4 * Processor and Memory placement constraints for sets of tasks.
6 * Copyright (C) 2003 BULL SA.
7 * Copyright (C) 2004-2007 Silicon Graphics, Inc.
8 * Copyright (C) 2006 Google, Inc
10 * Portions derived from Patrick Mochel's sysfs code.
11 * sysfs is Copyright (c) 2001-3 Patrick Mochel
13 * 2003-10-10 Written by Simon Derr.
14 * 2003-10-22 Updates by Stephen Hemminger.
15 * 2004 May-July Rework by Paul Jackson.
16 * 2006 Rework by Paul Menage to use generic cgroups
17 * 2008 Rework of the scheduler domains and CPU hotplug handling
20 * This file is subject to the terms and conditions of the GNU General Public
21 * License. See the file COPYING in the main directory of the Linux
22 * distribution for more details.
25 #include <linux/cpu.h>
26 #include <linux/cpumask.h>
27 #include <linux/cpuset.h>
28 #include <linux/err.h>
29 #include <linux/errno.h>
30 #include <linux/file.h>
32 #include <linux/init.h>
33 #include <linux/interrupt.h>
34 #include <linux/kernel.h>
35 #include <linux/kmod.h>
36 #include <linux/list.h>
37 #include <linux/mempolicy.h>
39 #include <linux/memory.h>
40 #include <linux/export.h>
41 #include <linux/mount.h>
42 #include <linux/namei.h>
43 #include <linux/pagemap.h>
44 #include <linux/proc_fs.h>
45 #include <linux/rcupdate.h>
46 #include <linux/sched.h>
47 #include <linux/seq_file.h>
48 #include <linux/security.h>
49 #include <linux/slab.h>
50 #include <linux/spinlock.h>
51 #include <linux/stat.h>
52 #include <linux/string.h>
53 #include <linux/time.h>
54 #include <linux/backing-dev.h>
55 #include <linux/sort.h>
57 #include <asm/uaccess.h>
58 #include <linux/atomic.h>
59 #include <linux/mutex.h>
60 #include <linux/workqueue.h>
61 #include <linux/cgroup.h>
62 #include <linux/wait.h>
65 * Tracks how many cpusets are currently defined in system.
66 * When there is only one cpuset (the root cpuset) we can
67 * short circuit some hooks.
69 int number_of_cpusets __read_mostly;
71 /* Forward declare cgroup structures */
72 struct cgroup_subsys cpuset_subsys;
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 */
91 * This is old Memory Nodes tasks took on.
93 * - top_cpuset.old_mems_allowed is initialized to mems_allowed.
94 * - A new cpuset's old_mems_allowed is initialized when some
95 * task is moved into it.
96 * - old_mems_allowed is used in cpuset_migrate_mm() when we change
97 * cpuset.mems_allowed and have tasks' nodemask updated, and
98 * then old_mems_allowed is updated to mems_allowed.
100 nodemask_t old_mems_allowed;
102 struct fmeter fmeter; /* memory_pressure filter */
105 * Tasks are being attached to this cpuset. Used to prevent
106 * zeroing cpus/mems_allowed between ->can_attach() and ->attach().
108 int attach_in_progress;
110 /* partition number for rebuild_sched_domains() */
113 /* for custom sched domain */
114 int relax_domain_level;
117 static inline struct cpuset *css_cs(struct cgroup_subsys_state *css)
119 return css ? container_of(css, struct cpuset, css) : NULL;
122 /* Retrieve the cpuset for a cgroup */
123 static inline struct cpuset *cgroup_cs(struct cgroup *cgrp)
125 return css_cs(cgroup_css(cgrp, cpuset_subsys_id));
128 /* Retrieve the cpuset for a task */
129 static inline struct cpuset *task_cs(struct task_struct *task)
131 return css_cs(task_css(task, cpuset_subsys_id));
134 static inline struct cpuset *parent_cs(struct cpuset *cs)
136 return css_cs(css_parent(&cs->css));
140 static inline bool task_has_mempolicy(struct task_struct *task)
142 return task->mempolicy;
145 static inline bool task_has_mempolicy(struct task_struct *task)
152 /* bits in struct cpuset flags field */
159 CS_SCHED_LOAD_BALANCE,
164 /* convenient tests for these bits */
165 static inline bool is_cpuset_online(const struct cpuset *cs)
167 return test_bit(CS_ONLINE, &cs->flags);
170 static inline int is_cpu_exclusive(const struct cpuset *cs)
172 return test_bit(CS_CPU_EXCLUSIVE, &cs->flags);
175 static inline int is_mem_exclusive(const struct cpuset *cs)
177 return test_bit(CS_MEM_EXCLUSIVE, &cs->flags);
180 static inline int is_mem_hardwall(const struct cpuset *cs)
182 return test_bit(CS_MEM_HARDWALL, &cs->flags);
185 static inline int is_sched_load_balance(const struct cpuset *cs)
187 return test_bit(CS_SCHED_LOAD_BALANCE, &cs->flags);
190 static inline int is_memory_migrate(const struct cpuset *cs)
192 return test_bit(CS_MEMORY_MIGRATE, &cs->flags);
195 static inline int is_spread_page(const struct cpuset *cs)
197 return test_bit(CS_SPREAD_PAGE, &cs->flags);
200 static inline int is_spread_slab(const struct cpuset *cs)
202 return test_bit(CS_SPREAD_SLAB, &cs->flags);
205 static struct cpuset top_cpuset = {
206 .flags = ((1 << CS_ONLINE) | (1 << CS_CPU_EXCLUSIVE) |
207 (1 << CS_MEM_EXCLUSIVE)),
211 * cpuset_for_each_child - traverse online children of a cpuset
212 * @child_cs: loop cursor pointing to the current child
213 * @pos_css: used for iteration
214 * @parent_cs: target cpuset to walk children of
216 * Walk @child_cs through the online children of @parent_cs. Must be used
217 * with RCU read locked.
219 #define cpuset_for_each_child(child_cs, pos_css, parent_cs) \
220 css_for_each_child((pos_css), &(parent_cs)->css) \
221 if (is_cpuset_online(((child_cs) = css_cs((pos_css)))))
224 * cpuset_for_each_descendant_pre - pre-order walk of a cpuset's descendants
225 * @des_cs: loop cursor pointing to the current descendant
226 * @pos_css: used for iteration
227 * @root_cs: target cpuset to walk ancestor of
229 * Walk @des_cs through the online descendants of @root_cs. Must be used
230 * with RCU read locked. The caller may modify @pos_css by calling
231 * css_rightmost_descendant() to skip subtree.
233 #define cpuset_for_each_descendant_pre(des_cs, pos_css, root_cs) \
234 css_for_each_descendant_pre((pos_css), &(root_cs)->css) \
235 if (is_cpuset_online(((des_cs) = css_cs((pos_css)))))
238 * There are two global mutexes guarding cpuset structures - cpuset_mutex
239 * and callback_mutex. The latter may nest inside the former. We also
240 * require taking task_lock() when dereferencing a task's cpuset pointer.
241 * See "The task_lock() exception", at the end of this comment.
243 * A task must hold both mutexes to modify cpusets. If a task holds
244 * cpuset_mutex, then it blocks others wanting that mutex, ensuring that it
245 * is the only task able to also acquire callback_mutex and be able to
246 * modify cpusets. It can perform various checks on the cpuset structure
247 * first, knowing nothing will change. It can also allocate memory while
248 * just holding cpuset_mutex. While it is performing these checks, various
249 * callback routines can briefly acquire callback_mutex to query cpusets.
250 * Once it is ready to make the changes, it takes callback_mutex, blocking
253 * Calls to the kernel memory allocator can not be made while holding
254 * callback_mutex, as that would risk double tripping on callback_mutex
255 * from one of the callbacks into the cpuset code from within
258 * If a task is only holding callback_mutex, then it has read-only
261 * Now, the task_struct fields mems_allowed and mempolicy may be changed
262 * by other task, we use alloc_lock in the task_struct fields to protect
265 * The cpuset_common_file_read() handlers only hold callback_mutex across
266 * small pieces of code, such as when reading out possibly multi-word
267 * cpumasks and nodemasks.
269 * Accessing a task's cpuset should be done in accordance with the
270 * guidelines for accessing subsystem state in kernel/cgroup.c
273 static DEFINE_MUTEX(cpuset_mutex);
274 static DEFINE_MUTEX(callback_mutex);
277 * CPU / memory hotplug is handled asynchronously.
279 static void cpuset_hotplug_workfn(struct work_struct *work);
280 static DECLARE_WORK(cpuset_hotplug_work, cpuset_hotplug_workfn);
282 static DECLARE_WAIT_QUEUE_HEAD(cpuset_attach_wq);
285 * This is ugly, but preserves the userspace API for existing cpuset
286 * users. If someone tries to mount the "cpuset" filesystem, we
287 * silently switch it to mount "cgroup" instead
289 static struct dentry *cpuset_mount(struct file_system_type *fs_type,
290 int flags, const char *unused_dev_name, void *data)
292 struct file_system_type *cgroup_fs = get_fs_type("cgroup");
293 struct dentry *ret = ERR_PTR(-ENODEV);
297 "release_agent=/sbin/cpuset_release_agent";
298 ret = cgroup_fs->mount(cgroup_fs, flags,
299 unused_dev_name, mountopts);
300 put_filesystem(cgroup_fs);
305 static struct file_system_type cpuset_fs_type = {
307 .mount = cpuset_mount,
311 * Return in pmask the portion of a cpusets's cpus_allowed that
312 * are online. If none are online, walk up the cpuset hierarchy
313 * until we find one that does have some online cpus. The top
314 * cpuset always has some cpus online.
316 * One way or another, we guarantee to return some non-empty subset
317 * of cpu_online_mask.
319 * Call with callback_mutex held.
321 static void guarantee_online_cpus(struct cpuset *cs, struct cpumask *pmask)
323 while (!cpumask_intersects(cs->cpus_allowed, cpu_online_mask))
325 cpumask_and(pmask, cs->cpus_allowed, cpu_online_mask);
329 * Return in *pmask the portion of a cpusets's mems_allowed that
330 * are online, with memory. If none are online with memory, walk
331 * up the cpuset hierarchy until we find one that does have some
332 * online mems. The top cpuset always has some mems online.
334 * One way or another, we guarantee to return some non-empty subset
335 * of node_states[N_MEMORY].
337 * Call with callback_mutex held.
339 static void guarantee_online_mems(struct cpuset *cs, nodemask_t *pmask)
341 while (!nodes_intersects(cs->mems_allowed, node_states[N_MEMORY]))
343 nodes_and(*pmask, cs->mems_allowed, node_states[N_MEMORY]);
347 * update task's spread flag if cpuset's page/slab spread flag is set
349 * Called with callback_mutex/cpuset_mutex held
351 static void cpuset_update_task_spread_flag(struct cpuset *cs,
352 struct task_struct *tsk)
354 if (is_spread_page(cs))
355 tsk->flags |= PF_SPREAD_PAGE;
357 tsk->flags &= ~PF_SPREAD_PAGE;
358 if (is_spread_slab(cs))
359 tsk->flags |= PF_SPREAD_SLAB;
361 tsk->flags &= ~PF_SPREAD_SLAB;
365 * is_cpuset_subset(p, q) - Is cpuset p a subset of cpuset q?
367 * One cpuset is a subset of another if all its allowed CPUs and
368 * Memory Nodes are a subset of the other, and its exclusive flags
369 * are only set if the other's are set. Call holding cpuset_mutex.
372 static int is_cpuset_subset(const struct cpuset *p, const struct cpuset *q)
374 return cpumask_subset(p->cpus_allowed, q->cpus_allowed) &&
375 nodes_subset(p->mems_allowed, q->mems_allowed) &&
376 is_cpu_exclusive(p) <= is_cpu_exclusive(q) &&
377 is_mem_exclusive(p) <= is_mem_exclusive(q);
381 * alloc_trial_cpuset - allocate a trial cpuset
382 * @cs: the cpuset that the trial cpuset duplicates
384 static struct cpuset *alloc_trial_cpuset(struct cpuset *cs)
386 struct cpuset *trial;
388 trial = kmemdup(cs, sizeof(*cs), GFP_KERNEL);
392 if (!alloc_cpumask_var(&trial->cpus_allowed, GFP_KERNEL)) {
396 cpumask_copy(trial->cpus_allowed, cs->cpus_allowed);
402 * free_trial_cpuset - free the trial cpuset
403 * @trial: the trial cpuset to be freed
405 static void free_trial_cpuset(struct cpuset *trial)
407 free_cpumask_var(trial->cpus_allowed);
412 * validate_change() - Used to validate that any proposed cpuset change
413 * follows the structural rules for cpusets.
415 * If we replaced the flag and mask values of the current cpuset
416 * (cur) with those values in the trial cpuset (trial), would
417 * our various subset and exclusive rules still be valid? Presumes
420 * 'cur' is the address of an actual, in-use cpuset. Operations
421 * such as list traversal that depend on the actual address of the
422 * cpuset in the list must use cur below, not trial.
424 * 'trial' is the address of bulk structure copy of cur, with
425 * perhaps one or more of the fields cpus_allowed, mems_allowed,
426 * or flags changed to new, trial values.
428 * Return 0 if valid, -errno if not.
431 static int validate_change(struct cpuset *cur, struct cpuset *trial)
433 struct cgroup_subsys_state *css;
434 struct cpuset *c, *par;
439 /* Each of our child cpusets must be a subset of us */
441 cpuset_for_each_child(c, css, cur)
442 if (!is_cpuset_subset(c, trial))
445 /* Remaining checks don't apply to root cpuset */
447 if (cur == &top_cpuset)
450 par = parent_cs(cur);
452 /* We must be a subset of our parent cpuset */
454 if (!is_cpuset_subset(trial, par))
458 * If either I or some sibling (!= me) is exclusive, we can't
462 cpuset_for_each_child(c, css, par) {
463 if ((is_cpu_exclusive(trial) || is_cpu_exclusive(c)) &&
465 cpumask_intersects(trial->cpus_allowed, c->cpus_allowed))
467 if ((is_mem_exclusive(trial) || is_mem_exclusive(c)) &&
469 nodes_intersects(trial->mems_allowed, c->mems_allowed))
474 * Cpusets with tasks - existing or newly being attached - can't
475 * have empty cpus_allowed or mems_allowed.
478 if ((cgroup_task_count(cur->css.cgroup) || cur->attach_in_progress) &&
479 (cpumask_empty(trial->cpus_allowed) &&
480 nodes_empty(trial->mems_allowed)))
491 * Helper routine for generate_sched_domains().
492 * Do cpusets a, b have overlapping cpus_allowed masks?
494 static int cpusets_overlap(struct cpuset *a, struct cpuset *b)
496 return cpumask_intersects(a->cpus_allowed, b->cpus_allowed);
500 update_domain_attr(struct sched_domain_attr *dattr, struct cpuset *c)
502 if (dattr->relax_domain_level < c->relax_domain_level)
503 dattr->relax_domain_level = c->relax_domain_level;
507 static void update_domain_attr_tree(struct sched_domain_attr *dattr,
508 struct cpuset *root_cs)
511 struct cgroup_subsys_state *pos_css;
514 cpuset_for_each_descendant_pre(cp, pos_css, root_cs) {
515 /* skip the whole subtree if @cp doesn't have any CPU */
516 if (cpumask_empty(cp->cpus_allowed)) {
517 pos_css = css_rightmost_descendant(pos_css);
521 if (is_sched_load_balance(cp))
522 update_domain_attr(dattr, cp);
528 * generate_sched_domains()
530 * This function builds a partial partition of the systems CPUs
531 * A 'partial partition' is a set of non-overlapping subsets whose
532 * union is a subset of that set.
533 * The output of this function needs to be passed to kernel/sched/core.c
534 * partition_sched_domains() routine, which will rebuild the scheduler's
535 * load balancing domains (sched domains) as specified by that partial
538 * See "What is sched_load_balance" in Documentation/cgroups/cpusets.txt
539 * for a background explanation of this.
541 * Does not return errors, on the theory that the callers of this
542 * routine would rather not worry about failures to rebuild sched
543 * domains when operating in the severe memory shortage situations
544 * that could cause allocation failures below.
546 * Must be called with cpuset_mutex held.
548 * The three key local variables below are:
549 * q - a linked-list queue of cpuset pointers, used to implement a
550 * top-down scan of all cpusets. This scan loads a pointer
551 * to each cpuset marked is_sched_load_balance into the
552 * array 'csa'. For our purposes, rebuilding the schedulers
553 * sched domains, we can ignore !is_sched_load_balance cpusets.
554 * csa - (for CpuSet Array) Array of pointers to all the cpusets
555 * that need to be load balanced, for convenient iterative
556 * access by the subsequent code that finds the best partition,
557 * i.e the set of domains (subsets) of CPUs such that the
558 * cpus_allowed of every cpuset marked is_sched_load_balance
559 * is a subset of one of these domains, while there are as
560 * many such domains as possible, each as small as possible.
561 * doms - Conversion of 'csa' to an array of cpumasks, for passing to
562 * the kernel/sched/core.c routine partition_sched_domains() in a
563 * convenient format, that can be easily compared to the prior
564 * value to determine what partition elements (sched domains)
565 * were changed (added or removed.)
567 * Finding the best partition (set of domains):
568 * The triple nested loops below over i, j, k scan over the
569 * load balanced cpusets (using the array of cpuset pointers in
570 * csa[]) looking for pairs of cpusets that have overlapping
571 * cpus_allowed, but which don't have the same 'pn' partition
572 * number and gives them in the same partition number. It keeps
573 * looping on the 'restart' label until it can no longer find
576 * The union of the cpus_allowed masks from the set of
577 * all cpusets having the same 'pn' value then form the one
578 * element of the partition (one sched domain) to be passed to
579 * partition_sched_domains().
581 static int generate_sched_domains(cpumask_var_t **domains,
582 struct sched_domain_attr **attributes)
584 struct cpuset *cp; /* scans q */
585 struct cpuset **csa; /* array of all cpuset ptrs */
586 int csn; /* how many cpuset ptrs in csa so far */
587 int i, j, k; /* indices for partition finding loops */
588 cpumask_var_t *doms; /* resulting partition; i.e. sched domains */
589 struct sched_domain_attr *dattr; /* attributes for custom domains */
590 int ndoms = 0; /* number of sched domains in result */
591 int nslot; /* next empty doms[] struct cpumask slot */
592 struct cgroup_subsys_state *pos_css;
598 /* Special case for the 99% of systems with one, full, sched domain */
599 if (is_sched_load_balance(&top_cpuset)) {
601 doms = alloc_sched_domains(ndoms);
605 dattr = kmalloc(sizeof(struct sched_domain_attr), GFP_KERNEL);
607 *dattr = SD_ATTR_INIT;
608 update_domain_attr_tree(dattr, &top_cpuset);
610 cpumask_copy(doms[0], top_cpuset.cpus_allowed);
615 csa = kmalloc(number_of_cpusets * sizeof(cp), GFP_KERNEL);
621 cpuset_for_each_descendant_pre(cp, pos_css, &top_cpuset) {
623 * Continue traversing beyond @cp iff @cp has some CPUs and
624 * isn't load balancing. The former is obvious. The
625 * latter: All child cpusets contain a subset of the
626 * parent's cpus, so just skip them, and then we call
627 * update_domain_attr_tree() to calc relax_domain_level of
628 * the corresponding sched domain.
630 if (!cpumask_empty(cp->cpus_allowed) &&
631 !is_sched_load_balance(cp))
634 if (is_sched_load_balance(cp))
637 /* skip @cp's subtree */
638 pos_css = css_rightmost_descendant(pos_css);
642 for (i = 0; i < csn; i++)
647 /* Find the best partition (set of sched domains) */
648 for (i = 0; i < csn; i++) {
649 struct cpuset *a = csa[i];
652 for (j = 0; j < csn; j++) {
653 struct cpuset *b = csa[j];
656 if (apn != bpn && cpusets_overlap(a, b)) {
657 for (k = 0; k < csn; k++) {
658 struct cpuset *c = csa[k];
663 ndoms--; /* one less element */
670 * Now we know how many domains to create.
671 * Convert <csn, csa> to <ndoms, doms> and populate cpu masks.
673 doms = alloc_sched_domains(ndoms);
678 * The rest of the code, including the scheduler, can deal with
679 * dattr==NULL case. No need to abort if alloc fails.
681 dattr = kmalloc(ndoms * sizeof(struct sched_domain_attr), GFP_KERNEL);
683 for (nslot = 0, i = 0; i < csn; i++) {
684 struct cpuset *a = csa[i];
689 /* Skip completed partitions */
695 if (nslot == ndoms) {
696 static int warnings = 10;
699 "rebuild_sched_domains confused:"
700 " nslot %d, ndoms %d, csn %d, i %d,"
702 nslot, ndoms, csn, i, apn);
710 *(dattr + nslot) = SD_ATTR_INIT;
711 for (j = i; j < csn; j++) {
712 struct cpuset *b = csa[j];
715 cpumask_or(dp, dp, b->cpus_allowed);
717 update_domain_attr_tree(dattr + nslot, b);
719 /* Done with this partition */
725 BUG_ON(nslot != ndoms);
731 * Fallback to the default domain if kmalloc() failed.
732 * See comments in partition_sched_domains().
743 * Rebuild scheduler domains.
745 * If the flag 'sched_load_balance' of any cpuset with non-empty
746 * 'cpus' changes, or if the 'cpus' allowed changes in any cpuset
747 * which has that flag enabled, or if any cpuset with a non-empty
748 * 'cpus' is removed, then call this routine to rebuild the
749 * scheduler's dynamic sched domains.
751 * Call with cpuset_mutex held. Takes get_online_cpus().
753 static void rebuild_sched_domains_locked(void)
755 struct sched_domain_attr *attr;
759 lockdep_assert_held(&cpuset_mutex);
763 * We have raced with CPU hotplug. Don't do anything to avoid
764 * passing doms with offlined cpu to partition_sched_domains().
765 * Anyways, hotplug work item will rebuild sched domains.
767 if (!cpumask_equal(top_cpuset.cpus_allowed, cpu_active_mask))
770 /* Generate domain masks and attrs */
771 ndoms = generate_sched_domains(&doms, &attr);
773 /* Have scheduler rebuild the domains */
774 partition_sched_domains(ndoms, doms, attr);
778 #else /* !CONFIG_SMP */
779 static void rebuild_sched_domains_locked(void)
782 #endif /* CONFIG_SMP */
784 void rebuild_sched_domains(void)
786 mutex_lock(&cpuset_mutex);
787 rebuild_sched_domains_locked();
788 mutex_unlock(&cpuset_mutex);
792 * effective_cpumask_cpuset - return nearest ancestor with non-empty cpus
793 * @cs: the cpuset in interest
795 * A cpuset's effective cpumask is the cpumask of the nearest ancestor
796 * with non-empty cpus. We use effective cpumask whenever:
797 * - we update tasks' cpus_allowed. (they take on the ancestor's cpumask
798 * if the cpuset they reside in has no cpus)
799 * - we want to retrieve task_cs(tsk)'s cpus_allowed.
801 * Called with cpuset_mutex held. cpuset_cpus_allowed_fallback() is an
802 * exception. See comments there.
804 static struct cpuset *effective_cpumask_cpuset(struct cpuset *cs)
806 while (cpumask_empty(cs->cpus_allowed))
812 * effective_nodemask_cpuset - return nearest ancestor with non-empty mems
813 * @cs: the cpuset in interest
815 * A cpuset's effective nodemask is the nodemask of the nearest ancestor
816 * with non-empty memss. We use effective nodemask whenever:
817 * - we update tasks' mems_allowed. (they take on the ancestor's nodemask
818 * if the cpuset they reside in has no mems)
819 * - we want to retrieve task_cs(tsk)'s mems_allowed.
821 * Called with cpuset_mutex held.
823 static struct cpuset *effective_nodemask_cpuset(struct cpuset *cs)
825 while (nodes_empty(cs->mems_allowed))
831 * cpuset_change_cpumask - make a task's cpus_allowed the same as its cpuset's
833 * @scan: struct cgroup_scanner containing the cgroup of the task
835 * Called by cgroup_scan_tasks() for each task in a cgroup whose
836 * cpus_allowed mask needs to be changed.
838 * We don't need to re-check for the cgroup/cpuset membership, since we're
839 * holding cpuset_mutex at this point.
841 static void cpuset_change_cpumask(struct task_struct *tsk,
842 struct cgroup_scanner *scan)
844 struct cpuset *cpus_cs;
846 cpus_cs = effective_cpumask_cpuset(cgroup_cs(scan->cgrp));
847 set_cpus_allowed_ptr(tsk, cpus_cs->cpus_allowed);
851 * update_tasks_cpumask - Update the cpumasks of tasks in the cpuset.
852 * @cs: the cpuset in which each task's cpus_allowed mask needs to be changed
853 * @heap: if NULL, defer allocating heap memory to cgroup_scan_tasks()
855 * Called with cpuset_mutex held
857 * The cgroup_scan_tasks() function will scan all the tasks in a cgroup,
858 * calling callback functions for each.
860 * No return value. It's guaranteed that cgroup_scan_tasks() always returns 0
863 static void update_tasks_cpumask(struct cpuset *cs, struct ptr_heap *heap)
865 struct cgroup_scanner scan;
867 scan.cgrp = cs->css.cgroup;
868 scan.test_task = NULL;
869 scan.process_task = cpuset_change_cpumask;
871 cgroup_scan_tasks(&scan);
875 * update_tasks_cpumask_hier - Update the cpumasks of tasks in the hierarchy.
876 * @root_cs: the root cpuset of the hierarchy
877 * @update_root: update root cpuset or not?
878 * @heap: the heap used by cgroup_scan_tasks()
880 * This will update cpumasks of tasks in @root_cs and all other empty cpusets
881 * which take on cpumask of @root_cs.
883 * Called with cpuset_mutex held
885 static void update_tasks_cpumask_hier(struct cpuset *root_cs,
886 bool update_root, struct ptr_heap *heap)
889 struct cgroup_subsys_state *pos_css;
892 update_tasks_cpumask(root_cs, heap);
895 cpuset_for_each_descendant_pre(cp, pos_css, root_cs) {
896 /* skip the whole subtree if @cp have some CPU */
897 if (!cpumask_empty(cp->cpus_allowed)) {
898 pos_css = css_rightmost_descendant(pos_css);
901 if (!css_tryget(&cp->css))
905 update_tasks_cpumask(cp, heap);
914 * update_cpumask - update the cpus_allowed mask of a cpuset and all tasks in it
915 * @cs: the cpuset to consider
916 * @buf: buffer of cpu numbers written to this cpuset
918 static int update_cpumask(struct cpuset *cs, struct cpuset *trialcs,
921 struct ptr_heap heap;
923 int is_load_balanced;
925 /* top_cpuset.cpus_allowed tracks cpu_online_mask; it's read-only */
926 if (cs == &top_cpuset)
930 * An empty cpus_allowed is ok only if the cpuset has no tasks.
931 * Since cpulist_parse() fails on an empty mask, we special case
932 * that parsing. The validate_change() call ensures that cpusets
933 * with tasks have cpus.
936 cpumask_clear(trialcs->cpus_allowed);
938 retval = cpulist_parse(buf, trialcs->cpus_allowed);
942 if (!cpumask_subset(trialcs->cpus_allowed, cpu_active_mask))
946 /* Nothing to do if the cpus didn't change */
947 if (cpumask_equal(cs->cpus_allowed, trialcs->cpus_allowed))
950 retval = validate_change(cs, trialcs);
954 retval = heap_init(&heap, PAGE_SIZE, GFP_KERNEL, NULL);
958 is_load_balanced = is_sched_load_balance(trialcs);
960 mutex_lock(&callback_mutex);
961 cpumask_copy(cs->cpus_allowed, trialcs->cpus_allowed);
962 mutex_unlock(&callback_mutex);
964 update_tasks_cpumask_hier(cs, true, &heap);
968 if (is_load_balanced)
969 rebuild_sched_domains_locked();
976 * Migrate memory region from one set of nodes to another.
978 * Temporarilly set tasks mems_allowed to target nodes of migration,
979 * so that the migration code can allocate pages on these nodes.
981 * Call holding cpuset_mutex, so current's cpuset won't change
982 * during this call, as manage_mutex holds off any cpuset_attach()
983 * calls. Therefore we don't need to take task_lock around the
984 * call to guarantee_online_mems(), as we know no one is changing
987 * While the mm_struct we are migrating is typically from some
988 * other task, the task_struct mems_allowed that we are hacking
989 * is for our current task, which must allocate new pages for that
990 * migrating memory region.
993 static void cpuset_migrate_mm(struct mm_struct *mm, const nodemask_t *from,
994 const nodemask_t *to)
996 struct task_struct *tsk = current;
997 struct cpuset *mems_cs;
999 tsk->mems_allowed = *to;
1001 do_migrate_pages(mm, from, to, MPOL_MF_MOVE_ALL);
1003 mems_cs = effective_nodemask_cpuset(task_cs(tsk));
1004 guarantee_online_mems(mems_cs, &tsk->mems_allowed);
1008 * cpuset_change_task_nodemask - change task's mems_allowed and mempolicy
1009 * @tsk: the task to change
1010 * @newmems: new nodes that the task will be set
1012 * In order to avoid seeing no nodes if the old and new nodes are disjoint,
1013 * we structure updates as setting all new allowed nodes, then clearing newly
1016 static void cpuset_change_task_nodemask(struct task_struct *tsk,
1017 nodemask_t *newmems)
1022 * Allow tasks that have access to memory reserves because they have
1023 * been OOM killed to get memory anywhere.
1025 if (unlikely(test_thread_flag(TIF_MEMDIE)))
1027 if (current->flags & PF_EXITING) /* Let dying task have memory */
1032 * Determine if a loop is necessary if another thread is doing
1033 * get_mems_allowed(). If at least one node remains unchanged and
1034 * tsk does not have a mempolicy, then an empty nodemask will not be
1035 * possible when mems_allowed is larger than a word.
1037 need_loop = task_has_mempolicy(tsk) ||
1038 !nodes_intersects(*newmems, tsk->mems_allowed);
1041 write_seqcount_begin(&tsk->mems_allowed_seq);
1043 nodes_or(tsk->mems_allowed, tsk->mems_allowed, *newmems);
1044 mpol_rebind_task(tsk, newmems, MPOL_REBIND_STEP1);
1046 mpol_rebind_task(tsk, newmems, MPOL_REBIND_STEP2);
1047 tsk->mems_allowed = *newmems;
1050 write_seqcount_end(&tsk->mems_allowed_seq);
1056 * Update task's mems_allowed and rebind its mempolicy and vmas' mempolicy
1057 * of it to cpuset's new mems_allowed, and migrate pages to new nodes if
1058 * memory_migrate flag is set. Called with cpuset_mutex held.
1060 static void cpuset_change_nodemask(struct task_struct *p,
1061 struct cgroup_scanner *scan)
1063 struct cpuset *cs = cgroup_cs(scan->cgrp);
1064 struct mm_struct *mm;
1066 nodemask_t *newmems = scan->data;
1068 cpuset_change_task_nodemask(p, newmems);
1070 mm = get_task_mm(p);
1074 migrate = is_memory_migrate(cs);
1076 mpol_rebind_mm(mm, &cs->mems_allowed);
1078 cpuset_migrate_mm(mm, &cs->old_mems_allowed, newmems);
1082 static void *cpuset_being_rebound;
1085 * update_tasks_nodemask - Update the nodemasks of tasks in the cpuset.
1086 * @cs: the cpuset in which each task's mems_allowed mask needs to be changed
1087 * @heap: if NULL, defer allocating heap memory to cgroup_scan_tasks()
1089 * Called with cpuset_mutex held
1090 * No return value. It's guaranteed that cgroup_scan_tasks() always returns 0
1093 static void update_tasks_nodemask(struct cpuset *cs, struct ptr_heap *heap)
1095 static nodemask_t newmems; /* protected by cpuset_mutex */
1096 struct cgroup_scanner scan;
1097 struct cpuset *mems_cs = effective_nodemask_cpuset(cs);
1099 cpuset_being_rebound = cs; /* causes mpol_dup() rebind */
1101 guarantee_online_mems(mems_cs, &newmems);
1103 scan.cgrp = cs->css.cgroup;
1104 scan.test_task = NULL;
1105 scan.process_task = cpuset_change_nodemask;
1107 scan.data = &newmems;
1110 * The mpol_rebind_mm() call takes mmap_sem, which we couldn't
1111 * take while holding tasklist_lock. Forks can happen - the
1112 * mpol_dup() cpuset_being_rebound check will catch such forks,
1113 * and rebind their vma mempolicies too. Because we still hold
1114 * the global cpuset_mutex, we know that no other rebind effort
1115 * will be contending for the global variable cpuset_being_rebound.
1116 * It's ok if we rebind the same mm twice; mpol_rebind_mm()
1117 * is idempotent. Also migrate pages in each mm to new nodes.
1119 cgroup_scan_tasks(&scan);
1122 * All the tasks' nodemasks have been updated, update
1123 * cs->old_mems_allowed.
1125 cs->old_mems_allowed = newmems;
1127 /* We're done rebinding vmas to this cpuset's new mems_allowed. */
1128 cpuset_being_rebound = NULL;
1132 * update_tasks_nodemask_hier - Update the nodemasks of tasks in the hierarchy.
1133 * @cs: the root cpuset of the hierarchy
1134 * @update_root: update the root cpuset or not?
1135 * @heap: the heap used by cgroup_scan_tasks()
1137 * This will update nodemasks of tasks in @root_cs and all other empty cpusets
1138 * which take on nodemask of @root_cs.
1140 * Called with cpuset_mutex held
1142 static void update_tasks_nodemask_hier(struct cpuset *root_cs,
1143 bool update_root, struct ptr_heap *heap)
1146 struct cgroup_subsys_state *pos_css;
1149 update_tasks_nodemask(root_cs, heap);
1152 cpuset_for_each_descendant_pre(cp, pos_css, root_cs) {
1153 /* skip the whole subtree if @cp have some CPU */
1154 if (!nodes_empty(cp->mems_allowed)) {
1155 pos_css = css_rightmost_descendant(pos_css);
1158 if (!css_tryget(&cp->css))
1162 update_tasks_nodemask(cp, heap);
1171 * Handle user request to change the 'mems' memory placement
1172 * of a cpuset. Needs to validate the request, update the
1173 * cpusets mems_allowed, and for each task in the cpuset,
1174 * update mems_allowed and rebind task's mempolicy and any vma
1175 * mempolicies and if the cpuset is marked 'memory_migrate',
1176 * migrate the tasks pages to the new memory.
1178 * Call with cpuset_mutex held. May take callback_mutex during call.
1179 * Will take tasklist_lock, scan tasklist for tasks in cpuset cs,
1180 * lock each such tasks mm->mmap_sem, scan its vma's and rebind
1181 * their mempolicies to the cpusets new mems_allowed.
1183 static int update_nodemask(struct cpuset *cs, struct cpuset *trialcs,
1187 struct ptr_heap heap;
1190 * top_cpuset.mems_allowed tracks node_stats[N_MEMORY];
1193 if (cs == &top_cpuset) {
1199 * An empty mems_allowed is ok iff there are no tasks in the cpuset.
1200 * Since nodelist_parse() fails on an empty mask, we special case
1201 * that parsing. The validate_change() call ensures that cpusets
1202 * with tasks have memory.
1205 nodes_clear(trialcs->mems_allowed);
1207 retval = nodelist_parse(buf, trialcs->mems_allowed);
1211 if (!nodes_subset(trialcs->mems_allowed,
1212 node_states[N_MEMORY])) {
1218 if (nodes_equal(cs->mems_allowed, trialcs->mems_allowed)) {
1219 retval = 0; /* Too easy - nothing to do */
1222 retval = validate_change(cs, trialcs);
1226 retval = heap_init(&heap, PAGE_SIZE, GFP_KERNEL, NULL);
1230 mutex_lock(&callback_mutex);
1231 cs->mems_allowed = trialcs->mems_allowed;
1232 mutex_unlock(&callback_mutex);
1234 update_tasks_nodemask_hier(cs, true, &heap);
1241 int current_cpuset_is_being_rebound(void)
1243 return task_cs(current) == cpuset_being_rebound;
1246 static int update_relax_domain_level(struct cpuset *cs, s64 val)
1249 if (val < -1 || val >= sched_domain_level_max)
1253 if (val != cs->relax_domain_level) {
1254 cs->relax_domain_level = val;
1255 if (!cpumask_empty(cs->cpus_allowed) &&
1256 is_sched_load_balance(cs))
1257 rebuild_sched_domains_locked();
1264 * cpuset_change_flag - make a task's spread flags the same as its cpuset's
1265 * @tsk: task to be updated
1266 * @scan: struct cgroup_scanner containing the cgroup of the task
1268 * Called by cgroup_scan_tasks() for each task in a cgroup.
1270 * We don't need to re-check for the cgroup/cpuset membership, since we're
1271 * holding cpuset_mutex at this point.
1273 static void cpuset_change_flag(struct task_struct *tsk,
1274 struct cgroup_scanner *scan)
1276 cpuset_update_task_spread_flag(cgroup_cs(scan->cgrp), tsk);
1280 * update_tasks_flags - update the spread flags of tasks in the cpuset.
1281 * @cs: the cpuset in which each task's spread flags needs to be changed
1282 * @heap: if NULL, defer allocating heap memory to cgroup_scan_tasks()
1284 * Called with cpuset_mutex held
1286 * The cgroup_scan_tasks() function will scan all the tasks in a cgroup,
1287 * calling callback functions for each.
1289 * No return value. It's guaranteed that cgroup_scan_tasks() always returns 0
1292 static void update_tasks_flags(struct cpuset *cs, struct ptr_heap *heap)
1294 struct cgroup_scanner scan;
1296 scan.cgrp = cs->css.cgroup;
1297 scan.test_task = NULL;
1298 scan.process_task = cpuset_change_flag;
1300 cgroup_scan_tasks(&scan);
1304 * update_flag - read a 0 or a 1 in a file and update associated flag
1305 * bit: the bit to update (see cpuset_flagbits_t)
1306 * cs: the cpuset to update
1307 * turning_on: whether the flag is being set or cleared
1309 * Call with cpuset_mutex held.
1312 static int update_flag(cpuset_flagbits_t bit, struct cpuset *cs,
1315 struct cpuset *trialcs;
1316 int balance_flag_changed;
1317 int spread_flag_changed;
1318 struct ptr_heap heap;
1321 trialcs = alloc_trial_cpuset(cs);
1326 set_bit(bit, &trialcs->flags);
1328 clear_bit(bit, &trialcs->flags);
1330 err = validate_change(cs, trialcs);
1334 err = heap_init(&heap, PAGE_SIZE, GFP_KERNEL, NULL);
1338 balance_flag_changed = (is_sched_load_balance(cs) !=
1339 is_sched_load_balance(trialcs));
1341 spread_flag_changed = ((is_spread_slab(cs) != is_spread_slab(trialcs))
1342 || (is_spread_page(cs) != is_spread_page(trialcs)));
1344 mutex_lock(&callback_mutex);
1345 cs->flags = trialcs->flags;
1346 mutex_unlock(&callback_mutex);
1348 if (!cpumask_empty(trialcs->cpus_allowed) && balance_flag_changed)
1349 rebuild_sched_domains_locked();
1351 if (spread_flag_changed)
1352 update_tasks_flags(cs, &heap);
1355 free_trial_cpuset(trialcs);
1360 * Frequency meter - How fast is some event occurring?
1362 * These routines manage a digitally filtered, constant time based,
1363 * event frequency meter. There are four routines:
1364 * fmeter_init() - initialize a frequency meter.
1365 * fmeter_markevent() - called each time the event happens.
1366 * fmeter_getrate() - returns the recent rate of such events.
1367 * fmeter_update() - internal routine used to update fmeter.
1369 * A common data structure is passed to each of these routines,
1370 * which is used to keep track of the state required to manage the
1371 * frequency meter and its digital filter.
1373 * The filter works on the number of events marked per unit time.
1374 * The filter is single-pole low-pass recursive (IIR). The time unit
1375 * is 1 second. Arithmetic is done using 32-bit integers scaled to
1376 * simulate 3 decimal digits of precision (multiplied by 1000).
1378 * With an FM_COEF of 933, and a time base of 1 second, the filter
1379 * has a half-life of 10 seconds, meaning that if the events quit
1380 * happening, then the rate returned from the fmeter_getrate()
1381 * will be cut in half each 10 seconds, until it converges to zero.
1383 * It is not worth doing a real infinitely recursive filter. If more
1384 * than FM_MAXTICKS ticks have elapsed since the last filter event,
1385 * just compute FM_MAXTICKS ticks worth, by which point the level
1388 * Limit the count of unprocessed events to FM_MAXCNT, so as to avoid
1389 * arithmetic overflow in the fmeter_update() routine.
1391 * Given the simple 32 bit integer arithmetic used, this meter works
1392 * best for reporting rates between one per millisecond (msec) and
1393 * one per 32 (approx) seconds. At constant rates faster than one
1394 * per msec it maxes out at values just under 1,000,000. At constant
1395 * rates between one per msec, and one per second it will stabilize
1396 * to a value N*1000, where N is the rate of events per second.
1397 * At constant rates between one per second and one per 32 seconds,
1398 * it will be choppy, moving up on the seconds that have an event,
1399 * and then decaying until the next event. At rates slower than
1400 * about one in 32 seconds, it decays all the way back to zero between
1404 #define FM_COEF 933 /* coefficient for half-life of 10 secs */
1405 #define FM_MAXTICKS ((time_t)99) /* useless computing more ticks than this */
1406 #define FM_MAXCNT 1000000 /* limit cnt to avoid overflow */
1407 #define FM_SCALE 1000 /* faux fixed point scale */
1409 /* Initialize a frequency meter */
1410 static void fmeter_init(struct fmeter *fmp)
1415 spin_lock_init(&fmp->lock);
1418 /* Internal meter update - process cnt events and update value */
1419 static void fmeter_update(struct fmeter *fmp)
1421 time_t now = get_seconds();
1422 time_t ticks = now - fmp->time;
1427 ticks = min(FM_MAXTICKS, ticks);
1429 fmp->val = (FM_COEF * fmp->val) / FM_SCALE;
1432 fmp->val += ((FM_SCALE - FM_COEF) * fmp->cnt) / FM_SCALE;
1436 /* Process any previous ticks, then bump cnt by one (times scale). */
1437 static void fmeter_markevent(struct fmeter *fmp)
1439 spin_lock(&fmp->lock);
1441 fmp->cnt = min(FM_MAXCNT, fmp->cnt + FM_SCALE);
1442 spin_unlock(&fmp->lock);
1445 /* Process any previous ticks, then return current value. */
1446 static int fmeter_getrate(struct fmeter *fmp)
1450 spin_lock(&fmp->lock);
1453 spin_unlock(&fmp->lock);
1457 /* Called by cgroups to determine if a cpuset is usable; cpuset_mutex held */
1458 static int cpuset_can_attach(struct cgroup_subsys_state *css,
1459 struct cgroup_taskset *tset)
1461 struct cpuset *cs = css_cs(css);
1462 struct task_struct *task;
1465 mutex_lock(&cpuset_mutex);
1468 * We allow to move tasks into an empty cpuset if sane_behavior
1472 if (!cgroup_sane_behavior(css->cgroup) &&
1473 (cpumask_empty(cs->cpus_allowed) || nodes_empty(cs->mems_allowed)))
1476 cgroup_taskset_for_each(task, css->cgroup, tset) {
1478 * Kthreads which disallow setaffinity shouldn't be moved
1479 * to a new cpuset; we don't want to change their cpu
1480 * affinity and isolating such threads by their set of
1481 * allowed nodes is unnecessary. Thus, cpusets are not
1482 * applicable for such threads. This prevents checking for
1483 * success of set_cpus_allowed_ptr() on all attached tasks
1484 * before cpus_allowed may be changed.
1487 if (task->flags & PF_NO_SETAFFINITY)
1489 ret = security_task_setscheduler(task);
1495 * Mark attach is in progress. This makes validate_change() fail
1496 * changes which zero cpus/mems_allowed.
1498 cs->attach_in_progress++;
1501 mutex_unlock(&cpuset_mutex);
1505 static void cpuset_cancel_attach(struct cgroup_subsys_state *css,
1506 struct cgroup_taskset *tset)
1508 mutex_lock(&cpuset_mutex);
1509 css_cs(css)->attach_in_progress--;
1510 mutex_unlock(&cpuset_mutex);
1514 * Protected by cpuset_mutex. cpus_attach is used only by cpuset_attach()
1515 * but we can't allocate it dynamically there. Define it global and
1516 * allocate from cpuset_init().
1518 static cpumask_var_t cpus_attach;
1520 static void cpuset_attach(struct cgroup_subsys_state *css,
1521 struct cgroup_taskset *tset)
1523 /* static buf protected by cpuset_mutex */
1524 static nodemask_t cpuset_attach_nodemask_to;
1525 struct mm_struct *mm;
1526 struct task_struct *task;
1527 struct task_struct *leader = cgroup_taskset_first(tset);
1528 struct cgroup *oldcgrp = cgroup_taskset_cur_cgroup(tset);
1529 struct cpuset *cs = css_cs(css);
1530 struct cpuset *oldcs = cgroup_cs(oldcgrp);
1531 struct cpuset *cpus_cs = effective_cpumask_cpuset(cs);
1532 struct cpuset *mems_cs = effective_nodemask_cpuset(cs);
1534 mutex_lock(&cpuset_mutex);
1536 /* prepare for attach */
1537 if (cs == &top_cpuset)
1538 cpumask_copy(cpus_attach, cpu_possible_mask);
1540 guarantee_online_cpus(cpus_cs, cpus_attach);
1542 guarantee_online_mems(mems_cs, &cpuset_attach_nodemask_to);
1544 cgroup_taskset_for_each(task, css->cgroup, tset) {
1546 * can_attach beforehand should guarantee that this doesn't
1547 * fail. TODO: have a better way to handle failure here
1549 WARN_ON_ONCE(set_cpus_allowed_ptr(task, cpus_attach));
1551 cpuset_change_task_nodemask(task, &cpuset_attach_nodemask_to);
1552 cpuset_update_task_spread_flag(cs, task);
1556 * Change mm, possibly for multiple threads in a threadgroup. This is
1557 * expensive and may sleep.
1559 cpuset_attach_nodemask_to = cs->mems_allowed;
1560 mm = get_task_mm(leader);
1562 struct cpuset *mems_oldcs = effective_nodemask_cpuset(oldcs);
1564 mpol_rebind_mm(mm, &cpuset_attach_nodemask_to);
1567 * old_mems_allowed is the same with mems_allowed here, except
1568 * if this task is being moved automatically due to hotplug.
1569 * In that case @mems_allowed has been updated and is empty,
1570 * so @old_mems_allowed is the right nodesets that we migrate
1573 if (is_memory_migrate(cs)) {
1574 cpuset_migrate_mm(mm, &mems_oldcs->old_mems_allowed,
1575 &cpuset_attach_nodemask_to);
1580 cs->old_mems_allowed = cpuset_attach_nodemask_to;
1582 cs->attach_in_progress--;
1583 if (!cs->attach_in_progress)
1584 wake_up(&cpuset_attach_wq);
1586 mutex_unlock(&cpuset_mutex);
1589 /* The various types of files and directories in a cpuset file system */
1592 FILE_MEMORY_MIGRATE,
1598 FILE_SCHED_LOAD_BALANCE,
1599 FILE_SCHED_RELAX_DOMAIN_LEVEL,
1600 FILE_MEMORY_PRESSURE_ENABLED,
1601 FILE_MEMORY_PRESSURE,
1604 } cpuset_filetype_t;
1606 static int cpuset_write_u64(struct cgroup_subsys_state *css, struct cftype *cft,
1609 struct cpuset *cs = css_cs(css);
1610 cpuset_filetype_t type = cft->private;
1611 int retval = -ENODEV;
1613 mutex_lock(&cpuset_mutex);
1614 if (!is_cpuset_online(cs))
1618 case FILE_CPU_EXCLUSIVE:
1619 retval = update_flag(CS_CPU_EXCLUSIVE, cs, val);
1621 case FILE_MEM_EXCLUSIVE:
1622 retval = update_flag(CS_MEM_EXCLUSIVE, cs, val);
1624 case FILE_MEM_HARDWALL:
1625 retval = update_flag(CS_MEM_HARDWALL, cs, val);
1627 case FILE_SCHED_LOAD_BALANCE:
1628 retval = update_flag(CS_SCHED_LOAD_BALANCE, cs, val);
1630 case FILE_MEMORY_MIGRATE:
1631 retval = update_flag(CS_MEMORY_MIGRATE, cs, val);
1633 case FILE_MEMORY_PRESSURE_ENABLED:
1634 cpuset_memory_pressure_enabled = !!val;
1636 case FILE_MEMORY_PRESSURE:
1639 case FILE_SPREAD_PAGE:
1640 retval = update_flag(CS_SPREAD_PAGE, cs, val);
1642 case FILE_SPREAD_SLAB:
1643 retval = update_flag(CS_SPREAD_SLAB, cs, val);
1650 mutex_unlock(&cpuset_mutex);
1654 static int cpuset_write_s64(struct cgroup_subsys_state *css, struct cftype *cft,
1657 struct cpuset *cs = css_cs(css);
1658 cpuset_filetype_t type = cft->private;
1659 int retval = -ENODEV;
1661 mutex_lock(&cpuset_mutex);
1662 if (!is_cpuset_online(cs))
1666 case FILE_SCHED_RELAX_DOMAIN_LEVEL:
1667 retval = update_relax_domain_level(cs, val);
1674 mutex_unlock(&cpuset_mutex);
1679 * Common handling for a write to a "cpus" or "mems" file.
1681 static int cpuset_write_resmask(struct cgroup_subsys_state *css,
1682 struct cftype *cft, const char *buf)
1684 struct cpuset *cs = css_cs(css);
1685 struct cpuset *trialcs;
1686 int retval = -ENODEV;
1689 * CPU or memory hotunplug may leave @cs w/o any execution
1690 * resources, in which case the hotplug code asynchronously updates
1691 * configuration and transfers all tasks to the nearest ancestor
1692 * which can execute.
1694 * As writes to "cpus" or "mems" may restore @cs's execution
1695 * resources, wait for the previously scheduled operations before
1696 * proceeding, so that we don't end up keep removing tasks added
1697 * after execution capability is restored.
1699 flush_work(&cpuset_hotplug_work);
1701 mutex_lock(&cpuset_mutex);
1702 if (!is_cpuset_online(cs))
1705 trialcs = alloc_trial_cpuset(cs);
1711 switch (cft->private) {
1713 retval = update_cpumask(cs, trialcs, buf);
1716 retval = update_nodemask(cs, trialcs, buf);
1723 free_trial_cpuset(trialcs);
1725 mutex_unlock(&cpuset_mutex);
1730 * These ascii lists should be read in a single call, by using a user
1731 * buffer large enough to hold the entire map. If read in smaller
1732 * chunks, there is no guarantee of atomicity. Since the display format
1733 * used, list of ranges of sequential numbers, is variable length,
1734 * and since these maps can change value dynamically, one could read
1735 * gibberish by doing partial reads while a list was changing.
1736 * A single large read to a buffer that crosses a page boundary is
1737 * ok, because the result being copied to user land is not recomputed
1738 * across a page fault.
1741 static size_t cpuset_sprintf_cpulist(char *page, struct cpuset *cs)
1745 mutex_lock(&callback_mutex);
1746 count = cpulist_scnprintf(page, PAGE_SIZE, cs->cpus_allowed);
1747 mutex_unlock(&callback_mutex);
1752 static size_t cpuset_sprintf_memlist(char *page, struct cpuset *cs)
1756 mutex_lock(&callback_mutex);
1757 count = nodelist_scnprintf(page, PAGE_SIZE, cs->mems_allowed);
1758 mutex_unlock(&callback_mutex);
1763 static ssize_t cpuset_common_file_read(struct cgroup_subsys_state *css,
1764 struct cftype *cft, struct file *file,
1765 char __user *buf, size_t nbytes,
1768 struct cpuset *cs = css_cs(css);
1769 cpuset_filetype_t type = cft->private;
1774 if (!(page = (char *)__get_free_page(GFP_TEMPORARY)))
1781 s += cpuset_sprintf_cpulist(s, cs);
1784 s += cpuset_sprintf_memlist(s, cs);
1792 retval = simple_read_from_buffer(buf, nbytes, ppos, page, s - page);
1794 free_page((unsigned long)page);
1798 static u64 cpuset_read_u64(struct cgroup_subsys_state *css, struct cftype *cft)
1800 struct cpuset *cs = css_cs(css);
1801 cpuset_filetype_t type = cft->private;
1803 case FILE_CPU_EXCLUSIVE:
1804 return is_cpu_exclusive(cs);
1805 case FILE_MEM_EXCLUSIVE:
1806 return is_mem_exclusive(cs);
1807 case FILE_MEM_HARDWALL:
1808 return is_mem_hardwall(cs);
1809 case FILE_SCHED_LOAD_BALANCE:
1810 return is_sched_load_balance(cs);
1811 case FILE_MEMORY_MIGRATE:
1812 return is_memory_migrate(cs);
1813 case FILE_MEMORY_PRESSURE_ENABLED:
1814 return cpuset_memory_pressure_enabled;
1815 case FILE_MEMORY_PRESSURE:
1816 return fmeter_getrate(&cs->fmeter);
1817 case FILE_SPREAD_PAGE:
1818 return is_spread_page(cs);
1819 case FILE_SPREAD_SLAB:
1820 return is_spread_slab(cs);
1825 /* Unreachable but makes gcc happy */
1829 static s64 cpuset_read_s64(struct cgroup_subsys_state *css, struct cftype *cft)
1831 struct cpuset *cs = css_cs(css);
1832 cpuset_filetype_t type = cft->private;
1834 case FILE_SCHED_RELAX_DOMAIN_LEVEL:
1835 return cs->relax_domain_level;
1840 /* Unrechable but makes gcc happy */
1846 * for the common functions, 'private' gives the type of file
1849 static struct cftype files[] = {
1852 .read = cpuset_common_file_read,
1853 .write_string = cpuset_write_resmask,
1854 .max_write_len = (100U + 6 * NR_CPUS),
1855 .private = FILE_CPULIST,
1860 .read = cpuset_common_file_read,
1861 .write_string = cpuset_write_resmask,
1862 .max_write_len = (100U + 6 * MAX_NUMNODES),
1863 .private = FILE_MEMLIST,
1867 .name = "cpu_exclusive",
1868 .read_u64 = cpuset_read_u64,
1869 .write_u64 = cpuset_write_u64,
1870 .private = FILE_CPU_EXCLUSIVE,
1874 .name = "mem_exclusive",
1875 .read_u64 = cpuset_read_u64,
1876 .write_u64 = cpuset_write_u64,
1877 .private = FILE_MEM_EXCLUSIVE,
1881 .name = "mem_hardwall",
1882 .read_u64 = cpuset_read_u64,
1883 .write_u64 = cpuset_write_u64,
1884 .private = FILE_MEM_HARDWALL,
1888 .name = "sched_load_balance",
1889 .read_u64 = cpuset_read_u64,
1890 .write_u64 = cpuset_write_u64,
1891 .private = FILE_SCHED_LOAD_BALANCE,
1895 .name = "sched_relax_domain_level",
1896 .read_s64 = cpuset_read_s64,
1897 .write_s64 = cpuset_write_s64,
1898 .private = FILE_SCHED_RELAX_DOMAIN_LEVEL,
1902 .name = "memory_migrate",
1903 .read_u64 = cpuset_read_u64,
1904 .write_u64 = cpuset_write_u64,
1905 .private = FILE_MEMORY_MIGRATE,
1909 .name = "memory_pressure",
1910 .read_u64 = cpuset_read_u64,
1911 .write_u64 = cpuset_write_u64,
1912 .private = FILE_MEMORY_PRESSURE,
1917 .name = "memory_spread_page",
1918 .read_u64 = cpuset_read_u64,
1919 .write_u64 = cpuset_write_u64,
1920 .private = FILE_SPREAD_PAGE,
1924 .name = "memory_spread_slab",
1925 .read_u64 = cpuset_read_u64,
1926 .write_u64 = cpuset_write_u64,
1927 .private = FILE_SPREAD_SLAB,
1931 .name = "memory_pressure_enabled",
1932 .flags = CFTYPE_ONLY_ON_ROOT,
1933 .read_u64 = cpuset_read_u64,
1934 .write_u64 = cpuset_write_u64,
1935 .private = FILE_MEMORY_PRESSURE_ENABLED,
1942 * cpuset_css_alloc - allocate a cpuset css
1943 * cgrp: control group that the new cpuset will be part of
1946 static struct cgroup_subsys_state *
1947 cpuset_css_alloc(struct cgroup_subsys_state *parent_css)
1952 return &top_cpuset.css;
1954 cs = kzalloc(sizeof(*cs), GFP_KERNEL);
1956 return ERR_PTR(-ENOMEM);
1957 if (!alloc_cpumask_var(&cs->cpus_allowed, GFP_KERNEL)) {
1959 return ERR_PTR(-ENOMEM);
1962 set_bit(CS_SCHED_LOAD_BALANCE, &cs->flags);
1963 cpumask_clear(cs->cpus_allowed);
1964 nodes_clear(cs->mems_allowed);
1965 fmeter_init(&cs->fmeter);
1966 cs->relax_domain_level = -1;
1971 static int cpuset_css_online(struct cgroup_subsys_state *css)
1973 struct cpuset *cs = css_cs(css);
1974 struct cpuset *parent = parent_cs(cs);
1975 struct cpuset *tmp_cs;
1976 struct cgroup_subsys_state *pos_css;
1981 mutex_lock(&cpuset_mutex);
1983 set_bit(CS_ONLINE, &cs->flags);
1984 if (is_spread_page(parent))
1985 set_bit(CS_SPREAD_PAGE, &cs->flags);
1986 if (is_spread_slab(parent))
1987 set_bit(CS_SPREAD_SLAB, &cs->flags);
1989 number_of_cpusets++;
1991 if (!test_bit(CGRP_CPUSET_CLONE_CHILDREN, &css->cgroup->flags))
1995 * Clone @parent's configuration if CGRP_CPUSET_CLONE_CHILDREN is
1996 * set. This flag handling is implemented in cgroup core for
1997 * histrical reasons - the flag may be specified during mount.
1999 * Currently, if any sibling cpusets have exclusive cpus or mem, we
2000 * refuse to clone the configuration - thereby refusing the task to
2001 * be entered, and as a result refusing the sys_unshare() or
2002 * clone() which initiated it. If this becomes a problem for some
2003 * users who wish to allow that scenario, then this could be
2004 * changed to grant parent->cpus_allowed-sibling_cpus_exclusive
2005 * (and likewise for mems) to the new cgroup.
2008 cpuset_for_each_child(tmp_cs, pos_css, parent) {
2009 if (is_mem_exclusive(tmp_cs) || is_cpu_exclusive(tmp_cs)) {
2016 mutex_lock(&callback_mutex);
2017 cs->mems_allowed = parent->mems_allowed;
2018 cpumask_copy(cs->cpus_allowed, parent->cpus_allowed);
2019 mutex_unlock(&callback_mutex);
2021 mutex_unlock(&cpuset_mutex);
2026 * If the cpuset being removed has its flag 'sched_load_balance'
2027 * enabled, then simulate turning sched_load_balance off, which
2028 * will call rebuild_sched_domains_locked().
2031 static void cpuset_css_offline(struct cgroup_subsys_state *css)
2033 struct cpuset *cs = css_cs(css);
2035 mutex_lock(&cpuset_mutex);
2037 if (is_sched_load_balance(cs))
2038 update_flag(CS_SCHED_LOAD_BALANCE, cs, 0);
2040 number_of_cpusets--;
2041 clear_bit(CS_ONLINE, &cs->flags);
2043 mutex_unlock(&cpuset_mutex);
2046 static void cpuset_css_free(struct cgroup_subsys_state *css)
2048 struct cpuset *cs = css_cs(css);
2050 free_cpumask_var(cs->cpus_allowed);
2054 struct cgroup_subsys cpuset_subsys = {
2056 .css_alloc = cpuset_css_alloc,
2057 .css_online = cpuset_css_online,
2058 .css_offline = cpuset_css_offline,
2059 .css_free = cpuset_css_free,
2060 .can_attach = cpuset_can_attach,
2061 .cancel_attach = cpuset_cancel_attach,
2062 .attach = cpuset_attach,
2063 .subsys_id = cpuset_subsys_id,
2064 .base_cftypes = files,
2069 * cpuset_init - initialize cpusets at system boot
2071 * Description: Initialize top_cpuset and the cpuset internal file system,
2074 int __init cpuset_init(void)
2078 if (!alloc_cpumask_var(&top_cpuset.cpus_allowed, GFP_KERNEL))
2081 cpumask_setall(top_cpuset.cpus_allowed);
2082 nodes_setall(top_cpuset.mems_allowed);
2084 fmeter_init(&top_cpuset.fmeter);
2085 set_bit(CS_SCHED_LOAD_BALANCE, &top_cpuset.flags);
2086 top_cpuset.relax_domain_level = -1;
2088 err = register_filesystem(&cpuset_fs_type);
2092 if (!alloc_cpumask_var(&cpus_attach, GFP_KERNEL))
2095 number_of_cpusets = 1;
2100 * If CPU and/or memory hotplug handlers, below, unplug any CPUs
2101 * or memory nodes, we need to walk over the cpuset hierarchy,
2102 * removing that CPU or node from all cpusets. If this removes the
2103 * last CPU or node from a cpuset, then move the tasks in the empty
2104 * cpuset to its next-highest non-empty parent.
2106 static void remove_tasks_in_empty_cpuset(struct cpuset *cs)
2108 struct cpuset *parent;
2111 * Find its next-highest non-empty parent, (top cpuset
2112 * has online cpus, so can't be empty).
2114 parent = parent_cs(cs);
2115 while (cpumask_empty(parent->cpus_allowed) ||
2116 nodes_empty(parent->mems_allowed))
2117 parent = parent_cs(parent);
2119 if (cgroup_transfer_tasks(parent->css.cgroup, cs->css.cgroup)) {
2121 printk(KERN_ERR "cpuset: failed to transfer tasks out of empty cpuset %s\n",
2122 cgroup_name(cs->css.cgroup));
2128 * cpuset_hotplug_update_tasks - update tasks in a cpuset for hotunplug
2129 * @cs: cpuset in interest
2131 * Compare @cs's cpu and mem masks against top_cpuset and if some have gone
2132 * offline, update @cs accordingly. If @cs ends up with no CPU or memory,
2133 * all its tasks are moved to the nearest ancestor with both resources.
2135 static void cpuset_hotplug_update_tasks(struct cpuset *cs)
2137 static cpumask_t off_cpus;
2138 static nodemask_t off_mems;
2140 bool sane = cgroup_sane_behavior(cs->css.cgroup);
2143 wait_event(cpuset_attach_wq, cs->attach_in_progress == 0);
2145 mutex_lock(&cpuset_mutex);
2148 * We have raced with task attaching. We wait until attaching
2149 * is finished, so we won't attach a task to an empty cpuset.
2151 if (cs->attach_in_progress) {
2152 mutex_unlock(&cpuset_mutex);
2156 cpumask_andnot(&off_cpus, cs->cpus_allowed, top_cpuset.cpus_allowed);
2157 nodes_andnot(off_mems, cs->mems_allowed, top_cpuset.mems_allowed);
2159 mutex_lock(&callback_mutex);
2160 cpumask_andnot(cs->cpus_allowed, cs->cpus_allowed, &off_cpus);
2161 mutex_unlock(&callback_mutex);
2164 * If sane_behavior flag is set, we need to update tasks' cpumask
2165 * for empty cpuset to take on ancestor's cpumask. Otherwise, don't
2166 * call update_tasks_cpumask() if the cpuset becomes empty, as
2167 * the tasks in it will be migrated to an ancestor.
2169 if ((sane && cpumask_empty(cs->cpus_allowed)) ||
2170 (!cpumask_empty(&off_cpus) && !cpumask_empty(cs->cpus_allowed)))
2171 update_tasks_cpumask(cs, NULL);
2173 mutex_lock(&callback_mutex);
2174 nodes_andnot(cs->mems_allowed, cs->mems_allowed, off_mems);
2175 mutex_unlock(&callback_mutex);
2178 * If sane_behavior flag is set, we need to update tasks' nodemask
2179 * for empty cpuset to take on ancestor's nodemask. Otherwise, don't
2180 * call update_tasks_nodemask() if the cpuset becomes empty, as
2181 * the tasks in it will be migratd to an ancestor.
2183 if ((sane && nodes_empty(cs->mems_allowed)) ||
2184 (!nodes_empty(off_mems) && !nodes_empty(cs->mems_allowed)))
2185 update_tasks_nodemask(cs, NULL);
2187 is_empty = cpumask_empty(cs->cpus_allowed) ||
2188 nodes_empty(cs->mems_allowed);
2190 mutex_unlock(&cpuset_mutex);
2193 * If sane_behavior flag is set, we'll keep tasks in empty cpusets.
2195 * Otherwise move tasks to the nearest ancestor with execution
2196 * resources. This is full cgroup operation which will
2197 * also call back into cpuset. Should be done outside any lock.
2199 if (!sane && is_empty)
2200 remove_tasks_in_empty_cpuset(cs);
2204 * cpuset_hotplug_workfn - handle CPU/memory hotunplug for a cpuset
2206 * This function is called after either CPU or memory configuration has
2207 * changed and updates cpuset accordingly. The top_cpuset is always
2208 * synchronized to cpu_active_mask and N_MEMORY, which is necessary in
2209 * order to make cpusets transparent (of no affect) on systems that are
2210 * actively using CPU hotplug but making no active use of cpusets.
2212 * Non-root cpusets are only affected by offlining. If any CPUs or memory
2213 * nodes have been taken down, cpuset_hotplug_update_tasks() is invoked on
2216 * Note that CPU offlining during suspend is ignored. We don't modify
2217 * cpusets across suspend/resume cycles at all.
2219 static void cpuset_hotplug_workfn(struct work_struct *work)
2221 static cpumask_t new_cpus;
2222 static nodemask_t new_mems;
2223 bool cpus_updated, mems_updated;
2225 mutex_lock(&cpuset_mutex);
2227 /* fetch the available cpus/mems and find out which changed how */
2228 cpumask_copy(&new_cpus, cpu_active_mask);
2229 new_mems = node_states[N_MEMORY];
2231 cpus_updated = !cpumask_equal(top_cpuset.cpus_allowed, &new_cpus);
2232 mems_updated = !nodes_equal(top_cpuset.mems_allowed, new_mems);
2234 /* synchronize cpus_allowed to cpu_active_mask */
2236 mutex_lock(&callback_mutex);
2237 cpumask_copy(top_cpuset.cpus_allowed, &new_cpus);
2238 mutex_unlock(&callback_mutex);
2239 /* we don't mess with cpumasks of tasks in top_cpuset */
2242 /* synchronize mems_allowed to N_MEMORY */
2244 mutex_lock(&callback_mutex);
2245 top_cpuset.mems_allowed = new_mems;
2246 mutex_unlock(&callback_mutex);
2247 update_tasks_nodemask(&top_cpuset, NULL);
2250 mutex_unlock(&cpuset_mutex);
2252 /* if cpus or mems changed, we need to propagate to descendants */
2253 if (cpus_updated || mems_updated) {
2255 struct cgroup_subsys_state *pos_css;
2258 cpuset_for_each_descendant_pre(cs, pos_css, &top_cpuset) {
2259 if (!css_tryget(&cs->css))
2263 cpuset_hotplug_update_tasks(cs);
2271 /* rebuild sched domains if cpus_allowed has changed */
2273 rebuild_sched_domains();
2276 void cpuset_update_active_cpus(bool cpu_online)
2279 * We're inside cpu hotplug critical region which usually nests
2280 * inside cgroup synchronization. Bounce actual hotplug processing
2281 * to a work item to avoid reverse locking order.
2283 * We still need to do partition_sched_domains() synchronously;
2284 * otherwise, the scheduler will get confused and put tasks to the
2285 * dead CPU. Fall back to the default single domain.
2286 * cpuset_hotplug_workfn() will rebuild it as necessary.
2288 partition_sched_domains(1, NULL, NULL);
2289 schedule_work(&cpuset_hotplug_work);
2293 * Keep top_cpuset.mems_allowed tracking node_states[N_MEMORY].
2294 * Call this routine anytime after node_states[N_MEMORY] changes.
2295 * See cpuset_update_active_cpus() for CPU hotplug handling.
2297 static int cpuset_track_online_nodes(struct notifier_block *self,
2298 unsigned long action, void *arg)
2300 schedule_work(&cpuset_hotplug_work);
2304 static struct notifier_block cpuset_track_online_nodes_nb = {
2305 .notifier_call = cpuset_track_online_nodes,
2306 .priority = 10, /* ??! */
2310 * cpuset_init_smp - initialize cpus_allowed
2312 * Description: Finish top cpuset after cpu, node maps are initialized
2314 void __init cpuset_init_smp(void)
2316 cpumask_copy(top_cpuset.cpus_allowed, cpu_active_mask);
2317 top_cpuset.mems_allowed = node_states[N_MEMORY];
2318 top_cpuset.old_mems_allowed = top_cpuset.mems_allowed;
2320 register_hotmemory_notifier(&cpuset_track_online_nodes_nb);
2324 * cpuset_cpus_allowed - return cpus_allowed mask from a tasks cpuset.
2325 * @tsk: pointer to task_struct from which to obtain cpuset->cpus_allowed.
2326 * @pmask: pointer to struct cpumask variable to receive cpus_allowed set.
2328 * Description: Returns the cpumask_var_t cpus_allowed of the cpuset
2329 * attached to the specified @tsk. Guaranteed to return some non-empty
2330 * subset of cpu_online_mask, even if this means going outside the
2334 void cpuset_cpus_allowed(struct task_struct *tsk, struct cpumask *pmask)
2336 struct cpuset *cpus_cs;
2338 mutex_lock(&callback_mutex);
2340 cpus_cs = effective_cpumask_cpuset(task_cs(tsk));
2341 guarantee_online_cpus(cpus_cs, pmask);
2343 mutex_unlock(&callback_mutex);
2346 void cpuset_cpus_allowed_fallback(struct task_struct *tsk)
2348 struct cpuset *cpus_cs;
2351 cpus_cs = effective_cpumask_cpuset(task_cs(tsk));
2352 do_set_cpus_allowed(tsk, cpus_cs->cpus_allowed);
2356 * We own tsk->cpus_allowed, nobody can change it under us.
2358 * But we used cs && cs->cpus_allowed lockless and thus can
2359 * race with cgroup_attach_task() or update_cpumask() and get
2360 * the wrong tsk->cpus_allowed. However, both cases imply the
2361 * subsequent cpuset_change_cpumask()->set_cpus_allowed_ptr()
2362 * which takes task_rq_lock().
2364 * If we are called after it dropped the lock we must see all
2365 * changes in tsk_cs()->cpus_allowed. Otherwise we can temporary
2366 * set any mask even if it is not right from task_cs() pov,
2367 * the pending set_cpus_allowed_ptr() will fix things.
2369 * select_fallback_rq() will fix things ups and set cpu_possible_mask
2374 void cpuset_init_current_mems_allowed(void)
2376 nodes_setall(current->mems_allowed);
2380 * cpuset_mems_allowed - return mems_allowed mask from a tasks cpuset.
2381 * @tsk: pointer to task_struct from which to obtain cpuset->mems_allowed.
2383 * Description: Returns the nodemask_t mems_allowed of the cpuset
2384 * attached to the specified @tsk. Guaranteed to return some non-empty
2385 * subset of node_states[N_MEMORY], even if this means going outside the
2389 nodemask_t cpuset_mems_allowed(struct task_struct *tsk)
2391 struct cpuset *mems_cs;
2394 mutex_lock(&callback_mutex);
2396 mems_cs = effective_nodemask_cpuset(task_cs(tsk));
2397 guarantee_online_mems(mems_cs, &mask);
2399 mutex_unlock(&callback_mutex);
2405 * cpuset_nodemask_valid_mems_allowed - check nodemask vs. curremt mems_allowed
2406 * @nodemask: the nodemask to be checked
2408 * Are any of the nodes in the nodemask allowed in current->mems_allowed?
2410 int cpuset_nodemask_valid_mems_allowed(nodemask_t *nodemask)
2412 return nodes_intersects(*nodemask, current->mems_allowed);
2416 * nearest_hardwall_ancestor() - Returns the nearest mem_exclusive or
2417 * mem_hardwall ancestor to the specified cpuset. Call holding
2418 * callback_mutex. If no ancestor is mem_exclusive or mem_hardwall
2419 * (an unusual configuration), then returns the root cpuset.
2421 static struct cpuset *nearest_hardwall_ancestor(struct cpuset *cs)
2423 while (!(is_mem_exclusive(cs) || is_mem_hardwall(cs)) && parent_cs(cs))
2429 * cpuset_node_allowed_softwall - Can we allocate on a memory node?
2430 * @node: is this an allowed node?
2431 * @gfp_mask: memory allocation flags
2433 * If we're in interrupt, yes, we can always allocate. If __GFP_THISNODE is
2434 * set, yes, we can always allocate. If node is in our task's mems_allowed,
2435 * yes. If it's not a __GFP_HARDWALL request and this node is in the nearest
2436 * hardwalled cpuset ancestor to this task's cpuset, yes. If the task has been
2437 * OOM killed and has access to memory reserves as specified by the TIF_MEMDIE
2441 * If __GFP_HARDWALL is set, cpuset_node_allowed_softwall() reduces to
2442 * cpuset_node_allowed_hardwall(). Otherwise, cpuset_node_allowed_softwall()
2443 * might sleep, and might allow a node from an enclosing cpuset.
2445 * cpuset_node_allowed_hardwall() only handles the simpler case of hardwall
2446 * cpusets, and never sleeps.
2448 * The __GFP_THISNODE placement logic is really handled elsewhere,
2449 * by forcibly using a zonelist starting at a specified node, and by
2450 * (in get_page_from_freelist()) refusing to consider the zones for
2451 * any node on the zonelist except the first. By the time any such
2452 * calls get to this routine, we should just shut up and say 'yes'.
2454 * GFP_USER allocations are marked with the __GFP_HARDWALL bit,
2455 * and do not allow allocations outside the current tasks cpuset
2456 * unless the task has been OOM killed as is marked TIF_MEMDIE.
2457 * GFP_KERNEL allocations are not so marked, so can escape to the
2458 * nearest enclosing hardwalled ancestor cpuset.
2460 * Scanning up parent cpusets requires callback_mutex. The
2461 * __alloc_pages() routine only calls here with __GFP_HARDWALL bit
2462 * _not_ set if it's a GFP_KERNEL allocation, and all nodes in the
2463 * current tasks mems_allowed came up empty on the first pass over
2464 * the zonelist. So only GFP_KERNEL allocations, if all nodes in the
2465 * cpuset are short of memory, might require taking the callback_mutex
2468 * The first call here from mm/page_alloc:get_page_from_freelist()
2469 * has __GFP_HARDWALL set in gfp_mask, enforcing hardwall cpusets,
2470 * so no allocation on a node outside the cpuset is allowed (unless
2471 * in interrupt, of course).
2473 * The second pass through get_page_from_freelist() doesn't even call
2474 * here for GFP_ATOMIC calls. For those calls, the __alloc_pages()
2475 * variable 'wait' is not set, and the bit ALLOC_CPUSET is not set
2476 * in alloc_flags. That logic and the checks below have the combined
2478 * in_interrupt - any node ok (current task context irrelevant)
2479 * GFP_ATOMIC - any node ok
2480 * TIF_MEMDIE - any node ok
2481 * GFP_KERNEL - any node in enclosing hardwalled cpuset ok
2482 * GFP_USER - only nodes in current tasks mems allowed ok.
2485 * Don't call cpuset_node_allowed_softwall if you can't sleep, unless you
2486 * pass in the __GFP_HARDWALL flag set in gfp_flag, which disables
2487 * the code that might scan up ancestor cpusets and sleep.
2489 int __cpuset_node_allowed_softwall(int node, gfp_t gfp_mask)
2491 struct cpuset *cs; /* current cpuset ancestors */
2492 int allowed; /* is allocation in zone z allowed? */
2494 if (in_interrupt() || (gfp_mask & __GFP_THISNODE))
2496 might_sleep_if(!(gfp_mask & __GFP_HARDWALL));
2497 if (node_isset(node, current->mems_allowed))
2500 * Allow tasks that have access to memory reserves because they have
2501 * been OOM killed to get memory anywhere.
2503 if (unlikely(test_thread_flag(TIF_MEMDIE)))
2505 if (gfp_mask & __GFP_HARDWALL) /* If hardwall request, stop here */
2508 if (current->flags & PF_EXITING) /* Let dying task have memory */
2511 /* Not hardwall and node outside mems_allowed: scan up cpusets */
2512 mutex_lock(&callback_mutex);
2515 cs = nearest_hardwall_ancestor(task_cs(current));
2516 task_unlock(current);
2518 allowed = node_isset(node, cs->mems_allowed);
2519 mutex_unlock(&callback_mutex);
2524 * cpuset_node_allowed_hardwall - Can we allocate on a memory node?
2525 * @node: is this an allowed node?
2526 * @gfp_mask: memory allocation flags
2528 * If we're in interrupt, yes, we can always allocate. If __GFP_THISNODE is
2529 * set, yes, we can always allocate. If node is in our task's mems_allowed,
2530 * yes. If the task has been OOM killed and has access to memory reserves as
2531 * specified by the TIF_MEMDIE flag, yes.
2534 * The __GFP_THISNODE placement logic is really handled elsewhere,
2535 * by forcibly using a zonelist starting at a specified node, and by
2536 * (in get_page_from_freelist()) refusing to consider the zones for
2537 * any node on the zonelist except the first. By the time any such
2538 * calls get to this routine, we should just shut up and say 'yes'.
2540 * Unlike the cpuset_node_allowed_softwall() variant, above,
2541 * this variant requires that the node be in the current task's
2542 * mems_allowed or that we're in interrupt. It does not scan up the
2543 * cpuset hierarchy for the nearest enclosing mem_exclusive cpuset.
2546 int __cpuset_node_allowed_hardwall(int node, gfp_t gfp_mask)
2548 if (in_interrupt() || (gfp_mask & __GFP_THISNODE))
2550 if (node_isset(node, current->mems_allowed))
2553 * Allow tasks that have access to memory reserves because they have
2554 * been OOM killed to get memory anywhere.
2556 if (unlikely(test_thread_flag(TIF_MEMDIE)))
2562 * cpuset_mem_spread_node() - On which node to begin search for a file page
2563 * cpuset_slab_spread_node() - On which node to begin search for a slab page
2565 * If a task is marked PF_SPREAD_PAGE or PF_SPREAD_SLAB (as for
2566 * tasks in a cpuset with is_spread_page or is_spread_slab set),
2567 * and if the memory allocation used cpuset_mem_spread_node()
2568 * to determine on which node to start looking, as it will for
2569 * certain page cache or slab cache pages such as used for file
2570 * system buffers and inode caches, then instead of starting on the
2571 * local node to look for a free page, rather spread the starting
2572 * node around the tasks mems_allowed nodes.
2574 * We don't have to worry about the returned node being offline
2575 * because "it can't happen", and even if it did, it would be ok.
2577 * The routines calling guarantee_online_mems() are careful to
2578 * only set nodes in task->mems_allowed that are online. So it
2579 * should not be possible for the following code to return an
2580 * offline node. But if it did, that would be ok, as this routine
2581 * is not returning the node where the allocation must be, only
2582 * the node where the search should start. The zonelist passed to
2583 * __alloc_pages() will include all nodes. If the slab allocator
2584 * is passed an offline node, it will fall back to the local node.
2585 * See kmem_cache_alloc_node().
2588 static int cpuset_spread_node(int *rotor)
2592 node = next_node(*rotor, current->mems_allowed);
2593 if (node == MAX_NUMNODES)
2594 node = first_node(current->mems_allowed);
2599 int cpuset_mem_spread_node(void)
2601 if (current->cpuset_mem_spread_rotor == NUMA_NO_NODE)
2602 current->cpuset_mem_spread_rotor =
2603 node_random(¤t->mems_allowed);
2605 return cpuset_spread_node(¤t->cpuset_mem_spread_rotor);
2608 int cpuset_slab_spread_node(void)
2610 if (current->cpuset_slab_spread_rotor == NUMA_NO_NODE)
2611 current->cpuset_slab_spread_rotor =
2612 node_random(¤t->mems_allowed);
2614 return cpuset_spread_node(¤t->cpuset_slab_spread_rotor);
2617 EXPORT_SYMBOL_GPL(cpuset_mem_spread_node);
2620 * cpuset_mems_allowed_intersects - Does @tsk1's mems_allowed intersect @tsk2's?
2621 * @tsk1: pointer to task_struct of some task.
2622 * @tsk2: pointer to task_struct of some other task.
2624 * Description: Return true if @tsk1's mems_allowed intersects the
2625 * mems_allowed of @tsk2. Used by the OOM killer to determine if
2626 * one of the task's memory usage might impact the memory available
2630 int cpuset_mems_allowed_intersects(const struct task_struct *tsk1,
2631 const struct task_struct *tsk2)
2633 return nodes_intersects(tsk1->mems_allowed, tsk2->mems_allowed);
2636 #define CPUSET_NODELIST_LEN (256)
2639 * cpuset_print_task_mems_allowed - prints task's cpuset and mems_allowed
2640 * @task: pointer to task_struct of some task.
2642 * Description: Prints @task's name, cpuset name, and cached copy of its
2643 * mems_allowed to the kernel log. Must hold task_lock(task) to allow
2644 * dereferencing task_cs(task).
2646 void cpuset_print_task_mems_allowed(struct task_struct *tsk)
2648 /* Statically allocated to prevent using excess stack. */
2649 static char cpuset_nodelist[CPUSET_NODELIST_LEN];
2650 static DEFINE_SPINLOCK(cpuset_buffer_lock);
2652 struct cgroup *cgrp = task_cs(tsk)->css.cgroup;
2655 spin_lock(&cpuset_buffer_lock);
2657 nodelist_scnprintf(cpuset_nodelist, CPUSET_NODELIST_LEN,
2659 printk(KERN_INFO "%s cpuset=%s mems_allowed=%s\n",
2660 tsk->comm, cgroup_name(cgrp), cpuset_nodelist);
2662 spin_unlock(&cpuset_buffer_lock);
2667 * Collection of memory_pressure is suppressed unless
2668 * this flag is enabled by writing "1" to the special
2669 * cpuset file 'memory_pressure_enabled' in the root cpuset.
2672 int cpuset_memory_pressure_enabled __read_mostly;
2675 * cpuset_memory_pressure_bump - keep stats of per-cpuset reclaims.
2677 * Keep a running average of the rate of synchronous (direct)
2678 * page reclaim efforts initiated by tasks in each cpuset.
2680 * This represents the rate at which some task in the cpuset
2681 * ran low on memory on all nodes it was allowed to use, and
2682 * had to enter the kernels page reclaim code in an effort to
2683 * create more free memory by tossing clean pages or swapping
2684 * or writing dirty pages.
2686 * Display to user space in the per-cpuset read-only file
2687 * "memory_pressure". Value displayed is an integer
2688 * representing the recent rate of entry into the synchronous
2689 * (direct) page reclaim by any task attached to the cpuset.
2692 void __cpuset_memory_pressure_bump(void)
2695 fmeter_markevent(&task_cs(current)->fmeter);
2696 task_unlock(current);
2699 #ifdef CONFIG_PROC_PID_CPUSET
2701 * proc_cpuset_show()
2702 * - Print tasks cpuset path into seq_file.
2703 * - Used for /proc/<pid>/cpuset.
2704 * - No need to task_lock(tsk) on this tsk->cpuset reference, as it
2705 * doesn't really matter if tsk->cpuset changes after we read it,
2706 * and we take cpuset_mutex, keeping cpuset_attach() from changing it
2709 int proc_cpuset_show(struct seq_file *m, void *unused_v)
2712 struct task_struct *tsk;
2714 struct cgroup_subsys_state *css;
2718 buf = kmalloc(PAGE_SIZE, GFP_KERNEL);
2724 tsk = get_pid_task(pid, PIDTYPE_PID);
2729 css = task_css(tsk, cpuset_subsys_id);
2730 retval = cgroup_path(css->cgroup, buf, PAGE_SIZE);
2737 put_task_struct(tsk);
2743 #endif /* CONFIG_PROC_PID_CPUSET */
2745 /* Display task mems_allowed in /proc/<pid>/status file. */
2746 void cpuset_task_status_allowed(struct seq_file *m, struct task_struct *task)
2748 seq_printf(m, "Mems_allowed:\t");
2749 seq_nodemask(m, &task->mems_allowed);
2750 seq_printf(m, "\n");
2751 seq_printf(m, "Mems_allowed_list:\t");
2752 seq_nodemask_list(m, &task->mems_allowed);
2753 seq_printf(m, "\n");