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_change_cpumask - make a task's cpus_allowed the same as its cpuset's
789 * @scan: struct cgroup_scanner containing the cgroup of the task
791 * Called by cgroup_scan_tasks() for each task in a cgroup whose
792 * cpus_allowed mask needs to be changed.
794 * We don't need to re-check for the cgroup/cpuset membership, since we're
795 * holding cpuset_mutex at this point.
797 static void cpuset_change_cpumask(struct task_struct *tsk,
798 struct cgroup_scanner *scan)
800 set_cpus_allowed_ptr(tsk, ((cgroup_cs(scan->cg))->cpus_allowed));
804 * update_tasks_cpumask - Update the cpumasks of tasks in the cpuset.
805 * @cs: the cpuset in which each task's cpus_allowed mask needs to be changed
806 * @heap: if NULL, defer allocating heap memory to cgroup_scan_tasks()
808 * Called with cpuset_mutex held
810 * The cgroup_scan_tasks() function will scan all the tasks in a cgroup,
811 * calling callback functions for each.
813 * No return value. It's guaranteed that cgroup_scan_tasks() always returns 0
816 static void update_tasks_cpumask(struct cpuset *cs, struct ptr_heap *heap)
818 struct cgroup_scanner scan;
820 scan.cg = cs->css.cgroup;
821 scan.test_task = NULL;
822 scan.process_task = cpuset_change_cpumask;
824 cgroup_scan_tasks(&scan);
828 * update_cpumask - update the cpus_allowed mask of a cpuset and all tasks in it
829 * @cs: the cpuset to consider
830 * @buf: buffer of cpu numbers written to this cpuset
832 static int update_cpumask(struct cpuset *cs, struct cpuset *trialcs,
835 struct ptr_heap heap;
837 int is_load_balanced;
839 /* top_cpuset.cpus_allowed tracks cpu_online_mask; it's read-only */
840 if (cs == &top_cpuset)
844 * An empty cpus_allowed is ok only if the cpuset has no tasks.
845 * Since cpulist_parse() fails on an empty mask, we special case
846 * that parsing. The validate_change() call ensures that cpusets
847 * with tasks have cpus.
850 cpumask_clear(trialcs->cpus_allowed);
852 retval = cpulist_parse(buf, trialcs->cpus_allowed);
856 if (!cpumask_subset(trialcs->cpus_allowed, cpu_active_mask))
859 retval = validate_change(cs, trialcs);
863 /* Nothing to do if the cpus didn't change */
864 if (cpumask_equal(cs->cpus_allowed, trialcs->cpus_allowed))
867 retval = heap_init(&heap, PAGE_SIZE, GFP_KERNEL, NULL);
871 is_load_balanced = is_sched_load_balance(trialcs);
873 mutex_lock(&callback_mutex);
874 cpumask_copy(cs->cpus_allowed, trialcs->cpus_allowed);
875 mutex_unlock(&callback_mutex);
878 * Scan tasks in the cpuset, and update the cpumasks of any
879 * that need an update.
881 update_tasks_cpumask(cs, &heap);
885 if (is_load_balanced)
886 rebuild_sched_domains_locked();
893 * Migrate memory region from one set of nodes to another.
895 * Temporarilly set tasks mems_allowed to target nodes of migration,
896 * so that the migration code can allocate pages on these nodes.
898 * Call holding cpuset_mutex, so current's cpuset won't change
899 * during this call, as manage_mutex holds off any cpuset_attach()
900 * calls. Therefore we don't need to take task_lock around the
901 * call to guarantee_online_mems(), as we know no one is changing
904 * While the mm_struct we are migrating is typically from some
905 * other task, the task_struct mems_allowed that we are hacking
906 * is for our current task, which must allocate new pages for that
907 * migrating memory region.
910 static void cpuset_migrate_mm(struct mm_struct *mm, const nodemask_t *from,
911 const nodemask_t *to)
913 struct task_struct *tsk = current;
915 tsk->mems_allowed = *to;
917 do_migrate_pages(mm, from, to, MPOL_MF_MOVE_ALL);
919 guarantee_online_mems(task_cs(tsk),&tsk->mems_allowed);
923 * cpuset_change_task_nodemask - change task's mems_allowed and mempolicy
924 * @tsk: the task to change
925 * @newmems: new nodes that the task will be set
927 * In order to avoid seeing no nodes if the old and new nodes are disjoint,
928 * we structure updates as setting all new allowed nodes, then clearing newly
931 static void cpuset_change_task_nodemask(struct task_struct *tsk,
937 * Allow tasks that have access to memory reserves because they have
938 * been OOM killed to get memory anywhere.
940 if (unlikely(test_thread_flag(TIF_MEMDIE)))
942 if (current->flags & PF_EXITING) /* Let dying task have memory */
947 * Determine if a loop is necessary if another thread is doing
948 * get_mems_allowed(). If at least one node remains unchanged and
949 * tsk does not have a mempolicy, then an empty nodemask will not be
950 * possible when mems_allowed is larger than a word.
952 need_loop = task_has_mempolicy(tsk) ||
953 !nodes_intersects(*newmems, tsk->mems_allowed);
956 write_seqcount_begin(&tsk->mems_allowed_seq);
958 nodes_or(tsk->mems_allowed, tsk->mems_allowed, *newmems);
959 mpol_rebind_task(tsk, newmems, MPOL_REBIND_STEP1);
961 mpol_rebind_task(tsk, newmems, MPOL_REBIND_STEP2);
962 tsk->mems_allowed = *newmems;
965 write_seqcount_end(&tsk->mems_allowed_seq);
971 * Update task's mems_allowed and rebind its mempolicy and vmas' mempolicy
972 * of it to cpuset's new mems_allowed, and migrate pages to new nodes if
973 * memory_migrate flag is set. Called with cpuset_mutex held.
975 static void cpuset_change_nodemask(struct task_struct *p,
976 struct cgroup_scanner *scan)
978 struct mm_struct *mm;
981 const nodemask_t *oldmem = scan->data;
982 static nodemask_t newmems; /* protected by cpuset_mutex */
984 cs = cgroup_cs(scan->cg);
985 guarantee_online_mems(cs, &newmems);
987 cpuset_change_task_nodemask(p, &newmems);
993 migrate = is_memory_migrate(cs);
995 mpol_rebind_mm(mm, &cs->mems_allowed);
997 cpuset_migrate_mm(mm, oldmem, &cs->mems_allowed);
1001 static void *cpuset_being_rebound;
1004 * update_tasks_nodemask - Update the nodemasks of tasks in the cpuset.
1005 * @cs: the cpuset in which each task's mems_allowed mask needs to be changed
1006 * @oldmem: old mems_allowed of cpuset cs
1007 * @heap: if NULL, defer allocating heap memory to cgroup_scan_tasks()
1009 * Called with cpuset_mutex held
1010 * No return value. It's guaranteed that cgroup_scan_tasks() always returns 0
1013 static void update_tasks_nodemask(struct cpuset *cs, const nodemask_t *oldmem,
1014 struct ptr_heap *heap)
1016 struct cgroup_scanner scan;
1018 cpuset_being_rebound = cs; /* causes mpol_dup() rebind */
1020 scan.cg = cs->css.cgroup;
1021 scan.test_task = NULL;
1022 scan.process_task = cpuset_change_nodemask;
1024 scan.data = (nodemask_t *)oldmem;
1027 * The mpol_rebind_mm() call takes mmap_sem, which we couldn't
1028 * take while holding tasklist_lock. Forks can happen - the
1029 * mpol_dup() cpuset_being_rebound check will catch such forks,
1030 * and rebind their vma mempolicies too. Because we still hold
1031 * the global cpuset_mutex, we know that no other rebind effort
1032 * will be contending for the global variable cpuset_being_rebound.
1033 * It's ok if we rebind the same mm twice; mpol_rebind_mm()
1034 * is idempotent. Also migrate pages in each mm to new nodes.
1036 cgroup_scan_tasks(&scan);
1038 /* We're done rebinding vmas to this cpuset's new mems_allowed. */
1039 cpuset_being_rebound = NULL;
1043 * Handle user request to change the 'mems' memory placement
1044 * of a cpuset. Needs to validate the request, update the
1045 * cpusets mems_allowed, and for each task in the cpuset,
1046 * update mems_allowed and rebind task's mempolicy and any vma
1047 * mempolicies and if the cpuset is marked 'memory_migrate',
1048 * migrate the tasks pages to the new memory.
1050 * Call with cpuset_mutex held. May take callback_mutex during call.
1051 * Will take tasklist_lock, scan tasklist for tasks in cpuset cs,
1052 * lock each such tasks mm->mmap_sem, scan its vma's and rebind
1053 * their mempolicies to the cpusets new mems_allowed.
1055 static int update_nodemask(struct cpuset *cs, struct cpuset *trialcs,
1058 NODEMASK_ALLOC(nodemask_t, oldmem, GFP_KERNEL);
1060 struct ptr_heap heap;
1066 * top_cpuset.mems_allowed tracks node_stats[N_MEMORY];
1069 if (cs == &top_cpuset) {
1075 * An empty mems_allowed is ok iff there are no tasks in the cpuset.
1076 * Since nodelist_parse() fails on an empty mask, we special case
1077 * that parsing. The validate_change() call ensures that cpusets
1078 * with tasks have memory.
1081 nodes_clear(trialcs->mems_allowed);
1083 retval = nodelist_parse(buf, trialcs->mems_allowed);
1087 if (!nodes_subset(trialcs->mems_allowed,
1088 node_states[N_MEMORY])) {
1093 *oldmem = cs->mems_allowed;
1094 if (nodes_equal(*oldmem, trialcs->mems_allowed)) {
1095 retval = 0; /* Too easy - nothing to do */
1098 retval = validate_change(cs, trialcs);
1102 retval = heap_init(&heap, PAGE_SIZE, GFP_KERNEL, NULL);
1106 mutex_lock(&callback_mutex);
1107 cs->mems_allowed = trialcs->mems_allowed;
1108 mutex_unlock(&callback_mutex);
1110 update_tasks_nodemask(cs, oldmem, &heap);
1114 NODEMASK_FREE(oldmem);
1118 int current_cpuset_is_being_rebound(void)
1120 return task_cs(current) == cpuset_being_rebound;
1123 static int update_relax_domain_level(struct cpuset *cs, s64 val)
1126 if (val < -1 || val >= sched_domain_level_max)
1130 if (val != cs->relax_domain_level) {
1131 cs->relax_domain_level = val;
1132 if (!cpumask_empty(cs->cpus_allowed) &&
1133 is_sched_load_balance(cs))
1134 rebuild_sched_domains_locked();
1141 * cpuset_change_flag - make a task's spread flags the same as its cpuset's
1142 * @tsk: task to be updated
1143 * @scan: struct cgroup_scanner containing the cgroup of the task
1145 * Called by cgroup_scan_tasks() for each task in a cgroup.
1147 * We don't need to re-check for the cgroup/cpuset membership, since we're
1148 * holding cpuset_mutex at this point.
1150 static void cpuset_change_flag(struct task_struct *tsk,
1151 struct cgroup_scanner *scan)
1153 cpuset_update_task_spread_flag(cgroup_cs(scan->cg), tsk);
1157 * update_tasks_flags - update the spread flags of tasks in the cpuset.
1158 * @cs: the cpuset in which each task's spread flags needs to be changed
1159 * @heap: if NULL, defer allocating heap memory to cgroup_scan_tasks()
1161 * Called with cpuset_mutex held
1163 * The cgroup_scan_tasks() function will scan all the tasks in a cgroup,
1164 * calling callback functions for each.
1166 * No return value. It's guaranteed that cgroup_scan_tasks() always returns 0
1169 static void update_tasks_flags(struct cpuset *cs, struct ptr_heap *heap)
1171 struct cgroup_scanner scan;
1173 scan.cg = cs->css.cgroup;
1174 scan.test_task = NULL;
1175 scan.process_task = cpuset_change_flag;
1177 cgroup_scan_tasks(&scan);
1181 * update_flag - read a 0 or a 1 in a file and update associated flag
1182 * bit: the bit to update (see cpuset_flagbits_t)
1183 * cs: the cpuset to update
1184 * turning_on: whether the flag is being set or cleared
1186 * Call with cpuset_mutex held.
1189 static int update_flag(cpuset_flagbits_t bit, struct cpuset *cs,
1192 struct cpuset *trialcs;
1193 int balance_flag_changed;
1194 int spread_flag_changed;
1195 struct ptr_heap heap;
1198 trialcs = alloc_trial_cpuset(cs);
1203 set_bit(bit, &trialcs->flags);
1205 clear_bit(bit, &trialcs->flags);
1207 err = validate_change(cs, trialcs);
1211 err = heap_init(&heap, PAGE_SIZE, GFP_KERNEL, NULL);
1215 balance_flag_changed = (is_sched_load_balance(cs) !=
1216 is_sched_load_balance(trialcs));
1218 spread_flag_changed = ((is_spread_slab(cs) != is_spread_slab(trialcs))
1219 || (is_spread_page(cs) != is_spread_page(trialcs)));
1221 mutex_lock(&callback_mutex);
1222 cs->flags = trialcs->flags;
1223 mutex_unlock(&callback_mutex);
1225 if (!cpumask_empty(trialcs->cpus_allowed) && balance_flag_changed)
1226 rebuild_sched_domains_locked();
1228 if (spread_flag_changed)
1229 update_tasks_flags(cs, &heap);
1232 free_trial_cpuset(trialcs);
1237 * Frequency meter - How fast is some event occurring?
1239 * These routines manage a digitally filtered, constant time based,
1240 * event frequency meter. There are four routines:
1241 * fmeter_init() - initialize a frequency meter.
1242 * fmeter_markevent() - called each time the event happens.
1243 * fmeter_getrate() - returns the recent rate of such events.
1244 * fmeter_update() - internal routine used to update fmeter.
1246 * A common data structure is passed to each of these routines,
1247 * which is used to keep track of the state required to manage the
1248 * frequency meter and its digital filter.
1250 * The filter works on the number of events marked per unit time.
1251 * The filter is single-pole low-pass recursive (IIR). The time unit
1252 * is 1 second. Arithmetic is done using 32-bit integers scaled to
1253 * simulate 3 decimal digits of precision (multiplied by 1000).
1255 * With an FM_COEF of 933, and a time base of 1 second, the filter
1256 * has a half-life of 10 seconds, meaning that if the events quit
1257 * happening, then the rate returned from the fmeter_getrate()
1258 * will be cut in half each 10 seconds, until it converges to zero.
1260 * It is not worth doing a real infinitely recursive filter. If more
1261 * than FM_MAXTICKS ticks have elapsed since the last filter event,
1262 * just compute FM_MAXTICKS ticks worth, by which point the level
1265 * Limit the count of unprocessed events to FM_MAXCNT, so as to avoid
1266 * arithmetic overflow in the fmeter_update() routine.
1268 * Given the simple 32 bit integer arithmetic used, this meter works
1269 * best for reporting rates between one per millisecond (msec) and
1270 * one per 32 (approx) seconds. At constant rates faster than one
1271 * per msec it maxes out at values just under 1,000,000. At constant
1272 * rates between one per msec, and one per second it will stabilize
1273 * to a value N*1000, where N is the rate of events per second.
1274 * At constant rates between one per second and one per 32 seconds,
1275 * it will be choppy, moving up on the seconds that have an event,
1276 * and then decaying until the next event. At rates slower than
1277 * about one in 32 seconds, it decays all the way back to zero between
1281 #define FM_COEF 933 /* coefficient for half-life of 10 secs */
1282 #define FM_MAXTICKS ((time_t)99) /* useless computing more ticks than this */
1283 #define FM_MAXCNT 1000000 /* limit cnt to avoid overflow */
1284 #define FM_SCALE 1000 /* faux fixed point scale */
1286 /* Initialize a frequency meter */
1287 static void fmeter_init(struct fmeter *fmp)
1292 spin_lock_init(&fmp->lock);
1295 /* Internal meter update - process cnt events and update value */
1296 static void fmeter_update(struct fmeter *fmp)
1298 time_t now = get_seconds();
1299 time_t ticks = now - fmp->time;
1304 ticks = min(FM_MAXTICKS, ticks);
1306 fmp->val = (FM_COEF * fmp->val) / FM_SCALE;
1309 fmp->val += ((FM_SCALE - FM_COEF) * fmp->cnt) / FM_SCALE;
1313 /* Process any previous ticks, then bump cnt by one (times scale). */
1314 static void fmeter_markevent(struct fmeter *fmp)
1316 spin_lock(&fmp->lock);
1318 fmp->cnt = min(FM_MAXCNT, fmp->cnt + FM_SCALE);
1319 spin_unlock(&fmp->lock);
1322 /* Process any previous ticks, then return current value. */
1323 static int fmeter_getrate(struct fmeter *fmp)
1327 spin_lock(&fmp->lock);
1330 spin_unlock(&fmp->lock);
1334 /* Called by cgroups to determine if a cpuset is usable; cpuset_mutex held */
1335 static int cpuset_can_attach(struct cgroup *cgrp, struct cgroup_taskset *tset)
1337 struct cpuset *cs = cgroup_cs(cgrp);
1338 struct task_struct *task;
1341 mutex_lock(&cpuset_mutex);
1344 if (cpumask_empty(cs->cpus_allowed) || nodes_empty(cs->mems_allowed))
1347 cgroup_taskset_for_each(task, cgrp, tset) {
1349 * Kthreads which disallow setaffinity shouldn't be moved
1350 * to a new cpuset; we don't want to change their cpu
1351 * affinity and isolating such threads by their set of
1352 * allowed nodes is unnecessary. Thus, cpusets are not
1353 * applicable for such threads. This prevents checking for
1354 * success of set_cpus_allowed_ptr() on all attached tasks
1355 * before cpus_allowed may be changed.
1358 if (task->flags & PF_NO_SETAFFINITY)
1360 ret = security_task_setscheduler(task);
1366 * Mark attach is in progress. This makes validate_change() fail
1367 * changes which zero cpus/mems_allowed.
1369 cs->attach_in_progress++;
1372 mutex_unlock(&cpuset_mutex);
1376 static void cpuset_cancel_attach(struct cgroup *cgrp,
1377 struct cgroup_taskset *tset)
1379 mutex_lock(&cpuset_mutex);
1380 cgroup_cs(cgrp)->attach_in_progress--;
1381 mutex_unlock(&cpuset_mutex);
1385 * Protected by cpuset_mutex. cpus_attach is used only by cpuset_attach()
1386 * but we can't allocate it dynamically there. Define it global and
1387 * allocate from cpuset_init().
1389 static cpumask_var_t cpus_attach;
1391 static void cpuset_attach(struct cgroup *cgrp, struct cgroup_taskset *tset)
1393 /* static buf protected by cpuset_mutex */
1394 static nodemask_t cpuset_attach_nodemask_to;
1395 struct mm_struct *mm;
1396 struct task_struct *task;
1397 struct task_struct *leader = cgroup_taskset_first(tset);
1398 struct cgroup *oldcgrp = cgroup_taskset_cur_cgroup(tset);
1399 struct cpuset *cs = cgroup_cs(cgrp);
1400 struct cpuset *oldcs = cgroup_cs(oldcgrp);
1402 mutex_lock(&cpuset_mutex);
1404 /* prepare for attach */
1405 if (cs == &top_cpuset)
1406 cpumask_copy(cpus_attach, cpu_possible_mask);
1408 guarantee_online_cpus(cs, cpus_attach);
1410 guarantee_online_mems(cs, &cpuset_attach_nodemask_to);
1412 cgroup_taskset_for_each(task, cgrp, tset) {
1414 * can_attach beforehand should guarantee that this doesn't
1415 * fail. TODO: have a better way to handle failure here
1417 WARN_ON_ONCE(set_cpus_allowed_ptr(task, cpus_attach));
1419 cpuset_change_task_nodemask(task, &cpuset_attach_nodemask_to);
1420 cpuset_update_task_spread_flag(cs, task);
1424 * Change mm, possibly for multiple threads in a threadgroup. This is
1425 * expensive and may sleep.
1427 cpuset_attach_nodemask_to = cs->mems_allowed;
1428 mm = get_task_mm(leader);
1430 mpol_rebind_mm(mm, &cpuset_attach_nodemask_to);
1431 if (is_memory_migrate(cs))
1432 cpuset_migrate_mm(mm, &oldcs->mems_allowed,
1433 &cpuset_attach_nodemask_to);
1437 cs->attach_in_progress--;
1440 * We may have raced with CPU/memory hotunplug. Trigger hotplug
1441 * propagation if @cs doesn't have any CPU or memory. It will move
1442 * the newly added tasks to the nearest parent which can execute.
1444 if (cpumask_empty(cs->cpus_allowed) || nodes_empty(cs->mems_allowed))
1445 schedule_cpuset_propagate_hotplug(cs);
1447 mutex_unlock(&cpuset_mutex);
1450 /* The various types of files and directories in a cpuset file system */
1453 FILE_MEMORY_MIGRATE,
1459 FILE_SCHED_LOAD_BALANCE,
1460 FILE_SCHED_RELAX_DOMAIN_LEVEL,
1461 FILE_MEMORY_PRESSURE_ENABLED,
1462 FILE_MEMORY_PRESSURE,
1465 } cpuset_filetype_t;
1467 static int cpuset_write_u64(struct cgroup *cgrp, struct cftype *cft, u64 val)
1469 struct cpuset *cs = cgroup_cs(cgrp);
1470 cpuset_filetype_t type = cft->private;
1471 int retval = -ENODEV;
1473 mutex_lock(&cpuset_mutex);
1474 if (!is_cpuset_online(cs))
1478 case FILE_CPU_EXCLUSIVE:
1479 retval = update_flag(CS_CPU_EXCLUSIVE, cs, val);
1481 case FILE_MEM_EXCLUSIVE:
1482 retval = update_flag(CS_MEM_EXCLUSIVE, cs, val);
1484 case FILE_MEM_HARDWALL:
1485 retval = update_flag(CS_MEM_HARDWALL, cs, val);
1487 case FILE_SCHED_LOAD_BALANCE:
1488 retval = update_flag(CS_SCHED_LOAD_BALANCE, cs, val);
1490 case FILE_MEMORY_MIGRATE:
1491 retval = update_flag(CS_MEMORY_MIGRATE, cs, val);
1493 case FILE_MEMORY_PRESSURE_ENABLED:
1494 cpuset_memory_pressure_enabled = !!val;
1496 case FILE_MEMORY_PRESSURE:
1499 case FILE_SPREAD_PAGE:
1500 retval = update_flag(CS_SPREAD_PAGE, cs, val);
1502 case FILE_SPREAD_SLAB:
1503 retval = update_flag(CS_SPREAD_SLAB, cs, val);
1510 mutex_unlock(&cpuset_mutex);
1514 static int cpuset_write_s64(struct cgroup *cgrp, struct cftype *cft, s64 val)
1516 struct cpuset *cs = cgroup_cs(cgrp);
1517 cpuset_filetype_t type = cft->private;
1518 int retval = -ENODEV;
1520 mutex_lock(&cpuset_mutex);
1521 if (!is_cpuset_online(cs))
1525 case FILE_SCHED_RELAX_DOMAIN_LEVEL:
1526 retval = update_relax_domain_level(cs, val);
1533 mutex_unlock(&cpuset_mutex);
1538 * Common handling for a write to a "cpus" or "mems" file.
1540 static int cpuset_write_resmask(struct cgroup *cgrp, struct cftype *cft,
1543 struct cpuset *cs = cgroup_cs(cgrp);
1544 struct cpuset *trialcs;
1545 int retval = -ENODEV;
1548 * CPU or memory hotunplug may leave @cs w/o any execution
1549 * resources, in which case the hotplug code asynchronously updates
1550 * configuration and transfers all tasks to the nearest ancestor
1551 * which can execute.
1553 * As writes to "cpus" or "mems" may restore @cs's execution
1554 * resources, wait for the previously scheduled operations before
1555 * proceeding, so that we don't end up keep removing tasks added
1556 * after execution capability is restored.
1558 * Flushing cpuset_hotplug_work is enough to synchronize against
1559 * hotplug hanlding; however, cpuset_attach() may schedule
1560 * propagation work directly. Flush the workqueue too.
1562 flush_work(&cpuset_hotplug_work);
1563 flush_workqueue(cpuset_propagate_hotplug_wq);
1565 mutex_lock(&cpuset_mutex);
1566 if (!is_cpuset_online(cs))
1569 trialcs = alloc_trial_cpuset(cs);
1575 switch (cft->private) {
1577 retval = update_cpumask(cs, trialcs, buf);
1580 retval = update_nodemask(cs, trialcs, buf);
1587 free_trial_cpuset(trialcs);
1589 mutex_unlock(&cpuset_mutex);
1594 * These ascii lists should be read in a single call, by using a user
1595 * buffer large enough to hold the entire map. If read in smaller
1596 * chunks, there is no guarantee of atomicity. Since the display format
1597 * used, list of ranges of sequential numbers, is variable length,
1598 * and since these maps can change value dynamically, one could read
1599 * gibberish by doing partial reads while a list was changing.
1600 * A single large read to a buffer that crosses a page boundary is
1601 * ok, because the result being copied to user land is not recomputed
1602 * across a page fault.
1605 static size_t cpuset_sprintf_cpulist(char *page, struct cpuset *cs)
1609 mutex_lock(&callback_mutex);
1610 count = cpulist_scnprintf(page, PAGE_SIZE, cs->cpus_allowed);
1611 mutex_unlock(&callback_mutex);
1616 static size_t cpuset_sprintf_memlist(char *page, struct cpuset *cs)
1620 mutex_lock(&callback_mutex);
1621 count = nodelist_scnprintf(page, PAGE_SIZE, cs->mems_allowed);
1622 mutex_unlock(&callback_mutex);
1627 static ssize_t cpuset_common_file_read(struct cgroup *cont,
1631 size_t nbytes, loff_t *ppos)
1633 struct cpuset *cs = cgroup_cs(cont);
1634 cpuset_filetype_t type = cft->private;
1639 if (!(page = (char *)__get_free_page(GFP_TEMPORARY)))
1646 s += cpuset_sprintf_cpulist(s, cs);
1649 s += cpuset_sprintf_memlist(s, cs);
1657 retval = simple_read_from_buffer(buf, nbytes, ppos, page, s - page);
1659 free_page((unsigned long)page);
1663 static u64 cpuset_read_u64(struct cgroup *cont, struct cftype *cft)
1665 struct cpuset *cs = cgroup_cs(cont);
1666 cpuset_filetype_t type = cft->private;
1668 case FILE_CPU_EXCLUSIVE:
1669 return is_cpu_exclusive(cs);
1670 case FILE_MEM_EXCLUSIVE:
1671 return is_mem_exclusive(cs);
1672 case FILE_MEM_HARDWALL:
1673 return is_mem_hardwall(cs);
1674 case FILE_SCHED_LOAD_BALANCE:
1675 return is_sched_load_balance(cs);
1676 case FILE_MEMORY_MIGRATE:
1677 return is_memory_migrate(cs);
1678 case FILE_MEMORY_PRESSURE_ENABLED:
1679 return cpuset_memory_pressure_enabled;
1680 case FILE_MEMORY_PRESSURE:
1681 return fmeter_getrate(&cs->fmeter);
1682 case FILE_SPREAD_PAGE:
1683 return is_spread_page(cs);
1684 case FILE_SPREAD_SLAB:
1685 return is_spread_slab(cs);
1690 /* Unreachable but makes gcc happy */
1694 static s64 cpuset_read_s64(struct cgroup *cont, struct cftype *cft)
1696 struct cpuset *cs = cgroup_cs(cont);
1697 cpuset_filetype_t type = cft->private;
1699 case FILE_SCHED_RELAX_DOMAIN_LEVEL:
1700 return cs->relax_domain_level;
1705 /* Unrechable but makes gcc happy */
1711 * for the common functions, 'private' gives the type of file
1714 static struct cftype files[] = {
1717 .read = cpuset_common_file_read,
1718 .write_string = cpuset_write_resmask,
1719 .max_write_len = (100U + 6 * NR_CPUS),
1720 .private = FILE_CPULIST,
1725 .read = cpuset_common_file_read,
1726 .write_string = cpuset_write_resmask,
1727 .max_write_len = (100U + 6 * MAX_NUMNODES),
1728 .private = FILE_MEMLIST,
1732 .name = "cpu_exclusive",
1733 .read_u64 = cpuset_read_u64,
1734 .write_u64 = cpuset_write_u64,
1735 .private = FILE_CPU_EXCLUSIVE,
1739 .name = "mem_exclusive",
1740 .read_u64 = cpuset_read_u64,
1741 .write_u64 = cpuset_write_u64,
1742 .private = FILE_MEM_EXCLUSIVE,
1746 .name = "mem_hardwall",
1747 .read_u64 = cpuset_read_u64,
1748 .write_u64 = cpuset_write_u64,
1749 .private = FILE_MEM_HARDWALL,
1753 .name = "sched_load_balance",
1754 .read_u64 = cpuset_read_u64,
1755 .write_u64 = cpuset_write_u64,
1756 .private = FILE_SCHED_LOAD_BALANCE,
1760 .name = "sched_relax_domain_level",
1761 .read_s64 = cpuset_read_s64,
1762 .write_s64 = cpuset_write_s64,
1763 .private = FILE_SCHED_RELAX_DOMAIN_LEVEL,
1767 .name = "memory_migrate",
1768 .read_u64 = cpuset_read_u64,
1769 .write_u64 = cpuset_write_u64,
1770 .private = FILE_MEMORY_MIGRATE,
1774 .name = "memory_pressure",
1775 .read_u64 = cpuset_read_u64,
1776 .write_u64 = cpuset_write_u64,
1777 .private = FILE_MEMORY_PRESSURE,
1782 .name = "memory_spread_page",
1783 .read_u64 = cpuset_read_u64,
1784 .write_u64 = cpuset_write_u64,
1785 .private = FILE_SPREAD_PAGE,
1789 .name = "memory_spread_slab",
1790 .read_u64 = cpuset_read_u64,
1791 .write_u64 = cpuset_write_u64,
1792 .private = FILE_SPREAD_SLAB,
1796 .name = "memory_pressure_enabled",
1797 .flags = CFTYPE_ONLY_ON_ROOT,
1798 .read_u64 = cpuset_read_u64,
1799 .write_u64 = cpuset_write_u64,
1800 .private = FILE_MEMORY_PRESSURE_ENABLED,
1807 * cpuset_css_alloc - allocate a cpuset css
1808 * cont: control group that the new cpuset will be part of
1811 static struct cgroup_subsys_state *cpuset_css_alloc(struct cgroup *cont)
1816 return &top_cpuset.css;
1818 cs = kzalloc(sizeof(*cs), GFP_KERNEL);
1820 return ERR_PTR(-ENOMEM);
1821 if (!alloc_cpumask_var(&cs->cpus_allowed, GFP_KERNEL)) {
1823 return ERR_PTR(-ENOMEM);
1826 set_bit(CS_SCHED_LOAD_BALANCE, &cs->flags);
1827 cpumask_clear(cs->cpus_allowed);
1828 nodes_clear(cs->mems_allowed);
1829 fmeter_init(&cs->fmeter);
1830 INIT_WORK(&cs->hotplug_work, cpuset_propagate_hotplug_workfn);
1831 cs->relax_domain_level = -1;
1836 static int cpuset_css_online(struct cgroup *cgrp)
1838 struct cpuset *cs = cgroup_cs(cgrp);
1839 struct cpuset *parent = parent_cs(cs);
1840 struct cpuset *tmp_cs;
1841 struct cgroup *pos_cg;
1846 mutex_lock(&cpuset_mutex);
1848 set_bit(CS_ONLINE, &cs->flags);
1849 if (is_spread_page(parent))
1850 set_bit(CS_SPREAD_PAGE, &cs->flags);
1851 if (is_spread_slab(parent))
1852 set_bit(CS_SPREAD_SLAB, &cs->flags);
1854 number_of_cpusets++;
1856 if (!test_bit(CGRP_CPUSET_CLONE_CHILDREN, &cgrp->flags))
1860 * Clone @parent's configuration if CGRP_CPUSET_CLONE_CHILDREN is
1861 * set. This flag handling is implemented in cgroup core for
1862 * histrical reasons - the flag may be specified during mount.
1864 * Currently, if any sibling cpusets have exclusive cpus or mem, we
1865 * refuse to clone the configuration - thereby refusing the task to
1866 * be entered, and as a result refusing the sys_unshare() or
1867 * clone() which initiated it. If this becomes a problem for some
1868 * users who wish to allow that scenario, then this could be
1869 * changed to grant parent->cpus_allowed-sibling_cpus_exclusive
1870 * (and likewise for mems) to the new cgroup.
1873 cpuset_for_each_child(tmp_cs, pos_cg, parent) {
1874 if (is_mem_exclusive(tmp_cs) || is_cpu_exclusive(tmp_cs)) {
1881 mutex_lock(&callback_mutex);
1882 cs->mems_allowed = parent->mems_allowed;
1883 cpumask_copy(cs->cpus_allowed, parent->cpus_allowed);
1884 mutex_unlock(&callback_mutex);
1886 mutex_unlock(&cpuset_mutex);
1890 static void cpuset_css_offline(struct cgroup *cgrp)
1892 struct cpuset *cs = cgroup_cs(cgrp);
1894 mutex_lock(&cpuset_mutex);
1896 if (is_sched_load_balance(cs))
1897 update_flag(CS_SCHED_LOAD_BALANCE, cs, 0);
1899 number_of_cpusets--;
1900 clear_bit(CS_ONLINE, &cs->flags);
1902 mutex_unlock(&cpuset_mutex);
1906 * If the cpuset being removed has its flag 'sched_load_balance'
1907 * enabled, then simulate turning sched_load_balance off, which
1908 * will call rebuild_sched_domains_locked().
1911 static void cpuset_css_free(struct cgroup *cont)
1913 struct cpuset *cs = cgroup_cs(cont);
1915 free_cpumask_var(cs->cpus_allowed);
1919 struct cgroup_subsys cpuset_subsys = {
1921 .css_alloc = cpuset_css_alloc,
1922 .css_online = cpuset_css_online,
1923 .css_offline = cpuset_css_offline,
1924 .css_free = cpuset_css_free,
1925 .can_attach = cpuset_can_attach,
1926 .cancel_attach = cpuset_cancel_attach,
1927 .attach = cpuset_attach,
1928 .subsys_id = cpuset_subsys_id,
1929 .base_cftypes = files,
1934 * cpuset_init - initialize cpusets at system boot
1936 * Description: Initialize top_cpuset and the cpuset internal file system,
1939 int __init cpuset_init(void)
1943 if (!alloc_cpumask_var(&top_cpuset.cpus_allowed, GFP_KERNEL))
1946 cpumask_setall(top_cpuset.cpus_allowed);
1947 nodes_setall(top_cpuset.mems_allowed);
1949 fmeter_init(&top_cpuset.fmeter);
1950 set_bit(CS_SCHED_LOAD_BALANCE, &top_cpuset.flags);
1951 top_cpuset.relax_domain_level = -1;
1953 err = register_filesystem(&cpuset_fs_type);
1957 if (!alloc_cpumask_var(&cpus_attach, GFP_KERNEL))
1960 number_of_cpusets = 1;
1965 * If CPU and/or memory hotplug handlers, below, unplug any CPUs
1966 * or memory nodes, we need to walk over the cpuset hierarchy,
1967 * removing that CPU or node from all cpusets. If this removes the
1968 * last CPU or node from a cpuset, then move the tasks in the empty
1969 * cpuset to its next-highest non-empty parent.
1971 static void remove_tasks_in_empty_cpuset(struct cpuset *cs)
1973 struct cpuset *parent;
1976 * Find its next-highest non-empty parent, (top cpuset
1977 * has online cpus, so can't be empty).
1979 parent = parent_cs(cs);
1980 while (cpumask_empty(parent->cpus_allowed) ||
1981 nodes_empty(parent->mems_allowed))
1982 parent = parent_cs(parent);
1984 if (cgroup_transfer_tasks(parent->css.cgroup, cs->css.cgroup)) {
1986 printk(KERN_ERR "cpuset: failed to transfer tasks out of empty cpuset %s\n",
1987 cgroup_name(cs->css.cgroup));
1993 * cpuset_propagate_hotplug_workfn - propagate CPU/memory hotplug to a cpuset
1994 * @cs: cpuset in interest
1996 * Compare @cs's cpu and mem masks against top_cpuset and if some have gone
1997 * offline, update @cs accordingly. If @cs ends up with no CPU or memory,
1998 * all its tasks are moved to the nearest ancestor with both resources.
2000 static void cpuset_propagate_hotplug_workfn(struct work_struct *work)
2002 static cpumask_t off_cpus;
2003 static nodemask_t off_mems, tmp_mems;
2004 struct cpuset *cs = container_of(work, struct cpuset, hotplug_work);
2007 mutex_lock(&cpuset_mutex);
2009 cpumask_andnot(&off_cpus, cs->cpus_allowed, top_cpuset.cpus_allowed);
2010 nodes_andnot(off_mems, cs->mems_allowed, top_cpuset.mems_allowed);
2012 /* remove offline cpus from @cs */
2013 if (!cpumask_empty(&off_cpus)) {
2014 mutex_lock(&callback_mutex);
2015 cpumask_andnot(cs->cpus_allowed, cs->cpus_allowed, &off_cpus);
2016 mutex_unlock(&callback_mutex);
2017 update_tasks_cpumask(cs, NULL);
2020 /* remove offline mems from @cs */
2021 if (!nodes_empty(off_mems)) {
2022 tmp_mems = cs->mems_allowed;
2023 mutex_lock(&callback_mutex);
2024 nodes_andnot(cs->mems_allowed, cs->mems_allowed, off_mems);
2025 mutex_unlock(&callback_mutex);
2026 update_tasks_nodemask(cs, &tmp_mems, NULL);
2029 is_empty = cpumask_empty(cs->cpus_allowed) ||
2030 nodes_empty(cs->mems_allowed);
2032 mutex_unlock(&cpuset_mutex);
2035 * If @cs became empty, move tasks to the nearest ancestor with
2036 * execution resources. This is full cgroup operation which will
2037 * also call back into cpuset. Should be done outside any lock.
2040 remove_tasks_in_empty_cpuset(cs);
2042 /* the following may free @cs, should be the last operation */
2047 * schedule_cpuset_propagate_hotplug - schedule hotplug propagation to a cpuset
2048 * @cs: cpuset of interest
2050 * Schedule cpuset_propagate_hotplug_workfn() which will update CPU and
2051 * memory masks according to top_cpuset.
2053 static void schedule_cpuset_propagate_hotplug(struct cpuset *cs)
2056 * Pin @cs. The refcnt will be released when the work item
2057 * finishes executing.
2059 if (!css_tryget(&cs->css))
2063 * Queue @cs->hotplug_work. If already pending, lose the css ref.
2064 * cpuset_propagate_hotplug_wq is ordered and propagation will
2065 * happen in the order this function is called.
2067 if (!queue_work(cpuset_propagate_hotplug_wq, &cs->hotplug_work))
2072 * cpuset_hotplug_workfn - handle CPU/memory hotunplug for a cpuset
2074 * This function is called after either CPU or memory configuration has
2075 * changed and updates cpuset accordingly. The top_cpuset is always
2076 * synchronized to cpu_active_mask and N_MEMORY, which is necessary in
2077 * order to make cpusets transparent (of no affect) on systems that are
2078 * actively using CPU hotplug but making no active use of cpusets.
2080 * Non-root cpusets are only affected by offlining. If any CPUs or memory
2081 * nodes have been taken down, cpuset_propagate_hotplug() is invoked on all
2084 * Note that CPU offlining during suspend is ignored. We don't modify
2085 * cpusets across suspend/resume cycles at all.
2087 static void cpuset_hotplug_workfn(struct work_struct *work)
2089 static cpumask_t new_cpus, tmp_cpus;
2090 static nodemask_t new_mems, tmp_mems;
2091 bool cpus_updated, mems_updated;
2092 bool cpus_offlined, mems_offlined;
2094 mutex_lock(&cpuset_mutex);
2096 /* fetch the available cpus/mems and find out which changed how */
2097 cpumask_copy(&new_cpus, cpu_active_mask);
2098 new_mems = node_states[N_MEMORY];
2100 cpus_updated = !cpumask_equal(top_cpuset.cpus_allowed, &new_cpus);
2101 cpus_offlined = cpumask_andnot(&tmp_cpus, top_cpuset.cpus_allowed,
2104 mems_updated = !nodes_equal(top_cpuset.mems_allowed, new_mems);
2105 nodes_andnot(tmp_mems, top_cpuset.mems_allowed, new_mems);
2106 mems_offlined = !nodes_empty(tmp_mems);
2108 /* synchronize cpus_allowed to cpu_active_mask */
2110 mutex_lock(&callback_mutex);
2111 cpumask_copy(top_cpuset.cpus_allowed, &new_cpus);
2112 mutex_unlock(&callback_mutex);
2113 /* we don't mess with cpumasks of tasks in top_cpuset */
2116 /* synchronize mems_allowed to N_MEMORY */
2118 tmp_mems = top_cpuset.mems_allowed;
2119 mutex_lock(&callback_mutex);
2120 top_cpuset.mems_allowed = new_mems;
2121 mutex_unlock(&callback_mutex);
2122 update_tasks_nodemask(&top_cpuset, &tmp_mems, NULL);
2125 /* if cpus or mems went down, we need to propagate to descendants */
2126 if (cpus_offlined || mems_offlined) {
2128 struct cgroup *pos_cgrp;
2131 cpuset_for_each_descendant_pre(cs, pos_cgrp, &top_cpuset)
2132 schedule_cpuset_propagate_hotplug(cs);
2136 mutex_unlock(&cpuset_mutex);
2138 /* wait for propagations to finish */
2139 flush_workqueue(cpuset_propagate_hotplug_wq);
2141 /* rebuild sched domains if cpus_allowed has changed */
2143 rebuild_sched_domains();
2146 void cpuset_update_active_cpus(bool cpu_online)
2149 * We're inside cpu hotplug critical region which usually nests
2150 * inside cgroup synchronization. Bounce actual hotplug processing
2151 * to a work item to avoid reverse locking order.
2153 * We still need to do partition_sched_domains() synchronously;
2154 * otherwise, the scheduler will get confused and put tasks to the
2155 * dead CPU. Fall back to the default single domain.
2156 * cpuset_hotplug_workfn() will rebuild it as necessary.
2158 partition_sched_domains(1, NULL, NULL);
2159 schedule_work(&cpuset_hotplug_work);
2163 * Keep top_cpuset.mems_allowed tracking node_states[N_MEMORY].
2164 * Call this routine anytime after node_states[N_MEMORY] changes.
2165 * See cpuset_update_active_cpus() for CPU hotplug handling.
2167 static int cpuset_track_online_nodes(struct notifier_block *self,
2168 unsigned long action, void *arg)
2170 schedule_work(&cpuset_hotplug_work);
2174 static struct notifier_block cpuset_track_online_nodes_nb = {
2175 .notifier_call = cpuset_track_online_nodes,
2176 .priority = 10, /* ??! */
2180 * cpuset_init_smp - initialize cpus_allowed
2182 * Description: Finish top cpuset after cpu, node maps are initialized
2184 void __init cpuset_init_smp(void)
2186 cpumask_copy(top_cpuset.cpus_allowed, cpu_active_mask);
2187 top_cpuset.mems_allowed = node_states[N_MEMORY];
2189 register_hotmemory_notifier(&cpuset_track_online_nodes_nb);
2191 cpuset_propagate_hotplug_wq =
2192 alloc_ordered_workqueue("cpuset_hotplug", 0);
2193 BUG_ON(!cpuset_propagate_hotplug_wq);
2197 * cpuset_cpus_allowed - return cpus_allowed mask from a tasks cpuset.
2198 * @tsk: pointer to task_struct from which to obtain cpuset->cpus_allowed.
2199 * @pmask: pointer to struct cpumask variable to receive cpus_allowed set.
2201 * Description: Returns the cpumask_var_t cpus_allowed of the cpuset
2202 * attached to the specified @tsk. Guaranteed to return some non-empty
2203 * subset of cpu_online_mask, even if this means going outside the
2207 void cpuset_cpus_allowed(struct task_struct *tsk, struct cpumask *pmask)
2209 mutex_lock(&callback_mutex);
2211 guarantee_online_cpus(task_cs(tsk), pmask);
2213 mutex_unlock(&callback_mutex);
2216 void cpuset_cpus_allowed_fallback(struct task_struct *tsk)
2218 const struct cpuset *cs;
2222 do_set_cpus_allowed(tsk, cs->cpus_allowed);
2226 * We own tsk->cpus_allowed, nobody can change it under us.
2228 * But we used cs && cs->cpus_allowed lockless and thus can
2229 * race with cgroup_attach_task() or update_cpumask() and get
2230 * the wrong tsk->cpus_allowed. However, both cases imply the
2231 * subsequent cpuset_change_cpumask()->set_cpus_allowed_ptr()
2232 * which takes task_rq_lock().
2234 * If we are called after it dropped the lock we must see all
2235 * changes in tsk_cs()->cpus_allowed. Otherwise we can temporary
2236 * set any mask even if it is not right from task_cs() pov,
2237 * the pending set_cpus_allowed_ptr() will fix things.
2239 * select_fallback_rq() will fix things ups and set cpu_possible_mask
2244 void cpuset_init_current_mems_allowed(void)
2246 nodes_setall(current->mems_allowed);
2250 * cpuset_mems_allowed - return mems_allowed mask from a tasks cpuset.
2251 * @tsk: pointer to task_struct from which to obtain cpuset->mems_allowed.
2253 * Description: Returns the nodemask_t mems_allowed of the cpuset
2254 * attached to the specified @tsk. Guaranteed to return some non-empty
2255 * subset of node_states[N_MEMORY], even if this means going outside the
2259 nodemask_t cpuset_mems_allowed(struct task_struct *tsk)
2263 mutex_lock(&callback_mutex);
2265 guarantee_online_mems(task_cs(tsk), &mask);
2267 mutex_unlock(&callback_mutex);
2273 * cpuset_nodemask_valid_mems_allowed - check nodemask vs. curremt mems_allowed
2274 * @nodemask: the nodemask to be checked
2276 * Are any of the nodes in the nodemask allowed in current->mems_allowed?
2278 int cpuset_nodemask_valid_mems_allowed(nodemask_t *nodemask)
2280 return nodes_intersects(*nodemask, current->mems_allowed);
2284 * nearest_hardwall_ancestor() - Returns the nearest mem_exclusive or
2285 * mem_hardwall ancestor to the specified cpuset. Call holding
2286 * callback_mutex. If no ancestor is mem_exclusive or mem_hardwall
2287 * (an unusual configuration), then returns the root cpuset.
2289 static const struct cpuset *nearest_hardwall_ancestor(const struct cpuset *cs)
2291 while (!(is_mem_exclusive(cs) || is_mem_hardwall(cs)) && parent_cs(cs))
2297 * cpuset_node_allowed_softwall - Can we allocate on a memory node?
2298 * @node: is this an allowed node?
2299 * @gfp_mask: memory allocation flags
2301 * If we're in interrupt, yes, we can always allocate. If __GFP_THISNODE is
2302 * set, yes, we can always allocate. If node is in our task's mems_allowed,
2303 * yes. If it's not a __GFP_HARDWALL request and this node is in the nearest
2304 * hardwalled cpuset ancestor to this task's cpuset, yes. If the task has been
2305 * OOM killed and has access to memory reserves as specified by the TIF_MEMDIE
2309 * If __GFP_HARDWALL is set, cpuset_node_allowed_softwall() reduces to
2310 * cpuset_node_allowed_hardwall(). Otherwise, cpuset_node_allowed_softwall()
2311 * might sleep, and might allow a node from an enclosing cpuset.
2313 * cpuset_node_allowed_hardwall() only handles the simpler case of hardwall
2314 * cpusets, and never sleeps.
2316 * The __GFP_THISNODE placement logic is really handled elsewhere,
2317 * by forcibly using a zonelist starting at a specified node, and by
2318 * (in get_page_from_freelist()) refusing to consider the zones for
2319 * any node on the zonelist except the first. By the time any such
2320 * calls get to this routine, we should just shut up and say 'yes'.
2322 * GFP_USER allocations are marked with the __GFP_HARDWALL bit,
2323 * and do not allow allocations outside the current tasks cpuset
2324 * unless the task has been OOM killed as is marked TIF_MEMDIE.
2325 * GFP_KERNEL allocations are not so marked, so can escape to the
2326 * nearest enclosing hardwalled ancestor cpuset.
2328 * Scanning up parent cpusets requires callback_mutex. The
2329 * __alloc_pages() routine only calls here with __GFP_HARDWALL bit
2330 * _not_ set if it's a GFP_KERNEL allocation, and all nodes in the
2331 * current tasks mems_allowed came up empty on the first pass over
2332 * the zonelist. So only GFP_KERNEL allocations, if all nodes in the
2333 * cpuset are short of memory, might require taking the callback_mutex
2336 * The first call here from mm/page_alloc:get_page_from_freelist()
2337 * has __GFP_HARDWALL set in gfp_mask, enforcing hardwall cpusets,
2338 * so no allocation on a node outside the cpuset is allowed (unless
2339 * in interrupt, of course).
2341 * The second pass through get_page_from_freelist() doesn't even call
2342 * here for GFP_ATOMIC calls. For those calls, the __alloc_pages()
2343 * variable 'wait' is not set, and the bit ALLOC_CPUSET is not set
2344 * in alloc_flags. That logic and the checks below have the combined
2346 * in_interrupt - any node ok (current task context irrelevant)
2347 * GFP_ATOMIC - any node ok
2348 * TIF_MEMDIE - any node ok
2349 * GFP_KERNEL - any node in enclosing hardwalled cpuset ok
2350 * GFP_USER - only nodes in current tasks mems allowed ok.
2353 * Don't call cpuset_node_allowed_softwall if you can't sleep, unless you
2354 * pass in the __GFP_HARDWALL flag set in gfp_flag, which disables
2355 * the code that might scan up ancestor cpusets and sleep.
2357 int __cpuset_node_allowed_softwall(int node, gfp_t gfp_mask)
2359 const struct cpuset *cs; /* current cpuset ancestors */
2360 int allowed; /* is allocation in zone z allowed? */
2362 if (in_interrupt() || (gfp_mask & __GFP_THISNODE))
2364 might_sleep_if(!(gfp_mask & __GFP_HARDWALL));
2365 if (node_isset(node, current->mems_allowed))
2368 * Allow tasks that have access to memory reserves because they have
2369 * been OOM killed to get memory anywhere.
2371 if (unlikely(test_thread_flag(TIF_MEMDIE)))
2373 if (gfp_mask & __GFP_HARDWALL) /* If hardwall request, stop here */
2376 if (current->flags & PF_EXITING) /* Let dying task have memory */
2379 /* Not hardwall and node outside mems_allowed: scan up cpusets */
2380 mutex_lock(&callback_mutex);
2383 cs = nearest_hardwall_ancestor(task_cs(current));
2384 task_unlock(current);
2386 allowed = node_isset(node, cs->mems_allowed);
2387 mutex_unlock(&callback_mutex);
2392 * cpuset_node_allowed_hardwall - Can we allocate on a memory node?
2393 * @node: is this an allowed node?
2394 * @gfp_mask: memory allocation flags
2396 * If we're in interrupt, yes, we can always allocate. If __GFP_THISNODE is
2397 * set, yes, we can always allocate. If node is in our task's mems_allowed,
2398 * yes. If the task has been OOM killed and has access to memory reserves as
2399 * specified by the TIF_MEMDIE flag, yes.
2402 * The __GFP_THISNODE placement logic is really handled elsewhere,
2403 * by forcibly using a zonelist starting at a specified node, and by
2404 * (in get_page_from_freelist()) refusing to consider the zones for
2405 * any node on the zonelist except the first. By the time any such
2406 * calls get to this routine, we should just shut up and say 'yes'.
2408 * Unlike the cpuset_node_allowed_softwall() variant, above,
2409 * this variant requires that the node be in the current task's
2410 * mems_allowed or that we're in interrupt. It does not scan up the
2411 * cpuset hierarchy for the nearest enclosing mem_exclusive cpuset.
2414 int __cpuset_node_allowed_hardwall(int node, gfp_t gfp_mask)
2416 if (in_interrupt() || (gfp_mask & __GFP_THISNODE))
2418 if (node_isset(node, current->mems_allowed))
2421 * Allow tasks that have access to memory reserves because they have
2422 * been OOM killed to get memory anywhere.
2424 if (unlikely(test_thread_flag(TIF_MEMDIE)))
2430 * cpuset_mem_spread_node() - On which node to begin search for a file page
2431 * cpuset_slab_spread_node() - On which node to begin search for a slab page
2433 * If a task is marked PF_SPREAD_PAGE or PF_SPREAD_SLAB (as for
2434 * tasks in a cpuset with is_spread_page or is_spread_slab set),
2435 * and if the memory allocation used cpuset_mem_spread_node()
2436 * to determine on which node to start looking, as it will for
2437 * certain page cache or slab cache pages such as used for file
2438 * system buffers and inode caches, then instead of starting on the
2439 * local node to look for a free page, rather spread the starting
2440 * node around the tasks mems_allowed nodes.
2442 * We don't have to worry about the returned node being offline
2443 * because "it can't happen", and even if it did, it would be ok.
2445 * The routines calling guarantee_online_mems() are careful to
2446 * only set nodes in task->mems_allowed that are online. So it
2447 * should not be possible for the following code to return an
2448 * offline node. But if it did, that would be ok, as this routine
2449 * is not returning the node where the allocation must be, only
2450 * the node where the search should start. The zonelist passed to
2451 * __alloc_pages() will include all nodes. If the slab allocator
2452 * is passed an offline node, it will fall back to the local node.
2453 * See kmem_cache_alloc_node().
2456 static int cpuset_spread_node(int *rotor)
2460 node = next_node(*rotor, current->mems_allowed);
2461 if (node == MAX_NUMNODES)
2462 node = first_node(current->mems_allowed);
2467 int cpuset_mem_spread_node(void)
2469 if (current->cpuset_mem_spread_rotor == NUMA_NO_NODE)
2470 current->cpuset_mem_spread_rotor =
2471 node_random(¤t->mems_allowed);
2473 return cpuset_spread_node(¤t->cpuset_mem_spread_rotor);
2476 int cpuset_slab_spread_node(void)
2478 if (current->cpuset_slab_spread_rotor == NUMA_NO_NODE)
2479 current->cpuset_slab_spread_rotor =
2480 node_random(¤t->mems_allowed);
2482 return cpuset_spread_node(¤t->cpuset_slab_spread_rotor);
2485 EXPORT_SYMBOL_GPL(cpuset_mem_spread_node);
2488 * cpuset_mems_allowed_intersects - Does @tsk1's mems_allowed intersect @tsk2's?
2489 * @tsk1: pointer to task_struct of some task.
2490 * @tsk2: pointer to task_struct of some other task.
2492 * Description: Return true if @tsk1's mems_allowed intersects the
2493 * mems_allowed of @tsk2. Used by the OOM killer to determine if
2494 * one of the task's memory usage might impact the memory available
2498 int cpuset_mems_allowed_intersects(const struct task_struct *tsk1,
2499 const struct task_struct *tsk2)
2501 return nodes_intersects(tsk1->mems_allowed, tsk2->mems_allowed);
2504 #define CPUSET_NODELIST_LEN (256)
2507 * cpuset_print_task_mems_allowed - prints task's cpuset and mems_allowed
2508 * @task: pointer to task_struct of some task.
2510 * Description: Prints @task's name, cpuset name, and cached copy of its
2511 * mems_allowed to the kernel log. Must hold task_lock(task) to allow
2512 * dereferencing task_cs(task).
2514 void cpuset_print_task_mems_allowed(struct task_struct *tsk)
2516 /* Statically allocated to prevent using excess stack. */
2517 static char cpuset_nodelist[CPUSET_NODELIST_LEN];
2518 static DEFINE_SPINLOCK(cpuset_buffer_lock);
2520 struct cgroup *cgrp = task_cs(tsk)->css.cgroup;
2523 spin_lock(&cpuset_buffer_lock);
2525 nodelist_scnprintf(cpuset_nodelist, CPUSET_NODELIST_LEN,
2527 printk(KERN_INFO "%s cpuset=%s mems_allowed=%s\n",
2528 tsk->comm, cgroup_name(cgrp), cpuset_nodelist);
2530 spin_unlock(&cpuset_buffer_lock);
2535 * Collection of memory_pressure is suppressed unless
2536 * this flag is enabled by writing "1" to the special
2537 * cpuset file 'memory_pressure_enabled' in the root cpuset.
2540 int cpuset_memory_pressure_enabled __read_mostly;
2543 * cpuset_memory_pressure_bump - keep stats of per-cpuset reclaims.
2545 * Keep a running average of the rate of synchronous (direct)
2546 * page reclaim efforts initiated by tasks in each cpuset.
2548 * This represents the rate at which some task in the cpuset
2549 * ran low on memory on all nodes it was allowed to use, and
2550 * had to enter the kernels page reclaim code in an effort to
2551 * create more free memory by tossing clean pages or swapping
2552 * or writing dirty pages.
2554 * Display to user space in the per-cpuset read-only file
2555 * "memory_pressure". Value displayed is an integer
2556 * representing the recent rate of entry into the synchronous
2557 * (direct) page reclaim by any task attached to the cpuset.
2560 void __cpuset_memory_pressure_bump(void)
2563 fmeter_markevent(&task_cs(current)->fmeter);
2564 task_unlock(current);
2567 #ifdef CONFIG_PROC_PID_CPUSET
2569 * proc_cpuset_show()
2570 * - Print tasks cpuset path into seq_file.
2571 * - Used for /proc/<pid>/cpuset.
2572 * - No need to task_lock(tsk) on this tsk->cpuset reference, as it
2573 * doesn't really matter if tsk->cpuset changes after we read it,
2574 * and we take cpuset_mutex, keeping cpuset_attach() from changing it
2577 int proc_cpuset_show(struct seq_file *m, void *unused_v)
2580 struct task_struct *tsk;
2582 struct cgroup_subsys_state *css;
2586 buf = kmalloc(PAGE_SIZE, GFP_KERNEL);
2592 tsk = get_pid_task(pid, PIDTYPE_PID);
2597 css = task_subsys_state(tsk, cpuset_subsys_id);
2598 retval = cgroup_path(css->cgroup, buf, PAGE_SIZE);
2605 put_task_struct(tsk);
2611 #endif /* CONFIG_PROC_PID_CPUSET */
2613 /* Display task mems_allowed in /proc/<pid>/status file. */
2614 void cpuset_task_status_allowed(struct seq_file *m, struct task_struct *task)
2616 seq_printf(m, "Mems_allowed:\t");
2617 seq_nodemask(m, &task->mems_allowed);
2618 seq_printf(m, "\n");
2619 seq_printf(m, "Mems_allowed_list:\t");
2620 seq_nodemask_list(m, &task->mems_allowed);
2621 seq_printf(m, "\n");