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. The top
308 * cpuset always has some cpus online.
310 * One way or another, we guarantee to return some non-empty subset
311 * of cpu_online_mask.
313 * Call with callback_mutex held.
315 static void guarantee_online_cpus(const struct cpuset *cs,
316 struct cpumask *pmask)
318 while (!cpumask_intersects(cs->cpus_allowed, cpu_online_mask))
320 cpumask_and(pmask, cs->cpus_allowed, cpu_online_mask);
324 * Return in *pmask the portion of a cpusets's mems_allowed that
325 * are online, with memory. If none are online with memory, walk
326 * up the cpuset hierarchy until we find one that does have some
327 * online mems. The top cpuset always has some mems online.
329 * One way or another, we guarantee to return some non-empty subset
330 * of node_states[N_MEMORY].
332 * Call with callback_mutex held.
334 static void guarantee_online_mems(const struct cpuset *cs, nodemask_t *pmask)
336 while (!nodes_intersects(cs->mems_allowed, node_states[N_MEMORY]))
338 nodes_and(*pmask, cs->mems_allowed, node_states[N_MEMORY]);
342 * update task's spread flag if cpuset's page/slab spread flag is set
344 * Called with callback_mutex/cpuset_mutex held
346 static void cpuset_update_task_spread_flag(struct cpuset *cs,
347 struct task_struct *tsk)
349 if (is_spread_page(cs))
350 tsk->flags |= PF_SPREAD_PAGE;
352 tsk->flags &= ~PF_SPREAD_PAGE;
353 if (is_spread_slab(cs))
354 tsk->flags |= PF_SPREAD_SLAB;
356 tsk->flags &= ~PF_SPREAD_SLAB;
360 * is_cpuset_subset(p, q) - Is cpuset p a subset of cpuset q?
362 * One cpuset is a subset of another if all its allowed CPUs and
363 * Memory Nodes are a subset of the other, and its exclusive flags
364 * are only set if the other's are set. Call holding cpuset_mutex.
367 static int is_cpuset_subset(const struct cpuset *p, const struct cpuset *q)
369 return cpumask_subset(p->cpus_allowed, q->cpus_allowed) &&
370 nodes_subset(p->mems_allowed, q->mems_allowed) &&
371 is_cpu_exclusive(p) <= is_cpu_exclusive(q) &&
372 is_mem_exclusive(p) <= is_mem_exclusive(q);
376 * alloc_trial_cpuset - allocate a trial cpuset
377 * @cs: the cpuset that the trial cpuset duplicates
379 static struct cpuset *alloc_trial_cpuset(const struct cpuset *cs)
381 struct cpuset *trial;
383 trial = kmemdup(cs, sizeof(*cs), GFP_KERNEL);
387 if (!alloc_cpumask_var(&trial->cpus_allowed, GFP_KERNEL)) {
391 cpumask_copy(trial->cpus_allowed, cs->cpus_allowed);
397 * free_trial_cpuset - free the trial cpuset
398 * @trial: the trial cpuset to be freed
400 static void free_trial_cpuset(struct cpuset *trial)
402 free_cpumask_var(trial->cpus_allowed);
407 * validate_change() - Used to validate that any proposed cpuset change
408 * follows the structural rules for cpusets.
410 * If we replaced the flag and mask values of the current cpuset
411 * (cur) with those values in the trial cpuset (trial), would
412 * our various subset and exclusive rules still be valid? Presumes
415 * 'cur' is the address of an actual, in-use cpuset. Operations
416 * such as list traversal that depend on the actual address of the
417 * cpuset in the list must use cur below, not trial.
419 * 'trial' is the address of bulk structure copy of cur, with
420 * perhaps one or more of the fields cpus_allowed, mems_allowed,
421 * or flags changed to new, trial values.
423 * Return 0 if valid, -errno if not.
426 static int validate_change(const struct cpuset *cur, const struct cpuset *trial)
429 struct cpuset *c, *par;
434 /* Each of our child cpusets must be a subset of us */
436 cpuset_for_each_child(c, cont, cur)
437 if (!is_cpuset_subset(c, trial))
440 /* Remaining checks don't apply to root cpuset */
442 if (cur == &top_cpuset)
445 par = parent_cs(cur);
447 /* We must be a subset of our parent cpuset */
449 if (!is_cpuset_subset(trial, par))
453 * If either I or some sibling (!= me) is exclusive, we can't
457 cpuset_for_each_child(c, cont, par) {
458 if ((is_cpu_exclusive(trial) || is_cpu_exclusive(c)) &&
460 cpumask_intersects(trial->cpus_allowed, c->cpus_allowed))
462 if ((is_mem_exclusive(trial) || is_mem_exclusive(c)) &&
464 nodes_intersects(trial->mems_allowed, c->mems_allowed))
469 * Cpusets with tasks - existing or newly being attached - can't
470 * have empty cpus_allowed or mems_allowed.
473 if ((cgroup_task_count(cur->css.cgroup) || cur->attach_in_progress) &&
474 (cpumask_empty(trial->cpus_allowed) ||
475 nodes_empty(trial->mems_allowed)))
486 * Helper routine for generate_sched_domains().
487 * Do cpusets a, b have overlapping cpus_allowed masks?
489 static int cpusets_overlap(struct cpuset *a, struct cpuset *b)
491 return cpumask_intersects(a->cpus_allowed, b->cpus_allowed);
495 update_domain_attr(struct sched_domain_attr *dattr, struct cpuset *c)
497 if (dattr->relax_domain_level < c->relax_domain_level)
498 dattr->relax_domain_level = c->relax_domain_level;
502 static void update_domain_attr_tree(struct sched_domain_attr *dattr,
503 struct cpuset *root_cs)
506 struct cgroup *pos_cgrp;
509 cpuset_for_each_descendant_pre(cp, pos_cgrp, root_cs) {
510 /* skip the whole subtree if @cp doesn't have any CPU */
511 if (cpumask_empty(cp->cpus_allowed)) {
512 pos_cgrp = cgroup_rightmost_descendant(pos_cgrp);
516 if (is_sched_load_balance(cp))
517 update_domain_attr(dattr, cp);
523 * generate_sched_domains()
525 * This function builds a partial partition of the systems CPUs
526 * A 'partial partition' is a set of non-overlapping subsets whose
527 * union is a subset of that set.
528 * The output of this function needs to be passed to kernel/sched.c
529 * partition_sched_domains() routine, which will rebuild the scheduler's
530 * load balancing domains (sched domains) as specified by that partial
533 * See "What is sched_load_balance" in Documentation/cgroups/cpusets.txt
534 * for a background explanation of this.
536 * Does not return errors, on the theory that the callers of this
537 * routine would rather not worry about failures to rebuild sched
538 * domains when operating in the severe memory shortage situations
539 * that could cause allocation failures below.
541 * Must be called with cpuset_mutex held.
543 * The three key local variables below are:
544 * q - a linked-list queue of cpuset pointers, used to implement a
545 * top-down scan of all cpusets. This scan loads a pointer
546 * to each cpuset marked is_sched_load_balance into the
547 * array 'csa'. For our purposes, rebuilding the schedulers
548 * sched domains, we can ignore !is_sched_load_balance cpusets.
549 * csa - (for CpuSet Array) Array of pointers to all the cpusets
550 * that need to be load balanced, for convenient iterative
551 * access by the subsequent code that finds the best partition,
552 * i.e the set of domains (subsets) of CPUs such that the
553 * cpus_allowed of every cpuset marked is_sched_load_balance
554 * is a subset of one of these domains, while there are as
555 * many such domains as possible, each as small as possible.
556 * doms - Conversion of 'csa' to an array of cpumasks, for passing to
557 * the kernel/sched.c routine partition_sched_domains() in a
558 * convenient format, that can be easily compared to the prior
559 * value to determine what partition elements (sched domains)
560 * were changed (added or removed.)
562 * Finding the best partition (set of domains):
563 * The triple nested loops below over i, j, k scan over the
564 * load balanced cpusets (using the array of cpuset pointers in
565 * csa[]) looking for pairs of cpusets that have overlapping
566 * cpus_allowed, but which don't have the same 'pn' partition
567 * number and gives them in the same partition number. It keeps
568 * looping on the 'restart' label until it can no longer find
571 * The union of the cpus_allowed masks from the set of
572 * all cpusets having the same 'pn' value then form the one
573 * element of the partition (one sched domain) to be passed to
574 * partition_sched_domains().
576 static int generate_sched_domains(cpumask_var_t **domains,
577 struct sched_domain_attr **attributes)
579 struct cpuset *cp; /* scans q */
580 struct cpuset **csa; /* array of all cpuset ptrs */
581 int csn; /* how many cpuset ptrs in csa so far */
582 int i, j, k; /* indices for partition finding loops */
583 cpumask_var_t *doms; /* resulting partition; i.e. sched domains */
584 struct sched_domain_attr *dattr; /* attributes for custom domains */
585 int ndoms = 0; /* number of sched domains in result */
586 int nslot; /* next empty doms[] struct cpumask slot */
587 struct cgroup *pos_cgrp;
593 /* Special case for the 99% of systems with one, full, sched domain */
594 if (is_sched_load_balance(&top_cpuset)) {
596 doms = alloc_sched_domains(ndoms);
600 dattr = kmalloc(sizeof(struct sched_domain_attr), GFP_KERNEL);
602 *dattr = SD_ATTR_INIT;
603 update_domain_attr_tree(dattr, &top_cpuset);
605 cpumask_copy(doms[0], top_cpuset.cpus_allowed);
610 csa = kmalloc(number_of_cpusets * sizeof(cp), GFP_KERNEL);
616 cpuset_for_each_descendant_pre(cp, pos_cgrp, &top_cpuset) {
618 * Continue traversing beyond @cp iff @cp has some CPUs and
619 * isn't load balancing. The former is obvious. The
620 * latter: All child cpusets contain a subset of the
621 * parent's cpus, so just skip them, and then we call
622 * update_domain_attr_tree() to calc relax_domain_level of
623 * the corresponding sched domain.
625 if (!cpumask_empty(cp->cpus_allowed) &&
626 !is_sched_load_balance(cp))
629 if (is_sched_load_balance(cp))
632 /* skip @cp's subtree */
633 pos_cgrp = cgroup_rightmost_descendant(pos_cgrp);
637 for (i = 0; i < csn; i++)
642 /* Find the best partition (set of sched domains) */
643 for (i = 0; i < csn; i++) {
644 struct cpuset *a = csa[i];
647 for (j = 0; j < csn; j++) {
648 struct cpuset *b = csa[j];
651 if (apn != bpn && cpusets_overlap(a, b)) {
652 for (k = 0; k < csn; k++) {
653 struct cpuset *c = csa[k];
658 ndoms--; /* one less element */
665 * Now we know how many domains to create.
666 * Convert <csn, csa> to <ndoms, doms> and populate cpu masks.
668 doms = alloc_sched_domains(ndoms);
673 * The rest of the code, including the scheduler, can deal with
674 * dattr==NULL case. No need to abort if alloc fails.
676 dattr = kmalloc(ndoms * sizeof(struct sched_domain_attr), GFP_KERNEL);
678 for (nslot = 0, i = 0; i < csn; i++) {
679 struct cpuset *a = csa[i];
684 /* Skip completed partitions */
690 if (nslot == ndoms) {
691 static int warnings = 10;
694 "rebuild_sched_domains confused:"
695 " nslot %d, ndoms %d, csn %d, i %d,"
697 nslot, ndoms, csn, i, apn);
705 *(dattr + nslot) = SD_ATTR_INIT;
706 for (j = i; j < csn; j++) {
707 struct cpuset *b = csa[j];
710 cpumask_or(dp, dp, b->cpus_allowed);
712 update_domain_attr_tree(dattr + nslot, b);
714 /* Done with this partition */
720 BUG_ON(nslot != ndoms);
726 * Fallback to the default domain if kmalloc() failed.
727 * See comments in partition_sched_domains().
738 * Rebuild scheduler domains.
740 * If the flag 'sched_load_balance' of any cpuset with non-empty
741 * 'cpus' changes, or if the 'cpus' allowed changes in any cpuset
742 * which has that flag enabled, or if any cpuset with a non-empty
743 * 'cpus' is removed, then call this routine to rebuild the
744 * scheduler's dynamic sched domains.
746 * Call with cpuset_mutex held. Takes get_online_cpus().
748 static void rebuild_sched_domains_locked(void)
750 struct sched_domain_attr *attr;
754 lockdep_assert_held(&cpuset_mutex);
758 * We have raced with CPU hotplug. Don't do anything to avoid
759 * passing doms with offlined cpu to partition_sched_domains().
760 * Anyways, hotplug work item will rebuild sched domains.
762 if (!cpumask_equal(top_cpuset.cpus_allowed, cpu_active_mask))
765 /* Generate domain masks and attrs */
766 ndoms = generate_sched_domains(&doms, &attr);
768 /* Have scheduler rebuild the domains */
769 partition_sched_domains(ndoms, doms, attr);
773 #else /* !CONFIG_SMP */
774 static void rebuild_sched_domains_locked(void)
777 #endif /* CONFIG_SMP */
779 void rebuild_sched_domains(void)
781 mutex_lock(&cpuset_mutex);
782 rebuild_sched_domains_locked();
783 mutex_unlock(&cpuset_mutex);
787 * cpuset_test_cpumask - test a task's cpus_allowed versus its cpuset's
789 * @scan: struct cgroup_scanner contained in its struct cpuset_hotplug_scanner
791 * Call with cpuset_mutex held. May take callback_mutex during call.
792 * Called for each task in a cgroup by cgroup_scan_tasks().
793 * Return nonzero if this tasks's cpus_allowed mask should be changed (in other
794 * words, if its mask is not equal to its cpuset's mask).
796 static int cpuset_test_cpumask(struct task_struct *tsk,
797 struct cgroup_scanner *scan)
799 return !cpumask_equal(&tsk->cpus_allowed,
800 (cgroup_cs(scan->cg))->cpus_allowed);
804 * cpuset_change_cpumask - make a task's cpus_allowed the same as its cpuset's
806 * @scan: struct cgroup_scanner containing the cgroup of the task
808 * Called by cgroup_scan_tasks() for each task in a cgroup whose
809 * cpus_allowed mask needs to be changed.
811 * We don't need to re-check for the cgroup/cpuset membership, since we're
812 * holding cpuset_mutex at this point.
814 static void cpuset_change_cpumask(struct task_struct *tsk,
815 struct cgroup_scanner *scan)
817 set_cpus_allowed_ptr(tsk, ((cgroup_cs(scan->cg))->cpus_allowed));
821 * update_tasks_cpumask - Update the cpumasks of tasks in the cpuset.
822 * @cs: the cpuset in which each task's cpus_allowed mask needs to be changed
823 * @heap: if NULL, defer allocating heap memory to cgroup_scan_tasks()
825 * Called with cpuset_mutex held
827 * The cgroup_scan_tasks() function will scan all the tasks in a cgroup,
828 * calling callback functions for each.
830 * No return value. It's guaranteed that cgroup_scan_tasks() always returns 0
833 static void update_tasks_cpumask(struct cpuset *cs, struct ptr_heap *heap)
835 struct cgroup_scanner scan;
837 scan.cg = cs->css.cgroup;
838 scan.test_task = cpuset_test_cpumask;
839 scan.process_task = cpuset_change_cpumask;
841 cgroup_scan_tasks(&scan);
845 * update_cpumask - update the cpus_allowed mask of a cpuset and all tasks in it
846 * @cs: the cpuset to consider
847 * @buf: buffer of cpu numbers written to this cpuset
849 static int update_cpumask(struct cpuset *cs, struct cpuset *trialcs,
852 struct ptr_heap heap;
854 int is_load_balanced;
856 /* top_cpuset.cpus_allowed tracks cpu_online_mask; it's read-only */
857 if (cs == &top_cpuset)
861 * An empty cpus_allowed is ok only if the cpuset has no tasks.
862 * Since cpulist_parse() fails on an empty mask, we special case
863 * that parsing. The validate_change() call ensures that cpusets
864 * with tasks have cpus.
867 cpumask_clear(trialcs->cpus_allowed);
869 retval = cpulist_parse(buf, trialcs->cpus_allowed);
873 if (!cpumask_subset(trialcs->cpus_allowed, cpu_active_mask))
876 retval = validate_change(cs, trialcs);
880 /* Nothing to do if the cpus didn't change */
881 if (cpumask_equal(cs->cpus_allowed, trialcs->cpus_allowed))
884 retval = heap_init(&heap, PAGE_SIZE, GFP_KERNEL, NULL);
888 is_load_balanced = is_sched_load_balance(trialcs);
890 mutex_lock(&callback_mutex);
891 cpumask_copy(cs->cpus_allowed, trialcs->cpus_allowed);
892 mutex_unlock(&callback_mutex);
895 * Scan tasks in the cpuset, and update the cpumasks of any
896 * that need an update.
898 update_tasks_cpumask(cs, &heap);
902 if (is_load_balanced)
903 rebuild_sched_domains_locked();
910 * Migrate memory region from one set of nodes to another.
912 * Temporarilly set tasks mems_allowed to target nodes of migration,
913 * so that the migration code can allocate pages on these nodes.
915 * Call holding cpuset_mutex, so current's cpuset won't change
916 * during this call, as manage_mutex holds off any cpuset_attach()
917 * calls. Therefore we don't need to take task_lock around the
918 * call to guarantee_online_mems(), as we know no one is changing
921 * While the mm_struct we are migrating is typically from some
922 * other task, the task_struct mems_allowed that we are hacking
923 * is for our current task, which must allocate new pages for that
924 * migrating memory region.
927 static void cpuset_migrate_mm(struct mm_struct *mm, const nodemask_t *from,
928 const nodemask_t *to)
930 struct task_struct *tsk = current;
932 tsk->mems_allowed = *to;
934 do_migrate_pages(mm, from, to, MPOL_MF_MOVE_ALL);
936 guarantee_online_mems(task_cs(tsk),&tsk->mems_allowed);
940 * cpuset_change_task_nodemask - change task's mems_allowed and mempolicy
941 * @tsk: the task to change
942 * @newmems: new nodes that the task will be set
944 * In order to avoid seeing no nodes if the old and new nodes are disjoint,
945 * we structure updates as setting all new allowed nodes, then clearing newly
948 static void cpuset_change_task_nodemask(struct task_struct *tsk,
954 * Allow tasks that have access to memory reserves because they have
955 * been OOM killed to get memory anywhere.
957 if (unlikely(test_thread_flag(TIF_MEMDIE)))
959 if (current->flags & PF_EXITING) /* Let dying task have memory */
964 * Determine if a loop is necessary if another thread is doing
965 * get_mems_allowed(). If at least one node remains unchanged and
966 * tsk does not have a mempolicy, then an empty nodemask will not be
967 * possible when mems_allowed is larger than a word.
969 need_loop = task_has_mempolicy(tsk) ||
970 !nodes_intersects(*newmems, tsk->mems_allowed);
973 write_seqcount_begin(&tsk->mems_allowed_seq);
975 nodes_or(tsk->mems_allowed, tsk->mems_allowed, *newmems);
976 mpol_rebind_task(tsk, newmems, MPOL_REBIND_STEP1);
978 mpol_rebind_task(tsk, newmems, MPOL_REBIND_STEP2);
979 tsk->mems_allowed = *newmems;
982 write_seqcount_end(&tsk->mems_allowed_seq);
988 * Update task's mems_allowed and rebind its mempolicy and vmas' mempolicy
989 * of it to cpuset's new mems_allowed, and migrate pages to new nodes if
990 * memory_migrate flag is set. Called with cpuset_mutex held.
992 static void cpuset_change_nodemask(struct task_struct *p,
993 struct cgroup_scanner *scan)
995 struct mm_struct *mm;
998 const nodemask_t *oldmem = scan->data;
999 static nodemask_t newmems; /* protected by cpuset_mutex */
1001 cs = cgroup_cs(scan->cg);
1002 guarantee_online_mems(cs, &newmems);
1004 cpuset_change_task_nodemask(p, &newmems);
1006 mm = get_task_mm(p);
1010 migrate = is_memory_migrate(cs);
1012 mpol_rebind_mm(mm, &cs->mems_allowed);
1014 cpuset_migrate_mm(mm, oldmem, &cs->mems_allowed);
1018 static void *cpuset_being_rebound;
1021 * update_tasks_nodemask - Update the nodemasks of tasks in the cpuset.
1022 * @cs: the cpuset in which each task's mems_allowed mask needs to be changed
1023 * @oldmem: old mems_allowed of cpuset cs
1024 * @heap: if NULL, defer allocating heap memory to cgroup_scan_tasks()
1026 * Called with cpuset_mutex held
1027 * No return value. It's guaranteed that cgroup_scan_tasks() always returns 0
1030 static void update_tasks_nodemask(struct cpuset *cs, const nodemask_t *oldmem,
1031 struct ptr_heap *heap)
1033 struct cgroup_scanner scan;
1035 cpuset_being_rebound = cs; /* causes mpol_dup() rebind */
1037 scan.cg = cs->css.cgroup;
1038 scan.test_task = NULL;
1039 scan.process_task = cpuset_change_nodemask;
1041 scan.data = (nodemask_t *)oldmem;
1044 * The mpol_rebind_mm() call takes mmap_sem, which we couldn't
1045 * take while holding tasklist_lock. Forks can happen - the
1046 * mpol_dup() cpuset_being_rebound check will catch such forks,
1047 * and rebind their vma mempolicies too. Because we still hold
1048 * the global cpuset_mutex, we know that no other rebind effort
1049 * will be contending for the global variable cpuset_being_rebound.
1050 * It's ok if we rebind the same mm twice; mpol_rebind_mm()
1051 * is idempotent. Also migrate pages in each mm to new nodes.
1053 cgroup_scan_tasks(&scan);
1055 /* We're done rebinding vmas to this cpuset's new mems_allowed. */
1056 cpuset_being_rebound = NULL;
1060 * Handle user request to change the 'mems' memory placement
1061 * of a cpuset. Needs to validate the request, update the
1062 * cpusets mems_allowed, and for each task in the cpuset,
1063 * update mems_allowed and rebind task's mempolicy and any vma
1064 * mempolicies and if the cpuset is marked 'memory_migrate',
1065 * migrate the tasks pages to the new memory.
1067 * Call with cpuset_mutex held. May take callback_mutex during call.
1068 * Will take tasklist_lock, scan tasklist for tasks in cpuset cs,
1069 * lock each such tasks mm->mmap_sem, scan its vma's and rebind
1070 * their mempolicies to the cpusets new mems_allowed.
1072 static int update_nodemask(struct cpuset *cs, struct cpuset *trialcs,
1075 NODEMASK_ALLOC(nodemask_t, oldmem, GFP_KERNEL);
1077 struct ptr_heap heap;
1083 * top_cpuset.mems_allowed tracks node_stats[N_MEMORY];
1086 if (cs == &top_cpuset) {
1092 * An empty mems_allowed is ok iff there are no tasks in the cpuset.
1093 * Since nodelist_parse() fails on an empty mask, we special case
1094 * that parsing. The validate_change() call ensures that cpusets
1095 * with tasks have memory.
1098 nodes_clear(trialcs->mems_allowed);
1100 retval = nodelist_parse(buf, trialcs->mems_allowed);
1104 if (!nodes_subset(trialcs->mems_allowed,
1105 node_states[N_MEMORY])) {
1110 *oldmem = cs->mems_allowed;
1111 if (nodes_equal(*oldmem, trialcs->mems_allowed)) {
1112 retval = 0; /* Too easy - nothing to do */
1115 retval = validate_change(cs, trialcs);
1119 retval = heap_init(&heap, PAGE_SIZE, GFP_KERNEL, NULL);
1123 mutex_lock(&callback_mutex);
1124 cs->mems_allowed = trialcs->mems_allowed;
1125 mutex_unlock(&callback_mutex);
1127 update_tasks_nodemask(cs, oldmem, &heap);
1131 NODEMASK_FREE(oldmem);
1135 int current_cpuset_is_being_rebound(void)
1137 return task_cs(current) == cpuset_being_rebound;
1140 static int update_relax_domain_level(struct cpuset *cs, s64 val)
1143 if (val < -1 || val >= sched_domain_level_max)
1147 if (val != cs->relax_domain_level) {
1148 cs->relax_domain_level = val;
1149 if (!cpumask_empty(cs->cpus_allowed) &&
1150 is_sched_load_balance(cs))
1151 rebuild_sched_domains_locked();
1158 * cpuset_change_flag - make a task's spread flags the same as its cpuset's
1159 * @tsk: task to be updated
1160 * @scan: struct cgroup_scanner containing the cgroup of the task
1162 * Called by cgroup_scan_tasks() for each task in a cgroup.
1164 * We don't need to re-check for the cgroup/cpuset membership, since we're
1165 * holding cpuset_mutex at this point.
1167 static void cpuset_change_flag(struct task_struct *tsk,
1168 struct cgroup_scanner *scan)
1170 cpuset_update_task_spread_flag(cgroup_cs(scan->cg), tsk);
1174 * update_tasks_flags - update the spread flags of tasks in the cpuset.
1175 * @cs: the cpuset in which each task's spread flags needs to be changed
1176 * @heap: if NULL, defer allocating heap memory to cgroup_scan_tasks()
1178 * Called with cpuset_mutex held
1180 * The cgroup_scan_tasks() function will scan all the tasks in a cgroup,
1181 * calling callback functions for each.
1183 * No return value. It's guaranteed that cgroup_scan_tasks() always returns 0
1186 static void update_tasks_flags(struct cpuset *cs, struct ptr_heap *heap)
1188 struct cgroup_scanner scan;
1190 scan.cg = cs->css.cgroup;
1191 scan.test_task = NULL;
1192 scan.process_task = cpuset_change_flag;
1194 cgroup_scan_tasks(&scan);
1198 * update_flag - read a 0 or a 1 in a file and update associated flag
1199 * bit: the bit to update (see cpuset_flagbits_t)
1200 * cs: the cpuset to update
1201 * turning_on: whether the flag is being set or cleared
1203 * Call with cpuset_mutex held.
1206 static int update_flag(cpuset_flagbits_t bit, struct cpuset *cs,
1209 struct cpuset *trialcs;
1210 int balance_flag_changed;
1211 int spread_flag_changed;
1212 struct ptr_heap heap;
1215 trialcs = alloc_trial_cpuset(cs);
1220 set_bit(bit, &trialcs->flags);
1222 clear_bit(bit, &trialcs->flags);
1224 err = validate_change(cs, trialcs);
1228 err = heap_init(&heap, PAGE_SIZE, GFP_KERNEL, NULL);
1232 balance_flag_changed = (is_sched_load_balance(cs) !=
1233 is_sched_load_balance(trialcs));
1235 spread_flag_changed = ((is_spread_slab(cs) != is_spread_slab(trialcs))
1236 || (is_spread_page(cs) != is_spread_page(trialcs)));
1238 mutex_lock(&callback_mutex);
1239 cs->flags = trialcs->flags;
1240 mutex_unlock(&callback_mutex);
1242 if (!cpumask_empty(trialcs->cpus_allowed) && balance_flag_changed)
1243 rebuild_sched_domains_locked();
1245 if (spread_flag_changed)
1246 update_tasks_flags(cs, &heap);
1249 free_trial_cpuset(trialcs);
1254 * Frequency meter - How fast is some event occurring?
1256 * These routines manage a digitally filtered, constant time based,
1257 * event frequency meter. There are four routines:
1258 * fmeter_init() - initialize a frequency meter.
1259 * fmeter_markevent() - called each time the event happens.
1260 * fmeter_getrate() - returns the recent rate of such events.
1261 * fmeter_update() - internal routine used to update fmeter.
1263 * A common data structure is passed to each of these routines,
1264 * which is used to keep track of the state required to manage the
1265 * frequency meter and its digital filter.
1267 * The filter works on the number of events marked per unit time.
1268 * The filter is single-pole low-pass recursive (IIR). The time unit
1269 * is 1 second. Arithmetic is done using 32-bit integers scaled to
1270 * simulate 3 decimal digits of precision (multiplied by 1000).
1272 * With an FM_COEF of 933, and a time base of 1 second, the filter
1273 * has a half-life of 10 seconds, meaning that if the events quit
1274 * happening, then the rate returned from the fmeter_getrate()
1275 * will be cut in half each 10 seconds, until it converges to zero.
1277 * It is not worth doing a real infinitely recursive filter. If more
1278 * than FM_MAXTICKS ticks have elapsed since the last filter event,
1279 * just compute FM_MAXTICKS ticks worth, by which point the level
1282 * Limit the count of unprocessed events to FM_MAXCNT, so as to avoid
1283 * arithmetic overflow in the fmeter_update() routine.
1285 * Given the simple 32 bit integer arithmetic used, this meter works
1286 * best for reporting rates between one per millisecond (msec) and
1287 * one per 32 (approx) seconds. At constant rates faster than one
1288 * per msec it maxes out at values just under 1,000,000. At constant
1289 * rates between one per msec, and one per second it will stabilize
1290 * to a value N*1000, where N is the rate of events per second.
1291 * At constant rates between one per second and one per 32 seconds,
1292 * it will be choppy, moving up on the seconds that have an event,
1293 * and then decaying until the next event. At rates slower than
1294 * about one in 32 seconds, it decays all the way back to zero between
1298 #define FM_COEF 933 /* coefficient for half-life of 10 secs */
1299 #define FM_MAXTICKS ((time_t)99) /* useless computing more ticks than this */
1300 #define FM_MAXCNT 1000000 /* limit cnt to avoid overflow */
1301 #define FM_SCALE 1000 /* faux fixed point scale */
1303 /* Initialize a frequency meter */
1304 static void fmeter_init(struct fmeter *fmp)
1309 spin_lock_init(&fmp->lock);
1312 /* Internal meter update - process cnt events and update value */
1313 static void fmeter_update(struct fmeter *fmp)
1315 time_t now = get_seconds();
1316 time_t ticks = now - fmp->time;
1321 ticks = min(FM_MAXTICKS, ticks);
1323 fmp->val = (FM_COEF * fmp->val) / FM_SCALE;
1326 fmp->val += ((FM_SCALE - FM_COEF) * fmp->cnt) / FM_SCALE;
1330 /* Process any previous ticks, then bump cnt by one (times scale). */
1331 static void fmeter_markevent(struct fmeter *fmp)
1333 spin_lock(&fmp->lock);
1335 fmp->cnt = min(FM_MAXCNT, fmp->cnt + FM_SCALE);
1336 spin_unlock(&fmp->lock);
1339 /* Process any previous ticks, then return current value. */
1340 static int fmeter_getrate(struct fmeter *fmp)
1344 spin_lock(&fmp->lock);
1347 spin_unlock(&fmp->lock);
1351 /* Called by cgroups to determine if a cpuset is usable; cpuset_mutex held */
1352 static int cpuset_can_attach(struct cgroup *cgrp, struct cgroup_taskset *tset)
1354 struct cpuset *cs = cgroup_cs(cgrp);
1355 struct task_struct *task;
1358 mutex_lock(&cpuset_mutex);
1361 if (cpumask_empty(cs->cpus_allowed) || nodes_empty(cs->mems_allowed))
1364 cgroup_taskset_for_each(task, cgrp, tset) {
1366 * Kthreads which disallow setaffinity shouldn't be moved
1367 * to a new cpuset; we don't want to change their cpu
1368 * affinity and isolating such threads by their set of
1369 * allowed nodes is unnecessary. Thus, cpusets are not
1370 * applicable for such threads. This prevents checking for
1371 * success of set_cpus_allowed_ptr() on all attached tasks
1372 * before cpus_allowed may be changed.
1375 if (task->flags & PF_NO_SETAFFINITY)
1377 ret = security_task_setscheduler(task);
1383 * Mark attach is in progress. This makes validate_change() fail
1384 * changes which zero cpus/mems_allowed.
1386 cs->attach_in_progress++;
1389 mutex_unlock(&cpuset_mutex);
1393 static void cpuset_cancel_attach(struct cgroup *cgrp,
1394 struct cgroup_taskset *tset)
1396 mutex_lock(&cpuset_mutex);
1397 cgroup_cs(cgrp)->attach_in_progress--;
1398 mutex_unlock(&cpuset_mutex);
1402 * Protected by cpuset_mutex. cpus_attach is used only by cpuset_attach()
1403 * but we can't allocate it dynamically there. Define it global and
1404 * allocate from cpuset_init().
1406 static cpumask_var_t cpus_attach;
1408 static void cpuset_attach(struct cgroup *cgrp, struct cgroup_taskset *tset)
1410 /* static bufs protected by cpuset_mutex */
1411 static nodemask_t cpuset_attach_nodemask_from;
1412 static nodemask_t cpuset_attach_nodemask_to;
1413 struct mm_struct *mm;
1414 struct task_struct *task;
1415 struct task_struct *leader = cgroup_taskset_first(tset);
1416 struct cgroup *oldcgrp = cgroup_taskset_cur_cgroup(tset);
1417 struct cpuset *cs = cgroup_cs(cgrp);
1418 struct cpuset *oldcs = cgroup_cs(oldcgrp);
1420 mutex_lock(&cpuset_mutex);
1422 /* prepare for attach */
1423 if (cs == &top_cpuset)
1424 cpumask_copy(cpus_attach, cpu_possible_mask);
1426 guarantee_online_cpus(cs, cpus_attach);
1428 guarantee_online_mems(cs, &cpuset_attach_nodemask_to);
1430 cgroup_taskset_for_each(task, cgrp, tset) {
1432 * can_attach beforehand should guarantee that this doesn't
1433 * fail. TODO: have a better way to handle failure here
1435 WARN_ON_ONCE(set_cpus_allowed_ptr(task, cpus_attach));
1437 cpuset_change_task_nodemask(task, &cpuset_attach_nodemask_to);
1438 cpuset_update_task_spread_flag(cs, task);
1442 * Change mm, possibly for multiple threads in a threadgroup. This is
1443 * expensive and may sleep.
1445 cpuset_attach_nodemask_from = oldcs->mems_allowed;
1446 cpuset_attach_nodemask_to = cs->mems_allowed;
1447 mm = get_task_mm(leader);
1449 mpol_rebind_mm(mm, &cpuset_attach_nodemask_to);
1450 if (is_memory_migrate(cs))
1451 cpuset_migrate_mm(mm, &cpuset_attach_nodemask_from,
1452 &cpuset_attach_nodemask_to);
1456 cs->attach_in_progress--;
1459 * We may have raced with CPU/memory hotunplug. Trigger hotplug
1460 * propagation if @cs doesn't have any CPU or memory. It will move
1461 * the newly added tasks to the nearest parent which can execute.
1463 if (cpumask_empty(cs->cpus_allowed) || nodes_empty(cs->mems_allowed))
1464 schedule_cpuset_propagate_hotplug(cs);
1466 mutex_unlock(&cpuset_mutex);
1469 /* The various types of files and directories in a cpuset file system */
1472 FILE_MEMORY_MIGRATE,
1478 FILE_SCHED_LOAD_BALANCE,
1479 FILE_SCHED_RELAX_DOMAIN_LEVEL,
1480 FILE_MEMORY_PRESSURE_ENABLED,
1481 FILE_MEMORY_PRESSURE,
1484 } cpuset_filetype_t;
1486 static int cpuset_write_u64(struct cgroup *cgrp, struct cftype *cft, u64 val)
1488 struct cpuset *cs = cgroup_cs(cgrp);
1489 cpuset_filetype_t type = cft->private;
1490 int retval = -ENODEV;
1492 mutex_lock(&cpuset_mutex);
1493 if (!is_cpuset_online(cs))
1497 case FILE_CPU_EXCLUSIVE:
1498 retval = update_flag(CS_CPU_EXCLUSIVE, cs, val);
1500 case FILE_MEM_EXCLUSIVE:
1501 retval = update_flag(CS_MEM_EXCLUSIVE, cs, val);
1503 case FILE_MEM_HARDWALL:
1504 retval = update_flag(CS_MEM_HARDWALL, cs, val);
1506 case FILE_SCHED_LOAD_BALANCE:
1507 retval = update_flag(CS_SCHED_LOAD_BALANCE, cs, val);
1509 case FILE_MEMORY_MIGRATE:
1510 retval = update_flag(CS_MEMORY_MIGRATE, cs, val);
1512 case FILE_MEMORY_PRESSURE_ENABLED:
1513 cpuset_memory_pressure_enabled = !!val;
1515 case FILE_MEMORY_PRESSURE:
1518 case FILE_SPREAD_PAGE:
1519 retval = update_flag(CS_SPREAD_PAGE, cs, val);
1521 case FILE_SPREAD_SLAB:
1522 retval = update_flag(CS_SPREAD_SLAB, cs, val);
1529 mutex_unlock(&cpuset_mutex);
1533 static int cpuset_write_s64(struct cgroup *cgrp, struct cftype *cft, s64 val)
1535 struct cpuset *cs = cgroup_cs(cgrp);
1536 cpuset_filetype_t type = cft->private;
1537 int retval = -ENODEV;
1539 mutex_lock(&cpuset_mutex);
1540 if (!is_cpuset_online(cs))
1544 case FILE_SCHED_RELAX_DOMAIN_LEVEL:
1545 retval = update_relax_domain_level(cs, val);
1552 mutex_unlock(&cpuset_mutex);
1557 * Common handling for a write to a "cpus" or "mems" file.
1559 static int cpuset_write_resmask(struct cgroup *cgrp, struct cftype *cft,
1562 struct cpuset *cs = cgroup_cs(cgrp);
1563 struct cpuset *trialcs;
1564 int retval = -ENODEV;
1567 * CPU or memory hotunplug may leave @cs w/o any execution
1568 * resources, in which case the hotplug code asynchronously updates
1569 * configuration and transfers all tasks to the nearest ancestor
1570 * which can execute.
1572 * As writes to "cpus" or "mems" may restore @cs's execution
1573 * resources, wait for the previously scheduled operations before
1574 * proceeding, so that we don't end up keep removing tasks added
1575 * after execution capability is restored.
1577 * Flushing cpuset_hotplug_work is enough to synchronize against
1578 * hotplug hanlding; however, cpuset_attach() may schedule
1579 * propagation work directly. Flush the workqueue too.
1581 flush_work(&cpuset_hotplug_work);
1582 flush_workqueue(cpuset_propagate_hotplug_wq);
1584 mutex_lock(&cpuset_mutex);
1585 if (!is_cpuset_online(cs))
1588 trialcs = alloc_trial_cpuset(cs);
1594 switch (cft->private) {
1596 retval = update_cpumask(cs, trialcs, buf);
1599 retval = update_nodemask(cs, trialcs, buf);
1606 free_trial_cpuset(trialcs);
1608 mutex_unlock(&cpuset_mutex);
1613 * These ascii lists should be read in a single call, by using a user
1614 * buffer large enough to hold the entire map. If read in smaller
1615 * chunks, there is no guarantee of atomicity. Since the display format
1616 * used, list of ranges of sequential numbers, is variable length,
1617 * and since these maps can change value dynamically, one could read
1618 * gibberish by doing partial reads while a list was changing.
1619 * A single large read to a buffer that crosses a page boundary is
1620 * ok, because the result being copied to user land is not recomputed
1621 * across a page fault.
1624 static size_t cpuset_sprintf_cpulist(char *page, struct cpuset *cs)
1628 mutex_lock(&callback_mutex);
1629 count = cpulist_scnprintf(page, PAGE_SIZE, cs->cpus_allowed);
1630 mutex_unlock(&callback_mutex);
1635 static size_t cpuset_sprintf_memlist(char *page, struct cpuset *cs)
1639 mutex_lock(&callback_mutex);
1640 count = nodelist_scnprintf(page, PAGE_SIZE, cs->mems_allowed);
1641 mutex_unlock(&callback_mutex);
1646 static ssize_t cpuset_common_file_read(struct cgroup *cont,
1650 size_t nbytes, loff_t *ppos)
1652 struct cpuset *cs = cgroup_cs(cont);
1653 cpuset_filetype_t type = cft->private;
1658 if (!(page = (char *)__get_free_page(GFP_TEMPORARY)))
1665 s += cpuset_sprintf_cpulist(s, cs);
1668 s += cpuset_sprintf_memlist(s, cs);
1676 retval = simple_read_from_buffer(buf, nbytes, ppos, page, s - page);
1678 free_page((unsigned long)page);
1682 static u64 cpuset_read_u64(struct cgroup *cont, struct cftype *cft)
1684 struct cpuset *cs = cgroup_cs(cont);
1685 cpuset_filetype_t type = cft->private;
1687 case FILE_CPU_EXCLUSIVE:
1688 return is_cpu_exclusive(cs);
1689 case FILE_MEM_EXCLUSIVE:
1690 return is_mem_exclusive(cs);
1691 case FILE_MEM_HARDWALL:
1692 return is_mem_hardwall(cs);
1693 case FILE_SCHED_LOAD_BALANCE:
1694 return is_sched_load_balance(cs);
1695 case FILE_MEMORY_MIGRATE:
1696 return is_memory_migrate(cs);
1697 case FILE_MEMORY_PRESSURE_ENABLED:
1698 return cpuset_memory_pressure_enabled;
1699 case FILE_MEMORY_PRESSURE:
1700 return fmeter_getrate(&cs->fmeter);
1701 case FILE_SPREAD_PAGE:
1702 return is_spread_page(cs);
1703 case FILE_SPREAD_SLAB:
1704 return is_spread_slab(cs);
1709 /* Unreachable but makes gcc happy */
1713 static s64 cpuset_read_s64(struct cgroup *cont, struct cftype *cft)
1715 struct cpuset *cs = cgroup_cs(cont);
1716 cpuset_filetype_t type = cft->private;
1718 case FILE_SCHED_RELAX_DOMAIN_LEVEL:
1719 return cs->relax_domain_level;
1724 /* Unrechable but makes gcc happy */
1730 * for the common functions, 'private' gives the type of file
1733 static struct cftype files[] = {
1736 .read = cpuset_common_file_read,
1737 .write_string = cpuset_write_resmask,
1738 .max_write_len = (100U + 6 * NR_CPUS),
1739 .private = FILE_CPULIST,
1744 .read = cpuset_common_file_read,
1745 .write_string = cpuset_write_resmask,
1746 .max_write_len = (100U + 6 * MAX_NUMNODES),
1747 .private = FILE_MEMLIST,
1751 .name = "cpu_exclusive",
1752 .read_u64 = cpuset_read_u64,
1753 .write_u64 = cpuset_write_u64,
1754 .private = FILE_CPU_EXCLUSIVE,
1758 .name = "mem_exclusive",
1759 .read_u64 = cpuset_read_u64,
1760 .write_u64 = cpuset_write_u64,
1761 .private = FILE_MEM_EXCLUSIVE,
1765 .name = "mem_hardwall",
1766 .read_u64 = cpuset_read_u64,
1767 .write_u64 = cpuset_write_u64,
1768 .private = FILE_MEM_HARDWALL,
1772 .name = "sched_load_balance",
1773 .read_u64 = cpuset_read_u64,
1774 .write_u64 = cpuset_write_u64,
1775 .private = FILE_SCHED_LOAD_BALANCE,
1779 .name = "sched_relax_domain_level",
1780 .read_s64 = cpuset_read_s64,
1781 .write_s64 = cpuset_write_s64,
1782 .private = FILE_SCHED_RELAX_DOMAIN_LEVEL,
1786 .name = "memory_migrate",
1787 .read_u64 = cpuset_read_u64,
1788 .write_u64 = cpuset_write_u64,
1789 .private = FILE_MEMORY_MIGRATE,
1793 .name = "memory_pressure",
1794 .read_u64 = cpuset_read_u64,
1795 .write_u64 = cpuset_write_u64,
1796 .private = FILE_MEMORY_PRESSURE,
1801 .name = "memory_spread_page",
1802 .read_u64 = cpuset_read_u64,
1803 .write_u64 = cpuset_write_u64,
1804 .private = FILE_SPREAD_PAGE,
1808 .name = "memory_spread_slab",
1809 .read_u64 = cpuset_read_u64,
1810 .write_u64 = cpuset_write_u64,
1811 .private = FILE_SPREAD_SLAB,
1815 .name = "memory_pressure_enabled",
1816 .flags = CFTYPE_ONLY_ON_ROOT,
1817 .read_u64 = cpuset_read_u64,
1818 .write_u64 = cpuset_write_u64,
1819 .private = FILE_MEMORY_PRESSURE_ENABLED,
1826 * cpuset_css_alloc - allocate a cpuset css
1827 * cont: control group that the new cpuset will be part of
1830 static struct cgroup_subsys_state *cpuset_css_alloc(struct cgroup *cont)
1835 return &top_cpuset.css;
1837 cs = kzalloc(sizeof(*cs), GFP_KERNEL);
1839 return ERR_PTR(-ENOMEM);
1840 if (!alloc_cpumask_var(&cs->cpus_allowed, GFP_KERNEL)) {
1842 return ERR_PTR(-ENOMEM);
1845 set_bit(CS_SCHED_LOAD_BALANCE, &cs->flags);
1846 cpumask_clear(cs->cpus_allowed);
1847 nodes_clear(cs->mems_allowed);
1848 fmeter_init(&cs->fmeter);
1849 INIT_WORK(&cs->hotplug_work, cpuset_propagate_hotplug_workfn);
1850 cs->relax_domain_level = -1;
1855 static int cpuset_css_online(struct cgroup *cgrp)
1857 struct cpuset *cs = cgroup_cs(cgrp);
1858 struct cpuset *parent = parent_cs(cs);
1859 struct cpuset *tmp_cs;
1860 struct cgroup *pos_cg;
1865 mutex_lock(&cpuset_mutex);
1867 set_bit(CS_ONLINE, &cs->flags);
1868 if (is_spread_page(parent))
1869 set_bit(CS_SPREAD_PAGE, &cs->flags);
1870 if (is_spread_slab(parent))
1871 set_bit(CS_SPREAD_SLAB, &cs->flags);
1873 number_of_cpusets++;
1875 if (!test_bit(CGRP_CPUSET_CLONE_CHILDREN, &cgrp->flags))
1879 * Clone @parent's configuration if CGRP_CPUSET_CLONE_CHILDREN is
1880 * set. This flag handling is implemented in cgroup core for
1881 * histrical reasons - the flag may be specified during mount.
1883 * Currently, if any sibling cpusets have exclusive cpus or mem, we
1884 * refuse to clone the configuration - thereby refusing the task to
1885 * be entered, and as a result refusing the sys_unshare() or
1886 * clone() which initiated it. If this becomes a problem for some
1887 * users who wish to allow that scenario, then this could be
1888 * changed to grant parent->cpus_allowed-sibling_cpus_exclusive
1889 * (and likewise for mems) to the new cgroup.
1892 cpuset_for_each_child(tmp_cs, pos_cg, parent) {
1893 if (is_mem_exclusive(tmp_cs) || is_cpu_exclusive(tmp_cs)) {
1900 mutex_lock(&callback_mutex);
1901 cs->mems_allowed = parent->mems_allowed;
1902 cpumask_copy(cs->cpus_allowed, parent->cpus_allowed);
1903 mutex_unlock(&callback_mutex);
1905 mutex_unlock(&cpuset_mutex);
1909 static void cpuset_css_offline(struct cgroup *cgrp)
1911 struct cpuset *cs = cgroup_cs(cgrp);
1913 mutex_lock(&cpuset_mutex);
1915 if (is_sched_load_balance(cs))
1916 update_flag(CS_SCHED_LOAD_BALANCE, cs, 0);
1918 number_of_cpusets--;
1919 clear_bit(CS_ONLINE, &cs->flags);
1921 mutex_unlock(&cpuset_mutex);
1925 * If the cpuset being removed has its flag 'sched_load_balance'
1926 * enabled, then simulate turning sched_load_balance off, which
1927 * will call rebuild_sched_domains_locked().
1930 static void cpuset_css_free(struct cgroup *cont)
1932 struct cpuset *cs = cgroup_cs(cont);
1934 free_cpumask_var(cs->cpus_allowed);
1938 struct cgroup_subsys cpuset_subsys = {
1940 .css_alloc = cpuset_css_alloc,
1941 .css_online = cpuset_css_online,
1942 .css_offline = cpuset_css_offline,
1943 .css_free = cpuset_css_free,
1944 .can_attach = cpuset_can_attach,
1945 .cancel_attach = cpuset_cancel_attach,
1946 .attach = cpuset_attach,
1947 .subsys_id = cpuset_subsys_id,
1948 .base_cftypes = files,
1953 * cpuset_init - initialize cpusets at system boot
1955 * Description: Initialize top_cpuset and the cpuset internal file system,
1958 int __init cpuset_init(void)
1962 if (!alloc_cpumask_var(&top_cpuset.cpus_allowed, GFP_KERNEL))
1965 cpumask_setall(top_cpuset.cpus_allowed);
1966 nodes_setall(top_cpuset.mems_allowed);
1968 fmeter_init(&top_cpuset.fmeter);
1969 set_bit(CS_SCHED_LOAD_BALANCE, &top_cpuset.flags);
1970 top_cpuset.relax_domain_level = -1;
1972 err = register_filesystem(&cpuset_fs_type);
1976 if (!alloc_cpumask_var(&cpus_attach, GFP_KERNEL))
1979 number_of_cpusets = 1;
1984 * If CPU and/or memory hotplug handlers, below, unplug any CPUs
1985 * or memory nodes, we need to walk over the cpuset hierarchy,
1986 * removing that CPU or node from all cpusets. If this removes the
1987 * last CPU or node from a cpuset, then move the tasks in the empty
1988 * cpuset to its next-highest non-empty parent.
1990 static void remove_tasks_in_empty_cpuset(struct cpuset *cs)
1992 struct cpuset *parent;
1995 * Find its next-highest non-empty parent, (top cpuset
1996 * has online cpus, so can't be empty).
1998 parent = parent_cs(cs);
1999 while (cpumask_empty(parent->cpus_allowed) ||
2000 nodes_empty(parent->mems_allowed))
2001 parent = parent_cs(parent);
2003 if (cgroup_transfer_tasks(parent->css.cgroup, cs->css.cgroup)) {
2005 printk(KERN_ERR "cpuset: failed to transfer tasks out of empty cpuset %s\n",
2006 cgroup_name(cs->css.cgroup));
2012 * cpuset_propagate_hotplug_workfn - propagate CPU/memory hotplug to a cpuset
2013 * @cs: cpuset in interest
2015 * Compare @cs's cpu and mem masks against top_cpuset and if some have gone
2016 * offline, update @cs accordingly. If @cs ends up with no CPU or memory,
2017 * all its tasks are moved to the nearest ancestor with both resources.
2019 static void cpuset_propagate_hotplug_workfn(struct work_struct *work)
2021 static cpumask_t off_cpus;
2022 static nodemask_t off_mems, tmp_mems;
2023 struct cpuset *cs = container_of(work, struct cpuset, hotplug_work);
2026 mutex_lock(&cpuset_mutex);
2028 cpumask_andnot(&off_cpus, cs->cpus_allowed, top_cpuset.cpus_allowed);
2029 nodes_andnot(off_mems, cs->mems_allowed, top_cpuset.mems_allowed);
2031 /* remove offline cpus from @cs */
2032 if (!cpumask_empty(&off_cpus)) {
2033 mutex_lock(&callback_mutex);
2034 cpumask_andnot(cs->cpus_allowed, cs->cpus_allowed, &off_cpus);
2035 mutex_unlock(&callback_mutex);
2036 update_tasks_cpumask(cs, NULL);
2039 /* remove offline mems from @cs */
2040 if (!nodes_empty(off_mems)) {
2041 tmp_mems = cs->mems_allowed;
2042 mutex_lock(&callback_mutex);
2043 nodes_andnot(cs->mems_allowed, cs->mems_allowed, off_mems);
2044 mutex_unlock(&callback_mutex);
2045 update_tasks_nodemask(cs, &tmp_mems, NULL);
2048 is_empty = cpumask_empty(cs->cpus_allowed) ||
2049 nodes_empty(cs->mems_allowed);
2051 mutex_unlock(&cpuset_mutex);
2054 * If @cs became empty, move tasks to the nearest ancestor with
2055 * execution resources. This is full cgroup operation which will
2056 * also call back into cpuset. Should be done outside any lock.
2059 remove_tasks_in_empty_cpuset(cs);
2061 /* the following may free @cs, should be the last operation */
2066 * schedule_cpuset_propagate_hotplug - schedule hotplug propagation to a cpuset
2067 * @cs: cpuset of interest
2069 * Schedule cpuset_propagate_hotplug_workfn() which will update CPU and
2070 * memory masks according to top_cpuset.
2072 static void schedule_cpuset_propagate_hotplug(struct cpuset *cs)
2075 * Pin @cs. The refcnt will be released when the work item
2076 * finishes executing.
2078 if (!css_tryget(&cs->css))
2082 * Queue @cs->hotplug_work. If already pending, lose the css ref.
2083 * cpuset_propagate_hotplug_wq is ordered and propagation will
2084 * happen in the order this function is called.
2086 if (!queue_work(cpuset_propagate_hotplug_wq, &cs->hotplug_work))
2091 * cpuset_hotplug_workfn - handle CPU/memory hotunplug for a cpuset
2093 * This function is called after either CPU or memory configuration has
2094 * changed and updates cpuset accordingly. The top_cpuset is always
2095 * synchronized to cpu_active_mask and N_MEMORY, which is necessary in
2096 * order to make cpusets transparent (of no affect) on systems that are
2097 * actively using CPU hotplug but making no active use of cpusets.
2099 * Non-root cpusets are only affected by offlining. If any CPUs or memory
2100 * nodes have been taken down, cpuset_propagate_hotplug() is invoked on all
2103 * Note that CPU offlining during suspend is ignored. We don't modify
2104 * cpusets across suspend/resume cycles at all.
2106 static void cpuset_hotplug_workfn(struct work_struct *work)
2108 static cpumask_t new_cpus, tmp_cpus;
2109 static nodemask_t new_mems, tmp_mems;
2110 bool cpus_updated, mems_updated;
2111 bool cpus_offlined, mems_offlined;
2113 mutex_lock(&cpuset_mutex);
2115 /* fetch the available cpus/mems and find out which changed how */
2116 cpumask_copy(&new_cpus, cpu_active_mask);
2117 new_mems = node_states[N_MEMORY];
2119 cpus_updated = !cpumask_equal(top_cpuset.cpus_allowed, &new_cpus);
2120 cpus_offlined = cpumask_andnot(&tmp_cpus, top_cpuset.cpus_allowed,
2123 mems_updated = !nodes_equal(top_cpuset.mems_allowed, new_mems);
2124 nodes_andnot(tmp_mems, top_cpuset.mems_allowed, new_mems);
2125 mems_offlined = !nodes_empty(tmp_mems);
2127 /* synchronize cpus_allowed to cpu_active_mask */
2129 mutex_lock(&callback_mutex);
2130 cpumask_copy(top_cpuset.cpus_allowed, &new_cpus);
2131 mutex_unlock(&callback_mutex);
2132 /* we don't mess with cpumasks of tasks in top_cpuset */
2135 /* synchronize mems_allowed to N_MEMORY */
2137 tmp_mems = top_cpuset.mems_allowed;
2138 mutex_lock(&callback_mutex);
2139 top_cpuset.mems_allowed = new_mems;
2140 mutex_unlock(&callback_mutex);
2141 update_tasks_nodemask(&top_cpuset, &tmp_mems, NULL);
2144 /* if cpus or mems went down, we need to propagate to descendants */
2145 if (cpus_offlined || mems_offlined) {
2147 struct cgroup *pos_cgrp;
2150 cpuset_for_each_descendant_pre(cs, pos_cgrp, &top_cpuset)
2151 schedule_cpuset_propagate_hotplug(cs);
2155 mutex_unlock(&cpuset_mutex);
2157 /* wait for propagations to finish */
2158 flush_workqueue(cpuset_propagate_hotplug_wq);
2160 /* rebuild sched domains if cpus_allowed has changed */
2162 rebuild_sched_domains();
2165 void cpuset_update_active_cpus(bool cpu_online)
2168 * We're inside cpu hotplug critical region which usually nests
2169 * inside cgroup synchronization. Bounce actual hotplug processing
2170 * to a work item to avoid reverse locking order.
2172 * We still need to do partition_sched_domains() synchronously;
2173 * otherwise, the scheduler will get confused and put tasks to the
2174 * dead CPU. Fall back to the default single domain.
2175 * cpuset_hotplug_workfn() will rebuild it as necessary.
2177 partition_sched_domains(1, NULL, NULL);
2178 schedule_work(&cpuset_hotplug_work);
2182 * Keep top_cpuset.mems_allowed tracking node_states[N_MEMORY].
2183 * Call this routine anytime after node_states[N_MEMORY] changes.
2184 * See cpuset_update_active_cpus() for CPU hotplug handling.
2186 static int cpuset_track_online_nodes(struct notifier_block *self,
2187 unsigned long action, void *arg)
2189 schedule_work(&cpuset_hotplug_work);
2193 static struct notifier_block cpuset_track_online_nodes_nb = {
2194 .notifier_call = cpuset_track_online_nodes,
2195 .priority = 10, /* ??! */
2199 * cpuset_init_smp - initialize cpus_allowed
2201 * Description: Finish top cpuset after cpu, node maps are initialized
2203 void __init cpuset_init_smp(void)
2205 cpumask_copy(top_cpuset.cpus_allowed, cpu_active_mask);
2206 top_cpuset.mems_allowed = node_states[N_MEMORY];
2208 register_hotmemory_notifier(&cpuset_track_online_nodes_nb);
2210 cpuset_propagate_hotplug_wq =
2211 alloc_ordered_workqueue("cpuset_hotplug", 0);
2212 BUG_ON(!cpuset_propagate_hotplug_wq);
2216 * cpuset_cpus_allowed - return cpus_allowed mask from a tasks cpuset.
2217 * @tsk: pointer to task_struct from which to obtain cpuset->cpus_allowed.
2218 * @pmask: pointer to struct cpumask variable to receive cpus_allowed set.
2220 * Description: Returns the cpumask_var_t cpus_allowed of the cpuset
2221 * attached to the specified @tsk. Guaranteed to return some non-empty
2222 * subset of cpu_online_mask, even if this means going outside the
2226 void cpuset_cpus_allowed(struct task_struct *tsk, struct cpumask *pmask)
2228 mutex_lock(&callback_mutex);
2230 guarantee_online_cpus(task_cs(tsk), pmask);
2232 mutex_unlock(&callback_mutex);
2235 void cpuset_cpus_allowed_fallback(struct task_struct *tsk)
2237 const struct cpuset *cs;
2241 do_set_cpus_allowed(tsk, cs->cpus_allowed);
2245 * We own tsk->cpus_allowed, nobody can change it under us.
2247 * But we used cs && cs->cpus_allowed lockless and thus can
2248 * race with cgroup_attach_task() or update_cpumask() and get
2249 * the wrong tsk->cpus_allowed. However, both cases imply the
2250 * subsequent cpuset_change_cpumask()->set_cpus_allowed_ptr()
2251 * which takes task_rq_lock().
2253 * If we are called after it dropped the lock we must see all
2254 * changes in tsk_cs()->cpus_allowed. Otherwise we can temporary
2255 * set any mask even if it is not right from task_cs() pov,
2256 * the pending set_cpus_allowed_ptr() will fix things.
2258 * select_fallback_rq() will fix things ups and set cpu_possible_mask
2263 void cpuset_init_current_mems_allowed(void)
2265 nodes_setall(current->mems_allowed);
2269 * cpuset_mems_allowed - return mems_allowed mask from a tasks cpuset.
2270 * @tsk: pointer to task_struct from which to obtain cpuset->mems_allowed.
2272 * Description: Returns the nodemask_t mems_allowed of the cpuset
2273 * attached to the specified @tsk. Guaranteed to return some non-empty
2274 * subset of node_states[N_MEMORY], even if this means going outside the
2278 nodemask_t cpuset_mems_allowed(struct task_struct *tsk)
2282 mutex_lock(&callback_mutex);
2284 guarantee_online_mems(task_cs(tsk), &mask);
2286 mutex_unlock(&callback_mutex);
2292 * cpuset_nodemask_valid_mems_allowed - check nodemask vs. curremt mems_allowed
2293 * @nodemask: the nodemask to be checked
2295 * Are any of the nodes in the nodemask allowed in current->mems_allowed?
2297 int cpuset_nodemask_valid_mems_allowed(nodemask_t *nodemask)
2299 return nodes_intersects(*nodemask, current->mems_allowed);
2303 * nearest_hardwall_ancestor() - Returns the nearest mem_exclusive or
2304 * mem_hardwall ancestor to the specified cpuset. Call holding
2305 * callback_mutex. If no ancestor is mem_exclusive or mem_hardwall
2306 * (an unusual configuration), then returns the root cpuset.
2308 static const struct cpuset *nearest_hardwall_ancestor(const struct cpuset *cs)
2310 while (!(is_mem_exclusive(cs) || is_mem_hardwall(cs)) && parent_cs(cs))
2316 * cpuset_node_allowed_softwall - Can we allocate on a memory node?
2317 * @node: is this an allowed node?
2318 * @gfp_mask: memory allocation flags
2320 * If we're in interrupt, yes, we can always allocate. If __GFP_THISNODE is
2321 * set, yes, we can always allocate. If node is in our task's mems_allowed,
2322 * yes. If it's not a __GFP_HARDWALL request and this node is in the nearest
2323 * hardwalled cpuset ancestor to this task's cpuset, yes. If the task has been
2324 * OOM killed and has access to memory reserves as specified by the TIF_MEMDIE
2328 * If __GFP_HARDWALL is set, cpuset_node_allowed_softwall() reduces to
2329 * cpuset_node_allowed_hardwall(). Otherwise, cpuset_node_allowed_softwall()
2330 * might sleep, and might allow a node from an enclosing cpuset.
2332 * cpuset_node_allowed_hardwall() only handles the simpler case of hardwall
2333 * cpusets, and never sleeps.
2335 * The __GFP_THISNODE placement logic is really handled elsewhere,
2336 * by forcibly using a zonelist starting at a specified node, and by
2337 * (in get_page_from_freelist()) refusing to consider the zones for
2338 * any node on the zonelist except the first. By the time any such
2339 * calls get to this routine, we should just shut up and say 'yes'.
2341 * GFP_USER allocations are marked with the __GFP_HARDWALL bit,
2342 * and do not allow allocations outside the current tasks cpuset
2343 * unless the task has been OOM killed as is marked TIF_MEMDIE.
2344 * GFP_KERNEL allocations are not so marked, so can escape to the
2345 * nearest enclosing hardwalled ancestor cpuset.
2347 * Scanning up parent cpusets requires callback_mutex. The
2348 * __alloc_pages() routine only calls here with __GFP_HARDWALL bit
2349 * _not_ set if it's a GFP_KERNEL allocation, and all nodes in the
2350 * current tasks mems_allowed came up empty on the first pass over
2351 * the zonelist. So only GFP_KERNEL allocations, if all nodes in the
2352 * cpuset are short of memory, might require taking the callback_mutex
2355 * The first call here from mm/page_alloc:get_page_from_freelist()
2356 * has __GFP_HARDWALL set in gfp_mask, enforcing hardwall cpusets,
2357 * so no allocation on a node outside the cpuset is allowed (unless
2358 * in interrupt, of course).
2360 * The second pass through get_page_from_freelist() doesn't even call
2361 * here for GFP_ATOMIC calls. For those calls, the __alloc_pages()
2362 * variable 'wait' is not set, and the bit ALLOC_CPUSET is not set
2363 * in alloc_flags. That logic and the checks below have the combined
2365 * in_interrupt - any node ok (current task context irrelevant)
2366 * GFP_ATOMIC - any node ok
2367 * TIF_MEMDIE - any node ok
2368 * GFP_KERNEL - any node in enclosing hardwalled cpuset ok
2369 * GFP_USER - only nodes in current tasks mems allowed ok.
2372 * Don't call cpuset_node_allowed_softwall if you can't sleep, unless you
2373 * pass in the __GFP_HARDWALL flag set in gfp_flag, which disables
2374 * the code that might scan up ancestor cpusets and sleep.
2376 int __cpuset_node_allowed_softwall(int node, gfp_t gfp_mask)
2378 const struct cpuset *cs; /* current cpuset ancestors */
2379 int allowed; /* is allocation in zone z allowed? */
2381 if (in_interrupt() || (gfp_mask & __GFP_THISNODE))
2383 might_sleep_if(!(gfp_mask & __GFP_HARDWALL));
2384 if (node_isset(node, current->mems_allowed))
2387 * Allow tasks that have access to memory reserves because they have
2388 * been OOM killed to get memory anywhere.
2390 if (unlikely(test_thread_flag(TIF_MEMDIE)))
2392 if (gfp_mask & __GFP_HARDWALL) /* If hardwall request, stop here */
2395 if (current->flags & PF_EXITING) /* Let dying task have memory */
2398 /* Not hardwall and node outside mems_allowed: scan up cpusets */
2399 mutex_lock(&callback_mutex);
2402 cs = nearest_hardwall_ancestor(task_cs(current));
2403 task_unlock(current);
2405 allowed = node_isset(node, cs->mems_allowed);
2406 mutex_unlock(&callback_mutex);
2411 * cpuset_node_allowed_hardwall - Can we allocate on a memory node?
2412 * @node: is this an allowed node?
2413 * @gfp_mask: memory allocation flags
2415 * If we're in interrupt, yes, we can always allocate. If __GFP_THISNODE is
2416 * set, yes, we can always allocate. If node is in our task's mems_allowed,
2417 * yes. If the task has been OOM killed and has access to memory reserves as
2418 * specified by the TIF_MEMDIE flag, yes.
2421 * The __GFP_THISNODE placement logic is really handled elsewhere,
2422 * by forcibly using a zonelist starting at a specified node, and by
2423 * (in get_page_from_freelist()) refusing to consider the zones for
2424 * any node on the zonelist except the first. By the time any such
2425 * calls get to this routine, we should just shut up and say 'yes'.
2427 * Unlike the cpuset_node_allowed_softwall() variant, above,
2428 * this variant requires that the node be in the current task's
2429 * mems_allowed or that we're in interrupt. It does not scan up the
2430 * cpuset hierarchy for the nearest enclosing mem_exclusive cpuset.
2433 int __cpuset_node_allowed_hardwall(int node, gfp_t gfp_mask)
2435 if (in_interrupt() || (gfp_mask & __GFP_THISNODE))
2437 if (node_isset(node, current->mems_allowed))
2440 * Allow tasks that have access to memory reserves because they have
2441 * been OOM killed to get memory anywhere.
2443 if (unlikely(test_thread_flag(TIF_MEMDIE)))
2449 * cpuset_mem_spread_node() - On which node to begin search for a file page
2450 * cpuset_slab_spread_node() - On which node to begin search for a slab page
2452 * If a task is marked PF_SPREAD_PAGE or PF_SPREAD_SLAB (as for
2453 * tasks in a cpuset with is_spread_page or is_spread_slab set),
2454 * and if the memory allocation used cpuset_mem_spread_node()
2455 * to determine on which node to start looking, as it will for
2456 * certain page cache or slab cache pages such as used for file
2457 * system buffers and inode caches, then instead of starting on the
2458 * local node to look for a free page, rather spread the starting
2459 * node around the tasks mems_allowed nodes.
2461 * We don't have to worry about the returned node being offline
2462 * because "it can't happen", and even if it did, it would be ok.
2464 * The routines calling guarantee_online_mems() are careful to
2465 * only set nodes in task->mems_allowed that are online. So it
2466 * should not be possible for the following code to return an
2467 * offline node. But if it did, that would be ok, as this routine
2468 * is not returning the node where the allocation must be, only
2469 * the node where the search should start. The zonelist passed to
2470 * __alloc_pages() will include all nodes. If the slab allocator
2471 * is passed an offline node, it will fall back to the local node.
2472 * See kmem_cache_alloc_node().
2475 static int cpuset_spread_node(int *rotor)
2479 node = next_node(*rotor, current->mems_allowed);
2480 if (node == MAX_NUMNODES)
2481 node = first_node(current->mems_allowed);
2486 int cpuset_mem_spread_node(void)
2488 if (current->cpuset_mem_spread_rotor == NUMA_NO_NODE)
2489 current->cpuset_mem_spread_rotor =
2490 node_random(¤t->mems_allowed);
2492 return cpuset_spread_node(¤t->cpuset_mem_spread_rotor);
2495 int cpuset_slab_spread_node(void)
2497 if (current->cpuset_slab_spread_rotor == NUMA_NO_NODE)
2498 current->cpuset_slab_spread_rotor =
2499 node_random(¤t->mems_allowed);
2501 return cpuset_spread_node(¤t->cpuset_slab_spread_rotor);
2504 EXPORT_SYMBOL_GPL(cpuset_mem_spread_node);
2507 * cpuset_mems_allowed_intersects - Does @tsk1's mems_allowed intersect @tsk2's?
2508 * @tsk1: pointer to task_struct of some task.
2509 * @tsk2: pointer to task_struct of some other task.
2511 * Description: Return true if @tsk1's mems_allowed intersects the
2512 * mems_allowed of @tsk2. Used by the OOM killer to determine if
2513 * one of the task's memory usage might impact the memory available
2517 int cpuset_mems_allowed_intersects(const struct task_struct *tsk1,
2518 const struct task_struct *tsk2)
2520 return nodes_intersects(tsk1->mems_allowed, tsk2->mems_allowed);
2523 #define CPUSET_NODELIST_LEN (256)
2526 * cpuset_print_task_mems_allowed - prints task's cpuset and mems_allowed
2527 * @task: pointer to task_struct of some task.
2529 * Description: Prints @task's name, cpuset name, and cached copy of its
2530 * mems_allowed to the kernel log. Must hold task_lock(task) to allow
2531 * dereferencing task_cs(task).
2533 void cpuset_print_task_mems_allowed(struct task_struct *tsk)
2535 /* Statically allocated to prevent using excess stack. */
2536 static char cpuset_nodelist[CPUSET_NODELIST_LEN];
2537 static DEFINE_SPINLOCK(cpuset_buffer_lock);
2539 struct cgroup *cgrp = task_cs(tsk)->css.cgroup;
2542 spin_lock(&cpuset_buffer_lock);
2544 nodelist_scnprintf(cpuset_nodelist, CPUSET_NODELIST_LEN,
2546 printk(KERN_INFO "%s cpuset=%s mems_allowed=%s\n",
2547 tsk->comm, cgroup_name(cgrp), cpuset_nodelist);
2549 spin_unlock(&cpuset_buffer_lock);
2554 * Collection of memory_pressure is suppressed unless
2555 * this flag is enabled by writing "1" to the special
2556 * cpuset file 'memory_pressure_enabled' in the root cpuset.
2559 int cpuset_memory_pressure_enabled __read_mostly;
2562 * cpuset_memory_pressure_bump - keep stats of per-cpuset reclaims.
2564 * Keep a running average of the rate of synchronous (direct)
2565 * page reclaim efforts initiated by tasks in each cpuset.
2567 * This represents the rate at which some task in the cpuset
2568 * ran low on memory on all nodes it was allowed to use, and
2569 * had to enter the kernels page reclaim code in an effort to
2570 * create more free memory by tossing clean pages or swapping
2571 * or writing dirty pages.
2573 * Display to user space in the per-cpuset read-only file
2574 * "memory_pressure". Value displayed is an integer
2575 * representing the recent rate of entry into the synchronous
2576 * (direct) page reclaim by any task attached to the cpuset.
2579 void __cpuset_memory_pressure_bump(void)
2582 fmeter_markevent(&task_cs(current)->fmeter);
2583 task_unlock(current);
2586 #ifdef CONFIG_PROC_PID_CPUSET
2588 * proc_cpuset_show()
2589 * - Print tasks cpuset path into seq_file.
2590 * - Used for /proc/<pid>/cpuset.
2591 * - No need to task_lock(tsk) on this tsk->cpuset reference, as it
2592 * doesn't really matter if tsk->cpuset changes after we read it,
2593 * and we take cpuset_mutex, keeping cpuset_attach() from changing it
2596 int proc_cpuset_show(struct seq_file *m, void *unused_v)
2599 struct task_struct *tsk;
2601 struct cgroup_subsys_state *css;
2605 buf = kmalloc(PAGE_SIZE, GFP_KERNEL);
2611 tsk = get_pid_task(pid, PIDTYPE_PID);
2616 css = task_subsys_state(tsk, cpuset_subsys_id);
2617 retval = cgroup_path(css->cgroup, buf, PAGE_SIZE);
2624 put_task_struct(tsk);
2630 #endif /* CONFIG_PROC_PID_CPUSET */
2632 /* Display task mems_allowed in /proc/<pid>/status file. */
2633 void cpuset_task_status_allowed(struct seq_file *m, struct task_struct *task)
2635 seq_printf(m, "Mems_allowed:\t");
2636 seq_nodemask(m, &task->mems_allowed);
2637 seq_printf(m, "\n");
2638 seq_printf(m, "Mems_allowed_list:\t");
2639 seq_nodemask_list(m, &task->mems_allowed);
2640 seq_printf(m, "\n");