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>
64 * Tracks how many cpusets are currently defined in system.
65 * When there is only one cpuset (the root cpuset) we can
66 * short circuit some hooks.
68 int number_of_cpusets __read_mostly;
70 /* Forward declare cgroup structures */
71 struct cgroup_subsys cpuset_subsys;
74 /* See "Frequency meter" comments, below. */
77 int cnt; /* unprocessed events count */
78 int val; /* most recent output value */
79 time_t time; /* clock (secs) when val computed */
80 spinlock_t lock; /* guards read or write of above */
84 struct cgroup_subsys_state css;
86 unsigned long flags; /* "unsigned long" so bitops work */
87 cpumask_var_t cpus_allowed; /* CPUs allowed to tasks in cpuset */
88 nodemask_t mems_allowed; /* Memory Nodes allowed to tasks */
90 struct fmeter fmeter; /* memory_pressure filter */
93 * Tasks are being attached to this cpuset. Used to prevent
94 * zeroing cpus/mems_allowed between ->can_attach() and ->attach().
96 int attach_in_progress;
98 /* partition number for rebuild_sched_domains() */
101 /* for custom sched domain */
102 int relax_domain_level;
104 struct work_struct hotplug_work;
107 /* Retrieve the cpuset for a cgroup */
108 static inline struct cpuset *cgroup_cs(struct cgroup *cont)
110 return container_of(cgroup_subsys_state(cont, cpuset_subsys_id),
114 /* Retrieve the cpuset for a task */
115 static inline struct cpuset *task_cs(struct task_struct *task)
117 return container_of(task_subsys_state(task, cpuset_subsys_id),
121 static inline struct cpuset *parent_cs(const struct cpuset *cs)
123 struct cgroup *pcgrp = cs->css.cgroup->parent;
126 return cgroup_cs(pcgrp);
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_cgrp: 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_cgrp, parent_cs) \
211 cgroup_for_each_child((pos_cgrp), (parent_cs)->css.cgroup) \
212 if (is_cpuset_online(((child_cs) = cgroup_cs((pos_cgrp)))))
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_cgrp: 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_cgrp by calling
222 * cgroup_rightmost_descendant() to skip subtree.
224 #define cpuset_for_each_descendant_pre(des_cs, pos_cgrp, root_cs) \
225 cgroup_for_each_descendant_pre((pos_cgrp), (root_cs)->css.cgroup) \
226 if (is_cpuset_online(((des_cs) = cgroup_cs((pos_cgrp)))))
229 * There are two global mutexes guarding cpuset structures - cpuset_mutex
230 * and callback_mutex. The latter may nest inside the former. We also
231 * require taking task_lock() when dereferencing a task's cpuset pointer.
232 * See "The task_lock() exception", at the end of this comment.
234 * A task must hold both mutexes to modify cpusets. If a task holds
235 * cpuset_mutex, then it blocks others wanting that mutex, ensuring that it
236 * is the only task able to also acquire callback_mutex and be able to
237 * modify cpusets. It can perform various checks on the cpuset structure
238 * first, knowing nothing will change. It can also allocate memory while
239 * just holding cpuset_mutex. While it is performing these checks, various
240 * callback routines can briefly acquire callback_mutex to query cpusets.
241 * Once it is ready to make the changes, it takes callback_mutex, blocking
244 * Calls to the kernel memory allocator can not be made while holding
245 * callback_mutex, as that would risk double tripping on callback_mutex
246 * from one of the callbacks into the cpuset code from within
249 * If a task is only holding callback_mutex, then it has read-only
252 * Now, the task_struct fields mems_allowed and mempolicy may be changed
253 * by other task, we use alloc_lock in the task_struct fields to protect
256 * The cpuset_common_file_read() handlers only hold callback_mutex across
257 * small pieces of code, such as when reading out possibly multi-word
258 * cpumasks and nodemasks.
260 * Accessing a task's cpuset should be done in accordance with the
261 * guidelines for accessing subsystem state in kernel/cgroup.c
264 static DEFINE_MUTEX(cpuset_mutex);
265 static DEFINE_MUTEX(callback_mutex);
268 * CPU / memory hotplug is handled asynchronously.
270 static struct workqueue_struct *cpuset_propagate_hotplug_wq;
272 static void cpuset_hotplug_workfn(struct work_struct *work);
273 static void cpuset_propagate_hotplug_workfn(struct work_struct *work);
274 static void schedule_cpuset_propagate_hotplug(struct cpuset *cs);
276 static DECLARE_WORK(cpuset_hotplug_work, cpuset_hotplug_workfn);
279 * This is ugly, but preserves the userspace API for existing cpuset
280 * users. If someone tries to mount the "cpuset" filesystem, we
281 * silently switch it to mount "cgroup" instead
283 static struct dentry *cpuset_mount(struct file_system_type *fs_type,
284 int flags, const char *unused_dev_name, void *data)
286 struct file_system_type *cgroup_fs = get_fs_type("cgroup");
287 struct dentry *ret = ERR_PTR(-ENODEV);
291 "release_agent=/sbin/cpuset_release_agent";
292 ret = cgroup_fs->mount(cgroup_fs, flags,
293 unused_dev_name, mountopts);
294 put_filesystem(cgroup_fs);
299 static struct file_system_type cpuset_fs_type = {
301 .mount = cpuset_mount,
305 * Return in pmask the portion of a cpusets's cpus_allowed that
306 * are online. If none are online, walk up the cpuset hierarchy
307 * until we find one that does have some online cpus. If we get
308 * all the way to the top and still haven't found any online cpus,
309 * return cpu_online_mask. Or if passed a NULL cs from an exit'ing
310 * task, return cpu_online_mask.
312 * One way or another, we guarantee to return some non-empty subset
313 * of cpu_online_mask.
315 * Call with callback_mutex held.
318 static void guarantee_online_cpus(const struct cpuset *cs,
319 struct cpumask *pmask)
321 while (cs && !cpumask_intersects(cs->cpus_allowed, cpu_online_mask))
324 cpumask_and(pmask, cs->cpus_allowed, cpu_online_mask);
326 cpumask_copy(pmask, cpu_online_mask);
327 BUG_ON(!cpumask_intersects(pmask, cpu_online_mask));
331 * Return in *pmask the portion of a cpusets's mems_allowed that
332 * are online, with memory. If none are online with memory, walk
333 * up the cpuset hierarchy until we find one that does have some
334 * online mems. If we get all the way to the top and still haven't
335 * found any online mems, return node_states[N_MEMORY].
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.
343 static void guarantee_online_mems(const struct cpuset *cs, nodemask_t *pmask)
345 while (cs && !nodes_intersects(cs->mems_allowed,
346 node_states[N_MEMORY]))
349 nodes_and(*pmask, cs->mems_allowed,
350 node_states[N_MEMORY]);
352 *pmask = node_states[N_MEMORY];
353 BUG_ON(!nodes_intersects(*pmask, node_states[N_MEMORY]));
357 * update task's spread flag if cpuset's page/slab spread flag is set
359 * Called with callback_mutex/cpuset_mutex held
361 static void cpuset_update_task_spread_flag(struct cpuset *cs,
362 struct task_struct *tsk)
364 if (is_spread_page(cs))
365 tsk->flags |= PF_SPREAD_PAGE;
367 tsk->flags &= ~PF_SPREAD_PAGE;
368 if (is_spread_slab(cs))
369 tsk->flags |= PF_SPREAD_SLAB;
371 tsk->flags &= ~PF_SPREAD_SLAB;
375 * is_cpuset_subset(p, q) - Is cpuset p a subset of cpuset q?
377 * One cpuset is a subset of another if all its allowed CPUs and
378 * Memory Nodes are a subset of the other, and its exclusive flags
379 * are only set if the other's are set. Call holding cpuset_mutex.
382 static int is_cpuset_subset(const struct cpuset *p, const struct cpuset *q)
384 return cpumask_subset(p->cpus_allowed, q->cpus_allowed) &&
385 nodes_subset(p->mems_allowed, q->mems_allowed) &&
386 is_cpu_exclusive(p) <= is_cpu_exclusive(q) &&
387 is_mem_exclusive(p) <= is_mem_exclusive(q);
391 * alloc_trial_cpuset - allocate a trial cpuset
392 * @cs: the cpuset that the trial cpuset duplicates
394 static struct cpuset *alloc_trial_cpuset(const struct cpuset *cs)
396 struct cpuset *trial;
398 trial = kmemdup(cs, sizeof(*cs), GFP_KERNEL);
402 if (!alloc_cpumask_var(&trial->cpus_allowed, GFP_KERNEL)) {
406 cpumask_copy(trial->cpus_allowed, cs->cpus_allowed);
412 * free_trial_cpuset - free the trial cpuset
413 * @trial: the trial cpuset to be freed
415 static void free_trial_cpuset(struct cpuset *trial)
417 free_cpumask_var(trial->cpus_allowed);
422 * validate_change() - Used to validate that any proposed cpuset change
423 * follows the structural rules for cpusets.
425 * If we replaced the flag and mask values of the current cpuset
426 * (cur) with those values in the trial cpuset (trial), would
427 * our various subset and exclusive rules still be valid? Presumes
430 * 'cur' is the address of an actual, in-use cpuset. Operations
431 * such as list traversal that depend on the actual address of the
432 * cpuset in the list must use cur below, not trial.
434 * 'trial' is the address of bulk structure copy of cur, with
435 * perhaps one or more of the fields cpus_allowed, mems_allowed,
436 * or flags changed to new, trial values.
438 * Return 0 if valid, -errno if not.
441 static int validate_change(const struct cpuset *cur, const struct cpuset *trial)
444 struct cpuset *c, *par;
449 /* Each of our child cpusets must be a subset of us */
451 cpuset_for_each_child(c, cont, cur)
452 if (!is_cpuset_subset(c, trial))
455 /* Remaining checks don't apply to root cpuset */
457 if (cur == &top_cpuset)
460 par = parent_cs(cur);
462 /* We must be a subset of our parent cpuset */
464 if (!is_cpuset_subset(trial, par))
468 * If either I or some sibling (!= me) is exclusive, we can't
472 cpuset_for_each_child(c, cont, par) {
473 if ((is_cpu_exclusive(trial) || is_cpu_exclusive(c)) &&
475 cpumask_intersects(trial->cpus_allowed, c->cpus_allowed))
477 if ((is_mem_exclusive(trial) || is_mem_exclusive(c)) &&
479 nodes_intersects(trial->mems_allowed, c->mems_allowed))
484 * Cpusets with tasks - existing or newly being attached - can't
485 * have empty cpus_allowed or mems_allowed.
488 if ((cgroup_task_count(cur->css.cgroup) || cur->attach_in_progress) &&
489 (cpumask_empty(trial->cpus_allowed) ||
490 nodes_empty(trial->mems_allowed)))
501 * Helper routine for generate_sched_domains().
502 * Do cpusets a, b have overlapping cpus_allowed masks?
504 static int cpusets_overlap(struct cpuset *a, struct cpuset *b)
506 return cpumask_intersects(a->cpus_allowed, b->cpus_allowed);
510 update_domain_attr(struct sched_domain_attr *dattr, struct cpuset *c)
512 if (dattr->relax_domain_level < c->relax_domain_level)
513 dattr->relax_domain_level = c->relax_domain_level;
517 static void update_domain_attr_tree(struct sched_domain_attr *dattr,
518 struct cpuset *root_cs)
521 struct cgroup *pos_cgrp;
524 cpuset_for_each_descendant_pre(cp, pos_cgrp, root_cs) {
525 /* skip the whole subtree if @cp doesn't have any CPU */
526 if (cpumask_empty(cp->cpus_allowed)) {
527 pos_cgrp = cgroup_rightmost_descendant(pos_cgrp);
531 if (is_sched_load_balance(cp))
532 update_domain_attr(dattr, cp);
538 * generate_sched_domains()
540 * This function builds a partial partition of the systems CPUs
541 * A 'partial partition' is a set of non-overlapping subsets whose
542 * union is a subset of that set.
543 * The output of this function needs to be passed to kernel/sched.c
544 * partition_sched_domains() routine, which will rebuild the scheduler's
545 * load balancing domains (sched domains) as specified by that partial
548 * See "What is sched_load_balance" in Documentation/cgroups/cpusets.txt
549 * for a background explanation of this.
551 * Does not return errors, on the theory that the callers of this
552 * routine would rather not worry about failures to rebuild sched
553 * domains when operating in the severe memory shortage situations
554 * that could cause allocation failures below.
556 * Must be called with cpuset_mutex held.
558 * The three key local variables below are:
559 * q - a linked-list queue of cpuset pointers, used to implement a
560 * top-down scan of all cpusets. This scan loads a pointer
561 * to each cpuset marked is_sched_load_balance into the
562 * array 'csa'. For our purposes, rebuilding the schedulers
563 * sched domains, we can ignore !is_sched_load_balance cpusets.
564 * csa - (for CpuSet Array) Array of pointers to all the cpusets
565 * that need to be load balanced, for convenient iterative
566 * access by the subsequent code that finds the best partition,
567 * i.e the set of domains (subsets) of CPUs such that the
568 * cpus_allowed of every cpuset marked is_sched_load_balance
569 * is a subset of one of these domains, while there are as
570 * many such domains as possible, each as small as possible.
571 * doms - Conversion of 'csa' to an array of cpumasks, for passing to
572 * the kernel/sched.c routine partition_sched_domains() in a
573 * convenient format, that can be easily compared to the prior
574 * value to determine what partition elements (sched domains)
575 * were changed (added or removed.)
577 * Finding the best partition (set of domains):
578 * The triple nested loops below over i, j, k scan over the
579 * load balanced cpusets (using the array of cpuset pointers in
580 * csa[]) looking for pairs of cpusets that have overlapping
581 * cpus_allowed, but which don't have the same 'pn' partition
582 * number and gives them in the same partition number. It keeps
583 * looping on the 'restart' label until it can no longer find
586 * The union of the cpus_allowed masks from the set of
587 * all cpusets having the same 'pn' value then form the one
588 * element of the partition (one sched domain) to be passed to
589 * partition_sched_domains().
591 static int generate_sched_domains(cpumask_var_t **domains,
592 struct sched_domain_attr **attributes)
594 struct cpuset *cp; /* scans q */
595 struct cpuset **csa; /* array of all cpuset ptrs */
596 int csn; /* how many cpuset ptrs in csa so far */
597 int i, j, k; /* indices for partition finding loops */
598 cpumask_var_t *doms; /* resulting partition; i.e. sched domains */
599 struct sched_domain_attr *dattr; /* attributes for custom domains */
600 int ndoms = 0; /* number of sched domains in result */
601 int nslot; /* next empty doms[] struct cpumask slot */
602 struct cgroup *pos_cgrp;
608 /* Special case for the 99% of systems with one, full, sched domain */
609 if (is_sched_load_balance(&top_cpuset)) {
611 doms = alloc_sched_domains(ndoms);
615 dattr = kmalloc(sizeof(struct sched_domain_attr), GFP_KERNEL);
617 *dattr = SD_ATTR_INIT;
618 update_domain_attr_tree(dattr, &top_cpuset);
620 cpumask_copy(doms[0], top_cpuset.cpus_allowed);
625 csa = kmalloc(number_of_cpusets * sizeof(cp), GFP_KERNEL);
631 cpuset_for_each_descendant_pre(cp, pos_cgrp, &top_cpuset) {
633 * Continue traversing beyond @cp iff @cp has some CPUs and
634 * isn't load balancing. The former is obvious. The
635 * latter: All child cpusets contain a subset of the
636 * parent's cpus, so just skip them, and then we call
637 * update_domain_attr_tree() to calc relax_domain_level of
638 * the corresponding sched domain.
640 if (!cpumask_empty(cp->cpus_allowed) &&
641 !is_sched_load_balance(cp))
644 if (is_sched_load_balance(cp))
647 /* skip @cp's subtree */
648 pos_cgrp = cgroup_rightmost_descendant(pos_cgrp);
652 for (i = 0; i < csn; i++)
657 /* Find the best partition (set of sched domains) */
658 for (i = 0; i < csn; i++) {
659 struct cpuset *a = csa[i];
662 for (j = 0; j < csn; j++) {
663 struct cpuset *b = csa[j];
666 if (apn != bpn && cpusets_overlap(a, b)) {
667 for (k = 0; k < csn; k++) {
668 struct cpuset *c = csa[k];
673 ndoms--; /* one less element */
680 * Now we know how many domains to create.
681 * Convert <csn, csa> to <ndoms, doms> and populate cpu masks.
683 doms = alloc_sched_domains(ndoms);
688 * The rest of the code, including the scheduler, can deal with
689 * dattr==NULL case. No need to abort if alloc fails.
691 dattr = kmalloc(ndoms * sizeof(struct sched_domain_attr), GFP_KERNEL);
693 for (nslot = 0, i = 0; i < csn; i++) {
694 struct cpuset *a = csa[i];
699 /* Skip completed partitions */
705 if (nslot == ndoms) {
706 static int warnings = 10;
709 "rebuild_sched_domains confused:"
710 " nslot %d, ndoms %d, csn %d, i %d,"
712 nslot, ndoms, csn, i, apn);
720 *(dattr + nslot) = SD_ATTR_INIT;
721 for (j = i; j < csn; j++) {
722 struct cpuset *b = csa[j];
725 cpumask_or(dp, dp, b->cpus_allowed);
727 update_domain_attr_tree(dattr + nslot, b);
729 /* Done with this partition */
735 BUG_ON(nslot != ndoms);
741 * Fallback to the default domain if kmalloc() failed.
742 * See comments in partition_sched_domains().
753 * Rebuild scheduler domains.
755 * If the flag 'sched_load_balance' of any cpuset with non-empty
756 * 'cpus' changes, or if the 'cpus' allowed changes in any cpuset
757 * which has that flag enabled, or if any cpuset with a non-empty
758 * 'cpus' is removed, then call this routine to rebuild the
759 * scheduler's dynamic sched domains.
761 * Call with cpuset_mutex held. Takes get_online_cpus().
763 static void rebuild_sched_domains_locked(void)
765 struct sched_domain_attr *attr;
769 lockdep_assert_held(&cpuset_mutex);
773 * We have raced with CPU hotplug. Don't do anything to avoid
774 * passing doms with offlined cpu to partition_sched_domains().
775 * Anyways, hotplug work item will rebuild sched domains.
777 if (!cpumask_equal(top_cpuset.cpus_allowed, cpu_active_mask))
780 /* Generate domain masks and attrs */
781 ndoms = generate_sched_domains(&doms, &attr);
783 /* Have scheduler rebuild the domains */
784 partition_sched_domains(ndoms, doms, attr);
788 #else /* !CONFIG_SMP */
789 static void rebuild_sched_domains_locked(void)
792 #endif /* CONFIG_SMP */
794 void rebuild_sched_domains(void)
796 mutex_lock(&cpuset_mutex);
797 rebuild_sched_domains_locked();
798 mutex_unlock(&cpuset_mutex);
802 * cpuset_test_cpumask - test a task's cpus_allowed versus its cpuset's
804 * @scan: struct cgroup_scanner contained in its struct cpuset_hotplug_scanner
806 * Call with cpuset_mutex held. May take callback_mutex during call.
807 * Called for each task in a cgroup by cgroup_scan_tasks().
808 * Return nonzero if this tasks's cpus_allowed mask should be changed (in other
809 * words, if its mask is not equal to its cpuset's mask).
811 static int cpuset_test_cpumask(struct task_struct *tsk,
812 struct cgroup_scanner *scan)
814 return !cpumask_equal(&tsk->cpus_allowed,
815 (cgroup_cs(scan->cg))->cpus_allowed);
819 * cpuset_change_cpumask - make a task's cpus_allowed the same as its cpuset's
821 * @scan: struct cgroup_scanner containing the cgroup of the task
823 * Called by cgroup_scan_tasks() for each task in a cgroup whose
824 * cpus_allowed mask needs to be changed.
826 * We don't need to re-check for the cgroup/cpuset membership, since we're
827 * holding cpuset_mutex at this point.
829 static void cpuset_change_cpumask(struct task_struct *tsk,
830 struct cgroup_scanner *scan)
832 set_cpus_allowed_ptr(tsk, ((cgroup_cs(scan->cg))->cpus_allowed));
836 * update_tasks_cpumask - Update the cpumasks of tasks in the cpuset.
837 * @cs: the cpuset in which each task's cpus_allowed mask needs to be changed
838 * @heap: if NULL, defer allocating heap memory to cgroup_scan_tasks()
840 * Called with cpuset_mutex held
842 * The cgroup_scan_tasks() function will scan all the tasks in a cgroup,
843 * calling callback functions for each.
845 * No return value. It's guaranteed that cgroup_scan_tasks() always returns 0
848 static void update_tasks_cpumask(struct cpuset *cs, struct ptr_heap *heap)
850 struct cgroup_scanner scan;
852 scan.cg = cs->css.cgroup;
853 scan.test_task = cpuset_test_cpumask;
854 scan.process_task = cpuset_change_cpumask;
856 cgroup_scan_tasks(&scan);
860 * update_cpumask - update the cpus_allowed mask of a cpuset and all tasks in it
861 * @cs: the cpuset to consider
862 * @buf: buffer of cpu numbers written to this cpuset
864 static int update_cpumask(struct cpuset *cs, struct cpuset *trialcs,
867 struct ptr_heap heap;
869 int is_load_balanced;
871 /* top_cpuset.cpus_allowed tracks cpu_online_mask; it's read-only */
872 if (cs == &top_cpuset)
876 * An empty cpus_allowed is ok only if the cpuset has no tasks.
877 * Since cpulist_parse() fails on an empty mask, we special case
878 * that parsing. The validate_change() call ensures that cpusets
879 * with tasks have cpus.
882 cpumask_clear(trialcs->cpus_allowed);
884 retval = cpulist_parse(buf, trialcs->cpus_allowed);
888 if (!cpumask_subset(trialcs->cpus_allowed, cpu_active_mask))
891 retval = validate_change(cs, trialcs);
895 /* Nothing to do if the cpus didn't change */
896 if (cpumask_equal(cs->cpus_allowed, trialcs->cpus_allowed))
899 retval = heap_init(&heap, PAGE_SIZE, GFP_KERNEL, NULL);
903 is_load_balanced = is_sched_load_balance(trialcs);
905 mutex_lock(&callback_mutex);
906 cpumask_copy(cs->cpus_allowed, trialcs->cpus_allowed);
907 mutex_unlock(&callback_mutex);
910 * Scan tasks in the cpuset, and update the cpumasks of any
911 * that need an update.
913 update_tasks_cpumask(cs, &heap);
917 if (is_load_balanced)
918 rebuild_sched_domains_locked();
925 * Migrate memory region from one set of nodes to another.
927 * Temporarilly set tasks mems_allowed to target nodes of migration,
928 * so that the migration code can allocate pages on these nodes.
930 * Call holding cpuset_mutex, so current's cpuset won't change
931 * during this call, as manage_mutex holds off any cpuset_attach()
932 * calls. Therefore we don't need to take task_lock around the
933 * call to guarantee_online_mems(), as we know no one is changing
936 * While the mm_struct we are migrating is typically from some
937 * other task, the task_struct mems_allowed that we are hacking
938 * is for our current task, which must allocate new pages for that
939 * migrating memory region.
942 static void cpuset_migrate_mm(struct mm_struct *mm, const nodemask_t *from,
943 const nodemask_t *to)
945 struct task_struct *tsk = current;
947 tsk->mems_allowed = *to;
949 do_migrate_pages(mm, from, to, MPOL_MF_MOVE_ALL);
951 guarantee_online_mems(task_cs(tsk),&tsk->mems_allowed);
955 * cpuset_change_task_nodemask - change task's mems_allowed and mempolicy
956 * @tsk: the task to change
957 * @newmems: new nodes that the task will be set
959 * In order to avoid seeing no nodes if the old and new nodes are disjoint,
960 * we structure updates as setting all new allowed nodes, then clearing newly
963 static void cpuset_change_task_nodemask(struct task_struct *tsk,
969 * Allow tasks that have access to memory reserves because they have
970 * been OOM killed to get memory anywhere.
972 if (unlikely(test_thread_flag(TIF_MEMDIE)))
974 if (current->flags & PF_EXITING) /* Let dying task have memory */
979 * Determine if a loop is necessary if another thread is doing
980 * get_mems_allowed(). If at least one node remains unchanged and
981 * tsk does not have a mempolicy, then an empty nodemask will not be
982 * possible when mems_allowed is larger than a word.
984 need_loop = task_has_mempolicy(tsk) ||
985 !nodes_intersects(*newmems, tsk->mems_allowed);
989 write_seqcount_begin(&tsk->mems_allowed_seq);
992 nodes_or(tsk->mems_allowed, tsk->mems_allowed, *newmems);
993 mpol_rebind_task(tsk, newmems, MPOL_REBIND_STEP1);
995 mpol_rebind_task(tsk, newmems, MPOL_REBIND_STEP2);
996 tsk->mems_allowed = *newmems;
999 write_seqcount_end(&tsk->mems_allowed_seq);
1007 * Update task's mems_allowed and rebind its mempolicy and vmas' mempolicy
1008 * of it to cpuset's new mems_allowed, and migrate pages to new nodes if
1009 * memory_migrate flag is set. Called with cpuset_mutex held.
1011 static void cpuset_change_nodemask(struct task_struct *p,
1012 struct cgroup_scanner *scan)
1014 struct mm_struct *mm;
1017 const nodemask_t *oldmem = scan->data;
1018 static nodemask_t newmems; /* protected by cpuset_mutex */
1020 cs = cgroup_cs(scan->cg);
1021 guarantee_online_mems(cs, &newmems);
1023 cpuset_change_task_nodemask(p, &newmems);
1025 mm = get_task_mm(p);
1029 migrate = is_memory_migrate(cs);
1031 mpol_rebind_mm(mm, &cs->mems_allowed);
1033 cpuset_migrate_mm(mm, oldmem, &cs->mems_allowed);
1037 static void *cpuset_being_rebound;
1040 * update_tasks_nodemask - Update the nodemasks of tasks in the cpuset.
1041 * @cs: the cpuset in which each task's mems_allowed mask needs to be changed
1042 * @oldmem: old mems_allowed of cpuset cs
1043 * @heap: if NULL, defer allocating heap memory to cgroup_scan_tasks()
1045 * Called with cpuset_mutex held
1046 * No return value. It's guaranteed that cgroup_scan_tasks() always returns 0
1049 static void update_tasks_nodemask(struct cpuset *cs, const nodemask_t *oldmem,
1050 struct ptr_heap *heap)
1052 struct cgroup_scanner scan;
1054 cpuset_being_rebound = cs; /* causes mpol_dup() rebind */
1056 scan.cg = cs->css.cgroup;
1057 scan.test_task = NULL;
1058 scan.process_task = cpuset_change_nodemask;
1060 scan.data = (nodemask_t *)oldmem;
1063 * The mpol_rebind_mm() call takes mmap_sem, which we couldn't
1064 * take while holding tasklist_lock. Forks can happen - the
1065 * mpol_dup() cpuset_being_rebound check will catch such forks,
1066 * and rebind their vma mempolicies too. Because we still hold
1067 * the global cpuset_mutex, we know that no other rebind effort
1068 * will be contending for the global variable cpuset_being_rebound.
1069 * It's ok if we rebind the same mm twice; mpol_rebind_mm()
1070 * is idempotent. Also migrate pages in each mm to new nodes.
1072 cgroup_scan_tasks(&scan);
1074 /* We're done rebinding vmas to this cpuset's new mems_allowed. */
1075 cpuset_being_rebound = NULL;
1079 * Handle user request to change the 'mems' memory placement
1080 * of a cpuset. Needs to validate the request, update the
1081 * cpusets mems_allowed, and for each task in the cpuset,
1082 * update mems_allowed and rebind task's mempolicy and any vma
1083 * mempolicies and if the cpuset is marked 'memory_migrate',
1084 * migrate the tasks pages to the new memory.
1086 * Call with cpuset_mutex held. May take callback_mutex during call.
1087 * Will take tasklist_lock, scan tasklist for tasks in cpuset cs,
1088 * lock each such tasks mm->mmap_sem, scan its vma's and rebind
1089 * their mempolicies to the cpusets new mems_allowed.
1091 static int update_nodemask(struct cpuset *cs, struct cpuset *trialcs,
1094 NODEMASK_ALLOC(nodemask_t, oldmem, GFP_KERNEL);
1096 struct ptr_heap heap;
1102 * top_cpuset.mems_allowed tracks node_stats[N_MEMORY];
1105 if (cs == &top_cpuset) {
1111 * An empty mems_allowed is ok iff there are no tasks in the cpuset.
1112 * Since nodelist_parse() fails on an empty mask, we special case
1113 * that parsing. The validate_change() call ensures that cpusets
1114 * with tasks have memory.
1117 nodes_clear(trialcs->mems_allowed);
1119 retval = nodelist_parse(buf, trialcs->mems_allowed);
1123 if (!nodes_subset(trialcs->mems_allowed,
1124 node_states[N_MEMORY])) {
1129 *oldmem = cs->mems_allowed;
1130 if (nodes_equal(*oldmem, trialcs->mems_allowed)) {
1131 retval = 0; /* Too easy - nothing to do */
1134 retval = validate_change(cs, trialcs);
1138 retval = heap_init(&heap, PAGE_SIZE, GFP_KERNEL, NULL);
1142 mutex_lock(&callback_mutex);
1143 cs->mems_allowed = trialcs->mems_allowed;
1144 mutex_unlock(&callback_mutex);
1146 update_tasks_nodemask(cs, oldmem, &heap);
1150 NODEMASK_FREE(oldmem);
1154 int current_cpuset_is_being_rebound(void)
1156 return task_cs(current) == cpuset_being_rebound;
1159 static int update_relax_domain_level(struct cpuset *cs, s64 val)
1162 if (val < -1 || val >= sched_domain_level_max)
1166 if (val != cs->relax_domain_level) {
1167 cs->relax_domain_level = val;
1168 if (!cpumask_empty(cs->cpus_allowed) &&
1169 is_sched_load_balance(cs))
1170 rebuild_sched_domains_locked();
1177 * cpuset_change_flag - make a task's spread flags the same as its cpuset's
1178 * @tsk: task to be updated
1179 * @scan: struct cgroup_scanner containing the cgroup of the task
1181 * Called by cgroup_scan_tasks() for each task in a cgroup.
1183 * We don't need to re-check for the cgroup/cpuset membership, since we're
1184 * holding cpuset_mutex at this point.
1186 static void cpuset_change_flag(struct task_struct *tsk,
1187 struct cgroup_scanner *scan)
1189 cpuset_update_task_spread_flag(cgroup_cs(scan->cg), tsk);
1193 * update_tasks_flags - update the spread flags of tasks in the cpuset.
1194 * @cs: the cpuset in which each task's spread flags needs to be changed
1195 * @heap: if NULL, defer allocating heap memory to cgroup_scan_tasks()
1197 * Called with cpuset_mutex held
1199 * The cgroup_scan_tasks() function will scan all the tasks in a cgroup,
1200 * calling callback functions for each.
1202 * No return value. It's guaranteed that cgroup_scan_tasks() always returns 0
1205 static void update_tasks_flags(struct cpuset *cs, struct ptr_heap *heap)
1207 struct cgroup_scanner scan;
1209 scan.cg = cs->css.cgroup;
1210 scan.test_task = NULL;
1211 scan.process_task = cpuset_change_flag;
1213 cgroup_scan_tasks(&scan);
1217 * update_flag - read a 0 or a 1 in a file and update associated flag
1218 * bit: the bit to update (see cpuset_flagbits_t)
1219 * cs: the cpuset to update
1220 * turning_on: whether the flag is being set or cleared
1222 * Call with cpuset_mutex held.
1225 static int update_flag(cpuset_flagbits_t bit, struct cpuset *cs,
1228 struct cpuset *trialcs;
1229 int balance_flag_changed;
1230 int spread_flag_changed;
1231 struct ptr_heap heap;
1234 trialcs = alloc_trial_cpuset(cs);
1239 set_bit(bit, &trialcs->flags);
1241 clear_bit(bit, &trialcs->flags);
1243 err = validate_change(cs, trialcs);
1247 err = heap_init(&heap, PAGE_SIZE, GFP_KERNEL, NULL);
1251 balance_flag_changed = (is_sched_load_balance(cs) !=
1252 is_sched_load_balance(trialcs));
1254 spread_flag_changed = ((is_spread_slab(cs) != is_spread_slab(trialcs))
1255 || (is_spread_page(cs) != is_spread_page(trialcs)));
1257 mutex_lock(&callback_mutex);
1258 cs->flags = trialcs->flags;
1259 mutex_unlock(&callback_mutex);
1261 if (!cpumask_empty(trialcs->cpus_allowed) && balance_flag_changed)
1262 rebuild_sched_domains_locked();
1264 if (spread_flag_changed)
1265 update_tasks_flags(cs, &heap);
1268 free_trial_cpuset(trialcs);
1273 * Frequency meter - How fast is some event occurring?
1275 * These routines manage a digitally filtered, constant time based,
1276 * event frequency meter. There are four routines:
1277 * fmeter_init() - initialize a frequency meter.
1278 * fmeter_markevent() - called each time the event happens.
1279 * fmeter_getrate() - returns the recent rate of such events.
1280 * fmeter_update() - internal routine used to update fmeter.
1282 * A common data structure is passed to each of these routines,
1283 * which is used to keep track of the state required to manage the
1284 * frequency meter and its digital filter.
1286 * The filter works on the number of events marked per unit time.
1287 * The filter is single-pole low-pass recursive (IIR). The time unit
1288 * is 1 second. Arithmetic is done using 32-bit integers scaled to
1289 * simulate 3 decimal digits of precision (multiplied by 1000).
1291 * With an FM_COEF of 933, and a time base of 1 second, the filter
1292 * has a half-life of 10 seconds, meaning that if the events quit
1293 * happening, then the rate returned from the fmeter_getrate()
1294 * will be cut in half each 10 seconds, until it converges to zero.
1296 * It is not worth doing a real infinitely recursive filter. If more
1297 * than FM_MAXTICKS ticks have elapsed since the last filter event,
1298 * just compute FM_MAXTICKS ticks worth, by which point the level
1301 * Limit the count of unprocessed events to FM_MAXCNT, so as to avoid
1302 * arithmetic overflow in the fmeter_update() routine.
1304 * Given the simple 32 bit integer arithmetic used, this meter works
1305 * best for reporting rates between one per millisecond (msec) and
1306 * one per 32 (approx) seconds. At constant rates faster than one
1307 * per msec it maxes out at values just under 1,000,000. At constant
1308 * rates between one per msec, and one per second it will stabilize
1309 * to a value N*1000, where N is the rate of events per second.
1310 * At constant rates between one per second and one per 32 seconds,
1311 * it will be choppy, moving up on the seconds that have an event,
1312 * and then decaying until the next event. At rates slower than
1313 * about one in 32 seconds, it decays all the way back to zero between
1317 #define FM_COEF 933 /* coefficient for half-life of 10 secs */
1318 #define FM_MAXTICKS ((time_t)99) /* useless computing more ticks than this */
1319 #define FM_MAXCNT 1000000 /* limit cnt to avoid overflow */
1320 #define FM_SCALE 1000 /* faux fixed point scale */
1322 /* Initialize a frequency meter */
1323 static void fmeter_init(struct fmeter *fmp)
1328 spin_lock_init(&fmp->lock);
1331 /* Internal meter update - process cnt events and update value */
1332 static void fmeter_update(struct fmeter *fmp)
1334 time_t now = get_seconds();
1335 time_t ticks = now - fmp->time;
1340 ticks = min(FM_MAXTICKS, ticks);
1342 fmp->val = (FM_COEF * fmp->val) / FM_SCALE;
1345 fmp->val += ((FM_SCALE - FM_COEF) * fmp->cnt) / FM_SCALE;
1349 /* Process any previous ticks, then bump cnt by one (times scale). */
1350 static void fmeter_markevent(struct fmeter *fmp)
1352 spin_lock(&fmp->lock);
1354 fmp->cnt = min(FM_MAXCNT, fmp->cnt + FM_SCALE);
1355 spin_unlock(&fmp->lock);
1358 /* Process any previous ticks, then return current value. */
1359 static int fmeter_getrate(struct fmeter *fmp)
1363 spin_lock(&fmp->lock);
1366 spin_unlock(&fmp->lock);
1370 /* Called by cgroups to determine if a cpuset is usable; cpuset_mutex held */
1371 static int cpuset_can_attach(struct cgroup *cgrp, struct cgroup_taskset *tset)
1373 struct cpuset *cs = cgroup_cs(cgrp);
1374 struct task_struct *task;
1377 mutex_lock(&cpuset_mutex);
1380 if (cpumask_empty(cs->cpus_allowed) || nodes_empty(cs->mems_allowed))
1383 cgroup_taskset_for_each(task, cgrp, tset) {
1385 * Kthreads which disallow setaffinity shouldn't be moved
1386 * to a new cpuset; we don't want to change their cpu
1387 * affinity and isolating such threads by their set of
1388 * allowed nodes is unnecessary. Thus, cpusets are not
1389 * applicable for such threads. This prevents checking for
1390 * success of set_cpus_allowed_ptr() on all attached tasks
1391 * before cpus_allowed may be changed.
1394 if (task->flags & PF_NO_SETAFFINITY)
1396 ret = security_task_setscheduler(task);
1402 * Mark attach is in progress. This makes validate_change() fail
1403 * changes which zero cpus/mems_allowed.
1405 cs->attach_in_progress++;
1408 mutex_unlock(&cpuset_mutex);
1412 static void cpuset_cancel_attach(struct cgroup *cgrp,
1413 struct cgroup_taskset *tset)
1415 mutex_lock(&cpuset_mutex);
1416 cgroup_cs(cgrp)->attach_in_progress--;
1417 mutex_unlock(&cpuset_mutex);
1421 * Protected by cpuset_mutex. cpus_attach is used only by cpuset_attach()
1422 * but we can't allocate it dynamically there. Define it global and
1423 * allocate from cpuset_init().
1425 static cpumask_var_t cpus_attach;
1427 static void cpuset_attach(struct cgroup *cgrp, struct cgroup_taskset *tset)
1429 /* static bufs protected by cpuset_mutex */
1430 static nodemask_t cpuset_attach_nodemask_from;
1431 static nodemask_t cpuset_attach_nodemask_to;
1432 struct mm_struct *mm;
1433 struct task_struct *task;
1434 struct task_struct *leader = cgroup_taskset_first(tset);
1435 struct cgroup *oldcgrp = cgroup_taskset_cur_cgroup(tset);
1436 struct cpuset *cs = cgroup_cs(cgrp);
1437 struct cpuset *oldcs = cgroup_cs(oldcgrp);
1439 mutex_lock(&cpuset_mutex);
1441 /* prepare for attach */
1442 if (cs == &top_cpuset)
1443 cpumask_copy(cpus_attach, cpu_possible_mask);
1445 guarantee_online_cpus(cs, cpus_attach);
1447 guarantee_online_mems(cs, &cpuset_attach_nodemask_to);
1449 cgroup_taskset_for_each(task, cgrp, tset) {
1451 * can_attach beforehand should guarantee that this doesn't
1452 * fail. TODO: have a better way to handle failure here
1454 WARN_ON_ONCE(set_cpus_allowed_ptr(task, cpus_attach));
1456 cpuset_change_task_nodemask(task, &cpuset_attach_nodemask_to);
1457 cpuset_update_task_spread_flag(cs, task);
1461 * Change mm, possibly for multiple threads in a threadgroup. This is
1462 * expensive and may sleep.
1464 cpuset_attach_nodemask_from = oldcs->mems_allowed;
1465 cpuset_attach_nodemask_to = cs->mems_allowed;
1466 mm = get_task_mm(leader);
1468 mpol_rebind_mm(mm, &cpuset_attach_nodemask_to);
1469 if (is_memory_migrate(cs))
1470 cpuset_migrate_mm(mm, &cpuset_attach_nodemask_from,
1471 &cpuset_attach_nodemask_to);
1475 cs->attach_in_progress--;
1478 * We may have raced with CPU/memory hotunplug. Trigger hotplug
1479 * propagation if @cs doesn't have any CPU or memory. It will move
1480 * the newly added tasks to the nearest parent which can execute.
1482 if (cpumask_empty(cs->cpus_allowed) || nodes_empty(cs->mems_allowed))
1483 schedule_cpuset_propagate_hotplug(cs);
1485 mutex_unlock(&cpuset_mutex);
1488 /* The various types of files and directories in a cpuset file system */
1491 FILE_MEMORY_MIGRATE,
1497 FILE_SCHED_LOAD_BALANCE,
1498 FILE_SCHED_RELAX_DOMAIN_LEVEL,
1499 FILE_MEMORY_PRESSURE_ENABLED,
1500 FILE_MEMORY_PRESSURE,
1503 } cpuset_filetype_t;
1505 static int cpuset_write_u64(struct cgroup *cgrp, struct cftype *cft, u64 val)
1507 struct cpuset *cs = cgroup_cs(cgrp);
1508 cpuset_filetype_t type = cft->private;
1511 mutex_lock(&cpuset_mutex);
1512 if (!is_cpuset_online(cs)) {
1518 case FILE_CPU_EXCLUSIVE:
1519 retval = update_flag(CS_CPU_EXCLUSIVE, cs, val);
1521 case FILE_MEM_EXCLUSIVE:
1522 retval = update_flag(CS_MEM_EXCLUSIVE, cs, val);
1524 case FILE_MEM_HARDWALL:
1525 retval = update_flag(CS_MEM_HARDWALL, cs, val);
1527 case FILE_SCHED_LOAD_BALANCE:
1528 retval = update_flag(CS_SCHED_LOAD_BALANCE, cs, val);
1530 case FILE_MEMORY_MIGRATE:
1531 retval = update_flag(CS_MEMORY_MIGRATE, cs, val);
1533 case FILE_MEMORY_PRESSURE_ENABLED:
1534 cpuset_memory_pressure_enabled = !!val;
1536 case FILE_MEMORY_PRESSURE:
1539 case FILE_SPREAD_PAGE:
1540 retval = update_flag(CS_SPREAD_PAGE, cs, val);
1542 case FILE_SPREAD_SLAB:
1543 retval = update_flag(CS_SPREAD_SLAB, cs, val);
1550 mutex_unlock(&cpuset_mutex);
1554 static int cpuset_write_s64(struct cgroup *cgrp, struct cftype *cft, s64 val)
1556 struct cpuset *cs = cgroup_cs(cgrp);
1557 cpuset_filetype_t type = cft->private;
1558 int retval = -ENODEV;
1560 mutex_lock(&cpuset_mutex);
1561 if (!is_cpuset_online(cs))
1565 case FILE_SCHED_RELAX_DOMAIN_LEVEL:
1566 retval = update_relax_domain_level(cs, val);
1573 mutex_unlock(&cpuset_mutex);
1578 * Common handling for a write to a "cpus" or "mems" file.
1580 static int cpuset_write_resmask(struct cgroup *cgrp, struct cftype *cft,
1583 struct cpuset *cs = cgroup_cs(cgrp);
1584 struct cpuset *trialcs;
1585 int retval = -ENODEV;
1588 * CPU or memory hotunplug may leave @cs w/o any execution
1589 * resources, in which case the hotplug code asynchronously updates
1590 * configuration and transfers all tasks to the nearest ancestor
1591 * which can execute.
1593 * As writes to "cpus" or "mems" may restore @cs's execution
1594 * resources, wait for the previously scheduled operations before
1595 * proceeding, so that we don't end up keep removing tasks added
1596 * after execution capability is restored.
1598 * Flushing cpuset_hotplug_work is enough to synchronize against
1599 * hotplug hanlding; however, cpuset_attach() may schedule
1600 * propagation work directly. Flush the workqueue too.
1602 flush_work(&cpuset_hotplug_work);
1603 flush_workqueue(cpuset_propagate_hotplug_wq);
1605 mutex_lock(&cpuset_mutex);
1606 if (!is_cpuset_online(cs))
1609 trialcs = alloc_trial_cpuset(cs);
1615 switch (cft->private) {
1617 retval = update_cpumask(cs, trialcs, buf);
1620 retval = update_nodemask(cs, trialcs, buf);
1627 free_trial_cpuset(trialcs);
1629 mutex_unlock(&cpuset_mutex);
1634 * These ascii lists should be read in a single call, by using a user
1635 * buffer large enough to hold the entire map. If read in smaller
1636 * chunks, there is no guarantee of atomicity. Since the display format
1637 * used, list of ranges of sequential numbers, is variable length,
1638 * and since these maps can change value dynamically, one could read
1639 * gibberish by doing partial reads while a list was changing.
1640 * A single large read to a buffer that crosses a page boundary is
1641 * ok, because the result being copied to user land is not recomputed
1642 * across a page fault.
1645 static size_t cpuset_sprintf_cpulist(char *page, struct cpuset *cs)
1649 mutex_lock(&callback_mutex);
1650 count = cpulist_scnprintf(page, PAGE_SIZE, cs->cpus_allowed);
1651 mutex_unlock(&callback_mutex);
1656 static size_t cpuset_sprintf_memlist(char *page, struct cpuset *cs)
1660 mutex_lock(&callback_mutex);
1661 count = nodelist_scnprintf(page, PAGE_SIZE, cs->mems_allowed);
1662 mutex_unlock(&callback_mutex);
1667 static ssize_t cpuset_common_file_read(struct cgroup *cont,
1671 size_t nbytes, loff_t *ppos)
1673 struct cpuset *cs = cgroup_cs(cont);
1674 cpuset_filetype_t type = cft->private;
1679 if (!(page = (char *)__get_free_page(GFP_TEMPORARY)))
1686 s += cpuset_sprintf_cpulist(s, cs);
1689 s += cpuset_sprintf_memlist(s, cs);
1697 retval = simple_read_from_buffer(buf, nbytes, ppos, page, s - page);
1699 free_page((unsigned long)page);
1703 static u64 cpuset_read_u64(struct cgroup *cont, struct cftype *cft)
1705 struct cpuset *cs = cgroup_cs(cont);
1706 cpuset_filetype_t type = cft->private;
1708 case FILE_CPU_EXCLUSIVE:
1709 return is_cpu_exclusive(cs);
1710 case FILE_MEM_EXCLUSIVE:
1711 return is_mem_exclusive(cs);
1712 case FILE_MEM_HARDWALL:
1713 return is_mem_hardwall(cs);
1714 case FILE_SCHED_LOAD_BALANCE:
1715 return is_sched_load_balance(cs);
1716 case FILE_MEMORY_MIGRATE:
1717 return is_memory_migrate(cs);
1718 case FILE_MEMORY_PRESSURE_ENABLED:
1719 return cpuset_memory_pressure_enabled;
1720 case FILE_MEMORY_PRESSURE:
1721 return fmeter_getrate(&cs->fmeter);
1722 case FILE_SPREAD_PAGE:
1723 return is_spread_page(cs);
1724 case FILE_SPREAD_SLAB:
1725 return is_spread_slab(cs);
1730 /* Unreachable but makes gcc happy */
1734 static s64 cpuset_read_s64(struct cgroup *cont, struct cftype *cft)
1736 struct cpuset *cs = cgroup_cs(cont);
1737 cpuset_filetype_t type = cft->private;
1739 case FILE_SCHED_RELAX_DOMAIN_LEVEL:
1740 return cs->relax_domain_level;
1745 /* Unrechable but makes gcc happy */
1751 * for the common functions, 'private' gives the type of file
1754 static struct cftype files[] = {
1757 .read = cpuset_common_file_read,
1758 .write_string = cpuset_write_resmask,
1759 .max_write_len = (100U + 6 * NR_CPUS),
1760 .private = FILE_CPULIST,
1765 .read = cpuset_common_file_read,
1766 .write_string = cpuset_write_resmask,
1767 .max_write_len = (100U + 6 * MAX_NUMNODES),
1768 .private = FILE_MEMLIST,
1772 .name = "cpu_exclusive",
1773 .read_u64 = cpuset_read_u64,
1774 .write_u64 = cpuset_write_u64,
1775 .private = FILE_CPU_EXCLUSIVE,
1779 .name = "mem_exclusive",
1780 .read_u64 = cpuset_read_u64,
1781 .write_u64 = cpuset_write_u64,
1782 .private = FILE_MEM_EXCLUSIVE,
1786 .name = "mem_hardwall",
1787 .read_u64 = cpuset_read_u64,
1788 .write_u64 = cpuset_write_u64,
1789 .private = FILE_MEM_HARDWALL,
1793 .name = "sched_load_balance",
1794 .read_u64 = cpuset_read_u64,
1795 .write_u64 = cpuset_write_u64,
1796 .private = FILE_SCHED_LOAD_BALANCE,
1800 .name = "sched_relax_domain_level",
1801 .read_s64 = cpuset_read_s64,
1802 .write_s64 = cpuset_write_s64,
1803 .private = FILE_SCHED_RELAX_DOMAIN_LEVEL,
1807 .name = "memory_migrate",
1808 .read_u64 = cpuset_read_u64,
1809 .write_u64 = cpuset_write_u64,
1810 .private = FILE_MEMORY_MIGRATE,
1814 .name = "memory_pressure",
1815 .read_u64 = cpuset_read_u64,
1816 .write_u64 = cpuset_write_u64,
1817 .private = FILE_MEMORY_PRESSURE,
1822 .name = "memory_spread_page",
1823 .read_u64 = cpuset_read_u64,
1824 .write_u64 = cpuset_write_u64,
1825 .private = FILE_SPREAD_PAGE,
1829 .name = "memory_spread_slab",
1830 .read_u64 = cpuset_read_u64,
1831 .write_u64 = cpuset_write_u64,
1832 .private = FILE_SPREAD_SLAB,
1836 .name = "memory_pressure_enabled",
1837 .flags = CFTYPE_ONLY_ON_ROOT,
1838 .read_u64 = cpuset_read_u64,
1839 .write_u64 = cpuset_write_u64,
1840 .private = FILE_MEMORY_PRESSURE_ENABLED,
1847 * cpuset_css_alloc - allocate a cpuset css
1848 * cont: control group that the new cpuset will be part of
1851 static struct cgroup_subsys_state *cpuset_css_alloc(struct cgroup *cont)
1856 return &top_cpuset.css;
1858 cs = kzalloc(sizeof(*cs), GFP_KERNEL);
1860 return ERR_PTR(-ENOMEM);
1861 if (!alloc_cpumask_var(&cs->cpus_allowed, GFP_KERNEL)) {
1863 return ERR_PTR(-ENOMEM);
1866 set_bit(CS_SCHED_LOAD_BALANCE, &cs->flags);
1867 cpumask_clear(cs->cpus_allowed);
1868 nodes_clear(cs->mems_allowed);
1869 fmeter_init(&cs->fmeter);
1870 INIT_WORK(&cs->hotplug_work, cpuset_propagate_hotplug_workfn);
1871 cs->relax_domain_level = -1;
1876 static int cpuset_css_online(struct cgroup *cgrp)
1878 struct cpuset *cs = cgroup_cs(cgrp);
1879 struct cpuset *parent = parent_cs(cs);
1880 struct cpuset *tmp_cs;
1881 struct cgroup *pos_cg;
1886 mutex_lock(&cpuset_mutex);
1888 set_bit(CS_ONLINE, &cs->flags);
1889 if (is_spread_page(parent))
1890 set_bit(CS_SPREAD_PAGE, &cs->flags);
1891 if (is_spread_slab(parent))
1892 set_bit(CS_SPREAD_SLAB, &cs->flags);
1894 number_of_cpusets++;
1896 if (!test_bit(CGRP_CPUSET_CLONE_CHILDREN, &cgrp->flags))
1900 * Clone @parent's configuration if CGRP_CPUSET_CLONE_CHILDREN is
1901 * set. This flag handling is implemented in cgroup core for
1902 * histrical reasons - the flag may be specified during mount.
1904 * Currently, if any sibling cpusets have exclusive cpus or mem, we
1905 * refuse to clone the configuration - thereby refusing the task to
1906 * be entered, and as a result refusing the sys_unshare() or
1907 * clone() which initiated it. If this becomes a problem for some
1908 * users who wish to allow that scenario, then this could be
1909 * changed to grant parent->cpus_allowed-sibling_cpus_exclusive
1910 * (and likewise for mems) to the new cgroup.
1913 cpuset_for_each_child(tmp_cs, pos_cg, parent) {
1914 if (is_mem_exclusive(tmp_cs) || is_cpu_exclusive(tmp_cs)) {
1921 mutex_lock(&callback_mutex);
1922 cs->mems_allowed = parent->mems_allowed;
1923 cpumask_copy(cs->cpus_allowed, parent->cpus_allowed);
1924 mutex_unlock(&callback_mutex);
1926 mutex_unlock(&cpuset_mutex);
1930 static void cpuset_css_offline(struct cgroup *cgrp)
1932 struct cpuset *cs = cgroup_cs(cgrp);
1934 mutex_lock(&cpuset_mutex);
1936 if (is_sched_load_balance(cs))
1937 update_flag(CS_SCHED_LOAD_BALANCE, cs, 0);
1939 number_of_cpusets--;
1940 clear_bit(CS_ONLINE, &cs->flags);
1942 mutex_unlock(&cpuset_mutex);
1946 * If the cpuset being removed has its flag 'sched_load_balance'
1947 * enabled, then simulate turning sched_load_balance off, which
1948 * will call rebuild_sched_domains_locked().
1951 static void cpuset_css_free(struct cgroup *cont)
1953 struct cpuset *cs = cgroup_cs(cont);
1955 free_cpumask_var(cs->cpus_allowed);
1959 struct cgroup_subsys cpuset_subsys = {
1961 .css_alloc = cpuset_css_alloc,
1962 .css_online = cpuset_css_online,
1963 .css_offline = cpuset_css_offline,
1964 .css_free = cpuset_css_free,
1965 .can_attach = cpuset_can_attach,
1966 .cancel_attach = cpuset_cancel_attach,
1967 .attach = cpuset_attach,
1968 .subsys_id = cpuset_subsys_id,
1969 .base_cftypes = files,
1974 * cpuset_init - initialize cpusets at system boot
1976 * Description: Initialize top_cpuset and the cpuset internal file system,
1979 int __init cpuset_init(void)
1983 if (!alloc_cpumask_var(&top_cpuset.cpus_allowed, GFP_KERNEL))
1986 cpumask_setall(top_cpuset.cpus_allowed);
1987 nodes_setall(top_cpuset.mems_allowed);
1989 fmeter_init(&top_cpuset.fmeter);
1990 set_bit(CS_SCHED_LOAD_BALANCE, &top_cpuset.flags);
1991 top_cpuset.relax_domain_level = -1;
1993 err = register_filesystem(&cpuset_fs_type);
1997 if (!alloc_cpumask_var(&cpus_attach, GFP_KERNEL))
2000 number_of_cpusets = 1;
2005 * If CPU and/or memory hotplug handlers, below, unplug any CPUs
2006 * or memory nodes, we need to walk over the cpuset hierarchy,
2007 * removing that CPU or node from all cpusets. If this removes the
2008 * last CPU or node from a cpuset, then move the tasks in the empty
2009 * cpuset to its next-highest non-empty parent.
2011 static void remove_tasks_in_empty_cpuset(struct cpuset *cs)
2013 struct cpuset *parent;
2016 * Find its next-highest non-empty parent, (top cpuset
2017 * has online cpus, so can't be empty).
2019 parent = parent_cs(cs);
2020 while (cpumask_empty(parent->cpus_allowed) ||
2021 nodes_empty(parent->mems_allowed))
2022 parent = parent_cs(parent);
2024 if (cgroup_transfer_tasks(parent->css.cgroup, cs->css.cgroup)) {
2026 printk(KERN_ERR "cpuset: failed to transfer tasks out of empty cpuset %s\n",
2027 cgroup_name(cs->css.cgroup));
2033 * cpuset_propagate_hotplug_workfn - propagate CPU/memory hotplug to a cpuset
2034 * @cs: cpuset in interest
2036 * Compare @cs's cpu and mem masks against top_cpuset and if some have gone
2037 * offline, update @cs accordingly. If @cs ends up with no CPU or memory,
2038 * all its tasks are moved to the nearest ancestor with both resources.
2040 static void cpuset_propagate_hotplug_workfn(struct work_struct *work)
2042 static cpumask_t off_cpus;
2043 static nodemask_t off_mems, tmp_mems;
2044 struct cpuset *cs = container_of(work, struct cpuset, hotplug_work);
2047 mutex_lock(&cpuset_mutex);
2049 cpumask_andnot(&off_cpus, cs->cpus_allowed, top_cpuset.cpus_allowed);
2050 nodes_andnot(off_mems, cs->mems_allowed, top_cpuset.mems_allowed);
2052 /* remove offline cpus from @cs */
2053 if (!cpumask_empty(&off_cpus)) {
2054 mutex_lock(&callback_mutex);
2055 cpumask_andnot(cs->cpus_allowed, cs->cpus_allowed, &off_cpus);
2056 mutex_unlock(&callback_mutex);
2057 update_tasks_cpumask(cs, NULL);
2060 /* remove offline mems from @cs */
2061 if (!nodes_empty(off_mems)) {
2062 tmp_mems = cs->mems_allowed;
2063 mutex_lock(&callback_mutex);
2064 nodes_andnot(cs->mems_allowed, cs->mems_allowed, off_mems);
2065 mutex_unlock(&callback_mutex);
2066 update_tasks_nodemask(cs, &tmp_mems, NULL);
2069 is_empty = cpumask_empty(cs->cpus_allowed) ||
2070 nodes_empty(cs->mems_allowed);
2072 mutex_unlock(&cpuset_mutex);
2075 * If @cs became empty, move tasks to the nearest ancestor with
2076 * execution resources. This is full cgroup operation which will
2077 * also call back into cpuset. Should be done outside any lock.
2080 remove_tasks_in_empty_cpuset(cs);
2082 /* the following may free @cs, should be the last operation */
2087 * schedule_cpuset_propagate_hotplug - schedule hotplug propagation to a cpuset
2088 * @cs: cpuset of interest
2090 * Schedule cpuset_propagate_hotplug_workfn() which will update CPU and
2091 * memory masks according to top_cpuset.
2093 static void schedule_cpuset_propagate_hotplug(struct cpuset *cs)
2096 * Pin @cs. The refcnt will be released when the work item
2097 * finishes executing.
2099 if (!css_tryget(&cs->css))
2103 * Queue @cs->hotplug_work. If already pending, lose the css ref.
2104 * cpuset_propagate_hotplug_wq is ordered and propagation will
2105 * happen in the order this function is called.
2107 if (!queue_work(cpuset_propagate_hotplug_wq, &cs->hotplug_work))
2112 * cpuset_hotplug_workfn - handle CPU/memory hotunplug for a cpuset
2114 * This function is called after either CPU or memory configuration has
2115 * changed and updates cpuset accordingly. The top_cpuset is always
2116 * synchronized to cpu_active_mask and N_MEMORY, which is necessary in
2117 * order to make cpusets transparent (of no affect) on systems that are
2118 * actively using CPU hotplug but making no active use of cpusets.
2120 * Non-root cpusets are only affected by offlining. If any CPUs or memory
2121 * nodes have been taken down, cpuset_propagate_hotplug() is invoked on all
2124 * Note that CPU offlining during suspend is ignored. We don't modify
2125 * cpusets across suspend/resume cycles at all.
2127 static void cpuset_hotplug_workfn(struct work_struct *work)
2129 static cpumask_t new_cpus, tmp_cpus;
2130 static nodemask_t new_mems, tmp_mems;
2131 bool cpus_updated, mems_updated;
2132 bool cpus_offlined, mems_offlined;
2134 mutex_lock(&cpuset_mutex);
2136 /* fetch the available cpus/mems and find out which changed how */
2137 cpumask_copy(&new_cpus, cpu_active_mask);
2138 new_mems = node_states[N_MEMORY];
2140 cpus_updated = !cpumask_equal(top_cpuset.cpus_allowed, &new_cpus);
2141 cpus_offlined = cpumask_andnot(&tmp_cpus, top_cpuset.cpus_allowed,
2144 mems_updated = !nodes_equal(top_cpuset.mems_allowed, new_mems);
2145 nodes_andnot(tmp_mems, top_cpuset.mems_allowed, new_mems);
2146 mems_offlined = !nodes_empty(tmp_mems);
2148 /* synchronize cpus_allowed to cpu_active_mask */
2150 mutex_lock(&callback_mutex);
2151 cpumask_copy(top_cpuset.cpus_allowed, &new_cpus);
2152 mutex_unlock(&callback_mutex);
2153 /* we don't mess with cpumasks of tasks in top_cpuset */
2156 /* synchronize mems_allowed to N_MEMORY */
2158 tmp_mems = top_cpuset.mems_allowed;
2159 mutex_lock(&callback_mutex);
2160 top_cpuset.mems_allowed = new_mems;
2161 mutex_unlock(&callback_mutex);
2162 update_tasks_nodemask(&top_cpuset, &tmp_mems, NULL);
2165 /* if cpus or mems went down, we need to propagate to descendants */
2166 if (cpus_offlined || mems_offlined) {
2168 struct cgroup *pos_cgrp;
2171 cpuset_for_each_descendant_pre(cs, pos_cgrp, &top_cpuset)
2172 schedule_cpuset_propagate_hotplug(cs);
2176 mutex_unlock(&cpuset_mutex);
2178 /* wait for propagations to finish */
2179 flush_workqueue(cpuset_propagate_hotplug_wq);
2181 /* rebuild sched domains if cpus_allowed has changed */
2183 rebuild_sched_domains();
2186 void cpuset_update_active_cpus(bool cpu_online)
2189 * We're inside cpu hotplug critical region which usually nests
2190 * inside cgroup synchronization. Bounce actual hotplug processing
2191 * to a work item to avoid reverse locking order.
2193 * We still need to do partition_sched_domains() synchronously;
2194 * otherwise, the scheduler will get confused and put tasks to the
2195 * dead CPU. Fall back to the default single domain.
2196 * cpuset_hotplug_workfn() will rebuild it as necessary.
2198 partition_sched_domains(1, NULL, NULL);
2199 schedule_work(&cpuset_hotplug_work);
2203 * Keep top_cpuset.mems_allowed tracking node_states[N_MEMORY].
2204 * Call this routine anytime after node_states[N_MEMORY] changes.
2205 * See cpuset_update_active_cpus() for CPU hotplug handling.
2207 static int cpuset_track_online_nodes(struct notifier_block *self,
2208 unsigned long action, void *arg)
2210 schedule_work(&cpuset_hotplug_work);
2214 static struct notifier_block cpuset_track_online_nodes_nb = {
2215 .notifier_call = cpuset_track_online_nodes,
2216 .priority = 10, /* ??! */
2220 * cpuset_init_smp - initialize cpus_allowed
2222 * Description: Finish top cpuset after cpu, node maps are initialized
2224 void __init cpuset_init_smp(void)
2226 cpumask_copy(top_cpuset.cpus_allowed, cpu_active_mask);
2227 top_cpuset.mems_allowed = node_states[N_MEMORY];
2229 register_hotmemory_notifier(&cpuset_track_online_nodes_nb);
2231 cpuset_propagate_hotplug_wq =
2232 alloc_ordered_workqueue("cpuset_hotplug", 0);
2233 BUG_ON(!cpuset_propagate_hotplug_wq);
2237 * cpuset_cpus_allowed - return cpus_allowed mask from a tasks cpuset.
2238 * @tsk: pointer to task_struct from which to obtain cpuset->cpus_allowed.
2239 * @pmask: pointer to struct cpumask variable to receive cpus_allowed set.
2241 * Description: Returns the cpumask_var_t cpus_allowed of the cpuset
2242 * attached to the specified @tsk. Guaranteed to return some non-empty
2243 * subset of cpu_online_mask, even if this means going outside the
2247 void cpuset_cpus_allowed(struct task_struct *tsk, struct cpumask *pmask)
2249 mutex_lock(&callback_mutex);
2251 guarantee_online_cpus(task_cs(tsk), pmask);
2253 mutex_unlock(&callback_mutex);
2256 void cpuset_cpus_allowed_fallback(struct task_struct *tsk)
2258 const struct cpuset *cs;
2263 do_set_cpus_allowed(tsk, cs->cpus_allowed);
2267 * We own tsk->cpus_allowed, nobody can change it under us.
2269 * But we used cs && cs->cpus_allowed lockless and thus can
2270 * race with cgroup_attach_task() or update_cpumask() and get
2271 * the wrong tsk->cpus_allowed. However, both cases imply the
2272 * subsequent cpuset_change_cpumask()->set_cpus_allowed_ptr()
2273 * which takes task_rq_lock().
2275 * If we are called after it dropped the lock we must see all
2276 * changes in tsk_cs()->cpus_allowed. Otherwise we can temporary
2277 * set any mask even if it is not right from task_cs() pov,
2278 * the pending set_cpus_allowed_ptr() will fix things.
2280 * select_fallback_rq() will fix things ups and set cpu_possible_mask
2285 void cpuset_init_current_mems_allowed(void)
2287 nodes_setall(current->mems_allowed);
2291 * cpuset_mems_allowed - return mems_allowed mask from a tasks cpuset.
2292 * @tsk: pointer to task_struct from which to obtain cpuset->mems_allowed.
2294 * Description: Returns the nodemask_t mems_allowed of the cpuset
2295 * attached to the specified @tsk. Guaranteed to return some non-empty
2296 * subset of node_states[N_MEMORY], even if this means going outside the
2300 nodemask_t cpuset_mems_allowed(struct task_struct *tsk)
2304 mutex_lock(&callback_mutex);
2306 guarantee_online_mems(task_cs(tsk), &mask);
2308 mutex_unlock(&callback_mutex);
2314 * cpuset_nodemask_valid_mems_allowed - check nodemask vs. curremt mems_allowed
2315 * @nodemask: the nodemask to be checked
2317 * Are any of the nodes in the nodemask allowed in current->mems_allowed?
2319 int cpuset_nodemask_valid_mems_allowed(nodemask_t *nodemask)
2321 return nodes_intersects(*nodemask, current->mems_allowed);
2325 * nearest_hardwall_ancestor() - Returns the nearest mem_exclusive or
2326 * mem_hardwall ancestor to the specified cpuset. Call holding
2327 * callback_mutex. If no ancestor is mem_exclusive or mem_hardwall
2328 * (an unusual configuration), then returns the root cpuset.
2330 static const struct cpuset *nearest_hardwall_ancestor(const struct cpuset *cs)
2332 while (!(is_mem_exclusive(cs) || is_mem_hardwall(cs)) && parent_cs(cs))
2338 * cpuset_node_allowed_softwall - Can we allocate on a memory node?
2339 * @node: is this an allowed node?
2340 * @gfp_mask: memory allocation flags
2342 * If we're in interrupt, yes, we can always allocate. If __GFP_THISNODE is
2343 * set, yes, we can always allocate. If node is in our task's mems_allowed,
2344 * yes. If it's not a __GFP_HARDWALL request and this node is in the nearest
2345 * hardwalled cpuset ancestor to this task's cpuset, yes. If the task has been
2346 * OOM killed and has access to memory reserves as specified by the TIF_MEMDIE
2350 * If __GFP_HARDWALL is set, cpuset_node_allowed_softwall() reduces to
2351 * cpuset_node_allowed_hardwall(). Otherwise, cpuset_node_allowed_softwall()
2352 * might sleep, and might allow a node from an enclosing cpuset.
2354 * cpuset_node_allowed_hardwall() only handles the simpler case of hardwall
2355 * cpusets, and never sleeps.
2357 * The __GFP_THISNODE placement logic is really handled elsewhere,
2358 * by forcibly using a zonelist starting at a specified node, and by
2359 * (in get_page_from_freelist()) refusing to consider the zones for
2360 * any node on the zonelist except the first. By the time any such
2361 * calls get to this routine, we should just shut up and say 'yes'.
2363 * GFP_USER allocations are marked with the __GFP_HARDWALL bit,
2364 * and do not allow allocations outside the current tasks cpuset
2365 * unless the task has been OOM killed as is marked TIF_MEMDIE.
2366 * GFP_KERNEL allocations are not so marked, so can escape to the
2367 * nearest enclosing hardwalled ancestor cpuset.
2369 * Scanning up parent cpusets requires callback_mutex. The
2370 * __alloc_pages() routine only calls here with __GFP_HARDWALL bit
2371 * _not_ set if it's a GFP_KERNEL allocation, and all nodes in the
2372 * current tasks mems_allowed came up empty on the first pass over
2373 * the zonelist. So only GFP_KERNEL allocations, if all nodes in the
2374 * cpuset are short of memory, might require taking the callback_mutex
2377 * The first call here from mm/page_alloc:get_page_from_freelist()
2378 * has __GFP_HARDWALL set in gfp_mask, enforcing hardwall cpusets,
2379 * so no allocation on a node outside the cpuset is allowed (unless
2380 * in interrupt, of course).
2382 * The second pass through get_page_from_freelist() doesn't even call
2383 * here for GFP_ATOMIC calls. For those calls, the __alloc_pages()
2384 * variable 'wait' is not set, and the bit ALLOC_CPUSET is not set
2385 * in alloc_flags. That logic and the checks below have the combined
2387 * in_interrupt - any node ok (current task context irrelevant)
2388 * GFP_ATOMIC - any node ok
2389 * TIF_MEMDIE - any node ok
2390 * GFP_KERNEL - any node in enclosing hardwalled cpuset ok
2391 * GFP_USER - only nodes in current tasks mems allowed ok.
2394 * Don't call cpuset_node_allowed_softwall if you can't sleep, unless you
2395 * pass in the __GFP_HARDWALL flag set in gfp_flag, which disables
2396 * the code that might scan up ancestor cpusets and sleep.
2398 int __cpuset_node_allowed_softwall(int node, gfp_t gfp_mask)
2400 const struct cpuset *cs; /* current cpuset ancestors */
2401 int allowed; /* is allocation in zone z allowed? */
2403 if (in_interrupt() || (gfp_mask & __GFP_THISNODE))
2405 might_sleep_if(!(gfp_mask & __GFP_HARDWALL));
2406 if (node_isset(node, current->mems_allowed))
2409 * Allow tasks that have access to memory reserves because they have
2410 * been OOM killed to get memory anywhere.
2412 if (unlikely(test_thread_flag(TIF_MEMDIE)))
2414 if (gfp_mask & __GFP_HARDWALL) /* If hardwall request, stop here */
2417 if (current->flags & PF_EXITING) /* Let dying task have memory */
2420 /* Not hardwall and node outside mems_allowed: scan up cpusets */
2421 mutex_lock(&callback_mutex);
2424 cs = nearest_hardwall_ancestor(task_cs(current));
2425 allowed = node_isset(node, cs->mems_allowed);
2426 task_unlock(current);
2428 mutex_unlock(&callback_mutex);
2433 * cpuset_node_allowed_hardwall - Can we allocate on a memory node?
2434 * @node: is this an allowed node?
2435 * @gfp_mask: memory allocation flags
2437 * If we're in interrupt, yes, we can always allocate. If __GFP_THISNODE is
2438 * set, yes, we can always allocate. If node is in our task's mems_allowed,
2439 * yes. If the task has been OOM killed and has access to memory reserves as
2440 * specified by the TIF_MEMDIE flag, yes.
2443 * The __GFP_THISNODE placement logic is really handled elsewhere,
2444 * by forcibly using a zonelist starting at a specified node, and by
2445 * (in get_page_from_freelist()) refusing to consider the zones for
2446 * any node on the zonelist except the first. By the time any such
2447 * calls get to this routine, we should just shut up and say 'yes'.
2449 * Unlike the cpuset_node_allowed_softwall() variant, above,
2450 * this variant requires that the node be in the current task's
2451 * mems_allowed or that we're in interrupt. It does not scan up the
2452 * cpuset hierarchy for the nearest enclosing mem_exclusive cpuset.
2455 int __cpuset_node_allowed_hardwall(int node, gfp_t gfp_mask)
2457 if (in_interrupt() || (gfp_mask & __GFP_THISNODE))
2459 if (node_isset(node, current->mems_allowed))
2462 * Allow tasks that have access to memory reserves because they have
2463 * been OOM killed to get memory anywhere.
2465 if (unlikely(test_thread_flag(TIF_MEMDIE)))
2471 * cpuset_mem_spread_node() - On which node to begin search for a file page
2472 * cpuset_slab_spread_node() - On which node to begin search for a slab page
2474 * If a task is marked PF_SPREAD_PAGE or PF_SPREAD_SLAB (as for
2475 * tasks in a cpuset with is_spread_page or is_spread_slab set),
2476 * and if the memory allocation used cpuset_mem_spread_node()
2477 * to determine on which node to start looking, as it will for
2478 * certain page cache or slab cache pages such as used for file
2479 * system buffers and inode caches, then instead of starting on the
2480 * local node to look for a free page, rather spread the starting
2481 * node around the tasks mems_allowed nodes.
2483 * We don't have to worry about the returned node being offline
2484 * because "it can't happen", and even if it did, it would be ok.
2486 * The routines calling guarantee_online_mems() are careful to
2487 * only set nodes in task->mems_allowed that are online. So it
2488 * should not be possible for the following code to return an
2489 * offline node. But if it did, that would be ok, as this routine
2490 * is not returning the node where the allocation must be, only
2491 * the node where the search should start. The zonelist passed to
2492 * __alloc_pages() will include all nodes. If the slab allocator
2493 * is passed an offline node, it will fall back to the local node.
2494 * See kmem_cache_alloc_node().
2497 static int cpuset_spread_node(int *rotor)
2501 node = next_node(*rotor, current->mems_allowed);
2502 if (node == MAX_NUMNODES)
2503 node = first_node(current->mems_allowed);
2508 int cpuset_mem_spread_node(void)
2510 if (current->cpuset_mem_spread_rotor == NUMA_NO_NODE)
2511 current->cpuset_mem_spread_rotor =
2512 node_random(¤t->mems_allowed);
2514 return cpuset_spread_node(¤t->cpuset_mem_spread_rotor);
2517 int cpuset_slab_spread_node(void)
2519 if (current->cpuset_slab_spread_rotor == NUMA_NO_NODE)
2520 current->cpuset_slab_spread_rotor =
2521 node_random(¤t->mems_allowed);
2523 return cpuset_spread_node(¤t->cpuset_slab_spread_rotor);
2526 EXPORT_SYMBOL_GPL(cpuset_mem_spread_node);
2529 * cpuset_mems_allowed_intersects - Does @tsk1's mems_allowed intersect @tsk2's?
2530 * @tsk1: pointer to task_struct of some task.
2531 * @tsk2: pointer to task_struct of some other task.
2533 * Description: Return true if @tsk1's mems_allowed intersects the
2534 * mems_allowed of @tsk2. Used by the OOM killer to determine if
2535 * one of the task's memory usage might impact the memory available
2539 int cpuset_mems_allowed_intersects(const struct task_struct *tsk1,
2540 const struct task_struct *tsk2)
2542 return nodes_intersects(tsk1->mems_allowed, tsk2->mems_allowed);
2545 #define CPUSET_NODELIST_LEN (256)
2548 * cpuset_print_task_mems_allowed - prints task's cpuset and mems_allowed
2549 * @task: pointer to task_struct of some task.
2551 * Description: Prints @task's name, cpuset name, and cached copy of its
2552 * mems_allowed to the kernel log. Must hold task_lock(task) to allow
2553 * dereferencing task_cs(task).
2555 void cpuset_print_task_mems_allowed(struct task_struct *tsk)
2557 /* Statically allocated to prevent using excess stack. */
2558 static char cpuset_nodelist[CPUSET_NODELIST_LEN];
2559 static DEFINE_SPINLOCK(cpuset_buffer_lock);
2561 struct cgroup *cgrp = task_cs(tsk)->css.cgroup;
2564 spin_lock(&cpuset_buffer_lock);
2566 nodelist_scnprintf(cpuset_nodelist, CPUSET_NODELIST_LEN,
2568 printk(KERN_INFO "%s cpuset=%s mems_allowed=%s\n",
2569 tsk->comm, cgroup_name(cgrp), cpuset_nodelist);
2571 spin_unlock(&cpuset_buffer_lock);
2576 * Collection of memory_pressure is suppressed unless
2577 * this flag is enabled by writing "1" to the special
2578 * cpuset file 'memory_pressure_enabled' in the root cpuset.
2581 int cpuset_memory_pressure_enabled __read_mostly;
2584 * cpuset_memory_pressure_bump - keep stats of per-cpuset reclaims.
2586 * Keep a running average of the rate of synchronous (direct)
2587 * page reclaim efforts initiated by tasks in each cpuset.
2589 * This represents the rate at which some task in the cpuset
2590 * ran low on memory on all nodes it was allowed to use, and
2591 * had to enter the kernels page reclaim code in an effort to
2592 * create more free memory by tossing clean pages or swapping
2593 * or writing dirty pages.
2595 * Display to user space in the per-cpuset read-only file
2596 * "memory_pressure". Value displayed is an integer
2597 * representing the recent rate of entry into the synchronous
2598 * (direct) page reclaim by any task attached to the cpuset.
2601 void __cpuset_memory_pressure_bump(void)
2604 fmeter_markevent(&task_cs(current)->fmeter);
2605 task_unlock(current);
2608 #ifdef CONFIG_PROC_PID_CPUSET
2610 * proc_cpuset_show()
2611 * - Print tasks cpuset path into seq_file.
2612 * - Used for /proc/<pid>/cpuset.
2613 * - No need to task_lock(tsk) on this tsk->cpuset reference, as it
2614 * doesn't really matter if tsk->cpuset changes after we read it,
2615 * and we take cpuset_mutex, keeping cpuset_attach() from changing it
2618 int proc_cpuset_show(struct seq_file *m, void *unused_v)
2621 struct task_struct *tsk;
2623 struct cgroup_subsys_state *css;
2627 buf = kmalloc(PAGE_SIZE, GFP_KERNEL);
2633 tsk = get_pid_task(pid, PIDTYPE_PID);
2638 css = task_subsys_state(tsk, cpuset_subsys_id);
2639 retval = cgroup_path(css->cgroup, buf, PAGE_SIZE);
2646 put_task_struct(tsk);
2652 #endif /* CONFIG_PROC_PID_CPUSET */
2654 /* Display task mems_allowed in /proc/<pid>/status file. */
2655 void cpuset_task_status_allowed(struct seq_file *m, struct task_struct *task)
2657 seq_printf(m, "Mems_allowed:\t");
2658 seq_nodemask(m, &task->mems_allowed);
2659 seq_printf(m, "\n");
2660 seq_printf(m, "Mems_allowed_list:\t");
2661 seq_nodemask_list(m, &task->mems_allowed);
2662 seq_printf(m, "\n");