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/module.h>
41 #include <linux/mount.h>
42 #include <linux/namei.h>
43 #include <linux/pagemap.h>
44 #include <linux/proc_fs.h>
45 #include <linux/rcupdate.h>
46 #include <linux/sched.h>
47 #include <linux/seq_file.h>
48 #include <linux/security.h>
49 #include <linux/slab.h>
50 #include <linux/spinlock.h>
51 #include <linux/stat.h>
52 #include <linux/string.h>
53 #include <linux/time.h>
54 #include <linux/backing-dev.h>
55 #include <linux/sort.h>
57 #include <asm/uaccess.h>
58 #include <asm/atomic.h>
59 #include <linux/mutex.h>
60 #include <linux/workqueue.h>
61 #include <linux/cgroup.h>
64 * Workqueue for cpuset related tasks.
66 * Using kevent workqueue may cause deadlock when memory_migrate
67 * is set. So we create a separate workqueue thread for cpuset.
69 static struct workqueue_struct
*cpuset_wq
;
72 * Tracks how many cpusets are currently defined in system.
73 * When there is only one cpuset (the root cpuset) we can
74 * short circuit some hooks.
76 int number_of_cpusets __read_mostly
;
78 /* Forward declare cgroup structures */
79 struct cgroup_subsys cpuset_subsys
;
82 /* See "Frequency meter" comments, below. */
85 int cnt
; /* unprocessed events count */
86 int val
; /* most recent output value */
87 time_t time
; /* clock (secs) when val computed */
88 spinlock_t lock
; /* guards read or write of above */
92 struct cgroup_subsys_state css
;
94 unsigned long flags
; /* "unsigned long" so bitops work */
95 cpumask_var_t cpus_allowed
; /* CPUs allowed to tasks in cpuset */
96 nodemask_t mems_allowed
; /* Memory Nodes allowed to tasks */
98 struct cpuset
*parent
; /* my parent */
101 * Copy of global cpuset_mems_generation as of the most
102 * recent time this cpuset changed its mems_allowed.
106 struct fmeter fmeter
; /* memory_pressure filter */
108 /* partition number for rebuild_sched_domains() */
111 /* for custom sched domain */
112 int relax_domain_level
;
114 /* used for walking a cpuset heirarchy */
115 struct list_head stack_list
;
118 /* Retrieve the cpuset for a cgroup */
119 static inline struct cpuset
*cgroup_cs(struct cgroup
*cont
)
121 return container_of(cgroup_subsys_state(cont
, cpuset_subsys_id
),
125 /* Retrieve the cpuset for a task */
126 static inline struct cpuset
*task_cs(struct task_struct
*task
)
128 return container_of(task_subsys_state(task
, cpuset_subsys_id
),
131 struct cpuset_hotplug_scanner
{
132 struct cgroup_scanner scan
;
136 /* bits in struct cpuset flags field */
142 CS_SCHED_LOAD_BALANCE
,
147 /* convenient tests for these bits */
148 static inline int is_cpu_exclusive(const struct cpuset
*cs
)
150 return test_bit(CS_CPU_EXCLUSIVE
, &cs
->flags
);
153 static inline int is_mem_exclusive(const struct cpuset
*cs
)
155 return test_bit(CS_MEM_EXCLUSIVE
, &cs
->flags
);
158 static inline int is_mem_hardwall(const struct cpuset
*cs
)
160 return test_bit(CS_MEM_HARDWALL
, &cs
->flags
);
163 static inline int is_sched_load_balance(const struct cpuset
*cs
)
165 return test_bit(CS_SCHED_LOAD_BALANCE
, &cs
->flags
);
168 static inline int is_memory_migrate(const struct cpuset
*cs
)
170 return test_bit(CS_MEMORY_MIGRATE
, &cs
->flags
);
173 static inline int is_spread_page(const struct cpuset
*cs
)
175 return test_bit(CS_SPREAD_PAGE
, &cs
->flags
);
178 static inline int is_spread_slab(const struct cpuset
*cs
)
180 return test_bit(CS_SPREAD_SLAB
, &cs
->flags
);
184 * Increment this integer everytime any cpuset changes its
185 * mems_allowed value. Users of cpusets can track this generation
186 * number, and avoid having to lock and reload mems_allowed unless
187 * the cpuset they're using changes generation.
189 * A single, global generation is needed because cpuset_attach_task() could
190 * reattach a task to a different cpuset, which must not have its
191 * generation numbers aliased with those of that tasks previous cpuset.
193 * Generations are needed for mems_allowed because one task cannot
194 * modify another's memory placement. So we must enable every task,
195 * on every visit to __alloc_pages(), to efficiently check whether
196 * its current->cpuset->mems_allowed has changed, requiring an update
197 * of its current->mems_allowed.
199 * Since writes to cpuset_mems_generation are guarded by the cgroup lock
200 * there is no need to mark it atomic.
202 static int cpuset_mems_generation
;
204 static struct cpuset top_cpuset
= {
205 .flags
= ((1 << CS_CPU_EXCLUSIVE
) | (1 << CS_MEM_EXCLUSIVE
)),
209 * There are two global mutexes guarding cpuset structures. The first
210 * is the main control groups cgroup_mutex, accessed via
211 * cgroup_lock()/cgroup_unlock(). The second is the cpuset-specific
212 * callback_mutex, below. They can nest. It is ok to first take
213 * cgroup_mutex, then nest callback_mutex. We also require taking
214 * task_lock() when dereferencing a task's cpuset pointer. See "The
215 * task_lock() exception", at the end of this comment.
217 * A task must hold both mutexes to modify cpusets. If a task
218 * holds cgroup_mutex, then it blocks others wanting that mutex,
219 * ensuring that it is the only task able to also acquire callback_mutex
220 * and be able to modify cpusets. It can perform various checks on
221 * the cpuset structure first, knowing nothing will change. It can
222 * also allocate memory while just holding cgroup_mutex. While it is
223 * performing these checks, various callback routines can briefly
224 * acquire callback_mutex to query cpusets. Once it is ready to make
225 * the changes, it takes callback_mutex, blocking everyone else.
227 * Calls to the kernel memory allocator can not be made while holding
228 * callback_mutex, as that would risk double tripping on callback_mutex
229 * from one of the callbacks into the cpuset code from within
232 * If a task is only holding callback_mutex, then it has read-only
235 * The task_struct fields mems_allowed and mems_generation may only
236 * be accessed in the context of that task, so require no locks.
238 * The cpuset_common_file_read() handlers only hold callback_mutex across
239 * small pieces of code, such as when reading out possibly multi-word
240 * cpumasks and nodemasks.
242 * Accessing a task's cpuset should be done in accordance with the
243 * guidelines for accessing subsystem state in kernel/cgroup.c
246 static DEFINE_MUTEX(callback_mutex
);
249 * cpuset_buffer_lock protects both the cpuset_name and cpuset_nodelist
250 * buffers. They are statically allocated to prevent using excess stack
251 * when calling cpuset_print_task_mems_allowed().
253 #define CPUSET_NAME_LEN (128)
254 #define CPUSET_NODELIST_LEN (256)
255 static char cpuset_name
[CPUSET_NAME_LEN
];
256 static char cpuset_nodelist
[CPUSET_NODELIST_LEN
];
257 static DEFINE_SPINLOCK(cpuset_buffer_lock
);
260 * This is ugly, but preserves the userspace API for existing cpuset
261 * users. If someone tries to mount the "cpuset" filesystem, we
262 * silently switch it to mount "cgroup" instead
264 static int cpuset_get_sb(struct file_system_type
*fs_type
,
265 int flags
, const char *unused_dev_name
,
266 void *data
, struct vfsmount
*mnt
)
268 struct file_system_type
*cgroup_fs
= get_fs_type("cgroup");
273 "release_agent=/sbin/cpuset_release_agent";
274 ret
= cgroup_fs
->get_sb(cgroup_fs
, flags
,
275 unused_dev_name
, mountopts
, mnt
);
276 put_filesystem(cgroup_fs
);
281 static struct file_system_type cpuset_fs_type
= {
283 .get_sb
= cpuset_get_sb
,
287 * Return in pmask the portion of a cpusets's cpus_allowed that
288 * are online. If none are online, walk up the cpuset hierarchy
289 * until we find one that does have some online cpus. If we get
290 * all the way to the top and still haven't found any online cpus,
291 * return cpu_online_map. Or if passed a NULL cs from an exit'ing
292 * task, return cpu_online_map.
294 * One way or another, we guarantee to return some non-empty subset
297 * Call with callback_mutex held.
300 static void guarantee_online_cpus(const struct cpuset
*cs
,
301 struct cpumask
*pmask
)
303 while (cs
&& !cpumask_intersects(cs
->cpus_allowed
, cpu_online_mask
))
306 cpumask_and(pmask
, cs
->cpus_allowed
, cpu_online_mask
);
308 cpumask_copy(pmask
, cpu_online_mask
);
309 BUG_ON(!cpumask_intersects(pmask
, cpu_online_mask
));
313 * Return in *pmask the portion of a cpusets's mems_allowed that
314 * are online, with memory. If none are online with memory, walk
315 * up the cpuset hierarchy until we find one that does have some
316 * online mems. If we get all the way to the top and still haven't
317 * found any online mems, return node_states[N_HIGH_MEMORY].
319 * One way or another, we guarantee to return some non-empty subset
320 * of node_states[N_HIGH_MEMORY].
322 * Call with callback_mutex held.
325 static void guarantee_online_mems(const struct cpuset
*cs
, nodemask_t
*pmask
)
327 while (cs
&& !nodes_intersects(cs
->mems_allowed
,
328 node_states
[N_HIGH_MEMORY
]))
331 nodes_and(*pmask
, cs
->mems_allowed
,
332 node_states
[N_HIGH_MEMORY
]);
334 *pmask
= node_states
[N_HIGH_MEMORY
];
335 BUG_ON(!nodes_intersects(*pmask
, node_states
[N_HIGH_MEMORY
]));
339 * cpuset_update_task_memory_state - update task memory placement
341 * If the current tasks cpusets mems_allowed changed behind our
342 * backs, update current->mems_allowed, mems_generation and task NUMA
343 * mempolicy to the new value.
345 * Task mempolicy is updated by rebinding it relative to the
346 * current->cpuset if a task has its memory placement changed.
347 * Do not call this routine if in_interrupt().
349 * Call without callback_mutex or task_lock() held. May be
350 * called with or without cgroup_mutex held. Thanks in part to
351 * 'the_top_cpuset_hack', the task's cpuset pointer will never
352 * be NULL. This routine also might acquire callback_mutex during
355 * Reading current->cpuset->mems_generation doesn't need task_lock
356 * to guard the current->cpuset derefence, because it is guarded
357 * from concurrent freeing of current->cpuset using RCU.
359 * The rcu_dereference() is technically probably not needed,
360 * as I don't actually mind if I see a new cpuset pointer but
361 * an old value of mems_generation. However this really only
362 * matters on alpha systems using cpusets heavily. If I dropped
363 * that rcu_dereference(), it would save them a memory barrier.
364 * For all other arch's, rcu_dereference is a no-op anyway, and for
365 * alpha systems not using cpusets, another planned optimization,
366 * avoiding the rcu critical section for tasks in the root cpuset
367 * which is statically allocated, so can't vanish, will make this
368 * irrelevant. Better to use RCU as intended, than to engage in
369 * some cute trick to save a memory barrier that is impossible to
370 * test, for alpha systems using cpusets heavily, which might not
373 * This routine is needed to update the per-task mems_allowed data,
374 * within the tasks context, when it is trying to allocate memory
375 * (in various mm/mempolicy.c routines) and notices that some other
376 * task has been modifying its cpuset.
379 void cpuset_update_task_memory_state(void)
381 int my_cpusets_mem_gen
;
382 struct task_struct
*tsk
= current
;
386 my_cpusets_mem_gen
= task_cs(tsk
)->mems_generation
;
389 if (my_cpusets_mem_gen
!= tsk
->cpuset_mems_generation
) {
390 mutex_lock(&callback_mutex
);
392 cs
= task_cs(tsk
); /* Maybe changed when task not locked */
393 guarantee_online_mems(cs
, &tsk
->mems_allowed
);
394 tsk
->cpuset_mems_generation
= cs
->mems_generation
;
395 if (is_spread_page(cs
))
396 tsk
->flags
|= PF_SPREAD_PAGE
;
398 tsk
->flags
&= ~PF_SPREAD_PAGE
;
399 if (is_spread_slab(cs
))
400 tsk
->flags
|= PF_SPREAD_SLAB
;
402 tsk
->flags
&= ~PF_SPREAD_SLAB
;
404 mutex_unlock(&callback_mutex
);
405 mpol_rebind_task(tsk
, &tsk
->mems_allowed
);
410 * is_cpuset_subset(p, q) - Is cpuset p a subset of cpuset q?
412 * One cpuset is a subset of another if all its allowed CPUs and
413 * Memory Nodes are a subset of the other, and its exclusive flags
414 * are only set if the other's are set. Call holding cgroup_mutex.
417 static int is_cpuset_subset(const struct cpuset
*p
, const struct cpuset
*q
)
419 return cpumask_subset(p
->cpus_allowed
, q
->cpus_allowed
) &&
420 nodes_subset(p
->mems_allowed
, q
->mems_allowed
) &&
421 is_cpu_exclusive(p
) <= is_cpu_exclusive(q
) &&
422 is_mem_exclusive(p
) <= is_mem_exclusive(q
);
426 * alloc_trial_cpuset - allocate a trial cpuset
427 * @cs: the cpuset that the trial cpuset duplicates
429 static struct cpuset
*alloc_trial_cpuset(const struct cpuset
*cs
)
431 struct cpuset
*trial
;
433 trial
= kmemdup(cs
, sizeof(*cs
), GFP_KERNEL
);
437 if (!alloc_cpumask_var(&trial
->cpus_allowed
, GFP_KERNEL
)) {
441 cpumask_copy(trial
->cpus_allowed
, cs
->cpus_allowed
);
447 * free_trial_cpuset - free the trial cpuset
448 * @trial: the trial cpuset to be freed
450 static void free_trial_cpuset(struct cpuset
*trial
)
452 free_cpumask_var(trial
->cpus_allowed
);
457 * validate_change() - Used to validate that any proposed cpuset change
458 * follows the structural rules for cpusets.
460 * If we replaced the flag and mask values of the current cpuset
461 * (cur) with those values in the trial cpuset (trial), would
462 * our various subset and exclusive rules still be valid? Presumes
465 * 'cur' is the address of an actual, in-use cpuset. Operations
466 * such as list traversal that depend on the actual address of the
467 * cpuset in the list must use cur below, not trial.
469 * 'trial' is the address of bulk structure copy of cur, with
470 * perhaps one or more of the fields cpus_allowed, mems_allowed,
471 * or flags changed to new, trial values.
473 * Return 0 if valid, -errno if not.
476 static int validate_change(const struct cpuset
*cur
, const struct cpuset
*trial
)
479 struct cpuset
*c
, *par
;
481 /* Each of our child cpusets must be a subset of us */
482 list_for_each_entry(cont
, &cur
->css
.cgroup
->children
, sibling
) {
483 if (!is_cpuset_subset(cgroup_cs(cont
), trial
))
487 /* Remaining checks don't apply to root cpuset */
488 if (cur
== &top_cpuset
)
493 /* We must be a subset of our parent cpuset */
494 if (!is_cpuset_subset(trial
, par
))
498 * If either I or some sibling (!= me) is exclusive, we can't
501 list_for_each_entry(cont
, &par
->css
.cgroup
->children
, sibling
) {
503 if ((is_cpu_exclusive(trial
) || is_cpu_exclusive(c
)) &&
505 cpumask_intersects(trial
->cpus_allowed
, c
->cpus_allowed
))
507 if ((is_mem_exclusive(trial
) || is_mem_exclusive(c
)) &&
509 nodes_intersects(trial
->mems_allowed
, c
->mems_allowed
))
513 /* Cpusets with tasks can't have empty cpus_allowed or mems_allowed */
514 if (cgroup_task_count(cur
->css
.cgroup
)) {
515 if (cpumask_empty(trial
->cpus_allowed
) ||
516 nodes_empty(trial
->mems_allowed
)) {
525 * Helper routine for generate_sched_domains().
526 * Do cpusets a, b have overlapping cpus_allowed masks?
528 static int cpusets_overlap(struct cpuset
*a
, struct cpuset
*b
)
530 return cpumask_intersects(a
->cpus_allowed
, b
->cpus_allowed
);
534 update_domain_attr(struct sched_domain_attr
*dattr
, struct cpuset
*c
)
536 if (dattr
->relax_domain_level
< c
->relax_domain_level
)
537 dattr
->relax_domain_level
= c
->relax_domain_level
;
542 update_domain_attr_tree(struct sched_domain_attr
*dattr
, struct cpuset
*c
)
546 list_add(&c
->stack_list
, &q
);
547 while (!list_empty(&q
)) {
550 struct cpuset
*child
;
552 cp
= list_first_entry(&q
, struct cpuset
, stack_list
);
555 if (cpumask_empty(cp
->cpus_allowed
))
558 if (is_sched_load_balance(cp
))
559 update_domain_attr(dattr
, cp
);
561 list_for_each_entry(cont
, &cp
->css
.cgroup
->children
, sibling
) {
562 child
= cgroup_cs(cont
);
563 list_add_tail(&child
->stack_list
, &q
);
569 * generate_sched_domains()
571 * This function builds a partial partition of the systems CPUs
572 * A 'partial partition' is a set of non-overlapping subsets whose
573 * union is a subset of that set.
574 * The output of this function needs to be passed to kernel/sched.c
575 * partition_sched_domains() routine, which will rebuild the scheduler's
576 * load balancing domains (sched domains) as specified by that partial
579 * See "What is sched_load_balance" in Documentation/cgroups/cpusets.txt
580 * for a background explanation of this.
582 * Does not return errors, on the theory that the callers of this
583 * routine would rather not worry about failures to rebuild sched
584 * domains when operating in the severe memory shortage situations
585 * that could cause allocation failures below.
587 * Must be called with cgroup_lock held.
589 * The three key local variables below are:
590 * q - a linked-list queue of cpuset pointers, used to implement a
591 * top-down scan of all cpusets. This scan loads a pointer
592 * to each cpuset marked is_sched_load_balance into the
593 * array 'csa'. For our purposes, rebuilding the schedulers
594 * sched domains, we can ignore !is_sched_load_balance cpusets.
595 * csa - (for CpuSet Array) Array of pointers to all the cpusets
596 * that need to be load balanced, for convenient iterative
597 * access by the subsequent code that finds the best partition,
598 * i.e the set of domains (subsets) of CPUs such that the
599 * cpus_allowed of every cpuset marked is_sched_load_balance
600 * is a subset of one of these domains, while there are as
601 * many such domains as possible, each as small as possible.
602 * doms - Conversion of 'csa' to an array of cpumasks, for passing to
603 * the kernel/sched.c routine partition_sched_domains() in a
604 * convenient format, that can be easily compared to the prior
605 * value to determine what partition elements (sched domains)
606 * were changed (added or removed.)
608 * Finding the best partition (set of domains):
609 * The triple nested loops below over i, j, k scan over the
610 * load balanced cpusets (using the array of cpuset pointers in
611 * csa[]) looking for pairs of cpusets that have overlapping
612 * cpus_allowed, but which don't have the same 'pn' partition
613 * number and gives them in the same partition number. It keeps
614 * looping on the 'restart' label until it can no longer find
617 * The union of the cpus_allowed masks from the set of
618 * all cpusets having the same 'pn' value then form the one
619 * element of the partition (one sched domain) to be passed to
620 * partition_sched_domains().
622 /* FIXME: see the FIXME in partition_sched_domains() */
623 static int generate_sched_domains(struct cpumask
**domains
,
624 struct sched_domain_attr
**attributes
)
626 LIST_HEAD(q
); /* queue of cpusets to be scanned */
627 struct cpuset
*cp
; /* scans q */
628 struct cpuset
**csa
; /* array of all cpuset ptrs */
629 int csn
; /* how many cpuset ptrs in csa so far */
630 int i
, j
, k
; /* indices for partition finding loops */
631 struct cpumask
*doms
; /* resulting partition; i.e. sched domains */
632 struct sched_domain_attr
*dattr
; /* attributes for custom domains */
633 int ndoms
= 0; /* number of sched domains in result */
634 int nslot
; /* next empty doms[] struct cpumask slot */
640 /* Special case for the 99% of systems with one, full, sched domain */
641 if (is_sched_load_balance(&top_cpuset
)) {
642 doms
= kmalloc(cpumask_size(), GFP_KERNEL
);
646 dattr
= kmalloc(sizeof(struct sched_domain_attr
), GFP_KERNEL
);
648 *dattr
= SD_ATTR_INIT
;
649 update_domain_attr_tree(dattr
, &top_cpuset
);
651 cpumask_copy(doms
, top_cpuset
.cpus_allowed
);
657 csa
= kmalloc(number_of_cpusets
* sizeof(cp
), GFP_KERNEL
);
662 list_add(&top_cpuset
.stack_list
, &q
);
663 while (!list_empty(&q
)) {
665 struct cpuset
*child
; /* scans child cpusets of cp */
667 cp
= list_first_entry(&q
, struct cpuset
, stack_list
);
670 if (cpumask_empty(cp
->cpus_allowed
))
674 * All child cpusets contain a subset of the parent's cpus, so
675 * just skip them, and then we call update_domain_attr_tree()
676 * to calc relax_domain_level of the corresponding sched
679 if (is_sched_load_balance(cp
)) {
684 list_for_each_entry(cont
, &cp
->css
.cgroup
->children
, sibling
) {
685 child
= cgroup_cs(cont
);
686 list_add_tail(&child
->stack_list
, &q
);
690 for (i
= 0; i
< csn
; i
++)
695 /* Find the best partition (set of sched domains) */
696 for (i
= 0; i
< csn
; i
++) {
697 struct cpuset
*a
= csa
[i
];
700 for (j
= 0; j
< csn
; j
++) {
701 struct cpuset
*b
= csa
[j
];
704 if (apn
!= bpn
&& cpusets_overlap(a
, b
)) {
705 for (k
= 0; k
< csn
; k
++) {
706 struct cpuset
*c
= csa
[k
];
711 ndoms
--; /* one less element */
718 * Now we know how many domains to create.
719 * Convert <csn, csa> to <ndoms, doms> and populate cpu masks.
721 doms
= kmalloc(ndoms
* cpumask_size(), GFP_KERNEL
);
726 * The rest of the code, including the scheduler, can deal with
727 * dattr==NULL case. No need to abort if alloc fails.
729 dattr
= kmalloc(ndoms
* sizeof(struct sched_domain_attr
), GFP_KERNEL
);
731 for (nslot
= 0, i
= 0; i
< csn
; i
++) {
732 struct cpuset
*a
= csa
[i
];
737 /* Skip completed partitions */
743 if (nslot
== ndoms
) {
744 static int warnings
= 10;
747 "rebuild_sched_domains confused:"
748 " nslot %d, ndoms %d, csn %d, i %d,"
750 nslot
, ndoms
, csn
, i
, apn
);
758 *(dattr
+ nslot
) = SD_ATTR_INIT
;
759 for (j
= i
; j
< csn
; j
++) {
760 struct cpuset
*b
= csa
[j
];
763 cpumask_or(dp
, dp
, b
->cpus_allowed
);
765 update_domain_attr_tree(dattr
+ nslot
, b
);
767 /* Done with this partition */
773 BUG_ON(nslot
!= ndoms
);
779 * Fallback to the default domain if kmalloc() failed.
780 * See comments in partition_sched_domains().
791 * Rebuild scheduler domains.
793 * Call with neither cgroup_mutex held nor within get_online_cpus().
794 * Takes both cgroup_mutex and get_online_cpus().
796 * Cannot be directly called from cpuset code handling changes
797 * to the cpuset pseudo-filesystem, because it cannot be called
798 * from code that already holds cgroup_mutex.
800 static void do_rebuild_sched_domains(struct work_struct
*unused
)
802 struct sched_domain_attr
*attr
;
803 struct cpumask
*doms
;
808 /* Generate domain masks and attrs */
810 ndoms
= generate_sched_domains(&doms
, &attr
);
813 /* Have scheduler rebuild the domains */
814 partition_sched_domains(ndoms
, doms
, attr
);
819 static DECLARE_WORK(rebuild_sched_domains_work
, do_rebuild_sched_domains
);
822 * Rebuild scheduler domains, asynchronously via workqueue.
824 * If the flag 'sched_load_balance' of any cpuset with non-empty
825 * 'cpus' changes, or if the 'cpus' allowed changes in any cpuset
826 * which has that flag enabled, or if any cpuset with a non-empty
827 * 'cpus' is removed, then call this routine to rebuild the
828 * scheduler's dynamic sched domains.
830 * The rebuild_sched_domains() and partition_sched_domains()
831 * routines must nest cgroup_lock() inside get_online_cpus(),
832 * but such cpuset changes as these must nest that locking the
833 * other way, holding cgroup_lock() for much of the code.
835 * So in order to avoid an ABBA deadlock, the cpuset code handling
836 * these user changes delegates the actual sched domain rebuilding
837 * to a separate workqueue thread, which ends up processing the
838 * above do_rebuild_sched_domains() function.
840 static void async_rebuild_sched_domains(void)
842 queue_work(cpuset_wq
, &rebuild_sched_domains_work
);
846 * Accomplishes the same scheduler domain rebuild as the above
847 * async_rebuild_sched_domains(), however it directly calls the
848 * rebuild routine synchronously rather than calling it via an
849 * asynchronous work thread.
851 * This can only be called from code that is not holding
852 * cgroup_mutex (not nested in a cgroup_lock() call.)
854 void rebuild_sched_domains(void)
856 do_rebuild_sched_domains(NULL
);
860 * cpuset_test_cpumask - test a task's cpus_allowed versus its cpuset's
862 * @scan: struct cgroup_scanner contained in its struct cpuset_hotplug_scanner
864 * Call with cgroup_mutex held. May take callback_mutex during call.
865 * Called for each task in a cgroup by cgroup_scan_tasks().
866 * Return nonzero if this tasks's cpus_allowed mask should be changed (in other
867 * words, if its mask is not equal to its cpuset's mask).
869 static int cpuset_test_cpumask(struct task_struct
*tsk
,
870 struct cgroup_scanner
*scan
)
872 return !cpumask_equal(&tsk
->cpus_allowed
,
873 (cgroup_cs(scan
->cg
))->cpus_allowed
);
877 * cpuset_change_cpumask - make a task's cpus_allowed the same as its cpuset's
879 * @scan: struct cgroup_scanner containing the cgroup of the task
881 * Called by cgroup_scan_tasks() for each task in a cgroup whose
882 * cpus_allowed mask needs to be changed.
884 * We don't need to re-check for the cgroup/cpuset membership, since we're
885 * holding cgroup_lock() at this point.
887 static void cpuset_change_cpumask(struct task_struct
*tsk
,
888 struct cgroup_scanner
*scan
)
890 set_cpus_allowed_ptr(tsk
, ((cgroup_cs(scan
->cg
))->cpus_allowed
));
894 * update_tasks_cpumask - Update the cpumasks of tasks in the cpuset.
895 * @cs: the cpuset in which each task's cpus_allowed mask needs to be changed
896 * @heap: if NULL, defer allocating heap memory to cgroup_scan_tasks()
898 * Called with cgroup_mutex held
900 * The cgroup_scan_tasks() function will scan all the tasks in a cgroup,
901 * calling callback functions for each.
903 * No return value. It's guaranteed that cgroup_scan_tasks() always returns 0
906 static void update_tasks_cpumask(struct cpuset
*cs
, struct ptr_heap
*heap
)
908 struct cgroup_scanner scan
;
910 scan
.cg
= cs
->css
.cgroup
;
911 scan
.test_task
= cpuset_test_cpumask
;
912 scan
.process_task
= cpuset_change_cpumask
;
914 cgroup_scan_tasks(&scan
);
918 * update_cpumask - update the cpus_allowed mask of a cpuset and all tasks in it
919 * @cs: the cpuset to consider
920 * @buf: buffer of cpu numbers written to this cpuset
922 static int update_cpumask(struct cpuset
*cs
, struct cpuset
*trialcs
,
925 struct ptr_heap heap
;
927 int is_load_balanced
;
929 /* top_cpuset.cpus_allowed tracks cpu_online_map; it's read-only */
930 if (cs
== &top_cpuset
)
934 * An empty cpus_allowed is ok only if the cpuset has no tasks.
935 * Since cpulist_parse() fails on an empty mask, we special case
936 * that parsing. The validate_change() call ensures that cpusets
937 * with tasks have cpus.
940 cpumask_clear(trialcs
->cpus_allowed
);
942 retval
= cpulist_parse(buf
, trialcs
->cpus_allowed
);
946 if (!cpumask_subset(trialcs
->cpus_allowed
, cpu_online_mask
))
949 retval
= validate_change(cs
, trialcs
);
953 /* Nothing to do if the cpus didn't change */
954 if (cpumask_equal(cs
->cpus_allowed
, trialcs
->cpus_allowed
))
957 retval
= heap_init(&heap
, PAGE_SIZE
, GFP_KERNEL
, NULL
);
961 is_load_balanced
= is_sched_load_balance(trialcs
);
963 mutex_lock(&callback_mutex
);
964 cpumask_copy(cs
->cpus_allowed
, trialcs
->cpus_allowed
);
965 mutex_unlock(&callback_mutex
);
968 * Scan tasks in the cpuset, and update the cpumasks of any
969 * that need an update.
971 update_tasks_cpumask(cs
, &heap
);
975 if (is_load_balanced
)
976 async_rebuild_sched_domains();
983 * Migrate memory region from one set of nodes to another.
985 * Temporarilly set tasks mems_allowed to target nodes of migration,
986 * so that the migration code can allocate pages on these nodes.
988 * Call holding cgroup_mutex, so current's cpuset won't change
989 * during this call, as manage_mutex holds off any cpuset_attach()
990 * calls. Therefore we don't need to take task_lock around the
991 * call to guarantee_online_mems(), as we know no one is changing
994 * Hold callback_mutex around the two modifications of our tasks
995 * mems_allowed to synchronize with cpuset_mems_allowed().
997 * While the mm_struct we are migrating is typically from some
998 * other task, the task_struct mems_allowed that we are hacking
999 * is for our current task, which must allocate new pages for that
1000 * migrating memory region.
1002 * We call cpuset_update_task_memory_state() before hacking
1003 * our tasks mems_allowed, so that we are assured of being in
1004 * sync with our tasks cpuset, and in particular, callbacks to
1005 * cpuset_update_task_memory_state() from nested page allocations
1006 * won't see any mismatch of our cpuset and task mems_generation
1007 * values, so won't overwrite our hacked tasks mems_allowed
1011 static void cpuset_migrate_mm(struct mm_struct
*mm
, const nodemask_t
*from
,
1012 const nodemask_t
*to
)
1014 struct task_struct
*tsk
= current
;
1016 cpuset_update_task_memory_state();
1018 mutex_lock(&callback_mutex
);
1019 tsk
->mems_allowed
= *to
;
1020 mutex_unlock(&callback_mutex
);
1022 do_migrate_pages(mm
, from
, to
, MPOL_MF_MOVE_ALL
);
1024 mutex_lock(&callback_mutex
);
1025 guarantee_online_mems(task_cs(tsk
),&tsk
->mems_allowed
);
1026 mutex_unlock(&callback_mutex
);
1030 * Rebind task's vmas to cpuset's new mems_allowed, and migrate pages to new
1031 * nodes if memory_migrate flag is set. Called with cgroup_mutex held.
1033 static void cpuset_change_nodemask(struct task_struct
*p
,
1034 struct cgroup_scanner
*scan
)
1036 struct mm_struct
*mm
;
1039 const nodemask_t
*oldmem
= scan
->data
;
1041 mm
= get_task_mm(p
);
1045 cs
= cgroup_cs(scan
->cg
);
1046 migrate
= is_memory_migrate(cs
);
1048 mpol_rebind_mm(mm
, &cs
->mems_allowed
);
1050 cpuset_migrate_mm(mm
, oldmem
, &cs
->mems_allowed
);
1054 static void *cpuset_being_rebound
;
1057 * update_tasks_nodemask - Update the nodemasks of tasks in the cpuset.
1058 * @cs: the cpuset in which each task's mems_allowed mask needs to be changed
1059 * @oldmem: old mems_allowed of cpuset cs
1060 * @heap: if NULL, defer allocating heap memory to cgroup_scan_tasks()
1062 * Called with cgroup_mutex held
1063 * No return value. It's guaranteed that cgroup_scan_tasks() always returns 0
1066 static void update_tasks_nodemask(struct cpuset
*cs
, const nodemask_t
*oldmem
,
1067 struct ptr_heap
*heap
)
1069 struct cgroup_scanner scan
;
1071 cpuset_being_rebound
= cs
; /* causes mpol_dup() rebind */
1073 scan
.cg
= cs
->css
.cgroup
;
1074 scan
.test_task
= NULL
;
1075 scan
.process_task
= cpuset_change_nodemask
;
1077 scan
.data
= (nodemask_t
*)oldmem
;
1080 * The mpol_rebind_mm() call takes mmap_sem, which we couldn't
1081 * take while holding tasklist_lock. Forks can happen - the
1082 * mpol_dup() cpuset_being_rebound check will catch such forks,
1083 * and rebind their vma mempolicies too. Because we still hold
1084 * the global cgroup_mutex, we know that no other rebind effort
1085 * will be contending for the global variable cpuset_being_rebound.
1086 * It's ok if we rebind the same mm twice; mpol_rebind_mm()
1087 * is idempotent. Also migrate pages in each mm to new nodes.
1089 cgroup_scan_tasks(&scan
);
1091 /* We're done rebinding vmas to this cpuset's new mems_allowed. */
1092 cpuset_being_rebound
= NULL
;
1096 * Handle user request to change the 'mems' memory placement
1097 * of a cpuset. Needs to validate the request, update the
1098 * cpusets mems_allowed and mems_generation, and for each
1099 * task in the cpuset, rebind any vma mempolicies and if
1100 * the cpuset is marked 'memory_migrate', migrate the tasks
1101 * pages to the new memory.
1103 * Call with cgroup_mutex held. May take callback_mutex during call.
1104 * Will take tasklist_lock, scan tasklist for tasks in cpuset cs,
1105 * lock each such tasks mm->mmap_sem, scan its vma's and rebind
1106 * their mempolicies to the cpusets new mems_allowed.
1108 static int update_nodemask(struct cpuset
*cs
, struct cpuset
*trialcs
,
1113 struct ptr_heap heap
;
1116 * top_cpuset.mems_allowed tracks node_stats[N_HIGH_MEMORY];
1119 if (cs
== &top_cpuset
)
1123 * An empty mems_allowed is ok iff there are no tasks in the cpuset.
1124 * Since nodelist_parse() fails on an empty mask, we special case
1125 * that parsing. The validate_change() call ensures that cpusets
1126 * with tasks have memory.
1129 nodes_clear(trialcs
->mems_allowed
);
1131 retval
= nodelist_parse(buf
, trialcs
->mems_allowed
);
1135 if (!nodes_subset(trialcs
->mems_allowed
,
1136 node_states
[N_HIGH_MEMORY
]))
1139 oldmem
= cs
->mems_allowed
;
1140 if (nodes_equal(oldmem
, trialcs
->mems_allowed
)) {
1141 retval
= 0; /* Too easy - nothing to do */
1144 retval
= validate_change(cs
, trialcs
);
1148 retval
= heap_init(&heap
, PAGE_SIZE
, GFP_KERNEL
, NULL
);
1152 mutex_lock(&callback_mutex
);
1153 cs
->mems_allowed
= trialcs
->mems_allowed
;
1154 cs
->mems_generation
= cpuset_mems_generation
++;
1155 mutex_unlock(&callback_mutex
);
1157 update_tasks_nodemask(cs
, &oldmem
, &heap
);
1164 int current_cpuset_is_being_rebound(void)
1166 return task_cs(current
) == cpuset_being_rebound
;
1169 static int update_relax_domain_level(struct cpuset
*cs
, s64 val
)
1171 if (val
< -1 || val
>= SD_LV_MAX
)
1174 if (val
!= cs
->relax_domain_level
) {
1175 cs
->relax_domain_level
= val
;
1176 if (!cpumask_empty(cs
->cpus_allowed
) &&
1177 is_sched_load_balance(cs
))
1178 async_rebuild_sched_domains();
1185 * update_flag - read a 0 or a 1 in a file and update associated flag
1186 * bit: the bit to update (see cpuset_flagbits_t)
1187 * cs: the cpuset to update
1188 * turning_on: whether the flag is being set or cleared
1190 * Call with cgroup_mutex held.
1193 static int update_flag(cpuset_flagbits_t bit
, struct cpuset
*cs
,
1196 struct cpuset
*trialcs
;
1198 int balance_flag_changed
;
1200 trialcs
= alloc_trial_cpuset(cs
);
1205 set_bit(bit
, &trialcs
->flags
);
1207 clear_bit(bit
, &trialcs
->flags
);
1209 err
= validate_change(cs
, trialcs
);
1213 balance_flag_changed
= (is_sched_load_balance(cs
) !=
1214 is_sched_load_balance(trialcs
));
1216 mutex_lock(&callback_mutex
);
1217 cs
->flags
= trialcs
->flags
;
1218 mutex_unlock(&callback_mutex
);
1220 if (!cpumask_empty(trialcs
->cpus_allowed
) && balance_flag_changed
)
1221 async_rebuild_sched_domains();
1224 free_trial_cpuset(trialcs
);
1229 * Frequency meter - How fast is some event occurring?
1231 * These routines manage a digitally filtered, constant time based,
1232 * event frequency meter. There are four routines:
1233 * fmeter_init() - initialize a frequency meter.
1234 * fmeter_markevent() - called each time the event happens.
1235 * fmeter_getrate() - returns the recent rate of such events.
1236 * fmeter_update() - internal routine used to update fmeter.
1238 * A common data structure is passed to each of these routines,
1239 * which is used to keep track of the state required to manage the
1240 * frequency meter and its digital filter.
1242 * The filter works on the number of events marked per unit time.
1243 * The filter is single-pole low-pass recursive (IIR). The time unit
1244 * is 1 second. Arithmetic is done using 32-bit integers scaled to
1245 * simulate 3 decimal digits of precision (multiplied by 1000).
1247 * With an FM_COEF of 933, and a time base of 1 second, the filter
1248 * has a half-life of 10 seconds, meaning that if the events quit
1249 * happening, then the rate returned from the fmeter_getrate()
1250 * will be cut in half each 10 seconds, until it converges to zero.
1252 * It is not worth doing a real infinitely recursive filter. If more
1253 * than FM_MAXTICKS ticks have elapsed since the last filter event,
1254 * just compute FM_MAXTICKS ticks worth, by which point the level
1257 * Limit the count of unprocessed events to FM_MAXCNT, so as to avoid
1258 * arithmetic overflow in the fmeter_update() routine.
1260 * Given the simple 32 bit integer arithmetic used, this meter works
1261 * best for reporting rates between one per millisecond (msec) and
1262 * one per 32 (approx) seconds. At constant rates faster than one
1263 * per msec it maxes out at values just under 1,000,000. At constant
1264 * rates between one per msec, and one per second it will stabilize
1265 * to a value N*1000, where N is the rate of events per second.
1266 * At constant rates between one per second and one per 32 seconds,
1267 * it will be choppy, moving up on the seconds that have an event,
1268 * and then decaying until the next event. At rates slower than
1269 * about one in 32 seconds, it decays all the way back to zero between
1273 #define FM_COEF 933 /* coefficient for half-life of 10 secs */
1274 #define FM_MAXTICKS ((time_t)99) /* useless computing more ticks than this */
1275 #define FM_MAXCNT 1000000 /* limit cnt to avoid overflow */
1276 #define FM_SCALE 1000 /* faux fixed point scale */
1278 /* Initialize a frequency meter */
1279 static void fmeter_init(struct fmeter
*fmp
)
1284 spin_lock_init(&fmp
->lock
);
1287 /* Internal meter update - process cnt events and update value */
1288 static void fmeter_update(struct fmeter
*fmp
)
1290 time_t now
= get_seconds();
1291 time_t ticks
= now
- fmp
->time
;
1296 ticks
= min(FM_MAXTICKS
, ticks
);
1298 fmp
->val
= (FM_COEF
* fmp
->val
) / FM_SCALE
;
1301 fmp
->val
+= ((FM_SCALE
- FM_COEF
) * fmp
->cnt
) / FM_SCALE
;
1305 /* Process any previous ticks, then bump cnt by one (times scale). */
1306 static void fmeter_markevent(struct fmeter
*fmp
)
1308 spin_lock(&fmp
->lock
);
1310 fmp
->cnt
= min(FM_MAXCNT
, fmp
->cnt
+ FM_SCALE
);
1311 spin_unlock(&fmp
->lock
);
1314 /* Process any previous ticks, then return current value. */
1315 static int fmeter_getrate(struct fmeter
*fmp
)
1319 spin_lock(&fmp
->lock
);
1322 spin_unlock(&fmp
->lock
);
1326 /* Protected by cgroup_lock */
1327 static cpumask_var_t cpus_attach
;
1329 /* Called by cgroups to determine if a cpuset is usable; cgroup_mutex held */
1330 static int cpuset_can_attach(struct cgroup_subsys
*ss
,
1331 struct cgroup
*cont
, struct task_struct
*tsk
)
1333 struct cpuset
*cs
= cgroup_cs(cont
);
1336 if (cpumask_empty(cs
->cpus_allowed
) || nodes_empty(cs
->mems_allowed
))
1339 if (tsk
->flags
& PF_THREAD_BOUND
) {
1340 mutex_lock(&callback_mutex
);
1341 if (!cpumask_equal(&tsk
->cpus_allowed
, cs
->cpus_allowed
))
1343 mutex_unlock(&callback_mutex
);
1346 return ret
< 0 ? ret
: security_task_setscheduler(tsk
, 0, NULL
);
1349 static void cpuset_attach(struct cgroup_subsys
*ss
,
1350 struct cgroup
*cont
, struct cgroup
*oldcont
,
1351 struct task_struct
*tsk
)
1353 nodemask_t from
, to
;
1354 struct mm_struct
*mm
;
1355 struct cpuset
*cs
= cgroup_cs(cont
);
1356 struct cpuset
*oldcs
= cgroup_cs(oldcont
);
1359 if (cs
== &top_cpuset
) {
1360 cpumask_copy(cpus_attach
, cpu_possible_mask
);
1362 mutex_lock(&callback_mutex
);
1363 guarantee_online_cpus(cs
, cpus_attach
);
1364 mutex_unlock(&callback_mutex
);
1366 err
= set_cpus_allowed_ptr(tsk
, cpus_attach
);
1370 from
= oldcs
->mems_allowed
;
1371 to
= cs
->mems_allowed
;
1372 mm
= get_task_mm(tsk
);
1374 mpol_rebind_mm(mm
, &to
);
1375 if (is_memory_migrate(cs
))
1376 cpuset_migrate_mm(mm
, &from
, &to
);
1381 /* The various types of files and directories in a cpuset file system */
1384 FILE_MEMORY_MIGRATE
,
1390 FILE_SCHED_LOAD_BALANCE
,
1391 FILE_SCHED_RELAX_DOMAIN_LEVEL
,
1392 FILE_MEMORY_PRESSURE_ENABLED
,
1393 FILE_MEMORY_PRESSURE
,
1396 } cpuset_filetype_t
;
1398 static int cpuset_write_u64(struct cgroup
*cgrp
, struct cftype
*cft
, u64 val
)
1401 struct cpuset
*cs
= cgroup_cs(cgrp
);
1402 cpuset_filetype_t type
= cft
->private;
1404 if (!cgroup_lock_live_group(cgrp
))
1408 case FILE_CPU_EXCLUSIVE
:
1409 retval
= update_flag(CS_CPU_EXCLUSIVE
, cs
, val
);
1411 case FILE_MEM_EXCLUSIVE
:
1412 retval
= update_flag(CS_MEM_EXCLUSIVE
, cs
, val
);
1414 case FILE_MEM_HARDWALL
:
1415 retval
= update_flag(CS_MEM_HARDWALL
, cs
, val
);
1417 case FILE_SCHED_LOAD_BALANCE
:
1418 retval
= update_flag(CS_SCHED_LOAD_BALANCE
, cs
, val
);
1420 case FILE_MEMORY_MIGRATE
:
1421 retval
= update_flag(CS_MEMORY_MIGRATE
, cs
, val
);
1423 case FILE_MEMORY_PRESSURE_ENABLED
:
1424 cpuset_memory_pressure_enabled
= !!val
;
1426 case FILE_MEMORY_PRESSURE
:
1429 case FILE_SPREAD_PAGE
:
1430 retval
= update_flag(CS_SPREAD_PAGE
, cs
, val
);
1431 cs
->mems_generation
= cpuset_mems_generation
++;
1433 case FILE_SPREAD_SLAB
:
1434 retval
= update_flag(CS_SPREAD_SLAB
, cs
, val
);
1435 cs
->mems_generation
= cpuset_mems_generation
++;
1445 static int cpuset_write_s64(struct cgroup
*cgrp
, struct cftype
*cft
, s64 val
)
1448 struct cpuset
*cs
= cgroup_cs(cgrp
);
1449 cpuset_filetype_t type
= cft
->private;
1451 if (!cgroup_lock_live_group(cgrp
))
1455 case FILE_SCHED_RELAX_DOMAIN_LEVEL
:
1456 retval
= update_relax_domain_level(cs
, val
);
1467 * Common handling for a write to a "cpus" or "mems" file.
1469 static int cpuset_write_resmask(struct cgroup
*cgrp
, struct cftype
*cft
,
1473 struct cpuset
*cs
= cgroup_cs(cgrp
);
1474 struct cpuset
*trialcs
;
1476 if (!cgroup_lock_live_group(cgrp
))
1479 trialcs
= alloc_trial_cpuset(cs
);
1483 switch (cft
->private) {
1485 retval
= update_cpumask(cs
, trialcs
, buf
);
1488 retval
= update_nodemask(cs
, trialcs
, buf
);
1495 free_trial_cpuset(trialcs
);
1501 * These ascii lists should be read in a single call, by using a user
1502 * buffer large enough to hold the entire map. If read in smaller
1503 * chunks, there is no guarantee of atomicity. Since the display format
1504 * used, list of ranges of sequential numbers, is variable length,
1505 * and since these maps can change value dynamically, one could read
1506 * gibberish by doing partial reads while a list was changing.
1507 * A single large read to a buffer that crosses a page boundary is
1508 * ok, because the result being copied to user land is not recomputed
1509 * across a page fault.
1512 static int cpuset_sprintf_cpulist(char *page
, struct cpuset
*cs
)
1516 mutex_lock(&callback_mutex
);
1517 ret
= cpulist_scnprintf(page
, PAGE_SIZE
, cs
->cpus_allowed
);
1518 mutex_unlock(&callback_mutex
);
1523 static int cpuset_sprintf_memlist(char *page
, struct cpuset
*cs
)
1527 mutex_lock(&callback_mutex
);
1528 mask
= cs
->mems_allowed
;
1529 mutex_unlock(&callback_mutex
);
1531 return nodelist_scnprintf(page
, PAGE_SIZE
, mask
);
1534 static ssize_t
cpuset_common_file_read(struct cgroup
*cont
,
1538 size_t nbytes
, loff_t
*ppos
)
1540 struct cpuset
*cs
= cgroup_cs(cont
);
1541 cpuset_filetype_t type
= cft
->private;
1546 if (!(page
= (char *)__get_free_page(GFP_TEMPORARY
)))
1553 s
+= cpuset_sprintf_cpulist(s
, cs
);
1556 s
+= cpuset_sprintf_memlist(s
, cs
);
1564 retval
= simple_read_from_buffer(buf
, nbytes
, ppos
, page
, s
- page
);
1566 free_page((unsigned long)page
);
1570 static u64
cpuset_read_u64(struct cgroup
*cont
, struct cftype
*cft
)
1572 struct cpuset
*cs
= cgroup_cs(cont
);
1573 cpuset_filetype_t type
= cft
->private;
1575 case FILE_CPU_EXCLUSIVE
:
1576 return is_cpu_exclusive(cs
);
1577 case FILE_MEM_EXCLUSIVE
:
1578 return is_mem_exclusive(cs
);
1579 case FILE_MEM_HARDWALL
:
1580 return is_mem_hardwall(cs
);
1581 case FILE_SCHED_LOAD_BALANCE
:
1582 return is_sched_load_balance(cs
);
1583 case FILE_MEMORY_MIGRATE
:
1584 return is_memory_migrate(cs
);
1585 case FILE_MEMORY_PRESSURE_ENABLED
:
1586 return cpuset_memory_pressure_enabled
;
1587 case FILE_MEMORY_PRESSURE
:
1588 return fmeter_getrate(&cs
->fmeter
);
1589 case FILE_SPREAD_PAGE
:
1590 return is_spread_page(cs
);
1591 case FILE_SPREAD_SLAB
:
1592 return is_spread_slab(cs
);
1597 /* Unreachable but makes gcc happy */
1601 static s64
cpuset_read_s64(struct cgroup
*cont
, struct cftype
*cft
)
1603 struct cpuset
*cs
= cgroup_cs(cont
);
1604 cpuset_filetype_t type
= cft
->private;
1606 case FILE_SCHED_RELAX_DOMAIN_LEVEL
:
1607 return cs
->relax_domain_level
;
1612 /* Unrechable but makes gcc happy */
1618 * for the common functions, 'private' gives the type of file
1621 static struct cftype files
[] = {
1624 .read
= cpuset_common_file_read
,
1625 .write_string
= cpuset_write_resmask
,
1626 .max_write_len
= (100U + 6 * NR_CPUS
),
1627 .private = FILE_CPULIST
,
1632 .read
= cpuset_common_file_read
,
1633 .write_string
= cpuset_write_resmask
,
1634 .max_write_len
= (100U + 6 * MAX_NUMNODES
),
1635 .private = FILE_MEMLIST
,
1639 .name
= "cpu_exclusive",
1640 .read_u64
= cpuset_read_u64
,
1641 .write_u64
= cpuset_write_u64
,
1642 .private = FILE_CPU_EXCLUSIVE
,
1646 .name
= "mem_exclusive",
1647 .read_u64
= cpuset_read_u64
,
1648 .write_u64
= cpuset_write_u64
,
1649 .private = FILE_MEM_EXCLUSIVE
,
1653 .name
= "mem_hardwall",
1654 .read_u64
= cpuset_read_u64
,
1655 .write_u64
= cpuset_write_u64
,
1656 .private = FILE_MEM_HARDWALL
,
1660 .name
= "sched_load_balance",
1661 .read_u64
= cpuset_read_u64
,
1662 .write_u64
= cpuset_write_u64
,
1663 .private = FILE_SCHED_LOAD_BALANCE
,
1667 .name
= "sched_relax_domain_level",
1668 .read_s64
= cpuset_read_s64
,
1669 .write_s64
= cpuset_write_s64
,
1670 .private = FILE_SCHED_RELAX_DOMAIN_LEVEL
,
1674 .name
= "memory_migrate",
1675 .read_u64
= cpuset_read_u64
,
1676 .write_u64
= cpuset_write_u64
,
1677 .private = FILE_MEMORY_MIGRATE
,
1681 .name
= "memory_pressure",
1682 .read_u64
= cpuset_read_u64
,
1683 .write_u64
= cpuset_write_u64
,
1684 .private = FILE_MEMORY_PRESSURE
,
1689 .name
= "memory_spread_page",
1690 .read_u64
= cpuset_read_u64
,
1691 .write_u64
= cpuset_write_u64
,
1692 .private = FILE_SPREAD_PAGE
,
1696 .name
= "memory_spread_slab",
1697 .read_u64
= cpuset_read_u64
,
1698 .write_u64
= cpuset_write_u64
,
1699 .private = FILE_SPREAD_SLAB
,
1703 static struct cftype cft_memory_pressure_enabled
= {
1704 .name
= "memory_pressure_enabled",
1705 .read_u64
= cpuset_read_u64
,
1706 .write_u64
= cpuset_write_u64
,
1707 .private = FILE_MEMORY_PRESSURE_ENABLED
,
1710 static int cpuset_populate(struct cgroup_subsys
*ss
, struct cgroup
*cont
)
1714 err
= cgroup_add_files(cont
, ss
, files
, ARRAY_SIZE(files
));
1717 /* memory_pressure_enabled is in root cpuset only */
1719 err
= cgroup_add_file(cont
, ss
,
1720 &cft_memory_pressure_enabled
);
1725 * post_clone() is called at the end of cgroup_clone().
1726 * 'cgroup' was just created automatically as a result of
1727 * a cgroup_clone(), and the current task is about to
1728 * be moved into 'cgroup'.
1730 * Currently we refuse to set up the cgroup - thereby
1731 * refusing the task to be entered, and as a result refusing
1732 * the sys_unshare() or clone() which initiated it - if any
1733 * sibling cpusets have exclusive cpus or mem.
1735 * If this becomes a problem for some users who wish to
1736 * allow that scenario, then cpuset_post_clone() could be
1737 * changed to grant parent->cpus_allowed-sibling_cpus_exclusive
1738 * (and likewise for mems) to the new cgroup. Called with cgroup_mutex
1741 static void cpuset_post_clone(struct cgroup_subsys
*ss
,
1742 struct cgroup
*cgroup
)
1744 struct cgroup
*parent
, *child
;
1745 struct cpuset
*cs
, *parent_cs
;
1747 parent
= cgroup
->parent
;
1748 list_for_each_entry(child
, &parent
->children
, sibling
) {
1749 cs
= cgroup_cs(child
);
1750 if (is_mem_exclusive(cs
) || is_cpu_exclusive(cs
))
1753 cs
= cgroup_cs(cgroup
);
1754 parent_cs
= cgroup_cs(parent
);
1756 cs
->mems_allowed
= parent_cs
->mems_allowed
;
1757 cpumask_copy(cs
->cpus_allowed
, parent_cs
->cpus_allowed
);
1762 * cpuset_create - create a cpuset
1763 * ss: cpuset cgroup subsystem
1764 * cont: control group that the new cpuset will be part of
1767 static struct cgroup_subsys_state
*cpuset_create(
1768 struct cgroup_subsys
*ss
,
1769 struct cgroup
*cont
)
1772 struct cpuset
*parent
;
1774 if (!cont
->parent
) {
1775 /* This is early initialization for the top cgroup */
1776 top_cpuset
.mems_generation
= cpuset_mems_generation
++;
1777 return &top_cpuset
.css
;
1779 parent
= cgroup_cs(cont
->parent
);
1780 cs
= kmalloc(sizeof(*cs
), GFP_KERNEL
);
1782 return ERR_PTR(-ENOMEM
);
1783 if (!alloc_cpumask_var(&cs
->cpus_allowed
, GFP_KERNEL
)) {
1785 return ERR_PTR(-ENOMEM
);
1788 cpuset_update_task_memory_state();
1790 if (is_spread_page(parent
))
1791 set_bit(CS_SPREAD_PAGE
, &cs
->flags
);
1792 if (is_spread_slab(parent
))
1793 set_bit(CS_SPREAD_SLAB
, &cs
->flags
);
1794 set_bit(CS_SCHED_LOAD_BALANCE
, &cs
->flags
);
1795 cpumask_clear(cs
->cpus_allowed
);
1796 nodes_clear(cs
->mems_allowed
);
1797 cs
->mems_generation
= cpuset_mems_generation
++;
1798 fmeter_init(&cs
->fmeter
);
1799 cs
->relax_domain_level
= -1;
1801 cs
->parent
= parent
;
1802 number_of_cpusets
++;
1807 * If the cpuset being removed has its flag 'sched_load_balance'
1808 * enabled, then simulate turning sched_load_balance off, which
1809 * will call async_rebuild_sched_domains().
1812 static void cpuset_destroy(struct cgroup_subsys
*ss
, struct cgroup
*cont
)
1814 struct cpuset
*cs
= cgroup_cs(cont
);
1816 cpuset_update_task_memory_state();
1818 if (is_sched_load_balance(cs
))
1819 update_flag(CS_SCHED_LOAD_BALANCE
, cs
, 0);
1821 number_of_cpusets
--;
1822 free_cpumask_var(cs
->cpus_allowed
);
1826 struct cgroup_subsys cpuset_subsys
= {
1828 .create
= cpuset_create
,
1829 .destroy
= cpuset_destroy
,
1830 .can_attach
= cpuset_can_attach
,
1831 .attach
= cpuset_attach
,
1832 .populate
= cpuset_populate
,
1833 .post_clone
= cpuset_post_clone
,
1834 .subsys_id
= cpuset_subsys_id
,
1839 * cpuset_init_early - just enough so that the calls to
1840 * cpuset_update_task_memory_state() in early init code
1844 int __init
cpuset_init_early(void)
1846 alloc_bootmem_cpumask_var(&top_cpuset
.cpus_allowed
);
1848 top_cpuset
.mems_generation
= cpuset_mems_generation
++;
1854 * cpuset_init - initialize cpusets at system boot
1856 * Description: Initialize top_cpuset and the cpuset internal file system,
1859 int __init
cpuset_init(void)
1863 cpumask_setall(top_cpuset
.cpus_allowed
);
1864 nodes_setall(top_cpuset
.mems_allowed
);
1866 fmeter_init(&top_cpuset
.fmeter
);
1867 top_cpuset
.mems_generation
= cpuset_mems_generation
++;
1868 set_bit(CS_SCHED_LOAD_BALANCE
, &top_cpuset
.flags
);
1869 top_cpuset
.relax_domain_level
= -1;
1871 err
= register_filesystem(&cpuset_fs_type
);
1875 if (!alloc_cpumask_var(&cpus_attach
, GFP_KERNEL
))
1878 number_of_cpusets
= 1;
1883 * cpuset_do_move_task - move a given task to another cpuset
1884 * @tsk: pointer to task_struct the task to move
1885 * @scan: struct cgroup_scanner contained in its struct cpuset_hotplug_scanner
1887 * Called by cgroup_scan_tasks() for each task in a cgroup.
1888 * Return nonzero to stop the walk through the tasks.
1890 static void cpuset_do_move_task(struct task_struct
*tsk
,
1891 struct cgroup_scanner
*scan
)
1893 struct cpuset_hotplug_scanner
*chsp
;
1895 chsp
= container_of(scan
, struct cpuset_hotplug_scanner
, scan
);
1896 cgroup_attach_task(chsp
->to
, tsk
);
1900 * move_member_tasks_to_cpuset - move tasks from one cpuset to another
1901 * @from: cpuset in which the tasks currently reside
1902 * @to: cpuset to which the tasks will be moved
1904 * Called with cgroup_mutex held
1905 * callback_mutex must not be held, as cpuset_attach() will take it.
1907 * The cgroup_scan_tasks() function will scan all the tasks in a cgroup,
1908 * calling callback functions for each.
1910 static void move_member_tasks_to_cpuset(struct cpuset
*from
, struct cpuset
*to
)
1912 struct cpuset_hotplug_scanner scan
;
1914 scan
.scan
.cg
= from
->css
.cgroup
;
1915 scan
.scan
.test_task
= NULL
; /* select all tasks in cgroup */
1916 scan
.scan
.process_task
= cpuset_do_move_task
;
1917 scan
.scan
.heap
= NULL
;
1918 scan
.to
= to
->css
.cgroup
;
1920 if (cgroup_scan_tasks(&scan
.scan
))
1921 printk(KERN_ERR
"move_member_tasks_to_cpuset: "
1922 "cgroup_scan_tasks failed\n");
1926 * If CPU and/or memory hotplug handlers, below, unplug any CPUs
1927 * or memory nodes, we need to walk over the cpuset hierarchy,
1928 * removing that CPU or node from all cpusets. If this removes the
1929 * last CPU or node from a cpuset, then move the tasks in the empty
1930 * cpuset to its next-highest non-empty parent.
1932 * Called with cgroup_mutex held
1933 * callback_mutex must not be held, as cpuset_attach() will take it.
1935 static void remove_tasks_in_empty_cpuset(struct cpuset
*cs
)
1937 struct cpuset
*parent
;
1940 * The cgroup's css_sets list is in use if there are tasks
1941 * in the cpuset; the list is empty if there are none;
1942 * the cs->css.refcnt seems always 0.
1944 if (list_empty(&cs
->css
.cgroup
->css_sets
))
1948 * Find its next-highest non-empty parent, (top cpuset
1949 * has online cpus, so can't be empty).
1951 parent
= cs
->parent
;
1952 while (cpumask_empty(parent
->cpus_allowed
) ||
1953 nodes_empty(parent
->mems_allowed
))
1954 parent
= parent
->parent
;
1956 move_member_tasks_to_cpuset(cs
, parent
);
1960 * Walk the specified cpuset subtree and look for empty cpusets.
1961 * The tasks of such cpuset must be moved to a parent cpuset.
1963 * Called with cgroup_mutex held. We take callback_mutex to modify
1964 * cpus_allowed and mems_allowed.
1966 * This walk processes the tree from top to bottom, completing one layer
1967 * before dropping down to the next. It always processes a node before
1968 * any of its children.
1970 * For now, since we lack memory hot unplug, we'll never see a cpuset
1971 * that has tasks along with an empty 'mems'. But if we did see such
1972 * a cpuset, we'd handle it just like we do if its 'cpus' was empty.
1974 static void scan_for_empty_cpusets(struct cpuset
*root
)
1977 struct cpuset
*cp
; /* scans cpusets being updated */
1978 struct cpuset
*child
; /* scans child cpusets of cp */
1979 struct cgroup
*cont
;
1982 list_add_tail((struct list_head
*)&root
->stack_list
, &queue
);
1984 while (!list_empty(&queue
)) {
1985 cp
= list_first_entry(&queue
, struct cpuset
, stack_list
);
1986 list_del(queue
.next
);
1987 list_for_each_entry(cont
, &cp
->css
.cgroup
->children
, sibling
) {
1988 child
= cgroup_cs(cont
);
1989 list_add_tail(&child
->stack_list
, &queue
);
1992 /* Continue past cpusets with all cpus, mems online */
1993 if (cpumask_subset(cp
->cpus_allowed
, cpu_online_mask
) &&
1994 nodes_subset(cp
->mems_allowed
, node_states
[N_HIGH_MEMORY
]))
1997 oldmems
= cp
->mems_allowed
;
1999 /* Remove offline cpus and mems from this cpuset. */
2000 mutex_lock(&callback_mutex
);
2001 cpumask_and(cp
->cpus_allowed
, cp
->cpus_allowed
,
2003 nodes_and(cp
->mems_allowed
, cp
->mems_allowed
,
2004 node_states
[N_HIGH_MEMORY
]);
2005 mutex_unlock(&callback_mutex
);
2007 /* Move tasks from the empty cpuset to a parent */
2008 if (cpumask_empty(cp
->cpus_allowed
) ||
2009 nodes_empty(cp
->mems_allowed
))
2010 remove_tasks_in_empty_cpuset(cp
);
2012 update_tasks_cpumask(cp
, NULL
);
2013 update_tasks_nodemask(cp
, &oldmems
, NULL
);
2019 * The top_cpuset tracks what CPUs and Memory Nodes are online,
2020 * period. This is necessary in order to make cpusets transparent
2021 * (of no affect) on systems that are actively using CPU hotplug
2022 * but making no active use of cpusets.
2024 * This routine ensures that top_cpuset.cpus_allowed tracks
2025 * cpu_online_map on each CPU hotplug (cpuhp) event.
2027 * Called within get_online_cpus(). Needs to call cgroup_lock()
2028 * before calling generate_sched_domains().
2030 static int cpuset_track_online_cpus(struct notifier_block
*unused_nb
,
2031 unsigned long phase
, void *unused_cpu
)
2033 struct sched_domain_attr
*attr
;
2034 struct cpumask
*doms
;
2039 case CPU_ONLINE_FROZEN
:
2041 case CPU_DEAD_FROZEN
:
2049 mutex_lock(&callback_mutex
);
2050 cpumask_copy(top_cpuset
.cpus_allowed
, cpu_online_mask
);
2051 mutex_unlock(&callback_mutex
);
2052 scan_for_empty_cpusets(&top_cpuset
);
2053 ndoms
= generate_sched_domains(&doms
, &attr
);
2056 /* Have scheduler rebuild the domains */
2057 partition_sched_domains(ndoms
, doms
, attr
);
2062 #ifdef CONFIG_MEMORY_HOTPLUG
2064 * Keep top_cpuset.mems_allowed tracking node_states[N_HIGH_MEMORY].
2065 * Call this routine anytime after node_states[N_HIGH_MEMORY] changes.
2066 * See also the previous routine cpuset_track_online_cpus().
2068 static int cpuset_track_online_nodes(struct notifier_block
*self
,
2069 unsigned long action
, void *arg
)
2075 mutex_lock(&callback_mutex
);
2076 top_cpuset
.mems_allowed
= node_states
[N_HIGH_MEMORY
];
2077 mutex_unlock(&callback_mutex
);
2078 if (action
== MEM_OFFLINE
)
2079 scan_for_empty_cpusets(&top_cpuset
);
2090 * cpuset_init_smp - initialize cpus_allowed
2092 * Description: Finish top cpuset after cpu, node maps are initialized
2095 void __init
cpuset_init_smp(void)
2097 cpumask_copy(top_cpuset
.cpus_allowed
, cpu_online_mask
);
2098 top_cpuset
.mems_allowed
= node_states
[N_HIGH_MEMORY
];
2100 hotcpu_notifier(cpuset_track_online_cpus
, 0);
2101 hotplug_memory_notifier(cpuset_track_online_nodes
, 10);
2103 cpuset_wq
= create_singlethread_workqueue("cpuset");
2108 * cpuset_cpus_allowed - return cpus_allowed mask from a tasks cpuset.
2109 * @tsk: pointer to task_struct from which to obtain cpuset->cpus_allowed.
2110 * @pmask: pointer to struct cpumask variable to receive cpus_allowed set.
2112 * Description: Returns the cpumask_var_t cpus_allowed of the cpuset
2113 * attached to the specified @tsk. Guaranteed to return some non-empty
2114 * subset of cpu_online_map, even if this means going outside the
2118 void cpuset_cpus_allowed(struct task_struct
*tsk
, struct cpumask
*pmask
)
2120 mutex_lock(&callback_mutex
);
2121 cpuset_cpus_allowed_locked(tsk
, pmask
);
2122 mutex_unlock(&callback_mutex
);
2126 * cpuset_cpus_allowed_locked - return cpus_allowed mask from a tasks cpuset.
2127 * Must be called with callback_mutex held.
2129 void cpuset_cpus_allowed_locked(struct task_struct
*tsk
, struct cpumask
*pmask
)
2132 guarantee_online_cpus(task_cs(tsk
), pmask
);
2136 void cpuset_init_current_mems_allowed(void)
2138 nodes_setall(current
->mems_allowed
);
2142 * cpuset_mems_allowed - return mems_allowed mask from a tasks cpuset.
2143 * @tsk: pointer to task_struct from which to obtain cpuset->mems_allowed.
2145 * Description: Returns the nodemask_t mems_allowed of the cpuset
2146 * attached to the specified @tsk. Guaranteed to return some non-empty
2147 * subset of node_states[N_HIGH_MEMORY], even if this means going outside the
2151 nodemask_t
cpuset_mems_allowed(struct task_struct
*tsk
)
2155 mutex_lock(&callback_mutex
);
2157 guarantee_online_mems(task_cs(tsk
), &mask
);
2159 mutex_unlock(&callback_mutex
);
2165 * cpuset_nodemask_valid_mems_allowed - check nodemask vs. curremt mems_allowed
2166 * @nodemask: the nodemask to be checked
2168 * Are any of the nodes in the nodemask allowed in current->mems_allowed?
2170 int cpuset_nodemask_valid_mems_allowed(nodemask_t
*nodemask
)
2172 return nodes_intersects(*nodemask
, current
->mems_allowed
);
2176 * nearest_hardwall_ancestor() - Returns the nearest mem_exclusive or
2177 * mem_hardwall ancestor to the specified cpuset. Call holding
2178 * callback_mutex. If no ancestor is mem_exclusive or mem_hardwall
2179 * (an unusual configuration), then returns the root cpuset.
2181 static const struct cpuset
*nearest_hardwall_ancestor(const struct cpuset
*cs
)
2183 while (!(is_mem_exclusive(cs
) || is_mem_hardwall(cs
)) && cs
->parent
)
2189 * cpuset_zone_allowed_softwall - Can we allocate on zone z's memory node?
2190 * @z: is this zone on an allowed node?
2191 * @gfp_mask: memory allocation flags
2193 * If we're in interrupt, yes, we can always allocate. If
2194 * __GFP_THISNODE is set, yes, we can always allocate. If zone
2195 * z's node is in our tasks mems_allowed, yes. If it's not a
2196 * __GFP_HARDWALL request and this zone's nodes is in the nearest
2197 * hardwalled cpuset ancestor to this tasks cpuset, yes.
2198 * If the task has been OOM killed and has access to memory reserves
2199 * as specified by the TIF_MEMDIE flag, yes.
2202 * If __GFP_HARDWALL is set, cpuset_zone_allowed_softwall()
2203 * reduces to cpuset_zone_allowed_hardwall(). Otherwise,
2204 * cpuset_zone_allowed_softwall() might sleep, and might allow a zone
2205 * from an enclosing cpuset.
2207 * cpuset_zone_allowed_hardwall() only handles the simpler case of
2208 * hardwall cpusets, and never sleeps.
2210 * The __GFP_THISNODE placement logic is really handled elsewhere,
2211 * by forcibly using a zonelist starting at a specified node, and by
2212 * (in get_page_from_freelist()) refusing to consider the zones for
2213 * any node on the zonelist except the first. By the time any such
2214 * calls get to this routine, we should just shut up and say 'yes'.
2216 * GFP_USER allocations are marked with the __GFP_HARDWALL bit,
2217 * and do not allow allocations outside the current tasks cpuset
2218 * unless the task has been OOM killed as is marked TIF_MEMDIE.
2219 * GFP_KERNEL allocations are not so marked, so can escape to the
2220 * nearest enclosing hardwalled ancestor cpuset.
2222 * Scanning up parent cpusets requires callback_mutex. The
2223 * __alloc_pages() routine only calls here with __GFP_HARDWALL bit
2224 * _not_ set if it's a GFP_KERNEL allocation, and all nodes in the
2225 * current tasks mems_allowed came up empty on the first pass over
2226 * the zonelist. So only GFP_KERNEL allocations, if all nodes in the
2227 * cpuset are short of memory, might require taking the callback_mutex
2230 * The first call here from mm/page_alloc:get_page_from_freelist()
2231 * has __GFP_HARDWALL set in gfp_mask, enforcing hardwall cpusets,
2232 * so no allocation on a node outside the cpuset is allowed (unless
2233 * in interrupt, of course).
2235 * The second pass through get_page_from_freelist() doesn't even call
2236 * here for GFP_ATOMIC calls. For those calls, the __alloc_pages()
2237 * variable 'wait' is not set, and the bit ALLOC_CPUSET is not set
2238 * in alloc_flags. That logic and the checks below have the combined
2240 * in_interrupt - any node ok (current task context irrelevant)
2241 * GFP_ATOMIC - any node ok
2242 * TIF_MEMDIE - any node ok
2243 * GFP_KERNEL - any node in enclosing hardwalled cpuset ok
2244 * GFP_USER - only nodes in current tasks mems allowed ok.
2247 * Don't call cpuset_zone_allowed_softwall if you can't sleep, unless you
2248 * pass in the __GFP_HARDWALL flag set in gfp_flag, which disables
2249 * the code that might scan up ancestor cpusets and sleep.
2252 int __cpuset_zone_allowed_softwall(struct zone
*z
, gfp_t gfp_mask
)
2254 int node
; /* node that zone z is on */
2255 const struct cpuset
*cs
; /* current cpuset ancestors */
2256 int allowed
; /* is allocation in zone z allowed? */
2258 if (in_interrupt() || (gfp_mask
& __GFP_THISNODE
))
2260 node
= zone_to_nid(z
);
2261 might_sleep_if(!(gfp_mask
& __GFP_HARDWALL
));
2262 if (node_isset(node
, current
->mems_allowed
))
2265 * Allow tasks that have access to memory reserves because they have
2266 * been OOM killed to get memory anywhere.
2268 if (unlikely(test_thread_flag(TIF_MEMDIE
)))
2270 if (gfp_mask
& __GFP_HARDWALL
) /* If hardwall request, stop here */
2273 if (current
->flags
& PF_EXITING
) /* Let dying task have memory */
2276 /* Not hardwall and node outside mems_allowed: scan up cpusets */
2277 mutex_lock(&callback_mutex
);
2280 cs
= nearest_hardwall_ancestor(task_cs(current
));
2281 task_unlock(current
);
2283 allowed
= node_isset(node
, cs
->mems_allowed
);
2284 mutex_unlock(&callback_mutex
);
2289 * cpuset_zone_allowed_hardwall - Can we allocate on zone z's memory node?
2290 * @z: is this zone on an allowed node?
2291 * @gfp_mask: memory allocation flags
2293 * If we're in interrupt, yes, we can always allocate.
2294 * If __GFP_THISNODE is set, yes, we can always allocate. If zone
2295 * z's node is in our tasks mems_allowed, yes. If the task has been
2296 * OOM killed and has access to memory reserves as specified by the
2297 * TIF_MEMDIE flag, yes. Otherwise, no.
2299 * The __GFP_THISNODE placement logic is really handled elsewhere,
2300 * by forcibly using a zonelist starting at a specified node, and by
2301 * (in get_page_from_freelist()) refusing to consider the zones for
2302 * any node on the zonelist except the first. By the time any such
2303 * calls get to this routine, we should just shut up and say 'yes'.
2305 * Unlike the cpuset_zone_allowed_softwall() variant, above,
2306 * this variant requires that the zone be in the current tasks
2307 * mems_allowed or that we're in interrupt. It does not scan up the
2308 * cpuset hierarchy for the nearest enclosing mem_exclusive cpuset.
2312 int __cpuset_zone_allowed_hardwall(struct zone
*z
, gfp_t gfp_mask
)
2314 int node
; /* node that zone z is on */
2316 if (in_interrupt() || (gfp_mask
& __GFP_THISNODE
))
2318 node
= zone_to_nid(z
);
2319 if (node_isset(node
, current
->mems_allowed
))
2322 * Allow tasks that have access to memory reserves because they have
2323 * been OOM killed to get memory anywhere.
2325 if (unlikely(test_thread_flag(TIF_MEMDIE
)))
2331 * cpuset_lock - lock out any changes to cpuset structures
2333 * The out of memory (oom) code needs to mutex_lock cpusets
2334 * from being changed while it scans the tasklist looking for a
2335 * task in an overlapping cpuset. Expose callback_mutex via this
2336 * cpuset_lock() routine, so the oom code can lock it, before
2337 * locking the task list. The tasklist_lock is a spinlock, so
2338 * must be taken inside callback_mutex.
2341 void cpuset_lock(void)
2343 mutex_lock(&callback_mutex
);
2347 * cpuset_unlock - release lock on cpuset changes
2349 * Undo the lock taken in a previous cpuset_lock() call.
2352 void cpuset_unlock(void)
2354 mutex_unlock(&callback_mutex
);
2358 * cpuset_mem_spread_node() - On which node to begin search for a page
2360 * If a task is marked PF_SPREAD_PAGE or PF_SPREAD_SLAB (as for
2361 * tasks in a cpuset with is_spread_page or is_spread_slab set),
2362 * and if the memory allocation used cpuset_mem_spread_node()
2363 * to determine on which node to start looking, as it will for
2364 * certain page cache or slab cache pages such as used for file
2365 * system buffers and inode caches, then instead of starting on the
2366 * local node to look for a free page, rather spread the starting
2367 * node around the tasks mems_allowed nodes.
2369 * We don't have to worry about the returned node being offline
2370 * because "it can't happen", and even if it did, it would be ok.
2372 * The routines calling guarantee_online_mems() are careful to
2373 * only set nodes in task->mems_allowed that are online. So it
2374 * should not be possible for the following code to return an
2375 * offline node. But if it did, that would be ok, as this routine
2376 * is not returning the node where the allocation must be, only
2377 * the node where the search should start. The zonelist passed to
2378 * __alloc_pages() will include all nodes. If the slab allocator
2379 * is passed an offline node, it will fall back to the local node.
2380 * See kmem_cache_alloc_node().
2383 int cpuset_mem_spread_node(void)
2387 node
= next_node(current
->cpuset_mem_spread_rotor
, current
->mems_allowed
);
2388 if (node
== MAX_NUMNODES
)
2389 node
= first_node(current
->mems_allowed
);
2390 current
->cpuset_mem_spread_rotor
= node
;
2393 EXPORT_SYMBOL_GPL(cpuset_mem_spread_node
);
2396 * cpuset_mems_allowed_intersects - Does @tsk1's mems_allowed intersect @tsk2's?
2397 * @tsk1: pointer to task_struct of some task.
2398 * @tsk2: pointer to task_struct of some other task.
2400 * Description: Return true if @tsk1's mems_allowed intersects the
2401 * mems_allowed of @tsk2. Used by the OOM killer to determine if
2402 * one of the task's memory usage might impact the memory available
2406 int cpuset_mems_allowed_intersects(const struct task_struct
*tsk1
,
2407 const struct task_struct
*tsk2
)
2409 return nodes_intersects(tsk1
->mems_allowed
, tsk2
->mems_allowed
);
2413 * cpuset_print_task_mems_allowed - prints task's cpuset and mems_allowed
2414 * @task: pointer to task_struct of some task.
2416 * Description: Prints @task's name, cpuset name, and cached copy of its
2417 * mems_allowed to the kernel log. Must hold task_lock(task) to allow
2418 * dereferencing task_cs(task).
2420 void cpuset_print_task_mems_allowed(struct task_struct
*tsk
)
2422 struct dentry
*dentry
;
2424 dentry
= task_cs(tsk
)->css
.cgroup
->dentry
;
2425 spin_lock(&cpuset_buffer_lock
);
2426 snprintf(cpuset_name
, CPUSET_NAME_LEN
,
2427 dentry
? (const char *)dentry
->d_name
.name
: "/");
2428 nodelist_scnprintf(cpuset_nodelist
, CPUSET_NODELIST_LEN
,
2430 printk(KERN_INFO
"%s cpuset=%s mems_allowed=%s\n",
2431 tsk
->comm
, cpuset_name
, cpuset_nodelist
);
2432 spin_unlock(&cpuset_buffer_lock
);
2436 * Collection of memory_pressure is suppressed unless
2437 * this flag is enabled by writing "1" to the special
2438 * cpuset file 'memory_pressure_enabled' in the root cpuset.
2441 int cpuset_memory_pressure_enabled __read_mostly
;
2444 * cpuset_memory_pressure_bump - keep stats of per-cpuset reclaims.
2446 * Keep a running average of the rate of synchronous (direct)
2447 * page reclaim efforts initiated by tasks in each cpuset.
2449 * This represents the rate at which some task in the cpuset
2450 * ran low on memory on all nodes it was allowed to use, and
2451 * had to enter the kernels page reclaim code in an effort to
2452 * create more free memory by tossing clean pages or swapping
2453 * or writing dirty pages.
2455 * Display to user space in the per-cpuset read-only file
2456 * "memory_pressure". Value displayed is an integer
2457 * representing the recent rate of entry into the synchronous
2458 * (direct) page reclaim by any task attached to the cpuset.
2461 void __cpuset_memory_pressure_bump(void)
2464 fmeter_markevent(&task_cs(current
)->fmeter
);
2465 task_unlock(current
);
2468 #ifdef CONFIG_PROC_PID_CPUSET
2470 * proc_cpuset_show()
2471 * - Print tasks cpuset path into seq_file.
2472 * - Used for /proc/<pid>/cpuset.
2473 * - No need to task_lock(tsk) on this tsk->cpuset reference, as it
2474 * doesn't really matter if tsk->cpuset changes after we read it,
2475 * and we take cgroup_mutex, keeping cpuset_attach() from changing it
2478 static int proc_cpuset_show(struct seq_file
*m
, void *unused_v
)
2481 struct task_struct
*tsk
;
2483 struct cgroup_subsys_state
*css
;
2487 buf
= kmalloc(PAGE_SIZE
, GFP_KERNEL
);
2493 tsk
= get_pid_task(pid
, PIDTYPE_PID
);
2499 css
= task_subsys_state(tsk
, cpuset_subsys_id
);
2500 retval
= cgroup_path(css
->cgroup
, buf
, PAGE_SIZE
);
2507 put_task_struct(tsk
);
2514 static int cpuset_open(struct inode
*inode
, struct file
*file
)
2516 struct pid
*pid
= PROC_I(inode
)->pid
;
2517 return single_open(file
, proc_cpuset_show
, pid
);
2520 const struct file_operations proc_cpuset_operations
= {
2521 .open
= cpuset_open
,
2523 .llseek
= seq_lseek
,
2524 .release
= single_release
,
2526 #endif /* CONFIG_PROC_PID_CPUSET */
2528 /* Display task cpus_allowed, mems_allowed in /proc/<pid>/status file. */
2529 void cpuset_task_status_allowed(struct seq_file
*m
, struct task_struct
*task
)
2531 seq_printf(m
, "Cpus_allowed:\t");
2532 seq_cpumask(m
, &task
->cpus_allowed
);
2533 seq_printf(m
, "\n");
2534 seq_printf(m
, "Cpus_allowed_list:\t");
2535 seq_cpumask_list(m
, &task
->cpus_allowed
);
2536 seq_printf(m
, "\n");
2537 seq_printf(m
, "Mems_allowed:\t");
2538 seq_nodemask(m
, &task
->mems_allowed
);
2539 seq_printf(m
, "\n");
2540 seq_printf(m
, "Mems_allowed_list:\t");
2541 seq_nodemask_list(m
, &task
->mems_allowed
);
2542 seq_printf(m
, "\n");