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>
64 struct static_key cpusets_enabled_key __read_mostly
= STATIC_KEY_INIT_FALSE
;
66 /* See "Frequency meter" comments, below. */
69 int cnt
; /* unprocessed events count */
70 int val
; /* most recent output value */
71 time_t time
; /* clock (secs) when val computed */
72 spinlock_t lock
; /* guards read or write of above */
76 struct cgroup_subsys_state css
;
78 unsigned long flags
; /* "unsigned long" so bitops work */
80 /* user-configured CPUs and Memory Nodes allow to tasks */
81 cpumask_var_t cpus_allowed
;
82 nodemask_t mems_allowed
;
84 /* effective CPUs and Memory Nodes allow to tasks */
85 cpumask_var_t effective_cpus
;
86 nodemask_t effective_mems
;
89 * This is old Memory Nodes tasks took on.
91 * - top_cpuset.old_mems_allowed is initialized to mems_allowed.
92 * - A new cpuset's old_mems_allowed is initialized when some
93 * task is moved into it.
94 * - old_mems_allowed is used in cpuset_migrate_mm() when we change
95 * cpuset.mems_allowed and have tasks' nodemask updated, and
96 * then old_mems_allowed is updated to mems_allowed.
98 nodemask_t old_mems_allowed
;
100 struct fmeter fmeter
; /* memory_pressure filter */
103 * Tasks are being attached to this cpuset. Used to prevent
104 * zeroing cpus/mems_allowed between ->can_attach() and ->attach().
106 int attach_in_progress
;
108 /* partition number for rebuild_sched_domains() */
111 /* for custom sched domain */
112 int relax_domain_level
;
115 static inline struct cpuset
*css_cs(struct cgroup_subsys_state
*css
)
117 return css
? container_of(css
, struct cpuset
, css
) : NULL
;
120 /* Retrieve the cpuset for a task */
121 static inline struct cpuset
*task_cs(struct task_struct
*task
)
123 return css_cs(task_css(task
, cpuset_cgrp_id
));
126 static inline struct cpuset
*parent_cs(struct cpuset
*cs
)
128 return css_cs(cs
->css
.parent
);
132 static inline bool task_has_mempolicy(struct task_struct
*task
)
134 return task
->mempolicy
;
137 static inline bool task_has_mempolicy(struct task_struct
*task
)
144 /* bits in struct cpuset flags field */
151 CS_SCHED_LOAD_BALANCE
,
156 /* convenient tests for these bits */
157 static inline bool is_cpuset_online(const struct cpuset
*cs
)
159 return test_bit(CS_ONLINE
, &cs
->flags
);
162 static inline int is_cpu_exclusive(const struct cpuset
*cs
)
164 return test_bit(CS_CPU_EXCLUSIVE
, &cs
->flags
);
167 static inline int is_mem_exclusive(const struct cpuset
*cs
)
169 return test_bit(CS_MEM_EXCLUSIVE
, &cs
->flags
);
172 static inline int is_mem_hardwall(const struct cpuset
*cs
)
174 return test_bit(CS_MEM_HARDWALL
, &cs
->flags
);
177 static inline int is_sched_load_balance(const struct cpuset
*cs
)
179 return test_bit(CS_SCHED_LOAD_BALANCE
, &cs
->flags
);
182 static inline int is_memory_migrate(const struct cpuset
*cs
)
184 return test_bit(CS_MEMORY_MIGRATE
, &cs
->flags
);
187 static inline int is_spread_page(const struct cpuset
*cs
)
189 return test_bit(CS_SPREAD_PAGE
, &cs
->flags
);
192 static inline int is_spread_slab(const struct cpuset
*cs
)
194 return test_bit(CS_SPREAD_SLAB
, &cs
->flags
);
197 static struct cpuset top_cpuset
= {
198 .flags
= ((1 << CS_ONLINE
) | (1 << CS_CPU_EXCLUSIVE
) |
199 (1 << CS_MEM_EXCLUSIVE
)),
203 * cpuset_for_each_child - traverse online children of a cpuset
204 * @child_cs: loop cursor pointing to the current child
205 * @pos_css: used for iteration
206 * @parent_cs: target cpuset to walk children of
208 * Walk @child_cs through the online children of @parent_cs. Must be used
209 * with RCU read locked.
211 #define cpuset_for_each_child(child_cs, pos_css, parent_cs) \
212 css_for_each_child((pos_css), &(parent_cs)->css) \
213 if (is_cpuset_online(((child_cs) = css_cs((pos_css)))))
216 * cpuset_for_each_descendant_pre - pre-order walk of a cpuset's descendants
217 * @des_cs: loop cursor pointing to the current descendant
218 * @pos_css: used for iteration
219 * @root_cs: target cpuset to walk ancestor of
221 * Walk @des_cs through the online descendants of @root_cs. Must be used
222 * with RCU read locked. The caller may modify @pos_css by calling
223 * css_rightmost_descendant() to skip subtree. @root_cs is included in the
224 * iteration and the first node to be visited.
226 #define cpuset_for_each_descendant_pre(des_cs, pos_css, root_cs) \
227 css_for_each_descendant_pre((pos_css), &(root_cs)->css) \
228 if (is_cpuset_online(((des_cs) = css_cs((pos_css)))))
231 * There are two global mutexes guarding cpuset structures - cpuset_mutex
232 * and callback_mutex. The latter may nest inside the former. We also
233 * require taking task_lock() when dereferencing a task's cpuset pointer.
234 * See "The task_lock() exception", at the end of this comment.
236 * A task must hold both mutexes to modify cpusets. If a task holds
237 * cpuset_mutex, then it blocks others wanting that mutex, ensuring that it
238 * is the only task able to also acquire callback_mutex and be able to
239 * modify cpusets. It can perform various checks on the cpuset structure
240 * first, knowing nothing will change. It can also allocate memory while
241 * just holding cpuset_mutex. While it is performing these checks, various
242 * callback routines can briefly acquire callback_mutex to query cpusets.
243 * Once it is ready to make the changes, it takes callback_mutex, blocking
246 * Calls to the kernel memory allocator can not be made while holding
247 * callback_mutex, as that would risk double tripping on callback_mutex
248 * from one of the callbacks into the cpuset code from within
251 * If a task is only holding callback_mutex, then it has read-only
254 * Now, the task_struct fields mems_allowed and mempolicy may be changed
255 * by other task, we use alloc_lock in the task_struct fields to protect
258 * The cpuset_common_file_read() handlers only hold callback_mutex across
259 * small pieces of code, such as when reading out possibly multi-word
260 * cpumasks and nodemasks.
262 * Accessing a task's cpuset should be done in accordance with the
263 * guidelines for accessing subsystem state in kernel/cgroup.c
266 static DEFINE_MUTEX(cpuset_mutex
);
267 static DEFINE_MUTEX(callback_mutex
);
270 * CPU / memory hotplug is handled asynchronously.
272 static void cpuset_hotplug_workfn(struct work_struct
*work
);
273 static DECLARE_WORK(cpuset_hotplug_work
, cpuset_hotplug_workfn
);
275 static DECLARE_WAIT_QUEUE_HEAD(cpuset_attach_wq
);
278 * This is ugly, but preserves the userspace API for existing cpuset
279 * users. If someone tries to mount the "cpuset" filesystem, we
280 * silently switch it to mount "cgroup" instead
282 static struct dentry
*cpuset_mount(struct file_system_type
*fs_type
,
283 int flags
, const char *unused_dev_name
, void *data
)
285 struct file_system_type
*cgroup_fs
= get_fs_type("cgroup");
286 struct dentry
*ret
= ERR_PTR(-ENODEV
);
290 "release_agent=/sbin/cpuset_release_agent";
291 ret
= cgroup_fs
->mount(cgroup_fs
, flags
,
292 unused_dev_name
, mountopts
);
293 put_filesystem(cgroup_fs
);
298 static struct file_system_type cpuset_fs_type
= {
300 .mount
= cpuset_mount
,
304 * Return in pmask the portion of a cpusets's cpus_allowed that
305 * are online. If none are online, walk up the cpuset hierarchy
306 * until we find one that does have some online cpus. The top
307 * cpuset always has some cpus online.
309 * One way or another, we guarantee to return some non-empty subset
310 * of cpu_online_mask.
312 * Call with callback_mutex held.
314 static void guarantee_online_cpus(struct cpuset
*cs
, struct cpumask
*pmask
)
316 while (!cpumask_intersects(cs
->cpus_allowed
, cpu_online_mask
))
318 cpumask_and(pmask
, cs
->cpus_allowed
, cpu_online_mask
);
322 * Return in *pmask the portion of a cpusets's mems_allowed that
323 * are online, with memory. If none are online with memory, walk
324 * up the cpuset hierarchy until we find one that does have some
325 * online mems. The top cpuset always has some mems online.
327 * One way or another, we guarantee to return some non-empty subset
328 * of node_states[N_MEMORY].
330 * Call with callback_mutex held.
332 static void guarantee_online_mems(struct cpuset
*cs
, nodemask_t
*pmask
)
334 while (!nodes_intersects(cs
->mems_allowed
, node_states
[N_MEMORY
]))
336 nodes_and(*pmask
, cs
->mems_allowed
, node_states
[N_MEMORY
]);
340 * update task's spread flag if cpuset's page/slab spread flag is set
342 * Called with callback_mutex/cpuset_mutex held
344 static void cpuset_update_task_spread_flag(struct cpuset
*cs
,
345 struct task_struct
*tsk
)
347 if (is_spread_page(cs
))
348 tsk
->flags
|= PF_SPREAD_PAGE
;
350 tsk
->flags
&= ~PF_SPREAD_PAGE
;
351 if (is_spread_slab(cs
))
352 tsk
->flags
|= PF_SPREAD_SLAB
;
354 tsk
->flags
&= ~PF_SPREAD_SLAB
;
358 * is_cpuset_subset(p, q) - Is cpuset p a subset of cpuset q?
360 * One cpuset is a subset of another if all its allowed CPUs and
361 * Memory Nodes are a subset of the other, and its exclusive flags
362 * are only set if the other's are set. Call holding cpuset_mutex.
365 static int is_cpuset_subset(const struct cpuset
*p
, const struct cpuset
*q
)
367 return cpumask_subset(p
->cpus_allowed
, q
->cpus_allowed
) &&
368 nodes_subset(p
->mems_allowed
, q
->mems_allowed
) &&
369 is_cpu_exclusive(p
) <= is_cpu_exclusive(q
) &&
370 is_mem_exclusive(p
) <= is_mem_exclusive(q
);
374 * alloc_trial_cpuset - allocate a trial cpuset
375 * @cs: the cpuset that the trial cpuset duplicates
377 static struct cpuset
*alloc_trial_cpuset(struct cpuset
*cs
)
379 struct cpuset
*trial
;
381 trial
= kmemdup(cs
, sizeof(*cs
), GFP_KERNEL
);
385 if (!alloc_cpumask_var(&trial
->cpus_allowed
, GFP_KERNEL
))
387 if (!alloc_cpumask_var(&trial
->effective_cpus
, GFP_KERNEL
))
390 cpumask_copy(trial
->cpus_allowed
, cs
->cpus_allowed
);
391 cpumask_copy(trial
->effective_cpus
, cs
->effective_cpus
);
395 free_cpumask_var(trial
->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
->effective_cpus
);
408 free_cpumask_var(trial
->cpus_allowed
);
413 * validate_change() - Used to validate that any proposed cpuset change
414 * follows the structural rules for cpusets.
416 * If we replaced the flag and mask values of the current cpuset
417 * (cur) with those values in the trial cpuset (trial), would
418 * our various subset and exclusive rules still be valid? Presumes
421 * 'cur' is the address of an actual, in-use cpuset. Operations
422 * such as list traversal that depend on the actual address of the
423 * cpuset in the list must use cur below, not trial.
425 * 'trial' is the address of bulk structure copy of cur, with
426 * perhaps one or more of the fields cpus_allowed, mems_allowed,
427 * or flags changed to new, trial values.
429 * Return 0 if valid, -errno if not.
432 static int validate_change(struct cpuset
*cur
, struct cpuset
*trial
)
434 struct cgroup_subsys_state
*css
;
435 struct cpuset
*c
, *par
;
440 /* Each of our child cpusets must be a subset of us */
442 cpuset_for_each_child(c
, css
, cur
)
443 if (!is_cpuset_subset(c
, trial
))
446 /* Remaining checks don't apply to root cpuset */
448 if (cur
== &top_cpuset
)
451 par
= parent_cs(cur
);
453 /* We must be a subset of our parent cpuset */
455 if (!is_cpuset_subset(trial
, par
))
459 * If either I or some sibling (!= me) is exclusive, we can't
463 cpuset_for_each_child(c
, css
, par
) {
464 if ((is_cpu_exclusive(trial
) || is_cpu_exclusive(c
)) &&
466 cpumask_intersects(trial
->cpus_allowed
, c
->cpus_allowed
))
468 if ((is_mem_exclusive(trial
) || is_mem_exclusive(c
)) &&
470 nodes_intersects(trial
->mems_allowed
, c
->mems_allowed
))
475 * Cpusets with tasks - existing or newly being attached - can't
476 * be changed to have empty cpus_allowed or mems_allowed.
479 if ((cgroup_has_tasks(cur
->css
.cgroup
) || cur
->attach_in_progress
)) {
480 if (!cpumask_empty(cur
->cpus_allowed
) &&
481 cpumask_empty(trial
->cpus_allowed
))
483 if (!nodes_empty(cur
->mems_allowed
) &&
484 nodes_empty(trial
->mems_allowed
))
496 * Helper routine for generate_sched_domains().
497 * Do cpusets a, b have overlapping cpus_allowed masks?
499 static int cpusets_overlap(struct cpuset
*a
, struct cpuset
*b
)
501 return cpumask_intersects(a
->cpus_allowed
, b
->cpus_allowed
);
505 update_domain_attr(struct sched_domain_attr
*dattr
, struct cpuset
*c
)
507 if (dattr
->relax_domain_level
< c
->relax_domain_level
)
508 dattr
->relax_domain_level
= c
->relax_domain_level
;
512 static void update_domain_attr_tree(struct sched_domain_attr
*dattr
,
513 struct cpuset
*root_cs
)
516 struct cgroup_subsys_state
*pos_css
;
519 cpuset_for_each_descendant_pre(cp
, pos_css
, root_cs
) {
523 /* skip the whole subtree if @cp doesn't have any CPU */
524 if (cpumask_empty(cp
->cpus_allowed
)) {
525 pos_css
= css_rightmost_descendant(pos_css
);
529 if (is_sched_load_balance(cp
))
530 update_domain_attr(dattr
, cp
);
536 * generate_sched_domains()
538 * This function builds a partial partition of the systems CPUs
539 * A 'partial partition' is a set of non-overlapping subsets whose
540 * union is a subset of that set.
541 * The output of this function needs to be passed to kernel/sched/core.c
542 * partition_sched_domains() routine, which will rebuild the scheduler's
543 * load balancing domains (sched domains) as specified by that partial
546 * See "What is sched_load_balance" in Documentation/cgroups/cpusets.txt
547 * for a background explanation of this.
549 * Does not return errors, on the theory that the callers of this
550 * routine would rather not worry about failures to rebuild sched
551 * domains when operating in the severe memory shortage situations
552 * that could cause allocation failures below.
554 * Must be called with cpuset_mutex held.
556 * The three key local variables below are:
557 * q - a linked-list queue of cpuset pointers, used to implement a
558 * top-down scan of all cpusets. This scan loads a pointer
559 * to each cpuset marked is_sched_load_balance into the
560 * array 'csa'. For our purposes, rebuilding the schedulers
561 * sched domains, we can ignore !is_sched_load_balance cpusets.
562 * csa - (for CpuSet Array) Array of pointers to all the cpusets
563 * that need to be load balanced, for convenient iterative
564 * access by the subsequent code that finds the best partition,
565 * i.e the set of domains (subsets) of CPUs such that the
566 * cpus_allowed of every cpuset marked is_sched_load_balance
567 * is a subset of one of these domains, while there are as
568 * many such domains as possible, each as small as possible.
569 * doms - Conversion of 'csa' to an array of cpumasks, for passing to
570 * the kernel/sched/core.c routine partition_sched_domains() in a
571 * convenient format, that can be easily compared to the prior
572 * value to determine what partition elements (sched domains)
573 * were changed (added or removed.)
575 * Finding the best partition (set of domains):
576 * The triple nested loops below over i, j, k scan over the
577 * load balanced cpusets (using the array of cpuset pointers in
578 * csa[]) looking for pairs of cpusets that have overlapping
579 * cpus_allowed, but which don't have the same 'pn' partition
580 * number and gives them in the same partition number. It keeps
581 * looping on the 'restart' label until it can no longer find
584 * The union of the cpus_allowed masks from the set of
585 * all cpusets having the same 'pn' value then form the one
586 * element of the partition (one sched domain) to be passed to
587 * partition_sched_domains().
589 static int generate_sched_domains(cpumask_var_t
**domains
,
590 struct sched_domain_attr
**attributes
)
592 struct cpuset
*cp
; /* scans q */
593 struct cpuset
**csa
; /* array of all cpuset ptrs */
594 int csn
; /* how many cpuset ptrs in csa so far */
595 int i
, j
, k
; /* indices for partition finding loops */
596 cpumask_var_t
*doms
; /* resulting partition; i.e. sched domains */
597 struct sched_domain_attr
*dattr
; /* attributes for custom domains */
598 int ndoms
= 0; /* number of sched domains in result */
599 int nslot
; /* next empty doms[] struct cpumask slot */
600 struct cgroup_subsys_state
*pos_css
;
606 /* Special case for the 99% of systems with one, full, sched domain */
607 if (is_sched_load_balance(&top_cpuset
)) {
609 doms
= alloc_sched_domains(ndoms
);
613 dattr
= kmalloc(sizeof(struct sched_domain_attr
), GFP_KERNEL
);
615 *dattr
= SD_ATTR_INIT
;
616 update_domain_attr_tree(dattr
, &top_cpuset
);
618 cpumask_copy(doms
[0], top_cpuset
.cpus_allowed
);
623 csa
= kmalloc(nr_cpusets() * sizeof(cp
), GFP_KERNEL
);
629 cpuset_for_each_descendant_pre(cp
, pos_css
, &top_cpuset
) {
630 if (cp
== &top_cpuset
)
633 * Continue traversing beyond @cp iff @cp has some CPUs and
634 * isn't load balancing. The former is obvious. The
635 * latter: All child cpusets contain a subset of the
636 * parent's cpus, so just skip them, and then we call
637 * update_domain_attr_tree() to calc relax_domain_level of
638 * the corresponding sched domain.
640 if (!cpumask_empty(cp
->cpus_allowed
) &&
641 !is_sched_load_balance(cp
))
644 if (is_sched_load_balance(cp
))
647 /* skip @cp's subtree */
648 pos_css
= css_rightmost_descendant(pos_css
);
652 for (i
= 0; i
< csn
; i
++)
657 /* Find the best partition (set of sched domains) */
658 for (i
= 0; i
< csn
; i
++) {
659 struct cpuset
*a
= csa
[i
];
662 for (j
= 0; j
< csn
; j
++) {
663 struct cpuset
*b
= csa
[j
];
666 if (apn
!= bpn
&& cpusets_overlap(a
, b
)) {
667 for (k
= 0; k
< csn
; k
++) {
668 struct cpuset
*c
= csa
[k
];
673 ndoms
--; /* one less element */
680 * Now we know how many domains to create.
681 * Convert <csn, csa> to <ndoms, doms> and populate cpu masks.
683 doms
= alloc_sched_domains(ndoms
);
688 * The rest of the code, including the scheduler, can deal with
689 * dattr==NULL case. No need to abort if alloc fails.
691 dattr
= kmalloc(ndoms
* sizeof(struct sched_domain_attr
), GFP_KERNEL
);
693 for (nslot
= 0, i
= 0; i
< csn
; i
++) {
694 struct cpuset
*a
= csa
[i
];
699 /* Skip completed partitions */
705 if (nslot
== ndoms
) {
706 static int warnings
= 10;
708 pr_warn("rebuild_sched_domains confused: nslot %d, ndoms %d, csn %d, i %d, apn %d\n",
709 nslot
, ndoms
, csn
, i
, apn
);
717 *(dattr
+ nslot
) = SD_ATTR_INIT
;
718 for (j
= i
; j
< csn
; j
++) {
719 struct cpuset
*b
= csa
[j
];
722 cpumask_or(dp
, dp
, b
->cpus_allowed
);
724 update_domain_attr_tree(dattr
+ nslot
, b
);
726 /* Done with this partition */
732 BUG_ON(nslot
!= ndoms
);
738 * Fallback to the default domain if kmalloc() failed.
739 * See comments in partition_sched_domains().
750 * Rebuild scheduler domains.
752 * If the flag 'sched_load_balance' of any cpuset with non-empty
753 * 'cpus' changes, or if the 'cpus' allowed changes in any cpuset
754 * which has that flag enabled, or if any cpuset with a non-empty
755 * 'cpus' is removed, then call this routine to rebuild the
756 * scheduler's dynamic sched domains.
758 * Call with cpuset_mutex held. Takes get_online_cpus().
760 static void rebuild_sched_domains_locked(void)
762 struct sched_domain_attr
*attr
;
766 lockdep_assert_held(&cpuset_mutex
);
770 * We have raced with CPU hotplug. Don't do anything to avoid
771 * passing doms with offlined cpu to partition_sched_domains().
772 * Anyways, hotplug work item will rebuild sched domains.
774 if (!cpumask_equal(top_cpuset
.cpus_allowed
, cpu_active_mask
))
777 /* Generate domain masks and attrs */
778 ndoms
= generate_sched_domains(&doms
, &attr
);
780 /* Have scheduler rebuild the domains */
781 partition_sched_domains(ndoms
, doms
, attr
);
785 #else /* !CONFIG_SMP */
786 static void rebuild_sched_domains_locked(void)
789 #endif /* CONFIG_SMP */
791 void rebuild_sched_domains(void)
793 mutex_lock(&cpuset_mutex
);
794 rebuild_sched_domains_locked();
795 mutex_unlock(&cpuset_mutex
);
799 * effective_cpumask_cpuset - return nearest ancestor with non-empty cpus
800 * @cs: the cpuset in interest
802 * A cpuset's effective cpumask is the cpumask of the nearest ancestor
803 * with non-empty cpus. We use effective cpumask whenever:
804 * - we update tasks' cpus_allowed. (they take on the ancestor's cpumask
805 * if the cpuset they reside in has no cpus)
806 * - we want to retrieve task_cs(tsk)'s cpus_allowed.
808 * Called with cpuset_mutex held. cpuset_cpus_allowed_fallback() is an
809 * exception. See comments there.
811 static struct cpuset
*effective_cpumask_cpuset(struct cpuset
*cs
)
813 while (cpumask_empty(cs
->cpus_allowed
))
819 * effective_nodemask_cpuset - return nearest ancestor with non-empty mems
820 * @cs: the cpuset in interest
822 * A cpuset's effective nodemask is the nodemask of the nearest ancestor
823 * with non-empty memss. We use effective nodemask whenever:
824 * - we update tasks' mems_allowed. (they take on the ancestor's nodemask
825 * if the cpuset they reside in has no mems)
826 * - we want to retrieve task_cs(tsk)'s mems_allowed.
828 * Called with cpuset_mutex held.
830 static struct cpuset
*effective_nodemask_cpuset(struct cpuset
*cs
)
832 while (nodes_empty(cs
->mems_allowed
))
838 * update_tasks_cpumask - Update the cpumasks of tasks in the cpuset.
839 * @cs: the cpuset in which each task's cpus_allowed mask needs to be changed
841 * Iterate through each task of @cs updating its cpus_allowed to the
842 * effective cpuset's. As this function is called with cpuset_mutex held,
843 * cpuset membership stays stable.
845 static void update_tasks_cpumask(struct cpuset
*cs
)
847 struct cpuset
*cpus_cs
= effective_cpumask_cpuset(cs
);
848 struct css_task_iter it
;
849 struct task_struct
*task
;
851 css_task_iter_start(&cs
->css
, &it
);
852 while ((task
= css_task_iter_next(&it
)))
853 set_cpus_allowed_ptr(task
, cpus_cs
->cpus_allowed
);
854 css_task_iter_end(&it
);
858 * update_cpumasks_hier - Update effective cpumasks and tasks in the subtree
859 * @cs: the cpuset to consider
860 * @new_cpus: temp variable for calculating new effective_cpus
862 * When congifured cpumask is changed, the effective cpumasks of this cpuset
863 * and all its descendants need to be updated.
865 * On legacy hierachy, effective_cpus will be the same with cpu_allowed.
867 * Called with cpuset_mutex held
869 static void update_cpumasks_hier(struct cpuset
*cs
, struct cpumask
*new_cpus
)
872 struct cgroup_subsys_state
*pos_css
;
875 cpuset_for_each_descendant_pre(cp
, pos_css
, cs
) {
876 struct cpuset
*parent
= parent_cs(cp
);
878 cpumask_and(new_cpus
, cp
->cpus_allowed
, parent
->effective_cpus
);
881 * If it becomes empty, inherit the effective mask of the
882 * parent, which is guaranteed to have some CPUs.
884 if (cpumask_empty(new_cpus
))
885 cpumask_copy(new_cpus
, parent
->effective_cpus
);
887 /* Skip the whole subtree if the cpumask remains the same. */
888 if (cpumask_equal(new_cpus
, cp
->effective_cpus
)) {
889 pos_css
= css_rightmost_descendant(pos_css
);
893 if (!css_tryget_online(&cp
->css
))
897 mutex_lock(&callback_mutex
);
898 cpumask_copy(cp
->effective_cpus
, new_cpus
);
899 mutex_unlock(&callback_mutex
);
901 WARN_ON(!cgroup_on_dfl(cp
->css
.cgroup
) &&
902 !cpumask_equal(cp
->cpus_allowed
, cp
->effective_cpus
));
904 update_tasks_cpumask(cp
);
913 * update_cpumask - update the cpus_allowed mask of a cpuset and all tasks in it
914 * @cs: the cpuset to consider
915 * @trialcs: trial cpuset
916 * @buf: buffer of cpu numbers written to this cpuset
918 static int update_cpumask(struct cpuset
*cs
, struct cpuset
*trialcs
,
922 int is_load_balanced
;
924 /* top_cpuset.cpus_allowed tracks cpu_online_mask; it's read-only */
925 if (cs
== &top_cpuset
)
929 * An empty cpus_allowed is ok only if the cpuset has no tasks.
930 * Since cpulist_parse() fails on an empty mask, we special case
931 * that parsing. The validate_change() call ensures that cpusets
932 * with tasks have cpus.
935 cpumask_clear(trialcs
->cpus_allowed
);
937 retval
= cpulist_parse(buf
, trialcs
->cpus_allowed
);
941 if (!cpumask_subset(trialcs
->cpus_allowed
, cpu_active_mask
))
945 /* Nothing to do if the cpus didn't change */
946 if (cpumask_equal(cs
->cpus_allowed
, trialcs
->cpus_allowed
))
949 retval
= validate_change(cs
, trialcs
);
953 is_load_balanced
= is_sched_load_balance(trialcs
);
955 mutex_lock(&callback_mutex
);
956 cpumask_copy(cs
->cpus_allowed
, trialcs
->cpus_allowed
);
957 mutex_unlock(&callback_mutex
);
959 /* use trialcs->cpus_allowed as a temp variable */
960 update_cpumasks_hier(cs
, trialcs
->cpus_allowed
);
962 if (is_load_balanced
)
963 rebuild_sched_domains_locked();
970 * Migrate memory region from one set of nodes to another.
972 * Temporarilly set tasks mems_allowed to target nodes of migration,
973 * so that the migration code can allocate pages on these nodes.
975 * While the mm_struct we are migrating is typically from some
976 * other task, the task_struct mems_allowed that we are hacking
977 * is for our current task, which must allocate new pages for that
978 * migrating memory region.
981 static void cpuset_migrate_mm(struct mm_struct
*mm
, const nodemask_t
*from
,
982 const nodemask_t
*to
)
984 struct task_struct
*tsk
= current
;
985 struct cpuset
*mems_cs
;
987 tsk
->mems_allowed
= *to
;
989 do_migrate_pages(mm
, from
, to
, MPOL_MF_MOVE_ALL
);
992 mems_cs
= effective_nodemask_cpuset(task_cs(tsk
));
993 guarantee_online_mems(mems_cs
, &tsk
->mems_allowed
);
998 * cpuset_change_task_nodemask - change task's mems_allowed and mempolicy
999 * @tsk: the task to change
1000 * @newmems: new nodes that the task will be set
1002 * In order to avoid seeing no nodes if the old and new nodes are disjoint,
1003 * we structure updates as setting all new allowed nodes, then clearing newly
1006 static void cpuset_change_task_nodemask(struct task_struct
*tsk
,
1007 nodemask_t
*newmems
)
1012 * Allow tasks that have access to memory reserves because they have
1013 * been OOM killed to get memory anywhere.
1015 if (unlikely(test_thread_flag(TIF_MEMDIE
)))
1017 if (current
->flags
& PF_EXITING
) /* Let dying task have memory */
1022 * Determine if a loop is necessary if another thread is doing
1023 * read_mems_allowed_begin(). If at least one node remains unchanged and
1024 * tsk does not have a mempolicy, then an empty nodemask will not be
1025 * possible when mems_allowed is larger than a word.
1027 need_loop
= task_has_mempolicy(tsk
) ||
1028 !nodes_intersects(*newmems
, tsk
->mems_allowed
);
1031 local_irq_disable();
1032 write_seqcount_begin(&tsk
->mems_allowed_seq
);
1035 nodes_or(tsk
->mems_allowed
, tsk
->mems_allowed
, *newmems
);
1036 mpol_rebind_task(tsk
, newmems
, MPOL_REBIND_STEP1
);
1038 mpol_rebind_task(tsk
, newmems
, MPOL_REBIND_STEP2
);
1039 tsk
->mems_allowed
= *newmems
;
1042 write_seqcount_end(&tsk
->mems_allowed_seq
);
1049 static void *cpuset_being_rebound
;
1052 * update_tasks_nodemask - Update the nodemasks of tasks in the cpuset.
1053 * @cs: the cpuset in which each task's mems_allowed mask needs to be changed
1055 * Iterate through each task of @cs updating its mems_allowed to the
1056 * effective cpuset's. As this function is called with cpuset_mutex held,
1057 * cpuset membership stays stable.
1059 static void update_tasks_nodemask(struct cpuset
*cs
)
1061 static nodemask_t newmems
; /* protected by cpuset_mutex */
1062 struct cpuset
*mems_cs
= effective_nodemask_cpuset(cs
);
1063 struct css_task_iter it
;
1064 struct task_struct
*task
;
1066 cpuset_being_rebound
= cs
; /* causes mpol_dup() rebind */
1068 guarantee_online_mems(mems_cs
, &newmems
);
1071 * The mpol_rebind_mm() call takes mmap_sem, which we couldn't
1072 * take while holding tasklist_lock. Forks can happen - the
1073 * mpol_dup() cpuset_being_rebound check will catch such forks,
1074 * and rebind their vma mempolicies too. Because we still hold
1075 * the global cpuset_mutex, we know that no other rebind effort
1076 * will be contending for the global variable cpuset_being_rebound.
1077 * It's ok if we rebind the same mm twice; mpol_rebind_mm()
1078 * is idempotent. Also migrate pages in each mm to new nodes.
1080 css_task_iter_start(&cs
->css
, &it
);
1081 while ((task
= css_task_iter_next(&it
))) {
1082 struct mm_struct
*mm
;
1085 cpuset_change_task_nodemask(task
, &newmems
);
1087 mm
= get_task_mm(task
);
1091 migrate
= is_memory_migrate(cs
);
1093 mpol_rebind_mm(mm
, &cs
->mems_allowed
);
1095 cpuset_migrate_mm(mm
, &cs
->old_mems_allowed
, &newmems
);
1098 css_task_iter_end(&it
);
1101 * All the tasks' nodemasks have been updated, update
1102 * cs->old_mems_allowed.
1104 cs
->old_mems_allowed
= newmems
;
1106 /* We're done rebinding vmas to this cpuset's new mems_allowed. */
1107 cpuset_being_rebound
= NULL
;
1111 * update_nodemasks_hier - Update effective nodemasks and tasks in the subtree
1112 * @cs: the cpuset to consider
1113 * @new_mems: a temp variable for calculating new effective_mems
1115 * When configured nodemask is changed, the effective nodemasks of this cpuset
1116 * and all its descendants need to be updated.
1118 * On legacy hiearchy, effective_mems will be the same with mems_allowed.
1120 * Called with cpuset_mutex held
1122 static void update_nodemasks_hier(struct cpuset
*cs
, nodemask_t
*new_mems
)
1125 struct cgroup_subsys_state
*pos_css
;
1128 cpuset_for_each_descendant_pre(cp
, pos_css
, cs
) {
1129 struct cpuset
*parent
= parent_cs(cp
);
1131 nodes_and(*new_mems
, cp
->mems_allowed
, parent
->effective_mems
);
1134 * If it becomes empty, inherit the effective mask of the
1135 * parent, which is guaranteed to have some MEMs.
1137 if (nodes_empty(*new_mems
))
1138 *new_mems
= parent
->effective_mems
;
1140 /* Skip the whole subtree if the nodemask remains the same. */
1141 if (nodes_equal(*new_mems
, cp
->effective_mems
)) {
1142 pos_css
= css_rightmost_descendant(pos_css
);
1146 if (!css_tryget_online(&cp
->css
))
1150 mutex_lock(&callback_mutex
);
1151 cp
->effective_mems
= *new_mems
;
1152 mutex_unlock(&callback_mutex
);
1154 WARN_ON(!cgroup_on_dfl(cp
->css
.cgroup
) &&
1155 nodes_equal(cp
->mems_allowed
, cp
->effective_mems
));
1157 update_tasks_nodemask(cp
);
1166 * Handle user request to change the 'mems' memory placement
1167 * of a cpuset. Needs to validate the request, update the
1168 * cpusets mems_allowed, and for each task in the cpuset,
1169 * update mems_allowed and rebind task's mempolicy and any vma
1170 * mempolicies and if the cpuset is marked 'memory_migrate',
1171 * migrate the tasks pages to the new memory.
1173 * Call with cpuset_mutex held. May take callback_mutex during call.
1174 * Will take tasklist_lock, scan tasklist for tasks in cpuset cs,
1175 * lock each such tasks mm->mmap_sem, scan its vma's and rebind
1176 * their mempolicies to the cpusets new mems_allowed.
1178 static int update_nodemask(struct cpuset
*cs
, struct cpuset
*trialcs
,
1184 * top_cpuset.mems_allowed tracks node_stats[N_MEMORY];
1187 if (cs
== &top_cpuset
) {
1193 * An empty mems_allowed is ok iff there are no tasks in the cpuset.
1194 * Since nodelist_parse() fails on an empty mask, we special case
1195 * that parsing. The validate_change() call ensures that cpusets
1196 * with tasks have memory.
1199 nodes_clear(trialcs
->mems_allowed
);
1201 retval
= nodelist_parse(buf
, trialcs
->mems_allowed
);
1205 if (!nodes_subset(trialcs
->mems_allowed
,
1206 node_states
[N_MEMORY
])) {
1212 if (nodes_equal(cs
->mems_allowed
, trialcs
->mems_allowed
)) {
1213 retval
= 0; /* Too easy - nothing to do */
1216 retval
= validate_change(cs
, trialcs
);
1220 mutex_lock(&callback_mutex
);
1221 cs
->mems_allowed
= trialcs
->mems_allowed
;
1222 mutex_unlock(&callback_mutex
);
1224 /* use trialcs->mems_allowed as a temp variable */
1225 update_nodemasks_hier(cs
, &cs
->mems_allowed
);
1230 int current_cpuset_is_being_rebound(void)
1232 return task_cs(current
) == cpuset_being_rebound
;
1235 static int update_relax_domain_level(struct cpuset
*cs
, s64 val
)
1238 if (val
< -1 || val
>= sched_domain_level_max
)
1242 if (val
!= cs
->relax_domain_level
) {
1243 cs
->relax_domain_level
= val
;
1244 if (!cpumask_empty(cs
->cpus_allowed
) &&
1245 is_sched_load_balance(cs
))
1246 rebuild_sched_domains_locked();
1253 * update_tasks_flags - update the spread flags of tasks in the cpuset.
1254 * @cs: the cpuset in which each task's spread flags needs to be changed
1256 * Iterate through each task of @cs updating its spread flags. As this
1257 * function is called with cpuset_mutex held, cpuset membership stays
1260 static void update_tasks_flags(struct cpuset
*cs
)
1262 struct css_task_iter it
;
1263 struct task_struct
*task
;
1265 css_task_iter_start(&cs
->css
, &it
);
1266 while ((task
= css_task_iter_next(&it
)))
1267 cpuset_update_task_spread_flag(cs
, task
);
1268 css_task_iter_end(&it
);
1272 * update_flag - read a 0 or a 1 in a file and update associated flag
1273 * bit: the bit to update (see cpuset_flagbits_t)
1274 * cs: the cpuset to update
1275 * turning_on: whether the flag is being set or cleared
1277 * Call with cpuset_mutex held.
1280 static int update_flag(cpuset_flagbits_t bit
, struct cpuset
*cs
,
1283 struct cpuset
*trialcs
;
1284 int balance_flag_changed
;
1285 int spread_flag_changed
;
1288 trialcs
= alloc_trial_cpuset(cs
);
1293 set_bit(bit
, &trialcs
->flags
);
1295 clear_bit(bit
, &trialcs
->flags
);
1297 err
= validate_change(cs
, trialcs
);
1301 balance_flag_changed
= (is_sched_load_balance(cs
) !=
1302 is_sched_load_balance(trialcs
));
1304 spread_flag_changed
= ((is_spread_slab(cs
) != is_spread_slab(trialcs
))
1305 || (is_spread_page(cs
) != is_spread_page(trialcs
)));
1307 mutex_lock(&callback_mutex
);
1308 cs
->flags
= trialcs
->flags
;
1309 mutex_unlock(&callback_mutex
);
1311 if (!cpumask_empty(trialcs
->cpus_allowed
) && balance_flag_changed
)
1312 rebuild_sched_domains_locked();
1314 if (spread_flag_changed
)
1315 update_tasks_flags(cs
);
1317 free_trial_cpuset(trialcs
);
1322 * Frequency meter - How fast is some event occurring?
1324 * These routines manage a digitally filtered, constant time based,
1325 * event frequency meter. There are four routines:
1326 * fmeter_init() - initialize a frequency meter.
1327 * fmeter_markevent() - called each time the event happens.
1328 * fmeter_getrate() - returns the recent rate of such events.
1329 * fmeter_update() - internal routine used to update fmeter.
1331 * A common data structure is passed to each of these routines,
1332 * which is used to keep track of the state required to manage the
1333 * frequency meter and its digital filter.
1335 * The filter works on the number of events marked per unit time.
1336 * The filter is single-pole low-pass recursive (IIR). The time unit
1337 * is 1 second. Arithmetic is done using 32-bit integers scaled to
1338 * simulate 3 decimal digits of precision (multiplied by 1000).
1340 * With an FM_COEF of 933, and a time base of 1 second, the filter
1341 * has a half-life of 10 seconds, meaning that if the events quit
1342 * happening, then the rate returned from the fmeter_getrate()
1343 * will be cut in half each 10 seconds, until it converges to zero.
1345 * It is not worth doing a real infinitely recursive filter. If more
1346 * than FM_MAXTICKS ticks have elapsed since the last filter event,
1347 * just compute FM_MAXTICKS ticks worth, by which point the level
1350 * Limit the count of unprocessed events to FM_MAXCNT, so as to avoid
1351 * arithmetic overflow in the fmeter_update() routine.
1353 * Given the simple 32 bit integer arithmetic used, this meter works
1354 * best for reporting rates between one per millisecond (msec) and
1355 * one per 32 (approx) seconds. At constant rates faster than one
1356 * per msec it maxes out at values just under 1,000,000. At constant
1357 * rates between one per msec, and one per second it will stabilize
1358 * to a value N*1000, where N is the rate of events per second.
1359 * At constant rates between one per second and one per 32 seconds,
1360 * it will be choppy, moving up on the seconds that have an event,
1361 * and then decaying until the next event. At rates slower than
1362 * about one in 32 seconds, it decays all the way back to zero between
1366 #define FM_COEF 933 /* coefficient for half-life of 10 secs */
1367 #define FM_MAXTICKS ((time_t)99) /* useless computing more ticks than this */
1368 #define FM_MAXCNT 1000000 /* limit cnt to avoid overflow */
1369 #define FM_SCALE 1000 /* faux fixed point scale */
1371 /* Initialize a frequency meter */
1372 static void fmeter_init(struct fmeter
*fmp
)
1377 spin_lock_init(&fmp
->lock
);
1380 /* Internal meter update - process cnt events and update value */
1381 static void fmeter_update(struct fmeter
*fmp
)
1383 time_t now
= get_seconds();
1384 time_t ticks
= now
- fmp
->time
;
1389 ticks
= min(FM_MAXTICKS
, ticks
);
1391 fmp
->val
= (FM_COEF
* fmp
->val
) / FM_SCALE
;
1394 fmp
->val
+= ((FM_SCALE
- FM_COEF
) * fmp
->cnt
) / FM_SCALE
;
1398 /* Process any previous ticks, then bump cnt by one (times scale). */
1399 static void fmeter_markevent(struct fmeter
*fmp
)
1401 spin_lock(&fmp
->lock
);
1403 fmp
->cnt
= min(FM_MAXCNT
, fmp
->cnt
+ FM_SCALE
);
1404 spin_unlock(&fmp
->lock
);
1407 /* Process any previous ticks, then return current value. */
1408 static int fmeter_getrate(struct fmeter
*fmp
)
1412 spin_lock(&fmp
->lock
);
1415 spin_unlock(&fmp
->lock
);
1419 static struct cpuset
*cpuset_attach_old_cs
;
1421 /* Called by cgroups to determine if a cpuset is usable; cpuset_mutex held */
1422 static int cpuset_can_attach(struct cgroup_subsys_state
*css
,
1423 struct cgroup_taskset
*tset
)
1425 struct cpuset
*cs
= css_cs(css
);
1426 struct task_struct
*task
;
1429 /* used later by cpuset_attach() */
1430 cpuset_attach_old_cs
= task_cs(cgroup_taskset_first(tset
));
1432 mutex_lock(&cpuset_mutex
);
1434 /* allow moving tasks into an empty cpuset if on default hierarchy */
1436 if (!cgroup_on_dfl(css
->cgroup
) &&
1437 (cpumask_empty(cs
->cpus_allowed
) || nodes_empty(cs
->mems_allowed
)))
1440 cgroup_taskset_for_each(task
, tset
) {
1442 * Kthreads which disallow setaffinity shouldn't be moved
1443 * to a new cpuset; we don't want to change their cpu
1444 * affinity and isolating such threads by their set of
1445 * allowed nodes is unnecessary. Thus, cpusets are not
1446 * applicable for such threads. This prevents checking for
1447 * success of set_cpus_allowed_ptr() on all attached tasks
1448 * before cpus_allowed may be changed.
1451 if (task
->flags
& PF_NO_SETAFFINITY
)
1453 ret
= security_task_setscheduler(task
);
1459 * Mark attach is in progress. This makes validate_change() fail
1460 * changes which zero cpus/mems_allowed.
1462 cs
->attach_in_progress
++;
1465 mutex_unlock(&cpuset_mutex
);
1469 static void cpuset_cancel_attach(struct cgroup_subsys_state
*css
,
1470 struct cgroup_taskset
*tset
)
1472 mutex_lock(&cpuset_mutex
);
1473 css_cs(css
)->attach_in_progress
--;
1474 mutex_unlock(&cpuset_mutex
);
1478 * Protected by cpuset_mutex. cpus_attach is used only by cpuset_attach()
1479 * but we can't allocate it dynamically there. Define it global and
1480 * allocate from cpuset_init().
1482 static cpumask_var_t cpus_attach
;
1484 static void cpuset_attach(struct cgroup_subsys_state
*css
,
1485 struct cgroup_taskset
*tset
)
1487 /* static buf protected by cpuset_mutex */
1488 static nodemask_t cpuset_attach_nodemask_to
;
1489 struct mm_struct
*mm
;
1490 struct task_struct
*task
;
1491 struct task_struct
*leader
= cgroup_taskset_first(tset
);
1492 struct cpuset
*cs
= css_cs(css
);
1493 struct cpuset
*oldcs
= cpuset_attach_old_cs
;
1494 struct cpuset
*cpus_cs
= effective_cpumask_cpuset(cs
);
1495 struct cpuset
*mems_cs
= effective_nodemask_cpuset(cs
);
1497 mutex_lock(&cpuset_mutex
);
1499 /* prepare for attach */
1500 if (cs
== &top_cpuset
)
1501 cpumask_copy(cpus_attach
, cpu_possible_mask
);
1503 guarantee_online_cpus(cpus_cs
, cpus_attach
);
1505 guarantee_online_mems(mems_cs
, &cpuset_attach_nodemask_to
);
1507 cgroup_taskset_for_each(task
, tset
) {
1509 * can_attach beforehand should guarantee that this doesn't
1510 * fail. TODO: have a better way to handle failure here
1512 WARN_ON_ONCE(set_cpus_allowed_ptr(task
, cpus_attach
));
1514 cpuset_change_task_nodemask(task
, &cpuset_attach_nodemask_to
);
1515 cpuset_update_task_spread_flag(cs
, task
);
1519 * Change mm, possibly for multiple threads in a threadgroup. This is
1520 * expensive and may sleep.
1522 cpuset_attach_nodemask_to
= cs
->mems_allowed
;
1523 mm
= get_task_mm(leader
);
1525 struct cpuset
*mems_oldcs
= effective_nodemask_cpuset(oldcs
);
1527 mpol_rebind_mm(mm
, &cpuset_attach_nodemask_to
);
1530 * old_mems_allowed is the same with mems_allowed here, except
1531 * if this task is being moved automatically due to hotplug.
1532 * In that case @mems_allowed has been updated and is empty,
1533 * so @old_mems_allowed is the right nodesets that we migrate
1536 if (is_memory_migrate(cs
)) {
1537 cpuset_migrate_mm(mm
, &mems_oldcs
->old_mems_allowed
,
1538 &cpuset_attach_nodemask_to
);
1543 cs
->old_mems_allowed
= cpuset_attach_nodemask_to
;
1545 cs
->attach_in_progress
--;
1546 if (!cs
->attach_in_progress
)
1547 wake_up(&cpuset_attach_wq
);
1549 mutex_unlock(&cpuset_mutex
);
1552 /* The various types of files and directories in a cpuset file system */
1555 FILE_MEMORY_MIGRATE
,
1561 FILE_SCHED_LOAD_BALANCE
,
1562 FILE_SCHED_RELAX_DOMAIN_LEVEL
,
1563 FILE_MEMORY_PRESSURE_ENABLED
,
1564 FILE_MEMORY_PRESSURE
,
1567 } cpuset_filetype_t
;
1569 static int cpuset_write_u64(struct cgroup_subsys_state
*css
, struct cftype
*cft
,
1572 struct cpuset
*cs
= css_cs(css
);
1573 cpuset_filetype_t type
= cft
->private;
1576 mutex_lock(&cpuset_mutex
);
1577 if (!is_cpuset_online(cs
)) {
1583 case FILE_CPU_EXCLUSIVE
:
1584 retval
= update_flag(CS_CPU_EXCLUSIVE
, cs
, val
);
1586 case FILE_MEM_EXCLUSIVE
:
1587 retval
= update_flag(CS_MEM_EXCLUSIVE
, cs
, val
);
1589 case FILE_MEM_HARDWALL
:
1590 retval
= update_flag(CS_MEM_HARDWALL
, cs
, val
);
1592 case FILE_SCHED_LOAD_BALANCE
:
1593 retval
= update_flag(CS_SCHED_LOAD_BALANCE
, cs
, val
);
1595 case FILE_MEMORY_MIGRATE
:
1596 retval
= update_flag(CS_MEMORY_MIGRATE
, cs
, val
);
1598 case FILE_MEMORY_PRESSURE_ENABLED
:
1599 cpuset_memory_pressure_enabled
= !!val
;
1601 case FILE_MEMORY_PRESSURE
:
1604 case FILE_SPREAD_PAGE
:
1605 retval
= update_flag(CS_SPREAD_PAGE
, cs
, val
);
1607 case FILE_SPREAD_SLAB
:
1608 retval
= update_flag(CS_SPREAD_SLAB
, cs
, val
);
1615 mutex_unlock(&cpuset_mutex
);
1619 static int cpuset_write_s64(struct cgroup_subsys_state
*css
, struct cftype
*cft
,
1622 struct cpuset
*cs
= css_cs(css
);
1623 cpuset_filetype_t type
= cft
->private;
1624 int retval
= -ENODEV
;
1626 mutex_lock(&cpuset_mutex
);
1627 if (!is_cpuset_online(cs
))
1631 case FILE_SCHED_RELAX_DOMAIN_LEVEL
:
1632 retval
= update_relax_domain_level(cs
, val
);
1639 mutex_unlock(&cpuset_mutex
);
1644 * Common handling for a write to a "cpus" or "mems" file.
1646 static ssize_t
cpuset_write_resmask(struct kernfs_open_file
*of
,
1647 char *buf
, size_t nbytes
, loff_t off
)
1649 struct cpuset
*cs
= css_cs(of_css(of
));
1650 struct cpuset
*trialcs
;
1651 int retval
= -ENODEV
;
1653 buf
= strstrip(buf
);
1656 * CPU or memory hotunplug may leave @cs w/o any execution
1657 * resources, in which case the hotplug code asynchronously updates
1658 * configuration and transfers all tasks to the nearest ancestor
1659 * which can execute.
1661 * As writes to "cpus" or "mems" may restore @cs's execution
1662 * resources, wait for the previously scheduled operations before
1663 * proceeding, so that we don't end up keep removing tasks added
1664 * after execution capability is restored.
1666 flush_work(&cpuset_hotplug_work
);
1668 mutex_lock(&cpuset_mutex
);
1669 if (!is_cpuset_online(cs
))
1672 trialcs
= alloc_trial_cpuset(cs
);
1678 switch (of_cft(of
)->private) {
1680 retval
= update_cpumask(cs
, trialcs
, buf
);
1683 retval
= update_nodemask(cs
, trialcs
, buf
);
1690 free_trial_cpuset(trialcs
);
1692 mutex_unlock(&cpuset_mutex
);
1693 return retval
?: nbytes
;
1697 * These ascii lists should be read in a single call, by using a user
1698 * buffer large enough to hold the entire map. If read in smaller
1699 * chunks, there is no guarantee of atomicity. Since the display format
1700 * used, list of ranges of sequential numbers, is variable length,
1701 * and since these maps can change value dynamically, one could read
1702 * gibberish by doing partial reads while a list was changing.
1704 static int cpuset_common_seq_show(struct seq_file
*sf
, void *v
)
1706 struct cpuset
*cs
= css_cs(seq_css(sf
));
1707 cpuset_filetype_t type
= seq_cft(sf
)->private;
1712 count
= seq_get_buf(sf
, &buf
);
1715 mutex_lock(&callback_mutex
);
1719 s
+= cpulist_scnprintf(s
, count
, cs
->cpus_allowed
);
1722 s
+= nodelist_scnprintf(s
, count
, cs
->mems_allowed
);
1729 if (s
< buf
+ count
- 1) {
1731 seq_commit(sf
, s
- buf
);
1736 mutex_unlock(&callback_mutex
);
1740 static u64
cpuset_read_u64(struct cgroup_subsys_state
*css
, struct cftype
*cft
)
1742 struct cpuset
*cs
= css_cs(css
);
1743 cpuset_filetype_t type
= cft
->private;
1745 case FILE_CPU_EXCLUSIVE
:
1746 return is_cpu_exclusive(cs
);
1747 case FILE_MEM_EXCLUSIVE
:
1748 return is_mem_exclusive(cs
);
1749 case FILE_MEM_HARDWALL
:
1750 return is_mem_hardwall(cs
);
1751 case FILE_SCHED_LOAD_BALANCE
:
1752 return is_sched_load_balance(cs
);
1753 case FILE_MEMORY_MIGRATE
:
1754 return is_memory_migrate(cs
);
1755 case FILE_MEMORY_PRESSURE_ENABLED
:
1756 return cpuset_memory_pressure_enabled
;
1757 case FILE_MEMORY_PRESSURE
:
1758 return fmeter_getrate(&cs
->fmeter
);
1759 case FILE_SPREAD_PAGE
:
1760 return is_spread_page(cs
);
1761 case FILE_SPREAD_SLAB
:
1762 return is_spread_slab(cs
);
1767 /* Unreachable but makes gcc happy */
1771 static s64
cpuset_read_s64(struct cgroup_subsys_state
*css
, struct cftype
*cft
)
1773 struct cpuset
*cs
= css_cs(css
);
1774 cpuset_filetype_t type
= cft
->private;
1776 case FILE_SCHED_RELAX_DOMAIN_LEVEL
:
1777 return cs
->relax_domain_level
;
1782 /* Unrechable but makes gcc happy */
1788 * for the common functions, 'private' gives the type of file
1791 static struct cftype files
[] = {
1794 .seq_show
= cpuset_common_seq_show
,
1795 .write
= cpuset_write_resmask
,
1796 .max_write_len
= (100U + 6 * NR_CPUS
),
1797 .private = FILE_CPULIST
,
1802 .seq_show
= cpuset_common_seq_show
,
1803 .write
= cpuset_write_resmask
,
1804 .max_write_len
= (100U + 6 * MAX_NUMNODES
),
1805 .private = FILE_MEMLIST
,
1809 .name
= "cpu_exclusive",
1810 .read_u64
= cpuset_read_u64
,
1811 .write_u64
= cpuset_write_u64
,
1812 .private = FILE_CPU_EXCLUSIVE
,
1816 .name
= "mem_exclusive",
1817 .read_u64
= cpuset_read_u64
,
1818 .write_u64
= cpuset_write_u64
,
1819 .private = FILE_MEM_EXCLUSIVE
,
1823 .name
= "mem_hardwall",
1824 .read_u64
= cpuset_read_u64
,
1825 .write_u64
= cpuset_write_u64
,
1826 .private = FILE_MEM_HARDWALL
,
1830 .name
= "sched_load_balance",
1831 .read_u64
= cpuset_read_u64
,
1832 .write_u64
= cpuset_write_u64
,
1833 .private = FILE_SCHED_LOAD_BALANCE
,
1837 .name
= "sched_relax_domain_level",
1838 .read_s64
= cpuset_read_s64
,
1839 .write_s64
= cpuset_write_s64
,
1840 .private = FILE_SCHED_RELAX_DOMAIN_LEVEL
,
1844 .name
= "memory_migrate",
1845 .read_u64
= cpuset_read_u64
,
1846 .write_u64
= cpuset_write_u64
,
1847 .private = FILE_MEMORY_MIGRATE
,
1851 .name
= "memory_pressure",
1852 .read_u64
= cpuset_read_u64
,
1853 .write_u64
= cpuset_write_u64
,
1854 .private = FILE_MEMORY_PRESSURE
,
1859 .name
= "memory_spread_page",
1860 .read_u64
= cpuset_read_u64
,
1861 .write_u64
= cpuset_write_u64
,
1862 .private = FILE_SPREAD_PAGE
,
1866 .name
= "memory_spread_slab",
1867 .read_u64
= cpuset_read_u64
,
1868 .write_u64
= cpuset_write_u64
,
1869 .private = FILE_SPREAD_SLAB
,
1873 .name
= "memory_pressure_enabled",
1874 .flags
= CFTYPE_ONLY_ON_ROOT
,
1875 .read_u64
= cpuset_read_u64
,
1876 .write_u64
= cpuset_write_u64
,
1877 .private = FILE_MEMORY_PRESSURE_ENABLED
,
1884 * cpuset_css_alloc - allocate a cpuset css
1885 * cgrp: control group that the new cpuset will be part of
1888 static struct cgroup_subsys_state
*
1889 cpuset_css_alloc(struct cgroup_subsys_state
*parent_css
)
1894 return &top_cpuset
.css
;
1896 cs
= kzalloc(sizeof(*cs
), GFP_KERNEL
);
1898 return ERR_PTR(-ENOMEM
);
1899 if (!alloc_cpumask_var(&cs
->cpus_allowed
, GFP_KERNEL
))
1901 if (!alloc_cpumask_var(&cs
->effective_cpus
, GFP_KERNEL
))
1904 set_bit(CS_SCHED_LOAD_BALANCE
, &cs
->flags
);
1905 cpumask_clear(cs
->cpus_allowed
);
1906 nodes_clear(cs
->mems_allowed
);
1907 cpumask_clear(cs
->effective_cpus
);
1908 nodes_clear(cs
->effective_mems
);
1909 fmeter_init(&cs
->fmeter
);
1910 cs
->relax_domain_level
= -1;
1915 free_cpumask_var(cs
->cpus_allowed
);
1918 return ERR_PTR(-ENOMEM
);
1921 static int cpuset_css_online(struct cgroup_subsys_state
*css
)
1923 struct cpuset
*cs
= css_cs(css
);
1924 struct cpuset
*parent
= parent_cs(cs
);
1925 struct cpuset
*tmp_cs
;
1926 struct cgroup_subsys_state
*pos_css
;
1931 mutex_lock(&cpuset_mutex
);
1933 set_bit(CS_ONLINE
, &cs
->flags
);
1934 if (is_spread_page(parent
))
1935 set_bit(CS_SPREAD_PAGE
, &cs
->flags
);
1936 if (is_spread_slab(parent
))
1937 set_bit(CS_SPREAD_SLAB
, &cs
->flags
);
1941 mutex_lock(&callback_mutex
);
1942 if (cgroup_on_dfl(cs
->css
.cgroup
)) {
1943 cpumask_copy(cs
->effective_cpus
, parent
->effective_cpus
);
1944 cs
->effective_mems
= parent
->effective_mems
;
1946 mutex_unlock(&callback_mutex
);
1948 if (!test_bit(CGRP_CPUSET_CLONE_CHILDREN
, &css
->cgroup
->flags
))
1952 * Clone @parent's configuration if CGRP_CPUSET_CLONE_CHILDREN is
1953 * set. This flag handling is implemented in cgroup core for
1954 * histrical reasons - the flag may be specified during mount.
1956 * Currently, if any sibling cpusets have exclusive cpus or mem, we
1957 * refuse to clone the configuration - thereby refusing the task to
1958 * be entered, and as a result refusing the sys_unshare() or
1959 * clone() which initiated it. If this becomes a problem for some
1960 * users who wish to allow that scenario, then this could be
1961 * changed to grant parent->cpus_allowed-sibling_cpus_exclusive
1962 * (and likewise for mems) to the new cgroup.
1965 cpuset_for_each_child(tmp_cs
, pos_css
, parent
) {
1966 if (is_mem_exclusive(tmp_cs
) || is_cpu_exclusive(tmp_cs
)) {
1973 mutex_lock(&callback_mutex
);
1974 cs
->mems_allowed
= parent
->mems_allowed
;
1975 cpumask_copy(cs
->cpus_allowed
, parent
->cpus_allowed
);
1976 mutex_unlock(&callback_mutex
);
1978 mutex_unlock(&cpuset_mutex
);
1983 * If the cpuset being removed has its flag 'sched_load_balance'
1984 * enabled, then simulate turning sched_load_balance off, which
1985 * will call rebuild_sched_domains_locked().
1988 static void cpuset_css_offline(struct cgroup_subsys_state
*css
)
1990 struct cpuset
*cs
= css_cs(css
);
1992 mutex_lock(&cpuset_mutex
);
1994 if (is_sched_load_balance(cs
))
1995 update_flag(CS_SCHED_LOAD_BALANCE
, cs
, 0);
1998 clear_bit(CS_ONLINE
, &cs
->flags
);
2000 mutex_unlock(&cpuset_mutex
);
2003 static void cpuset_css_free(struct cgroup_subsys_state
*css
)
2005 struct cpuset
*cs
= css_cs(css
);
2007 free_cpumask_var(cs
->effective_cpus
);
2008 free_cpumask_var(cs
->cpus_allowed
);
2012 struct cgroup_subsys cpuset_cgrp_subsys
= {
2013 .css_alloc
= cpuset_css_alloc
,
2014 .css_online
= cpuset_css_online
,
2015 .css_offline
= cpuset_css_offline
,
2016 .css_free
= cpuset_css_free
,
2017 .can_attach
= cpuset_can_attach
,
2018 .cancel_attach
= cpuset_cancel_attach
,
2019 .attach
= cpuset_attach
,
2020 .base_cftypes
= files
,
2025 * cpuset_init - initialize cpusets at system boot
2027 * Description: Initialize top_cpuset and the cpuset internal file system,
2030 int __init
cpuset_init(void)
2034 if (!alloc_cpumask_var(&top_cpuset
.cpus_allowed
, GFP_KERNEL
))
2036 if (!alloc_cpumask_var(&top_cpuset
.effective_cpus
, GFP_KERNEL
))
2039 cpumask_setall(top_cpuset
.cpus_allowed
);
2040 nodes_setall(top_cpuset
.mems_allowed
);
2041 cpumask_setall(top_cpuset
.effective_cpus
);
2042 nodes_setall(top_cpuset
.effective_mems
);
2044 fmeter_init(&top_cpuset
.fmeter
);
2045 set_bit(CS_SCHED_LOAD_BALANCE
, &top_cpuset
.flags
);
2046 top_cpuset
.relax_domain_level
= -1;
2048 err
= register_filesystem(&cpuset_fs_type
);
2052 if (!alloc_cpumask_var(&cpus_attach
, GFP_KERNEL
))
2059 * If CPU and/or memory hotplug handlers, below, unplug any CPUs
2060 * or memory nodes, we need to walk over the cpuset hierarchy,
2061 * removing that CPU or node from all cpusets. If this removes the
2062 * last CPU or node from a cpuset, then move the tasks in the empty
2063 * cpuset to its next-highest non-empty parent.
2065 static void remove_tasks_in_empty_cpuset(struct cpuset
*cs
)
2067 struct cpuset
*parent
;
2070 * Find its next-highest non-empty parent, (top cpuset
2071 * has online cpus, so can't be empty).
2073 parent
= parent_cs(cs
);
2074 while (cpumask_empty(parent
->cpus_allowed
) ||
2075 nodes_empty(parent
->mems_allowed
))
2076 parent
= parent_cs(parent
);
2078 if (cgroup_transfer_tasks(parent
->css
.cgroup
, cs
->css
.cgroup
)) {
2079 pr_err("cpuset: failed to transfer tasks out of empty cpuset ");
2080 pr_cont_cgroup_name(cs
->css
.cgroup
);
2086 * cpuset_hotplug_update_tasks - update tasks in a cpuset for hotunplug
2087 * @cs: cpuset in interest
2089 * Compare @cs's cpu and mem masks against top_cpuset and if some have gone
2090 * offline, update @cs accordingly. If @cs ends up with no CPU or memory,
2091 * all its tasks are moved to the nearest ancestor with both resources.
2093 static void cpuset_hotplug_update_tasks(struct cpuset
*cs
)
2095 static cpumask_t off_cpus
;
2096 static nodemask_t off_mems
;
2098 bool on_dfl
= cgroup_on_dfl(cs
->css
.cgroup
);
2101 wait_event(cpuset_attach_wq
, cs
->attach_in_progress
== 0);
2103 mutex_lock(&cpuset_mutex
);
2106 * We have raced with task attaching. We wait until attaching
2107 * is finished, so we won't attach a task to an empty cpuset.
2109 if (cs
->attach_in_progress
) {
2110 mutex_unlock(&cpuset_mutex
);
2114 cpumask_andnot(&off_cpus
, cs
->cpus_allowed
, top_cpuset
.cpus_allowed
);
2115 nodes_andnot(off_mems
, cs
->mems_allowed
, top_cpuset
.mems_allowed
);
2117 mutex_lock(&callback_mutex
);
2118 cpumask_andnot(cs
->cpus_allowed
, cs
->cpus_allowed
, &off_cpus
);
2120 /* Inherit the effective mask of the parent, if it becomes empty. */
2121 cpumask_andnot(cs
->effective_cpus
, cs
->effective_cpus
, &off_cpus
);
2122 if (on_dfl
&& cpumask_empty(cs
->effective_cpus
))
2123 cpumask_copy(cs
->effective_cpus
, parent_cs(cs
)->effective_cpus
);
2124 mutex_unlock(&callback_mutex
);
2127 * If on_dfl, we need to update tasks' cpumask for empty cpuset to
2128 * take on ancestor's cpumask. Otherwise, don't call
2129 * update_tasks_cpumask() if the cpuset becomes empty, as the tasks
2130 * in it will be migrated to an ancestor.
2132 if ((on_dfl
&& cpumask_empty(cs
->cpus_allowed
)) ||
2133 (!cpumask_empty(&off_cpus
) && !cpumask_empty(cs
->cpus_allowed
)))
2134 update_tasks_cpumask(cs
);
2136 mutex_lock(&callback_mutex
);
2137 nodes_andnot(cs
->mems_allowed
, cs
->mems_allowed
, off_mems
);
2139 /* Inherit the effective mask of the parent, if it becomes empty */
2140 nodes_andnot(cs
->effective_mems
, cs
->effective_mems
, off_mems
);
2141 if (on_dfl
&& nodes_empty(cs
->effective_mems
))
2142 cs
->effective_mems
= parent_cs(cs
)->effective_mems
;
2143 mutex_unlock(&callback_mutex
);
2146 * If on_dfl, we need to update tasks' nodemask for empty cpuset to
2147 * take on ancestor's nodemask. Otherwise, don't call
2148 * update_tasks_nodemask() if the cpuset becomes empty, as the
2149 * tasks in it will be migratd to an ancestor.
2151 if ((on_dfl
&& nodes_empty(cs
->mems_allowed
)) ||
2152 (!nodes_empty(off_mems
) && !nodes_empty(cs
->mems_allowed
)))
2153 update_tasks_nodemask(cs
);
2155 is_empty
= cpumask_empty(cs
->cpus_allowed
) ||
2156 nodes_empty(cs
->mems_allowed
);
2158 mutex_unlock(&cpuset_mutex
);
2161 * If on_dfl, we'll keep tasks in empty cpusets.
2163 * Otherwise move tasks to the nearest ancestor with execution
2164 * resources. This is full cgroup operation which will
2165 * also call back into cpuset. Should be done outside any lock.
2167 if (!on_dfl
&& is_empty
)
2168 remove_tasks_in_empty_cpuset(cs
);
2172 * cpuset_hotplug_workfn - handle CPU/memory hotunplug for a cpuset
2174 * This function is called after either CPU or memory configuration has
2175 * changed and updates cpuset accordingly. The top_cpuset is always
2176 * synchronized to cpu_active_mask and N_MEMORY, which is necessary in
2177 * order to make cpusets transparent (of no affect) on systems that are
2178 * actively using CPU hotplug but making no active use of cpusets.
2180 * Non-root cpusets are only affected by offlining. If any CPUs or memory
2181 * nodes have been taken down, cpuset_hotplug_update_tasks() is invoked on
2184 * Note that CPU offlining during suspend is ignored. We don't modify
2185 * cpusets across suspend/resume cycles at all.
2187 static void cpuset_hotplug_workfn(struct work_struct
*work
)
2189 static cpumask_t new_cpus
;
2190 static nodemask_t new_mems
;
2191 bool cpus_updated
, mems_updated
;
2193 mutex_lock(&cpuset_mutex
);
2195 /* fetch the available cpus/mems and find out which changed how */
2196 cpumask_copy(&new_cpus
, cpu_active_mask
);
2197 new_mems
= node_states
[N_MEMORY
];
2199 cpus_updated
= !cpumask_equal(top_cpuset
.cpus_allowed
, &new_cpus
);
2200 mems_updated
= !nodes_equal(top_cpuset
.mems_allowed
, new_mems
);
2202 /* synchronize cpus_allowed to cpu_active_mask */
2204 mutex_lock(&callback_mutex
);
2205 cpumask_copy(top_cpuset
.cpus_allowed
, &new_cpus
);
2206 cpumask_copy(top_cpuset
.effective_cpus
, &new_cpus
);
2207 mutex_unlock(&callback_mutex
);
2208 /* we don't mess with cpumasks of tasks in top_cpuset */
2211 /* synchronize mems_allowed to N_MEMORY */
2213 mutex_lock(&callback_mutex
);
2214 top_cpuset
.mems_allowed
= new_mems
;
2215 top_cpuset
.effective_mems
= new_mems
;
2216 mutex_unlock(&callback_mutex
);
2217 update_tasks_nodemask(&top_cpuset
);
2220 mutex_unlock(&cpuset_mutex
);
2222 /* if cpus or mems changed, we need to propagate to descendants */
2223 if (cpus_updated
|| mems_updated
) {
2225 struct cgroup_subsys_state
*pos_css
;
2228 cpuset_for_each_descendant_pre(cs
, pos_css
, &top_cpuset
) {
2229 if (cs
== &top_cpuset
|| !css_tryget_online(&cs
->css
))
2233 cpuset_hotplug_update_tasks(cs
);
2241 /* rebuild sched domains if cpus_allowed has changed */
2243 rebuild_sched_domains();
2246 void cpuset_update_active_cpus(bool cpu_online
)
2249 * We're inside cpu hotplug critical region which usually nests
2250 * inside cgroup synchronization. Bounce actual hotplug processing
2251 * to a work item to avoid reverse locking order.
2253 * We still need to do partition_sched_domains() synchronously;
2254 * otherwise, the scheduler will get confused and put tasks to the
2255 * dead CPU. Fall back to the default single domain.
2256 * cpuset_hotplug_workfn() will rebuild it as necessary.
2258 partition_sched_domains(1, NULL
, NULL
);
2259 schedule_work(&cpuset_hotplug_work
);
2263 * Keep top_cpuset.mems_allowed tracking node_states[N_MEMORY].
2264 * Call this routine anytime after node_states[N_MEMORY] changes.
2265 * See cpuset_update_active_cpus() for CPU hotplug handling.
2267 static int cpuset_track_online_nodes(struct notifier_block
*self
,
2268 unsigned long action
, void *arg
)
2270 schedule_work(&cpuset_hotplug_work
);
2274 static struct notifier_block cpuset_track_online_nodes_nb
= {
2275 .notifier_call
= cpuset_track_online_nodes
,
2276 .priority
= 10, /* ??! */
2280 * cpuset_init_smp - initialize cpus_allowed
2282 * Description: Finish top cpuset after cpu, node maps are initialized
2284 void __init
cpuset_init_smp(void)
2286 cpumask_copy(top_cpuset
.cpus_allowed
, cpu_active_mask
);
2287 top_cpuset
.mems_allowed
= node_states
[N_MEMORY
];
2288 top_cpuset
.old_mems_allowed
= top_cpuset
.mems_allowed
;
2290 cpumask_copy(top_cpuset
.effective_cpus
, cpu_active_mask
);
2291 top_cpuset
.effective_mems
= node_states
[N_MEMORY
];
2293 register_hotmemory_notifier(&cpuset_track_online_nodes_nb
);
2297 * cpuset_cpus_allowed - return cpus_allowed mask from a tasks cpuset.
2298 * @tsk: pointer to task_struct from which to obtain cpuset->cpus_allowed.
2299 * @pmask: pointer to struct cpumask variable to receive cpus_allowed set.
2301 * Description: Returns the cpumask_var_t cpus_allowed of the cpuset
2302 * attached to the specified @tsk. Guaranteed to return some non-empty
2303 * subset of cpu_online_mask, even if this means going outside the
2307 void cpuset_cpus_allowed(struct task_struct
*tsk
, struct cpumask
*pmask
)
2309 struct cpuset
*cpus_cs
;
2311 mutex_lock(&callback_mutex
);
2313 cpus_cs
= effective_cpumask_cpuset(task_cs(tsk
));
2314 guarantee_online_cpus(cpus_cs
, pmask
);
2316 mutex_unlock(&callback_mutex
);
2319 void cpuset_cpus_allowed_fallback(struct task_struct
*tsk
)
2321 struct cpuset
*cpus_cs
;
2324 cpus_cs
= effective_cpumask_cpuset(task_cs(tsk
));
2325 do_set_cpus_allowed(tsk
, cpus_cs
->cpus_allowed
);
2329 * We own tsk->cpus_allowed, nobody can change it under us.
2331 * But we used cs && cs->cpus_allowed lockless and thus can
2332 * race with cgroup_attach_task() or update_cpumask() and get
2333 * the wrong tsk->cpus_allowed. However, both cases imply the
2334 * subsequent cpuset_change_cpumask()->set_cpus_allowed_ptr()
2335 * which takes task_rq_lock().
2337 * If we are called after it dropped the lock we must see all
2338 * changes in tsk_cs()->cpus_allowed. Otherwise we can temporary
2339 * set any mask even if it is not right from task_cs() pov,
2340 * the pending set_cpus_allowed_ptr() will fix things.
2342 * select_fallback_rq() will fix things ups and set cpu_possible_mask
2347 void cpuset_init_current_mems_allowed(void)
2349 nodes_setall(current
->mems_allowed
);
2353 * cpuset_mems_allowed - return mems_allowed mask from a tasks cpuset.
2354 * @tsk: pointer to task_struct from which to obtain cpuset->mems_allowed.
2356 * Description: Returns the nodemask_t mems_allowed of the cpuset
2357 * attached to the specified @tsk. Guaranteed to return some non-empty
2358 * subset of node_states[N_MEMORY], even if this means going outside the
2362 nodemask_t
cpuset_mems_allowed(struct task_struct
*tsk
)
2364 struct cpuset
*mems_cs
;
2367 mutex_lock(&callback_mutex
);
2369 mems_cs
= effective_nodemask_cpuset(task_cs(tsk
));
2370 guarantee_online_mems(mems_cs
, &mask
);
2372 mutex_unlock(&callback_mutex
);
2378 * cpuset_nodemask_valid_mems_allowed - check nodemask vs. curremt mems_allowed
2379 * @nodemask: the nodemask to be checked
2381 * Are any of the nodes in the nodemask allowed in current->mems_allowed?
2383 int cpuset_nodemask_valid_mems_allowed(nodemask_t
*nodemask
)
2385 return nodes_intersects(*nodemask
, current
->mems_allowed
);
2389 * nearest_hardwall_ancestor() - Returns the nearest mem_exclusive or
2390 * mem_hardwall ancestor to the specified cpuset. Call holding
2391 * callback_mutex. If no ancestor is mem_exclusive or mem_hardwall
2392 * (an unusual configuration), then returns the root cpuset.
2394 static struct cpuset
*nearest_hardwall_ancestor(struct cpuset
*cs
)
2396 while (!(is_mem_exclusive(cs
) || is_mem_hardwall(cs
)) && parent_cs(cs
))
2402 * cpuset_node_allowed_softwall - Can we allocate on a memory node?
2403 * @node: is this an allowed node?
2404 * @gfp_mask: memory allocation flags
2406 * If we're in interrupt, yes, we can always allocate. If __GFP_THISNODE is
2407 * set, yes, we can always allocate. If node is in our task's mems_allowed,
2408 * yes. If it's not a __GFP_HARDWALL request and this node is in the nearest
2409 * hardwalled cpuset ancestor to this task's cpuset, yes. If the task has been
2410 * OOM killed and has access to memory reserves as specified by the TIF_MEMDIE
2414 * If __GFP_HARDWALL is set, cpuset_node_allowed_softwall() reduces to
2415 * cpuset_node_allowed_hardwall(). Otherwise, cpuset_node_allowed_softwall()
2416 * might sleep, and might allow a node from an enclosing cpuset.
2418 * cpuset_node_allowed_hardwall() only handles the simpler case of hardwall
2419 * cpusets, and never sleeps.
2421 * The __GFP_THISNODE placement logic is really handled elsewhere,
2422 * by forcibly using a zonelist starting at a specified node, and by
2423 * (in get_page_from_freelist()) refusing to consider the zones for
2424 * any node on the zonelist except the first. By the time any such
2425 * calls get to this routine, we should just shut up and say 'yes'.
2427 * GFP_USER allocations are marked with the __GFP_HARDWALL bit,
2428 * and do not allow allocations outside the current tasks cpuset
2429 * unless the task has been OOM killed as is marked TIF_MEMDIE.
2430 * GFP_KERNEL allocations are not so marked, so can escape to the
2431 * nearest enclosing hardwalled ancestor cpuset.
2433 * Scanning up parent cpusets requires callback_mutex. The
2434 * __alloc_pages() routine only calls here with __GFP_HARDWALL bit
2435 * _not_ set if it's a GFP_KERNEL allocation, and all nodes in the
2436 * current tasks mems_allowed came up empty on the first pass over
2437 * the zonelist. So only GFP_KERNEL allocations, if all nodes in the
2438 * cpuset are short of memory, might require taking the callback_mutex
2441 * The first call here from mm/page_alloc:get_page_from_freelist()
2442 * has __GFP_HARDWALL set in gfp_mask, enforcing hardwall cpusets,
2443 * so no allocation on a node outside the cpuset is allowed (unless
2444 * in interrupt, of course).
2446 * The second pass through get_page_from_freelist() doesn't even call
2447 * here for GFP_ATOMIC calls. For those calls, the __alloc_pages()
2448 * variable 'wait' is not set, and the bit ALLOC_CPUSET is not set
2449 * in alloc_flags. That logic and the checks below have the combined
2451 * in_interrupt - any node ok (current task context irrelevant)
2452 * GFP_ATOMIC - any node ok
2453 * TIF_MEMDIE - any node ok
2454 * GFP_KERNEL - any node in enclosing hardwalled cpuset ok
2455 * GFP_USER - only nodes in current tasks mems allowed ok.
2458 * Don't call cpuset_node_allowed_softwall if you can't sleep, unless you
2459 * pass in the __GFP_HARDWALL flag set in gfp_flag, which disables
2460 * the code that might scan up ancestor cpusets and sleep.
2462 int __cpuset_node_allowed_softwall(int node
, gfp_t gfp_mask
)
2464 struct cpuset
*cs
; /* current cpuset ancestors */
2465 int allowed
; /* is allocation in zone z allowed? */
2467 if (in_interrupt() || (gfp_mask
& __GFP_THISNODE
))
2469 might_sleep_if(!(gfp_mask
& __GFP_HARDWALL
));
2470 if (node_isset(node
, current
->mems_allowed
))
2473 * Allow tasks that have access to memory reserves because they have
2474 * been OOM killed to get memory anywhere.
2476 if (unlikely(test_thread_flag(TIF_MEMDIE
)))
2478 if (gfp_mask
& __GFP_HARDWALL
) /* If hardwall request, stop here */
2481 if (current
->flags
& PF_EXITING
) /* Let dying task have memory */
2484 /* Not hardwall and node outside mems_allowed: scan up cpusets */
2485 mutex_lock(&callback_mutex
);
2488 cs
= nearest_hardwall_ancestor(task_cs(current
));
2489 allowed
= node_isset(node
, cs
->mems_allowed
);
2492 mutex_unlock(&callback_mutex
);
2497 * cpuset_node_allowed_hardwall - Can we allocate on a memory node?
2498 * @node: is this an allowed node?
2499 * @gfp_mask: memory allocation flags
2501 * If we're in interrupt, yes, we can always allocate. If __GFP_THISNODE is
2502 * set, yes, we can always allocate. If node is in our task's mems_allowed,
2503 * yes. If the task has been OOM killed and has access to memory reserves as
2504 * specified by the TIF_MEMDIE flag, yes.
2507 * The __GFP_THISNODE placement logic is really handled elsewhere,
2508 * by forcibly using a zonelist starting at a specified node, and by
2509 * (in get_page_from_freelist()) refusing to consider the zones for
2510 * any node on the zonelist except the first. By the time any such
2511 * calls get to this routine, we should just shut up and say 'yes'.
2513 * Unlike the cpuset_node_allowed_softwall() variant, above,
2514 * this variant requires that the node be in the current task's
2515 * mems_allowed or that we're in interrupt. It does not scan up the
2516 * cpuset hierarchy for the nearest enclosing mem_exclusive cpuset.
2519 int __cpuset_node_allowed_hardwall(int node
, gfp_t gfp_mask
)
2521 if (in_interrupt() || (gfp_mask
& __GFP_THISNODE
))
2523 if (node_isset(node
, current
->mems_allowed
))
2526 * Allow tasks that have access to memory reserves because they have
2527 * been OOM killed to get memory anywhere.
2529 if (unlikely(test_thread_flag(TIF_MEMDIE
)))
2535 * cpuset_mem_spread_node() - On which node to begin search for a file page
2536 * cpuset_slab_spread_node() - On which node to begin search for a slab page
2538 * If a task is marked PF_SPREAD_PAGE or PF_SPREAD_SLAB (as for
2539 * tasks in a cpuset with is_spread_page or is_spread_slab set),
2540 * and if the memory allocation used cpuset_mem_spread_node()
2541 * to determine on which node to start looking, as it will for
2542 * certain page cache or slab cache pages such as used for file
2543 * system buffers and inode caches, then instead of starting on the
2544 * local node to look for a free page, rather spread the starting
2545 * node around the tasks mems_allowed nodes.
2547 * We don't have to worry about the returned node being offline
2548 * because "it can't happen", and even if it did, it would be ok.
2550 * The routines calling guarantee_online_mems() are careful to
2551 * only set nodes in task->mems_allowed that are online. So it
2552 * should not be possible for the following code to return an
2553 * offline node. But if it did, that would be ok, as this routine
2554 * is not returning the node where the allocation must be, only
2555 * the node where the search should start. The zonelist passed to
2556 * __alloc_pages() will include all nodes. If the slab allocator
2557 * is passed an offline node, it will fall back to the local node.
2558 * See kmem_cache_alloc_node().
2561 static int cpuset_spread_node(int *rotor
)
2565 node
= next_node(*rotor
, current
->mems_allowed
);
2566 if (node
== MAX_NUMNODES
)
2567 node
= first_node(current
->mems_allowed
);
2572 int cpuset_mem_spread_node(void)
2574 if (current
->cpuset_mem_spread_rotor
== NUMA_NO_NODE
)
2575 current
->cpuset_mem_spread_rotor
=
2576 node_random(¤t
->mems_allowed
);
2578 return cpuset_spread_node(¤t
->cpuset_mem_spread_rotor
);
2581 int cpuset_slab_spread_node(void)
2583 if (current
->cpuset_slab_spread_rotor
== NUMA_NO_NODE
)
2584 current
->cpuset_slab_spread_rotor
=
2585 node_random(¤t
->mems_allowed
);
2587 return cpuset_spread_node(¤t
->cpuset_slab_spread_rotor
);
2590 EXPORT_SYMBOL_GPL(cpuset_mem_spread_node
);
2593 * cpuset_mems_allowed_intersects - Does @tsk1's mems_allowed intersect @tsk2's?
2594 * @tsk1: pointer to task_struct of some task.
2595 * @tsk2: pointer to task_struct of some other task.
2597 * Description: Return true if @tsk1's mems_allowed intersects the
2598 * mems_allowed of @tsk2. Used by the OOM killer to determine if
2599 * one of the task's memory usage might impact the memory available
2603 int cpuset_mems_allowed_intersects(const struct task_struct
*tsk1
,
2604 const struct task_struct
*tsk2
)
2606 return nodes_intersects(tsk1
->mems_allowed
, tsk2
->mems_allowed
);
2609 #define CPUSET_NODELIST_LEN (256)
2612 * cpuset_print_task_mems_allowed - prints task's cpuset and mems_allowed
2613 * @tsk: pointer to task_struct of some task.
2615 * Description: Prints @task's name, cpuset name, and cached copy of its
2616 * mems_allowed to the kernel log.
2618 void cpuset_print_task_mems_allowed(struct task_struct
*tsk
)
2620 /* Statically allocated to prevent using excess stack. */
2621 static char cpuset_nodelist
[CPUSET_NODELIST_LEN
];
2622 static DEFINE_SPINLOCK(cpuset_buffer_lock
);
2623 struct cgroup
*cgrp
;
2625 spin_lock(&cpuset_buffer_lock
);
2628 cgrp
= task_cs(tsk
)->css
.cgroup
;
2629 nodelist_scnprintf(cpuset_nodelist
, CPUSET_NODELIST_LEN
,
2631 pr_info("%s cpuset=", tsk
->comm
);
2632 pr_cont_cgroup_name(cgrp
);
2633 pr_cont(" mems_allowed=%s\n", cpuset_nodelist
);
2636 spin_unlock(&cpuset_buffer_lock
);
2640 * Collection of memory_pressure is suppressed unless
2641 * this flag is enabled by writing "1" to the special
2642 * cpuset file 'memory_pressure_enabled' in the root cpuset.
2645 int cpuset_memory_pressure_enabled __read_mostly
;
2648 * cpuset_memory_pressure_bump - keep stats of per-cpuset reclaims.
2650 * Keep a running average of the rate of synchronous (direct)
2651 * page reclaim efforts initiated by tasks in each cpuset.
2653 * This represents the rate at which some task in the cpuset
2654 * ran low on memory on all nodes it was allowed to use, and
2655 * had to enter the kernels page reclaim code in an effort to
2656 * create more free memory by tossing clean pages or swapping
2657 * or writing dirty pages.
2659 * Display to user space in the per-cpuset read-only file
2660 * "memory_pressure". Value displayed is an integer
2661 * representing the recent rate of entry into the synchronous
2662 * (direct) page reclaim by any task attached to the cpuset.
2665 void __cpuset_memory_pressure_bump(void)
2668 fmeter_markevent(&task_cs(current
)->fmeter
);
2672 #ifdef CONFIG_PROC_PID_CPUSET
2674 * proc_cpuset_show()
2675 * - Print tasks cpuset path into seq_file.
2676 * - Used for /proc/<pid>/cpuset.
2677 * - No need to task_lock(tsk) on this tsk->cpuset reference, as it
2678 * doesn't really matter if tsk->cpuset changes after we read it,
2679 * and we take cpuset_mutex, keeping cpuset_attach() from changing it
2682 int proc_cpuset_show(struct seq_file
*m
, void *unused_v
)
2685 struct task_struct
*tsk
;
2687 struct cgroup_subsys_state
*css
;
2691 buf
= kmalloc(PATH_MAX
, GFP_KERNEL
);
2697 tsk
= get_pid_task(pid
, PIDTYPE_PID
);
2701 retval
= -ENAMETOOLONG
;
2703 css
= task_css(tsk
, cpuset_cgrp_id
);
2704 p
= cgroup_path(css
->cgroup
, buf
, PATH_MAX
);
2712 put_task_struct(tsk
);
2718 #endif /* CONFIG_PROC_PID_CPUSET */
2720 /* Display task mems_allowed in /proc/<pid>/status file. */
2721 void cpuset_task_status_allowed(struct seq_file
*m
, struct task_struct
*task
)
2723 seq_puts(m
, "Mems_allowed:\t");
2724 seq_nodemask(m
, &task
->mems_allowed
);
2726 seq_puts(m
, "Mems_allowed_list:\t");
2727 seq_nodemask_list(m
, &task
->mems_allowed
);