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/export.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 <linux/atomic.h>
59 #include <linux/mutex.h>
60 #include <linux/workqueue.h>
61 #include <linux/cgroup.h>
62 #include <linux/wait.h>
65 * Tracks how many cpusets are currently defined in system.
66 * When there is only one cpuset (the root cpuset) we can
67 * short circuit some hooks.
69 int number_of_cpusets __read_mostly;
71 /* Forward declare cgroup structures */
72 struct cgroup_subsys cpuset_subsys;
75 /* See "Frequency meter" comments, below. */
78 int cnt; /* unprocessed events count */
79 int val; /* most recent output value */
80 time_t time; /* clock (secs) when val computed */
81 spinlock_t lock; /* guards read or write of above */
85 struct cgroup_subsys_state css;
87 unsigned long flags; /* "unsigned long" so bitops work */
88 cpumask_var_t cpus_allowed; /* CPUs allowed to tasks in cpuset */
89 nodemask_t mems_allowed; /* Memory Nodes allowed to tasks */
92 * This is old Memory Nodes tasks took on.
94 * - top_cpuset.old_mems_allowed is initialized to mems_allowed.
95 * - A new cpuset's old_mems_allowed is initialized when some
96 * task is moved into it.
97 * - old_mems_allowed is used in cpuset_migrate_mm() when we change
98 * cpuset.mems_allowed and have tasks' nodemask updated, and
99 * then old_mems_allowed is updated to mems_allowed.
101 nodemask_t old_mems_allowed;
103 struct fmeter fmeter; /* memory_pressure filter */
106 * Tasks are being attached to this cpuset. Used to prevent
107 * zeroing cpus/mems_allowed between ->can_attach() and ->attach().
109 int attach_in_progress;
111 /* partition number for rebuild_sched_domains() */
114 /* for custom sched domain */
115 int relax_domain_level;
118 /* Retrieve the cpuset for a cgroup */
119 static inline struct cpuset *cgroup_cs(struct cgroup *cont)
121 return container_of(cgroup_subsys_state(cont, cpuset_subsys_id),
125 /* Retrieve the cpuset for a task */
126 static inline struct cpuset *task_cs(struct task_struct *task)
128 return container_of(task_subsys_state(task, cpuset_subsys_id),
132 static inline struct cpuset *parent_cs(const struct cpuset *cs)
134 struct cgroup *pcgrp = cs->css.cgroup->parent;
137 return cgroup_cs(pcgrp);
142 static inline bool task_has_mempolicy(struct task_struct *task)
144 return task->mempolicy;
147 static inline bool task_has_mempolicy(struct task_struct *task)
154 /* bits in struct cpuset flags field */
161 CS_SCHED_LOAD_BALANCE,
166 /* convenient tests for these bits */
167 static inline bool is_cpuset_online(const struct cpuset *cs)
169 return test_bit(CS_ONLINE, &cs->flags);
172 static inline int is_cpu_exclusive(const struct cpuset *cs)
174 return test_bit(CS_CPU_EXCLUSIVE, &cs->flags);
177 static inline int is_mem_exclusive(const struct cpuset *cs)
179 return test_bit(CS_MEM_EXCLUSIVE, &cs->flags);
182 static inline int is_mem_hardwall(const struct cpuset *cs)
184 return test_bit(CS_MEM_HARDWALL, &cs->flags);
187 static inline int is_sched_load_balance(const struct cpuset *cs)
189 return test_bit(CS_SCHED_LOAD_BALANCE, &cs->flags);
192 static inline int is_memory_migrate(const struct cpuset *cs)
194 return test_bit(CS_MEMORY_MIGRATE, &cs->flags);
197 static inline int is_spread_page(const struct cpuset *cs)
199 return test_bit(CS_SPREAD_PAGE, &cs->flags);
202 static inline int is_spread_slab(const struct cpuset *cs)
204 return test_bit(CS_SPREAD_SLAB, &cs->flags);
207 static struct cpuset top_cpuset = {
208 .flags = ((1 << CS_ONLINE) | (1 << CS_CPU_EXCLUSIVE) |
209 (1 << CS_MEM_EXCLUSIVE)),
213 * cpuset_for_each_child - traverse online children of a cpuset
214 * @child_cs: loop cursor pointing to the current child
215 * @pos_cgrp: used for iteration
216 * @parent_cs: target cpuset to walk children of
218 * Walk @child_cs through the online children of @parent_cs. Must be used
219 * with RCU read locked.
221 #define cpuset_for_each_child(child_cs, pos_cgrp, parent_cs) \
222 cgroup_for_each_child((pos_cgrp), (parent_cs)->css.cgroup) \
223 if (is_cpuset_online(((child_cs) = cgroup_cs((pos_cgrp)))))
226 * cpuset_for_each_descendant_pre - pre-order walk of a cpuset's descendants
227 * @des_cs: loop cursor pointing to the current descendant
228 * @pos_cgrp: used for iteration
229 * @root_cs: target cpuset to walk ancestor of
231 * Walk @des_cs through the online descendants of @root_cs. Must be used
232 * with RCU read locked. The caller may modify @pos_cgrp by calling
233 * cgroup_rightmost_descendant() to skip subtree.
235 #define cpuset_for_each_descendant_pre(des_cs, pos_cgrp, root_cs) \
236 cgroup_for_each_descendant_pre((pos_cgrp), (root_cs)->css.cgroup) \
237 if (is_cpuset_online(((des_cs) = cgroup_cs((pos_cgrp)))))
240 * There are two global mutexes guarding cpuset structures - cpuset_mutex
241 * and callback_mutex. The latter may nest inside the former. We also
242 * require taking task_lock() when dereferencing a task's cpuset pointer.
243 * See "The task_lock() exception", at the end of this comment.
245 * A task must hold both mutexes to modify cpusets. If a task holds
246 * cpuset_mutex, then it blocks others wanting that mutex, ensuring that it
247 * is the only task able to also acquire callback_mutex and be able to
248 * modify cpusets. It can perform various checks on the cpuset structure
249 * first, knowing nothing will change. It can also allocate memory while
250 * just holding cpuset_mutex. While it is performing these checks, various
251 * callback routines can briefly acquire callback_mutex to query cpusets.
252 * Once it is ready to make the changes, it takes callback_mutex, blocking
255 * Calls to the kernel memory allocator can not be made while holding
256 * callback_mutex, as that would risk double tripping on callback_mutex
257 * from one of the callbacks into the cpuset code from within
260 * If a task is only holding callback_mutex, then it has read-only
263 * Now, the task_struct fields mems_allowed and mempolicy may be changed
264 * by other task, we use alloc_lock in the task_struct fields to protect
267 * The cpuset_common_file_read() handlers only hold callback_mutex across
268 * small pieces of code, such as when reading out possibly multi-word
269 * cpumasks and nodemasks.
271 * Accessing a task's cpuset should be done in accordance with the
272 * guidelines for accessing subsystem state in kernel/cgroup.c
275 static DEFINE_MUTEX(cpuset_mutex);
276 static DEFINE_MUTEX(callback_mutex);
279 * CPU / memory hotplug is handled asynchronously.
281 static void cpuset_hotplug_workfn(struct work_struct *work);
282 static DECLARE_WORK(cpuset_hotplug_work, cpuset_hotplug_workfn);
284 static DECLARE_WAIT_QUEUE_HEAD(cpuset_attach_wq);
287 * This is ugly, but preserves the userspace API for existing cpuset
288 * users. If someone tries to mount the "cpuset" filesystem, we
289 * silently switch it to mount "cgroup" instead
291 static struct dentry *cpuset_mount(struct file_system_type *fs_type,
292 int flags, const char *unused_dev_name, void *data)
294 struct file_system_type *cgroup_fs = get_fs_type("cgroup");
295 struct dentry *ret = ERR_PTR(-ENODEV);
299 "release_agent=/sbin/cpuset_release_agent";
300 ret = cgroup_fs->mount(cgroup_fs, flags,
301 unused_dev_name, mountopts);
302 put_filesystem(cgroup_fs);
307 static struct file_system_type cpuset_fs_type = {
309 .mount = cpuset_mount,
313 * Return in pmask the portion of a cpusets's cpus_allowed that
314 * are online. If none are online, walk up the cpuset hierarchy
315 * until we find one that does have some online cpus. The top
316 * cpuset always has some cpus online.
318 * One way or another, we guarantee to return some non-empty subset
319 * of cpu_online_mask.
321 * Call with callback_mutex held.
323 static void guarantee_online_cpus(const struct cpuset *cs,
324 struct cpumask *pmask)
326 while (!cpumask_intersects(cs->cpus_allowed, cpu_online_mask))
328 cpumask_and(pmask, cs->cpus_allowed, cpu_online_mask);
332 * Return in *pmask the portion of a cpusets's mems_allowed that
333 * are online, with memory. If none are online with memory, walk
334 * up the cpuset hierarchy until we find one that does have some
335 * online mems. The top cpuset always has some mems online.
337 * One way or another, we guarantee to return some non-empty subset
338 * of node_states[N_MEMORY].
340 * Call with callback_mutex held.
342 static void guarantee_online_mems(const struct cpuset *cs, nodemask_t *pmask)
344 while (!nodes_intersects(cs->mems_allowed, node_states[N_MEMORY]))
346 nodes_and(*pmask, cs->mems_allowed, node_states[N_MEMORY]);
350 * update task's spread flag if cpuset's page/slab spread flag is set
352 * Called with callback_mutex/cpuset_mutex held
354 static void cpuset_update_task_spread_flag(struct cpuset *cs,
355 struct task_struct *tsk)
357 if (is_spread_page(cs))
358 tsk->flags |= PF_SPREAD_PAGE;
360 tsk->flags &= ~PF_SPREAD_PAGE;
361 if (is_spread_slab(cs))
362 tsk->flags |= PF_SPREAD_SLAB;
364 tsk->flags &= ~PF_SPREAD_SLAB;
368 * is_cpuset_subset(p, q) - Is cpuset p a subset of cpuset q?
370 * One cpuset is a subset of another if all its allowed CPUs and
371 * Memory Nodes are a subset of the other, and its exclusive flags
372 * are only set if the other's are set. Call holding cpuset_mutex.
375 static int is_cpuset_subset(const struct cpuset *p, const struct cpuset *q)
377 return cpumask_subset(p->cpus_allowed, q->cpus_allowed) &&
378 nodes_subset(p->mems_allowed, q->mems_allowed) &&
379 is_cpu_exclusive(p) <= is_cpu_exclusive(q) &&
380 is_mem_exclusive(p) <= is_mem_exclusive(q);
384 * alloc_trial_cpuset - allocate a trial cpuset
385 * @cs: the cpuset that the trial cpuset duplicates
387 static struct cpuset *alloc_trial_cpuset(const struct cpuset *cs)
389 struct cpuset *trial;
391 trial = kmemdup(cs, sizeof(*cs), GFP_KERNEL);
395 if (!alloc_cpumask_var(&trial->cpus_allowed, GFP_KERNEL)) {
399 cpumask_copy(trial->cpus_allowed, cs->cpus_allowed);
405 * free_trial_cpuset - free the trial cpuset
406 * @trial: the trial cpuset to be freed
408 static void free_trial_cpuset(struct cpuset *trial)
410 free_cpumask_var(trial->cpus_allowed);
415 * validate_change() - Used to validate that any proposed cpuset change
416 * follows the structural rules for cpusets.
418 * If we replaced the flag and mask values of the current cpuset
419 * (cur) with those values in the trial cpuset (trial), would
420 * our various subset and exclusive rules still be valid? Presumes
423 * 'cur' is the address of an actual, in-use cpuset. Operations
424 * such as list traversal that depend on the actual address of the
425 * cpuset in the list must use cur below, not trial.
427 * 'trial' is the address of bulk structure copy of cur, with
428 * perhaps one or more of the fields cpus_allowed, mems_allowed,
429 * or flags changed to new, trial values.
431 * Return 0 if valid, -errno if not.
434 static int validate_change(const struct cpuset *cur, const struct cpuset *trial)
437 struct cpuset *c, *par;
442 /* Each of our child cpusets must be a subset of us */
444 cpuset_for_each_child(c, cont, cur)
445 if (!is_cpuset_subset(c, trial))
448 /* Remaining checks don't apply to root cpuset */
450 if (cur == &top_cpuset)
453 par = parent_cs(cur);
455 /* We must be a subset of our parent cpuset */
457 if (!is_cpuset_subset(trial, par))
461 * If either I or some sibling (!= me) is exclusive, we can't
465 cpuset_for_each_child(c, cont, par) {
466 if ((is_cpu_exclusive(trial) || is_cpu_exclusive(c)) &&
468 cpumask_intersects(trial->cpus_allowed, c->cpus_allowed))
470 if ((is_mem_exclusive(trial) || is_mem_exclusive(c)) &&
472 nodes_intersects(trial->mems_allowed, c->mems_allowed))
477 * Cpusets with tasks - existing or newly being attached - can't
478 * have empty cpus_allowed or mems_allowed.
481 if ((cgroup_task_count(cur->css.cgroup) || cur->attach_in_progress) &&
482 (cpumask_empty(trial->cpus_allowed) ||
483 nodes_empty(trial->mems_allowed)))
494 * Helper routine for generate_sched_domains().
495 * Do cpusets a, b have overlapping cpus_allowed masks?
497 static int cpusets_overlap(struct cpuset *a, struct cpuset *b)
499 return cpumask_intersects(a->cpus_allowed, b->cpus_allowed);
503 update_domain_attr(struct sched_domain_attr *dattr, struct cpuset *c)
505 if (dattr->relax_domain_level < c->relax_domain_level)
506 dattr->relax_domain_level = c->relax_domain_level;
510 static void update_domain_attr_tree(struct sched_domain_attr *dattr,
511 struct cpuset *root_cs)
514 struct cgroup *pos_cgrp;
517 cpuset_for_each_descendant_pre(cp, pos_cgrp, root_cs) {
518 /* skip the whole subtree if @cp doesn't have any CPU */
519 if (cpumask_empty(cp->cpus_allowed)) {
520 pos_cgrp = cgroup_rightmost_descendant(pos_cgrp);
524 if (is_sched_load_balance(cp))
525 update_domain_attr(dattr, cp);
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/cgroups/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 cpuset_mutex 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_var_t **domains,
585 struct sched_domain_attr **attributes)
587 struct cpuset *cp; /* scans q */
588 struct cpuset **csa; /* array of all cpuset ptrs */
589 int csn; /* how many cpuset ptrs in csa so far */
590 int i, j, k; /* indices for partition finding loops */
591 cpumask_var_t *doms; /* resulting partition; i.e. sched domains */
592 struct sched_domain_attr *dattr; /* attributes for custom domains */
593 int ndoms = 0; /* number of sched domains in result */
594 int nslot; /* next empty doms[] struct cpumask slot */
595 struct cgroup *pos_cgrp;
601 /* Special case for the 99% of systems with one, full, sched domain */
602 if (is_sched_load_balance(&top_cpuset)) {
604 doms = alloc_sched_domains(ndoms);
608 dattr = kmalloc(sizeof(struct sched_domain_attr), GFP_KERNEL);
610 *dattr = SD_ATTR_INIT;
611 update_domain_attr_tree(dattr, &top_cpuset);
613 cpumask_copy(doms[0], top_cpuset.cpus_allowed);
618 csa = kmalloc(number_of_cpusets * sizeof(cp), GFP_KERNEL);
624 cpuset_for_each_descendant_pre(cp, pos_cgrp, &top_cpuset) {
626 * Continue traversing beyond @cp iff @cp has some CPUs and
627 * isn't load balancing. The former is obvious. The
628 * latter: All child cpusets contain a subset of the
629 * parent's cpus, so just skip them, and then we call
630 * update_domain_attr_tree() to calc relax_domain_level of
631 * the corresponding sched domain.
633 if (!cpumask_empty(cp->cpus_allowed) &&
634 !is_sched_load_balance(cp))
637 if (is_sched_load_balance(cp))
640 /* skip @cp's subtree */
641 pos_cgrp = cgroup_rightmost_descendant(pos_cgrp);
645 for (i = 0; i < csn; i++)
650 /* Find the best partition (set of sched domains) */
651 for (i = 0; i < csn; i++) {
652 struct cpuset *a = csa[i];
655 for (j = 0; j < csn; j++) {
656 struct cpuset *b = csa[j];
659 if (apn != bpn && cpusets_overlap(a, b)) {
660 for (k = 0; k < csn; k++) {
661 struct cpuset *c = csa[k];
666 ndoms--; /* one less element */
673 * Now we know how many domains to create.
674 * Convert <csn, csa> to <ndoms, doms> and populate cpu masks.
676 doms = alloc_sched_domains(ndoms);
681 * The rest of the code, including the scheduler, can deal with
682 * dattr==NULL case. No need to abort if alloc fails.
684 dattr = kmalloc(ndoms * sizeof(struct sched_domain_attr), GFP_KERNEL);
686 for (nslot = 0, i = 0; i < csn; i++) {
687 struct cpuset *a = csa[i];
692 /* Skip completed partitions */
698 if (nslot == ndoms) {
699 static int warnings = 10;
702 "rebuild_sched_domains confused:"
703 " nslot %d, ndoms %d, csn %d, i %d,"
705 nslot, ndoms, csn, i, apn);
713 *(dattr + nslot) = SD_ATTR_INIT;
714 for (j = i; j < csn; j++) {
715 struct cpuset *b = csa[j];
718 cpumask_or(dp, dp, b->cpus_allowed);
720 update_domain_attr_tree(dattr + nslot, b);
722 /* Done with this partition */
728 BUG_ON(nslot != ndoms);
734 * Fallback to the default domain if kmalloc() failed.
735 * See comments in partition_sched_domains().
746 * Rebuild scheduler domains.
748 * If the flag 'sched_load_balance' of any cpuset with non-empty
749 * 'cpus' changes, or if the 'cpus' allowed changes in any cpuset
750 * which has that flag enabled, or if any cpuset with a non-empty
751 * 'cpus' is removed, then call this routine to rebuild the
752 * scheduler's dynamic sched domains.
754 * Call with cpuset_mutex held. Takes get_online_cpus().
756 static void rebuild_sched_domains_locked(void)
758 struct sched_domain_attr *attr;
762 lockdep_assert_held(&cpuset_mutex);
766 * We have raced with CPU hotplug. Don't do anything to avoid
767 * passing doms with offlined cpu to partition_sched_domains().
768 * Anyways, hotplug work item will rebuild sched domains.
770 if (!cpumask_equal(top_cpuset.cpus_allowed, cpu_active_mask))
773 /* Generate domain masks and attrs */
774 ndoms = generate_sched_domains(&doms, &attr);
776 /* Have scheduler rebuild the domains */
777 partition_sched_domains(ndoms, doms, attr);
781 #else /* !CONFIG_SMP */
782 static void rebuild_sched_domains_locked(void)
785 #endif /* CONFIG_SMP */
787 void rebuild_sched_domains(void)
789 mutex_lock(&cpuset_mutex);
790 rebuild_sched_domains_locked();
791 mutex_unlock(&cpuset_mutex);
795 * effective_cpumask_cpuset - return nearest ancestor with non-empty cpus
796 * @cs: the cpuset in interest
798 * A cpuset's effective cpumask is the cpumask of the nearest ancestor
799 * with non-empty cpus. We use effective cpumask whenever:
800 * - we update tasks' cpus_allowed. (they take on the ancestor's cpumask
801 * if the cpuset they reside in has no cpus)
802 * - we want to retrieve task_cs(tsk)'s cpus_allowed.
804 * Called with cpuset_mutex held. cpuset_cpus_allowed_fallback() is an
805 * exception. See comments there.
807 static struct cpuset *effective_cpumask_cpuset(struct cpuset *cs)
809 while (cpumask_empty(cs->cpus_allowed))
815 * effective_nodemask_cpuset - return nearest ancestor with non-empty mems
816 * @cs: the cpuset in interest
818 * A cpuset's effective nodemask is the nodemask of the nearest ancestor
819 * with non-empty memss. We use effective nodemask whenever:
820 * - we update tasks' mems_allowed. (they take on the ancestor's nodemask
821 * if the cpuset they reside in has no mems)
822 * - we want to retrieve task_cs(tsk)'s mems_allowed.
824 * Called with cpuset_mutex held.
826 static struct cpuset *effective_nodemask_cpuset(struct cpuset *cs)
828 while (nodes_empty(cs->mems_allowed))
834 * cpuset_change_cpumask - make a task's cpus_allowed the same as its cpuset's
836 * @scan: struct cgroup_scanner containing the cgroup of the task
838 * Called by cgroup_scan_tasks() for each task in a cgroup whose
839 * cpus_allowed mask needs to be changed.
841 * We don't need to re-check for the cgroup/cpuset membership, since we're
842 * holding cpuset_mutex at this point.
844 static void cpuset_change_cpumask(struct task_struct *tsk,
845 struct cgroup_scanner *scan)
847 struct cpuset *cpus_cs;
849 cpus_cs = effective_cpumask_cpuset(cgroup_cs(scan->cg));
850 set_cpus_allowed_ptr(tsk, cpus_cs->cpus_allowed);
854 * update_tasks_cpumask - Update the cpumasks of tasks in the cpuset.
855 * @cs: the cpuset in which each task's cpus_allowed mask needs to be changed
856 * @heap: if NULL, defer allocating heap memory to cgroup_scan_tasks()
858 * Called with cpuset_mutex held
860 * The cgroup_scan_tasks() function will scan all the tasks in a cgroup,
861 * calling callback functions for each.
863 * No return value. It's guaranteed that cgroup_scan_tasks() always returns 0
866 static void update_tasks_cpumask(struct cpuset *cs, struct ptr_heap *heap)
868 struct cgroup_scanner scan;
870 scan.cg = cs->css.cgroup;
871 scan.test_task = NULL;
872 scan.process_task = cpuset_change_cpumask;
874 cgroup_scan_tasks(&scan);
878 * update_cpumask - update the cpus_allowed mask of a cpuset and all tasks in it
879 * @cs: the cpuset to consider
880 * @buf: buffer of cpu numbers written to this cpuset
882 static int update_cpumask(struct cpuset *cs, struct cpuset *trialcs,
885 struct ptr_heap heap;
887 int is_load_balanced;
889 /* top_cpuset.cpus_allowed tracks cpu_online_mask; it's read-only */
890 if (cs == &top_cpuset)
894 * An empty cpus_allowed is ok only if the cpuset has no tasks.
895 * Since cpulist_parse() fails on an empty mask, we special case
896 * that parsing. The validate_change() call ensures that cpusets
897 * with tasks have cpus.
900 cpumask_clear(trialcs->cpus_allowed);
902 retval = cpulist_parse(buf, trialcs->cpus_allowed);
906 if (!cpumask_subset(trialcs->cpus_allowed, cpu_active_mask))
910 /* Nothing to do if the cpus didn't change */
911 if (cpumask_equal(cs->cpus_allowed, trialcs->cpus_allowed))
914 retval = validate_change(cs, trialcs);
918 retval = heap_init(&heap, PAGE_SIZE, GFP_KERNEL, NULL);
922 is_load_balanced = is_sched_load_balance(trialcs);
924 mutex_lock(&callback_mutex);
925 cpumask_copy(cs->cpus_allowed, trialcs->cpus_allowed);
926 mutex_unlock(&callback_mutex);
929 * Scan tasks in the cpuset, and update the cpumasks of any
930 * that need an update.
932 update_tasks_cpumask(cs, &heap);
936 if (is_load_balanced)
937 rebuild_sched_domains_locked();
944 * Migrate memory region from one set of nodes to another.
946 * Temporarilly set tasks mems_allowed to target nodes of migration,
947 * so that the migration code can allocate pages on these nodes.
949 * Call holding cpuset_mutex, so current's cpuset won't change
950 * during this call, as manage_mutex holds off any cpuset_attach()
951 * calls. Therefore we don't need to take task_lock around the
952 * call to guarantee_online_mems(), as we know no one is changing
955 * While the mm_struct we are migrating is typically from some
956 * other task, the task_struct mems_allowed that we are hacking
957 * is for our current task, which must allocate new pages for that
958 * migrating memory region.
961 static void cpuset_migrate_mm(struct mm_struct *mm, const nodemask_t *from,
962 const nodemask_t *to)
964 struct task_struct *tsk = current;
965 struct cpuset *mems_cs;
967 tsk->mems_allowed = *to;
969 do_migrate_pages(mm, from, to, MPOL_MF_MOVE_ALL);
971 mems_cs = effective_nodemask_cpuset(task_cs(tsk));
972 guarantee_online_mems(mems_cs, &tsk->mems_allowed);
976 * cpuset_change_task_nodemask - change task's mems_allowed and mempolicy
977 * @tsk: the task to change
978 * @newmems: new nodes that the task will be set
980 * In order to avoid seeing no nodes if the old and new nodes are disjoint,
981 * we structure updates as setting all new allowed nodes, then clearing newly
984 static void cpuset_change_task_nodemask(struct task_struct *tsk,
990 * Allow tasks that have access to memory reserves because they have
991 * been OOM killed to get memory anywhere.
993 if (unlikely(test_thread_flag(TIF_MEMDIE)))
995 if (current->flags & PF_EXITING) /* Let dying task have memory */
1000 * Determine if a loop is necessary if another thread is doing
1001 * get_mems_allowed(). If at least one node remains unchanged and
1002 * tsk does not have a mempolicy, then an empty nodemask will not be
1003 * possible when mems_allowed is larger than a word.
1005 need_loop = task_has_mempolicy(tsk) ||
1006 !nodes_intersects(*newmems, tsk->mems_allowed);
1009 write_seqcount_begin(&tsk->mems_allowed_seq);
1011 nodes_or(tsk->mems_allowed, tsk->mems_allowed, *newmems);
1012 mpol_rebind_task(tsk, newmems, MPOL_REBIND_STEP1);
1014 mpol_rebind_task(tsk, newmems, MPOL_REBIND_STEP2);
1015 tsk->mems_allowed = *newmems;
1018 write_seqcount_end(&tsk->mems_allowed_seq);
1024 * Update task's mems_allowed and rebind its mempolicy and vmas' mempolicy
1025 * of it to cpuset's new mems_allowed, and migrate pages to new nodes if
1026 * memory_migrate flag is set. Called with cpuset_mutex held.
1028 static void cpuset_change_nodemask(struct task_struct *p,
1029 struct cgroup_scanner *scan)
1031 struct cpuset *cs = cgroup_cs(scan->cg);
1032 struct mm_struct *mm;
1034 nodemask_t *newmems = scan->data;
1036 cpuset_change_task_nodemask(p, newmems);
1038 mm = get_task_mm(p);
1042 migrate = is_memory_migrate(cs);
1044 mpol_rebind_mm(mm, &cs->mems_allowed);
1046 cpuset_migrate_mm(mm, &cs->old_mems_allowed, newmems);
1050 static void *cpuset_being_rebound;
1053 * update_tasks_nodemask - Update the nodemasks of tasks in the cpuset.
1054 * @cs: the cpuset in which each task's mems_allowed mask needs to be changed
1055 * @heap: if NULL, defer allocating heap memory to cgroup_scan_tasks()
1057 * Called with cpuset_mutex held
1058 * No return value. It's guaranteed that cgroup_scan_tasks() always returns 0
1061 static void update_tasks_nodemask(struct cpuset *cs, struct ptr_heap *heap)
1063 static nodemask_t newmems; /* protected by cpuset_mutex */
1064 struct cgroup_scanner scan;
1065 struct cpuset *mems_cs = effective_nodemask_cpuset(cs);
1067 cpuset_being_rebound = cs; /* causes mpol_dup() rebind */
1069 guarantee_online_mems(mems_cs, &newmems);
1071 scan.cg = cs->css.cgroup;
1072 scan.test_task = NULL;
1073 scan.process_task = cpuset_change_nodemask;
1075 scan.data = &newmems;
1078 * The mpol_rebind_mm() call takes mmap_sem, which we couldn't
1079 * take while holding tasklist_lock. Forks can happen - the
1080 * mpol_dup() cpuset_being_rebound check will catch such forks,
1081 * and rebind their vma mempolicies too. Because we still hold
1082 * the global cpuset_mutex, we know that no other rebind effort
1083 * will be contending for the global variable cpuset_being_rebound.
1084 * It's ok if we rebind the same mm twice; mpol_rebind_mm()
1085 * is idempotent. Also migrate pages in each mm to new nodes.
1087 cgroup_scan_tasks(&scan);
1090 * All the tasks' nodemasks have been updated, update
1091 * cs->old_mems_allowed.
1093 cs->old_mems_allowed = newmems;
1095 /* We're done rebinding vmas to this cpuset's new mems_allowed. */
1096 cpuset_being_rebound = NULL;
1100 * Handle user request to change the 'mems' memory placement
1101 * of a cpuset. Needs to validate the request, update the
1102 * cpusets mems_allowed, and for each task in the cpuset,
1103 * update mems_allowed and rebind task's mempolicy and any vma
1104 * mempolicies and if the cpuset is marked 'memory_migrate',
1105 * migrate the tasks pages to the new memory.
1107 * Call with cpuset_mutex held. May take callback_mutex during call.
1108 * Will take tasklist_lock, scan tasklist for tasks in cpuset cs,
1109 * lock each such tasks mm->mmap_sem, scan its vma's and rebind
1110 * their mempolicies to the cpusets new mems_allowed.
1112 static int update_nodemask(struct cpuset *cs, struct cpuset *trialcs,
1116 struct ptr_heap heap;
1119 * top_cpuset.mems_allowed tracks node_stats[N_MEMORY];
1122 if (cs == &top_cpuset) {
1128 * An empty mems_allowed is ok iff there are no tasks in the cpuset.
1129 * Since nodelist_parse() fails on an empty mask, we special case
1130 * that parsing. The validate_change() call ensures that cpusets
1131 * with tasks have memory.
1134 nodes_clear(trialcs->mems_allowed);
1136 retval = nodelist_parse(buf, trialcs->mems_allowed);
1140 if (!nodes_subset(trialcs->mems_allowed,
1141 node_states[N_MEMORY])) {
1147 if (nodes_equal(cs->mems_allowed, trialcs->mems_allowed)) {
1148 retval = 0; /* Too easy - nothing to do */
1151 retval = validate_change(cs, trialcs);
1155 retval = heap_init(&heap, PAGE_SIZE, GFP_KERNEL, NULL);
1159 mutex_lock(&callback_mutex);
1160 cs->mems_allowed = trialcs->mems_allowed;
1161 mutex_unlock(&callback_mutex);
1163 update_tasks_nodemask(cs, &heap);
1170 int current_cpuset_is_being_rebound(void)
1172 return task_cs(current) == cpuset_being_rebound;
1175 static int update_relax_domain_level(struct cpuset *cs, s64 val)
1178 if (val < -1 || val >= sched_domain_level_max)
1182 if (val != cs->relax_domain_level) {
1183 cs->relax_domain_level = val;
1184 if (!cpumask_empty(cs->cpus_allowed) &&
1185 is_sched_load_balance(cs))
1186 rebuild_sched_domains_locked();
1193 * cpuset_change_flag - make a task's spread flags the same as its cpuset's
1194 * @tsk: task to be updated
1195 * @scan: struct cgroup_scanner containing the cgroup of the task
1197 * Called by cgroup_scan_tasks() for each task in a cgroup.
1199 * We don't need to re-check for the cgroup/cpuset membership, since we're
1200 * holding cpuset_mutex at this point.
1202 static void cpuset_change_flag(struct task_struct *tsk,
1203 struct cgroup_scanner *scan)
1205 cpuset_update_task_spread_flag(cgroup_cs(scan->cg), tsk);
1209 * update_tasks_flags - update the spread flags of tasks in the cpuset.
1210 * @cs: the cpuset in which each task's spread flags needs to be changed
1211 * @heap: if NULL, defer allocating heap memory to cgroup_scan_tasks()
1213 * Called with cpuset_mutex held
1215 * The cgroup_scan_tasks() function will scan all the tasks in a cgroup,
1216 * calling callback functions for each.
1218 * No return value. It's guaranteed that cgroup_scan_tasks() always returns 0
1221 static void update_tasks_flags(struct cpuset *cs, struct ptr_heap *heap)
1223 struct cgroup_scanner scan;
1225 scan.cg = cs->css.cgroup;
1226 scan.test_task = NULL;
1227 scan.process_task = cpuset_change_flag;
1229 cgroup_scan_tasks(&scan);
1233 * update_flag - read a 0 or a 1 in a file and update associated flag
1234 * bit: the bit to update (see cpuset_flagbits_t)
1235 * cs: the cpuset to update
1236 * turning_on: whether the flag is being set or cleared
1238 * Call with cpuset_mutex held.
1241 static int update_flag(cpuset_flagbits_t bit, struct cpuset *cs,
1244 struct cpuset *trialcs;
1245 int balance_flag_changed;
1246 int spread_flag_changed;
1247 struct ptr_heap heap;
1250 trialcs = alloc_trial_cpuset(cs);
1255 set_bit(bit, &trialcs->flags);
1257 clear_bit(bit, &trialcs->flags);
1259 err = validate_change(cs, trialcs);
1263 err = heap_init(&heap, PAGE_SIZE, GFP_KERNEL, NULL);
1267 balance_flag_changed = (is_sched_load_balance(cs) !=
1268 is_sched_load_balance(trialcs));
1270 spread_flag_changed = ((is_spread_slab(cs) != is_spread_slab(trialcs))
1271 || (is_spread_page(cs) != is_spread_page(trialcs)));
1273 mutex_lock(&callback_mutex);
1274 cs->flags = trialcs->flags;
1275 mutex_unlock(&callback_mutex);
1277 if (!cpumask_empty(trialcs->cpus_allowed) && balance_flag_changed)
1278 rebuild_sched_domains_locked();
1280 if (spread_flag_changed)
1281 update_tasks_flags(cs, &heap);
1284 free_trial_cpuset(trialcs);
1289 * Frequency meter - How fast is some event occurring?
1291 * These routines manage a digitally filtered, constant time based,
1292 * event frequency meter. There are four routines:
1293 * fmeter_init() - initialize a frequency meter.
1294 * fmeter_markevent() - called each time the event happens.
1295 * fmeter_getrate() - returns the recent rate of such events.
1296 * fmeter_update() - internal routine used to update fmeter.
1298 * A common data structure is passed to each of these routines,
1299 * which is used to keep track of the state required to manage the
1300 * frequency meter and its digital filter.
1302 * The filter works on the number of events marked per unit time.
1303 * The filter is single-pole low-pass recursive (IIR). The time unit
1304 * is 1 second. Arithmetic is done using 32-bit integers scaled to
1305 * simulate 3 decimal digits of precision (multiplied by 1000).
1307 * With an FM_COEF of 933, and a time base of 1 second, the filter
1308 * has a half-life of 10 seconds, meaning that if the events quit
1309 * happening, then the rate returned from the fmeter_getrate()
1310 * will be cut in half each 10 seconds, until it converges to zero.
1312 * It is not worth doing a real infinitely recursive filter. If more
1313 * than FM_MAXTICKS ticks have elapsed since the last filter event,
1314 * just compute FM_MAXTICKS ticks worth, by which point the level
1317 * Limit the count of unprocessed events to FM_MAXCNT, so as to avoid
1318 * arithmetic overflow in the fmeter_update() routine.
1320 * Given the simple 32 bit integer arithmetic used, this meter works
1321 * best for reporting rates between one per millisecond (msec) and
1322 * one per 32 (approx) seconds. At constant rates faster than one
1323 * per msec it maxes out at values just under 1,000,000. At constant
1324 * rates between one per msec, and one per second it will stabilize
1325 * to a value N*1000, where N is the rate of events per second.
1326 * At constant rates between one per second and one per 32 seconds,
1327 * it will be choppy, moving up on the seconds that have an event,
1328 * and then decaying until the next event. At rates slower than
1329 * about one in 32 seconds, it decays all the way back to zero between
1333 #define FM_COEF 933 /* coefficient for half-life of 10 secs */
1334 #define FM_MAXTICKS ((time_t)99) /* useless computing more ticks than this */
1335 #define FM_MAXCNT 1000000 /* limit cnt to avoid overflow */
1336 #define FM_SCALE 1000 /* faux fixed point scale */
1338 /* Initialize a frequency meter */
1339 static void fmeter_init(struct fmeter *fmp)
1344 spin_lock_init(&fmp->lock);
1347 /* Internal meter update - process cnt events and update value */
1348 static void fmeter_update(struct fmeter *fmp)
1350 time_t now = get_seconds();
1351 time_t ticks = now - fmp->time;
1356 ticks = min(FM_MAXTICKS, ticks);
1358 fmp->val = (FM_COEF * fmp->val) / FM_SCALE;
1361 fmp->val += ((FM_SCALE - FM_COEF) * fmp->cnt) / FM_SCALE;
1365 /* Process any previous ticks, then bump cnt by one (times scale). */
1366 static void fmeter_markevent(struct fmeter *fmp)
1368 spin_lock(&fmp->lock);
1370 fmp->cnt = min(FM_MAXCNT, fmp->cnt + FM_SCALE);
1371 spin_unlock(&fmp->lock);
1374 /* Process any previous ticks, then return current value. */
1375 static int fmeter_getrate(struct fmeter *fmp)
1379 spin_lock(&fmp->lock);
1382 spin_unlock(&fmp->lock);
1386 /* Called by cgroups to determine if a cpuset is usable; cpuset_mutex held */
1387 static int cpuset_can_attach(struct cgroup *cgrp, struct cgroup_taskset *tset)
1389 struct cpuset *cs = cgroup_cs(cgrp);
1390 struct task_struct *task;
1393 mutex_lock(&cpuset_mutex);
1396 if (cpumask_empty(cs->cpus_allowed) || nodes_empty(cs->mems_allowed))
1399 cgroup_taskset_for_each(task, cgrp, tset) {
1401 * Kthreads which disallow setaffinity shouldn't be moved
1402 * to a new cpuset; we don't want to change their cpu
1403 * affinity and isolating such threads by their set of
1404 * allowed nodes is unnecessary. Thus, cpusets are not
1405 * applicable for such threads. This prevents checking for
1406 * success of set_cpus_allowed_ptr() on all attached tasks
1407 * before cpus_allowed may be changed.
1410 if (task->flags & PF_NO_SETAFFINITY)
1412 ret = security_task_setscheduler(task);
1418 * Mark attach is in progress. This makes validate_change() fail
1419 * changes which zero cpus/mems_allowed.
1421 cs->attach_in_progress++;
1424 mutex_unlock(&cpuset_mutex);
1428 static void cpuset_cancel_attach(struct cgroup *cgrp,
1429 struct cgroup_taskset *tset)
1431 mutex_lock(&cpuset_mutex);
1432 cgroup_cs(cgrp)->attach_in_progress--;
1433 mutex_unlock(&cpuset_mutex);
1437 * Protected by cpuset_mutex. cpus_attach is used only by cpuset_attach()
1438 * but we can't allocate it dynamically there. Define it global and
1439 * allocate from cpuset_init().
1441 static cpumask_var_t cpus_attach;
1443 static void cpuset_attach(struct cgroup *cgrp, struct cgroup_taskset *tset)
1445 /* static buf protected by cpuset_mutex */
1446 static nodemask_t cpuset_attach_nodemask_to;
1447 struct mm_struct *mm;
1448 struct task_struct *task;
1449 struct task_struct *leader = cgroup_taskset_first(tset);
1450 struct cgroup *oldcgrp = cgroup_taskset_cur_cgroup(tset);
1451 struct cpuset *cs = cgroup_cs(cgrp);
1452 struct cpuset *oldcs = cgroup_cs(oldcgrp);
1453 struct cpuset *cpus_cs = effective_cpumask_cpuset(cs);
1454 struct cpuset *mems_cs = effective_nodemask_cpuset(cs);
1456 mutex_lock(&cpuset_mutex);
1458 /* prepare for attach */
1459 if (cs == &top_cpuset)
1460 cpumask_copy(cpus_attach, cpu_possible_mask);
1462 guarantee_online_cpus(cpus_cs, cpus_attach);
1464 guarantee_online_mems(mems_cs, &cpuset_attach_nodemask_to);
1466 cgroup_taskset_for_each(task, cgrp, tset) {
1468 * can_attach beforehand should guarantee that this doesn't
1469 * fail. TODO: have a better way to handle failure here
1471 WARN_ON_ONCE(set_cpus_allowed_ptr(task, cpus_attach));
1473 cpuset_change_task_nodemask(task, &cpuset_attach_nodemask_to);
1474 cpuset_update_task_spread_flag(cs, task);
1478 * Change mm, possibly for multiple threads in a threadgroup. This is
1479 * expensive and may sleep.
1481 cpuset_attach_nodemask_to = cs->mems_allowed;
1482 mm = get_task_mm(leader);
1484 struct cpuset *mems_oldcs = effective_nodemask_cpuset(oldcs);
1486 mpol_rebind_mm(mm, &cpuset_attach_nodemask_to);
1487 if (is_memory_migrate(cs))
1488 cpuset_migrate_mm(mm, &mems_oldcs->mems_allowed,
1489 &cpuset_attach_nodemask_to);
1493 cs->old_mems_allowed = cpuset_attach_nodemask_to;
1495 cs->attach_in_progress--;
1496 if (!cs->attach_in_progress)
1497 wake_up(&cpuset_attach_wq);
1499 mutex_unlock(&cpuset_mutex);
1502 /* The various types of files and directories in a cpuset file system */
1505 FILE_MEMORY_MIGRATE,
1511 FILE_SCHED_LOAD_BALANCE,
1512 FILE_SCHED_RELAX_DOMAIN_LEVEL,
1513 FILE_MEMORY_PRESSURE_ENABLED,
1514 FILE_MEMORY_PRESSURE,
1517 } cpuset_filetype_t;
1519 static int cpuset_write_u64(struct cgroup *cgrp, struct cftype *cft, u64 val)
1521 struct cpuset *cs = cgroup_cs(cgrp);
1522 cpuset_filetype_t type = cft->private;
1523 int retval = -ENODEV;
1525 mutex_lock(&cpuset_mutex);
1526 if (!is_cpuset_online(cs))
1530 case FILE_CPU_EXCLUSIVE:
1531 retval = update_flag(CS_CPU_EXCLUSIVE, cs, val);
1533 case FILE_MEM_EXCLUSIVE:
1534 retval = update_flag(CS_MEM_EXCLUSIVE, cs, val);
1536 case FILE_MEM_HARDWALL:
1537 retval = update_flag(CS_MEM_HARDWALL, cs, val);
1539 case FILE_SCHED_LOAD_BALANCE:
1540 retval = update_flag(CS_SCHED_LOAD_BALANCE, cs, val);
1542 case FILE_MEMORY_MIGRATE:
1543 retval = update_flag(CS_MEMORY_MIGRATE, cs, val);
1545 case FILE_MEMORY_PRESSURE_ENABLED:
1546 cpuset_memory_pressure_enabled = !!val;
1548 case FILE_MEMORY_PRESSURE:
1551 case FILE_SPREAD_PAGE:
1552 retval = update_flag(CS_SPREAD_PAGE, cs, val);
1554 case FILE_SPREAD_SLAB:
1555 retval = update_flag(CS_SPREAD_SLAB, cs, val);
1562 mutex_unlock(&cpuset_mutex);
1566 static int cpuset_write_s64(struct cgroup *cgrp, struct cftype *cft, s64 val)
1568 struct cpuset *cs = cgroup_cs(cgrp);
1569 cpuset_filetype_t type = cft->private;
1570 int retval = -ENODEV;
1572 mutex_lock(&cpuset_mutex);
1573 if (!is_cpuset_online(cs))
1577 case FILE_SCHED_RELAX_DOMAIN_LEVEL:
1578 retval = update_relax_domain_level(cs, val);
1585 mutex_unlock(&cpuset_mutex);
1590 * Common handling for a write to a "cpus" or "mems" file.
1592 static int cpuset_write_resmask(struct cgroup *cgrp, struct cftype *cft,
1595 struct cpuset *cs = cgroup_cs(cgrp);
1596 struct cpuset *trialcs;
1597 int retval = -ENODEV;
1600 * CPU or memory hotunplug may leave @cs w/o any execution
1601 * resources, in which case the hotplug code asynchronously updates
1602 * configuration and transfers all tasks to the nearest ancestor
1603 * which can execute.
1605 * As writes to "cpus" or "mems" may restore @cs's execution
1606 * resources, wait for the previously scheduled operations before
1607 * proceeding, so that we don't end up keep removing tasks added
1608 * after execution capability is restored.
1610 flush_work(&cpuset_hotplug_work);
1612 mutex_lock(&cpuset_mutex);
1613 if (!is_cpuset_online(cs))
1616 trialcs = alloc_trial_cpuset(cs);
1622 switch (cft->private) {
1624 retval = update_cpumask(cs, trialcs, buf);
1627 retval = update_nodemask(cs, trialcs, buf);
1634 free_trial_cpuset(trialcs);
1636 mutex_unlock(&cpuset_mutex);
1641 * These ascii lists should be read in a single call, by using a user
1642 * buffer large enough to hold the entire map. If read in smaller
1643 * chunks, there is no guarantee of atomicity. Since the display format
1644 * used, list of ranges of sequential numbers, is variable length,
1645 * and since these maps can change value dynamically, one could read
1646 * gibberish by doing partial reads while a list was changing.
1647 * A single large read to a buffer that crosses a page boundary is
1648 * ok, because the result being copied to user land is not recomputed
1649 * across a page fault.
1652 static size_t cpuset_sprintf_cpulist(char *page, struct cpuset *cs)
1656 mutex_lock(&callback_mutex);
1657 count = cpulist_scnprintf(page, PAGE_SIZE, cs->cpus_allowed);
1658 mutex_unlock(&callback_mutex);
1663 static size_t cpuset_sprintf_memlist(char *page, struct cpuset *cs)
1667 mutex_lock(&callback_mutex);
1668 count = nodelist_scnprintf(page, PAGE_SIZE, cs->mems_allowed);
1669 mutex_unlock(&callback_mutex);
1674 static ssize_t cpuset_common_file_read(struct cgroup *cont,
1678 size_t nbytes, loff_t *ppos)
1680 struct cpuset *cs = cgroup_cs(cont);
1681 cpuset_filetype_t type = cft->private;
1686 if (!(page = (char *)__get_free_page(GFP_TEMPORARY)))
1693 s += cpuset_sprintf_cpulist(s, cs);
1696 s += cpuset_sprintf_memlist(s, cs);
1704 retval = simple_read_from_buffer(buf, nbytes, ppos, page, s - page);
1706 free_page((unsigned long)page);
1710 static u64 cpuset_read_u64(struct cgroup *cont, struct cftype *cft)
1712 struct cpuset *cs = cgroup_cs(cont);
1713 cpuset_filetype_t type = cft->private;
1715 case FILE_CPU_EXCLUSIVE:
1716 return is_cpu_exclusive(cs);
1717 case FILE_MEM_EXCLUSIVE:
1718 return is_mem_exclusive(cs);
1719 case FILE_MEM_HARDWALL:
1720 return is_mem_hardwall(cs);
1721 case FILE_SCHED_LOAD_BALANCE:
1722 return is_sched_load_balance(cs);
1723 case FILE_MEMORY_MIGRATE:
1724 return is_memory_migrate(cs);
1725 case FILE_MEMORY_PRESSURE_ENABLED:
1726 return cpuset_memory_pressure_enabled;
1727 case FILE_MEMORY_PRESSURE:
1728 return fmeter_getrate(&cs->fmeter);
1729 case FILE_SPREAD_PAGE:
1730 return is_spread_page(cs);
1731 case FILE_SPREAD_SLAB:
1732 return is_spread_slab(cs);
1737 /* Unreachable but makes gcc happy */
1741 static s64 cpuset_read_s64(struct cgroup *cont, struct cftype *cft)
1743 struct cpuset *cs = cgroup_cs(cont);
1744 cpuset_filetype_t type = cft->private;
1746 case FILE_SCHED_RELAX_DOMAIN_LEVEL:
1747 return cs->relax_domain_level;
1752 /* Unrechable but makes gcc happy */
1758 * for the common functions, 'private' gives the type of file
1761 static struct cftype files[] = {
1764 .read = cpuset_common_file_read,
1765 .write_string = cpuset_write_resmask,
1766 .max_write_len = (100U + 6 * NR_CPUS),
1767 .private = FILE_CPULIST,
1772 .read = cpuset_common_file_read,
1773 .write_string = cpuset_write_resmask,
1774 .max_write_len = (100U + 6 * MAX_NUMNODES),
1775 .private = FILE_MEMLIST,
1779 .name = "cpu_exclusive",
1780 .read_u64 = cpuset_read_u64,
1781 .write_u64 = cpuset_write_u64,
1782 .private = FILE_CPU_EXCLUSIVE,
1786 .name = "mem_exclusive",
1787 .read_u64 = cpuset_read_u64,
1788 .write_u64 = cpuset_write_u64,
1789 .private = FILE_MEM_EXCLUSIVE,
1793 .name = "mem_hardwall",
1794 .read_u64 = cpuset_read_u64,
1795 .write_u64 = cpuset_write_u64,
1796 .private = FILE_MEM_HARDWALL,
1800 .name = "sched_load_balance",
1801 .read_u64 = cpuset_read_u64,
1802 .write_u64 = cpuset_write_u64,
1803 .private = FILE_SCHED_LOAD_BALANCE,
1807 .name = "sched_relax_domain_level",
1808 .read_s64 = cpuset_read_s64,
1809 .write_s64 = cpuset_write_s64,
1810 .private = FILE_SCHED_RELAX_DOMAIN_LEVEL,
1814 .name = "memory_migrate",
1815 .read_u64 = cpuset_read_u64,
1816 .write_u64 = cpuset_write_u64,
1817 .private = FILE_MEMORY_MIGRATE,
1821 .name = "memory_pressure",
1822 .read_u64 = cpuset_read_u64,
1823 .write_u64 = cpuset_write_u64,
1824 .private = FILE_MEMORY_PRESSURE,
1829 .name = "memory_spread_page",
1830 .read_u64 = cpuset_read_u64,
1831 .write_u64 = cpuset_write_u64,
1832 .private = FILE_SPREAD_PAGE,
1836 .name = "memory_spread_slab",
1837 .read_u64 = cpuset_read_u64,
1838 .write_u64 = cpuset_write_u64,
1839 .private = FILE_SPREAD_SLAB,
1843 .name = "memory_pressure_enabled",
1844 .flags = CFTYPE_ONLY_ON_ROOT,
1845 .read_u64 = cpuset_read_u64,
1846 .write_u64 = cpuset_write_u64,
1847 .private = FILE_MEMORY_PRESSURE_ENABLED,
1854 * cpuset_css_alloc - allocate a cpuset css
1855 * cont: control group that the new cpuset will be part of
1858 static struct cgroup_subsys_state *cpuset_css_alloc(struct cgroup *cont)
1863 return &top_cpuset.css;
1865 cs = kzalloc(sizeof(*cs), GFP_KERNEL);
1867 return ERR_PTR(-ENOMEM);
1868 if (!alloc_cpumask_var(&cs->cpus_allowed, GFP_KERNEL)) {
1870 return ERR_PTR(-ENOMEM);
1873 set_bit(CS_SCHED_LOAD_BALANCE, &cs->flags);
1874 cpumask_clear(cs->cpus_allowed);
1875 nodes_clear(cs->mems_allowed);
1876 fmeter_init(&cs->fmeter);
1877 cs->relax_domain_level = -1;
1882 static int cpuset_css_online(struct cgroup *cgrp)
1884 struct cpuset *cs = cgroup_cs(cgrp);
1885 struct cpuset *parent = parent_cs(cs);
1886 struct cpuset *tmp_cs;
1887 struct cgroup *pos_cg;
1892 mutex_lock(&cpuset_mutex);
1894 set_bit(CS_ONLINE, &cs->flags);
1895 if (is_spread_page(parent))
1896 set_bit(CS_SPREAD_PAGE, &cs->flags);
1897 if (is_spread_slab(parent))
1898 set_bit(CS_SPREAD_SLAB, &cs->flags);
1900 number_of_cpusets++;
1902 if (!test_bit(CGRP_CPUSET_CLONE_CHILDREN, &cgrp->flags))
1906 * Clone @parent's configuration if CGRP_CPUSET_CLONE_CHILDREN is
1907 * set. This flag handling is implemented in cgroup core for
1908 * histrical reasons - the flag may be specified during mount.
1910 * Currently, if any sibling cpusets have exclusive cpus or mem, we
1911 * refuse to clone the configuration - thereby refusing the task to
1912 * be entered, and as a result refusing the sys_unshare() or
1913 * clone() which initiated it. If this becomes a problem for some
1914 * users who wish to allow that scenario, then this could be
1915 * changed to grant parent->cpus_allowed-sibling_cpus_exclusive
1916 * (and likewise for mems) to the new cgroup.
1919 cpuset_for_each_child(tmp_cs, pos_cg, parent) {
1920 if (is_mem_exclusive(tmp_cs) || is_cpu_exclusive(tmp_cs)) {
1927 mutex_lock(&callback_mutex);
1928 cs->mems_allowed = parent->mems_allowed;
1929 cpumask_copy(cs->cpus_allowed, parent->cpus_allowed);
1930 mutex_unlock(&callback_mutex);
1932 mutex_unlock(&cpuset_mutex);
1936 static void cpuset_css_offline(struct cgroup *cgrp)
1938 struct cpuset *cs = cgroup_cs(cgrp);
1940 mutex_lock(&cpuset_mutex);
1942 if (is_sched_load_balance(cs))
1943 update_flag(CS_SCHED_LOAD_BALANCE, cs, 0);
1945 number_of_cpusets--;
1946 clear_bit(CS_ONLINE, &cs->flags);
1948 mutex_unlock(&cpuset_mutex);
1952 * If the cpuset being removed has its flag 'sched_load_balance'
1953 * enabled, then simulate turning sched_load_balance off, which
1954 * will call rebuild_sched_domains_locked().
1957 static void cpuset_css_free(struct cgroup *cont)
1959 struct cpuset *cs = cgroup_cs(cont);
1961 free_cpumask_var(cs->cpus_allowed);
1965 struct cgroup_subsys cpuset_subsys = {
1967 .css_alloc = cpuset_css_alloc,
1968 .css_online = cpuset_css_online,
1969 .css_offline = cpuset_css_offline,
1970 .css_free = cpuset_css_free,
1971 .can_attach = cpuset_can_attach,
1972 .cancel_attach = cpuset_cancel_attach,
1973 .attach = cpuset_attach,
1974 .subsys_id = cpuset_subsys_id,
1975 .base_cftypes = files,
1980 * cpuset_init - initialize cpusets at system boot
1982 * Description: Initialize top_cpuset and the cpuset internal file system,
1985 int __init cpuset_init(void)
1989 if (!alloc_cpumask_var(&top_cpuset.cpus_allowed, GFP_KERNEL))
1992 cpumask_setall(top_cpuset.cpus_allowed);
1993 nodes_setall(top_cpuset.mems_allowed);
1995 fmeter_init(&top_cpuset.fmeter);
1996 set_bit(CS_SCHED_LOAD_BALANCE, &top_cpuset.flags);
1997 top_cpuset.relax_domain_level = -1;
1999 err = register_filesystem(&cpuset_fs_type);
2003 if (!alloc_cpumask_var(&cpus_attach, GFP_KERNEL))
2006 number_of_cpusets = 1;
2011 * If CPU and/or memory hotplug handlers, below, unplug any CPUs
2012 * or memory nodes, we need to walk over the cpuset hierarchy,
2013 * removing that CPU or node from all cpusets. If this removes the
2014 * last CPU or node from a cpuset, then move the tasks in the empty
2015 * cpuset to its next-highest non-empty parent.
2017 static void remove_tasks_in_empty_cpuset(struct cpuset *cs)
2019 struct cpuset *parent;
2022 * Find its next-highest non-empty parent, (top cpuset
2023 * has online cpus, so can't be empty).
2025 parent = parent_cs(cs);
2026 while (cpumask_empty(parent->cpus_allowed) ||
2027 nodes_empty(parent->mems_allowed))
2028 parent = parent_cs(parent);
2030 if (cgroup_transfer_tasks(parent->css.cgroup, cs->css.cgroup)) {
2032 printk(KERN_ERR "cpuset: failed to transfer tasks out of empty cpuset %s\n",
2033 cgroup_name(cs->css.cgroup));
2039 * cpuset_hotplug_update_tasks - update tasks in a cpuset for hotunplug
2040 * @cs: cpuset in interest
2042 * Compare @cs's cpu and mem masks against top_cpuset and if some have gone
2043 * offline, update @cs accordingly. If @cs ends up with no CPU or memory,
2044 * all its tasks are moved to the nearest ancestor with both resources.
2046 static void cpuset_hotplug_update_tasks(struct cpuset *cs)
2048 static cpumask_t off_cpus;
2049 static nodemask_t off_mems;
2053 wait_event(cpuset_attach_wq, cs->attach_in_progress == 0);
2055 mutex_lock(&cpuset_mutex);
2058 * We have raced with task attaching. We wait until attaching
2059 * is finished, so we won't attach a task to an empty cpuset.
2061 if (cs->attach_in_progress) {
2062 mutex_unlock(&cpuset_mutex);
2066 cpumask_andnot(&off_cpus, cs->cpus_allowed, top_cpuset.cpus_allowed);
2067 nodes_andnot(off_mems, cs->mems_allowed, top_cpuset.mems_allowed);
2069 /* remove offline cpus from @cs */
2070 if (!cpumask_empty(&off_cpus)) {
2071 mutex_lock(&callback_mutex);
2072 cpumask_andnot(cs->cpus_allowed, cs->cpus_allowed, &off_cpus);
2073 mutex_unlock(&callback_mutex);
2074 update_tasks_cpumask(cs, NULL);
2077 /* remove offline mems from @cs */
2078 if (!nodes_empty(off_mems)) {
2079 mutex_lock(&callback_mutex);
2080 nodes_andnot(cs->mems_allowed, cs->mems_allowed, off_mems);
2081 mutex_unlock(&callback_mutex);
2082 update_tasks_nodemask(cs, NULL);
2085 is_empty = cpumask_empty(cs->cpus_allowed) ||
2086 nodes_empty(cs->mems_allowed);
2088 mutex_unlock(&cpuset_mutex);
2091 * If @cs became empty, move tasks to the nearest ancestor with
2092 * execution resources. This is full cgroup operation which will
2093 * also call back into cpuset. Should be done outside any lock.
2096 remove_tasks_in_empty_cpuset(cs);
2100 * cpuset_hotplug_workfn - handle CPU/memory hotunplug for a cpuset
2102 * This function is called after either CPU or memory configuration has
2103 * changed and updates cpuset accordingly. The top_cpuset is always
2104 * synchronized to cpu_active_mask and N_MEMORY, which is necessary in
2105 * order to make cpusets transparent (of no affect) on systems that are
2106 * actively using CPU hotplug but making no active use of cpusets.
2108 * Non-root cpusets are only affected by offlining. If any CPUs or memory
2109 * nodes have been taken down, cpuset_hotplug_update_tasks() is invoked on
2112 * Note that CPU offlining during suspend is ignored. We don't modify
2113 * cpusets across suspend/resume cycles at all.
2115 static void cpuset_hotplug_workfn(struct work_struct *work)
2117 static cpumask_t new_cpus, tmp_cpus;
2118 static nodemask_t new_mems, tmp_mems;
2119 bool cpus_updated, mems_updated;
2120 bool cpus_offlined, mems_offlined;
2122 mutex_lock(&cpuset_mutex);
2124 /* fetch the available cpus/mems and find out which changed how */
2125 cpumask_copy(&new_cpus, cpu_active_mask);
2126 new_mems = node_states[N_MEMORY];
2128 cpus_updated = !cpumask_equal(top_cpuset.cpus_allowed, &new_cpus);
2129 cpus_offlined = cpumask_andnot(&tmp_cpus, top_cpuset.cpus_allowed,
2132 mems_updated = !nodes_equal(top_cpuset.mems_allowed, new_mems);
2133 nodes_andnot(tmp_mems, top_cpuset.mems_allowed, new_mems);
2134 mems_offlined = !nodes_empty(tmp_mems);
2136 /* synchronize cpus_allowed to cpu_active_mask */
2138 mutex_lock(&callback_mutex);
2139 cpumask_copy(top_cpuset.cpus_allowed, &new_cpus);
2140 mutex_unlock(&callback_mutex);
2141 /* we don't mess with cpumasks of tasks in top_cpuset */
2144 /* synchronize mems_allowed to N_MEMORY */
2146 mutex_lock(&callback_mutex);
2147 top_cpuset.mems_allowed = new_mems;
2148 mutex_unlock(&callback_mutex);
2149 update_tasks_nodemask(&top_cpuset, NULL);
2152 mutex_unlock(&cpuset_mutex);
2154 /* if cpus or mems went down, we need to propagate to descendants */
2155 if (cpus_offlined || mems_offlined) {
2157 struct cgroup *pos_cgrp;
2160 cpuset_for_each_descendant_pre(cs, pos_cgrp, &top_cpuset) {
2161 if (!css_tryget(&cs->css))
2165 cpuset_hotplug_update_tasks(cs);
2173 /* rebuild sched domains if cpus_allowed has changed */
2175 rebuild_sched_domains();
2178 void cpuset_update_active_cpus(bool cpu_online)
2181 * We're inside cpu hotplug critical region which usually nests
2182 * inside cgroup synchronization. Bounce actual hotplug processing
2183 * to a work item to avoid reverse locking order.
2185 * We still need to do partition_sched_domains() synchronously;
2186 * otherwise, the scheduler will get confused and put tasks to the
2187 * dead CPU. Fall back to the default single domain.
2188 * cpuset_hotplug_workfn() will rebuild it as necessary.
2190 partition_sched_domains(1, NULL, NULL);
2191 schedule_work(&cpuset_hotplug_work);
2195 * Keep top_cpuset.mems_allowed tracking node_states[N_MEMORY].
2196 * Call this routine anytime after node_states[N_MEMORY] changes.
2197 * See cpuset_update_active_cpus() for CPU hotplug handling.
2199 static int cpuset_track_online_nodes(struct notifier_block *self,
2200 unsigned long action, void *arg)
2202 schedule_work(&cpuset_hotplug_work);
2206 static struct notifier_block cpuset_track_online_nodes_nb = {
2207 .notifier_call = cpuset_track_online_nodes,
2208 .priority = 10, /* ??! */
2212 * cpuset_init_smp - initialize cpus_allowed
2214 * Description: Finish top cpuset after cpu, node maps are initialized
2216 void __init cpuset_init_smp(void)
2218 cpumask_copy(top_cpuset.cpus_allowed, cpu_active_mask);
2219 top_cpuset.mems_allowed = node_states[N_MEMORY];
2220 top_cpuset.old_mems_allowed = top_cpuset.mems_allowed;
2222 register_hotmemory_notifier(&cpuset_track_online_nodes_nb);
2226 * cpuset_cpus_allowed - return cpus_allowed mask from a tasks cpuset.
2227 * @tsk: pointer to task_struct from which to obtain cpuset->cpus_allowed.
2228 * @pmask: pointer to struct cpumask variable to receive cpus_allowed set.
2230 * Description: Returns the cpumask_var_t cpus_allowed of the cpuset
2231 * attached to the specified @tsk. Guaranteed to return some non-empty
2232 * subset of cpu_online_mask, even if this means going outside the
2236 void cpuset_cpus_allowed(struct task_struct *tsk, struct cpumask *pmask)
2238 struct cpuset *cpus_cs;
2240 mutex_lock(&callback_mutex);
2242 cpus_cs = effective_cpumask_cpuset(task_cs(tsk));
2243 guarantee_online_cpus(cpus_cs, pmask);
2245 mutex_unlock(&callback_mutex);
2248 void cpuset_cpus_allowed_fallback(struct task_struct *tsk)
2250 const struct cpuset *cpus_cs;
2253 cpus_cs = effective_cpumask_cpuset(task_cs(tsk));
2254 do_set_cpus_allowed(tsk, cpus_cs->cpus_allowed);
2258 * We own tsk->cpus_allowed, nobody can change it under us.
2260 * But we used cs && cs->cpus_allowed lockless and thus can
2261 * race with cgroup_attach_task() or update_cpumask() and get
2262 * the wrong tsk->cpus_allowed. However, both cases imply the
2263 * subsequent cpuset_change_cpumask()->set_cpus_allowed_ptr()
2264 * which takes task_rq_lock().
2266 * If we are called after it dropped the lock we must see all
2267 * changes in tsk_cs()->cpus_allowed. Otherwise we can temporary
2268 * set any mask even if it is not right from task_cs() pov,
2269 * the pending set_cpus_allowed_ptr() will fix things.
2271 * select_fallback_rq() will fix things ups and set cpu_possible_mask
2276 void cpuset_init_current_mems_allowed(void)
2278 nodes_setall(current->mems_allowed);
2282 * cpuset_mems_allowed - return mems_allowed mask from a tasks cpuset.
2283 * @tsk: pointer to task_struct from which to obtain cpuset->mems_allowed.
2285 * Description: Returns the nodemask_t mems_allowed of the cpuset
2286 * attached to the specified @tsk. Guaranteed to return some non-empty
2287 * subset of node_states[N_MEMORY], even if this means going outside the
2291 nodemask_t cpuset_mems_allowed(struct task_struct *tsk)
2293 struct cpuset *mems_cs;
2296 mutex_lock(&callback_mutex);
2298 mems_cs = effective_nodemask_cpuset(task_cs(tsk));
2299 guarantee_online_mems(mems_cs, &mask);
2301 mutex_unlock(&callback_mutex);
2307 * cpuset_nodemask_valid_mems_allowed - check nodemask vs. curremt mems_allowed
2308 * @nodemask: the nodemask to be checked
2310 * Are any of the nodes in the nodemask allowed in current->mems_allowed?
2312 int cpuset_nodemask_valid_mems_allowed(nodemask_t *nodemask)
2314 return nodes_intersects(*nodemask, current->mems_allowed);
2318 * nearest_hardwall_ancestor() - Returns the nearest mem_exclusive or
2319 * mem_hardwall ancestor to the specified cpuset. Call holding
2320 * callback_mutex. If no ancestor is mem_exclusive or mem_hardwall
2321 * (an unusual configuration), then returns the root cpuset.
2323 static const struct cpuset *nearest_hardwall_ancestor(const struct cpuset *cs)
2325 while (!(is_mem_exclusive(cs) || is_mem_hardwall(cs)) && parent_cs(cs))
2331 * cpuset_node_allowed_softwall - Can we allocate on a memory node?
2332 * @node: is this an allowed node?
2333 * @gfp_mask: memory allocation flags
2335 * If we're in interrupt, yes, we can always allocate. If __GFP_THISNODE is
2336 * set, yes, we can always allocate. If node is in our task's mems_allowed,
2337 * yes. If it's not a __GFP_HARDWALL request and this node is in the nearest
2338 * hardwalled cpuset ancestor to this task's cpuset, yes. If the task has been
2339 * OOM killed and has access to memory reserves as specified by the TIF_MEMDIE
2343 * If __GFP_HARDWALL is set, cpuset_node_allowed_softwall() reduces to
2344 * cpuset_node_allowed_hardwall(). Otherwise, cpuset_node_allowed_softwall()
2345 * might sleep, and might allow a node from an enclosing cpuset.
2347 * cpuset_node_allowed_hardwall() only handles the simpler case of hardwall
2348 * cpusets, and never sleeps.
2350 * The __GFP_THISNODE placement logic is really handled elsewhere,
2351 * by forcibly using a zonelist starting at a specified node, and by
2352 * (in get_page_from_freelist()) refusing to consider the zones for
2353 * any node on the zonelist except the first. By the time any such
2354 * calls get to this routine, we should just shut up and say 'yes'.
2356 * GFP_USER allocations are marked with the __GFP_HARDWALL bit,
2357 * and do not allow allocations outside the current tasks cpuset
2358 * unless the task has been OOM killed as is marked TIF_MEMDIE.
2359 * GFP_KERNEL allocations are not so marked, so can escape to the
2360 * nearest enclosing hardwalled ancestor cpuset.
2362 * Scanning up parent cpusets requires callback_mutex. The
2363 * __alloc_pages() routine only calls here with __GFP_HARDWALL bit
2364 * _not_ set if it's a GFP_KERNEL allocation, and all nodes in the
2365 * current tasks mems_allowed came up empty on the first pass over
2366 * the zonelist. So only GFP_KERNEL allocations, if all nodes in the
2367 * cpuset are short of memory, might require taking the callback_mutex
2370 * The first call here from mm/page_alloc:get_page_from_freelist()
2371 * has __GFP_HARDWALL set in gfp_mask, enforcing hardwall cpusets,
2372 * so no allocation on a node outside the cpuset is allowed (unless
2373 * in interrupt, of course).
2375 * The second pass through get_page_from_freelist() doesn't even call
2376 * here for GFP_ATOMIC calls. For those calls, the __alloc_pages()
2377 * variable 'wait' is not set, and the bit ALLOC_CPUSET is not set
2378 * in alloc_flags. That logic and the checks below have the combined
2380 * in_interrupt - any node ok (current task context irrelevant)
2381 * GFP_ATOMIC - any node ok
2382 * TIF_MEMDIE - any node ok
2383 * GFP_KERNEL - any node in enclosing hardwalled cpuset ok
2384 * GFP_USER - only nodes in current tasks mems allowed ok.
2387 * Don't call cpuset_node_allowed_softwall if you can't sleep, unless you
2388 * pass in the __GFP_HARDWALL flag set in gfp_flag, which disables
2389 * the code that might scan up ancestor cpusets and sleep.
2391 int __cpuset_node_allowed_softwall(int node, gfp_t gfp_mask)
2393 const struct cpuset *cs; /* current cpuset ancestors */
2394 int allowed; /* is allocation in zone z allowed? */
2396 if (in_interrupt() || (gfp_mask & __GFP_THISNODE))
2398 might_sleep_if(!(gfp_mask & __GFP_HARDWALL));
2399 if (node_isset(node, current->mems_allowed))
2402 * Allow tasks that have access to memory reserves because they have
2403 * been OOM killed to get memory anywhere.
2405 if (unlikely(test_thread_flag(TIF_MEMDIE)))
2407 if (gfp_mask & __GFP_HARDWALL) /* If hardwall request, stop here */
2410 if (current->flags & PF_EXITING) /* Let dying task have memory */
2413 /* Not hardwall and node outside mems_allowed: scan up cpusets */
2414 mutex_lock(&callback_mutex);
2417 cs = nearest_hardwall_ancestor(task_cs(current));
2418 task_unlock(current);
2420 allowed = node_isset(node, cs->mems_allowed);
2421 mutex_unlock(&callback_mutex);
2426 * cpuset_node_allowed_hardwall - Can we allocate on a memory node?
2427 * @node: is this an allowed node?
2428 * @gfp_mask: memory allocation flags
2430 * If we're in interrupt, yes, we can always allocate. If __GFP_THISNODE is
2431 * set, yes, we can always allocate. If node is in our task's mems_allowed,
2432 * yes. If the task has been OOM killed and has access to memory reserves as
2433 * specified by the TIF_MEMDIE flag, yes.
2436 * The __GFP_THISNODE placement logic is really handled elsewhere,
2437 * by forcibly using a zonelist starting at a specified node, and by
2438 * (in get_page_from_freelist()) refusing to consider the zones for
2439 * any node on the zonelist except the first. By the time any such
2440 * calls get to this routine, we should just shut up and say 'yes'.
2442 * Unlike the cpuset_node_allowed_softwall() variant, above,
2443 * this variant requires that the node be in the current task's
2444 * mems_allowed or that we're in interrupt. It does not scan up the
2445 * cpuset hierarchy for the nearest enclosing mem_exclusive cpuset.
2448 int __cpuset_node_allowed_hardwall(int node, gfp_t gfp_mask)
2450 if (in_interrupt() || (gfp_mask & __GFP_THISNODE))
2452 if (node_isset(node, current->mems_allowed))
2455 * Allow tasks that have access to memory reserves because they have
2456 * been OOM killed to get memory anywhere.
2458 if (unlikely(test_thread_flag(TIF_MEMDIE)))
2464 * cpuset_mem_spread_node() - On which node to begin search for a file page
2465 * cpuset_slab_spread_node() - On which node to begin search for a slab page
2467 * If a task is marked PF_SPREAD_PAGE or PF_SPREAD_SLAB (as for
2468 * tasks in a cpuset with is_spread_page or is_spread_slab set),
2469 * and if the memory allocation used cpuset_mem_spread_node()
2470 * to determine on which node to start looking, as it will for
2471 * certain page cache or slab cache pages such as used for file
2472 * system buffers and inode caches, then instead of starting on the
2473 * local node to look for a free page, rather spread the starting
2474 * node around the tasks mems_allowed nodes.
2476 * We don't have to worry about the returned node being offline
2477 * because "it can't happen", and even if it did, it would be ok.
2479 * The routines calling guarantee_online_mems() are careful to
2480 * only set nodes in task->mems_allowed that are online. So it
2481 * should not be possible for the following code to return an
2482 * offline node. But if it did, that would be ok, as this routine
2483 * is not returning the node where the allocation must be, only
2484 * the node where the search should start. The zonelist passed to
2485 * __alloc_pages() will include all nodes. If the slab allocator
2486 * is passed an offline node, it will fall back to the local node.
2487 * See kmem_cache_alloc_node().
2490 static int cpuset_spread_node(int *rotor)
2494 node = next_node(*rotor, current->mems_allowed);
2495 if (node == MAX_NUMNODES)
2496 node = first_node(current->mems_allowed);
2501 int cpuset_mem_spread_node(void)
2503 if (current->cpuset_mem_spread_rotor == NUMA_NO_NODE)
2504 current->cpuset_mem_spread_rotor =
2505 node_random(¤t->mems_allowed);
2507 return cpuset_spread_node(¤t->cpuset_mem_spread_rotor);
2510 int cpuset_slab_spread_node(void)
2512 if (current->cpuset_slab_spread_rotor == NUMA_NO_NODE)
2513 current->cpuset_slab_spread_rotor =
2514 node_random(¤t->mems_allowed);
2516 return cpuset_spread_node(¤t->cpuset_slab_spread_rotor);
2519 EXPORT_SYMBOL_GPL(cpuset_mem_spread_node);
2522 * cpuset_mems_allowed_intersects - Does @tsk1's mems_allowed intersect @tsk2's?
2523 * @tsk1: pointer to task_struct of some task.
2524 * @tsk2: pointer to task_struct of some other task.
2526 * Description: Return true if @tsk1's mems_allowed intersects the
2527 * mems_allowed of @tsk2. Used by the OOM killer to determine if
2528 * one of the task's memory usage might impact the memory available
2532 int cpuset_mems_allowed_intersects(const struct task_struct *tsk1,
2533 const struct task_struct *tsk2)
2535 return nodes_intersects(tsk1->mems_allowed, tsk2->mems_allowed);
2538 #define CPUSET_NODELIST_LEN (256)
2541 * cpuset_print_task_mems_allowed - prints task's cpuset and mems_allowed
2542 * @task: pointer to task_struct of some task.
2544 * Description: Prints @task's name, cpuset name, and cached copy of its
2545 * mems_allowed to the kernel log. Must hold task_lock(task) to allow
2546 * dereferencing task_cs(task).
2548 void cpuset_print_task_mems_allowed(struct task_struct *tsk)
2550 /* Statically allocated to prevent using excess stack. */
2551 static char cpuset_nodelist[CPUSET_NODELIST_LEN];
2552 static DEFINE_SPINLOCK(cpuset_buffer_lock);
2554 struct cgroup *cgrp = task_cs(tsk)->css.cgroup;
2557 spin_lock(&cpuset_buffer_lock);
2559 nodelist_scnprintf(cpuset_nodelist, CPUSET_NODELIST_LEN,
2561 printk(KERN_INFO "%s cpuset=%s mems_allowed=%s\n",
2562 tsk->comm, cgroup_name(cgrp), cpuset_nodelist);
2564 spin_unlock(&cpuset_buffer_lock);
2569 * Collection of memory_pressure is suppressed unless
2570 * this flag is enabled by writing "1" to the special
2571 * cpuset file 'memory_pressure_enabled' in the root cpuset.
2574 int cpuset_memory_pressure_enabled __read_mostly;
2577 * cpuset_memory_pressure_bump - keep stats of per-cpuset reclaims.
2579 * Keep a running average of the rate of synchronous (direct)
2580 * page reclaim efforts initiated by tasks in each cpuset.
2582 * This represents the rate at which some task in the cpuset
2583 * ran low on memory on all nodes it was allowed to use, and
2584 * had to enter the kernels page reclaim code in an effort to
2585 * create more free memory by tossing clean pages or swapping
2586 * or writing dirty pages.
2588 * Display to user space in the per-cpuset read-only file
2589 * "memory_pressure". Value displayed is an integer
2590 * representing the recent rate of entry into the synchronous
2591 * (direct) page reclaim by any task attached to the cpuset.
2594 void __cpuset_memory_pressure_bump(void)
2597 fmeter_markevent(&task_cs(current)->fmeter);
2598 task_unlock(current);
2601 #ifdef CONFIG_PROC_PID_CPUSET
2603 * proc_cpuset_show()
2604 * - Print tasks cpuset path into seq_file.
2605 * - Used for /proc/<pid>/cpuset.
2606 * - No need to task_lock(tsk) on this tsk->cpuset reference, as it
2607 * doesn't really matter if tsk->cpuset changes after we read it,
2608 * and we take cpuset_mutex, keeping cpuset_attach() from changing it
2611 int proc_cpuset_show(struct seq_file *m, void *unused_v)
2614 struct task_struct *tsk;
2616 struct cgroup_subsys_state *css;
2620 buf = kmalloc(PAGE_SIZE, GFP_KERNEL);
2626 tsk = get_pid_task(pid, PIDTYPE_PID);
2631 css = task_subsys_state(tsk, cpuset_subsys_id);
2632 retval = cgroup_path(css->cgroup, buf, PAGE_SIZE);
2639 put_task_struct(tsk);
2645 #endif /* CONFIG_PROC_PID_CPUSET */
2647 /* Display task mems_allowed in /proc/<pid>/status file. */
2648 void cpuset_task_status_allowed(struct seq_file *m, struct task_struct *task)
2650 seq_printf(m, "Mems_allowed:\t");
2651 seq_nodemask(m, &task->mems_allowed);
2652 seq_printf(m, "\n");
2653 seq_printf(m, "Mems_allowed_list:\t");
2654 seq_nodemask_list(m, &task->mems_allowed);
2655 seq_printf(m, "\n");