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
18 * This file is subject to the terms and conditions of the GNU General Public
19 * License. See the file COPYING in the main directory of the Linux
20 * distribution for more details.
23 #include <linux/cpu.h>
24 #include <linux/cpumask.h>
25 #include <linux/cpuset.h>
26 #include <linux/err.h>
27 #include <linux/errno.h>
28 #include <linux/file.h>
30 #include <linux/init.h>
31 #include <linux/interrupt.h>
32 #include <linux/kernel.h>
33 #include <linux/kmod.h>
34 #include <linux/list.h>
35 #include <linux/mempolicy.h>
37 #include <linux/module.h>
38 #include <linux/mount.h>
39 #include <linux/namei.h>
40 #include <linux/pagemap.h>
41 #include <linux/proc_fs.h>
42 #include <linux/rcupdate.h>
43 #include <linux/sched.h>
44 #include <linux/seq_file.h>
45 #include <linux/security.h>
46 #include <linux/slab.h>
47 #include <linux/spinlock.h>
48 #include <linux/stat.h>
49 #include <linux/string.h>
50 #include <linux/time.h>
51 #include <linux/backing-dev.h>
52 #include <linux/sort.h>
54 #include <asm/uaccess.h>
55 #include <asm/atomic.h>
56 #include <linux/mutex.h>
57 #include <linux/kfifo.h>
58 #include <linux/workqueue.h>
59 #include <linux/cgroup.h>
62 * Tracks how many cpusets are currently defined in system.
63 * When there is only one cpuset (the root cpuset) we can
64 * short circuit some hooks.
66 int number_of_cpusets __read_mostly
;
68 /* Forward declare cgroup structures */
69 struct cgroup_subsys cpuset_subsys
;
72 /* See "Frequency meter" comments, below. */
75 int cnt
; /* unprocessed events count */
76 int val
; /* most recent output value */
77 time_t time
; /* clock (secs) when val computed */
78 spinlock_t lock
; /* guards read or write of above */
82 struct cgroup_subsys_state css
;
84 unsigned long flags
; /* "unsigned long" so bitops work */
85 cpumask_t cpus_allowed
; /* CPUs allowed to tasks in cpuset */
86 nodemask_t mems_allowed
; /* Memory Nodes allowed to tasks */
88 struct cpuset
*parent
; /* my parent */
91 * Copy of global cpuset_mems_generation as of the most
92 * recent time this cpuset changed its mems_allowed.
96 struct fmeter fmeter
; /* memory_pressure filter */
98 /* partition number for rebuild_sched_domains() */
101 /* for custom sched domain */
102 int relax_domain_level
;
104 /* used for walking a cpuset heirarchy */
105 struct list_head stack_list
;
108 /* Retrieve the cpuset for a cgroup */
109 static inline struct cpuset
*cgroup_cs(struct cgroup
*cont
)
111 return container_of(cgroup_subsys_state(cont
, cpuset_subsys_id
),
115 /* Retrieve the cpuset for a task */
116 static inline struct cpuset
*task_cs(struct task_struct
*task
)
118 return container_of(task_subsys_state(task
, cpuset_subsys_id
),
121 struct cpuset_hotplug_scanner
{
122 struct cgroup_scanner scan
;
126 /* bits in struct cpuset flags field */
132 CS_SCHED_LOAD_BALANCE
,
137 /* convenient tests for these bits */
138 static inline int is_cpu_exclusive(const struct cpuset
*cs
)
140 return test_bit(CS_CPU_EXCLUSIVE
, &cs
->flags
);
143 static inline int is_mem_exclusive(const struct cpuset
*cs
)
145 return test_bit(CS_MEM_EXCLUSIVE
, &cs
->flags
);
148 static inline int is_mem_hardwall(const struct cpuset
*cs
)
150 return test_bit(CS_MEM_HARDWALL
, &cs
->flags
);
153 static inline int is_sched_load_balance(const struct cpuset
*cs
)
155 return test_bit(CS_SCHED_LOAD_BALANCE
, &cs
->flags
);
158 static inline int is_memory_migrate(const struct cpuset
*cs
)
160 return test_bit(CS_MEMORY_MIGRATE
, &cs
->flags
);
163 static inline int is_spread_page(const struct cpuset
*cs
)
165 return test_bit(CS_SPREAD_PAGE
, &cs
->flags
);
168 static inline int is_spread_slab(const struct cpuset
*cs
)
170 return test_bit(CS_SPREAD_SLAB
, &cs
->flags
);
174 * Increment this integer everytime any cpuset changes its
175 * mems_allowed value. Users of cpusets can track this generation
176 * number, and avoid having to lock and reload mems_allowed unless
177 * the cpuset they're using changes generation.
179 * A single, global generation is needed because cpuset_attach_task() could
180 * reattach a task to a different cpuset, which must not have its
181 * generation numbers aliased with those of that tasks previous cpuset.
183 * Generations are needed for mems_allowed because one task cannot
184 * modify another's memory placement. So we must enable every task,
185 * on every visit to __alloc_pages(), to efficiently check whether
186 * its current->cpuset->mems_allowed has changed, requiring an update
187 * of its current->mems_allowed.
189 * Since writes to cpuset_mems_generation are guarded by the cgroup lock
190 * there is no need to mark it atomic.
192 static int cpuset_mems_generation
;
194 static struct cpuset top_cpuset
= {
195 .flags
= ((1 << CS_CPU_EXCLUSIVE
) | (1 << CS_MEM_EXCLUSIVE
)),
196 .cpus_allowed
= CPU_MASK_ALL
,
197 .mems_allowed
= NODE_MASK_ALL
,
201 * There are two global mutexes guarding cpuset structures. The first
202 * is the main control groups cgroup_mutex, accessed via
203 * cgroup_lock()/cgroup_unlock(). The second is the cpuset-specific
204 * callback_mutex, below. They can nest. It is ok to first take
205 * cgroup_mutex, then nest callback_mutex. We also require taking
206 * task_lock() when dereferencing a task's cpuset pointer. See "The
207 * task_lock() exception", at the end of this comment.
209 * A task must hold both mutexes to modify cpusets. If a task
210 * holds cgroup_mutex, then it blocks others wanting that mutex,
211 * ensuring that it is the only task able to also acquire callback_mutex
212 * and be able to modify cpusets. It can perform various checks on
213 * the cpuset structure first, knowing nothing will change. It can
214 * also allocate memory while just holding cgroup_mutex. While it is
215 * performing these checks, various callback routines can briefly
216 * acquire callback_mutex to query cpusets. Once it is ready to make
217 * the changes, it takes callback_mutex, blocking everyone else.
219 * Calls to the kernel memory allocator can not be made while holding
220 * callback_mutex, as that would risk double tripping on callback_mutex
221 * from one of the callbacks into the cpuset code from within
224 * If a task is only holding callback_mutex, then it has read-only
227 * The task_struct fields mems_allowed and mems_generation may only
228 * be accessed in the context of that task, so require no locks.
230 * The cpuset_common_file_write handler for operations that modify
231 * the cpuset hierarchy holds cgroup_mutex across the entire operation,
232 * single threading all such cpuset modifications across the system.
234 * The cpuset_common_file_read() handlers only hold callback_mutex across
235 * small pieces of code, such as when reading out possibly multi-word
236 * cpumasks and nodemasks.
238 * Accessing a task's cpuset should be done in accordance with the
239 * guidelines for accessing subsystem state in kernel/cgroup.c
242 static DEFINE_MUTEX(callback_mutex
);
244 /* This is ugly, but preserves the userspace API for existing cpuset
245 * users. If someone tries to mount the "cpuset" filesystem, we
246 * silently switch it to mount "cgroup" instead */
247 static int cpuset_get_sb(struct file_system_type
*fs_type
,
248 int flags
, const char *unused_dev_name
,
249 void *data
, struct vfsmount
*mnt
)
251 struct file_system_type
*cgroup_fs
= get_fs_type("cgroup");
256 "release_agent=/sbin/cpuset_release_agent";
257 ret
= cgroup_fs
->get_sb(cgroup_fs
, flags
,
258 unused_dev_name
, mountopts
, mnt
);
259 put_filesystem(cgroup_fs
);
264 static struct file_system_type cpuset_fs_type
= {
266 .get_sb
= cpuset_get_sb
,
270 * Return in *pmask the portion of a cpusets's cpus_allowed that
271 * are online. If none are online, walk up the cpuset hierarchy
272 * until we find one that does have some online cpus. If we get
273 * all the way to the top and still haven't found any online cpus,
274 * return cpu_online_map. Or if passed a NULL cs from an exit'ing
275 * task, return cpu_online_map.
277 * One way or another, we guarantee to return some non-empty subset
280 * Call with callback_mutex held.
283 static void guarantee_online_cpus(const struct cpuset
*cs
, cpumask_t
*pmask
)
285 while (cs
&& !cpus_intersects(cs
->cpus_allowed
, cpu_online_map
))
288 cpus_and(*pmask
, cs
->cpus_allowed
, cpu_online_map
);
290 *pmask
= cpu_online_map
;
291 BUG_ON(!cpus_intersects(*pmask
, cpu_online_map
));
295 * Return in *pmask the portion of a cpusets's mems_allowed that
296 * are online, with memory. If none are online with memory, walk
297 * up the cpuset hierarchy until we find one that does have some
298 * online mems. If we get all the way to the top and still haven't
299 * found any online mems, return node_states[N_HIGH_MEMORY].
301 * One way or another, we guarantee to return some non-empty subset
302 * of node_states[N_HIGH_MEMORY].
304 * Call with callback_mutex held.
307 static void guarantee_online_mems(const struct cpuset
*cs
, nodemask_t
*pmask
)
309 while (cs
&& !nodes_intersects(cs
->mems_allowed
,
310 node_states
[N_HIGH_MEMORY
]))
313 nodes_and(*pmask
, cs
->mems_allowed
,
314 node_states
[N_HIGH_MEMORY
]);
316 *pmask
= node_states
[N_HIGH_MEMORY
];
317 BUG_ON(!nodes_intersects(*pmask
, node_states
[N_HIGH_MEMORY
]));
321 * cpuset_update_task_memory_state - update task memory placement
323 * If the current tasks cpusets mems_allowed changed behind our
324 * backs, update current->mems_allowed, mems_generation and task NUMA
325 * mempolicy to the new value.
327 * Task mempolicy is updated by rebinding it relative to the
328 * current->cpuset if a task has its memory placement changed.
329 * Do not call this routine if in_interrupt().
331 * Call without callback_mutex or task_lock() held. May be
332 * called with or without cgroup_mutex held. Thanks in part to
333 * 'the_top_cpuset_hack', the task's cpuset pointer will never
334 * be NULL. This routine also might acquire callback_mutex during
337 * Reading current->cpuset->mems_generation doesn't need task_lock
338 * to guard the current->cpuset derefence, because it is guarded
339 * from concurrent freeing of current->cpuset using RCU.
341 * The rcu_dereference() is technically probably not needed,
342 * as I don't actually mind if I see a new cpuset pointer but
343 * an old value of mems_generation. However this really only
344 * matters on alpha systems using cpusets heavily. If I dropped
345 * that rcu_dereference(), it would save them a memory barrier.
346 * For all other arch's, rcu_dereference is a no-op anyway, and for
347 * alpha systems not using cpusets, another planned optimization,
348 * avoiding the rcu critical section for tasks in the root cpuset
349 * which is statically allocated, so can't vanish, will make this
350 * irrelevant. Better to use RCU as intended, than to engage in
351 * some cute trick to save a memory barrier that is impossible to
352 * test, for alpha systems using cpusets heavily, which might not
355 * This routine is needed to update the per-task mems_allowed data,
356 * within the tasks context, when it is trying to allocate memory
357 * (in various mm/mempolicy.c routines) and notices that some other
358 * task has been modifying its cpuset.
361 void cpuset_update_task_memory_state(void)
363 int my_cpusets_mem_gen
;
364 struct task_struct
*tsk
= current
;
367 if (task_cs(tsk
) == &top_cpuset
) {
368 /* Don't need rcu for top_cpuset. It's never freed. */
369 my_cpusets_mem_gen
= top_cpuset
.mems_generation
;
372 my_cpusets_mem_gen
= task_cs(current
)->mems_generation
;
376 if (my_cpusets_mem_gen
!= tsk
->cpuset_mems_generation
) {
377 mutex_lock(&callback_mutex
);
379 cs
= task_cs(tsk
); /* Maybe changed when task not locked */
380 guarantee_online_mems(cs
, &tsk
->mems_allowed
);
381 tsk
->cpuset_mems_generation
= cs
->mems_generation
;
382 if (is_spread_page(cs
))
383 tsk
->flags
|= PF_SPREAD_PAGE
;
385 tsk
->flags
&= ~PF_SPREAD_PAGE
;
386 if (is_spread_slab(cs
))
387 tsk
->flags
|= PF_SPREAD_SLAB
;
389 tsk
->flags
&= ~PF_SPREAD_SLAB
;
391 mutex_unlock(&callback_mutex
);
392 mpol_rebind_task(tsk
, &tsk
->mems_allowed
);
397 * is_cpuset_subset(p, q) - Is cpuset p a subset of cpuset q?
399 * One cpuset is a subset of another if all its allowed CPUs and
400 * Memory Nodes are a subset of the other, and its exclusive flags
401 * are only set if the other's are set. Call holding cgroup_mutex.
404 static int is_cpuset_subset(const struct cpuset
*p
, const struct cpuset
*q
)
406 return cpus_subset(p
->cpus_allowed
, q
->cpus_allowed
) &&
407 nodes_subset(p
->mems_allowed
, q
->mems_allowed
) &&
408 is_cpu_exclusive(p
) <= is_cpu_exclusive(q
) &&
409 is_mem_exclusive(p
) <= is_mem_exclusive(q
);
413 * validate_change() - Used to validate that any proposed cpuset change
414 * follows the structural rules for cpusets.
416 * If we replaced the flag and mask values of the current cpuset
417 * (cur) with those values in the trial cpuset (trial), would
418 * our various subset and exclusive rules still be valid? Presumes
421 * 'cur' is the address of an actual, in-use cpuset. Operations
422 * such as list traversal that depend on the actual address of the
423 * cpuset in the list must use cur below, not trial.
425 * 'trial' is the address of bulk structure copy of cur, with
426 * perhaps one or more of the fields cpus_allowed, mems_allowed,
427 * or flags changed to new, trial values.
429 * Return 0 if valid, -errno if not.
432 static int validate_change(const struct cpuset
*cur
, const struct cpuset
*trial
)
435 struct cpuset
*c
, *par
;
437 /* Each of our child cpusets must be a subset of us */
438 list_for_each_entry(cont
, &cur
->css
.cgroup
->children
, sibling
) {
439 if (!is_cpuset_subset(cgroup_cs(cont
), trial
))
443 /* Remaining checks don't apply to root cpuset */
444 if (cur
== &top_cpuset
)
449 /* We must be a subset of our parent cpuset */
450 if (!is_cpuset_subset(trial
, par
))
454 * If either I or some sibling (!= me) is exclusive, we can't
457 list_for_each_entry(cont
, &par
->css
.cgroup
->children
, sibling
) {
459 if ((is_cpu_exclusive(trial
) || is_cpu_exclusive(c
)) &&
461 cpus_intersects(trial
->cpus_allowed
, c
->cpus_allowed
))
463 if ((is_mem_exclusive(trial
) || is_mem_exclusive(c
)) &&
465 nodes_intersects(trial
->mems_allowed
, c
->mems_allowed
))
469 /* Cpusets with tasks can't have empty cpus_allowed or mems_allowed */
470 if (cgroup_task_count(cur
->css
.cgroup
)) {
471 if (cpus_empty(trial
->cpus_allowed
) ||
472 nodes_empty(trial
->mems_allowed
)) {
481 * Helper routine for rebuild_sched_domains().
482 * Do cpusets a, b have overlapping cpus_allowed masks?
485 static int cpusets_overlap(struct cpuset
*a
, struct cpuset
*b
)
487 return cpus_intersects(a
->cpus_allowed
, b
->cpus_allowed
);
491 update_domain_attr(struct sched_domain_attr
*dattr
, struct cpuset
*c
)
495 if (dattr
->relax_domain_level
< c
->relax_domain_level
)
496 dattr
->relax_domain_level
= c
->relax_domain_level
;
501 * rebuild_sched_domains()
503 * If the flag 'sched_load_balance' of any cpuset with non-empty
504 * 'cpus' changes, or if the 'cpus' allowed changes in any cpuset
505 * which has that flag enabled, or if any cpuset with a non-empty
506 * 'cpus' is removed, then call this routine to rebuild the
507 * scheduler's dynamic sched domains.
509 * This routine builds a partial partition of the systems CPUs
510 * (the set of non-overlappping cpumask_t's in the array 'part'
511 * below), and passes that partial partition to the kernel/sched.c
512 * partition_sched_domains() routine, which will rebuild the
513 * schedulers load balancing domains (sched domains) as specified
514 * by that partial partition. A 'partial partition' is a set of
515 * non-overlapping subsets whose union is a subset of that set.
517 * See "What is sched_load_balance" in Documentation/cpusets.txt
518 * for a background explanation of this.
520 * Does not return errors, on the theory that the callers of this
521 * routine would rather not worry about failures to rebuild sched
522 * domains when operating in the severe memory shortage situations
523 * that could cause allocation failures below.
525 * Call with cgroup_mutex held. May take callback_mutex during
526 * call due to the kfifo_alloc() and kmalloc() calls. May nest
527 * a call to the get_online_cpus()/put_online_cpus() pair.
528 * Must not be called holding callback_mutex, because we must not
529 * call get_online_cpus() while holding callback_mutex. Elsewhere
530 * the kernel nests callback_mutex inside get_online_cpus() calls.
531 * So the reverse nesting would risk an ABBA deadlock.
533 * The three key local variables below are:
534 * q - a kfifo queue of cpuset pointers, used to implement a
535 * top-down scan of all cpusets. This scan loads a pointer
536 * to each cpuset marked is_sched_load_balance into the
537 * array 'csa'. For our purposes, rebuilding the schedulers
538 * sched domains, we can ignore !is_sched_load_balance cpusets.
539 * csa - (for CpuSet Array) Array of pointers to all the cpusets
540 * that need to be load balanced, for convenient iterative
541 * access by the subsequent code that finds the best partition,
542 * i.e the set of domains (subsets) of CPUs such that the
543 * cpus_allowed of every cpuset marked is_sched_load_balance
544 * is a subset of one of these domains, while there are as
545 * many such domains as possible, each as small as possible.
546 * doms - Conversion of 'csa' to an array of cpumasks, for passing to
547 * the kernel/sched.c routine partition_sched_domains() in a
548 * convenient format, that can be easily compared to the prior
549 * value to determine what partition elements (sched domains)
550 * were changed (added or removed.)
552 * Finding the best partition (set of domains):
553 * The triple nested loops below over i, j, k scan over the
554 * load balanced cpusets (using the array of cpuset pointers in
555 * csa[]) looking for pairs of cpusets that have overlapping
556 * cpus_allowed, but which don't have the same 'pn' partition
557 * number and gives them in the same partition number. It keeps
558 * looping on the 'restart' label until it can no longer find
561 * The union of the cpus_allowed masks from the set of
562 * all cpusets having the same 'pn' value then form the one
563 * element of the partition (one sched domain) to be passed to
564 * partition_sched_domains().
567 static void rebuild_sched_domains(void)
569 struct kfifo
*q
; /* queue of cpusets to be scanned */
570 struct cpuset
*cp
; /* scans q */
571 struct cpuset
**csa
; /* array of all cpuset ptrs */
572 int csn
; /* how many cpuset ptrs in csa so far */
573 int i
, j
, k
; /* indices for partition finding loops */
574 cpumask_t
*doms
; /* resulting partition; i.e. sched domains */
575 struct sched_domain_attr
*dattr
; /* attributes for custom domains */
576 int ndoms
; /* number of sched domains in result */
577 int nslot
; /* next empty doms[] cpumask_t slot */
584 /* Special case for the 99% of systems with one, full, sched domain */
585 if (is_sched_load_balance(&top_cpuset
)) {
587 doms
= kmalloc(sizeof(cpumask_t
), GFP_KERNEL
);
590 dattr
= kmalloc(sizeof(struct sched_domain_attr
), GFP_KERNEL
);
592 *dattr
= SD_ATTR_INIT
;
593 update_domain_attr(dattr
, &top_cpuset
);
595 *doms
= top_cpuset
.cpus_allowed
;
599 q
= kfifo_alloc(number_of_cpusets
* sizeof(cp
), GFP_KERNEL
, NULL
);
602 csa
= kmalloc(number_of_cpusets
* sizeof(cp
), GFP_KERNEL
);
608 __kfifo_put(q
, (void *)&cp
, sizeof(cp
));
609 while (__kfifo_get(q
, (void *)&cp
, sizeof(cp
))) {
611 struct cpuset
*child
; /* scans child cpusets of cp */
612 if (is_sched_load_balance(cp
))
614 list_for_each_entry(cont
, &cp
->css
.cgroup
->children
, sibling
) {
615 child
= cgroup_cs(cont
);
616 __kfifo_put(q
, (void *)&child
, sizeof(cp
));
620 for (i
= 0; i
< csn
; i
++)
625 /* Find the best partition (set of sched domains) */
626 for (i
= 0; i
< csn
; i
++) {
627 struct cpuset
*a
= csa
[i
];
630 for (j
= 0; j
< csn
; j
++) {
631 struct cpuset
*b
= csa
[j
];
634 if (apn
!= bpn
&& cpusets_overlap(a
, b
)) {
635 for (k
= 0; k
< csn
; k
++) {
636 struct cpuset
*c
= csa
[k
];
641 ndoms
--; /* one less element */
647 /* Convert <csn, csa> to <ndoms, doms> */
648 doms
= kmalloc(ndoms
* sizeof(cpumask_t
), GFP_KERNEL
);
651 dattr
= kmalloc(ndoms
* sizeof(struct sched_domain_attr
), GFP_KERNEL
);
653 for (nslot
= 0, i
= 0; i
< csn
; i
++) {
654 struct cpuset
*a
= csa
[i
];
658 cpumask_t
*dp
= doms
+ nslot
;
660 if (nslot
== ndoms
) {
661 static int warnings
= 10;
664 "rebuild_sched_domains confused:"
665 " nslot %d, ndoms %d, csn %d, i %d,"
667 nslot
, ndoms
, csn
, i
, apn
);
675 *(dattr
+ nslot
) = SD_ATTR_INIT
;
676 for (j
= i
; j
< csn
; j
++) {
677 struct cpuset
*b
= csa
[j
];
680 cpus_or(*dp
, *dp
, b
->cpus_allowed
);
682 update_domain_attr(dattr
, b
);
688 BUG_ON(nslot
!= ndoms
);
691 /* Have scheduler rebuild sched domains */
693 partition_sched_domains(ndoms
, doms
, dattr
);
700 /* Don't kfree(doms) -- partition_sched_domains() does that. */
701 /* Don't kfree(dattr) -- partition_sched_domains() does that. */
704 static inline int started_after_time(struct task_struct
*t1
,
705 struct timespec
*time
,
706 struct task_struct
*t2
)
708 int start_diff
= timespec_compare(&t1
->start_time
, time
);
709 if (start_diff
> 0) {
711 } else if (start_diff
< 0) {
715 * Arbitrarily, if two processes started at the same
716 * time, we'll say that the lower pointer value
717 * started first. Note that t2 may have exited by now
718 * so this may not be a valid pointer any longer, but
719 * that's fine - it still serves to distinguish
720 * between two tasks started (effectively)
727 static inline int started_after(void *p1
, void *p2
)
729 struct task_struct
*t1
= p1
;
730 struct task_struct
*t2
= p2
;
731 return started_after_time(t1
, &t2
->start_time
, t2
);
735 * cpuset_test_cpumask - test a task's cpus_allowed versus its cpuset's
737 * @scan: struct cgroup_scanner contained in its struct cpuset_hotplug_scanner
739 * Call with cgroup_mutex held. May take callback_mutex during call.
740 * Called for each task in a cgroup by cgroup_scan_tasks().
741 * Return nonzero if this tasks's cpus_allowed mask should be changed (in other
742 * words, if its mask is not equal to its cpuset's mask).
744 static int cpuset_test_cpumask(struct task_struct
*tsk
,
745 struct cgroup_scanner
*scan
)
747 return !cpus_equal(tsk
->cpus_allowed
,
748 (cgroup_cs(scan
->cg
))->cpus_allowed
);
752 * cpuset_change_cpumask - make a task's cpus_allowed the same as its cpuset's
754 * @scan: struct cgroup_scanner containing the cgroup of the task
756 * Called by cgroup_scan_tasks() for each task in a cgroup whose
757 * cpus_allowed mask needs to be changed.
759 * We don't need to re-check for the cgroup/cpuset membership, since we're
760 * holding cgroup_lock() at this point.
762 static void cpuset_change_cpumask(struct task_struct
*tsk
,
763 struct cgroup_scanner
*scan
)
765 set_cpus_allowed_ptr(tsk
, &((cgroup_cs(scan
->cg
))->cpus_allowed
));
769 * update_cpumask - update the cpus_allowed mask of a cpuset and all tasks in it
770 * @cs: the cpuset to consider
771 * @buf: buffer of cpu numbers written to this cpuset
773 static int update_cpumask(struct cpuset
*cs
, char *buf
)
775 struct cpuset trialcs
;
776 struct cgroup_scanner scan
;
777 struct ptr_heap heap
;
779 int is_load_balanced
;
781 /* top_cpuset.cpus_allowed tracks cpu_online_map; it's read-only */
782 if (cs
== &top_cpuset
)
788 * An empty cpus_allowed is ok only if the cpuset has no tasks.
789 * Since cpulist_parse() fails on an empty mask, we special case
790 * that parsing. The validate_change() call ensures that cpusets
791 * with tasks have cpus.
795 cpus_clear(trialcs
.cpus_allowed
);
797 retval
= cpulist_parse(buf
, trialcs
.cpus_allowed
);
801 cpus_and(trialcs
.cpus_allowed
, trialcs
.cpus_allowed
, cpu_online_map
);
802 retval
= validate_change(cs
, &trialcs
);
806 /* Nothing to do if the cpus didn't change */
807 if (cpus_equal(cs
->cpus_allowed
, trialcs
.cpus_allowed
))
810 retval
= heap_init(&heap
, PAGE_SIZE
, GFP_KERNEL
, &started_after
);
814 is_load_balanced
= is_sched_load_balance(&trialcs
);
816 mutex_lock(&callback_mutex
);
817 cs
->cpus_allowed
= trialcs
.cpus_allowed
;
818 mutex_unlock(&callback_mutex
);
821 * Scan tasks in the cpuset, and update the cpumasks of any
822 * that need an update.
824 scan
.cg
= cs
->css
.cgroup
;
825 scan
.test_task
= cpuset_test_cpumask
;
826 scan
.process_task
= cpuset_change_cpumask
;
828 cgroup_scan_tasks(&scan
);
831 if (is_load_balanced
)
832 rebuild_sched_domains();
839 * Migrate memory region from one set of nodes to another.
841 * Temporarilly set tasks mems_allowed to target nodes of migration,
842 * so that the migration code can allocate pages on these nodes.
844 * Call holding cgroup_mutex, so current's cpuset won't change
845 * during this call, as manage_mutex holds off any cpuset_attach()
846 * calls. Therefore we don't need to take task_lock around the
847 * call to guarantee_online_mems(), as we know no one is changing
850 * Hold callback_mutex around the two modifications of our tasks
851 * mems_allowed to synchronize with cpuset_mems_allowed().
853 * While the mm_struct we are migrating is typically from some
854 * other task, the task_struct mems_allowed that we are hacking
855 * is for our current task, which must allocate new pages for that
856 * migrating memory region.
858 * We call cpuset_update_task_memory_state() before hacking
859 * our tasks mems_allowed, so that we are assured of being in
860 * sync with our tasks cpuset, and in particular, callbacks to
861 * cpuset_update_task_memory_state() from nested page allocations
862 * won't see any mismatch of our cpuset and task mems_generation
863 * values, so won't overwrite our hacked tasks mems_allowed
867 static void cpuset_migrate_mm(struct mm_struct
*mm
, const nodemask_t
*from
,
868 const nodemask_t
*to
)
870 struct task_struct
*tsk
= current
;
872 cpuset_update_task_memory_state();
874 mutex_lock(&callback_mutex
);
875 tsk
->mems_allowed
= *to
;
876 mutex_unlock(&callback_mutex
);
878 do_migrate_pages(mm
, from
, to
, MPOL_MF_MOVE_ALL
);
880 mutex_lock(&callback_mutex
);
881 guarantee_online_mems(task_cs(tsk
),&tsk
->mems_allowed
);
882 mutex_unlock(&callback_mutex
);
886 * Handle user request to change the 'mems' memory placement
887 * of a cpuset. Needs to validate the request, update the
888 * cpusets mems_allowed and mems_generation, and for each
889 * task in the cpuset, rebind any vma mempolicies and if
890 * the cpuset is marked 'memory_migrate', migrate the tasks
891 * pages to the new memory.
893 * Call with cgroup_mutex held. May take callback_mutex during call.
894 * Will take tasklist_lock, scan tasklist for tasks in cpuset cs,
895 * lock each such tasks mm->mmap_sem, scan its vma's and rebind
896 * their mempolicies to the cpusets new mems_allowed.
899 static void *cpuset_being_rebound
;
901 static int update_nodemask(struct cpuset
*cs
, char *buf
)
903 struct cpuset trialcs
;
905 struct task_struct
*p
;
906 struct mm_struct
**mmarray
;
911 struct cgroup_iter it
;
914 * top_cpuset.mems_allowed tracks node_stats[N_HIGH_MEMORY];
917 if (cs
== &top_cpuset
)
923 * An empty mems_allowed is ok iff there are no tasks in the cpuset.
924 * Since nodelist_parse() fails on an empty mask, we special case
925 * that parsing. The validate_change() call ensures that cpusets
926 * with tasks have memory.
930 nodes_clear(trialcs
.mems_allowed
);
932 retval
= nodelist_parse(buf
, trialcs
.mems_allowed
);
936 nodes_and(trialcs
.mems_allowed
, trialcs
.mems_allowed
,
937 node_states
[N_HIGH_MEMORY
]);
938 oldmem
= cs
->mems_allowed
;
939 if (nodes_equal(oldmem
, trialcs
.mems_allowed
)) {
940 retval
= 0; /* Too easy - nothing to do */
943 retval
= validate_change(cs
, &trialcs
);
947 mutex_lock(&callback_mutex
);
948 cs
->mems_allowed
= trialcs
.mems_allowed
;
949 cs
->mems_generation
= cpuset_mems_generation
++;
950 mutex_unlock(&callback_mutex
);
952 cpuset_being_rebound
= cs
; /* causes mpol_dup() rebind */
954 fudge
= 10; /* spare mmarray[] slots */
955 fudge
+= cpus_weight(cs
->cpus_allowed
); /* imagine one fork-bomb/cpu */
959 * Allocate mmarray[] to hold mm reference for each task
960 * in cpuset cs. Can't kmalloc GFP_KERNEL while holding
961 * tasklist_lock. We could use GFP_ATOMIC, but with a
962 * few more lines of code, we can retry until we get a big
963 * enough mmarray[] w/o using GFP_ATOMIC.
966 ntasks
= cgroup_task_count(cs
->css
.cgroup
); /* guess */
968 mmarray
= kmalloc(ntasks
* sizeof(*mmarray
), GFP_KERNEL
);
971 read_lock(&tasklist_lock
); /* block fork */
972 if (cgroup_task_count(cs
->css
.cgroup
) <= ntasks
)
973 break; /* got enough */
974 read_unlock(&tasklist_lock
); /* try again */
980 /* Load up mmarray[] with mm reference for each task in cpuset. */
981 cgroup_iter_start(cs
->css
.cgroup
, &it
);
982 while ((p
= cgroup_iter_next(cs
->css
.cgroup
, &it
))) {
983 struct mm_struct
*mm
;
987 "Cpuset mempolicy rebind incomplete.\n");
995 cgroup_iter_end(cs
->css
.cgroup
, &it
);
996 read_unlock(&tasklist_lock
);
999 * Now that we've dropped the tasklist spinlock, we can
1000 * rebind the vma mempolicies of each mm in mmarray[] to their
1001 * new cpuset, and release that mm. The mpol_rebind_mm()
1002 * call takes mmap_sem, which we couldn't take while holding
1003 * tasklist_lock. Forks can happen again now - the mpol_dup()
1004 * cpuset_being_rebound check will catch such forks, and rebind
1005 * their vma mempolicies too. Because we still hold the global
1006 * cgroup_mutex, we know that no other rebind effort will
1007 * be contending for the global variable cpuset_being_rebound.
1008 * It's ok if we rebind the same mm twice; mpol_rebind_mm()
1009 * is idempotent. Also migrate pages in each mm to new nodes.
1011 migrate
= is_memory_migrate(cs
);
1012 for (i
= 0; i
< n
; i
++) {
1013 struct mm_struct
*mm
= mmarray
[i
];
1015 mpol_rebind_mm(mm
, &cs
->mems_allowed
);
1017 cpuset_migrate_mm(mm
, &oldmem
, &cs
->mems_allowed
);
1021 /* We're done rebinding vmas to this cpuset's new mems_allowed. */
1023 cpuset_being_rebound
= NULL
;
1029 int current_cpuset_is_being_rebound(void)
1031 return task_cs(current
) == cpuset_being_rebound
;
1034 static int update_relax_domain_level(struct cpuset
*cs
, s64 val
)
1039 if (val
!= cs
->relax_domain_level
) {
1040 cs
->relax_domain_level
= val
;
1041 rebuild_sched_domains();
1048 * update_flag - read a 0 or a 1 in a file and update associated flag
1049 * bit: the bit to update (see cpuset_flagbits_t)
1050 * cs: the cpuset to update
1051 * turning_on: whether the flag is being set or cleared
1053 * Call with cgroup_mutex held.
1056 static int update_flag(cpuset_flagbits_t bit
, struct cpuset
*cs
,
1059 struct cpuset trialcs
;
1061 int cpus_nonempty
, balance_flag_changed
;
1065 set_bit(bit
, &trialcs
.flags
);
1067 clear_bit(bit
, &trialcs
.flags
);
1069 err
= validate_change(cs
, &trialcs
);
1073 cpus_nonempty
= !cpus_empty(trialcs
.cpus_allowed
);
1074 balance_flag_changed
= (is_sched_load_balance(cs
) !=
1075 is_sched_load_balance(&trialcs
));
1077 mutex_lock(&callback_mutex
);
1078 cs
->flags
= trialcs
.flags
;
1079 mutex_unlock(&callback_mutex
);
1081 if (cpus_nonempty
&& balance_flag_changed
)
1082 rebuild_sched_domains();
1088 * Frequency meter - How fast is some event occurring?
1090 * These routines manage a digitally filtered, constant time based,
1091 * event frequency meter. There are four routines:
1092 * fmeter_init() - initialize a frequency meter.
1093 * fmeter_markevent() - called each time the event happens.
1094 * fmeter_getrate() - returns the recent rate of such events.
1095 * fmeter_update() - internal routine used to update fmeter.
1097 * A common data structure is passed to each of these routines,
1098 * which is used to keep track of the state required to manage the
1099 * frequency meter and its digital filter.
1101 * The filter works on the number of events marked per unit time.
1102 * The filter is single-pole low-pass recursive (IIR). The time unit
1103 * is 1 second. Arithmetic is done using 32-bit integers scaled to
1104 * simulate 3 decimal digits of precision (multiplied by 1000).
1106 * With an FM_COEF of 933, and a time base of 1 second, the filter
1107 * has a half-life of 10 seconds, meaning that if the events quit
1108 * happening, then the rate returned from the fmeter_getrate()
1109 * will be cut in half each 10 seconds, until it converges to zero.
1111 * It is not worth doing a real infinitely recursive filter. If more
1112 * than FM_MAXTICKS ticks have elapsed since the last filter event,
1113 * just compute FM_MAXTICKS ticks worth, by which point the level
1116 * Limit the count of unprocessed events to FM_MAXCNT, so as to avoid
1117 * arithmetic overflow in the fmeter_update() routine.
1119 * Given the simple 32 bit integer arithmetic used, this meter works
1120 * best for reporting rates between one per millisecond (msec) and
1121 * one per 32 (approx) seconds. At constant rates faster than one
1122 * per msec it maxes out at values just under 1,000,000. At constant
1123 * rates between one per msec, and one per second it will stabilize
1124 * to a value N*1000, where N is the rate of events per second.
1125 * At constant rates between one per second and one per 32 seconds,
1126 * it will be choppy, moving up on the seconds that have an event,
1127 * and then decaying until the next event. At rates slower than
1128 * about one in 32 seconds, it decays all the way back to zero between
1132 #define FM_COEF 933 /* coefficient for half-life of 10 secs */
1133 #define FM_MAXTICKS ((time_t)99) /* useless computing more ticks than this */
1134 #define FM_MAXCNT 1000000 /* limit cnt to avoid overflow */
1135 #define FM_SCALE 1000 /* faux fixed point scale */
1137 /* Initialize a frequency meter */
1138 static void fmeter_init(struct fmeter
*fmp
)
1143 spin_lock_init(&fmp
->lock
);
1146 /* Internal meter update - process cnt events and update value */
1147 static void fmeter_update(struct fmeter
*fmp
)
1149 time_t now
= get_seconds();
1150 time_t ticks
= now
- fmp
->time
;
1155 ticks
= min(FM_MAXTICKS
, ticks
);
1157 fmp
->val
= (FM_COEF
* fmp
->val
) / FM_SCALE
;
1160 fmp
->val
+= ((FM_SCALE
- FM_COEF
) * fmp
->cnt
) / FM_SCALE
;
1164 /* Process any previous ticks, then bump cnt by one (times scale). */
1165 static void fmeter_markevent(struct fmeter
*fmp
)
1167 spin_lock(&fmp
->lock
);
1169 fmp
->cnt
= min(FM_MAXCNT
, fmp
->cnt
+ FM_SCALE
);
1170 spin_unlock(&fmp
->lock
);
1173 /* Process any previous ticks, then return current value. */
1174 static int fmeter_getrate(struct fmeter
*fmp
)
1178 spin_lock(&fmp
->lock
);
1181 spin_unlock(&fmp
->lock
);
1185 /* Called by cgroups to determine if a cpuset is usable; cgroup_mutex held */
1186 static int cpuset_can_attach(struct cgroup_subsys
*ss
,
1187 struct cgroup
*cont
, struct task_struct
*tsk
)
1189 struct cpuset
*cs
= cgroup_cs(cont
);
1191 if (cpus_empty(cs
->cpus_allowed
) || nodes_empty(cs
->mems_allowed
))
1193 if (tsk
->flags
& PF_THREAD_BOUND
) {
1196 mutex_lock(&callback_mutex
);
1197 mask
= cs
->cpus_allowed
;
1198 mutex_unlock(&callback_mutex
);
1199 if (!cpus_equal(tsk
->cpus_allowed
, mask
))
1203 return security_task_setscheduler(tsk
, 0, NULL
);
1206 static void cpuset_attach(struct cgroup_subsys
*ss
,
1207 struct cgroup
*cont
, struct cgroup
*oldcont
,
1208 struct task_struct
*tsk
)
1211 nodemask_t from
, to
;
1212 struct mm_struct
*mm
;
1213 struct cpuset
*cs
= cgroup_cs(cont
);
1214 struct cpuset
*oldcs
= cgroup_cs(oldcont
);
1217 mutex_lock(&callback_mutex
);
1218 guarantee_online_cpus(cs
, &cpus
);
1219 err
= set_cpus_allowed_ptr(tsk
, &cpus
);
1220 mutex_unlock(&callback_mutex
);
1224 from
= oldcs
->mems_allowed
;
1225 to
= cs
->mems_allowed
;
1226 mm
= get_task_mm(tsk
);
1228 mpol_rebind_mm(mm
, &to
);
1229 if (is_memory_migrate(cs
))
1230 cpuset_migrate_mm(mm
, &from
, &to
);
1236 /* The various types of files and directories in a cpuset file system */
1239 FILE_MEMORY_MIGRATE
,
1245 FILE_SCHED_LOAD_BALANCE
,
1246 FILE_SCHED_RELAX_DOMAIN_LEVEL
,
1247 FILE_MEMORY_PRESSURE_ENABLED
,
1248 FILE_MEMORY_PRESSURE
,
1251 } cpuset_filetype_t
;
1253 static ssize_t
cpuset_common_file_write(struct cgroup
*cont
,
1256 const char __user
*userbuf
,
1257 size_t nbytes
, loff_t
*unused_ppos
)
1259 struct cpuset
*cs
= cgroup_cs(cont
);
1260 cpuset_filetype_t type
= cft
->private;
1264 /* Crude upper limit on largest legitimate cpulist user might write. */
1265 if (nbytes
> 100U + 6 * max(NR_CPUS
, MAX_NUMNODES
))
1268 /* +1 for nul-terminator */
1269 buffer
= kmalloc(nbytes
+ 1, GFP_KERNEL
);
1273 if (copy_from_user(buffer
, userbuf
, nbytes
)) {
1277 buffer
[nbytes
] = 0; /* nul-terminate */
1281 if (cgroup_is_removed(cont
)) {
1288 retval
= update_cpumask(cs
, buffer
);
1291 retval
= update_nodemask(cs
, buffer
);
1307 static int cpuset_write_u64(struct cgroup
*cgrp
, struct cftype
*cft
, u64 val
)
1310 struct cpuset
*cs
= cgroup_cs(cgrp
);
1311 cpuset_filetype_t type
= cft
->private;
1315 if (cgroup_is_removed(cgrp
)) {
1321 case FILE_CPU_EXCLUSIVE
:
1322 retval
= update_flag(CS_CPU_EXCLUSIVE
, cs
, val
);
1324 case FILE_MEM_EXCLUSIVE
:
1325 retval
= update_flag(CS_MEM_EXCLUSIVE
, cs
, val
);
1327 case FILE_MEM_HARDWALL
:
1328 retval
= update_flag(CS_MEM_HARDWALL
, cs
, val
);
1330 case FILE_SCHED_LOAD_BALANCE
:
1331 retval
= update_flag(CS_SCHED_LOAD_BALANCE
, cs
, val
);
1333 case FILE_MEMORY_MIGRATE
:
1334 retval
= update_flag(CS_MEMORY_MIGRATE
, cs
, val
);
1336 case FILE_MEMORY_PRESSURE_ENABLED
:
1337 cpuset_memory_pressure_enabled
= !!val
;
1339 case FILE_MEMORY_PRESSURE
:
1342 case FILE_SPREAD_PAGE
:
1343 retval
= update_flag(CS_SPREAD_PAGE
, cs
, val
);
1344 cs
->mems_generation
= cpuset_mems_generation
++;
1346 case FILE_SPREAD_SLAB
:
1347 retval
= update_flag(CS_SPREAD_SLAB
, cs
, val
);
1348 cs
->mems_generation
= cpuset_mems_generation
++;
1358 static int cpuset_write_s64(struct cgroup
*cgrp
, struct cftype
*cft
, s64 val
)
1361 struct cpuset
*cs
= cgroup_cs(cgrp
);
1362 cpuset_filetype_t type
= cft
->private;
1366 if (cgroup_is_removed(cgrp
)) {
1371 case FILE_SCHED_RELAX_DOMAIN_LEVEL
:
1372 retval
= update_relax_domain_level(cs
, val
);
1383 * These ascii lists should be read in a single call, by using a user
1384 * buffer large enough to hold the entire map. If read in smaller
1385 * chunks, there is no guarantee of atomicity. Since the display format
1386 * used, list of ranges of sequential numbers, is variable length,
1387 * and since these maps can change value dynamically, one could read
1388 * gibberish by doing partial reads while a list was changing.
1389 * A single large read to a buffer that crosses a page boundary is
1390 * ok, because the result being copied to user land is not recomputed
1391 * across a page fault.
1394 static int cpuset_sprintf_cpulist(char *page
, struct cpuset
*cs
)
1398 mutex_lock(&callback_mutex
);
1399 mask
= cs
->cpus_allowed
;
1400 mutex_unlock(&callback_mutex
);
1402 return cpulist_scnprintf(page
, PAGE_SIZE
, mask
);
1405 static int cpuset_sprintf_memlist(char *page
, struct cpuset
*cs
)
1409 mutex_lock(&callback_mutex
);
1410 mask
= cs
->mems_allowed
;
1411 mutex_unlock(&callback_mutex
);
1413 return nodelist_scnprintf(page
, PAGE_SIZE
, mask
);
1416 static ssize_t
cpuset_common_file_read(struct cgroup
*cont
,
1420 size_t nbytes
, loff_t
*ppos
)
1422 struct cpuset
*cs
= cgroup_cs(cont
);
1423 cpuset_filetype_t type
= cft
->private;
1428 if (!(page
= (char *)__get_free_page(GFP_TEMPORARY
)))
1435 s
+= cpuset_sprintf_cpulist(s
, cs
);
1438 s
+= cpuset_sprintf_memlist(s
, cs
);
1446 retval
= simple_read_from_buffer(buf
, nbytes
, ppos
, page
, s
- page
);
1448 free_page((unsigned long)page
);
1452 static u64
cpuset_read_u64(struct cgroup
*cont
, struct cftype
*cft
)
1454 struct cpuset
*cs
= cgroup_cs(cont
);
1455 cpuset_filetype_t type
= cft
->private;
1457 case FILE_CPU_EXCLUSIVE
:
1458 return is_cpu_exclusive(cs
);
1459 case FILE_MEM_EXCLUSIVE
:
1460 return is_mem_exclusive(cs
);
1461 case FILE_MEM_HARDWALL
:
1462 return is_mem_hardwall(cs
);
1463 case FILE_SCHED_LOAD_BALANCE
:
1464 return is_sched_load_balance(cs
);
1465 case FILE_MEMORY_MIGRATE
:
1466 return is_memory_migrate(cs
);
1467 case FILE_MEMORY_PRESSURE_ENABLED
:
1468 return cpuset_memory_pressure_enabled
;
1469 case FILE_MEMORY_PRESSURE
:
1470 return fmeter_getrate(&cs
->fmeter
);
1471 case FILE_SPREAD_PAGE
:
1472 return is_spread_page(cs
);
1473 case FILE_SPREAD_SLAB
:
1474 return is_spread_slab(cs
);
1480 static s64
cpuset_read_s64(struct cgroup
*cont
, struct cftype
*cft
)
1482 struct cpuset
*cs
= cgroup_cs(cont
);
1483 cpuset_filetype_t type
= cft
->private;
1485 case FILE_SCHED_RELAX_DOMAIN_LEVEL
:
1486 return cs
->relax_domain_level
;
1494 * for the common functions, 'private' gives the type of file
1497 static struct cftype files
[] = {
1500 .read
= cpuset_common_file_read
,
1501 .write
= cpuset_common_file_write
,
1502 .private = FILE_CPULIST
,
1507 .read
= cpuset_common_file_read
,
1508 .write
= cpuset_common_file_write
,
1509 .private = FILE_MEMLIST
,
1513 .name
= "cpu_exclusive",
1514 .read_u64
= cpuset_read_u64
,
1515 .write_u64
= cpuset_write_u64
,
1516 .private = FILE_CPU_EXCLUSIVE
,
1520 .name
= "mem_exclusive",
1521 .read_u64
= cpuset_read_u64
,
1522 .write_u64
= cpuset_write_u64
,
1523 .private = FILE_MEM_EXCLUSIVE
,
1527 .name
= "mem_hardwall",
1528 .read_u64
= cpuset_read_u64
,
1529 .write_u64
= cpuset_write_u64
,
1530 .private = FILE_MEM_HARDWALL
,
1534 .name
= "sched_load_balance",
1535 .read_u64
= cpuset_read_u64
,
1536 .write_u64
= cpuset_write_u64
,
1537 .private = FILE_SCHED_LOAD_BALANCE
,
1541 .name
= "sched_relax_domain_level",
1542 .read_s64
= cpuset_read_s64
,
1543 .write_s64
= cpuset_write_s64
,
1544 .private = FILE_SCHED_RELAX_DOMAIN_LEVEL
,
1548 .name
= "memory_migrate",
1549 .read_u64
= cpuset_read_u64
,
1550 .write_u64
= cpuset_write_u64
,
1551 .private = FILE_MEMORY_MIGRATE
,
1555 .name
= "memory_pressure",
1556 .read_u64
= cpuset_read_u64
,
1557 .write_u64
= cpuset_write_u64
,
1558 .private = FILE_MEMORY_PRESSURE
,
1562 .name
= "memory_spread_page",
1563 .read_u64
= cpuset_read_u64
,
1564 .write_u64
= cpuset_write_u64
,
1565 .private = FILE_SPREAD_PAGE
,
1569 .name
= "memory_spread_slab",
1570 .read_u64
= cpuset_read_u64
,
1571 .write_u64
= cpuset_write_u64
,
1572 .private = FILE_SPREAD_SLAB
,
1576 static struct cftype cft_memory_pressure_enabled
= {
1577 .name
= "memory_pressure_enabled",
1578 .read_u64
= cpuset_read_u64
,
1579 .write_u64
= cpuset_write_u64
,
1580 .private = FILE_MEMORY_PRESSURE_ENABLED
,
1583 static int cpuset_populate(struct cgroup_subsys
*ss
, struct cgroup
*cont
)
1587 err
= cgroup_add_files(cont
, ss
, files
, ARRAY_SIZE(files
));
1590 /* memory_pressure_enabled is in root cpuset only */
1592 err
= cgroup_add_file(cont
, ss
,
1593 &cft_memory_pressure_enabled
);
1598 * post_clone() is called at the end of cgroup_clone().
1599 * 'cgroup' was just created automatically as a result of
1600 * a cgroup_clone(), and the current task is about to
1601 * be moved into 'cgroup'.
1603 * Currently we refuse to set up the cgroup - thereby
1604 * refusing the task to be entered, and as a result refusing
1605 * the sys_unshare() or clone() which initiated it - if any
1606 * sibling cpusets have exclusive cpus or mem.
1608 * If this becomes a problem for some users who wish to
1609 * allow that scenario, then cpuset_post_clone() could be
1610 * changed to grant parent->cpus_allowed-sibling_cpus_exclusive
1611 * (and likewise for mems) to the new cgroup. Called with cgroup_mutex
1614 static void cpuset_post_clone(struct cgroup_subsys
*ss
,
1615 struct cgroup
*cgroup
)
1617 struct cgroup
*parent
, *child
;
1618 struct cpuset
*cs
, *parent_cs
;
1620 parent
= cgroup
->parent
;
1621 list_for_each_entry(child
, &parent
->children
, sibling
) {
1622 cs
= cgroup_cs(child
);
1623 if (is_mem_exclusive(cs
) || is_cpu_exclusive(cs
))
1626 cs
= cgroup_cs(cgroup
);
1627 parent_cs
= cgroup_cs(parent
);
1629 cs
->mems_allowed
= parent_cs
->mems_allowed
;
1630 cs
->cpus_allowed
= parent_cs
->cpus_allowed
;
1635 * cpuset_create - create a cpuset
1636 * ss: cpuset cgroup subsystem
1637 * cont: control group that the new cpuset will be part of
1640 static struct cgroup_subsys_state
*cpuset_create(
1641 struct cgroup_subsys
*ss
,
1642 struct cgroup
*cont
)
1645 struct cpuset
*parent
;
1647 if (!cont
->parent
) {
1648 /* This is early initialization for the top cgroup */
1649 top_cpuset
.mems_generation
= cpuset_mems_generation
++;
1650 return &top_cpuset
.css
;
1652 parent
= cgroup_cs(cont
->parent
);
1653 cs
= kmalloc(sizeof(*cs
), GFP_KERNEL
);
1655 return ERR_PTR(-ENOMEM
);
1657 cpuset_update_task_memory_state();
1659 if (is_spread_page(parent
))
1660 set_bit(CS_SPREAD_PAGE
, &cs
->flags
);
1661 if (is_spread_slab(parent
))
1662 set_bit(CS_SPREAD_SLAB
, &cs
->flags
);
1663 set_bit(CS_SCHED_LOAD_BALANCE
, &cs
->flags
);
1664 cpus_clear(cs
->cpus_allowed
);
1665 nodes_clear(cs
->mems_allowed
);
1666 cs
->mems_generation
= cpuset_mems_generation
++;
1667 fmeter_init(&cs
->fmeter
);
1668 cs
->relax_domain_level
= -1;
1670 cs
->parent
= parent
;
1671 number_of_cpusets
++;
1676 * Locking note on the strange update_flag() call below:
1678 * If the cpuset being removed has its flag 'sched_load_balance'
1679 * enabled, then simulate turning sched_load_balance off, which
1680 * will call rebuild_sched_domains(). The get_online_cpus()
1681 * call in rebuild_sched_domains() must not be made while holding
1682 * callback_mutex. Elsewhere the kernel nests callback_mutex inside
1683 * get_online_cpus() calls. So the reverse nesting would risk an
1687 static void cpuset_destroy(struct cgroup_subsys
*ss
, struct cgroup
*cont
)
1689 struct cpuset
*cs
= cgroup_cs(cont
);
1691 cpuset_update_task_memory_state();
1693 if (is_sched_load_balance(cs
))
1694 update_flag(CS_SCHED_LOAD_BALANCE
, cs
, 0);
1696 number_of_cpusets
--;
1700 struct cgroup_subsys cpuset_subsys
= {
1702 .create
= cpuset_create
,
1703 .destroy
= cpuset_destroy
,
1704 .can_attach
= cpuset_can_attach
,
1705 .attach
= cpuset_attach
,
1706 .populate
= cpuset_populate
,
1707 .post_clone
= cpuset_post_clone
,
1708 .subsys_id
= cpuset_subsys_id
,
1713 * cpuset_init_early - just enough so that the calls to
1714 * cpuset_update_task_memory_state() in early init code
1718 int __init
cpuset_init_early(void)
1720 top_cpuset
.mems_generation
= cpuset_mems_generation
++;
1726 * cpuset_init - initialize cpusets at system boot
1728 * Description: Initialize top_cpuset and the cpuset internal file system,
1731 int __init
cpuset_init(void)
1735 cpus_setall(top_cpuset
.cpus_allowed
);
1736 nodes_setall(top_cpuset
.mems_allowed
);
1738 fmeter_init(&top_cpuset
.fmeter
);
1739 top_cpuset
.mems_generation
= cpuset_mems_generation
++;
1740 set_bit(CS_SCHED_LOAD_BALANCE
, &top_cpuset
.flags
);
1741 top_cpuset
.relax_domain_level
= -1;
1743 err
= register_filesystem(&cpuset_fs_type
);
1747 number_of_cpusets
= 1;
1752 * cpuset_do_move_task - move a given task to another cpuset
1753 * @tsk: pointer to task_struct the task to move
1754 * @scan: struct cgroup_scanner contained in its struct cpuset_hotplug_scanner
1756 * Called by cgroup_scan_tasks() for each task in a cgroup.
1757 * Return nonzero to stop the walk through the tasks.
1759 static void cpuset_do_move_task(struct task_struct
*tsk
,
1760 struct cgroup_scanner
*scan
)
1762 struct cpuset_hotplug_scanner
*chsp
;
1764 chsp
= container_of(scan
, struct cpuset_hotplug_scanner
, scan
);
1765 cgroup_attach_task(chsp
->to
, tsk
);
1769 * move_member_tasks_to_cpuset - move tasks from one cpuset to another
1770 * @from: cpuset in which the tasks currently reside
1771 * @to: cpuset to which the tasks will be moved
1773 * Called with cgroup_mutex held
1774 * callback_mutex must not be held, as cpuset_attach() will take it.
1776 * The cgroup_scan_tasks() function will scan all the tasks in a cgroup,
1777 * calling callback functions for each.
1779 static void move_member_tasks_to_cpuset(struct cpuset
*from
, struct cpuset
*to
)
1781 struct cpuset_hotplug_scanner scan
;
1783 scan
.scan
.cg
= from
->css
.cgroup
;
1784 scan
.scan
.test_task
= NULL
; /* select all tasks in cgroup */
1785 scan
.scan
.process_task
= cpuset_do_move_task
;
1786 scan
.scan
.heap
= NULL
;
1787 scan
.to
= to
->css
.cgroup
;
1789 if (cgroup_scan_tasks((struct cgroup_scanner
*)&scan
))
1790 printk(KERN_ERR
"move_member_tasks_to_cpuset: "
1791 "cgroup_scan_tasks failed\n");
1795 * If common_cpu_mem_hotplug_unplug(), below, unplugs any CPUs
1796 * or memory nodes, we need to walk over the cpuset hierarchy,
1797 * removing that CPU or node from all cpusets. If this removes the
1798 * last CPU or node from a cpuset, then move the tasks in the empty
1799 * cpuset to its next-highest non-empty parent.
1801 * Called with cgroup_mutex held
1802 * callback_mutex must not be held, as cpuset_attach() will take it.
1804 static void remove_tasks_in_empty_cpuset(struct cpuset
*cs
)
1806 struct cpuset
*parent
;
1809 * The cgroup's css_sets list is in use if there are tasks
1810 * in the cpuset; the list is empty if there are none;
1811 * the cs->css.refcnt seems always 0.
1813 if (list_empty(&cs
->css
.cgroup
->css_sets
))
1817 * Find its next-highest non-empty parent, (top cpuset
1818 * has online cpus, so can't be empty).
1820 parent
= cs
->parent
;
1821 while (cpus_empty(parent
->cpus_allowed
) ||
1822 nodes_empty(parent
->mems_allowed
))
1823 parent
= parent
->parent
;
1825 move_member_tasks_to_cpuset(cs
, parent
);
1829 * Walk the specified cpuset subtree and look for empty cpusets.
1830 * The tasks of such cpuset must be moved to a parent cpuset.
1832 * Called with cgroup_mutex held. We take callback_mutex to modify
1833 * cpus_allowed and mems_allowed.
1835 * This walk processes the tree from top to bottom, completing one layer
1836 * before dropping down to the next. It always processes a node before
1837 * any of its children.
1839 * For now, since we lack memory hot unplug, we'll never see a cpuset
1840 * that has tasks along with an empty 'mems'. But if we did see such
1841 * a cpuset, we'd handle it just like we do if its 'cpus' was empty.
1843 static void scan_for_empty_cpusets(const struct cpuset
*root
)
1845 struct cpuset
*cp
; /* scans cpusets being updated */
1846 struct cpuset
*child
; /* scans child cpusets of cp */
1847 struct list_head queue
;
1848 struct cgroup
*cont
;
1850 INIT_LIST_HEAD(&queue
);
1852 list_add_tail((struct list_head
*)&root
->stack_list
, &queue
);
1854 while (!list_empty(&queue
)) {
1855 cp
= container_of(queue
.next
, struct cpuset
, stack_list
);
1856 list_del(queue
.next
);
1857 list_for_each_entry(cont
, &cp
->css
.cgroup
->children
, sibling
) {
1858 child
= cgroup_cs(cont
);
1859 list_add_tail(&child
->stack_list
, &queue
);
1861 cont
= cp
->css
.cgroup
;
1863 /* Continue past cpusets with all cpus, mems online */
1864 if (cpus_subset(cp
->cpus_allowed
, cpu_online_map
) &&
1865 nodes_subset(cp
->mems_allowed
, node_states
[N_HIGH_MEMORY
]))
1868 /* Remove offline cpus and mems from this cpuset. */
1869 mutex_lock(&callback_mutex
);
1870 cpus_and(cp
->cpus_allowed
, cp
->cpus_allowed
, cpu_online_map
);
1871 nodes_and(cp
->mems_allowed
, cp
->mems_allowed
,
1872 node_states
[N_HIGH_MEMORY
]);
1873 mutex_unlock(&callback_mutex
);
1875 /* Move tasks from the empty cpuset to a parent */
1876 if (cpus_empty(cp
->cpus_allowed
) ||
1877 nodes_empty(cp
->mems_allowed
))
1878 remove_tasks_in_empty_cpuset(cp
);
1883 * The cpus_allowed and mems_allowed nodemasks in the top_cpuset track
1884 * cpu_online_map and node_states[N_HIGH_MEMORY]. Force the top cpuset to
1885 * track what's online after any CPU or memory node hotplug or unplug event.
1887 * Since there are two callers of this routine, one for CPU hotplug
1888 * events and one for memory node hotplug events, we could have coded
1889 * two separate routines here. We code it as a single common routine
1890 * in order to minimize text size.
1893 static void common_cpu_mem_hotplug_unplug(void)
1897 top_cpuset
.cpus_allowed
= cpu_online_map
;
1898 top_cpuset
.mems_allowed
= node_states
[N_HIGH_MEMORY
];
1899 scan_for_empty_cpusets(&top_cpuset
);
1902 * Scheduler destroys domains on hotplug events.
1903 * Rebuild them based on the current settings.
1905 rebuild_sched_domains();
1911 * The top_cpuset tracks what CPUs and Memory Nodes are online,
1912 * period. This is necessary in order to make cpusets transparent
1913 * (of no affect) on systems that are actively using CPU hotplug
1914 * but making no active use of cpusets.
1916 * This routine ensures that top_cpuset.cpus_allowed tracks
1917 * cpu_online_map on each CPU hotplug (cpuhp) event.
1920 static int cpuset_handle_cpuhp(struct notifier_block
*unused_nb
,
1921 unsigned long phase
, void *unused_cpu
)
1923 if (phase
== CPU_DYING
|| phase
== CPU_DYING_FROZEN
)
1926 common_cpu_mem_hotplug_unplug();
1930 #ifdef CONFIG_MEMORY_HOTPLUG
1932 * Keep top_cpuset.mems_allowed tracking node_states[N_HIGH_MEMORY].
1933 * Call this routine anytime after you change
1934 * node_states[N_HIGH_MEMORY].
1935 * See also the previous routine cpuset_handle_cpuhp().
1938 void cpuset_track_online_nodes(void)
1940 common_cpu_mem_hotplug_unplug();
1945 * cpuset_init_smp - initialize cpus_allowed
1947 * Description: Finish top cpuset after cpu, node maps are initialized
1950 void __init
cpuset_init_smp(void)
1952 top_cpuset
.cpus_allowed
= cpu_online_map
;
1953 top_cpuset
.mems_allowed
= node_states
[N_HIGH_MEMORY
];
1955 hotcpu_notifier(cpuset_handle_cpuhp
, 0);
1960 * cpuset_cpus_allowed - return cpus_allowed mask from a tasks cpuset.
1961 * @tsk: pointer to task_struct from which to obtain cpuset->cpus_allowed.
1962 * @pmask: pointer to cpumask_t variable to receive cpus_allowed set.
1964 * Description: Returns the cpumask_t cpus_allowed of the cpuset
1965 * attached to the specified @tsk. Guaranteed to return some non-empty
1966 * subset of cpu_online_map, even if this means going outside the
1970 void cpuset_cpus_allowed(struct task_struct
*tsk
, cpumask_t
*pmask
)
1972 mutex_lock(&callback_mutex
);
1973 cpuset_cpus_allowed_locked(tsk
, pmask
);
1974 mutex_unlock(&callback_mutex
);
1978 * cpuset_cpus_allowed_locked - return cpus_allowed mask from a tasks cpuset.
1979 * Must be called with callback_mutex held.
1981 void cpuset_cpus_allowed_locked(struct task_struct
*tsk
, cpumask_t
*pmask
)
1984 guarantee_online_cpus(task_cs(tsk
), pmask
);
1988 void cpuset_init_current_mems_allowed(void)
1990 nodes_setall(current
->mems_allowed
);
1994 * cpuset_mems_allowed - return mems_allowed mask from a tasks cpuset.
1995 * @tsk: pointer to task_struct from which to obtain cpuset->mems_allowed.
1997 * Description: Returns the nodemask_t mems_allowed of the cpuset
1998 * attached to the specified @tsk. Guaranteed to return some non-empty
1999 * subset of node_states[N_HIGH_MEMORY], even if this means going outside the
2003 nodemask_t
cpuset_mems_allowed(struct task_struct
*tsk
)
2007 mutex_lock(&callback_mutex
);
2009 guarantee_online_mems(task_cs(tsk
), &mask
);
2011 mutex_unlock(&callback_mutex
);
2017 * cpuset_nodemask_valid_mems_allowed - check nodemask vs. curremt mems_allowed
2018 * @nodemask: the nodemask to be checked
2020 * Are any of the nodes in the nodemask allowed in current->mems_allowed?
2022 int cpuset_nodemask_valid_mems_allowed(nodemask_t
*nodemask
)
2024 return nodes_intersects(*nodemask
, current
->mems_allowed
);
2028 * nearest_hardwall_ancestor() - Returns the nearest mem_exclusive or
2029 * mem_hardwall ancestor to the specified cpuset. Call holding
2030 * callback_mutex. If no ancestor is mem_exclusive or mem_hardwall
2031 * (an unusual configuration), then returns the root cpuset.
2033 static const struct cpuset
*nearest_hardwall_ancestor(const struct cpuset
*cs
)
2035 while (!(is_mem_exclusive(cs
) || is_mem_hardwall(cs
)) && cs
->parent
)
2041 * cpuset_zone_allowed_softwall - Can we allocate on zone z's memory node?
2042 * @z: is this zone on an allowed node?
2043 * @gfp_mask: memory allocation flags
2045 * If we're in interrupt, yes, we can always allocate. If
2046 * __GFP_THISNODE is set, yes, we can always allocate. If zone
2047 * z's node is in our tasks mems_allowed, yes. If it's not a
2048 * __GFP_HARDWALL request and this zone's nodes is in the nearest
2049 * hardwalled cpuset ancestor to this tasks cpuset, yes.
2050 * If the task has been OOM killed and has access to memory reserves
2051 * as specified by the TIF_MEMDIE flag, yes.
2054 * If __GFP_HARDWALL is set, cpuset_zone_allowed_softwall()
2055 * reduces to cpuset_zone_allowed_hardwall(). Otherwise,
2056 * cpuset_zone_allowed_softwall() might sleep, and might allow a zone
2057 * from an enclosing cpuset.
2059 * cpuset_zone_allowed_hardwall() only handles the simpler case of
2060 * hardwall cpusets, and never sleeps.
2062 * The __GFP_THISNODE placement logic is really handled elsewhere,
2063 * by forcibly using a zonelist starting at a specified node, and by
2064 * (in get_page_from_freelist()) refusing to consider the zones for
2065 * any node on the zonelist except the first. By the time any such
2066 * calls get to this routine, we should just shut up and say 'yes'.
2068 * GFP_USER allocations are marked with the __GFP_HARDWALL bit,
2069 * and do not allow allocations outside the current tasks cpuset
2070 * unless the task has been OOM killed as is marked TIF_MEMDIE.
2071 * GFP_KERNEL allocations are not so marked, so can escape to the
2072 * nearest enclosing hardwalled ancestor cpuset.
2074 * Scanning up parent cpusets requires callback_mutex. The
2075 * __alloc_pages() routine only calls here with __GFP_HARDWALL bit
2076 * _not_ set if it's a GFP_KERNEL allocation, and all nodes in the
2077 * current tasks mems_allowed came up empty on the first pass over
2078 * the zonelist. So only GFP_KERNEL allocations, if all nodes in the
2079 * cpuset are short of memory, might require taking the callback_mutex
2082 * The first call here from mm/page_alloc:get_page_from_freelist()
2083 * has __GFP_HARDWALL set in gfp_mask, enforcing hardwall cpusets,
2084 * so no allocation on a node outside the cpuset is allowed (unless
2085 * in interrupt, of course).
2087 * The second pass through get_page_from_freelist() doesn't even call
2088 * here for GFP_ATOMIC calls. For those calls, the __alloc_pages()
2089 * variable 'wait' is not set, and the bit ALLOC_CPUSET is not set
2090 * in alloc_flags. That logic and the checks below have the combined
2092 * in_interrupt - any node ok (current task context irrelevant)
2093 * GFP_ATOMIC - any node ok
2094 * TIF_MEMDIE - any node ok
2095 * GFP_KERNEL - any node in enclosing hardwalled cpuset ok
2096 * GFP_USER - only nodes in current tasks mems allowed ok.
2099 * Don't call cpuset_zone_allowed_softwall if you can't sleep, unless you
2100 * pass in the __GFP_HARDWALL flag set in gfp_flag, which disables
2101 * the code that might scan up ancestor cpusets and sleep.
2104 int __cpuset_zone_allowed_softwall(struct zone
*z
, gfp_t gfp_mask
)
2106 int node
; /* node that zone z is on */
2107 const struct cpuset
*cs
; /* current cpuset ancestors */
2108 int allowed
; /* is allocation in zone z allowed? */
2110 if (in_interrupt() || (gfp_mask
& __GFP_THISNODE
))
2112 node
= zone_to_nid(z
);
2113 might_sleep_if(!(gfp_mask
& __GFP_HARDWALL
));
2114 if (node_isset(node
, current
->mems_allowed
))
2117 * Allow tasks that have access to memory reserves because they have
2118 * been OOM killed to get memory anywhere.
2120 if (unlikely(test_thread_flag(TIF_MEMDIE
)))
2122 if (gfp_mask
& __GFP_HARDWALL
) /* If hardwall request, stop here */
2125 if (current
->flags
& PF_EXITING
) /* Let dying task have memory */
2128 /* Not hardwall and node outside mems_allowed: scan up cpusets */
2129 mutex_lock(&callback_mutex
);
2132 cs
= nearest_hardwall_ancestor(task_cs(current
));
2133 task_unlock(current
);
2135 allowed
= node_isset(node
, cs
->mems_allowed
);
2136 mutex_unlock(&callback_mutex
);
2141 * cpuset_zone_allowed_hardwall - Can we allocate on zone z's memory node?
2142 * @z: is this zone on an allowed node?
2143 * @gfp_mask: memory allocation flags
2145 * If we're in interrupt, yes, we can always allocate.
2146 * If __GFP_THISNODE is set, yes, we can always allocate. If zone
2147 * z's node is in our tasks mems_allowed, yes. If the task has been
2148 * OOM killed and has access to memory reserves as specified by the
2149 * TIF_MEMDIE flag, yes. Otherwise, no.
2151 * The __GFP_THISNODE placement logic is really handled elsewhere,
2152 * by forcibly using a zonelist starting at a specified node, and by
2153 * (in get_page_from_freelist()) refusing to consider the zones for
2154 * any node on the zonelist except the first. By the time any such
2155 * calls get to this routine, we should just shut up and say 'yes'.
2157 * Unlike the cpuset_zone_allowed_softwall() variant, above,
2158 * this variant requires that the zone be in the current tasks
2159 * mems_allowed or that we're in interrupt. It does not scan up the
2160 * cpuset hierarchy for the nearest enclosing mem_exclusive cpuset.
2164 int __cpuset_zone_allowed_hardwall(struct zone
*z
, gfp_t gfp_mask
)
2166 int node
; /* node that zone z is on */
2168 if (in_interrupt() || (gfp_mask
& __GFP_THISNODE
))
2170 node
= zone_to_nid(z
);
2171 if (node_isset(node
, current
->mems_allowed
))
2174 * Allow tasks that have access to memory reserves because they have
2175 * been OOM killed to get memory anywhere.
2177 if (unlikely(test_thread_flag(TIF_MEMDIE
)))
2183 * cpuset_lock - lock out any changes to cpuset structures
2185 * The out of memory (oom) code needs to mutex_lock cpusets
2186 * from being changed while it scans the tasklist looking for a
2187 * task in an overlapping cpuset. Expose callback_mutex via this
2188 * cpuset_lock() routine, so the oom code can lock it, before
2189 * locking the task list. The tasklist_lock is a spinlock, so
2190 * must be taken inside callback_mutex.
2193 void cpuset_lock(void)
2195 mutex_lock(&callback_mutex
);
2199 * cpuset_unlock - release lock on cpuset changes
2201 * Undo the lock taken in a previous cpuset_lock() call.
2204 void cpuset_unlock(void)
2206 mutex_unlock(&callback_mutex
);
2210 * cpuset_mem_spread_node() - On which node to begin search for a page
2212 * If a task is marked PF_SPREAD_PAGE or PF_SPREAD_SLAB (as for
2213 * tasks in a cpuset with is_spread_page or is_spread_slab set),
2214 * and if the memory allocation used cpuset_mem_spread_node()
2215 * to determine on which node to start looking, as it will for
2216 * certain page cache or slab cache pages such as used for file
2217 * system buffers and inode caches, then instead of starting on the
2218 * local node to look for a free page, rather spread the starting
2219 * node around the tasks mems_allowed nodes.
2221 * We don't have to worry about the returned node being offline
2222 * because "it can't happen", and even if it did, it would be ok.
2224 * The routines calling guarantee_online_mems() are careful to
2225 * only set nodes in task->mems_allowed that are online. So it
2226 * should not be possible for the following code to return an
2227 * offline node. But if it did, that would be ok, as this routine
2228 * is not returning the node where the allocation must be, only
2229 * the node where the search should start. The zonelist passed to
2230 * __alloc_pages() will include all nodes. If the slab allocator
2231 * is passed an offline node, it will fall back to the local node.
2232 * See kmem_cache_alloc_node().
2235 int cpuset_mem_spread_node(void)
2239 node
= next_node(current
->cpuset_mem_spread_rotor
, current
->mems_allowed
);
2240 if (node
== MAX_NUMNODES
)
2241 node
= first_node(current
->mems_allowed
);
2242 current
->cpuset_mem_spread_rotor
= node
;
2245 EXPORT_SYMBOL_GPL(cpuset_mem_spread_node
);
2248 * cpuset_mems_allowed_intersects - Does @tsk1's mems_allowed intersect @tsk2's?
2249 * @tsk1: pointer to task_struct of some task.
2250 * @tsk2: pointer to task_struct of some other task.
2252 * Description: Return true if @tsk1's mems_allowed intersects the
2253 * mems_allowed of @tsk2. Used by the OOM killer to determine if
2254 * one of the task's memory usage might impact the memory available
2258 int cpuset_mems_allowed_intersects(const struct task_struct
*tsk1
,
2259 const struct task_struct
*tsk2
)
2261 return nodes_intersects(tsk1
->mems_allowed
, tsk2
->mems_allowed
);
2265 * Collection of memory_pressure is suppressed unless
2266 * this flag is enabled by writing "1" to the special
2267 * cpuset file 'memory_pressure_enabled' in the root cpuset.
2270 int cpuset_memory_pressure_enabled __read_mostly
;
2273 * cpuset_memory_pressure_bump - keep stats of per-cpuset reclaims.
2275 * Keep a running average of the rate of synchronous (direct)
2276 * page reclaim efforts initiated by tasks in each cpuset.
2278 * This represents the rate at which some task in the cpuset
2279 * ran low on memory on all nodes it was allowed to use, and
2280 * had to enter the kernels page reclaim code in an effort to
2281 * create more free memory by tossing clean pages or swapping
2282 * or writing dirty pages.
2284 * Display to user space in the per-cpuset read-only file
2285 * "memory_pressure". Value displayed is an integer
2286 * representing the recent rate of entry into the synchronous
2287 * (direct) page reclaim by any task attached to the cpuset.
2290 void __cpuset_memory_pressure_bump(void)
2293 fmeter_markevent(&task_cs(current
)->fmeter
);
2294 task_unlock(current
);
2297 #ifdef CONFIG_PROC_PID_CPUSET
2299 * proc_cpuset_show()
2300 * - Print tasks cpuset path into seq_file.
2301 * - Used for /proc/<pid>/cpuset.
2302 * - No need to task_lock(tsk) on this tsk->cpuset reference, as it
2303 * doesn't really matter if tsk->cpuset changes after we read it,
2304 * and we take cgroup_mutex, keeping cpuset_attach() from changing it
2307 static int proc_cpuset_show(struct seq_file
*m
, void *unused_v
)
2310 struct task_struct
*tsk
;
2312 struct cgroup_subsys_state
*css
;
2316 buf
= kmalloc(PAGE_SIZE
, GFP_KERNEL
);
2322 tsk
= get_pid_task(pid
, PIDTYPE_PID
);
2328 css
= task_subsys_state(tsk
, cpuset_subsys_id
);
2329 retval
= cgroup_path(css
->cgroup
, buf
, PAGE_SIZE
);
2336 put_task_struct(tsk
);
2343 static int cpuset_open(struct inode
*inode
, struct file
*file
)
2345 struct pid
*pid
= PROC_I(inode
)->pid
;
2346 return single_open(file
, proc_cpuset_show
, pid
);
2349 const struct file_operations proc_cpuset_operations
= {
2350 .open
= cpuset_open
,
2352 .llseek
= seq_lseek
,
2353 .release
= single_release
,
2355 #endif /* CONFIG_PROC_PID_CPUSET */
2357 /* Display task cpus_allowed, mems_allowed in /proc/<pid>/status file. */
2358 void cpuset_task_status_allowed(struct seq_file
*m
, struct task_struct
*task
)
2360 seq_printf(m
, "Cpus_allowed:\t");
2361 m
->count
+= cpumask_scnprintf(m
->buf
+ m
->count
, m
->size
- m
->count
,
2362 task
->cpus_allowed
);
2363 seq_printf(m
, "\n");
2364 seq_printf(m
, "Cpus_allowed_list:\t");
2365 m
->count
+= cpulist_scnprintf(m
->buf
+ m
->count
, m
->size
- m
->count
,
2366 task
->cpus_allowed
);
2367 seq_printf(m
, "\n");
2368 seq_printf(m
, "Mems_allowed:\t");
2369 m
->count
+= nodemask_scnprintf(m
->buf
+ m
->count
, m
->size
- m
->count
,
2370 task
->mems_allowed
);
2371 seq_printf(m
, "\n");
2372 seq_printf(m
, "Mems_allowed_list:\t");
2373 m
->count
+= nodelist_scnprintf(m
->buf
+ m
->count
, m
->size
- m
->count
,
2374 task
->mems_allowed
);
2375 seq_printf(m
, "\n");