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/module.h>
40 #include <linux/mount.h>
41 #include <linux/namei.h>
42 #include <linux/pagemap.h>
43 #include <linux/proc_fs.h>
44 #include <linux/rcupdate.h>
45 #include <linux/sched.h>
46 #include <linux/seq_file.h>
47 #include <linux/security.h>
48 #include <linux/slab.h>
49 #include <linux/spinlock.h>
50 #include <linux/stat.h>
51 #include <linux/string.h>
52 #include <linux/time.h>
53 #include <linux/backing-dev.h>
54 #include <linux/sort.h>
56 #include <asm/uaccess.h>
57 #include <asm/atomic.h>
58 #include <linux/mutex.h>
59 #include <linux/workqueue.h>
60 #include <linux/cgroup.h>
63 * Tracks how many cpusets are currently defined in system.
64 * When there is only one cpuset (the root cpuset) we can
65 * short circuit some hooks.
67 int number_of_cpusets __read_mostly
;
69 /* Forward declare cgroup structures */
70 struct cgroup_subsys cpuset_subsys
;
73 /* See "Frequency meter" comments, below. */
76 int cnt
; /* unprocessed events count */
77 int val
; /* most recent output value */
78 time_t time
; /* clock (secs) when val computed */
79 spinlock_t lock
; /* guards read or write of above */
83 struct cgroup_subsys_state css
;
85 unsigned long flags
; /* "unsigned long" so bitops work */
86 cpumask_t cpus_allowed
; /* CPUs allowed to tasks in cpuset */
87 nodemask_t mems_allowed
; /* Memory Nodes allowed to tasks */
89 struct cpuset
*parent
; /* my parent */
92 * Copy of global cpuset_mems_generation as of the most
93 * recent time this cpuset changed its mems_allowed.
97 struct fmeter fmeter
; /* memory_pressure filter */
99 /* partition number for rebuild_sched_domains() */
102 /* for custom sched domain */
103 int relax_domain_level
;
105 /* used for walking a cpuset heirarchy */
106 struct list_head stack_list
;
109 /* Retrieve the cpuset for a cgroup */
110 static inline struct cpuset
*cgroup_cs(struct cgroup
*cont
)
112 return container_of(cgroup_subsys_state(cont
, cpuset_subsys_id
),
116 /* Retrieve the cpuset for a task */
117 static inline struct cpuset
*task_cs(struct task_struct
*task
)
119 return container_of(task_subsys_state(task
, cpuset_subsys_id
),
122 struct cpuset_hotplug_scanner
{
123 struct cgroup_scanner scan
;
127 /* bits in struct cpuset flags field */
133 CS_SCHED_LOAD_BALANCE
,
138 /* convenient tests for these bits */
139 static inline int is_cpu_exclusive(const struct cpuset
*cs
)
141 return test_bit(CS_CPU_EXCLUSIVE
, &cs
->flags
);
144 static inline int is_mem_exclusive(const struct cpuset
*cs
)
146 return test_bit(CS_MEM_EXCLUSIVE
, &cs
->flags
);
149 static inline int is_mem_hardwall(const struct cpuset
*cs
)
151 return test_bit(CS_MEM_HARDWALL
, &cs
->flags
);
154 static inline int is_sched_load_balance(const struct cpuset
*cs
)
156 return test_bit(CS_SCHED_LOAD_BALANCE
, &cs
->flags
);
159 static inline int is_memory_migrate(const struct cpuset
*cs
)
161 return test_bit(CS_MEMORY_MIGRATE
, &cs
->flags
);
164 static inline int is_spread_page(const struct cpuset
*cs
)
166 return test_bit(CS_SPREAD_PAGE
, &cs
->flags
);
169 static inline int is_spread_slab(const struct cpuset
*cs
)
171 return test_bit(CS_SPREAD_SLAB
, &cs
->flags
);
175 * Increment this integer everytime any cpuset changes its
176 * mems_allowed value. Users of cpusets can track this generation
177 * number, and avoid having to lock and reload mems_allowed unless
178 * the cpuset they're using changes generation.
180 * A single, global generation is needed because cpuset_attach_task() could
181 * reattach a task to a different cpuset, which must not have its
182 * generation numbers aliased with those of that tasks previous cpuset.
184 * Generations are needed for mems_allowed because one task cannot
185 * modify another's memory placement. So we must enable every task,
186 * on every visit to __alloc_pages(), to efficiently check whether
187 * its current->cpuset->mems_allowed has changed, requiring an update
188 * of its current->mems_allowed.
190 * Since writes to cpuset_mems_generation are guarded by the cgroup lock
191 * there is no need to mark it atomic.
193 static int cpuset_mems_generation
;
195 static struct cpuset top_cpuset
= {
196 .flags
= ((1 << CS_CPU_EXCLUSIVE
) | (1 << CS_MEM_EXCLUSIVE
)),
197 .cpus_allowed
= CPU_MASK_ALL
,
198 .mems_allowed
= NODE_MASK_ALL
,
202 * There are two global mutexes guarding cpuset structures. The first
203 * is the main control groups cgroup_mutex, accessed via
204 * cgroup_lock()/cgroup_unlock(). The second is the cpuset-specific
205 * callback_mutex, below. They can nest. It is ok to first take
206 * cgroup_mutex, then nest callback_mutex. We also require taking
207 * task_lock() when dereferencing a task's cpuset pointer. See "The
208 * task_lock() exception", at the end of this comment.
210 * A task must hold both mutexes to modify cpusets. If a task
211 * holds cgroup_mutex, then it blocks others wanting that mutex,
212 * ensuring that it is the only task able to also acquire callback_mutex
213 * and be able to modify cpusets. It can perform various checks on
214 * the cpuset structure first, knowing nothing will change. It can
215 * also allocate memory while just holding cgroup_mutex. While it is
216 * performing these checks, various callback routines can briefly
217 * acquire callback_mutex to query cpusets. Once it is ready to make
218 * the changes, it takes callback_mutex, blocking everyone else.
220 * Calls to the kernel memory allocator can not be made while holding
221 * callback_mutex, as that would risk double tripping on callback_mutex
222 * from one of the callbacks into the cpuset code from within
225 * If a task is only holding callback_mutex, then it has read-only
228 * The task_struct fields mems_allowed and mems_generation may only
229 * be accessed in the context of that task, so require no locks.
231 * The cpuset_common_file_read() handlers only hold callback_mutex across
232 * small pieces of code, such as when reading out possibly multi-word
233 * cpumasks and nodemasks.
235 * Accessing a task's cpuset should be done in accordance with the
236 * guidelines for accessing subsystem state in kernel/cgroup.c
239 static DEFINE_MUTEX(callback_mutex
);
242 * This is ugly, but preserves the userspace API for existing cpuset
243 * users. If someone tries to mount the "cpuset" filesystem, we
244 * silently switch it to mount "cgroup" instead
246 static int cpuset_get_sb(struct file_system_type
*fs_type
,
247 int flags
, const char *unused_dev_name
,
248 void *data
, struct vfsmount
*mnt
)
250 struct file_system_type
*cgroup_fs
= get_fs_type("cgroup");
255 "release_agent=/sbin/cpuset_release_agent";
256 ret
= cgroup_fs
->get_sb(cgroup_fs
, flags
,
257 unused_dev_name
, mountopts
, mnt
);
258 put_filesystem(cgroup_fs
);
263 static struct file_system_type cpuset_fs_type
= {
265 .get_sb
= cpuset_get_sb
,
269 * Return in *pmask the portion of a cpusets's cpus_allowed that
270 * are online. If none are online, walk up the cpuset hierarchy
271 * until we find one that does have some online cpus. If we get
272 * all the way to the top and still haven't found any online cpus,
273 * return cpu_online_map. Or if passed a NULL cs from an exit'ing
274 * task, return cpu_online_map.
276 * One way or another, we guarantee to return some non-empty subset
279 * Call with callback_mutex held.
282 static void guarantee_online_cpus(const struct cpuset
*cs
, cpumask_t
*pmask
)
284 while (cs
&& !cpus_intersects(cs
->cpus_allowed
, cpu_online_map
))
287 cpus_and(*pmask
, cs
->cpus_allowed
, cpu_online_map
);
289 *pmask
= cpu_online_map
;
290 BUG_ON(!cpus_intersects(*pmask
, cpu_online_map
));
294 * Return in *pmask the portion of a cpusets's mems_allowed that
295 * are online, with memory. If none are online with memory, walk
296 * up the cpuset hierarchy until we find one that does have some
297 * online mems. If we get all the way to the top and still haven't
298 * found any online mems, return node_states[N_HIGH_MEMORY].
300 * One way or another, we guarantee to return some non-empty subset
301 * of node_states[N_HIGH_MEMORY].
303 * Call with callback_mutex held.
306 static void guarantee_online_mems(const struct cpuset
*cs
, nodemask_t
*pmask
)
308 while (cs
&& !nodes_intersects(cs
->mems_allowed
,
309 node_states
[N_HIGH_MEMORY
]))
312 nodes_and(*pmask
, cs
->mems_allowed
,
313 node_states
[N_HIGH_MEMORY
]);
315 *pmask
= node_states
[N_HIGH_MEMORY
];
316 BUG_ON(!nodes_intersects(*pmask
, node_states
[N_HIGH_MEMORY
]));
320 * cpuset_update_task_memory_state - update task memory placement
322 * If the current tasks cpusets mems_allowed changed behind our
323 * backs, update current->mems_allowed, mems_generation and task NUMA
324 * mempolicy to the new value.
326 * Task mempolicy is updated by rebinding it relative to the
327 * current->cpuset if a task has its memory placement changed.
328 * Do not call this routine if in_interrupt().
330 * Call without callback_mutex or task_lock() held. May be
331 * called with or without cgroup_mutex held. Thanks in part to
332 * 'the_top_cpuset_hack', the task's cpuset pointer will never
333 * be NULL. This routine also might acquire callback_mutex during
336 * Reading current->cpuset->mems_generation doesn't need task_lock
337 * to guard the current->cpuset derefence, because it is guarded
338 * from concurrent freeing of current->cpuset using RCU.
340 * The rcu_dereference() is technically probably not needed,
341 * as I don't actually mind if I see a new cpuset pointer but
342 * an old value of mems_generation. However this really only
343 * matters on alpha systems using cpusets heavily. If I dropped
344 * that rcu_dereference(), it would save them a memory barrier.
345 * For all other arch's, rcu_dereference is a no-op anyway, and for
346 * alpha systems not using cpusets, another planned optimization,
347 * avoiding the rcu critical section for tasks in the root cpuset
348 * which is statically allocated, so can't vanish, will make this
349 * irrelevant. Better to use RCU as intended, than to engage in
350 * some cute trick to save a memory barrier that is impossible to
351 * test, for alpha systems using cpusets heavily, which might not
354 * This routine is needed to update the per-task mems_allowed data,
355 * within the tasks context, when it is trying to allocate memory
356 * (in various mm/mempolicy.c routines) and notices that some other
357 * task has been modifying its cpuset.
360 void cpuset_update_task_memory_state(void)
362 int my_cpusets_mem_gen
;
363 struct task_struct
*tsk
= current
;
366 if (task_cs(tsk
) == &top_cpuset
) {
367 /* Don't need rcu for top_cpuset. It's never freed. */
368 my_cpusets_mem_gen
= top_cpuset
.mems_generation
;
371 my_cpusets_mem_gen
= task_cs(tsk
)->mems_generation
;
375 if (my_cpusets_mem_gen
!= tsk
->cpuset_mems_generation
) {
376 mutex_lock(&callback_mutex
);
378 cs
= task_cs(tsk
); /* Maybe changed when task not locked */
379 guarantee_online_mems(cs
, &tsk
->mems_allowed
);
380 tsk
->cpuset_mems_generation
= cs
->mems_generation
;
381 if (is_spread_page(cs
))
382 tsk
->flags
|= PF_SPREAD_PAGE
;
384 tsk
->flags
&= ~PF_SPREAD_PAGE
;
385 if (is_spread_slab(cs
))
386 tsk
->flags
|= PF_SPREAD_SLAB
;
388 tsk
->flags
&= ~PF_SPREAD_SLAB
;
390 mutex_unlock(&callback_mutex
);
391 mpol_rebind_task(tsk
, &tsk
->mems_allowed
);
396 * is_cpuset_subset(p, q) - Is cpuset p a subset of cpuset q?
398 * One cpuset is a subset of another if all its allowed CPUs and
399 * Memory Nodes are a subset of the other, and its exclusive flags
400 * are only set if the other's are set. Call holding cgroup_mutex.
403 static int is_cpuset_subset(const struct cpuset
*p
, const struct cpuset
*q
)
405 return cpus_subset(p
->cpus_allowed
, q
->cpus_allowed
) &&
406 nodes_subset(p
->mems_allowed
, q
->mems_allowed
) &&
407 is_cpu_exclusive(p
) <= is_cpu_exclusive(q
) &&
408 is_mem_exclusive(p
) <= is_mem_exclusive(q
);
412 * validate_change() - Used to validate that any proposed cpuset change
413 * follows the structural rules for cpusets.
415 * If we replaced the flag and mask values of the current cpuset
416 * (cur) with those values in the trial cpuset (trial), would
417 * our various subset and exclusive rules still be valid? Presumes
420 * 'cur' is the address of an actual, in-use cpuset. Operations
421 * such as list traversal that depend on the actual address of the
422 * cpuset in the list must use cur below, not trial.
424 * 'trial' is the address of bulk structure copy of cur, with
425 * perhaps one or more of the fields cpus_allowed, mems_allowed,
426 * or flags changed to new, trial values.
428 * Return 0 if valid, -errno if not.
431 static int validate_change(const struct cpuset
*cur
, const struct cpuset
*trial
)
434 struct cpuset
*c
, *par
;
436 /* Each of our child cpusets must be a subset of us */
437 list_for_each_entry(cont
, &cur
->css
.cgroup
->children
, sibling
) {
438 if (!is_cpuset_subset(cgroup_cs(cont
), trial
))
442 /* Remaining checks don't apply to root cpuset */
443 if (cur
== &top_cpuset
)
448 /* We must be a subset of our parent cpuset */
449 if (!is_cpuset_subset(trial
, par
))
453 * If either I or some sibling (!= me) is exclusive, we can't
456 list_for_each_entry(cont
, &par
->css
.cgroup
->children
, sibling
) {
458 if ((is_cpu_exclusive(trial
) || is_cpu_exclusive(c
)) &&
460 cpus_intersects(trial
->cpus_allowed
, c
->cpus_allowed
))
462 if ((is_mem_exclusive(trial
) || is_mem_exclusive(c
)) &&
464 nodes_intersects(trial
->mems_allowed
, c
->mems_allowed
))
468 /* Cpusets with tasks can't have empty cpus_allowed or mems_allowed */
469 if (cgroup_task_count(cur
->css
.cgroup
)) {
470 if (cpus_empty(trial
->cpus_allowed
) ||
471 nodes_empty(trial
->mems_allowed
)) {
480 * Helper routine for generate_sched_domains().
481 * Do cpusets a, b have overlapping cpus_allowed masks?
483 static int cpusets_overlap(struct cpuset
*a
, struct cpuset
*b
)
485 return cpus_intersects(a
->cpus_allowed
, b
->cpus_allowed
);
489 update_domain_attr(struct sched_domain_attr
*dattr
, struct cpuset
*c
)
491 if (dattr
->relax_domain_level
< c
->relax_domain_level
)
492 dattr
->relax_domain_level
= c
->relax_domain_level
;
497 update_domain_attr_tree(struct sched_domain_attr
*dattr
, struct cpuset
*c
)
501 list_add(&c
->stack_list
, &q
);
502 while (!list_empty(&q
)) {
505 struct cpuset
*child
;
507 cp
= list_first_entry(&q
, struct cpuset
, stack_list
);
510 if (cpus_empty(cp
->cpus_allowed
))
513 if (is_sched_load_balance(cp
))
514 update_domain_attr(dattr
, cp
);
516 list_for_each_entry(cont
, &cp
->css
.cgroup
->children
, sibling
) {
517 child
= cgroup_cs(cont
);
518 list_add_tail(&child
->stack_list
, &q
);
524 * generate_sched_domains()
526 * This function builds a partial partition of the systems CPUs
527 * A 'partial partition' is a set of non-overlapping subsets whose
528 * union is a subset of that set.
529 * The output of this function needs to be passed to kernel/sched.c
530 * partition_sched_domains() routine, which will rebuild the scheduler's
531 * load balancing domains (sched domains) as specified by that partial
534 * See "What is sched_load_balance" in Documentation/cpusets.txt
535 * for a background explanation of this.
537 * Does not return errors, on the theory that the callers of this
538 * routine would rather not worry about failures to rebuild sched
539 * domains when operating in the severe memory shortage situations
540 * that could cause allocation failures below.
542 * Must be called with cgroup_lock held.
544 * The three key local variables below are:
545 * q - a linked-list queue of cpuset pointers, used to implement a
546 * top-down scan of all cpusets. This scan loads a pointer
547 * to each cpuset marked is_sched_load_balance into the
548 * array 'csa'. For our purposes, rebuilding the schedulers
549 * sched domains, we can ignore !is_sched_load_balance cpusets.
550 * csa - (for CpuSet Array) Array of pointers to all the cpusets
551 * that need to be load balanced, for convenient iterative
552 * access by the subsequent code that finds the best partition,
553 * i.e the set of domains (subsets) of CPUs such that the
554 * cpus_allowed of every cpuset marked is_sched_load_balance
555 * is a subset of one of these domains, while there are as
556 * many such domains as possible, each as small as possible.
557 * doms - Conversion of 'csa' to an array of cpumasks, for passing to
558 * the kernel/sched.c routine partition_sched_domains() in a
559 * convenient format, that can be easily compared to the prior
560 * value to determine what partition elements (sched domains)
561 * were changed (added or removed.)
563 * Finding the best partition (set of domains):
564 * The triple nested loops below over i, j, k scan over the
565 * load balanced cpusets (using the array of cpuset pointers in
566 * csa[]) looking for pairs of cpusets that have overlapping
567 * cpus_allowed, but which don't have the same 'pn' partition
568 * number and gives them in the same partition number. It keeps
569 * looping on the 'restart' label until it can no longer find
572 * The union of the cpus_allowed masks from the set of
573 * all cpusets having the same 'pn' value then form the one
574 * element of the partition (one sched domain) to be passed to
575 * partition_sched_domains().
577 static int generate_sched_domains(cpumask_t
**domains
,
578 struct sched_domain_attr
**attributes
)
580 LIST_HEAD(q
); /* queue of cpusets to be scanned */
581 struct cpuset
*cp
; /* scans q */
582 struct cpuset
**csa
; /* array of all cpuset ptrs */
583 int csn
; /* how many cpuset ptrs in csa so far */
584 int i
, j
, k
; /* indices for partition finding loops */
585 cpumask_t
*doms
; /* resulting partition; i.e. sched domains */
586 struct sched_domain_attr
*dattr
; /* attributes for custom domains */
587 int ndoms
; /* number of sched domains in result */
588 int nslot
; /* next empty doms[] cpumask_t slot */
594 /* Special case for the 99% of systems with one, full, sched domain */
595 if (is_sched_load_balance(&top_cpuset
)) {
596 doms
= kmalloc(sizeof(cpumask_t
), GFP_KERNEL
);
600 dattr
= kmalloc(sizeof(struct sched_domain_attr
), GFP_KERNEL
);
602 *dattr
= SD_ATTR_INIT
;
603 update_domain_attr_tree(dattr
, &top_cpuset
);
605 *doms
= top_cpuset
.cpus_allowed
;
611 csa
= kmalloc(number_of_cpusets
* sizeof(cp
), GFP_KERNEL
);
616 list_add(&top_cpuset
.stack_list
, &q
);
617 while (!list_empty(&q
)) {
619 struct cpuset
*child
; /* scans child cpusets of cp */
621 cp
= list_first_entry(&q
, struct cpuset
, stack_list
);
624 if (cpus_empty(cp
->cpus_allowed
))
628 * All child cpusets contain a subset of the parent's cpus, so
629 * just skip them, and then we call update_domain_attr_tree()
630 * to calc relax_domain_level of the corresponding sched
633 if (is_sched_load_balance(cp
)) {
638 list_for_each_entry(cont
, &cp
->css
.cgroup
->children
, sibling
) {
639 child
= cgroup_cs(cont
);
640 list_add_tail(&child
->stack_list
, &q
);
644 for (i
= 0; i
< csn
; i
++)
649 /* Find the best partition (set of sched domains) */
650 for (i
= 0; i
< csn
; i
++) {
651 struct cpuset
*a
= csa
[i
];
654 for (j
= 0; j
< csn
; j
++) {
655 struct cpuset
*b
= csa
[j
];
658 if (apn
!= bpn
&& cpusets_overlap(a
, b
)) {
659 for (k
= 0; k
< csn
; k
++) {
660 struct cpuset
*c
= csa
[k
];
665 ndoms
--; /* one less element */
672 * Now we know how many domains to create.
673 * Convert <csn, csa> to <ndoms, doms> and populate cpu masks.
675 doms
= kmalloc(ndoms
* sizeof(cpumask_t
), GFP_KERNEL
);
680 * The rest of the code, including the scheduler, can deal with
681 * dattr==NULL case. No need to abort if alloc fails.
683 dattr
= kmalloc(ndoms
* sizeof(struct sched_domain_attr
), GFP_KERNEL
);
685 for (nslot
= 0, i
= 0; i
< csn
; i
++) {
686 struct cpuset
*a
= csa
[i
];
691 /* Skip completed partitions */
697 if (nslot
== ndoms
) {
698 static int warnings
= 10;
701 "rebuild_sched_domains confused:"
702 " nslot %d, ndoms %d, csn %d, i %d,"
704 nslot
, ndoms
, csn
, i
, apn
);
712 *(dattr
+ nslot
) = SD_ATTR_INIT
;
713 for (j
= i
; j
< csn
; j
++) {
714 struct cpuset
*b
= csa
[j
];
717 cpus_or(*dp
, *dp
, b
->cpus_allowed
);
719 update_domain_attr_tree(dattr
+ nslot
, b
);
721 /* Done with this partition */
727 BUG_ON(nslot
!= ndoms
);
733 * Fallback to the default domain if kmalloc() failed.
734 * See comments in partition_sched_domains().
745 * Rebuild scheduler domains.
747 * Call with neither cgroup_mutex held nor within get_online_cpus().
748 * Takes both cgroup_mutex and get_online_cpus().
750 * Cannot be directly called from cpuset code handling changes
751 * to the cpuset pseudo-filesystem, because it cannot be called
752 * from code that already holds cgroup_mutex.
754 static void do_rebuild_sched_domains(struct work_struct
*unused
)
756 struct sched_domain_attr
*attr
;
762 /* Generate domain masks and attrs */
764 ndoms
= generate_sched_domains(&doms
, &attr
);
767 /* Have scheduler rebuild the domains */
768 partition_sched_domains(ndoms
, doms
, attr
);
773 static DECLARE_WORK(rebuild_sched_domains_work
, do_rebuild_sched_domains
);
776 * Rebuild scheduler domains, asynchronously via workqueue.
778 * If the flag 'sched_load_balance' of any cpuset with non-empty
779 * 'cpus' changes, or if the 'cpus' allowed changes in any cpuset
780 * which has that flag enabled, or if any cpuset with a non-empty
781 * 'cpus' is removed, then call this routine to rebuild the
782 * scheduler's dynamic sched domains.
784 * The rebuild_sched_domains() and partition_sched_domains()
785 * routines must nest cgroup_lock() inside get_online_cpus(),
786 * but such cpuset changes as these must nest that locking the
787 * other way, holding cgroup_lock() for much of the code.
789 * So in order to avoid an ABBA deadlock, the cpuset code handling
790 * these user changes delegates the actual sched domain rebuilding
791 * to a separate workqueue thread, which ends up processing the
792 * above do_rebuild_sched_domains() function.
794 static void async_rebuild_sched_domains(void)
796 schedule_work(&rebuild_sched_domains_work
);
800 * Accomplishes the same scheduler domain rebuild as the above
801 * async_rebuild_sched_domains(), however it directly calls the
802 * rebuild routine synchronously rather than calling it via an
803 * asynchronous work thread.
805 * This can only be called from code that is not holding
806 * cgroup_mutex (not nested in a cgroup_lock() call.)
808 void rebuild_sched_domains(void)
810 do_rebuild_sched_domains(NULL
);
814 * cpuset_test_cpumask - test a task's cpus_allowed versus its cpuset's
816 * @scan: struct cgroup_scanner contained in its struct cpuset_hotplug_scanner
818 * Call with cgroup_mutex held. May take callback_mutex during call.
819 * Called for each task in a cgroup by cgroup_scan_tasks().
820 * Return nonzero if this tasks's cpus_allowed mask should be changed (in other
821 * words, if its mask is not equal to its cpuset's mask).
823 static int cpuset_test_cpumask(struct task_struct
*tsk
,
824 struct cgroup_scanner
*scan
)
826 return !cpus_equal(tsk
->cpus_allowed
,
827 (cgroup_cs(scan
->cg
))->cpus_allowed
);
831 * cpuset_change_cpumask - make a task's cpus_allowed the same as its cpuset's
833 * @scan: struct cgroup_scanner containing the cgroup of the task
835 * Called by cgroup_scan_tasks() for each task in a cgroup whose
836 * cpus_allowed mask needs to be changed.
838 * We don't need to re-check for the cgroup/cpuset membership, since we're
839 * holding cgroup_lock() at this point.
841 static void cpuset_change_cpumask(struct task_struct
*tsk
,
842 struct cgroup_scanner
*scan
)
844 set_cpus_allowed_ptr(tsk
, &((cgroup_cs(scan
->cg
))->cpus_allowed
));
848 * update_tasks_cpumask - Update the cpumasks of tasks in the cpuset.
849 * @cs: the cpuset in which each task's cpus_allowed mask needs to be changed
850 * @heap: if NULL, defer allocating heap memory to cgroup_scan_tasks()
852 * Called with cgroup_mutex held
854 * The cgroup_scan_tasks() function will scan all the tasks in a cgroup,
855 * calling callback functions for each.
857 * No return value. It's guaranteed that cgroup_scan_tasks() always returns 0
860 static void update_tasks_cpumask(struct cpuset
*cs
, struct ptr_heap
*heap
)
862 struct cgroup_scanner scan
;
864 scan
.cg
= cs
->css
.cgroup
;
865 scan
.test_task
= cpuset_test_cpumask
;
866 scan
.process_task
= cpuset_change_cpumask
;
868 cgroup_scan_tasks(&scan
);
872 * update_cpumask - update the cpus_allowed mask of a cpuset and all tasks in it
873 * @cs: the cpuset to consider
874 * @buf: buffer of cpu numbers written to this cpuset
876 static int update_cpumask(struct cpuset
*cs
, const char *buf
)
878 struct ptr_heap heap
;
879 struct cpuset trialcs
;
881 int is_load_balanced
;
883 /* top_cpuset.cpus_allowed tracks cpu_online_map; it's read-only */
884 if (cs
== &top_cpuset
)
890 * An empty cpus_allowed is ok only if the cpuset has no tasks.
891 * Since cpulist_parse() fails on an empty mask, we special case
892 * that parsing. The validate_change() call ensures that cpusets
893 * with tasks have cpus.
896 cpus_clear(trialcs
.cpus_allowed
);
898 retval
= cpulist_parse(buf
, trialcs
.cpus_allowed
);
902 if (!cpus_subset(trialcs
.cpus_allowed
, cpu_online_map
))
905 retval
= validate_change(cs
, &trialcs
);
909 /* Nothing to do if the cpus didn't change */
910 if (cpus_equal(cs
->cpus_allowed
, trialcs
.cpus_allowed
))
913 retval
= heap_init(&heap
, PAGE_SIZE
, GFP_KERNEL
, NULL
);
917 is_load_balanced
= is_sched_load_balance(&trialcs
);
919 mutex_lock(&callback_mutex
);
920 cs
->cpus_allowed
= trialcs
.cpus_allowed
;
921 mutex_unlock(&callback_mutex
);
924 * Scan tasks in the cpuset, and update the cpumasks of any
925 * that need an update.
927 update_tasks_cpumask(cs
, &heap
);
931 if (is_load_balanced
)
932 async_rebuild_sched_domains();
939 * Migrate memory region from one set of nodes to another.
941 * Temporarilly set tasks mems_allowed to target nodes of migration,
942 * so that the migration code can allocate pages on these nodes.
944 * Call holding cgroup_mutex, so current's cpuset won't change
945 * during this call, as manage_mutex holds off any cpuset_attach()
946 * calls. Therefore we don't need to take task_lock around the
947 * call to guarantee_online_mems(), as we know no one is changing
950 * Hold callback_mutex around the two modifications of our tasks
951 * mems_allowed to synchronize with cpuset_mems_allowed().
953 * While the mm_struct we are migrating is typically from some
954 * other task, the task_struct mems_allowed that we are hacking
955 * is for our current task, which must allocate new pages for that
956 * migrating memory region.
958 * We call cpuset_update_task_memory_state() before hacking
959 * our tasks mems_allowed, so that we are assured of being in
960 * sync with our tasks cpuset, and in particular, callbacks to
961 * cpuset_update_task_memory_state() from nested page allocations
962 * won't see any mismatch of our cpuset and task mems_generation
963 * values, so won't overwrite our hacked tasks mems_allowed
967 static void cpuset_migrate_mm(struct mm_struct
*mm
, const nodemask_t
*from
,
968 const nodemask_t
*to
)
970 struct task_struct
*tsk
= current
;
972 cpuset_update_task_memory_state();
974 mutex_lock(&callback_mutex
);
975 tsk
->mems_allowed
= *to
;
976 mutex_unlock(&callback_mutex
);
978 do_migrate_pages(mm
, from
, to
, MPOL_MF_MOVE_ALL
);
980 mutex_lock(&callback_mutex
);
981 guarantee_online_mems(task_cs(tsk
),&tsk
->mems_allowed
);
982 mutex_unlock(&callback_mutex
);
985 static void *cpuset_being_rebound
;
988 * update_tasks_nodemask - Update the nodemasks of tasks in the cpuset.
989 * @cs: the cpuset in which each task's mems_allowed mask needs to be changed
990 * @oldmem: old mems_allowed of cpuset cs
992 * Called with cgroup_mutex held
993 * Return 0 if successful, -errno if not.
995 static int update_tasks_nodemask(struct cpuset
*cs
, const nodemask_t
*oldmem
)
997 struct task_struct
*p
;
998 struct mm_struct
**mmarray
;
1002 struct cgroup_iter it
;
1005 cpuset_being_rebound
= cs
; /* causes mpol_dup() rebind */
1007 fudge
= 10; /* spare mmarray[] slots */
1008 fudge
+= cpus_weight(cs
->cpus_allowed
); /* imagine one fork-bomb/cpu */
1012 * Allocate mmarray[] to hold mm reference for each task
1013 * in cpuset cs. Can't kmalloc GFP_KERNEL while holding
1014 * tasklist_lock. We could use GFP_ATOMIC, but with a
1015 * few more lines of code, we can retry until we get a big
1016 * enough mmarray[] w/o using GFP_ATOMIC.
1019 ntasks
= cgroup_task_count(cs
->css
.cgroup
); /* guess */
1021 mmarray
= kmalloc(ntasks
* sizeof(*mmarray
), GFP_KERNEL
);
1024 read_lock(&tasklist_lock
); /* block fork */
1025 if (cgroup_task_count(cs
->css
.cgroup
) <= ntasks
)
1026 break; /* got enough */
1027 read_unlock(&tasklist_lock
); /* try again */
1033 /* Load up mmarray[] with mm reference for each task in cpuset. */
1034 cgroup_iter_start(cs
->css
.cgroup
, &it
);
1035 while ((p
= cgroup_iter_next(cs
->css
.cgroup
, &it
))) {
1036 struct mm_struct
*mm
;
1040 "Cpuset mempolicy rebind incomplete.\n");
1043 mm
= get_task_mm(p
);
1048 cgroup_iter_end(cs
->css
.cgroup
, &it
);
1049 read_unlock(&tasklist_lock
);
1052 * Now that we've dropped the tasklist spinlock, we can
1053 * rebind the vma mempolicies of each mm in mmarray[] to their
1054 * new cpuset, and release that mm. The mpol_rebind_mm()
1055 * call takes mmap_sem, which we couldn't take while holding
1056 * tasklist_lock. Forks can happen again now - the mpol_dup()
1057 * cpuset_being_rebound check will catch such forks, and rebind
1058 * their vma mempolicies too. Because we still hold the global
1059 * cgroup_mutex, we know that no other rebind effort will
1060 * be contending for the global variable cpuset_being_rebound.
1061 * It's ok if we rebind the same mm twice; mpol_rebind_mm()
1062 * is idempotent. Also migrate pages in each mm to new nodes.
1064 migrate
= is_memory_migrate(cs
);
1065 for (i
= 0; i
< n
; i
++) {
1066 struct mm_struct
*mm
= mmarray
[i
];
1068 mpol_rebind_mm(mm
, &cs
->mems_allowed
);
1070 cpuset_migrate_mm(mm
, oldmem
, &cs
->mems_allowed
);
1074 /* We're done rebinding vmas to this cpuset's new mems_allowed. */
1076 cpuset_being_rebound
= NULL
;
1083 * Handle user request to change the 'mems' memory placement
1084 * of a cpuset. Needs to validate the request, update the
1085 * cpusets mems_allowed and mems_generation, and for each
1086 * task in the cpuset, rebind any vma mempolicies and if
1087 * the cpuset is marked 'memory_migrate', migrate the tasks
1088 * pages to the new memory.
1090 * Call with cgroup_mutex held. May take callback_mutex during call.
1091 * Will take tasklist_lock, scan tasklist for tasks in cpuset cs,
1092 * lock each such tasks mm->mmap_sem, scan its vma's and rebind
1093 * their mempolicies to the cpusets new mems_allowed.
1095 static int update_nodemask(struct cpuset
*cs
, const char *buf
)
1097 struct cpuset trialcs
;
1102 * top_cpuset.mems_allowed tracks node_stats[N_HIGH_MEMORY];
1105 if (cs
== &top_cpuset
)
1111 * An empty mems_allowed is ok iff there are no tasks in the cpuset.
1112 * Since nodelist_parse() fails on an empty mask, we special case
1113 * that parsing. The validate_change() call ensures that cpusets
1114 * with tasks have memory.
1117 nodes_clear(trialcs
.mems_allowed
);
1119 retval
= nodelist_parse(buf
, trialcs
.mems_allowed
);
1123 if (!nodes_subset(trialcs
.mems_allowed
,
1124 node_states
[N_HIGH_MEMORY
]))
1127 oldmem
= cs
->mems_allowed
;
1128 if (nodes_equal(oldmem
, trialcs
.mems_allowed
)) {
1129 retval
= 0; /* Too easy - nothing to do */
1132 retval
= validate_change(cs
, &trialcs
);
1136 mutex_lock(&callback_mutex
);
1137 cs
->mems_allowed
= trialcs
.mems_allowed
;
1138 cs
->mems_generation
= cpuset_mems_generation
++;
1139 mutex_unlock(&callback_mutex
);
1141 retval
= update_tasks_nodemask(cs
, &oldmem
);
1146 int current_cpuset_is_being_rebound(void)
1148 return task_cs(current
) == cpuset_being_rebound
;
1151 static int update_relax_domain_level(struct cpuset
*cs
, s64 val
)
1153 if (val
< -1 || val
>= SD_LV_MAX
)
1156 if (val
!= cs
->relax_domain_level
) {
1157 cs
->relax_domain_level
= val
;
1158 if (!cpus_empty(cs
->cpus_allowed
) && is_sched_load_balance(cs
))
1159 async_rebuild_sched_domains();
1166 * update_flag - read a 0 or a 1 in a file and update associated flag
1167 * bit: the bit to update (see cpuset_flagbits_t)
1168 * cs: the cpuset to update
1169 * turning_on: whether the flag is being set or cleared
1171 * Call with cgroup_mutex held.
1174 static int update_flag(cpuset_flagbits_t bit
, struct cpuset
*cs
,
1177 struct cpuset trialcs
;
1179 int cpus_nonempty
, balance_flag_changed
;
1183 set_bit(bit
, &trialcs
.flags
);
1185 clear_bit(bit
, &trialcs
.flags
);
1187 err
= validate_change(cs
, &trialcs
);
1191 cpus_nonempty
= !cpus_empty(trialcs
.cpus_allowed
);
1192 balance_flag_changed
= (is_sched_load_balance(cs
) !=
1193 is_sched_load_balance(&trialcs
));
1195 mutex_lock(&callback_mutex
);
1196 cs
->flags
= trialcs
.flags
;
1197 mutex_unlock(&callback_mutex
);
1199 if (cpus_nonempty
&& balance_flag_changed
)
1200 async_rebuild_sched_domains();
1206 * Frequency meter - How fast is some event occurring?
1208 * These routines manage a digitally filtered, constant time based,
1209 * event frequency meter. There are four routines:
1210 * fmeter_init() - initialize a frequency meter.
1211 * fmeter_markevent() - called each time the event happens.
1212 * fmeter_getrate() - returns the recent rate of such events.
1213 * fmeter_update() - internal routine used to update fmeter.
1215 * A common data structure is passed to each of these routines,
1216 * which is used to keep track of the state required to manage the
1217 * frequency meter and its digital filter.
1219 * The filter works on the number of events marked per unit time.
1220 * The filter is single-pole low-pass recursive (IIR). The time unit
1221 * is 1 second. Arithmetic is done using 32-bit integers scaled to
1222 * simulate 3 decimal digits of precision (multiplied by 1000).
1224 * With an FM_COEF of 933, and a time base of 1 second, the filter
1225 * has a half-life of 10 seconds, meaning that if the events quit
1226 * happening, then the rate returned from the fmeter_getrate()
1227 * will be cut in half each 10 seconds, until it converges to zero.
1229 * It is not worth doing a real infinitely recursive filter. If more
1230 * than FM_MAXTICKS ticks have elapsed since the last filter event,
1231 * just compute FM_MAXTICKS ticks worth, by which point the level
1234 * Limit the count of unprocessed events to FM_MAXCNT, so as to avoid
1235 * arithmetic overflow in the fmeter_update() routine.
1237 * Given the simple 32 bit integer arithmetic used, this meter works
1238 * best for reporting rates between one per millisecond (msec) and
1239 * one per 32 (approx) seconds. At constant rates faster than one
1240 * per msec it maxes out at values just under 1,000,000. At constant
1241 * rates between one per msec, and one per second it will stabilize
1242 * to a value N*1000, where N is the rate of events per second.
1243 * At constant rates between one per second and one per 32 seconds,
1244 * it will be choppy, moving up on the seconds that have an event,
1245 * and then decaying until the next event. At rates slower than
1246 * about one in 32 seconds, it decays all the way back to zero between
1250 #define FM_COEF 933 /* coefficient for half-life of 10 secs */
1251 #define FM_MAXTICKS ((time_t)99) /* useless computing more ticks than this */
1252 #define FM_MAXCNT 1000000 /* limit cnt to avoid overflow */
1253 #define FM_SCALE 1000 /* faux fixed point scale */
1255 /* Initialize a frequency meter */
1256 static void fmeter_init(struct fmeter
*fmp
)
1261 spin_lock_init(&fmp
->lock
);
1264 /* Internal meter update - process cnt events and update value */
1265 static void fmeter_update(struct fmeter
*fmp
)
1267 time_t now
= get_seconds();
1268 time_t ticks
= now
- fmp
->time
;
1273 ticks
= min(FM_MAXTICKS
, ticks
);
1275 fmp
->val
= (FM_COEF
* fmp
->val
) / FM_SCALE
;
1278 fmp
->val
+= ((FM_SCALE
- FM_COEF
) * fmp
->cnt
) / FM_SCALE
;
1282 /* Process any previous ticks, then bump cnt by one (times scale). */
1283 static void fmeter_markevent(struct fmeter
*fmp
)
1285 spin_lock(&fmp
->lock
);
1287 fmp
->cnt
= min(FM_MAXCNT
, fmp
->cnt
+ FM_SCALE
);
1288 spin_unlock(&fmp
->lock
);
1291 /* Process any previous ticks, then return current value. */
1292 static int fmeter_getrate(struct fmeter
*fmp
)
1296 spin_lock(&fmp
->lock
);
1299 spin_unlock(&fmp
->lock
);
1303 /* Called by cgroups to determine if a cpuset is usable; cgroup_mutex held */
1304 static int cpuset_can_attach(struct cgroup_subsys
*ss
,
1305 struct cgroup
*cont
, struct task_struct
*tsk
)
1307 struct cpuset
*cs
= cgroup_cs(cont
);
1309 if (cpus_empty(cs
->cpus_allowed
) || nodes_empty(cs
->mems_allowed
))
1311 if (tsk
->flags
& PF_THREAD_BOUND
) {
1314 mutex_lock(&callback_mutex
);
1315 mask
= cs
->cpus_allowed
;
1316 mutex_unlock(&callback_mutex
);
1317 if (!cpus_equal(tsk
->cpus_allowed
, mask
))
1321 return security_task_setscheduler(tsk
, 0, NULL
);
1324 static void cpuset_attach(struct cgroup_subsys
*ss
,
1325 struct cgroup
*cont
, struct cgroup
*oldcont
,
1326 struct task_struct
*tsk
)
1329 nodemask_t from
, to
;
1330 struct mm_struct
*mm
;
1331 struct cpuset
*cs
= cgroup_cs(cont
);
1332 struct cpuset
*oldcs
= cgroup_cs(oldcont
);
1335 mutex_lock(&callback_mutex
);
1336 guarantee_online_cpus(cs
, &cpus
);
1337 err
= set_cpus_allowed_ptr(tsk
, &cpus
);
1338 mutex_unlock(&callback_mutex
);
1342 from
= oldcs
->mems_allowed
;
1343 to
= cs
->mems_allowed
;
1344 mm
= get_task_mm(tsk
);
1346 mpol_rebind_mm(mm
, &to
);
1347 if (is_memory_migrate(cs
))
1348 cpuset_migrate_mm(mm
, &from
, &to
);
1354 /* The various types of files and directories in a cpuset file system */
1357 FILE_MEMORY_MIGRATE
,
1363 FILE_SCHED_LOAD_BALANCE
,
1364 FILE_SCHED_RELAX_DOMAIN_LEVEL
,
1365 FILE_MEMORY_PRESSURE_ENABLED
,
1366 FILE_MEMORY_PRESSURE
,
1369 } cpuset_filetype_t
;
1371 static int cpuset_write_u64(struct cgroup
*cgrp
, struct cftype
*cft
, u64 val
)
1374 struct cpuset
*cs
= cgroup_cs(cgrp
);
1375 cpuset_filetype_t type
= cft
->private;
1377 if (!cgroup_lock_live_group(cgrp
))
1381 case FILE_CPU_EXCLUSIVE
:
1382 retval
= update_flag(CS_CPU_EXCLUSIVE
, cs
, val
);
1384 case FILE_MEM_EXCLUSIVE
:
1385 retval
= update_flag(CS_MEM_EXCLUSIVE
, cs
, val
);
1387 case FILE_MEM_HARDWALL
:
1388 retval
= update_flag(CS_MEM_HARDWALL
, cs
, val
);
1390 case FILE_SCHED_LOAD_BALANCE
:
1391 retval
= update_flag(CS_SCHED_LOAD_BALANCE
, cs
, val
);
1393 case FILE_MEMORY_MIGRATE
:
1394 retval
= update_flag(CS_MEMORY_MIGRATE
, cs
, val
);
1396 case FILE_MEMORY_PRESSURE_ENABLED
:
1397 cpuset_memory_pressure_enabled
= !!val
;
1399 case FILE_MEMORY_PRESSURE
:
1402 case FILE_SPREAD_PAGE
:
1403 retval
= update_flag(CS_SPREAD_PAGE
, cs
, val
);
1404 cs
->mems_generation
= cpuset_mems_generation
++;
1406 case FILE_SPREAD_SLAB
:
1407 retval
= update_flag(CS_SPREAD_SLAB
, cs
, val
);
1408 cs
->mems_generation
= cpuset_mems_generation
++;
1418 static int cpuset_write_s64(struct cgroup
*cgrp
, struct cftype
*cft
, s64 val
)
1421 struct cpuset
*cs
= cgroup_cs(cgrp
);
1422 cpuset_filetype_t type
= cft
->private;
1424 if (!cgroup_lock_live_group(cgrp
))
1428 case FILE_SCHED_RELAX_DOMAIN_LEVEL
:
1429 retval
= update_relax_domain_level(cs
, val
);
1440 * Common handling for a write to a "cpus" or "mems" file.
1442 static int cpuset_write_resmask(struct cgroup
*cgrp
, struct cftype
*cft
,
1447 if (!cgroup_lock_live_group(cgrp
))
1450 switch (cft
->private) {
1452 retval
= update_cpumask(cgroup_cs(cgrp
), buf
);
1455 retval
= update_nodemask(cgroup_cs(cgrp
), buf
);
1466 * These ascii lists should be read in a single call, by using a user
1467 * buffer large enough to hold the entire map. If read in smaller
1468 * chunks, there is no guarantee of atomicity. Since the display format
1469 * used, list of ranges of sequential numbers, is variable length,
1470 * and since these maps can change value dynamically, one could read
1471 * gibberish by doing partial reads while a list was changing.
1472 * A single large read to a buffer that crosses a page boundary is
1473 * ok, because the result being copied to user land is not recomputed
1474 * across a page fault.
1477 static int cpuset_sprintf_cpulist(char *page
, struct cpuset
*cs
)
1481 mutex_lock(&callback_mutex
);
1482 mask
= cs
->cpus_allowed
;
1483 mutex_unlock(&callback_mutex
);
1485 return cpulist_scnprintf(page
, PAGE_SIZE
, mask
);
1488 static int cpuset_sprintf_memlist(char *page
, struct cpuset
*cs
)
1492 mutex_lock(&callback_mutex
);
1493 mask
= cs
->mems_allowed
;
1494 mutex_unlock(&callback_mutex
);
1496 return nodelist_scnprintf(page
, PAGE_SIZE
, mask
);
1499 static ssize_t
cpuset_common_file_read(struct cgroup
*cont
,
1503 size_t nbytes
, loff_t
*ppos
)
1505 struct cpuset
*cs
= cgroup_cs(cont
);
1506 cpuset_filetype_t type
= cft
->private;
1511 if (!(page
= (char *)__get_free_page(GFP_TEMPORARY
)))
1518 s
+= cpuset_sprintf_cpulist(s
, cs
);
1521 s
+= cpuset_sprintf_memlist(s
, cs
);
1529 retval
= simple_read_from_buffer(buf
, nbytes
, ppos
, page
, s
- page
);
1531 free_page((unsigned long)page
);
1535 static u64
cpuset_read_u64(struct cgroup
*cont
, struct cftype
*cft
)
1537 struct cpuset
*cs
= cgroup_cs(cont
);
1538 cpuset_filetype_t type
= cft
->private;
1540 case FILE_CPU_EXCLUSIVE
:
1541 return is_cpu_exclusive(cs
);
1542 case FILE_MEM_EXCLUSIVE
:
1543 return is_mem_exclusive(cs
);
1544 case FILE_MEM_HARDWALL
:
1545 return is_mem_hardwall(cs
);
1546 case FILE_SCHED_LOAD_BALANCE
:
1547 return is_sched_load_balance(cs
);
1548 case FILE_MEMORY_MIGRATE
:
1549 return is_memory_migrate(cs
);
1550 case FILE_MEMORY_PRESSURE_ENABLED
:
1551 return cpuset_memory_pressure_enabled
;
1552 case FILE_MEMORY_PRESSURE
:
1553 return fmeter_getrate(&cs
->fmeter
);
1554 case FILE_SPREAD_PAGE
:
1555 return is_spread_page(cs
);
1556 case FILE_SPREAD_SLAB
:
1557 return is_spread_slab(cs
);
1562 /* Unreachable but makes gcc happy */
1566 static s64
cpuset_read_s64(struct cgroup
*cont
, struct cftype
*cft
)
1568 struct cpuset
*cs
= cgroup_cs(cont
);
1569 cpuset_filetype_t type
= cft
->private;
1571 case FILE_SCHED_RELAX_DOMAIN_LEVEL
:
1572 return cs
->relax_domain_level
;
1577 /* Unrechable but makes gcc happy */
1583 * for the common functions, 'private' gives the type of file
1586 static struct cftype files
[] = {
1589 .read
= cpuset_common_file_read
,
1590 .write_string
= cpuset_write_resmask
,
1591 .max_write_len
= (100U + 6 * NR_CPUS
),
1592 .private = FILE_CPULIST
,
1597 .read
= cpuset_common_file_read
,
1598 .write_string
= cpuset_write_resmask
,
1599 .max_write_len
= (100U + 6 * MAX_NUMNODES
),
1600 .private = FILE_MEMLIST
,
1604 .name
= "cpu_exclusive",
1605 .read_u64
= cpuset_read_u64
,
1606 .write_u64
= cpuset_write_u64
,
1607 .private = FILE_CPU_EXCLUSIVE
,
1611 .name
= "mem_exclusive",
1612 .read_u64
= cpuset_read_u64
,
1613 .write_u64
= cpuset_write_u64
,
1614 .private = FILE_MEM_EXCLUSIVE
,
1618 .name
= "mem_hardwall",
1619 .read_u64
= cpuset_read_u64
,
1620 .write_u64
= cpuset_write_u64
,
1621 .private = FILE_MEM_HARDWALL
,
1625 .name
= "sched_load_balance",
1626 .read_u64
= cpuset_read_u64
,
1627 .write_u64
= cpuset_write_u64
,
1628 .private = FILE_SCHED_LOAD_BALANCE
,
1632 .name
= "sched_relax_domain_level",
1633 .read_s64
= cpuset_read_s64
,
1634 .write_s64
= cpuset_write_s64
,
1635 .private = FILE_SCHED_RELAX_DOMAIN_LEVEL
,
1639 .name
= "memory_migrate",
1640 .read_u64
= cpuset_read_u64
,
1641 .write_u64
= cpuset_write_u64
,
1642 .private = FILE_MEMORY_MIGRATE
,
1646 .name
= "memory_pressure",
1647 .read_u64
= cpuset_read_u64
,
1648 .write_u64
= cpuset_write_u64
,
1649 .private = FILE_MEMORY_PRESSURE
,
1653 .name
= "memory_spread_page",
1654 .read_u64
= cpuset_read_u64
,
1655 .write_u64
= cpuset_write_u64
,
1656 .private = FILE_SPREAD_PAGE
,
1660 .name
= "memory_spread_slab",
1661 .read_u64
= cpuset_read_u64
,
1662 .write_u64
= cpuset_write_u64
,
1663 .private = FILE_SPREAD_SLAB
,
1667 static struct cftype cft_memory_pressure_enabled
= {
1668 .name
= "memory_pressure_enabled",
1669 .read_u64
= cpuset_read_u64
,
1670 .write_u64
= cpuset_write_u64
,
1671 .private = FILE_MEMORY_PRESSURE_ENABLED
,
1674 static int cpuset_populate(struct cgroup_subsys
*ss
, struct cgroup
*cont
)
1678 err
= cgroup_add_files(cont
, ss
, files
, ARRAY_SIZE(files
));
1681 /* memory_pressure_enabled is in root cpuset only */
1683 err
= cgroup_add_file(cont
, ss
,
1684 &cft_memory_pressure_enabled
);
1689 * post_clone() is called at the end of cgroup_clone().
1690 * 'cgroup' was just created automatically as a result of
1691 * a cgroup_clone(), and the current task is about to
1692 * be moved into 'cgroup'.
1694 * Currently we refuse to set up the cgroup - thereby
1695 * refusing the task to be entered, and as a result refusing
1696 * the sys_unshare() or clone() which initiated it - if any
1697 * sibling cpusets have exclusive cpus or mem.
1699 * If this becomes a problem for some users who wish to
1700 * allow that scenario, then cpuset_post_clone() could be
1701 * changed to grant parent->cpus_allowed-sibling_cpus_exclusive
1702 * (and likewise for mems) to the new cgroup. Called with cgroup_mutex
1705 static void cpuset_post_clone(struct cgroup_subsys
*ss
,
1706 struct cgroup
*cgroup
)
1708 struct cgroup
*parent
, *child
;
1709 struct cpuset
*cs
, *parent_cs
;
1711 parent
= cgroup
->parent
;
1712 list_for_each_entry(child
, &parent
->children
, sibling
) {
1713 cs
= cgroup_cs(child
);
1714 if (is_mem_exclusive(cs
) || is_cpu_exclusive(cs
))
1717 cs
= cgroup_cs(cgroup
);
1718 parent_cs
= cgroup_cs(parent
);
1720 cs
->mems_allowed
= parent_cs
->mems_allowed
;
1721 cs
->cpus_allowed
= parent_cs
->cpus_allowed
;
1726 * cpuset_create - create a cpuset
1727 * ss: cpuset cgroup subsystem
1728 * cont: control group that the new cpuset will be part of
1731 static struct cgroup_subsys_state
*cpuset_create(
1732 struct cgroup_subsys
*ss
,
1733 struct cgroup
*cont
)
1736 struct cpuset
*parent
;
1738 if (!cont
->parent
) {
1739 /* This is early initialization for the top cgroup */
1740 top_cpuset
.mems_generation
= cpuset_mems_generation
++;
1741 return &top_cpuset
.css
;
1743 parent
= cgroup_cs(cont
->parent
);
1744 cs
= kmalloc(sizeof(*cs
), GFP_KERNEL
);
1746 return ERR_PTR(-ENOMEM
);
1748 cpuset_update_task_memory_state();
1750 if (is_spread_page(parent
))
1751 set_bit(CS_SPREAD_PAGE
, &cs
->flags
);
1752 if (is_spread_slab(parent
))
1753 set_bit(CS_SPREAD_SLAB
, &cs
->flags
);
1754 set_bit(CS_SCHED_LOAD_BALANCE
, &cs
->flags
);
1755 cpus_clear(cs
->cpus_allowed
);
1756 nodes_clear(cs
->mems_allowed
);
1757 cs
->mems_generation
= cpuset_mems_generation
++;
1758 fmeter_init(&cs
->fmeter
);
1759 cs
->relax_domain_level
= -1;
1761 cs
->parent
= parent
;
1762 number_of_cpusets
++;
1767 * If the cpuset being removed has its flag 'sched_load_balance'
1768 * enabled, then simulate turning sched_load_balance off, which
1769 * will call async_rebuild_sched_domains().
1772 static void cpuset_destroy(struct cgroup_subsys
*ss
, struct cgroup
*cont
)
1774 struct cpuset
*cs
= cgroup_cs(cont
);
1776 cpuset_update_task_memory_state();
1778 if (is_sched_load_balance(cs
))
1779 update_flag(CS_SCHED_LOAD_BALANCE
, cs
, 0);
1781 number_of_cpusets
--;
1785 struct cgroup_subsys cpuset_subsys
= {
1787 .create
= cpuset_create
,
1788 .destroy
= cpuset_destroy
,
1789 .can_attach
= cpuset_can_attach
,
1790 .attach
= cpuset_attach
,
1791 .populate
= cpuset_populate
,
1792 .post_clone
= cpuset_post_clone
,
1793 .subsys_id
= cpuset_subsys_id
,
1798 * cpuset_init_early - just enough so that the calls to
1799 * cpuset_update_task_memory_state() in early init code
1803 int __init
cpuset_init_early(void)
1805 top_cpuset
.mems_generation
= cpuset_mems_generation
++;
1811 * cpuset_init - initialize cpusets at system boot
1813 * Description: Initialize top_cpuset and the cpuset internal file system,
1816 int __init
cpuset_init(void)
1820 cpus_setall(top_cpuset
.cpus_allowed
);
1821 nodes_setall(top_cpuset
.mems_allowed
);
1823 fmeter_init(&top_cpuset
.fmeter
);
1824 top_cpuset
.mems_generation
= cpuset_mems_generation
++;
1825 set_bit(CS_SCHED_LOAD_BALANCE
, &top_cpuset
.flags
);
1826 top_cpuset
.relax_domain_level
= -1;
1828 err
= register_filesystem(&cpuset_fs_type
);
1832 number_of_cpusets
= 1;
1837 * cpuset_do_move_task - move a given task to another cpuset
1838 * @tsk: pointer to task_struct the task to move
1839 * @scan: struct cgroup_scanner contained in its struct cpuset_hotplug_scanner
1841 * Called by cgroup_scan_tasks() for each task in a cgroup.
1842 * Return nonzero to stop the walk through the tasks.
1844 static void cpuset_do_move_task(struct task_struct
*tsk
,
1845 struct cgroup_scanner
*scan
)
1847 struct cpuset_hotplug_scanner
*chsp
;
1849 chsp
= container_of(scan
, struct cpuset_hotplug_scanner
, scan
);
1850 cgroup_attach_task(chsp
->to
, tsk
);
1854 * move_member_tasks_to_cpuset - move tasks from one cpuset to another
1855 * @from: cpuset in which the tasks currently reside
1856 * @to: cpuset to which the tasks will be moved
1858 * Called with cgroup_mutex held
1859 * callback_mutex must not be held, as cpuset_attach() will take it.
1861 * The cgroup_scan_tasks() function will scan all the tasks in a cgroup,
1862 * calling callback functions for each.
1864 static void move_member_tasks_to_cpuset(struct cpuset
*from
, struct cpuset
*to
)
1866 struct cpuset_hotplug_scanner scan
;
1868 scan
.scan
.cg
= from
->css
.cgroup
;
1869 scan
.scan
.test_task
= NULL
; /* select all tasks in cgroup */
1870 scan
.scan
.process_task
= cpuset_do_move_task
;
1871 scan
.scan
.heap
= NULL
;
1872 scan
.to
= to
->css
.cgroup
;
1874 if (cgroup_scan_tasks(&scan
.scan
))
1875 printk(KERN_ERR
"move_member_tasks_to_cpuset: "
1876 "cgroup_scan_tasks failed\n");
1880 * If CPU and/or memory hotplug handlers, below, unplug any CPUs
1881 * or memory nodes, we need to walk over the cpuset hierarchy,
1882 * removing that CPU or node from all cpusets. If this removes the
1883 * last CPU or node from a cpuset, then move the tasks in the empty
1884 * cpuset to its next-highest non-empty parent.
1886 * Called with cgroup_mutex held
1887 * callback_mutex must not be held, as cpuset_attach() will take it.
1889 static void remove_tasks_in_empty_cpuset(struct cpuset
*cs
)
1891 struct cpuset
*parent
;
1894 * The cgroup's css_sets list is in use if there are tasks
1895 * in the cpuset; the list is empty if there are none;
1896 * the cs->css.refcnt seems always 0.
1898 if (list_empty(&cs
->css
.cgroup
->css_sets
))
1902 * Find its next-highest non-empty parent, (top cpuset
1903 * has online cpus, so can't be empty).
1905 parent
= cs
->parent
;
1906 while (cpus_empty(parent
->cpus_allowed
) ||
1907 nodes_empty(parent
->mems_allowed
))
1908 parent
= parent
->parent
;
1910 move_member_tasks_to_cpuset(cs
, parent
);
1914 * Walk the specified cpuset subtree and look for empty cpusets.
1915 * The tasks of such cpuset must be moved to a parent cpuset.
1917 * Called with cgroup_mutex held. We take callback_mutex to modify
1918 * cpus_allowed and mems_allowed.
1920 * This walk processes the tree from top to bottom, completing one layer
1921 * before dropping down to the next. It always processes a node before
1922 * any of its children.
1924 * For now, since we lack memory hot unplug, we'll never see a cpuset
1925 * that has tasks along with an empty 'mems'. But if we did see such
1926 * a cpuset, we'd handle it just like we do if its 'cpus' was empty.
1928 static void scan_for_empty_cpusets(const struct cpuset
*root
)
1931 struct cpuset
*cp
; /* scans cpusets being updated */
1932 struct cpuset
*child
; /* scans child cpusets of cp */
1933 struct cgroup
*cont
;
1936 list_add_tail((struct list_head
*)&root
->stack_list
, &queue
);
1938 while (!list_empty(&queue
)) {
1939 cp
= list_first_entry(&queue
, struct cpuset
, stack_list
);
1940 list_del(queue
.next
);
1941 list_for_each_entry(cont
, &cp
->css
.cgroup
->children
, sibling
) {
1942 child
= cgroup_cs(cont
);
1943 list_add_tail(&child
->stack_list
, &queue
);
1946 /* Continue past cpusets with all cpus, mems online */
1947 if (cpus_subset(cp
->cpus_allowed
, cpu_online_map
) &&
1948 nodes_subset(cp
->mems_allowed
, node_states
[N_HIGH_MEMORY
]))
1951 oldmems
= cp
->mems_allowed
;
1953 /* Remove offline cpus and mems from this cpuset. */
1954 mutex_lock(&callback_mutex
);
1955 cpus_and(cp
->cpus_allowed
, cp
->cpus_allowed
, cpu_online_map
);
1956 nodes_and(cp
->mems_allowed
, cp
->mems_allowed
,
1957 node_states
[N_HIGH_MEMORY
]);
1958 mutex_unlock(&callback_mutex
);
1960 /* Move tasks from the empty cpuset to a parent */
1961 if (cpus_empty(cp
->cpus_allowed
) ||
1962 nodes_empty(cp
->mems_allowed
))
1963 remove_tasks_in_empty_cpuset(cp
);
1965 update_tasks_cpumask(cp
, NULL
);
1966 update_tasks_nodemask(cp
, &oldmems
);
1972 * The top_cpuset tracks what CPUs and Memory Nodes are online,
1973 * period. This is necessary in order to make cpusets transparent
1974 * (of no affect) on systems that are actively using CPU hotplug
1975 * but making no active use of cpusets.
1977 * This routine ensures that top_cpuset.cpus_allowed tracks
1978 * cpu_online_map on each CPU hotplug (cpuhp) event.
1980 * Called within get_online_cpus(). Needs to call cgroup_lock()
1981 * before calling generate_sched_domains().
1983 static int cpuset_track_online_cpus(struct notifier_block
*unused_nb
,
1984 unsigned long phase
, void *unused_cpu
)
1986 struct sched_domain_attr
*attr
;
1992 case CPU_ONLINE_FROZEN
:
1994 case CPU_DEAD_FROZEN
:
2002 top_cpuset
.cpus_allowed
= cpu_online_map
;
2003 scan_for_empty_cpusets(&top_cpuset
);
2004 ndoms
= generate_sched_domains(&doms
, &attr
);
2007 /* Have scheduler rebuild the domains */
2008 partition_sched_domains(ndoms
, doms
, attr
);
2013 #ifdef CONFIG_MEMORY_HOTPLUG
2015 * Keep top_cpuset.mems_allowed tracking node_states[N_HIGH_MEMORY].
2016 * Call this routine anytime after node_states[N_HIGH_MEMORY] changes.
2017 * See also the previous routine cpuset_track_online_cpus().
2019 void cpuset_track_online_nodes(void)
2022 top_cpuset
.mems_allowed
= node_states
[N_HIGH_MEMORY
];
2023 scan_for_empty_cpusets(&top_cpuset
);
2029 * cpuset_init_smp - initialize cpus_allowed
2031 * Description: Finish top cpuset after cpu, node maps are initialized
2034 void __init
cpuset_init_smp(void)
2036 top_cpuset
.cpus_allowed
= cpu_online_map
;
2037 top_cpuset
.mems_allowed
= node_states
[N_HIGH_MEMORY
];
2039 hotcpu_notifier(cpuset_track_online_cpus
, 0);
2043 * cpuset_cpus_allowed - return cpus_allowed mask from a tasks cpuset.
2044 * @tsk: pointer to task_struct from which to obtain cpuset->cpus_allowed.
2045 * @pmask: pointer to cpumask_t variable to receive cpus_allowed set.
2047 * Description: Returns the cpumask_t cpus_allowed of the cpuset
2048 * attached to the specified @tsk. Guaranteed to return some non-empty
2049 * subset of cpu_online_map, even if this means going outside the
2053 void cpuset_cpus_allowed(struct task_struct
*tsk
, cpumask_t
*pmask
)
2055 mutex_lock(&callback_mutex
);
2056 cpuset_cpus_allowed_locked(tsk
, pmask
);
2057 mutex_unlock(&callback_mutex
);
2061 * cpuset_cpus_allowed_locked - return cpus_allowed mask from a tasks cpuset.
2062 * Must be called with callback_mutex held.
2064 void cpuset_cpus_allowed_locked(struct task_struct
*tsk
, cpumask_t
*pmask
)
2067 guarantee_online_cpus(task_cs(tsk
), pmask
);
2071 void cpuset_init_current_mems_allowed(void)
2073 nodes_setall(current
->mems_allowed
);
2077 * cpuset_mems_allowed - return mems_allowed mask from a tasks cpuset.
2078 * @tsk: pointer to task_struct from which to obtain cpuset->mems_allowed.
2080 * Description: Returns the nodemask_t mems_allowed of the cpuset
2081 * attached to the specified @tsk. Guaranteed to return some non-empty
2082 * subset of node_states[N_HIGH_MEMORY], even if this means going outside the
2086 nodemask_t
cpuset_mems_allowed(struct task_struct
*tsk
)
2090 mutex_lock(&callback_mutex
);
2092 guarantee_online_mems(task_cs(tsk
), &mask
);
2094 mutex_unlock(&callback_mutex
);
2100 * cpuset_nodemask_valid_mems_allowed - check nodemask vs. curremt mems_allowed
2101 * @nodemask: the nodemask to be checked
2103 * Are any of the nodes in the nodemask allowed in current->mems_allowed?
2105 int cpuset_nodemask_valid_mems_allowed(nodemask_t
*nodemask
)
2107 return nodes_intersects(*nodemask
, current
->mems_allowed
);
2111 * nearest_hardwall_ancestor() - Returns the nearest mem_exclusive or
2112 * mem_hardwall ancestor to the specified cpuset. Call holding
2113 * callback_mutex. If no ancestor is mem_exclusive or mem_hardwall
2114 * (an unusual configuration), then returns the root cpuset.
2116 static const struct cpuset
*nearest_hardwall_ancestor(const struct cpuset
*cs
)
2118 while (!(is_mem_exclusive(cs
) || is_mem_hardwall(cs
)) && cs
->parent
)
2124 * cpuset_zone_allowed_softwall - Can we allocate on zone z's memory node?
2125 * @z: is this zone on an allowed node?
2126 * @gfp_mask: memory allocation flags
2128 * If we're in interrupt, yes, we can always allocate. If
2129 * __GFP_THISNODE is set, yes, we can always allocate. If zone
2130 * z's node is in our tasks mems_allowed, yes. If it's not a
2131 * __GFP_HARDWALL request and this zone's nodes is in the nearest
2132 * hardwalled cpuset ancestor to this tasks cpuset, yes.
2133 * If the task has been OOM killed and has access to memory reserves
2134 * as specified by the TIF_MEMDIE flag, yes.
2137 * If __GFP_HARDWALL is set, cpuset_zone_allowed_softwall()
2138 * reduces to cpuset_zone_allowed_hardwall(). Otherwise,
2139 * cpuset_zone_allowed_softwall() might sleep, and might allow a zone
2140 * from an enclosing cpuset.
2142 * cpuset_zone_allowed_hardwall() only handles the simpler case of
2143 * hardwall cpusets, and never sleeps.
2145 * The __GFP_THISNODE placement logic is really handled elsewhere,
2146 * by forcibly using a zonelist starting at a specified node, and by
2147 * (in get_page_from_freelist()) refusing to consider the zones for
2148 * any node on the zonelist except the first. By the time any such
2149 * calls get to this routine, we should just shut up and say 'yes'.
2151 * GFP_USER allocations are marked with the __GFP_HARDWALL bit,
2152 * and do not allow allocations outside the current tasks cpuset
2153 * unless the task has been OOM killed as is marked TIF_MEMDIE.
2154 * GFP_KERNEL allocations are not so marked, so can escape to the
2155 * nearest enclosing hardwalled ancestor cpuset.
2157 * Scanning up parent cpusets requires callback_mutex. The
2158 * __alloc_pages() routine only calls here with __GFP_HARDWALL bit
2159 * _not_ set if it's a GFP_KERNEL allocation, and all nodes in the
2160 * current tasks mems_allowed came up empty on the first pass over
2161 * the zonelist. So only GFP_KERNEL allocations, if all nodes in the
2162 * cpuset are short of memory, might require taking the callback_mutex
2165 * The first call here from mm/page_alloc:get_page_from_freelist()
2166 * has __GFP_HARDWALL set in gfp_mask, enforcing hardwall cpusets,
2167 * so no allocation on a node outside the cpuset is allowed (unless
2168 * in interrupt, of course).
2170 * The second pass through get_page_from_freelist() doesn't even call
2171 * here for GFP_ATOMIC calls. For those calls, the __alloc_pages()
2172 * variable 'wait' is not set, and the bit ALLOC_CPUSET is not set
2173 * in alloc_flags. That logic and the checks below have the combined
2175 * in_interrupt - any node ok (current task context irrelevant)
2176 * GFP_ATOMIC - any node ok
2177 * TIF_MEMDIE - any node ok
2178 * GFP_KERNEL - any node in enclosing hardwalled cpuset ok
2179 * GFP_USER - only nodes in current tasks mems allowed ok.
2182 * Don't call cpuset_zone_allowed_softwall if you can't sleep, unless you
2183 * pass in the __GFP_HARDWALL flag set in gfp_flag, which disables
2184 * the code that might scan up ancestor cpusets and sleep.
2187 int __cpuset_zone_allowed_softwall(struct zone
*z
, gfp_t gfp_mask
)
2189 int node
; /* node that zone z is on */
2190 const struct cpuset
*cs
; /* current cpuset ancestors */
2191 int allowed
; /* is allocation in zone z allowed? */
2193 if (in_interrupt() || (gfp_mask
& __GFP_THISNODE
))
2195 node
= zone_to_nid(z
);
2196 might_sleep_if(!(gfp_mask
& __GFP_HARDWALL
));
2197 if (node_isset(node
, current
->mems_allowed
))
2200 * Allow tasks that have access to memory reserves because they have
2201 * been OOM killed to get memory anywhere.
2203 if (unlikely(test_thread_flag(TIF_MEMDIE
)))
2205 if (gfp_mask
& __GFP_HARDWALL
) /* If hardwall request, stop here */
2208 if (current
->flags
& PF_EXITING
) /* Let dying task have memory */
2211 /* Not hardwall and node outside mems_allowed: scan up cpusets */
2212 mutex_lock(&callback_mutex
);
2215 cs
= nearest_hardwall_ancestor(task_cs(current
));
2216 task_unlock(current
);
2218 allowed
= node_isset(node
, cs
->mems_allowed
);
2219 mutex_unlock(&callback_mutex
);
2224 * cpuset_zone_allowed_hardwall - Can we allocate on zone z's memory node?
2225 * @z: is this zone on an allowed node?
2226 * @gfp_mask: memory allocation flags
2228 * If we're in interrupt, yes, we can always allocate.
2229 * If __GFP_THISNODE is set, yes, we can always allocate. If zone
2230 * z's node is in our tasks mems_allowed, yes. If the task has been
2231 * OOM killed and has access to memory reserves as specified by the
2232 * TIF_MEMDIE flag, yes. Otherwise, no.
2234 * The __GFP_THISNODE placement logic is really handled elsewhere,
2235 * by forcibly using a zonelist starting at a specified node, and by
2236 * (in get_page_from_freelist()) refusing to consider the zones for
2237 * any node on the zonelist except the first. By the time any such
2238 * calls get to this routine, we should just shut up and say 'yes'.
2240 * Unlike the cpuset_zone_allowed_softwall() variant, above,
2241 * this variant requires that the zone be in the current tasks
2242 * mems_allowed or that we're in interrupt. It does not scan up the
2243 * cpuset hierarchy for the nearest enclosing mem_exclusive cpuset.
2247 int __cpuset_zone_allowed_hardwall(struct zone
*z
, gfp_t gfp_mask
)
2249 int node
; /* node that zone z is on */
2251 if (in_interrupt() || (gfp_mask
& __GFP_THISNODE
))
2253 node
= zone_to_nid(z
);
2254 if (node_isset(node
, current
->mems_allowed
))
2257 * Allow tasks that have access to memory reserves because they have
2258 * been OOM killed to get memory anywhere.
2260 if (unlikely(test_thread_flag(TIF_MEMDIE
)))
2266 * cpuset_lock - lock out any changes to cpuset structures
2268 * The out of memory (oom) code needs to mutex_lock cpusets
2269 * from being changed while it scans the tasklist looking for a
2270 * task in an overlapping cpuset. Expose callback_mutex via this
2271 * cpuset_lock() routine, so the oom code can lock it, before
2272 * locking the task list. The tasklist_lock is a spinlock, so
2273 * must be taken inside callback_mutex.
2276 void cpuset_lock(void)
2278 mutex_lock(&callback_mutex
);
2282 * cpuset_unlock - release lock on cpuset changes
2284 * Undo the lock taken in a previous cpuset_lock() call.
2287 void cpuset_unlock(void)
2289 mutex_unlock(&callback_mutex
);
2293 * cpuset_mem_spread_node() - On which node to begin search for a page
2295 * If a task is marked PF_SPREAD_PAGE or PF_SPREAD_SLAB (as for
2296 * tasks in a cpuset with is_spread_page or is_spread_slab set),
2297 * and if the memory allocation used cpuset_mem_spread_node()
2298 * to determine on which node to start looking, as it will for
2299 * certain page cache or slab cache pages such as used for file
2300 * system buffers and inode caches, then instead of starting on the
2301 * local node to look for a free page, rather spread the starting
2302 * node around the tasks mems_allowed nodes.
2304 * We don't have to worry about the returned node being offline
2305 * because "it can't happen", and even if it did, it would be ok.
2307 * The routines calling guarantee_online_mems() are careful to
2308 * only set nodes in task->mems_allowed that are online. So it
2309 * should not be possible for the following code to return an
2310 * offline node. But if it did, that would be ok, as this routine
2311 * is not returning the node where the allocation must be, only
2312 * the node where the search should start. The zonelist passed to
2313 * __alloc_pages() will include all nodes. If the slab allocator
2314 * is passed an offline node, it will fall back to the local node.
2315 * See kmem_cache_alloc_node().
2318 int cpuset_mem_spread_node(void)
2322 node
= next_node(current
->cpuset_mem_spread_rotor
, current
->mems_allowed
);
2323 if (node
== MAX_NUMNODES
)
2324 node
= first_node(current
->mems_allowed
);
2325 current
->cpuset_mem_spread_rotor
= node
;
2328 EXPORT_SYMBOL_GPL(cpuset_mem_spread_node
);
2331 * cpuset_mems_allowed_intersects - Does @tsk1's mems_allowed intersect @tsk2's?
2332 * @tsk1: pointer to task_struct of some task.
2333 * @tsk2: pointer to task_struct of some other task.
2335 * Description: Return true if @tsk1's mems_allowed intersects the
2336 * mems_allowed of @tsk2. Used by the OOM killer to determine if
2337 * one of the task's memory usage might impact the memory available
2341 int cpuset_mems_allowed_intersects(const struct task_struct
*tsk1
,
2342 const struct task_struct
*tsk2
)
2344 return nodes_intersects(tsk1
->mems_allowed
, tsk2
->mems_allowed
);
2348 * Collection of memory_pressure is suppressed unless
2349 * this flag is enabled by writing "1" to the special
2350 * cpuset file 'memory_pressure_enabled' in the root cpuset.
2353 int cpuset_memory_pressure_enabled __read_mostly
;
2356 * cpuset_memory_pressure_bump - keep stats of per-cpuset reclaims.
2358 * Keep a running average of the rate of synchronous (direct)
2359 * page reclaim efforts initiated by tasks in each cpuset.
2361 * This represents the rate at which some task in the cpuset
2362 * ran low on memory on all nodes it was allowed to use, and
2363 * had to enter the kernels page reclaim code in an effort to
2364 * create more free memory by tossing clean pages or swapping
2365 * or writing dirty pages.
2367 * Display to user space in the per-cpuset read-only file
2368 * "memory_pressure". Value displayed is an integer
2369 * representing the recent rate of entry into the synchronous
2370 * (direct) page reclaim by any task attached to the cpuset.
2373 void __cpuset_memory_pressure_bump(void)
2376 fmeter_markevent(&task_cs(current
)->fmeter
);
2377 task_unlock(current
);
2380 #ifdef CONFIG_PROC_PID_CPUSET
2382 * proc_cpuset_show()
2383 * - Print tasks cpuset path into seq_file.
2384 * - Used for /proc/<pid>/cpuset.
2385 * - No need to task_lock(tsk) on this tsk->cpuset reference, as it
2386 * doesn't really matter if tsk->cpuset changes after we read it,
2387 * and we take cgroup_mutex, keeping cpuset_attach() from changing it
2390 static int proc_cpuset_show(struct seq_file
*m
, void *unused_v
)
2393 struct task_struct
*tsk
;
2395 struct cgroup_subsys_state
*css
;
2399 buf
= kmalloc(PAGE_SIZE
, GFP_KERNEL
);
2405 tsk
= get_pid_task(pid
, PIDTYPE_PID
);
2411 css
= task_subsys_state(tsk
, cpuset_subsys_id
);
2412 retval
= cgroup_path(css
->cgroup
, buf
, PAGE_SIZE
);
2419 put_task_struct(tsk
);
2426 static int cpuset_open(struct inode
*inode
, struct file
*file
)
2428 struct pid
*pid
= PROC_I(inode
)->pid
;
2429 return single_open(file
, proc_cpuset_show
, pid
);
2432 const struct file_operations proc_cpuset_operations
= {
2433 .open
= cpuset_open
,
2435 .llseek
= seq_lseek
,
2436 .release
= single_release
,
2438 #endif /* CONFIG_PROC_PID_CPUSET */
2440 /* Display task cpus_allowed, mems_allowed in /proc/<pid>/status file. */
2441 void cpuset_task_status_allowed(struct seq_file
*m
, struct task_struct
*task
)
2443 seq_printf(m
, "Cpus_allowed:\t");
2444 m
->count
+= cpumask_scnprintf(m
->buf
+ m
->count
, m
->size
- m
->count
,
2445 task
->cpus_allowed
);
2446 seq_printf(m
, "\n");
2447 seq_printf(m
, "Cpus_allowed_list:\t");
2448 m
->count
+= cpulist_scnprintf(m
->buf
+ m
->count
, m
->size
- m
->count
,
2449 task
->cpus_allowed
);
2450 seq_printf(m
, "\n");
2451 seq_printf(m
, "Mems_allowed:\t");
2452 m
->count
+= nodemask_scnprintf(m
->buf
+ m
->count
, m
->size
- m
->count
,
2453 task
->mems_allowed
);
2454 seq_printf(m
, "\n");
2455 seq_printf(m
, "Mems_allowed_list:\t");
2456 m
->count
+= nodelist_scnprintf(m
->buf
+ m
->count
, m
->size
- m
->count
,
2457 task
->mems_allowed
);
2458 seq_printf(m
, "\n");