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/sched/mm.h>
48 #include <linux/sched/task.h>
49 #include <linux/seq_file.h>
50 #include <linux/security.h>
51 #include <linux/slab.h>
52 #include <linux/spinlock.h>
53 #include <linux/stat.h>
54 #include <linux/string.h>
55 #include <linux/time.h>
56 #include <linux/time64.h>
57 #include <linux/backing-dev.h>
58 #include <linux/sort.h>
60 #include <linux/uaccess.h>
61 #include <linux/atomic.h>
62 #include <linux/mutex.h>
63 #include <linux/cgroup.h>
64 #include <linux/wait.h>
66 DEFINE_STATIC_KEY_FALSE(cpusets_pre_enable_key
);
67 DEFINE_STATIC_KEY_FALSE(cpusets_enabled_key
);
69 /* See "Frequency meter" comments, below. */
72 int cnt
; /* unprocessed events count */
73 int val
; /* most recent output value */
74 time64_t time
; /* clock (secs) when val computed */
75 spinlock_t lock
; /* guards read or write of above */
79 struct cgroup_subsys_state css
;
81 unsigned long flags
; /* "unsigned long" so bitops work */
84 * On default hierarchy:
86 * The user-configured masks can only be changed by writing to
87 * cpuset.cpus and cpuset.mems, and won't be limited by the
90 * The effective masks is the real masks that apply to the tasks
91 * in the cpuset. They may be changed if the configured masks are
92 * changed or hotplug happens.
94 * effective_mask == configured_mask & parent's effective_mask,
95 * and if it ends up empty, it will inherit the parent's mask.
100 * The user-configured masks are always the same with effective masks.
103 /* user-configured CPUs and Memory Nodes allow to tasks */
104 cpumask_var_t cpus_allowed
;
105 nodemask_t mems_allowed
;
107 /* effective CPUs and Memory Nodes allow to tasks */
108 cpumask_var_t effective_cpus
;
109 nodemask_t effective_mems
;
112 * This is old Memory Nodes tasks took on.
114 * - top_cpuset.old_mems_allowed is initialized to mems_allowed.
115 * - A new cpuset's old_mems_allowed is initialized when some
116 * task is moved into it.
117 * - old_mems_allowed is used in cpuset_migrate_mm() when we change
118 * cpuset.mems_allowed and have tasks' nodemask updated, and
119 * then old_mems_allowed is updated to mems_allowed.
121 nodemask_t old_mems_allowed
;
123 struct fmeter fmeter
; /* memory_pressure filter */
126 * Tasks are being attached to this cpuset. Used to prevent
127 * zeroing cpus/mems_allowed between ->can_attach() and ->attach().
129 int attach_in_progress
;
131 /* partition number for rebuild_sched_domains() */
134 /* for custom sched domain */
135 int relax_domain_level
;
138 static inline struct cpuset
*css_cs(struct cgroup_subsys_state
*css
)
140 return css
? container_of(css
, struct cpuset
, css
) : NULL
;
143 /* Retrieve the cpuset for a task */
144 static inline struct cpuset
*task_cs(struct task_struct
*task
)
146 return css_cs(task_css(task
, cpuset_cgrp_id
));
149 static inline struct cpuset
*parent_cs(struct cpuset
*cs
)
151 return css_cs(cs
->css
.parent
);
155 static inline bool task_has_mempolicy(struct task_struct
*task
)
157 return task
->mempolicy
;
160 static inline bool task_has_mempolicy(struct task_struct
*task
)
167 /* bits in struct cpuset flags field */
174 CS_SCHED_LOAD_BALANCE
,
179 /* convenient tests for these bits */
180 static inline bool is_cpuset_online(struct cpuset
*cs
)
182 return test_bit(CS_ONLINE
, &cs
->flags
) && !css_is_dying(&cs
->css
);
185 static inline int is_cpu_exclusive(const struct cpuset
*cs
)
187 return test_bit(CS_CPU_EXCLUSIVE
, &cs
->flags
);
190 static inline int is_mem_exclusive(const struct cpuset
*cs
)
192 return test_bit(CS_MEM_EXCLUSIVE
, &cs
->flags
);
195 static inline int is_mem_hardwall(const struct cpuset
*cs
)
197 return test_bit(CS_MEM_HARDWALL
, &cs
->flags
);
200 static inline int is_sched_load_balance(const struct cpuset
*cs
)
202 return test_bit(CS_SCHED_LOAD_BALANCE
, &cs
->flags
);
205 static inline int is_memory_migrate(const struct cpuset
*cs
)
207 return test_bit(CS_MEMORY_MIGRATE
, &cs
->flags
);
210 static inline int is_spread_page(const struct cpuset
*cs
)
212 return test_bit(CS_SPREAD_PAGE
, &cs
->flags
);
215 static inline int is_spread_slab(const struct cpuset
*cs
)
217 return test_bit(CS_SPREAD_SLAB
, &cs
->flags
);
220 static struct cpuset top_cpuset
= {
221 .flags
= ((1 << CS_ONLINE
) | (1 << CS_CPU_EXCLUSIVE
) |
222 (1 << CS_MEM_EXCLUSIVE
)),
226 * cpuset_for_each_child - traverse online children of a cpuset
227 * @child_cs: loop cursor pointing to the current child
228 * @pos_css: used for iteration
229 * @parent_cs: target cpuset to walk children of
231 * Walk @child_cs through the online children of @parent_cs. Must be used
232 * with RCU read locked.
234 #define cpuset_for_each_child(child_cs, pos_css, parent_cs) \
235 css_for_each_child((pos_css), &(parent_cs)->css) \
236 if (is_cpuset_online(((child_cs) = css_cs((pos_css)))))
239 * cpuset_for_each_descendant_pre - pre-order walk of a cpuset's descendants
240 * @des_cs: loop cursor pointing to the current descendant
241 * @pos_css: used for iteration
242 * @root_cs: target cpuset to walk ancestor of
244 * Walk @des_cs through the online descendants of @root_cs. Must be used
245 * with RCU read locked. The caller may modify @pos_css by calling
246 * css_rightmost_descendant() to skip subtree. @root_cs is included in the
247 * iteration and the first node to be visited.
249 #define cpuset_for_each_descendant_pre(des_cs, pos_css, root_cs) \
250 css_for_each_descendant_pre((pos_css), &(root_cs)->css) \
251 if (is_cpuset_online(((des_cs) = css_cs((pos_css)))))
254 * There are two global locks guarding cpuset structures - cpuset_mutex and
255 * callback_lock. We also require taking task_lock() when dereferencing a
256 * task's cpuset pointer. See "The task_lock() exception", at the end of this
259 * A task must hold both locks to modify cpusets. If a task holds
260 * cpuset_mutex, then it blocks others wanting that mutex, ensuring that it
261 * is the only task able to also acquire callback_lock and be able to
262 * modify cpusets. It can perform various checks on the cpuset structure
263 * first, knowing nothing will change. It can also allocate memory while
264 * just holding cpuset_mutex. While it is performing these checks, various
265 * callback routines can briefly acquire callback_lock to query cpusets.
266 * Once it is ready to make the changes, it takes callback_lock, blocking
269 * Calls to the kernel memory allocator can not be made while holding
270 * callback_lock, as that would risk double tripping on callback_lock
271 * from one of the callbacks into the cpuset code from within
274 * If a task is only holding callback_lock, then it has read-only
277 * Now, the task_struct fields mems_allowed and mempolicy may be changed
278 * by other task, we use alloc_lock in the task_struct fields to protect
281 * The cpuset_common_file_read() handlers only hold callback_lock across
282 * small pieces of code, such as when reading out possibly multi-word
283 * cpumasks and nodemasks.
285 * Accessing a task's cpuset should be done in accordance with the
286 * guidelines for accessing subsystem state in kernel/cgroup.c
289 static DEFINE_MUTEX(cpuset_mutex
);
290 static DEFINE_SPINLOCK(callback_lock
);
292 static struct workqueue_struct
*cpuset_migrate_mm_wq
;
295 * CPU / memory hotplug is handled asynchronously.
297 static void cpuset_hotplug_workfn(struct work_struct
*work
);
298 static DECLARE_WORK(cpuset_hotplug_work
, cpuset_hotplug_workfn
);
300 static DECLARE_WAIT_QUEUE_HEAD(cpuset_attach_wq
);
303 * This is ugly, but preserves the userspace API for existing cpuset
304 * users. If someone tries to mount the "cpuset" filesystem, we
305 * silently switch it to mount "cgroup" instead
307 static struct dentry
*cpuset_mount(struct file_system_type
*fs_type
,
308 int flags
, const char *unused_dev_name
, void *data
)
310 struct file_system_type
*cgroup_fs
= get_fs_type("cgroup");
311 struct dentry
*ret
= ERR_PTR(-ENODEV
);
315 "release_agent=/sbin/cpuset_release_agent";
316 ret
= cgroup_fs
->mount(cgroup_fs
, flags
,
317 unused_dev_name
, mountopts
);
318 put_filesystem(cgroup_fs
);
323 static struct file_system_type cpuset_fs_type
= {
325 .mount
= cpuset_mount
,
329 * Return in pmask the portion of a cpusets's cpus_allowed that
330 * are online. If none are online, walk up the cpuset hierarchy
331 * until we find one that does have some online cpus.
333 * One way or another, we guarantee to return some non-empty subset
334 * of cpu_online_mask.
336 * Call with callback_lock or cpuset_mutex held.
338 static void guarantee_online_cpus(struct cpuset
*cs
, struct cpumask
*pmask
)
340 while (!cpumask_intersects(cs
->effective_cpus
, cpu_online_mask
)) {
344 * The top cpuset doesn't have any online cpu as a
345 * consequence of a race between cpuset_hotplug_work
346 * and cpu hotplug notifier. But we know the top
347 * cpuset's effective_cpus is on its way to to be
348 * identical to cpu_online_mask.
350 cpumask_copy(pmask
, cpu_online_mask
);
354 cpumask_and(pmask
, cs
->effective_cpus
, cpu_online_mask
);
358 * Return in *pmask the portion of a cpusets's mems_allowed that
359 * are online, with memory. If none are online with memory, walk
360 * up the cpuset hierarchy until we find one that does have some
361 * online mems. The top cpuset always has some mems online.
363 * One way or another, we guarantee to return some non-empty subset
364 * of node_states[N_MEMORY].
366 * Call with callback_lock or cpuset_mutex held.
368 static void guarantee_online_mems(struct cpuset
*cs
, nodemask_t
*pmask
)
370 while (!nodes_intersects(cs
->effective_mems
, node_states
[N_MEMORY
]))
372 nodes_and(*pmask
, cs
->effective_mems
, node_states
[N_MEMORY
]);
376 * update task's spread flag if cpuset's page/slab spread flag is set
378 * Call with callback_lock or cpuset_mutex held.
380 static void cpuset_update_task_spread_flag(struct cpuset
*cs
,
381 struct task_struct
*tsk
)
383 if (is_spread_page(cs
))
384 task_set_spread_page(tsk
);
386 task_clear_spread_page(tsk
);
388 if (is_spread_slab(cs
))
389 task_set_spread_slab(tsk
);
391 task_clear_spread_slab(tsk
);
395 * is_cpuset_subset(p, q) - Is cpuset p a subset of cpuset q?
397 * One cpuset is a subset of another if all its allowed CPUs and
398 * Memory Nodes are a subset of the other, and its exclusive flags
399 * are only set if the other's are set. Call holding cpuset_mutex.
402 static int is_cpuset_subset(const struct cpuset
*p
, const struct cpuset
*q
)
404 return cpumask_subset(p
->cpus_allowed
, q
->cpus_allowed
) &&
405 nodes_subset(p
->mems_allowed
, q
->mems_allowed
) &&
406 is_cpu_exclusive(p
) <= is_cpu_exclusive(q
) &&
407 is_mem_exclusive(p
) <= is_mem_exclusive(q
);
411 * alloc_trial_cpuset - allocate a trial cpuset
412 * @cs: the cpuset that the trial cpuset duplicates
414 static struct cpuset
*alloc_trial_cpuset(struct cpuset
*cs
)
416 struct cpuset
*trial
;
418 trial
= kmemdup(cs
, sizeof(*cs
), GFP_KERNEL
);
422 if (!alloc_cpumask_var(&trial
->cpus_allowed
, GFP_KERNEL
))
424 if (!alloc_cpumask_var(&trial
->effective_cpus
, GFP_KERNEL
))
427 cpumask_copy(trial
->cpus_allowed
, cs
->cpus_allowed
);
428 cpumask_copy(trial
->effective_cpus
, cs
->effective_cpus
);
432 free_cpumask_var(trial
->cpus_allowed
);
439 * free_trial_cpuset - free the trial cpuset
440 * @trial: the trial cpuset to be freed
442 static void free_trial_cpuset(struct cpuset
*trial
)
444 free_cpumask_var(trial
->effective_cpus
);
445 free_cpumask_var(trial
->cpus_allowed
);
450 * validate_change() - Used to validate that any proposed cpuset change
451 * follows the structural rules for cpusets.
453 * If we replaced the flag and mask values of the current cpuset
454 * (cur) with those values in the trial cpuset (trial), would
455 * our various subset and exclusive rules still be valid? Presumes
458 * 'cur' is the address of an actual, in-use cpuset. Operations
459 * such as list traversal that depend on the actual address of the
460 * cpuset in the list must use cur below, not trial.
462 * 'trial' is the address of bulk structure copy of cur, with
463 * perhaps one or more of the fields cpus_allowed, mems_allowed,
464 * or flags changed to new, trial values.
466 * Return 0 if valid, -errno if not.
469 static int validate_change(struct cpuset
*cur
, struct cpuset
*trial
)
471 struct cgroup_subsys_state
*css
;
472 struct cpuset
*c
, *par
;
477 /* Each of our child cpusets must be a subset of us */
479 cpuset_for_each_child(c
, css
, cur
)
480 if (!is_cpuset_subset(c
, trial
))
483 /* Remaining checks don't apply to root cpuset */
485 if (cur
== &top_cpuset
)
488 par
= parent_cs(cur
);
490 /* On legacy hiearchy, we must be a subset of our parent cpuset. */
492 if (!cgroup_subsys_on_dfl(cpuset_cgrp_subsys
) &&
493 !is_cpuset_subset(trial
, par
))
497 * If either I or some sibling (!= me) is exclusive, we can't
501 cpuset_for_each_child(c
, css
, par
) {
502 if ((is_cpu_exclusive(trial
) || is_cpu_exclusive(c
)) &&
504 cpumask_intersects(trial
->cpus_allowed
, c
->cpus_allowed
))
506 if ((is_mem_exclusive(trial
) || is_mem_exclusive(c
)) &&
508 nodes_intersects(trial
->mems_allowed
, c
->mems_allowed
))
513 * Cpusets with tasks - existing or newly being attached - can't
514 * be changed to have empty cpus_allowed or mems_allowed.
517 if ((cgroup_is_populated(cur
->css
.cgroup
) || cur
->attach_in_progress
)) {
518 if (!cpumask_empty(cur
->cpus_allowed
) &&
519 cpumask_empty(trial
->cpus_allowed
))
521 if (!nodes_empty(cur
->mems_allowed
) &&
522 nodes_empty(trial
->mems_allowed
))
527 * We can't shrink if we won't have enough room for SCHED_DEADLINE
531 if (is_cpu_exclusive(cur
) &&
532 !cpuset_cpumask_can_shrink(cur
->cpus_allowed
,
533 trial
->cpus_allowed
))
544 * Helper routine for generate_sched_domains().
545 * Do cpusets a, b have overlapping effective cpus_allowed masks?
547 static int cpusets_overlap(struct cpuset
*a
, struct cpuset
*b
)
549 return cpumask_intersects(a
->effective_cpus
, b
->effective_cpus
);
553 update_domain_attr(struct sched_domain_attr
*dattr
, struct cpuset
*c
)
555 if (dattr
->relax_domain_level
< c
->relax_domain_level
)
556 dattr
->relax_domain_level
= c
->relax_domain_level
;
560 static void update_domain_attr_tree(struct sched_domain_attr
*dattr
,
561 struct cpuset
*root_cs
)
564 struct cgroup_subsys_state
*pos_css
;
567 cpuset_for_each_descendant_pre(cp
, pos_css
, root_cs
) {
568 /* skip the whole subtree if @cp doesn't have any CPU */
569 if (cpumask_empty(cp
->cpus_allowed
)) {
570 pos_css
= css_rightmost_descendant(pos_css
);
574 if (is_sched_load_balance(cp
))
575 update_domain_attr(dattr
, cp
);
581 * generate_sched_domains()
583 * This function builds a partial partition of the systems CPUs
584 * A 'partial partition' is a set of non-overlapping subsets whose
585 * union is a subset of that set.
586 * The output of this function needs to be passed to kernel/sched/core.c
587 * partition_sched_domains() routine, which will rebuild the scheduler's
588 * load balancing domains (sched domains) as specified by that partial
591 * See "What is sched_load_balance" in Documentation/cgroups/cpusets.txt
592 * for a background explanation of this.
594 * Does not return errors, on the theory that the callers of this
595 * routine would rather not worry about failures to rebuild sched
596 * domains when operating in the severe memory shortage situations
597 * that could cause allocation failures below.
599 * Must be called with cpuset_mutex held.
601 * The three key local variables below are:
602 * q - a linked-list queue of cpuset pointers, used to implement a
603 * top-down scan of all cpusets. This scan loads a pointer
604 * to each cpuset marked is_sched_load_balance into the
605 * array 'csa'. For our purposes, rebuilding the schedulers
606 * sched domains, we can ignore !is_sched_load_balance cpusets.
607 * csa - (for CpuSet Array) Array of pointers to all the cpusets
608 * that need to be load balanced, for convenient iterative
609 * access by the subsequent code that finds the best partition,
610 * i.e the set of domains (subsets) of CPUs such that the
611 * cpus_allowed of every cpuset marked is_sched_load_balance
612 * is a subset of one of these domains, while there are as
613 * many such domains as possible, each as small as possible.
614 * doms - Conversion of 'csa' to an array of cpumasks, for passing to
615 * the kernel/sched/core.c routine partition_sched_domains() in a
616 * convenient format, that can be easily compared to the prior
617 * value to determine what partition elements (sched domains)
618 * were changed (added or removed.)
620 * Finding the best partition (set of domains):
621 * The triple nested loops below over i, j, k scan over the
622 * load balanced cpusets (using the array of cpuset pointers in
623 * csa[]) looking for pairs of cpusets that have overlapping
624 * cpus_allowed, but which don't have the same 'pn' partition
625 * number and gives them in the same partition number. It keeps
626 * looping on the 'restart' label until it can no longer find
629 * The union of the cpus_allowed masks from the set of
630 * all cpusets having the same 'pn' value then form the one
631 * element of the partition (one sched domain) to be passed to
632 * partition_sched_domains().
634 static int generate_sched_domains(cpumask_var_t
**domains
,
635 struct sched_domain_attr
**attributes
)
637 struct cpuset
*cp
; /* scans q */
638 struct cpuset
**csa
; /* array of all cpuset ptrs */
639 int csn
; /* how many cpuset ptrs in csa so far */
640 int i
, j
, k
; /* indices for partition finding loops */
641 cpumask_var_t
*doms
; /* resulting partition; i.e. sched domains */
642 cpumask_var_t non_isolated_cpus
; /* load balanced CPUs */
643 struct sched_domain_attr
*dattr
; /* attributes for custom domains */
644 int ndoms
= 0; /* number of sched domains in result */
645 int nslot
; /* next empty doms[] struct cpumask slot */
646 struct cgroup_subsys_state
*pos_css
;
652 if (!alloc_cpumask_var(&non_isolated_cpus
, GFP_KERNEL
))
654 cpumask_andnot(non_isolated_cpus
, cpu_possible_mask
, cpu_isolated_map
);
656 /* Special case for the 99% of systems with one, full, sched domain */
657 if (is_sched_load_balance(&top_cpuset
)) {
659 doms
= alloc_sched_domains(ndoms
);
663 dattr
= kmalloc(sizeof(struct sched_domain_attr
), GFP_KERNEL
);
665 *dattr
= SD_ATTR_INIT
;
666 update_domain_attr_tree(dattr
, &top_cpuset
);
668 cpumask_and(doms
[0], top_cpuset
.effective_cpus
,
674 csa
= kmalloc(nr_cpusets() * sizeof(cp
), GFP_KERNEL
);
680 cpuset_for_each_descendant_pre(cp
, pos_css
, &top_cpuset
) {
681 if (cp
== &top_cpuset
)
684 * Continue traversing beyond @cp iff @cp has some CPUs and
685 * isn't load balancing. The former is obvious. The
686 * latter: All child cpusets contain a subset of the
687 * parent's cpus, so just skip them, and then we call
688 * update_domain_attr_tree() to calc relax_domain_level of
689 * the corresponding sched domain.
691 if (!cpumask_empty(cp
->cpus_allowed
) &&
692 !(is_sched_load_balance(cp
) &&
693 cpumask_intersects(cp
->cpus_allowed
, non_isolated_cpus
)))
696 if (is_sched_load_balance(cp
))
699 /* skip @cp's subtree */
700 pos_css
= css_rightmost_descendant(pos_css
);
704 for (i
= 0; i
< csn
; i
++)
709 /* Find the best partition (set of sched domains) */
710 for (i
= 0; i
< csn
; i
++) {
711 struct cpuset
*a
= csa
[i
];
714 for (j
= 0; j
< csn
; j
++) {
715 struct cpuset
*b
= csa
[j
];
718 if (apn
!= bpn
&& cpusets_overlap(a
, b
)) {
719 for (k
= 0; k
< csn
; k
++) {
720 struct cpuset
*c
= csa
[k
];
725 ndoms
--; /* one less element */
732 * Now we know how many domains to create.
733 * Convert <csn, csa> to <ndoms, doms> and populate cpu masks.
735 doms
= alloc_sched_domains(ndoms
);
740 * The rest of the code, including the scheduler, can deal with
741 * dattr==NULL case. No need to abort if alloc fails.
743 dattr
= kmalloc(ndoms
* sizeof(struct sched_domain_attr
), GFP_KERNEL
);
745 for (nslot
= 0, i
= 0; i
< csn
; i
++) {
746 struct cpuset
*a
= csa
[i
];
751 /* Skip completed partitions */
757 if (nslot
== ndoms
) {
758 static int warnings
= 10;
760 pr_warn("rebuild_sched_domains confused: nslot %d, ndoms %d, csn %d, i %d, apn %d\n",
761 nslot
, ndoms
, csn
, i
, apn
);
769 *(dattr
+ nslot
) = SD_ATTR_INIT
;
770 for (j
= i
; j
< csn
; j
++) {
771 struct cpuset
*b
= csa
[j
];
774 cpumask_or(dp
, dp
, b
->effective_cpus
);
775 cpumask_and(dp
, dp
, non_isolated_cpus
);
777 update_domain_attr_tree(dattr
+ nslot
, b
);
779 /* Done with this partition */
785 BUG_ON(nslot
!= ndoms
);
788 free_cpumask_var(non_isolated_cpus
);
792 * Fallback to the default domain if kmalloc() failed.
793 * See comments in partition_sched_domains().
804 * Rebuild scheduler domains.
806 * If the flag 'sched_load_balance' of any cpuset with non-empty
807 * 'cpus' changes, or if the 'cpus' allowed changes in any cpuset
808 * which has that flag enabled, or if any cpuset with a non-empty
809 * 'cpus' is removed, then call this routine to rebuild the
810 * scheduler's dynamic sched domains.
812 * Call with cpuset_mutex held. Takes get_online_cpus().
814 static void rebuild_sched_domains_locked(void)
816 struct sched_domain_attr
*attr
;
820 lockdep_assert_held(&cpuset_mutex
);
824 * We have raced with CPU hotplug. Don't do anything to avoid
825 * passing doms with offlined cpu to partition_sched_domains().
826 * Anyways, hotplug work item will rebuild sched domains.
828 if (!cpumask_equal(top_cpuset
.effective_cpus
, cpu_active_mask
))
831 /* Generate domain masks and attrs */
832 ndoms
= generate_sched_domains(&doms
, &attr
);
834 /* Have scheduler rebuild the domains */
835 partition_sched_domains(ndoms
, doms
, attr
);
839 #else /* !CONFIG_SMP */
840 static void rebuild_sched_domains_locked(void)
843 #endif /* CONFIG_SMP */
845 void rebuild_sched_domains(void)
847 mutex_lock(&cpuset_mutex
);
848 rebuild_sched_domains_locked();
849 mutex_unlock(&cpuset_mutex
);
853 * update_tasks_cpumask - Update the cpumasks of tasks in the cpuset.
854 * @cs: the cpuset in which each task's cpus_allowed mask needs to be changed
856 * Iterate through each task of @cs updating its cpus_allowed to the
857 * effective cpuset's. As this function is called with cpuset_mutex held,
858 * cpuset membership stays stable.
860 static void update_tasks_cpumask(struct cpuset
*cs
)
862 struct css_task_iter it
;
863 struct task_struct
*task
;
865 css_task_iter_start(&cs
->css
, &it
);
866 while ((task
= css_task_iter_next(&it
)))
867 set_cpus_allowed_ptr(task
, cs
->effective_cpus
);
868 css_task_iter_end(&it
);
872 * update_cpumasks_hier - Update effective cpumasks and tasks in the subtree
873 * @cs: the cpuset to consider
874 * @new_cpus: temp variable for calculating new effective_cpus
876 * When congifured cpumask is changed, the effective cpumasks of this cpuset
877 * and all its descendants need to be updated.
879 * On legacy hierachy, effective_cpus will be the same with cpu_allowed.
881 * Called with cpuset_mutex held
883 static void update_cpumasks_hier(struct cpuset
*cs
, struct cpumask
*new_cpus
)
886 struct cgroup_subsys_state
*pos_css
;
887 bool need_rebuild_sched_domains
= false;
890 cpuset_for_each_descendant_pre(cp
, pos_css
, cs
) {
891 struct cpuset
*parent
= parent_cs(cp
);
893 cpumask_and(new_cpus
, cp
->cpus_allowed
, parent
->effective_cpus
);
896 * If it becomes empty, inherit the effective mask of the
897 * parent, which is guaranteed to have some CPUs.
899 if (cgroup_subsys_on_dfl(cpuset_cgrp_subsys
) &&
900 cpumask_empty(new_cpus
))
901 cpumask_copy(new_cpus
, parent
->effective_cpus
);
903 /* Skip the whole subtree if the cpumask remains the same. */
904 if (cpumask_equal(new_cpus
, cp
->effective_cpus
)) {
905 pos_css
= css_rightmost_descendant(pos_css
);
909 if (!css_tryget_online(&cp
->css
))
913 spin_lock_irq(&callback_lock
);
914 cpumask_copy(cp
->effective_cpus
, new_cpus
);
915 spin_unlock_irq(&callback_lock
);
917 WARN_ON(!cgroup_subsys_on_dfl(cpuset_cgrp_subsys
) &&
918 !cpumask_equal(cp
->cpus_allowed
, cp
->effective_cpus
));
920 update_tasks_cpumask(cp
);
923 * If the effective cpumask of any non-empty cpuset is changed,
924 * we need to rebuild sched domains.
926 if (!cpumask_empty(cp
->cpus_allowed
) &&
927 is_sched_load_balance(cp
))
928 need_rebuild_sched_domains
= true;
935 if (need_rebuild_sched_domains
)
936 rebuild_sched_domains_locked();
940 * update_cpumask - update the cpus_allowed mask of a cpuset and all tasks in it
941 * @cs: the cpuset to consider
942 * @trialcs: trial cpuset
943 * @buf: buffer of cpu numbers written to this cpuset
945 static int update_cpumask(struct cpuset
*cs
, struct cpuset
*trialcs
,
950 /* top_cpuset.cpus_allowed tracks cpu_online_mask; it's read-only */
951 if (cs
== &top_cpuset
)
955 * An empty cpus_allowed is ok only if the cpuset has no tasks.
956 * Since cpulist_parse() fails on an empty mask, we special case
957 * that parsing. The validate_change() call ensures that cpusets
958 * with tasks have cpus.
961 cpumask_clear(trialcs
->cpus_allowed
);
963 retval
= cpulist_parse(buf
, trialcs
->cpus_allowed
);
967 if (!cpumask_subset(trialcs
->cpus_allowed
,
968 top_cpuset
.cpus_allowed
))
972 /* Nothing to do if the cpus didn't change */
973 if (cpumask_equal(cs
->cpus_allowed
, trialcs
->cpus_allowed
))
976 retval
= validate_change(cs
, trialcs
);
980 spin_lock_irq(&callback_lock
);
981 cpumask_copy(cs
->cpus_allowed
, trialcs
->cpus_allowed
);
982 spin_unlock_irq(&callback_lock
);
984 /* use trialcs->cpus_allowed as a temp variable */
985 update_cpumasks_hier(cs
, trialcs
->cpus_allowed
);
990 * Migrate memory region from one set of nodes to another. This is
991 * performed asynchronously as it can be called from process migration path
992 * holding locks involved in process management. All mm migrations are
993 * performed in the queued order and can be waited for by flushing
994 * cpuset_migrate_mm_wq.
997 struct cpuset_migrate_mm_work
{
998 struct work_struct work
;
999 struct mm_struct
*mm
;
1004 static void cpuset_migrate_mm_workfn(struct work_struct
*work
)
1006 struct cpuset_migrate_mm_work
*mwork
=
1007 container_of(work
, struct cpuset_migrate_mm_work
, work
);
1009 /* on a wq worker, no need to worry about %current's mems_allowed */
1010 do_migrate_pages(mwork
->mm
, &mwork
->from
, &mwork
->to
, MPOL_MF_MOVE_ALL
);
1015 static void cpuset_migrate_mm(struct mm_struct
*mm
, const nodemask_t
*from
,
1016 const nodemask_t
*to
)
1018 struct cpuset_migrate_mm_work
*mwork
;
1020 mwork
= kzalloc(sizeof(*mwork
), GFP_KERNEL
);
1023 mwork
->from
= *from
;
1025 INIT_WORK(&mwork
->work
, cpuset_migrate_mm_workfn
);
1026 queue_work(cpuset_migrate_mm_wq
, &mwork
->work
);
1032 static void cpuset_post_attach(void)
1034 flush_workqueue(cpuset_migrate_mm_wq
);
1038 * cpuset_change_task_nodemask - change task's mems_allowed and mempolicy
1039 * @tsk: the task to change
1040 * @newmems: new nodes that the task will be set
1042 * We use the mems_allowed_seq seqlock to safely update both tsk->mems_allowed
1043 * and rebind an eventual tasks' mempolicy. If the task is allocating in
1044 * parallel, it might temporarily see an empty intersection, which results in
1045 * a seqlock check and retry before OOM or allocation failure.
1047 static void cpuset_change_task_nodemask(struct task_struct
*tsk
,
1048 nodemask_t
*newmems
)
1052 local_irq_disable();
1053 write_seqcount_begin(&tsk
->mems_allowed_seq
);
1055 nodes_or(tsk
->mems_allowed
, tsk
->mems_allowed
, *newmems
);
1056 mpol_rebind_task(tsk
, newmems
);
1057 tsk
->mems_allowed
= *newmems
;
1059 write_seqcount_end(&tsk
->mems_allowed_seq
);
1065 static void *cpuset_being_rebound
;
1068 * update_tasks_nodemask - Update the nodemasks of tasks in the cpuset.
1069 * @cs: the cpuset in which each task's mems_allowed mask needs to be changed
1071 * Iterate through each task of @cs updating its mems_allowed to the
1072 * effective cpuset's. As this function is called with cpuset_mutex held,
1073 * cpuset membership stays stable.
1075 static void update_tasks_nodemask(struct cpuset
*cs
)
1077 static nodemask_t newmems
; /* protected by cpuset_mutex */
1078 struct css_task_iter it
;
1079 struct task_struct
*task
;
1081 cpuset_being_rebound
= cs
; /* causes mpol_dup() rebind */
1083 guarantee_online_mems(cs
, &newmems
);
1086 * The mpol_rebind_mm() call takes mmap_sem, which we couldn't
1087 * take while holding tasklist_lock. Forks can happen - the
1088 * mpol_dup() cpuset_being_rebound check will catch such forks,
1089 * and rebind their vma mempolicies too. Because we still hold
1090 * the global cpuset_mutex, we know that no other rebind effort
1091 * will be contending for the global variable cpuset_being_rebound.
1092 * It's ok if we rebind the same mm twice; mpol_rebind_mm()
1093 * is idempotent. Also migrate pages in each mm to new nodes.
1095 css_task_iter_start(&cs
->css
, &it
);
1096 while ((task
= css_task_iter_next(&it
))) {
1097 struct mm_struct
*mm
;
1100 cpuset_change_task_nodemask(task
, &newmems
);
1102 mm
= get_task_mm(task
);
1106 migrate
= is_memory_migrate(cs
);
1108 mpol_rebind_mm(mm
, &cs
->mems_allowed
);
1110 cpuset_migrate_mm(mm
, &cs
->old_mems_allowed
, &newmems
);
1114 css_task_iter_end(&it
);
1117 * All the tasks' nodemasks have been updated, update
1118 * cs->old_mems_allowed.
1120 cs
->old_mems_allowed
= newmems
;
1122 /* We're done rebinding vmas to this cpuset's new mems_allowed. */
1123 cpuset_being_rebound
= NULL
;
1127 * update_nodemasks_hier - Update effective nodemasks and tasks in the subtree
1128 * @cs: the cpuset to consider
1129 * @new_mems: a temp variable for calculating new effective_mems
1131 * When configured nodemask is changed, the effective nodemasks of this cpuset
1132 * and all its descendants need to be updated.
1134 * On legacy hiearchy, effective_mems will be the same with mems_allowed.
1136 * Called with cpuset_mutex held
1138 static void update_nodemasks_hier(struct cpuset
*cs
, nodemask_t
*new_mems
)
1141 struct cgroup_subsys_state
*pos_css
;
1144 cpuset_for_each_descendant_pre(cp
, pos_css
, cs
) {
1145 struct cpuset
*parent
= parent_cs(cp
);
1147 nodes_and(*new_mems
, cp
->mems_allowed
, parent
->effective_mems
);
1150 * If it becomes empty, inherit the effective mask of the
1151 * parent, which is guaranteed to have some MEMs.
1153 if (cgroup_subsys_on_dfl(cpuset_cgrp_subsys
) &&
1154 nodes_empty(*new_mems
))
1155 *new_mems
= parent
->effective_mems
;
1157 /* Skip the whole subtree if the nodemask remains the same. */
1158 if (nodes_equal(*new_mems
, cp
->effective_mems
)) {
1159 pos_css
= css_rightmost_descendant(pos_css
);
1163 if (!css_tryget_online(&cp
->css
))
1167 spin_lock_irq(&callback_lock
);
1168 cp
->effective_mems
= *new_mems
;
1169 spin_unlock_irq(&callback_lock
);
1171 WARN_ON(!cgroup_subsys_on_dfl(cpuset_cgrp_subsys
) &&
1172 !nodes_equal(cp
->mems_allowed
, cp
->effective_mems
));
1174 update_tasks_nodemask(cp
);
1183 * Handle user request to change the 'mems' memory placement
1184 * of a cpuset. Needs to validate the request, update the
1185 * cpusets mems_allowed, and for each task in the cpuset,
1186 * update mems_allowed and rebind task's mempolicy and any vma
1187 * mempolicies and if the cpuset is marked 'memory_migrate',
1188 * migrate the tasks pages to the new memory.
1190 * Call with cpuset_mutex held. May take callback_lock during call.
1191 * Will take tasklist_lock, scan tasklist for tasks in cpuset cs,
1192 * lock each such tasks mm->mmap_sem, scan its vma's and rebind
1193 * their mempolicies to the cpusets new mems_allowed.
1195 static int update_nodemask(struct cpuset
*cs
, struct cpuset
*trialcs
,
1201 * top_cpuset.mems_allowed tracks node_stats[N_MEMORY];
1204 if (cs
== &top_cpuset
) {
1210 * An empty mems_allowed is ok iff there are no tasks in the cpuset.
1211 * Since nodelist_parse() fails on an empty mask, we special case
1212 * that parsing. The validate_change() call ensures that cpusets
1213 * with tasks have memory.
1216 nodes_clear(trialcs
->mems_allowed
);
1218 retval
= nodelist_parse(buf
, trialcs
->mems_allowed
);
1222 if (!nodes_subset(trialcs
->mems_allowed
,
1223 top_cpuset
.mems_allowed
)) {
1229 if (nodes_equal(cs
->mems_allowed
, trialcs
->mems_allowed
)) {
1230 retval
= 0; /* Too easy - nothing to do */
1233 retval
= validate_change(cs
, trialcs
);
1237 spin_lock_irq(&callback_lock
);
1238 cs
->mems_allowed
= trialcs
->mems_allowed
;
1239 spin_unlock_irq(&callback_lock
);
1241 /* use trialcs->mems_allowed as a temp variable */
1242 update_nodemasks_hier(cs
, &trialcs
->mems_allowed
);
1247 int current_cpuset_is_being_rebound(void)
1252 ret
= task_cs(current
) == cpuset_being_rebound
;
1258 static int update_relax_domain_level(struct cpuset
*cs
, s64 val
)
1261 if (val
< -1 || val
>= sched_domain_level_max
)
1265 if (val
!= cs
->relax_domain_level
) {
1266 cs
->relax_domain_level
= val
;
1267 if (!cpumask_empty(cs
->cpus_allowed
) &&
1268 is_sched_load_balance(cs
))
1269 rebuild_sched_domains_locked();
1276 * update_tasks_flags - update the spread flags of tasks in the cpuset.
1277 * @cs: the cpuset in which each task's spread flags needs to be changed
1279 * Iterate through each task of @cs updating its spread flags. As this
1280 * function is called with cpuset_mutex held, cpuset membership stays
1283 static void update_tasks_flags(struct cpuset
*cs
)
1285 struct css_task_iter it
;
1286 struct task_struct
*task
;
1288 css_task_iter_start(&cs
->css
, &it
);
1289 while ((task
= css_task_iter_next(&it
)))
1290 cpuset_update_task_spread_flag(cs
, task
);
1291 css_task_iter_end(&it
);
1295 * update_flag - read a 0 or a 1 in a file and update associated flag
1296 * bit: the bit to update (see cpuset_flagbits_t)
1297 * cs: the cpuset to update
1298 * turning_on: whether the flag is being set or cleared
1300 * Call with cpuset_mutex held.
1303 static int update_flag(cpuset_flagbits_t bit
, struct cpuset
*cs
,
1306 struct cpuset
*trialcs
;
1307 int balance_flag_changed
;
1308 int spread_flag_changed
;
1311 trialcs
= alloc_trial_cpuset(cs
);
1316 set_bit(bit
, &trialcs
->flags
);
1318 clear_bit(bit
, &trialcs
->flags
);
1320 err
= validate_change(cs
, trialcs
);
1324 balance_flag_changed
= (is_sched_load_balance(cs
) !=
1325 is_sched_load_balance(trialcs
));
1327 spread_flag_changed
= ((is_spread_slab(cs
) != is_spread_slab(trialcs
))
1328 || (is_spread_page(cs
) != is_spread_page(trialcs
)));
1330 spin_lock_irq(&callback_lock
);
1331 cs
->flags
= trialcs
->flags
;
1332 spin_unlock_irq(&callback_lock
);
1334 if (!cpumask_empty(trialcs
->cpus_allowed
) && balance_flag_changed
)
1335 rebuild_sched_domains_locked();
1337 if (spread_flag_changed
)
1338 update_tasks_flags(cs
);
1340 free_trial_cpuset(trialcs
);
1345 * Frequency meter - How fast is some event occurring?
1347 * These routines manage a digitally filtered, constant time based,
1348 * event frequency meter. There are four routines:
1349 * fmeter_init() - initialize a frequency meter.
1350 * fmeter_markevent() - called each time the event happens.
1351 * fmeter_getrate() - returns the recent rate of such events.
1352 * fmeter_update() - internal routine used to update fmeter.
1354 * A common data structure is passed to each of these routines,
1355 * which is used to keep track of the state required to manage the
1356 * frequency meter and its digital filter.
1358 * The filter works on the number of events marked per unit time.
1359 * The filter is single-pole low-pass recursive (IIR). The time unit
1360 * is 1 second. Arithmetic is done using 32-bit integers scaled to
1361 * simulate 3 decimal digits of precision (multiplied by 1000).
1363 * With an FM_COEF of 933, and a time base of 1 second, the filter
1364 * has a half-life of 10 seconds, meaning that if the events quit
1365 * happening, then the rate returned from the fmeter_getrate()
1366 * will be cut in half each 10 seconds, until it converges to zero.
1368 * It is not worth doing a real infinitely recursive filter. If more
1369 * than FM_MAXTICKS ticks have elapsed since the last filter event,
1370 * just compute FM_MAXTICKS ticks worth, by which point the level
1373 * Limit the count of unprocessed events to FM_MAXCNT, so as to avoid
1374 * arithmetic overflow in the fmeter_update() routine.
1376 * Given the simple 32 bit integer arithmetic used, this meter works
1377 * best for reporting rates between one per millisecond (msec) and
1378 * one per 32 (approx) seconds. At constant rates faster than one
1379 * per msec it maxes out at values just under 1,000,000. At constant
1380 * rates between one per msec, and one per second it will stabilize
1381 * to a value N*1000, where N is the rate of events per second.
1382 * At constant rates between one per second and one per 32 seconds,
1383 * it will be choppy, moving up on the seconds that have an event,
1384 * and then decaying until the next event. At rates slower than
1385 * about one in 32 seconds, it decays all the way back to zero between
1389 #define FM_COEF 933 /* coefficient for half-life of 10 secs */
1390 #define FM_MAXTICKS ((u32)99) /* useless computing more ticks than this */
1391 #define FM_MAXCNT 1000000 /* limit cnt to avoid overflow */
1392 #define FM_SCALE 1000 /* faux fixed point scale */
1394 /* Initialize a frequency meter */
1395 static void fmeter_init(struct fmeter
*fmp
)
1400 spin_lock_init(&fmp
->lock
);
1403 /* Internal meter update - process cnt events and update value */
1404 static void fmeter_update(struct fmeter
*fmp
)
1409 now
= ktime_get_seconds();
1410 ticks
= now
- fmp
->time
;
1415 ticks
= min(FM_MAXTICKS
, ticks
);
1417 fmp
->val
= (FM_COEF
* fmp
->val
) / FM_SCALE
;
1420 fmp
->val
+= ((FM_SCALE
- FM_COEF
) * fmp
->cnt
) / FM_SCALE
;
1424 /* Process any previous ticks, then bump cnt by one (times scale). */
1425 static void fmeter_markevent(struct fmeter
*fmp
)
1427 spin_lock(&fmp
->lock
);
1429 fmp
->cnt
= min(FM_MAXCNT
, fmp
->cnt
+ FM_SCALE
);
1430 spin_unlock(&fmp
->lock
);
1433 /* Process any previous ticks, then return current value. */
1434 static int fmeter_getrate(struct fmeter
*fmp
)
1438 spin_lock(&fmp
->lock
);
1441 spin_unlock(&fmp
->lock
);
1445 static struct cpuset
*cpuset_attach_old_cs
;
1447 /* Called by cgroups to determine if a cpuset is usable; cpuset_mutex held */
1448 static int cpuset_can_attach(struct cgroup_taskset
*tset
)
1450 struct cgroup_subsys_state
*css
;
1452 struct task_struct
*task
;
1455 /* used later by cpuset_attach() */
1456 cpuset_attach_old_cs
= task_cs(cgroup_taskset_first(tset
, &css
));
1459 mutex_lock(&cpuset_mutex
);
1461 /* allow moving tasks into an empty cpuset if on default hierarchy */
1463 if (!cgroup_subsys_on_dfl(cpuset_cgrp_subsys
) &&
1464 (cpumask_empty(cs
->cpus_allowed
) || nodes_empty(cs
->mems_allowed
)))
1467 cgroup_taskset_for_each(task
, css
, tset
) {
1468 ret
= task_can_attach(task
, cs
->cpus_allowed
);
1471 ret
= security_task_setscheduler(task
);
1477 * Mark attach is in progress. This makes validate_change() fail
1478 * changes which zero cpus/mems_allowed.
1480 cs
->attach_in_progress
++;
1483 mutex_unlock(&cpuset_mutex
);
1487 static void cpuset_cancel_attach(struct cgroup_taskset
*tset
)
1489 struct cgroup_subsys_state
*css
;
1492 cgroup_taskset_first(tset
, &css
);
1495 mutex_lock(&cpuset_mutex
);
1496 css_cs(css
)->attach_in_progress
--;
1497 mutex_unlock(&cpuset_mutex
);
1501 * Protected by cpuset_mutex. cpus_attach is used only by cpuset_attach()
1502 * but we can't allocate it dynamically there. Define it global and
1503 * allocate from cpuset_init().
1505 static cpumask_var_t cpus_attach
;
1507 static void cpuset_attach(struct cgroup_taskset
*tset
)
1509 /* static buf protected by cpuset_mutex */
1510 static nodemask_t cpuset_attach_nodemask_to
;
1511 struct task_struct
*task
;
1512 struct task_struct
*leader
;
1513 struct cgroup_subsys_state
*css
;
1515 struct cpuset
*oldcs
= cpuset_attach_old_cs
;
1517 cgroup_taskset_first(tset
, &css
);
1520 mutex_lock(&cpuset_mutex
);
1522 /* prepare for attach */
1523 if (cs
== &top_cpuset
)
1524 cpumask_copy(cpus_attach
, cpu_possible_mask
);
1526 guarantee_online_cpus(cs
, cpus_attach
);
1528 guarantee_online_mems(cs
, &cpuset_attach_nodemask_to
);
1530 cgroup_taskset_for_each(task
, css
, tset
) {
1532 * can_attach beforehand should guarantee that this doesn't
1533 * fail. TODO: have a better way to handle failure here
1535 WARN_ON_ONCE(set_cpus_allowed_ptr(task
, cpus_attach
));
1537 cpuset_change_task_nodemask(task
, &cpuset_attach_nodemask_to
);
1538 cpuset_update_task_spread_flag(cs
, task
);
1542 * Change mm for all threadgroup leaders. This is expensive and may
1543 * sleep and should be moved outside migration path proper.
1545 cpuset_attach_nodemask_to
= cs
->effective_mems
;
1546 cgroup_taskset_for_each_leader(leader
, css
, tset
) {
1547 struct mm_struct
*mm
= get_task_mm(leader
);
1550 mpol_rebind_mm(mm
, &cpuset_attach_nodemask_to
);
1553 * old_mems_allowed is the same with mems_allowed
1554 * here, except if this task is being moved
1555 * automatically due to hotplug. In that case
1556 * @mems_allowed has been updated and is empty, so
1557 * @old_mems_allowed is the right nodesets that we
1560 if (is_memory_migrate(cs
))
1561 cpuset_migrate_mm(mm
, &oldcs
->old_mems_allowed
,
1562 &cpuset_attach_nodemask_to
);
1568 cs
->old_mems_allowed
= cpuset_attach_nodemask_to
;
1570 cs
->attach_in_progress
--;
1571 if (!cs
->attach_in_progress
)
1572 wake_up(&cpuset_attach_wq
);
1574 mutex_unlock(&cpuset_mutex
);
1577 /* The various types of files and directories in a cpuset file system */
1580 FILE_MEMORY_MIGRATE
,
1583 FILE_EFFECTIVE_CPULIST
,
1584 FILE_EFFECTIVE_MEMLIST
,
1588 FILE_SCHED_LOAD_BALANCE
,
1589 FILE_SCHED_RELAX_DOMAIN_LEVEL
,
1590 FILE_MEMORY_PRESSURE_ENABLED
,
1591 FILE_MEMORY_PRESSURE
,
1594 } cpuset_filetype_t
;
1596 static int cpuset_write_u64(struct cgroup_subsys_state
*css
, struct cftype
*cft
,
1599 struct cpuset
*cs
= css_cs(css
);
1600 cpuset_filetype_t type
= cft
->private;
1603 mutex_lock(&cpuset_mutex
);
1604 if (!is_cpuset_online(cs
)) {
1610 case FILE_CPU_EXCLUSIVE
:
1611 retval
= update_flag(CS_CPU_EXCLUSIVE
, cs
, val
);
1613 case FILE_MEM_EXCLUSIVE
:
1614 retval
= update_flag(CS_MEM_EXCLUSIVE
, cs
, val
);
1616 case FILE_MEM_HARDWALL
:
1617 retval
= update_flag(CS_MEM_HARDWALL
, cs
, val
);
1619 case FILE_SCHED_LOAD_BALANCE
:
1620 retval
= update_flag(CS_SCHED_LOAD_BALANCE
, cs
, val
);
1622 case FILE_MEMORY_MIGRATE
:
1623 retval
= update_flag(CS_MEMORY_MIGRATE
, cs
, val
);
1625 case FILE_MEMORY_PRESSURE_ENABLED
:
1626 cpuset_memory_pressure_enabled
= !!val
;
1628 case FILE_SPREAD_PAGE
:
1629 retval
= update_flag(CS_SPREAD_PAGE
, cs
, val
);
1631 case FILE_SPREAD_SLAB
:
1632 retval
= update_flag(CS_SPREAD_SLAB
, cs
, val
);
1639 mutex_unlock(&cpuset_mutex
);
1643 static int cpuset_write_s64(struct cgroup_subsys_state
*css
, struct cftype
*cft
,
1646 struct cpuset
*cs
= css_cs(css
);
1647 cpuset_filetype_t type
= cft
->private;
1648 int retval
= -ENODEV
;
1650 mutex_lock(&cpuset_mutex
);
1651 if (!is_cpuset_online(cs
))
1655 case FILE_SCHED_RELAX_DOMAIN_LEVEL
:
1656 retval
= update_relax_domain_level(cs
, val
);
1663 mutex_unlock(&cpuset_mutex
);
1668 * Common handling for a write to a "cpus" or "mems" file.
1670 static ssize_t
cpuset_write_resmask(struct kernfs_open_file
*of
,
1671 char *buf
, size_t nbytes
, loff_t off
)
1673 struct cpuset
*cs
= css_cs(of_css(of
));
1674 struct cpuset
*trialcs
;
1675 int retval
= -ENODEV
;
1677 buf
= strstrip(buf
);
1680 * CPU or memory hotunplug may leave @cs w/o any execution
1681 * resources, in which case the hotplug code asynchronously updates
1682 * configuration and transfers all tasks to the nearest ancestor
1683 * which can execute.
1685 * As writes to "cpus" or "mems" may restore @cs's execution
1686 * resources, wait for the previously scheduled operations before
1687 * proceeding, so that we don't end up keep removing tasks added
1688 * after execution capability is restored.
1690 * cpuset_hotplug_work calls back into cgroup core via
1691 * cgroup_transfer_tasks() and waiting for it from a cgroupfs
1692 * operation like this one can lead to a deadlock through kernfs
1693 * active_ref protection. Let's break the protection. Losing the
1694 * protection is okay as we check whether @cs is online after
1695 * grabbing cpuset_mutex anyway. This only happens on the legacy
1699 kernfs_break_active_protection(of
->kn
);
1700 flush_work(&cpuset_hotplug_work
);
1702 mutex_lock(&cpuset_mutex
);
1703 if (!is_cpuset_online(cs
))
1706 trialcs
= alloc_trial_cpuset(cs
);
1712 switch (of_cft(of
)->private) {
1714 retval
= update_cpumask(cs
, trialcs
, buf
);
1717 retval
= update_nodemask(cs
, trialcs
, buf
);
1724 free_trial_cpuset(trialcs
);
1726 mutex_unlock(&cpuset_mutex
);
1727 kernfs_unbreak_active_protection(of
->kn
);
1729 flush_workqueue(cpuset_migrate_mm_wq
);
1730 return retval
?: nbytes
;
1734 * These ascii lists should be read in a single call, by using a user
1735 * buffer large enough to hold the entire map. If read in smaller
1736 * chunks, there is no guarantee of atomicity. Since the display format
1737 * used, list of ranges of sequential numbers, is variable length,
1738 * and since these maps can change value dynamically, one could read
1739 * gibberish by doing partial reads while a list was changing.
1741 static int cpuset_common_seq_show(struct seq_file
*sf
, void *v
)
1743 struct cpuset
*cs
= css_cs(seq_css(sf
));
1744 cpuset_filetype_t type
= seq_cft(sf
)->private;
1747 spin_lock_irq(&callback_lock
);
1751 seq_printf(sf
, "%*pbl\n", cpumask_pr_args(cs
->cpus_allowed
));
1754 seq_printf(sf
, "%*pbl\n", nodemask_pr_args(&cs
->mems_allowed
));
1756 case FILE_EFFECTIVE_CPULIST
:
1757 seq_printf(sf
, "%*pbl\n", cpumask_pr_args(cs
->effective_cpus
));
1759 case FILE_EFFECTIVE_MEMLIST
:
1760 seq_printf(sf
, "%*pbl\n", nodemask_pr_args(&cs
->effective_mems
));
1766 spin_unlock_irq(&callback_lock
);
1770 static u64
cpuset_read_u64(struct cgroup_subsys_state
*css
, struct cftype
*cft
)
1772 struct cpuset
*cs
= css_cs(css
);
1773 cpuset_filetype_t type
= cft
->private;
1775 case FILE_CPU_EXCLUSIVE
:
1776 return is_cpu_exclusive(cs
);
1777 case FILE_MEM_EXCLUSIVE
:
1778 return is_mem_exclusive(cs
);
1779 case FILE_MEM_HARDWALL
:
1780 return is_mem_hardwall(cs
);
1781 case FILE_SCHED_LOAD_BALANCE
:
1782 return is_sched_load_balance(cs
);
1783 case FILE_MEMORY_MIGRATE
:
1784 return is_memory_migrate(cs
);
1785 case FILE_MEMORY_PRESSURE_ENABLED
:
1786 return cpuset_memory_pressure_enabled
;
1787 case FILE_MEMORY_PRESSURE
:
1788 return fmeter_getrate(&cs
->fmeter
);
1789 case FILE_SPREAD_PAGE
:
1790 return is_spread_page(cs
);
1791 case FILE_SPREAD_SLAB
:
1792 return is_spread_slab(cs
);
1797 /* Unreachable but makes gcc happy */
1801 static s64
cpuset_read_s64(struct cgroup_subsys_state
*css
, struct cftype
*cft
)
1803 struct cpuset
*cs
= css_cs(css
);
1804 cpuset_filetype_t type
= cft
->private;
1806 case FILE_SCHED_RELAX_DOMAIN_LEVEL
:
1807 return cs
->relax_domain_level
;
1812 /* Unrechable but makes gcc happy */
1818 * for the common functions, 'private' gives the type of file
1821 static struct cftype files
[] = {
1824 .seq_show
= cpuset_common_seq_show
,
1825 .write
= cpuset_write_resmask
,
1826 .max_write_len
= (100U + 6 * NR_CPUS
),
1827 .private = FILE_CPULIST
,
1832 .seq_show
= cpuset_common_seq_show
,
1833 .write
= cpuset_write_resmask
,
1834 .max_write_len
= (100U + 6 * MAX_NUMNODES
),
1835 .private = FILE_MEMLIST
,
1839 .name
= "effective_cpus",
1840 .seq_show
= cpuset_common_seq_show
,
1841 .private = FILE_EFFECTIVE_CPULIST
,
1845 .name
= "effective_mems",
1846 .seq_show
= cpuset_common_seq_show
,
1847 .private = FILE_EFFECTIVE_MEMLIST
,
1851 .name
= "cpu_exclusive",
1852 .read_u64
= cpuset_read_u64
,
1853 .write_u64
= cpuset_write_u64
,
1854 .private = FILE_CPU_EXCLUSIVE
,
1858 .name
= "mem_exclusive",
1859 .read_u64
= cpuset_read_u64
,
1860 .write_u64
= cpuset_write_u64
,
1861 .private = FILE_MEM_EXCLUSIVE
,
1865 .name
= "mem_hardwall",
1866 .read_u64
= cpuset_read_u64
,
1867 .write_u64
= cpuset_write_u64
,
1868 .private = FILE_MEM_HARDWALL
,
1872 .name
= "sched_load_balance",
1873 .read_u64
= cpuset_read_u64
,
1874 .write_u64
= cpuset_write_u64
,
1875 .private = FILE_SCHED_LOAD_BALANCE
,
1879 .name
= "sched_relax_domain_level",
1880 .read_s64
= cpuset_read_s64
,
1881 .write_s64
= cpuset_write_s64
,
1882 .private = FILE_SCHED_RELAX_DOMAIN_LEVEL
,
1886 .name
= "memory_migrate",
1887 .read_u64
= cpuset_read_u64
,
1888 .write_u64
= cpuset_write_u64
,
1889 .private = FILE_MEMORY_MIGRATE
,
1893 .name
= "memory_pressure",
1894 .read_u64
= cpuset_read_u64
,
1895 .private = FILE_MEMORY_PRESSURE
,
1899 .name
= "memory_spread_page",
1900 .read_u64
= cpuset_read_u64
,
1901 .write_u64
= cpuset_write_u64
,
1902 .private = FILE_SPREAD_PAGE
,
1906 .name
= "memory_spread_slab",
1907 .read_u64
= cpuset_read_u64
,
1908 .write_u64
= cpuset_write_u64
,
1909 .private = FILE_SPREAD_SLAB
,
1913 .name
= "memory_pressure_enabled",
1914 .flags
= CFTYPE_ONLY_ON_ROOT
,
1915 .read_u64
= cpuset_read_u64
,
1916 .write_u64
= cpuset_write_u64
,
1917 .private = FILE_MEMORY_PRESSURE_ENABLED
,
1924 * cpuset_css_alloc - allocate a cpuset css
1925 * cgrp: control group that the new cpuset will be part of
1928 static struct cgroup_subsys_state
*
1929 cpuset_css_alloc(struct cgroup_subsys_state
*parent_css
)
1934 return &top_cpuset
.css
;
1936 cs
= kzalloc(sizeof(*cs
), GFP_KERNEL
);
1938 return ERR_PTR(-ENOMEM
);
1939 if (!alloc_cpumask_var(&cs
->cpus_allowed
, GFP_KERNEL
))
1941 if (!alloc_cpumask_var(&cs
->effective_cpus
, GFP_KERNEL
))
1944 set_bit(CS_SCHED_LOAD_BALANCE
, &cs
->flags
);
1945 cpumask_clear(cs
->cpus_allowed
);
1946 nodes_clear(cs
->mems_allowed
);
1947 cpumask_clear(cs
->effective_cpus
);
1948 nodes_clear(cs
->effective_mems
);
1949 fmeter_init(&cs
->fmeter
);
1950 cs
->relax_domain_level
= -1;
1955 free_cpumask_var(cs
->cpus_allowed
);
1958 return ERR_PTR(-ENOMEM
);
1961 static int cpuset_css_online(struct cgroup_subsys_state
*css
)
1963 struct cpuset
*cs
= css_cs(css
);
1964 struct cpuset
*parent
= parent_cs(cs
);
1965 struct cpuset
*tmp_cs
;
1966 struct cgroup_subsys_state
*pos_css
;
1971 mutex_lock(&cpuset_mutex
);
1973 set_bit(CS_ONLINE
, &cs
->flags
);
1974 if (is_spread_page(parent
))
1975 set_bit(CS_SPREAD_PAGE
, &cs
->flags
);
1976 if (is_spread_slab(parent
))
1977 set_bit(CS_SPREAD_SLAB
, &cs
->flags
);
1981 spin_lock_irq(&callback_lock
);
1982 if (cgroup_subsys_on_dfl(cpuset_cgrp_subsys
)) {
1983 cpumask_copy(cs
->effective_cpus
, parent
->effective_cpus
);
1984 cs
->effective_mems
= parent
->effective_mems
;
1986 spin_unlock_irq(&callback_lock
);
1988 if (!test_bit(CGRP_CPUSET_CLONE_CHILDREN
, &css
->cgroup
->flags
))
1992 * Clone @parent's configuration if CGRP_CPUSET_CLONE_CHILDREN is
1993 * set. This flag handling is implemented in cgroup core for
1994 * histrical reasons - the flag may be specified during mount.
1996 * Currently, if any sibling cpusets have exclusive cpus or mem, we
1997 * refuse to clone the configuration - thereby refusing the task to
1998 * be entered, and as a result refusing the sys_unshare() or
1999 * clone() which initiated it. If this becomes a problem for some
2000 * users who wish to allow that scenario, then this could be
2001 * changed to grant parent->cpus_allowed-sibling_cpus_exclusive
2002 * (and likewise for mems) to the new cgroup.
2005 cpuset_for_each_child(tmp_cs
, pos_css
, parent
) {
2006 if (is_mem_exclusive(tmp_cs
) || is_cpu_exclusive(tmp_cs
)) {
2013 spin_lock_irq(&callback_lock
);
2014 cs
->mems_allowed
= parent
->mems_allowed
;
2015 cs
->effective_mems
= parent
->mems_allowed
;
2016 cpumask_copy(cs
->cpus_allowed
, parent
->cpus_allowed
);
2017 cpumask_copy(cs
->effective_cpus
, parent
->cpus_allowed
);
2018 spin_unlock_irq(&callback_lock
);
2020 mutex_unlock(&cpuset_mutex
);
2025 * If the cpuset being removed has its flag 'sched_load_balance'
2026 * enabled, then simulate turning sched_load_balance off, which
2027 * will call rebuild_sched_domains_locked().
2030 static void cpuset_css_offline(struct cgroup_subsys_state
*css
)
2032 struct cpuset
*cs
= css_cs(css
);
2034 mutex_lock(&cpuset_mutex
);
2036 if (is_sched_load_balance(cs
))
2037 update_flag(CS_SCHED_LOAD_BALANCE
, cs
, 0);
2040 clear_bit(CS_ONLINE
, &cs
->flags
);
2042 mutex_unlock(&cpuset_mutex
);
2045 static void cpuset_css_free(struct cgroup_subsys_state
*css
)
2047 struct cpuset
*cs
= css_cs(css
);
2049 free_cpumask_var(cs
->effective_cpus
);
2050 free_cpumask_var(cs
->cpus_allowed
);
2054 static void cpuset_bind(struct cgroup_subsys_state
*root_css
)
2056 mutex_lock(&cpuset_mutex
);
2057 spin_lock_irq(&callback_lock
);
2059 if (cgroup_subsys_on_dfl(cpuset_cgrp_subsys
)) {
2060 cpumask_copy(top_cpuset
.cpus_allowed
, cpu_possible_mask
);
2061 top_cpuset
.mems_allowed
= node_possible_map
;
2063 cpumask_copy(top_cpuset
.cpus_allowed
,
2064 top_cpuset
.effective_cpus
);
2065 top_cpuset
.mems_allowed
= top_cpuset
.effective_mems
;
2068 spin_unlock_irq(&callback_lock
);
2069 mutex_unlock(&cpuset_mutex
);
2073 * Make sure the new task conform to the current state of its parent,
2074 * which could have been changed by cpuset just after it inherits the
2075 * state from the parent and before it sits on the cgroup's task list.
2077 static void cpuset_fork(struct task_struct
*task
)
2079 if (task_css_is_root(task
, cpuset_cgrp_id
))
2082 set_cpus_allowed_ptr(task
, ¤t
->cpus_allowed
);
2083 task
->mems_allowed
= current
->mems_allowed
;
2086 struct cgroup_subsys cpuset_cgrp_subsys
= {
2087 .css_alloc
= cpuset_css_alloc
,
2088 .css_online
= cpuset_css_online
,
2089 .css_offline
= cpuset_css_offline
,
2090 .css_free
= cpuset_css_free
,
2091 .can_attach
= cpuset_can_attach
,
2092 .cancel_attach
= cpuset_cancel_attach
,
2093 .attach
= cpuset_attach
,
2094 .post_attach
= cpuset_post_attach
,
2095 .bind
= cpuset_bind
,
2096 .fork
= cpuset_fork
,
2097 .legacy_cftypes
= files
,
2102 * cpuset_init - initialize cpusets at system boot
2104 * Description: Initialize top_cpuset and the cpuset internal file system,
2107 int __init
cpuset_init(void)
2111 BUG_ON(!alloc_cpumask_var(&top_cpuset
.cpus_allowed
, GFP_KERNEL
));
2112 BUG_ON(!alloc_cpumask_var(&top_cpuset
.effective_cpus
, GFP_KERNEL
));
2114 cpumask_setall(top_cpuset
.cpus_allowed
);
2115 nodes_setall(top_cpuset
.mems_allowed
);
2116 cpumask_setall(top_cpuset
.effective_cpus
);
2117 nodes_setall(top_cpuset
.effective_mems
);
2119 fmeter_init(&top_cpuset
.fmeter
);
2120 set_bit(CS_SCHED_LOAD_BALANCE
, &top_cpuset
.flags
);
2121 top_cpuset
.relax_domain_level
= -1;
2123 err
= register_filesystem(&cpuset_fs_type
);
2127 BUG_ON(!alloc_cpumask_var(&cpus_attach
, GFP_KERNEL
));
2133 * If CPU and/or memory hotplug handlers, below, unplug any CPUs
2134 * or memory nodes, we need to walk over the cpuset hierarchy,
2135 * removing that CPU or node from all cpusets. If this removes the
2136 * last CPU or node from a cpuset, then move the tasks in the empty
2137 * cpuset to its next-highest non-empty parent.
2139 static void remove_tasks_in_empty_cpuset(struct cpuset
*cs
)
2141 struct cpuset
*parent
;
2144 * Find its next-highest non-empty parent, (top cpuset
2145 * has online cpus, so can't be empty).
2147 parent
= parent_cs(cs
);
2148 while (cpumask_empty(parent
->cpus_allowed
) ||
2149 nodes_empty(parent
->mems_allowed
))
2150 parent
= parent_cs(parent
);
2152 if (cgroup_transfer_tasks(parent
->css
.cgroup
, cs
->css
.cgroup
)) {
2153 pr_err("cpuset: failed to transfer tasks out of empty cpuset ");
2154 pr_cont_cgroup_name(cs
->css
.cgroup
);
2160 hotplug_update_tasks_legacy(struct cpuset
*cs
,
2161 struct cpumask
*new_cpus
, nodemask_t
*new_mems
,
2162 bool cpus_updated
, bool mems_updated
)
2166 spin_lock_irq(&callback_lock
);
2167 cpumask_copy(cs
->cpus_allowed
, new_cpus
);
2168 cpumask_copy(cs
->effective_cpus
, new_cpus
);
2169 cs
->mems_allowed
= *new_mems
;
2170 cs
->effective_mems
= *new_mems
;
2171 spin_unlock_irq(&callback_lock
);
2174 * Don't call update_tasks_cpumask() if the cpuset becomes empty,
2175 * as the tasks will be migratecd to an ancestor.
2177 if (cpus_updated
&& !cpumask_empty(cs
->cpus_allowed
))
2178 update_tasks_cpumask(cs
);
2179 if (mems_updated
&& !nodes_empty(cs
->mems_allowed
))
2180 update_tasks_nodemask(cs
);
2182 is_empty
= cpumask_empty(cs
->cpus_allowed
) ||
2183 nodes_empty(cs
->mems_allowed
);
2185 mutex_unlock(&cpuset_mutex
);
2188 * Move tasks to the nearest ancestor with execution resources,
2189 * This is full cgroup operation which will also call back into
2190 * cpuset. Should be done outside any lock.
2193 remove_tasks_in_empty_cpuset(cs
);
2195 mutex_lock(&cpuset_mutex
);
2199 hotplug_update_tasks(struct cpuset
*cs
,
2200 struct cpumask
*new_cpus
, nodemask_t
*new_mems
,
2201 bool cpus_updated
, bool mems_updated
)
2203 if (cpumask_empty(new_cpus
))
2204 cpumask_copy(new_cpus
, parent_cs(cs
)->effective_cpus
);
2205 if (nodes_empty(*new_mems
))
2206 *new_mems
= parent_cs(cs
)->effective_mems
;
2208 spin_lock_irq(&callback_lock
);
2209 cpumask_copy(cs
->effective_cpus
, new_cpus
);
2210 cs
->effective_mems
= *new_mems
;
2211 spin_unlock_irq(&callback_lock
);
2214 update_tasks_cpumask(cs
);
2216 update_tasks_nodemask(cs
);
2220 * cpuset_hotplug_update_tasks - update tasks in a cpuset for hotunplug
2221 * @cs: cpuset in interest
2223 * Compare @cs's cpu and mem masks against top_cpuset and if some have gone
2224 * offline, update @cs accordingly. If @cs ends up with no CPU or memory,
2225 * all its tasks are moved to the nearest ancestor with both resources.
2227 static void cpuset_hotplug_update_tasks(struct cpuset
*cs
)
2229 static cpumask_t new_cpus
;
2230 static nodemask_t new_mems
;
2234 wait_event(cpuset_attach_wq
, cs
->attach_in_progress
== 0);
2236 mutex_lock(&cpuset_mutex
);
2239 * We have raced with task attaching. We wait until attaching
2240 * is finished, so we won't attach a task to an empty cpuset.
2242 if (cs
->attach_in_progress
) {
2243 mutex_unlock(&cpuset_mutex
);
2247 cpumask_and(&new_cpus
, cs
->cpus_allowed
, parent_cs(cs
)->effective_cpus
);
2248 nodes_and(new_mems
, cs
->mems_allowed
, parent_cs(cs
)->effective_mems
);
2250 cpus_updated
= !cpumask_equal(&new_cpus
, cs
->effective_cpus
);
2251 mems_updated
= !nodes_equal(new_mems
, cs
->effective_mems
);
2253 if (cgroup_subsys_on_dfl(cpuset_cgrp_subsys
))
2254 hotplug_update_tasks(cs
, &new_cpus
, &new_mems
,
2255 cpus_updated
, mems_updated
);
2257 hotplug_update_tasks_legacy(cs
, &new_cpus
, &new_mems
,
2258 cpus_updated
, mems_updated
);
2260 mutex_unlock(&cpuset_mutex
);
2263 static bool force_rebuild
;
2265 void cpuset_force_rebuild(void)
2267 force_rebuild
= true;
2271 * cpuset_hotplug_workfn - handle CPU/memory hotunplug for a cpuset
2273 * This function is called after either CPU or memory configuration has
2274 * changed and updates cpuset accordingly. The top_cpuset is always
2275 * synchronized to cpu_active_mask and N_MEMORY, which is necessary in
2276 * order to make cpusets transparent (of no affect) on systems that are
2277 * actively using CPU hotplug but making no active use of cpusets.
2279 * Non-root cpusets are only affected by offlining. If any CPUs or memory
2280 * nodes have been taken down, cpuset_hotplug_update_tasks() is invoked on
2283 * Note that CPU offlining during suspend is ignored. We don't modify
2284 * cpusets across suspend/resume cycles at all.
2286 static void cpuset_hotplug_workfn(struct work_struct
*work
)
2288 static cpumask_t new_cpus
;
2289 static nodemask_t new_mems
;
2290 bool cpus_updated
, mems_updated
;
2291 bool on_dfl
= cgroup_subsys_on_dfl(cpuset_cgrp_subsys
);
2293 mutex_lock(&cpuset_mutex
);
2295 /* fetch the available cpus/mems and find out which changed how */
2296 cpumask_copy(&new_cpus
, cpu_active_mask
);
2297 new_mems
= node_states
[N_MEMORY
];
2299 cpus_updated
= !cpumask_equal(top_cpuset
.effective_cpus
, &new_cpus
);
2300 mems_updated
= !nodes_equal(top_cpuset
.effective_mems
, new_mems
);
2302 /* synchronize cpus_allowed to cpu_active_mask */
2304 spin_lock_irq(&callback_lock
);
2306 cpumask_copy(top_cpuset
.cpus_allowed
, &new_cpus
);
2307 cpumask_copy(top_cpuset
.effective_cpus
, &new_cpus
);
2308 spin_unlock_irq(&callback_lock
);
2309 /* we don't mess with cpumasks of tasks in top_cpuset */
2312 /* synchronize mems_allowed to N_MEMORY */
2314 spin_lock_irq(&callback_lock
);
2316 top_cpuset
.mems_allowed
= new_mems
;
2317 top_cpuset
.effective_mems
= new_mems
;
2318 spin_unlock_irq(&callback_lock
);
2319 update_tasks_nodemask(&top_cpuset
);
2322 mutex_unlock(&cpuset_mutex
);
2324 /* if cpus or mems changed, we need to propagate to descendants */
2325 if (cpus_updated
|| mems_updated
) {
2327 struct cgroup_subsys_state
*pos_css
;
2330 cpuset_for_each_descendant_pre(cs
, pos_css
, &top_cpuset
) {
2331 if (cs
== &top_cpuset
|| !css_tryget_online(&cs
->css
))
2335 cpuset_hotplug_update_tasks(cs
);
2343 /* rebuild sched domains if cpus_allowed has changed */
2344 if (cpus_updated
|| force_rebuild
) {
2345 force_rebuild
= false;
2346 rebuild_sched_domains();
2350 void cpuset_update_active_cpus(void)
2353 * We're inside cpu hotplug critical region which usually nests
2354 * inside cgroup synchronization. Bounce actual hotplug processing
2355 * to a work item to avoid reverse locking order.
2357 * We still need to do partition_sched_domains() synchronously;
2358 * otherwise, the scheduler will get confused and put tasks to the
2359 * dead CPU. Fall back to the default single domain.
2360 * cpuset_hotplug_workfn() will rebuild it as necessary.
2362 partition_sched_domains(1, NULL
, NULL
);
2363 schedule_work(&cpuset_hotplug_work
);
2366 void cpuset_wait_for_hotplug(void)
2368 flush_work(&cpuset_hotplug_work
);
2372 * Keep top_cpuset.mems_allowed tracking node_states[N_MEMORY].
2373 * Call this routine anytime after node_states[N_MEMORY] changes.
2374 * See cpuset_update_active_cpus() for CPU hotplug handling.
2376 static int cpuset_track_online_nodes(struct notifier_block
*self
,
2377 unsigned long action
, void *arg
)
2379 schedule_work(&cpuset_hotplug_work
);
2383 static struct notifier_block cpuset_track_online_nodes_nb
= {
2384 .notifier_call
= cpuset_track_online_nodes
,
2385 .priority
= 10, /* ??! */
2389 * cpuset_init_smp - initialize cpus_allowed
2391 * Description: Finish top cpuset after cpu, node maps are initialized
2393 void __init
cpuset_init_smp(void)
2395 cpumask_copy(top_cpuset
.cpus_allowed
, cpu_active_mask
);
2396 top_cpuset
.mems_allowed
= node_states
[N_MEMORY
];
2397 top_cpuset
.old_mems_allowed
= top_cpuset
.mems_allowed
;
2399 cpumask_copy(top_cpuset
.effective_cpus
, cpu_active_mask
);
2400 top_cpuset
.effective_mems
= node_states
[N_MEMORY
];
2402 register_hotmemory_notifier(&cpuset_track_online_nodes_nb
);
2404 cpuset_migrate_mm_wq
= alloc_ordered_workqueue("cpuset_migrate_mm", 0);
2405 BUG_ON(!cpuset_migrate_mm_wq
);
2409 * cpuset_cpus_allowed - return cpus_allowed mask from a tasks cpuset.
2410 * @tsk: pointer to task_struct from which to obtain cpuset->cpus_allowed.
2411 * @pmask: pointer to struct cpumask variable to receive cpus_allowed set.
2413 * Description: Returns the cpumask_var_t cpus_allowed of the cpuset
2414 * attached to the specified @tsk. Guaranteed to return some non-empty
2415 * subset of cpu_online_mask, even if this means going outside the
2419 void cpuset_cpus_allowed(struct task_struct
*tsk
, struct cpumask
*pmask
)
2421 unsigned long flags
;
2423 spin_lock_irqsave(&callback_lock
, flags
);
2425 guarantee_online_cpus(task_cs(tsk
), pmask
);
2427 spin_unlock_irqrestore(&callback_lock
, flags
);
2430 void cpuset_cpus_allowed_fallback(struct task_struct
*tsk
)
2433 do_set_cpus_allowed(tsk
, task_cs(tsk
)->effective_cpus
);
2437 * We own tsk->cpus_allowed, nobody can change it under us.
2439 * But we used cs && cs->cpus_allowed lockless and thus can
2440 * race with cgroup_attach_task() or update_cpumask() and get
2441 * the wrong tsk->cpus_allowed. However, both cases imply the
2442 * subsequent cpuset_change_cpumask()->set_cpus_allowed_ptr()
2443 * which takes task_rq_lock().
2445 * If we are called after it dropped the lock we must see all
2446 * changes in tsk_cs()->cpus_allowed. Otherwise we can temporary
2447 * set any mask even if it is not right from task_cs() pov,
2448 * the pending set_cpus_allowed_ptr() will fix things.
2450 * select_fallback_rq() will fix things ups and set cpu_possible_mask
2455 void __init
cpuset_init_current_mems_allowed(void)
2457 nodes_setall(current
->mems_allowed
);
2461 * cpuset_mems_allowed - return mems_allowed mask from a tasks cpuset.
2462 * @tsk: pointer to task_struct from which to obtain cpuset->mems_allowed.
2464 * Description: Returns the nodemask_t mems_allowed of the cpuset
2465 * attached to the specified @tsk. Guaranteed to return some non-empty
2466 * subset of node_states[N_MEMORY], even if this means going outside the
2470 nodemask_t
cpuset_mems_allowed(struct task_struct
*tsk
)
2473 unsigned long flags
;
2475 spin_lock_irqsave(&callback_lock
, flags
);
2477 guarantee_online_mems(task_cs(tsk
), &mask
);
2479 spin_unlock_irqrestore(&callback_lock
, flags
);
2485 * cpuset_nodemask_valid_mems_allowed - check nodemask vs. curremt mems_allowed
2486 * @nodemask: the nodemask to be checked
2488 * Are any of the nodes in the nodemask allowed in current->mems_allowed?
2490 int cpuset_nodemask_valid_mems_allowed(nodemask_t
*nodemask
)
2492 return nodes_intersects(*nodemask
, current
->mems_allowed
);
2496 * nearest_hardwall_ancestor() - Returns the nearest mem_exclusive or
2497 * mem_hardwall ancestor to the specified cpuset. Call holding
2498 * callback_lock. If no ancestor is mem_exclusive or mem_hardwall
2499 * (an unusual configuration), then returns the root cpuset.
2501 static struct cpuset
*nearest_hardwall_ancestor(struct cpuset
*cs
)
2503 while (!(is_mem_exclusive(cs
) || is_mem_hardwall(cs
)) && parent_cs(cs
))
2509 * cpuset_node_allowed - Can we allocate on a memory node?
2510 * @node: is this an allowed node?
2511 * @gfp_mask: memory allocation flags
2513 * If we're in interrupt, yes, we can always allocate. If @node is set in
2514 * current's mems_allowed, yes. If it's not a __GFP_HARDWALL request and this
2515 * node is set in the nearest hardwalled cpuset ancestor to current's cpuset,
2516 * yes. If current has access to memory reserves due to TIF_MEMDIE, yes.
2519 * GFP_USER allocations are marked with the __GFP_HARDWALL bit,
2520 * and do not allow allocations outside the current tasks cpuset
2521 * unless the task has been OOM killed as is marked TIF_MEMDIE.
2522 * GFP_KERNEL allocations are not so marked, so can escape to the
2523 * nearest enclosing hardwalled ancestor cpuset.
2525 * Scanning up parent cpusets requires callback_lock. The
2526 * __alloc_pages() routine only calls here with __GFP_HARDWALL bit
2527 * _not_ set if it's a GFP_KERNEL allocation, and all nodes in the
2528 * current tasks mems_allowed came up empty on the first pass over
2529 * the zonelist. So only GFP_KERNEL allocations, if all nodes in the
2530 * cpuset are short of memory, might require taking the callback_lock.
2532 * The first call here from mm/page_alloc:get_page_from_freelist()
2533 * has __GFP_HARDWALL set in gfp_mask, enforcing hardwall cpusets,
2534 * so no allocation on a node outside the cpuset is allowed (unless
2535 * in interrupt, of course).
2537 * The second pass through get_page_from_freelist() doesn't even call
2538 * here for GFP_ATOMIC calls. For those calls, the __alloc_pages()
2539 * variable 'wait' is not set, and the bit ALLOC_CPUSET is not set
2540 * in alloc_flags. That logic and the checks below have the combined
2542 * in_interrupt - any node ok (current task context irrelevant)
2543 * GFP_ATOMIC - any node ok
2544 * TIF_MEMDIE - any node ok
2545 * GFP_KERNEL - any node in enclosing hardwalled cpuset ok
2546 * GFP_USER - only nodes in current tasks mems allowed ok.
2548 bool __cpuset_node_allowed(int node
, gfp_t gfp_mask
)
2550 struct cpuset
*cs
; /* current cpuset ancestors */
2551 int allowed
; /* is allocation in zone z allowed? */
2552 unsigned long flags
;
2556 if (node_isset(node
, current
->mems_allowed
))
2559 * Allow tasks that have access to memory reserves because they have
2560 * been OOM killed to get memory anywhere.
2562 if (unlikely(test_thread_flag(TIF_MEMDIE
)))
2564 if (gfp_mask
& __GFP_HARDWALL
) /* If hardwall request, stop here */
2567 if (current
->flags
& PF_EXITING
) /* Let dying task have memory */
2570 /* Not hardwall and node outside mems_allowed: scan up cpusets */
2571 spin_lock_irqsave(&callback_lock
, flags
);
2574 cs
= nearest_hardwall_ancestor(task_cs(current
));
2575 allowed
= node_isset(node
, cs
->mems_allowed
);
2578 spin_unlock_irqrestore(&callback_lock
, flags
);
2583 * cpuset_mem_spread_node() - On which node to begin search for a file page
2584 * cpuset_slab_spread_node() - On which node to begin search for a slab page
2586 * If a task is marked PF_SPREAD_PAGE or PF_SPREAD_SLAB (as for
2587 * tasks in a cpuset with is_spread_page or is_spread_slab set),
2588 * and if the memory allocation used cpuset_mem_spread_node()
2589 * to determine on which node to start looking, as it will for
2590 * certain page cache or slab cache pages such as used for file
2591 * system buffers and inode caches, then instead of starting on the
2592 * local node to look for a free page, rather spread the starting
2593 * node around the tasks mems_allowed nodes.
2595 * We don't have to worry about the returned node being offline
2596 * because "it can't happen", and even if it did, it would be ok.
2598 * The routines calling guarantee_online_mems() are careful to
2599 * only set nodes in task->mems_allowed that are online. So it
2600 * should not be possible for the following code to return an
2601 * offline node. But if it did, that would be ok, as this routine
2602 * is not returning the node where the allocation must be, only
2603 * the node where the search should start. The zonelist passed to
2604 * __alloc_pages() will include all nodes. If the slab allocator
2605 * is passed an offline node, it will fall back to the local node.
2606 * See kmem_cache_alloc_node().
2609 static int cpuset_spread_node(int *rotor
)
2611 return *rotor
= next_node_in(*rotor
, current
->mems_allowed
);
2614 int cpuset_mem_spread_node(void)
2616 if (current
->cpuset_mem_spread_rotor
== NUMA_NO_NODE
)
2617 current
->cpuset_mem_spread_rotor
=
2618 node_random(¤t
->mems_allowed
);
2620 return cpuset_spread_node(¤t
->cpuset_mem_spread_rotor
);
2623 int cpuset_slab_spread_node(void)
2625 if (current
->cpuset_slab_spread_rotor
== NUMA_NO_NODE
)
2626 current
->cpuset_slab_spread_rotor
=
2627 node_random(¤t
->mems_allowed
);
2629 return cpuset_spread_node(¤t
->cpuset_slab_spread_rotor
);
2632 EXPORT_SYMBOL_GPL(cpuset_mem_spread_node
);
2635 * cpuset_mems_allowed_intersects - Does @tsk1's mems_allowed intersect @tsk2's?
2636 * @tsk1: pointer to task_struct of some task.
2637 * @tsk2: pointer to task_struct of some other task.
2639 * Description: Return true if @tsk1's mems_allowed intersects the
2640 * mems_allowed of @tsk2. Used by the OOM killer to determine if
2641 * one of the task's memory usage might impact the memory available
2645 int cpuset_mems_allowed_intersects(const struct task_struct
*tsk1
,
2646 const struct task_struct
*tsk2
)
2648 return nodes_intersects(tsk1
->mems_allowed
, tsk2
->mems_allowed
);
2652 * cpuset_print_current_mems_allowed - prints current's cpuset and mems_allowed
2654 * Description: Prints current's name, cpuset name, and cached copy of its
2655 * mems_allowed to the kernel log.
2657 void cpuset_print_current_mems_allowed(void)
2659 struct cgroup
*cgrp
;
2663 cgrp
= task_cs(current
)->css
.cgroup
;
2664 pr_info("%s cpuset=", current
->comm
);
2665 pr_cont_cgroup_name(cgrp
);
2666 pr_cont(" mems_allowed=%*pbl\n",
2667 nodemask_pr_args(¤t
->mems_allowed
));
2673 * Collection of memory_pressure is suppressed unless
2674 * this flag is enabled by writing "1" to the special
2675 * cpuset file 'memory_pressure_enabled' in the root cpuset.
2678 int cpuset_memory_pressure_enabled __read_mostly
;
2681 * cpuset_memory_pressure_bump - keep stats of per-cpuset reclaims.
2683 * Keep a running average of the rate of synchronous (direct)
2684 * page reclaim efforts initiated by tasks in each cpuset.
2686 * This represents the rate at which some task in the cpuset
2687 * ran low on memory on all nodes it was allowed to use, and
2688 * had to enter the kernels page reclaim code in an effort to
2689 * create more free memory by tossing clean pages or swapping
2690 * or writing dirty pages.
2692 * Display to user space in the per-cpuset read-only file
2693 * "memory_pressure". Value displayed is an integer
2694 * representing the recent rate of entry into the synchronous
2695 * (direct) page reclaim by any task attached to the cpuset.
2698 void __cpuset_memory_pressure_bump(void)
2701 fmeter_markevent(&task_cs(current
)->fmeter
);
2705 #ifdef CONFIG_PROC_PID_CPUSET
2707 * proc_cpuset_show()
2708 * - Print tasks cpuset path into seq_file.
2709 * - Used for /proc/<pid>/cpuset.
2710 * - No need to task_lock(tsk) on this tsk->cpuset reference, as it
2711 * doesn't really matter if tsk->cpuset changes after we read it,
2712 * and we take cpuset_mutex, keeping cpuset_attach() from changing it
2715 int proc_cpuset_show(struct seq_file
*m
, struct pid_namespace
*ns
,
2716 struct pid
*pid
, struct task_struct
*tsk
)
2719 struct cgroup_subsys_state
*css
;
2723 buf
= kmalloc(PATH_MAX
, GFP_KERNEL
);
2727 css
= task_get_css(tsk
, cpuset_cgrp_id
);
2728 retval
= cgroup_path_ns(css
->cgroup
, buf
, PATH_MAX
,
2729 current
->nsproxy
->cgroup_ns
);
2731 if (retval
>= PATH_MAX
)
2732 retval
= -ENAMETOOLONG
;
2743 #endif /* CONFIG_PROC_PID_CPUSET */
2745 /* Display task mems_allowed in /proc/<pid>/status file. */
2746 void cpuset_task_status_allowed(struct seq_file
*m
, struct task_struct
*task
)
2748 seq_printf(m
, "Mems_allowed:\t%*pb\n",
2749 nodemask_pr_args(&task
->mems_allowed
));
2750 seq_printf(m
, "Mems_allowed_list:\t%*pbl\n",
2751 nodemask_pr_args(&task
->mems_allowed
));