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 */
146 CS_SCHED_LOAD_BALANCE
,
151 /* the type of hotplug event */
157 /* convenient tests for these bits */
158 static inline bool is_cpuset_online(const struct cpuset
*cs
)
160 return test_bit(CS_ONLINE
, &cs
->flags
);
163 static inline int is_cpu_exclusive(const struct cpuset
*cs
)
165 return test_bit(CS_CPU_EXCLUSIVE
, &cs
->flags
);
168 static inline int is_mem_exclusive(const struct cpuset
*cs
)
170 return test_bit(CS_MEM_EXCLUSIVE
, &cs
->flags
);
173 static inline int is_mem_hardwall(const struct cpuset
*cs
)
175 return test_bit(CS_MEM_HARDWALL
, &cs
->flags
);
178 static inline int is_sched_load_balance(const struct cpuset
*cs
)
180 return test_bit(CS_SCHED_LOAD_BALANCE
, &cs
->flags
);
183 static inline int is_memory_migrate(const struct cpuset
*cs
)
185 return test_bit(CS_MEMORY_MIGRATE
, &cs
->flags
);
188 static inline int is_spread_page(const struct cpuset
*cs
)
190 return test_bit(CS_SPREAD_PAGE
, &cs
->flags
);
193 static inline int is_spread_slab(const struct cpuset
*cs
)
195 return test_bit(CS_SPREAD_SLAB
, &cs
->flags
);
198 static struct cpuset top_cpuset
= {
199 .flags
= ((1 << CS_ONLINE
) | (1 << CS_CPU_EXCLUSIVE
) |
200 (1 << CS_MEM_EXCLUSIVE
)),
204 * cpuset_for_each_child - traverse online children of a cpuset
205 * @child_cs: loop cursor pointing to the current child
206 * @pos_cgrp: used for iteration
207 * @parent_cs: target cpuset to walk children of
209 * Walk @child_cs through the online children of @parent_cs. Must be used
210 * with RCU read locked.
212 #define cpuset_for_each_child(child_cs, pos_cgrp, parent_cs) \
213 cgroup_for_each_child((pos_cgrp), (parent_cs)->css.cgroup) \
214 if (is_cpuset_online(((child_cs) = cgroup_cs((pos_cgrp)))))
217 * There are two global mutexes guarding cpuset structures. The first
218 * is the main control groups cgroup_mutex, accessed via
219 * cgroup_lock()/cgroup_unlock(). The second is the cpuset-specific
220 * callback_mutex, below. They can nest. It is ok to first take
221 * cgroup_mutex, then nest callback_mutex. We also require taking
222 * task_lock() when dereferencing a task's cpuset pointer. See "The
223 * task_lock() exception", at the end of this comment.
225 * A task must hold both mutexes to modify cpusets. If a task
226 * holds cgroup_mutex, then it blocks others wanting that mutex,
227 * ensuring that it is the only task able to also acquire callback_mutex
228 * and be able to modify cpusets. It can perform various checks on
229 * the cpuset structure first, knowing nothing will change. It can
230 * also allocate memory while just holding cgroup_mutex. While it is
231 * performing these checks, various callback routines can briefly
232 * acquire callback_mutex to query cpusets. Once it is ready to make
233 * the changes, it takes callback_mutex, blocking everyone else.
235 * Calls to the kernel memory allocator can not be made while holding
236 * callback_mutex, as that would risk double tripping on callback_mutex
237 * from one of the callbacks into the cpuset code from within
240 * If a task is only holding callback_mutex, then it has read-only
243 * Now, the task_struct fields mems_allowed and mempolicy may be changed
244 * by other task, we use alloc_lock in the task_struct fields to protect
247 * The cpuset_common_file_read() handlers only hold callback_mutex across
248 * small pieces of code, such as when reading out possibly multi-word
249 * cpumasks and nodemasks.
251 * Accessing a task's cpuset should be done in accordance with the
252 * guidelines for accessing subsystem state in kernel/cgroup.c
255 static DEFINE_MUTEX(callback_mutex
);
258 * cpuset_buffer_lock protects both the cpuset_name and cpuset_nodelist
259 * buffers. They are statically allocated to prevent using excess stack
260 * when calling cpuset_print_task_mems_allowed().
262 #define CPUSET_NAME_LEN (128)
263 #define CPUSET_NODELIST_LEN (256)
264 static char cpuset_name
[CPUSET_NAME_LEN
];
265 static char cpuset_nodelist
[CPUSET_NODELIST_LEN
];
266 static DEFINE_SPINLOCK(cpuset_buffer_lock
);
269 * This is ugly, but preserves the userspace API for existing cpuset
270 * users. If someone tries to mount the "cpuset" filesystem, we
271 * silently switch it to mount "cgroup" instead
273 static struct dentry
*cpuset_mount(struct file_system_type
*fs_type
,
274 int flags
, const char *unused_dev_name
, void *data
)
276 struct file_system_type
*cgroup_fs
= get_fs_type("cgroup");
277 struct dentry
*ret
= ERR_PTR(-ENODEV
);
281 "release_agent=/sbin/cpuset_release_agent";
282 ret
= cgroup_fs
->mount(cgroup_fs
, flags
,
283 unused_dev_name
, mountopts
);
284 put_filesystem(cgroup_fs
);
289 static struct file_system_type cpuset_fs_type
= {
291 .mount
= cpuset_mount
,
295 * Return in pmask the portion of a cpusets's cpus_allowed that
296 * are online. If none are online, walk up the cpuset hierarchy
297 * until we find one that does have some online cpus. If we get
298 * all the way to the top and still haven't found any online cpus,
299 * return cpu_online_mask. Or if passed a NULL cs from an exit'ing
300 * task, return cpu_online_mask.
302 * One way or another, we guarantee to return some non-empty subset
303 * of cpu_online_mask.
305 * Call with callback_mutex held.
308 static void guarantee_online_cpus(const struct cpuset
*cs
,
309 struct cpumask
*pmask
)
311 while (cs
&& !cpumask_intersects(cs
->cpus_allowed
, cpu_online_mask
))
314 cpumask_and(pmask
, cs
->cpus_allowed
, cpu_online_mask
);
316 cpumask_copy(pmask
, cpu_online_mask
);
317 BUG_ON(!cpumask_intersects(pmask
, cpu_online_mask
));
321 * Return in *pmask the portion of a cpusets's mems_allowed that
322 * are online, with memory. If none are online with memory, walk
323 * up the cpuset hierarchy until we find one that does have some
324 * online mems. If we get all the way to the top and still haven't
325 * found any online mems, return node_states[N_MEMORY].
327 * One way or another, we guarantee to return some non-empty subset
328 * of node_states[N_MEMORY].
330 * Call with callback_mutex held.
333 static void guarantee_online_mems(const struct cpuset
*cs
, nodemask_t
*pmask
)
335 while (cs
&& !nodes_intersects(cs
->mems_allowed
,
336 node_states
[N_MEMORY
]))
339 nodes_and(*pmask
, cs
->mems_allowed
,
340 node_states
[N_MEMORY
]);
342 *pmask
= node_states
[N_MEMORY
];
343 BUG_ON(!nodes_intersects(*pmask
, node_states
[N_MEMORY
]));
347 * update task's spread flag if cpuset's page/slab spread flag is set
349 * Called with callback_mutex/cgroup_mutex held
351 static void cpuset_update_task_spread_flag(struct cpuset
*cs
,
352 struct task_struct
*tsk
)
354 if (is_spread_page(cs
))
355 tsk
->flags
|= PF_SPREAD_PAGE
;
357 tsk
->flags
&= ~PF_SPREAD_PAGE
;
358 if (is_spread_slab(cs
))
359 tsk
->flags
|= PF_SPREAD_SLAB
;
361 tsk
->flags
&= ~PF_SPREAD_SLAB
;
365 * is_cpuset_subset(p, q) - Is cpuset p a subset of cpuset q?
367 * One cpuset is a subset of another if all its allowed CPUs and
368 * Memory Nodes are a subset of the other, and its exclusive flags
369 * are only set if the other's are set. Call holding cgroup_mutex.
372 static int is_cpuset_subset(const struct cpuset
*p
, const struct cpuset
*q
)
374 return cpumask_subset(p
->cpus_allowed
, q
->cpus_allowed
) &&
375 nodes_subset(p
->mems_allowed
, q
->mems_allowed
) &&
376 is_cpu_exclusive(p
) <= is_cpu_exclusive(q
) &&
377 is_mem_exclusive(p
) <= is_mem_exclusive(q
);
381 * alloc_trial_cpuset - allocate a trial cpuset
382 * @cs: the cpuset that the trial cpuset duplicates
384 static struct cpuset
*alloc_trial_cpuset(const struct cpuset
*cs
)
386 struct cpuset
*trial
;
388 trial
= kmemdup(cs
, sizeof(*cs
), GFP_KERNEL
);
392 if (!alloc_cpumask_var(&trial
->cpus_allowed
, GFP_KERNEL
)) {
396 cpumask_copy(trial
->cpus_allowed
, cs
->cpus_allowed
);
402 * free_trial_cpuset - free the trial cpuset
403 * @trial: the trial cpuset to be freed
405 static void free_trial_cpuset(struct cpuset
*trial
)
407 free_cpumask_var(trial
->cpus_allowed
);
412 * validate_change() - Used to validate that any proposed cpuset change
413 * follows the structural rules for cpusets.
415 * If we replaced the flag and mask values of the current cpuset
416 * (cur) with those values in the trial cpuset (trial), would
417 * our various subset and exclusive rules still be valid? Presumes
420 * 'cur' is the address of an actual, in-use cpuset. Operations
421 * such as list traversal that depend on the actual address of the
422 * cpuset in the list must use cur below, not trial.
424 * 'trial' is the address of bulk structure copy of cur, with
425 * perhaps one or more of the fields cpus_allowed, mems_allowed,
426 * or flags changed to new, trial values.
428 * Return 0 if valid, -errno if not.
431 static int validate_change(const struct cpuset
*cur
, const struct cpuset
*trial
)
434 struct cpuset
*c
, *par
;
439 /* Each of our child cpusets must be a subset of us */
441 cpuset_for_each_child(c
, cont
, cur
)
442 if (!is_cpuset_subset(c
, trial
))
445 /* Remaining checks don't apply to root cpuset */
447 if (cur
== &top_cpuset
)
452 /* We must be a subset of our parent cpuset */
454 if (!is_cpuset_subset(trial
, par
))
458 * If either I or some sibling (!= me) is exclusive, we can't
462 cpuset_for_each_child(c
, cont
, par
) {
463 if ((is_cpu_exclusive(trial
) || is_cpu_exclusive(c
)) &&
465 cpumask_intersects(trial
->cpus_allowed
, c
->cpus_allowed
))
467 if ((is_mem_exclusive(trial
) || is_mem_exclusive(c
)) &&
469 nodes_intersects(trial
->mems_allowed
, c
->mems_allowed
))
473 /* Cpusets with tasks can't have empty cpus_allowed or mems_allowed */
475 if (cgroup_task_count(cur
->css
.cgroup
) &&
476 (cpumask_empty(trial
->cpus_allowed
) ||
477 nodes_empty(trial
->mems_allowed
)))
488 * Helper routine for generate_sched_domains().
489 * Do cpusets a, b have overlapping cpus_allowed masks?
491 static int cpusets_overlap(struct cpuset
*a
, struct cpuset
*b
)
493 return cpumask_intersects(a
->cpus_allowed
, b
->cpus_allowed
);
497 update_domain_attr(struct sched_domain_attr
*dattr
, struct cpuset
*c
)
499 if (dattr
->relax_domain_level
< c
->relax_domain_level
)
500 dattr
->relax_domain_level
= c
->relax_domain_level
;
505 update_domain_attr_tree(struct sched_domain_attr
*dattr
, struct cpuset
*c
)
509 list_add(&c
->stack_list
, &q
);
510 while (!list_empty(&q
)) {
513 struct cpuset
*child
;
515 cp
= list_first_entry(&q
, struct cpuset
, stack_list
);
518 if (cpumask_empty(cp
->cpus_allowed
))
521 if (is_sched_load_balance(cp
))
522 update_domain_attr(dattr
, cp
);
525 cpuset_for_each_child(child
, cont
, cp
)
526 list_add_tail(&child
->stack_list
, &q
);
532 * generate_sched_domains()
534 * This function builds a partial partition of the systems CPUs
535 * A 'partial partition' is a set of non-overlapping subsets whose
536 * union is a subset of that set.
537 * The output of this function needs to be passed to kernel/sched.c
538 * partition_sched_domains() routine, which will rebuild the scheduler's
539 * load balancing domains (sched domains) as specified by that partial
542 * See "What is sched_load_balance" in Documentation/cgroups/cpusets.txt
543 * for a background explanation of this.
545 * Does not return errors, on the theory that the callers of this
546 * routine would rather not worry about failures to rebuild sched
547 * domains when operating in the severe memory shortage situations
548 * that could cause allocation failures below.
550 * Must be called with cgroup_lock held.
552 * The three key local variables below are:
553 * q - a linked-list queue of cpuset pointers, used to implement a
554 * top-down scan of all cpusets. This scan loads a pointer
555 * to each cpuset marked is_sched_load_balance into the
556 * array 'csa'. For our purposes, rebuilding the schedulers
557 * sched domains, we can ignore !is_sched_load_balance cpusets.
558 * csa - (for CpuSet Array) Array of pointers to all the cpusets
559 * that need to be load balanced, for convenient iterative
560 * access by the subsequent code that finds the best partition,
561 * i.e the set of domains (subsets) of CPUs such that the
562 * cpus_allowed of every cpuset marked is_sched_load_balance
563 * is a subset of one of these domains, while there are as
564 * many such domains as possible, each as small as possible.
565 * doms - Conversion of 'csa' to an array of cpumasks, for passing to
566 * the kernel/sched.c routine partition_sched_domains() in a
567 * convenient format, that can be easily compared to the prior
568 * value to determine what partition elements (sched domains)
569 * were changed (added or removed.)
571 * Finding the best partition (set of domains):
572 * The triple nested loops below over i, j, k scan over the
573 * load balanced cpusets (using the array of cpuset pointers in
574 * csa[]) looking for pairs of cpusets that have overlapping
575 * cpus_allowed, but which don't have the same 'pn' partition
576 * number and gives them in the same partition number. It keeps
577 * looping on the 'restart' label until it can no longer find
580 * The union of the cpus_allowed masks from the set of
581 * all cpusets having the same 'pn' value then form the one
582 * element of the partition (one sched domain) to be passed to
583 * partition_sched_domains().
585 static int generate_sched_domains(cpumask_var_t
**domains
,
586 struct sched_domain_attr
**attributes
)
588 LIST_HEAD(q
); /* queue of cpusets to be scanned */
589 struct cpuset
*cp
; /* scans q */
590 struct cpuset
**csa
; /* array of all cpuset ptrs */
591 int csn
; /* how many cpuset ptrs in csa so far */
592 int i
, j
, k
; /* indices for partition finding loops */
593 cpumask_var_t
*doms
; /* resulting partition; i.e. sched domains */
594 struct sched_domain_attr
*dattr
; /* attributes for custom domains */
595 int ndoms
= 0; /* number of sched domains in result */
596 int nslot
; /* next empty doms[] struct cpumask slot */
602 /* Special case for the 99% of systems with one, full, sched domain */
603 if (is_sched_load_balance(&top_cpuset
)) {
605 doms
= alloc_sched_domains(ndoms
);
609 dattr
= kmalloc(sizeof(struct sched_domain_attr
), GFP_KERNEL
);
611 *dattr
= SD_ATTR_INIT
;
612 update_domain_attr_tree(dattr
, &top_cpuset
);
614 cpumask_copy(doms
[0], top_cpuset
.cpus_allowed
);
619 csa
= kmalloc(number_of_cpusets
* sizeof(cp
), GFP_KERNEL
);
624 list_add(&top_cpuset
.stack_list
, &q
);
625 while (!list_empty(&q
)) {
627 struct cpuset
*child
; /* scans child cpusets of cp */
629 cp
= list_first_entry(&q
, struct cpuset
, stack_list
);
632 if (cpumask_empty(cp
->cpus_allowed
))
636 * All child cpusets contain a subset of the parent's cpus, so
637 * just skip them, and then we call update_domain_attr_tree()
638 * to calc relax_domain_level of the corresponding sched
641 if (is_sched_load_balance(cp
)) {
647 cpuset_for_each_child(child
, cont
, cp
)
648 list_add_tail(&child
->stack_list
, &q
);
652 for (i
= 0; i
< csn
; i
++)
657 /* Find the best partition (set of sched domains) */
658 for (i
= 0; i
< csn
; i
++) {
659 struct cpuset
*a
= csa
[i
];
662 for (j
= 0; j
< csn
; j
++) {
663 struct cpuset
*b
= csa
[j
];
666 if (apn
!= bpn
&& cpusets_overlap(a
, b
)) {
667 for (k
= 0; k
< csn
; k
++) {
668 struct cpuset
*c
= csa
[k
];
673 ndoms
--; /* one less element */
680 * Now we know how many domains to create.
681 * Convert <csn, csa> to <ndoms, doms> and populate cpu masks.
683 doms
= alloc_sched_domains(ndoms
);
688 * The rest of the code, including the scheduler, can deal with
689 * dattr==NULL case. No need to abort if alloc fails.
691 dattr
= kmalloc(ndoms
* sizeof(struct sched_domain_attr
), GFP_KERNEL
);
693 for (nslot
= 0, i
= 0; i
< csn
; i
++) {
694 struct cpuset
*a
= csa
[i
];
699 /* Skip completed partitions */
705 if (nslot
== ndoms
) {
706 static int warnings
= 10;
709 "rebuild_sched_domains confused:"
710 " nslot %d, ndoms %d, csn %d, i %d,"
712 nslot
, ndoms
, csn
, i
, apn
);
720 *(dattr
+ nslot
) = SD_ATTR_INIT
;
721 for (j
= i
; j
< csn
; j
++) {
722 struct cpuset
*b
= csa
[j
];
725 cpumask_or(dp
, dp
, b
->cpus_allowed
);
727 update_domain_attr_tree(dattr
+ nslot
, b
);
729 /* Done with this partition */
735 BUG_ON(nslot
!= ndoms
);
741 * Fallback to the default domain if kmalloc() failed.
742 * See comments in partition_sched_domains().
753 * Rebuild scheduler domains.
755 * Call with neither cgroup_mutex held nor within get_online_cpus().
756 * Takes both cgroup_mutex and get_online_cpus().
758 * Cannot be directly called from cpuset code handling changes
759 * to the cpuset pseudo-filesystem, because it cannot be called
760 * from code that already holds cgroup_mutex.
762 static void do_rebuild_sched_domains(struct work_struct
*unused
)
764 struct sched_domain_attr
*attr
;
770 /* Generate domain masks and attrs */
772 ndoms
= generate_sched_domains(&doms
, &attr
);
775 /* Have scheduler rebuild the domains */
776 partition_sched_domains(ndoms
, doms
, attr
);
780 #else /* !CONFIG_SMP */
781 static void do_rebuild_sched_domains(struct work_struct
*unused
)
785 static int generate_sched_domains(cpumask_var_t
**domains
,
786 struct sched_domain_attr
**attributes
)
791 #endif /* CONFIG_SMP */
793 static DECLARE_WORK(rebuild_sched_domains_work
, do_rebuild_sched_domains
);
796 * Rebuild scheduler domains, asynchronously via workqueue.
798 * If the flag 'sched_load_balance' of any cpuset with non-empty
799 * 'cpus' changes, or if the 'cpus' allowed changes in any cpuset
800 * which has that flag enabled, or if any cpuset with a non-empty
801 * 'cpus' is removed, then call this routine to rebuild the
802 * scheduler's dynamic sched domains.
804 * The rebuild_sched_domains() and partition_sched_domains()
805 * routines must nest cgroup_lock() inside get_online_cpus(),
806 * but such cpuset changes as these must nest that locking the
807 * other way, holding cgroup_lock() for much of the code.
809 * So in order to avoid an ABBA deadlock, the cpuset code handling
810 * these user changes delegates the actual sched domain rebuilding
811 * to a separate workqueue thread, which ends up processing the
812 * above do_rebuild_sched_domains() function.
814 static void async_rebuild_sched_domains(void)
816 queue_work(cpuset_wq
, &rebuild_sched_domains_work
);
820 * Accomplishes the same scheduler domain rebuild as the above
821 * async_rebuild_sched_domains(), however it directly calls the
822 * rebuild routine synchronously rather than calling it via an
823 * asynchronous work thread.
825 * This can only be called from code that is not holding
826 * cgroup_mutex (not nested in a cgroup_lock() call.)
828 void rebuild_sched_domains(void)
830 do_rebuild_sched_domains(NULL
);
834 * cpuset_test_cpumask - test a task's cpus_allowed versus its cpuset's
836 * @scan: struct cgroup_scanner contained in its struct cpuset_hotplug_scanner
838 * Call with cgroup_mutex held. May take callback_mutex during call.
839 * Called for each task in a cgroup by cgroup_scan_tasks().
840 * Return nonzero if this tasks's cpus_allowed mask should be changed (in other
841 * words, if its mask is not equal to its cpuset's mask).
843 static int cpuset_test_cpumask(struct task_struct
*tsk
,
844 struct cgroup_scanner
*scan
)
846 return !cpumask_equal(&tsk
->cpus_allowed
,
847 (cgroup_cs(scan
->cg
))->cpus_allowed
);
851 * cpuset_change_cpumask - make a task's cpus_allowed the same as its cpuset's
853 * @scan: struct cgroup_scanner containing the cgroup of the task
855 * Called by cgroup_scan_tasks() for each task in a cgroup whose
856 * cpus_allowed mask needs to be changed.
858 * We don't need to re-check for the cgroup/cpuset membership, since we're
859 * holding cgroup_lock() at this point.
861 static void cpuset_change_cpumask(struct task_struct
*tsk
,
862 struct cgroup_scanner
*scan
)
864 set_cpus_allowed_ptr(tsk
, ((cgroup_cs(scan
->cg
))->cpus_allowed
));
868 * update_tasks_cpumask - Update the cpumasks of tasks in the cpuset.
869 * @cs: the cpuset in which each task's cpus_allowed mask needs to be changed
870 * @heap: if NULL, defer allocating heap memory to cgroup_scan_tasks()
872 * Called with cgroup_mutex held
874 * The cgroup_scan_tasks() function will scan all the tasks in a cgroup,
875 * calling callback functions for each.
877 * No return value. It's guaranteed that cgroup_scan_tasks() always returns 0
880 static void update_tasks_cpumask(struct cpuset
*cs
, struct ptr_heap
*heap
)
882 struct cgroup_scanner scan
;
884 scan
.cg
= cs
->css
.cgroup
;
885 scan
.test_task
= cpuset_test_cpumask
;
886 scan
.process_task
= cpuset_change_cpumask
;
888 cgroup_scan_tasks(&scan
);
892 * update_cpumask - update the cpus_allowed mask of a cpuset and all tasks in it
893 * @cs: the cpuset to consider
894 * @buf: buffer of cpu numbers written to this cpuset
896 static int update_cpumask(struct cpuset
*cs
, struct cpuset
*trialcs
,
899 struct ptr_heap heap
;
901 int is_load_balanced
;
903 /* top_cpuset.cpus_allowed tracks cpu_online_mask; it's read-only */
904 if (cs
== &top_cpuset
)
908 * An empty cpus_allowed is ok only if the cpuset has no tasks.
909 * Since cpulist_parse() fails on an empty mask, we special case
910 * that parsing. The validate_change() call ensures that cpusets
911 * with tasks have cpus.
914 cpumask_clear(trialcs
->cpus_allowed
);
916 retval
= cpulist_parse(buf
, trialcs
->cpus_allowed
);
920 if (!cpumask_subset(trialcs
->cpus_allowed
, cpu_active_mask
))
923 retval
= validate_change(cs
, trialcs
);
927 /* Nothing to do if the cpus didn't change */
928 if (cpumask_equal(cs
->cpus_allowed
, trialcs
->cpus_allowed
))
931 retval
= heap_init(&heap
, PAGE_SIZE
, GFP_KERNEL
, NULL
);
935 is_load_balanced
= is_sched_load_balance(trialcs
);
937 mutex_lock(&callback_mutex
);
938 cpumask_copy(cs
->cpus_allowed
, trialcs
->cpus_allowed
);
939 mutex_unlock(&callback_mutex
);
942 * Scan tasks in the cpuset, and update the cpumasks of any
943 * that need an update.
945 update_tasks_cpumask(cs
, &heap
);
949 if (is_load_balanced
)
950 async_rebuild_sched_domains();
957 * Migrate memory region from one set of nodes to another.
959 * Temporarilly set tasks mems_allowed to target nodes of migration,
960 * so that the migration code can allocate pages on these nodes.
962 * Call holding cgroup_mutex, so current's cpuset won't change
963 * during this call, as manage_mutex holds off any cpuset_attach()
964 * calls. Therefore we don't need to take task_lock around the
965 * call to guarantee_online_mems(), as we know no one is changing
968 * While the mm_struct we are migrating is typically from some
969 * other task, the task_struct mems_allowed that we are hacking
970 * is for our current task, which must allocate new pages for that
971 * migrating memory region.
974 static void cpuset_migrate_mm(struct mm_struct
*mm
, const nodemask_t
*from
,
975 const nodemask_t
*to
)
977 struct task_struct
*tsk
= current
;
979 tsk
->mems_allowed
= *to
;
981 do_migrate_pages(mm
, from
, to
, MPOL_MF_MOVE_ALL
);
983 guarantee_online_mems(task_cs(tsk
),&tsk
->mems_allowed
);
987 * cpuset_change_task_nodemask - change task's mems_allowed and mempolicy
988 * @tsk: the task to change
989 * @newmems: new nodes that the task will be set
991 * In order to avoid seeing no nodes if the old and new nodes are disjoint,
992 * we structure updates as setting all new allowed nodes, then clearing newly
995 static void cpuset_change_task_nodemask(struct task_struct
*tsk
,
1001 * Allow tasks that have access to memory reserves because they have
1002 * been OOM killed to get memory anywhere.
1004 if (unlikely(test_thread_flag(TIF_MEMDIE
)))
1006 if (current
->flags
& PF_EXITING
) /* Let dying task have memory */
1011 * Determine if a loop is necessary if another thread is doing
1012 * get_mems_allowed(). If at least one node remains unchanged and
1013 * tsk does not have a mempolicy, then an empty nodemask will not be
1014 * possible when mems_allowed is larger than a word.
1016 need_loop
= task_has_mempolicy(tsk
) ||
1017 !nodes_intersects(*newmems
, tsk
->mems_allowed
);
1020 write_seqcount_begin(&tsk
->mems_allowed_seq
);
1022 nodes_or(tsk
->mems_allowed
, tsk
->mems_allowed
, *newmems
);
1023 mpol_rebind_task(tsk
, newmems
, MPOL_REBIND_STEP1
);
1025 mpol_rebind_task(tsk
, newmems
, MPOL_REBIND_STEP2
);
1026 tsk
->mems_allowed
= *newmems
;
1029 write_seqcount_end(&tsk
->mems_allowed_seq
);
1035 * Update task's mems_allowed and rebind its mempolicy and vmas' mempolicy
1036 * of it to cpuset's new mems_allowed, and migrate pages to new nodes if
1037 * memory_migrate flag is set. Called with cgroup_mutex held.
1039 static void cpuset_change_nodemask(struct task_struct
*p
,
1040 struct cgroup_scanner
*scan
)
1042 struct mm_struct
*mm
;
1045 const nodemask_t
*oldmem
= scan
->data
;
1046 static nodemask_t newmems
; /* protected by cgroup_mutex */
1048 cs
= cgroup_cs(scan
->cg
);
1049 guarantee_online_mems(cs
, &newmems
);
1051 cpuset_change_task_nodemask(p
, &newmems
);
1053 mm
= get_task_mm(p
);
1057 migrate
= is_memory_migrate(cs
);
1059 mpol_rebind_mm(mm
, &cs
->mems_allowed
);
1061 cpuset_migrate_mm(mm
, oldmem
, &cs
->mems_allowed
);
1065 static void *cpuset_being_rebound
;
1068 * update_tasks_nodemask - Update the nodemasks of tasks in the cpuset.
1069 * @cs: the cpuset in which each task's mems_allowed mask needs to be changed
1070 * @oldmem: old mems_allowed of cpuset cs
1071 * @heap: if NULL, defer allocating heap memory to cgroup_scan_tasks()
1073 * Called with cgroup_mutex held
1074 * No return value. It's guaranteed that cgroup_scan_tasks() always returns 0
1077 static void update_tasks_nodemask(struct cpuset
*cs
, const nodemask_t
*oldmem
,
1078 struct ptr_heap
*heap
)
1080 struct cgroup_scanner scan
;
1082 cpuset_being_rebound
= cs
; /* causes mpol_dup() rebind */
1084 scan
.cg
= cs
->css
.cgroup
;
1085 scan
.test_task
= NULL
;
1086 scan
.process_task
= cpuset_change_nodemask
;
1088 scan
.data
= (nodemask_t
*)oldmem
;
1091 * The mpol_rebind_mm() call takes mmap_sem, which we couldn't
1092 * take while holding tasklist_lock. Forks can happen - the
1093 * mpol_dup() cpuset_being_rebound check will catch such forks,
1094 * and rebind their vma mempolicies too. Because we still hold
1095 * the global cgroup_mutex, we know that no other rebind effort
1096 * will be contending for the global variable cpuset_being_rebound.
1097 * It's ok if we rebind the same mm twice; mpol_rebind_mm()
1098 * is idempotent. Also migrate pages in each mm to new nodes.
1100 cgroup_scan_tasks(&scan
);
1102 /* We're done rebinding vmas to this cpuset's new mems_allowed. */
1103 cpuset_being_rebound
= NULL
;
1107 * Handle user request to change the 'mems' memory placement
1108 * of a cpuset. Needs to validate the request, update the
1109 * cpusets mems_allowed, and for each task in the cpuset,
1110 * update mems_allowed and rebind task's mempolicy and any vma
1111 * mempolicies and if the cpuset is marked 'memory_migrate',
1112 * migrate the tasks pages to the new memory.
1114 * Call with cgroup_mutex held. May take callback_mutex during call.
1115 * Will take tasklist_lock, scan tasklist for tasks in cpuset cs,
1116 * lock each such tasks mm->mmap_sem, scan its vma's and rebind
1117 * their mempolicies to the cpusets new mems_allowed.
1119 static int update_nodemask(struct cpuset
*cs
, struct cpuset
*trialcs
,
1122 NODEMASK_ALLOC(nodemask_t
, oldmem
, GFP_KERNEL
);
1124 struct ptr_heap heap
;
1130 * top_cpuset.mems_allowed tracks node_stats[N_MEMORY];
1133 if (cs
== &top_cpuset
) {
1139 * An empty mems_allowed is ok iff there are no tasks in the cpuset.
1140 * Since nodelist_parse() fails on an empty mask, we special case
1141 * that parsing. The validate_change() call ensures that cpusets
1142 * with tasks have memory.
1145 nodes_clear(trialcs
->mems_allowed
);
1147 retval
= nodelist_parse(buf
, trialcs
->mems_allowed
);
1151 if (!nodes_subset(trialcs
->mems_allowed
,
1152 node_states
[N_MEMORY
])) {
1157 *oldmem
= cs
->mems_allowed
;
1158 if (nodes_equal(*oldmem
, trialcs
->mems_allowed
)) {
1159 retval
= 0; /* Too easy - nothing to do */
1162 retval
= validate_change(cs
, trialcs
);
1166 retval
= heap_init(&heap
, PAGE_SIZE
, GFP_KERNEL
, NULL
);
1170 mutex_lock(&callback_mutex
);
1171 cs
->mems_allowed
= trialcs
->mems_allowed
;
1172 mutex_unlock(&callback_mutex
);
1174 update_tasks_nodemask(cs
, oldmem
, &heap
);
1178 NODEMASK_FREE(oldmem
);
1182 int current_cpuset_is_being_rebound(void)
1184 return task_cs(current
) == cpuset_being_rebound
;
1187 static int update_relax_domain_level(struct cpuset
*cs
, s64 val
)
1190 if (val
< -1 || val
>= sched_domain_level_max
)
1194 if (val
!= cs
->relax_domain_level
) {
1195 cs
->relax_domain_level
= val
;
1196 if (!cpumask_empty(cs
->cpus_allowed
) &&
1197 is_sched_load_balance(cs
))
1198 async_rebuild_sched_domains();
1205 * cpuset_change_flag - make a task's spread flags the same as its cpuset's
1206 * @tsk: task to be updated
1207 * @scan: struct cgroup_scanner containing the cgroup of the task
1209 * Called by cgroup_scan_tasks() for each task in a cgroup.
1211 * We don't need to re-check for the cgroup/cpuset membership, since we're
1212 * holding cgroup_lock() at this point.
1214 static void cpuset_change_flag(struct task_struct
*tsk
,
1215 struct cgroup_scanner
*scan
)
1217 cpuset_update_task_spread_flag(cgroup_cs(scan
->cg
), tsk
);
1221 * update_tasks_flags - update the spread flags of tasks in the cpuset.
1222 * @cs: the cpuset in which each task's spread flags needs to be changed
1223 * @heap: if NULL, defer allocating heap memory to cgroup_scan_tasks()
1225 * Called with cgroup_mutex held
1227 * The cgroup_scan_tasks() function will scan all the tasks in a cgroup,
1228 * calling callback functions for each.
1230 * No return value. It's guaranteed that cgroup_scan_tasks() always returns 0
1233 static void update_tasks_flags(struct cpuset
*cs
, struct ptr_heap
*heap
)
1235 struct cgroup_scanner scan
;
1237 scan
.cg
= cs
->css
.cgroup
;
1238 scan
.test_task
= NULL
;
1239 scan
.process_task
= cpuset_change_flag
;
1241 cgroup_scan_tasks(&scan
);
1245 * update_flag - read a 0 or a 1 in a file and update associated flag
1246 * bit: the bit to update (see cpuset_flagbits_t)
1247 * cs: the cpuset to update
1248 * turning_on: whether the flag is being set or cleared
1250 * Call with cgroup_mutex held.
1253 static int update_flag(cpuset_flagbits_t bit
, struct cpuset
*cs
,
1256 struct cpuset
*trialcs
;
1257 int balance_flag_changed
;
1258 int spread_flag_changed
;
1259 struct ptr_heap heap
;
1262 trialcs
= alloc_trial_cpuset(cs
);
1267 set_bit(bit
, &trialcs
->flags
);
1269 clear_bit(bit
, &trialcs
->flags
);
1271 err
= validate_change(cs
, trialcs
);
1275 err
= heap_init(&heap
, PAGE_SIZE
, GFP_KERNEL
, NULL
);
1279 balance_flag_changed
= (is_sched_load_balance(cs
) !=
1280 is_sched_load_balance(trialcs
));
1282 spread_flag_changed
= ((is_spread_slab(cs
) != is_spread_slab(trialcs
))
1283 || (is_spread_page(cs
) != is_spread_page(trialcs
)));
1285 mutex_lock(&callback_mutex
);
1286 cs
->flags
= trialcs
->flags
;
1287 mutex_unlock(&callback_mutex
);
1289 if (!cpumask_empty(trialcs
->cpus_allowed
) && balance_flag_changed
)
1290 async_rebuild_sched_domains();
1292 if (spread_flag_changed
)
1293 update_tasks_flags(cs
, &heap
);
1296 free_trial_cpuset(trialcs
);
1301 * Frequency meter - How fast is some event occurring?
1303 * These routines manage a digitally filtered, constant time based,
1304 * event frequency meter. There are four routines:
1305 * fmeter_init() - initialize a frequency meter.
1306 * fmeter_markevent() - called each time the event happens.
1307 * fmeter_getrate() - returns the recent rate of such events.
1308 * fmeter_update() - internal routine used to update fmeter.
1310 * A common data structure is passed to each of these routines,
1311 * which is used to keep track of the state required to manage the
1312 * frequency meter and its digital filter.
1314 * The filter works on the number of events marked per unit time.
1315 * The filter is single-pole low-pass recursive (IIR). The time unit
1316 * is 1 second. Arithmetic is done using 32-bit integers scaled to
1317 * simulate 3 decimal digits of precision (multiplied by 1000).
1319 * With an FM_COEF of 933, and a time base of 1 second, the filter
1320 * has a half-life of 10 seconds, meaning that if the events quit
1321 * happening, then the rate returned from the fmeter_getrate()
1322 * will be cut in half each 10 seconds, until it converges to zero.
1324 * It is not worth doing a real infinitely recursive filter. If more
1325 * than FM_MAXTICKS ticks have elapsed since the last filter event,
1326 * just compute FM_MAXTICKS ticks worth, by which point the level
1329 * Limit the count of unprocessed events to FM_MAXCNT, so as to avoid
1330 * arithmetic overflow in the fmeter_update() routine.
1332 * Given the simple 32 bit integer arithmetic used, this meter works
1333 * best for reporting rates between one per millisecond (msec) and
1334 * one per 32 (approx) seconds. At constant rates faster than one
1335 * per msec it maxes out at values just under 1,000,000. At constant
1336 * rates between one per msec, and one per second it will stabilize
1337 * to a value N*1000, where N is the rate of events per second.
1338 * At constant rates between one per second and one per 32 seconds,
1339 * it will be choppy, moving up on the seconds that have an event,
1340 * and then decaying until the next event. At rates slower than
1341 * about one in 32 seconds, it decays all the way back to zero between
1345 #define FM_COEF 933 /* coefficient for half-life of 10 secs */
1346 #define FM_MAXTICKS ((time_t)99) /* useless computing more ticks than this */
1347 #define FM_MAXCNT 1000000 /* limit cnt to avoid overflow */
1348 #define FM_SCALE 1000 /* faux fixed point scale */
1350 /* Initialize a frequency meter */
1351 static void fmeter_init(struct fmeter
*fmp
)
1356 spin_lock_init(&fmp
->lock
);
1359 /* Internal meter update - process cnt events and update value */
1360 static void fmeter_update(struct fmeter
*fmp
)
1362 time_t now
= get_seconds();
1363 time_t ticks
= now
- fmp
->time
;
1368 ticks
= min(FM_MAXTICKS
, ticks
);
1370 fmp
->val
= (FM_COEF
* fmp
->val
) / FM_SCALE
;
1373 fmp
->val
+= ((FM_SCALE
- FM_COEF
) * fmp
->cnt
) / FM_SCALE
;
1377 /* Process any previous ticks, then bump cnt by one (times scale). */
1378 static void fmeter_markevent(struct fmeter
*fmp
)
1380 spin_lock(&fmp
->lock
);
1382 fmp
->cnt
= min(FM_MAXCNT
, fmp
->cnt
+ FM_SCALE
);
1383 spin_unlock(&fmp
->lock
);
1386 /* Process any previous ticks, then return current value. */
1387 static int fmeter_getrate(struct fmeter
*fmp
)
1391 spin_lock(&fmp
->lock
);
1394 spin_unlock(&fmp
->lock
);
1398 /* Called by cgroups to determine if a cpuset is usable; cgroup_mutex held */
1399 static int cpuset_can_attach(struct cgroup
*cgrp
, struct cgroup_taskset
*tset
)
1401 struct cpuset
*cs
= cgroup_cs(cgrp
);
1402 struct task_struct
*task
;
1405 if (cpumask_empty(cs
->cpus_allowed
) || nodes_empty(cs
->mems_allowed
))
1408 cgroup_taskset_for_each(task
, cgrp
, tset
) {
1410 * Kthreads bound to specific cpus cannot be moved to a new
1411 * cpuset; we cannot change their cpu affinity and
1412 * isolating such threads by their set of allowed nodes is
1413 * unnecessary. Thus, cpusets are not applicable for such
1414 * threads. This prevents checking for success of
1415 * set_cpus_allowed_ptr() on all attached tasks before
1416 * cpus_allowed may be changed.
1418 if (task
->flags
& PF_THREAD_BOUND
)
1420 if ((ret
= security_task_setscheduler(task
)))
1428 * Protected by cgroup_mutex. cpus_attach is used only by cpuset_attach()
1429 * but we can't allocate it dynamically there. Define it global and
1430 * allocate from cpuset_init().
1432 static cpumask_var_t cpus_attach
;
1434 static void cpuset_attach(struct cgroup
*cgrp
, struct cgroup_taskset
*tset
)
1436 /* static bufs protected by cgroup_mutex */
1437 static nodemask_t cpuset_attach_nodemask_from
;
1438 static nodemask_t cpuset_attach_nodemask_to
;
1439 struct mm_struct
*mm
;
1440 struct task_struct
*task
;
1441 struct task_struct
*leader
= cgroup_taskset_first(tset
);
1442 struct cgroup
*oldcgrp
= cgroup_taskset_cur_cgroup(tset
);
1443 struct cpuset
*cs
= cgroup_cs(cgrp
);
1444 struct cpuset
*oldcs
= cgroup_cs(oldcgrp
);
1446 /* prepare for attach */
1447 if (cs
== &top_cpuset
)
1448 cpumask_copy(cpus_attach
, cpu_possible_mask
);
1450 guarantee_online_cpus(cs
, cpus_attach
);
1452 guarantee_online_mems(cs
, &cpuset_attach_nodemask_to
);
1454 cgroup_taskset_for_each(task
, cgrp
, tset
) {
1456 * can_attach beforehand should guarantee that this doesn't
1457 * fail. TODO: have a better way to handle failure here
1459 WARN_ON_ONCE(set_cpus_allowed_ptr(task
, cpus_attach
));
1461 cpuset_change_task_nodemask(task
, &cpuset_attach_nodemask_to
);
1462 cpuset_update_task_spread_flag(cs
, task
);
1466 * Change mm, possibly for multiple threads in a threadgroup. This is
1467 * expensive and may sleep.
1469 cpuset_attach_nodemask_from
= oldcs
->mems_allowed
;
1470 cpuset_attach_nodemask_to
= cs
->mems_allowed
;
1471 mm
= get_task_mm(leader
);
1473 mpol_rebind_mm(mm
, &cpuset_attach_nodemask_to
);
1474 if (is_memory_migrate(cs
))
1475 cpuset_migrate_mm(mm
, &cpuset_attach_nodemask_from
,
1476 &cpuset_attach_nodemask_to
);
1481 /* The various types of files and directories in a cpuset file system */
1484 FILE_MEMORY_MIGRATE
,
1490 FILE_SCHED_LOAD_BALANCE
,
1491 FILE_SCHED_RELAX_DOMAIN_LEVEL
,
1492 FILE_MEMORY_PRESSURE_ENABLED
,
1493 FILE_MEMORY_PRESSURE
,
1496 } cpuset_filetype_t
;
1498 static int cpuset_write_u64(struct cgroup
*cgrp
, struct cftype
*cft
, u64 val
)
1501 struct cpuset
*cs
= cgroup_cs(cgrp
);
1502 cpuset_filetype_t type
= cft
->private;
1504 if (!cgroup_lock_live_group(cgrp
))
1508 case FILE_CPU_EXCLUSIVE
:
1509 retval
= update_flag(CS_CPU_EXCLUSIVE
, cs
, val
);
1511 case FILE_MEM_EXCLUSIVE
:
1512 retval
= update_flag(CS_MEM_EXCLUSIVE
, cs
, val
);
1514 case FILE_MEM_HARDWALL
:
1515 retval
= update_flag(CS_MEM_HARDWALL
, cs
, val
);
1517 case FILE_SCHED_LOAD_BALANCE
:
1518 retval
= update_flag(CS_SCHED_LOAD_BALANCE
, cs
, val
);
1520 case FILE_MEMORY_MIGRATE
:
1521 retval
= update_flag(CS_MEMORY_MIGRATE
, cs
, val
);
1523 case FILE_MEMORY_PRESSURE_ENABLED
:
1524 cpuset_memory_pressure_enabled
= !!val
;
1526 case FILE_MEMORY_PRESSURE
:
1529 case FILE_SPREAD_PAGE
:
1530 retval
= update_flag(CS_SPREAD_PAGE
, cs
, val
);
1532 case FILE_SPREAD_SLAB
:
1533 retval
= update_flag(CS_SPREAD_SLAB
, cs
, val
);
1543 static int cpuset_write_s64(struct cgroup
*cgrp
, struct cftype
*cft
, s64 val
)
1546 struct cpuset
*cs
= cgroup_cs(cgrp
);
1547 cpuset_filetype_t type
= cft
->private;
1549 if (!cgroup_lock_live_group(cgrp
))
1553 case FILE_SCHED_RELAX_DOMAIN_LEVEL
:
1554 retval
= update_relax_domain_level(cs
, val
);
1565 * Common handling for a write to a "cpus" or "mems" file.
1567 static int cpuset_write_resmask(struct cgroup
*cgrp
, struct cftype
*cft
,
1571 struct cpuset
*cs
= cgroup_cs(cgrp
);
1572 struct cpuset
*trialcs
;
1574 if (!cgroup_lock_live_group(cgrp
))
1577 trialcs
= alloc_trial_cpuset(cs
);
1583 switch (cft
->private) {
1585 retval
= update_cpumask(cs
, trialcs
, buf
);
1588 retval
= update_nodemask(cs
, trialcs
, buf
);
1595 free_trial_cpuset(trialcs
);
1602 * These ascii lists should be read in a single call, by using a user
1603 * buffer large enough to hold the entire map. If read in smaller
1604 * chunks, there is no guarantee of atomicity. Since the display format
1605 * used, list of ranges of sequential numbers, is variable length,
1606 * and since these maps can change value dynamically, one could read
1607 * gibberish by doing partial reads while a list was changing.
1608 * A single large read to a buffer that crosses a page boundary is
1609 * ok, because the result being copied to user land is not recomputed
1610 * across a page fault.
1613 static size_t cpuset_sprintf_cpulist(char *page
, struct cpuset
*cs
)
1617 mutex_lock(&callback_mutex
);
1618 count
= cpulist_scnprintf(page
, PAGE_SIZE
, cs
->cpus_allowed
);
1619 mutex_unlock(&callback_mutex
);
1624 static size_t cpuset_sprintf_memlist(char *page
, struct cpuset
*cs
)
1628 mutex_lock(&callback_mutex
);
1629 count
= nodelist_scnprintf(page
, PAGE_SIZE
, cs
->mems_allowed
);
1630 mutex_unlock(&callback_mutex
);
1635 static ssize_t
cpuset_common_file_read(struct cgroup
*cont
,
1639 size_t nbytes
, loff_t
*ppos
)
1641 struct cpuset
*cs
= cgroup_cs(cont
);
1642 cpuset_filetype_t type
= cft
->private;
1647 if (!(page
= (char *)__get_free_page(GFP_TEMPORARY
)))
1654 s
+= cpuset_sprintf_cpulist(s
, cs
);
1657 s
+= cpuset_sprintf_memlist(s
, cs
);
1665 retval
= simple_read_from_buffer(buf
, nbytes
, ppos
, page
, s
- page
);
1667 free_page((unsigned long)page
);
1671 static u64
cpuset_read_u64(struct cgroup
*cont
, struct cftype
*cft
)
1673 struct cpuset
*cs
= cgroup_cs(cont
);
1674 cpuset_filetype_t type
= cft
->private;
1676 case FILE_CPU_EXCLUSIVE
:
1677 return is_cpu_exclusive(cs
);
1678 case FILE_MEM_EXCLUSIVE
:
1679 return is_mem_exclusive(cs
);
1680 case FILE_MEM_HARDWALL
:
1681 return is_mem_hardwall(cs
);
1682 case FILE_SCHED_LOAD_BALANCE
:
1683 return is_sched_load_balance(cs
);
1684 case FILE_MEMORY_MIGRATE
:
1685 return is_memory_migrate(cs
);
1686 case FILE_MEMORY_PRESSURE_ENABLED
:
1687 return cpuset_memory_pressure_enabled
;
1688 case FILE_MEMORY_PRESSURE
:
1689 return fmeter_getrate(&cs
->fmeter
);
1690 case FILE_SPREAD_PAGE
:
1691 return is_spread_page(cs
);
1692 case FILE_SPREAD_SLAB
:
1693 return is_spread_slab(cs
);
1698 /* Unreachable but makes gcc happy */
1702 static s64
cpuset_read_s64(struct cgroup
*cont
, struct cftype
*cft
)
1704 struct cpuset
*cs
= cgroup_cs(cont
);
1705 cpuset_filetype_t type
= cft
->private;
1707 case FILE_SCHED_RELAX_DOMAIN_LEVEL
:
1708 return cs
->relax_domain_level
;
1713 /* Unrechable but makes gcc happy */
1719 * for the common functions, 'private' gives the type of file
1722 static struct cftype files
[] = {
1725 .read
= cpuset_common_file_read
,
1726 .write_string
= cpuset_write_resmask
,
1727 .max_write_len
= (100U + 6 * NR_CPUS
),
1728 .private = FILE_CPULIST
,
1733 .read
= cpuset_common_file_read
,
1734 .write_string
= cpuset_write_resmask
,
1735 .max_write_len
= (100U + 6 * MAX_NUMNODES
),
1736 .private = FILE_MEMLIST
,
1740 .name
= "cpu_exclusive",
1741 .read_u64
= cpuset_read_u64
,
1742 .write_u64
= cpuset_write_u64
,
1743 .private = FILE_CPU_EXCLUSIVE
,
1747 .name
= "mem_exclusive",
1748 .read_u64
= cpuset_read_u64
,
1749 .write_u64
= cpuset_write_u64
,
1750 .private = FILE_MEM_EXCLUSIVE
,
1754 .name
= "mem_hardwall",
1755 .read_u64
= cpuset_read_u64
,
1756 .write_u64
= cpuset_write_u64
,
1757 .private = FILE_MEM_HARDWALL
,
1761 .name
= "sched_load_balance",
1762 .read_u64
= cpuset_read_u64
,
1763 .write_u64
= cpuset_write_u64
,
1764 .private = FILE_SCHED_LOAD_BALANCE
,
1768 .name
= "sched_relax_domain_level",
1769 .read_s64
= cpuset_read_s64
,
1770 .write_s64
= cpuset_write_s64
,
1771 .private = FILE_SCHED_RELAX_DOMAIN_LEVEL
,
1775 .name
= "memory_migrate",
1776 .read_u64
= cpuset_read_u64
,
1777 .write_u64
= cpuset_write_u64
,
1778 .private = FILE_MEMORY_MIGRATE
,
1782 .name
= "memory_pressure",
1783 .read_u64
= cpuset_read_u64
,
1784 .write_u64
= cpuset_write_u64
,
1785 .private = FILE_MEMORY_PRESSURE
,
1790 .name
= "memory_spread_page",
1791 .read_u64
= cpuset_read_u64
,
1792 .write_u64
= cpuset_write_u64
,
1793 .private = FILE_SPREAD_PAGE
,
1797 .name
= "memory_spread_slab",
1798 .read_u64
= cpuset_read_u64
,
1799 .write_u64
= cpuset_write_u64
,
1800 .private = FILE_SPREAD_SLAB
,
1804 .name
= "memory_pressure_enabled",
1805 .flags
= CFTYPE_ONLY_ON_ROOT
,
1806 .read_u64
= cpuset_read_u64
,
1807 .write_u64
= cpuset_write_u64
,
1808 .private = FILE_MEMORY_PRESSURE_ENABLED
,
1815 * cpuset_css_alloc - allocate a cpuset css
1816 * cont: control group that the new cpuset will be part of
1819 static struct cgroup_subsys_state
*cpuset_css_alloc(struct cgroup
*cont
)
1824 return &top_cpuset
.css
;
1826 cs
= kzalloc(sizeof(*cs
), GFP_KERNEL
);
1828 return ERR_PTR(-ENOMEM
);
1829 if (!alloc_cpumask_var(&cs
->cpus_allowed
, GFP_KERNEL
)) {
1831 return ERR_PTR(-ENOMEM
);
1834 set_bit(CS_SCHED_LOAD_BALANCE
, &cs
->flags
);
1835 cpumask_clear(cs
->cpus_allowed
);
1836 nodes_clear(cs
->mems_allowed
);
1837 fmeter_init(&cs
->fmeter
);
1838 cs
->relax_domain_level
= -1;
1839 cs
->parent
= cgroup_cs(cont
->parent
);
1844 static int cpuset_css_online(struct cgroup
*cgrp
)
1846 struct cpuset
*cs
= cgroup_cs(cgrp
);
1847 struct cpuset
*parent
= cs
->parent
;
1848 struct cpuset
*tmp_cs
;
1849 struct cgroup
*pos_cg
;
1854 set_bit(CS_ONLINE
, &cs
->flags
);
1855 if (is_spread_page(parent
))
1856 set_bit(CS_SPREAD_PAGE
, &cs
->flags
);
1857 if (is_spread_slab(parent
))
1858 set_bit(CS_SPREAD_SLAB
, &cs
->flags
);
1860 number_of_cpusets
++;
1862 if (!test_bit(CGRP_CPUSET_CLONE_CHILDREN
, &cgrp
->flags
))
1866 * Clone @parent's configuration if CGRP_CPUSET_CLONE_CHILDREN is
1867 * set. This flag handling is implemented in cgroup core for
1868 * histrical reasons - the flag may be specified during mount.
1870 * Currently, if any sibling cpusets have exclusive cpus or mem, we
1871 * refuse to clone the configuration - thereby refusing the task to
1872 * be entered, and as a result refusing the sys_unshare() or
1873 * clone() which initiated it. If this becomes a problem for some
1874 * users who wish to allow that scenario, then this could be
1875 * changed to grant parent->cpus_allowed-sibling_cpus_exclusive
1876 * (and likewise for mems) to the new cgroup.
1879 cpuset_for_each_child(tmp_cs
, pos_cg
, parent
) {
1880 if (is_mem_exclusive(tmp_cs
) || is_cpu_exclusive(tmp_cs
)) {
1887 mutex_lock(&callback_mutex
);
1888 cs
->mems_allowed
= parent
->mems_allowed
;
1889 cpumask_copy(cs
->cpus_allowed
, parent
->cpus_allowed
);
1890 mutex_unlock(&callback_mutex
);
1895 static void cpuset_css_offline(struct cgroup
*cgrp
)
1897 struct cpuset
*cs
= cgroup_cs(cgrp
);
1899 /* css_offline is called w/o cgroup_mutex, grab it */
1902 if (is_sched_load_balance(cs
))
1903 update_flag(CS_SCHED_LOAD_BALANCE
, cs
, 0);
1905 number_of_cpusets
--;
1906 clear_bit(CS_ONLINE
, &cs
->flags
);
1912 * If the cpuset being removed has its flag 'sched_load_balance'
1913 * enabled, then simulate turning sched_load_balance off, which
1914 * will call async_rebuild_sched_domains().
1917 static void cpuset_css_free(struct cgroup
*cont
)
1919 struct cpuset
*cs
= cgroup_cs(cont
);
1921 free_cpumask_var(cs
->cpus_allowed
);
1925 struct cgroup_subsys cpuset_subsys
= {
1927 .css_alloc
= cpuset_css_alloc
,
1928 .css_online
= cpuset_css_online
,
1929 .css_offline
= cpuset_css_offline
,
1930 .css_free
= cpuset_css_free
,
1931 .can_attach
= cpuset_can_attach
,
1932 .attach
= cpuset_attach
,
1933 .subsys_id
= cpuset_subsys_id
,
1934 .base_cftypes
= files
,
1939 * cpuset_init - initialize cpusets at system boot
1941 * Description: Initialize top_cpuset and the cpuset internal file system,
1944 int __init
cpuset_init(void)
1948 if (!alloc_cpumask_var(&top_cpuset
.cpus_allowed
, GFP_KERNEL
))
1951 cpumask_setall(top_cpuset
.cpus_allowed
);
1952 nodes_setall(top_cpuset
.mems_allowed
);
1954 fmeter_init(&top_cpuset
.fmeter
);
1955 set_bit(CS_SCHED_LOAD_BALANCE
, &top_cpuset
.flags
);
1956 top_cpuset
.relax_domain_level
= -1;
1958 err
= register_filesystem(&cpuset_fs_type
);
1962 if (!alloc_cpumask_var(&cpus_attach
, GFP_KERNEL
))
1965 number_of_cpusets
= 1;
1970 * cpuset_do_move_task - move a given task to another cpuset
1971 * @tsk: pointer to task_struct the task to move
1972 * @scan: struct cgroup_scanner contained in its struct cpuset_hotplug_scanner
1974 * Called by cgroup_scan_tasks() for each task in a cgroup.
1975 * Return nonzero to stop the walk through the tasks.
1977 static void cpuset_do_move_task(struct task_struct
*tsk
,
1978 struct cgroup_scanner
*scan
)
1980 struct cgroup
*new_cgroup
= scan
->data
;
1982 cgroup_attach_task(new_cgroup
, tsk
);
1986 * move_member_tasks_to_cpuset - move tasks from one cpuset to another
1987 * @from: cpuset in which the tasks currently reside
1988 * @to: cpuset to which the tasks will be moved
1990 * Called with cgroup_mutex held
1991 * callback_mutex must not be held, as cpuset_attach() will take it.
1993 * The cgroup_scan_tasks() function will scan all the tasks in a cgroup,
1994 * calling callback functions for each.
1996 static void move_member_tasks_to_cpuset(struct cpuset
*from
, struct cpuset
*to
)
1998 struct cgroup_scanner scan
;
2000 scan
.cg
= from
->css
.cgroup
;
2001 scan
.test_task
= NULL
; /* select all tasks in cgroup */
2002 scan
.process_task
= cpuset_do_move_task
;
2004 scan
.data
= to
->css
.cgroup
;
2006 if (cgroup_scan_tasks(&scan
))
2007 printk(KERN_ERR
"move_member_tasks_to_cpuset: "
2008 "cgroup_scan_tasks failed\n");
2012 * If CPU and/or memory hotplug handlers, below, unplug any CPUs
2013 * or memory nodes, we need to walk over the cpuset hierarchy,
2014 * removing that CPU or node from all cpusets. If this removes the
2015 * last CPU or node from a cpuset, then move the tasks in the empty
2016 * cpuset to its next-highest non-empty parent.
2018 * Called with cgroup_mutex held
2019 * callback_mutex must not be held, as cpuset_attach() will take it.
2021 static void remove_tasks_in_empty_cpuset(struct cpuset
*cs
)
2023 struct cpuset
*parent
;
2026 * Find its next-highest non-empty parent, (top cpuset
2027 * has online cpus, so can't be empty).
2029 parent
= cs
->parent
;
2030 while (cpumask_empty(parent
->cpus_allowed
) ||
2031 nodes_empty(parent
->mems_allowed
))
2032 parent
= parent
->parent
;
2034 move_member_tasks_to_cpuset(cs
, parent
);
2038 * Helper function to traverse cpusets.
2039 * It can be used to walk the cpuset tree from top to bottom, completing
2040 * one layer before dropping down to the next (thus always processing a
2041 * node before any of its children).
2043 static struct cpuset
*cpuset_next(struct list_head
*queue
)
2046 struct cpuset
*child
; /* scans child cpusets of cp */
2047 struct cgroup
*cont
;
2049 if (list_empty(queue
))
2052 cp
= list_first_entry(queue
, struct cpuset
, stack_list
);
2053 list_del(queue
->next
);
2055 cpuset_for_each_child(child
, cont
, cp
)
2056 list_add_tail(&child
->stack_list
, queue
);
2064 * Walk the specified cpuset subtree upon a hotplug operation (CPU/Memory
2065 * online/offline) and update the cpusets accordingly.
2066 * For regular CPU/Mem hotplug, look for empty cpusets; the tasks of such
2067 * cpuset must be moved to a parent cpuset.
2069 * Called with cgroup_mutex held. We take callback_mutex to modify
2070 * cpus_allowed and mems_allowed.
2072 * This walk processes the tree from top to bottom, completing one layer
2073 * before dropping down to the next. It always processes a node before
2074 * any of its children.
2076 * In the case of memory hot-unplug, it will remove nodes from N_MEMORY
2077 * if all present pages from a node are offlined.
2080 scan_cpusets_upon_hotplug(struct cpuset
*root
, enum hotplug_event event
)
2083 struct cpuset
*cp
; /* scans cpusets being updated */
2084 static nodemask_t oldmems
; /* protected by cgroup_mutex */
2086 list_add_tail((struct list_head
*)&root
->stack_list
, &queue
);
2089 case CPUSET_CPU_OFFLINE
:
2090 while ((cp
= cpuset_next(&queue
)) != NULL
) {
2092 /* Continue past cpusets with all cpus online */
2093 if (cpumask_subset(cp
->cpus_allowed
, cpu_active_mask
))
2096 /* Remove offline cpus from this cpuset. */
2097 mutex_lock(&callback_mutex
);
2098 cpumask_and(cp
->cpus_allowed
, cp
->cpus_allowed
,
2100 mutex_unlock(&callback_mutex
);
2102 /* Move tasks from the empty cpuset to a parent */
2103 if (cpumask_empty(cp
->cpus_allowed
))
2104 remove_tasks_in_empty_cpuset(cp
);
2106 update_tasks_cpumask(cp
, NULL
);
2110 case CPUSET_MEM_OFFLINE
:
2111 while ((cp
= cpuset_next(&queue
)) != NULL
) {
2113 /* Continue past cpusets with all mems online */
2114 if (nodes_subset(cp
->mems_allowed
,
2115 node_states
[N_MEMORY
]))
2118 oldmems
= cp
->mems_allowed
;
2120 /* Remove offline mems from this cpuset. */
2121 mutex_lock(&callback_mutex
);
2122 nodes_and(cp
->mems_allowed
, cp
->mems_allowed
,
2123 node_states
[N_MEMORY
]);
2124 mutex_unlock(&callback_mutex
);
2126 /* Move tasks from the empty cpuset to a parent */
2127 if (nodes_empty(cp
->mems_allowed
))
2128 remove_tasks_in_empty_cpuset(cp
);
2130 update_tasks_nodemask(cp
, &oldmems
, NULL
);
2136 * The top_cpuset tracks what CPUs and Memory Nodes are online,
2137 * period. This is necessary in order to make cpusets transparent
2138 * (of no affect) on systems that are actively using CPU hotplug
2139 * but making no active use of cpusets.
2141 * The only exception to this is suspend/resume, where we don't
2142 * modify cpusets at all.
2144 * This routine ensures that top_cpuset.cpus_allowed tracks
2145 * cpu_active_mask on each CPU hotplug (cpuhp) event.
2147 * Called within get_online_cpus(). Needs to call cgroup_lock()
2148 * before calling generate_sched_domains().
2150 * @cpu_online: Indicates whether this is a CPU online event (true) or
2151 * a CPU offline event (false).
2153 void cpuset_update_active_cpus(bool cpu_online
)
2155 struct sched_domain_attr
*attr
;
2156 cpumask_var_t
*doms
;
2160 mutex_lock(&callback_mutex
);
2161 cpumask_copy(top_cpuset
.cpus_allowed
, cpu_active_mask
);
2162 mutex_unlock(&callback_mutex
);
2165 scan_cpusets_upon_hotplug(&top_cpuset
, CPUSET_CPU_OFFLINE
);
2167 ndoms
= generate_sched_domains(&doms
, &attr
);
2170 /* Have scheduler rebuild the domains */
2171 partition_sched_domains(ndoms
, doms
, attr
);
2174 #ifdef CONFIG_MEMORY_HOTPLUG
2176 * Keep top_cpuset.mems_allowed tracking node_states[N_MEMORY].
2177 * Call this routine anytime after node_states[N_MEMORY] changes.
2178 * See cpuset_update_active_cpus() for CPU hotplug handling.
2180 static int cpuset_track_online_nodes(struct notifier_block
*self
,
2181 unsigned long action
, void *arg
)
2183 static nodemask_t oldmems
; /* protected by cgroup_mutex */
2188 oldmems
= top_cpuset
.mems_allowed
;
2189 mutex_lock(&callback_mutex
);
2190 top_cpuset
.mems_allowed
= node_states
[N_MEMORY
];
2191 mutex_unlock(&callback_mutex
);
2192 update_tasks_nodemask(&top_cpuset
, &oldmems
, NULL
);
2196 * needn't update top_cpuset.mems_allowed explicitly because
2197 * scan_cpusets_upon_hotplug() will update it.
2199 scan_cpusets_upon_hotplug(&top_cpuset
, CPUSET_MEM_OFFLINE
);
2211 * cpuset_init_smp - initialize cpus_allowed
2213 * Description: Finish top cpuset after cpu, node maps are initialized
2216 void __init
cpuset_init_smp(void)
2218 cpumask_copy(top_cpuset
.cpus_allowed
, cpu_active_mask
);
2219 top_cpuset
.mems_allowed
= node_states
[N_MEMORY
];
2221 hotplug_memory_notifier(cpuset_track_online_nodes
, 10);
2223 cpuset_wq
= create_singlethread_workqueue("cpuset");
2228 * cpuset_cpus_allowed - return cpus_allowed mask from a tasks cpuset.
2229 * @tsk: pointer to task_struct from which to obtain cpuset->cpus_allowed.
2230 * @pmask: pointer to struct cpumask variable to receive cpus_allowed set.
2232 * Description: Returns the cpumask_var_t cpus_allowed of the cpuset
2233 * attached to the specified @tsk. Guaranteed to return some non-empty
2234 * subset of cpu_online_mask, even if this means going outside the
2238 void cpuset_cpus_allowed(struct task_struct
*tsk
, struct cpumask
*pmask
)
2240 mutex_lock(&callback_mutex
);
2242 guarantee_online_cpus(task_cs(tsk
), pmask
);
2244 mutex_unlock(&callback_mutex
);
2247 void cpuset_cpus_allowed_fallback(struct task_struct
*tsk
)
2249 const struct cpuset
*cs
;
2254 do_set_cpus_allowed(tsk
, cs
->cpus_allowed
);
2258 * We own tsk->cpus_allowed, nobody can change it under us.
2260 * But we used cs && cs->cpus_allowed lockless and thus can
2261 * race with cgroup_attach_task() or update_cpumask() and get
2262 * the wrong tsk->cpus_allowed. However, both cases imply the
2263 * subsequent cpuset_change_cpumask()->set_cpus_allowed_ptr()
2264 * which takes task_rq_lock().
2266 * If we are called after it dropped the lock we must see all
2267 * changes in tsk_cs()->cpus_allowed. Otherwise we can temporary
2268 * set any mask even if it is not right from task_cs() pov,
2269 * the pending set_cpus_allowed_ptr() will fix things.
2271 * select_fallback_rq() will fix things ups and set cpu_possible_mask
2276 void cpuset_init_current_mems_allowed(void)
2278 nodes_setall(current
->mems_allowed
);
2282 * cpuset_mems_allowed - return mems_allowed mask from a tasks cpuset.
2283 * @tsk: pointer to task_struct from which to obtain cpuset->mems_allowed.
2285 * Description: Returns the nodemask_t mems_allowed of the cpuset
2286 * attached to the specified @tsk. Guaranteed to return some non-empty
2287 * subset of node_states[N_MEMORY], even if this means going outside the
2291 nodemask_t
cpuset_mems_allowed(struct task_struct
*tsk
)
2295 mutex_lock(&callback_mutex
);
2297 guarantee_online_mems(task_cs(tsk
), &mask
);
2299 mutex_unlock(&callback_mutex
);
2305 * cpuset_nodemask_valid_mems_allowed - check nodemask vs. curremt mems_allowed
2306 * @nodemask: the nodemask to be checked
2308 * Are any of the nodes in the nodemask allowed in current->mems_allowed?
2310 int cpuset_nodemask_valid_mems_allowed(nodemask_t
*nodemask
)
2312 return nodes_intersects(*nodemask
, current
->mems_allowed
);
2316 * nearest_hardwall_ancestor() - Returns the nearest mem_exclusive or
2317 * mem_hardwall ancestor to the specified cpuset. Call holding
2318 * callback_mutex. If no ancestor is mem_exclusive or mem_hardwall
2319 * (an unusual configuration), then returns the root cpuset.
2321 static const struct cpuset
*nearest_hardwall_ancestor(const struct cpuset
*cs
)
2323 while (!(is_mem_exclusive(cs
) || is_mem_hardwall(cs
)) && cs
->parent
)
2329 * cpuset_node_allowed_softwall - Can we allocate on a memory node?
2330 * @node: is this an allowed node?
2331 * @gfp_mask: memory allocation flags
2333 * If we're in interrupt, yes, we can always allocate. If __GFP_THISNODE is
2334 * set, yes, we can always allocate. If node is in our task's mems_allowed,
2335 * yes. If it's not a __GFP_HARDWALL request and this node is in the nearest
2336 * hardwalled cpuset ancestor to this task's cpuset, yes. If the task has been
2337 * OOM killed and has access to memory reserves as specified by the TIF_MEMDIE
2341 * If __GFP_HARDWALL is set, cpuset_node_allowed_softwall() reduces to
2342 * cpuset_node_allowed_hardwall(). Otherwise, cpuset_node_allowed_softwall()
2343 * might sleep, and might allow a node from an enclosing cpuset.
2345 * cpuset_node_allowed_hardwall() only handles the simpler case of hardwall
2346 * cpusets, and never sleeps.
2348 * The __GFP_THISNODE placement logic is really handled elsewhere,
2349 * by forcibly using a zonelist starting at a specified node, and by
2350 * (in get_page_from_freelist()) refusing to consider the zones for
2351 * any node on the zonelist except the first. By the time any such
2352 * calls get to this routine, we should just shut up and say 'yes'.
2354 * GFP_USER allocations are marked with the __GFP_HARDWALL bit,
2355 * and do not allow allocations outside the current tasks cpuset
2356 * unless the task has been OOM killed as is marked TIF_MEMDIE.
2357 * GFP_KERNEL allocations are not so marked, so can escape to the
2358 * nearest enclosing hardwalled ancestor cpuset.
2360 * Scanning up parent cpusets requires callback_mutex. The
2361 * __alloc_pages() routine only calls here with __GFP_HARDWALL bit
2362 * _not_ set if it's a GFP_KERNEL allocation, and all nodes in the
2363 * current tasks mems_allowed came up empty on the first pass over
2364 * the zonelist. So only GFP_KERNEL allocations, if all nodes in the
2365 * cpuset are short of memory, might require taking the callback_mutex
2368 * The first call here from mm/page_alloc:get_page_from_freelist()
2369 * has __GFP_HARDWALL set in gfp_mask, enforcing hardwall cpusets,
2370 * so no allocation on a node outside the cpuset is allowed (unless
2371 * in interrupt, of course).
2373 * The second pass through get_page_from_freelist() doesn't even call
2374 * here for GFP_ATOMIC calls. For those calls, the __alloc_pages()
2375 * variable 'wait' is not set, and the bit ALLOC_CPUSET is not set
2376 * in alloc_flags. That logic and the checks below have the combined
2378 * in_interrupt - any node ok (current task context irrelevant)
2379 * GFP_ATOMIC - any node ok
2380 * TIF_MEMDIE - any node ok
2381 * GFP_KERNEL - any node in enclosing hardwalled cpuset ok
2382 * GFP_USER - only nodes in current tasks mems allowed ok.
2385 * Don't call cpuset_node_allowed_softwall if you can't sleep, unless you
2386 * pass in the __GFP_HARDWALL flag set in gfp_flag, which disables
2387 * the code that might scan up ancestor cpusets and sleep.
2389 int __cpuset_node_allowed_softwall(int node
, gfp_t gfp_mask
)
2391 const struct cpuset
*cs
; /* current cpuset ancestors */
2392 int allowed
; /* is allocation in zone z allowed? */
2394 if (in_interrupt() || (gfp_mask
& __GFP_THISNODE
))
2396 might_sleep_if(!(gfp_mask
& __GFP_HARDWALL
));
2397 if (node_isset(node
, current
->mems_allowed
))
2400 * Allow tasks that have access to memory reserves because they have
2401 * been OOM killed to get memory anywhere.
2403 if (unlikely(test_thread_flag(TIF_MEMDIE
)))
2405 if (gfp_mask
& __GFP_HARDWALL
) /* If hardwall request, stop here */
2408 if (current
->flags
& PF_EXITING
) /* Let dying task have memory */
2411 /* Not hardwall and node outside mems_allowed: scan up cpusets */
2412 mutex_lock(&callback_mutex
);
2415 cs
= nearest_hardwall_ancestor(task_cs(current
));
2416 task_unlock(current
);
2418 allowed
= node_isset(node
, cs
->mems_allowed
);
2419 mutex_unlock(&callback_mutex
);
2424 * cpuset_node_allowed_hardwall - Can we allocate on a memory node?
2425 * @node: is this an allowed node?
2426 * @gfp_mask: memory allocation flags
2428 * If we're in interrupt, yes, we can always allocate. If __GFP_THISNODE is
2429 * set, yes, we can always allocate. If node is in our task's mems_allowed,
2430 * yes. If the task has been OOM killed and has access to memory reserves as
2431 * specified by the TIF_MEMDIE flag, yes.
2434 * The __GFP_THISNODE placement logic is really handled elsewhere,
2435 * by forcibly using a zonelist starting at a specified node, and by
2436 * (in get_page_from_freelist()) refusing to consider the zones for
2437 * any node on the zonelist except the first. By the time any such
2438 * calls get to this routine, we should just shut up and say 'yes'.
2440 * Unlike the cpuset_node_allowed_softwall() variant, above,
2441 * this variant requires that the node be in the current task's
2442 * mems_allowed or that we're in interrupt. It does not scan up the
2443 * cpuset hierarchy for the nearest enclosing mem_exclusive cpuset.
2446 int __cpuset_node_allowed_hardwall(int node
, gfp_t gfp_mask
)
2448 if (in_interrupt() || (gfp_mask
& __GFP_THISNODE
))
2450 if (node_isset(node
, current
->mems_allowed
))
2453 * Allow tasks that have access to memory reserves because they have
2454 * been OOM killed to get memory anywhere.
2456 if (unlikely(test_thread_flag(TIF_MEMDIE
)))
2462 * cpuset_mem_spread_node() - On which node to begin search for a file page
2463 * cpuset_slab_spread_node() - On which node to begin search for a slab page
2465 * If a task is marked PF_SPREAD_PAGE or PF_SPREAD_SLAB (as for
2466 * tasks in a cpuset with is_spread_page or is_spread_slab set),
2467 * and if the memory allocation used cpuset_mem_spread_node()
2468 * to determine on which node to start looking, as it will for
2469 * certain page cache or slab cache pages such as used for file
2470 * system buffers and inode caches, then instead of starting on the
2471 * local node to look for a free page, rather spread the starting
2472 * node around the tasks mems_allowed nodes.
2474 * We don't have to worry about the returned node being offline
2475 * because "it can't happen", and even if it did, it would be ok.
2477 * The routines calling guarantee_online_mems() are careful to
2478 * only set nodes in task->mems_allowed that are online. So it
2479 * should not be possible for the following code to return an
2480 * offline node. But if it did, that would be ok, as this routine
2481 * is not returning the node where the allocation must be, only
2482 * the node where the search should start. The zonelist passed to
2483 * __alloc_pages() will include all nodes. If the slab allocator
2484 * is passed an offline node, it will fall back to the local node.
2485 * See kmem_cache_alloc_node().
2488 static int cpuset_spread_node(int *rotor
)
2492 node
= next_node(*rotor
, current
->mems_allowed
);
2493 if (node
== MAX_NUMNODES
)
2494 node
= first_node(current
->mems_allowed
);
2499 int cpuset_mem_spread_node(void)
2501 if (current
->cpuset_mem_spread_rotor
== NUMA_NO_NODE
)
2502 current
->cpuset_mem_spread_rotor
=
2503 node_random(¤t
->mems_allowed
);
2505 return cpuset_spread_node(¤t
->cpuset_mem_spread_rotor
);
2508 int cpuset_slab_spread_node(void)
2510 if (current
->cpuset_slab_spread_rotor
== NUMA_NO_NODE
)
2511 current
->cpuset_slab_spread_rotor
=
2512 node_random(¤t
->mems_allowed
);
2514 return cpuset_spread_node(¤t
->cpuset_slab_spread_rotor
);
2517 EXPORT_SYMBOL_GPL(cpuset_mem_spread_node
);
2520 * cpuset_mems_allowed_intersects - Does @tsk1's mems_allowed intersect @tsk2's?
2521 * @tsk1: pointer to task_struct of some task.
2522 * @tsk2: pointer to task_struct of some other task.
2524 * Description: Return true if @tsk1's mems_allowed intersects the
2525 * mems_allowed of @tsk2. Used by the OOM killer to determine if
2526 * one of the task's memory usage might impact the memory available
2530 int cpuset_mems_allowed_intersects(const struct task_struct
*tsk1
,
2531 const struct task_struct
*tsk2
)
2533 return nodes_intersects(tsk1
->mems_allowed
, tsk2
->mems_allowed
);
2537 * cpuset_print_task_mems_allowed - prints task's cpuset and mems_allowed
2538 * @task: pointer to task_struct of some task.
2540 * Description: Prints @task's name, cpuset name, and cached copy of its
2541 * mems_allowed to the kernel log. Must hold task_lock(task) to allow
2542 * dereferencing task_cs(task).
2544 void cpuset_print_task_mems_allowed(struct task_struct
*tsk
)
2546 struct dentry
*dentry
;
2548 dentry
= task_cs(tsk
)->css
.cgroup
->dentry
;
2549 spin_lock(&cpuset_buffer_lock
);
2550 snprintf(cpuset_name
, CPUSET_NAME_LEN
,
2551 dentry
? (const char *)dentry
->d_name
.name
: "/");
2552 nodelist_scnprintf(cpuset_nodelist
, CPUSET_NODELIST_LEN
,
2554 printk(KERN_INFO
"%s cpuset=%s mems_allowed=%s\n",
2555 tsk
->comm
, cpuset_name
, cpuset_nodelist
);
2556 spin_unlock(&cpuset_buffer_lock
);
2560 * Collection of memory_pressure is suppressed unless
2561 * this flag is enabled by writing "1" to the special
2562 * cpuset file 'memory_pressure_enabled' in the root cpuset.
2565 int cpuset_memory_pressure_enabled __read_mostly
;
2568 * cpuset_memory_pressure_bump - keep stats of per-cpuset reclaims.
2570 * Keep a running average of the rate of synchronous (direct)
2571 * page reclaim efforts initiated by tasks in each cpuset.
2573 * This represents the rate at which some task in the cpuset
2574 * ran low on memory on all nodes it was allowed to use, and
2575 * had to enter the kernels page reclaim code in an effort to
2576 * create more free memory by tossing clean pages or swapping
2577 * or writing dirty pages.
2579 * Display to user space in the per-cpuset read-only file
2580 * "memory_pressure". Value displayed is an integer
2581 * representing the recent rate of entry into the synchronous
2582 * (direct) page reclaim by any task attached to the cpuset.
2585 void __cpuset_memory_pressure_bump(void)
2588 fmeter_markevent(&task_cs(current
)->fmeter
);
2589 task_unlock(current
);
2592 #ifdef CONFIG_PROC_PID_CPUSET
2594 * proc_cpuset_show()
2595 * - Print tasks cpuset path into seq_file.
2596 * - Used for /proc/<pid>/cpuset.
2597 * - No need to task_lock(tsk) on this tsk->cpuset reference, as it
2598 * doesn't really matter if tsk->cpuset changes after we read it,
2599 * and we take cgroup_mutex, keeping cpuset_attach() from changing it
2602 static int proc_cpuset_show(struct seq_file
*m
, void *unused_v
)
2605 struct task_struct
*tsk
;
2607 struct cgroup_subsys_state
*css
;
2611 buf
= kmalloc(PAGE_SIZE
, GFP_KERNEL
);
2617 tsk
= get_pid_task(pid
, PIDTYPE_PID
);
2623 css
= task_subsys_state(tsk
, cpuset_subsys_id
);
2624 retval
= cgroup_path(css
->cgroup
, buf
, PAGE_SIZE
);
2631 put_task_struct(tsk
);
2638 static int cpuset_open(struct inode
*inode
, struct file
*file
)
2640 struct pid
*pid
= PROC_I(inode
)->pid
;
2641 return single_open(file
, proc_cpuset_show
, pid
);
2644 const struct file_operations proc_cpuset_operations
= {
2645 .open
= cpuset_open
,
2647 .llseek
= seq_lseek
,
2648 .release
= single_release
,
2650 #endif /* CONFIG_PROC_PID_CPUSET */
2652 /* Display task mems_allowed in /proc/<pid>/status file. */
2653 void cpuset_task_status_allowed(struct seq_file
*m
, struct task_struct
*task
)
2655 seq_printf(m
, "Mems_allowed:\t");
2656 seq_nodemask(m
, &task
->mems_allowed
);
2657 seq_printf(m
, "\n");
2658 seq_printf(m
, "Mems_allowed_list:\t");
2659 seq_nodemask_list(m
, &task
->mems_allowed
);
2660 seq_printf(m
, "\n");