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 /* See "Frequency meter" comments, below. */
74 int cnt; /* unprocessed events count */
75 int val; /* most recent output value */
76 time_t time; /* clock (secs) when val computed */
77 spinlock_t lock; /* guards read or write of above */
81 struct cgroup_subsys_state css;
83 unsigned long flags; /* "unsigned long" so bitops work */
84 cpumask_var_t cpus_allowed; /* CPUs allowed to tasks in cpuset */
85 nodemask_t mems_allowed; /* Memory Nodes allowed to tasks */
88 * This is old Memory Nodes tasks took on.
90 * - top_cpuset.old_mems_allowed is initialized to mems_allowed.
91 * - A new cpuset's old_mems_allowed is initialized when some
92 * task is moved into it.
93 * - old_mems_allowed is used in cpuset_migrate_mm() when we change
94 * cpuset.mems_allowed and have tasks' nodemask updated, and
95 * then old_mems_allowed is updated to mems_allowed.
97 nodemask_t old_mems_allowed;
99 struct fmeter fmeter; /* memory_pressure filter */
102 * Tasks are being attached to this cpuset. Used to prevent
103 * zeroing cpus/mems_allowed between ->can_attach() and ->attach().
105 int attach_in_progress;
107 /* partition number for rebuild_sched_domains() */
110 /* for custom sched domain */
111 int relax_domain_level;
114 static inline struct cpuset *css_cs(struct cgroup_subsys_state *css)
116 return css ? container_of(css, struct cpuset, css) : NULL;
119 /* Retrieve the cpuset for a task */
120 static inline struct cpuset *task_cs(struct task_struct *task)
122 return css_cs(task_css(task, cpuset_cgrp_id));
125 static inline struct cpuset *parent_cs(struct cpuset *cs)
127 return css_cs(css_parent(&cs->css));
131 static inline bool task_has_mempolicy(struct task_struct *task)
133 return task->mempolicy;
136 static inline bool task_has_mempolicy(struct task_struct *task)
143 /* bits in struct cpuset flags field */
150 CS_SCHED_LOAD_BALANCE,
155 /* convenient tests for these bits */
156 static inline bool is_cpuset_online(const struct cpuset *cs)
158 return test_bit(CS_ONLINE, &cs->flags);
161 static inline int is_cpu_exclusive(const struct cpuset *cs)
163 return test_bit(CS_CPU_EXCLUSIVE, &cs->flags);
166 static inline int is_mem_exclusive(const struct cpuset *cs)
168 return test_bit(CS_MEM_EXCLUSIVE, &cs->flags);
171 static inline int is_mem_hardwall(const struct cpuset *cs)
173 return test_bit(CS_MEM_HARDWALL, &cs->flags);
176 static inline int is_sched_load_balance(const struct cpuset *cs)
178 return test_bit(CS_SCHED_LOAD_BALANCE, &cs->flags);
181 static inline int is_memory_migrate(const struct cpuset *cs)
183 return test_bit(CS_MEMORY_MIGRATE, &cs->flags);
186 static inline int is_spread_page(const struct cpuset *cs)
188 return test_bit(CS_SPREAD_PAGE, &cs->flags);
191 static inline int is_spread_slab(const struct cpuset *cs)
193 return test_bit(CS_SPREAD_SLAB, &cs->flags);
196 static struct cpuset top_cpuset = {
197 .flags = ((1 << CS_ONLINE) | (1 << CS_CPU_EXCLUSIVE) |
198 (1 << CS_MEM_EXCLUSIVE)),
202 * cpuset_for_each_child - traverse online children of a cpuset
203 * @child_cs: loop cursor pointing to the current child
204 * @pos_css: used for iteration
205 * @parent_cs: target cpuset to walk children of
207 * Walk @child_cs through the online children of @parent_cs. Must be used
208 * with RCU read locked.
210 #define cpuset_for_each_child(child_cs, pos_css, parent_cs) \
211 css_for_each_child((pos_css), &(parent_cs)->css) \
212 if (is_cpuset_online(((child_cs) = css_cs((pos_css)))))
215 * cpuset_for_each_descendant_pre - pre-order walk of a cpuset's descendants
216 * @des_cs: loop cursor pointing to the current descendant
217 * @pos_css: used for iteration
218 * @root_cs: target cpuset to walk ancestor of
220 * Walk @des_cs through the online descendants of @root_cs. Must be used
221 * with RCU read locked. The caller may modify @pos_css by calling
222 * css_rightmost_descendant() to skip subtree. @root_cs is included in the
223 * iteration and the first node to be visited.
225 #define cpuset_for_each_descendant_pre(des_cs, pos_css, root_cs) \
226 css_for_each_descendant_pre((pos_css), &(root_cs)->css) \
227 if (is_cpuset_online(((des_cs) = css_cs((pos_css)))))
230 * There are two global mutexes guarding cpuset structures - cpuset_mutex
231 * and callback_mutex. The latter may nest inside the former. We also
232 * require taking task_lock() when dereferencing a task's cpuset pointer.
233 * See "The task_lock() exception", at the end of this comment.
235 * A task must hold both mutexes to modify cpusets. If a task holds
236 * cpuset_mutex, then it blocks others wanting that mutex, ensuring that it
237 * is the only task able to also acquire callback_mutex and be able to
238 * modify cpusets. It can perform various checks on the cpuset structure
239 * first, knowing nothing will change. It can also allocate memory while
240 * just holding cpuset_mutex. While it is performing these checks, various
241 * callback routines can briefly acquire callback_mutex to query cpusets.
242 * Once it is ready to make the changes, it takes callback_mutex, blocking
245 * Calls to the kernel memory allocator can not be made while holding
246 * callback_mutex, as that would risk double tripping on callback_mutex
247 * from one of the callbacks into the cpuset code from within
250 * If a task is only holding callback_mutex, then it has read-only
253 * Now, the task_struct fields mems_allowed and mempolicy may be changed
254 * by other task, we use alloc_lock in the task_struct fields to protect
257 * The cpuset_common_file_read() handlers only hold callback_mutex across
258 * small pieces of code, such as when reading out possibly multi-word
259 * cpumasks and nodemasks.
261 * Accessing a task's cpuset should be done in accordance with the
262 * guidelines for accessing subsystem state in kernel/cgroup.c
265 static DEFINE_MUTEX(cpuset_mutex);
266 static DEFINE_MUTEX(callback_mutex);
269 * CPU / memory hotplug is handled asynchronously.
271 static void cpuset_hotplug_workfn(struct work_struct *work);
272 static DECLARE_WORK(cpuset_hotplug_work, cpuset_hotplug_workfn);
274 static DECLARE_WAIT_QUEUE_HEAD(cpuset_attach_wq);
277 * This is ugly, but preserves the userspace API for existing cpuset
278 * users. If someone tries to mount the "cpuset" filesystem, we
279 * silently switch it to mount "cgroup" instead
281 static struct dentry *cpuset_mount(struct file_system_type *fs_type,
282 int flags, const char *unused_dev_name, void *data)
284 struct file_system_type *cgroup_fs = get_fs_type("cgroup");
285 struct dentry *ret = ERR_PTR(-ENODEV);
289 "release_agent=/sbin/cpuset_release_agent";
290 ret = cgroup_fs->mount(cgroup_fs, flags,
291 unused_dev_name, mountopts);
292 put_filesystem(cgroup_fs);
297 static struct file_system_type cpuset_fs_type = {
299 .mount = cpuset_mount,
303 * Return in pmask the portion of a cpusets's cpus_allowed that
304 * are online. If none are online, walk up the cpuset hierarchy
305 * until we find one that does have some online cpus. The top
306 * cpuset always has some cpus online.
308 * One way or another, we guarantee to return some non-empty subset
309 * of cpu_online_mask.
311 * Call with callback_mutex held.
313 static void guarantee_online_cpus(struct cpuset *cs, struct cpumask *pmask)
315 while (!cpumask_intersects(cs->cpus_allowed, cpu_online_mask))
317 cpumask_and(pmask, cs->cpus_allowed, cpu_online_mask);
321 * Return in *pmask the portion of a cpusets's mems_allowed that
322 * are online, with memory. If none are online with memory, walk
323 * up the cpuset hierarchy until we find one that does have some
324 * online mems. The top cpuset always has some mems online.
326 * One way or another, we guarantee to return some non-empty subset
327 * of node_states[N_MEMORY].
329 * Call with callback_mutex held.
331 static void guarantee_online_mems(struct cpuset *cs, nodemask_t *pmask)
333 while (!nodes_intersects(cs->mems_allowed, node_states[N_MEMORY]))
335 nodes_and(*pmask, cs->mems_allowed, node_states[N_MEMORY]);
339 * update task's spread flag if cpuset's page/slab spread flag is set
341 * Called with callback_mutex/cpuset_mutex held
343 static void cpuset_update_task_spread_flag(struct cpuset *cs,
344 struct task_struct *tsk)
346 if (is_spread_page(cs))
347 tsk->flags |= PF_SPREAD_PAGE;
349 tsk->flags &= ~PF_SPREAD_PAGE;
350 if (is_spread_slab(cs))
351 tsk->flags |= PF_SPREAD_SLAB;
353 tsk->flags &= ~PF_SPREAD_SLAB;
357 * is_cpuset_subset(p, q) - Is cpuset p a subset of cpuset q?
359 * One cpuset is a subset of another if all its allowed CPUs and
360 * Memory Nodes are a subset of the other, and its exclusive flags
361 * are only set if the other's are set. Call holding cpuset_mutex.
364 static int is_cpuset_subset(const struct cpuset *p, const struct cpuset *q)
366 return cpumask_subset(p->cpus_allowed, q->cpus_allowed) &&
367 nodes_subset(p->mems_allowed, q->mems_allowed) &&
368 is_cpu_exclusive(p) <= is_cpu_exclusive(q) &&
369 is_mem_exclusive(p) <= is_mem_exclusive(q);
373 * alloc_trial_cpuset - allocate a trial cpuset
374 * @cs: the cpuset that the trial cpuset duplicates
376 static struct cpuset *alloc_trial_cpuset(struct cpuset *cs)
378 struct cpuset *trial;
380 trial = kmemdup(cs, sizeof(*cs), GFP_KERNEL);
384 if (!alloc_cpumask_var(&trial->cpus_allowed, GFP_KERNEL)) {
388 cpumask_copy(trial->cpus_allowed, cs->cpus_allowed);
394 * free_trial_cpuset - free the trial cpuset
395 * @trial: the trial cpuset to be freed
397 static void free_trial_cpuset(struct cpuset *trial)
399 free_cpumask_var(trial->cpus_allowed);
404 * validate_change() - Used to validate that any proposed cpuset change
405 * follows the structural rules for cpusets.
407 * If we replaced the flag and mask values of the current cpuset
408 * (cur) with those values in the trial cpuset (trial), would
409 * our various subset and exclusive rules still be valid? Presumes
412 * 'cur' is the address of an actual, in-use cpuset. Operations
413 * such as list traversal that depend on the actual address of the
414 * cpuset in the list must use cur below, not trial.
416 * 'trial' is the address of bulk structure copy of cur, with
417 * perhaps one or more of the fields cpus_allowed, mems_allowed,
418 * or flags changed to new, trial values.
420 * Return 0 if valid, -errno if not.
423 static int validate_change(struct cpuset *cur, struct cpuset *trial)
425 struct cgroup_subsys_state *css;
426 struct cpuset *c, *par;
431 /* Each of our child cpusets must be a subset of us */
433 cpuset_for_each_child(c, css, cur)
434 if (!is_cpuset_subset(c, trial))
437 /* Remaining checks don't apply to root cpuset */
439 if (cur == &top_cpuset)
442 par = parent_cs(cur);
444 /* We must be a subset of our parent cpuset */
446 if (!is_cpuset_subset(trial, par))
450 * If either I or some sibling (!= me) is exclusive, we can't
454 cpuset_for_each_child(c, css, par) {
455 if ((is_cpu_exclusive(trial) || is_cpu_exclusive(c)) &&
457 cpumask_intersects(trial->cpus_allowed, c->cpus_allowed))
459 if ((is_mem_exclusive(trial) || is_mem_exclusive(c)) &&
461 nodes_intersects(trial->mems_allowed, c->mems_allowed))
466 * Cpusets with tasks - existing or newly being attached - can't
467 * be changed to have empty cpus_allowed or mems_allowed.
470 if ((cgroup_has_tasks(cur->css.cgroup) || cur->attach_in_progress)) {
471 if (!cpumask_empty(cur->cpus_allowed) &&
472 cpumask_empty(trial->cpus_allowed))
474 if (!nodes_empty(cur->mems_allowed) &&
475 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 cpumask_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;
503 static void update_domain_attr_tree(struct sched_domain_attr *dattr,
504 struct cpuset *root_cs)
507 struct cgroup_subsys_state *pos_css;
510 cpuset_for_each_descendant_pre(cp, pos_css, root_cs) {
514 /* skip the whole subtree if @cp doesn't have any CPU */
515 if (cpumask_empty(cp->cpus_allowed)) {
516 pos_css = css_rightmost_descendant(pos_css);
520 if (is_sched_load_balance(cp))
521 update_domain_attr(dattr, cp);
527 * generate_sched_domains()
529 * This function builds a partial partition of the systems CPUs
530 * A 'partial partition' is a set of non-overlapping subsets whose
531 * union is a subset of that set.
532 * The output of this function needs to be passed to kernel/sched/core.c
533 * partition_sched_domains() routine, which will rebuild the scheduler's
534 * load balancing domains (sched domains) as specified by that partial
537 * See "What is sched_load_balance" in Documentation/cgroups/cpusets.txt
538 * for a background explanation of this.
540 * Does not return errors, on the theory that the callers of this
541 * routine would rather not worry about failures to rebuild sched
542 * domains when operating in the severe memory shortage situations
543 * that could cause allocation failures below.
545 * Must be called with cpuset_mutex held.
547 * The three key local variables below are:
548 * q - a linked-list queue of cpuset pointers, used to implement a
549 * top-down scan of all cpusets. This scan loads a pointer
550 * to each cpuset marked is_sched_load_balance into the
551 * array 'csa'. For our purposes, rebuilding the schedulers
552 * sched domains, we can ignore !is_sched_load_balance cpusets.
553 * csa - (for CpuSet Array) Array of pointers to all the cpusets
554 * that need to be load balanced, for convenient iterative
555 * access by the subsequent code that finds the best partition,
556 * i.e the set of domains (subsets) of CPUs such that the
557 * cpus_allowed of every cpuset marked is_sched_load_balance
558 * is a subset of one of these domains, while there are as
559 * many such domains as possible, each as small as possible.
560 * doms - Conversion of 'csa' to an array of cpumasks, for passing to
561 * the kernel/sched/core.c routine partition_sched_domains() in a
562 * convenient format, that can be easily compared to the prior
563 * value to determine what partition elements (sched domains)
564 * were changed (added or removed.)
566 * Finding the best partition (set of domains):
567 * The triple nested loops below over i, j, k scan over the
568 * load balanced cpusets (using the array of cpuset pointers in
569 * csa[]) looking for pairs of cpusets that have overlapping
570 * cpus_allowed, but which don't have the same 'pn' partition
571 * number and gives them in the same partition number. It keeps
572 * looping on the 'restart' label until it can no longer find
575 * The union of the cpus_allowed masks from the set of
576 * all cpusets having the same 'pn' value then form the one
577 * element of the partition (one sched domain) to be passed to
578 * partition_sched_domains().
580 static int generate_sched_domains(cpumask_var_t **domains,
581 struct sched_domain_attr **attributes)
583 struct cpuset *cp; /* scans q */
584 struct cpuset **csa; /* array of all cpuset ptrs */
585 int csn; /* how many cpuset ptrs in csa so far */
586 int i, j, k; /* indices for partition finding loops */
587 cpumask_var_t *doms; /* resulting partition; i.e. sched domains */
588 struct sched_domain_attr *dattr; /* attributes for custom domains */
589 int ndoms = 0; /* number of sched domains in result */
590 int nslot; /* next empty doms[] struct cpumask slot */
591 struct cgroup_subsys_state *pos_css;
597 /* Special case for the 99% of systems with one, full, sched domain */
598 if (is_sched_load_balance(&top_cpuset)) {
600 doms = alloc_sched_domains(ndoms);
604 dattr = kmalloc(sizeof(struct sched_domain_attr), GFP_KERNEL);
606 *dattr = SD_ATTR_INIT;
607 update_domain_attr_tree(dattr, &top_cpuset);
609 cpumask_copy(doms[0], top_cpuset.cpus_allowed);
614 csa = kmalloc(number_of_cpusets * sizeof(cp), GFP_KERNEL);
620 cpuset_for_each_descendant_pre(cp, pos_css, &top_cpuset) {
621 if (cp == &top_cpuset)
624 * Continue traversing beyond @cp iff @cp has some CPUs and
625 * isn't load balancing. The former is obvious. The
626 * latter: All child cpusets contain a subset of the
627 * parent's cpus, so just skip them, and then we call
628 * update_domain_attr_tree() to calc relax_domain_level of
629 * the corresponding sched domain.
631 if (!cpumask_empty(cp->cpus_allowed) &&
632 !is_sched_load_balance(cp))
635 if (is_sched_load_balance(cp))
638 /* skip @cp's subtree */
639 pos_css = css_rightmost_descendant(pos_css);
643 for (i = 0; i < csn; i++)
648 /* Find the best partition (set of sched domains) */
649 for (i = 0; i < csn; i++) {
650 struct cpuset *a = csa[i];
653 for (j = 0; j < csn; j++) {
654 struct cpuset *b = csa[j];
657 if (apn != bpn && cpusets_overlap(a, b)) {
658 for (k = 0; k < csn; k++) {
659 struct cpuset *c = csa[k];
664 ndoms--; /* one less element */
671 * Now we know how many domains to create.
672 * Convert <csn, csa> to <ndoms, doms> and populate cpu masks.
674 doms = alloc_sched_domains(ndoms);
679 * The rest of the code, including the scheduler, can deal with
680 * dattr==NULL case. No need to abort if alloc fails.
682 dattr = kmalloc(ndoms * sizeof(struct sched_domain_attr), GFP_KERNEL);
684 for (nslot = 0, i = 0; i < csn; i++) {
685 struct cpuset *a = csa[i];
690 /* Skip completed partitions */
696 if (nslot == ndoms) {
697 static int warnings = 10;
700 "rebuild_sched_domains confused:"
701 " nslot %d, ndoms %d, csn %d, i %d,"
703 nslot, ndoms, csn, i, apn);
711 *(dattr + nslot) = SD_ATTR_INIT;
712 for (j = i; j < csn; j++) {
713 struct cpuset *b = csa[j];
716 cpumask_or(dp, dp, b->cpus_allowed);
718 update_domain_attr_tree(dattr + nslot, b);
720 /* Done with this partition */
726 BUG_ON(nslot != ndoms);
732 * Fallback to the default domain if kmalloc() failed.
733 * See comments in partition_sched_domains().
744 * Rebuild scheduler domains.
746 * If the flag 'sched_load_balance' of any cpuset with non-empty
747 * 'cpus' changes, or if the 'cpus' allowed changes in any cpuset
748 * which has that flag enabled, or if any cpuset with a non-empty
749 * 'cpus' is removed, then call this routine to rebuild the
750 * scheduler's dynamic sched domains.
752 * Call with cpuset_mutex held. Takes get_online_cpus().
754 static void rebuild_sched_domains_locked(void)
756 struct sched_domain_attr *attr;
760 lockdep_assert_held(&cpuset_mutex);
764 * We have raced with CPU hotplug. Don't do anything to avoid
765 * passing doms with offlined cpu to partition_sched_domains().
766 * Anyways, hotplug work item will rebuild sched domains.
768 if (!cpumask_equal(top_cpuset.cpus_allowed, cpu_active_mask))
771 /* Generate domain masks and attrs */
772 ndoms = generate_sched_domains(&doms, &attr);
774 /* Have scheduler rebuild the domains */
775 partition_sched_domains(ndoms, doms, attr);
779 #else /* !CONFIG_SMP */
780 static void rebuild_sched_domains_locked(void)
783 #endif /* CONFIG_SMP */
785 void rebuild_sched_domains(void)
787 mutex_lock(&cpuset_mutex);
788 rebuild_sched_domains_locked();
789 mutex_unlock(&cpuset_mutex);
793 * effective_cpumask_cpuset - return nearest ancestor with non-empty cpus
794 * @cs: the cpuset in interest
796 * A cpuset's effective cpumask is the cpumask of the nearest ancestor
797 * with non-empty cpus. We use effective cpumask whenever:
798 * - we update tasks' cpus_allowed. (they take on the ancestor's cpumask
799 * if the cpuset they reside in has no cpus)
800 * - we want to retrieve task_cs(tsk)'s cpus_allowed.
802 * Called with cpuset_mutex held. cpuset_cpus_allowed_fallback() is an
803 * exception. See comments there.
805 static struct cpuset *effective_cpumask_cpuset(struct cpuset *cs)
807 while (cpumask_empty(cs->cpus_allowed))
813 * effective_nodemask_cpuset - return nearest ancestor with non-empty mems
814 * @cs: the cpuset in interest
816 * A cpuset's effective nodemask is the nodemask of the nearest ancestor
817 * with non-empty memss. We use effective nodemask whenever:
818 * - we update tasks' mems_allowed. (they take on the ancestor's nodemask
819 * if the cpuset they reside in has no mems)
820 * - we want to retrieve task_cs(tsk)'s mems_allowed.
822 * Called with cpuset_mutex held.
824 static struct cpuset *effective_nodemask_cpuset(struct cpuset *cs)
826 while (nodes_empty(cs->mems_allowed))
832 * cpuset_change_cpumask - make a task's cpus_allowed the same as its cpuset's
834 * @data: cpuset to @tsk belongs to
836 * Called by css_scan_tasks() for each task in a cgroup whose cpus_allowed
837 * mask needs to be changed.
839 * We don't need to re-check for the cgroup/cpuset membership, since we're
840 * holding cpuset_mutex at this point.
842 static void cpuset_change_cpumask(struct task_struct *tsk, void *data)
844 struct cpuset *cs = data;
845 struct cpuset *cpus_cs = effective_cpumask_cpuset(cs);
847 set_cpus_allowed_ptr(tsk, cpus_cs->cpus_allowed);
851 * update_tasks_cpumask - Update the cpumasks of tasks in the cpuset.
852 * @cs: the cpuset in which each task's cpus_allowed mask needs to be changed
853 * @heap: if NULL, defer allocating heap memory to css_scan_tasks()
855 * Called with cpuset_mutex held
857 * The css_scan_tasks() function will scan all the tasks in a cgroup,
858 * calling callback functions for each.
860 * No return value. It's guaranteed that css_scan_tasks() always returns 0
863 static void update_tasks_cpumask(struct cpuset *cs, struct ptr_heap *heap)
865 css_scan_tasks(&cs->css, NULL, cpuset_change_cpumask, cs, heap);
869 * update_tasks_cpumask_hier - Update the cpumasks of tasks in the hierarchy.
870 * @root_cs: the root cpuset of the hierarchy
871 * @update_root: update root cpuset or not?
872 * @heap: the heap used by css_scan_tasks()
874 * This will update cpumasks of tasks in @root_cs and all other empty cpusets
875 * which take on cpumask of @root_cs.
877 * Called with cpuset_mutex held
879 static void update_tasks_cpumask_hier(struct cpuset *root_cs,
880 bool update_root, struct ptr_heap *heap)
883 struct cgroup_subsys_state *pos_css;
886 cpuset_for_each_descendant_pre(cp, pos_css, root_cs) {
891 /* skip the whole subtree if @cp have some CPU */
892 if (!cpumask_empty(cp->cpus_allowed)) {
893 pos_css = css_rightmost_descendant(pos_css);
897 if (!css_tryget(&cp->css))
901 update_tasks_cpumask(cp, heap);
910 * update_cpumask - update the cpus_allowed mask of a cpuset and all tasks in it
911 * @cs: the cpuset to consider
912 * @buf: buffer of cpu numbers written to this cpuset
914 static int update_cpumask(struct cpuset *cs, struct cpuset *trialcs,
917 struct ptr_heap heap;
919 int is_load_balanced;
921 /* top_cpuset.cpus_allowed tracks cpu_online_mask; it's read-only */
922 if (cs == &top_cpuset)
926 * An empty cpus_allowed is ok only if the cpuset has no tasks.
927 * Since cpulist_parse() fails on an empty mask, we special case
928 * that parsing. The validate_change() call ensures that cpusets
929 * with tasks have cpus.
932 cpumask_clear(trialcs->cpus_allowed);
934 retval = cpulist_parse(buf, trialcs->cpus_allowed);
938 if (!cpumask_subset(trialcs->cpus_allowed, cpu_active_mask))
942 /* Nothing to do if the cpus didn't change */
943 if (cpumask_equal(cs->cpus_allowed, trialcs->cpus_allowed))
946 retval = validate_change(cs, trialcs);
950 retval = heap_init(&heap, PAGE_SIZE, GFP_KERNEL, NULL);
954 is_load_balanced = is_sched_load_balance(trialcs);
956 mutex_lock(&callback_mutex);
957 cpumask_copy(cs->cpus_allowed, trialcs->cpus_allowed);
958 mutex_unlock(&callback_mutex);
960 update_tasks_cpumask_hier(cs, true, &heap);
964 if (is_load_balanced)
965 rebuild_sched_domains_locked();
972 * Migrate memory region from one set of nodes to another.
974 * Temporarilly set tasks mems_allowed to target nodes of migration,
975 * so that the migration code can allocate pages on these nodes.
977 * Call holding cpuset_mutex, so current's cpuset won't change
978 * during this call, as manage_mutex holds off any cpuset_attach()
979 * calls. Therefore we don't need to take task_lock around the
980 * call to guarantee_online_mems(), as we know no one is changing
983 * While the mm_struct we are migrating is typically from some
984 * other task, the task_struct mems_allowed that we are hacking
985 * is for our current task, which must allocate new pages for that
986 * migrating memory region.
989 static void cpuset_migrate_mm(struct mm_struct *mm, const nodemask_t *from,
990 const nodemask_t *to)
992 struct task_struct *tsk = current;
993 struct cpuset *mems_cs;
995 tsk->mems_allowed = *to;
997 do_migrate_pages(mm, from, to, MPOL_MF_MOVE_ALL);
999 mems_cs = effective_nodemask_cpuset(task_cs(tsk));
1000 guarantee_online_mems(mems_cs, &tsk->mems_allowed);
1004 * cpuset_change_task_nodemask - change task's mems_allowed and mempolicy
1005 * @tsk: the task to change
1006 * @newmems: new nodes that the task will be set
1008 * In order to avoid seeing no nodes if the old and new nodes are disjoint,
1009 * we structure updates as setting all new allowed nodes, then clearing newly
1012 static void cpuset_change_task_nodemask(struct task_struct *tsk,
1013 nodemask_t *newmems)
1018 * Allow tasks that have access to memory reserves because they have
1019 * been OOM killed to get memory anywhere.
1021 if (unlikely(test_thread_flag(TIF_MEMDIE)))
1023 if (current->flags & PF_EXITING) /* Let dying task have memory */
1028 * Determine if a loop is necessary if another thread is doing
1029 * get_mems_allowed(). If at least one node remains unchanged and
1030 * tsk does not have a mempolicy, then an empty nodemask will not be
1031 * possible when mems_allowed is larger than a word.
1033 need_loop = task_has_mempolicy(tsk) ||
1034 !nodes_intersects(*newmems, tsk->mems_allowed);
1037 local_irq_disable();
1038 write_seqcount_begin(&tsk->mems_allowed_seq);
1041 nodes_or(tsk->mems_allowed, tsk->mems_allowed, *newmems);
1042 mpol_rebind_task(tsk, newmems, MPOL_REBIND_STEP1);
1044 mpol_rebind_task(tsk, newmems, MPOL_REBIND_STEP2);
1045 tsk->mems_allowed = *newmems;
1048 write_seqcount_end(&tsk->mems_allowed_seq);
1055 struct cpuset_change_nodemask_arg {
1057 nodemask_t *newmems;
1061 * Update task's mems_allowed and rebind its mempolicy and vmas' mempolicy
1062 * of it to cpuset's new mems_allowed, and migrate pages to new nodes if
1063 * memory_migrate flag is set. Called with cpuset_mutex held.
1065 static void cpuset_change_nodemask(struct task_struct *p, void *data)
1067 struct cpuset_change_nodemask_arg *arg = data;
1068 struct cpuset *cs = arg->cs;
1069 struct mm_struct *mm;
1072 cpuset_change_task_nodemask(p, arg->newmems);
1074 mm = get_task_mm(p);
1078 migrate = is_memory_migrate(cs);
1080 mpol_rebind_mm(mm, &cs->mems_allowed);
1082 cpuset_migrate_mm(mm, &cs->old_mems_allowed, arg->newmems);
1086 static void *cpuset_being_rebound;
1089 * update_tasks_nodemask - Update the nodemasks of tasks in the cpuset.
1090 * @cs: the cpuset in which each task's mems_allowed mask needs to be changed
1091 * @heap: if NULL, defer allocating heap memory to css_scan_tasks()
1093 * Called with cpuset_mutex held. No return value. It's guaranteed that
1094 * css_scan_tasks() always returns 0 if @heap != NULL.
1096 static void update_tasks_nodemask(struct cpuset *cs, struct ptr_heap *heap)
1098 static nodemask_t newmems; /* protected by cpuset_mutex */
1099 struct cpuset *mems_cs = effective_nodemask_cpuset(cs);
1100 struct cpuset_change_nodemask_arg arg = { .cs = cs,
1101 .newmems = &newmems };
1103 cpuset_being_rebound = cs; /* causes mpol_dup() rebind */
1105 guarantee_online_mems(mems_cs, &newmems);
1108 * The mpol_rebind_mm() call takes mmap_sem, which we couldn't
1109 * take while holding tasklist_lock. Forks can happen - the
1110 * mpol_dup() cpuset_being_rebound check will catch such forks,
1111 * and rebind their vma mempolicies too. Because we still hold
1112 * the global cpuset_mutex, we know that no other rebind effort
1113 * will be contending for the global variable cpuset_being_rebound.
1114 * It's ok if we rebind the same mm twice; mpol_rebind_mm()
1115 * is idempotent. Also migrate pages in each mm to new nodes.
1117 css_scan_tasks(&cs->css, NULL, cpuset_change_nodemask, &arg, heap);
1120 * All the tasks' nodemasks have been updated, update
1121 * cs->old_mems_allowed.
1123 cs->old_mems_allowed = newmems;
1125 /* We're done rebinding vmas to this cpuset's new mems_allowed. */
1126 cpuset_being_rebound = NULL;
1130 * update_tasks_nodemask_hier - Update the nodemasks of tasks in the hierarchy.
1131 * @cs: the root cpuset of the hierarchy
1132 * @update_root: update the root cpuset or not?
1133 * @heap: the heap used by css_scan_tasks()
1135 * This will update nodemasks of tasks in @root_cs and all other empty cpusets
1136 * which take on nodemask of @root_cs.
1138 * Called with cpuset_mutex held
1140 static void update_tasks_nodemask_hier(struct cpuset *root_cs,
1141 bool update_root, struct ptr_heap *heap)
1144 struct cgroup_subsys_state *pos_css;
1147 cpuset_for_each_descendant_pre(cp, pos_css, root_cs) {
1148 if (cp == root_cs) {
1152 /* skip the whole subtree if @cp have some CPU */
1153 if (!nodes_empty(cp->mems_allowed)) {
1154 pos_css = css_rightmost_descendant(pos_css);
1158 if (!css_tryget(&cp->css))
1162 update_tasks_nodemask(cp, heap);
1171 * Handle user request to change the 'mems' memory placement
1172 * of a cpuset. Needs to validate the request, update the
1173 * cpusets mems_allowed, and for each task in the cpuset,
1174 * update mems_allowed and rebind task's mempolicy and any vma
1175 * mempolicies and if the cpuset is marked 'memory_migrate',
1176 * migrate the tasks pages to the new memory.
1178 * Call with cpuset_mutex held. May take callback_mutex during call.
1179 * Will take tasklist_lock, scan tasklist for tasks in cpuset cs,
1180 * lock each such tasks mm->mmap_sem, scan its vma's and rebind
1181 * their mempolicies to the cpusets new mems_allowed.
1183 static int update_nodemask(struct cpuset *cs, struct cpuset *trialcs,
1187 struct ptr_heap heap;
1190 * top_cpuset.mems_allowed tracks node_stats[N_MEMORY];
1193 if (cs == &top_cpuset) {
1199 * An empty mems_allowed is ok iff there are no tasks in the cpuset.
1200 * Since nodelist_parse() fails on an empty mask, we special case
1201 * that parsing. The validate_change() call ensures that cpusets
1202 * with tasks have memory.
1205 nodes_clear(trialcs->mems_allowed);
1207 retval = nodelist_parse(buf, trialcs->mems_allowed);
1211 if (!nodes_subset(trialcs->mems_allowed,
1212 node_states[N_MEMORY])) {
1218 if (nodes_equal(cs->mems_allowed, trialcs->mems_allowed)) {
1219 retval = 0; /* Too easy - nothing to do */
1222 retval = validate_change(cs, trialcs);
1226 retval = heap_init(&heap, PAGE_SIZE, GFP_KERNEL, NULL);
1230 mutex_lock(&callback_mutex);
1231 cs->mems_allowed = trialcs->mems_allowed;
1232 mutex_unlock(&callback_mutex);
1234 update_tasks_nodemask_hier(cs, true, &heap);
1241 int current_cpuset_is_being_rebound(void)
1243 return task_cs(current) == cpuset_being_rebound;
1246 static int update_relax_domain_level(struct cpuset *cs, s64 val)
1249 if (val < -1 || val >= sched_domain_level_max)
1253 if (val != cs->relax_domain_level) {
1254 cs->relax_domain_level = val;
1255 if (!cpumask_empty(cs->cpus_allowed) &&
1256 is_sched_load_balance(cs))
1257 rebuild_sched_domains_locked();
1264 * cpuset_change_flag - make a task's spread flags the same as its cpuset's
1265 * @tsk: task to be updated
1266 * @data: cpuset to @tsk belongs to
1268 * Called by css_scan_tasks() for each task in a cgroup.
1270 * We don't need to re-check for the cgroup/cpuset membership, since we're
1271 * holding cpuset_mutex at this point.
1273 static void cpuset_change_flag(struct task_struct *tsk, void *data)
1275 struct cpuset *cs = data;
1277 cpuset_update_task_spread_flag(cs, tsk);
1281 * update_tasks_flags - update the spread flags of tasks in the cpuset.
1282 * @cs: the cpuset in which each task's spread flags needs to be changed
1283 * @heap: if NULL, defer allocating heap memory to css_scan_tasks()
1285 * Called with cpuset_mutex held
1287 * The css_scan_tasks() function will scan all the tasks in a cgroup,
1288 * calling callback functions for each.
1290 * No return value. It's guaranteed that css_scan_tasks() always returns 0
1293 static void update_tasks_flags(struct cpuset *cs, struct ptr_heap *heap)
1295 css_scan_tasks(&cs->css, NULL, cpuset_change_flag, cs, heap);
1299 * update_flag - read a 0 or a 1 in a file and update associated flag
1300 * bit: the bit to update (see cpuset_flagbits_t)
1301 * cs: the cpuset to update
1302 * turning_on: whether the flag is being set or cleared
1304 * Call with cpuset_mutex held.
1307 static int update_flag(cpuset_flagbits_t bit, struct cpuset *cs,
1310 struct cpuset *trialcs;
1311 int balance_flag_changed;
1312 int spread_flag_changed;
1313 struct ptr_heap heap;
1316 trialcs = alloc_trial_cpuset(cs);
1321 set_bit(bit, &trialcs->flags);
1323 clear_bit(bit, &trialcs->flags);
1325 err = validate_change(cs, trialcs);
1329 err = heap_init(&heap, PAGE_SIZE, GFP_KERNEL, NULL);
1333 balance_flag_changed = (is_sched_load_balance(cs) !=
1334 is_sched_load_balance(trialcs));
1336 spread_flag_changed = ((is_spread_slab(cs) != is_spread_slab(trialcs))
1337 || (is_spread_page(cs) != is_spread_page(trialcs)));
1339 mutex_lock(&callback_mutex);
1340 cs->flags = trialcs->flags;
1341 mutex_unlock(&callback_mutex);
1343 if (!cpumask_empty(trialcs->cpus_allowed) && balance_flag_changed)
1344 rebuild_sched_domains_locked();
1346 if (spread_flag_changed)
1347 update_tasks_flags(cs, &heap);
1350 free_trial_cpuset(trialcs);
1355 * Frequency meter - How fast is some event occurring?
1357 * These routines manage a digitally filtered, constant time based,
1358 * event frequency meter. There are four routines:
1359 * fmeter_init() - initialize a frequency meter.
1360 * fmeter_markevent() - called each time the event happens.
1361 * fmeter_getrate() - returns the recent rate of such events.
1362 * fmeter_update() - internal routine used to update fmeter.
1364 * A common data structure is passed to each of these routines,
1365 * which is used to keep track of the state required to manage the
1366 * frequency meter and its digital filter.
1368 * The filter works on the number of events marked per unit time.
1369 * The filter is single-pole low-pass recursive (IIR). The time unit
1370 * is 1 second. Arithmetic is done using 32-bit integers scaled to
1371 * simulate 3 decimal digits of precision (multiplied by 1000).
1373 * With an FM_COEF of 933, and a time base of 1 second, the filter
1374 * has a half-life of 10 seconds, meaning that if the events quit
1375 * happening, then the rate returned from the fmeter_getrate()
1376 * will be cut in half each 10 seconds, until it converges to zero.
1378 * It is not worth doing a real infinitely recursive filter. If more
1379 * than FM_MAXTICKS ticks have elapsed since the last filter event,
1380 * just compute FM_MAXTICKS ticks worth, by which point the level
1383 * Limit the count of unprocessed events to FM_MAXCNT, so as to avoid
1384 * arithmetic overflow in the fmeter_update() routine.
1386 * Given the simple 32 bit integer arithmetic used, this meter works
1387 * best for reporting rates between one per millisecond (msec) and
1388 * one per 32 (approx) seconds. At constant rates faster than one
1389 * per msec it maxes out at values just under 1,000,000. At constant
1390 * rates between one per msec, and one per second it will stabilize
1391 * to a value N*1000, where N is the rate of events per second.
1392 * At constant rates between one per second and one per 32 seconds,
1393 * it will be choppy, moving up on the seconds that have an event,
1394 * and then decaying until the next event. At rates slower than
1395 * about one in 32 seconds, it decays all the way back to zero between
1399 #define FM_COEF 933 /* coefficient for half-life of 10 secs */
1400 #define FM_MAXTICKS ((time_t)99) /* useless computing more ticks than this */
1401 #define FM_MAXCNT 1000000 /* limit cnt to avoid overflow */
1402 #define FM_SCALE 1000 /* faux fixed point scale */
1404 /* Initialize a frequency meter */
1405 static void fmeter_init(struct fmeter *fmp)
1410 spin_lock_init(&fmp->lock);
1413 /* Internal meter update - process cnt events and update value */
1414 static void fmeter_update(struct fmeter *fmp)
1416 time_t now = get_seconds();
1417 time_t ticks = now - fmp->time;
1422 ticks = min(FM_MAXTICKS, ticks);
1424 fmp->val = (FM_COEF * fmp->val) / FM_SCALE;
1427 fmp->val += ((FM_SCALE - FM_COEF) * fmp->cnt) / FM_SCALE;
1431 /* Process any previous ticks, then bump cnt by one (times scale). */
1432 static void fmeter_markevent(struct fmeter *fmp)
1434 spin_lock(&fmp->lock);
1436 fmp->cnt = min(FM_MAXCNT, fmp->cnt + FM_SCALE);
1437 spin_unlock(&fmp->lock);
1440 /* Process any previous ticks, then return current value. */
1441 static int fmeter_getrate(struct fmeter *fmp)
1445 spin_lock(&fmp->lock);
1448 spin_unlock(&fmp->lock);
1452 /* Called by cgroups to determine if a cpuset is usable; cpuset_mutex held */
1453 static int cpuset_can_attach(struct cgroup_subsys_state *css,
1454 struct cgroup_taskset *tset)
1456 struct cpuset *cs = css_cs(css);
1457 struct task_struct *task;
1460 mutex_lock(&cpuset_mutex);
1463 * We allow to move tasks into an empty cpuset if sane_behavior
1467 if (!cgroup_sane_behavior(css->cgroup) &&
1468 (cpumask_empty(cs->cpus_allowed) || nodes_empty(cs->mems_allowed)))
1471 cgroup_taskset_for_each(task, css, tset) {
1473 * Kthreads which disallow setaffinity shouldn't be moved
1474 * to a new cpuset; we don't want to change their cpu
1475 * affinity and isolating such threads by their set of
1476 * allowed nodes is unnecessary. Thus, cpusets are not
1477 * applicable for such threads. This prevents checking for
1478 * success of set_cpus_allowed_ptr() on all attached tasks
1479 * before cpus_allowed may be changed.
1482 if (task->flags & PF_NO_SETAFFINITY)
1484 ret = security_task_setscheduler(task);
1490 * Mark attach is in progress. This makes validate_change() fail
1491 * changes which zero cpus/mems_allowed.
1493 cs->attach_in_progress++;
1496 mutex_unlock(&cpuset_mutex);
1500 static void cpuset_cancel_attach(struct cgroup_subsys_state *css,
1501 struct cgroup_taskset *tset)
1503 mutex_lock(&cpuset_mutex);
1504 css_cs(css)->attach_in_progress--;
1505 mutex_unlock(&cpuset_mutex);
1509 * Protected by cpuset_mutex. cpus_attach is used only by cpuset_attach()
1510 * but we can't allocate it dynamically there. Define it global and
1511 * allocate from cpuset_init().
1513 static cpumask_var_t cpus_attach;
1515 static void cpuset_attach(struct cgroup_subsys_state *css,
1516 struct cgroup_taskset *tset)
1518 /* static buf protected by cpuset_mutex */
1519 static nodemask_t cpuset_attach_nodemask_to;
1520 struct mm_struct *mm;
1521 struct task_struct *task;
1522 struct task_struct *leader = cgroup_taskset_first(tset);
1523 struct cgroup_subsys_state *oldcss = cgroup_taskset_cur_css(tset,
1525 struct cpuset *cs = css_cs(css);
1526 struct cpuset *oldcs = css_cs(oldcss);
1527 struct cpuset *cpus_cs = effective_cpumask_cpuset(cs);
1528 struct cpuset *mems_cs = effective_nodemask_cpuset(cs);
1530 mutex_lock(&cpuset_mutex);
1532 /* prepare for attach */
1533 if (cs == &top_cpuset)
1534 cpumask_copy(cpus_attach, cpu_possible_mask);
1536 guarantee_online_cpus(cpus_cs, cpus_attach);
1538 guarantee_online_mems(mems_cs, &cpuset_attach_nodemask_to);
1540 cgroup_taskset_for_each(task, css, tset) {
1542 * can_attach beforehand should guarantee that this doesn't
1543 * fail. TODO: have a better way to handle failure here
1545 WARN_ON_ONCE(set_cpus_allowed_ptr(task, cpus_attach));
1547 cpuset_change_task_nodemask(task, &cpuset_attach_nodemask_to);
1548 cpuset_update_task_spread_flag(cs, task);
1552 * Change mm, possibly for multiple threads in a threadgroup. This is
1553 * expensive and may sleep.
1555 cpuset_attach_nodemask_to = cs->mems_allowed;
1556 mm = get_task_mm(leader);
1558 struct cpuset *mems_oldcs = effective_nodemask_cpuset(oldcs);
1560 mpol_rebind_mm(mm, &cpuset_attach_nodemask_to);
1563 * old_mems_allowed is the same with mems_allowed here, except
1564 * if this task is being moved automatically due to hotplug.
1565 * In that case @mems_allowed has been updated and is empty,
1566 * so @old_mems_allowed is the right nodesets that we migrate
1569 if (is_memory_migrate(cs)) {
1570 cpuset_migrate_mm(mm, &mems_oldcs->old_mems_allowed,
1571 &cpuset_attach_nodemask_to);
1576 cs->old_mems_allowed = cpuset_attach_nodemask_to;
1578 cs->attach_in_progress--;
1579 if (!cs->attach_in_progress)
1580 wake_up(&cpuset_attach_wq);
1582 mutex_unlock(&cpuset_mutex);
1585 /* The various types of files and directories in a cpuset file system */
1588 FILE_MEMORY_MIGRATE,
1594 FILE_SCHED_LOAD_BALANCE,
1595 FILE_SCHED_RELAX_DOMAIN_LEVEL,
1596 FILE_MEMORY_PRESSURE_ENABLED,
1597 FILE_MEMORY_PRESSURE,
1600 } cpuset_filetype_t;
1602 static int cpuset_write_u64(struct cgroup_subsys_state *css, struct cftype *cft,
1605 struct cpuset *cs = css_cs(css);
1606 cpuset_filetype_t type = cft->private;
1609 mutex_lock(&cpuset_mutex);
1610 if (!is_cpuset_online(cs)) {
1616 case FILE_CPU_EXCLUSIVE:
1617 retval = update_flag(CS_CPU_EXCLUSIVE, cs, val);
1619 case FILE_MEM_EXCLUSIVE:
1620 retval = update_flag(CS_MEM_EXCLUSIVE, cs, val);
1622 case FILE_MEM_HARDWALL:
1623 retval = update_flag(CS_MEM_HARDWALL, cs, val);
1625 case FILE_SCHED_LOAD_BALANCE:
1626 retval = update_flag(CS_SCHED_LOAD_BALANCE, cs, val);
1628 case FILE_MEMORY_MIGRATE:
1629 retval = update_flag(CS_MEMORY_MIGRATE, cs, val);
1631 case FILE_MEMORY_PRESSURE_ENABLED:
1632 cpuset_memory_pressure_enabled = !!val;
1634 case FILE_MEMORY_PRESSURE:
1637 case FILE_SPREAD_PAGE:
1638 retval = update_flag(CS_SPREAD_PAGE, cs, val);
1640 case FILE_SPREAD_SLAB:
1641 retval = update_flag(CS_SPREAD_SLAB, cs, val);
1648 mutex_unlock(&cpuset_mutex);
1652 static int cpuset_write_s64(struct cgroup_subsys_state *css, struct cftype *cft,
1655 struct cpuset *cs = css_cs(css);
1656 cpuset_filetype_t type = cft->private;
1657 int retval = -ENODEV;
1659 mutex_lock(&cpuset_mutex);
1660 if (!is_cpuset_online(cs))
1664 case FILE_SCHED_RELAX_DOMAIN_LEVEL:
1665 retval = update_relax_domain_level(cs, val);
1672 mutex_unlock(&cpuset_mutex);
1677 * Common handling for a write to a "cpus" or "mems" file.
1679 static int cpuset_write_resmask(struct cgroup_subsys_state *css,
1680 struct cftype *cft, const char *buf)
1682 struct cpuset *cs = css_cs(css);
1683 struct cpuset *trialcs;
1684 int retval = -ENODEV;
1687 * CPU or memory hotunplug may leave @cs w/o any execution
1688 * resources, in which case the hotplug code asynchronously updates
1689 * configuration and transfers all tasks to the nearest ancestor
1690 * which can execute.
1692 * As writes to "cpus" or "mems" may restore @cs's execution
1693 * resources, wait for the previously scheduled operations before
1694 * proceeding, so that we don't end up keep removing tasks added
1695 * after execution capability is restored.
1697 flush_work(&cpuset_hotplug_work);
1699 mutex_lock(&cpuset_mutex);
1700 if (!is_cpuset_online(cs))
1703 trialcs = alloc_trial_cpuset(cs);
1709 switch (cft->private) {
1711 retval = update_cpumask(cs, trialcs, buf);
1714 retval = update_nodemask(cs, trialcs, buf);
1721 free_trial_cpuset(trialcs);
1723 mutex_unlock(&cpuset_mutex);
1728 * These ascii lists should be read in a single call, by using a user
1729 * buffer large enough to hold the entire map. If read in smaller
1730 * chunks, there is no guarantee of atomicity. Since the display format
1731 * used, list of ranges of sequential numbers, is variable length,
1732 * and since these maps can change value dynamically, one could read
1733 * gibberish by doing partial reads while a list was changing.
1735 static int cpuset_common_seq_show(struct seq_file *sf, void *v)
1737 struct cpuset *cs = css_cs(seq_css(sf));
1738 cpuset_filetype_t type = seq_cft(sf)->private;
1743 count = seq_get_buf(sf, &buf);
1746 mutex_lock(&callback_mutex);
1750 s += cpulist_scnprintf(s, count, cs->cpus_allowed);
1753 s += nodelist_scnprintf(s, count, cs->mems_allowed);
1760 if (s < buf + count - 1) {
1762 seq_commit(sf, s - buf);
1767 mutex_unlock(&callback_mutex);
1771 static u64 cpuset_read_u64(struct cgroup_subsys_state *css, struct cftype *cft)
1773 struct cpuset *cs = css_cs(css);
1774 cpuset_filetype_t type = cft->private;
1776 case FILE_CPU_EXCLUSIVE:
1777 return is_cpu_exclusive(cs);
1778 case FILE_MEM_EXCLUSIVE:
1779 return is_mem_exclusive(cs);
1780 case FILE_MEM_HARDWALL:
1781 return is_mem_hardwall(cs);
1782 case FILE_SCHED_LOAD_BALANCE:
1783 return is_sched_load_balance(cs);
1784 case FILE_MEMORY_MIGRATE:
1785 return is_memory_migrate(cs);
1786 case FILE_MEMORY_PRESSURE_ENABLED:
1787 return cpuset_memory_pressure_enabled;
1788 case FILE_MEMORY_PRESSURE:
1789 return fmeter_getrate(&cs->fmeter);
1790 case FILE_SPREAD_PAGE:
1791 return is_spread_page(cs);
1792 case FILE_SPREAD_SLAB:
1793 return is_spread_slab(cs);
1798 /* Unreachable but makes gcc happy */
1802 static s64 cpuset_read_s64(struct cgroup_subsys_state *css, struct cftype *cft)
1804 struct cpuset *cs = css_cs(css);
1805 cpuset_filetype_t type = cft->private;
1807 case FILE_SCHED_RELAX_DOMAIN_LEVEL:
1808 return cs->relax_domain_level;
1813 /* Unrechable but makes gcc happy */
1819 * for the common functions, 'private' gives the type of file
1822 static struct cftype files[] = {
1825 .seq_show = cpuset_common_seq_show,
1826 .write_string = cpuset_write_resmask,
1827 .max_write_len = (100U + 6 * NR_CPUS),
1828 .private = FILE_CPULIST,
1833 .seq_show = cpuset_common_seq_show,
1834 .write_string = cpuset_write_resmask,
1835 .max_write_len = (100U + 6 * MAX_NUMNODES),
1836 .private = FILE_MEMLIST,
1840 .name = "cpu_exclusive",
1841 .read_u64 = cpuset_read_u64,
1842 .write_u64 = cpuset_write_u64,
1843 .private = FILE_CPU_EXCLUSIVE,
1847 .name = "mem_exclusive",
1848 .read_u64 = cpuset_read_u64,
1849 .write_u64 = cpuset_write_u64,
1850 .private = FILE_MEM_EXCLUSIVE,
1854 .name = "mem_hardwall",
1855 .read_u64 = cpuset_read_u64,
1856 .write_u64 = cpuset_write_u64,
1857 .private = FILE_MEM_HARDWALL,
1861 .name = "sched_load_balance",
1862 .read_u64 = cpuset_read_u64,
1863 .write_u64 = cpuset_write_u64,
1864 .private = FILE_SCHED_LOAD_BALANCE,
1868 .name = "sched_relax_domain_level",
1869 .read_s64 = cpuset_read_s64,
1870 .write_s64 = cpuset_write_s64,
1871 .private = FILE_SCHED_RELAX_DOMAIN_LEVEL,
1875 .name = "memory_migrate",
1876 .read_u64 = cpuset_read_u64,
1877 .write_u64 = cpuset_write_u64,
1878 .private = FILE_MEMORY_MIGRATE,
1882 .name = "memory_pressure",
1883 .read_u64 = cpuset_read_u64,
1884 .write_u64 = cpuset_write_u64,
1885 .private = FILE_MEMORY_PRESSURE,
1890 .name = "memory_spread_page",
1891 .read_u64 = cpuset_read_u64,
1892 .write_u64 = cpuset_write_u64,
1893 .private = FILE_SPREAD_PAGE,
1897 .name = "memory_spread_slab",
1898 .read_u64 = cpuset_read_u64,
1899 .write_u64 = cpuset_write_u64,
1900 .private = FILE_SPREAD_SLAB,
1904 .name = "memory_pressure_enabled",
1905 .flags = CFTYPE_ONLY_ON_ROOT,
1906 .read_u64 = cpuset_read_u64,
1907 .write_u64 = cpuset_write_u64,
1908 .private = FILE_MEMORY_PRESSURE_ENABLED,
1915 * cpuset_css_alloc - allocate a cpuset css
1916 * cgrp: control group that the new cpuset will be part of
1919 static struct cgroup_subsys_state *
1920 cpuset_css_alloc(struct cgroup_subsys_state *parent_css)
1925 return &top_cpuset.css;
1927 cs = kzalloc(sizeof(*cs), GFP_KERNEL);
1929 return ERR_PTR(-ENOMEM);
1930 if (!alloc_cpumask_var(&cs->cpus_allowed, GFP_KERNEL)) {
1932 return ERR_PTR(-ENOMEM);
1935 set_bit(CS_SCHED_LOAD_BALANCE, &cs->flags);
1936 cpumask_clear(cs->cpus_allowed);
1937 nodes_clear(cs->mems_allowed);
1938 fmeter_init(&cs->fmeter);
1939 cs->relax_domain_level = -1;
1944 static int cpuset_css_online(struct cgroup_subsys_state *css)
1946 struct cpuset *cs = css_cs(css);
1947 struct cpuset *parent = parent_cs(cs);
1948 struct cpuset *tmp_cs;
1949 struct cgroup_subsys_state *pos_css;
1954 mutex_lock(&cpuset_mutex);
1956 set_bit(CS_ONLINE, &cs->flags);
1957 if (is_spread_page(parent))
1958 set_bit(CS_SPREAD_PAGE, &cs->flags);
1959 if (is_spread_slab(parent))
1960 set_bit(CS_SPREAD_SLAB, &cs->flags);
1962 number_of_cpusets++;
1964 if (!test_bit(CGRP_CPUSET_CLONE_CHILDREN, &css->cgroup->flags))
1968 * Clone @parent's configuration if CGRP_CPUSET_CLONE_CHILDREN is
1969 * set. This flag handling is implemented in cgroup core for
1970 * histrical reasons - the flag may be specified during mount.
1972 * Currently, if any sibling cpusets have exclusive cpus or mem, we
1973 * refuse to clone the configuration - thereby refusing the task to
1974 * be entered, and as a result refusing the sys_unshare() or
1975 * clone() which initiated it. If this becomes a problem for some
1976 * users who wish to allow that scenario, then this could be
1977 * changed to grant parent->cpus_allowed-sibling_cpus_exclusive
1978 * (and likewise for mems) to the new cgroup.
1981 cpuset_for_each_child(tmp_cs, pos_css, parent) {
1982 if (is_mem_exclusive(tmp_cs) || is_cpu_exclusive(tmp_cs)) {
1989 mutex_lock(&callback_mutex);
1990 cs->mems_allowed = parent->mems_allowed;
1991 cpumask_copy(cs->cpus_allowed, parent->cpus_allowed);
1992 mutex_unlock(&callback_mutex);
1994 mutex_unlock(&cpuset_mutex);
1999 * If the cpuset being removed has its flag 'sched_load_balance'
2000 * enabled, then simulate turning sched_load_balance off, which
2001 * will call rebuild_sched_domains_locked().
2004 static void cpuset_css_offline(struct cgroup_subsys_state *css)
2006 struct cpuset *cs = css_cs(css);
2008 mutex_lock(&cpuset_mutex);
2010 if (is_sched_load_balance(cs))
2011 update_flag(CS_SCHED_LOAD_BALANCE, cs, 0);
2013 number_of_cpusets--;
2014 clear_bit(CS_ONLINE, &cs->flags);
2016 mutex_unlock(&cpuset_mutex);
2019 static void cpuset_css_free(struct cgroup_subsys_state *css)
2021 struct cpuset *cs = css_cs(css);
2023 free_cpumask_var(cs->cpus_allowed);
2027 struct cgroup_subsys cpuset_cgrp_subsys = {
2028 .css_alloc = cpuset_css_alloc,
2029 .css_online = cpuset_css_online,
2030 .css_offline = cpuset_css_offline,
2031 .css_free = cpuset_css_free,
2032 .can_attach = cpuset_can_attach,
2033 .cancel_attach = cpuset_cancel_attach,
2034 .attach = cpuset_attach,
2035 .base_cftypes = files,
2040 * cpuset_init - initialize cpusets at system boot
2042 * Description: Initialize top_cpuset and the cpuset internal file system,
2045 int __init cpuset_init(void)
2049 if (!alloc_cpumask_var(&top_cpuset.cpus_allowed, GFP_KERNEL))
2052 cpumask_setall(top_cpuset.cpus_allowed);
2053 nodes_setall(top_cpuset.mems_allowed);
2055 fmeter_init(&top_cpuset.fmeter);
2056 set_bit(CS_SCHED_LOAD_BALANCE, &top_cpuset.flags);
2057 top_cpuset.relax_domain_level = -1;
2059 err = register_filesystem(&cpuset_fs_type);
2063 if (!alloc_cpumask_var(&cpus_attach, GFP_KERNEL))
2066 number_of_cpusets = 1;
2071 * If CPU and/or memory hotplug handlers, below, unplug any CPUs
2072 * or memory nodes, we need to walk over the cpuset hierarchy,
2073 * removing that CPU or node from all cpusets. If this removes the
2074 * last CPU or node from a cpuset, then move the tasks in the empty
2075 * cpuset to its next-highest non-empty parent.
2077 static void remove_tasks_in_empty_cpuset(struct cpuset *cs)
2079 struct cpuset *parent;
2082 * Find its next-highest non-empty parent, (top cpuset
2083 * has online cpus, so can't be empty).
2085 parent = parent_cs(cs);
2086 while (cpumask_empty(parent->cpus_allowed) ||
2087 nodes_empty(parent->mems_allowed))
2088 parent = parent_cs(parent);
2090 if (cgroup_transfer_tasks(parent->css.cgroup, cs->css.cgroup)) {
2091 printk(KERN_ERR "cpuset: failed to transfer tasks out of empty cpuset ");
2092 pr_cont_cgroup_name(cs->css.cgroup);
2098 * cpuset_hotplug_update_tasks - update tasks in a cpuset for hotunplug
2099 * @cs: cpuset in interest
2101 * Compare @cs's cpu and mem masks against top_cpuset and if some have gone
2102 * offline, update @cs accordingly. If @cs ends up with no CPU or memory,
2103 * all its tasks are moved to the nearest ancestor with both resources.
2105 static void cpuset_hotplug_update_tasks(struct cpuset *cs)
2107 static cpumask_t off_cpus;
2108 static nodemask_t off_mems;
2110 bool sane = cgroup_sane_behavior(cs->css.cgroup);
2113 wait_event(cpuset_attach_wq, cs->attach_in_progress == 0);
2115 mutex_lock(&cpuset_mutex);
2118 * We have raced with task attaching. We wait until attaching
2119 * is finished, so we won't attach a task to an empty cpuset.
2121 if (cs->attach_in_progress) {
2122 mutex_unlock(&cpuset_mutex);
2126 cpumask_andnot(&off_cpus, cs->cpus_allowed, top_cpuset.cpus_allowed);
2127 nodes_andnot(off_mems, cs->mems_allowed, top_cpuset.mems_allowed);
2129 mutex_lock(&callback_mutex);
2130 cpumask_andnot(cs->cpus_allowed, cs->cpus_allowed, &off_cpus);
2131 mutex_unlock(&callback_mutex);
2134 * If sane_behavior flag is set, we need to update tasks' cpumask
2135 * for empty cpuset to take on ancestor's cpumask. Otherwise, don't
2136 * call update_tasks_cpumask() if the cpuset becomes empty, as
2137 * the tasks in it will be migrated to an ancestor.
2139 if ((sane && cpumask_empty(cs->cpus_allowed)) ||
2140 (!cpumask_empty(&off_cpus) && !cpumask_empty(cs->cpus_allowed)))
2141 update_tasks_cpumask(cs, NULL);
2143 mutex_lock(&callback_mutex);
2144 nodes_andnot(cs->mems_allowed, cs->mems_allowed, off_mems);
2145 mutex_unlock(&callback_mutex);
2148 * If sane_behavior flag is set, we need to update tasks' nodemask
2149 * for empty cpuset to take on ancestor's nodemask. Otherwise, don't
2150 * call update_tasks_nodemask() if the cpuset becomes empty, as
2151 * the tasks in it will be migratd to an ancestor.
2153 if ((sane && nodes_empty(cs->mems_allowed)) ||
2154 (!nodes_empty(off_mems) && !nodes_empty(cs->mems_allowed)))
2155 update_tasks_nodemask(cs, NULL);
2157 is_empty = cpumask_empty(cs->cpus_allowed) ||
2158 nodes_empty(cs->mems_allowed);
2160 mutex_unlock(&cpuset_mutex);
2163 * If sane_behavior flag is set, we'll keep tasks in empty cpusets.
2165 * Otherwise move tasks to the nearest ancestor with execution
2166 * resources. This is full cgroup operation which will
2167 * also call back into cpuset. Should be done outside any lock.
2169 if (!sane && is_empty)
2170 remove_tasks_in_empty_cpuset(cs);
2174 * cpuset_hotplug_workfn - handle CPU/memory hotunplug for a cpuset
2176 * This function is called after either CPU or memory configuration has
2177 * changed and updates cpuset accordingly. The top_cpuset is always
2178 * synchronized to cpu_active_mask and N_MEMORY, which is necessary in
2179 * order to make cpusets transparent (of no affect) on systems that are
2180 * actively using CPU hotplug but making no active use of cpusets.
2182 * Non-root cpusets are only affected by offlining. If any CPUs or memory
2183 * nodes have been taken down, cpuset_hotplug_update_tasks() is invoked on
2186 * Note that CPU offlining during suspend is ignored. We don't modify
2187 * cpusets across suspend/resume cycles at all.
2189 static void cpuset_hotplug_workfn(struct work_struct *work)
2191 static cpumask_t new_cpus;
2192 static nodemask_t new_mems;
2193 bool cpus_updated, mems_updated;
2195 mutex_lock(&cpuset_mutex);
2197 /* fetch the available cpus/mems and find out which changed how */
2198 cpumask_copy(&new_cpus, cpu_active_mask);
2199 new_mems = node_states[N_MEMORY];
2201 cpus_updated = !cpumask_equal(top_cpuset.cpus_allowed, &new_cpus);
2202 mems_updated = !nodes_equal(top_cpuset.mems_allowed, new_mems);
2204 /* synchronize cpus_allowed to cpu_active_mask */
2206 mutex_lock(&callback_mutex);
2207 cpumask_copy(top_cpuset.cpus_allowed, &new_cpus);
2208 mutex_unlock(&callback_mutex);
2209 /* we don't mess with cpumasks of tasks in top_cpuset */
2212 /* synchronize mems_allowed to N_MEMORY */
2214 mutex_lock(&callback_mutex);
2215 top_cpuset.mems_allowed = new_mems;
2216 mutex_unlock(&callback_mutex);
2217 update_tasks_nodemask(&top_cpuset, NULL);
2220 mutex_unlock(&cpuset_mutex);
2222 /* if cpus or mems changed, we need to propagate to descendants */
2223 if (cpus_updated || mems_updated) {
2225 struct cgroup_subsys_state *pos_css;
2228 cpuset_for_each_descendant_pre(cs, pos_css, &top_cpuset) {
2229 if (cs == &top_cpuset || !css_tryget(&cs->css))
2233 cpuset_hotplug_update_tasks(cs);
2241 /* rebuild sched domains if cpus_allowed has changed */
2243 rebuild_sched_domains();
2246 void cpuset_update_active_cpus(bool cpu_online)
2249 * We're inside cpu hotplug critical region which usually nests
2250 * inside cgroup synchronization. Bounce actual hotplug processing
2251 * to a work item to avoid reverse locking order.
2253 * We still need to do partition_sched_domains() synchronously;
2254 * otherwise, the scheduler will get confused and put tasks to the
2255 * dead CPU. Fall back to the default single domain.
2256 * cpuset_hotplug_workfn() will rebuild it as necessary.
2258 partition_sched_domains(1, NULL, NULL);
2259 schedule_work(&cpuset_hotplug_work);
2263 * Keep top_cpuset.mems_allowed tracking node_states[N_MEMORY].
2264 * Call this routine anytime after node_states[N_MEMORY] changes.
2265 * See cpuset_update_active_cpus() for CPU hotplug handling.
2267 static int cpuset_track_online_nodes(struct notifier_block *self,
2268 unsigned long action, void *arg)
2270 schedule_work(&cpuset_hotplug_work);
2274 static struct notifier_block cpuset_track_online_nodes_nb = {
2275 .notifier_call = cpuset_track_online_nodes,
2276 .priority = 10, /* ??! */
2280 * cpuset_init_smp - initialize cpus_allowed
2282 * Description: Finish top cpuset after cpu, node maps are initialized
2284 void __init cpuset_init_smp(void)
2286 cpumask_copy(top_cpuset.cpus_allowed, cpu_active_mask);
2287 top_cpuset.mems_allowed = node_states[N_MEMORY];
2288 top_cpuset.old_mems_allowed = top_cpuset.mems_allowed;
2290 register_hotmemory_notifier(&cpuset_track_online_nodes_nb);
2294 * cpuset_cpus_allowed - return cpus_allowed mask from a tasks cpuset.
2295 * @tsk: pointer to task_struct from which to obtain cpuset->cpus_allowed.
2296 * @pmask: pointer to struct cpumask variable to receive cpus_allowed set.
2298 * Description: Returns the cpumask_var_t cpus_allowed of the cpuset
2299 * attached to the specified @tsk. Guaranteed to return some non-empty
2300 * subset of cpu_online_mask, even if this means going outside the
2304 void cpuset_cpus_allowed(struct task_struct *tsk, struct cpumask *pmask)
2306 struct cpuset *cpus_cs;
2308 mutex_lock(&callback_mutex);
2310 cpus_cs = effective_cpumask_cpuset(task_cs(tsk));
2311 guarantee_online_cpus(cpus_cs, pmask);
2313 mutex_unlock(&callback_mutex);
2316 void cpuset_cpus_allowed_fallback(struct task_struct *tsk)
2318 struct cpuset *cpus_cs;
2321 cpus_cs = effective_cpumask_cpuset(task_cs(tsk));
2322 do_set_cpus_allowed(tsk, cpus_cs->cpus_allowed);
2326 * We own tsk->cpus_allowed, nobody can change it under us.
2328 * But we used cs && cs->cpus_allowed lockless and thus can
2329 * race with cgroup_attach_task() or update_cpumask() and get
2330 * the wrong tsk->cpus_allowed. However, both cases imply the
2331 * subsequent cpuset_change_cpumask()->set_cpus_allowed_ptr()
2332 * which takes task_rq_lock().
2334 * If we are called after it dropped the lock we must see all
2335 * changes in tsk_cs()->cpus_allowed. Otherwise we can temporary
2336 * set any mask even if it is not right from task_cs() pov,
2337 * the pending set_cpus_allowed_ptr() will fix things.
2339 * select_fallback_rq() will fix things ups and set cpu_possible_mask
2344 void cpuset_init_current_mems_allowed(void)
2346 nodes_setall(current->mems_allowed);
2350 * cpuset_mems_allowed - return mems_allowed mask from a tasks cpuset.
2351 * @tsk: pointer to task_struct from which to obtain cpuset->mems_allowed.
2353 * Description: Returns the nodemask_t mems_allowed of the cpuset
2354 * attached to the specified @tsk. Guaranteed to return some non-empty
2355 * subset of node_states[N_MEMORY], even if this means going outside the
2359 nodemask_t cpuset_mems_allowed(struct task_struct *tsk)
2361 struct cpuset *mems_cs;
2364 mutex_lock(&callback_mutex);
2366 mems_cs = effective_nodemask_cpuset(task_cs(tsk));
2367 guarantee_online_mems(mems_cs, &mask);
2369 mutex_unlock(&callback_mutex);
2375 * cpuset_nodemask_valid_mems_allowed - check nodemask vs. curremt mems_allowed
2376 * @nodemask: the nodemask to be checked
2378 * Are any of the nodes in the nodemask allowed in current->mems_allowed?
2380 int cpuset_nodemask_valid_mems_allowed(nodemask_t *nodemask)
2382 return nodes_intersects(*nodemask, current->mems_allowed);
2386 * nearest_hardwall_ancestor() - Returns the nearest mem_exclusive or
2387 * mem_hardwall ancestor to the specified cpuset. Call holding
2388 * callback_mutex. If no ancestor is mem_exclusive or mem_hardwall
2389 * (an unusual configuration), then returns the root cpuset.
2391 static struct cpuset *nearest_hardwall_ancestor(struct cpuset *cs)
2393 while (!(is_mem_exclusive(cs) || is_mem_hardwall(cs)) && parent_cs(cs))
2399 * cpuset_node_allowed_softwall - Can we allocate on a memory node?
2400 * @node: is this an allowed node?
2401 * @gfp_mask: memory allocation flags
2403 * If we're in interrupt, yes, we can always allocate. If __GFP_THISNODE is
2404 * set, yes, we can always allocate. If node is in our task's mems_allowed,
2405 * yes. If it's not a __GFP_HARDWALL request and this node is in the nearest
2406 * hardwalled cpuset ancestor to this task's cpuset, yes. If the task has been
2407 * OOM killed and has access to memory reserves as specified by the TIF_MEMDIE
2411 * If __GFP_HARDWALL is set, cpuset_node_allowed_softwall() reduces to
2412 * cpuset_node_allowed_hardwall(). Otherwise, cpuset_node_allowed_softwall()
2413 * might sleep, and might allow a node from an enclosing cpuset.
2415 * cpuset_node_allowed_hardwall() only handles the simpler case of hardwall
2416 * cpusets, and never sleeps.
2418 * The __GFP_THISNODE placement logic is really handled elsewhere,
2419 * by forcibly using a zonelist starting at a specified node, and by
2420 * (in get_page_from_freelist()) refusing to consider the zones for
2421 * any node on the zonelist except the first. By the time any such
2422 * calls get to this routine, we should just shut up and say 'yes'.
2424 * GFP_USER allocations are marked with the __GFP_HARDWALL bit,
2425 * and do not allow allocations outside the current tasks cpuset
2426 * unless the task has been OOM killed as is marked TIF_MEMDIE.
2427 * GFP_KERNEL allocations are not so marked, so can escape to the
2428 * nearest enclosing hardwalled ancestor cpuset.
2430 * Scanning up parent cpusets requires callback_mutex. The
2431 * __alloc_pages() routine only calls here with __GFP_HARDWALL bit
2432 * _not_ set if it's a GFP_KERNEL allocation, and all nodes in the
2433 * current tasks mems_allowed came up empty on the first pass over
2434 * the zonelist. So only GFP_KERNEL allocations, if all nodes in the
2435 * cpuset are short of memory, might require taking the callback_mutex
2438 * The first call here from mm/page_alloc:get_page_from_freelist()
2439 * has __GFP_HARDWALL set in gfp_mask, enforcing hardwall cpusets,
2440 * so no allocation on a node outside the cpuset is allowed (unless
2441 * in interrupt, of course).
2443 * The second pass through get_page_from_freelist() doesn't even call
2444 * here for GFP_ATOMIC calls. For those calls, the __alloc_pages()
2445 * variable 'wait' is not set, and the bit ALLOC_CPUSET is not set
2446 * in alloc_flags. That logic and the checks below have the combined
2448 * in_interrupt - any node ok (current task context irrelevant)
2449 * GFP_ATOMIC - any node ok
2450 * TIF_MEMDIE - any node ok
2451 * GFP_KERNEL - any node in enclosing hardwalled cpuset ok
2452 * GFP_USER - only nodes in current tasks mems allowed ok.
2455 * Don't call cpuset_node_allowed_softwall if you can't sleep, unless you
2456 * pass in the __GFP_HARDWALL flag set in gfp_flag, which disables
2457 * the code that might scan up ancestor cpusets and sleep.
2459 int __cpuset_node_allowed_softwall(int node, gfp_t gfp_mask)
2461 struct cpuset *cs; /* current cpuset ancestors */
2462 int allowed; /* is allocation in zone z allowed? */
2464 if (in_interrupt() || (gfp_mask & __GFP_THISNODE))
2466 might_sleep_if(!(gfp_mask & __GFP_HARDWALL));
2467 if (node_isset(node, current->mems_allowed))
2470 * Allow tasks that have access to memory reserves because they have
2471 * been OOM killed to get memory anywhere.
2473 if (unlikely(test_thread_flag(TIF_MEMDIE)))
2475 if (gfp_mask & __GFP_HARDWALL) /* If hardwall request, stop here */
2478 if (current->flags & PF_EXITING) /* Let dying task have memory */
2481 /* Not hardwall and node outside mems_allowed: scan up cpusets */
2482 mutex_lock(&callback_mutex);
2485 cs = nearest_hardwall_ancestor(task_cs(current));
2486 task_unlock(current);
2488 allowed = node_isset(node, cs->mems_allowed);
2489 mutex_unlock(&callback_mutex);
2494 * cpuset_node_allowed_hardwall - Can we allocate on a memory node?
2495 * @node: is this an allowed node?
2496 * @gfp_mask: memory allocation flags
2498 * If we're in interrupt, yes, we can always allocate. If __GFP_THISNODE is
2499 * set, yes, we can always allocate. If node is in our task's mems_allowed,
2500 * yes. If the task has been OOM killed and has access to memory reserves as
2501 * specified by the TIF_MEMDIE flag, yes.
2504 * The __GFP_THISNODE placement logic is really handled elsewhere,
2505 * by forcibly using a zonelist starting at a specified node, and by
2506 * (in get_page_from_freelist()) refusing to consider the zones for
2507 * any node on the zonelist except the first. By the time any such
2508 * calls get to this routine, we should just shut up and say 'yes'.
2510 * Unlike the cpuset_node_allowed_softwall() variant, above,
2511 * this variant requires that the node be in the current task's
2512 * mems_allowed or that we're in interrupt. It does not scan up the
2513 * cpuset hierarchy for the nearest enclosing mem_exclusive cpuset.
2516 int __cpuset_node_allowed_hardwall(int node, gfp_t gfp_mask)
2518 if (in_interrupt() || (gfp_mask & __GFP_THISNODE))
2520 if (node_isset(node, current->mems_allowed))
2523 * Allow tasks that have access to memory reserves because they have
2524 * been OOM killed to get memory anywhere.
2526 if (unlikely(test_thread_flag(TIF_MEMDIE)))
2532 * cpuset_mem_spread_node() - On which node to begin search for a file page
2533 * cpuset_slab_spread_node() - On which node to begin search for a slab page
2535 * If a task is marked PF_SPREAD_PAGE or PF_SPREAD_SLAB (as for
2536 * tasks in a cpuset with is_spread_page or is_spread_slab set),
2537 * and if the memory allocation used cpuset_mem_spread_node()
2538 * to determine on which node to start looking, as it will for
2539 * certain page cache or slab cache pages such as used for file
2540 * system buffers and inode caches, then instead of starting on the
2541 * local node to look for a free page, rather spread the starting
2542 * node around the tasks mems_allowed nodes.
2544 * We don't have to worry about the returned node being offline
2545 * because "it can't happen", and even if it did, it would be ok.
2547 * The routines calling guarantee_online_mems() are careful to
2548 * only set nodes in task->mems_allowed that are online. So it
2549 * should not be possible for the following code to return an
2550 * offline node. But if it did, that would be ok, as this routine
2551 * is not returning the node where the allocation must be, only
2552 * the node where the search should start. The zonelist passed to
2553 * __alloc_pages() will include all nodes. If the slab allocator
2554 * is passed an offline node, it will fall back to the local node.
2555 * See kmem_cache_alloc_node().
2558 static int cpuset_spread_node(int *rotor)
2562 node = next_node(*rotor, current->mems_allowed);
2563 if (node == MAX_NUMNODES)
2564 node = first_node(current->mems_allowed);
2569 int cpuset_mem_spread_node(void)
2571 if (current->cpuset_mem_spread_rotor == NUMA_NO_NODE)
2572 current->cpuset_mem_spread_rotor =
2573 node_random(¤t->mems_allowed);
2575 return cpuset_spread_node(¤t->cpuset_mem_spread_rotor);
2578 int cpuset_slab_spread_node(void)
2580 if (current->cpuset_slab_spread_rotor == NUMA_NO_NODE)
2581 current->cpuset_slab_spread_rotor =
2582 node_random(¤t->mems_allowed);
2584 return cpuset_spread_node(¤t->cpuset_slab_spread_rotor);
2587 EXPORT_SYMBOL_GPL(cpuset_mem_spread_node);
2590 * cpuset_mems_allowed_intersects - Does @tsk1's mems_allowed intersect @tsk2's?
2591 * @tsk1: pointer to task_struct of some task.
2592 * @tsk2: pointer to task_struct of some other task.
2594 * Description: Return true if @tsk1's mems_allowed intersects the
2595 * mems_allowed of @tsk2. Used by the OOM killer to determine if
2596 * one of the task's memory usage might impact the memory available
2600 int cpuset_mems_allowed_intersects(const struct task_struct *tsk1,
2601 const struct task_struct *tsk2)
2603 return nodes_intersects(tsk1->mems_allowed, tsk2->mems_allowed);
2606 #define CPUSET_NODELIST_LEN (256)
2609 * cpuset_print_task_mems_allowed - prints task's cpuset and mems_allowed
2610 * @task: pointer to task_struct of some task.
2612 * Description: Prints @task's name, cpuset name, and cached copy of its
2613 * mems_allowed to the kernel log. Must hold task_lock(task) to allow
2614 * dereferencing task_cs(task).
2616 void cpuset_print_task_mems_allowed(struct task_struct *tsk)
2618 /* Statically allocated to prevent using excess stack. */
2619 static char cpuset_nodelist[CPUSET_NODELIST_LEN];
2620 static DEFINE_SPINLOCK(cpuset_buffer_lock);
2621 struct cgroup *cgrp = task_cs(tsk)->css.cgroup;
2623 spin_lock(&cpuset_buffer_lock);
2625 nodelist_scnprintf(cpuset_nodelist, CPUSET_NODELIST_LEN,
2627 printk(KERN_INFO "%s cpuset=", tsk->comm);
2628 pr_cont_cgroup_name(cgrp);
2629 pr_cont(" mems_allowed=%s\n", cpuset_nodelist);
2631 spin_unlock(&cpuset_buffer_lock);
2635 * Collection of memory_pressure is suppressed unless
2636 * this flag is enabled by writing "1" to the special
2637 * cpuset file 'memory_pressure_enabled' in the root cpuset.
2640 int cpuset_memory_pressure_enabled __read_mostly;
2643 * cpuset_memory_pressure_bump - keep stats of per-cpuset reclaims.
2645 * Keep a running average of the rate of synchronous (direct)
2646 * page reclaim efforts initiated by tasks in each cpuset.
2648 * This represents the rate at which some task in the cpuset
2649 * ran low on memory on all nodes it was allowed to use, and
2650 * had to enter the kernels page reclaim code in an effort to
2651 * create more free memory by tossing clean pages or swapping
2652 * or writing dirty pages.
2654 * Display to user space in the per-cpuset read-only file
2655 * "memory_pressure". Value displayed is an integer
2656 * representing the recent rate of entry into the synchronous
2657 * (direct) page reclaim by any task attached to the cpuset.
2660 void __cpuset_memory_pressure_bump(void)
2663 fmeter_markevent(&task_cs(current)->fmeter);
2664 task_unlock(current);
2667 #ifdef CONFIG_PROC_PID_CPUSET
2669 * proc_cpuset_show()
2670 * - Print tasks cpuset path into seq_file.
2671 * - Used for /proc/<pid>/cpuset.
2672 * - No need to task_lock(tsk) on this tsk->cpuset reference, as it
2673 * doesn't really matter if tsk->cpuset changes after we read it,
2674 * and we take cpuset_mutex, keeping cpuset_attach() from changing it
2677 int proc_cpuset_show(struct seq_file *m, void *unused_v)
2680 struct task_struct *tsk;
2682 struct cgroup_subsys_state *css;
2686 buf = kmalloc(PATH_MAX, GFP_KERNEL);
2692 tsk = get_pid_task(pid, PIDTYPE_PID);
2696 retval = -ENAMETOOLONG;
2698 css = task_css(tsk, cpuset_cgrp_id);
2699 p = cgroup_path(css->cgroup, buf, PATH_MAX);
2707 put_task_struct(tsk);
2713 #endif /* CONFIG_PROC_PID_CPUSET */
2715 /* Display task mems_allowed in /proc/<pid>/status file. */
2716 void cpuset_task_status_allowed(struct seq_file *m, struct task_struct *task)
2718 seq_printf(m, "Mems_allowed:\t");
2719 seq_nodemask(m, &task->mems_allowed);
2720 seq_printf(m, "\n");
2721 seq_printf(m, "Mems_allowed_list:\t");
2722 seq_nodemask_list(m, &task->mems_allowed);
2723 seq_printf(m, "\n");