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>
59 #include <linux/oom.h>
60 #include <linux/sched/isolation.h>
61 #include <linux/uaccess.h>
62 #include <linux/atomic.h>
63 #include <linux/mutex.h>
64 #include <linux/cgroup.h>
65 #include <linux/wait.h>
67 DEFINE_STATIC_KEY_FALSE(cpusets_pre_enable_key
);
68 DEFINE_STATIC_KEY_FALSE(cpusets_enabled_key
);
70 /* See "Frequency meter" comments, below. */
73 int cnt
; /* unprocessed events count */
74 int val
; /* most recent output value */
75 time64_t time
; /* clock (secs) when val computed */
76 spinlock_t lock
; /* guards read or write of above */
80 struct cgroup_subsys_state css
;
82 unsigned long flags
; /* "unsigned long" so bitops work */
85 * On default hierarchy:
87 * The user-configured masks can only be changed by writing to
88 * cpuset.cpus and cpuset.mems, and won't be limited by the
91 * The effective masks is the real masks that apply to the tasks
92 * in the cpuset. They may be changed if the configured masks are
93 * changed or hotplug happens.
95 * effective_mask == configured_mask & parent's effective_mask,
96 * and if it ends up empty, it will inherit the parent's mask.
101 * The user-configured masks are always the same with effective masks.
104 /* user-configured CPUs and Memory Nodes allow to tasks */
105 cpumask_var_t cpus_allowed
;
106 nodemask_t mems_allowed
;
108 /* effective CPUs and Memory Nodes allow to tasks */
109 cpumask_var_t effective_cpus
;
110 nodemask_t effective_mems
;
113 * CPUs allocated to child sub-partitions (default hierarchy only)
114 * - CPUs granted by the parent = effective_cpus U subparts_cpus
115 * - effective_cpus and subparts_cpus are mutually exclusive.
117 cpumask_var_t subparts_cpus
;
120 * This is old Memory Nodes tasks took on.
122 * - top_cpuset.old_mems_allowed is initialized to mems_allowed.
123 * - A new cpuset's old_mems_allowed is initialized when some
124 * task is moved into it.
125 * - old_mems_allowed is used in cpuset_migrate_mm() when we change
126 * cpuset.mems_allowed and have tasks' nodemask updated, and
127 * then old_mems_allowed is updated to mems_allowed.
129 nodemask_t old_mems_allowed
;
131 struct fmeter fmeter
; /* memory_pressure filter */
134 * Tasks are being attached to this cpuset. Used to prevent
135 * zeroing cpus/mems_allowed between ->can_attach() and ->attach().
137 int attach_in_progress
;
139 /* partition number for rebuild_sched_domains() */
142 /* for custom sched domain */
143 int relax_domain_level
;
145 /* number of CPUs in subparts_cpus */
146 int nr_subparts_cpus
;
148 /* partition root state */
149 int partition_root_state
;
153 * Partition root states:
155 * 0 - not a partition root
159 * -1 - invalid partition root
160 * None of the cpus in cpus_allowed can be put into the parent's
161 * subparts_cpus. In this case, the cpuset is not a real partition
162 * root anymore. However, the CPU_EXCLUSIVE bit will still be set
163 * and the cpuset can be restored back to a partition root if the
164 * parent cpuset can give more CPUs back to this child cpuset.
166 #define PRS_DISABLED 0
167 #define PRS_ENABLED 1
171 * Temporary cpumasks for working with partitions that are passed among
172 * functions to avoid memory allocation in inner functions.
175 cpumask_var_t addmask
, delmask
; /* For partition root */
176 cpumask_var_t new_cpus
; /* For update_cpumasks_hier() */
179 static inline struct cpuset
*css_cs(struct cgroup_subsys_state
*css
)
181 return css
? container_of(css
, struct cpuset
, css
) : NULL
;
184 /* Retrieve the cpuset for a task */
185 static inline struct cpuset
*task_cs(struct task_struct
*task
)
187 return css_cs(task_css(task
, cpuset_cgrp_id
));
190 static inline struct cpuset
*parent_cs(struct cpuset
*cs
)
192 return css_cs(cs
->css
.parent
);
196 static inline bool task_has_mempolicy(struct task_struct
*task
)
198 return task
->mempolicy
;
201 static inline bool task_has_mempolicy(struct task_struct
*task
)
208 /* bits in struct cpuset flags field */
215 CS_SCHED_LOAD_BALANCE
,
220 /* convenient tests for these bits */
221 static inline bool is_cpuset_online(struct cpuset
*cs
)
223 return test_bit(CS_ONLINE
, &cs
->flags
) && !css_is_dying(&cs
->css
);
226 static inline int is_cpu_exclusive(const struct cpuset
*cs
)
228 return test_bit(CS_CPU_EXCLUSIVE
, &cs
->flags
);
231 static inline int is_mem_exclusive(const struct cpuset
*cs
)
233 return test_bit(CS_MEM_EXCLUSIVE
, &cs
->flags
);
236 static inline int is_mem_hardwall(const struct cpuset
*cs
)
238 return test_bit(CS_MEM_HARDWALL
, &cs
->flags
);
241 static inline int is_sched_load_balance(const struct cpuset
*cs
)
243 return test_bit(CS_SCHED_LOAD_BALANCE
, &cs
->flags
);
246 static inline int is_memory_migrate(const struct cpuset
*cs
)
248 return test_bit(CS_MEMORY_MIGRATE
, &cs
->flags
);
251 static inline int is_spread_page(const struct cpuset
*cs
)
253 return test_bit(CS_SPREAD_PAGE
, &cs
->flags
);
256 static inline int is_spread_slab(const struct cpuset
*cs
)
258 return test_bit(CS_SPREAD_SLAB
, &cs
->flags
);
261 static inline int is_partition_root(const struct cpuset
*cs
)
263 return cs
->partition_root_state
> 0;
266 static struct cpuset top_cpuset
= {
267 .flags
= ((1 << CS_ONLINE
) | (1 << CS_CPU_EXCLUSIVE
) |
268 (1 << CS_MEM_EXCLUSIVE
)),
269 .partition_root_state
= PRS_ENABLED
,
273 * cpuset_for_each_child - traverse online children of a cpuset
274 * @child_cs: loop cursor pointing to the current child
275 * @pos_css: used for iteration
276 * @parent_cs: target cpuset to walk children of
278 * Walk @child_cs through the online children of @parent_cs. Must be used
279 * with RCU read locked.
281 #define cpuset_for_each_child(child_cs, pos_css, parent_cs) \
282 css_for_each_child((pos_css), &(parent_cs)->css) \
283 if (is_cpuset_online(((child_cs) = css_cs((pos_css)))))
286 * cpuset_for_each_descendant_pre - pre-order walk of a cpuset's descendants
287 * @des_cs: loop cursor pointing to the current descendant
288 * @pos_css: used for iteration
289 * @root_cs: target cpuset to walk ancestor of
291 * Walk @des_cs through the online descendants of @root_cs. Must be used
292 * with RCU read locked. The caller may modify @pos_css by calling
293 * css_rightmost_descendant() to skip subtree. @root_cs is included in the
294 * iteration and the first node to be visited.
296 #define cpuset_for_each_descendant_pre(des_cs, pos_css, root_cs) \
297 css_for_each_descendant_pre((pos_css), &(root_cs)->css) \
298 if (is_cpuset_online(((des_cs) = css_cs((pos_css)))))
301 * There are two global locks guarding cpuset structures - cpuset_mutex and
302 * callback_lock. We also require taking task_lock() when dereferencing a
303 * task's cpuset pointer. See "The task_lock() exception", at the end of this
306 * A task must hold both locks to modify cpusets. If a task holds
307 * cpuset_mutex, then it blocks others wanting that mutex, ensuring that it
308 * is the only task able to also acquire callback_lock and be able to
309 * modify cpusets. It can perform various checks on the cpuset structure
310 * first, knowing nothing will change. It can also allocate memory while
311 * just holding cpuset_mutex. While it is performing these checks, various
312 * callback routines can briefly acquire callback_lock to query cpusets.
313 * Once it is ready to make the changes, it takes callback_lock, blocking
316 * Calls to the kernel memory allocator can not be made while holding
317 * callback_lock, as that would risk double tripping on callback_lock
318 * from one of the callbacks into the cpuset code from within
321 * If a task is only holding callback_lock, then it has read-only
324 * Now, the task_struct fields mems_allowed and mempolicy may be changed
325 * by other task, we use alloc_lock in the task_struct fields to protect
328 * The cpuset_common_file_read() handlers only hold callback_lock across
329 * small pieces of code, such as when reading out possibly multi-word
330 * cpumasks and nodemasks.
332 * Accessing a task's cpuset should be done in accordance with the
333 * guidelines for accessing subsystem state in kernel/cgroup.c
336 static DEFINE_MUTEX(cpuset_mutex
);
337 static DEFINE_SPINLOCK(callback_lock
);
339 static struct workqueue_struct
*cpuset_migrate_mm_wq
;
342 * CPU / memory hotplug is handled asynchronously.
344 static void cpuset_hotplug_workfn(struct work_struct
*work
);
345 static DECLARE_WORK(cpuset_hotplug_work
, cpuset_hotplug_workfn
);
347 static DECLARE_WAIT_QUEUE_HEAD(cpuset_attach_wq
);
350 * Cgroup v2 behavior is used when on default hierarchy or the
351 * cgroup_v2_mode flag is set.
353 static inline bool is_in_v2_mode(void)
355 return cgroup_subsys_on_dfl(cpuset_cgrp_subsys
) ||
356 (cpuset_cgrp_subsys
.root
->flags
& CGRP_ROOT_CPUSET_V2_MODE
);
360 * This is ugly, but preserves the userspace API for existing cpuset
361 * users. If someone tries to mount the "cpuset" filesystem, we
362 * silently switch it to mount "cgroup" instead
364 static struct dentry
*cpuset_mount(struct file_system_type
*fs_type
,
365 int flags
, const char *unused_dev_name
, void *data
)
367 struct file_system_type
*cgroup_fs
= get_fs_type("cgroup");
368 struct dentry
*ret
= ERR_PTR(-ENODEV
);
372 "release_agent=/sbin/cpuset_release_agent";
373 ret
= cgroup_fs
->mount(cgroup_fs
, flags
,
374 unused_dev_name
, mountopts
);
375 put_filesystem(cgroup_fs
);
380 static struct file_system_type cpuset_fs_type
= {
382 .mount
= cpuset_mount
,
386 * Return in pmask the portion of a cpusets's cpus_allowed that
387 * are online. If none are online, walk up the cpuset hierarchy
388 * until we find one that does have some online cpus.
390 * One way or another, we guarantee to return some non-empty subset
391 * of cpu_online_mask.
393 * Call with callback_lock or cpuset_mutex held.
395 static void guarantee_online_cpus(struct cpuset
*cs
, struct cpumask
*pmask
)
397 while (!cpumask_intersects(cs
->effective_cpus
, cpu_online_mask
)) {
401 * The top cpuset doesn't have any online cpu as a
402 * consequence of a race between cpuset_hotplug_work
403 * and cpu hotplug notifier. But we know the top
404 * cpuset's effective_cpus is on its way to to be
405 * identical to cpu_online_mask.
407 cpumask_copy(pmask
, cpu_online_mask
);
411 cpumask_and(pmask
, cs
->effective_cpus
, cpu_online_mask
);
415 * Return in *pmask the portion of a cpusets's mems_allowed that
416 * are online, with memory. If none are online with memory, walk
417 * up the cpuset hierarchy until we find one that does have some
418 * online mems. The top cpuset always has some mems online.
420 * One way or another, we guarantee to return some non-empty subset
421 * of node_states[N_MEMORY].
423 * Call with callback_lock or cpuset_mutex held.
425 static void guarantee_online_mems(struct cpuset
*cs
, nodemask_t
*pmask
)
427 while (!nodes_intersects(cs
->effective_mems
, node_states
[N_MEMORY
]))
429 nodes_and(*pmask
, cs
->effective_mems
, node_states
[N_MEMORY
]);
433 * update task's spread flag if cpuset's page/slab spread flag is set
435 * Call with callback_lock or cpuset_mutex held.
437 static void cpuset_update_task_spread_flag(struct cpuset
*cs
,
438 struct task_struct
*tsk
)
440 if (is_spread_page(cs
))
441 task_set_spread_page(tsk
);
443 task_clear_spread_page(tsk
);
445 if (is_spread_slab(cs
))
446 task_set_spread_slab(tsk
);
448 task_clear_spread_slab(tsk
);
452 * is_cpuset_subset(p, q) - Is cpuset p a subset of cpuset q?
454 * One cpuset is a subset of another if all its allowed CPUs and
455 * Memory Nodes are a subset of the other, and its exclusive flags
456 * are only set if the other's are set. Call holding cpuset_mutex.
459 static int is_cpuset_subset(const struct cpuset
*p
, const struct cpuset
*q
)
461 return cpumask_subset(p
->cpus_allowed
, q
->cpus_allowed
) &&
462 nodes_subset(p
->mems_allowed
, q
->mems_allowed
) &&
463 is_cpu_exclusive(p
) <= is_cpu_exclusive(q
) &&
464 is_mem_exclusive(p
) <= is_mem_exclusive(q
);
468 * alloc_cpumasks - allocate three cpumasks for cpuset
469 * @cs: the cpuset that have cpumasks to be allocated.
470 * @tmp: the tmpmasks structure pointer
471 * Return: 0 if successful, -ENOMEM otherwise.
473 * Only one of the two input arguments should be non-NULL.
475 static inline int alloc_cpumasks(struct cpuset
*cs
, struct tmpmasks
*tmp
)
477 cpumask_var_t
*pmask1
, *pmask2
, *pmask3
;
480 pmask1
= &cs
->cpus_allowed
;
481 pmask2
= &cs
->effective_cpus
;
482 pmask3
= &cs
->subparts_cpus
;
484 pmask1
= &tmp
->new_cpus
;
485 pmask2
= &tmp
->addmask
;
486 pmask3
= &tmp
->delmask
;
489 if (!zalloc_cpumask_var(pmask1
, GFP_KERNEL
))
492 if (!zalloc_cpumask_var(pmask2
, GFP_KERNEL
))
495 if (!zalloc_cpumask_var(pmask3
, GFP_KERNEL
))
501 free_cpumask_var(*pmask2
);
503 free_cpumask_var(*pmask1
);
508 * free_cpumasks - free cpumasks in a tmpmasks structure
509 * @cs: the cpuset that have cpumasks to be free.
510 * @tmp: the tmpmasks structure pointer
512 static inline void free_cpumasks(struct cpuset
*cs
, struct tmpmasks
*tmp
)
515 free_cpumask_var(cs
->cpus_allowed
);
516 free_cpumask_var(cs
->effective_cpus
);
517 free_cpumask_var(cs
->subparts_cpus
);
520 free_cpumask_var(tmp
->new_cpus
);
521 free_cpumask_var(tmp
->addmask
);
522 free_cpumask_var(tmp
->delmask
);
527 * alloc_trial_cpuset - allocate a trial cpuset
528 * @cs: the cpuset that the trial cpuset duplicates
530 static struct cpuset
*alloc_trial_cpuset(struct cpuset
*cs
)
532 struct cpuset
*trial
;
534 trial
= kmemdup(cs
, sizeof(*cs
), GFP_KERNEL
);
538 if (alloc_cpumasks(trial
, NULL
)) {
543 cpumask_copy(trial
->cpus_allowed
, cs
->cpus_allowed
);
544 cpumask_copy(trial
->effective_cpus
, cs
->effective_cpus
);
549 * free_cpuset - free the cpuset
550 * @cs: the cpuset to be freed
552 static inline void free_cpuset(struct cpuset
*cs
)
554 free_cpumasks(cs
, NULL
);
559 * validate_change() - Used to validate that any proposed cpuset change
560 * follows the structural rules for cpusets.
562 * If we replaced the flag and mask values of the current cpuset
563 * (cur) with those values in the trial cpuset (trial), would
564 * our various subset and exclusive rules still be valid? Presumes
567 * 'cur' is the address of an actual, in-use cpuset. Operations
568 * such as list traversal that depend on the actual address of the
569 * cpuset in the list must use cur below, not trial.
571 * 'trial' is the address of bulk structure copy of cur, with
572 * perhaps one or more of the fields cpus_allowed, mems_allowed,
573 * or flags changed to new, trial values.
575 * Return 0 if valid, -errno if not.
578 static int validate_change(struct cpuset
*cur
, struct cpuset
*trial
)
580 struct cgroup_subsys_state
*css
;
581 struct cpuset
*c
, *par
;
586 /* Each of our child cpusets must be a subset of us */
588 cpuset_for_each_child(c
, css
, cur
)
589 if (!is_cpuset_subset(c
, trial
))
592 /* Remaining checks don't apply to root cpuset */
594 if (cur
== &top_cpuset
)
597 par
= parent_cs(cur
);
599 /* On legacy hiearchy, we must be a subset of our parent cpuset. */
601 if (!is_in_v2_mode() && !is_cpuset_subset(trial
, par
))
605 * If either I or some sibling (!= me) is exclusive, we can't
609 cpuset_for_each_child(c
, css
, par
) {
610 if ((is_cpu_exclusive(trial
) || is_cpu_exclusive(c
)) &&
612 cpumask_intersects(trial
->cpus_allowed
, c
->cpus_allowed
))
614 if ((is_mem_exclusive(trial
) || is_mem_exclusive(c
)) &&
616 nodes_intersects(trial
->mems_allowed
, c
->mems_allowed
))
621 * Cpusets with tasks - existing or newly being attached - can't
622 * be changed to have empty cpus_allowed or mems_allowed.
625 if ((cgroup_is_populated(cur
->css
.cgroup
) || cur
->attach_in_progress
)) {
626 if (!cpumask_empty(cur
->cpus_allowed
) &&
627 cpumask_empty(trial
->cpus_allowed
))
629 if (!nodes_empty(cur
->mems_allowed
) &&
630 nodes_empty(trial
->mems_allowed
))
635 * We can't shrink if we won't have enough room for SCHED_DEADLINE
639 if (is_cpu_exclusive(cur
) &&
640 !cpuset_cpumask_can_shrink(cur
->cpus_allowed
,
641 trial
->cpus_allowed
))
652 * Helper routine for generate_sched_domains().
653 * Do cpusets a, b have overlapping effective cpus_allowed masks?
655 static int cpusets_overlap(struct cpuset
*a
, struct cpuset
*b
)
657 return cpumask_intersects(a
->effective_cpus
, b
->effective_cpus
);
661 update_domain_attr(struct sched_domain_attr
*dattr
, struct cpuset
*c
)
663 if (dattr
->relax_domain_level
< c
->relax_domain_level
)
664 dattr
->relax_domain_level
= c
->relax_domain_level
;
668 static void update_domain_attr_tree(struct sched_domain_attr
*dattr
,
669 struct cpuset
*root_cs
)
672 struct cgroup_subsys_state
*pos_css
;
675 cpuset_for_each_descendant_pre(cp
, pos_css
, root_cs
) {
676 /* skip the whole subtree if @cp doesn't have any CPU */
677 if (cpumask_empty(cp
->cpus_allowed
)) {
678 pos_css
= css_rightmost_descendant(pos_css
);
682 if (is_sched_load_balance(cp
))
683 update_domain_attr(dattr
, cp
);
688 /* Must be called with cpuset_mutex held. */
689 static inline int nr_cpusets(void)
691 /* jump label reference count + the top-level cpuset */
692 return static_key_count(&cpusets_enabled_key
.key
) + 1;
696 * generate_sched_domains()
698 * This function builds a partial partition of the systems CPUs
699 * A 'partial partition' is a set of non-overlapping subsets whose
700 * union is a subset of that set.
701 * The output of this function needs to be passed to kernel/sched/core.c
702 * partition_sched_domains() routine, which will rebuild the scheduler's
703 * load balancing domains (sched domains) as specified by that partial
706 * See "What is sched_load_balance" in Documentation/cgroup-v1/cpusets.txt
707 * for a background explanation of this.
709 * Does not return errors, on the theory that the callers of this
710 * routine would rather not worry about failures to rebuild sched
711 * domains when operating in the severe memory shortage situations
712 * that could cause allocation failures below.
714 * Must be called with cpuset_mutex held.
716 * The three key local variables below are:
717 * q - a linked-list queue of cpuset pointers, used to implement a
718 * top-down scan of all cpusets. This scan loads a pointer
719 * to each cpuset marked is_sched_load_balance into the
720 * array 'csa'. For our purposes, rebuilding the schedulers
721 * sched domains, we can ignore !is_sched_load_balance cpusets.
722 * csa - (for CpuSet Array) Array of pointers to all the cpusets
723 * that need to be load balanced, for convenient iterative
724 * access by the subsequent code that finds the best partition,
725 * i.e the set of domains (subsets) of CPUs such that the
726 * cpus_allowed of every cpuset marked is_sched_load_balance
727 * is a subset of one of these domains, while there are as
728 * many such domains as possible, each as small as possible.
729 * doms - Conversion of 'csa' to an array of cpumasks, for passing to
730 * the kernel/sched/core.c routine partition_sched_domains() in a
731 * convenient format, that can be easily compared to the prior
732 * value to determine what partition elements (sched domains)
733 * were changed (added or removed.)
735 * Finding the best partition (set of domains):
736 * The triple nested loops below over i, j, k scan over the
737 * load balanced cpusets (using the array of cpuset pointers in
738 * csa[]) looking for pairs of cpusets that have overlapping
739 * cpus_allowed, but which don't have the same 'pn' partition
740 * number and gives them in the same partition number. It keeps
741 * looping on the 'restart' label until it can no longer find
744 * The union of the cpus_allowed masks from the set of
745 * all cpusets having the same 'pn' value then form the one
746 * element of the partition (one sched domain) to be passed to
747 * partition_sched_domains().
749 static int generate_sched_domains(cpumask_var_t
**domains
,
750 struct sched_domain_attr
**attributes
)
752 struct cpuset
*cp
; /* scans q */
753 struct cpuset
**csa
; /* array of all cpuset ptrs */
754 int csn
; /* how many cpuset ptrs in csa so far */
755 int i
, j
, k
; /* indices for partition finding loops */
756 cpumask_var_t
*doms
; /* resulting partition; i.e. sched domains */
757 struct sched_domain_attr
*dattr
; /* attributes for custom domains */
758 int ndoms
= 0; /* number of sched domains in result */
759 int nslot
; /* next empty doms[] struct cpumask slot */
760 struct cgroup_subsys_state
*pos_css
;
766 /* Special case for the 99% of systems with one, full, sched domain */
767 if (is_sched_load_balance(&top_cpuset
)) {
769 doms
= alloc_sched_domains(ndoms
);
773 dattr
= kmalloc(sizeof(struct sched_domain_attr
), GFP_KERNEL
);
775 *dattr
= SD_ATTR_INIT
;
776 update_domain_attr_tree(dattr
, &top_cpuset
);
778 cpumask_and(doms
[0], top_cpuset
.effective_cpus
,
779 housekeeping_cpumask(HK_FLAG_DOMAIN
));
784 csa
= kmalloc_array(nr_cpusets(), sizeof(cp
), GFP_KERNEL
);
790 cpuset_for_each_descendant_pre(cp
, pos_css
, &top_cpuset
) {
791 if (cp
== &top_cpuset
)
794 * Continue traversing beyond @cp iff @cp has some CPUs and
795 * isn't load balancing. The former is obvious. The
796 * latter: All child cpusets contain a subset of the
797 * parent's cpus, so just skip them, and then we call
798 * update_domain_attr_tree() to calc relax_domain_level of
799 * the corresponding sched domain.
801 if (!cpumask_empty(cp
->cpus_allowed
) &&
802 !(is_sched_load_balance(cp
) &&
803 cpumask_intersects(cp
->cpus_allowed
,
804 housekeeping_cpumask(HK_FLAG_DOMAIN
))))
807 if (is_sched_load_balance(cp
))
810 /* skip @cp's subtree */
811 pos_css
= css_rightmost_descendant(pos_css
);
815 for (i
= 0; i
< csn
; i
++)
820 /* Find the best partition (set of sched domains) */
821 for (i
= 0; i
< csn
; i
++) {
822 struct cpuset
*a
= csa
[i
];
825 for (j
= 0; j
< csn
; j
++) {
826 struct cpuset
*b
= csa
[j
];
829 if (apn
!= bpn
&& cpusets_overlap(a
, b
)) {
830 for (k
= 0; k
< csn
; k
++) {
831 struct cpuset
*c
= csa
[k
];
836 ndoms
--; /* one less element */
843 * Now we know how many domains to create.
844 * Convert <csn, csa> to <ndoms, doms> and populate cpu masks.
846 doms
= alloc_sched_domains(ndoms
);
851 * The rest of the code, including the scheduler, can deal with
852 * dattr==NULL case. No need to abort if alloc fails.
854 dattr
= kmalloc_array(ndoms
, sizeof(struct sched_domain_attr
),
857 for (nslot
= 0, i
= 0; i
< csn
; i
++) {
858 struct cpuset
*a
= csa
[i
];
863 /* Skip completed partitions */
869 if (nslot
== ndoms
) {
870 static int warnings
= 10;
872 pr_warn("rebuild_sched_domains confused: nslot %d, ndoms %d, csn %d, i %d, apn %d\n",
873 nslot
, ndoms
, csn
, i
, apn
);
881 *(dattr
+ nslot
) = SD_ATTR_INIT
;
882 for (j
= i
; j
< csn
; j
++) {
883 struct cpuset
*b
= csa
[j
];
886 cpumask_or(dp
, dp
, b
->effective_cpus
);
887 cpumask_and(dp
, dp
, housekeeping_cpumask(HK_FLAG_DOMAIN
));
889 update_domain_attr_tree(dattr
+ nslot
, b
);
891 /* Done with this partition */
897 BUG_ON(nslot
!= ndoms
);
903 * Fallback to the default domain if kmalloc() failed.
904 * See comments in partition_sched_domains().
915 * Rebuild scheduler domains.
917 * If the flag 'sched_load_balance' of any cpuset with non-empty
918 * 'cpus' changes, or if the 'cpus' allowed changes in any cpuset
919 * which has that flag enabled, or if any cpuset with a non-empty
920 * 'cpus' is removed, then call this routine to rebuild the
921 * scheduler's dynamic sched domains.
923 * Call with cpuset_mutex held. Takes get_online_cpus().
925 static void rebuild_sched_domains_locked(void)
927 struct sched_domain_attr
*attr
;
931 lockdep_assert_held(&cpuset_mutex
);
935 * We have raced with CPU hotplug. Don't do anything to avoid
936 * passing doms with offlined cpu to partition_sched_domains().
937 * Anyways, hotplug work item will rebuild sched domains.
939 if (!cpumask_equal(top_cpuset
.effective_cpus
, cpu_active_mask
))
942 /* Generate domain masks and attrs */
943 ndoms
= generate_sched_domains(&doms
, &attr
);
945 /* Have scheduler rebuild the domains */
946 partition_sched_domains(ndoms
, doms
, attr
);
950 #else /* !CONFIG_SMP */
951 static void rebuild_sched_domains_locked(void)
954 #endif /* CONFIG_SMP */
956 void rebuild_sched_domains(void)
958 mutex_lock(&cpuset_mutex
);
959 rebuild_sched_domains_locked();
960 mutex_unlock(&cpuset_mutex
);
964 * update_tasks_cpumask - Update the cpumasks of tasks in the cpuset.
965 * @cs: the cpuset in which each task's cpus_allowed mask needs to be changed
967 * Iterate through each task of @cs updating its cpus_allowed to the
968 * effective cpuset's. As this function is called with cpuset_mutex held,
969 * cpuset membership stays stable.
971 static void update_tasks_cpumask(struct cpuset
*cs
)
973 struct css_task_iter it
;
974 struct task_struct
*task
;
976 css_task_iter_start(&cs
->css
, 0, &it
);
977 while ((task
= css_task_iter_next(&it
)))
978 set_cpus_allowed_ptr(task
, cs
->effective_cpus
);
979 css_task_iter_end(&it
);
983 * compute_effective_cpumask - Compute the effective cpumask of the cpuset
984 * @new_cpus: the temp variable for the new effective_cpus mask
985 * @cs: the cpuset the need to recompute the new effective_cpus mask
986 * @parent: the parent cpuset
988 * If the parent has subpartition CPUs, include them in the list of
989 * allowable CPUs in computing the new effective_cpus mask.
991 static void compute_effective_cpumask(struct cpumask
*new_cpus
,
992 struct cpuset
*cs
, struct cpuset
*parent
)
994 if (parent
->nr_subparts_cpus
) {
995 cpumask_or(new_cpus
, parent
->effective_cpus
,
996 parent
->subparts_cpus
);
997 cpumask_and(new_cpus
, new_cpus
, cs
->cpus_allowed
);
999 cpumask_and(new_cpus
, cs
->cpus_allowed
, parent
->effective_cpus
);
1004 * Commands for update_parent_subparts_cpumask
1007 partcmd_enable
, /* Enable partition root */
1008 partcmd_disable
, /* Disable partition root */
1009 partcmd_update
, /* Update parent's subparts_cpus */
1013 * update_parent_subparts_cpumask - update subparts_cpus mask of parent cpuset
1014 * @cpuset: The cpuset that requests change in partition root state
1015 * @cmd: Partition root state change command
1016 * @newmask: Optional new cpumask for partcmd_update
1017 * @tmp: Temporary addmask and delmask
1018 * Return: 0, 1 or an error code
1020 * For partcmd_enable, the cpuset is being transformed from a non-partition
1021 * root to a partition root. The cpus_allowed mask of the given cpuset will
1022 * be put into parent's subparts_cpus and taken away from parent's
1023 * effective_cpus. The function will return 0 if all the CPUs listed in
1024 * cpus_allowed can be granted or an error code will be returned.
1026 * For partcmd_disable, the cpuset is being transofrmed from a partition
1027 * root back to a non-partition root. any CPUs in cpus_allowed that are in
1028 * parent's subparts_cpus will be taken away from that cpumask and put back
1029 * into parent's effective_cpus. 0 should always be returned.
1031 * For partcmd_update, if the optional newmask is specified, the cpu
1032 * list is to be changed from cpus_allowed to newmask. Otherwise,
1033 * cpus_allowed is assumed to remain the same. The cpuset should either
1034 * be a partition root or an invalid partition root. The partition root
1035 * state may change if newmask is NULL and none of the requested CPUs can
1036 * be granted by the parent. The function will return 1 if changes to
1037 * parent's subparts_cpus and effective_cpus happen or 0 otherwise.
1038 * Error code should only be returned when newmask is non-NULL.
1040 * The partcmd_enable and partcmd_disable commands are used by
1041 * update_prstate(). The partcmd_update command is used by
1042 * update_cpumasks_hier() with newmask NULL and update_cpumask() with
1045 * The checking is more strict when enabling partition root than the
1046 * other two commands.
1048 * Because of the implicit cpu exclusive nature of a partition root,
1049 * cpumask changes that violates the cpu exclusivity rule will not be
1050 * permitted when checked by validate_change(). The validate_change()
1051 * function will also prevent any changes to the cpu list if it is not
1052 * a superset of children's cpu lists.
1054 static int update_parent_subparts_cpumask(struct cpuset
*cpuset
, int cmd
,
1055 struct cpumask
*newmask
,
1056 struct tmpmasks
*tmp
)
1058 struct cpuset
*parent
= parent_cs(cpuset
);
1059 int adding
; /* Moving cpus from effective_cpus to subparts_cpus */
1060 int deleting
; /* Moving cpus from subparts_cpus to effective_cpus */
1061 bool part_error
= false; /* Partition error? */
1063 lockdep_assert_held(&cpuset_mutex
);
1066 * The parent must be a partition root.
1067 * The new cpumask, if present, or the current cpus_allowed must
1070 if (!is_partition_root(parent
) ||
1071 (newmask
&& cpumask_empty(newmask
)) ||
1072 (!newmask
&& cpumask_empty(cpuset
->cpus_allowed
)))
1076 * Enabling/disabling partition root is not allowed if there are
1079 if ((cmd
!= partcmd_update
) && css_has_online_children(&cpuset
->css
))
1083 * Enabling partition root is not allowed if not all the CPUs
1084 * can be granted from parent's effective_cpus or at least one
1085 * CPU will be left after that.
1087 if ((cmd
== partcmd_enable
) &&
1088 (!cpumask_subset(cpuset
->cpus_allowed
, parent
->effective_cpus
) ||
1089 cpumask_equal(cpuset
->cpus_allowed
, parent
->effective_cpus
)))
1093 * A cpumask update cannot make parent's effective_cpus become empty.
1095 adding
= deleting
= false;
1096 if (cmd
== partcmd_enable
) {
1097 cpumask_copy(tmp
->addmask
, cpuset
->cpus_allowed
);
1099 } else if (cmd
== partcmd_disable
) {
1100 deleting
= cpumask_and(tmp
->delmask
, cpuset
->cpus_allowed
,
1101 parent
->subparts_cpus
);
1102 } else if (newmask
) {
1104 * partcmd_update with newmask:
1106 * delmask = cpus_allowed & ~newmask & parent->subparts_cpus
1107 * addmask = newmask & parent->effective_cpus
1108 * & ~parent->subparts_cpus
1110 cpumask_andnot(tmp
->delmask
, cpuset
->cpus_allowed
, newmask
);
1111 deleting
= cpumask_and(tmp
->delmask
, tmp
->delmask
,
1112 parent
->subparts_cpus
);
1114 cpumask_and(tmp
->addmask
, newmask
, parent
->effective_cpus
);
1115 adding
= cpumask_andnot(tmp
->addmask
, tmp
->addmask
,
1116 parent
->subparts_cpus
);
1118 * Return error if the new effective_cpus could become empty.
1120 if (adding
&& !deleting
&&
1121 cpumask_equal(parent
->effective_cpus
, tmp
->addmask
))
1125 * partcmd_update w/o newmask:
1127 * addmask = cpus_allowed & parent->effectiveb_cpus
1129 * Note that parent's subparts_cpus may have been
1130 * pre-shrunk in case there is a change in the cpu list.
1131 * So no deletion is needed.
1133 adding
= cpumask_and(tmp
->addmask
, cpuset
->cpus_allowed
,
1134 parent
->effective_cpus
);
1135 part_error
= cpumask_equal(tmp
->addmask
,
1136 parent
->effective_cpus
);
1139 if (cmd
== partcmd_update
) {
1140 int prev_prs
= cpuset
->partition_root_state
;
1143 * Check for possible transition between PRS_ENABLED
1146 switch (cpuset
->partition_root_state
) {
1149 cpuset
->partition_root_state
= PRS_ERROR
;
1153 cpuset
->partition_root_state
= PRS_ENABLED
;
1157 * Set part_error if previously in invalid state.
1159 part_error
= (prev_prs
== PRS_ERROR
);
1162 if (!part_error
&& (cpuset
->partition_root_state
== PRS_ERROR
))
1163 return 0; /* Nothing need to be done */
1165 if (cpuset
->partition_root_state
== PRS_ERROR
) {
1167 * Remove all its cpus from parent's subparts_cpus.
1170 deleting
= cpumask_and(tmp
->delmask
, cpuset
->cpus_allowed
,
1171 parent
->subparts_cpus
);
1174 if (!adding
&& !deleting
)
1178 * Change the parent's subparts_cpus.
1179 * Newly added CPUs will be removed from effective_cpus and
1180 * newly deleted ones will be added back to effective_cpus.
1182 spin_lock_irq(&callback_lock
);
1184 cpumask_or(parent
->subparts_cpus
,
1185 parent
->subparts_cpus
, tmp
->addmask
);
1186 cpumask_andnot(parent
->effective_cpus
,
1187 parent
->effective_cpus
, tmp
->addmask
);
1190 cpumask_andnot(parent
->subparts_cpus
,
1191 parent
->subparts_cpus
, tmp
->delmask
);
1192 cpumask_or(parent
->effective_cpus
,
1193 parent
->effective_cpus
, tmp
->delmask
);
1196 parent
->nr_subparts_cpus
= cpumask_weight(parent
->subparts_cpus
);
1197 spin_unlock_irq(&callback_lock
);
1199 return cmd
== partcmd_update
;
1203 * update_cpumasks_hier - Update effective cpumasks and tasks in the subtree
1204 * @cs: the cpuset to consider
1205 * @tmp: temp variables for calculating effective_cpus & partition setup
1207 * When congifured cpumask is changed, the effective cpumasks of this cpuset
1208 * and all its descendants need to be updated.
1210 * On legacy hierachy, effective_cpus will be the same with cpu_allowed.
1212 * Called with cpuset_mutex held
1214 static void update_cpumasks_hier(struct cpuset
*cs
, struct tmpmasks
*tmp
)
1217 struct cgroup_subsys_state
*pos_css
;
1218 bool need_rebuild_sched_domains
= false;
1221 cpuset_for_each_descendant_pre(cp
, pos_css
, cs
) {
1222 struct cpuset
*parent
= parent_cs(cp
);
1224 compute_effective_cpumask(tmp
->new_cpus
, cp
, parent
);
1227 * If it becomes empty, inherit the effective mask of the
1228 * parent, which is guaranteed to have some CPUs.
1230 if (is_in_v2_mode() && cpumask_empty(tmp
->new_cpus
))
1231 cpumask_copy(tmp
->new_cpus
, parent
->effective_cpus
);
1234 * Skip the whole subtree if the cpumask remains the same
1235 * and has no partition root state.
1237 if (!cp
->partition_root_state
&&
1238 cpumask_equal(tmp
->new_cpus
, cp
->effective_cpus
)) {
1239 pos_css
= css_rightmost_descendant(pos_css
);
1244 * update_parent_subparts_cpumask() should have been called
1245 * for cs already in update_cpumask(). We should also call
1246 * update_tasks_cpumask() again for tasks in the parent
1247 * cpuset if the parent's subparts_cpus changes.
1249 if ((cp
!= cs
) && cp
->partition_root_state
) {
1250 switch (parent
->partition_root_state
) {
1253 * If parent is not a partition root or an
1254 * invalid partition root, clear the state
1255 * state and the CS_CPU_EXCLUSIVE flag.
1257 WARN_ON_ONCE(cp
->partition_root_state
1259 cp
->partition_root_state
= 0;
1262 * clear_bit() is an atomic operation and
1263 * readers aren't interested in the state
1264 * of CS_CPU_EXCLUSIVE anyway. So we can
1265 * just update the flag without holding
1266 * the callback_lock.
1268 clear_bit(CS_CPU_EXCLUSIVE
, &cp
->flags
);
1272 if (update_parent_subparts_cpumask(cp
, partcmd_update
, NULL
, tmp
))
1273 update_tasks_cpumask(parent
);
1278 * When parent is invalid, it has to be too.
1280 cp
->partition_root_state
= PRS_ERROR
;
1281 if (cp
->nr_subparts_cpus
) {
1282 cp
->nr_subparts_cpus
= 0;
1283 cpumask_clear(cp
->subparts_cpus
);
1289 if (!css_tryget_online(&cp
->css
))
1293 spin_lock_irq(&callback_lock
);
1295 cpumask_copy(cp
->effective_cpus
, tmp
->new_cpus
);
1296 if (cp
->nr_subparts_cpus
&&
1297 (cp
->partition_root_state
!= PRS_ENABLED
)) {
1298 cp
->nr_subparts_cpus
= 0;
1299 cpumask_clear(cp
->subparts_cpus
);
1300 } else if (cp
->nr_subparts_cpus
) {
1302 * Make sure that effective_cpus & subparts_cpus
1303 * are mutually exclusive.
1305 * In the unlikely event that effective_cpus
1306 * becomes empty. we clear cp->nr_subparts_cpus and
1307 * let its child partition roots to compete for
1310 cpumask_andnot(cp
->effective_cpus
, cp
->effective_cpus
,
1312 if (cpumask_empty(cp
->effective_cpus
)) {
1313 cpumask_copy(cp
->effective_cpus
, tmp
->new_cpus
);
1314 cpumask_clear(cp
->subparts_cpus
);
1315 cp
->nr_subparts_cpus
= 0;
1316 } else if (!cpumask_subset(cp
->subparts_cpus
,
1318 cpumask_andnot(cp
->subparts_cpus
,
1319 cp
->subparts_cpus
, tmp
->new_cpus
);
1320 cp
->nr_subparts_cpus
1321 = cpumask_weight(cp
->subparts_cpus
);
1324 spin_unlock_irq(&callback_lock
);
1326 WARN_ON(!is_in_v2_mode() &&
1327 !cpumask_equal(cp
->cpus_allowed
, cp
->effective_cpus
));
1329 update_tasks_cpumask(cp
);
1332 * If the effective cpumask of any non-empty cpuset is changed,
1333 * we need to rebuild sched domains.
1335 if (!cpumask_empty(cp
->cpus_allowed
) &&
1336 is_sched_load_balance(cp
))
1337 need_rebuild_sched_domains
= true;
1344 if (need_rebuild_sched_domains
)
1345 rebuild_sched_domains_locked();
1349 * update_cpumask - update the cpus_allowed mask of a cpuset and all tasks in it
1350 * @cs: the cpuset to consider
1351 * @trialcs: trial cpuset
1352 * @buf: buffer of cpu numbers written to this cpuset
1354 static int update_cpumask(struct cpuset
*cs
, struct cpuset
*trialcs
,
1358 struct tmpmasks tmp
;
1360 /* top_cpuset.cpus_allowed tracks cpu_online_mask; it's read-only */
1361 if (cs
== &top_cpuset
)
1365 * An empty cpus_allowed is ok only if the cpuset has no tasks.
1366 * Since cpulist_parse() fails on an empty mask, we special case
1367 * that parsing. The validate_change() call ensures that cpusets
1368 * with tasks have cpus.
1371 cpumask_clear(trialcs
->cpus_allowed
);
1373 retval
= cpulist_parse(buf
, trialcs
->cpus_allowed
);
1377 if (!cpumask_subset(trialcs
->cpus_allowed
,
1378 top_cpuset
.cpus_allowed
))
1382 /* Nothing to do if the cpus didn't change */
1383 if (cpumask_equal(cs
->cpus_allowed
, trialcs
->cpus_allowed
))
1386 retval
= validate_change(cs
, trialcs
);
1390 #ifdef CONFIG_CPUMASK_OFFSTACK
1392 * Use the cpumasks in trialcs for tmpmasks when they are pointers
1393 * to allocated cpumasks.
1395 tmp
.addmask
= trialcs
->subparts_cpus
;
1396 tmp
.delmask
= trialcs
->effective_cpus
;
1397 tmp
.new_cpus
= trialcs
->cpus_allowed
;
1400 if (cs
->partition_root_state
) {
1401 /* Cpumask of a partition root cannot be empty */
1402 if (cpumask_empty(trialcs
->cpus_allowed
))
1404 if (update_parent_subparts_cpumask(cs
, partcmd_update
,
1405 trialcs
->cpus_allowed
, &tmp
) < 0)
1409 spin_lock_irq(&callback_lock
);
1410 cpumask_copy(cs
->cpus_allowed
, trialcs
->cpus_allowed
);
1413 * Make sure that subparts_cpus is a subset of cpus_allowed.
1415 if (cs
->nr_subparts_cpus
) {
1416 cpumask_andnot(cs
->subparts_cpus
, cs
->subparts_cpus
,
1418 cs
->nr_subparts_cpus
= cpumask_weight(cs
->subparts_cpus
);
1420 spin_unlock_irq(&callback_lock
);
1422 update_cpumasks_hier(cs
, &tmp
);
1427 * Migrate memory region from one set of nodes to another. This is
1428 * performed asynchronously as it can be called from process migration path
1429 * holding locks involved in process management. All mm migrations are
1430 * performed in the queued order and can be waited for by flushing
1431 * cpuset_migrate_mm_wq.
1434 struct cpuset_migrate_mm_work
{
1435 struct work_struct work
;
1436 struct mm_struct
*mm
;
1441 static void cpuset_migrate_mm_workfn(struct work_struct
*work
)
1443 struct cpuset_migrate_mm_work
*mwork
=
1444 container_of(work
, struct cpuset_migrate_mm_work
, work
);
1446 /* on a wq worker, no need to worry about %current's mems_allowed */
1447 do_migrate_pages(mwork
->mm
, &mwork
->from
, &mwork
->to
, MPOL_MF_MOVE_ALL
);
1452 static void cpuset_migrate_mm(struct mm_struct
*mm
, const nodemask_t
*from
,
1453 const nodemask_t
*to
)
1455 struct cpuset_migrate_mm_work
*mwork
;
1457 mwork
= kzalloc(sizeof(*mwork
), GFP_KERNEL
);
1460 mwork
->from
= *from
;
1462 INIT_WORK(&mwork
->work
, cpuset_migrate_mm_workfn
);
1463 queue_work(cpuset_migrate_mm_wq
, &mwork
->work
);
1469 static void cpuset_post_attach(void)
1471 flush_workqueue(cpuset_migrate_mm_wq
);
1475 * cpuset_change_task_nodemask - change task's mems_allowed and mempolicy
1476 * @tsk: the task to change
1477 * @newmems: new nodes that the task will be set
1479 * We use the mems_allowed_seq seqlock to safely update both tsk->mems_allowed
1480 * and rebind an eventual tasks' mempolicy. If the task is allocating in
1481 * parallel, it might temporarily see an empty intersection, which results in
1482 * a seqlock check and retry before OOM or allocation failure.
1484 static void cpuset_change_task_nodemask(struct task_struct
*tsk
,
1485 nodemask_t
*newmems
)
1489 local_irq_disable();
1490 write_seqcount_begin(&tsk
->mems_allowed_seq
);
1492 nodes_or(tsk
->mems_allowed
, tsk
->mems_allowed
, *newmems
);
1493 mpol_rebind_task(tsk
, newmems
);
1494 tsk
->mems_allowed
= *newmems
;
1496 write_seqcount_end(&tsk
->mems_allowed_seq
);
1502 static void *cpuset_being_rebound
;
1505 * update_tasks_nodemask - Update the nodemasks of tasks in the cpuset.
1506 * @cs: the cpuset in which each task's mems_allowed mask needs to be changed
1508 * Iterate through each task of @cs updating its mems_allowed to the
1509 * effective cpuset's. As this function is called with cpuset_mutex held,
1510 * cpuset membership stays stable.
1512 static void update_tasks_nodemask(struct cpuset
*cs
)
1514 static nodemask_t newmems
; /* protected by cpuset_mutex */
1515 struct css_task_iter it
;
1516 struct task_struct
*task
;
1518 cpuset_being_rebound
= cs
; /* causes mpol_dup() rebind */
1520 guarantee_online_mems(cs
, &newmems
);
1523 * The mpol_rebind_mm() call takes mmap_sem, which we couldn't
1524 * take while holding tasklist_lock. Forks can happen - the
1525 * mpol_dup() cpuset_being_rebound check will catch such forks,
1526 * and rebind their vma mempolicies too. Because we still hold
1527 * the global cpuset_mutex, we know that no other rebind effort
1528 * will be contending for the global variable cpuset_being_rebound.
1529 * It's ok if we rebind the same mm twice; mpol_rebind_mm()
1530 * is idempotent. Also migrate pages in each mm to new nodes.
1532 css_task_iter_start(&cs
->css
, 0, &it
);
1533 while ((task
= css_task_iter_next(&it
))) {
1534 struct mm_struct
*mm
;
1537 cpuset_change_task_nodemask(task
, &newmems
);
1539 mm
= get_task_mm(task
);
1543 migrate
= is_memory_migrate(cs
);
1545 mpol_rebind_mm(mm
, &cs
->mems_allowed
);
1547 cpuset_migrate_mm(mm
, &cs
->old_mems_allowed
, &newmems
);
1551 css_task_iter_end(&it
);
1554 * All the tasks' nodemasks have been updated, update
1555 * cs->old_mems_allowed.
1557 cs
->old_mems_allowed
= newmems
;
1559 /* We're done rebinding vmas to this cpuset's new mems_allowed. */
1560 cpuset_being_rebound
= NULL
;
1564 * update_nodemasks_hier - Update effective nodemasks and tasks in the subtree
1565 * @cs: the cpuset to consider
1566 * @new_mems: a temp variable for calculating new effective_mems
1568 * When configured nodemask is changed, the effective nodemasks of this cpuset
1569 * and all its descendants need to be updated.
1571 * On legacy hiearchy, effective_mems will be the same with mems_allowed.
1573 * Called with cpuset_mutex held
1575 static void update_nodemasks_hier(struct cpuset
*cs
, nodemask_t
*new_mems
)
1578 struct cgroup_subsys_state
*pos_css
;
1581 cpuset_for_each_descendant_pre(cp
, pos_css
, cs
) {
1582 struct cpuset
*parent
= parent_cs(cp
);
1584 nodes_and(*new_mems
, cp
->mems_allowed
, parent
->effective_mems
);
1587 * If it becomes empty, inherit the effective mask of the
1588 * parent, which is guaranteed to have some MEMs.
1590 if (is_in_v2_mode() && nodes_empty(*new_mems
))
1591 *new_mems
= parent
->effective_mems
;
1593 /* Skip the whole subtree if the nodemask remains the same. */
1594 if (nodes_equal(*new_mems
, cp
->effective_mems
)) {
1595 pos_css
= css_rightmost_descendant(pos_css
);
1599 if (!css_tryget_online(&cp
->css
))
1603 spin_lock_irq(&callback_lock
);
1604 cp
->effective_mems
= *new_mems
;
1605 spin_unlock_irq(&callback_lock
);
1607 WARN_ON(!is_in_v2_mode() &&
1608 !nodes_equal(cp
->mems_allowed
, cp
->effective_mems
));
1610 update_tasks_nodemask(cp
);
1619 * Handle user request to change the 'mems' memory placement
1620 * of a cpuset. Needs to validate the request, update the
1621 * cpusets mems_allowed, and for each task in the cpuset,
1622 * update mems_allowed and rebind task's mempolicy and any vma
1623 * mempolicies and if the cpuset is marked 'memory_migrate',
1624 * migrate the tasks pages to the new memory.
1626 * Call with cpuset_mutex held. May take callback_lock during call.
1627 * Will take tasklist_lock, scan tasklist for tasks in cpuset cs,
1628 * lock each such tasks mm->mmap_sem, scan its vma's and rebind
1629 * their mempolicies to the cpusets new mems_allowed.
1631 static int update_nodemask(struct cpuset
*cs
, struct cpuset
*trialcs
,
1637 * top_cpuset.mems_allowed tracks node_stats[N_MEMORY];
1640 if (cs
== &top_cpuset
) {
1646 * An empty mems_allowed is ok iff there are no tasks in the cpuset.
1647 * Since nodelist_parse() fails on an empty mask, we special case
1648 * that parsing. The validate_change() call ensures that cpusets
1649 * with tasks have memory.
1652 nodes_clear(trialcs
->mems_allowed
);
1654 retval
= nodelist_parse(buf
, trialcs
->mems_allowed
);
1658 if (!nodes_subset(trialcs
->mems_allowed
,
1659 top_cpuset
.mems_allowed
)) {
1665 if (nodes_equal(cs
->mems_allowed
, trialcs
->mems_allowed
)) {
1666 retval
= 0; /* Too easy - nothing to do */
1669 retval
= validate_change(cs
, trialcs
);
1673 spin_lock_irq(&callback_lock
);
1674 cs
->mems_allowed
= trialcs
->mems_allowed
;
1675 spin_unlock_irq(&callback_lock
);
1677 /* use trialcs->mems_allowed as a temp variable */
1678 update_nodemasks_hier(cs
, &trialcs
->mems_allowed
);
1683 bool current_cpuset_is_being_rebound(void)
1688 ret
= task_cs(current
) == cpuset_being_rebound
;
1694 static int update_relax_domain_level(struct cpuset
*cs
, s64 val
)
1697 if (val
< -1 || val
>= sched_domain_level_max
)
1701 if (val
!= cs
->relax_domain_level
) {
1702 cs
->relax_domain_level
= val
;
1703 if (!cpumask_empty(cs
->cpus_allowed
) &&
1704 is_sched_load_balance(cs
))
1705 rebuild_sched_domains_locked();
1712 * update_tasks_flags - update the spread flags of tasks in the cpuset.
1713 * @cs: the cpuset in which each task's spread flags needs to be changed
1715 * Iterate through each task of @cs updating its spread flags. As this
1716 * function is called with cpuset_mutex held, cpuset membership stays
1719 static void update_tasks_flags(struct cpuset
*cs
)
1721 struct css_task_iter it
;
1722 struct task_struct
*task
;
1724 css_task_iter_start(&cs
->css
, 0, &it
);
1725 while ((task
= css_task_iter_next(&it
)))
1726 cpuset_update_task_spread_flag(cs
, task
);
1727 css_task_iter_end(&it
);
1731 * update_flag - read a 0 or a 1 in a file and update associated flag
1732 * bit: the bit to update (see cpuset_flagbits_t)
1733 * cs: the cpuset to update
1734 * turning_on: whether the flag is being set or cleared
1736 * Call with cpuset_mutex held.
1739 static int update_flag(cpuset_flagbits_t bit
, struct cpuset
*cs
,
1742 struct cpuset
*trialcs
;
1743 int balance_flag_changed
;
1744 int spread_flag_changed
;
1747 trialcs
= alloc_trial_cpuset(cs
);
1752 set_bit(bit
, &trialcs
->flags
);
1754 clear_bit(bit
, &trialcs
->flags
);
1756 err
= validate_change(cs
, trialcs
);
1760 balance_flag_changed
= (is_sched_load_balance(cs
) !=
1761 is_sched_load_balance(trialcs
));
1763 spread_flag_changed
= ((is_spread_slab(cs
) != is_spread_slab(trialcs
))
1764 || (is_spread_page(cs
) != is_spread_page(trialcs
)));
1766 spin_lock_irq(&callback_lock
);
1767 cs
->flags
= trialcs
->flags
;
1768 spin_unlock_irq(&callback_lock
);
1770 if (!cpumask_empty(trialcs
->cpus_allowed
) && balance_flag_changed
)
1771 rebuild_sched_domains_locked();
1773 if (spread_flag_changed
)
1774 update_tasks_flags(cs
);
1776 free_cpuset(trialcs
);
1781 * update_prstate - update partititon_root_state
1782 * cs: the cpuset to update
1783 * val: 0 - disabled, 1 - enabled
1785 * Call with cpuset_mutex held.
1787 static int update_prstate(struct cpuset
*cs
, int val
)
1790 struct cpuset
*parent
= parent_cs(cs
);
1791 struct tmpmasks tmp
;
1793 if ((val
!= 0) && (val
!= 1))
1795 if (val
== cs
->partition_root_state
)
1799 * Cannot force a partial or invalid partition root to a full
1802 if (val
&& cs
->partition_root_state
)
1805 if (alloc_cpumasks(NULL
, &tmp
))
1809 if (!cs
->partition_root_state
) {
1811 * Turning on partition root requires setting the
1812 * CS_CPU_EXCLUSIVE bit implicitly as well and cpus_allowed
1815 if (cpumask_empty(cs
->cpus_allowed
))
1818 err
= update_flag(CS_CPU_EXCLUSIVE
, cs
, 1);
1822 err
= update_parent_subparts_cpumask(cs
, partcmd_enable
,
1825 update_flag(CS_CPU_EXCLUSIVE
, cs
, 0);
1828 cs
->partition_root_state
= PRS_ENABLED
;
1831 * Turning off partition root will clear the
1832 * CS_CPU_EXCLUSIVE bit.
1834 if (cs
->partition_root_state
== PRS_ERROR
) {
1835 cs
->partition_root_state
= 0;
1836 update_flag(CS_CPU_EXCLUSIVE
, cs
, 0);
1841 err
= update_parent_subparts_cpumask(cs
, partcmd_disable
,
1846 cs
->partition_root_state
= 0;
1848 /* Turning off CS_CPU_EXCLUSIVE will not return error */
1849 update_flag(CS_CPU_EXCLUSIVE
, cs
, 0);
1853 * Update cpumask of parent's tasks except when it is the top
1854 * cpuset as some system daemons cannot be mapped to other CPUs.
1856 if (parent
!= &top_cpuset
)
1857 update_tasks_cpumask(parent
);
1859 rebuild_sched_domains_locked();
1861 free_cpumasks(NULL
, &tmp
);
1866 * Frequency meter - How fast is some event occurring?
1868 * These routines manage a digitally filtered, constant time based,
1869 * event frequency meter. There are four routines:
1870 * fmeter_init() - initialize a frequency meter.
1871 * fmeter_markevent() - called each time the event happens.
1872 * fmeter_getrate() - returns the recent rate of such events.
1873 * fmeter_update() - internal routine used to update fmeter.
1875 * A common data structure is passed to each of these routines,
1876 * which is used to keep track of the state required to manage the
1877 * frequency meter and its digital filter.
1879 * The filter works on the number of events marked per unit time.
1880 * The filter is single-pole low-pass recursive (IIR). The time unit
1881 * is 1 second. Arithmetic is done using 32-bit integers scaled to
1882 * simulate 3 decimal digits of precision (multiplied by 1000).
1884 * With an FM_COEF of 933, and a time base of 1 second, the filter
1885 * has a half-life of 10 seconds, meaning that if the events quit
1886 * happening, then the rate returned from the fmeter_getrate()
1887 * will be cut in half each 10 seconds, until it converges to zero.
1889 * It is not worth doing a real infinitely recursive filter. If more
1890 * than FM_MAXTICKS ticks have elapsed since the last filter event,
1891 * just compute FM_MAXTICKS ticks worth, by which point the level
1894 * Limit the count of unprocessed events to FM_MAXCNT, so as to avoid
1895 * arithmetic overflow in the fmeter_update() routine.
1897 * Given the simple 32 bit integer arithmetic used, this meter works
1898 * best for reporting rates between one per millisecond (msec) and
1899 * one per 32 (approx) seconds. At constant rates faster than one
1900 * per msec it maxes out at values just under 1,000,000. At constant
1901 * rates between one per msec, and one per second it will stabilize
1902 * to a value N*1000, where N is the rate of events per second.
1903 * At constant rates between one per second and one per 32 seconds,
1904 * it will be choppy, moving up on the seconds that have an event,
1905 * and then decaying until the next event. At rates slower than
1906 * about one in 32 seconds, it decays all the way back to zero between
1910 #define FM_COEF 933 /* coefficient for half-life of 10 secs */
1911 #define FM_MAXTICKS ((u32)99) /* useless computing more ticks than this */
1912 #define FM_MAXCNT 1000000 /* limit cnt to avoid overflow */
1913 #define FM_SCALE 1000 /* faux fixed point scale */
1915 /* Initialize a frequency meter */
1916 static void fmeter_init(struct fmeter
*fmp
)
1921 spin_lock_init(&fmp
->lock
);
1924 /* Internal meter update - process cnt events and update value */
1925 static void fmeter_update(struct fmeter
*fmp
)
1930 now
= ktime_get_seconds();
1931 ticks
= now
- fmp
->time
;
1936 ticks
= min(FM_MAXTICKS
, ticks
);
1938 fmp
->val
= (FM_COEF
* fmp
->val
) / FM_SCALE
;
1941 fmp
->val
+= ((FM_SCALE
- FM_COEF
) * fmp
->cnt
) / FM_SCALE
;
1945 /* Process any previous ticks, then bump cnt by one (times scale). */
1946 static void fmeter_markevent(struct fmeter
*fmp
)
1948 spin_lock(&fmp
->lock
);
1950 fmp
->cnt
= min(FM_MAXCNT
, fmp
->cnt
+ FM_SCALE
);
1951 spin_unlock(&fmp
->lock
);
1954 /* Process any previous ticks, then return current value. */
1955 static int fmeter_getrate(struct fmeter
*fmp
)
1959 spin_lock(&fmp
->lock
);
1962 spin_unlock(&fmp
->lock
);
1966 static struct cpuset
*cpuset_attach_old_cs
;
1968 /* Called by cgroups to determine if a cpuset is usable; cpuset_mutex held */
1969 static int cpuset_can_attach(struct cgroup_taskset
*tset
)
1971 struct cgroup_subsys_state
*css
;
1973 struct task_struct
*task
;
1976 /* used later by cpuset_attach() */
1977 cpuset_attach_old_cs
= task_cs(cgroup_taskset_first(tset
, &css
));
1980 mutex_lock(&cpuset_mutex
);
1982 /* allow moving tasks into an empty cpuset if on default hierarchy */
1984 if (!is_in_v2_mode() &&
1985 (cpumask_empty(cs
->cpus_allowed
) || nodes_empty(cs
->mems_allowed
)))
1988 cgroup_taskset_for_each(task
, css
, tset
) {
1989 ret
= task_can_attach(task
, cs
->cpus_allowed
);
1992 ret
= security_task_setscheduler(task
);
1998 * Mark attach is in progress. This makes validate_change() fail
1999 * changes which zero cpus/mems_allowed.
2001 cs
->attach_in_progress
++;
2004 mutex_unlock(&cpuset_mutex
);
2008 static void cpuset_cancel_attach(struct cgroup_taskset
*tset
)
2010 struct cgroup_subsys_state
*css
;
2013 cgroup_taskset_first(tset
, &css
);
2016 mutex_lock(&cpuset_mutex
);
2017 css_cs(css
)->attach_in_progress
--;
2018 mutex_unlock(&cpuset_mutex
);
2022 * Protected by cpuset_mutex. cpus_attach is used only by cpuset_attach()
2023 * but we can't allocate it dynamically there. Define it global and
2024 * allocate from cpuset_init().
2026 static cpumask_var_t cpus_attach
;
2028 static void cpuset_attach(struct cgroup_taskset
*tset
)
2030 /* static buf protected by cpuset_mutex */
2031 static nodemask_t cpuset_attach_nodemask_to
;
2032 struct task_struct
*task
;
2033 struct task_struct
*leader
;
2034 struct cgroup_subsys_state
*css
;
2036 struct cpuset
*oldcs
= cpuset_attach_old_cs
;
2038 cgroup_taskset_first(tset
, &css
);
2041 mutex_lock(&cpuset_mutex
);
2043 /* prepare for attach */
2044 if (cs
== &top_cpuset
)
2045 cpumask_copy(cpus_attach
, cpu_possible_mask
);
2047 guarantee_online_cpus(cs
, cpus_attach
);
2049 guarantee_online_mems(cs
, &cpuset_attach_nodemask_to
);
2051 cgroup_taskset_for_each(task
, css
, tset
) {
2053 * can_attach beforehand should guarantee that this doesn't
2054 * fail. TODO: have a better way to handle failure here
2056 WARN_ON_ONCE(set_cpus_allowed_ptr(task
, cpus_attach
));
2058 cpuset_change_task_nodemask(task
, &cpuset_attach_nodemask_to
);
2059 cpuset_update_task_spread_flag(cs
, task
);
2063 * Change mm for all threadgroup leaders. This is expensive and may
2064 * sleep and should be moved outside migration path proper.
2066 cpuset_attach_nodemask_to
= cs
->effective_mems
;
2067 cgroup_taskset_for_each_leader(leader
, css
, tset
) {
2068 struct mm_struct
*mm
= get_task_mm(leader
);
2071 mpol_rebind_mm(mm
, &cpuset_attach_nodemask_to
);
2074 * old_mems_allowed is the same with mems_allowed
2075 * here, except if this task is being moved
2076 * automatically due to hotplug. In that case
2077 * @mems_allowed has been updated and is empty, so
2078 * @old_mems_allowed is the right nodesets that we
2081 if (is_memory_migrate(cs
))
2082 cpuset_migrate_mm(mm
, &oldcs
->old_mems_allowed
,
2083 &cpuset_attach_nodemask_to
);
2089 cs
->old_mems_allowed
= cpuset_attach_nodemask_to
;
2091 cs
->attach_in_progress
--;
2092 if (!cs
->attach_in_progress
)
2093 wake_up(&cpuset_attach_wq
);
2095 mutex_unlock(&cpuset_mutex
);
2098 /* The various types of files and directories in a cpuset file system */
2101 FILE_MEMORY_MIGRATE
,
2104 FILE_EFFECTIVE_CPULIST
,
2105 FILE_EFFECTIVE_MEMLIST
,
2109 FILE_SCHED_LOAD_BALANCE
,
2110 FILE_PARTITION_ROOT
,
2111 FILE_SCHED_RELAX_DOMAIN_LEVEL
,
2112 FILE_MEMORY_PRESSURE_ENABLED
,
2113 FILE_MEMORY_PRESSURE
,
2116 } cpuset_filetype_t
;
2118 static int cpuset_write_u64(struct cgroup_subsys_state
*css
, struct cftype
*cft
,
2121 struct cpuset
*cs
= css_cs(css
);
2122 cpuset_filetype_t type
= cft
->private;
2125 mutex_lock(&cpuset_mutex
);
2126 if (!is_cpuset_online(cs
)) {
2132 case FILE_CPU_EXCLUSIVE
:
2133 retval
= update_flag(CS_CPU_EXCLUSIVE
, cs
, val
);
2135 case FILE_MEM_EXCLUSIVE
:
2136 retval
= update_flag(CS_MEM_EXCLUSIVE
, cs
, val
);
2138 case FILE_MEM_HARDWALL
:
2139 retval
= update_flag(CS_MEM_HARDWALL
, cs
, val
);
2141 case FILE_SCHED_LOAD_BALANCE
:
2142 retval
= update_flag(CS_SCHED_LOAD_BALANCE
, cs
, val
);
2144 case FILE_MEMORY_MIGRATE
:
2145 retval
= update_flag(CS_MEMORY_MIGRATE
, cs
, val
);
2147 case FILE_MEMORY_PRESSURE_ENABLED
:
2148 cpuset_memory_pressure_enabled
= !!val
;
2150 case FILE_SPREAD_PAGE
:
2151 retval
= update_flag(CS_SPREAD_PAGE
, cs
, val
);
2153 case FILE_SPREAD_SLAB
:
2154 retval
= update_flag(CS_SPREAD_SLAB
, cs
, val
);
2161 mutex_unlock(&cpuset_mutex
);
2165 static int cpuset_write_s64(struct cgroup_subsys_state
*css
, struct cftype
*cft
,
2168 struct cpuset
*cs
= css_cs(css
);
2169 cpuset_filetype_t type
= cft
->private;
2170 int retval
= -ENODEV
;
2172 mutex_lock(&cpuset_mutex
);
2173 if (!is_cpuset_online(cs
))
2177 case FILE_SCHED_RELAX_DOMAIN_LEVEL
:
2178 retval
= update_relax_domain_level(cs
, val
);
2180 case FILE_PARTITION_ROOT
:
2181 retval
= update_prstate(cs
, val
);
2188 mutex_unlock(&cpuset_mutex
);
2193 * Common handling for a write to a "cpus" or "mems" file.
2195 static ssize_t
cpuset_write_resmask(struct kernfs_open_file
*of
,
2196 char *buf
, size_t nbytes
, loff_t off
)
2198 struct cpuset
*cs
= css_cs(of_css(of
));
2199 struct cpuset
*trialcs
;
2200 int retval
= -ENODEV
;
2202 buf
= strstrip(buf
);
2205 * CPU or memory hotunplug may leave @cs w/o any execution
2206 * resources, in which case the hotplug code asynchronously updates
2207 * configuration and transfers all tasks to the nearest ancestor
2208 * which can execute.
2210 * As writes to "cpus" or "mems" may restore @cs's execution
2211 * resources, wait for the previously scheduled operations before
2212 * proceeding, so that we don't end up keep removing tasks added
2213 * after execution capability is restored.
2215 * cpuset_hotplug_work calls back into cgroup core via
2216 * cgroup_transfer_tasks() and waiting for it from a cgroupfs
2217 * operation like this one can lead to a deadlock through kernfs
2218 * active_ref protection. Let's break the protection. Losing the
2219 * protection is okay as we check whether @cs is online after
2220 * grabbing cpuset_mutex anyway. This only happens on the legacy
2224 kernfs_break_active_protection(of
->kn
);
2225 flush_work(&cpuset_hotplug_work
);
2227 mutex_lock(&cpuset_mutex
);
2228 if (!is_cpuset_online(cs
))
2231 trialcs
= alloc_trial_cpuset(cs
);
2237 switch (of_cft(of
)->private) {
2239 retval
= update_cpumask(cs
, trialcs
, buf
);
2242 retval
= update_nodemask(cs
, trialcs
, buf
);
2249 free_cpuset(trialcs
);
2251 mutex_unlock(&cpuset_mutex
);
2252 kernfs_unbreak_active_protection(of
->kn
);
2254 flush_workqueue(cpuset_migrate_mm_wq
);
2255 return retval
?: nbytes
;
2259 * These ascii lists should be read in a single call, by using a user
2260 * buffer large enough to hold the entire map. If read in smaller
2261 * chunks, there is no guarantee of atomicity. Since the display format
2262 * used, list of ranges of sequential numbers, is variable length,
2263 * and since these maps can change value dynamically, one could read
2264 * gibberish by doing partial reads while a list was changing.
2266 static int cpuset_common_seq_show(struct seq_file
*sf
, void *v
)
2268 struct cpuset
*cs
= css_cs(seq_css(sf
));
2269 cpuset_filetype_t type
= seq_cft(sf
)->private;
2272 spin_lock_irq(&callback_lock
);
2276 seq_printf(sf
, "%*pbl\n", cpumask_pr_args(cs
->cpus_allowed
));
2279 seq_printf(sf
, "%*pbl\n", nodemask_pr_args(&cs
->mems_allowed
));
2281 case FILE_EFFECTIVE_CPULIST
:
2282 seq_printf(sf
, "%*pbl\n", cpumask_pr_args(cs
->effective_cpus
));
2284 case FILE_EFFECTIVE_MEMLIST
:
2285 seq_printf(sf
, "%*pbl\n", nodemask_pr_args(&cs
->effective_mems
));
2291 spin_unlock_irq(&callback_lock
);
2295 static u64
cpuset_read_u64(struct cgroup_subsys_state
*css
, struct cftype
*cft
)
2297 struct cpuset
*cs
= css_cs(css
);
2298 cpuset_filetype_t type
= cft
->private;
2300 case FILE_CPU_EXCLUSIVE
:
2301 return is_cpu_exclusive(cs
);
2302 case FILE_MEM_EXCLUSIVE
:
2303 return is_mem_exclusive(cs
);
2304 case FILE_MEM_HARDWALL
:
2305 return is_mem_hardwall(cs
);
2306 case FILE_SCHED_LOAD_BALANCE
:
2307 return is_sched_load_balance(cs
);
2308 case FILE_MEMORY_MIGRATE
:
2309 return is_memory_migrate(cs
);
2310 case FILE_MEMORY_PRESSURE_ENABLED
:
2311 return cpuset_memory_pressure_enabled
;
2312 case FILE_MEMORY_PRESSURE
:
2313 return fmeter_getrate(&cs
->fmeter
);
2314 case FILE_SPREAD_PAGE
:
2315 return is_spread_page(cs
);
2316 case FILE_SPREAD_SLAB
:
2317 return is_spread_slab(cs
);
2322 /* Unreachable but makes gcc happy */
2326 static s64
cpuset_read_s64(struct cgroup_subsys_state
*css
, struct cftype
*cft
)
2328 struct cpuset
*cs
= css_cs(css
);
2329 cpuset_filetype_t type
= cft
->private;
2331 case FILE_SCHED_RELAX_DOMAIN_LEVEL
:
2332 return cs
->relax_domain_level
;
2333 case FILE_PARTITION_ROOT
:
2334 return cs
->partition_root_state
;
2339 /* Unrechable but makes gcc happy */
2344 * for the common functions, 'private' gives the type of file
2347 static struct cftype legacy_files
[] = {
2350 .seq_show
= cpuset_common_seq_show
,
2351 .write
= cpuset_write_resmask
,
2352 .max_write_len
= (100U + 6 * NR_CPUS
),
2353 .private = FILE_CPULIST
,
2358 .seq_show
= cpuset_common_seq_show
,
2359 .write
= cpuset_write_resmask
,
2360 .max_write_len
= (100U + 6 * MAX_NUMNODES
),
2361 .private = FILE_MEMLIST
,
2365 .name
= "effective_cpus",
2366 .seq_show
= cpuset_common_seq_show
,
2367 .private = FILE_EFFECTIVE_CPULIST
,
2371 .name
= "effective_mems",
2372 .seq_show
= cpuset_common_seq_show
,
2373 .private = FILE_EFFECTIVE_MEMLIST
,
2377 .name
= "cpu_exclusive",
2378 .read_u64
= cpuset_read_u64
,
2379 .write_u64
= cpuset_write_u64
,
2380 .private = FILE_CPU_EXCLUSIVE
,
2384 .name
= "mem_exclusive",
2385 .read_u64
= cpuset_read_u64
,
2386 .write_u64
= cpuset_write_u64
,
2387 .private = FILE_MEM_EXCLUSIVE
,
2391 .name
= "mem_hardwall",
2392 .read_u64
= cpuset_read_u64
,
2393 .write_u64
= cpuset_write_u64
,
2394 .private = FILE_MEM_HARDWALL
,
2398 .name
= "sched_load_balance",
2399 .read_u64
= cpuset_read_u64
,
2400 .write_u64
= cpuset_write_u64
,
2401 .private = FILE_SCHED_LOAD_BALANCE
,
2405 .name
= "sched_relax_domain_level",
2406 .read_s64
= cpuset_read_s64
,
2407 .write_s64
= cpuset_write_s64
,
2408 .private = FILE_SCHED_RELAX_DOMAIN_LEVEL
,
2412 .name
= "memory_migrate",
2413 .read_u64
= cpuset_read_u64
,
2414 .write_u64
= cpuset_write_u64
,
2415 .private = FILE_MEMORY_MIGRATE
,
2419 .name
= "memory_pressure",
2420 .read_u64
= cpuset_read_u64
,
2421 .private = FILE_MEMORY_PRESSURE
,
2425 .name
= "memory_spread_page",
2426 .read_u64
= cpuset_read_u64
,
2427 .write_u64
= cpuset_write_u64
,
2428 .private = FILE_SPREAD_PAGE
,
2432 .name
= "memory_spread_slab",
2433 .read_u64
= cpuset_read_u64
,
2434 .write_u64
= cpuset_write_u64
,
2435 .private = FILE_SPREAD_SLAB
,
2439 .name
= "memory_pressure_enabled",
2440 .flags
= CFTYPE_ONLY_ON_ROOT
,
2441 .read_u64
= cpuset_read_u64
,
2442 .write_u64
= cpuset_write_u64
,
2443 .private = FILE_MEMORY_PRESSURE_ENABLED
,
2450 * This is currently a minimal set for the default hierarchy. It can be
2451 * expanded later on by migrating more features and control files from v1.
2453 static struct cftype dfl_files
[] = {
2456 .seq_show
= cpuset_common_seq_show
,
2457 .write
= cpuset_write_resmask
,
2458 .max_write_len
= (100U + 6 * NR_CPUS
),
2459 .private = FILE_CPULIST
,
2460 .flags
= CFTYPE_NOT_ON_ROOT
,
2465 .seq_show
= cpuset_common_seq_show
,
2466 .write
= cpuset_write_resmask
,
2467 .max_write_len
= (100U + 6 * MAX_NUMNODES
),
2468 .private = FILE_MEMLIST
,
2469 .flags
= CFTYPE_NOT_ON_ROOT
,
2473 .name
= "cpus.effective",
2474 .seq_show
= cpuset_common_seq_show
,
2475 .private = FILE_EFFECTIVE_CPULIST
,
2476 .flags
= CFTYPE_NOT_ON_ROOT
,
2480 .name
= "mems.effective",
2481 .seq_show
= cpuset_common_seq_show
,
2482 .private = FILE_EFFECTIVE_MEMLIST
,
2483 .flags
= CFTYPE_NOT_ON_ROOT
,
2487 .name
= "sched.partition",
2488 .read_s64
= cpuset_read_s64
,
2489 .write_s64
= cpuset_write_s64
,
2490 .private = FILE_PARTITION_ROOT
,
2491 .flags
= CFTYPE_NOT_ON_ROOT
,
2499 * cpuset_css_alloc - allocate a cpuset css
2500 * cgrp: control group that the new cpuset will be part of
2503 static struct cgroup_subsys_state
*
2504 cpuset_css_alloc(struct cgroup_subsys_state
*parent_css
)
2509 return &top_cpuset
.css
;
2511 cs
= kzalloc(sizeof(*cs
), GFP_KERNEL
);
2513 return ERR_PTR(-ENOMEM
);
2515 if (alloc_cpumasks(cs
, NULL
)) {
2517 return ERR_PTR(-ENOMEM
);
2520 set_bit(CS_SCHED_LOAD_BALANCE
, &cs
->flags
);
2521 nodes_clear(cs
->mems_allowed
);
2522 nodes_clear(cs
->effective_mems
);
2523 fmeter_init(&cs
->fmeter
);
2524 cs
->relax_domain_level
= -1;
2529 static int cpuset_css_online(struct cgroup_subsys_state
*css
)
2531 struct cpuset
*cs
= css_cs(css
);
2532 struct cpuset
*parent
= parent_cs(cs
);
2533 struct cpuset
*tmp_cs
;
2534 struct cgroup_subsys_state
*pos_css
;
2539 mutex_lock(&cpuset_mutex
);
2541 set_bit(CS_ONLINE
, &cs
->flags
);
2542 if (is_spread_page(parent
))
2543 set_bit(CS_SPREAD_PAGE
, &cs
->flags
);
2544 if (is_spread_slab(parent
))
2545 set_bit(CS_SPREAD_SLAB
, &cs
->flags
);
2549 spin_lock_irq(&callback_lock
);
2550 if (is_in_v2_mode()) {
2551 cpumask_copy(cs
->effective_cpus
, parent
->effective_cpus
);
2552 cs
->effective_mems
= parent
->effective_mems
;
2554 spin_unlock_irq(&callback_lock
);
2556 if (!test_bit(CGRP_CPUSET_CLONE_CHILDREN
, &css
->cgroup
->flags
))
2560 * Clone @parent's configuration if CGRP_CPUSET_CLONE_CHILDREN is
2561 * set. This flag handling is implemented in cgroup core for
2562 * histrical reasons - the flag may be specified during mount.
2564 * Currently, if any sibling cpusets have exclusive cpus or mem, we
2565 * refuse to clone the configuration - thereby refusing the task to
2566 * be entered, and as a result refusing the sys_unshare() or
2567 * clone() which initiated it. If this becomes a problem for some
2568 * users who wish to allow that scenario, then this could be
2569 * changed to grant parent->cpus_allowed-sibling_cpus_exclusive
2570 * (and likewise for mems) to the new cgroup.
2573 cpuset_for_each_child(tmp_cs
, pos_css
, parent
) {
2574 if (is_mem_exclusive(tmp_cs
) || is_cpu_exclusive(tmp_cs
)) {
2581 spin_lock_irq(&callback_lock
);
2582 cs
->mems_allowed
= parent
->mems_allowed
;
2583 cs
->effective_mems
= parent
->mems_allowed
;
2584 cpumask_copy(cs
->cpus_allowed
, parent
->cpus_allowed
);
2585 cpumask_copy(cs
->effective_cpus
, parent
->cpus_allowed
);
2586 spin_unlock_irq(&callback_lock
);
2588 mutex_unlock(&cpuset_mutex
);
2593 * If the cpuset being removed has its flag 'sched_load_balance'
2594 * enabled, then simulate turning sched_load_balance off, which
2595 * will call rebuild_sched_domains_locked(). That is not needed
2596 * in the default hierarchy where only changes in partition
2597 * will cause repartitioning.
2599 * If the cpuset has the 'sched.partition' flag enabled, simulate
2600 * turning 'sched.partition" off.
2603 static void cpuset_css_offline(struct cgroup_subsys_state
*css
)
2605 struct cpuset
*cs
= css_cs(css
);
2607 mutex_lock(&cpuset_mutex
);
2609 if (is_partition_root(cs
))
2610 update_prstate(cs
, 0);
2612 if (!cgroup_subsys_on_dfl(cpuset_cgrp_subsys
) &&
2613 is_sched_load_balance(cs
))
2614 update_flag(CS_SCHED_LOAD_BALANCE
, cs
, 0);
2617 clear_bit(CS_ONLINE
, &cs
->flags
);
2619 mutex_unlock(&cpuset_mutex
);
2622 static void cpuset_css_free(struct cgroup_subsys_state
*css
)
2624 struct cpuset
*cs
= css_cs(css
);
2629 static void cpuset_bind(struct cgroup_subsys_state
*root_css
)
2631 mutex_lock(&cpuset_mutex
);
2632 spin_lock_irq(&callback_lock
);
2634 if (is_in_v2_mode()) {
2635 cpumask_copy(top_cpuset
.cpus_allowed
, cpu_possible_mask
);
2636 top_cpuset
.mems_allowed
= node_possible_map
;
2638 cpumask_copy(top_cpuset
.cpus_allowed
,
2639 top_cpuset
.effective_cpus
);
2640 top_cpuset
.mems_allowed
= top_cpuset
.effective_mems
;
2643 spin_unlock_irq(&callback_lock
);
2644 mutex_unlock(&cpuset_mutex
);
2648 * Make sure the new task conform to the current state of its parent,
2649 * which could have been changed by cpuset just after it inherits the
2650 * state from the parent and before it sits on the cgroup's task list.
2652 static void cpuset_fork(struct task_struct
*task
)
2654 if (task_css_is_root(task
, cpuset_cgrp_id
))
2657 set_cpus_allowed_ptr(task
, ¤t
->cpus_allowed
);
2658 task
->mems_allowed
= current
->mems_allowed
;
2661 struct cgroup_subsys cpuset_cgrp_subsys
= {
2662 .css_alloc
= cpuset_css_alloc
,
2663 .css_online
= cpuset_css_online
,
2664 .css_offline
= cpuset_css_offline
,
2665 .css_free
= cpuset_css_free
,
2666 .can_attach
= cpuset_can_attach
,
2667 .cancel_attach
= cpuset_cancel_attach
,
2668 .attach
= cpuset_attach
,
2669 .post_attach
= cpuset_post_attach
,
2670 .bind
= cpuset_bind
,
2671 .fork
= cpuset_fork
,
2672 .legacy_cftypes
= legacy_files
,
2673 .dfl_cftypes
= dfl_files
,
2679 * cpuset_init - initialize cpusets at system boot
2681 * Description: Initialize top_cpuset and the cpuset internal file system,
2684 int __init
cpuset_init(void)
2688 BUG_ON(!alloc_cpumask_var(&top_cpuset
.cpus_allowed
, GFP_KERNEL
));
2689 BUG_ON(!alloc_cpumask_var(&top_cpuset
.effective_cpus
, GFP_KERNEL
));
2690 BUG_ON(!zalloc_cpumask_var(&top_cpuset
.subparts_cpus
, GFP_KERNEL
));
2692 cpumask_setall(top_cpuset
.cpus_allowed
);
2693 nodes_setall(top_cpuset
.mems_allowed
);
2694 cpumask_setall(top_cpuset
.effective_cpus
);
2695 nodes_setall(top_cpuset
.effective_mems
);
2697 fmeter_init(&top_cpuset
.fmeter
);
2698 set_bit(CS_SCHED_LOAD_BALANCE
, &top_cpuset
.flags
);
2699 top_cpuset
.relax_domain_level
= -1;
2701 err
= register_filesystem(&cpuset_fs_type
);
2705 BUG_ON(!alloc_cpumask_var(&cpus_attach
, GFP_KERNEL
));
2711 * If CPU and/or memory hotplug handlers, below, unplug any CPUs
2712 * or memory nodes, we need to walk over the cpuset hierarchy,
2713 * removing that CPU or node from all cpusets. If this removes the
2714 * last CPU or node from a cpuset, then move the tasks in the empty
2715 * cpuset to its next-highest non-empty parent.
2717 static void remove_tasks_in_empty_cpuset(struct cpuset
*cs
)
2719 struct cpuset
*parent
;
2722 * Find its next-highest non-empty parent, (top cpuset
2723 * has online cpus, so can't be empty).
2725 parent
= parent_cs(cs
);
2726 while (cpumask_empty(parent
->cpus_allowed
) ||
2727 nodes_empty(parent
->mems_allowed
))
2728 parent
= parent_cs(parent
);
2730 if (cgroup_transfer_tasks(parent
->css
.cgroup
, cs
->css
.cgroup
)) {
2731 pr_err("cpuset: failed to transfer tasks out of empty cpuset ");
2732 pr_cont_cgroup_name(cs
->css
.cgroup
);
2738 hotplug_update_tasks_legacy(struct cpuset
*cs
,
2739 struct cpumask
*new_cpus
, nodemask_t
*new_mems
,
2740 bool cpus_updated
, bool mems_updated
)
2744 spin_lock_irq(&callback_lock
);
2745 cpumask_copy(cs
->cpus_allowed
, new_cpus
);
2746 cpumask_copy(cs
->effective_cpus
, new_cpus
);
2747 cs
->mems_allowed
= *new_mems
;
2748 cs
->effective_mems
= *new_mems
;
2749 spin_unlock_irq(&callback_lock
);
2752 * Don't call update_tasks_cpumask() if the cpuset becomes empty,
2753 * as the tasks will be migratecd to an ancestor.
2755 if (cpus_updated
&& !cpumask_empty(cs
->cpus_allowed
))
2756 update_tasks_cpumask(cs
);
2757 if (mems_updated
&& !nodes_empty(cs
->mems_allowed
))
2758 update_tasks_nodemask(cs
);
2760 is_empty
= cpumask_empty(cs
->cpus_allowed
) ||
2761 nodes_empty(cs
->mems_allowed
);
2763 mutex_unlock(&cpuset_mutex
);
2766 * Move tasks to the nearest ancestor with execution resources,
2767 * This is full cgroup operation which will also call back into
2768 * cpuset. Should be done outside any lock.
2771 remove_tasks_in_empty_cpuset(cs
);
2773 mutex_lock(&cpuset_mutex
);
2777 hotplug_update_tasks(struct cpuset
*cs
,
2778 struct cpumask
*new_cpus
, nodemask_t
*new_mems
,
2779 bool cpus_updated
, bool mems_updated
)
2781 if (cpumask_empty(new_cpus
))
2782 cpumask_copy(new_cpus
, parent_cs(cs
)->effective_cpus
);
2783 if (nodes_empty(*new_mems
))
2784 *new_mems
= parent_cs(cs
)->effective_mems
;
2786 spin_lock_irq(&callback_lock
);
2787 cpumask_copy(cs
->effective_cpus
, new_cpus
);
2788 cs
->effective_mems
= *new_mems
;
2789 spin_unlock_irq(&callback_lock
);
2792 update_tasks_cpumask(cs
);
2794 update_tasks_nodemask(cs
);
2798 * cpuset_hotplug_update_tasks - update tasks in a cpuset for hotunplug
2799 * @cs: cpuset in interest
2801 * Compare @cs's cpu and mem masks against top_cpuset and if some have gone
2802 * offline, update @cs accordingly. If @cs ends up with no CPU or memory,
2803 * all its tasks are moved to the nearest ancestor with both resources.
2805 static void cpuset_hotplug_update_tasks(struct cpuset
*cs
)
2807 static cpumask_t new_cpus
;
2808 static nodemask_t new_mems
;
2812 wait_event(cpuset_attach_wq
, cs
->attach_in_progress
== 0);
2814 mutex_lock(&cpuset_mutex
);
2817 * We have raced with task attaching. We wait until attaching
2818 * is finished, so we won't attach a task to an empty cpuset.
2820 if (cs
->attach_in_progress
) {
2821 mutex_unlock(&cpuset_mutex
);
2825 cpumask_and(&new_cpus
, cs
->cpus_allowed
, parent_cs(cs
)->effective_cpus
);
2826 nodes_and(new_mems
, cs
->mems_allowed
, parent_cs(cs
)->effective_mems
);
2828 cpus_updated
= !cpumask_equal(&new_cpus
, cs
->effective_cpus
);
2829 mems_updated
= !nodes_equal(new_mems
, cs
->effective_mems
);
2831 if (is_in_v2_mode())
2832 hotplug_update_tasks(cs
, &new_cpus
, &new_mems
,
2833 cpus_updated
, mems_updated
);
2835 hotplug_update_tasks_legacy(cs
, &new_cpus
, &new_mems
,
2836 cpus_updated
, mems_updated
);
2838 mutex_unlock(&cpuset_mutex
);
2841 static bool force_rebuild
;
2843 void cpuset_force_rebuild(void)
2845 force_rebuild
= true;
2849 * cpuset_hotplug_workfn - handle CPU/memory hotunplug for a cpuset
2851 * This function is called after either CPU or memory configuration has
2852 * changed and updates cpuset accordingly. The top_cpuset is always
2853 * synchronized to cpu_active_mask and N_MEMORY, which is necessary in
2854 * order to make cpusets transparent (of no affect) on systems that are
2855 * actively using CPU hotplug but making no active use of cpusets.
2857 * Non-root cpusets are only affected by offlining. If any CPUs or memory
2858 * nodes have been taken down, cpuset_hotplug_update_tasks() is invoked on
2861 * Note that CPU offlining during suspend is ignored. We don't modify
2862 * cpusets across suspend/resume cycles at all.
2864 static void cpuset_hotplug_workfn(struct work_struct
*work
)
2866 static cpumask_t new_cpus
;
2867 static nodemask_t new_mems
;
2868 bool cpus_updated
, mems_updated
;
2869 bool on_dfl
= is_in_v2_mode();
2871 mutex_lock(&cpuset_mutex
);
2873 /* fetch the available cpus/mems and find out which changed how */
2874 cpumask_copy(&new_cpus
, cpu_active_mask
);
2875 new_mems
= node_states
[N_MEMORY
];
2877 cpus_updated
= !cpumask_equal(top_cpuset
.effective_cpus
, &new_cpus
);
2878 mems_updated
= !nodes_equal(top_cpuset
.effective_mems
, new_mems
);
2880 /* synchronize cpus_allowed to cpu_active_mask */
2882 spin_lock_irq(&callback_lock
);
2884 cpumask_copy(top_cpuset
.cpus_allowed
, &new_cpus
);
2885 cpumask_copy(top_cpuset
.effective_cpus
, &new_cpus
);
2886 spin_unlock_irq(&callback_lock
);
2887 /* we don't mess with cpumasks of tasks in top_cpuset */
2890 /* synchronize mems_allowed to N_MEMORY */
2892 spin_lock_irq(&callback_lock
);
2894 top_cpuset
.mems_allowed
= new_mems
;
2895 top_cpuset
.effective_mems
= new_mems
;
2896 spin_unlock_irq(&callback_lock
);
2897 update_tasks_nodemask(&top_cpuset
);
2900 mutex_unlock(&cpuset_mutex
);
2902 /* if cpus or mems changed, we need to propagate to descendants */
2903 if (cpus_updated
|| mems_updated
) {
2905 struct cgroup_subsys_state
*pos_css
;
2908 cpuset_for_each_descendant_pre(cs
, pos_css
, &top_cpuset
) {
2909 if (cs
== &top_cpuset
|| !css_tryget_online(&cs
->css
))
2913 cpuset_hotplug_update_tasks(cs
);
2921 /* rebuild sched domains if cpus_allowed has changed */
2922 if (cpus_updated
|| force_rebuild
) {
2923 force_rebuild
= false;
2924 rebuild_sched_domains();
2928 void cpuset_update_active_cpus(void)
2931 * We're inside cpu hotplug critical region which usually nests
2932 * inside cgroup synchronization. Bounce actual hotplug processing
2933 * to a work item to avoid reverse locking order.
2935 schedule_work(&cpuset_hotplug_work
);
2938 void cpuset_wait_for_hotplug(void)
2940 flush_work(&cpuset_hotplug_work
);
2944 * Keep top_cpuset.mems_allowed tracking node_states[N_MEMORY].
2945 * Call this routine anytime after node_states[N_MEMORY] changes.
2946 * See cpuset_update_active_cpus() for CPU hotplug handling.
2948 static int cpuset_track_online_nodes(struct notifier_block
*self
,
2949 unsigned long action
, void *arg
)
2951 schedule_work(&cpuset_hotplug_work
);
2955 static struct notifier_block cpuset_track_online_nodes_nb
= {
2956 .notifier_call
= cpuset_track_online_nodes
,
2957 .priority
= 10, /* ??! */
2961 * cpuset_init_smp - initialize cpus_allowed
2963 * Description: Finish top cpuset after cpu, node maps are initialized
2965 void __init
cpuset_init_smp(void)
2967 cpumask_copy(top_cpuset
.cpus_allowed
, cpu_active_mask
);
2968 top_cpuset
.mems_allowed
= node_states
[N_MEMORY
];
2969 top_cpuset
.old_mems_allowed
= top_cpuset
.mems_allowed
;
2971 cpumask_copy(top_cpuset
.effective_cpus
, cpu_active_mask
);
2972 top_cpuset
.effective_mems
= node_states
[N_MEMORY
];
2974 register_hotmemory_notifier(&cpuset_track_online_nodes_nb
);
2976 cpuset_migrate_mm_wq
= alloc_ordered_workqueue("cpuset_migrate_mm", 0);
2977 BUG_ON(!cpuset_migrate_mm_wq
);
2981 * cpuset_cpus_allowed - return cpus_allowed mask from a tasks cpuset.
2982 * @tsk: pointer to task_struct from which to obtain cpuset->cpus_allowed.
2983 * @pmask: pointer to struct cpumask variable to receive cpus_allowed set.
2985 * Description: Returns the cpumask_var_t cpus_allowed of the cpuset
2986 * attached to the specified @tsk. Guaranteed to return some non-empty
2987 * subset of cpu_online_mask, even if this means going outside the
2991 void cpuset_cpus_allowed(struct task_struct
*tsk
, struct cpumask
*pmask
)
2993 unsigned long flags
;
2995 spin_lock_irqsave(&callback_lock
, flags
);
2997 guarantee_online_cpus(task_cs(tsk
), pmask
);
2999 spin_unlock_irqrestore(&callback_lock
, flags
);
3002 void cpuset_cpus_allowed_fallback(struct task_struct
*tsk
)
3005 do_set_cpus_allowed(tsk
, task_cs(tsk
)->effective_cpus
);
3009 * We own tsk->cpus_allowed, nobody can change it under us.
3011 * But we used cs && cs->cpus_allowed lockless and thus can
3012 * race with cgroup_attach_task() or update_cpumask() and get
3013 * the wrong tsk->cpus_allowed. However, both cases imply the
3014 * subsequent cpuset_change_cpumask()->set_cpus_allowed_ptr()
3015 * which takes task_rq_lock().
3017 * If we are called after it dropped the lock we must see all
3018 * changes in tsk_cs()->cpus_allowed. Otherwise we can temporary
3019 * set any mask even if it is not right from task_cs() pov,
3020 * the pending set_cpus_allowed_ptr() will fix things.
3022 * select_fallback_rq() will fix things ups and set cpu_possible_mask
3027 void __init
cpuset_init_current_mems_allowed(void)
3029 nodes_setall(current
->mems_allowed
);
3033 * cpuset_mems_allowed - return mems_allowed mask from a tasks cpuset.
3034 * @tsk: pointer to task_struct from which to obtain cpuset->mems_allowed.
3036 * Description: Returns the nodemask_t mems_allowed of the cpuset
3037 * attached to the specified @tsk. Guaranteed to return some non-empty
3038 * subset of node_states[N_MEMORY], even if this means going outside the
3042 nodemask_t
cpuset_mems_allowed(struct task_struct
*tsk
)
3045 unsigned long flags
;
3047 spin_lock_irqsave(&callback_lock
, flags
);
3049 guarantee_online_mems(task_cs(tsk
), &mask
);
3051 spin_unlock_irqrestore(&callback_lock
, flags
);
3057 * cpuset_nodemask_valid_mems_allowed - check nodemask vs. curremt mems_allowed
3058 * @nodemask: the nodemask to be checked
3060 * Are any of the nodes in the nodemask allowed in current->mems_allowed?
3062 int cpuset_nodemask_valid_mems_allowed(nodemask_t
*nodemask
)
3064 return nodes_intersects(*nodemask
, current
->mems_allowed
);
3068 * nearest_hardwall_ancestor() - Returns the nearest mem_exclusive or
3069 * mem_hardwall ancestor to the specified cpuset. Call holding
3070 * callback_lock. If no ancestor is mem_exclusive or mem_hardwall
3071 * (an unusual configuration), then returns the root cpuset.
3073 static struct cpuset
*nearest_hardwall_ancestor(struct cpuset
*cs
)
3075 while (!(is_mem_exclusive(cs
) || is_mem_hardwall(cs
)) && parent_cs(cs
))
3081 * cpuset_node_allowed - Can we allocate on a memory node?
3082 * @node: is this an allowed node?
3083 * @gfp_mask: memory allocation flags
3085 * If we're in interrupt, yes, we can always allocate. If @node is set in
3086 * current's mems_allowed, yes. If it's not a __GFP_HARDWALL request and this
3087 * node is set in the nearest hardwalled cpuset ancestor to current's cpuset,
3088 * yes. If current has access to memory reserves as an oom victim, yes.
3091 * GFP_USER allocations are marked with the __GFP_HARDWALL bit,
3092 * and do not allow allocations outside the current tasks cpuset
3093 * unless the task has been OOM killed.
3094 * GFP_KERNEL allocations are not so marked, so can escape to the
3095 * nearest enclosing hardwalled ancestor cpuset.
3097 * Scanning up parent cpusets requires callback_lock. The
3098 * __alloc_pages() routine only calls here with __GFP_HARDWALL bit
3099 * _not_ set if it's a GFP_KERNEL allocation, and all nodes in the
3100 * current tasks mems_allowed came up empty on the first pass over
3101 * the zonelist. So only GFP_KERNEL allocations, if all nodes in the
3102 * cpuset are short of memory, might require taking the callback_lock.
3104 * The first call here from mm/page_alloc:get_page_from_freelist()
3105 * has __GFP_HARDWALL set in gfp_mask, enforcing hardwall cpusets,
3106 * so no allocation on a node outside the cpuset is allowed (unless
3107 * in interrupt, of course).
3109 * The second pass through get_page_from_freelist() doesn't even call
3110 * here for GFP_ATOMIC calls. For those calls, the __alloc_pages()
3111 * variable 'wait' is not set, and the bit ALLOC_CPUSET is not set
3112 * in alloc_flags. That logic and the checks below have the combined
3114 * in_interrupt - any node ok (current task context irrelevant)
3115 * GFP_ATOMIC - any node ok
3116 * tsk_is_oom_victim - any node ok
3117 * GFP_KERNEL - any node in enclosing hardwalled cpuset ok
3118 * GFP_USER - only nodes in current tasks mems allowed ok.
3120 bool __cpuset_node_allowed(int node
, gfp_t gfp_mask
)
3122 struct cpuset
*cs
; /* current cpuset ancestors */
3123 int allowed
; /* is allocation in zone z allowed? */
3124 unsigned long flags
;
3128 if (node_isset(node
, current
->mems_allowed
))
3131 * Allow tasks that have access to memory reserves because they have
3132 * been OOM killed to get memory anywhere.
3134 if (unlikely(tsk_is_oom_victim(current
)))
3136 if (gfp_mask
& __GFP_HARDWALL
) /* If hardwall request, stop here */
3139 if (current
->flags
& PF_EXITING
) /* Let dying task have memory */
3142 /* Not hardwall and node outside mems_allowed: scan up cpusets */
3143 spin_lock_irqsave(&callback_lock
, flags
);
3146 cs
= nearest_hardwall_ancestor(task_cs(current
));
3147 allowed
= node_isset(node
, cs
->mems_allowed
);
3150 spin_unlock_irqrestore(&callback_lock
, flags
);
3155 * cpuset_mem_spread_node() - On which node to begin search for a file page
3156 * cpuset_slab_spread_node() - On which node to begin search for a slab page
3158 * If a task is marked PF_SPREAD_PAGE or PF_SPREAD_SLAB (as for
3159 * tasks in a cpuset with is_spread_page or is_spread_slab set),
3160 * and if the memory allocation used cpuset_mem_spread_node()
3161 * to determine on which node to start looking, as it will for
3162 * certain page cache or slab cache pages such as used for file
3163 * system buffers and inode caches, then instead of starting on the
3164 * local node to look for a free page, rather spread the starting
3165 * node around the tasks mems_allowed nodes.
3167 * We don't have to worry about the returned node being offline
3168 * because "it can't happen", and even if it did, it would be ok.
3170 * The routines calling guarantee_online_mems() are careful to
3171 * only set nodes in task->mems_allowed that are online. So it
3172 * should not be possible for the following code to return an
3173 * offline node. But if it did, that would be ok, as this routine
3174 * is not returning the node where the allocation must be, only
3175 * the node where the search should start. The zonelist passed to
3176 * __alloc_pages() will include all nodes. If the slab allocator
3177 * is passed an offline node, it will fall back to the local node.
3178 * See kmem_cache_alloc_node().
3181 static int cpuset_spread_node(int *rotor
)
3183 return *rotor
= next_node_in(*rotor
, current
->mems_allowed
);
3186 int cpuset_mem_spread_node(void)
3188 if (current
->cpuset_mem_spread_rotor
== NUMA_NO_NODE
)
3189 current
->cpuset_mem_spread_rotor
=
3190 node_random(¤t
->mems_allowed
);
3192 return cpuset_spread_node(¤t
->cpuset_mem_spread_rotor
);
3195 int cpuset_slab_spread_node(void)
3197 if (current
->cpuset_slab_spread_rotor
== NUMA_NO_NODE
)
3198 current
->cpuset_slab_spread_rotor
=
3199 node_random(¤t
->mems_allowed
);
3201 return cpuset_spread_node(¤t
->cpuset_slab_spread_rotor
);
3204 EXPORT_SYMBOL_GPL(cpuset_mem_spread_node
);
3207 * cpuset_mems_allowed_intersects - Does @tsk1's mems_allowed intersect @tsk2's?
3208 * @tsk1: pointer to task_struct of some task.
3209 * @tsk2: pointer to task_struct of some other task.
3211 * Description: Return true if @tsk1's mems_allowed intersects the
3212 * mems_allowed of @tsk2. Used by the OOM killer to determine if
3213 * one of the task's memory usage might impact the memory available
3217 int cpuset_mems_allowed_intersects(const struct task_struct
*tsk1
,
3218 const struct task_struct
*tsk2
)
3220 return nodes_intersects(tsk1
->mems_allowed
, tsk2
->mems_allowed
);
3224 * cpuset_print_current_mems_allowed - prints current's cpuset and mems_allowed
3226 * Description: Prints current's name, cpuset name, and cached copy of its
3227 * mems_allowed to the kernel log.
3229 void cpuset_print_current_mems_allowed(void)
3231 struct cgroup
*cgrp
;
3235 cgrp
= task_cs(current
)->css
.cgroup
;
3236 pr_info("%s cpuset=", current
->comm
);
3237 pr_cont_cgroup_name(cgrp
);
3238 pr_cont(" mems_allowed=%*pbl\n",
3239 nodemask_pr_args(¤t
->mems_allowed
));
3245 * Collection of memory_pressure is suppressed unless
3246 * this flag is enabled by writing "1" to the special
3247 * cpuset file 'memory_pressure_enabled' in the root cpuset.
3250 int cpuset_memory_pressure_enabled __read_mostly
;
3253 * cpuset_memory_pressure_bump - keep stats of per-cpuset reclaims.
3255 * Keep a running average of the rate of synchronous (direct)
3256 * page reclaim efforts initiated by tasks in each cpuset.
3258 * This represents the rate at which some task in the cpuset
3259 * ran low on memory on all nodes it was allowed to use, and
3260 * had to enter the kernels page reclaim code in an effort to
3261 * create more free memory by tossing clean pages or swapping
3262 * or writing dirty pages.
3264 * Display to user space in the per-cpuset read-only file
3265 * "memory_pressure". Value displayed is an integer
3266 * representing the recent rate of entry into the synchronous
3267 * (direct) page reclaim by any task attached to the cpuset.
3270 void __cpuset_memory_pressure_bump(void)
3273 fmeter_markevent(&task_cs(current
)->fmeter
);
3277 #ifdef CONFIG_PROC_PID_CPUSET
3279 * proc_cpuset_show()
3280 * - Print tasks cpuset path into seq_file.
3281 * - Used for /proc/<pid>/cpuset.
3282 * - No need to task_lock(tsk) on this tsk->cpuset reference, as it
3283 * doesn't really matter if tsk->cpuset changes after we read it,
3284 * and we take cpuset_mutex, keeping cpuset_attach() from changing it
3287 int proc_cpuset_show(struct seq_file
*m
, struct pid_namespace
*ns
,
3288 struct pid
*pid
, struct task_struct
*tsk
)
3291 struct cgroup_subsys_state
*css
;
3295 buf
= kmalloc(PATH_MAX
, GFP_KERNEL
);
3299 css
= task_get_css(tsk
, cpuset_cgrp_id
);
3300 retval
= cgroup_path_ns(css
->cgroup
, buf
, PATH_MAX
,
3301 current
->nsproxy
->cgroup_ns
);
3303 if (retval
>= PATH_MAX
)
3304 retval
= -ENAMETOOLONG
;
3315 #endif /* CONFIG_PROC_PID_CPUSET */
3317 /* Display task mems_allowed in /proc/<pid>/status file. */
3318 void cpuset_task_status_allowed(struct seq_file
*m
, struct task_struct
*task
)
3320 seq_printf(m
, "Mems_allowed:\t%*pb\n",
3321 nodemask_pr_args(&task
->mems_allowed
));
3322 seq_printf(m
, "Mems_allowed_list:\t%*pbl\n",
3323 nodemask_pr_args(&task
->mems_allowed
));