4 * Processor and Memory placement constraints for sets of tasks.
6 * Copyright (C) 2003 BULL SA.
7 * Copyright (C) 2004-2007 Silicon Graphics, Inc.
8 * Copyright (C) 2006 Google, Inc
10 * Portions derived from Patrick Mochel's sysfs code.
11 * sysfs is Copyright (c) 2001-3 Patrick Mochel
13 * 2003-10-10 Written by Simon Derr.
14 * 2003-10-22 Updates by Stephen Hemminger.
15 * 2004 May-July Rework by Paul Jackson.
16 * 2006 Rework by Paul Menage to use generic cgroups
17 * 2008 Rework of the scheduler domains and CPU hotplug handling
20 * This file is subject to the terms and conditions of the GNU General Public
21 * License. See the file COPYING in the main directory of the Linux
22 * distribution for more details.
25 #include <linux/cpu.h>
26 #include <linux/cpumask.h>
27 #include <linux/cpuset.h>
28 #include <linux/err.h>
29 #include <linux/errno.h>
30 #include <linux/file.h>
32 #include <linux/init.h>
33 #include <linux/interrupt.h>
34 #include <linux/kernel.h>
35 #include <linux/kmod.h>
36 #include <linux/list.h>
37 #include <linux/mempolicy.h>
39 #include <linux/memory.h>
40 #include <linux/export.h>
41 #include <linux/mount.h>
42 #include <linux/namei.h>
43 #include <linux/pagemap.h>
44 #include <linux/proc_fs.h>
45 #include <linux/rcupdate.h>
46 #include <linux/sched.h>
47 #include <linux/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 <linux/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 hierarchy */
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
),
127 static inline bool task_has_mempolicy(struct task_struct
*task
)
129 return task
->mempolicy
;
132 static inline bool task_has_mempolicy(struct task_struct
*task
)
139 /* bits in struct cpuset flags field */
145 CS_SCHED_LOAD_BALANCE
,
150 /* convenient tests for these bits */
151 static inline int is_cpu_exclusive(const struct cpuset
*cs
)
153 return test_bit(CS_CPU_EXCLUSIVE
, &cs
->flags
);
156 static inline int is_mem_exclusive(const struct cpuset
*cs
)
158 return test_bit(CS_MEM_EXCLUSIVE
, &cs
->flags
);
161 static inline int is_mem_hardwall(const struct cpuset
*cs
)
163 return test_bit(CS_MEM_HARDWALL
, &cs
->flags
);
166 static inline int is_sched_load_balance(const struct cpuset
*cs
)
168 return test_bit(CS_SCHED_LOAD_BALANCE
, &cs
->flags
);
171 static inline int is_memory_migrate(const struct cpuset
*cs
)
173 return test_bit(CS_MEMORY_MIGRATE
, &cs
->flags
);
176 static inline int is_spread_page(const struct cpuset
*cs
)
178 return test_bit(CS_SPREAD_PAGE
, &cs
->flags
);
181 static inline int is_spread_slab(const struct cpuset
*cs
)
183 return test_bit(CS_SPREAD_SLAB
, &cs
->flags
);
186 static struct cpuset top_cpuset
= {
187 .flags
= ((1 << CS_CPU_EXCLUSIVE
) | (1 << CS_MEM_EXCLUSIVE
)),
191 * There are two global mutexes guarding cpuset structures. The first
192 * is the main control groups cgroup_mutex, accessed via
193 * cgroup_lock()/cgroup_unlock(). The second is the cpuset-specific
194 * callback_mutex, below. They can nest. It is ok to first take
195 * cgroup_mutex, then nest callback_mutex. We also require taking
196 * task_lock() when dereferencing a task's cpuset pointer. See "The
197 * task_lock() exception", at the end of this comment.
199 * A task must hold both mutexes to modify cpusets. If a task
200 * holds cgroup_mutex, then it blocks others wanting that mutex,
201 * ensuring that it is the only task able to also acquire callback_mutex
202 * and be able to modify cpusets. It can perform various checks on
203 * the cpuset structure first, knowing nothing will change. It can
204 * also allocate memory while just holding cgroup_mutex. While it is
205 * performing these checks, various callback routines can briefly
206 * acquire callback_mutex to query cpusets. Once it is ready to make
207 * the changes, it takes callback_mutex, blocking everyone else.
209 * Calls to the kernel memory allocator can not be made while holding
210 * callback_mutex, as that would risk double tripping on callback_mutex
211 * from one of the callbacks into the cpuset code from within
214 * If a task is only holding callback_mutex, then it has read-only
217 * Now, the task_struct fields mems_allowed and mempolicy may be changed
218 * by other task, we use alloc_lock in the task_struct fields to protect
221 * The cpuset_common_file_read() handlers only hold callback_mutex across
222 * small pieces of code, such as when reading out possibly multi-word
223 * cpumasks and nodemasks.
225 * Accessing a task's cpuset should be done in accordance with the
226 * guidelines for accessing subsystem state in kernel/cgroup.c
229 static DEFINE_MUTEX(callback_mutex
);
232 * cpuset_buffer_lock protects both the cpuset_name and cpuset_nodelist
233 * buffers. They are statically allocated to prevent using excess stack
234 * when calling cpuset_print_task_mems_allowed().
236 #define CPUSET_NAME_LEN (128)
237 #define CPUSET_NODELIST_LEN (256)
238 static char cpuset_name
[CPUSET_NAME_LEN
];
239 static char cpuset_nodelist
[CPUSET_NODELIST_LEN
];
240 static DEFINE_SPINLOCK(cpuset_buffer_lock
);
243 * This is ugly, but preserves the userspace API for existing cpuset
244 * users. If someone tries to mount the "cpuset" filesystem, we
245 * silently switch it to mount "cgroup" instead
247 static struct dentry
*cpuset_mount(struct file_system_type
*fs_type
,
248 int flags
, const char *unused_dev_name
, void *data
)
250 struct file_system_type
*cgroup_fs
= get_fs_type("cgroup");
251 struct dentry
*ret
= ERR_PTR(-ENODEV
);
255 "release_agent=/sbin/cpuset_release_agent";
256 ret
= cgroup_fs
->mount(cgroup_fs
, flags
,
257 unused_dev_name
, mountopts
);
258 put_filesystem(cgroup_fs
);
263 static struct file_system_type cpuset_fs_type
= {
265 .mount
= cpuset_mount
,
269 * Return in pmask the portion of a cpusets's cpus_allowed that
270 * are online. If none are online, walk up the cpuset hierarchy
271 * until we find one that does have some online cpus. If we get
272 * all the way to the top and still haven't found any online cpus,
273 * return cpu_online_map. Or if passed a NULL cs from an exit'ing
274 * task, return cpu_online_map.
276 * One way or another, we guarantee to return some non-empty subset
279 * Call with callback_mutex held.
282 static void guarantee_online_cpus(const struct cpuset
*cs
,
283 struct cpumask
*pmask
)
285 while (cs
&& !cpumask_intersects(cs
->cpus_allowed
, cpu_online_mask
))
288 cpumask_and(pmask
, cs
->cpus_allowed
, cpu_online_mask
);
290 cpumask_copy(pmask
, cpu_online_mask
);
291 BUG_ON(!cpumask_intersects(pmask
, cpu_online_mask
));
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 * update task's spread flag if cpuset's page/slab spread flag is set
323 * Called with callback_mutex/cgroup_mutex held
325 static void cpuset_update_task_spread_flag(struct cpuset
*cs
,
326 struct task_struct
*tsk
)
328 if (is_spread_page(cs
))
329 tsk
->flags
|= PF_SPREAD_PAGE
;
331 tsk
->flags
&= ~PF_SPREAD_PAGE
;
332 if (is_spread_slab(cs
))
333 tsk
->flags
|= PF_SPREAD_SLAB
;
335 tsk
->flags
&= ~PF_SPREAD_SLAB
;
339 * is_cpuset_subset(p, q) - Is cpuset p a subset of cpuset q?
341 * One cpuset is a subset of another if all its allowed CPUs and
342 * Memory Nodes are a subset of the other, and its exclusive flags
343 * are only set if the other's are set. Call holding cgroup_mutex.
346 static int is_cpuset_subset(const struct cpuset
*p
, const struct cpuset
*q
)
348 return cpumask_subset(p
->cpus_allowed
, q
->cpus_allowed
) &&
349 nodes_subset(p
->mems_allowed
, q
->mems_allowed
) &&
350 is_cpu_exclusive(p
) <= is_cpu_exclusive(q
) &&
351 is_mem_exclusive(p
) <= is_mem_exclusive(q
);
355 * alloc_trial_cpuset - allocate a trial cpuset
356 * @cs: the cpuset that the trial cpuset duplicates
358 static struct cpuset
*alloc_trial_cpuset(const struct cpuset
*cs
)
360 struct cpuset
*trial
;
362 trial
= kmemdup(cs
, sizeof(*cs
), GFP_KERNEL
);
366 if (!alloc_cpumask_var(&trial
->cpus_allowed
, GFP_KERNEL
)) {
370 cpumask_copy(trial
->cpus_allowed
, cs
->cpus_allowed
);
376 * free_trial_cpuset - free the trial cpuset
377 * @trial: the trial cpuset to be freed
379 static void free_trial_cpuset(struct cpuset
*trial
)
381 free_cpumask_var(trial
->cpus_allowed
);
386 * validate_change() - Used to validate that any proposed cpuset change
387 * follows the structural rules for cpusets.
389 * If we replaced the flag and mask values of the current cpuset
390 * (cur) with those values in the trial cpuset (trial), would
391 * our various subset and exclusive rules still be valid? Presumes
394 * 'cur' is the address of an actual, in-use cpuset. Operations
395 * such as list traversal that depend on the actual address of the
396 * cpuset in the list must use cur below, not trial.
398 * 'trial' is the address of bulk structure copy of cur, with
399 * perhaps one or more of the fields cpus_allowed, mems_allowed,
400 * or flags changed to new, trial values.
402 * Return 0 if valid, -errno if not.
405 static int validate_change(const struct cpuset
*cur
, const struct cpuset
*trial
)
408 struct cpuset
*c
, *par
;
410 /* Each of our child cpusets must be a subset of us */
411 list_for_each_entry(cont
, &cur
->css
.cgroup
->children
, sibling
) {
412 if (!is_cpuset_subset(cgroup_cs(cont
), trial
))
416 /* Remaining checks don't apply to root cpuset */
417 if (cur
== &top_cpuset
)
422 /* We must be a subset of our parent cpuset */
423 if (!is_cpuset_subset(trial
, par
))
427 * If either I or some sibling (!= me) is exclusive, we can't
430 list_for_each_entry(cont
, &par
->css
.cgroup
->children
, sibling
) {
432 if ((is_cpu_exclusive(trial
) || is_cpu_exclusive(c
)) &&
434 cpumask_intersects(trial
->cpus_allowed
, c
->cpus_allowed
))
436 if ((is_mem_exclusive(trial
) || is_mem_exclusive(c
)) &&
438 nodes_intersects(trial
->mems_allowed
, c
->mems_allowed
))
442 /* Cpusets with tasks can't have empty cpus_allowed or mems_allowed */
443 if (cgroup_task_count(cur
->css
.cgroup
)) {
444 if (cpumask_empty(trial
->cpus_allowed
) ||
445 nodes_empty(trial
->mems_allowed
)) {
455 * Helper routine for generate_sched_domains().
456 * Do cpusets a, b have overlapping cpus_allowed masks?
458 static int cpusets_overlap(struct cpuset
*a
, struct cpuset
*b
)
460 return cpumask_intersects(a
->cpus_allowed
, b
->cpus_allowed
);
464 update_domain_attr(struct sched_domain_attr
*dattr
, struct cpuset
*c
)
466 if (dattr
->relax_domain_level
< c
->relax_domain_level
)
467 dattr
->relax_domain_level
= c
->relax_domain_level
;
472 update_domain_attr_tree(struct sched_domain_attr
*dattr
, struct cpuset
*c
)
476 list_add(&c
->stack_list
, &q
);
477 while (!list_empty(&q
)) {
480 struct cpuset
*child
;
482 cp
= list_first_entry(&q
, struct cpuset
, stack_list
);
485 if (cpumask_empty(cp
->cpus_allowed
))
488 if (is_sched_load_balance(cp
))
489 update_domain_attr(dattr
, cp
);
491 list_for_each_entry(cont
, &cp
->css
.cgroup
->children
, sibling
) {
492 child
= cgroup_cs(cont
);
493 list_add_tail(&child
->stack_list
, &q
);
499 * generate_sched_domains()
501 * This function builds a partial partition of the systems CPUs
502 * A 'partial partition' is a set of non-overlapping subsets whose
503 * union is a subset of that set.
504 * The output of this function needs to be passed to kernel/sched.c
505 * partition_sched_domains() routine, which will rebuild the scheduler's
506 * load balancing domains (sched domains) as specified by that partial
509 * See "What is sched_load_balance" in Documentation/cgroups/cpusets.txt
510 * for a background explanation of this.
512 * Does not return errors, on the theory that the callers of this
513 * routine would rather not worry about failures to rebuild sched
514 * domains when operating in the severe memory shortage situations
515 * that could cause allocation failures below.
517 * Must be called with cgroup_lock held.
519 * The three key local variables below are:
520 * q - a linked-list queue of cpuset pointers, used to implement a
521 * top-down scan of all cpusets. This scan loads a pointer
522 * to each cpuset marked is_sched_load_balance into the
523 * array 'csa'. For our purposes, rebuilding the schedulers
524 * sched domains, we can ignore !is_sched_load_balance cpusets.
525 * csa - (for CpuSet Array) Array of pointers to all the cpusets
526 * that need to be load balanced, for convenient iterative
527 * access by the subsequent code that finds the best partition,
528 * i.e the set of domains (subsets) of CPUs such that the
529 * cpus_allowed of every cpuset marked is_sched_load_balance
530 * is a subset of one of these domains, while there are as
531 * many such domains as possible, each as small as possible.
532 * doms - Conversion of 'csa' to an array of cpumasks, for passing to
533 * the kernel/sched.c routine partition_sched_domains() in a
534 * convenient format, that can be easily compared to the prior
535 * value to determine what partition elements (sched domains)
536 * were changed (added or removed.)
538 * Finding the best partition (set of domains):
539 * The triple nested loops below over i, j, k scan over the
540 * load balanced cpusets (using the array of cpuset pointers in
541 * csa[]) looking for pairs of cpusets that have overlapping
542 * cpus_allowed, but which don't have the same 'pn' partition
543 * number and gives them in the same partition number. It keeps
544 * looping on the 'restart' label until it can no longer find
547 * The union of the cpus_allowed masks from the set of
548 * all cpusets having the same 'pn' value then form the one
549 * element of the partition (one sched domain) to be passed to
550 * partition_sched_domains().
552 static int generate_sched_domains(cpumask_var_t
**domains
,
553 struct sched_domain_attr
**attributes
)
555 LIST_HEAD(q
); /* queue of cpusets to be scanned */
556 struct cpuset
*cp
; /* scans q */
557 struct cpuset
**csa
; /* array of all cpuset ptrs */
558 int csn
; /* how many cpuset ptrs in csa so far */
559 int i
, j
, k
; /* indices for partition finding loops */
560 cpumask_var_t
*doms
; /* resulting partition; i.e. sched domains */
561 struct sched_domain_attr
*dattr
; /* attributes for custom domains */
562 int ndoms
= 0; /* number of sched domains in result */
563 int nslot
; /* next empty doms[] struct cpumask slot */
569 /* Special case for the 99% of systems with one, full, sched domain */
570 if (is_sched_load_balance(&top_cpuset
)) {
572 doms
= alloc_sched_domains(ndoms
);
576 dattr
= kmalloc(sizeof(struct sched_domain_attr
), GFP_KERNEL
);
578 *dattr
= SD_ATTR_INIT
;
579 update_domain_attr_tree(dattr
, &top_cpuset
);
581 cpumask_copy(doms
[0], top_cpuset
.cpus_allowed
);
586 csa
= kmalloc(number_of_cpusets
* sizeof(cp
), GFP_KERNEL
);
591 list_add(&top_cpuset
.stack_list
, &q
);
592 while (!list_empty(&q
)) {
594 struct cpuset
*child
; /* scans child cpusets of cp */
596 cp
= list_first_entry(&q
, struct cpuset
, stack_list
);
599 if (cpumask_empty(cp
->cpus_allowed
))
603 * All child cpusets contain a subset of the parent's cpus, so
604 * just skip them, and then we call update_domain_attr_tree()
605 * to calc relax_domain_level of the corresponding sched
608 if (is_sched_load_balance(cp
)) {
613 list_for_each_entry(cont
, &cp
->css
.cgroup
->children
, sibling
) {
614 child
= cgroup_cs(cont
);
615 list_add_tail(&child
->stack_list
, &q
);
619 for (i
= 0; i
< csn
; i
++)
624 /* Find the best partition (set of sched domains) */
625 for (i
= 0; i
< csn
; i
++) {
626 struct cpuset
*a
= csa
[i
];
629 for (j
= 0; j
< csn
; j
++) {
630 struct cpuset
*b
= csa
[j
];
633 if (apn
!= bpn
&& cpusets_overlap(a
, b
)) {
634 for (k
= 0; k
< csn
; k
++) {
635 struct cpuset
*c
= csa
[k
];
640 ndoms
--; /* one less element */
647 * Now we know how many domains to create.
648 * Convert <csn, csa> to <ndoms, doms> and populate cpu masks.
650 doms
= alloc_sched_domains(ndoms
);
655 * The rest of the code, including the scheduler, can deal with
656 * dattr==NULL case. No need to abort if alloc fails.
658 dattr
= kmalloc(ndoms
* sizeof(struct sched_domain_attr
), GFP_KERNEL
);
660 for (nslot
= 0, i
= 0; i
< csn
; i
++) {
661 struct cpuset
*a
= csa
[i
];
666 /* Skip completed partitions */
672 if (nslot
== ndoms
) {
673 static int warnings
= 10;
676 "rebuild_sched_domains confused:"
677 " nslot %d, ndoms %d, csn %d, i %d,"
679 nslot
, ndoms
, csn
, i
, apn
);
687 *(dattr
+ nslot
) = SD_ATTR_INIT
;
688 for (j
= i
; j
< csn
; j
++) {
689 struct cpuset
*b
= csa
[j
];
692 cpumask_or(dp
, dp
, b
->cpus_allowed
);
694 update_domain_attr_tree(dattr
+ nslot
, b
);
696 /* Done with this partition */
702 BUG_ON(nslot
!= ndoms
);
708 * Fallback to the default domain if kmalloc() failed.
709 * See comments in partition_sched_domains().
720 * Rebuild scheduler domains.
722 * Call with neither cgroup_mutex held nor within get_online_cpus().
723 * Takes both cgroup_mutex and get_online_cpus().
725 * Cannot be directly called from cpuset code handling changes
726 * to the cpuset pseudo-filesystem, because it cannot be called
727 * from code that already holds cgroup_mutex.
729 static void do_rebuild_sched_domains(struct work_struct
*unused
)
731 struct sched_domain_attr
*attr
;
737 /* Generate domain masks and attrs */
739 ndoms
= generate_sched_domains(&doms
, &attr
);
742 /* Have scheduler rebuild the domains */
743 partition_sched_domains(ndoms
, doms
, attr
);
747 #else /* !CONFIG_SMP */
748 static void do_rebuild_sched_domains(struct work_struct
*unused
)
752 static int generate_sched_domains(cpumask_var_t
**domains
,
753 struct sched_domain_attr
**attributes
)
758 #endif /* CONFIG_SMP */
760 static DECLARE_WORK(rebuild_sched_domains_work
, do_rebuild_sched_domains
);
763 * Rebuild scheduler domains, asynchronously via workqueue.
765 * If the flag 'sched_load_balance' of any cpuset with non-empty
766 * 'cpus' changes, or if the 'cpus' allowed changes in any cpuset
767 * which has that flag enabled, or if any cpuset with a non-empty
768 * 'cpus' is removed, then call this routine to rebuild the
769 * scheduler's dynamic sched domains.
771 * The rebuild_sched_domains() and partition_sched_domains()
772 * routines must nest cgroup_lock() inside get_online_cpus(),
773 * but such cpuset changes as these must nest that locking the
774 * other way, holding cgroup_lock() for much of the code.
776 * So in order to avoid an ABBA deadlock, the cpuset code handling
777 * these user changes delegates the actual sched domain rebuilding
778 * to a separate workqueue thread, which ends up processing the
779 * above do_rebuild_sched_domains() function.
781 static void async_rebuild_sched_domains(void)
783 queue_work(cpuset_wq
, &rebuild_sched_domains_work
);
787 * Accomplishes the same scheduler domain rebuild as the above
788 * async_rebuild_sched_domains(), however it directly calls the
789 * rebuild routine synchronously rather than calling it via an
790 * asynchronous work thread.
792 * This can only be called from code that is not holding
793 * cgroup_mutex (not nested in a cgroup_lock() call.)
795 void rebuild_sched_domains(void)
797 do_rebuild_sched_domains(NULL
);
801 * cpuset_test_cpumask - test a task's cpus_allowed versus its cpuset's
803 * @scan: struct cgroup_scanner contained in its struct cpuset_hotplug_scanner
805 * Call with cgroup_mutex held. May take callback_mutex during call.
806 * Called for each task in a cgroup by cgroup_scan_tasks().
807 * Return nonzero if this tasks's cpus_allowed mask should be changed (in other
808 * words, if its mask is not equal to its cpuset's mask).
810 static int cpuset_test_cpumask(struct task_struct
*tsk
,
811 struct cgroup_scanner
*scan
)
813 return !cpumask_equal(&tsk
->cpus_allowed
,
814 (cgroup_cs(scan
->cg
))->cpus_allowed
);
818 * cpuset_change_cpumask - make a task's cpus_allowed the same as its cpuset's
820 * @scan: struct cgroup_scanner containing the cgroup of the task
822 * Called by cgroup_scan_tasks() for each task in a cgroup whose
823 * cpus_allowed mask needs to be changed.
825 * We don't need to re-check for the cgroup/cpuset membership, since we're
826 * holding cgroup_lock() at this point.
828 static void cpuset_change_cpumask(struct task_struct
*tsk
,
829 struct cgroup_scanner
*scan
)
831 set_cpus_allowed_ptr(tsk
, ((cgroup_cs(scan
->cg
))->cpus_allowed
));
835 * update_tasks_cpumask - Update the cpumasks of tasks in the cpuset.
836 * @cs: the cpuset in which each task's cpus_allowed mask needs to be changed
837 * @heap: if NULL, defer allocating heap memory to cgroup_scan_tasks()
839 * Called with cgroup_mutex held
841 * The cgroup_scan_tasks() function will scan all the tasks in a cgroup,
842 * calling callback functions for each.
844 * No return value. It's guaranteed that cgroup_scan_tasks() always returns 0
847 static void update_tasks_cpumask(struct cpuset
*cs
, struct ptr_heap
*heap
)
849 struct cgroup_scanner scan
;
851 scan
.cg
= cs
->css
.cgroup
;
852 scan
.test_task
= cpuset_test_cpumask
;
853 scan
.process_task
= cpuset_change_cpumask
;
855 cgroup_scan_tasks(&scan
);
859 * update_cpumask - update the cpus_allowed mask of a cpuset and all tasks in it
860 * @cs: the cpuset to consider
861 * @buf: buffer of cpu numbers written to this cpuset
863 static int update_cpumask(struct cpuset
*cs
, struct cpuset
*trialcs
,
866 struct ptr_heap heap
;
868 int is_load_balanced
;
870 /* top_cpuset.cpus_allowed tracks cpu_online_map; it's read-only */
871 if (cs
== &top_cpuset
)
875 * An empty cpus_allowed is ok only if the cpuset has no tasks.
876 * Since cpulist_parse() fails on an empty mask, we special case
877 * that parsing. The validate_change() call ensures that cpusets
878 * with tasks have cpus.
881 cpumask_clear(trialcs
->cpus_allowed
);
883 retval
= cpulist_parse(buf
, trialcs
->cpus_allowed
);
887 if (!cpumask_subset(trialcs
->cpus_allowed
, cpu_active_mask
))
890 retval
= validate_change(cs
, trialcs
);
894 /* Nothing to do if the cpus didn't change */
895 if (cpumask_equal(cs
->cpus_allowed
, trialcs
->cpus_allowed
))
898 retval
= heap_init(&heap
, PAGE_SIZE
, GFP_KERNEL
, NULL
);
902 is_load_balanced
= is_sched_load_balance(trialcs
);
904 mutex_lock(&callback_mutex
);
905 cpumask_copy(cs
->cpus_allowed
, trialcs
->cpus_allowed
);
906 mutex_unlock(&callback_mutex
);
909 * Scan tasks in the cpuset, and update the cpumasks of any
910 * that need an update.
912 update_tasks_cpumask(cs
, &heap
);
916 if (is_load_balanced
)
917 async_rebuild_sched_domains();
924 * Migrate memory region from one set of nodes to another.
926 * Temporarilly set tasks mems_allowed to target nodes of migration,
927 * so that the migration code can allocate pages on these nodes.
929 * Call holding cgroup_mutex, so current's cpuset won't change
930 * during this call, as manage_mutex holds off any cpuset_attach()
931 * calls. Therefore we don't need to take task_lock around the
932 * call to guarantee_online_mems(), as we know no one is changing
935 * While the mm_struct we are migrating is typically from some
936 * other task, the task_struct mems_allowed that we are hacking
937 * is for our current task, which must allocate new pages for that
938 * migrating memory region.
941 static void cpuset_migrate_mm(struct mm_struct
*mm
, const nodemask_t
*from
,
942 const nodemask_t
*to
)
944 struct task_struct
*tsk
= current
;
946 tsk
->mems_allowed
= *to
;
948 do_migrate_pages(mm
, from
, to
, MPOL_MF_MOVE_ALL
);
950 guarantee_online_mems(task_cs(tsk
),&tsk
->mems_allowed
);
954 * cpuset_change_task_nodemask - change task's mems_allowed and mempolicy
955 * @tsk: the task to change
956 * @newmems: new nodes that the task will be set
958 * In order to avoid seeing no nodes if the old and new nodes are disjoint,
959 * we structure updates as setting all new allowed nodes, then clearing newly
962 static void cpuset_change_task_nodemask(struct task_struct
*tsk
,
969 * Allow tasks that have access to memory reserves because they have
970 * been OOM killed to get memory anywhere.
972 if (unlikely(test_thread_flag(TIF_MEMDIE
)))
974 if (current
->flags
& PF_EXITING
) /* Let dying task have memory */
979 * Determine if a loop is necessary if another thread is doing
980 * get_mems_allowed(). If at least one node remains unchanged and
981 * tsk does not have a mempolicy, then an empty nodemask will not be
982 * possible when mems_allowed is larger than a word.
984 need_loop
= task_has_mempolicy(tsk
) ||
985 !nodes_intersects(*newmems
, tsk
->mems_allowed
);
986 nodes_or(tsk
->mems_allowed
, tsk
->mems_allowed
, *newmems
);
987 mpol_rebind_task(tsk
, newmems
, MPOL_REBIND_STEP1
);
990 * ensure checking ->mems_allowed_change_disable after setting all new
993 * the read-side task can see an nodemask with new allowed nodes and
994 * old allowed nodes. and if it allocates page when cpuset clears newly
995 * disallowed ones continuous, it can see the new allowed bits.
997 * And if setting all new allowed nodes is after the checking, setting
998 * all new allowed nodes and clearing newly disallowed ones will be done
999 * continuous, and the read-side task may find no node to alloc page.
1004 * Allocation of memory is very fast, we needn't sleep when waiting
1005 * for the read-side.
1007 while (need_loop
&& ACCESS_ONCE(tsk
->mems_allowed_change_disable
)) {
1009 if (!task_curr(tsk
))
1015 * ensure checking ->mems_allowed_change_disable before clearing all new
1018 * if clearing newly disallowed bits before the checking, the read-side
1019 * task may find no node to alloc page.
1023 mpol_rebind_task(tsk
, newmems
, MPOL_REBIND_STEP2
);
1024 tsk
->mems_allowed
= *newmems
;
1029 * Update task's mems_allowed and rebind its mempolicy and vmas' mempolicy
1030 * of it to cpuset's new mems_allowed, and migrate pages to new nodes if
1031 * memory_migrate flag is set. Called with cgroup_mutex held.
1033 static void cpuset_change_nodemask(struct task_struct
*p
,
1034 struct cgroup_scanner
*scan
)
1036 struct mm_struct
*mm
;
1039 const nodemask_t
*oldmem
= scan
->data
;
1040 static nodemask_t newmems
; /* protected by cgroup_mutex */
1042 cs
= cgroup_cs(scan
->cg
);
1043 guarantee_online_mems(cs
, &newmems
);
1045 cpuset_change_task_nodemask(p
, &newmems
);
1047 mm
= get_task_mm(p
);
1051 migrate
= is_memory_migrate(cs
);
1053 mpol_rebind_mm(mm
, &cs
->mems_allowed
);
1055 cpuset_migrate_mm(mm
, oldmem
, &cs
->mems_allowed
);
1059 static void *cpuset_being_rebound
;
1062 * update_tasks_nodemask - Update the nodemasks of tasks in the cpuset.
1063 * @cs: the cpuset in which each task's mems_allowed mask needs to be changed
1064 * @oldmem: old mems_allowed of cpuset cs
1065 * @heap: if NULL, defer allocating heap memory to cgroup_scan_tasks()
1067 * Called with cgroup_mutex held
1068 * No return value. It's guaranteed that cgroup_scan_tasks() always returns 0
1071 static void update_tasks_nodemask(struct cpuset
*cs
, const nodemask_t
*oldmem
,
1072 struct ptr_heap
*heap
)
1074 struct cgroup_scanner scan
;
1076 cpuset_being_rebound
= cs
; /* causes mpol_dup() rebind */
1078 scan
.cg
= cs
->css
.cgroup
;
1079 scan
.test_task
= NULL
;
1080 scan
.process_task
= cpuset_change_nodemask
;
1082 scan
.data
= (nodemask_t
*)oldmem
;
1085 * The mpol_rebind_mm() call takes mmap_sem, which we couldn't
1086 * take while holding tasklist_lock. Forks can happen - the
1087 * mpol_dup() cpuset_being_rebound check will catch such forks,
1088 * and rebind their vma mempolicies too. Because we still hold
1089 * the global cgroup_mutex, we know that no other rebind effort
1090 * will be contending for the global variable cpuset_being_rebound.
1091 * It's ok if we rebind the same mm twice; mpol_rebind_mm()
1092 * is idempotent. Also migrate pages in each mm to new nodes.
1094 cgroup_scan_tasks(&scan
);
1096 /* We're done rebinding vmas to this cpuset's new mems_allowed. */
1097 cpuset_being_rebound
= NULL
;
1101 * Handle user request to change the 'mems' memory placement
1102 * of a cpuset. Needs to validate the request, update the
1103 * cpusets mems_allowed, and for each task in the cpuset,
1104 * update mems_allowed and rebind task's mempolicy and any vma
1105 * mempolicies and if the cpuset is marked 'memory_migrate',
1106 * migrate the tasks pages to the new memory.
1108 * Call with cgroup_mutex held. May take callback_mutex during call.
1109 * Will take tasklist_lock, scan tasklist for tasks in cpuset cs,
1110 * lock each such tasks mm->mmap_sem, scan its vma's and rebind
1111 * their mempolicies to the cpusets new mems_allowed.
1113 static int update_nodemask(struct cpuset
*cs
, struct cpuset
*trialcs
,
1116 NODEMASK_ALLOC(nodemask_t
, oldmem
, GFP_KERNEL
);
1118 struct ptr_heap heap
;
1124 * top_cpuset.mems_allowed tracks node_stats[N_HIGH_MEMORY];
1127 if (cs
== &top_cpuset
) {
1133 * An empty mems_allowed is ok iff there are no tasks in the cpuset.
1134 * Since nodelist_parse() fails on an empty mask, we special case
1135 * that parsing. The validate_change() call ensures that cpusets
1136 * with tasks have memory.
1139 nodes_clear(trialcs
->mems_allowed
);
1141 retval
= nodelist_parse(buf
, trialcs
->mems_allowed
);
1145 if (!nodes_subset(trialcs
->mems_allowed
,
1146 node_states
[N_HIGH_MEMORY
])) {
1151 *oldmem
= cs
->mems_allowed
;
1152 if (nodes_equal(*oldmem
, trialcs
->mems_allowed
)) {
1153 retval
= 0; /* Too easy - nothing to do */
1156 retval
= validate_change(cs
, trialcs
);
1160 retval
= heap_init(&heap
, PAGE_SIZE
, GFP_KERNEL
, NULL
);
1164 mutex_lock(&callback_mutex
);
1165 cs
->mems_allowed
= trialcs
->mems_allowed
;
1166 mutex_unlock(&callback_mutex
);
1168 update_tasks_nodemask(cs
, oldmem
, &heap
);
1172 NODEMASK_FREE(oldmem
);
1176 int current_cpuset_is_being_rebound(void)
1178 return task_cs(current
) == cpuset_being_rebound
;
1181 static int update_relax_domain_level(struct cpuset
*cs
, s64 val
)
1184 if (val
< -1 || val
>= sched_domain_level_max
)
1188 if (val
!= cs
->relax_domain_level
) {
1189 cs
->relax_domain_level
= val
;
1190 if (!cpumask_empty(cs
->cpus_allowed
) &&
1191 is_sched_load_balance(cs
))
1192 async_rebuild_sched_domains();
1199 * cpuset_change_flag - make a task's spread flags the same as its cpuset's
1200 * @tsk: task to be updated
1201 * @scan: struct cgroup_scanner containing the cgroup of the task
1203 * Called by cgroup_scan_tasks() for each task in a cgroup.
1205 * We don't need to re-check for the cgroup/cpuset membership, since we're
1206 * holding cgroup_lock() at this point.
1208 static void cpuset_change_flag(struct task_struct
*tsk
,
1209 struct cgroup_scanner
*scan
)
1211 cpuset_update_task_spread_flag(cgroup_cs(scan
->cg
), tsk
);
1215 * update_tasks_flags - update the spread flags of tasks in the cpuset.
1216 * @cs: the cpuset in which each task's spread flags needs to be changed
1217 * @heap: if NULL, defer allocating heap memory to cgroup_scan_tasks()
1219 * Called with cgroup_mutex held
1221 * The cgroup_scan_tasks() function will scan all the tasks in a cgroup,
1222 * calling callback functions for each.
1224 * No return value. It's guaranteed that cgroup_scan_tasks() always returns 0
1227 static void update_tasks_flags(struct cpuset
*cs
, struct ptr_heap
*heap
)
1229 struct cgroup_scanner scan
;
1231 scan
.cg
= cs
->css
.cgroup
;
1232 scan
.test_task
= NULL
;
1233 scan
.process_task
= cpuset_change_flag
;
1235 cgroup_scan_tasks(&scan
);
1239 * update_flag - read a 0 or a 1 in a file and update associated flag
1240 * bit: the bit to update (see cpuset_flagbits_t)
1241 * cs: the cpuset to update
1242 * turning_on: whether the flag is being set or cleared
1244 * Call with cgroup_mutex held.
1247 static int update_flag(cpuset_flagbits_t bit
, struct cpuset
*cs
,
1250 struct cpuset
*trialcs
;
1251 int balance_flag_changed
;
1252 int spread_flag_changed
;
1253 struct ptr_heap heap
;
1256 trialcs
= alloc_trial_cpuset(cs
);
1261 set_bit(bit
, &trialcs
->flags
);
1263 clear_bit(bit
, &trialcs
->flags
);
1265 err
= validate_change(cs
, trialcs
);
1269 err
= heap_init(&heap
, PAGE_SIZE
, GFP_KERNEL
, NULL
);
1273 balance_flag_changed
= (is_sched_load_balance(cs
) !=
1274 is_sched_load_balance(trialcs
));
1276 spread_flag_changed
= ((is_spread_slab(cs
) != is_spread_slab(trialcs
))
1277 || (is_spread_page(cs
) != is_spread_page(trialcs
)));
1279 mutex_lock(&callback_mutex
);
1280 cs
->flags
= trialcs
->flags
;
1281 mutex_unlock(&callback_mutex
);
1283 if (!cpumask_empty(trialcs
->cpus_allowed
) && balance_flag_changed
)
1284 async_rebuild_sched_domains();
1286 if (spread_flag_changed
)
1287 update_tasks_flags(cs
, &heap
);
1290 free_trial_cpuset(trialcs
);
1295 * Frequency meter - How fast is some event occurring?
1297 * These routines manage a digitally filtered, constant time based,
1298 * event frequency meter. There are four routines:
1299 * fmeter_init() - initialize a frequency meter.
1300 * fmeter_markevent() - called each time the event happens.
1301 * fmeter_getrate() - returns the recent rate of such events.
1302 * fmeter_update() - internal routine used to update fmeter.
1304 * A common data structure is passed to each of these routines,
1305 * which is used to keep track of the state required to manage the
1306 * frequency meter and its digital filter.
1308 * The filter works on the number of events marked per unit time.
1309 * The filter is single-pole low-pass recursive (IIR). The time unit
1310 * is 1 second. Arithmetic is done using 32-bit integers scaled to
1311 * simulate 3 decimal digits of precision (multiplied by 1000).
1313 * With an FM_COEF of 933, and a time base of 1 second, the filter
1314 * has a half-life of 10 seconds, meaning that if the events quit
1315 * happening, then the rate returned from the fmeter_getrate()
1316 * will be cut in half each 10 seconds, until it converges to zero.
1318 * It is not worth doing a real infinitely recursive filter. If more
1319 * than FM_MAXTICKS ticks have elapsed since the last filter event,
1320 * just compute FM_MAXTICKS ticks worth, by which point the level
1323 * Limit the count of unprocessed events to FM_MAXCNT, so as to avoid
1324 * arithmetic overflow in the fmeter_update() routine.
1326 * Given the simple 32 bit integer arithmetic used, this meter works
1327 * best for reporting rates between one per millisecond (msec) and
1328 * one per 32 (approx) seconds. At constant rates faster than one
1329 * per msec it maxes out at values just under 1,000,000. At constant
1330 * rates between one per msec, and one per second it will stabilize
1331 * to a value N*1000, where N is the rate of events per second.
1332 * At constant rates between one per second and one per 32 seconds,
1333 * it will be choppy, moving up on the seconds that have an event,
1334 * and then decaying until the next event. At rates slower than
1335 * about one in 32 seconds, it decays all the way back to zero between
1339 #define FM_COEF 933 /* coefficient for half-life of 10 secs */
1340 #define FM_MAXTICKS ((time_t)99) /* useless computing more ticks than this */
1341 #define FM_MAXCNT 1000000 /* limit cnt to avoid overflow */
1342 #define FM_SCALE 1000 /* faux fixed point scale */
1344 /* Initialize a frequency meter */
1345 static void fmeter_init(struct fmeter
*fmp
)
1350 spin_lock_init(&fmp
->lock
);
1353 /* Internal meter update - process cnt events and update value */
1354 static void fmeter_update(struct fmeter
*fmp
)
1356 time_t now
= get_seconds();
1357 time_t ticks
= now
- fmp
->time
;
1362 ticks
= min(FM_MAXTICKS
, ticks
);
1364 fmp
->val
= (FM_COEF
* fmp
->val
) / FM_SCALE
;
1367 fmp
->val
+= ((FM_SCALE
- FM_COEF
) * fmp
->cnt
) / FM_SCALE
;
1371 /* Process any previous ticks, then bump cnt by one (times scale). */
1372 static void fmeter_markevent(struct fmeter
*fmp
)
1374 spin_lock(&fmp
->lock
);
1376 fmp
->cnt
= min(FM_MAXCNT
, fmp
->cnt
+ FM_SCALE
);
1377 spin_unlock(&fmp
->lock
);
1380 /* Process any previous ticks, then return current value. */
1381 static int fmeter_getrate(struct fmeter
*fmp
)
1385 spin_lock(&fmp
->lock
);
1388 spin_unlock(&fmp
->lock
);
1393 * Protected by cgroup_lock. The nodemasks must be stored globally because
1394 * dynamically allocating them is not allowed in can_attach, and they must
1395 * persist until attach.
1397 static cpumask_var_t cpus_attach
;
1398 static nodemask_t cpuset_attach_nodemask_from
;
1399 static nodemask_t cpuset_attach_nodemask_to
;
1401 /* Called by cgroups to determine if a cpuset is usable; cgroup_mutex held */
1402 static int cpuset_can_attach(struct cgroup
*cgrp
, struct cgroup_taskset
*tset
)
1404 struct cpuset
*cs
= cgroup_cs(cgrp
);
1405 struct task_struct
*task
;
1408 if (cpumask_empty(cs
->cpus_allowed
) || nodes_empty(cs
->mems_allowed
))
1411 cgroup_taskset_for_each(task
, cgrp
, tset
) {
1413 * Kthreads bound to specific cpus cannot be moved to a new
1414 * cpuset; we cannot change their cpu affinity and
1415 * isolating such threads by their set of allowed nodes is
1416 * unnecessary. Thus, cpusets are not applicable for such
1417 * threads. This prevents checking for success of
1418 * set_cpus_allowed_ptr() on all attached tasks before
1419 * cpus_allowed may be changed.
1421 if (task
->flags
& PF_THREAD_BOUND
)
1423 if ((ret
= security_task_setscheduler(task
)))
1427 /* prepare for attach */
1428 if (cs
== &top_cpuset
)
1429 cpumask_copy(cpus_attach
, cpu_possible_mask
);
1431 guarantee_online_cpus(cs
, cpus_attach
);
1433 guarantee_online_mems(cs
, &cpuset_attach_nodemask_to
);
1438 static void cpuset_attach(struct cgroup
*cgrp
, struct cgroup_taskset
*tset
)
1440 struct mm_struct
*mm
;
1441 struct task_struct
*task
;
1442 struct task_struct
*leader
= cgroup_taskset_first(tset
);
1443 struct cgroup
*oldcgrp
= cgroup_taskset_cur_cgroup(tset
);
1444 struct cpuset
*cs
= cgroup_cs(cgrp
);
1445 struct cpuset
*oldcs
= cgroup_cs(oldcgrp
);
1447 cgroup_taskset_for_each(task
, cgrp
, tset
) {
1449 * can_attach beforehand should guarantee that this doesn't
1450 * fail. TODO: have a better way to handle failure here
1452 WARN_ON_ONCE(set_cpus_allowed_ptr(task
, cpus_attach
));
1454 cpuset_change_task_nodemask(task
, &cpuset_attach_nodemask_to
);
1455 cpuset_update_task_spread_flag(cs
, task
);
1459 * Change mm, possibly for multiple threads in a threadgroup. This is
1460 * expensive and may sleep.
1462 cpuset_attach_nodemask_from
= oldcs
->mems_allowed
;
1463 cpuset_attach_nodemask_to
= cs
->mems_allowed
;
1464 mm
= get_task_mm(leader
);
1466 mpol_rebind_mm(mm
, &cpuset_attach_nodemask_to
);
1467 if (is_memory_migrate(cs
))
1468 cpuset_migrate_mm(mm
, &cpuset_attach_nodemask_from
,
1469 &cpuset_attach_nodemask_to
);
1474 /* The various types of files and directories in a cpuset file system */
1477 FILE_MEMORY_MIGRATE
,
1483 FILE_SCHED_LOAD_BALANCE
,
1484 FILE_SCHED_RELAX_DOMAIN_LEVEL
,
1485 FILE_MEMORY_PRESSURE_ENABLED
,
1486 FILE_MEMORY_PRESSURE
,
1489 } cpuset_filetype_t
;
1491 static int cpuset_write_u64(struct cgroup
*cgrp
, struct cftype
*cft
, u64 val
)
1494 struct cpuset
*cs
= cgroup_cs(cgrp
);
1495 cpuset_filetype_t type
= cft
->private;
1497 if (!cgroup_lock_live_group(cgrp
))
1501 case FILE_CPU_EXCLUSIVE
:
1502 retval
= update_flag(CS_CPU_EXCLUSIVE
, cs
, val
);
1504 case FILE_MEM_EXCLUSIVE
:
1505 retval
= update_flag(CS_MEM_EXCLUSIVE
, cs
, val
);
1507 case FILE_MEM_HARDWALL
:
1508 retval
= update_flag(CS_MEM_HARDWALL
, cs
, val
);
1510 case FILE_SCHED_LOAD_BALANCE
:
1511 retval
= update_flag(CS_SCHED_LOAD_BALANCE
, cs
, val
);
1513 case FILE_MEMORY_MIGRATE
:
1514 retval
= update_flag(CS_MEMORY_MIGRATE
, cs
, val
);
1516 case FILE_MEMORY_PRESSURE_ENABLED
:
1517 cpuset_memory_pressure_enabled
= !!val
;
1519 case FILE_MEMORY_PRESSURE
:
1522 case FILE_SPREAD_PAGE
:
1523 retval
= update_flag(CS_SPREAD_PAGE
, cs
, val
);
1525 case FILE_SPREAD_SLAB
:
1526 retval
= update_flag(CS_SPREAD_SLAB
, cs
, val
);
1536 static int cpuset_write_s64(struct cgroup
*cgrp
, struct cftype
*cft
, s64 val
)
1539 struct cpuset
*cs
= cgroup_cs(cgrp
);
1540 cpuset_filetype_t type
= cft
->private;
1542 if (!cgroup_lock_live_group(cgrp
))
1546 case FILE_SCHED_RELAX_DOMAIN_LEVEL
:
1547 retval
= update_relax_domain_level(cs
, val
);
1558 * Common handling for a write to a "cpus" or "mems" file.
1560 static int cpuset_write_resmask(struct cgroup
*cgrp
, struct cftype
*cft
,
1564 struct cpuset
*cs
= cgroup_cs(cgrp
);
1565 struct cpuset
*trialcs
;
1567 if (!cgroup_lock_live_group(cgrp
))
1570 trialcs
= alloc_trial_cpuset(cs
);
1576 switch (cft
->private) {
1578 retval
= update_cpumask(cs
, trialcs
, buf
);
1581 retval
= update_nodemask(cs
, trialcs
, buf
);
1588 free_trial_cpuset(trialcs
);
1595 * These ascii lists should be read in a single call, by using a user
1596 * buffer large enough to hold the entire map. If read in smaller
1597 * chunks, there is no guarantee of atomicity. Since the display format
1598 * used, list of ranges of sequential numbers, is variable length,
1599 * and since these maps can change value dynamically, one could read
1600 * gibberish by doing partial reads while a list was changing.
1601 * A single large read to a buffer that crosses a page boundary is
1602 * ok, because the result being copied to user land is not recomputed
1603 * across a page fault.
1606 static size_t cpuset_sprintf_cpulist(char *page
, struct cpuset
*cs
)
1610 mutex_lock(&callback_mutex
);
1611 count
= cpulist_scnprintf(page
, PAGE_SIZE
, cs
->cpus_allowed
);
1612 mutex_unlock(&callback_mutex
);
1617 static size_t cpuset_sprintf_memlist(char *page
, struct cpuset
*cs
)
1621 mutex_lock(&callback_mutex
);
1622 count
= nodelist_scnprintf(page
, PAGE_SIZE
, cs
->mems_allowed
);
1623 mutex_unlock(&callback_mutex
);
1628 static ssize_t
cpuset_common_file_read(struct cgroup
*cont
,
1632 size_t nbytes
, loff_t
*ppos
)
1634 struct cpuset
*cs
= cgroup_cs(cont
);
1635 cpuset_filetype_t type
= cft
->private;
1640 if (!(page
= (char *)__get_free_page(GFP_TEMPORARY
)))
1647 s
+= cpuset_sprintf_cpulist(s
, cs
);
1650 s
+= cpuset_sprintf_memlist(s
, cs
);
1658 retval
= simple_read_from_buffer(buf
, nbytes
, ppos
, page
, s
- page
);
1660 free_page((unsigned long)page
);
1664 static u64
cpuset_read_u64(struct cgroup
*cont
, struct cftype
*cft
)
1666 struct cpuset
*cs
= cgroup_cs(cont
);
1667 cpuset_filetype_t type
= cft
->private;
1669 case FILE_CPU_EXCLUSIVE
:
1670 return is_cpu_exclusive(cs
);
1671 case FILE_MEM_EXCLUSIVE
:
1672 return is_mem_exclusive(cs
);
1673 case FILE_MEM_HARDWALL
:
1674 return is_mem_hardwall(cs
);
1675 case FILE_SCHED_LOAD_BALANCE
:
1676 return is_sched_load_balance(cs
);
1677 case FILE_MEMORY_MIGRATE
:
1678 return is_memory_migrate(cs
);
1679 case FILE_MEMORY_PRESSURE_ENABLED
:
1680 return cpuset_memory_pressure_enabled
;
1681 case FILE_MEMORY_PRESSURE
:
1682 return fmeter_getrate(&cs
->fmeter
);
1683 case FILE_SPREAD_PAGE
:
1684 return is_spread_page(cs
);
1685 case FILE_SPREAD_SLAB
:
1686 return is_spread_slab(cs
);
1691 /* Unreachable but makes gcc happy */
1695 static s64
cpuset_read_s64(struct cgroup
*cont
, struct cftype
*cft
)
1697 struct cpuset
*cs
= cgroup_cs(cont
);
1698 cpuset_filetype_t type
= cft
->private;
1700 case FILE_SCHED_RELAX_DOMAIN_LEVEL
:
1701 return cs
->relax_domain_level
;
1706 /* Unrechable but makes gcc happy */
1712 * for the common functions, 'private' gives the type of file
1715 static struct cftype files
[] = {
1718 .read
= cpuset_common_file_read
,
1719 .write_string
= cpuset_write_resmask
,
1720 .max_write_len
= (100U + 6 * NR_CPUS
),
1721 .private = FILE_CPULIST
,
1726 .read
= cpuset_common_file_read
,
1727 .write_string
= cpuset_write_resmask
,
1728 .max_write_len
= (100U + 6 * MAX_NUMNODES
),
1729 .private = FILE_MEMLIST
,
1733 .name
= "cpu_exclusive",
1734 .read_u64
= cpuset_read_u64
,
1735 .write_u64
= cpuset_write_u64
,
1736 .private = FILE_CPU_EXCLUSIVE
,
1740 .name
= "mem_exclusive",
1741 .read_u64
= cpuset_read_u64
,
1742 .write_u64
= cpuset_write_u64
,
1743 .private = FILE_MEM_EXCLUSIVE
,
1747 .name
= "mem_hardwall",
1748 .read_u64
= cpuset_read_u64
,
1749 .write_u64
= cpuset_write_u64
,
1750 .private = FILE_MEM_HARDWALL
,
1754 .name
= "sched_load_balance",
1755 .read_u64
= cpuset_read_u64
,
1756 .write_u64
= cpuset_write_u64
,
1757 .private = FILE_SCHED_LOAD_BALANCE
,
1761 .name
= "sched_relax_domain_level",
1762 .read_s64
= cpuset_read_s64
,
1763 .write_s64
= cpuset_write_s64
,
1764 .private = FILE_SCHED_RELAX_DOMAIN_LEVEL
,
1768 .name
= "memory_migrate",
1769 .read_u64
= cpuset_read_u64
,
1770 .write_u64
= cpuset_write_u64
,
1771 .private = FILE_MEMORY_MIGRATE
,
1775 .name
= "memory_pressure",
1776 .read_u64
= cpuset_read_u64
,
1777 .write_u64
= cpuset_write_u64
,
1778 .private = FILE_MEMORY_PRESSURE
,
1783 .name
= "memory_spread_page",
1784 .read_u64
= cpuset_read_u64
,
1785 .write_u64
= cpuset_write_u64
,
1786 .private = FILE_SPREAD_PAGE
,
1790 .name
= "memory_spread_slab",
1791 .read_u64
= cpuset_read_u64
,
1792 .write_u64
= cpuset_write_u64
,
1793 .private = FILE_SPREAD_SLAB
,
1797 static struct cftype cft_memory_pressure_enabled
= {
1798 .name
= "memory_pressure_enabled",
1799 .read_u64
= cpuset_read_u64
,
1800 .write_u64
= cpuset_write_u64
,
1801 .private = FILE_MEMORY_PRESSURE_ENABLED
,
1804 static int cpuset_populate(struct cgroup_subsys
*ss
, struct cgroup
*cont
)
1808 err
= cgroup_add_files(cont
, ss
, files
, ARRAY_SIZE(files
));
1811 /* memory_pressure_enabled is in root cpuset only */
1813 err
= cgroup_add_file(cont
, ss
,
1814 &cft_memory_pressure_enabled
);
1819 * post_clone() is called during cgroup_create() when the
1820 * clone_children mount argument was specified. The cgroup
1821 * can not yet have any tasks.
1823 * Currently we refuse to set up the cgroup - thereby
1824 * refusing the task to be entered, and as a result refusing
1825 * the sys_unshare() or clone() which initiated it - if any
1826 * sibling cpusets have exclusive cpus or mem.
1828 * If this becomes a problem for some users who wish to
1829 * allow that scenario, then cpuset_post_clone() could be
1830 * changed to grant parent->cpus_allowed-sibling_cpus_exclusive
1831 * (and likewise for mems) to the new cgroup. Called with cgroup_mutex
1834 static void cpuset_post_clone(struct cgroup
*cgroup
)
1836 struct cgroup
*parent
, *child
;
1837 struct cpuset
*cs
, *parent_cs
;
1839 parent
= cgroup
->parent
;
1840 list_for_each_entry(child
, &parent
->children
, sibling
) {
1841 cs
= cgroup_cs(child
);
1842 if (is_mem_exclusive(cs
) || is_cpu_exclusive(cs
))
1845 cs
= cgroup_cs(cgroup
);
1846 parent_cs
= cgroup_cs(parent
);
1848 mutex_lock(&callback_mutex
);
1849 cs
->mems_allowed
= parent_cs
->mems_allowed
;
1850 cpumask_copy(cs
->cpus_allowed
, parent_cs
->cpus_allowed
);
1851 mutex_unlock(&callback_mutex
);
1856 * cpuset_create - create a cpuset
1857 * cont: control group that the new cpuset will be part of
1860 static struct cgroup_subsys_state
*cpuset_create(struct cgroup
*cont
)
1863 struct cpuset
*parent
;
1865 if (!cont
->parent
) {
1866 return &top_cpuset
.css
;
1868 parent
= cgroup_cs(cont
->parent
);
1869 cs
= kmalloc(sizeof(*cs
), GFP_KERNEL
);
1871 return ERR_PTR(-ENOMEM
);
1872 if (!alloc_cpumask_var(&cs
->cpus_allowed
, GFP_KERNEL
)) {
1874 return ERR_PTR(-ENOMEM
);
1878 if (is_spread_page(parent
))
1879 set_bit(CS_SPREAD_PAGE
, &cs
->flags
);
1880 if (is_spread_slab(parent
))
1881 set_bit(CS_SPREAD_SLAB
, &cs
->flags
);
1882 set_bit(CS_SCHED_LOAD_BALANCE
, &cs
->flags
);
1883 cpumask_clear(cs
->cpus_allowed
);
1884 nodes_clear(cs
->mems_allowed
);
1885 fmeter_init(&cs
->fmeter
);
1886 cs
->relax_domain_level
= -1;
1888 cs
->parent
= parent
;
1889 number_of_cpusets
++;
1894 * If the cpuset being removed has its flag 'sched_load_balance'
1895 * enabled, then simulate turning sched_load_balance off, which
1896 * will call async_rebuild_sched_domains().
1899 static void cpuset_destroy(struct cgroup
*cont
)
1901 struct cpuset
*cs
= cgroup_cs(cont
);
1903 if (is_sched_load_balance(cs
))
1904 update_flag(CS_SCHED_LOAD_BALANCE
, cs
, 0);
1906 number_of_cpusets
--;
1907 free_cpumask_var(cs
->cpus_allowed
);
1911 struct cgroup_subsys cpuset_subsys
= {
1913 .create
= cpuset_create
,
1914 .destroy
= cpuset_destroy
,
1915 .can_attach
= cpuset_can_attach
,
1916 .attach
= cpuset_attach
,
1917 .populate
= cpuset_populate
,
1918 .post_clone
= cpuset_post_clone
,
1919 .subsys_id
= cpuset_subsys_id
,
1924 * cpuset_init - initialize cpusets at system boot
1926 * Description: Initialize top_cpuset and the cpuset internal file system,
1929 int __init
cpuset_init(void)
1933 if (!alloc_cpumask_var(&top_cpuset
.cpus_allowed
, GFP_KERNEL
))
1936 cpumask_setall(top_cpuset
.cpus_allowed
);
1937 nodes_setall(top_cpuset
.mems_allowed
);
1939 fmeter_init(&top_cpuset
.fmeter
);
1940 set_bit(CS_SCHED_LOAD_BALANCE
, &top_cpuset
.flags
);
1941 top_cpuset
.relax_domain_level
= -1;
1943 err
= register_filesystem(&cpuset_fs_type
);
1947 if (!alloc_cpumask_var(&cpus_attach
, GFP_KERNEL
))
1950 number_of_cpusets
= 1;
1955 * cpuset_do_move_task - move a given task to another cpuset
1956 * @tsk: pointer to task_struct the task to move
1957 * @scan: struct cgroup_scanner contained in its struct cpuset_hotplug_scanner
1959 * Called by cgroup_scan_tasks() for each task in a cgroup.
1960 * Return nonzero to stop the walk through the tasks.
1962 static void cpuset_do_move_task(struct task_struct
*tsk
,
1963 struct cgroup_scanner
*scan
)
1965 struct cgroup
*new_cgroup
= scan
->data
;
1967 cgroup_attach_task(new_cgroup
, tsk
);
1971 * move_member_tasks_to_cpuset - move tasks from one cpuset to another
1972 * @from: cpuset in which the tasks currently reside
1973 * @to: cpuset to which the tasks will be moved
1975 * Called with cgroup_mutex held
1976 * callback_mutex must not be held, as cpuset_attach() will take it.
1978 * The cgroup_scan_tasks() function will scan all the tasks in a cgroup,
1979 * calling callback functions for each.
1981 static void move_member_tasks_to_cpuset(struct cpuset
*from
, struct cpuset
*to
)
1983 struct cgroup_scanner scan
;
1985 scan
.cg
= from
->css
.cgroup
;
1986 scan
.test_task
= NULL
; /* select all tasks in cgroup */
1987 scan
.process_task
= cpuset_do_move_task
;
1989 scan
.data
= to
->css
.cgroup
;
1991 if (cgroup_scan_tasks(&scan
))
1992 printk(KERN_ERR
"move_member_tasks_to_cpuset: "
1993 "cgroup_scan_tasks failed\n");
1997 * If CPU and/or memory hotplug handlers, below, unplug any CPUs
1998 * or memory nodes, we need to walk over the cpuset hierarchy,
1999 * removing that CPU or node from all cpusets. If this removes the
2000 * last CPU or node from a cpuset, then move the tasks in the empty
2001 * cpuset to its next-highest non-empty parent.
2003 * Called with cgroup_mutex held
2004 * callback_mutex must not be held, as cpuset_attach() will take it.
2006 static void remove_tasks_in_empty_cpuset(struct cpuset
*cs
)
2008 struct cpuset
*parent
;
2011 * The cgroup's css_sets list is in use if there are tasks
2012 * in the cpuset; the list is empty if there are none;
2013 * the cs->css.refcnt seems always 0.
2015 if (list_empty(&cs
->css
.cgroup
->css_sets
))
2019 * Find its next-highest non-empty parent, (top cpuset
2020 * has online cpus, so can't be empty).
2022 parent
= cs
->parent
;
2023 while (cpumask_empty(parent
->cpus_allowed
) ||
2024 nodes_empty(parent
->mems_allowed
))
2025 parent
= parent
->parent
;
2027 move_member_tasks_to_cpuset(cs
, parent
);
2031 * Walk the specified cpuset subtree and look for empty cpusets.
2032 * The tasks of such cpuset must be moved to a parent cpuset.
2034 * Called with cgroup_mutex held. We take callback_mutex to modify
2035 * cpus_allowed and mems_allowed.
2037 * This walk processes the tree from top to bottom, completing one layer
2038 * before dropping down to the next. It always processes a node before
2039 * any of its children.
2041 * For now, since we lack memory hot unplug, we'll never see a cpuset
2042 * that has tasks along with an empty 'mems'. But if we did see such
2043 * a cpuset, we'd handle it just like we do if its 'cpus' was empty.
2045 static void scan_for_empty_cpusets(struct cpuset
*root
)
2048 struct cpuset
*cp
; /* scans cpusets being updated */
2049 struct cpuset
*child
; /* scans child cpusets of cp */
2050 struct cgroup
*cont
;
2051 static nodemask_t oldmems
; /* protected by cgroup_mutex */
2053 list_add_tail((struct list_head
*)&root
->stack_list
, &queue
);
2055 while (!list_empty(&queue
)) {
2056 cp
= list_first_entry(&queue
, struct cpuset
, stack_list
);
2057 list_del(queue
.next
);
2058 list_for_each_entry(cont
, &cp
->css
.cgroup
->children
, sibling
) {
2059 child
= cgroup_cs(cont
);
2060 list_add_tail(&child
->stack_list
, &queue
);
2063 /* Continue past cpusets with all cpus, mems online */
2064 if (cpumask_subset(cp
->cpus_allowed
, cpu_active_mask
) &&
2065 nodes_subset(cp
->mems_allowed
, node_states
[N_HIGH_MEMORY
]))
2068 oldmems
= cp
->mems_allowed
;
2070 /* Remove offline cpus and mems from this cpuset. */
2071 mutex_lock(&callback_mutex
);
2072 cpumask_and(cp
->cpus_allowed
, cp
->cpus_allowed
,
2074 nodes_and(cp
->mems_allowed
, cp
->mems_allowed
,
2075 node_states
[N_HIGH_MEMORY
]);
2076 mutex_unlock(&callback_mutex
);
2078 /* Move tasks from the empty cpuset to a parent */
2079 if (cpumask_empty(cp
->cpus_allowed
) ||
2080 nodes_empty(cp
->mems_allowed
))
2081 remove_tasks_in_empty_cpuset(cp
);
2083 update_tasks_cpumask(cp
, NULL
);
2084 update_tasks_nodemask(cp
, &oldmems
, NULL
);
2090 * The top_cpuset tracks what CPUs and Memory Nodes are online,
2091 * period. This is necessary in order to make cpusets transparent
2092 * (of no affect) on systems that are actively using CPU hotplug
2093 * but making no active use of cpusets.
2095 * This routine ensures that top_cpuset.cpus_allowed tracks
2096 * cpu_active_mask on each CPU hotplug (cpuhp) event.
2098 * Called within get_online_cpus(). Needs to call cgroup_lock()
2099 * before calling generate_sched_domains().
2101 void cpuset_update_active_cpus(void)
2103 struct sched_domain_attr
*attr
;
2104 cpumask_var_t
*doms
;
2108 mutex_lock(&callback_mutex
);
2109 cpumask_copy(top_cpuset
.cpus_allowed
, cpu_active_mask
);
2110 mutex_unlock(&callback_mutex
);
2111 scan_for_empty_cpusets(&top_cpuset
);
2112 ndoms
= generate_sched_domains(&doms
, &attr
);
2115 /* Have scheduler rebuild the domains */
2116 partition_sched_domains(ndoms
, doms
, attr
);
2119 #ifdef CONFIG_MEMORY_HOTPLUG
2121 * Keep top_cpuset.mems_allowed tracking node_states[N_HIGH_MEMORY].
2122 * Call this routine anytime after node_states[N_HIGH_MEMORY] changes.
2123 * See also the previous routine cpuset_track_online_cpus().
2125 static int cpuset_track_online_nodes(struct notifier_block
*self
,
2126 unsigned long action
, void *arg
)
2128 static nodemask_t oldmems
; /* protected by cgroup_mutex */
2133 oldmems
= top_cpuset
.mems_allowed
;
2134 mutex_lock(&callback_mutex
);
2135 top_cpuset
.mems_allowed
= node_states
[N_HIGH_MEMORY
];
2136 mutex_unlock(&callback_mutex
);
2137 update_tasks_nodemask(&top_cpuset
, &oldmems
, NULL
);
2141 * needn't update top_cpuset.mems_allowed explicitly because
2142 * scan_for_empty_cpusets() will update it.
2144 scan_for_empty_cpusets(&top_cpuset
);
2156 * cpuset_init_smp - initialize cpus_allowed
2158 * Description: Finish top cpuset after cpu, node maps are initialized
2161 void __init
cpuset_init_smp(void)
2163 cpumask_copy(top_cpuset
.cpus_allowed
, cpu_active_mask
);
2164 top_cpuset
.mems_allowed
= node_states
[N_HIGH_MEMORY
];
2166 hotplug_memory_notifier(cpuset_track_online_nodes
, 10);
2168 cpuset_wq
= create_singlethread_workqueue("cpuset");
2173 * cpuset_cpus_allowed - return cpus_allowed mask from a tasks cpuset.
2174 * @tsk: pointer to task_struct from which to obtain cpuset->cpus_allowed.
2175 * @pmask: pointer to struct cpumask variable to receive cpus_allowed set.
2177 * Description: Returns the cpumask_var_t cpus_allowed of the cpuset
2178 * attached to the specified @tsk. Guaranteed to return some non-empty
2179 * subset of cpu_online_map, even if this means going outside the
2183 void cpuset_cpus_allowed(struct task_struct
*tsk
, struct cpumask
*pmask
)
2185 mutex_lock(&callback_mutex
);
2187 guarantee_online_cpus(task_cs(tsk
), pmask
);
2189 mutex_unlock(&callback_mutex
);
2192 int cpuset_cpus_allowed_fallback(struct task_struct
*tsk
)
2194 const struct cpuset
*cs
;
2200 do_set_cpus_allowed(tsk
, cs
->cpus_allowed
);
2204 * We own tsk->cpus_allowed, nobody can change it under us.
2206 * But we used cs && cs->cpus_allowed lockless and thus can
2207 * race with cgroup_attach_task() or update_cpumask() and get
2208 * the wrong tsk->cpus_allowed. However, both cases imply the
2209 * subsequent cpuset_change_cpumask()->set_cpus_allowed_ptr()
2210 * which takes task_rq_lock().
2212 * If we are called after it dropped the lock we must see all
2213 * changes in tsk_cs()->cpus_allowed. Otherwise we can temporary
2214 * set any mask even if it is not right from task_cs() pov,
2215 * the pending set_cpus_allowed_ptr() will fix things.
2218 cpu
= cpumask_any_and(&tsk
->cpus_allowed
, cpu_active_mask
);
2219 if (cpu
>= nr_cpu_ids
) {
2221 * Either tsk->cpus_allowed is wrong (see above) or it
2222 * is actually empty. The latter case is only possible
2223 * if we are racing with remove_tasks_in_empty_cpuset().
2224 * Like above we can temporary set any mask and rely on
2225 * set_cpus_allowed_ptr() as synchronization point.
2227 do_set_cpus_allowed(tsk
, cpu_possible_mask
);
2228 cpu
= cpumask_any(cpu_active_mask
);
2234 void cpuset_init_current_mems_allowed(void)
2236 nodes_setall(current
->mems_allowed
);
2240 * cpuset_mems_allowed - return mems_allowed mask from a tasks cpuset.
2241 * @tsk: pointer to task_struct from which to obtain cpuset->mems_allowed.
2243 * Description: Returns the nodemask_t mems_allowed of the cpuset
2244 * attached to the specified @tsk. Guaranteed to return some non-empty
2245 * subset of node_states[N_HIGH_MEMORY], even if this means going outside the
2249 nodemask_t
cpuset_mems_allowed(struct task_struct
*tsk
)
2253 mutex_lock(&callback_mutex
);
2255 guarantee_online_mems(task_cs(tsk
), &mask
);
2257 mutex_unlock(&callback_mutex
);
2263 * cpuset_nodemask_valid_mems_allowed - check nodemask vs. curremt mems_allowed
2264 * @nodemask: the nodemask to be checked
2266 * Are any of the nodes in the nodemask allowed in current->mems_allowed?
2268 int cpuset_nodemask_valid_mems_allowed(nodemask_t
*nodemask
)
2270 return nodes_intersects(*nodemask
, current
->mems_allowed
);
2274 * nearest_hardwall_ancestor() - Returns the nearest mem_exclusive or
2275 * mem_hardwall ancestor to the specified cpuset. Call holding
2276 * callback_mutex. If no ancestor is mem_exclusive or mem_hardwall
2277 * (an unusual configuration), then returns the root cpuset.
2279 static const struct cpuset
*nearest_hardwall_ancestor(const struct cpuset
*cs
)
2281 while (!(is_mem_exclusive(cs
) || is_mem_hardwall(cs
)) && cs
->parent
)
2287 * cpuset_node_allowed_softwall - Can we allocate on a memory node?
2288 * @node: is this an allowed node?
2289 * @gfp_mask: memory allocation flags
2291 * If we're in interrupt, yes, we can always allocate. If __GFP_THISNODE is
2292 * set, yes, we can always allocate. If node is in our task's mems_allowed,
2293 * yes. If it's not a __GFP_HARDWALL request and this node is in the nearest
2294 * hardwalled cpuset ancestor to this task's cpuset, yes. If the task has been
2295 * OOM killed and has access to memory reserves as specified by the TIF_MEMDIE
2299 * If __GFP_HARDWALL is set, cpuset_node_allowed_softwall() reduces to
2300 * cpuset_node_allowed_hardwall(). Otherwise, cpuset_node_allowed_softwall()
2301 * might sleep, and might allow a node from an enclosing cpuset.
2303 * cpuset_node_allowed_hardwall() only handles the simpler case of hardwall
2304 * cpusets, and never sleeps.
2306 * The __GFP_THISNODE placement logic is really handled elsewhere,
2307 * by forcibly using a zonelist starting at a specified node, and by
2308 * (in get_page_from_freelist()) refusing to consider the zones for
2309 * any node on the zonelist except the first. By the time any such
2310 * calls get to this routine, we should just shut up and say 'yes'.
2312 * GFP_USER allocations are marked with the __GFP_HARDWALL bit,
2313 * and do not allow allocations outside the current tasks cpuset
2314 * unless the task has been OOM killed as is marked TIF_MEMDIE.
2315 * GFP_KERNEL allocations are not so marked, so can escape to the
2316 * nearest enclosing hardwalled ancestor cpuset.
2318 * Scanning up parent cpusets requires callback_mutex. The
2319 * __alloc_pages() routine only calls here with __GFP_HARDWALL bit
2320 * _not_ set if it's a GFP_KERNEL allocation, and all nodes in the
2321 * current tasks mems_allowed came up empty on the first pass over
2322 * the zonelist. So only GFP_KERNEL allocations, if all nodes in the
2323 * cpuset are short of memory, might require taking the callback_mutex
2326 * The first call here from mm/page_alloc:get_page_from_freelist()
2327 * has __GFP_HARDWALL set in gfp_mask, enforcing hardwall cpusets,
2328 * so no allocation on a node outside the cpuset is allowed (unless
2329 * in interrupt, of course).
2331 * The second pass through get_page_from_freelist() doesn't even call
2332 * here for GFP_ATOMIC calls. For those calls, the __alloc_pages()
2333 * variable 'wait' is not set, and the bit ALLOC_CPUSET is not set
2334 * in alloc_flags. That logic and the checks below have the combined
2336 * in_interrupt - any node ok (current task context irrelevant)
2337 * GFP_ATOMIC - any node ok
2338 * TIF_MEMDIE - any node ok
2339 * GFP_KERNEL - any node in enclosing hardwalled cpuset ok
2340 * GFP_USER - only nodes in current tasks mems allowed ok.
2343 * Don't call cpuset_node_allowed_softwall if you can't sleep, unless you
2344 * pass in the __GFP_HARDWALL flag set in gfp_flag, which disables
2345 * the code that might scan up ancestor cpusets and sleep.
2347 int __cpuset_node_allowed_softwall(int node
, gfp_t gfp_mask
)
2349 const struct cpuset
*cs
; /* current cpuset ancestors */
2350 int allowed
; /* is allocation in zone z allowed? */
2352 if (in_interrupt() || (gfp_mask
& __GFP_THISNODE
))
2354 might_sleep_if(!(gfp_mask
& __GFP_HARDWALL
));
2355 if (node_isset(node
, current
->mems_allowed
))
2358 * Allow tasks that have access to memory reserves because they have
2359 * been OOM killed to get memory anywhere.
2361 if (unlikely(test_thread_flag(TIF_MEMDIE
)))
2363 if (gfp_mask
& __GFP_HARDWALL
) /* If hardwall request, stop here */
2366 if (current
->flags
& PF_EXITING
) /* Let dying task have memory */
2369 /* Not hardwall and node outside mems_allowed: scan up cpusets */
2370 mutex_lock(&callback_mutex
);
2373 cs
= nearest_hardwall_ancestor(task_cs(current
));
2374 task_unlock(current
);
2376 allowed
= node_isset(node
, cs
->mems_allowed
);
2377 mutex_unlock(&callback_mutex
);
2382 * cpuset_node_allowed_hardwall - Can we allocate on a memory node?
2383 * @node: is this an allowed node?
2384 * @gfp_mask: memory allocation flags
2386 * If we're in interrupt, yes, we can always allocate. If __GFP_THISNODE is
2387 * set, yes, we can always allocate. If node is in our task's mems_allowed,
2388 * yes. If the task has been OOM killed and has access to memory reserves as
2389 * specified by the TIF_MEMDIE flag, yes.
2392 * The __GFP_THISNODE placement logic is really handled elsewhere,
2393 * by forcibly using a zonelist starting at a specified node, and by
2394 * (in get_page_from_freelist()) refusing to consider the zones for
2395 * any node on the zonelist except the first. By the time any such
2396 * calls get to this routine, we should just shut up and say 'yes'.
2398 * Unlike the cpuset_node_allowed_softwall() variant, above,
2399 * this variant requires that the node be in the current task's
2400 * mems_allowed or that we're in interrupt. It does not scan up the
2401 * cpuset hierarchy for the nearest enclosing mem_exclusive cpuset.
2404 int __cpuset_node_allowed_hardwall(int node
, gfp_t gfp_mask
)
2406 if (in_interrupt() || (gfp_mask
& __GFP_THISNODE
))
2408 if (node_isset(node
, current
->mems_allowed
))
2411 * Allow tasks that have access to memory reserves because they have
2412 * been OOM killed to get memory anywhere.
2414 if (unlikely(test_thread_flag(TIF_MEMDIE
)))
2420 * cpuset_unlock - release lock on cpuset changes
2422 * Undo the lock taken in a previous cpuset_lock() call.
2425 void cpuset_unlock(void)
2427 mutex_unlock(&callback_mutex
);
2431 * cpuset_mem_spread_node() - On which node to begin search for a file page
2432 * cpuset_slab_spread_node() - On which node to begin search for a slab page
2434 * If a task is marked PF_SPREAD_PAGE or PF_SPREAD_SLAB (as for
2435 * tasks in a cpuset with is_spread_page or is_spread_slab set),
2436 * and if the memory allocation used cpuset_mem_spread_node()
2437 * to determine on which node to start looking, as it will for
2438 * certain page cache or slab cache pages such as used for file
2439 * system buffers and inode caches, then instead of starting on the
2440 * local node to look for a free page, rather spread the starting
2441 * node around the tasks mems_allowed nodes.
2443 * We don't have to worry about the returned node being offline
2444 * because "it can't happen", and even if it did, it would be ok.
2446 * The routines calling guarantee_online_mems() are careful to
2447 * only set nodes in task->mems_allowed that are online. So it
2448 * should not be possible for the following code to return an
2449 * offline node. But if it did, that would be ok, as this routine
2450 * is not returning the node where the allocation must be, only
2451 * the node where the search should start. The zonelist passed to
2452 * __alloc_pages() will include all nodes. If the slab allocator
2453 * is passed an offline node, it will fall back to the local node.
2454 * See kmem_cache_alloc_node().
2457 static int cpuset_spread_node(int *rotor
)
2461 node
= next_node(*rotor
, current
->mems_allowed
);
2462 if (node
== MAX_NUMNODES
)
2463 node
= first_node(current
->mems_allowed
);
2468 int cpuset_mem_spread_node(void)
2470 if (current
->cpuset_mem_spread_rotor
== NUMA_NO_NODE
)
2471 current
->cpuset_mem_spread_rotor
=
2472 node_random(¤t
->mems_allowed
);
2474 return cpuset_spread_node(¤t
->cpuset_mem_spread_rotor
);
2477 int cpuset_slab_spread_node(void)
2479 if (current
->cpuset_slab_spread_rotor
== NUMA_NO_NODE
)
2480 current
->cpuset_slab_spread_rotor
=
2481 node_random(¤t
->mems_allowed
);
2483 return cpuset_spread_node(¤t
->cpuset_slab_spread_rotor
);
2486 EXPORT_SYMBOL_GPL(cpuset_mem_spread_node
);
2489 * cpuset_mems_allowed_intersects - Does @tsk1's mems_allowed intersect @tsk2's?
2490 * @tsk1: pointer to task_struct of some task.
2491 * @tsk2: pointer to task_struct of some other task.
2493 * Description: Return true if @tsk1's mems_allowed intersects the
2494 * mems_allowed of @tsk2. Used by the OOM killer to determine if
2495 * one of the task's memory usage might impact the memory available
2499 int cpuset_mems_allowed_intersects(const struct task_struct
*tsk1
,
2500 const struct task_struct
*tsk2
)
2502 return nodes_intersects(tsk1
->mems_allowed
, tsk2
->mems_allowed
);
2506 * cpuset_print_task_mems_allowed - prints task's cpuset and mems_allowed
2507 * @task: pointer to task_struct of some task.
2509 * Description: Prints @task's name, cpuset name, and cached copy of its
2510 * mems_allowed to the kernel log. Must hold task_lock(task) to allow
2511 * dereferencing task_cs(task).
2513 void cpuset_print_task_mems_allowed(struct task_struct
*tsk
)
2515 struct dentry
*dentry
;
2517 dentry
= task_cs(tsk
)->css
.cgroup
->dentry
;
2518 spin_lock(&cpuset_buffer_lock
);
2519 snprintf(cpuset_name
, CPUSET_NAME_LEN
,
2520 dentry
? (const char *)dentry
->d_name
.name
: "/");
2521 nodelist_scnprintf(cpuset_nodelist
, CPUSET_NODELIST_LEN
,
2523 printk(KERN_INFO
"%s cpuset=%s mems_allowed=%s\n",
2524 tsk
->comm
, cpuset_name
, cpuset_nodelist
);
2525 spin_unlock(&cpuset_buffer_lock
);
2529 * Collection of memory_pressure is suppressed unless
2530 * this flag is enabled by writing "1" to the special
2531 * cpuset file 'memory_pressure_enabled' in the root cpuset.
2534 int cpuset_memory_pressure_enabled __read_mostly
;
2537 * cpuset_memory_pressure_bump - keep stats of per-cpuset reclaims.
2539 * Keep a running average of the rate of synchronous (direct)
2540 * page reclaim efforts initiated by tasks in each cpuset.
2542 * This represents the rate at which some task in the cpuset
2543 * ran low on memory on all nodes it was allowed to use, and
2544 * had to enter the kernels page reclaim code in an effort to
2545 * create more free memory by tossing clean pages or swapping
2546 * or writing dirty pages.
2548 * Display to user space in the per-cpuset read-only file
2549 * "memory_pressure". Value displayed is an integer
2550 * representing the recent rate of entry into the synchronous
2551 * (direct) page reclaim by any task attached to the cpuset.
2554 void __cpuset_memory_pressure_bump(void)
2557 fmeter_markevent(&task_cs(current
)->fmeter
);
2558 task_unlock(current
);
2561 #ifdef CONFIG_PROC_PID_CPUSET
2563 * proc_cpuset_show()
2564 * - Print tasks cpuset path into seq_file.
2565 * - Used for /proc/<pid>/cpuset.
2566 * - No need to task_lock(tsk) on this tsk->cpuset reference, as it
2567 * doesn't really matter if tsk->cpuset changes after we read it,
2568 * and we take cgroup_mutex, keeping cpuset_attach() from changing it
2571 static int proc_cpuset_show(struct seq_file
*m
, void *unused_v
)
2574 struct task_struct
*tsk
;
2576 struct cgroup_subsys_state
*css
;
2580 buf
= kmalloc(PAGE_SIZE
, GFP_KERNEL
);
2586 tsk
= get_pid_task(pid
, PIDTYPE_PID
);
2592 css
= task_subsys_state(tsk
, cpuset_subsys_id
);
2593 retval
= cgroup_path(css
->cgroup
, buf
, PAGE_SIZE
);
2600 put_task_struct(tsk
);
2607 static int cpuset_open(struct inode
*inode
, struct file
*file
)
2609 struct pid
*pid
= PROC_I(inode
)->pid
;
2610 return single_open(file
, proc_cpuset_show
, pid
);
2613 const struct file_operations proc_cpuset_operations
= {
2614 .open
= cpuset_open
,
2616 .llseek
= seq_lseek
,
2617 .release
= single_release
,
2619 #endif /* CONFIG_PROC_PID_CPUSET */
2621 /* Display task mems_allowed in /proc/<pid>/status file. */
2622 void cpuset_task_status_allowed(struct seq_file
*m
, struct task_struct
*task
)
2624 seq_printf(m
, "Mems_allowed:\t");
2625 seq_nodemask(m
, &task
->mems_allowed
);
2626 seq_printf(m
, "\n");
2627 seq_printf(m
, "Mems_allowed_list:\t");
2628 seq_nodemask_list(m
, &task
->mems_allowed
);
2629 seq_printf(m
, "\n");