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/time64.h>
55 #include <linux/backing-dev.h>
56 #include <linux/sort.h>
58 #include <asm/uaccess.h>
59 #include <linux/atomic.h>
60 #include <linux/mutex.h>
61 #include <linux/workqueue.h>
62 #include <linux/cgroup.h>
63 #include <linux/wait.h>
65 struct static_key cpusets_enabled_key __read_mostly
= STATIC_KEY_INIT_FALSE
;
67 /* See "Frequency meter" comments, below. */
70 int cnt
; /* unprocessed events count */
71 int val
; /* most recent output value */
72 time64_t time
; /* clock (secs) when val computed */
73 spinlock_t lock
; /* guards read or write of above */
77 struct cgroup_subsys_state css
;
79 unsigned long flags
; /* "unsigned long" so bitops work */
82 * On default hierarchy:
84 * The user-configured masks can only be changed by writing to
85 * cpuset.cpus and cpuset.mems, and won't be limited by the
88 * The effective masks is the real masks that apply to the tasks
89 * in the cpuset. They may be changed if the configured masks are
90 * changed or hotplug happens.
92 * effective_mask == configured_mask & parent's effective_mask,
93 * and if it ends up empty, it will inherit the parent's mask.
98 * The user-configured masks are always the same with effective masks.
101 /* user-configured CPUs and Memory Nodes allow to tasks */
102 cpumask_var_t cpus_allowed
;
103 nodemask_t mems_allowed
;
105 /* effective CPUs and Memory Nodes allow to tasks */
106 cpumask_var_t effective_cpus
;
107 nodemask_t effective_mems
;
110 * This is old Memory Nodes tasks took on.
112 * - top_cpuset.old_mems_allowed is initialized to mems_allowed.
113 * - A new cpuset's old_mems_allowed is initialized when some
114 * task is moved into it.
115 * - old_mems_allowed is used in cpuset_migrate_mm() when we change
116 * cpuset.mems_allowed and have tasks' nodemask updated, and
117 * then old_mems_allowed is updated to mems_allowed.
119 nodemask_t old_mems_allowed
;
121 struct fmeter fmeter
; /* memory_pressure filter */
124 * Tasks are being attached to this cpuset. Used to prevent
125 * zeroing cpus/mems_allowed between ->can_attach() and ->attach().
127 int attach_in_progress
;
129 /* partition number for rebuild_sched_domains() */
132 /* for custom sched domain */
133 int relax_domain_level
;
136 static inline struct cpuset
*css_cs(struct cgroup_subsys_state
*css
)
138 return css
? container_of(css
, struct cpuset
, css
) : NULL
;
141 /* Retrieve the cpuset for a task */
142 static inline struct cpuset
*task_cs(struct task_struct
*task
)
144 return css_cs(task_css(task
, cpuset_cgrp_id
));
147 static inline struct cpuset
*parent_cs(struct cpuset
*cs
)
149 return css_cs(cs
->css
.parent
);
153 static inline bool task_has_mempolicy(struct task_struct
*task
)
155 return task
->mempolicy
;
158 static inline bool task_has_mempolicy(struct task_struct
*task
)
165 /* bits in struct cpuset flags field */
172 CS_SCHED_LOAD_BALANCE
,
177 /* convenient tests for these bits */
178 static inline bool is_cpuset_online(const struct cpuset
*cs
)
180 return test_bit(CS_ONLINE
, &cs
->flags
);
183 static inline int is_cpu_exclusive(const struct cpuset
*cs
)
185 return test_bit(CS_CPU_EXCLUSIVE
, &cs
->flags
);
188 static inline int is_mem_exclusive(const struct cpuset
*cs
)
190 return test_bit(CS_MEM_EXCLUSIVE
, &cs
->flags
);
193 static inline int is_mem_hardwall(const struct cpuset
*cs
)
195 return test_bit(CS_MEM_HARDWALL
, &cs
->flags
);
198 static inline int is_sched_load_balance(const struct cpuset
*cs
)
200 return test_bit(CS_SCHED_LOAD_BALANCE
, &cs
->flags
);
203 static inline int is_memory_migrate(const struct cpuset
*cs
)
205 return test_bit(CS_MEMORY_MIGRATE
, &cs
->flags
);
208 static inline int is_spread_page(const struct cpuset
*cs
)
210 return test_bit(CS_SPREAD_PAGE
, &cs
->flags
);
213 static inline int is_spread_slab(const struct cpuset
*cs
)
215 return test_bit(CS_SPREAD_SLAB
, &cs
->flags
);
218 static struct cpuset top_cpuset
= {
219 .flags
= ((1 << CS_ONLINE
) | (1 << CS_CPU_EXCLUSIVE
) |
220 (1 << CS_MEM_EXCLUSIVE
)),
224 * cpuset_for_each_child - traverse online children of a cpuset
225 * @child_cs: loop cursor pointing to the current child
226 * @pos_css: used for iteration
227 * @parent_cs: target cpuset to walk children of
229 * Walk @child_cs through the online children of @parent_cs. Must be used
230 * with RCU read locked.
232 #define cpuset_for_each_child(child_cs, pos_css, parent_cs) \
233 css_for_each_child((pos_css), &(parent_cs)->css) \
234 if (is_cpuset_online(((child_cs) = css_cs((pos_css)))))
237 * cpuset_for_each_descendant_pre - pre-order walk of a cpuset's descendants
238 * @des_cs: loop cursor pointing to the current descendant
239 * @pos_css: used for iteration
240 * @root_cs: target cpuset to walk ancestor of
242 * Walk @des_cs through the online descendants of @root_cs. Must be used
243 * with RCU read locked. The caller may modify @pos_css by calling
244 * css_rightmost_descendant() to skip subtree. @root_cs is included in the
245 * iteration and the first node to be visited.
247 #define cpuset_for_each_descendant_pre(des_cs, pos_css, root_cs) \
248 css_for_each_descendant_pre((pos_css), &(root_cs)->css) \
249 if (is_cpuset_online(((des_cs) = css_cs((pos_css)))))
252 * There are two global locks guarding cpuset structures - cpuset_mutex and
253 * callback_lock. We also require taking task_lock() when dereferencing a
254 * task's cpuset pointer. See "The task_lock() exception", at the end of this
257 * A task must hold both locks to modify cpusets. If a task holds
258 * cpuset_mutex, then it blocks others wanting that mutex, ensuring that it
259 * is the only task able to also acquire callback_lock and be able to
260 * modify cpusets. It can perform various checks on the cpuset structure
261 * first, knowing nothing will change. It can also allocate memory while
262 * just holding cpuset_mutex. While it is performing these checks, various
263 * callback routines can briefly acquire callback_lock to query cpusets.
264 * Once it is ready to make the changes, it takes callback_lock, blocking
267 * Calls to the kernel memory allocator can not be made while holding
268 * callback_lock, as that would risk double tripping on callback_lock
269 * from one of the callbacks into the cpuset code from within
272 * If a task is only holding callback_lock, then it has read-only
275 * Now, the task_struct fields mems_allowed and mempolicy may be changed
276 * by other task, we use alloc_lock in the task_struct fields to protect
279 * The cpuset_common_file_read() handlers only hold callback_lock across
280 * small pieces of code, such as when reading out possibly multi-word
281 * cpumasks and nodemasks.
283 * Accessing a task's cpuset should be done in accordance with the
284 * guidelines for accessing subsystem state in kernel/cgroup.c
287 static DEFINE_MUTEX(cpuset_mutex
);
288 static DEFINE_SPINLOCK(callback_lock
);
291 * CPU / memory hotplug is handled asynchronously.
293 static void cpuset_hotplug_workfn(struct work_struct
*work
);
294 static DECLARE_WORK(cpuset_hotplug_work
, cpuset_hotplug_workfn
);
296 static DECLARE_WAIT_QUEUE_HEAD(cpuset_attach_wq
);
299 * This is ugly, but preserves the userspace API for existing cpuset
300 * users. If someone tries to mount the "cpuset" filesystem, we
301 * silently switch it to mount "cgroup" instead
303 static struct dentry
*cpuset_mount(struct file_system_type
*fs_type
,
304 int flags
, const char *unused_dev_name
, void *data
)
306 struct file_system_type
*cgroup_fs
= get_fs_type("cgroup");
307 struct dentry
*ret
= ERR_PTR(-ENODEV
);
311 "release_agent=/sbin/cpuset_release_agent";
312 ret
= cgroup_fs
->mount(cgroup_fs
, flags
,
313 unused_dev_name
, mountopts
);
314 put_filesystem(cgroup_fs
);
319 static struct file_system_type cpuset_fs_type
= {
321 .mount
= cpuset_mount
,
325 * Return in pmask the portion of a cpusets's cpus_allowed that
326 * are online. If none are online, walk up the cpuset hierarchy
327 * until we find one that does have some online cpus. The top
328 * cpuset always has some cpus online.
330 * One way or another, we guarantee to return some non-empty subset
331 * of cpu_online_mask.
333 * Call with callback_lock or cpuset_mutex held.
335 static void guarantee_online_cpus(struct cpuset
*cs
, struct cpumask
*pmask
)
337 while (!cpumask_intersects(cs
->effective_cpus
, cpu_online_mask
))
339 cpumask_and(pmask
, cs
->effective_cpus
, cpu_online_mask
);
343 * Return in *pmask the portion of a cpusets's mems_allowed that
344 * are online, with memory. If none are online with memory, walk
345 * up the cpuset hierarchy until we find one that does have some
346 * online mems. The top cpuset always has some mems online.
348 * One way or another, we guarantee to return some non-empty subset
349 * of node_states[N_MEMORY].
351 * Call with callback_lock or cpuset_mutex held.
353 static void guarantee_online_mems(struct cpuset
*cs
, nodemask_t
*pmask
)
355 while (!nodes_intersects(cs
->effective_mems
, node_states
[N_MEMORY
]))
357 nodes_and(*pmask
, cs
->effective_mems
, node_states
[N_MEMORY
]);
361 * update task's spread flag if cpuset's page/slab spread flag is set
363 * Call with callback_lock or cpuset_mutex held.
365 static void cpuset_update_task_spread_flag(struct cpuset
*cs
,
366 struct task_struct
*tsk
)
368 if (is_spread_page(cs
))
369 task_set_spread_page(tsk
);
371 task_clear_spread_page(tsk
);
373 if (is_spread_slab(cs
))
374 task_set_spread_slab(tsk
);
376 task_clear_spread_slab(tsk
);
380 * is_cpuset_subset(p, q) - Is cpuset p a subset of cpuset q?
382 * One cpuset is a subset of another if all its allowed CPUs and
383 * Memory Nodes are a subset of the other, and its exclusive flags
384 * are only set if the other's are set. Call holding cpuset_mutex.
387 static int is_cpuset_subset(const struct cpuset
*p
, const struct cpuset
*q
)
389 return cpumask_subset(p
->cpus_allowed
, q
->cpus_allowed
) &&
390 nodes_subset(p
->mems_allowed
, q
->mems_allowed
) &&
391 is_cpu_exclusive(p
) <= is_cpu_exclusive(q
) &&
392 is_mem_exclusive(p
) <= is_mem_exclusive(q
);
396 * alloc_trial_cpuset - allocate a trial cpuset
397 * @cs: the cpuset that the trial cpuset duplicates
399 static struct cpuset
*alloc_trial_cpuset(struct cpuset
*cs
)
401 struct cpuset
*trial
;
403 trial
= kmemdup(cs
, sizeof(*cs
), GFP_KERNEL
);
407 if (!alloc_cpumask_var(&trial
->cpus_allowed
, GFP_KERNEL
))
409 if (!alloc_cpumask_var(&trial
->effective_cpus
, GFP_KERNEL
))
412 cpumask_copy(trial
->cpus_allowed
, cs
->cpus_allowed
);
413 cpumask_copy(trial
->effective_cpus
, cs
->effective_cpus
);
417 free_cpumask_var(trial
->cpus_allowed
);
424 * free_trial_cpuset - free the trial cpuset
425 * @trial: the trial cpuset to be freed
427 static void free_trial_cpuset(struct cpuset
*trial
)
429 free_cpumask_var(trial
->effective_cpus
);
430 free_cpumask_var(trial
->cpus_allowed
);
435 * validate_change() - Used to validate that any proposed cpuset change
436 * follows the structural rules for cpusets.
438 * If we replaced the flag and mask values of the current cpuset
439 * (cur) with those values in the trial cpuset (trial), would
440 * our various subset and exclusive rules still be valid? Presumes
443 * 'cur' is the address of an actual, in-use cpuset. Operations
444 * such as list traversal that depend on the actual address of the
445 * cpuset in the list must use cur below, not trial.
447 * 'trial' is the address of bulk structure copy of cur, with
448 * perhaps one or more of the fields cpus_allowed, mems_allowed,
449 * or flags changed to new, trial values.
451 * Return 0 if valid, -errno if not.
454 static int validate_change(struct cpuset
*cur
, struct cpuset
*trial
)
456 struct cgroup_subsys_state
*css
;
457 struct cpuset
*c
, *par
;
462 /* Each of our child cpusets must be a subset of us */
464 cpuset_for_each_child(c
, css
, cur
)
465 if (!is_cpuset_subset(c
, trial
))
468 /* Remaining checks don't apply to root cpuset */
470 if (cur
== &top_cpuset
)
473 par
= parent_cs(cur
);
475 /* On legacy hiearchy, we must be a subset of our parent cpuset. */
477 if (!cgroup_subsys_on_dfl(cpuset_cgrp_subsys
) &&
478 !is_cpuset_subset(trial
, par
))
482 * If either I or some sibling (!= me) is exclusive, we can't
486 cpuset_for_each_child(c
, css
, par
) {
487 if ((is_cpu_exclusive(trial
) || is_cpu_exclusive(c
)) &&
489 cpumask_intersects(trial
->cpus_allowed
, c
->cpus_allowed
))
491 if ((is_mem_exclusive(trial
) || is_mem_exclusive(c
)) &&
493 nodes_intersects(trial
->mems_allowed
, c
->mems_allowed
))
498 * Cpusets with tasks - existing or newly being attached - can't
499 * be changed to have empty cpus_allowed or mems_allowed.
502 if ((cgroup_is_populated(cur
->css
.cgroup
) || cur
->attach_in_progress
)) {
503 if (!cpumask_empty(cur
->cpus_allowed
) &&
504 cpumask_empty(trial
->cpus_allowed
))
506 if (!nodes_empty(cur
->mems_allowed
) &&
507 nodes_empty(trial
->mems_allowed
))
512 * We can't shrink if we won't have enough room for SCHED_DEADLINE
516 if (is_cpu_exclusive(cur
) &&
517 !cpuset_cpumask_can_shrink(cur
->cpus_allowed
,
518 trial
->cpus_allowed
))
529 * Helper routine for generate_sched_domains().
530 * Do cpusets a, b have overlapping effective cpus_allowed masks?
532 static int cpusets_overlap(struct cpuset
*a
, struct cpuset
*b
)
534 return cpumask_intersects(a
->effective_cpus
, b
->effective_cpus
);
538 update_domain_attr(struct sched_domain_attr
*dattr
, struct cpuset
*c
)
540 if (dattr
->relax_domain_level
< c
->relax_domain_level
)
541 dattr
->relax_domain_level
= c
->relax_domain_level
;
545 static void update_domain_attr_tree(struct sched_domain_attr
*dattr
,
546 struct cpuset
*root_cs
)
549 struct cgroup_subsys_state
*pos_css
;
552 cpuset_for_each_descendant_pre(cp
, pos_css
, root_cs
) {
553 /* skip the whole subtree if @cp doesn't have any CPU */
554 if (cpumask_empty(cp
->cpus_allowed
)) {
555 pos_css
= css_rightmost_descendant(pos_css
);
559 if (is_sched_load_balance(cp
))
560 update_domain_attr(dattr
, cp
);
566 * generate_sched_domains()
568 * This function builds a partial partition of the systems CPUs
569 * A 'partial partition' is a set of non-overlapping subsets whose
570 * union is a subset of that set.
571 * The output of this function needs to be passed to kernel/sched/core.c
572 * partition_sched_domains() routine, which will rebuild the scheduler's
573 * load balancing domains (sched domains) as specified by that partial
576 * See "What is sched_load_balance" in Documentation/cgroups/cpusets.txt
577 * for a background explanation of this.
579 * Does not return errors, on the theory that the callers of this
580 * routine would rather not worry about failures to rebuild sched
581 * domains when operating in the severe memory shortage situations
582 * that could cause allocation failures below.
584 * Must be called with cpuset_mutex held.
586 * The three key local variables below are:
587 * q - a linked-list queue of cpuset pointers, used to implement a
588 * top-down scan of all cpusets. This scan loads a pointer
589 * to each cpuset marked is_sched_load_balance into the
590 * array 'csa'. For our purposes, rebuilding the schedulers
591 * sched domains, we can ignore !is_sched_load_balance cpusets.
592 * csa - (for CpuSet Array) Array of pointers to all the cpusets
593 * that need to be load balanced, for convenient iterative
594 * access by the subsequent code that finds the best partition,
595 * i.e the set of domains (subsets) of CPUs such that the
596 * cpus_allowed of every cpuset marked is_sched_load_balance
597 * is a subset of one of these domains, while there are as
598 * many such domains as possible, each as small as possible.
599 * doms - Conversion of 'csa' to an array of cpumasks, for passing to
600 * the kernel/sched/core.c routine partition_sched_domains() in a
601 * convenient format, that can be easily compared to the prior
602 * value to determine what partition elements (sched domains)
603 * were changed (added or removed.)
605 * Finding the best partition (set of domains):
606 * The triple nested loops below over i, j, k scan over the
607 * load balanced cpusets (using the array of cpuset pointers in
608 * csa[]) looking for pairs of cpusets that have overlapping
609 * cpus_allowed, but which don't have the same 'pn' partition
610 * number and gives them in the same partition number. It keeps
611 * looping on the 'restart' label until it can no longer find
614 * The union of the cpus_allowed masks from the set of
615 * all cpusets having the same 'pn' value then form the one
616 * element of the partition (one sched domain) to be passed to
617 * partition_sched_domains().
619 static int generate_sched_domains(cpumask_var_t
**domains
,
620 struct sched_domain_attr
**attributes
)
622 struct cpuset
*cp
; /* scans q */
623 struct cpuset
**csa
; /* array of all cpuset ptrs */
624 int csn
; /* how many cpuset ptrs in csa so far */
625 int i
, j
, k
; /* indices for partition finding loops */
626 cpumask_var_t
*doms
; /* resulting partition; i.e. sched domains */
627 cpumask_var_t non_isolated_cpus
; /* load balanced CPUs */
628 struct sched_domain_attr
*dattr
; /* attributes for custom domains */
629 int ndoms
= 0; /* number of sched domains in result */
630 int nslot
; /* next empty doms[] struct cpumask slot */
631 struct cgroup_subsys_state
*pos_css
;
637 if (!alloc_cpumask_var(&non_isolated_cpus
, GFP_KERNEL
))
639 cpumask_andnot(non_isolated_cpus
, cpu_possible_mask
, cpu_isolated_map
);
641 /* Special case for the 99% of systems with one, full, sched domain */
642 if (is_sched_load_balance(&top_cpuset
)) {
644 doms
= alloc_sched_domains(ndoms
);
648 dattr
= kmalloc(sizeof(struct sched_domain_attr
), GFP_KERNEL
);
650 *dattr
= SD_ATTR_INIT
;
651 update_domain_attr_tree(dattr
, &top_cpuset
);
653 cpumask_and(doms
[0], top_cpuset
.effective_cpus
,
659 csa
= kmalloc(nr_cpusets() * sizeof(cp
), GFP_KERNEL
);
665 cpuset_for_each_descendant_pre(cp
, pos_css
, &top_cpuset
) {
666 if (cp
== &top_cpuset
)
669 * Continue traversing beyond @cp iff @cp has some CPUs and
670 * isn't load balancing. The former is obvious. The
671 * latter: All child cpusets contain a subset of the
672 * parent's cpus, so just skip them, and then we call
673 * update_domain_attr_tree() to calc relax_domain_level of
674 * the corresponding sched domain.
676 if (!cpumask_empty(cp
->cpus_allowed
) &&
677 !(is_sched_load_balance(cp
) &&
678 cpumask_intersects(cp
->cpus_allowed
, non_isolated_cpus
)))
681 if (is_sched_load_balance(cp
))
684 /* skip @cp's subtree */
685 pos_css
= css_rightmost_descendant(pos_css
);
689 for (i
= 0; i
< csn
; i
++)
694 /* Find the best partition (set of sched domains) */
695 for (i
= 0; i
< csn
; i
++) {
696 struct cpuset
*a
= csa
[i
];
699 for (j
= 0; j
< csn
; j
++) {
700 struct cpuset
*b
= csa
[j
];
703 if (apn
!= bpn
&& cpusets_overlap(a
, b
)) {
704 for (k
= 0; k
< csn
; k
++) {
705 struct cpuset
*c
= csa
[k
];
710 ndoms
--; /* one less element */
717 * Now we know how many domains to create.
718 * Convert <csn, csa> to <ndoms, doms> and populate cpu masks.
720 doms
= alloc_sched_domains(ndoms
);
725 * The rest of the code, including the scheduler, can deal with
726 * dattr==NULL case. No need to abort if alloc fails.
728 dattr
= kmalloc(ndoms
* sizeof(struct sched_domain_attr
), GFP_KERNEL
);
730 for (nslot
= 0, i
= 0; i
< csn
; i
++) {
731 struct cpuset
*a
= csa
[i
];
736 /* Skip completed partitions */
742 if (nslot
== ndoms
) {
743 static int warnings
= 10;
745 pr_warn("rebuild_sched_domains confused: nslot %d, ndoms %d, csn %d, i %d, apn %d\n",
746 nslot
, ndoms
, csn
, i
, apn
);
754 *(dattr
+ nslot
) = SD_ATTR_INIT
;
755 for (j
= i
; j
< csn
; j
++) {
756 struct cpuset
*b
= csa
[j
];
759 cpumask_or(dp
, dp
, b
->effective_cpus
);
760 cpumask_and(dp
, dp
, non_isolated_cpus
);
762 update_domain_attr_tree(dattr
+ nslot
, b
);
764 /* Done with this partition */
770 BUG_ON(nslot
!= ndoms
);
773 free_cpumask_var(non_isolated_cpus
);
777 * Fallback to the default domain if kmalloc() failed.
778 * See comments in partition_sched_domains().
789 * Rebuild scheduler domains.
791 * If the flag 'sched_load_balance' of any cpuset with non-empty
792 * 'cpus' changes, or if the 'cpus' allowed changes in any cpuset
793 * which has that flag enabled, or if any cpuset with a non-empty
794 * 'cpus' is removed, then call this routine to rebuild the
795 * scheduler's dynamic sched domains.
797 * Call with cpuset_mutex held. Takes get_online_cpus().
799 static void rebuild_sched_domains_locked(void)
801 struct sched_domain_attr
*attr
;
805 lockdep_assert_held(&cpuset_mutex
);
809 * We have raced with CPU hotplug. Don't do anything to avoid
810 * passing doms with offlined cpu to partition_sched_domains().
811 * Anyways, hotplug work item will rebuild sched domains.
813 if (!cpumask_equal(top_cpuset
.effective_cpus
, cpu_active_mask
))
816 /* Generate domain masks and attrs */
817 ndoms
= generate_sched_domains(&doms
, &attr
);
819 /* Have scheduler rebuild the domains */
820 partition_sched_domains(ndoms
, doms
, attr
);
824 #else /* !CONFIG_SMP */
825 static void rebuild_sched_domains_locked(void)
828 #endif /* CONFIG_SMP */
830 void rebuild_sched_domains(void)
832 mutex_lock(&cpuset_mutex
);
833 rebuild_sched_domains_locked();
834 mutex_unlock(&cpuset_mutex
);
838 * update_tasks_cpumask - Update the cpumasks of tasks in the cpuset.
839 * @cs: the cpuset in which each task's cpus_allowed mask needs to be changed
841 * Iterate through each task of @cs updating its cpus_allowed to the
842 * effective cpuset's. As this function is called with cpuset_mutex held,
843 * cpuset membership stays stable.
845 static void update_tasks_cpumask(struct cpuset
*cs
)
847 struct css_task_iter it
;
848 struct task_struct
*task
;
850 css_task_iter_start(&cs
->css
, &it
);
851 while ((task
= css_task_iter_next(&it
)))
852 set_cpus_allowed_ptr(task
, cs
->effective_cpus
);
853 css_task_iter_end(&it
);
857 * update_cpumasks_hier - Update effective cpumasks and tasks in the subtree
858 * @cs: the cpuset to consider
859 * @new_cpus: temp variable for calculating new effective_cpus
861 * When congifured cpumask is changed, the effective cpumasks of this cpuset
862 * and all its descendants need to be updated.
864 * On legacy hierachy, effective_cpus will be the same with cpu_allowed.
866 * Called with cpuset_mutex held
868 static void update_cpumasks_hier(struct cpuset
*cs
, struct cpumask
*new_cpus
)
871 struct cgroup_subsys_state
*pos_css
;
872 bool need_rebuild_sched_domains
= false;
875 cpuset_for_each_descendant_pre(cp
, pos_css
, cs
) {
876 struct cpuset
*parent
= parent_cs(cp
);
878 cpumask_and(new_cpus
, cp
->cpus_allowed
, parent
->effective_cpus
);
881 * If it becomes empty, inherit the effective mask of the
882 * parent, which is guaranteed to have some CPUs.
884 if (cgroup_subsys_on_dfl(cpuset_cgrp_subsys
) &&
885 cpumask_empty(new_cpus
))
886 cpumask_copy(new_cpus
, parent
->effective_cpus
);
888 /* Skip the whole subtree if the cpumask remains the same. */
889 if (cpumask_equal(new_cpus
, cp
->effective_cpus
)) {
890 pos_css
= css_rightmost_descendant(pos_css
);
894 if (!css_tryget_online(&cp
->css
))
898 spin_lock_irq(&callback_lock
);
899 cpumask_copy(cp
->effective_cpus
, new_cpus
);
900 spin_unlock_irq(&callback_lock
);
902 WARN_ON(!cgroup_subsys_on_dfl(cpuset_cgrp_subsys
) &&
903 !cpumask_equal(cp
->cpus_allowed
, cp
->effective_cpus
));
905 update_tasks_cpumask(cp
);
908 * If the effective cpumask of any non-empty cpuset is changed,
909 * we need to rebuild sched domains.
911 if (!cpumask_empty(cp
->cpus_allowed
) &&
912 is_sched_load_balance(cp
))
913 need_rebuild_sched_domains
= true;
920 if (need_rebuild_sched_domains
)
921 rebuild_sched_domains_locked();
925 * update_cpumask - update the cpus_allowed mask of a cpuset and all tasks in it
926 * @cs: the cpuset to consider
927 * @trialcs: trial cpuset
928 * @buf: buffer of cpu numbers written to this cpuset
930 static int update_cpumask(struct cpuset
*cs
, struct cpuset
*trialcs
,
935 /* top_cpuset.cpus_allowed tracks cpu_online_mask; it's read-only */
936 if (cs
== &top_cpuset
)
940 * An empty cpus_allowed is ok only if the cpuset has no tasks.
941 * Since cpulist_parse() fails on an empty mask, we special case
942 * that parsing. The validate_change() call ensures that cpusets
943 * with tasks have cpus.
946 cpumask_clear(trialcs
->cpus_allowed
);
948 retval
= cpulist_parse(buf
, trialcs
->cpus_allowed
);
952 if (!cpumask_subset(trialcs
->cpus_allowed
,
953 top_cpuset
.cpus_allowed
))
957 /* Nothing to do if the cpus didn't change */
958 if (cpumask_equal(cs
->cpus_allowed
, trialcs
->cpus_allowed
))
961 retval
= validate_change(cs
, trialcs
);
965 spin_lock_irq(&callback_lock
);
966 cpumask_copy(cs
->cpus_allowed
, trialcs
->cpus_allowed
);
967 spin_unlock_irq(&callback_lock
);
969 /* use trialcs->cpus_allowed as a temp variable */
970 update_cpumasks_hier(cs
, trialcs
->cpus_allowed
);
977 * Migrate memory region from one set of nodes to another.
979 * Temporarilly set tasks mems_allowed to target nodes of migration,
980 * so that the migration code can allocate pages on these nodes.
982 * While the mm_struct we are migrating is typically from some
983 * other task, the task_struct mems_allowed that we are hacking
984 * is for our current task, which must allocate new pages for that
985 * migrating memory region.
988 static void cpuset_migrate_mm(struct mm_struct
*mm
, const nodemask_t
*from
,
989 const nodemask_t
*to
)
991 struct task_struct
*tsk
= current
;
993 tsk
->mems_allowed
= *to
;
995 do_migrate_pages(mm
, from
, to
, MPOL_MF_MOVE_ALL
);
998 guarantee_online_mems(task_cs(tsk
), &tsk
->mems_allowed
);
1003 * cpuset_change_task_nodemask - change task's mems_allowed and mempolicy
1004 * @tsk: the task to change
1005 * @newmems: new nodes that the task will be set
1007 * In order to avoid seeing no nodes if the old and new nodes are disjoint,
1008 * we structure updates as setting all new allowed nodes, then clearing newly
1011 static void cpuset_change_task_nodemask(struct task_struct
*tsk
,
1012 nodemask_t
*newmems
)
1017 * Allow tasks that have access to memory reserves because they have
1018 * been OOM killed to get memory anywhere.
1020 if (unlikely(test_thread_flag(TIF_MEMDIE
)))
1022 if (current
->flags
& PF_EXITING
) /* Let dying task have memory */
1027 * Determine if a loop is necessary if another thread is doing
1028 * read_mems_allowed_begin(). If at least one node remains unchanged and
1029 * tsk does not have a mempolicy, then an empty nodemask will not be
1030 * possible when mems_allowed is larger than a word.
1032 need_loop
= task_has_mempolicy(tsk
) ||
1033 !nodes_intersects(*newmems
, tsk
->mems_allowed
);
1036 local_irq_disable();
1037 write_seqcount_begin(&tsk
->mems_allowed_seq
);
1040 nodes_or(tsk
->mems_allowed
, tsk
->mems_allowed
, *newmems
);
1041 mpol_rebind_task(tsk
, newmems
, MPOL_REBIND_STEP1
);
1043 mpol_rebind_task(tsk
, newmems
, MPOL_REBIND_STEP2
);
1044 tsk
->mems_allowed
= *newmems
;
1047 write_seqcount_end(&tsk
->mems_allowed_seq
);
1054 static void *cpuset_being_rebound
;
1057 * update_tasks_nodemask - Update the nodemasks of tasks in the cpuset.
1058 * @cs: the cpuset in which each task's mems_allowed mask needs to be changed
1060 * Iterate through each task of @cs updating its mems_allowed to the
1061 * effective cpuset's. As this function is called with cpuset_mutex held,
1062 * cpuset membership stays stable.
1064 static void update_tasks_nodemask(struct cpuset
*cs
)
1066 static nodemask_t newmems
; /* protected by cpuset_mutex */
1067 struct css_task_iter it
;
1068 struct task_struct
*task
;
1070 cpuset_being_rebound
= cs
; /* causes mpol_dup() rebind */
1072 guarantee_online_mems(cs
, &newmems
);
1075 * The mpol_rebind_mm() call takes mmap_sem, which we couldn't
1076 * take while holding tasklist_lock. Forks can happen - the
1077 * mpol_dup() cpuset_being_rebound check will catch such forks,
1078 * and rebind their vma mempolicies too. Because we still hold
1079 * the global cpuset_mutex, we know that no other rebind effort
1080 * will be contending for the global variable cpuset_being_rebound.
1081 * It's ok if we rebind the same mm twice; mpol_rebind_mm()
1082 * is idempotent. Also migrate pages in each mm to new nodes.
1084 css_task_iter_start(&cs
->css
, &it
);
1085 while ((task
= css_task_iter_next(&it
))) {
1086 struct mm_struct
*mm
;
1089 cpuset_change_task_nodemask(task
, &newmems
);
1091 mm
= get_task_mm(task
);
1095 migrate
= is_memory_migrate(cs
);
1097 mpol_rebind_mm(mm
, &cs
->mems_allowed
);
1099 cpuset_migrate_mm(mm
, &cs
->old_mems_allowed
, &newmems
);
1102 css_task_iter_end(&it
);
1105 * All the tasks' nodemasks have been updated, update
1106 * cs->old_mems_allowed.
1108 cs
->old_mems_allowed
= newmems
;
1110 /* We're done rebinding vmas to this cpuset's new mems_allowed. */
1111 cpuset_being_rebound
= NULL
;
1115 * update_nodemasks_hier - Update effective nodemasks and tasks in the subtree
1116 * @cs: the cpuset to consider
1117 * @new_mems: a temp variable for calculating new effective_mems
1119 * When configured nodemask is changed, the effective nodemasks of this cpuset
1120 * and all its descendants need to be updated.
1122 * On legacy hiearchy, effective_mems will be the same with mems_allowed.
1124 * Called with cpuset_mutex held
1126 static void update_nodemasks_hier(struct cpuset
*cs
, nodemask_t
*new_mems
)
1129 struct cgroup_subsys_state
*pos_css
;
1132 cpuset_for_each_descendant_pre(cp
, pos_css
, cs
) {
1133 struct cpuset
*parent
= parent_cs(cp
);
1135 nodes_and(*new_mems
, cp
->mems_allowed
, parent
->effective_mems
);
1138 * If it becomes empty, inherit the effective mask of the
1139 * parent, which is guaranteed to have some MEMs.
1141 if (cgroup_subsys_on_dfl(cpuset_cgrp_subsys
) &&
1142 nodes_empty(*new_mems
))
1143 *new_mems
= parent
->effective_mems
;
1145 /* Skip the whole subtree if the nodemask remains the same. */
1146 if (nodes_equal(*new_mems
, cp
->effective_mems
)) {
1147 pos_css
= css_rightmost_descendant(pos_css
);
1151 if (!css_tryget_online(&cp
->css
))
1155 spin_lock_irq(&callback_lock
);
1156 cp
->effective_mems
= *new_mems
;
1157 spin_unlock_irq(&callback_lock
);
1159 WARN_ON(!cgroup_subsys_on_dfl(cpuset_cgrp_subsys
) &&
1160 !nodes_equal(cp
->mems_allowed
, cp
->effective_mems
));
1162 update_tasks_nodemask(cp
);
1171 * Handle user request to change the 'mems' memory placement
1172 * of a cpuset. Needs to validate the request, update the
1173 * cpusets mems_allowed, and for each task in the cpuset,
1174 * update mems_allowed and rebind task's mempolicy and any vma
1175 * mempolicies and if the cpuset is marked 'memory_migrate',
1176 * migrate the tasks pages to the new memory.
1178 * Call with cpuset_mutex held. May take callback_lock during call.
1179 * Will take tasklist_lock, scan tasklist for tasks in cpuset cs,
1180 * lock each such tasks mm->mmap_sem, scan its vma's and rebind
1181 * their mempolicies to the cpusets new mems_allowed.
1183 static int update_nodemask(struct cpuset
*cs
, struct cpuset
*trialcs
,
1189 * top_cpuset.mems_allowed tracks node_stats[N_MEMORY];
1192 if (cs
== &top_cpuset
) {
1198 * An empty mems_allowed is ok iff there are no tasks in the cpuset.
1199 * Since nodelist_parse() fails on an empty mask, we special case
1200 * that parsing. The validate_change() call ensures that cpusets
1201 * with tasks have memory.
1204 nodes_clear(trialcs
->mems_allowed
);
1206 retval
= nodelist_parse(buf
, trialcs
->mems_allowed
);
1210 if (!nodes_subset(trialcs
->mems_allowed
,
1211 top_cpuset
.mems_allowed
)) {
1217 if (nodes_equal(cs
->mems_allowed
, trialcs
->mems_allowed
)) {
1218 retval
= 0; /* Too easy - nothing to do */
1221 retval
= validate_change(cs
, trialcs
);
1225 spin_lock_irq(&callback_lock
);
1226 cs
->mems_allowed
= trialcs
->mems_allowed
;
1227 spin_unlock_irq(&callback_lock
);
1229 /* use trialcs->mems_allowed as a temp variable */
1230 update_nodemasks_hier(cs
, &trialcs
->mems_allowed
);
1235 int current_cpuset_is_being_rebound(void)
1240 ret
= task_cs(current
) == cpuset_being_rebound
;
1246 static int update_relax_domain_level(struct cpuset
*cs
, s64 val
)
1249 if (val
< -1 || val
>= sched_domain_level_max
)
1253 if (val
!= cs
->relax_domain_level
) {
1254 cs
->relax_domain_level
= val
;
1255 if (!cpumask_empty(cs
->cpus_allowed
) &&
1256 is_sched_load_balance(cs
))
1257 rebuild_sched_domains_locked();
1264 * update_tasks_flags - update the spread flags of tasks in the cpuset.
1265 * @cs: the cpuset in which each task's spread flags needs to be changed
1267 * Iterate through each task of @cs updating its spread flags. As this
1268 * function is called with cpuset_mutex held, cpuset membership stays
1271 static void update_tasks_flags(struct cpuset
*cs
)
1273 struct css_task_iter it
;
1274 struct task_struct
*task
;
1276 css_task_iter_start(&cs
->css
, &it
);
1277 while ((task
= css_task_iter_next(&it
)))
1278 cpuset_update_task_spread_flag(cs
, task
);
1279 css_task_iter_end(&it
);
1283 * update_flag - read a 0 or a 1 in a file and update associated flag
1284 * bit: the bit to update (see cpuset_flagbits_t)
1285 * cs: the cpuset to update
1286 * turning_on: whether the flag is being set or cleared
1288 * Call with cpuset_mutex held.
1291 static int update_flag(cpuset_flagbits_t bit
, struct cpuset
*cs
,
1294 struct cpuset
*trialcs
;
1295 int balance_flag_changed
;
1296 int spread_flag_changed
;
1299 trialcs
= alloc_trial_cpuset(cs
);
1304 set_bit(bit
, &trialcs
->flags
);
1306 clear_bit(bit
, &trialcs
->flags
);
1308 err
= validate_change(cs
, trialcs
);
1312 balance_flag_changed
= (is_sched_load_balance(cs
) !=
1313 is_sched_load_balance(trialcs
));
1315 spread_flag_changed
= ((is_spread_slab(cs
) != is_spread_slab(trialcs
))
1316 || (is_spread_page(cs
) != is_spread_page(trialcs
)));
1318 spin_lock_irq(&callback_lock
);
1319 cs
->flags
= trialcs
->flags
;
1320 spin_unlock_irq(&callback_lock
);
1322 if (!cpumask_empty(trialcs
->cpus_allowed
) && balance_flag_changed
)
1323 rebuild_sched_domains_locked();
1325 if (spread_flag_changed
)
1326 update_tasks_flags(cs
);
1328 free_trial_cpuset(trialcs
);
1333 * Frequency meter - How fast is some event occurring?
1335 * These routines manage a digitally filtered, constant time based,
1336 * event frequency meter. There are four routines:
1337 * fmeter_init() - initialize a frequency meter.
1338 * fmeter_markevent() - called each time the event happens.
1339 * fmeter_getrate() - returns the recent rate of such events.
1340 * fmeter_update() - internal routine used to update fmeter.
1342 * A common data structure is passed to each of these routines,
1343 * which is used to keep track of the state required to manage the
1344 * frequency meter and its digital filter.
1346 * The filter works on the number of events marked per unit time.
1347 * The filter is single-pole low-pass recursive (IIR). The time unit
1348 * is 1 second. Arithmetic is done using 32-bit integers scaled to
1349 * simulate 3 decimal digits of precision (multiplied by 1000).
1351 * With an FM_COEF of 933, and a time base of 1 second, the filter
1352 * has a half-life of 10 seconds, meaning that if the events quit
1353 * happening, then the rate returned from the fmeter_getrate()
1354 * will be cut in half each 10 seconds, until it converges to zero.
1356 * It is not worth doing a real infinitely recursive filter. If more
1357 * than FM_MAXTICKS ticks have elapsed since the last filter event,
1358 * just compute FM_MAXTICKS ticks worth, by which point the level
1361 * Limit the count of unprocessed events to FM_MAXCNT, so as to avoid
1362 * arithmetic overflow in the fmeter_update() routine.
1364 * Given the simple 32 bit integer arithmetic used, this meter works
1365 * best for reporting rates between one per millisecond (msec) and
1366 * one per 32 (approx) seconds. At constant rates faster than one
1367 * per msec it maxes out at values just under 1,000,000. At constant
1368 * rates between one per msec, and one per second it will stabilize
1369 * to a value N*1000, where N is the rate of events per second.
1370 * At constant rates between one per second and one per 32 seconds,
1371 * it will be choppy, moving up on the seconds that have an event,
1372 * and then decaying until the next event. At rates slower than
1373 * about one in 32 seconds, it decays all the way back to zero between
1377 #define FM_COEF 933 /* coefficient for half-life of 10 secs */
1378 #define FM_MAXTICKS ((u32)99) /* useless computing more ticks than this */
1379 #define FM_MAXCNT 1000000 /* limit cnt to avoid overflow */
1380 #define FM_SCALE 1000 /* faux fixed point scale */
1382 /* Initialize a frequency meter */
1383 static void fmeter_init(struct fmeter
*fmp
)
1388 spin_lock_init(&fmp
->lock
);
1391 /* Internal meter update - process cnt events and update value */
1392 static void fmeter_update(struct fmeter
*fmp
)
1397 now
= ktime_get_seconds();
1398 ticks
= now
- fmp
->time
;
1403 ticks
= min(FM_MAXTICKS
, ticks
);
1405 fmp
->val
= (FM_COEF
* fmp
->val
) / FM_SCALE
;
1408 fmp
->val
+= ((FM_SCALE
- FM_COEF
) * fmp
->cnt
) / FM_SCALE
;
1412 /* Process any previous ticks, then bump cnt by one (times scale). */
1413 static void fmeter_markevent(struct fmeter
*fmp
)
1415 spin_lock(&fmp
->lock
);
1417 fmp
->cnt
= min(FM_MAXCNT
, fmp
->cnt
+ FM_SCALE
);
1418 spin_unlock(&fmp
->lock
);
1421 /* Process any previous ticks, then return current value. */
1422 static int fmeter_getrate(struct fmeter
*fmp
)
1426 spin_lock(&fmp
->lock
);
1429 spin_unlock(&fmp
->lock
);
1433 static struct cpuset
*cpuset_attach_old_cs
;
1435 /* Called by cgroups to determine if a cpuset is usable; cpuset_mutex held */
1436 static int cpuset_can_attach(struct cgroup_taskset
*tset
)
1438 struct cgroup_subsys_state
*css
;
1440 struct task_struct
*task
;
1443 /* used later by cpuset_attach() */
1444 cpuset_attach_old_cs
= task_cs(cgroup_taskset_first(tset
, &css
));
1447 mutex_lock(&cpuset_mutex
);
1449 /* allow moving tasks into an empty cpuset if on default hierarchy */
1451 if (!cgroup_subsys_on_dfl(cpuset_cgrp_subsys
) &&
1452 (cpumask_empty(cs
->cpus_allowed
) || nodes_empty(cs
->mems_allowed
)))
1455 cgroup_taskset_for_each(task
, css
, tset
) {
1456 ret
= task_can_attach(task
, cs
->cpus_allowed
);
1459 ret
= security_task_setscheduler(task
);
1465 * Mark attach is in progress. This makes validate_change() fail
1466 * changes which zero cpus/mems_allowed.
1468 cs
->attach_in_progress
++;
1471 mutex_unlock(&cpuset_mutex
);
1475 static void cpuset_cancel_attach(struct cgroup_taskset
*tset
)
1477 struct cgroup_subsys_state
*css
;
1480 cgroup_taskset_first(tset
, &css
);
1483 mutex_lock(&cpuset_mutex
);
1484 css_cs(css
)->attach_in_progress
--;
1485 mutex_unlock(&cpuset_mutex
);
1489 * Protected by cpuset_mutex. cpus_attach is used only by cpuset_attach()
1490 * but we can't allocate it dynamically there. Define it global and
1491 * allocate from cpuset_init().
1493 static cpumask_var_t cpus_attach
;
1495 static void cpuset_attach(struct cgroup_taskset
*tset
)
1497 /* static buf protected by cpuset_mutex */
1498 static nodemask_t cpuset_attach_nodemask_to
;
1499 struct task_struct
*task
;
1500 struct task_struct
*leader
;
1501 struct cgroup_subsys_state
*css
;
1503 struct cpuset
*oldcs
= cpuset_attach_old_cs
;
1505 cgroup_taskset_first(tset
, &css
);
1508 mutex_lock(&cpuset_mutex
);
1510 /* prepare for attach */
1511 if (cs
== &top_cpuset
)
1512 cpumask_copy(cpus_attach
, cpu_possible_mask
);
1514 guarantee_online_cpus(cs
, cpus_attach
);
1516 guarantee_online_mems(cs
, &cpuset_attach_nodemask_to
);
1518 cgroup_taskset_for_each(task
, css
, tset
) {
1520 * can_attach beforehand should guarantee that this doesn't
1521 * fail. TODO: have a better way to handle failure here
1523 WARN_ON_ONCE(set_cpus_allowed_ptr(task
, cpus_attach
));
1525 cpuset_change_task_nodemask(task
, &cpuset_attach_nodemask_to
);
1526 cpuset_update_task_spread_flag(cs
, task
);
1530 * Change mm for all threadgroup leaders. This is expensive and may
1531 * sleep and should be moved outside migration path proper.
1533 cpuset_attach_nodemask_to
= cs
->effective_mems
;
1534 cgroup_taskset_for_each_leader(leader
, css
, tset
) {
1535 struct mm_struct
*mm
= get_task_mm(leader
);
1538 mpol_rebind_mm(mm
, &cpuset_attach_nodemask_to
);
1541 * old_mems_allowed is the same with mems_allowed
1542 * here, except if this task is being moved
1543 * automatically due to hotplug. In that case
1544 * @mems_allowed has been updated and is empty, so
1545 * @old_mems_allowed is the right nodesets that we
1548 if (is_memory_migrate(cs
)) {
1549 cpuset_migrate_mm(mm
, &oldcs
->old_mems_allowed
,
1550 &cpuset_attach_nodemask_to
);
1556 cs
->old_mems_allowed
= cpuset_attach_nodemask_to
;
1558 cs
->attach_in_progress
--;
1559 if (!cs
->attach_in_progress
)
1560 wake_up(&cpuset_attach_wq
);
1562 mutex_unlock(&cpuset_mutex
);
1565 /* The various types of files and directories in a cpuset file system */
1568 FILE_MEMORY_MIGRATE
,
1571 FILE_EFFECTIVE_CPULIST
,
1572 FILE_EFFECTIVE_MEMLIST
,
1576 FILE_SCHED_LOAD_BALANCE
,
1577 FILE_SCHED_RELAX_DOMAIN_LEVEL
,
1578 FILE_MEMORY_PRESSURE_ENABLED
,
1579 FILE_MEMORY_PRESSURE
,
1582 } cpuset_filetype_t
;
1584 static int cpuset_write_u64(struct cgroup_subsys_state
*css
, struct cftype
*cft
,
1587 struct cpuset
*cs
= css_cs(css
);
1588 cpuset_filetype_t type
= cft
->private;
1591 mutex_lock(&cpuset_mutex
);
1592 if (!is_cpuset_online(cs
)) {
1598 case FILE_CPU_EXCLUSIVE
:
1599 retval
= update_flag(CS_CPU_EXCLUSIVE
, cs
, val
);
1601 case FILE_MEM_EXCLUSIVE
:
1602 retval
= update_flag(CS_MEM_EXCLUSIVE
, cs
, val
);
1604 case FILE_MEM_HARDWALL
:
1605 retval
= update_flag(CS_MEM_HARDWALL
, cs
, val
);
1607 case FILE_SCHED_LOAD_BALANCE
:
1608 retval
= update_flag(CS_SCHED_LOAD_BALANCE
, cs
, val
);
1610 case FILE_MEMORY_MIGRATE
:
1611 retval
= update_flag(CS_MEMORY_MIGRATE
, cs
, val
);
1613 case FILE_MEMORY_PRESSURE_ENABLED
:
1614 cpuset_memory_pressure_enabled
= !!val
;
1616 case FILE_SPREAD_PAGE
:
1617 retval
= update_flag(CS_SPREAD_PAGE
, cs
, val
);
1619 case FILE_SPREAD_SLAB
:
1620 retval
= update_flag(CS_SPREAD_SLAB
, cs
, val
);
1627 mutex_unlock(&cpuset_mutex
);
1631 static int cpuset_write_s64(struct cgroup_subsys_state
*css
, struct cftype
*cft
,
1634 struct cpuset
*cs
= css_cs(css
);
1635 cpuset_filetype_t type
= cft
->private;
1636 int retval
= -ENODEV
;
1638 mutex_lock(&cpuset_mutex
);
1639 if (!is_cpuset_online(cs
))
1643 case FILE_SCHED_RELAX_DOMAIN_LEVEL
:
1644 retval
= update_relax_domain_level(cs
, val
);
1651 mutex_unlock(&cpuset_mutex
);
1656 * Common handling for a write to a "cpus" or "mems" file.
1658 static ssize_t
cpuset_write_resmask(struct kernfs_open_file
*of
,
1659 char *buf
, size_t nbytes
, loff_t off
)
1661 struct cpuset
*cs
= css_cs(of_css(of
));
1662 struct cpuset
*trialcs
;
1663 int retval
= -ENODEV
;
1665 buf
= strstrip(buf
);
1668 * CPU or memory hotunplug may leave @cs w/o any execution
1669 * resources, in which case the hotplug code asynchronously updates
1670 * configuration and transfers all tasks to the nearest ancestor
1671 * which can execute.
1673 * As writes to "cpus" or "mems" may restore @cs's execution
1674 * resources, wait for the previously scheduled operations before
1675 * proceeding, so that we don't end up keep removing tasks added
1676 * after execution capability is restored.
1678 * cpuset_hotplug_work calls back into cgroup core via
1679 * cgroup_transfer_tasks() and waiting for it from a cgroupfs
1680 * operation like this one can lead to a deadlock through kernfs
1681 * active_ref protection. Let's break the protection. Losing the
1682 * protection is okay as we check whether @cs is online after
1683 * grabbing cpuset_mutex anyway. This only happens on the legacy
1687 kernfs_break_active_protection(of
->kn
);
1688 flush_work(&cpuset_hotplug_work
);
1690 mutex_lock(&cpuset_mutex
);
1691 if (!is_cpuset_online(cs
))
1694 trialcs
= alloc_trial_cpuset(cs
);
1700 switch (of_cft(of
)->private) {
1702 retval
= update_cpumask(cs
, trialcs
, buf
);
1705 retval
= update_nodemask(cs
, trialcs
, buf
);
1712 free_trial_cpuset(trialcs
);
1714 mutex_unlock(&cpuset_mutex
);
1715 kernfs_unbreak_active_protection(of
->kn
);
1717 return retval
?: nbytes
;
1721 * These ascii lists should be read in a single call, by using a user
1722 * buffer large enough to hold the entire map. If read in smaller
1723 * chunks, there is no guarantee of atomicity. Since the display format
1724 * used, list of ranges of sequential numbers, is variable length,
1725 * and since these maps can change value dynamically, one could read
1726 * gibberish by doing partial reads while a list was changing.
1728 static int cpuset_common_seq_show(struct seq_file
*sf
, void *v
)
1730 struct cpuset
*cs
= css_cs(seq_css(sf
));
1731 cpuset_filetype_t type
= seq_cft(sf
)->private;
1734 spin_lock_irq(&callback_lock
);
1738 seq_printf(sf
, "%*pbl\n", cpumask_pr_args(cs
->cpus_allowed
));
1741 seq_printf(sf
, "%*pbl\n", nodemask_pr_args(&cs
->mems_allowed
));
1743 case FILE_EFFECTIVE_CPULIST
:
1744 seq_printf(sf
, "%*pbl\n", cpumask_pr_args(cs
->effective_cpus
));
1746 case FILE_EFFECTIVE_MEMLIST
:
1747 seq_printf(sf
, "%*pbl\n", nodemask_pr_args(&cs
->effective_mems
));
1753 spin_unlock_irq(&callback_lock
);
1757 static u64
cpuset_read_u64(struct cgroup_subsys_state
*css
, struct cftype
*cft
)
1759 struct cpuset
*cs
= css_cs(css
);
1760 cpuset_filetype_t type
= cft
->private;
1762 case FILE_CPU_EXCLUSIVE
:
1763 return is_cpu_exclusive(cs
);
1764 case FILE_MEM_EXCLUSIVE
:
1765 return is_mem_exclusive(cs
);
1766 case FILE_MEM_HARDWALL
:
1767 return is_mem_hardwall(cs
);
1768 case FILE_SCHED_LOAD_BALANCE
:
1769 return is_sched_load_balance(cs
);
1770 case FILE_MEMORY_MIGRATE
:
1771 return is_memory_migrate(cs
);
1772 case FILE_MEMORY_PRESSURE_ENABLED
:
1773 return cpuset_memory_pressure_enabled
;
1774 case FILE_MEMORY_PRESSURE
:
1775 return fmeter_getrate(&cs
->fmeter
);
1776 case FILE_SPREAD_PAGE
:
1777 return is_spread_page(cs
);
1778 case FILE_SPREAD_SLAB
:
1779 return is_spread_slab(cs
);
1784 /* Unreachable but makes gcc happy */
1788 static s64
cpuset_read_s64(struct cgroup_subsys_state
*css
, struct cftype
*cft
)
1790 struct cpuset
*cs
= css_cs(css
);
1791 cpuset_filetype_t type
= cft
->private;
1793 case FILE_SCHED_RELAX_DOMAIN_LEVEL
:
1794 return cs
->relax_domain_level
;
1799 /* Unrechable but makes gcc happy */
1805 * for the common functions, 'private' gives the type of file
1808 static struct cftype files
[] = {
1811 .seq_show
= cpuset_common_seq_show
,
1812 .write
= cpuset_write_resmask
,
1813 .max_write_len
= (100U + 6 * NR_CPUS
),
1814 .private = FILE_CPULIST
,
1819 .seq_show
= cpuset_common_seq_show
,
1820 .write
= cpuset_write_resmask
,
1821 .max_write_len
= (100U + 6 * MAX_NUMNODES
),
1822 .private = FILE_MEMLIST
,
1826 .name
= "effective_cpus",
1827 .seq_show
= cpuset_common_seq_show
,
1828 .private = FILE_EFFECTIVE_CPULIST
,
1832 .name
= "effective_mems",
1833 .seq_show
= cpuset_common_seq_show
,
1834 .private = FILE_EFFECTIVE_MEMLIST
,
1838 .name
= "cpu_exclusive",
1839 .read_u64
= cpuset_read_u64
,
1840 .write_u64
= cpuset_write_u64
,
1841 .private = FILE_CPU_EXCLUSIVE
,
1845 .name
= "mem_exclusive",
1846 .read_u64
= cpuset_read_u64
,
1847 .write_u64
= cpuset_write_u64
,
1848 .private = FILE_MEM_EXCLUSIVE
,
1852 .name
= "mem_hardwall",
1853 .read_u64
= cpuset_read_u64
,
1854 .write_u64
= cpuset_write_u64
,
1855 .private = FILE_MEM_HARDWALL
,
1859 .name
= "sched_load_balance",
1860 .read_u64
= cpuset_read_u64
,
1861 .write_u64
= cpuset_write_u64
,
1862 .private = FILE_SCHED_LOAD_BALANCE
,
1866 .name
= "sched_relax_domain_level",
1867 .read_s64
= cpuset_read_s64
,
1868 .write_s64
= cpuset_write_s64
,
1869 .private = FILE_SCHED_RELAX_DOMAIN_LEVEL
,
1873 .name
= "memory_migrate",
1874 .read_u64
= cpuset_read_u64
,
1875 .write_u64
= cpuset_write_u64
,
1876 .private = FILE_MEMORY_MIGRATE
,
1880 .name
= "memory_pressure",
1881 .read_u64
= cpuset_read_u64
,
1885 .name
= "memory_spread_page",
1886 .read_u64
= cpuset_read_u64
,
1887 .write_u64
= cpuset_write_u64
,
1888 .private = FILE_SPREAD_PAGE
,
1892 .name
= "memory_spread_slab",
1893 .read_u64
= cpuset_read_u64
,
1894 .write_u64
= cpuset_write_u64
,
1895 .private = FILE_SPREAD_SLAB
,
1899 .name
= "memory_pressure_enabled",
1900 .flags
= CFTYPE_ONLY_ON_ROOT
,
1901 .read_u64
= cpuset_read_u64
,
1902 .write_u64
= cpuset_write_u64
,
1903 .private = FILE_MEMORY_PRESSURE_ENABLED
,
1910 * cpuset_css_alloc - allocate a cpuset css
1911 * cgrp: control group that the new cpuset will be part of
1914 static struct cgroup_subsys_state
*
1915 cpuset_css_alloc(struct cgroup_subsys_state
*parent_css
)
1920 return &top_cpuset
.css
;
1922 cs
= kzalloc(sizeof(*cs
), GFP_KERNEL
);
1924 return ERR_PTR(-ENOMEM
);
1925 if (!alloc_cpumask_var(&cs
->cpus_allowed
, GFP_KERNEL
))
1927 if (!alloc_cpumask_var(&cs
->effective_cpus
, GFP_KERNEL
))
1930 set_bit(CS_SCHED_LOAD_BALANCE
, &cs
->flags
);
1931 cpumask_clear(cs
->cpus_allowed
);
1932 nodes_clear(cs
->mems_allowed
);
1933 cpumask_clear(cs
->effective_cpus
);
1934 nodes_clear(cs
->effective_mems
);
1935 fmeter_init(&cs
->fmeter
);
1936 cs
->relax_domain_level
= -1;
1941 free_cpumask_var(cs
->cpus_allowed
);
1944 return ERR_PTR(-ENOMEM
);
1947 static int cpuset_css_online(struct cgroup_subsys_state
*css
)
1949 struct cpuset
*cs
= css_cs(css
);
1950 struct cpuset
*parent
= parent_cs(cs
);
1951 struct cpuset
*tmp_cs
;
1952 struct cgroup_subsys_state
*pos_css
;
1957 mutex_lock(&cpuset_mutex
);
1959 set_bit(CS_ONLINE
, &cs
->flags
);
1960 if (is_spread_page(parent
))
1961 set_bit(CS_SPREAD_PAGE
, &cs
->flags
);
1962 if (is_spread_slab(parent
))
1963 set_bit(CS_SPREAD_SLAB
, &cs
->flags
);
1967 spin_lock_irq(&callback_lock
);
1968 if (cgroup_subsys_on_dfl(cpuset_cgrp_subsys
)) {
1969 cpumask_copy(cs
->effective_cpus
, parent
->effective_cpus
);
1970 cs
->effective_mems
= parent
->effective_mems
;
1972 spin_unlock_irq(&callback_lock
);
1974 if (!test_bit(CGRP_CPUSET_CLONE_CHILDREN
, &css
->cgroup
->flags
))
1978 * Clone @parent's configuration if CGRP_CPUSET_CLONE_CHILDREN is
1979 * set. This flag handling is implemented in cgroup core for
1980 * histrical reasons - the flag may be specified during mount.
1982 * Currently, if any sibling cpusets have exclusive cpus or mem, we
1983 * refuse to clone the configuration - thereby refusing the task to
1984 * be entered, and as a result refusing the sys_unshare() or
1985 * clone() which initiated it. If this becomes a problem for some
1986 * users who wish to allow that scenario, then this could be
1987 * changed to grant parent->cpus_allowed-sibling_cpus_exclusive
1988 * (and likewise for mems) to the new cgroup.
1991 cpuset_for_each_child(tmp_cs
, pos_css
, parent
) {
1992 if (is_mem_exclusive(tmp_cs
) || is_cpu_exclusive(tmp_cs
)) {
1999 spin_lock_irq(&callback_lock
);
2000 cs
->mems_allowed
= parent
->mems_allowed
;
2001 cs
->effective_mems
= parent
->mems_allowed
;
2002 cpumask_copy(cs
->cpus_allowed
, parent
->cpus_allowed
);
2003 cpumask_copy(cs
->effective_cpus
, parent
->cpus_allowed
);
2004 spin_unlock_irq(&callback_lock
);
2006 mutex_unlock(&cpuset_mutex
);
2011 * If the cpuset being removed has its flag 'sched_load_balance'
2012 * enabled, then simulate turning sched_load_balance off, which
2013 * will call rebuild_sched_domains_locked().
2016 static void cpuset_css_offline(struct cgroup_subsys_state
*css
)
2018 struct cpuset
*cs
= css_cs(css
);
2020 mutex_lock(&cpuset_mutex
);
2022 if (is_sched_load_balance(cs
))
2023 update_flag(CS_SCHED_LOAD_BALANCE
, cs
, 0);
2026 clear_bit(CS_ONLINE
, &cs
->flags
);
2028 mutex_unlock(&cpuset_mutex
);
2031 static void cpuset_css_free(struct cgroup_subsys_state
*css
)
2033 struct cpuset
*cs
= css_cs(css
);
2035 free_cpumask_var(cs
->effective_cpus
);
2036 free_cpumask_var(cs
->cpus_allowed
);
2040 static void cpuset_bind(struct cgroup_subsys_state
*root_css
)
2042 mutex_lock(&cpuset_mutex
);
2043 spin_lock_irq(&callback_lock
);
2045 if (cgroup_subsys_on_dfl(cpuset_cgrp_subsys
)) {
2046 cpumask_copy(top_cpuset
.cpus_allowed
, cpu_possible_mask
);
2047 top_cpuset
.mems_allowed
= node_possible_map
;
2049 cpumask_copy(top_cpuset
.cpus_allowed
,
2050 top_cpuset
.effective_cpus
);
2051 top_cpuset
.mems_allowed
= top_cpuset
.effective_mems
;
2054 spin_unlock_irq(&callback_lock
);
2055 mutex_unlock(&cpuset_mutex
);
2058 struct cgroup_subsys cpuset_cgrp_subsys
= {
2059 .css_alloc
= cpuset_css_alloc
,
2060 .css_online
= cpuset_css_online
,
2061 .css_offline
= cpuset_css_offline
,
2062 .css_free
= cpuset_css_free
,
2063 .can_attach
= cpuset_can_attach
,
2064 .cancel_attach
= cpuset_cancel_attach
,
2065 .attach
= cpuset_attach
,
2066 .bind
= cpuset_bind
,
2067 .legacy_cftypes
= files
,
2072 * cpuset_init - initialize cpusets at system boot
2074 * Description: Initialize top_cpuset and the cpuset internal file system,
2077 int __init
cpuset_init(void)
2081 if (!alloc_cpumask_var(&top_cpuset
.cpus_allowed
, GFP_KERNEL
))
2083 if (!alloc_cpumask_var(&top_cpuset
.effective_cpus
, GFP_KERNEL
))
2086 cpumask_setall(top_cpuset
.cpus_allowed
);
2087 nodes_setall(top_cpuset
.mems_allowed
);
2088 cpumask_setall(top_cpuset
.effective_cpus
);
2089 nodes_setall(top_cpuset
.effective_mems
);
2091 fmeter_init(&top_cpuset
.fmeter
);
2092 set_bit(CS_SCHED_LOAD_BALANCE
, &top_cpuset
.flags
);
2093 top_cpuset
.relax_domain_level
= -1;
2095 err
= register_filesystem(&cpuset_fs_type
);
2099 if (!alloc_cpumask_var(&cpus_attach
, GFP_KERNEL
))
2106 * If CPU and/or memory hotplug handlers, below, unplug any CPUs
2107 * or memory nodes, we need to walk over the cpuset hierarchy,
2108 * removing that CPU or node from all cpusets. If this removes the
2109 * last CPU or node from a cpuset, then move the tasks in the empty
2110 * cpuset to its next-highest non-empty parent.
2112 static void remove_tasks_in_empty_cpuset(struct cpuset
*cs
)
2114 struct cpuset
*parent
;
2117 * Find its next-highest non-empty parent, (top cpuset
2118 * has online cpus, so can't be empty).
2120 parent
= parent_cs(cs
);
2121 while (cpumask_empty(parent
->cpus_allowed
) ||
2122 nodes_empty(parent
->mems_allowed
))
2123 parent
= parent_cs(parent
);
2125 if (cgroup_transfer_tasks(parent
->css
.cgroup
, cs
->css
.cgroup
)) {
2126 pr_err("cpuset: failed to transfer tasks out of empty cpuset ");
2127 pr_cont_cgroup_name(cs
->css
.cgroup
);
2133 hotplug_update_tasks_legacy(struct cpuset
*cs
,
2134 struct cpumask
*new_cpus
, nodemask_t
*new_mems
,
2135 bool cpus_updated
, bool mems_updated
)
2139 spin_lock_irq(&callback_lock
);
2140 cpumask_copy(cs
->cpus_allowed
, new_cpus
);
2141 cpumask_copy(cs
->effective_cpus
, new_cpus
);
2142 cs
->mems_allowed
= *new_mems
;
2143 cs
->effective_mems
= *new_mems
;
2144 spin_unlock_irq(&callback_lock
);
2147 * Don't call update_tasks_cpumask() if the cpuset becomes empty,
2148 * as the tasks will be migratecd to an ancestor.
2150 if (cpus_updated
&& !cpumask_empty(cs
->cpus_allowed
))
2151 update_tasks_cpumask(cs
);
2152 if (mems_updated
&& !nodes_empty(cs
->mems_allowed
))
2153 update_tasks_nodemask(cs
);
2155 is_empty
= cpumask_empty(cs
->cpus_allowed
) ||
2156 nodes_empty(cs
->mems_allowed
);
2158 mutex_unlock(&cpuset_mutex
);
2161 * Move tasks to the nearest ancestor with execution resources,
2162 * This is full cgroup operation which will also call back into
2163 * cpuset. Should be done outside any lock.
2166 remove_tasks_in_empty_cpuset(cs
);
2168 mutex_lock(&cpuset_mutex
);
2172 hotplug_update_tasks(struct cpuset
*cs
,
2173 struct cpumask
*new_cpus
, nodemask_t
*new_mems
,
2174 bool cpus_updated
, bool mems_updated
)
2176 if (cpumask_empty(new_cpus
))
2177 cpumask_copy(new_cpus
, parent_cs(cs
)->effective_cpus
);
2178 if (nodes_empty(*new_mems
))
2179 *new_mems
= parent_cs(cs
)->effective_mems
;
2181 spin_lock_irq(&callback_lock
);
2182 cpumask_copy(cs
->effective_cpus
, new_cpus
);
2183 cs
->effective_mems
= *new_mems
;
2184 spin_unlock_irq(&callback_lock
);
2187 update_tasks_cpumask(cs
);
2189 update_tasks_nodemask(cs
);
2193 * cpuset_hotplug_update_tasks - update tasks in a cpuset for hotunplug
2194 * @cs: cpuset in interest
2196 * Compare @cs's cpu and mem masks against top_cpuset and if some have gone
2197 * offline, update @cs accordingly. If @cs ends up with no CPU or memory,
2198 * all its tasks are moved to the nearest ancestor with both resources.
2200 static void cpuset_hotplug_update_tasks(struct cpuset
*cs
)
2202 static cpumask_t new_cpus
;
2203 static nodemask_t new_mems
;
2207 wait_event(cpuset_attach_wq
, cs
->attach_in_progress
== 0);
2209 mutex_lock(&cpuset_mutex
);
2212 * We have raced with task attaching. We wait until attaching
2213 * is finished, so we won't attach a task to an empty cpuset.
2215 if (cs
->attach_in_progress
) {
2216 mutex_unlock(&cpuset_mutex
);
2220 cpumask_and(&new_cpus
, cs
->cpus_allowed
, parent_cs(cs
)->effective_cpus
);
2221 nodes_and(new_mems
, cs
->mems_allowed
, parent_cs(cs
)->effective_mems
);
2223 cpus_updated
= !cpumask_equal(&new_cpus
, cs
->effective_cpus
);
2224 mems_updated
= !nodes_equal(new_mems
, cs
->effective_mems
);
2226 if (cgroup_subsys_on_dfl(cpuset_cgrp_subsys
))
2227 hotplug_update_tasks(cs
, &new_cpus
, &new_mems
,
2228 cpus_updated
, mems_updated
);
2230 hotplug_update_tasks_legacy(cs
, &new_cpus
, &new_mems
,
2231 cpus_updated
, mems_updated
);
2233 mutex_unlock(&cpuset_mutex
);
2237 * cpuset_hotplug_workfn - handle CPU/memory hotunplug for a cpuset
2239 * This function is called after either CPU or memory configuration has
2240 * changed and updates cpuset accordingly. The top_cpuset is always
2241 * synchronized to cpu_active_mask and N_MEMORY, which is necessary in
2242 * order to make cpusets transparent (of no affect) on systems that are
2243 * actively using CPU hotplug but making no active use of cpusets.
2245 * Non-root cpusets are only affected by offlining. If any CPUs or memory
2246 * nodes have been taken down, cpuset_hotplug_update_tasks() is invoked on
2249 * Note that CPU offlining during suspend is ignored. We don't modify
2250 * cpusets across suspend/resume cycles at all.
2252 static void cpuset_hotplug_workfn(struct work_struct
*work
)
2254 static cpumask_t new_cpus
;
2255 static nodemask_t new_mems
;
2256 bool cpus_updated
, mems_updated
;
2257 bool on_dfl
= cgroup_subsys_on_dfl(cpuset_cgrp_subsys
);
2259 mutex_lock(&cpuset_mutex
);
2261 /* fetch the available cpus/mems and find out which changed how */
2262 cpumask_copy(&new_cpus
, cpu_active_mask
);
2263 new_mems
= node_states
[N_MEMORY
];
2265 cpus_updated
= !cpumask_equal(top_cpuset
.effective_cpus
, &new_cpus
);
2266 mems_updated
= !nodes_equal(top_cpuset
.effective_mems
, new_mems
);
2268 /* synchronize cpus_allowed to cpu_active_mask */
2270 spin_lock_irq(&callback_lock
);
2272 cpumask_copy(top_cpuset
.cpus_allowed
, &new_cpus
);
2273 cpumask_copy(top_cpuset
.effective_cpus
, &new_cpus
);
2274 spin_unlock_irq(&callback_lock
);
2275 /* we don't mess with cpumasks of tasks in top_cpuset */
2278 /* synchronize mems_allowed to N_MEMORY */
2280 spin_lock_irq(&callback_lock
);
2282 top_cpuset
.mems_allowed
= new_mems
;
2283 top_cpuset
.effective_mems
= new_mems
;
2284 spin_unlock_irq(&callback_lock
);
2285 update_tasks_nodemask(&top_cpuset
);
2288 mutex_unlock(&cpuset_mutex
);
2290 /* if cpus or mems changed, we need to propagate to descendants */
2291 if (cpus_updated
|| mems_updated
) {
2293 struct cgroup_subsys_state
*pos_css
;
2296 cpuset_for_each_descendant_pre(cs
, pos_css
, &top_cpuset
) {
2297 if (cs
== &top_cpuset
|| !css_tryget_online(&cs
->css
))
2301 cpuset_hotplug_update_tasks(cs
);
2309 /* rebuild sched domains if cpus_allowed has changed */
2311 rebuild_sched_domains();
2314 void cpuset_update_active_cpus(bool cpu_online
)
2317 * We're inside cpu hotplug critical region which usually nests
2318 * inside cgroup synchronization. Bounce actual hotplug processing
2319 * to a work item to avoid reverse locking order.
2321 * We still need to do partition_sched_domains() synchronously;
2322 * otherwise, the scheduler will get confused and put tasks to the
2323 * dead CPU. Fall back to the default single domain.
2324 * cpuset_hotplug_workfn() will rebuild it as necessary.
2326 partition_sched_domains(1, NULL
, NULL
);
2327 schedule_work(&cpuset_hotplug_work
);
2331 * Keep top_cpuset.mems_allowed tracking node_states[N_MEMORY].
2332 * Call this routine anytime after node_states[N_MEMORY] changes.
2333 * See cpuset_update_active_cpus() for CPU hotplug handling.
2335 static int cpuset_track_online_nodes(struct notifier_block
*self
,
2336 unsigned long action
, void *arg
)
2338 schedule_work(&cpuset_hotplug_work
);
2342 static struct notifier_block cpuset_track_online_nodes_nb
= {
2343 .notifier_call
= cpuset_track_online_nodes
,
2344 .priority
= 10, /* ??! */
2348 * cpuset_init_smp - initialize cpus_allowed
2350 * Description: Finish top cpuset after cpu, node maps are initialized
2352 void __init
cpuset_init_smp(void)
2354 cpumask_copy(top_cpuset
.cpus_allowed
, cpu_active_mask
);
2355 top_cpuset
.mems_allowed
= node_states
[N_MEMORY
];
2356 top_cpuset
.old_mems_allowed
= top_cpuset
.mems_allowed
;
2358 cpumask_copy(top_cpuset
.effective_cpus
, cpu_active_mask
);
2359 top_cpuset
.effective_mems
= node_states
[N_MEMORY
];
2361 register_hotmemory_notifier(&cpuset_track_online_nodes_nb
);
2365 * cpuset_cpus_allowed - return cpus_allowed mask from a tasks cpuset.
2366 * @tsk: pointer to task_struct from which to obtain cpuset->cpus_allowed.
2367 * @pmask: pointer to struct cpumask variable to receive cpus_allowed set.
2369 * Description: Returns the cpumask_var_t cpus_allowed of the cpuset
2370 * attached to the specified @tsk. Guaranteed to return some non-empty
2371 * subset of cpu_online_mask, even if this means going outside the
2375 void cpuset_cpus_allowed(struct task_struct
*tsk
, struct cpumask
*pmask
)
2377 unsigned long flags
;
2379 spin_lock_irqsave(&callback_lock
, flags
);
2381 guarantee_online_cpus(task_cs(tsk
), pmask
);
2383 spin_unlock_irqrestore(&callback_lock
, flags
);
2386 void cpuset_cpus_allowed_fallback(struct task_struct
*tsk
)
2389 do_set_cpus_allowed(tsk
, task_cs(tsk
)->effective_cpus
);
2393 * We own tsk->cpus_allowed, nobody can change it under us.
2395 * But we used cs && cs->cpus_allowed lockless and thus can
2396 * race with cgroup_attach_task() or update_cpumask() and get
2397 * the wrong tsk->cpus_allowed. However, both cases imply the
2398 * subsequent cpuset_change_cpumask()->set_cpus_allowed_ptr()
2399 * which takes task_rq_lock().
2401 * If we are called after it dropped the lock we must see all
2402 * changes in tsk_cs()->cpus_allowed. Otherwise we can temporary
2403 * set any mask even if it is not right from task_cs() pov,
2404 * the pending set_cpus_allowed_ptr() will fix things.
2406 * select_fallback_rq() will fix things ups and set cpu_possible_mask
2411 void __init
cpuset_init_current_mems_allowed(void)
2413 nodes_setall(current
->mems_allowed
);
2417 * cpuset_mems_allowed - return mems_allowed mask from a tasks cpuset.
2418 * @tsk: pointer to task_struct from which to obtain cpuset->mems_allowed.
2420 * Description: Returns the nodemask_t mems_allowed of the cpuset
2421 * attached to the specified @tsk. Guaranteed to return some non-empty
2422 * subset of node_states[N_MEMORY], even if this means going outside the
2426 nodemask_t
cpuset_mems_allowed(struct task_struct
*tsk
)
2429 unsigned long flags
;
2431 spin_lock_irqsave(&callback_lock
, flags
);
2433 guarantee_online_mems(task_cs(tsk
), &mask
);
2435 spin_unlock_irqrestore(&callback_lock
, flags
);
2441 * cpuset_nodemask_valid_mems_allowed - check nodemask vs. curremt mems_allowed
2442 * @nodemask: the nodemask to be checked
2444 * Are any of the nodes in the nodemask allowed in current->mems_allowed?
2446 int cpuset_nodemask_valid_mems_allowed(nodemask_t
*nodemask
)
2448 return nodes_intersects(*nodemask
, current
->mems_allowed
);
2452 * nearest_hardwall_ancestor() - Returns the nearest mem_exclusive or
2453 * mem_hardwall ancestor to the specified cpuset. Call holding
2454 * callback_lock. If no ancestor is mem_exclusive or mem_hardwall
2455 * (an unusual configuration), then returns the root cpuset.
2457 static struct cpuset
*nearest_hardwall_ancestor(struct cpuset
*cs
)
2459 while (!(is_mem_exclusive(cs
) || is_mem_hardwall(cs
)) && parent_cs(cs
))
2465 * cpuset_node_allowed - Can we allocate on a memory node?
2466 * @node: is this an allowed node?
2467 * @gfp_mask: memory allocation flags
2469 * If we're in interrupt, yes, we can always allocate. If @node is set in
2470 * current's mems_allowed, yes. If it's not a __GFP_HARDWALL request and this
2471 * node is set in the nearest hardwalled cpuset ancestor to current's cpuset,
2472 * yes. If current has access to memory reserves due to TIF_MEMDIE, yes.
2475 * GFP_USER allocations are marked with the __GFP_HARDWALL bit,
2476 * and do not allow allocations outside the current tasks cpuset
2477 * unless the task has been OOM killed as is marked TIF_MEMDIE.
2478 * GFP_KERNEL allocations are not so marked, so can escape to the
2479 * nearest enclosing hardwalled ancestor cpuset.
2481 * Scanning up parent cpusets requires callback_lock. The
2482 * __alloc_pages() routine only calls here with __GFP_HARDWALL bit
2483 * _not_ set if it's a GFP_KERNEL allocation, and all nodes in the
2484 * current tasks mems_allowed came up empty on the first pass over
2485 * the zonelist. So only GFP_KERNEL allocations, if all nodes in the
2486 * cpuset are short of memory, might require taking the callback_lock.
2488 * The first call here from mm/page_alloc:get_page_from_freelist()
2489 * has __GFP_HARDWALL set in gfp_mask, enforcing hardwall cpusets,
2490 * so no allocation on a node outside the cpuset is allowed (unless
2491 * in interrupt, of course).
2493 * The second pass through get_page_from_freelist() doesn't even call
2494 * here for GFP_ATOMIC calls. For those calls, the __alloc_pages()
2495 * variable 'wait' is not set, and the bit ALLOC_CPUSET is not set
2496 * in alloc_flags. That logic and the checks below have the combined
2498 * in_interrupt - any node ok (current task context irrelevant)
2499 * GFP_ATOMIC - any node ok
2500 * TIF_MEMDIE - any node ok
2501 * GFP_KERNEL - any node in enclosing hardwalled cpuset ok
2502 * GFP_USER - only nodes in current tasks mems allowed ok.
2504 int __cpuset_node_allowed(int node
, gfp_t gfp_mask
)
2506 struct cpuset
*cs
; /* current cpuset ancestors */
2507 int allowed
; /* is allocation in zone z allowed? */
2508 unsigned long flags
;
2512 if (node_isset(node
, current
->mems_allowed
))
2515 * Allow tasks that have access to memory reserves because they have
2516 * been OOM killed to get memory anywhere.
2518 if (unlikely(test_thread_flag(TIF_MEMDIE
)))
2520 if (gfp_mask
& __GFP_HARDWALL
) /* If hardwall request, stop here */
2523 if (current
->flags
& PF_EXITING
) /* Let dying task have memory */
2526 /* Not hardwall and node outside mems_allowed: scan up cpusets */
2527 spin_lock_irqsave(&callback_lock
, flags
);
2530 cs
= nearest_hardwall_ancestor(task_cs(current
));
2531 allowed
= node_isset(node
, cs
->mems_allowed
);
2534 spin_unlock_irqrestore(&callback_lock
, flags
);
2539 * cpuset_mem_spread_node() - On which node to begin search for a file page
2540 * cpuset_slab_spread_node() - On which node to begin search for a slab page
2542 * If a task is marked PF_SPREAD_PAGE or PF_SPREAD_SLAB (as for
2543 * tasks in a cpuset with is_spread_page or is_spread_slab set),
2544 * and if the memory allocation used cpuset_mem_spread_node()
2545 * to determine on which node to start looking, as it will for
2546 * certain page cache or slab cache pages such as used for file
2547 * system buffers and inode caches, then instead of starting on the
2548 * local node to look for a free page, rather spread the starting
2549 * node around the tasks mems_allowed nodes.
2551 * We don't have to worry about the returned node being offline
2552 * because "it can't happen", and even if it did, it would be ok.
2554 * The routines calling guarantee_online_mems() are careful to
2555 * only set nodes in task->mems_allowed that are online. So it
2556 * should not be possible for the following code to return an
2557 * offline node. But if it did, that would be ok, as this routine
2558 * is not returning the node where the allocation must be, only
2559 * the node where the search should start. The zonelist passed to
2560 * __alloc_pages() will include all nodes. If the slab allocator
2561 * is passed an offline node, it will fall back to the local node.
2562 * See kmem_cache_alloc_node().
2565 static int cpuset_spread_node(int *rotor
)
2569 node
= next_node(*rotor
, current
->mems_allowed
);
2570 if (node
== MAX_NUMNODES
)
2571 node
= first_node(current
->mems_allowed
);
2576 int cpuset_mem_spread_node(void)
2578 if (current
->cpuset_mem_spread_rotor
== NUMA_NO_NODE
)
2579 current
->cpuset_mem_spread_rotor
=
2580 node_random(¤t
->mems_allowed
);
2582 return cpuset_spread_node(¤t
->cpuset_mem_spread_rotor
);
2585 int cpuset_slab_spread_node(void)
2587 if (current
->cpuset_slab_spread_rotor
== NUMA_NO_NODE
)
2588 current
->cpuset_slab_spread_rotor
=
2589 node_random(¤t
->mems_allowed
);
2591 return cpuset_spread_node(¤t
->cpuset_slab_spread_rotor
);
2594 EXPORT_SYMBOL_GPL(cpuset_mem_spread_node
);
2597 * cpuset_mems_allowed_intersects - Does @tsk1's mems_allowed intersect @tsk2's?
2598 * @tsk1: pointer to task_struct of some task.
2599 * @tsk2: pointer to task_struct of some other task.
2601 * Description: Return true if @tsk1's mems_allowed intersects the
2602 * mems_allowed of @tsk2. Used by the OOM killer to determine if
2603 * one of the task's memory usage might impact the memory available
2607 int cpuset_mems_allowed_intersects(const struct task_struct
*tsk1
,
2608 const struct task_struct
*tsk2
)
2610 return nodes_intersects(tsk1
->mems_allowed
, tsk2
->mems_allowed
);
2614 * cpuset_print_current_mems_allowed - prints current's cpuset and mems_allowed
2616 * Description: Prints current's name, cpuset name, and cached copy of its
2617 * mems_allowed to the kernel log.
2619 void cpuset_print_current_mems_allowed(void)
2621 struct cgroup
*cgrp
;
2625 cgrp
= task_cs(current
)->css
.cgroup
;
2626 pr_info("%s cpuset=", current
->comm
);
2627 pr_cont_cgroup_name(cgrp
);
2628 pr_cont(" mems_allowed=%*pbl\n",
2629 nodemask_pr_args(¤t
->mems_allowed
));
2635 * Collection of memory_pressure is suppressed unless
2636 * this flag is enabled by writing "1" to the special
2637 * cpuset file 'memory_pressure_enabled' in the root cpuset.
2640 int cpuset_memory_pressure_enabled __read_mostly
;
2643 * cpuset_memory_pressure_bump - keep stats of per-cpuset reclaims.
2645 * Keep a running average of the rate of synchronous (direct)
2646 * page reclaim efforts initiated by tasks in each cpuset.
2648 * This represents the rate at which some task in the cpuset
2649 * ran low on memory on all nodes it was allowed to use, and
2650 * had to enter the kernels page reclaim code in an effort to
2651 * create more free memory by tossing clean pages or swapping
2652 * or writing dirty pages.
2654 * Display to user space in the per-cpuset read-only file
2655 * "memory_pressure". Value displayed is an integer
2656 * representing the recent rate of entry into the synchronous
2657 * (direct) page reclaim by any task attached to the cpuset.
2660 void __cpuset_memory_pressure_bump(void)
2663 fmeter_markevent(&task_cs(current
)->fmeter
);
2667 #ifdef CONFIG_PROC_PID_CPUSET
2669 * proc_cpuset_show()
2670 * - Print tasks cpuset path into seq_file.
2671 * - Used for /proc/<pid>/cpuset.
2672 * - No need to task_lock(tsk) on this tsk->cpuset reference, as it
2673 * doesn't really matter if tsk->cpuset changes after we read it,
2674 * and we take cpuset_mutex, keeping cpuset_attach() from changing it
2677 int proc_cpuset_show(struct seq_file
*m
, struct pid_namespace
*ns
,
2678 struct pid
*pid
, struct task_struct
*tsk
)
2681 struct cgroup_subsys_state
*css
;
2685 buf
= kmalloc(PATH_MAX
, GFP_KERNEL
);
2689 retval
= -ENAMETOOLONG
;
2691 css
= task_css(tsk
, cpuset_cgrp_id
);
2692 p
= cgroup_path(css
->cgroup
, buf
, PATH_MAX
);
2704 #endif /* CONFIG_PROC_PID_CPUSET */
2706 /* Display task mems_allowed in /proc/<pid>/status file. */
2707 void cpuset_task_status_allowed(struct seq_file
*m
, struct task_struct
*task
)
2709 seq_printf(m
, "Mems_allowed:\t%*pb\n",
2710 nodemask_pr_args(&task
->mems_allowed
));
2711 seq_printf(m
, "Mems_allowed_list:\t%*pbl\n",
2712 nodemask_pr_args(&task
->mems_allowed
));