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 */
81 * On default hierarchy:
83 * The user-configured masks can only be changed by writing to
84 * cpuset.cpus and cpuset.mems, and won't be limited by the
87 * The effective masks is the real masks that apply to the tasks
88 * in the cpuset. They may be changed if the configured masks are
89 * changed or hotplug happens.
91 * effective_mask == configured_mask & parent's effective_mask,
92 * and if it ends up empty, it will inherit the parent's mask.
97 * The user-configured masks are always the same with effective masks.
100 /* user-configured CPUs and Memory Nodes allow to tasks */
101 cpumask_var_t cpus_allowed
;
102 nodemask_t mems_allowed
;
104 /* effective CPUs and Memory Nodes allow to tasks */
105 cpumask_var_t effective_cpus
;
106 nodemask_t effective_mems
;
109 * This is old Memory Nodes tasks took on.
111 * - top_cpuset.old_mems_allowed is initialized to mems_allowed.
112 * - A new cpuset's old_mems_allowed is initialized when some
113 * task is moved into it.
114 * - old_mems_allowed is used in cpuset_migrate_mm() when we change
115 * cpuset.mems_allowed and have tasks' nodemask updated, and
116 * then old_mems_allowed is updated to mems_allowed.
118 nodemask_t old_mems_allowed
;
120 struct fmeter fmeter
; /* memory_pressure filter */
123 * Tasks are being attached to this cpuset. Used to prevent
124 * zeroing cpus/mems_allowed between ->can_attach() and ->attach().
126 int attach_in_progress
;
128 /* partition number for rebuild_sched_domains() */
131 /* for custom sched domain */
132 int relax_domain_level
;
135 static inline struct cpuset
*css_cs(struct cgroup_subsys_state
*css
)
137 return css
? container_of(css
, struct cpuset
, css
) : NULL
;
140 /* Retrieve the cpuset for a task */
141 static inline struct cpuset
*task_cs(struct task_struct
*task
)
143 return css_cs(task_css(task
, cpuset_cgrp_id
));
146 static inline struct cpuset
*parent_cs(struct cpuset
*cs
)
148 return css_cs(cs
->css
.parent
);
152 static inline bool task_has_mempolicy(struct task_struct
*task
)
154 return task
->mempolicy
;
157 static inline bool task_has_mempolicy(struct task_struct
*task
)
164 /* bits in struct cpuset flags field */
171 CS_SCHED_LOAD_BALANCE
,
176 /* convenient tests for these bits */
177 static inline bool is_cpuset_online(const struct cpuset
*cs
)
179 return test_bit(CS_ONLINE
, &cs
->flags
);
182 static inline int is_cpu_exclusive(const struct cpuset
*cs
)
184 return test_bit(CS_CPU_EXCLUSIVE
, &cs
->flags
);
187 static inline int is_mem_exclusive(const struct cpuset
*cs
)
189 return test_bit(CS_MEM_EXCLUSIVE
, &cs
->flags
);
192 static inline int is_mem_hardwall(const struct cpuset
*cs
)
194 return test_bit(CS_MEM_HARDWALL
, &cs
->flags
);
197 static inline int is_sched_load_balance(const struct cpuset
*cs
)
199 return test_bit(CS_SCHED_LOAD_BALANCE
, &cs
->flags
);
202 static inline int is_memory_migrate(const struct cpuset
*cs
)
204 return test_bit(CS_MEMORY_MIGRATE
, &cs
->flags
);
207 static inline int is_spread_page(const struct cpuset
*cs
)
209 return test_bit(CS_SPREAD_PAGE
, &cs
->flags
);
212 static inline int is_spread_slab(const struct cpuset
*cs
)
214 return test_bit(CS_SPREAD_SLAB
, &cs
->flags
);
217 static struct cpuset top_cpuset
= {
218 .flags
= ((1 << CS_ONLINE
) | (1 << CS_CPU_EXCLUSIVE
) |
219 (1 << CS_MEM_EXCLUSIVE
)),
223 * cpuset_for_each_child - traverse online children of a cpuset
224 * @child_cs: loop cursor pointing to the current child
225 * @pos_css: used for iteration
226 * @parent_cs: target cpuset to walk children of
228 * Walk @child_cs through the online children of @parent_cs. Must be used
229 * with RCU read locked.
231 #define cpuset_for_each_child(child_cs, pos_css, parent_cs) \
232 css_for_each_child((pos_css), &(parent_cs)->css) \
233 if (is_cpuset_online(((child_cs) = css_cs((pos_css)))))
236 * cpuset_for_each_descendant_pre - pre-order walk of a cpuset's descendants
237 * @des_cs: loop cursor pointing to the current descendant
238 * @pos_css: used for iteration
239 * @root_cs: target cpuset to walk ancestor of
241 * Walk @des_cs through the online descendants of @root_cs. Must be used
242 * with RCU read locked. The caller may modify @pos_css by calling
243 * css_rightmost_descendant() to skip subtree. @root_cs is included in the
244 * iteration and the first node to be visited.
246 #define cpuset_for_each_descendant_pre(des_cs, pos_css, root_cs) \
247 css_for_each_descendant_pre((pos_css), &(root_cs)->css) \
248 if (is_cpuset_online(((des_cs) = css_cs((pos_css)))))
251 * There are two global mutexes guarding cpuset structures - cpuset_mutex
252 * and callback_mutex. The latter may nest inside the former. We also
253 * require taking task_lock() when dereferencing a task's cpuset pointer.
254 * See "The task_lock() exception", at the end of this comment.
256 * A task must hold both mutexes to modify cpusets. If a task holds
257 * cpuset_mutex, then it blocks others wanting that mutex, ensuring that it
258 * is the only task able to also acquire callback_mutex and be able to
259 * modify cpusets. It can perform various checks on the cpuset structure
260 * first, knowing nothing will change. It can also allocate memory while
261 * just holding cpuset_mutex. While it is performing these checks, various
262 * callback routines can briefly acquire callback_mutex to query cpusets.
263 * Once it is ready to make the changes, it takes callback_mutex, blocking
266 * Calls to the kernel memory allocator can not be made while holding
267 * callback_mutex, as that would risk double tripping on callback_mutex
268 * from one of the callbacks into the cpuset code from within
271 * If a task is only holding callback_mutex, then it has read-only
274 * Now, the task_struct fields mems_allowed and mempolicy may be changed
275 * by other task, we use alloc_lock in the task_struct fields to protect
278 * The cpuset_common_file_read() handlers only hold callback_mutex across
279 * small pieces of code, such as when reading out possibly multi-word
280 * cpumasks and nodemasks.
282 * Accessing a task's cpuset should be done in accordance with the
283 * guidelines for accessing subsystem state in kernel/cgroup.c
286 static DEFINE_MUTEX(cpuset_mutex
);
287 static DEFINE_MUTEX(callback_mutex
);
290 * CPU / memory hotplug is handled asynchronously.
292 static void cpuset_hotplug_workfn(struct work_struct
*work
);
293 static DECLARE_WORK(cpuset_hotplug_work
, cpuset_hotplug_workfn
);
295 static DECLARE_WAIT_QUEUE_HEAD(cpuset_attach_wq
);
298 * This is ugly, but preserves the userspace API for existing cpuset
299 * users. If someone tries to mount the "cpuset" filesystem, we
300 * silently switch it to mount "cgroup" instead
302 static struct dentry
*cpuset_mount(struct file_system_type
*fs_type
,
303 int flags
, const char *unused_dev_name
, void *data
)
305 struct file_system_type
*cgroup_fs
= get_fs_type("cgroup");
306 struct dentry
*ret
= ERR_PTR(-ENODEV
);
310 "release_agent=/sbin/cpuset_release_agent";
311 ret
= cgroup_fs
->mount(cgroup_fs
, flags
,
312 unused_dev_name
, mountopts
);
313 put_filesystem(cgroup_fs
);
318 static struct file_system_type cpuset_fs_type
= {
320 .mount
= cpuset_mount
,
324 * Return in pmask the portion of a cpusets's cpus_allowed that
325 * are online. If none are online, walk up the cpuset hierarchy
326 * until we find one that does have some online cpus. The top
327 * cpuset always has some cpus online.
329 * One way or another, we guarantee to return some non-empty subset
330 * of cpu_online_mask.
332 * Call with callback_mutex held.
334 static void guarantee_online_cpus(struct cpuset
*cs
, struct cpumask
*pmask
)
336 while (!cpumask_intersects(cs
->effective_cpus
, cpu_online_mask
))
338 cpumask_and(pmask
, cs
->effective_cpus
, cpu_online_mask
);
342 * Return in *pmask the portion of a cpusets's mems_allowed that
343 * are online, with memory. If none are online with memory, walk
344 * up the cpuset hierarchy until we find one that does have some
345 * online mems. The top cpuset always has some mems online.
347 * One way or another, we guarantee to return some non-empty subset
348 * of node_states[N_MEMORY].
350 * Call with callback_mutex held.
352 static void guarantee_online_mems(struct cpuset
*cs
, nodemask_t
*pmask
)
354 while (!nodes_intersects(cs
->effective_mems
, node_states
[N_MEMORY
]))
356 nodes_and(*pmask
, cs
->effective_mems
, node_states
[N_MEMORY
]);
360 * update task's spread flag if cpuset's page/slab spread flag is set
362 * Called with callback_mutex/cpuset_mutex held
364 static void cpuset_update_task_spread_flag(struct cpuset
*cs
,
365 struct task_struct
*tsk
)
367 if (is_spread_page(cs
))
368 tsk
->flags
|= PF_SPREAD_PAGE
;
370 tsk
->flags
&= ~PF_SPREAD_PAGE
;
371 if (is_spread_slab(cs
))
372 tsk
->flags
|= PF_SPREAD_SLAB
;
374 tsk
->flags
&= ~PF_SPREAD_SLAB
;
378 * is_cpuset_subset(p, q) - Is cpuset p a subset of cpuset q?
380 * One cpuset is a subset of another if all its allowed CPUs and
381 * Memory Nodes are a subset of the other, and its exclusive flags
382 * are only set if the other's are set. Call holding cpuset_mutex.
385 static int is_cpuset_subset(const struct cpuset
*p
, const struct cpuset
*q
)
387 return cpumask_subset(p
->cpus_allowed
, q
->cpus_allowed
) &&
388 nodes_subset(p
->mems_allowed
, q
->mems_allowed
) &&
389 is_cpu_exclusive(p
) <= is_cpu_exclusive(q
) &&
390 is_mem_exclusive(p
) <= is_mem_exclusive(q
);
394 * alloc_trial_cpuset - allocate a trial cpuset
395 * @cs: the cpuset that the trial cpuset duplicates
397 static struct cpuset
*alloc_trial_cpuset(struct cpuset
*cs
)
399 struct cpuset
*trial
;
401 trial
= kmemdup(cs
, sizeof(*cs
), GFP_KERNEL
);
405 if (!alloc_cpumask_var(&trial
->cpus_allowed
, GFP_KERNEL
))
407 if (!alloc_cpumask_var(&trial
->effective_cpus
, GFP_KERNEL
))
410 cpumask_copy(trial
->cpus_allowed
, cs
->cpus_allowed
);
411 cpumask_copy(trial
->effective_cpus
, cs
->effective_cpus
);
415 free_cpumask_var(trial
->cpus_allowed
);
422 * free_trial_cpuset - free the trial cpuset
423 * @trial: the trial cpuset to be freed
425 static void free_trial_cpuset(struct cpuset
*trial
)
427 free_cpumask_var(trial
->effective_cpus
);
428 free_cpumask_var(trial
->cpus_allowed
);
433 * validate_change() - Used to validate that any proposed cpuset change
434 * follows the structural rules for cpusets.
436 * If we replaced the flag and mask values of the current cpuset
437 * (cur) with those values in the trial cpuset (trial), would
438 * our various subset and exclusive rules still be valid? Presumes
441 * 'cur' is the address of an actual, in-use cpuset. Operations
442 * such as list traversal that depend on the actual address of the
443 * cpuset in the list must use cur below, not trial.
445 * 'trial' is the address of bulk structure copy of cur, with
446 * perhaps one or more of the fields cpus_allowed, mems_allowed,
447 * or flags changed to new, trial values.
449 * Return 0 if valid, -errno if not.
452 static int validate_change(struct cpuset
*cur
, struct cpuset
*trial
)
454 struct cgroup_subsys_state
*css
;
455 struct cpuset
*c
, *par
;
460 /* Each of our child cpusets must be a subset of us */
462 cpuset_for_each_child(c
, css
, cur
)
463 if (!is_cpuset_subset(c
, trial
))
466 /* Remaining checks don't apply to root cpuset */
468 if (cur
== &top_cpuset
)
471 par
= parent_cs(cur
);
473 /* On legacy hiearchy, we must be a subset of our parent cpuset. */
475 if (!cgroup_on_dfl(cur
->css
.cgroup
) && !is_cpuset_subset(trial
, par
))
479 * If either I or some sibling (!= me) is exclusive, we can't
483 cpuset_for_each_child(c
, css
, par
) {
484 if ((is_cpu_exclusive(trial
) || is_cpu_exclusive(c
)) &&
486 cpumask_intersects(trial
->cpus_allowed
, c
->cpus_allowed
))
488 if ((is_mem_exclusive(trial
) || is_mem_exclusive(c
)) &&
490 nodes_intersects(trial
->mems_allowed
, c
->mems_allowed
))
495 * Cpusets with tasks - existing or newly being attached - can't
496 * be changed to have empty cpus_allowed or mems_allowed.
499 if ((cgroup_has_tasks(cur
->css
.cgroup
) || cur
->attach_in_progress
)) {
500 if (!cpumask_empty(cur
->cpus_allowed
) &&
501 cpumask_empty(trial
->cpus_allowed
))
503 if (!nodes_empty(cur
->mems_allowed
) &&
504 nodes_empty(trial
->mems_allowed
))
516 * Helper routine for generate_sched_domains().
517 * Do cpusets a, b have overlapping effective cpus_allowed masks?
519 static int cpusets_overlap(struct cpuset
*a
, struct cpuset
*b
)
521 return cpumask_intersects(a
->effective_cpus
, b
->effective_cpus
);
525 update_domain_attr(struct sched_domain_attr
*dattr
, struct cpuset
*c
)
527 if (dattr
->relax_domain_level
< c
->relax_domain_level
)
528 dattr
->relax_domain_level
= c
->relax_domain_level
;
532 static void update_domain_attr_tree(struct sched_domain_attr
*dattr
,
533 struct cpuset
*root_cs
)
536 struct cgroup_subsys_state
*pos_css
;
539 cpuset_for_each_descendant_pre(cp
, pos_css
, root_cs
) {
543 /* skip the whole subtree if @cp doesn't have any CPU */
544 if (cpumask_empty(cp
->cpus_allowed
)) {
545 pos_css
= css_rightmost_descendant(pos_css
);
549 if (is_sched_load_balance(cp
))
550 update_domain_attr(dattr
, cp
);
556 * generate_sched_domains()
558 * This function builds a partial partition of the systems CPUs
559 * A 'partial partition' is a set of non-overlapping subsets whose
560 * union is a subset of that set.
561 * The output of this function needs to be passed to kernel/sched/core.c
562 * partition_sched_domains() routine, which will rebuild the scheduler's
563 * load balancing domains (sched domains) as specified by that partial
566 * See "What is sched_load_balance" in Documentation/cgroups/cpusets.txt
567 * for a background explanation of this.
569 * Does not return errors, on the theory that the callers of this
570 * routine would rather not worry about failures to rebuild sched
571 * domains when operating in the severe memory shortage situations
572 * that could cause allocation failures below.
574 * Must be called with cpuset_mutex held.
576 * The three key local variables below are:
577 * q - a linked-list queue of cpuset pointers, used to implement a
578 * top-down scan of all cpusets. This scan loads a pointer
579 * to each cpuset marked is_sched_load_balance into the
580 * array 'csa'. For our purposes, rebuilding the schedulers
581 * sched domains, we can ignore !is_sched_load_balance cpusets.
582 * csa - (for CpuSet Array) Array of pointers to all the cpusets
583 * that need to be load balanced, for convenient iterative
584 * access by the subsequent code that finds the best partition,
585 * i.e the set of domains (subsets) of CPUs such that the
586 * cpus_allowed of every cpuset marked is_sched_load_balance
587 * is a subset of one of these domains, while there are as
588 * many such domains as possible, each as small as possible.
589 * doms - Conversion of 'csa' to an array of cpumasks, for passing to
590 * the kernel/sched/core.c routine partition_sched_domains() in a
591 * convenient format, that can be easily compared to the prior
592 * value to determine what partition elements (sched domains)
593 * were changed (added or removed.)
595 * Finding the best partition (set of domains):
596 * The triple nested loops below over i, j, k scan over the
597 * load balanced cpusets (using the array of cpuset pointers in
598 * csa[]) looking for pairs of cpusets that have overlapping
599 * cpus_allowed, but which don't have the same 'pn' partition
600 * number and gives them in the same partition number. It keeps
601 * looping on the 'restart' label until it can no longer find
604 * The union of the cpus_allowed masks from the set of
605 * all cpusets having the same 'pn' value then form the one
606 * element of the partition (one sched domain) to be passed to
607 * partition_sched_domains().
609 static int generate_sched_domains(cpumask_var_t
**domains
,
610 struct sched_domain_attr
**attributes
)
612 struct cpuset
*cp
; /* scans q */
613 struct cpuset
**csa
; /* array of all cpuset ptrs */
614 int csn
; /* how many cpuset ptrs in csa so far */
615 int i
, j
, k
; /* indices for partition finding loops */
616 cpumask_var_t
*doms
; /* resulting partition; i.e. sched domains */
617 struct sched_domain_attr
*dattr
; /* attributes for custom domains */
618 int ndoms
= 0; /* number of sched domains in result */
619 int nslot
; /* next empty doms[] struct cpumask slot */
620 struct cgroup_subsys_state
*pos_css
;
626 /* Special case for the 99% of systems with one, full, sched domain */
627 if (is_sched_load_balance(&top_cpuset
)) {
629 doms
= alloc_sched_domains(ndoms
);
633 dattr
= kmalloc(sizeof(struct sched_domain_attr
), GFP_KERNEL
);
635 *dattr
= SD_ATTR_INIT
;
636 update_domain_attr_tree(dattr
, &top_cpuset
);
638 cpumask_copy(doms
[0], top_cpuset
.effective_cpus
);
643 csa
= kmalloc(nr_cpusets() * sizeof(cp
), GFP_KERNEL
);
649 cpuset_for_each_descendant_pre(cp
, pos_css
, &top_cpuset
) {
650 if (cp
== &top_cpuset
)
653 * Continue traversing beyond @cp iff @cp has some CPUs and
654 * isn't load balancing. The former is obvious. The
655 * latter: All child cpusets contain a subset of the
656 * parent's cpus, so just skip them, and then we call
657 * update_domain_attr_tree() to calc relax_domain_level of
658 * the corresponding sched domain.
660 if (!cpumask_empty(cp
->cpus_allowed
) &&
661 !is_sched_load_balance(cp
))
664 if (is_sched_load_balance(cp
))
667 /* skip @cp's subtree */
668 pos_css
= css_rightmost_descendant(pos_css
);
672 for (i
= 0; i
< csn
; i
++)
677 /* Find the best partition (set of sched domains) */
678 for (i
= 0; i
< csn
; i
++) {
679 struct cpuset
*a
= csa
[i
];
682 for (j
= 0; j
< csn
; j
++) {
683 struct cpuset
*b
= csa
[j
];
686 if (apn
!= bpn
&& cpusets_overlap(a
, b
)) {
687 for (k
= 0; k
< csn
; k
++) {
688 struct cpuset
*c
= csa
[k
];
693 ndoms
--; /* one less element */
700 * Now we know how many domains to create.
701 * Convert <csn, csa> to <ndoms, doms> and populate cpu masks.
703 doms
= alloc_sched_domains(ndoms
);
708 * The rest of the code, including the scheduler, can deal with
709 * dattr==NULL case. No need to abort if alloc fails.
711 dattr
= kmalloc(ndoms
* sizeof(struct sched_domain_attr
), GFP_KERNEL
);
713 for (nslot
= 0, i
= 0; i
< csn
; i
++) {
714 struct cpuset
*a
= csa
[i
];
719 /* Skip completed partitions */
725 if (nslot
== ndoms
) {
726 static int warnings
= 10;
728 pr_warn("rebuild_sched_domains confused: nslot %d, ndoms %d, csn %d, i %d, apn %d\n",
729 nslot
, ndoms
, csn
, i
, apn
);
737 *(dattr
+ nslot
) = SD_ATTR_INIT
;
738 for (j
= i
; j
< csn
; j
++) {
739 struct cpuset
*b
= csa
[j
];
742 cpumask_or(dp
, dp
, b
->effective_cpus
);
744 update_domain_attr_tree(dattr
+ nslot
, b
);
746 /* Done with this partition */
752 BUG_ON(nslot
!= ndoms
);
758 * Fallback to the default domain if kmalloc() failed.
759 * See comments in partition_sched_domains().
770 * Rebuild scheduler domains.
772 * If the flag 'sched_load_balance' of any cpuset with non-empty
773 * 'cpus' changes, or if the 'cpus' allowed changes in any cpuset
774 * which has that flag enabled, or if any cpuset with a non-empty
775 * 'cpus' is removed, then call this routine to rebuild the
776 * scheduler's dynamic sched domains.
778 * Call with cpuset_mutex held. Takes get_online_cpus().
780 static void rebuild_sched_domains_locked(void)
782 struct sched_domain_attr
*attr
;
786 lockdep_assert_held(&cpuset_mutex
);
790 * We have raced with CPU hotplug. Don't do anything to avoid
791 * passing doms with offlined cpu to partition_sched_domains().
792 * Anyways, hotplug work item will rebuild sched domains.
794 if (!cpumask_equal(top_cpuset
.effective_cpus
, cpu_active_mask
))
797 /* Generate domain masks and attrs */
798 ndoms
= generate_sched_domains(&doms
, &attr
);
800 /* Have scheduler rebuild the domains */
801 partition_sched_domains(ndoms
, doms
, attr
);
805 #else /* !CONFIG_SMP */
806 static void rebuild_sched_domains_locked(void)
809 #endif /* CONFIG_SMP */
811 void rebuild_sched_domains(void)
813 mutex_lock(&cpuset_mutex
);
814 rebuild_sched_domains_locked();
815 mutex_unlock(&cpuset_mutex
);
819 * update_tasks_cpumask - Update the cpumasks of tasks in the cpuset.
820 * @cs: the cpuset in which each task's cpus_allowed mask needs to be changed
822 * Iterate through each task of @cs updating its cpus_allowed to the
823 * effective cpuset's. As this function is called with cpuset_mutex held,
824 * cpuset membership stays stable.
826 static void update_tasks_cpumask(struct cpuset
*cs
)
828 struct css_task_iter it
;
829 struct task_struct
*task
;
831 css_task_iter_start(&cs
->css
, &it
);
832 while ((task
= css_task_iter_next(&it
)))
833 set_cpus_allowed_ptr(task
, cs
->effective_cpus
);
834 css_task_iter_end(&it
);
838 * update_cpumasks_hier - Update effective cpumasks and tasks in the subtree
839 * @cs: the cpuset to consider
840 * @new_cpus: temp variable for calculating new effective_cpus
842 * When congifured cpumask is changed, the effective cpumasks of this cpuset
843 * and all its descendants need to be updated.
845 * On legacy hierachy, effective_cpus will be the same with cpu_allowed.
847 * Called with cpuset_mutex held
849 static void update_cpumasks_hier(struct cpuset
*cs
, struct cpumask
*new_cpus
)
852 struct cgroup_subsys_state
*pos_css
;
853 bool need_rebuild_sched_domains
= false;
856 cpuset_for_each_descendant_pre(cp
, pos_css
, cs
) {
857 struct cpuset
*parent
= parent_cs(cp
);
859 cpumask_and(new_cpus
, cp
->cpus_allowed
, parent
->effective_cpus
);
862 * If it becomes empty, inherit the effective mask of the
863 * parent, which is guaranteed to have some CPUs.
865 if (cpumask_empty(new_cpus
))
866 cpumask_copy(new_cpus
, parent
->effective_cpus
);
868 /* Skip the whole subtree if the cpumask remains the same. */
869 if (cpumask_equal(new_cpus
, cp
->effective_cpus
)) {
870 pos_css
= css_rightmost_descendant(pos_css
);
874 if (!css_tryget_online(&cp
->css
))
878 mutex_lock(&callback_mutex
);
879 cpumask_copy(cp
->effective_cpus
, new_cpus
);
880 mutex_unlock(&callback_mutex
);
882 WARN_ON(!cgroup_on_dfl(cp
->css
.cgroup
) &&
883 !cpumask_equal(cp
->cpus_allowed
, cp
->effective_cpus
));
885 update_tasks_cpumask(cp
);
888 * If the effective cpumask of any non-empty cpuset is changed,
889 * we need to rebuild sched domains.
891 if (!cpumask_empty(cp
->cpus_allowed
) &&
892 is_sched_load_balance(cp
))
893 need_rebuild_sched_domains
= true;
900 if (need_rebuild_sched_domains
)
901 rebuild_sched_domains_locked();
905 * update_cpumask - update the cpus_allowed mask of a cpuset and all tasks in it
906 * @cs: the cpuset to consider
907 * @trialcs: trial cpuset
908 * @buf: buffer of cpu numbers written to this cpuset
910 static int update_cpumask(struct cpuset
*cs
, struct cpuset
*trialcs
,
915 /* top_cpuset.cpus_allowed tracks cpu_online_mask; it's read-only */
916 if (cs
== &top_cpuset
)
920 * An empty cpus_allowed is ok only if the cpuset has no tasks.
921 * Since cpulist_parse() fails on an empty mask, we special case
922 * that parsing. The validate_change() call ensures that cpusets
923 * with tasks have cpus.
926 cpumask_clear(trialcs
->cpus_allowed
);
928 retval
= cpulist_parse(buf
, trialcs
->cpus_allowed
);
932 if (!cpumask_subset(trialcs
->cpus_allowed
, cpu_active_mask
))
936 /* Nothing to do if the cpus didn't change */
937 if (cpumask_equal(cs
->cpus_allowed
, trialcs
->cpus_allowed
))
940 retval
= validate_change(cs
, trialcs
);
944 mutex_lock(&callback_mutex
);
945 cpumask_copy(cs
->cpus_allowed
, trialcs
->cpus_allowed
);
946 mutex_unlock(&callback_mutex
);
948 /* use trialcs->cpus_allowed as a temp variable */
949 update_cpumasks_hier(cs
, trialcs
->cpus_allowed
);
956 * Migrate memory region from one set of nodes to another.
958 * Temporarilly set tasks mems_allowed to target nodes of migration,
959 * so that the migration code can allocate pages on these nodes.
961 * While the mm_struct we are migrating is typically from some
962 * other task, the task_struct mems_allowed that we are hacking
963 * is for our current task, which must allocate new pages for that
964 * migrating memory region.
967 static void cpuset_migrate_mm(struct mm_struct
*mm
, const nodemask_t
*from
,
968 const nodemask_t
*to
)
970 struct task_struct
*tsk
= current
;
972 tsk
->mems_allowed
= *to
;
974 do_migrate_pages(mm
, from
, to
, MPOL_MF_MOVE_ALL
);
977 guarantee_online_mems(task_cs(tsk
), &tsk
->mems_allowed
);
982 * cpuset_change_task_nodemask - change task's mems_allowed and mempolicy
983 * @tsk: the task to change
984 * @newmems: new nodes that the task will be set
986 * In order to avoid seeing no nodes if the old and new nodes are disjoint,
987 * we structure updates as setting all new allowed nodes, then clearing newly
990 static void cpuset_change_task_nodemask(struct task_struct
*tsk
,
996 * Allow tasks that have access to memory reserves because they have
997 * been OOM killed to get memory anywhere.
999 if (unlikely(test_thread_flag(TIF_MEMDIE
)))
1001 if (current
->flags
& PF_EXITING
) /* Let dying task have memory */
1006 * Determine if a loop is necessary if another thread is doing
1007 * read_mems_allowed_begin(). If at least one node remains unchanged and
1008 * tsk does not have a mempolicy, then an empty nodemask will not be
1009 * possible when mems_allowed is larger than a word.
1011 need_loop
= task_has_mempolicy(tsk
) ||
1012 !nodes_intersects(*newmems
, tsk
->mems_allowed
);
1015 local_irq_disable();
1016 write_seqcount_begin(&tsk
->mems_allowed_seq
);
1019 nodes_or(tsk
->mems_allowed
, tsk
->mems_allowed
, *newmems
);
1020 mpol_rebind_task(tsk
, newmems
, MPOL_REBIND_STEP1
);
1022 mpol_rebind_task(tsk
, newmems
, MPOL_REBIND_STEP2
);
1023 tsk
->mems_allowed
= *newmems
;
1026 write_seqcount_end(&tsk
->mems_allowed_seq
);
1033 static void *cpuset_being_rebound
;
1036 * update_tasks_nodemask - Update the nodemasks of tasks in the cpuset.
1037 * @cs: the cpuset in which each task's mems_allowed mask needs to be changed
1039 * Iterate through each task of @cs updating its mems_allowed to the
1040 * effective cpuset's. As this function is called with cpuset_mutex held,
1041 * cpuset membership stays stable.
1043 static void update_tasks_nodemask(struct cpuset
*cs
)
1045 static nodemask_t newmems
; /* protected by cpuset_mutex */
1046 struct css_task_iter it
;
1047 struct task_struct
*task
;
1049 cpuset_being_rebound
= cs
; /* causes mpol_dup() rebind */
1051 guarantee_online_mems(cs
, &newmems
);
1054 * The mpol_rebind_mm() call takes mmap_sem, which we couldn't
1055 * take while holding tasklist_lock. Forks can happen - the
1056 * mpol_dup() cpuset_being_rebound check will catch such forks,
1057 * and rebind their vma mempolicies too. Because we still hold
1058 * the global cpuset_mutex, we know that no other rebind effort
1059 * will be contending for the global variable cpuset_being_rebound.
1060 * It's ok if we rebind the same mm twice; mpol_rebind_mm()
1061 * is idempotent. Also migrate pages in each mm to new nodes.
1063 css_task_iter_start(&cs
->css
, &it
);
1064 while ((task
= css_task_iter_next(&it
))) {
1065 struct mm_struct
*mm
;
1068 cpuset_change_task_nodemask(task
, &newmems
);
1070 mm
= get_task_mm(task
);
1074 migrate
= is_memory_migrate(cs
);
1076 mpol_rebind_mm(mm
, &cs
->mems_allowed
);
1078 cpuset_migrate_mm(mm
, &cs
->old_mems_allowed
, &newmems
);
1081 css_task_iter_end(&it
);
1084 * All the tasks' nodemasks have been updated, update
1085 * cs->old_mems_allowed.
1087 cs
->old_mems_allowed
= newmems
;
1089 /* We're done rebinding vmas to this cpuset's new mems_allowed. */
1090 cpuset_being_rebound
= NULL
;
1094 * update_nodemasks_hier - Update effective nodemasks and tasks in the subtree
1095 * @cs: the cpuset to consider
1096 * @new_mems: a temp variable for calculating new effective_mems
1098 * When configured nodemask is changed, the effective nodemasks of this cpuset
1099 * and all its descendants need to be updated.
1101 * On legacy hiearchy, effective_mems will be the same with mems_allowed.
1103 * Called with cpuset_mutex held
1105 static void update_nodemasks_hier(struct cpuset
*cs
, nodemask_t
*new_mems
)
1108 struct cgroup_subsys_state
*pos_css
;
1111 cpuset_for_each_descendant_pre(cp
, pos_css
, cs
) {
1112 struct cpuset
*parent
= parent_cs(cp
);
1114 nodes_and(*new_mems
, cp
->mems_allowed
, parent
->effective_mems
);
1117 * If it becomes empty, inherit the effective mask of the
1118 * parent, which is guaranteed to have some MEMs.
1120 if (nodes_empty(*new_mems
))
1121 *new_mems
= parent
->effective_mems
;
1123 /* Skip the whole subtree if the nodemask remains the same. */
1124 if (nodes_equal(*new_mems
, cp
->effective_mems
)) {
1125 pos_css
= css_rightmost_descendant(pos_css
);
1129 if (!css_tryget_online(&cp
->css
))
1133 mutex_lock(&callback_mutex
);
1134 cp
->effective_mems
= *new_mems
;
1135 mutex_unlock(&callback_mutex
);
1137 WARN_ON(!cgroup_on_dfl(cp
->css
.cgroup
) &&
1138 nodes_equal(cp
->mems_allowed
, cp
->effective_mems
));
1140 update_tasks_nodemask(cp
);
1149 * Handle user request to change the 'mems' memory placement
1150 * of a cpuset. Needs to validate the request, update the
1151 * cpusets mems_allowed, and for each task in the cpuset,
1152 * update mems_allowed and rebind task's mempolicy and any vma
1153 * mempolicies and if the cpuset is marked 'memory_migrate',
1154 * migrate the tasks pages to the new memory.
1156 * Call with cpuset_mutex held. May take callback_mutex during call.
1157 * Will take tasklist_lock, scan tasklist for tasks in cpuset cs,
1158 * lock each such tasks mm->mmap_sem, scan its vma's and rebind
1159 * their mempolicies to the cpusets new mems_allowed.
1161 static int update_nodemask(struct cpuset
*cs
, struct cpuset
*trialcs
,
1167 * top_cpuset.mems_allowed tracks node_stats[N_MEMORY];
1170 if (cs
== &top_cpuset
) {
1176 * An empty mems_allowed is ok iff there are no tasks in the cpuset.
1177 * Since nodelist_parse() fails on an empty mask, we special case
1178 * that parsing. The validate_change() call ensures that cpusets
1179 * with tasks have memory.
1182 nodes_clear(trialcs
->mems_allowed
);
1184 retval
= nodelist_parse(buf
, trialcs
->mems_allowed
);
1188 if (!nodes_subset(trialcs
->mems_allowed
,
1189 node_states
[N_MEMORY
])) {
1195 if (nodes_equal(cs
->mems_allowed
, trialcs
->mems_allowed
)) {
1196 retval
= 0; /* Too easy - nothing to do */
1199 retval
= validate_change(cs
, trialcs
);
1203 mutex_lock(&callback_mutex
);
1204 cs
->mems_allowed
= trialcs
->mems_allowed
;
1205 mutex_unlock(&callback_mutex
);
1207 /* use trialcs->mems_allowed as a temp variable */
1208 update_nodemasks_hier(cs
, &cs
->mems_allowed
);
1213 int current_cpuset_is_being_rebound(void)
1215 return task_cs(current
) == cpuset_being_rebound
;
1218 static int update_relax_domain_level(struct cpuset
*cs
, s64 val
)
1221 if (val
< -1 || val
>= sched_domain_level_max
)
1225 if (val
!= cs
->relax_domain_level
) {
1226 cs
->relax_domain_level
= val
;
1227 if (!cpumask_empty(cs
->cpus_allowed
) &&
1228 is_sched_load_balance(cs
))
1229 rebuild_sched_domains_locked();
1236 * update_tasks_flags - update the spread flags of tasks in the cpuset.
1237 * @cs: the cpuset in which each task's spread flags needs to be changed
1239 * Iterate through each task of @cs updating its spread flags. As this
1240 * function is called with cpuset_mutex held, cpuset membership stays
1243 static void update_tasks_flags(struct cpuset
*cs
)
1245 struct css_task_iter it
;
1246 struct task_struct
*task
;
1248 css_task_iter_start(&cs
->css
, &it
);
1249 while ((task
= css_task_iter_next(&it
)))
1250 cpuset_update_task_spread_flag(cs
, task
);
1251 css_task_iter_end(&it
);
1255 * update_flag - read a 0 or a 1 in a file and update associated flag
1256 * bit: the bit to update (see cpuset_flagbits_t)
1257 * cs: the cpuset to update
1258 * turning_on: whether the flag is being set or cleared
1260 * Call with cpuset_mutex held.
1263 static int update_flag(cpuset_flagbits_t bit
, struct cpuset
*cs
,
1266 struct cpuset
*trialcs
;
1267 int balance_flag_changed
;
1268 int spread_flag_changed
;
1271 trialcs
= alloc_trial_cpuset(cs
);
1276 set_bit(bit
, &trialcs
->flags
);
1278 clear_bit(bit
, &trialcs
->flags
);
1280 err
= validate_change(cs
, trialcs
);
1284 balance_flag_changed
= (is_sched_load_balance(cs
) !=
1285 is_sched_load_balance(trialcs
));
1287 spread_flag_changed
= ((is_spread_slab(cs
) != is_spread_slab(trialcs
))
1288 || (is_spread_page(cs
) != is_spread_page(trialcs
)));
1290 mutex_lock(&callback_mutex
);
1291 cs
->flags
= trialcs
->flags
;
1292 mutex_unlock(&callback_mutex
);
1294 if (!cpumask_empty(trialcs
->cpus_allowed
) && balance_flag_changed
)
1295 rebuild_sched_domains_locked();
1297 if (spread_flag_changed
)
1298 update_tasks_flags(cs
);
1300 free_trial_cpuset(trialcs
);
1305 * Frequency meter - How fast is some event occurring?
1307 * These routines manage a digitally filtered, constant time based,
1308 * event frequency meter. There are four routines:
1309 * fmeter_init() - initialize a frequency meter.
1310 * fmeter_markevent() - called each time the event happens.
1311 * fmeter_getrate() - returns the recent rate of such events.
1312 * fmeter_update() - internal routine used to update fmeter.
1314 * A common data structure is passed to each of these routines,
1315 * which is used to keep track of the state required to manage the
1316 * frequency meter and its digital filter.
1318 * The filter works on the number of events marked per unit time.
1319 * The filter is single-pole low-pass recursive (IIR). The time unit
1320 * is 1 second. Arithmetic is done using 32-bit integers scaled to
1321 * simulate 3 decimal digits of precision (multiplied by 1000).
1323 * With an FM_COEF of 933, and a time base of 1 second, the filter
1324 * has a half-life of 10 seconds, meaning that if the events quit
1325 * happening, then the rate returned from the fmeter_getrate()
1326 * will be cut in half each 10 seconds, until it converges to zero.
1328 * It is not worth doing a real infinitely recursive filter. If more
1329 * than FM_MAXTICKS ticks have elapsed since the last filter event,
1330 * just compute FM_MAXTICKS ticks worth, by which point the level
1333 * Limit the count of unprocessed events to FM_MAXCNT, so as to avoid
1334 * arithmetic overflow in the fmeter_update() routine.
1336 * Given the simple 32 bit integer arithmetic used, this meter works
1337 * best for reporting rates between one per millisecond (msec) and
1338 * one per 32 (approx) seconds. At constant rates faster than one
1339 * per msec it maxes out at values just under 1,000,000. At constant
1340 * rates between one per msec, and one per second it will stabilize
1341 * to a value N*1000, where N is the rate of events per second.
1342 * At constant rates between one per second and one per 32 seconds,
1343 * it will be choppy, moving up on the seconds that have an event,
1344 * and then decaying until the next event. At rates slower than
1345 * about one in 32 seconds, it decays all the way back to zero between
1349 #define FM_COEF 933 /* coefficient for half-life of 10 secs */
1350 #define FM_MAXTICKS ((time_t)99) /* useless computing more ticks than this */
1351 #define FM_MAXCNT 1000000 /* limit cnt to avoid overflow */
1352 #define FM_SCALE 1000 /* faux fixed point scale */
1354 /* Initialize a frequency meter */
1355 static void fmeter_init(struct fmeter
*fmp
)
1360 spin_lock_init(&fmp
->lock
);
1363 /* Internal meter update - process cnt events and update value */
1364 static void fmeter_update(struct fmeter
*fmp
)
1366 time_t now
= get_seconds();
1367 time_t ticks
= now
- fmp
->time
;
1372 ticks
= min(FM_MAXTICKS
, ticks
);
1374 fmp
->val
= (FM_COEF
* fmp
->val
) / FM_SCALE
;
1377 fmp
->val
+= ((FM_SCALE
- FM_COEF
) * fmp
->cnt
) / FM_SCALE
;
1381 /* Process any previous ticks, then bump cnt by one (times scale). */
1382 static void fmeter_markevent(struct fmeter
*fmp
)
1384 spin_lock(&fmp
->lock
);
1386 fmp
->cnt
= min(FM_MAXCNT
, fmp
->cnt
+ FM_SCALE
);
1387 spin_unlock(&fmp
->lock
);
1390 /* Process any previous ticks, then return current value. */
1391 static int fmeter_getrate(struct fmeter
*fmp
)
1395 spin_lock(&fmp
->lock
);
1398 spin_unlock(&fmp
->lock
);
1402 static struct cpuset
*cpuset_attach_old_cs
;
1404 /* Called by cgroups to determine if a cpuset is usable; cpuset_mutex held */
1405 static int cpuset_can_attach(struct cgroup_subsys_state
*css
,
1406 struct cgroup_taskset
*tset
)
1408 struct cpuset
*cs
= css_cs(css
);
1409 struct task_struct
*task
;
1412 /* used later by cpuset_attach() */
1413 cpuset_attach_old_cs
= task_cs(cgroup_taskset_first(tset
));
1415 mutex_lock(&cpuset_mutex
);
1417 /* allow moving tasks into an empty cpuset if on default hierarchy */
1419 if (!cgroup_on_dfl(css
->cgroup
) &&
1420 (cpumask_empty(cs
->cpus_allowed
) || nodes_empty(cs
->mems_allowed
)))
1423 cgroup_taskset_for_each(task
, tset
) {
1425 * Kthreads which disallow setaffinity shouldn't be moved
1426 * to a new cpuset; we don't want to change their cpu
1427 * affinity and isolating such threads by their set of
1428 * allowed nodes is unnecessary. Thus, cpusets are not
1429 * applicable for such threads. This prevents checking for
1430 * success of set_cpus_allowed_ptr() on all attached tasks
1431 * before cpus_allowed may be changed.
1434 if (task
->flags
& PF_NO_SETAFFINITY
)
1436 ret
= security_task_setscheduler(task
);
1442 * Mark attach is in progress. This makes validate_change() fail
1443 * changes which zero cpus/mems_allowed.
1445 cs
->attach_in_progress
++;
1448 mutex_unlock(&cpuset_mutex
);
1452 static void cpuset_cancel_attach(struct cgroup_subsys_state
*css
,
1453 struct cgroup_taskset
*tset
)
1455 mutex_lock(&cpuset_mutex
);
1456 css_cs(css
)->attach_in_progress
--;
1457 mutex_unlock(&cpuset_mutex
);
1461 * Protected by cpuset_mutex. cpus_attach is used only by cpuset_attach()
1462 * but we can't allocate it dynamically there. Define it global and
1463 * allocate from cpuset_init().
1465 static cpumask_var_t cpus_attach
;
1467 static void cpuset_attach(struct cgroup_subsys_state
*css
,
1468 struct cgroup_taskset
*tset
)
1470 /* static buf protected by cpuset_mutex */
1471 static nodemask_t cpuset_attach_nodemask_to
;
1472 struct mm_struct
*mm
;
1473 struct task_struct
*task
;
1474 struct task_struct
*leader
= cgroup_taskset_first(tset
);
1475 struct cpuset
*cs
= css_cs(css
);
1476 struct cpuset
*oldcs
= cpuset_attach_old_cs
;
1478 mutex_lock(&cpuset_mutex
);
1480 /* prepare for attach */
1481 if (cs
== &top_cpuset
)
1482 cpumask_copy(cpus_attach
, cpu_possible_mask
);
1484 guarantee_online_cpus(cs
, cpus_attach
);
1486 guarantee_online_mems(cs
, &cpuset_attach_nodemask_to
);
1488 cgroup_taskset_for_each(task
, tset
) {
1490 * can_attach beforehand should guarantee that this doesn't
1491 * fail. TODO: have a better way to handle failure here
1493 WARN_ON_ONCE(set_cpus_allowed_ptr(task
, cpus_attach
));
1495 cpuset_change_task_nodemask(task
, &cpuset_attach_nodemask_to
);
1496 cpuset_update_task_spread_flag(cs
, task
);
1500 * Change mm, possibly for multiple threads in a threadgroup. This is
1501 * expensive and may sleep.
1503 cpuset_attach_nodemask_to
= cs
->effective_mems
;
1504 mm
= get_task_mm(leader
);
1506 mpol_rebind_mm(mm
, &cpuset_attach_nodemask_to
);
1509 * old_mems_allowed is the same with mems_allowed here, except
1510 * if this task is being moved automatically due to hotplug.
1511 * In that case @mems_allowed has been updated and is empty,
1512 * so @old_mems_allowed is the right nodesets that we migrate
1515 if (is_memory_migrate(cs
)) {
1516 cpuset_migrate_mm(mm
, &oldcs
->old_mems_allowed
,
1517 &cpuset_attach_nodemask_to
);
1522 cs
->old_mems_allowed
= cpuset_attach_nodemask_to
;
1524 cs
->attach_in_progress
--;
1525 if (!cs
->attach_in_progress
)
1526 wake_up(&cpuset_attach_wq
);
1528 mutex_unlock(&cpuset_mutex
);
1531 /* The various types of files and directories in a cpuset file system */
1534 FILE_MEMORY_MIGRATE
,
1540 FILE_SCHED_LOAD_BALANCE
,
1541 FILE_SCHED_RELAX_DOMAIN_LEVEL
,
1542 FILE_MEMORY_PRESSURE_ENABLED
,
1543 FILE_MEMORY_PRESSURE
,
1546 } cpuset_filetype_t
;
1548 static int cpuset_write_u64(struct cgroup_subsys_state
*css
, struct cftype
*cft
,
1551 struct cpuset
*cs
= css_cs(css
);
1552 cpuset_filetype_t type
= cft
->private;
1555 mutex_lock(&cpuset_mutex
);
1556 if (!is_cpuset_online(cs
)) {
1562 case FILE_CPU_EXCLUSIVE
:
1563 retval
= update_flag(CS_CPU_EXCLUSIVE
, cs
, val
);
1565 case FILE_MEM_EXCLUSIVE
:
1566 retval
= update_flag(CS_MEM_EXCLUSIVE
, cs
, val
);
1568 case FILE_MEM_HARDWALL
:
1569 retval
= update_flag(CS_MEM_HARDWALL
, cs
, val
);
1571 case FILE_SCHED_LOAD_BALANCE
:
1572 retval
= update_flag(CS_SCHED_LOAD_BALANCE
, cs
, val
);
1574 case FILE_MEMORY_MIGRATE
:
1575 retval
= update_flag(CS_MEMORY_MIGRATE
, cs
, val
);
1577 case FILE_MEMORY_PRESSURE_ENABLED
:
1578 cpuset_memory_pressure_enabled
= !!val
;
1580 case FILE_MEMORY_PRESSURE
:
1583 case FILE_SPREAD_PAGE
:
1584 retval
= update_flag(CS_SPREAD_PAGE
, cs
, val
);
1586 case FILE_SPREAD_SLAB
:
1587 retval
= update_flag(CS_SPREAD_SLAB
, cs
, val
);
1594 mutex_unlock(&cpuset_mutex
);
1598 static int cpuset_write_s64(struct cgroup_subsys_state
*css
, struct cftype
*cft
,
1601 struct cpuset
*cs
= css_cs(css
);
1602 cpuset_filetype_t type
= cft
->private;
1603 int retval
= -ENODEV
;
1605 mutex_lock(&cpuset_mutex
);
1606 if (!is_cpuset_online(cs
))
1610 case FILE_SCHED_RELAX_DOMAIN_LEVEL
:
1611 retval
= update_relax_domain_level(cs
, val
);
1618 mutex_unlock(&cpuset_mutex
);
1623 * Common handling for a write to a "cpus" or "mems" file.
1625 static ssize_t
cpuset_write_resmask(struct kernfs_open_file
*of
,
1626 char *buf
, size_t nbytes
, loff_t off
)
1628 struct cpuset
*cs
= css_cs(of_css(of
));
1629 struct cpuset
*trialcs
;
1630 int retval
= -ENODEV
;
1632 buf
= strstrip(buf
);
1635 * CPU or memory hotunplug may leave @cs w/o any execution
1636 * resources, in which case the hotplug code asynchronously updates
1637 * configuration and transfers all tasks to the nearest ancestor
1638 * which can execute.
1640 * As writes to "cpus" or "mems" may restore @cs's execution
1641 * resources, wait for the previously scheduled operations before
1642 * proceeding, so that we don't end up keep removing tasks added
1643 * after execution capability is restored.
1645 flush_work(&cpuset_hotplug_work
);
1647 mutex_lock(&cpuset_mutex
);
1648 if (!is_cpuset_online(cs
))
1651 trialcs
= alloc_trial_cpuset(cs
);
1657 switch (of_cft(of
)->private) {
1659 retval
= update_cpumask(cs
, trialcs
, buf
);
1662 retval
= update_nodemask(cs
, trialcs
, buf
);
1669 free_trial_cpuset(trialcs
);
1671 mutex_unlock(&cpuset_mutex
);
1672 return retval
?: nbytes
;
1676 * These ascii lists should be read in a single call, by using a user
1677 * buffer large enough to hold the entire map. If read in smaller
1678 * chunks, there is no guarantee of atomicity. Since the display format
1679 * used, list of ranges of sequential numbers, is variable length,
1680 * and since these maps can change value dynamically, one could read
1681 * gibberish by doing partial reads while a list was changing.
1683 static int cpuset_common_seq_show(struct seq_file
*sf
, void *v
)
1685 struct cpuset
*cs
= css_cs(seq_css(sf
));
1686 cpuset_filetype_t type
= seq_cft(sf
)->private;
1691 count
= seq_get_buf(sf
, &buf
);
1694 mutex_lock(&callback_mutex
);
1698 s
+= cpulist_scnprintf(s
, count
, cs
->cpus_allowed
);
1701 s
+= nodelist_scnprintf(s
, count
, cs
->mems_allowed
);
1708 if (s
< buf
+ count
- 1) {
1710 seq_commit(sf
, s
- buf
);
1715 mutex_unlock(&callback_mutex
);
1719 static u64
cpuset_read_u64(struct cgroup_subsys_state
*css
, struct cftype
*cft
)
1721 struct cpuset
*cs
= css_cs(css
);
1722 cpuset_filetype_t type
= cft
->private;
1724 case FILE_CPU_EXCLUSIVE
:
1725 return is_cpu_exclusive(cs
);
1726 case FILE_MEM_EXCLUSIVE
:
1727 return is_mem_exclusive(cs
);
1728 case FILE_MEM_HARDWALL
:
1729 return is_mem_hardwall(cs
);
1730 case FILE_SCHED_LOAD_BALANCE
:
1731 return is_sched_load_balance(cs
);
1732 case FILE_MEMORY_MIGRATE
:
1733 return is_memory_migrate(cs
);
1734 case FILE_MEMORY_PRESSURE_ENABLED
:
1735 return cpuset_memory_pressure_enabled
;
1736 case FILE_MEMORY_PRESSURE
:
1737 return fmeter_getrate(&cs
->fmeter
);
1738 case FILE_SPREAD_PAGE
:
1739 return is_spread_page(cs
);
1740 case FILE_SPREAD_SLAB
:
1741 return is_spread_slab(cs
);
1746 /* Unreachable but makes gcc happy */
1750 static s64
cpuset_read_s64(struct cgroup_subsys_state
*css
, struct cftype
*cft
)
1752 struct cpuset
*cs
= css_cs(css
);
1753 cpuset_filetype_t type
= cft
->private;
1755 case FILE_SCHED_RELAX_DOMAIN_LEVEL
:
1756 return cs
->relax_domain_level
;
1761 /* Unrechable but makes gcc happy */
1767 * for the common functions, 'private' gives the type of file
1770 static struct cftype files
[] = {
1773 .seq_show
= cpuset_common_seq_show
,
1774 .write
= cpuset_write_resmask
,
1775 .max_write_len
= (100U + 6 * NR_CPUS
),
1776 .private = FILE_CPULIST
,
1781 .seq_show
= cpuset_common_seq_show
,
1782 .write
= cpuset_write_resmask
,
1783 .max_write_len
= (100U + 6 * MAX_NUMNODES
),
1784 .private = FILE_MEMLIST
,
1788 .name
= "cpu_exclusive",
1789 .read_u64
= cpuset_read_u64
,
1790 .write_u64
= cpuset_write_u64
,
1791 .private = FILE_CPU_EXCLUSIVE
,
1795 .name
= "mem_exclusive",
1796 .read_u64
= cpuset_read_u64
,
1797 .write_u64
= cpuset_write_u64
,
1798 .private = FILE_MEM_EXCLUSIVE
,
1802 .name
= "mem_hardwall",
1803 .read_u64
= cpuset_read_u64
,
1804 .write_u64
= cpuset_write_u64
,
1805 .private = FILE_MEM_HARDWALL
,
1809 .name
= "sched_load_balance",
1810 .read_u64
= cpuset_read_u64
,
1811 .write_u64
= cpuset_write_u64
,
1812 .private = FILE_SCHED_LOAD_BALANCE
,
1816 .name
= "sched_relax_domain_level",
1817 .read_s64
= cpuset_read_s64
,
1818 .write_s64
= cpuset_write_s64
,
1819 .private = FILE_SCHED_RELAX_DOMAIN_LEVEL
,
1823 .name
= "memory_migrate",
1824 .read_u64
= cpuset_read_u64
,
1825 .write_u64
= cpuset_write_u64
,
1826 .private = FILE_MEMORY_MIGRATE
,
1830 .name
= "memory_pressure",
1831 .read_u64
= cpuset_read_u64
,
1832 .write_u64
= cpuset_write_u64
,
1833 .private = FILE_MEMORY_PRESSURE
,
1838 .name
= "memory_spread_page",
1839 .read_u64
= cpuset_read_u64
,
1840 .write_u64
= cpuset_write_u64
,
1841 .private = FILE_SPREAD_PAGE
,
1845 .name
= "memory_spread_slab",
1846 .read_u64
= cpuset_read_u64
,
1847 .write_u64
= cpuset_write_u64
,
1848 .private = FILE_SPREAD_SLAB
,
1852 .name
= "memory_pressure_enabled",
1853 .flags
= CFTYPE_ONLY_ON_ROOT
,
1854 .read_u64
= cpuset_read_u64
,
1855 .write_u64
= cpuset_write_u64
,
1856 .private = FILE_MEMORY_PRESSURE_ENABLED
,
1863 * cpuset_css_alloc - allocate a cpuset css
1864 * cgrp: control group that the new cpuset will be part of
1867 static struct cgroup_subsys_state
*
1868 cpuset_css_alloc(struct cgroup_subsys_state
*parent_css
)
1873 return &top_cpuset
.css
;
1875 cs
= kzalloc(sizeof(*cs
), GFP_KERNEL
);
1877 return ERR_PTR(-ENOMEM
);
1878 if (!alloc_cpumask_var(&cs
->cpus_allowed
, GFP_KERNEL
))
1880 if (!alloc_cpumask_var(&cs
->effective_cpus
, GFP_KERNEL
))
1883 set_bit(CS_SCHED_LOAD_BALANCE
, &cs
->flags
);
1884 cpumask_clear(cs
->cpus_allowed
);
1885 nodes_clear(cs
->mems_allowed
);
1886 cpumask_clear(cs
->effective_cpus
);
1887 nodes_clear(cs
->effective_mems
);
1888 fmeter_init(&cs
->fmeter
);
1889 cs
->relax_domain_level
= -1;
1894 free_cpumask_var(cs
->cpus_allowed
);
1897 return ERR_PTR(-ENOMEM
);
1900 static int cpuset_css_online(struct cgroup_subsys_state
*css
)
1902 struct cpuset
*cs
= css_cs(css
);
1903 struct cpuset
*parent
= parent_cs(cs
);
1904 struct cpuset
*tmp_cs
;
1905 struct cgroup_subsys_state
*pos_css
;
1910 mutex_lock(&cpuset_mutex
);
1912 set_bit(CS_ONLINE
, &cs
->flags
);
1913 if (is_spread_page(parent
))
1914 set_bit(CS_SPREAD_PAGE
, &cs
->flags
);
1915 if (is_spread_slab(parent
))
1916 set_bit(CS_SPREAD_SLAB
, &cs
->flags
);
1920 mutex_lock(&callback_mutex
);
1921 if (cgroup_on_dfl(cs
->css
.cgroup
)) {
1922 cpumask_copy(cs
->effective_cpus
, parent
->effective_cpus
);
1923 cs
->effective_mems
= parent
->effective_mems
;
1925 mutex_unlock(&callback_mutex
);
1927 if (!test_bit(CGRP_CPUSET_CLONE_CHILDREN
, &css
->cgroup
->flags
))
1931 * Clone @parent's configuration if CGRP_CPUSET_CLONE_CHILDREN is
1932 * set. This flag handling is implemented in cgroup core for
1933 * histrical reasons - the flag may be specified during mount.
1935 * Currently, if any sibling cpusets have exclusive cpus or mem, we
1936 * refuse to clone the configuration - thereby refusing the task to
1937 * be entered, and as a result refusing the sys_unshare() or
1938 * clone() which initiated it. If this becomes a problem for some
1939 * users who wish to allow that scenario, then this could be
1940 * changed to grant parent->cpus_allowed-sibling_cpus_exclusive
1941 * (and likewise for mems) to the new cgroup.
1944 cpuset_for_each_child(tmp_cs
, pos_css
, parent
) {
1945 if (is_mem_exclusive(tmp_cs
) || is_cpu_exclusive(tmp_cs
)) {
1952 mutex_lock(&callback_mutex
);
1953 cs
->mems_allowed
= parent
->mems_allowed
;
1954 cpumask_copy(cs
->cpus_allowed
, parent
->cpus_allowed
);
1955 mutex_unlock(&callback_mutex
);
1957 mutex_unlock(&cpuset_mutex
);
1962 * If the cpuset being removed has its flag 'sched_load_balance'
1963 * enabled, then simulate turning sched_load_balance off, which
1964 * will call rebuild_sched_domains_locked().
1967 static void cpuset_css_offline(struct cgroup_subsys_state
*css
)
1969 struct cpuset
*cs
= css_cs(css
);
1971 mutex_lock(&cpuset_mutex
);
1973 if (is_sched_load_balance(cs
))
1974 update_flag(CS_SCHED_LOAD_BALANCE
, cs
, 0);
1977 clear_bit(CS_ONLINE
, &cs
->flags
);
1979 mutex_unlock(&cpuset_mutex
);
1982 static void cpuset_css_free(struct cgroup_subsys_state
*css
)
1984 struct cpuset
*cs
= css_cs(css
);
1986 free_cpumask_var(cs
->effective_cpus
);
1987 free_cpumask_var(cs
->cpus_allowed
);
1991 static void cpuset_bind(struct cgroup_subsys_state
*root_css
)
1993 mutex_lock(&cpuset_mutex
);
1994 mutex_lock(&callback_mutex
);
1996 if (cgroup_on_dfl(root_css
->cgroup
)) {
1997 cpumask_copy(top_cpuset
.cpus_allowed
, cpu_possible_mask
);
1998 top_cpuset
.mems_allowed
= node_possible_map
;
2000 cpumask_copy(top_cpuset
.cpus_allowed
,
2001 top_cpuset
.effective_cpus
);
2002 top_cpuset
.mems_allowed
= top_cpuset
.effective_mems
;
2005 mutex_unlock(&callback_mutex
);
2006 mutex_unlock(&cpuset_mutex
);
2009 struct cgroup_subsys cpuset_cgrp_subsys
= {
2010 .css_alloc
= cpuset_css_alloc
,
2011 .css_online
= cpuset_css_online
,
2012 .css_offline
= cpuset_css_offline
,
2013 .css_free
= cpuset_css_free
,
2014 .can_attach
= cpuset_can_attach
,
2015 .cancel_attach
= cpuset_cancel_attach
,
2016 .attach
= cpuset_attach
,
2017 .bind
= cpuset_bind
,
2018 .base_cftypes
= files
,
2023 * cpuset_init - initialize cpusets at system boot
2025 * Description: Initialize top_cpuset and the cpuset internal file system,
2028 int __init
cpuset_init(void)
2032 if (!alloc_cpumask_var(&top_cpuset
.cpus_allowed
, GFP_KERNEL
))
2034 if (!alloc_cpumask_var(&top_cpuset
.effective_cpus
, GFP_KERNEL
))
2037 cpumask_setall(top_cpuset
.cpus_allowed
);
2038 nodes_setall(top_cpuset
.mems_allowed
);
2039 cpumask_setall(top_cpuset
.effective_cpus
);
2040 nodes_setall(top_cpuset
.effective_mems
);
2042 fmeter_init(&top_cpuset
.fmeter
);
2043 set_bit(CS_SCHED_LOAD_BALANCE
, &top_cpuset
.flags
);
2044 top_cpuset
.relax_domain_level
= -1;
2046 err
= register_filesystem(&cpuset_fs_type
);
2050 if (!alloc_cpumask_var(&cpus_attach
, GFP_KERNEL
))
2057 * If CPU and/or memory hotplug handlers, below, unplug any CPUs
2058 * or memory nodes, we need to walk over the cpuset hierarchy,
2059 * removing that CPU or node from all cpusets. If this removes the
2060 * last CPU or node from a cpuset, then move the tasks in the empty
2061 * cpuset to its next-highest non-empty parent.
2063 static void remove_tasks_in_empty_cpuset(struct cpuset
*cs
)
2065 struct cpuset
*parent
;
2068 * Find its next-highest non-empty parent, (top cpuset
2069 * has online cpus, so can't be empty).
2071 parent
= parent_cs(cs
);
2072 while (cpumask_empty(parent
->cpus_allowed
) ||
2073 nodes_empty(parent
->mems_allowed
))
2074 parent
= parent_cs(parent
);
2076 if (cgroup_transfer_tasks(parent
->css
.cgroup
, cs
->css
.cgroup
)) {
2077 pr_err("cpuset: failed to transfer tasks out of empty cpuset ");
2078 pr_cont_cgroup_name(cs
->css
.cgroup
);
2084 * cpuset_hotplug_update_tasks - update tasks in a cpuset for hotunplug
2085 * @cs: cpuset in interest
2087 * Compare @cs's cpu and mem masks against top_cpuset and if some have gone
2088 * offline, update @cs accordingly. If @cs ends up with no CPU or memory,
2089 * all its tasks are moved to the nearest ancestor with both resources.
2091 static void cpuset_hotplug_update_tasks(struct cpuset
*cs
)
2093 static cpumask_t off_cpus
;
2094 static nodemask_t off_mems
;
2096 bool on_dfl
= cgroup_on_dfl(cs
->css
.cgroup
);
2099 wait_event(cpuset_attach_wq
, cs
->attach_in_progress
== 0);
2101 mutex_lock(&cpuset_mutex
);
2104 * We have raced with task attaching. We wait until attaching
2105 * is finished, so we won't attach a task to an empty cpuset.
2107 if (cs
->attach_in_progress
) {
2108 mutex_unlock(&cpuset_mutex
);
2112 cpumask_andnot(&off_cpus
, cs
->cpus_allowed
, top_cpuset
.cpus_allowed
);
2113 nodes_andnot(off_mems
, cs
->mems_allowed
, top_cpuset
.mems_allowed
);
2115 mutex_lock(&callback_mutex
);
2116 cpumask_andnot(cs
->cpus_allowed
, cs
->cpus_allowed
, &off_cpus
);
2118 /* Inherit the effective mask of the parent, if it becomes empty. */
2119 cpumask_andnot(cs
->effective_cpus
, cs
->effective_cpus
, &off_cpus
);
2120 if (on_dfl
&& cpumask_empty(cs
->effective_cpus
))
2121 cpumask_copy(cs
->effective_cpus
, parent_cs(cs
)->effective_cpus
);
2122 mutex_unlock(&callback_mutex
);
2125 * If on_dfl, we need to update tasks' cpumask for empty cpuset to
2126 * take on ancestor's cpumask. Otherwise, don't call
2127 * update_tasks_cpumask() if the cpuset becomes empty, as the tasks
2128 * in it will be migrated to an ancestor.
2130 if ((on_dfl
&& cpumask_empty(cs
->cpus_allowed
)) ||
2131 (!cpumask_empty(&off_cpus
) && !cpumask_empty(cs
->cpus_allowed
)))
2132 update_tasks_cpumask(cs
);
2134 mutex_lock(&callback_mutex
);
2135 nodes_andnot(cs
->mems_allowed
, cs
->mems_allowed
, off_mems
);
2137 /* Inherit the effective mask of the parent, if it becomes empty */
2138 nodes_andnot(cs
->effective_mems
, cs
->effective_mems
, off_mems
);
2139 if (on_dfl
&& nodes_empty(cs
->effective_mems
))
2140 cs
->effective_mems
= parent_cs(cs
)->effective_mems
;
2141 mutex_unlock(&callback_mutex
);
2144 * If on_dfl, we need to update tasks' nodemask for empty cpuset to
2145 * take on ancestor's nodemask. Otherwise, don't call
2146 * update_tasks_nodemask() if the cpuset becomes empty, as the
2147 * tasks in it will be migratd to an ancestor.
2149 if ((on_dfl
&& nodes_empty(cs
->mems_allowed
)) ||
2150 (!nodes_empty(off_mems
) && !nodes_empty(cs
->mems_allowed
)))
2151 update_tasks_nodemask(cs
);
2153 is_empty
= cpumask_empty(cs
->cpus_allowed
) ||
2154 nodes_empty(cs
->mems_allowed
);
2156 mutex_unlock(&cpuset_mutex
);
2159 * If on_dfl, we'll keep tasks in empty cpusets.
2161 * Otherwise move tasks to the nearest ancestor with execution
2162 * resources. This is full cgroup operation which will
2163 * also call back into cpuset. Should be done outside any lock.
2165 if (!on_dfl
&& is_empty
)
2166 remove_tasks_in_empty_cpuset(cs
);
2170 * cpuset_hotplug_workfn - handle CPU/memory hotunplug for a cpuset
2172 * This function is called after either CPU or memory configuration has
2173 * changed and updates cpuset accordingly. The top_cpuset is always
2174 * synchronized to cpu_active_mask and N_MEMORY, which is necessary in
2175 * order to make cpusets transparent (of no affect) on systems that are
2176 * actively using CPU hotplug but making no active use of cpusets.
2178 * Non-root cpusets are only affected by offlining. If any CPUs or memory
2179 * nodes have been taken down, cpuset_hotplug_update_tasks() is invoked on
2182 * Note that CPU offlining during suspend is ignored. We don't modify
2183 * cpusets across suspend/resume cycles at all.
2185 static void cpuset_hotplug_workfn(struct work_struct
*work
)
2187 static cpumask_t new_cpus
;
2188 static nodemask_t new_mems
;
2189 bool cpus_updated
, mems_updated
;
2190 bool on_dfl
= cgroup_on_dfl(top_cpuset
.css
.cgroup
);
2192 mutex_lock(&cpuset_mutex
);
2194 /* fetch the available cpus/mems and find out which changed how */
2195 cpumask_copy(&new_cpus
, cpu_active_mask
);
2196 new_mems
= node_states
[N_MEMORY
];
2198 cpus_updated
= !cpumask_equal(top_cpuset
.effective_cpus
, &new_cpus
);
2199 mems_updated
= !nodes_equal(top_cpuset
.effective_mems
, new_mems
);
2201 /* synchronize cpus_allowed to cpu_active_mask */
2203 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
);
2215 top_cpuset
.mems_allowed
= new_mems
;
2216 top_cpuset
.effective_mems
= new_mems
;
2217 mutex_unlock(&callback_mutex
);
2218 update_tasks_nodemask(&top_cpuset
);
2221 mutex_unlock(&cpuset_mutex
);
2223 /* if cpus or mems changed, we need to propagate to descendants */
2224 if (cpus_updated
|| mems_updated
) {
2226 struct cgroup_subsys_state
*pos_css
;
2229 cpuset_for_each_descendant_pre(cs
, pos_css
, &top_cpuset
) {
2230 if (cs
== &top_cpuset
|| !css_tryget_online(&cs
->css
))
2234 cpuset_hotplug_update_tasks(cs
);
2242 /* rebuild sched domains if cpus_allowed has changed */
2244 rebuild_sched_domains();
2247 void cpuset_update_active_cpus(bool cpu_online
)
2250 * We're inside cpu hotplug critical region which usually nests
2251 * inside cgroup synchronization. Bounce actual hotplug processing
2252 * to a work item to avoid reverse locking order.
2254 * We still need to do partition_sched_domains() synchronously;
2255 * otherwise, the scheduler will get confused and put tasks to the
2256 * dead CPU. Fall back to the default single domain.
2257 * cpuset_hotplug_workfn() will rebuild it as necessary.
2259 partition_sched_domains(1, NULL
, NULL
);
2260 schedule_work(&cpuset_hotplug_work
);
2264 * Keep top_cpuset.mems_allowed tracking node_states[N_MEMORY].
2265 * Call this routine anytime after node_states[N_MEMORY] changes.
2266 * See cpuset_update_active_cpus() for CPU hotplug handling.
2268 static int cpuset_track_online_nodes(struct notifier_block
*self
,
2269 unsigned long action
, void *arg
)
2271 schedule_work(&cpuset_hotplug_work
);
2275 static struct notifier_block cpuset_track_online_nodes_nb
= {
2276 .notifier_call
= cpuset_track_online_nodes
,
2277 .priority
= 10, /* ??! */
2281 * cpuset_init_smp - initialize cpus_allowed
2283 * Description: Finish top cpuset after cpu, node maps are initialized
2285 void __init
cpuset_init_smp(void)
2287 cpumask_copy(top_cpuset
.cpus_allowed
, cpu_active_mask
);
2288 top_cpuset
.mems_allowed
= node_states
[N_MEMORY
];
2289 top_cpuset
.old_mems_allowed
= top_cpuset
.mems_allowed
;
2291 cpumask_copy(top_cpuset
.effective_cpus
, cpu_active_mask
);
2292 top_cpuset
.effective_mems
= node_states
[N_MEMORY
];
2294 register_hotmemory_notifier(&cpuset_track_online_nodes_nb
);
2298 * cpuset_cpus_allowed - return cpus_allowed mask from a tasks cpuset.
2299 * @tsk: pointer to task_struct from which to obtain cpuset->cpus_allowed.
2300 * @pmask: pointer to struct cpumask variable to receive cpus_allowed set.
2302 * Description: Returns the cpumask_var_t cpus_allowed of the cpuset
2303 * attached to the specified @tsk. Guaranteed to return some non-empty
2304 * subset of cpu_online_mask, even if this means going outside the
2308 void cpuset_cpus_allowed(struct task_struct
*tsk
, struct cpumask
*pmask
)
2310 mutex_lock(&callback_mutex
);
2312 guarantee_online_cpus(task_cs(tsk
), pmask
);
2314 mutex_unlock(&callback_mutex
);
2317 void cpuset_cpus_allowed_fallback(struct task_struct
*tsk
)
2320 do_set_cpus_allowed(tsk
, task_cs(tsk
)->effective_cpus
);
2324 * We own tsk->cpus_allowed, nobody can change it under us.
2326 * But we used cs && cs->cpus_allowed lockless and thus can
2327 * race with cgroup_attach_task() or update_cpumask() and get
2328 * the wrong tsk->cpus_allowed. However, both cases imply the
2329 * subsequent cpuset_change_cpumask()->set_cpus_allowed_ptr()
2330 * which takes task_rq_lock().
2332 * If we are called after it dropped the lock we must see all
2333 * changes in tsk_cs()->cpus_allowed. Otherwise we can temporary
2334 * set any mask even if it is not right from task_cs() pov,
2335 * the pending set_cpus_allowed_ptr() will fix things.
2337 * select_fallback_rq() will fix things ups and set cpu_possible_mask
2342 void cpuset_init_current_mems_allowed(void)
2344 nodes_setall(current
->mems_allowed
);
2348 * cpuset_mems_allowed - return mems_allowed mask from a tasks cpuset.
2349 * @tsk: pointer to task_struct from which to obtain cpuset->mems_allowed.
2351 * Description: Returns the nodemask_t mems_allowed of the cpuset
2352 * attached to the specified @tsk. Guaranteed to return some non-empty
2353 * subset of node_states[N_MEMORY], even if this means going outside the
2357 nodemask_t
cpuset_mems_allowed(struct task_struct
*tsk
)
2361 mutex_lock(&callback_mutex
);
2363 guarantee_online_mems(task_cs(tsk
), &mask
);
2365 mutex_unlock(&callback_mutex
);
2371 * cpuset_nodemask_valid_mems_allowed - check nodemask vs. curremt mems_allowed
2372 * @nodemask: the nodemask to be checked
2374 * Are any of the nodes in the nodemask allowed in current->mems_allowed?
2376 int cpuset_nodemask_valid_mems_allowed(nodemask_t
*nodemask
)
2378 return nodes_intersects(*nodemask
, current
->mems_allowed
);
2382 * nearest_hardwall_ancestor() - Returns the nearest mem_exclusive or
2383 * mem_hardwall ancestor to the specified cpuset. Call holding
2384 * callback_mutex. If no ancestor is mem_exclusive or mem_hardwall
2385 * (an unusual configuration), then returns the root cpuset.
2387 static struct cpuset
*nearest_hardwall_ancestor(struct cpuset
*cs
)
2389 while (!(is_mem_exclusive(cs
) || is_mem_hardwall(cs
)) && parent_cs(cs
))
2395 * cpuset_node_allowed_softwall - Can we allocate on a memory node?
2396 * @node: is this an allowed node?
2397 * @gfp_mask: memory allocation flags
2399 * If we're in interrupt, yes, we can always allocate. If __GFP_THISNODE is
2400 * set, yes, we can always allocate. If node is in our task's mems_allowed,
2401 * yes. If it's not a __GFP_HARDWALL request and this node is in the nearest
2402 * hardwalled cpuset ancestor to this task's cpuset, yes. If the task has been
2403 * OOM killed and has access to memory reserves as specified by the TIF_MEMDIE
2407 * If __GFP_HARDWALL is set, cpuset_node_allowed_softwall() reduces to
2408 * cpuset_node_allowed_hardwall(). Otherwise, cpuset_node_allowed_softwall()
2409 * might sleep, and might allow a node from an enclosing cpuset.
2411 * cpuset_node_allowed_hardwall() only handles the simpler case of hardwall
2412 * cpusets, and never sleeps.
2414 * The __GFP_THISNODE placement logic is really handled elsewhere,
2415 * by forcibly using a zonelist starting at a specified node, and by
2416 * (in get_page_from_freelist()) refusing to consider the zones for
2417 * any node on the zonelist except the first. By the time any such
2418 * calls get to this routine, we should just shut up and say 'yes'.
2420 * GFP_USER allocations are marked with the __GFP_HARDWALL bit,
2421 * and do not allow allocations outside the current tasks cpuset
2422 * unless the task has been OOM killed as is marked TIF_MEMDIE.
2423 * GFP_KERNEL allocations are not so marked, so can escape to the
2424 * nearest enclosing hardwalled ancestor cpuset.
2426 * Scanning up parent cpusets requires callback_mutex. The
2427 * __alloc_pages() routine only calls here with __GFP_HARDWALL bit
2428 * _not_ set if it's a GFP_KERNEL allocation, and all nodes in the
2429 * current tasks mems_allowed came up empty on the first pass over
2430 * the zonelist. So only GFP_KERNEL allocations, if all nodes in the
2431 * cpuset are short of memory, might require taking the callback_mutex
2434 * The first call here from mm/page_alloc:get_page_from_freelist()
2435 * has __GFP_HARDWALL set in gfp_mask, enforcing hardwall cpusets,
2436 * so no allocation on a node outside the cpuset is allowed (unless
2437 * in interrupt, of course).
2439 * The second pass through get_page_from_freelist() doesn't even call
2440 * here for GFP_ATOMIC calls. For those calls, the __alloc_pages()
2441 * variable 'wait' is not set, and the bit ALLOC_CPUSET is not set
2442 * in alloc_flags. That logic and the checks below have the combined
2444 * in_interrupt - any node ok (current task context irrelevant)
2445 * GFP_ATOMIC - any node ok
2446 * TIF_MEMDIE - any node ok
2447 * GFP_KERNEL - any node in enclosing hardwalled cpuset ok
2448 * GFP_USER - only nodes in current tasks mems allowed ok.
2451 * Don't call cpuset_node_allowed_softwall if you can't sleep, unless you
2452 * pass in the __GFP_HARDWALL flag set in gfp_flag, which disables
2453 * the code that might scan up ancestor cpusets and sleep.
2455 int __cpuset_node_allowed_softwall(int node
, gfp_t gfp_mask
)
2457 struct cpuset
*cs
; /* current cpuset ancestors */
2458 int allowed
; /* is allocation in zone z allowed? */
2460 if (in_interrupt() || (gfp_mask
& __GFP_THISNODE
))
2462 might_sleep_if(!(gfp_mask
& __GFP_HARDWALL
));
2463 if (node_isset(node
, current
->mems_allowed
))
2466 * Allow tasks that have access to memory reserves because they have
2467 * been OOM killed to get memory anywhere.
2469 if (unlikely(test_thread_flag(TIF_MEMDIE
)))
2471 if (gfp_mask
& __GFP_HARDWALL
) /* If hardwall request, stop here */
2474 if (current
->flags
& PF_EXITING
) /* Let dying task have memory */
2477 /* Not hardwall and node outside mems_allowed: scan up cpusets */
2478 mutex_lock(&callback_mutex
);
2481 cs
= nearest_hardwall_ancestor(task_cs(current
));
2482 allowed
= node_isset(node
, cs
->mems_allowed
);
2485 mutex_unlock(&callback_mutex
);
2490 * cpuset_node_allowed_hardwall - Can we allocate on a memory node?
2491 * @node: is this an allowed node?
2492 * @gfp_mask: memory allocation flags
2494 * If we're in interrupt, yes, we can always allocate. If __GFP_THISNODE is
2495 * set, yes, we can always allocate. If node is in our task's mems_allowed,
2496 * yes. If the task has been OOM killed and has access to memory reserves as
2497 * specified by the TIF_MEMDIE flag, yes.
2500 * The __GFP_THISNODE placement logic is really handled elsewhere,
2501 * by forcibly using a zonelist starting at a specified node, and by
2502 * (in get_page_from_freelist()) refusing to consider the zones for
2503 * any node on the zonelist except the first. By the time any such
2504 * calls get to this routine, we should just shut up and say 'yes'.
2506 * Unlike the cpuset_node_allowed_softwall() variant, above,
2507 * this variant requires that the node be in the current task's
2508 * mems_allowed or that we're in interrupt. It does not scan up the
2509 * cpuset hierarchy for the nearest enclosing mem_exclusive cpuset.
2512 int __cpuset_node_allowed_hardwall(int node
, gfp_t gfp_mask
)
2514 if (in_interrupt() || (gfp_mask
& __GFP_THISNODE
))
2516 if (node_isset(node
, current
->mems_allowed
))
2519 * Allow tasks that have access to memory reserves because they have
2520 * been OOM killed to get memory anywhere.
2522 if (unlikely(test_thread_flag(TIF_MEMDIE
)))
2528 * cpuset_mem_spread_node() - On which node to begin search for a file page
2529 * cpuset_slab_spread_node() - On which node to begin search for a slab page
2531 * If a task is marked PF_SPREAD_PAGE or PF_SPREAD_SLAB (as for
2532 * tasks in a cpuset with is_spread_page or is_spread_slab set),
2533 * and if the memory allocation used cpuset_mem_spread_node()
2534 * to determine on which node to start looking, as it will for
2535 * certain page cache or slab cache pages such as used for file
2536 * system buffers and inode caches, then instead of starting on the
2537 * local node to look for a free page, rather spread the starting
2538 * node around the tasks mems_allowed nodes.
2540 * We don't have to worry about the returned node being offline
2541 * because "it can't happen", and even if it did, it would be ok.
2543 * The routines calling guarantee_online_mems() are careful to
2544 * only set nodes in task->mems_allowed that are online. So it
2545 * should not be possible for the following code to return an
2546 * offline node. But if it did, that would be ok, as this routine
2547 * is not returning the node where the allocation must be, only
2548 * the node where the search should start. The zonelist passed to
2549 * __alloc_pages() will include all nodes. If the slab allocator
2550 * is passed an offline node, it will fall back to the local node.
2551 * See kmem_cache_alloc_node().
2554 static int cpuset_spread_node(int *rotor
)
2558 node
= next_node(*rotor
, current
->mems_allowed
);
2559 if (node
== MAX_NUMNODES
)
2560 node
= first_node(current
->mems_allowed
);
2565 int cpuset_mem_spread_node(void)
2567 if (current
->cpuset_mem_spread_rotor
== NUMA_NO_NODE
)
2568 current
->cpuset_mem_spread_rotor
=
2569 node_random(¤t
->mems_allowed
);
2571 return cpuset_spread_node(¤t
->cpuset_mem_spread_rotor
);
2574 int cpuset_slab_spread_node(void)
2576 if (current
->cpuset_slab_spread_rotor
== NUMA_NO_NODE
)
2577 current
->cpuset_slab_spread_rotor
=
2578 node_random(¤t
->mems_allowed
);
2580 return cpuset_spread_node(¤t
->cpuset_slab_spread_rotor
);
2583 EXPORT_SYMBOL_GPL(cpuset_mem_spread_node
);
2586 * cpuset_mems_allowed_intersects - Does @tsk1's mems_allowed intersect @tsk2's?
2587 * @tsk1: pointer to task_struct of some task.
2588 * @tsk2: pointer to task_struct of some other task.
2590 * Description: Return true if @tsk1's mems_allowed intersects the
2591 * mems_allowed of @tsk2. Used by the OOM killer to determine if
2592 * one of the task's memory usage might impact the memory available
2596 int cpuset_mems_allowed_intersects(const struct task_struct
*tsk1
,
2597 const struct task_struct
*tsk2
)
2599 return nodes_intersects(tsk1
->mems_allowed
, tsk2
->mems_allowed
);
2602 #define CPUSET_NODELIST_LEN (256)
2605 * cpuset_print_task_mems_allowed - prints task's cpuset and mems_allowed
2606 * @tsk: pointer to task_struct of some task.
2608 * Description: Prints @task's name, cpuset name, and cached copy of its
2609 * mems_allowed to the kernel log.
2611 void cpuset_print_task_mems_allowed(struct task_struct
*tsk
)
2613 /* Statically allocated to prevent using excess stack. */
2614 static char cpuset_nodelist
[CPUSET_NODELIST_LEN
];
2615 static DEFINE_SPINLOCK(cpuset_buffer_lock
);
2616 struct cgroup
*cgrp
;
2618 spin_lock(&cpuset_buffer_lock
);
2621 cgrp
= task_cs(tsk
)->css
.cgroup
;
2622 nodelist_scnprintf(cpuset_nodelist
, CPUSET_NODELIST_LEN
,
2624 pr_info("%s cpuset=", tsk
->comm
);
2625 pr_cont_cgroup_name(cgrp
);
2626 pr_cont(" mems_allowed=%s\n", cpuset_nodelist
);
2629 spin_unlock(&cpuset_buffer_lock
);
2633 * Collection of memory_pressure is suppressed unless
2634 * this flag is enabled by writing "1" to the special
2635 * cpuset file 'memory_pressure_enabled' in the root cpuset.
2638 int cpuset_memory_pressure_enabled __read_mostly
;
2641 * cpuset_memory_pressure_bump - keep stats of per-cpuset reclaims.
2643 * Keep a running average of the rate of synchronous (direct)
2644 * page reclaim efforts initiated by tasks in each cpuset.
2646 * This represents the rate at which some task in the cpuset
2647 * ran low on memory on all nodes it was allowed to use, and
2648 * had to enter the kernels page reclaim code in an effort to
2649 * create more free memory by tossing clean pages or swapping
2650 * or writing dirty pages.
2652 * Display to user space in the per-cpuset read-only file
2653 * "memory_pressure". Value displayed is an integer
2654 * representing the recent rate of entry into the synchronous
2655 * (direct) page reclaim by any task attached to the cpuset.
2658 void __cpuset_memory_pressure_bump(void)
2661 fmeter_markevent(&task_cs(current
)->fmeter
);
2665 #ifdef CONFIG_PROC_PID_CPUSET
2667 * proc_cpuset_show()
2668 * - Print tasks cpuset path into seq_file.
2669 * - Used for /proc/<pid>/cpuset.
2670 * - No need to task_lock(tsk) on this tsk->cpuset reference, as it
2671 * doesn't really matter if tsk->cpuset changes after we read it,
2672 * and we take cpuset_mutex, keeping cpuset_attach() from changing it
2675 int proc_cpuset_show(struct seq_file
*m
, void *unused_v
)
2678 struct task_struct
*tsk
;
2680 struct cgroup_subsys_state
*css
;
2684 buf
= kmalloc(PATH_MAX
, GFP_KERNEL
);
2690 tsk
= get_pid_task(pid
, PIDTYPE_PID
);
2694 retval
= -ENAMETOOLONG
;
2696 css
= task_css(tsk
, cpuset_cgrp_id
);
2697 p
= cgroup_path(css
->cgroup
, buf
, PATH_MAX
);
2705 put_task_struct(tsk
);
2711 #endif /* CONFIG_PROC_PID_CPUSET */
2713 /* Display task mems_allowed in /proc/<pid>/status file. */
2714 void cpuset_task_status_allowed(struct seq_file
*m
, struct task_struct
*task
)
2716 seq_puts(m
, "Mems_allowed:\t");
2717 seq_nodemask(m
, &task
->mems_allowed
);
2719 seq_puts(m
, "Mems_allowed_list:\t");
2720 seq_nodemask_list(m
, &task
->mems_allowed
);