csa; /* array of all cpuset ptrs */
590 int csn; /* how many cpuset ptrs in csa so far */
591 int i, j, k; /* indices for partition finding loops */
592 cpumask_t *doms; /* resulting partition; i.e. sched domains */
593 struct sched_domain_attr *dattr; /* attributes for custom domains */
594 int ndoms = 0; /* number of sched domains in result */
595 int nslot; /* next empty doms[] cpumask_t slot */
601 /* Special case for the 99% of systems with one, full, sched domain */
602 if (is_sched_load_balance(&top_cpuset)) {
603 doms = kmalloc(sizeof(cpumask_t), GFP_KERNEL);
607 dattr = kmalloc(sizeof(struct sched_domain_attr), GFP_KERNEL);
609 *dattr = SD_ATTR_INIT;
610 update_domain_attr_tree(dattr, &top_cpuset);
612 *doms = top_cpuset.cpus_allowed;
618 csa = kmalloc(number_of_cpusets * sizeof(cp), GFP_KERNEL);
623 list_add(&top_cpuset.stack_list, &q);
624 while (!list_empty(&q)) {
626 struct cpuset *child; /* scans child cpusets of cp */
628 cp = list_first_entry(&q, struct cpuset, stack_list);
631 if (cpus_empty(cp->cpus_allowed))
635 * All child cpusets contain a subset of the parent's cpus, so
636 * just skip them, and then we call update_domain_attr_tree()
637 * to calc relax_domain_level of the corresponding sched
640 if (is_sched_load_balance(cp)) {
645 list_for_each_entry(cont, &cp->css.cgroup->children, sibling) {
646 child = cgroup_cs(cont);
647 list_add_tail(&child->stack_list, &q);
651 for (i = 0; i < csn; i++)
656 /* Find the best partition (set of sched domains) */
657 for (i = 0; i < csn; i++) {
658 struct cpuset *a = csa[i];
661 for (j = 0; j < csn; j++) {
662 struct cpuset *b = csa[j];
665 if (apn != bpn && cpusets_overlap(a, b)) {
666 for (k = 0; k < csn; k++) {
667 struct cpuset *c = csa[k];
672 ndoms--; /* one less element */
679 * Now we know how many domains to create.
680 * Convert <csn, csa> to <ndoms, doms> and populate cpu masks.
682 doms = kmalloc(ndoms * sizeof(cpumask_t), GFP_KERNEL);
687 * The rest of the code, including the scheduler, can deal with
688 * dattr==NULL case. No need to abort if alloc fails.
690 dattr = kmalloc(ndoms * sizeof(struct sched_domain_attr), GFP_KERNEL);
692 for (nslot = 0, i = 0; i < csn; i++) {
693 struct cpuset *a = csa[i];
698 /* Skip completed partitions */
704 if (nslot == ndoms) {
705 static int warnings = 10;
708 "rebuild_sched_domains confused:"
709 " nslot %d, ndoms %d, csn %d, i %d,"
711 nslot, ndoms, csn, i, apn);
719 *(dattr + nslot) = SD_ATTR_INIT;
720 for (j = i; j < csn; j++) {
721 struct cpuset *b = csa[j];
724 cpus_or(*dp, *dp, b->cpus_allowed);
726 update_domain_attr_tree(dattr + nslot, b);
728 /* Done with this partition */
734 BUG_ON(nslot != ndoms);
740 * Fallback to the default domain if kmalloc() failed.
741 * See comments in partition_sched_domains().
752 * Rebuild scheduler domains.
754 * Call with neither cgroup_mutex held nor within get_online_cpus().
755 * Takes both cgroup_mutex and get_online_cpus().
757 * Cannot be directly called from cpuset code handling changes
758 * to the cpuset pseudo-filesystem, because it cannot be called
759 * from code that already holds cgroup_mutex.
761 static void do_rebuild_sched_domains(struct work_struct *unused)
763 struct sched_domain_attr *attr;
769 /* Generate domain masks and attrs */
771 ndoms = generate_sched_domains(&doms, &attr);
774 /* Have scheduler rebuild the domains */
775 partition_sched_domains(ndoms, doms, attr);
780 static DECLARE_WORK(rebuild_sched_domains_work, do_rebuild_sched_domains);
783 * Rebuild scheduler domains, asynchronously via workqueue.
785 * If the flag 'sched_load_balance' of any cpuset with non-empty
786 * 'cpus' changes, or if the 'cpus' allowed changes in any cpuset
787 * which has that flag enabled, or if any cpuset with a non-empty
788 * 'cpus' is removed, then call this routine to rebuild the
789 * scheduler's dynamic sched domains.
791 * The rebuild_sched_domains() and partition_sched_domains()
792 * routines must nest cgroup_lock() inside get_online_cpus(),
793 * but such cpuset changes as these must nest that locking the
794 * other way, holding cgroup_lock() for much of the code.
796 * So in order to avoid an ABBA deadlock, the cpuset code handling
797 * these user changes delegates the actual sched domain rebuilding
798 * to a separate workqueue thread, which ends up processing the
799 * above do_rebuild_sched_domains() function.
801 static void async_rebuild_sched_domains(void)
803 schedule_work(&rebuild_sched_domains_work);
807 * Accomplishes the same scheduler domain rebuild as the above
808 * async_rebuild_sched_domains(), however it directly calls the
809 * rebuild routine synchronously rather than calling it via an
810 * asynchronous work thread.
812 * This can only be called from code that is not holding
813 * cgroup_mutex (not nested in a cgroup_lock() call.)
815 void rebuild_sched_domains(void)
817 do_rebuild_sched_domains(NULL);
821 * cpuset_test_cpumask - test a task's cpus_allowed versus its cpuset's
823 * @scan: struct cgroup_scanner contained in its struct cpuset_hotplug_scanner
825 * Call with cgroup_mutex held. May take callback_mutex during call.
826 * Called for each task in a cgroup by cgroup_scan_tasks().
827 * Return nonzero if this tasks's cpus_allowed mask should be changed (in other
828 * words, if its mask is not equal to its cpuset's mask).
830 static int cpuset_test_cpumask(struct task_struct *tsk,
831 struct cgroup_scanner *scan)
833 return !cpus_equal(tsk->cpus_allowed,
834 (cgroup_cs(scan->cg))->cpus_allowed);
838 * cpuset_change_cpumask - make a task's cpus_allowed the same as its cpuset's
840 * @scan: struct cgroup_scanner containing the cgroup of the task
842 * Called by cgroup_scan_tasks() for each task in a cgroup whose
843 * cpus_allowed mask needs to be changed.
845 * We don't need to re-check for the cgroup/cpuset membership, since we're
846 * holding cgroup_lock() at this point.
848 static void cpuset_change_cpumask(struct task_struct *tsk,
849 struct cgroup_scanner *scan)
851 set_cpus_allowed_ptr(tsk, &((cgroup_cs(scan->cg))->cpus_allowed));
855 * update_tasks_cpumask - Update the cpumasks of tasks in the cpuset.
856 * @cs: the cpuset in which each task's cpus_allowed mask needs to be changed
857 * @heap: if NULL, defer allocating heap memory to cgroup_scan_tasks()
859 * Called with cgroup_mutex held
861 * The cgroup_scan_tasks() function will scan all the tasks in a cgroup,
862 * calling callback functions for each.
864 * No return value. It's guaranteed that cgroup_scan_tasks() always returns 0
867 static void update_tasks_cpumask(struct cpuset *cs, struct ptr_heap *heap)
869 struct cgroup_scanner scan;
871 scan.cg = cs->css.cgroup;
872 scan.test_task = cpuset_test_cpumask;
873 scan.process_task = cpuset_change_cpumask;
875 cgroup_scan_tasks(&scan);
879 * update_cpumask - update the cpus_allowed mask of a cpuset and all tasks in it
880 * @cs: the cpuset to consider
881 * @buf: buffer of cpu numbers written to this cpuset
883 static int update_cpumask(struct cpuset *cs, const char *buf)
885 struct ptr_heap heap;
886 struct cpuset trialcs;
888 int is_load_balanced;
890 /* top_cpuset.cpus_allowed tracks cpu_online_map; it's read-only */
891 if (cs == &top_cpuset)
897 * An empty cpus_allowed is ok only if the cpuset has no tasks.
898 * Since cpulist_parse() fails on an empty mask, we special case
899 * that parsing. The validate_change() call ensures that cpusets
900 * with tasks have cpus.
903 cpus_clear(trialcs.cpus_allowed);
905 retval = cpulist_parse(buf, &trialcs.cpus_allowed);
909 if (!cpus_subset(trialcs.cpus_allowed, cpu_online_map))
912 retval = validate_change(cs, &trialcs);
916 /* Nothing to do if the cpus didn't change */
917 if (cpus_equal(cs->cpus_allowed, trialcs.cpus_allowed))
920 retval = heap_init(&heap, PAGE_SIZE, GFP_KERNEL, NULL);
924 is_load_balanced = is_sched_load_balance(&trialcs);
926 mutex_lock(&callback_mutex);
927 cs->cpus_allowed = trialcs.cpus_allowed;
928 mutex_unlock(&callback_mutex);
931 * Scan tasks in the cpuset, and update the cpumasks of any
932 * that need an update.
934 update_tasks_cpumask(cs, &heap);
938 if (is_load_balanced)
939 async_rebuild_sched_domains();
946 * Migrate memory region from one set of nodes to another.
948 * Temporarilly set tasks mems_allowed to target nodes of migration,
949 * so that the migration code can allocate pages on these nodes.
951 * Call holding cgroup_mutex, so current's cpuset won't change
952 * during this call, as manage_mutex holds off any cpuset_attach()
953 * calls. Therefore we don't need to take task_lock around the
954 * call to guarantee_online_mems(), as we know no one is changing
957 * Hold callback_mutex around the two modifications of our tasks
958 * mems_allowed to synchronize with cpuset_mems_allowed().
960 * While the mm_struct we are migrating is typically from some
961 * other task, the task_struct mems_allowed that we are hacking
962 * is for our current task, which must allocate new pages for that
963 * migrating memory region.
965 * We call cpuset_update_task_memory_state() before hacking
966 * our tasks mems_allowed, so that we are assured of being in
967 * sync with our tasks cpuset, and in particular, callbacks to
968 * cpuset_update_task_memory_state() from nested page allocations
969 * won't see any mismatch of our cpuset and task mems_generation
970 * values, so won't overwrite our hacked tasks mems_allowed
974 static void cpuset_migrate_mm(struct mm_struct *mm, const nodemask_t *from,
975 const nodemask_t *to)
977 struct task_struct *tsk = current;
979 cpuset_update_task_memory_state();
981 mutex_lock(&callback_mutex);
982 tsk->mems_allowed = *to;
983 mutex_unlock(&callback_mutex);
985 do_migrate_pages(mm, from, to, MPOL_MF_MOVE_ALL);
987 mutex_lock(&callback_mutex);
988 guarantee_online_mems(task_cs(tsk),&tsk->mems_allowed);
989 mutex_unlock(&callback_mutex);
992 static void *cpuset_being_rebound;
995 * update_tasks_nodemask - Update the nodemasks of tasks in the cpuset.
996 * @cs: the cpuset in which each task's mems_allowed mask needs to be changed
997 * @oldmem: old mems_allowed of cpuset cs
999 * Called with cgroup_mutex held
1000 * Return 0 if successful, -errno if not.
1002 static int update_tasks_nodemask(struct cpuset *cs, const nodemask_t *oldmem)
1004 struct task_struct *p;
1005 struct mm_struct **mmarray;
1009 struct cgroup_iter it;
1012 cpuset_being_rebound = cs; /* causes mpol_dup() rebind */
1014 fudge = 10; /* spare mmarray[] slots */
1015 fudge += cpus_weight(cs->cpus_allowed); /* imagine one fork-bomb/cpu */
1019 * Allocate mmarray[] to hold mm reference for each task
1020 * in cpuset cs. Can't kmalloc GFP_KERNEL while holding
1021 * tasklist_lock. We could use GFP_ATOMIC, but with a
1022 * few more lines of code, we can retry until we get a big
1023 * enough mmarray[] w/o using GFP_ATOMIC.
1026 ntasks = cgroup_task_count(cs->css.cgroup); /* guess */
1028 mmarray = kmalloc(ntasks * sizeof(*mmarray), GFP_KERNEL);
1031 read_lock(&tasklist_lock); /* block fork */
1032 if (cgroup_task_count(cs->css.cgroup) <= ntasks)
1033 break; /* got enough */
1034 read_unlock(&tasklist_lock); /* try again */
1040 /* Load up mmarray[] with mm reference for each task in cpuset. */
1041 cgroup_iter_start(cs->css.cgroup, &it);
1042 while ((p = cgroup_iter_next(cs->css.cgroup, &it))) {
1043 struct mm_struct *mm;
1047 "Cpuset mempolicy rebind incomplete.\n");
1050 mm = get_task_mm(p);
1055 cgroup_iter_end(cs->css.cgroup, &it);
1056 read_unlock(&tasklist_lock);
1059 * Now that we've dropped the tasklist spinlock, we can
1060 * rebind the vma mempolicies of each mm in mmarray[] to their
1061 * new cpuset, and release that mm. The mpol_rebind_mm()
1062 * call takes mmap_sem, which we couldn't take while holding
1063 * tasklist_lock. Forks can happen again now - the mpol_dup()
1064 * cpuset_being_rebound check will catch such forks, and rebind
1065 * their vma mempolicies too. Because we still hold the global
1066 * cgroup_mutex, we know that no other rebind effort will
1067 * be contending for the global variable cpuset_being_rebound.
1068 * It's ok if we rebind the same mm twice; mpol_rebind_mm()
1069 * is idempotent. Also migrate pages in each mm to new nodes.
1071 migrate = is_memory_migrate(cs);
1072 for (i = 0; i < n; i++) {
1073 struct mm_struct *mm = mmarray[i];
1075 mpol_rebind_mm(mm, &cs->mems_allowed);
1077 cpuset_migrate_mm(mm, oldmem, &cs->mems_allowed);
1081 /* We're done rebinding vmas to this cpuset's new mems_allowed. */
1083 cpuset_being_rebound = NULL;
1090 * Handle user request to change the 'mems' memory placement
1091 * of a cpuset. Needs to validate the request, update the
1092 * cpusets mems_allowed and mems_generation, and for each
1093 * task in the cpuset, rebind any vma mempolicies and if
1094 * the cpuset is marked 'memory_migrate', migrate the tasks
1095 * pages to the new memory.
1097 * Call with cgroup_mutex held. May take callback_mutex during call.
1098 * Will take tasklist_lock, scan tasklist for tasks in cpuset cs,
1099 * lock each such tasks mm->mmap_sem, scan its vma's and rebind
1100 * their mempolicies to the cpusets new mems_allowed.
1102 static int update_nodemask(struct cpuset *cs, const char *buf)
1104 struct cpuset trialcs;
1109 * top_cpuset.mems_allowed tracks node_stats[N_HIGH_MEMORY];
1112 if (cs == &top_cpuset)
1118 * An empty mems_allowed is ok iff there are no tasks in the cpuset.
1119 * Since nodelist_parse() fails on an empty mask, we special case
1120 * that parsing. The validate_change() call ensures that cpusets
1121 * with tasks have memory.
1124 nodes_clear(trialcs.mems_allowed);
1126 retval = nodelist_parse(buf, trialcs.mems_allowed);
1130 if (!nodes_subset(trialcs.mems_allowed,
1131 node_states[N_HIGH_MEMORY]))
1134 oldmem = cs->mems_allowed;
1135 if (nodes_equal(oldmem, trialcs.mems_allowed)) {
1136 retval = 0; /* Too easy - nothing to do */
1139 retval = validate_change(cs, &trialcs);
1143 mutex_lock(&callback_mutex);
1144 cs->mems_allowed = trialcs.mems_allowed;
1145 cs->mems_generation = cpuset_mems_generation++;
1146 mutex_unlock(&callback_mutex);
1148 retval = update_tasks_nodemask(cs, &oldmem);
1153 int current_cpuset_is_being_rebound(void)
1155 return task_cs(current) == cpuset_being_rebound;
1158 static int update_relax_domain_level(struct cpuset *cs, s64 val)
1160 if (val < -1 || val >= SD_LV_MAX)
1163 if (val != cs->relax_domain_level) {
1164 cs->relax_domain_level = val;
1165 if (!cpus_empty(cs->cpus_allowed) && is_sched_load_balance(cs))
1166 async_rebuild_sched_domains();
1173 * update_flag - read a 0 or a 1 in a file and update associated flag
1174 * bit: the bit to update (see cpuset_flagbits_t)
1175 * cs: the cpuset to update
1176 * turning_on: whether the flag is being set or cleared
1178 * Call with cgroup_mutex held.
1181 static int update_flag(cpuset_flagbits_t bit, struct cpuset *cs,
1184 struct cpuset trialcs;
1186 int balance_flag_changed;
1190 set_bit(bit, &trialcs.flags);
1192 clear_bit(bit, &trialcs.flags);
1194 err = validate_change(cs, &trialcs);
1198 balance_flag_changed = (is_sched_load_balance(cs) !=
1199 is_sched_load_balance(&trialcs));
1201 mutex_lock(&callback_mutex);
1202 cs->flags = trialcs.flags;
1203 mutex_unlock(&callback_mutex);
1205 if (!cpus_empty(trialcs.cpus_allowed) && balance_flag_changed)
1206 async_rebuild_sched_domains();
1212 * Frequency meter - How fast is some event occurring?
1214 * These routines manage a digitally filtered, constant time based,
1215 * event frequency meter. There are four routines:
1216 * fmeter_init() - initialize a frequency meter.
1217 * fmeter_markevent() - called each time the event happens.
1218 * fmeter_getrate() - returns the recent rate of such events.
1219 * fmeter_update() - internal routine used to update fmeter.
1221 * A common data structure is passed to each of these routines,
1222 * which is used to keep track of the state required to manage the
1223 * frequency meter and its digital filter.
1225 * The filter works on the number of events marked per unit time.
1226 * The filter is single-pole low-pass recursive (IIR). The time unit
1227 * is 1 second. Arithmetic is done using 32-bit integers scaled to
1228 * simulate 3 decimal digits of precision (multiplied by 1000).
1230 * With an FM_COEF of 933, and a time base of 1 second, the filter
1231 * has a half-life of 10 seconds, meaning that if the events quit
1232 * happening, then the rate returned from the fmeter_getrate()
1233 * will be cut in half each 10 seconds, until it converges to zero.
1235 * It is not worth doing a real infinitely recursive filter. If more
1236 * than FM_MAXTICKS ticks have elapsed since the last filter event,
1237 * just compute FM_MAXTICKS ticks worth, by which point the level
1240 * Limit the count of unprocessed events to FM_MAXCNT, so as to avoid
1241 * arithmetic overflow in the fmeter_update() routine.
1243 * Given the simple 32 bit integer arithmetic used, this meter works
1244 * best for reporting rates between one per millisecond (msec) and
1245 * one per 32 (approx) seconds. At constant rates faster than one
1246 * per msec it maxes out at values just under 1,000,000. At constant
1247 * rates between one per msec, and one per second it will stabilize
1248 * to a value N*1000, where N is the rate of events per second.
1249 * At constant rates between one per second and one per 32 seconds,
1250 * it will be choppy, moving up on the seconds that have an event,
1251 * and then decaying until the next event. At rates slower than
1252 * about one in 32 seconds, it decays all the way back to zero between
1256 #define FM_COEF 933 /* coefficient for half-life of 10 secs */
1257 #define FM_MAXTICKS ((time_t)99) /* useless computing more ticks than this */
1258 #define FM_MAXCNT 1000000 /* limit cnt to avoid overflow */
1259 #define FM_SCALE 1000 /* faux fixed point scale */
1261 /* Initialize a frequency meter */
1262 static void fmeter_init(struct fmeter *fmp)
1267 spin_lock_init(&fmp->lock);
1270 /* Internal meter update - process cnt events and update value */
1271 static void fmeter_update(struct fmeter *fmp)
1273 time_t now = get_seconds();
1274 time_t ticks = now - fmp->time;
1279 ticks = min(FM_MAXTICKS, ticks);
1281 fmp->val = (FM_COEF * fmp->val) / FM_SCALE;
1284 fmp->val += ((FM_SCALE - FM_COEF) * fmp->cnt) / FM_SCALE;
1288 /* Process any previous ticks, then bump cnt by one (times scale). */
1289 static void fmeter_markevent(struct fmeter *fmp)
1291 spin_lock(&fmp->lock);
1293 fmp->cnt = min(FM_MAXCNT, fmp->cnt + FM_SCALE);
1294 spin_unlock(&fmp->lock);
1297 /* Process any previous ticks, then return current value. */
1298 static int fmeter_getrate(struct fmeter *fmp)
1302 spin_lock(&fmp->lock);
1305 spin_unlock(&fmp->lock);
1309 /* Called by cgroups to determine if a cpuset is usable; cgroup_mutex held */
1310 static int cpuset_can_attach(struct cgroup_subsys *ss,
1311 struct cgroup *cont, struct task_struct *tsk)
1313 struct cpuset *cs = cgroup_cs(cont);
1315 if (cpus_empty(cs->cpus_allowed) || nodes_empty(cs->mems_allowed))
1317 if (tsk->flags & PF_THREAD_BOUND) {
1320 mutex_lock(&callback_mutex);
1321 mask = cs->cpus_allowed;
1322 mutex_unlock(&callback_mutex);
1323 if (!cpus_equal(tsk->cpus_allowed, mask))
1327 return security_task_setscheduler(tsk, 0, NULL);
1330 static void cpuset_attach(struct cgroup_subsys *ss,
1331 struct cgroup *cont, struct cgroup *oldcont,
1332 struct task_struct *tsk)
1335 nodemask_t from, to;
1336 struct mm_struct *mm;
1337 struct cpuset *cs = cgroup_cs(cont);
1338 struct cpuset *oldcs = cgroup_cs(oldcont);
1341 mutex_lock(&callback_mutex);
1342 guarantee_online_cpus(cs, &cpus);
1343 err = set_cpus_allowed_ptr(tsk, &cpus);
1344 mutex_unlock(&callback_mutex);
1348 from = oldcs->mems_allowed;
1349 to = cs->mems_allowed;
1350 mm = get_task_mm(tsk);
1352 mpol_rebind_mm(mm, &to);
1353 if (is_memory_migrate(cs))
1354 cpuset_migrate_mm(mm, &from, &to);
1360 /* The various types of files and directories in a cpuset file system */
1363 FILE_MEMORY_MIGRATE,
1369 FILE_SCHED_LOAD_BALANCE,
1370 FILE_SCHED_RELAX_DOMAIN_LEVEL,
1371 FILE_MEMORY_PRESSURE_ENABLED,
1372 FILE_MEMORY_PRESSURE,
1375 } cpuset_filetype_t;
1377 static int cpuset_write_u64(struct cgroup *cgrp, struct cftype *cft, u64 val)
1380 struct cpuset *cs = cgroup_cs(cgrp);
1381 cpuset_filetype_t type = cft->private;
1383 if (!cgroup_lock_live_group(cgrp))
1387 case FILE_CPU_EXCLUSIVE:
1388 retval = update_flag(CS_CPU_EXCLUSIVE, cs, val);
1390 case FILE_MEM_EXCLUSIVE:
1391 retval = update_flag(CS_MEM_EXCLUSIVE, cs, val);
1393 case FILE_MEM_HARDWALL:
1394 retval = update_flag(CS_MEM_HARDWALL, cs, val);
1396 case FILE_SCHED_LOAD_BALANCE:
1397 retval = update_flag(CS_SCHED_LOAD_BALANCE, cs, val);
1399 case FILE_MEMORY_MIGRATE:
1400 retval = update_flag(CS_MEMORY_MIGRATE, cs, val);
1402 case FILE_MEMORY_PRESSURE_ENABLED:
1403 cpuset_memory_pressure_enabled = !!val;
1405 case FILE_MEMORY_PRESSURE:
1408 case FILE_SPREAD_PAGE:
1409 retval = update_flag(CS_SPREAD_PAGE, cs, val);
1410 cs->mems_generation = cpuset_mems_generation++;
1412 case FILE_SPREAD_SLAB:
1413 retval = update_flag(CS_SPREAD_SLAB, cs, val);
1414 cs->mems_generation = cpuset_mems_generation++;
1424 static int cpuset_write_s64(struct cgroup *cgrp, struct cftype *cft, s64 val)
1427 struct cpuset *cs = cgroup_cs(cgrp);
1428 cpuset_filetype_t type = cft->private;
1430 if (!cgroup_lock_live_group(cgrp))
1434 case FILE_SCHED_RELAX_DOMAIN_LEVEL:
1435 retval = update_relax_domain_level(cs, val);
1446 * Common handling for a write to a "cpus" or "mems" file.
1448 static int cpuset_write_resmask(struct cgroup *cgrp, struct cftype *cft,
1453 if (!cgroup_lock_live_group(cgrp))
1456 switch (cft->private) {
1458 retval = update_cpumask(cgroup_cs(cgrp), buf);
1461 retval = update_nodemask(cgroup_cs(cgrp), buf);
1472 * These ascii lists should be read in a single call, by using a user
1473 * buffer large enough to hold the entire map. If read in smaller
1474 * chunks, there is no guarantee of atomicity. Since the display format
1475 * used, list of ranges of sequential numbers, is variable length,
1476 * and since these maps can change value dynamically, one could read
1477 * gibberish by doing partial reads while a list was changing.
1478 * A single large read to a buffer that crosses a page boundary is
1479 * ok, because the result being copied to user land is not recomputed
1480 * across a page fault.
1483 static int cpuset_sprintf_cpulist(char *page, struct cpuset *cs)
1487 mutex_lock(&callback_mutex);
1488 mask = cs->cpus_allowed;
1489 mutex_unlock(&callback_mutex);
1491 return cpulist_scnprintf(page, PAGE_SIZE, &mask);
1494 static int cpuset_sprintf_memlist(char *page, struct cpuset *cs)
1498 mutex_lock(&callback_mutex);
1499 mask = cs->mems_allowed;
1500 mutex_unlock(&callback_mutex);
1502 return nodelist_scnprintf(page, PAGE_SIZE, mask);
1505 static ssize_t cpuset_common_file_read(struct cgroup *cont,
1509 size_t nbytes, loff_t *ppos)
1511 struct cpuset *cs = cgroup_cs(cont);
1512 cpuset_filetype_t type = cft->private;
1517 if (!(page = (char *)__get_free_page(GFP_TEMPORARY)))
1524 s += cpuset_sprintf_cpulist(s, cs);
1527 s += cpuset_sprintf_memlist(s, cs);
1535 retval = simple_read_from_buffer(buf, nbytes, ppos, page, s - page);
1537 free_page((unsigned long)page);
1541 static u64 cpuset_read_u64(struct cgroup *cont, struct cftype *cft)
1543 struct cpuset *cs = cgroup_cs(cont);
1544 cpuset_filetype_t type = cft->private;
1546 case FILE_CPU_EXCLUSIVE:
1547 return is_cpu_exclusive(cs);
1548 case FILE_MEM_EXCLUSIVE:
1549 return is_mem_exclusive(cs);
1550 case FILE_MEM_HARDWALL:
1551 return is_mem_hardwall(cs);
1552 case FILE_SCHED_LOAD_BALANCE:
1553 return is_sched_load_balance(cs);
1554 case FILE_MEMORY_MIGRATE:
1555 return is_memory_migrate(cs);
1556 case FILE_MEMORY_PRESSURE_ENABLED:
1557 return cpuset_memory_pressure_enabled;
1558 case FILE_MEMORY_PRESSURE:
1559 return fmeter_getrate(&cs->fmeter);
1560 case FILE_SPREAD_PAGE:
1561 return is_spread_page(cs);
1562 case FILE_SPREAD_SLAB:
1563 return is_spread_slab(cs);
1568 /* Unreachable but makes gcc happy */
1572 static s64 cpuset_read_s64(struct cgroup *cont, struct cftype *cft)
1574 struct cpuset *cs = cgroup_cs(cont);
1575 cpuset_filetype_t type = cft->private;
1577 case FILE_SCHED_RELAX_DOMAIN_LEVEL:
1578 return cs->relax_domain_level;
1583 /* Unrechable but makes gcc happy */
1589 * for the common functions, 'private' gives the type of file
1592 static struct cftype files[] = {
1595 .read = cpuset_common_file_read,
1596 .write_string = cpuset_write_resmask,
1597 .max_write_len = (100U + 6 * NR_CPUS),
1598 .private = FILE_CPULIST,
1603 .read = cpuset_common_file_read,
1604 .write_string = cpuset_write_resmask,
1605 .max_write_len = (100U + 6 * MAX_NUMNODES),
1606 .private = FILE_MEMLIST,
1610 .name = "cpu_exclusive",
1611 .read_u64 = cpuset_read_u64,
1612 .write_u64 = cpuset_write_u64,
1613 .private = FILE_CPU_EXCLUSIVE,
1617 .name = "mem_exclusive",
1618 .read_u64 = cpuset_read_u64,
1619 .write_u64 = cpuset_write_u64,
1620 .private = FILE_MEM_EXCLUSIVE,
1624 .name = "mem_hardwall",
1625 .read_u64 = cpuset_read_u64,
1626 .write_u64 = cpuset_write_u64,
1627 .private = FILE_MEM_HARDWALL,
1631 .name = "sched_load_balance",
1632 .read_u64 = cpuset_read_u64,
1633 .write_u64 = cpuset_write_u64,
1634 .private = FILE_SCHED_LOAD_BALANCE,
1638 .name = "sched_relax_domain_level",
1639 .read_s64 = cpuset_read_s64,
1640 .write_s64 = cpuset_write_s64,
1641 .private = FILE_SCHED_RELAX_DOMAIN_LEVEL,
1645 .name = "memory_migrate",
1646 .read_u64 = cpuset_read_u64,
1647 .write_u64 = cpuset_write_u64,
1648 .private = FILE_MEMORY_MIGRATE,
1652 .name = "memory_pressure",
1653 .read_u64 = cpuset_read_u64,
1654 .write_u64 = cpuset_write_u64,
1655 .private = FILE_MEMORY_PRESSURE,
1659 .name = "memory_spread_page",
1660 .read_u64 = cpuset_read_u64,
1661 .write_u64 = cpuset_write_u64,
1662 .private = FILE_SPREAD_PAGE,
1666 .name = "memory_spread_slab",
1667 .read_u64 = cpuset_read_u64,
1668 .write_u64 = cpuset_write_u64,
1669 .private = FILE_SPREAD_SLAB,
1673 static struct cftype cft_memory_pressure_enabled = {
1674 .name = "memory_pressure_enabled",
1675 .read_u64 = cpuset_read_u64,
1676 .write_u64 = cpuset_write_u64,
1677 .private = FILE_MEMORY_PRESSURE_ENABLED,
1680 static int cpuset_populate(struct cgroup_subsys *ss, struct cgroup *cont)
1684 err = cgroup_add_files(cont, ss, files, ARRAY_SIZE(files));
1687 /* memory_pressure_enabled is in root cpuset only */
1689 err = cgroup_add_file(cont, ss,
1690 &cft_memory_pressure_enabled);
1695 * post_clone() is called at the end of cgroup_clone().
1696 * 'cgroup' was just created automatically as a result of
1697 * a cgroup_clone(), and the current task is about to
1698 * be moved into 'cgroup'.
1700 * Currently we refuse to set up the cgroup - thereby
1701 * refusing the task to be entered, and as a result refusing
1702 * the sys_unshare() or clone() which initiated it - if any
1703 * sibling cpusets have exclusive cpus or mem.
1705 * If this becomes a problem for some users who wish to
1706 * allow that scenario, then cpuset_post_clone() could be
1707 * changed to grant parent->cpus_allowed-sibling_cpus_exclusive
1708 * (and likewise for mems) to the new cgroup. Called with cgroup_mutex
1711 static void cpuset_post_clone(struct cgroup_subsys *ss,
1712 struct cgroup *cgroup)
1714 struct cgroup *parent, *child;
1715 struct cpuset *cs, *parent_cs;
1717 parent = cgroup->parent;
1718 list_for_each_entry(child, &parent->children, sibling) {
1719 cs = cgroup_cs(child);
1720 if (is_mem_exclusive(cs) || is_cpu_exclusive(cs))
1723 cs = cgroup_cs(cgroup);
1724 parent_cs = cgroup_cs(parent);
1726 cs->mems_allowed = parent_cs->mems_allowed;
1727 cs->cpus_allowed = parent_cs->cpus_allowed;
1732 * cpuset_create - create a cpuset
1733 * ss: cpuset cgroup subsystem
1734 * cont: control group that the new cpuset will be part of
1737 static struct cgroup_subsys_state *cpuset_create(
1738 struct cgroup_subsys *ss,
1739 struct cgroup *cont)
1742 struct cpuset *parent;
1744 if (!cont->parent) {
1745 /* This is early initialization for the top cgroup */
1746 top_cpuset.mems_generation = cpuset_mems_generation++;
1747 return &top_cpuset.css;
1749 parent = cgroup_cs(cont->parent);
1750 cs = kmalloc(sizeof(*cs), GFP_KERNEL);
1752 return ERR_PTR(-ENOMEM);
1754 cpuset_update_task_memory_state();
1756 if (is_spread_page(parent))
1757 set_bit(CS_SPREAD_PAGE, &cs->flags);
1758 if (is_spread_slab(parent))
1759 set_bit(CS_SPREAD_SLAB, &cs->flags);
1760 set_bit(CS_SCHED_LOAD_BALANCE, &cs->flags);
1761 cpus_clear(cs->cpus_allowed);
1762 nodes_clear(cs->mems_allowed);
1763 cs->mems_generation = cpuset_mems_generation++;
1764 fmeter_init(&cs->fmeter);
1765 cs->relax_domain_level = -1;
1767 cs->parent = parent;
1768 number_of_cpusets++;
1773 * If the cpuset being removed has its flag 'sched_load_balance'
1774 * enabled, then simulate turning sched_load_balance off, which
1775 * will call async_rebuild_sched_domains().
1778 static void cpuset_destroy(struct cgroup_subsys *ss, struct cgroup *cont)
1780 struct cpuset *cs = cgroup_cs(cont);
1782 cpuset_update_task_memory_state();
1784 if (is_sched_load_balance(cs))
1785 update_flag(CS_SCHED_LOAD_BALANCE, cs, 0);
1787 number_of_cpusets--;
1791 struct cgroup_subsys cpuset_subsys = {
1793 .create = cpuset_create,
1794 .destroy = cpuset_destroy,
1795 .can_attach = cpuset_can_attach,
1796 .attach = cpuset_attach,
1797 .populate = cpuset_populate,
1798 .post_clone = cpuset_post_clone,
1799 .subsys_id = cpuset_subsys_id,
1804 * cpuset_init_early - just enough so that the calls to
1805 * cpuset_update_task_memory_state() in early init code
1809 int __init cpuset_init_early(void)
1811 top_cpuset.mems_generation = cpuset_mems_generation++;
1817 * cpuset_init - initialize cpusets at system boot
1819 * Description: Initialize top_cpuset and the cpuset internal file system,
1822 int __init cpuset_init(void)
1826 cpus_setall(top_cpuset.cpus_allowed);
1827 nodes_setall(top_cpuset.mems_allowed);
1829 fmeter_init(&top_cpuset.fmeter);
1830 top_cpuset.mems_generation = cpuset_mems_generation++;
1831 set_bit(CS_SCHED_LOAD_BALANCE, &top_cpuset.flags);
1832 top_cpuset.relax_domain_level = -1;
1834 err = register_filesystem(&cpuset_fs_type);
1838 number_of_cpusets = 1;
1843 * cpuset_do_move_task - move a given task to another cpuset
1844 * @tsk: pointer to task_struct the task to move
1845 * @scan: struct cgroup_scanner contained in its struct cpuset_hotplug_scanner
1847 * Called by cgroup_scan_tasks() for each task in a cgroup.
1848 * Return nonzero to stop the walk through the tasks.
1850 static void cpuset_do_move_task(struct task_struct *tsk,
1851 struct cgroup_scanner *scan)
1853 struct cpuset_hotplug_scanner *chsp;
1855 chsp = container_of(scan, struct cpuset_hotplug_scanner, scan);
1856 cgroup_attach_task(chsp->to, tsk);
1860 * move_member_tasks_to_cpuset - move tasks from one cpuset to another
1861 * @from: cpuset in which the tasks currently reside
1862 * @to: cpuset to which the tasks will be moved
1864 * Called with cgroup_mutex held
1865 * callback_mutex must not be held, as cpuset_attach() will take it.
1867 * The cgroup_scan_tasks() function will scan all the tasks in a cgroup,
1868 * calling callback functions for each.
1870 static void move_member_tasks_to_cpuset(struct cpuset *from, struct cpuset *to)
1872 struct cpuset_hotplug_scanner scan;
1874 scan.scan.cg = from->css.cgroup;
1875 scan.scan.test_task = NULL; /* select all tasks in cgroup */
1876 scan.scan.process_task = cpuset_do_move_task;
1877 scan.scan.heap = NULL;
1878 scan.to = to->css.cgroup;
1880 if (cgroup_scan_tasks(&scan.scan))
1881 printk(KERN_ERR "move_member_tasks_to_cpuset: "
1882 "cgroup_scan_tasks failed\n");
1886 * If CPU and/or memory hotplug handlers, below, unplug any CPUs
1887 * or memory nodes, we need to walk over the cpuset hierarchy,
1888 * removing that CPU or node from all cpusets. If this removes the
1889 * last CPU or node from a cpuset, then move the tasks in the empty
1890 * cpuset to its next-highest non-empty parent.
1892 * Called with cgroup_mutex held
1893 * callback_mutex must not be held, as cpuset_attach() will take it.
1895 static void remove_tasks_in_empty_cpuset(struct cpuset *cs)
1897 struct cpuset *parent;
1900 * The cgroup's css_sets list is in use if there are tasks
1901 * in the cpuset; the list is empty if there are none;
1902 * the cs->css.refcnt seems always 0.
1904 if (list_empty(&cs->css.cgroup->css_sets))
1908 * Find its next-highest non-empty parent, (top cpuset
1909 * has online cpus, so can't be empty).
1911 parent = cs->parent;
1912 while (cpus_empty(parent->cpus_allowed) ||
1913 nodes_empty(parent->mems_allowed))
1914 parent = parent->parent;
1916 move_member_tasks_to_cpuset(cs, parent);
1920 * Walk the specified cpuset subtree and look for empty cpusets.
1921 * The tasks of such cpuset must be moved to a parent cpuset.
1923 * Called with cgroup_mutex held. We take callback_mutex to modify
1924 * cpus_allowed and mems_allowed.
1926 * This walk processes the tree from top to bottom, completing one layer
1927 * before dropping down to the next. It always processes a node before
1928 * any of its children.
1930 * For now, since we lack memory hot unplug, we'll never see a cpuset
1931 * that has tasks along with an empty 'mems'. But if we did see such
1932 * a cpuset, we'd handle it just like we do if its 'cpus' was empty.
1934 static void scan_for_empty_cpusets(struct cpuset *root)
1937 struct cpuset *cp; /* scans cpusets being updated */
1938 struct cpuset *child; /* scans child cpusets of cp */
1939 struct cgroup *cont;
1942 list_add_tail((struct list_head *)&root->stack_list, &queue);
1944 while (!list_empty(&queue)) {
1945 cp = list_first_entry(&queue, struct cpuset, stack_list);
1946 list_del(queue.next);
1947 list_for_each_entry(cont, &cp->css.cgroup->children, sibling) {
1948 child = cgroup_cs(cont);
1949 list_add_tail(&child->stack_list, &queue);
1952 /* Continue past cpusets with all cpus, mems online */
1953 if (cpus_subset(cp->cpus_allowed, cpu_online_map) &&
1954 nodes_subset(cp->mems_allowed, node_states[N_HIGH_MEMORY]))
1957 oldmems = cp->mems_allowed;
1959 /* Remove offline cpus and mems from this cpuset. */
1960 mutex_lock(&callback_mutex);
1961 cpus_and(cp->cpus_allowed, cp->cpus_allowed, cpu_online_map);
1962 nodes_and(cp->mems_allowed, cp->mems_allowed,
1963 node_states[N_HIGH_MEMORY]);
1964 mutex_unlock(&callback_mutex);
1966 /* Move tasks from the empty cpuset to a parent */
1967 if (cpus_empty(cp->cpus_allowed) ||
1968 nodes_empty(cp->mems_allowed))
1969 remove_tasks_in_empty_cpuset(cp);
1971 update_tasks_cpumask(cp, NULL);
1972 update_tasks_nodemask(cp, &oldmems);
1978 * The top_cpuset tracks what CPUs and Memory Nodes are online,
1979 * period. This is necessary in order to make cpusets transparent
1980 * (of no affect) on systems that are actively using CPU hotplug
1981 * but making no active use of cpusets.
1983 * This routine ensures that top_cpuset.cpus_allowed tracks
1984 * cpu_online_map on each CPU hotplug (cpuhp) event.
1986 * Called within get_online_cpus(). Needs to call cgroup_lock()
1987 * before calling generate_sched_domains().
1989 static int cpuset_track_online_cpus(struct notifier_block *unused_nb,
1990 unsigned long phase, void *unused_cpu)
1992 struct sched_domain_attr *attr;
1998 case CPU_ONLINE_FROZEN:
2000 case CPU_DEAD_FROZEN:
2008 top_cpuset.cpus_allowed = cpu_online_map;
2009 scan_for_empty_cpusets(&top_cpuset);
2010 ndoms = generate_sched_domains(&doms, &attr);
2013 /* Have scheduler rebuild the domains */
2014 partition_sched_domains(ndoms, doms, attr);
2019 #ifdef CONFIG_MEMORY_HOTPLUG
2021 * Keep top_cpuset.mems_allowed tracking node_states[N_HIGH_MEMORY].
2022 * Call this routine anytime after node_states[N_HIGH_MEMORY] changes.
2023 * See also the previous routine cpuset_track_online_cpus().
2025 static int cpuset_track_online_nodes(struct notifier_block *self,
2026 unsigned long action, void *arg)
2031 top_cpuset.mems_allowed = node_states[N_HIGH_MEMORY];
2034 top_cpuset.mems_allowed = node_states[N_HIGH_MEMORY];
2035 scan_for_empty_cpusets(&top_cpuset);
2046 * cpuset_init_smp - initialize cpus_allowed
2048 * Description: Finish top cpuset after cpu, node maps are initialized
2051 void __init cpuset_init_smp(void)
2053 top_cpuset.cpus_allowed = cpu_online_map;
2054 top_cpuset.mems_allowed = node_states[N_HIGH_MEMORY];
2056 hotcpu_notifier(cpuset_track_online_cpus, 0);
2057 hotplug_memory_notifier(cpuset_track_online_nodes, 10);
2061 * cpuset_cpus_allowed - return cpus_allowed mask from a tasks cpuset.
2062 * @tsk: pointer to task_struct from which to obtain cpuset->cpus_allowed.
2063 * @pmask: pointer to cpumask_t variable to receive cpus_allowed set.
2065 * Description: Returns the cpumask_t cpus_allowed of the cpuset
2066 * attached to the specified @tsk. Guaranteed to return some non-empty
2067 * subset of cpu_online_map, even if this means going outside the
2071 void cpuset_cpus_allowed(struct task_struct *tsk, cpumask_t *pmask)
2073 mutex_lock(&callback_mutex);
2074 cpuset_cpus_allowed_locked(tsk, pmask);
2075 mutex_unlock(&callback_mutex);
2079 * cpuset_cpus_allowed_locked - return cpus_allowed mask from a tasks cpuset.
2080 * Must be called with callback_mutex held.
2082 void cpuset_cpus_allowed_locked(struct task_struct *tsk, cpumask_t *pmask)
2085 guarantee_online_cpus(task_cs(tsk), pmask);
2089 void cpuset_init_current_mems_allowed(void)
2091 nodes_setall(current->mems_allowed);
2095 * cpuset_mems_allowed - return mems_allowed mask from a tasks cpuset.
2096 * @tsk: pointer to task_struct from which to obtain cpuset->mems_allowed.
2098 * Description: Returns the nodemask_t mems_allowed of the cpuset
2099 * attached to the specified @tsk. Guaranteed to return some non-empty
2100 * subset of node_states[N_HIGH_MEMORY], even if this means going outside the
2104 nodemask_t cpuset_mems_allowed(struct task_struct *tsk)
2108 mutex_lock(&callback_mutex);
2110 guarantee_online_mems(task_cs(tsk), &mask);
2112 mutex_unlock(&callback_mutex);
2118 * cpuset_nodemask_valid_mems_allowed - check nodemask vs. curremt mems_allowed
2119 * @nodemask: the nodemask to be checked
2121 * Are any of the nodes in the nodemask allowed in current->mems_allowed?
2123 int cpuset_nodemask_valid_mems_allowed(nodemask_t *nodemask)
2125 return nodes_intersects(*nodemask, current->mems_allowed);
2129 * nearest_hardwall_ancestor() - Returns the nearest mem_exclusive or
2130 * mem_hardwall ancestor to the specified cpuset. Call holding
2131 * callback_mutex. If no ancestor is mem_exclusive or mem_hardwall
2132 * (an unusual configuration), then returns the root cpuset.
2134 static const struct cpuset *nearest_hardwall_ancestor(const struct cpuset *cs)
2136 while (!(is_mem_exclusive(cs) || is_mem_hardwall(cs)) && cs->parent)
2142 * cpuset_zone_allowed_softwall - Can we allocate on zone z's memory node?
2143 * @z: is this zone on an allowed node?
2144 * @gfp_mask: memory allocation flags
2146 * If we're in interrupt, yes, we can always allocate. If
2147 * __GFP_THISNODE is set, yes, we can always allocate. If zone
2148 * z's node is in our tasks mems_allowed, yes. If it's not a
2149 * __GFP_HARDWALL request and this zone's nodes is in the nearest
2150 * hardwalled cpuset ancestor to this tasks cpuset, yes.
2151 * If the task has been OOM killed and has access to memory reserves
2152 * as specified by the TIF_MEMDIE flag, yes.
2155 * If __GFP_HARDWALL is set, cpuset_zone_allowed_softwall()
2156 * reduces to cpuset_zone_allowed_hardwall(). Otherwise,
2157 * cpuset_zone_allowed_softwall() might sleep, and might allow a zone
2158 * from an enclosing cpuset.
2160 * cpuset_zone_allowed_hardwall() only handles the simpler case of
2161 * hardwall cpusets, and never sleeps.
2163 * The __GFP_THISNODE placement logic is really handled elsewhere,
2164 * by forcibly using a zonelist starting at a specified node, and by
2165 * (in get_page_from_freelist()) refusing to consider the zones for
2166 * any node on the zonelist except the first. By the time any such
2167 * calls get to this routine, we should just shut up and say 'yes'.
2169 * GFP_USER allocations are marked with the __GFP_HARDWALL bit,
2170 * and do not allow allocations outside the current tasks cpuset
2171 * unless the task has been OOM killed as is marked TIF_MEMDIE.
2172 * GFP_KERNEL allocations are not so marked, so can escape to the
2173 * nearest enclosing hardwalled ancestor cpuset.
2175 * Scanning up parent cpusets requires callback_mutex. The
2176 * __alloc_pages() routine only calls here with __GFP_HARDWALL bit
2177 * _not_ set if it's a GFP_KERNEL allocation, and all nodes in the
2178 * current tasks mems_allowed came up empty on the first pass over
2179 * the zonelist. So only GFP_KERNEL allocations, if all nodes in the
2180 * cpuset are short of memory, might require taking the callback_mutex
2183 * The first call here from mm/page_alloc:get_page_from_freelist()
2184 * has __GFP_HARDWALL set in gfp_mask, enforcing hardwall cpusets,
2185 * so no allocation on a node outside the cpuset is allowed (unless
2186 * in interrupt, of course).
2188 * The second pass through get_page_from_freelist() doesn't even call
2189 * here for GFP_ATOMIC calls. For those calls, the __alloc_pages()
2190 * variable 'wait' is not set, and the bit ALLOC_CPUSET is not set
2191 * in alloc_flags. That logic and the checks below have the combined
2193 * in_interrupt - any node ok (current task context irrelevant)
2194 * GFP_ATOMIC - any node ok
2195 * TIF_MEMDIE - any node ok
2196 * GFP_KERNEL - any node in enclosing hardwalled cpuset ok
2197 * GFP_USER - only nodes in current tasks mems allowed ok.
2200 * Don't call cpuset_zone_allowed_softwall if you can't sleep, unless you
2201 * pass in the __GFP_HARDWALL flag set in gfp_flag, which disables
2202 * the code that might scan up ancestor cpusets and sleep.
2205 int __cpuset_zone_allowed_softwall(struct zone *z, gfp_t gfp_mask)
2207 int node; /* node that zone z is on */
2208 const struct cpuset *cs; /* current cpuset ancestors */
2209 int allowed; /* is allocation in zone z allowed? */
2211 if (in_interrupt() || (gfp_mask & __GFP_THISNODE))
2213 node = zone_to_nid(z);
2214 might_sleep_if(!(gfp_mask & __GFP_HARDWALL));
2215 if (node_isset(node, current->mems_allowed))
2218 * Allow tasks that have access to memory reserves because they have
2219 * been OOM killed to get memory anywhere.
2221 if (unlikely(test_thread_flag(TIF_MEMDIE)))
2223 if (gfp_mask & __GFP_HARDWALL) /* If hardwall request, stop here */
2226 if (current->flags & PF_EXITING) /* Let dying task have memory */
2229 /* Not hardwall and node outside mems_allowed: scan up cpusets */
2230 mutex_lock(&callback_mutex);
2233 cs = nearest_hardwall_ancestor(task_cs(current));
2234 task_unlock(current);
2236 allowed = node_isset(node, cs->mems_allowed);
2237 mutex_unlock(&callback_mutex);
2242 * cpuset_zone_allowed_hardwall - Can we allocate on zone z's memory node?
2243 * @z: is this zone on an allowed node?
2244 * @gfp_mask: memory allocation flags
2246 * If we're in interrupt, yes, we can always allocate.
2247 * If __GFP_THISNODE is set, yes, we can always allocate. If zone
2248 * z's node is in our tasks mems_allowed, yes. If the task has been
2249 * OOM killed and has access to memory reserves as specified by the
2250 * TIF_MEMDIE flag, yes. Otherwise, no.
2252 * The __GFP_THISNODE placement logic is really handled elsewhere,
2253 * by forcibly using a zonelist starting at a specified node, and by
2254 * (in get_page_from_freelist()) refusing to consider the zones for
2255 * any node on the zonelist except the first. By the time any such
2256 * calls get to this routine, we should just shut up and say 'yes'.
2258 * Unlike the cpuset_zone_allowed_softwall() variant, above,
2259 * this variant requires that the zone be in the current tasks
2260 * mems_allowed or that we're in interrupt. It does not scan up the
2261 * cpuset hierarchy for the nearest enclosing mem_exclusive cpuset.
2265 int __cpuset_zone_allowed_hardwall(struct zone *z, gfp_t gfp_mask)
2267 int node; /* node that zone z is on */
2269 if (in_interrupt() || (gfp_mask & __GFP_THISNODE))
2271 node = zone_to_nid(z);
2272 if (node_isset(node, current->mems_allowed))
2275 * Allow tasks that have access to memory reserves because they have
2276 * been OOM killed to get memory anywhere.
2278 if (unlikely(test_thread_flag(TIF_MEMDIE)))
2284 * cpuset_lock - lock out any changes to cpuset structures
2286 * The out of memory (oom) code needs to mutex_lock cpusets
2287 * from being changed while it scans the tasklist looking for a
2288 * task in an overlapping cpuset. Expose callback_mutex via this
2289 * cpuset_lock() routine, so the oom code can lock it, before
2290 * locking the task list. The tasklist_lock is a spinlock, so
2291 * must be taken inside callback_mutex.
2294 void cpuset_lock(void)
2296 mutex_lock(&callback_mutex);
2300 * cpuset_unlock - release lock on cpuset changes
2302 * Undo the lock taken in a previous cpuset_lock() call.
2305 void cpuset_unlock(void)
2307 mutex_unlock(&callback_mutex);
2311 * cpuset_mem_spread_node() - On which node to begin search for a page
2313 * If a task is marked PF_SPREAD_PAGE or PF_SPREAD_SLAB (as for
2314 * tasks in a cpuset with is_spread_page or is_spread_slab set),
2315 * and if the memory allocation used cpuset_mem_spread_node()
2316 * to determine on which node to start looking, as it will for
2317 * certain page cache or slab cache pages such as used for file
2318 * system buffers and inode caches, then instead of starting on the
2319 * local node to look for a free page, rather spread the starting
2320 * node around the tasks mems_allowed nodes.
2322 * We don't have to worry about the returned node being offline
2323 * because "it can't happen", and even if it did, it would be ok.
2325 * The routines calling guarantee_online_mems() are careful to
2326 * only set nodes in task->mems_allowed that are online. So it
2327 * should not be possible for the following code to return an
2328 * offline node. But if it did, that would be ok, as this routine
2329 * is not returning the node where the allocation must be, only
2330 * the node where the search should start. The zonelist passed to
2331 * __alloc_pages() will include all nodes. If the slab allocator
2332 * is passed an offline node, it will fall back to the local node.
2333 * See kmem_cache_alloc_node().
2336 int cpuset_mem_spread_node(void)
2340 node = next_node(current->cpuset_mem_spread_rotor, current->mems_allowed);
2341 if (node == MAX_NUMNODES)
2342 node = first_node(current->mems_allowed);
2343 current->cpuset_mem_spread_rotor = node;
2346 EXPORT_SYMBOL_GPL(cpuset_mem_spread_node);
2349 * cpuset_mems_allowed_intersects - Does @tsk1's mems_allowed intersect @tsk2's?
2350 * @tsk1: pointer to task_struct of some task.
2351 * @tsk2: pointer to task_struct of some other task.
2353 * Description: Return true if @tsk1's mems_allowed intersects the
2354 * mems_allowed of @tsk2. Used by the OOM killer to determine if
2355 * one of the task's memory usage might impact the memory available
2359 int cpuset_mems_allowed_intersects(const struct task_struct *tsk1,
2360 const struct task_struct *tsk2)
2362 return nodes_intersects(tsk1->mems_allowed, tsk2->mems_allowed);
2366 * cpuset_print_task_mems_allowed - prints task's cpuset and mems_allowed
2367 * @task: pointer to task_struct of some task.
2369 * Description: Prints @task's name, cpuset name, and cached copy of its
2370 * mems_allowed to the kernel log. Must hold task_lock(task) to allow
2371 * dereferencing task_cs(task).
2373 void cpuset_print_task_mems_allowed(struct task_struct *tsk)
2375 struct dentry *dentry;
2377 dentry = task_cs(tsk)->css.cgroup->dentry;
2378 spin_lock(&cpuset_buffer_lock);
2379 snprintf(cpuset_name, CPUSET_NAME_LEN,
2380 dentry ? (const char *)dentry->d_name.name : "/");
2381 nodelist_scnprintf(cpuset_nodelist, CPUSET_NODELIST_LEN,
2383 printk(KERN_INFO "%s cpuset=%s mems_allowed=%s\n",
2384 tsk->comm, cpuset_name, cpuset_nodelist);
2385 spin_unlock(&cpuset_buffer_lock);
2389 * Collection of memory_pressure is suppressed unless
2390 * this flag is enabled by writing "1" to the special
2391 * cpuset file 'memory_pressure_enabled' in the root cpuset.
2394 int cpuset_memory_pressure_enabled __read_mostly;
2397 * cpuset_memory_pressure_bump - keep stats of per-cpuset reclaims.
2399 * Keep a running average of the rate of synchronous (direct)
2400 * page reclaim efforts initiated by tasks in each cpuset.
2402 * This represents the rate at which some task in the cpuset
2403 * ran low on memory on all nodes it was allowed to use, and
2404 * had to enter the kernels page reclaim code in an effort to
2405 * create more free memory by tossing clean pages or swapping
2406 * or writing dirty pages.
2408 * Display to user space in the per-cpuset read-only file
2409 * "memory_pressure". Value displayed is an integer
2410 * representing the recent rate of entry into the synchronous
2411 * (direct) page reclaim by any task attached to the cpuset.
2414 void __cpuset_memory_pressure_bump(void)
2417 fmeter_markevent(&task_cs(current)->fmeter);
2418 task_unlock(current);
2421 #ifdef CONFIG_PROC_PID_CPUSET
2423 * proc_cpuset_show()
2424 * - Print tasks cpuset path into seq_file.
2425 * - Used for /proc/<pid>/cpuset.
2426 * - No need to task_lock(tsk) on this tsk->cpuset reference, as it
2427 * doesn't really matter if tsk->cpuset changes after we read it,
2428 * and we take cgroup_mutex, keeping cpuset_attach() from changing it
2431 static int proc_cpuset_show(struct seq_file *m, void *unused_v)
2434 struct task_struct *tsk;
2436 struct cgroup_subsys_state *css;
2440 buf = kmalloc(PAGE_SIZE, GFP_KERNEL);
2446 tsk = get_pid_task(pid, PIDTYPE_PID);
2452 css = task_subsys_state(tsk, cpuset_subsys_id);
2453 retval = cgroup_path(css->cgroup, buf, PAGE_SIZE);
2460 put_task_struct(tsk);
2467 static int cpuset_open(struct inode *inode, struct file *file)
2469 struct pid *pid = PROC_I(inode)->pid;
2470 return single_open(file, proc_cpuset_show, pid);
2473 const struct file_operations proc_cpuset_operations = {
2474 .open = cpuset_open,
2476 .llseek = seq_lseek,
2477 .release = single_release,
2479 #endif /* CONFIG_PROC_PID_CPUSET */
2481 /* Display task cpus_allowed, mems_allowed in /proc/<pid>/status file. */
2482 void cpuset_task_status_allowed(struct seq_file *m, struct task_struct *task)
2484 seq_printf(m, "Cpus_allowed:\t");
2485 seq_cpumask(m, &task->cpus_allowed);
2486 seq_printf(m, "\n");
2487 seq_printf(m, "Cpus_allowed_list:\t");
2488 seq_cpumask_list(m, &task->cpus_allowed);
2489 seq_printf(m, "\n");
2490 seq_printf(m, "Mems_allowed:\t");
2491 seq_nodemask(m, &task->mems_allowed);
2492 seq_printf(m, "\n");
2493 seq_printf(m, "Mems_allowed_list:\t");
2494 seq_nodemask_list(m, &task->mems_allowed);
2495 seq_printf(m, "\n");