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 */
100 struct fmeter fmeter
; /* memory_pressure filter */
102 /* partition number for rebuild_sched_domains() */
105 /* for custom sched domain */
106 int relax_domain_level
;
108 /* used for walking a cpuset heirarchy */
109 struct list_head stack_list
;
112 /* Retrieve the cpuset for a cgroup */
113 static inline struct cpuset
*cgroup_cs(struct cgroup
*cont
)
115 return container_of(cgroup_subsys_state(cont
, cpuset_subsys_id
),
119 /* Retrieve the cpuset for a task */
120 static inline struct cpuset
*task_cs(struct task_struct
*task
)
122 return container_of(task_subsys_state(task
, cpuset_subsys_id
),
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
);
173 static struct cpuset top_cpuset
= {
174 .flags
= ((1 << CS_CPU_EXCLUSIVE
) | (1 << CS_MEM_EXCLUSIVE
)),
178 * There are two global mutexes guarding cpuset structures. The first
179 * is the main control groups cgroup_mutex, accessed via
180 * cgroup_lock()/cgroup_unlock(). The second is the cpuset-specific
181 * callback_mutex, below. They can nest. It is ok to first take
182 * cgroup_mutex, then nest callback_mutex. We also require taking
183 * task_lock() when dereferencing a task's cpuset pointer. See "The
184 * task_lock() exception", at the end of this comment.
186 * A task must hold both mutexes to modify cpusets. If a task
187 * holds cgroup_mutex, then it blocks others wanting that mutex,
188 * ensuring that it is the only task able to also acquire callback_mutex
189 * and be able to modify cpusets. It can perform various checks on
190 * the cpuset structure first, knowing nothing will change. It can
191 * also allocate memory while just holding cgroup_mutex. While it is
192 * performing these checks, various callback routines can briefly
193 * acquire callback_mutex to query cpusets. Once it is ready to make
194 * the changes, it takes callback_mutex, blocking everyone else.
196 * Calls to the kernel memory allocator can not be made while holding
197 * callback_mutex, as that would risk double tripping on callback_mutex
198 * from one of the callbacks into the cpuset code from within
201 * If a task is only holding callback_mutex, then it has read-only
204 * Now, the task_struct fields mems_allowed and mempolicy may be changed
205 * by other task, we use alloc_lock in the task_struct fields to protect
208 * The cpuset_common_file_read() handlers only hold callback_mutex across
209 * small pieces of code, such as when reading out possibly multi-word
210 * cpumasks and nodemasks.
212 * Accessing a task's cpuset should be done in accordance with the
213 * guidelines for accessing subsystem state in kernel/cgroup.c
216 static DEFINE_MUTEX(callback_mutex
);
219 * cpuset_buffer_lock protects both the cpuset_name and cpuset_nodelist
220 * buffers. They are statically allocated to prevent using excess stack
221 * when calling cpuset_print_task_mems_allowed().
223 #define CPUSET_NAME_LEN (128)
224 #define CPUSET_NODELIST_LEN (256)
225 static char cpuset_name
[CPUSET_NAME_LEN
];
226 static char cpuset_nodelist
[CPUSET_NODELIST_LEN
];
227 static DEFINE_SPINLOCK(cpuset_buffer_lock
);
230 * This is ugly, but preserves the userspace API for existing cpuset
231 * users. If someone tries to mount the "cpuset" filesystem, we
232 * silently switch it to mount "cgroup" instead
234 static int cpuset_get_sb(struct file_system_type
*fs_type
,
235 int flags
, const char *unused_dev_name
,
236 void *data
, struct vfsmount
*mnt
)
238 struct file_system_type
*cgroup_fs
= get_fs_type("cgroup");
243 "release_agent=/sbin/cpuset_release_agent";
244 ret
= cgroup_fs
->get_sb(cgroup_fs
, flags
,
245 unused_dev_name
, mountopts
, mnt
);
246 put_filesystem(cgroup_fs
);
251 static struct file_system_type cpuset_fs_type
= {
253 .get_sb
= cpuset_get_sb
,
257 * Return in pmask the portion of a cpusets's cpus_allowed that
258 * are online. If none are online, walk up the cpuset hierarchy
259 * until we find one that does have some online cpus. If we get
260 * all the way to the top and still haven't found any online cpus,
261 * return cpu_online_map. Or if passed a NULL cs from an exit'ing
262 * task, return cpu_online_map.
264 * One way or another, we guarantee to return some non-empty subset
267 * Call with callback_mutex held.
270 static void guarantee_online_cpus(const struct cpuset
*cs
,
271 struct cpumask
*pmask
)
273 while (cs
&& !cpumask_intersects(cs
->cpus_allowed
, cpu_online_mask
))
276 cpumask_and(pmask
, cs
->cpus_allowed
, cpu_online_mask
);
278 cpumask_copy(pmask
, cpu_online_mask
);
279 BUG_ON(!cpumask_intersects(pmask
, cpu_online_mask
));
283 * Return in *pmask the portion of a cpusets's mems_allowed that
284 * are online, with memory. If none are online with memory, walk
285 * up the cpuset hierarchy until we find one that does have some
286 * online mems. If we get all the way to the top and still haven't
287 * found any online mems, return node_states[N_HIGH_MEMORY].
289 * One way or another, we guarantee to return some non-empty subset
290 * of node_states[N_HIGH_MEMORY].
292 * Call with callback_mutex held.
295 static void guarantee_online_mems(const struct cpuset
*cs
, nodemask_t
*pmask
)
297 while (cs
&& !nodes_intersects(cs
->mems_allowed
,
298 node_states
[N_HIGH_MEMORY
]))
301 nodes_and(*pmask
, cs
->mems_allowed
,
302 node_states
[N_HIGH_MEMORY
]);
304 *pmask
= node_states
[N_HIGH_MEMORY
];
305 BUG_ON(!nodes_intersects(*pmask
, node_states
[N_HIGH_MEMORY
]));
309 * update task's spread flag if cpuset's page/slab spread flag is set
311 * Called with callback_mutex/cgroup_mutex held
313 static void cpuset_update_task_spread_flag(struct cpuset
*cs
,
314 struct task_struct
*tsk
)
316 if (is_spread_page(cs
))
317 tsk
->flags
|= PF_SPREAD_PAGE
;
319 tsk
->flags
&= ~PF_SPREAD_PAGE
;
320 if (is_spread_slab(cs
))
321 tsk
->flags
|= PF_SPREAD_SLAB
;
323 tsk
->flags
&= ~PF_SPREAD_SLAB
;
327 * is_cpuset_subset(p, q) - Is cpuset p a subset of cpuset q?
329 * One cpuset is a subset of another if all its allowed CPUs and
330 * Memory Nodes are a subset of the other, and its exclusive flags
331 * are only set if the other's are set. Call holding cgroup_mutex.
334 static int is_cpuset_subset(const struct cpuset
*p
, const struct cpuset
*q
)
336 return cpumask_subset(p
->cpus_allowed
, q
->cpus_allowed
) &&
337 nodes_subset(p
->mems_allowed
, q
->mems_allowed
) &&
338 is_cpu_exclusive(p
) <= is_cpu_exclusive(q
) &&
339 is_mem_exclusive(p
) <= is_mem_exclusive(q
);
343 * alloc_trial_cpuset - allocate a trial cpuset
344 * @cs: the cpuset that the trial cpuset duplicates
346 static struct cpuset
*alloc_trial_cpuset(const struct cpuset
*cs
)
348 struct cpuset
*trial
;
350 trial
= kmemdup(cs
, sizeof(*cs
), GFP_KERNEL
);
354 if (!alloc_cpumask_var(&trial
->cpus_allowed
, GFP_KERNEL
)) {
358 cpumask_copy(trial
->cpus_allowed
, cs
->cpus_allowed
);
364 * free_trial_cpuset - free the trial cpuset
365 * @trial: the trial cpuset to be freed
367 static void free_trial_cpuset(struct cpuset
*trial
)
369 free_cpumask_var(trial
->cpus_allowed
);
374 * validate_change() - Used to validate that any proposed cpuset change
375 * follows the structural rules for cpusets.
377 * If we replaced the flag and mask values of the current cpuset
378 * (cur) with those values in the trial cpuset (trial), would
379 * our various subset and exclusive rules still be valid? Presumes
382 * 'cur' is the address of an actual, in-use cpuset. Operations
383 * such as list traversal that depend on the actual address of the
384 * cpuset in the list must use cur below, not trial.
386 * 'trial' is the address of bulk structure copy of cur, with
387 * perhaps one or more of the fields cpus_allowed, mems_allowed,
388 * or flags changed to new, trial values.
390 * Return 0 if valid, -errno if not.
393 static int validate_change(const struct cpuset
*cur
, const struct cpuset
*trial
)
396 struct cpuset
*c
, *par
;
398 /* Each of our child cpusets must be a subset of us */
399 list_for_each_entry(cont
, &cur
->css
.cgroup
->children
, sibling
) {
400 if (!is_cpuset_subset(cgroup_cs(cont
), trial
))
404 /* Remaining checks don't apply to root cpuset */
405 if (cur
== &top_cpuset
)
410 /* We must be a subset of our parent cpuset */
411 if (!is_cpuset_subset(trial
, par
))
415 * If either I or some sibling (!= me) is exclusive, we can't
418 list_for_each_entry(cont
, &par
->css
.cgroup
->children
, sibling
) {
420 if ((is_cpu_exclusive(trial
) || is_cpu_exclusive(c
)) &&
422 cpumask_intersects(trial
->cpus_allowed
, c
->cpus_allowed
))
424 if ((is_mem_exclusive(trial
) || is_mem_exclusive(c
)) &&
426 nodes_intersects(trial
->mems_allowed
, c
->mems_allowed
))
430 /* Cpusets with tasks can't have empty cpus_allowed or mems_allowed */
431 if (cgroup_task_count(cur
->css
.cgroup
)) {
432 if (cpumask_empty(trial
->cpus_allowed
) ||
433 nodes_empty(trial
->mems_allowed
)) {
443 * Helper routine for generate_sched_domains().
444 * Do cpusets a, b have overlapping cpus_allowed masks?
446 static int cpusets_overlap(struct cpuset
*a
, struct cpuset
*b
)
448 return cpumask_intersects(a
->cpus_allowed
, b
->cpus_allowed
);
452 update_domain_attr(struct sched_domain_attr
*dattr
, struct cpuset
*c
)
454 if (dattr
->relax_domain_level
< c
->relax_domain_level
)
455 dattr
->relax_domain_level
= c
->relax_domain_level
;
460 update_domain_attr_tree(struct sched_domain_attr
*dattr
, struct cpuset
*c
)
464 list_add(&c
->stack_list
, &q
);
465 while (!list_empty(&q
)) {
468 struct cpuset
*child
;
470 cp
= list_first_entry(&q
, struct cpuset
, stack_list
);
473 if (cpumask_empty(cp
->cpus_allowed
))
476 if (is_sched_load_balance(cp
))
477 update_domain_attr(dattr
, cp
);
479 list_for_each_entry(cont
, &cp
->css
.cgroup
->children
, sibling
) {
480 child
= cgroup_cs(cont
);
481 list_add_tail(&child
->stack_list
, &q
);
487 * generate_sched_domains()
489 * This function builds a partial partition of the systems CPUs
490 * A 'partial partition' is a set of non-overlapping subsets whose
491 * union is a subset of that set.
492 * The output of this function needs to be passed to kernel/sched.c
493 * partition_sched_domains() routine, which will rebuild the scheduler's
494 * load balancing domains (sched domains) as specified by that partial
497 * See "What is sched_load_balance" in Documentation/cgroups/cpusets.txt
498 * for a background explanation of this.
500 * Does not return errors, on the theory that the callers of this
501 * routine would rather not worry about failures to rebuild sched
502 * domains when operating in the severe memory shortage situations
503 * that could cause allocation failures below.
505 * Must be called with cgroup_lock held.
507 * The three key local variables below are:
508 * q - a linked-list queue of cpuset pointers, used to implement a
509 * top-down scan of all cpusets. This scan loads a pointer
510 * to each cpuset marked is_sched_load_balance into the
511 * array 'csa'. For our purposes, rebuilding the schedulers
512 * sched domains, we can ignore !is_sched_load_balance cpusets.
513 * csa - (for CpuSet Array) Array of pointers to all the cpusets
514 * that need to be load balanced, for convenient iterative
515 * access by the subsequent code that finds the best partition,
516 * i.e the set of domains (subsets) of CPUs such that the
517 * cpus_allowed of every cpuset marked is_sched_load_balance
518 * is a subset of one of these domains, while there are as
519 * many such domains as possible, each as small as possible.
520 * doms - Conversion of 'csa' to an array of cpumasks, for passing to
521 * the kernel/sched.c routine partition_sched_domains() in a
522 * convenient format, that can be easily compared to the prior
523 * value to determine what partition elements (sched domains)
524 * were changed (added or removed.)
526 * Finding the best partition (set of domains):
527 * The triple nested loops below over i, j, k scan over the
528 * load balanced cpusets (using the array of cpuset pointers in
529 * csa[]) looking for pairs of cpusets that have overlapping
530 * cpus_allowed, but which don't have the same 'pn' partition
531 * number and gives them in the same partition number. It keeps
532 * looping on the 'restart' label until it can no longer find
535 * The union of the cpus_allowed masks from the set of
536 * all cpusets having the same 'pn' value then form the one
537 * element of the partition (one sched domain) to be passed to
538 * partition_sched_domains().
540 /* FIXME: see the FIXME in partition_sched_domains() */
541 static int generate_sched_domains(struct cpumask
**domains
,
542 struct sched_domain_attr
**attributes
)
544 LIST_HEAD(q
); /* queue of cpusets to be scanned */
545 struct cpuset
*cp
; /* scans q */
546 struct cpuset
**csa
; /* array of all cpuset ptrs */
547 int csn
; /* how many cpuset ptrs in csa so far */
548 int i
, j
, k
; /* indices for partition finding loops */
549 struct cpumask
*doms
; /* resulting partition; i.e. sched domains */
550 struct sched_domain_attr
*dattr
; /* attributes for custom domains */
551 int ndoms
= 0; /* number of sched domains in result */
552 int nslot
; /* next empty doms[] struct cpumask slot */
558 /* Special case for the 99% of systems with one, full, sched domain */
559 if (is_sched_load_balance(&top_cpuset
)) {
560 doms
= kmalloc(cpumask_size(), GFP_KERNEL
);
564 dattr
= kmalloc(sizeof(struct sched_domain_attr
), GFP_KERNEL
);
566 *dattr
= SD_ATTR_INIT
;
567 update_domain_attr_tree(dattr
, &top_cpuset
);
569 cpumask_copy(doms
, top_cpuset
.cpus_allowed
);
575 csa
= kmalloc(number_of_cpusets
* sizeof(cp
), GFP_KERNEL
);
580 list_add(&top_cpuset
.stack_list
, &q
);
581 while (!list_empty(&q
)) {
583 struct cpuset
*child
; /* scans child cpusets of cp */
585 cp
= list_first_entry(&q
, struct cpuset
, stack_list
);
588 if (cpumask_empty(cp
->cpus_allowed
))
592 * All child cpusets contain a subset of the parent's cpus, so
593 * just skip them, and then we call update_domain_attr_tree()
594 * to calc relax_domain_level of the corresponding sched
597 if (is_sched_load_balance(cp
)) {
602 list_for_each_entry(cont
, &cp
->css
.cgroup
->children
, sibling
) {
603 child
= cgroup_cs(cont
);
604 list_add_tail(&child
->stack_list
, &q
);
608 for (i
= 0; i
< csn
; i
++)
613 /* Find the best partition (set of sched domains) */
614 for (i
= 0; i
< csn
; i
++) {
615 struct cpuset
*a
= csa
[i
];
618 for (j
= 0; j
< csn
; j
++) {
619 struct cpuset
*b
= csa
[j
];
622 if (apn
!= bpn
&& cpusets_overlap(a
, b
)) {
623 for (k
= 0; k
< csn
; k
++) {
624 struct cpuset
*c
= csa
[k
];
629 ndoms
--; /* one less element */
636 * Now we know how many domains to create.
637 * Convert <csn, csa> to <ndoms, doms> and populate cpu masks.
639 doms
= kmalloc(ndoms
* cpumask_size(), GFP_KERNEL
);
644 * The rest of the code, including the scheduler, can deal with
645 * dattr==NULL case. No need to abort if alloc fails.
647 dattr
= kmalloc(ndoms
* sizeof(struct sched_domain_attr
), GFP_KERNEL
);
649 for (nslot
= 0, i
= 0; i
< csn
; i
++) {
650 struct cpuset
*a
= csa
[i
];
655 /* Skip completed partitions */
661 if (nslot
== ndoms
) {
662 static int warnings
= 10;
665 "rebuild_sched_domains confused:"
666 " nslot %d, ndoms %d, csn %d, i %d,"
668 nslot
, ndoms
, csn
, i
, apn
);
676 *(dattr
+ nslot
) = SD_ATTR_INIT
;
677 for (j
= i
; j
< csn
; j
++) {
678 struct cpuset
*b
= csa
[j
];
681 cpumask_or(dp
, dp
, b
->cpus_allowed
);
683 update_domain_attr_tree(dattr
+ nslot
, b
);
685 /* Done with this partition */
691 BUG_ON(nslot
!= ndoms
);
697 * Fallback to the default domain if kmalloc() failed.
698 * See comments in partition_sched_domains().
709 * Rebuild scheduler domains.
711 * Call with neither cgroup_mutex held nor within get_online_cpus().
712 * Takes both cgroup_mutex and get_online_cpus().
714 * Cannot be directly called from cpuset code handling changes
715 * to the cpuset pseudo-filesystem, because it cannot be called
716 * from code that already holds cgroup_mutex.
718 static void do_rebuild_sched_domains(struct work_struct
*unused
)
720 struct sched_domain_attr
*attr
;
721 struct cpumask
*doms
;
726 /* Generate domain masks and attrs */
728 ndoms
= generate_sched_domains(&doms
, &attr
);
731 /* Have scheduler rebuild the domains */
732 partition_sched_domains(ndoms
, doms
, attr
);
736 #else /* !CONFIG_SMP */
737 static void do_rebuild_sched_domains(struct work_struct
*unused
)
741 static int generate_sched_domains(struct cpumask
**domains
,
742 struct sched_domain_attr
**attributes
)
747 #endif /* CONFIG_SMP */
749 static DECLARE_WORK(rebuild_sched_domains_work
, do_rebuild_sched_domains
);
752 * Rebuild scheduler domains, asynchronously via workqueue.
754 * If the flag 'sched_load_balance' of any cpuset with non-empty
755 * 'cpus' changes, or if the 'cpus' allowed changes in any cpuset
756 * which has that flag enabled, or if any cpuset with a non-empty
757 * 'cpus' is removed, then call this routine to rebuild the
758 * scheduler's dynamic sched domains.
760 * The rebuild_sched_domains() and partition_sched_domains()
761 * routines must nest cgroup_lock() inside get_online_cpus(),
762 * but such cpuset changes as these must nest that locking the
763 * other way, holding cgroup_lock() for much of the code.
765 * So in order to avoid an ABBA deadlock, the cpuset code handling
766 * these user changes delegates the actual sched domain rebuilding
767 * to a separate workqueue thread, which ends up processing the
768 * above do_rebuild_sched_domains() function.
770 static void async_rebuild_sched_domains(void)
772 queue_work(cpuset_wq
, &rebuild_sched_domains_work
);
776 * Accomplishes the same scheduler domain rebuild as the above
777 * async_rebuild_sched_domains(), however it directly calls the
778 * rebuild routine synchronously rather than calling it via an
779 * asynchronous work thread.
781 * This can only be called from code that is not holding
782 * cgroup_mutex (not nested in a cgroup_lock() call.)
784 void rebuild_sched_domains(void)
786 do_rebuild_sched_domains(NULL
);
790 * cpuset_test_cpumask - test a task's cpus_allowed versus its cpuset's
792 * @scan: struct cgroup_scanner contained in its struct cpuset_hotplug_scanner
794 * Call with cgroup_mutex held. May take callback_mutex during call.
795 * Called for each task in a cgroup by cgroup_scan_tasks().
796 * Return nonzero if this tasks's cpus_allowed mask should be changed (in other
797 * words, if its mask is not equal to its cpuset's mask).
799 static int cpuset_test_cpumask(struct task_struct
*tsk
,
800 struct cgroup_scanner
*scan
)
802 return !cpumask_equal(&tsk
->cpus_allowed
,
803 (cgroup_cs(scan
->cg
))->cpus_allowed
);
807 * cpuset_change_cpumask - make a task's cpus_allowed the same as its cpuset's
809 * @scan: struct cgroup_scanner containing the cgroup of the task
811 * Called by cgroup_scan_tasks() for each task in a cgroup whose
812 * cpus_allowed mask needs to be changed.
814 * We don't need to re-check for the cgroup/cpuset membership, since we're
815 * holding cgroup_lock() at this point.
817 static void cpuset_change_cpumask(struct task_struct
*tsk
,
818 struct cgroup_scanner
*scan
)
820 set_cpus_allowed_ptr(tsk
, ((cgroup_cs(scan
->cg
))->cpus_allowed
));
824 * update_tasks_cpumask - Update the cpumasks of tasks in the cpuset.
825 * @cs: the cpuset in which each task's cpus_allowed mask needs to be changed
826 * @heap: if NULL, defer allocating heap memory to cgroup_scan_tasks()
828 * Called with cgroup_mutex held
830 * The cgroup_scan_tasks() function will scan all the tasks in a cgroup,
831 * calling callback functions for each.
833 * No return value. It's guaranteed that cgroup_scan_tasks() always returns 0
836 static void update_tasks_cpumask(struct cpuset
*cs
, struct ptr_heap
*heap
)
838 struct cgroup_scanner scan
;
840 scan
.cg
= cs
->css
.cgroup
;
841 scan
.test_task
= cpuset_test_cpumask
;
842 scan
.process_task
= cpuset_change_cpumask
;
844 cgroup_scan_tasks(&scan
);
848 * update_cpumask - update the cpus_allowed mask of a cpuset and all tasks in it
849 * @cs: the cpuset to consider
850 * @buf: buffer of cpu numbers written to this cpuset
852 static int update_cpumask(struct cpuset
*cs
, struct cpuset
*trialcs
,
855 struct ptr_heap heap
;
857 int is_load_balanced
;
859 /* top_cpuset.cpus_allowed tracks cpu_online_map; it's read-only */
860 if (cs
== &top_cpuset
)
864 * An empty cpus_allowed is ok only if the cpuset has no tasks.
865 * Since cpulist_parse() fails on an empty mask, we special case
866 * that parsing. The validate_change() call ensures that cpusets
867 * with tasks have cpus.
870 cpumask_clear(trialcs
->cpus_allowed
);
872 retval
= cpulist_parse(buf
, trialcs
->cpus_allowed
);
876 if (!cpumask_subset(trialcs
->cpus_allowed
, cpu_online_mask
))
879 retval
= validate_change(cs
, trialcs
);
883 /* Nothing to do if the cpus didn't change */
884 if (cpumask_equal(cs
->cpus_allowed
, trialcs
->cpus_allowed
))
887 retval
= heap_init(&heap
, PAGE_SIZE
, GFP_KERNEL
, NULL
);
891 is_load_balanced
= is_sched_load_balance(trialcs
);
893 mutex_lock(&callback_mutex
);
894 cpumask_copy(cs
->cpus_allowed
, trialcs
->cpus_allowed
);
895 mutex_unlock(&callback_mutex
);
898 * Scan tasks in the cpuset, and update the cpumasks of any
899 * that need an update.
901 update_tasks_cpumask(cs
, &heap
);
905 if (is_load_balanced
)
906 async_rebuild_sched_domains();
913 * Migrate memory region from one set of nodes to another.
915 * Temporarilly set tasks mems_allowed to target nodes of migration,
916 * so that the migration code can allocate pages on these nodes.
918 * Call holding cgroup_mutex, so current's cpuset won't change
919 * during this call, as manage_mutex holds off any cpuset_attach()
920 * calls. Therefore we don't need to take task_lock around the
921 * call to guarantee_online_mems(), as we know no one is changing
924 * Hold callback_mutex around the two modifications of our tasks
925 * mems_allowed to synchronize with cpuset_mems_allowed().
927 * While the mm_struct we are migrating is typically from some
928 * other task, the task_struct mems_allowed that we are hacking
929 * is for our current task, which must allocate new pages for that
930 * migrating memory region.
933 static void cpuset_migrate_mm(struct mm_struct
*mm
, const nodemask_t
*from
,
934 const nodemask_t
*to
)
936 struct task_struct
*tsk
= current
;
938 tsk
->mems_allowed
= *to
;
940 do_migrate_pages(mm
, from
, to
, MPOL_MF_MOVE_ALL
);
942 guarantee_online_mems(task_cs(tsk
),&tsk
->mems_allowed
);
946 * cpuset_change_task_nodemask - change task's mems_allowed and mempolicy
947 * @tsk: the task to change
948 * @newmems: new nodes that the task will be set
950 * In order to avoid seeing no nodes if the old and new nodes are disjoint,
951 * we structure updates as setting all new allowed nodes, then clearing newly
954 * Called with task's alloc_lock held
956 static void cpuset_change_task_nodemask(struct task_struct
*tsk
,
959 nodes_or(tsk
->mems_allowed
, tsk
->mems_allowed
, *newmems
);
960 mpol_rebind_task(tsk
, &tsk
->mems_allowed
);
961 mpol_rebind_task(tsk
, newmems
);
962 tsk
->mems_allowed
= *newmems
;
966 * Update task's mems_allowed and rebind its mempolicy and vmas' mempolicy
967 * of it to cpuset's new mems_allowed, and migrate pages to new nodes if
968 * memory_migrate flag is set. Called with cgroup_mutex held.
970 static void cpuset_change_nodemask(struct task_struct
*p
,
971 struct cgroup_scanner
*scan
)
973 struct mm_struct
*mm
;
976 const nodemask_t
*oldmem
= scan
->data
;
979 cs
= cgroup_cs(scan
->cg
);
980 guarantee_online_mems(cs
, &newmems
);
983 cpuset_change_task_nodemask(p
, &newmems
);
990 migrate
= is_memory_migrate(cs
);
992 mpol_rebind_mm(mm
, &cs
->mems_allowed
);
994 cpuset_migrate_mm(mm
, oldmem
, &cs
->mems_allowed
);
998 static void *cpuset_being_rebound
;
1001 * update_tasks_nodemask - Update the nodemasks of tasks in the cpuset.
1002 * @cs: the cpuset in which each task's mems_allowed mask needs to be changed
1003 * @oldmem: old mems_allowed of cpuset cs
1004 * @heap: if NULL, defer allocating heap memory to cgroup_scan_tasks()
1006 * Called with cgroup_mutex held
1007 * No return value. It's guaranteed that cgroup_scan_tasks() always returns 0
1010 static void update_tasks_nodemask(struct cpuset
*cs
, const nodemask_t
*oldmem
,
1011 struct ptr_heap
*heap
)
1013 struct cgroup_scanner scan
;
1015 cpuset_being_rebound
= cs
; /* causes mpol_dup() rebind */
1017 scan
.cg
= cs
->css
.cgroup
;
1018 scan
.test_task
= NULL
;
1019 scan
.process_task
= cpuset_change_nodemask
;
1021 scan
.data
= (nodemask_t
*)oldmem
;
1024 * The mpol_rebind_mm() call takes mmap_sem, which we couldn't
1025 * take while holding tasklist_lock. Forks can happen - the
1026 * mpol_dup() cpuset_being_rebound check will catch such forks,
1027 * and rebind their vma mempolicies too. Because we still hold
1028 * the global cgroup_mutex, we know that no other rebind effort
1029 * will be contending for the global variable cpuset_being_rebound.
1030 * It's ok if we rebind the same mm twice; mpol_rebind_mm()
1031 * is idempotent. Also migrate pages in each mm to new nodes.
1033 cgroup_scan_tasks(&scan
);
1035 /* We're done rebinding vmas to this cpuset's new mems_allowed. */
1036 cpuset_being_rebound
= NULL
;
1040 * Handle user request to change the 'mems' memory placement
1041 * of a cpuset. Needs to validate the request, update the
1042 * cpusets mems_allowed, and for each task in the cpuset,
1043 * update mems_allowed and rebind task's mempolicy and any vma
1044 * mempolicies and if the cpuset is marked 'memory_migrate',
1045 * migrate the tasks pages to the new memory.
1047 * Call with cgroup_mutex held. May take callback_mutex during call.
1048 * Will take tasklist_lock, scan tasklist for tasks in cpuset cs,
1049 * lock each such tasks mm->mmap_sem, scan its vma's and rebind
1050 * their mempolicies to the cpusets new mems_allowed.
1052 static int update_nodemask(struct cpuset
*cs
, struct cpuset
*trialcs
,
1057 struct ptr_heap heap
;
1060 * top_cpuset.mems_allowed tracks node_stats[N_HIGH_MEMORY];
1063 if (cs
== &top_cpuset
)
1067 * An empty mems_allowed is ok iff there are no tasks in the cpuset.
1068 * Since nodelist_parse() fails on an empty mask, we special case
1069 * that parsing. The validate_change() call ensures that cpusets
1070 * with tasks have memory.
1073 nodes_clear(trialcs
->mems_allowed
);
1075 retval
= nodelist_parse(buf
, trialcs
->mems_allowed
);
1079 if (!nodes_subset(trialcs
->mems_allowed
,
1080 node_states
[N_HIGH_MEMORY
]))
1083 oldmem
= cs
->mems_allowed
;
1084 if (nodes_equal(oldmem
, trialcs
->mems_allowed
)) {
1085 retval
= 0; /* Too easy - nothing to do */
1088 retval
= validate_change(cs
, trialcs
);
1092 retval
= heap_init(&heap
, PAGE_SIZE
, GFP_KERNEL
, NULL
);
1096 mutex_lock(&callback_mutex
);
1097 cs
->mems_allowed
= trialcs
->mems_allowed
;
1098 mutex_unlock(&callback_mutex
);
1100 update_tasks_nodemask(cs
, &oldmem
, &heap
);
1107 int current_cpuset_is_being_rebound(void)
1109 return task_cs(current
) == cpuset_being_rebound
;
1112 static int update_relax_domain_level(struct cpuset
*cs
, s64 val
)
1115 if (val
< -1 || val
>= SD_LV_MAX
)
1119 if (val
!= cs
->relax_domain_level
) {
1120 cs
->relax_domain_level
= val
;
1121 if (!cpumask_empty(cs
->cpus_allowed
) &&
1122 is_sched_load_balance(cs
))
1123 async_rebuild_sched_domains();
1130 * cpuset_change_flag - make a task's spread flags the same as its cpuset's
1131 * @tsk: task to be updated
1132 * @scan: struct cgroup_scanner containing the cgroup of the task
1134 * Called by cgroup_scan_tasks() for each task in a cgroup.
1136 * We don't need to re-check for the cgroup/cpuset membership, since we're
1137 * holding cgroup_lock() at this point.
1139 static void cpuset_change_flag(struct task_struct
*tsk
,
1140 struct cgroup_scanner
*scan
)
1142 cpuset_update_task_spread_flag(cgroup_cs(scan
->cg
), tsk
);
1146 * update_tasks_flags - update the spread flags of tasks in the cpuset.
1147 * @cs: the cpuset in which each task's spread flags needs to be changed
1148 * @heap: if NULL, defer allocating heap memory to cgroup_scan_tasks()
1150 * Called with cgroup_mutex held
1152 * The cgroup_scan_tasks() function will scan all the tasks in a cgroup,
1153 * calling callback functions for each.
1155 * No return value. It's guaranteed that cgroup_scan_tasks() always returns 0
1158 static void update_tasks_flags(struct cpuset
*cs
, struct ptr_heap
*heap
)
1160 struct cgroup_scanner scan
;
1162 scan
.cg
= cs
->css
.cgroup
;
1163 scan
.test_task
= NULL
;
1164 scan
.process_task
= cpuset_change_flag
;
1166 cgroup_scan_tasks(&scan
);
1170 * update_flag - read a 0 or a 1 in a file and update associated flag
1171 * bit: the bit to update (see cpuset_flagbits_t)
1172 * cs: the cpuset to update
1173 * turning_on: whether the flag is being set or cleared
1175 * Call with cgroup_mutex held.
1178 static int update_flag(cpuset_flagbits_t bit
, struct cpuset
*cs
,
1181 struct cpuset
*trialcs
;
1182 int balance_flag_changed
;
1183 int spread_flag_changed
;
1184 struct ptr_heap heap
;
1187 trialcs
= alloc_trial_cpuset(cs
);
1192 set_bit(bit
, &trialcs
->flags
);
1194 clear_bit(bit
, &trialcs
->flags
);
1196 err
= validate_change(cs
, trialcs
);
1200 err
= heap_init(&heap
, PAGE_SIZE
, GFP_KERNEL
, NULL
);
1204 balance_flag_changed
= (is_sched_load_balance(cs
) !=
1205 is_sched_load_balance(trialcs
));
1207 spread_flag_changed
= ((is_spread_slab(cs
) != is_spread_slab(trialcs
))
1208 || (is_spread_page(cs
) != is_spread_page(trialcs
)));
1210 mutex_lock(&callback_mutex
);
1211 cs
->flags
= trialcs
->flags
;
1212 mutex_unlock(&callback_mutex
);
1214 if (!cpumask_empty(trialcs
->cpus_allowed
) && balance_flag_changed
)
1215 async_rebuild_sched_domains();
1217 if (spread_flag_changed
)
1218 update_tasks_flags(cs
, &heap
);
1221 free_trial_cpuset(trialcs
);
1226 * Frequency meter - How fast is some event occurring?
1228 * These routines manage a digitally filtered, constant time based,
1229 * event frequency meter. There are four routines:
1230 * fmeter_init() - initialize a frequency meter.
1231 * fmeter_markevent() - called each time the event happens.
1232 * fmeter_getrate() - returns the recent rate of such events.
1233 * fmeter_update() - internal routine used to update fmeter.
1235 * A common data structure is passed to each of these routines,
1236 * which is used to keep track of the state required to manage the
1237 * frequency meter and its digital filter.
1239 * The filter works on the number of events marked per unit time.
1240 * The filter is single-pole low-pass recursive (IIR). The time unit
1241 * is 1 second. Arithmetic is done using 32-bit integers scaled to
1242 * simulate 3 decimal digits of precision (multiplied by 1000).
1244 * With an FM_COEF of 933, and a time base of 1 second, the filter
1245 * has a half-life of 10 seconds, meaning that if the events quit
1246 * happening, then the rate returned from the fmeter_getrate()
1247 * will be cut in half each 10 seconds, until it converges to zero.
1249 * It is not worth doing a real infinitely recursive filter. If more
1250 * than FM_MAXTICKS ticks have elapsed since the last filter event,
1251 * just compute FM_MAXTICKS ticks worth, by which point the level
1254 * Limit the count of unprocessed events to FM_MAXCNT, so as to avoid
1255 * arithmetic overflow in the fmeter_update() routine.
1257 * Given the simple 32 bit integer arithmetic used, this meter works
1258 * best for reporting rates between one per millisecond (msec) and
1259 * one per 32 (approx) seconds. At constant rates faster than one
1260 * per msec it maxes out at values just under 1,000,000. At constant
1261 * rates between one per msec, and one per second it will stabilize
1262 * to a value N*1000, where N is the rate of events per second.
1263 * At constant rates between one per second and one per 32 seconds,
1264 * it will be choppy, moving up on the seconds that have an event,
1265 * and then decaying until the next event. At rates slower than
1266 * about one in 32 seconds, it decays all the way back to zero between
1270 #define FM_COEF 933 /* coefficient for half-life of 10 secs */
1271 #define FM_MAXTICKS ((time_t)99) /* useless computing more ticks than this */
1272 #define FM_MAXCNT 1000000 /* limit cnt to avoid overflow */
1273 #define FM_SCALE 1000 /* faux fixed point scale */
1275 /* Initialize a frequency meter */
1276 static void fmeter_init(struct fmeter
*fmp
)
1281 spin_lock_init(&fmp
->lock
);
1284 /* Internal meter update - process cnt events and update value */
1285 static void fmeter_update(struct fmeter
*fmp
)
1287 time_t now
= get_seconds();
1288 time_t ticks
= now
- fmp
->time
;
1293 ticks
= min(FM_MAXTICKS
, ticks
);
1295 fmp
->val
= (FM_COEF
* fmp
->val
) / FM_SCALE
;
1298 fmp
->val
+= ((FM_SCALE
- FM_COEF
) * fmp
->cnt
) / FM_SCALE
;
1302 /* Process any previous ticks, then bump cnt by one (times scale). */
1303 static void fmeter_markevent(struct fmeter
*fmp
)
1305 spin_lock(&fmp
->lock
);
1307 fmp
->cnt
= min(FM_MAXCNT
, fmp
->cnt
+ FM_SCALE
);
1308 spin_unlock(&fmp
->lock
);
1311 /* Process any previous ticks, then return current value. */
1312 static int fmeter_getrate(struct fmeter
*fmp
)
1316 spin_lock(&fmp
->lock
);
1319 spin_unlock(&fmp
->lock
);
1323 /* Protected by cgroup_lock */
1324 static cpumask_var_t cpus_attach
;
1326 /* Called by cgroups to determine if a cpuset is usable; cgroup_mutex held */
1327 static int cpuset_can_attach(struct cgroup_subsys
*ss
, struct cgroup
*cont
,
1328 struct task_struct
*tsk
, bool threadgroup
)
1331 struct cpuset
*cs
= cgroup_cs(cont
);
1333 if (cpumask_empty(cs
->cpus_allowed
) || nodes_empty(cs
->mems_allowed
))
1337 * Kthreads bound to specific cpus cannot be moved to a new cpuset; we
1338 * cannot change their cpu affinity and isolating such threads by their
1339 * set of allowed nodes is unnecessary. Thus, cpusets are not
1340 * applicable for such threads. This prevents checking for success of
1341 * set_cpus_allowed_ptr() on all attached tasks before cpus_allowed may
1344 if (tsk
->flags
& PF_THREAD_BOUND
)
1347 ret
= security_task_setscheduler(tsk
, 0, NULL
);
1351 struct task_struct
*c
;
1354 list_for_each_entry_rcu(c
, &tsk
->thread_group
, thread_group
) {
1355 ret
= security_task_setscheduler(c
, 0, NULL
);
1366 static void cpuset_attach_task(struct task_struct
*tsk
, nodemask_t
*to
,
1371 * can_attach beforehand should guarantee that this doesn't fail.
1372 * TODO: have a better way to handle failure here
1374 err
= set_cpus_allowed_ptr(tsk
, cpus_attach
);
1378 cpuset_change_task_nodemask(tsk
, to
);
1380 cpuset_update_task_spread_flag(cs
, tsk
);
1384 static void cpuset_attach(struct cgroup_subsys
*ss
, struct cgroup
*cont
,
1385 struct cgroup
*oldcont
, struct task_struct
*tsk
,
1388 nodemask_t from
, to
;
1389 struct mm_struct
*mm
;
1390 struct cpuset
*cs
= cgroup_cs(cont
);
1391 struct cpuset
*oldcs
= cgroup_cs(oldcont
);
1393 if (cs
== &top_cpuset
) {
1394 cpumask_copy(cpus_attach
, cpu_possible_mask
);
1395 to
= node_possible_map
;
1397 guarantee_online_cpus(cs
, cpus_attach
);
1398 guarantee_online_mems(cs
, &to
);
1401 /* do per-task migration stuff possibly for each in the threadgroup */
1402 cpuset_attach_task(tsk
, &to
, cs
);
1404 struct task_struct
*c
;
1406 list_for_each_entry_rcu(c
, &tsk
->thread_group
, thread_group
) {
1407 cpuset_attach_task(c
, &to
, cs
);
1412 /* change mm; only needs to be done once even if threadgroup */
1413 from
= oldcs
->mems_allowed
;
1414 to
= cs
->mems_allowed
;
1415 mm
= get_task_mm(tsk
);
1417 mpol_rebind_mm(mm
, &to
);
1418 if (is_memory_migrate(cs
))
1419 cpuset_migrate_mm(mm
, &from
, &to
);
1424 /* The various types of files and directories in a cpuset file system */
1427 FILE_MEMORY_MIGRATE
,
1433 FILE_SCHED_LOAD_BALANCE
,
1434 FILE_SCHED_RELAX_DOMAIN_LEVEL
,
1435 FILE_MEMORY_PRESSURE_ENABLED
,
1436 FILE_MEMORY_PRESSURE
,
1439 } cpuset_filetype_t
;
1441 static int cpuset_write_u64(struct cgroup
*cgrp
, struct cftype
*cft
, u64 val
)
1444 struct cpuset
*cs
= cgroup_cs(cgrp
);
1445 cpuset_filetype_t type
= cft
->private;
1447 if (!cgroup_lock_live_group(cgrp
))
1451 case FILE_CPU_EXCLUSIVE
:
1452 retval
= update_flag(CS_CPU_EXCLUSIVE
, cs
, val
);
1454 case FILE_MEM_EXCLUSIVE
:
1455 retval
= update_flag(CS_MEM_EXCLUSIVE
, cs
, val
);
1457 case FILE_MEM_HARDWALL
:
1458 retval
= update_flag(CS_MEM_HARDWALL
, cs
, val
);
1460 case FILE_SCHED_LOAD_BALANCE
:
1461 retval
= update_flag(CS_SCHED_LOAD_BALANCE
, cs
, val
);
1463 case FILE_MEMORY_MIGRATE
:
1464 retval
= update_flag(CS_MEMORY_MIGRATE
, cs
, val
);
1466 case FILE_MEMORY_PRESSURE_ENABLED
:
1467 cpuset_memory_pressure_enabled
= !!val
;
1469 case FILE_MEMORY_PRESSURE
:
1472 case FILE_SPREAD_PAGE
:
1473 retval
= update_flag(CS_SPREAD_PAGE
, cs
, val
);
1475 case FILE_SPREAD_SLAB
:
1476 retval
= update_flag(CS_SPREAD_SLAB
, cs
, val
);
1486 static int cpuset_write_s64(struct cgroup
*cgrp
, struct cftype
*cft
, s64 val
)
1489 struct cpuset
*cs
= cgroup_cs(cgrp
);
1490 cpuset_filetype_t type
= cft
->private;
1492 if (!cgroup_lock_live_group(cgrp
))
1496 case FILE_SCHED_RELAX_DOMAIN_LEVEL
:
1497 retval
= update_relax_domain_level(cs
, val
);
1508 * Common handling for a write to a "cpus" or "mems" file.
1510 static int cpuset_write_resmask(struct cgroup
*cgrp
, struct cftype
*cft
,
1514 struct cpuset
*cs
= cgroup_cs(cgrp
);
1515 struct cpuset
*trialcs
;
1517 if (!cgroup_lock_live_group(cgrp
))
1520 trialcs
= alloc_trial_cpuset(cs
);
1524 switch (cft
->private) {
1526 retval
= update_cpumask(cs
, trialcs
, buf
);
1529 retval
= update_nodemask(cs
, trialcs
, buf
);
1536 free_trial_cpuset(trialcs
);
1542 * These ascii lists should be read in a single call, by using a user
1543 * buffer large enough to hold the entire map. If read in smaller
1544 * chunks, there is no guarantee of atomicity. Since the display format
1545 * used, list of ranges of sequential numbers, is variable length,
1546 * and since these maps can change value dynamically, one could read
1547 * gibberish by doing partial reads while a list was changing.
1548 * A single large read to a buffer that crosses a page boundary is
1549 * ok, because the result being copied to user land is not recomputed
1550 * across a page fault.
1553 static int cpuset_sprintf_cpulist(char *page
, struct cpuset
*cs
)
1557 mutex_lock(&callback_mutex
);
1558 ret
= cpulist_scnprintf(page
, PAGE_SIZE
, cs
->cpus_allowed
);
1559 mutex_unlock(&callback_mutex
);
1564 static int cpuset_sprintf_memlist(char *page
, struct cpuset
*cs
)
1568 mutex_lock(&callback_mutex
);
1569 mask
= cs
->mems_allowed
;
1570 mutex_unlock(&callback_mutex
);
1572 return nodelist_scnprintf(page
, PAGE_SIZE
, mask
);
1575 static ssize_t
cpuset_common_file_read(struct cgroup
*cont
,
1579 size_t nbytes
, loff_t
*ppos
)
1581 struct cpuset
*cs
= cgroup_cs(cont
);
1582 cpuset_filetype_t type
= cft
->private;
1587 if (!(page
= (char *)__get_free_page(GFP_TEMPORARY
)))
1594 s
+= cpuset_sprintf_cpulist(s
, cs
);
1597 s
+= cpuset_sprintf_memlist(s
, cs
);
1605 retval
= simple_read_from_buffer(buf
, nbytes
, ppos
, page
, s
- page
);
1607 free_page((unsigned long)page
);
1611 static u64
cpuset_read_u64(struct cgroup
*cont
, struct cftype
*cft
)
1613 struct cpuset
*cs
= cgroup_cs(cont
);
1614 cpuset_filetype_t type
= cft
->private;
1616 case FILE_CPU_EXCLUSIVE
:
1617 return is_cpu_exclusive(cs
);
1618 case FILE_MEM_EXCLUSIVE
:
1619 return is_mem_exclusive(cs
);
1620 case FILE_MEM_HARDWALL
:
1621 return is_mem_hardwall(cs
);
1622 case FILE_SCHED_LOAD_BALANCE
:
1623 return is_sched_load_balance(cs
);
1624 case FILE_MEMORY_MIGRATE
:
1625 return is_memory_migrate(cs
);
1626 case FILE_MEMORY_PRESSURE_ENABLED
:
1627 return cpuset_memory_pressure_enabled
;
1628 case FILE_MEMORY_PRESSURE
:
1629 return fmeter_getrate(&cs
->fmeter
);
1630 case FILE_SPREAD_PAGE
:
1631 return is_spread_page(cs
);
1632 case FILE_SPREAD_SLAB
:
1633 return is_spread_slab(cs
);
1638 /* Unreachable but makes gcc happy */
1642 static s64
cpuset_read_s64(struct cgroup
*cont
, struct cftype
*cft
)
1644 struct cpuset
*cs
= cgroup_cs(cont
);
1645 cpuset_filetype_t type
= cft
->private;
1647 case FILE_SCHED_RELAX_DOMAIN_LEVEL
:
1648 return cs
->relax_domain_level
;
1653 /* Unrechable but makes gcc happy */
1659 * for the common functions, 'private' gives the type of file
1662 static struct cftype files
[] = {
1665 .read
= cpuset_common_file_read
,
1666 .write_string
= cpuset_write_resmask
,
1667 .max_write_len
= (100U + 6 * NR_CPUS
),
1668 .private = FILE_CPULIST
,
1673 .read
= cpuset_common_file_read
,
1674 .write_string
= cpuset_write_resmask
,
1675 .max_write_len
= (100U + 6 * MAX_NUMNODES
),
1676 .private = FILE_MEMLIST
,
1680 .name
= "cpu_exclusive",
1681 .read_u64
= cpuset_read_u64
,
1682 .write_u64
= cpuset_write_u64
,
1683 .private = FILE_CPU_EXCLUSIVE
,
1687 .name
= "mem_exclusive",
1688 .read_u64
= cpuset_read_u64
,
1689 .write_u64
= cpuset_write_u64
,
1690 .private = FILE_MEM_EXCLUSIVE
,
1694 .name
= "mem_hardwall",
1695 .read_u64
= cpuset_read_u64
,
1696 .write_u64
= cpuset_write_u64
,
1697 .private = FILE_MEM_HARDWALL
,
1701 .name
= "sched_load_balance",
1702 .read_u64
= cpuset_read_u64
,
1703 .write_u64
= cpuset_write_u64
,
1704 .private = FILE_SCHED_LOAD_BALANCE
,
1708 .name
= "sched_relax_domain_level",
1709 .read_s64
= cpuset_read_s64
,
1710 .write_s64
= cpuset_write_s64
,
1711 .private = FILE_SCHED_RELAX_DOMAIN_LEVEL
,
1715 .name
= "memory_migrate",
1716 .read_u64
= cpuset_read_u64
,
1717 .write_u64
= cpuset_write_u64
,
1718 .private = FILE_MEMORY_MIGRATE
,
1722 .name
= "memory_pressure",
1723 .read_u64
= cpuset_read_u64
,
1724 .write_u64
= cpuset_write_u64
,
1725 .private = FILE_MEMORY_PRESSURE
,
1730 .name
= "memory_spread_page",
1731 .read_u64
= cpuset_read_u64
,
1732 .write_u64
= cpuset_write_u64
,
1733 .private = FILE_SPREAD_PAGE
,
1737 .name
= "memory_spread_slab",
1738 .read_u64
= cpuset_read_u64
,
1739 .write_u64
= cpuset_write_u64
,
1740 .private = FILE_SPREAD_SLAB
,
1744 static struct cftype cft_memory_pressure_enabled
= {
1745 .name
= "memory_pressure_enabled",
1746 .read_u64
= cpuset_read_u64
,
1747 .write_u64
= cpuset_write_u64
,
1748 .private = FILE_MEMORY_PRESSURE_ENABLED
,
1751 static int cpuset_populate(struct cgroup_subsys
*ss
, struct cgroup
*cont
)
1755 err
= cgroup_add_files(cont
, ss
, files
, ARRAY_SIZE(files
));
1758 /* memory_pressure_enabled is in root cpuset only */
1760 err
= cgroup_add_file(cont
, ss
,
1761 &cft_memory_pressure_enabled
);
1766 * post_clone() is called at the end of cgroup_clone().
1767 * 'cgroup' was just created automatically as a result of
1768 * a cgroup_clone(), and the current task is about to
1769 * be moved into 'cgroup'.
1771 * Currently we refuse to set up the cgroup - thereby
1772 * refusing the task to be entered, and as a result refusing
1773 * the sys_unshare() or clone() which initiated it - if any
1774 * sibling cpusets have exclusive cpus or mem.
1776 * If this becomes a problem for some users who wish to
1777 * allow that scenario, then cpuset_post_clone() could be
1778 * changed to grant parent->cpus_allowed-sibling_cpus_exclusive
1779 * (and likewise for mems) to the new cgroup. Called with cgroup_mutex
1782 static void cpuset_post_clone(struct cgroup_subsys
*ss
,
1783 struct cgroup
*cgroup
)
1785 struct cgroup
*parent
, *child
;
1786 struct cpuset
*cs
, *parent_cs
;
1788 parent
= cgroup
->parent
;
1789 list_for_each_entry(child
, &parent
->children
, sibling
) {
1790 cs
= cgroup_cs(child
);
1791 if (is_mem_exclusive(cs
) || is_cpu_exclusive(cs
))
1794 cs
= cgroup_cs(cgroup
);
1795 parent_cs
= cgroup_cs(parent
);
1797 cs
->mems_allowed
= parent_cs
->mems_allowed
;
1798 cpumask_copy(cs
->cpus_allowed
, parent_cs
->cpus_allowed
);
1803 * cpuset_create - create a cpuset
1804 * ss: cpuset cgroup subsystem
1805 * cont: control group that the new cpuset will be part of
1808 static struct cgroup_subsys_state
*cpuset_create(
1809 struct cgroup_subsys
*ss
,
1810 struct cgroup
*cont
)
1813 struct cpuset
*parent
;
1815 if (!cont
->parent
) {
1816 return &top_cpuset
.css
;
1818 parent
= cgroup_cs(cont
->parent
);
1819 cs
= kmalloc(sizeof(*cs
), GFP_KERNEL
);
1821 return ERR_PTR(-ENOMEM
);
1822 if (!alloc_cpumask_var(&cs
->cpus_allowed
, GFP_KERNEL
)) {
1824 return ERR_PTR(-ENOMEM
);
1828 if (is_spread_page(parent
))
1829 set_bit(CS_SPREAD_PAGE
, &cs
->flags
);
1830 if (is_spread_slab(parent
))
1831 set_bit(CS_SPREAD_SLAB
, &cs
->flags
);
1832 set_bit(CS_SCHED_LOAD_BALANCE
, &cs
->flags
);
1833 cpumask_clear(cs
->cpus_allowed
);
1834 nodes_clear(cs
->mems_allowed
);
1835 fmeter_init(&cs
->fmeter
);
1836 cs
->relax_domain_level
= -1;
1838 cs
->parent
= parent
;
1839 number_of_cpusets
++;
1844 * If the cpuset being removed has its flag 'sched_load_balance'
1845 * enabled, then simulate turning sched_load_balance off, which
1846 * will call async_rebuild_sched_domains().
1849 static void cpuset_destroy(struct cgroup_subsys
*ss
, struct cgroup
*cont
)
1851 struct cpuset
*cs
= cgroup_cs(cont
);
1853 if (is_sched_load_balance(cs
))
1854 update_flag(CS_SCHED_LOAD_BALANCE
, cs
, 0);
1856 number_of_cpusets
--;
1857 free_cpumask_var(cs
->cpus_allowed
);
1861 struct cgroup_subsys cpuset_subsys
= {
1863 .create
= cpuset_create
,
1864 .destroy
= cpuset_destroy
,
1865 .can_attach
= cpuset_can_attach
,
1866 .attach
= cpuset_attach
,
1867 .populate
= cpuset_populate
,
1868 .post_clone
= cpuset_post_clone
,
1869 .subsys_id
= cpuset_subsys_id
,
1874 * cpuset_init - initialize cpusets at system boot
1876 * Description: Initialize top_cpuset and the cpuset internal file system,
1879 int __init
cpuset_init(void)
1883 if (!alloc_cpumask_var(&top_cpuset
.cpus_allowed
, GFP_KERNEL
))
1886 cpumask_setall(top_cpuset
.cpus_allowed
);
1887 nodes_setall(top_cpuset
.mems_allowed
);
1889 fmeter_init(&top_cpuset
.fmeter
);
1890 set_bit(CS_SCHED_LOAD_BALANCE
, &top_cpuset
.flags
);
1891 top_cpuset
.relax_domain_level
= -1;
1893 err
= register_filesystem(&cpuset_fs_type
);
1897 if (!alloc_cpumask_var(&cpus_attach
, GFP_KERNEL
))
1900 number_of_cpusets
= 1;
1905 * cpuset_do_move_task - move a given task to another cpuset
1906 * @tsk: pointer to task_struct the task to move
1907 * @scan: struct cgroup_scanner contained in its struct cpuset_hotplug_scanner
1909 * Called by cgroup_scan_tasks() for each task in a cgroup.
1910 * Return nonzero to stop the walk through the tasks.
1912 static void cpuset_do_move_task(struct task_struct
*tsk
,
1913 struct cgroup_scanner
*scan
)
1915 struct cgroup
*new_cgroup
= scan
->data
;
1917 cgroup_attach_task(new_cgroup
, tsk
);
1921 * move_member_tasks_to_cpuset - move tasks from one cpuset to another
1922 * @from: cpuset in which the tasks currently reside
1923 * @to: cpuset to which the tasks will be moved
1925 * Called with cgroup_mutex held
1926 * callback_mutex must not be held, as cpuset_attach() will take it.
1928 * The cgroup_scan_tasks() function will scan all the tasks in a cgroup,
1929 * calling callback functions for each.
1931 static void move_member_tasks_to_cpuset(struct cpuset
*from
, struct cpuset
*to
)
1933 struct cgroup_scanner scan
;
1935 scan
.cg
= from
->css
.cgroup
;
1936 scan
.test_task
= NULL
; /* select all tasks in cgroup */
1937 scan
.process_task
= cpuset_do_move_task
;
1939 scan
.data
= to
->css
.cgroup
;
1941 if (cgroup_scan_tasks(&scan
))
1942 printk(KERN_ERR
"move_member_tasks_to_cpuset: "
1943 "cgroup_scan_tasks failed\n");
1947 * If CPU and/or memory hotplug handlers, below, unplug any CPUs
1948 * or memory nodes, we need to walk over the cpuset hierarchy,
1949 * removing that CPU or node from all cpusets. If this removes the
1950 * last CPU or node from a cpuset, then move the tasks in the empty
1951 * cpuset to its next-highest non-empty parent.
1953 * Called with cgroup_mutex held
1954 * callback_mutex must not be held, as cpuset_attach() will take it.
1956 static void remove_tasks_in_empty_cpuset(struct cpuset
*cs
)
1958 struct cpuset
*parent
;
1961 * The cgroup's css_sets list is in use if there are tasks
1962 * in the cpuset; the list is empty if there are none;
1963 * the cs->css.refcnt seems always 0.
1965 if (list_empty(&cs
->css
.cgroup
->css_sets
))
1969 * Find its next-highest non-empty parent, (top cpuset
1970 * has online cpus, so can't be empty).
1972 parent
= cs
->parent
;
1973 while (cpumask_empty(parent
->cpus_allowed
) ||
1974 nodes_empty(parent
->mems_allowed
))
1975 parent
= parent
->parent
;
1977 move_member_tasks_to_cpuset(cs
, parent
);
1981 * Walk the specified cpuset subtree and look for empty cpusets.
1982 * The tasks of such cpuset must be moved to a parent cpuset.
1984 * Called with cgroup_mutex held. We take callback_mutex to modify
1985 * cpus_allowed and mems_allowed.
1987 * This walk processes the tree from top to bottom, completing one layer
1988 * before dropping down to the next. It always processes a node before
1989 * any of its children.
1991 * For now, since we lack memory hot unplug, we'll never see a cpuset
1992 * that has tasks along with an empty 'mems'. But if we did see such
1993 * a cpuset, we'd handle it just like we do if its 'cpus' was empty.
1995 static void scan_for_empty_cpusets(struct cpuset
*root
)
1998 struct cpuset
*cp
; /* scans cpusets being updated */
1999 struct cpuset
*child
; /* scans child cpusets of cp */
2000 struct cgroup
*cont
;
2003 list_add_tail((struct list_head
*)&root
->stack_list
, &queue
);
2005 while (!list_empty(&queue
)) {
2006 cp
= list_first_entry(&queue
, struct cpuset
, stack_list
);
2007 list_del(queue
.next
);
2008 list_for_each_entry(cont
, &cp
->css
.cgroup
->children
, sibling
) {
2009 child
= cgroup_cs(cont
);
2010 list_add_tail(&child
->stack_list
, &queue
);
2013 /* Continue past cpusets with all cpus, mems online */
2014 if (cpumask_subset(cp
->cpus_allowed
, cpu_online_mask
) &&
2015 nodes_subset(cp
->mems_allowed
, node_states
[N_HIGH_MEMORY
]))
2018 oldmems
= cp
->mems_allowed
;
2020 /* Remove offline cpus and mems from this cpuset. */
2021 mutex_lock(&callback_mutex
);
2022 cpumask_and(cp
->cpus_allowed
, cp
->cpus_allowed
,
2024 nodes_and(cp
->mems_allowed
, cp
->mems_allowed
,
2025 node_states
[N_HIGH_MEMORY
]);
2026 mutex_unlock(&callback_mutex
);
2028 /* Move tasks from the empty cpuset to a parent */
2029 if (cpumask_empty(cp
->cpus_allowed
) ||
2030 nodes_empty(cp
->mems_allowed
))
2031 remove_tasks_in_empty_cpuset(cp
);
2033 update_tasks_cpumask(cp
, NULL
);
2034 update_tasks_nodemask(cp
, &oldmems
, NULL
);
2040 * The top_cpuset tracks what CPUs and Memory Nodes are online,
2041 * period. This is necessary in order to make cpusets transparent
2042 * (of no affect) on systems that are actively using CPU hotplug
2043 * but making no active use of cpusets.
2045 * This routine ensures that top_cpuset.cpus_allowed tracks
2046 * cpu_online_map on each CPU hotplug (cpuhp) event.
2048 * Called within get_online_cpus(). Needs to call cgroup_lock()
2049 * before calling generate_sched_domains().
2051 static int cpuset_track_online_cpus(struct notifier_block
*unused_nb
,
2052 unsigned long phase
, void *unused_cpu
)
2054 struct sched_domain_attr
*attr
;
2055 struct cpumask
*doms
;
2060 case CPU_ONLINE_FROZEN
:
2062 case CPU_DEAD_FROZEN
:
2070 mutex_lock(&callback_mutex
);
2071 cpumask_copy(top_cpuset
.cpus_allowed
, cpu_online_mask
);
2072 mutex_unlock(&callback_mutex
);
2073 scan_for_empty_cpusets(&top_cpuset
);
2074 ndoms
= generate_sched_domains(&doms
, &attr
);
2077 /* Have scheduler rebuild the domains */
2078 partition_sched_domains(ndoms
, doms
, attr
);
2083 #ifdef CONFIG_MEMORY_HOTPLUG
2085 * Keep top_cpuset.mems_allowed tracking node_states[N_HIGH_MEMORY].
2086 * Call this routine anytime after node_states[N_HIGH_MEMORY] changes.
2087 * See also the previous routine cpuset_track_online_cpus().
2089 static int cpuset_track_online_nodes(struct notifier_block
*self
,
2090 unsigned long action
, void *arg
)
2096 mutex_lock(&callback_mutex
);
2097 top_cpuset
.mems_allowed
= node_states
[N_HIGH_MEMORY
];
2098 mutex_unlock(&callback_mutex
);
2099 if (action
== MEM_OFFLINE
)
2100 scan_for_empty_cpusets(&top_cpuset
);
2111 * cpuset_init_smp - initialize cpus_allowed
2113 * Description: Finish top cpuset after cpu, node maps are initialized
2116 void __init
cpuset_init_smp(void)
2118 cpumask_copy(top_cpuset
.cpus_allowed
, cpu_online_mask
);
2119 top_cpuset
.mems_allowed
= node_states
[N_HIGH_MEMORY
];
2121 hotcpu_notifier(cpuset_track_online_cpus
, 0);
2122 hotplug_memory_notifier(cpuset_track_online_nodes
, 10);
2124 cpuset_wq
= create_singlethread_workqueue("cpuset");
2129 * cpuset_cpus_allowed - return cpus_allowed mask from a tasks cpuset.
2130 * @tsk: pointer to task_struct from which to obtain cpuset->cpus_allowed.
2131 * @pmask: pointer to struct cpumask variable to receive cpus_allowed set.
2133 * Description: Returns the cpumask_var_t cpus_allowed of the cpuset
2134 * attached to the specified @tsk. Guaranteed to return some non-empty
2135 * subset of cpu_online_map, even if this means going outside the
2139 void cpuset_cpus_allowed(struct task_struct
*tsk
, struct cpumask
*pmask
)
2141 mutex_lock(&callback_mutex
);
2142 cpuset_cpus_allowed_locked(tsk
, pmask
);
2143 mutex_unlock(&callback_mutex
);
2147 * cpuset_cpus_allowed_locked - return cpus_allowed mask from a tasks cpuset.
2148 * Must be called with callback_mutex held.
2150 void cpuset_cpus_allowed_locked(struct task_struct
*tsk
, struct cpumask
*pmask
)
2153 guarantee_online_cpus(task_cs(tsk
), pmask
);
2157 void cpuset_init_current_mems_allowed(void)
2159 nodes_setall(current
->mems_allowed
);
2163 * cpuset_mems_allowed - return mems_allowed mask from a tasks cpuset.
2164 * @tsk: pointer to task_struct from which to obtain cpuset->mems_allowed.
2166 * Description: Returns the nodemask_t mems_allowed of the cpuset
2167 * attached to the specified @tsk. Guaranteed to return some non-empty
2168 * subset of node_states[N_HIGH_MEMORY], even if this means going outside the
2172 nodemask_t
cpuset_mems_allowed(struct task_struct
*tsk
)
2176 mutex_lock(&callback_mutex
);
2178 guarantee_online_mems(task_cs(tsk
), &mask
);
2180 mutex_unlock(&callback_mutex
);
2186 * cpuset_nodemask_valid_mems_allowed - check nodemask vs. curremt mems_allowed
2187 * @nodemask: the nodemask to be checked
2189 * Are any of the nodes in the nodemask allowed in current->mems_allowed?
2191 int cpuset_nodemask_valid_mems_allowed(nodemask_t
*nodemask
)
2193 return nodes_intersects(*nodemask
, current
->mems_allowed
);
2197 * nearest_hardwall_ancestor() - Returns the nearest mem_exclusive or
2198 * mem_hardwall ancestor to the specified cpuset. Call holding
2199 * callback_mutex. If no ancestor is mem_exclusive or mem_hardwall
2200 * (an unusual configuration), then returns the root cpuset.
2202 static const struct cpuset
*nearest_hardwall_ancestor(const struct cpuset
*cs
)
2204 while (!(is_mem_exclusive(cs
) || is_mem_hardwall(cs
)) && cs
->parent
)
2210 * cpuset_node_allowed_softwall - Can we allocate on a memory node?
2211 * @node: is this an allowed node?
2212 * @gfp_mask: memory allocation flags
2214 * If we're in interrupt, yes, we can always allocate. If __GFP_THISNODE is
2215 * set, yes, we can always allocate. If node is in our task's mems_allowed,
2216 * yes. If it's not a __GFP_HARDWALL request and this node is in the nearest
2217 * hardwalled cpuset ancestor to this task's cpuset, yes. If the task has been
2218 * OOM killed and has access to memory reserves as specified by the TIF_MEMDIE
2222 * If __GFP_HARDWALL is set, cpuset_node_allowed_softwall() reduces to
2223 * cpuset_node_allowed_hardwall(). Otherwise, cpuset_node_allowed_softwall()
2224 * might sleep, and might allow a node from an enclosing cpuset.
2226 * cpuset_node_allowed_hardwall() only handles the simpler case of hardwall
2227 * cpusets, and never sleeps.
2229 * The __GFP_THISNODE placement logic is really handled elsewhere,
2230 * by forcibly using a zonelist starting at a specified node, and by
2231 * (in get_page_from_freelist()) refusing to consider the zones for
2232 * any node on the zonelist except the first. By the time any such
2233 * calls get to this routine, we should just shut up and say 'yes'.
2235 * GFP_USER allocations are marked with the __GFP_HARDWALL bit,
2236 * and do not allow allocations outside the current tasks cpuset
2237 * unless the task has been OOM killed as is marked TIF_MEMDIE.
2238 * GFP_KERNEL allocations are not so marked, so can escape to the
2239 * nearest enclosing hardwalled ancestor cpuset.
2241 * Scanning up parent cpusets requires callback_mutex. The
2242 * __alloc_pages() routine only calls here with __GFP_HARDWALL bit
2243 * _not_ set if it's a GFP_KERNEL allocation, and all nodes in the
2244 * current tasks mems_allowed came up empty on the first pass over
2245 * the zonelist. So only GFP_KERNEL allocations, if all nodes in the
2246 * cpuset are short of memory, might require taking the callback_mutex
2249 * The first call here from mm/page_alloc:get_page_from_freelist()
2250 * has __GFP_HARDWALL set in gfp_mask, enforcing hardwall cpusets,
2251 * so no allocation on a node outside the cpuset is allowed (unless
2252 * in interrupt, of course).
2254 * The second pass through get_page_from_freelist() doesn't even call
2255 * here for GFP_ATOMIC calls. For those calls, the __alloc_pages()
2256 * variable 'wait' is not set, and the bit ALLOC_CPUSET is not set
2257 * in alloc_flags. That logic and the checks below have the combined
2259 * in_interrupt - any node ok (current task context irrelevant)
2260 * GFP_ATOMIC - any node ok
2261 * TIF_MEMDIE - any node ok
2262 * GFP_KERNEL - any node in enclosing hardwalled cpuset ok
2263 * GFP_USER - only nodes in current tasks mems allowed ok.
2266 * Don't call cpuset_node_allowed_softwall if you can't sleep, unless you
2267 * pass in the __GFP_HARDWALL flag set in gfp_flag, which disables
2268 * the code that might scan up ancestor cpusets and sleep.
2270 int __cpuset_node_allowed_softwall(int node
, gfp_t gfp_mask
)
2272 const struct cpuset
*cs
; /* current cpuset ancestors */
2273 int allowed
; /* is allocation in zone z allowed? */
2275 if (in_interrupt() || (gfp_mask
& __GFP_THISNODE
))
2277 might_sleep_if(!(gfp_mask
& __GFP_HARDWALL
));
2278 if (node_isset(node
, current
->mems_allowed
))
2281 * Allow tasks that have access to memory reserves because they have
2282 * been OOM killed to get memory anywhere.
2284 if (unlikely(test_thread_flag(TIF_MEMDIE
)))
2286 if (gfp_mask
& __GFP_HARDWALL
) /* If hardwall request, stop here */
2289 if (current
->flags
& PF_EXITING
) /* Let dying task have memory */
2292 /* Not hardwall and node outside mems_allowed: scan up cpusets */
2293 mutex_lock(&callback_mutex
);
2296 cs
= nearest_hardwall_ancestor(task_cs(current
));
2297 task_unlock(current
);
2299 allowed
= node_isset(node
, cs
->mems_allowed
);
2300 mutex_unlock(&callback_mutex
);
2305 * cpuset_node_allowed_hardwall - Can we allocate on a memory node?
2306 * @node: is this an allowed node?
2307 * @gfp_mask: memory allocation flags
2309 * If we're in interrupt, yes, we can always allocate. If __GFP_THISNODE is
2310 * set, yes, we can always allocate. If node is in our task's mems_allowed,
2311 * yes. If the task has been OOM killed and has access to memory reserves as
2312 * specified by the TIF_MEMDIE flag, yes.
2315 * The __GFP_THISNODE placement logic is really handled elsewhere,
2316 * by forcibly using a zonelist starting at a specified node, and by
2317 * (in get_page_from_freelist()) refusing to consider the zones for
2318 * any node on the zonelist except the first. By the time any such
2319 * calls get to this routine, we should just shut up and say 'yes'.
2321 * Unlike the cpuset_node_allowed_softwall() variant, above,
2322 * this variant requires that the node be in the current task's
2323 * mems_allowed or that we're in interrupt. It does not scan up the
2324 * cpuset hierarchy for the nearest enclosing mem_exclusive cpuset.
2327 int __cpuset_node_allowed_hardwall(int node
, gfp_t gfp_mask
)
2329 if (in_interrupt() || (gfp_mask
& __GFP_THISNODE
))
2331 if (node_isset(node
, current
->mems_allowed
))
2334 * Allow tasks that have access to memory reserves because they have
2335 * been OOM killed to get memory anywhere.
2337 if (unlikely(test_thread_flag(TIF_MEMDIE
)))
2343 * cpuset_lock - lock out any changes to cpuset structures
2345 * The out of memory (oom) code needs to mutex_lock cpusets
2346 * from being changed while it scans the tasklist looking for a
2347 * task in an overlapping cpuset. Expose callback_mutex via this
2348 * cpuset_lock() routine, so the oom code can lock it, before
2349 * locking the task list. The tasklist_lock is a spinlock, so
2350 * must be taken inside callback_mutex.
2353 void cpuset_lock(void)
2355 mutex_lock(&callback_mutex
);
2359 * cpuset_unlock - release lock on cpuset changes
2361 * Undo the lock taken in a previous cpuset_lock() call.
2364 void cpuset_unlock(void)
2366 mutex_unlock(&callback_mutex
);
2370 * cpuset_mem_spread_node() - On which node to begin search for a page
2372 * If a task is marked PF_SPREAD_PAGE or PF_SPREAD_SLAB (as for
2373 * tasks in a cpuset with is_spread_page or is_spread_slab set),
2374 * and if the memory allocation used cpuset_mem_spread_node()
2375 * to determine on which node to start looking, as it will for
2376 * certain page cache or slab cache pages such as used for file
2377 * system buffers and inode caches, then instead of starting on the
2378 * local node to look for a free page, rather spread the starting
2379 * node around the tasks mems_allowed nodes.
2381 * We don't have to worry about the returned node being offline
2382 * because "it can't happen", and even if it did, it would be ok.
2384 * The routines calling guarantee_online_mems() are careful to
2385 * only set nodes in task->mems_allowed that are online. So it
2386 * should not be possible for the following code to return an
2387 * offline node. But if it did, that would be ok, as this routine
2388 * is not returning the node where the allocation must be, only
2389 * the node where the search should start. The zonelist passed to
2390 * __alloc_pages() will include all nodes. If the slab allocator
2391 * is passed an offline node, it will fall back to the local node.
2392 * See kmem_cache_alloc_node().
2395 int cpuset_mem_spread_node(void)
2399 node
= next_node(current
->cpuset_mem_spread_rotor
, current
->mems_allowed
);
2400 if (node
== MAX_NUMNODES
)
2401 node
= first_node(current
->mems_allowed
);
2402 current
->cpuset_mem_spread_rotor
= node
;
2405 EXPORT_SYMBOL_GPL(cpuset_mem_spread_node
);
2408 * cpuset_mems_allowed_intersects - Does @tsk1's mems_allowed intersect @tsk2's?
2409 * @tsk1: pointer to task_struct of some task.
2410 * @tsk2: pointer to task_struct of some other task.
2412 * Description: Return true if @tsk1's mems_allowed intersects the
2413 * mems_allowed of @tsk2. Used by the OOM killer to determine if
2414 * one of the task's memory usage might impact the memory available
2418 int cpuset_mems_allowed_intersects(const struct task_struct
*tsk1
,
2419 const struct task_struct
*tsk2
)
2421 return nodes_intersects(tsk1
->mems_allowed
, tsk2
->mems_allowed
);
2425 * cpuset_print_task_mems_allowed - prints task's cpuset and mems_allowed
2426 * @task: pointer to task_struct of some task.
2428 * Description: Prints @task's name, cpuset name, and cached copy of its
2429 * mems_allowed to the kernel log. Must hold task_lock(task) to allow
2430 * dereferencing task_cs(task).
2432 void cpuset_print_task_mems_allowed(struct task_struct
*tsk
)
2434 struct dentry
*dentry
;
2436 dentry
= task_cs(tsk
)->css
.cgroup
->dentry
;
2437 spin_lock(&cpuset_buffer_lock
);
2438 snprintf(cpuset_name
, CPUSET_NAME_LEN
,
2439 dentry
? (const char *)dentry
->d_name
.name
: "/");
2440 nodelist_scnprintf(cpuset_nodelist
, CPUSET_NODELIST_LEN
,
2442 printk(KERN_INFO
"%s cpuset=%s mems_allowed=%s\n",
2443 tsk
->comm
, cpuset_name
, cpuset_nodelist
);
2444 spin_unlock(&cpuset_buffer_lock
);
2448 * Collection of memory_pressure is suppressed unless
2449 * this flag is enabled by writing "1" to the special
2450 * cpuset file 'memory_pressure_enabled' in the root cpuset.
2453 int cpuset_memory_pressure_enabled __read_mostly
;
2456 * cpuset_memory_pressure_bump - keep stats of per-cpuset reclaims.
2458 * Keep a running average of the rate of synchronous (direct)
2459 * page reclaim efforts initiated by tasks in each cpuset.
2461 * This represents the rate at which some task in the cpuset
2462 * ran low on memory on all nodes it was allowed to use, and
2463 * had to enter the kernels page reclaim code in an effort to
2464 * create more free memory by tossing clean pages or swapping
2465 * or writing dirty pages.
2467 * Display to user space in the per-cpuset read-only file
2468 * "memory_pressure". Value displayed is an integer
2469 * representing the recent rate of entry into the synchronous
2470 * (direct) page reclaim by any task attached to the cpuset.
2473 void __cpuset_memory_pressure_bump(void)
2476 fmeter_markevent(&task_cs(current
)->fmeter
);
2477 task_unlock(current
);
2480 #ifdef CONFIG_PROC_PID_CPUSET
2482 * proc_cpuset_show()
2483 * - Print tasks cpuset path into seq_file.
2484 * - Used for /proc/<pid>/cpuset.
2485 * - No need to task_lock(tsk) on this tsk->cpuset reference, as it
2486 * doesn't really matter if tsk->cpuset changes after we read it,
2487 * and we take cgroup_mutex, keeping cpuset_attach() from changing it
2490 static int proc_cpuset_show(struct seq_file
*m
, void *unused_v
)
2493 struct task_struct
*tsk
;
2495 struct cgroup_subsys_state
*css
;
2499 buf
= kmalloc(PAGE_SIZE
, GFP_KERNEL
);
2505 tsk
= get_pid_task(pid
, PIDTYPE_PID
);
2511 css
= task_subsys_state(tsk
, cpuset_subsys_id
);
2512 retval
= cgroup_path(css
->cgroup
, buf
, PAGE_SIZE
);
2519 put_task_struct(tsk
);
2526 static int cpuset_open(struct inode
*inode
, struct file
*file
)
2528 struct pid
*pid
= PROC_I(inode
)->pid
;
2529 return single_open(file
, proc_cpuset_show
, pid
);
2532 const struct file_operations proc_cpuset_operations
= {
2533 .open
= cpuset_open
,
2535 .llseek
= seq_lseek
,
2536 .release
= single_release
,
2538 #endif /* CONFIG_PROC_PID_CPUSET */
2540 /* Display task mems_allowed in /proc/<pid>/status file. */
2541 void cpuset_task_status_allowed(struct seq_file
*m
, struct task_struct
*task
)
2543 seq_printf(m
, "Mems_allowed:\t");
2544 seq_nodemask(m
, &task
->mems_allowed
);
2545 seq_printf(m
, "\n");
2546 seq_printf(m
, "Mems_allowed_list:\t");
2547 seq_nodemask_list(m
, &task
->mems_allowed
);
2548 seq_printf(m
, "\n");