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 * Tracks how many cpusets are currently defined in system.
65 * When there is only one cpuset (the root cpuset) we can
66 * short circuit some hooks.
68 int number_of_cpusets __read_mostly
;
70 /* Forward declare cgroup structures */
71 struct cgroup_subsys cpuset_subsys
;
74 /* See "Frequency meter" comments, below. */
77 int cnt
; /* unprocessed events count */
78 int val
; /* most recent output value */
79 time_t time
; /* clock (secs) when val computed */
80 spinlock_t lock
; /* guards read or write of above */
84 struct cgroup_subsys_state css
;
86 unsigned long flags
; /* "unsigned long" so bitops work */
87 cpumask_var_t cpus_allowed
; /* CPUs allowed to tasks in cpuset */
88 nodemask_t mems_allowed
; /* Memory Nodes allowed to tasks */
90 struct fmeter fmeter
; /* memory_pressure filter */
93 * Tasks are being attached to this cpuset. Used to prevent
94 * zeroing cpus/mems_allowed between ->can_attach() and ->attach().
96 int attach_in_progress
;
98 /* partition number for rebuild_sched_domains() */
101 /* for custom sched domain */
102 int relax_domain_level
;
104 struct work_struct hotplug_work
;
107 /* Retrieve the cpuset for a cgroup */
108 static inline struct cpuset
*cgroup_cs(struct cgroup
*cont
)
110 return container_of(cgroup_subsys_state(cont
, cpuset_subsys_id
),
114 /* Retrieve the cpuset for a task */
115 static inline struct cpuset
*task_cs(struct task_struct
*task
)
117 return container_of(task_subsys_state(task
, cpuset_subsys_id
),
121 static inline struct cpuset
*parent_cs(const struct cpuset
*cs
)
123 struct cgroup
*pcgrp
= cs
->css
.cgroup
->parent
;
126 return cgroup_cs(pcgrp
);
131 static inline bool task_has_mempolicy(struct task_struct
*task
)
133 return task
->mempolicy
;
136 static inline bool task_has_mempolicy(struct task_struct
*task
)
143 /* bits in struct cpuset flags field */
150 CS_SCHED_LOAD_BALANCE
,
155 /* convenient tests for these bits */
156 static inline bool is_cpuset_online(const struct cpuset
*cs
)
158 return test_bit(CS_ONLINE
, &cs
->flags
);
161 static inline int is_cpu_exclusive(const struct cpuset
*cs
)
163 return test_bit(CS_CPU_EXCLUSIVE
, &cs
->flags
);
166 static inline int is_mem_exclusive(const struct cpuset
*cs
)
168 return test_bit(CS_MEM_EXCLUSIVE
, &cs
->flags
);
171 static inline int is_mem_hardwall(const struct cpuset
*cs
)
173 return test_bit(CS_MEM_HARDWALL
, &cs
->flags
);
176 static inline int is_sched_load_balance(const struct cpuset
*cs
)
178 return test_bit(CS_SCHED_LOAD_BALANCE
, &cs
->flags
);
181 static inline int is_memory_migrate(const struct cpuset
*cs
)
183 return test_bit(CS_MEMORY_MIGRATE
, &cs
->flags
);
186 static inline int is_spread_page(const struct cpuset
*cs
)
188 return test_bit(CS_SPREAD_PAGE
, &cs
->flags
);
191 static inline int is_spread_slab(const struct cpuset
*cs
)
193 return test_bit(CS_SPREAD_SLAB
, &cs
->flags
);
196 static struct cpuset top_cpuset
= {
197 .flags
= ((1 << CS_ONLINE
) | (1 << CS_CPU_EXCLUSIVE
) |
198 (1 << CS_MEM_EXCLUSIVE
)),
202 * cpuset_for_each_child - traverse online children of a cpuset
203 * @child_cs: loop cursor pointing to the current child
204 * @pos_cgrp: used for iteration
205 * @parent_cs: target cpuset to walk children of
207 * Walk @child_cs through the online children of @parent_cs. Must be used
208 * with RCU read locked.
210 #define cpuset_for_each_child(child_cs, pos_cgrp, parent_cs) \
211 cgroup_for_each_child((pos_cgrp), (parent_cs)->css.cgroup) \
212 if (is_cpuset_online(((child_cs) = cgroup_cs((pos_cgrp)))))
215 * cpuset_for_each_descendant_pre - pre-order walk of a cpuset's descendants
216 * @des_cs: loop cursor pointing to the current descendant
217 * @pos_cgrp: used for iteration
218 * @root_cs: target cpuset to walk ancestor of
220 * Walk @des_cs through the online descendants of @root_cs. Must be used
221 * with RCU read locked. The caller may modify @pos_cgrp by calling
222 * cgroup_rightmost_descendant() to skip subtree.
224 #define cpuset_for_each_descendant_pre(des_cs, pos_cgrp, root_cs) \
225 cgroup_for_each_descendant_pre((pos_cgrp), (root_cs)->css.cgroup) \
226 if (is_cpuset_online(((des_cs) = cgroup_cs((pos_cgrp)))))
229 * There are two global mutexes guarding cpuset structures - cpuset_mutex
230 * and callback_mutex. The latter may nest inside the former. We also
231 * require taking task_lock() when dereferencing a task's cpuset pointer.
232 * See "The task_lock() exception", at the end of this comment.
234 * A task must hold both mutexes to modify cpusets. If a task holds
235 * cpuset_mutex, then it blocks others wanting that mutex, ensuring that it
236 * is the only task able to also acquire callback_mutex and be able to
237 * modify cpusets. It can perform various checks on the cpuset structure
238 * first, knowing nothing will change. It can also allocate memory while
239 * just holding cpuset_mutex. While it is performing these checks, various
240 * callback routines can briefly acquire callback_mutex to query cpusets.
241 * Once it is ready to make the changes, it takes callback_mutex, blocking
244 * Calls to the kernel memory allocator can not be made while holding
245 * callback_mutex, as that would risk double tripping on callback_mutex
246 * from one of the callbacks into the cpuset code from within
249 * If a task is only holding callback_mutex, then it has read-only
252 * Now, the task_struct fields mems_allowed and mempolicy may be changed
253 * by other task, we use alloc_lock in the task_struct fields to protect
256 * The cpuset_common_file_read() handlers only hold callback_mutex across
257 * small pieces of code, such as when reading out possibly multi-word
258 * cpumasks and nodemasks.
260 * Accessing a task's cpuset should be done in accordance with the
261 * guidelines for accessing subsystem state in kernel/cgroup.c
264 static DEFINE_MUTEX(cpuset_mutex
);
265 static DEFINE_MUTEX(callback_mutex
);
268 * CPU / memory hotplug is handled asynchronously.
270 static struct workqueue_struct
*cpuset_propagate_hotplug_wq
;
272 static void cpuset_hotplug_workfn(struct work_struct
*work
);
273 static void cpuset_propagate_hotplug_workfn(struct work_struct
*work
);
274 static void schedule_cpuset_propagate_hotplug(struct cpuset
*cs
);
276 static DECLARE_WORK(cpuset_hotplug_work
, cpuset_hotplug_workfn
);
279 * This is ugly, but preserves the userspace API for existing cpuset
280 * users. If someone tries to mount the "cpuset" filesystem, we
281 * silently switch it to mount "cgroup" instead
283 static struct dentry
*cpuset_mount(struct file_system_type
*fs_type
,
284 int flags
, const char *unused_dev_name
, void *data
)
286 struct file_system_type
*cgroup_fs
= get_fs_type("cgroup");
287 struct dentry
*ret
= ERR_PTR(-ENODEV
);
291 "release_agent=/sbin/cpuset_release_agent";
292 ret
= cgroup_fs
->mount(cgroup_fs
, flags
,
293 unused_dev_name
, mountopts
);
294 put_filesystem(cgroup_fs
);
299 static struct file_system_type cpuset_fs_type
= {
301 .mount
= cpuset_mount
,
305 * Return in pmask the portion of a cpusets's cpus_allowed that
306 * are online. If none are online, walk up the cpuset hierarchy
307 * until we find one that does have some online cpus. If we get
308 * all the way to the top and still haven't found any online cpus,
309 * return cpu_online_mask. Or if passed a NULL cs from an exit'ing
310 * task, return cpu_online_mask.
312 * One way or another, we guarantee to return some non-empty subset
313 * of cpu_online_mask.
315 * Call with callback_mutex held.
318 static void guarantee_online_cpus(const struct cpuset
*cs
,
319 struct cpumask
*pmask
)
321 while (cs
&& !cpumask_intersects(cs
->cpus_allowed
, cpu_online_mask
))
324 cpumask_and(pmask
, cs
->cpus_allowed
, cpu_online_mask
);
326 cpumask_copy(pmask
, cpu_online_mask
);
327 BUG_ON(!cpumask_intersects(pmask
, cpu_online_mask
));
331 * Return in *pmask the portion of a cpusets's mems_allowed that
332 * are online, with memory. If none are online with memory, walk
333 * up the cpuset hierarchy until we find one that does have some
334 * online mems. If we get all the way to the top and still haven't
335 * found any online mems, return node_states[N_MEMORY].
337 * One way or another, we guarantee to return some non-empty subset
338 * of node_states[N_MEMORY].
340 * Call with callback_mutex held.
343 static void guarantee_online_mems(const struct cpuset
*cs
, nodemask_t
*pmask
)
345 while (cs
&& !nodes_intersects(cs
->mems_allowed
,
346 node_states
[N_MEMORY
]))
349 nodes_and(*pmask
, cs
->mems_allowed
,
350 node_states
[N_MEMORY
]);
352 *pmask
= node_states
[N_MEMORY
];
353 BUG_ON(!nodes_intersects(*pmask
, node_states
[N_MEMORY
]));
357 * update task's spread flag if cpuset's page/slab spread flag is set
359 * Called with callback_mutex/cpuset_mutex held
361 static void cpuset_update_task_spread_flag(struct cpuset
*cs
,
362 struct task_struct
*tsk
)
364 if (is_spread_page(cs
))
365 tsk
->flags
|= PF_SPREAD_PAGE
;
367 tsk
->flags
&= ~PF_SPREAD_PAGE
;
368 if (is_spread_slab(cs
))
369 tsk
->flags
|= PF_SPREAD_SLAB
;
371 tsk
->flags
&= ~PF_SPREAD_SLAB
;
375 * is_cpuset_subset(p, q) - Is cpuset p a subset of cpuset q?
377 * One cpuset is a subset of another if all its allowed CPUs and
378 * Memory Nodes are a subset of the other, and its exclusive flags
379 * are only set if the other's are set. Call holding cpuset_mutex.
382 static int is_cpuset_subset(const struct cpuset
*p
, const struct cpuset
*q
)
384 return cpumask_subset(p
->cpus_allowed
, q
->cpus_allowed
) &&
385 nodes_subset(p
->mems_allowed
, q
->mems_allowed
) &&
386 is_cpu_exclusive(p
) <= is_cpu_exclusive(q
) &&
387 is_mem_exclusive(p
) <= is_mem_exclusive(q
);
391 * alloc_trial_cpuset - allocate a trial cpuset
392 * @cs: the cpuset that the trial cpuset duplicates
394 static struct cpuset
*alloc_trial_cpuset(const struct cpuset
*cs
)
396 struct cpuset
*trial
;
398 trial
= kmemdup(cs
, sizeof(*cs
), GFP_KERNEL
);
402 if (!alloc_cpumask_var(&trial
->cpus_allowed
, GFP_KERNEL
)) {
406 cpumask_copy(trial
->cpus_allowed
, cs
->cpus_allowed
);
412 * free_trial_cpuset - free the trial cpuset
413 * @trial: the trial cpuset to be freed
415 static void free_trial_cpuset(struct cpuset
*trial
)
417 free_cpumask_var(trial
->cpus_allowed
);
422 * validate_change() - Used to validate that any proposed cpuset change
423 * follows the structural rules for cpusets.
425 * If we replaced the flag and mask values of the current cpuset
426 * (cur) with those values in the trial cpuset (trial), would
427 * our various subset and exclusive rules still be valid? Presumes
430 * 'cur' is the address of an actual, in-use cpuset. Operations
431 * such as list traversal that depend on the actual address of the
432 * cpuset in the list must use cur below, not trial.
434 * 'trial' is the address of bulk structure copy of cur, with
435 * perhaps one or more of the fields cpus_allowed, mems_allowed,
436 * or flags changed to new, trial values.
438 * Return 0 if valid, -errno if not.
441 static int validate_change(const struct cpuset
*cur
, const struct cpuset
*trial
)
444 struct cpuset
*c
, *par
;
449 /* Each of our child cpusets must be a subset of us */
451 cpuset_for_each_child(c
, cont
, cur
)
452 if (!is_cpuset_subset(c
, trial
))
455 /* Remaining checks don't apply to root cpuset */
457 if (cur
== &top_cpuset
)
460 par
= parent_cs(cur
);
462 /* We must be a subset of our parent cpuset */
464 if (!is_cpuset_subset(trial
, par
))
468 * If either I or some sibling (!= me) is exclusive, we can't
472 cpuset_for_each_child(c
, cont
, par
) {
473 if ((is_cpu_exclusive(trial
) || is_cpu_exclusive(c
)) &&
475 cpumask_intersects(trial
->cpus_allowed
, c
->cpus_allowed
))
477 if ((is_mem_exclusive(trial
) || is_mem_exclusive(c
)) &&
479 nodes_intersects(trial
->mems_allowed
, c
->mems_allowed
))
484 * Cpusets with tasks - existing or newly being attached - can't
485 * have empty cpus_allowed or mems_allowed.
488 if ((cgroup_task_count(cur
->css
.cgroup
) || cur
->attach_in_progress
) &&
489 (cpumask_empty(trial
->cpus_allowed
) ||
490 nodes_empty(trial
->mems_allowed
)))
501 * Helper routine for generate_sched_domains().
502 * Do cpusets a, b have overlapping cpus_allowed masks?
504 static int cpusets_overlap(struct cpuset
*a
, struct cpuset
*b
)
506 return cpumask_intersects(a
->cpus_allowed
, b
->cpus_allowed
);
510 update_domain_attr(struct sched_domain_attr
*dattr
, struct cpuset
*c
)
512 if (dattr
->relax_domain_level
< c
->relax_domain_level
)
513 dattr
->relax_domain_level
= c
->relax_domain_level
;
517 static void update_domain_attr_tree(struct sched_domain_attr
*dattr
,
518 struct cpuset
*root_cs
)
521 struct cgroup
*pos_cgrp
;
524 cpuset_for_each_descendant_pre(cp
, pos_cgrp
, root_cs
) {
525 /* skip the whole subtree if @cp doesn't have any CPU */
526 if (cpumask_empty(cp
->cpus_allowed
)) {
527 pos_cgrp
= cgroup_rightmost_descendant(pos_cgrp
);
531 if (is_sched_load_balance(cp
))
532 update_domain_attr(dattr
, cp
);
538 * generate_sched_domains()
540 * This function builds a partial partition of the systems CPUs
541 * A 'partial partition' is a set of non-overlapping subsets whose
542 * union is a subset of that set.
543 * The output of this function needs to be passed to kernel/sched.c
544 * partition_sched_domains() routine, which will rebuild the scheduler's
545 * load balancing domains (sched domains) as specified by that partial
548 * See "What is sched_load_balance" in Documentation/cgroups/cpusets.txt
549 * for a background explanation of this.
551 * Does not return errors, on the theory that the callers of this
552 * routine would rather not worry about failures to rebuild sched
553 * domains when operating in the severe memory shortage situations
554 * that could cause allocation failures below.
556 * Must be called with cpuset_mutex held.
558 * The three key local variables below are:
559 * q - a linked-list queue of cpuset pointers, used to implement a
560 * top-down scan of all cpusets. This scan loads a pointer
561 * to each cpuset marked is_sched_load_balance into the
562 * array 'csa'. For our purposes, rebuilding the schedulers
563 * sched domains, we can ignore !is_sched_load_balance cpusets.
564 * csa - (for CpuSet Array) Array of pointers to all the cpusets
565 * that need to be load balanced, for convenient iterative
566 * access by the subsequent code that finds the best partition,
567 * i.e the set of domains (subsets) of CPUs such that the
568 * cpus_allowed of every cpuset marked is_sched_load_balance
569 * is a subset of one of these domains, while there are as
570 * many such domains as possible, each as small as possible.
571 * doms - Conversion of 'csa' to an array of cpumasks, for passing to
572 * the kernel/sched.c routine partition_sched_domains() in a
573 * convenient format, that can be easily compared to the prior
574 * value to determine what partition elements (sched domains)
575 * were changed (added or removed.)
577 * Finding the best partition (set of domains):
578 * The triple nested loops below over i, j, k scan over the
579 * load balanced cpusets (using the array of cpuset pointers in
580 * csa[]) looking for pairs of cpusets that have overlapping
581 * cpus_allowed, but which don't have the same 'pn' partition
582 * number and gives them in the same partition number. It keeps
583 * looping on the 'restart' label until it can no longer find
586 * The union of the cpus_allowed masks from the set of
587 * all cpusets having the same 'pn' value then form the one
588 * element of the partition (one sched domain) to be passed to
589 * partition_sched_domains().
591 static int generate_sched_domains(cpumask_var_t
**domains
,
592 struct sched_domain_attr
**attributes
)
594 struct cpuset
*cp
; /* scans q */
595 struct cpuset
**csa
; /* array of all cpuset ptrs */
596 int csn
; /* how many cpuset ptrs in csa so far */
597 int i
, j
, k
; /* indices for partition finding loops */
598 cpumask_var_t
*doms
; /* resulting partition; i.e. sched domains */
599 struct sched_domain_attr
*dattr
; /* attributes for custom domains */
600 int ndoms
= 0; /* number of sched domains in result */
601 int nslot
; /* next empty doms[] struct cpumask slot */
602 struct cgroup
*pos_cgrp
;
608 /* Special case for the 99% of systems with one, full, sched domain */
609 if (is_sched_load_balance(&top_cpuset
)) {
611 doms
= alloc_sched_domains(ndoms
);
615 dattr
= kmalloc(sizeof(struct sched_domain_attr
), GFP_KERNEL
);
617 *dattr
= SD_ATTR_INIT
;
618 update_domain_attr_tree(dattr
, &top_cpuset
);
620 cpumask_copy(doms
[0], top_cpuset
.cpus_allowed
);
625 csa
= kmalloc(number_of_cpusets
* sizeof(cp
), GFP_KERNEL
);
631 cpuset_for_each_descendant_pre(cp
, pos_cgrp
, &top_cpuset
) {
633 * Continue traversing beyond @cp iff @cp has some CPUs and
634 * isn't load balancing. The former is obvious. The
635 * latter: All child cpusets contain a subset of the
636 * parent's cpus, so just skip them, and then we call
637 * update_domain_attr_tree() to calc relax_domain_level of
638 * the corresponding sched domain.
640 if (!cpumask_empty(cp
->cpus_allowed
) &&
641 !is_sched_load_balance(cp
))
644 if (is_sched_load_balance(cp
))
647 /* skip @cp's subtree */
648 pos_cgrp
= cgroup_rightmost_descendant(pos_cgrp
);
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 * If the flag 'sched_load_balance' of any cpuset with non-empty
756 * 'cpus' changes, or if the 'cpus' allowed changes in any cpuset
757 * which has that flag enabled, or if any cpuset with a non-empty
758 * 'cpus' is removed, then call this routine to rebuild the
759 * scheduler's dynamic sched domains.
761 * Call with cpuset_mutex held. Takes get_online_cpus().
763 static void rebuild_sched_domains_locked(void)
765 struct sched_domain_attr
*attr
;
769 lockdep_assert_held(&cpuset_mutex
);
773 * We have raced with CPU hotplug. Don't do anything to avoid
774 * passing doms with offlined cpu to partition_sched_domains().
775 * Anyways, hotplug work item will rebuild sched domains.
777 if (!cpumask_equal(top_cpuset
.cpus_allowed
, cpu_active_mask
))
780 /* Generate domain masks and attrs */
781 ndoms
= generate_sched_domains(&doms
, &attr
);
783 /* Have scheduler rebuild the domains */
784 partition_sched_domains(ndoms
, doms
, attr
);
788 #else /* !CONFIG_SMP */
789 static void rebuild_sched_domains_locked(void)
793 static int generate_sched_domains(cpumask_var_t
**domains
,
794 struct sched_domain_attr
**attributes
)
799 #endif /* CONFIG_SMP */
801 void rebuild_sched_domains(void)
803 mutex_lock(&cpuset_mutex
);
804 rebuild_sched_domains_locked();
805 mutex_unlock(&cpuset_mutex
);
809 * cpuset_test_cpumask - test a task's cpus_allowed versus its cpuset's
811 * @scan: struct cgroup_scanner contained in its struct cpuset_hotplug_scanner
813 * Call with cpuset_mutex held. May take callback_mutex during call.
814 * Called for each task in a cgroup by cgroup_scan_tasks().
815 * Return nonzero if this tasks's cpus_allowed mask should be changed (in other
816 * words, if its mask is not equal to its cpuset's mask).
818 static int cpuset_test_cpumask(struct task_struct
*tsk
,
819 struct cgroup_scanner
*scan
)
821 return !cpumask_equal(&tsk
->cpus_allowed
,
822 (cgroup_cs(scan
->cg
))->cpus_allowed
);
826 * cpuset_change_cpumask - make a task's cpus_allowed the same as its cpuset's
828 * @scan: struct cgroup_scanner containing the cgroup of the task
830 * Called by cgroup_scan_tasks() for each task in a cgroup whose
831 * cpus_allowed mask needs to be changed.
833 * We don't need to re-check for the cgroup/cpuset membership, since we're
834 * holding cpuset_mutex at this point.
836 static void cpuset_change_cpumask(struct task_struct
*tsk
,
837 struct cgroup_scanner
*scan
)
839 set_cpus_allowed_ptr(tsk
, ((cgroup_cs(scan
->cg
))->cpus_allowed
));
843 * update_tasks_cpumask - Update the cpumasks of tasks in the cpuset.
844 * @cs: the cpuset in which each task's cpus_allowed mask needs to be changed
845 * @heap: if NULL, defer allocating heap memory to cgroup_scan_tasks()
847 * Called with cpuset_mutex held
849 * The cgroup_scan_tasks() function will scan all the tasks in a cgroup,
850 * calling callback functions for each.
852 * No return value. It's guaranteed that cgroup_scan_tasks() always returns 0
855 static void update_tasks_cpumask(struct cpuset
*cs
, struct ptr_heap
*heap
)
857 struct cgroup_scanner scan
;
859 scan
.cg
= cs
->css
.cgroup
;
860 scan
.test_task
= cpuset_test_cpumask
;
861 scan
.process_task
= cpuset_change_cpumask
;
863 cgroup_scan_tasks(&scan
);
867 * update_cpumask - update the cpus_allowed mask of a cpuset and all tasks in it
868 * @cs: the cpuset to consider
869 * @buf: buffer of cpu numbers written to this cpuset
871 static int update_cpumask(struct cpuset
*cs
, struct cpuset
*trialcs
,
874 struct ptr_heap heap
;
876 int is_load_balanced
;
878 /* top_cpuset.cpus_allowed tracks cpu_online_mask; it's read-only */
879 if (cs
== &top_cpuset
)
883 * An empty cpus_allowed is ok only if the cpuset has no tasks.
884 * Since cpulist_parse() fails on an empty mask, we special case
885 * that parsing. The validate_change() call ensures that cpusets
886 * with tasks have cpus.
889 cpumask_clear(trialcs
->cpus_allowed
);
891 retval
= cpulist_parse(buf
, trialcs
->cpus_allowed
);
895 if (!cpumask_subset(trialcs
->cpus_allowed
, cpu_active_mask
))
898 retval
= validate_change(cs
, trialcs
);
902 /* Nothing to do if the cpus didn't change */
903 if (cpumask_equal(cs
->cpus_allowed
, trialcs
->cpus_allowed
))
906 retval
= heap_init(&heap
, PAGE_SIZE
, GFP_KERNEL
, NULL
);
910 is_load_balanced
= is_sched_load_balance(trialcs
);
912 mutex_lock(&callback_mutex
);
913 cpumask_copy(cs
->cpus_allowed
, trialcs
->cpus_allowed
);
914 mutex_unlock(&callback_mutex
);
917 * Scan tasks in the cpuset, and update the cpumasks of any
918 * that need an update.
920 update_tasks_cpumask(cs
, &heap
);
924 if (is_load_balanced
)
925 rebuild_sched_domains_locked();
932 * Migrate memory region from one set of nodes to another.
934 * Temporarilly set tasks mems_allowed to target nodes of migration,
935 * so that the migration code can allocate pages on these nodes.
937 * Call holding cpuset_mutex, so current's cpuset won't change
938 * during this call, as manage_mutex holds off any cpuset_attach()
939 * calls. Therefore we don't need to take task_lock around the
940 * call to guarantee_online_mems(), as we know no one is changing
943 * While the mm_struct we are migrating is typically from some
944 * other task, the task_struct mems_allowed that we are hacking
945 * is for our current task, which must allocate new pages for that
946 * migrating memory region.
949 static void cpuset_migrate_mm(struct mm_struct
*mm
, const nodemask_t
*from
,
950 const nodemask_t
*to
)
952 struct task_struct
*tsk
= current
;
954 tsk
->mems_allowed
= *to
;
956 do_migrate_pages(mm
, from
, to
, MPOL_MF_MOVE_ALL
);
958 guarantee_online_mems(task_cs(tsk
),&tsk
->mems_allowed
);
962 * cpuset_change_task_nodemask - change task's mems_allowed and mempolicy
963 * @tsk: the task to change
964 * @newmems: new nodes that the task will be set
966 * In order to avoid seeing no nodes if the old and new nodes are disjoint,
967 * we structure updates as setting all new allowed nodes, then clearing newly
970 static void cpuset_change_task_nodemask(struct task_struct
*tsk
,
976 * Allow tasks that have access to memory reserves because they have
977 * been OOM killed to get memory anywhere.
979 if (unlikely(test_thread_flag(TIF_MEMDIE
)))
981 if (current
->flags
& PF_EXITING
) /* Let dying task have memory */
986 * Determine if a loop is necessary if another thread is doing
987 * get_mems_allowed(). If at least one node remains unchanged and
988 * tsk does not have a mempolicy, then an empty nodemask will not be
989 * possible when mems_allowed is larger than a word.
991 need_loop
= task_has_mempolicy(tsk
) ||
992 !nodes_intersects(*newmems
, tsk
->mems_allowed
);
995 write_seqcount_begin(&tsk
->mems_allowed_seq
);
997 nodes_or(tsk
->mems_allowed
, tsk
->mems_allowed
, *newmems
);
998 mpol_rebind_task(tsk
, newmems
, MPOL_REBIND_STEP1
);
1000 mpol_rebind_task(tsk
, newmems
, MPOL_REBIND_STEP2
);
1001 tsk
->mems_allowed
= *newmems
;
1004 write_seqcount_end(&tsk
->mems_allowed_seq
);
1010 * Update task's mems_allowed and rebind its mempolicy and vmas' mempolicy
1011 * of it to cpuset's new mems_allowed, and migrate pages to new nodes if
1012 * memory_migrate flag is set. Called with cpuset_mutex held.
1014 static void cpuset_change_nodemask(struct task_struct
*p
,
1015 struct cgroup_scanner
*scan
)
1017 struct mm_struct
*mm
;
1020 const nodemask_t
*oldmem
= scan
->data
;
1021 static nodemask_t newmems
; /* protected by cpuset_mutex */
1023 cs
= cgroup_cs(scan
->cg
);
1024 guarantee_online_mems(cs
, &newmems
);
1026 cpuset_change_task_nodemask(p
, &newmems
);
1028 mm
= get_task_mm(p
);
1032 migrate
= is_memory_migrate(cs
);
1034 mpol_rebind_mm(mm
, &cs
->mems_allowed
);
1036 cpuset_migrate_mm(mm
, oldmem
, &cs
->mems_allowed
);
1040 static void *cpuset_being_rebound
;
1043 * update_tasks_nodemask - Update the nodemasks of tasks in the cpuset.
1044 * @cs: the cpuset in which each task's mems_allowed mask needs to be changed
1045 * @oldmem: old mems_allowed of cpuset cs
1046 * @heap: if NULL, defer allocating heap memory to cgroup_scan_tasks()
1048 * Called with cpuset_mutex held
1049 * No return value. It's guaranteed that cgroup_scan_tasks() always returns 0
1052 static void update_tasks_nodemask(struct cpuset
*cs
, const nodemask_t
*oldmem
,
1053 struct ptr_heap
*heap
)
1055 struct cgroup_scanner scan
;
1057 cpuset_being_rebound
= cs
; /* causes mpol_dup() rebind */
1059 scan
.cg
= cs
->css
.cgroup
;
1060 scan
.test_task
= NULL
;
1061 scan
.process_task
= cpuset_change_nodemask
;
1063 scan
.data
= (nodemask_t
*)oldmem
;
1066 * The mpol_rebind_mm() call takes mmap_sem, which we couldn't
1067 * take while holding tasklist_lock. Forks can happen - the
1068 * mpol_dup() cpuset_being_rebound check will catch such forks,
1069 * and rebind their vma mempolicies too. Because we still hold
1070 * the global cpuset_mutex, we know that no other rebind effort
1071 * will be contending for the global variable cpuset_being_rebound.
1072 * It's ok if we rebind the same mm twice; mpol_rebind_mm()
1073 * is idempotent. Also migrate pages in each mm to new nodes.
1075 cgroup_scan_tasks(&scan
);
1077 /* We're done rebinding vmas to this cpuset's new mems_allowed. */
1078 cpuset_being_rebound
= NULL
;
1082 * Handle user request to change the 'mems' memory placement
1083 * of a cpuset. Needs to validate the request, update the
1084 * cpusets mems_allowed, and for each task in the cpuset,
1085 * update mems_allowed and rebind task's mempolicy and any vma
1086 * mempolicies and if the cpuset is marked 'memory_migrate',
1087 * migrate the tasks pages to the new memory.
1089 * Call with cpuset_mutex held. May take callback_mutex during call.
1090 * Will take tasklist_lock, scan tasklist for tasks in cpuset cs,
1091 * lock each such tasks mm->mmap_sem, scan its vma's and rebind
1092 * their mempolicies to the cpusets new mems_allowed.
1094 static int update_nodemask(struct cpuset
*cs
, struct cpuset
*trialcs
,
1097 NODEMASK_ALLOC(nodemask_t
, oldmem
, GFP_KERNEL
);
1099 struct ptr_heap heap
;
1105 * top_cpuset.mems_allowed tracks node_stats[N_MEMORY];
1108 if (cs
== &top_cpuset
) {
1114 * An empty mems_allowed is ok iff there are no tasks in the cpuset.
1115 * Since nodelist_parse() fails on an empty mask, we special case
1116 * that parsing. The validate_change() call ensures that cpusets
1117 * with tasks have memory.
1120 nodes_clear(trialcs
->mems_allowed
);
1122 retval
= nodelist_parse(buf
, trialcs
->mems_allowed
);
1126 if (!nodes_subset(trialcs
->mems_allowed
,
1127 node_states
[N_MEMORY
])) {
1132 *oldmem
= cs
->mems_allowed
;
1133 if (nodes_equal(*oldmem
, trialcs
->mems_allowed
)) {
1134 retval
= 0; /* Too easy - nothing to do */
1137 retval
= validate_change(cs
, trialcs
);
1141 retval
= heap_init(&heap
, PAGE_SIZE
, GFP_KERNEL
, NULL
);
1145 mutex_lock(&callback_mutex
);
1146 cs
->mems_allowed
= trialcs
->mems_allowed
;
1147 mutex_unlock(&callback_mutex
);
1149 update_tasks_nodemask(cs
, oldmem
, &heap
);
1153 NODEMASK_FREE(oldmem
);
1157 int current_cpuset_is_being_rebound(void)
1159 return task_cs(current
) == cpuset_being_rebound
;
1162 static int update_relax_domain_level(struct cpuset
*cs
, s64 val
)
1165 if (val
< -1 || val
>= sched_domain_level_max
)
1169 if (val
!= cs
->relax_domain_level
) {
1170 cs
->relax_domain_level
= val
;
1171 if (!cpumask_empty(cs
->cpus_allowed
) &&
1172 is_sched_load_balance(cs
))
1173 rebuild_sched_domains_locked();
1180 * cpuset_change_flag - make a task's spread flags the same as its cpuset's
1181 * @tsk: task to be updated
1182 * @scan: struct cgroup_scanner containing the cgroup of the task
1184 * Called by cgroup_scan_tasks() for each task in a cgroup.
1186 * We don't need to re-check for the cgroup/cpuset membership, since we're
1187 * holding cpuset_mutex at this point.
1189 static void cpuset_change_flag(struct task_struct
*tsk
,
1190 struct cgroup_scanner
*scan
)
1192 cpuset_update_task_spread_flag(cgroup_cs(scan
->cg
), tsk
);
1196 * update_tasks_flags - update the spread flags of tasks in the cpuset.
1197 * @cs: the cpuset in which each task's spread flags needs to be changed
1198 * @heap: if NULL, defer allocating heap memory to cgroup_scan_tasks()
1200 * Called with cpuset_mutex held
1202 * The cgroup_scan_tasks() function will scan all the tasks in a cgroup,
1203 * calling callback functions for each.
1205 * No return value. It's guaranteed that cgroup_scan_tasks() always returns 0
1208 static void update_tasks_flags(struct cpuset
*cs
, struct ptr_heap
*heap
)
1210 struct cgroup_scanner scan
;
1212 scan
.cg
= cs
->css
.cgroup
;
1213 scan
.test_task
= NULL
;
1214 scan
.process_task
= cpuset_change_flag
;
1216 cgroup_scan_tasks(&scan
);
1220 * update_flag - read a 0 or a 1 in a file and update associated flag
1221 * bit: the bit to update (see cpuset_flagbits_t)
1222 * cs: the cpuset to update
1223 * turning_on: whether the flag is being set or cleared
1225 * Call with cpuset_mutex held.
1228 static int update_flag(cpuset_flagbits_t bit
, struct cpuset
*cs
,
1231 struct cpuset
*trialcs
;
1232 int balance_flag_changed
;
1233 int spread_flag_changed
;
1234 struct ptr_heap heap
;
1237 trialcs
= alloc_trial_cpuset(cs
);
1242 set_bit(bit
, &trialcs
->flags
);
1244 clear_bit(bit
, &trialcs
->flags
);
1246 err
= validate_change(cs
, trialcs
);
1250 err
= heap_init(&heap
, PAGE_SIZE
, GFP_KERNEL
, NULL
);
1254 balance_flag_changed
= (is_sched_load_balance(cs
) !=
1255 is_sched_load_balance(trialcs
));
1257 spread_flag_changed
= ((is_spread_slab(cs
) != is_spread_slab(trialcs
))
1258 || (is_spread_page(cs
) != is_spread_page(trialcs
)));
1260 mutex_lock(&callback_mutex
);
1261 cs
->flags
= trialcs
->flags
;
1262 mutex_unlock(&callback_mutex
);
1264 if (!cpumask_empty(trialcs
->cpus_allowed
) && balance_flag_changed
)
1265 rebuild_sched_domains_locked();
1267 if (spread_flag_changed
)
1268 update_tasks_flags(cs
, &heap
);
1271 free_trial_cpuset(trialcs
);
1276 * Frequency meter - How fast is some event occurring?
1278 * These routines manage a digitally filtered, constant time based,
1279 * event frequency meter. There are four routines:
1280 * fmeter_init() - initialize a frequency meter.
1281 * fmeter_markevent() - called each time the event happens.
1282 * fmeter_getrate() - returns the recent rate of such events.
1283 * fmeter_update() - internal routine used to update fmeter.
1285 * A common data structure is passed to each of these routines,
1286 * which is used to keep track of the state required to manage the
1287 * frequency meter and its digital filter.
1289 * The filter works on the number of events marked per unit time.
1290 * The filter is single-pole low-pass recursive (IIR). The time unit
1291 * is 1 second. Arithmetic is done using 32-bit integers scaled to
1292 * simulate 3 decimal digits of precision (multiplied by 1000).
1294 * With an FM_COEF of 933, and a time base of 1 second, the filter
1295 * has a half-life of 10 seconds, meaning that if the events quit
1296 * happening, then the rate returned from the fmeter_getrate()
1297 * will be cut in half each 10 seconds, until it converges to zero.
1299 * It is not worth doing a real infinitely recursive filter. If more
1300 * than FM_MAXTICKS ticks have elapsed since the last filter event,
1301 * just compute FM_MAXTICKS ticks worth, by which point the level
1304 * Limit the count of unprocessed events to FM_MAXCNT, so as to avoid
1305 * arithmetic overflow in the fmeter_update() routine.
1307 * Given the simple 32 bit integer arithmetic used, this meter works
1308 * best for reporting rates between one per millisecond (msec) and
1309 * one per 32 (approx) seconds. At constant rates faster than one
1310 * per msec it maxes out at values just under 1,000,000. At constant
1311 * rates between one per msec, and one per second it will stabilize
1312 * to a value N*1000, where N is the rate of events per second.
1313 * At constant rates between one per second and one per 32 seconds,
1314 * it will be choppy, moving up on the seconds that have an event,
1315 * and then decaying until the next event. At rates slower than
1316 * about one in 32 seconds, it decays all the way back to zero between
1320 #define FM_COEF 933 /* coefficient for half-life of 10 secs */
1321 #define FM_MAXTICKS ((time_t)99) /* useless computing more ticks than this */
1322 #define FM_MAXCNT 1000000 /* limit cnt to avoid overflow */
1323 #define FM_SCALE 1000 /* faux fixed point scale */
1325 /* Initialize a frequency meter */
1326 static void fmeter_init(struct fmeter
*fmp
)
1331 spin_lock_init(&fmp
->lock
);
1334 /* Internal meter update - process cnt events and update value */
1335 static void fmeter_update(struct fmeter
*fmp
)
1337 time_t now
= get_seconds();
1338 time_t ticks
= now
- fmp
->time
;
1343 ticks
= min(FM_MAXTICKS
, ticks
);
1345 fmp
->val
= (FM_COEF
* fmp
->val
) / FM_SCALE
;
1348 fmp
->val
+= ((FM_SCALE
- FM_COEF
) * fmp
->cnt
) / FM_SCALE
;
1352 /* Process any previous ticks, then bump cnt by one (times scale). */
1353 static void fmeter_markevent(struct fmeter
*fmp
)
1355 spin_lock(&fmp
->lock
);
1357 fmp
->cnt
= min(FM_MAXCNT
, fmp
->cnt
+ FM_SCALE
);
1358 spin_unlock(&fmp
->lock
);
1361 /* Process any previous ticks, then return current value. */
1362 static int fmeter_getrate(struct fmeter
*fmp
)
1366 spin_lock(&fmp
->lock
);
1369 spin_unlock(&fmp
->lock
);
1373 /* Called by cgroups to determine if a cpuset is usable; cpuset_mutex held */
1374 static int cpuset_can_attach(struct cgroup
*cgrp
, struct cgroup_taskset
*tset
)
1376 struct cpuset
*cs
= cgroup_cs(cgrp
);
1377 struct task_struct
*task
;
1380 mutex_lock(&cpuset_mutex
);
1383 if (cpumask_empty(cs
->cpus_allowed
) || nodes_empty(cs
->mems_allowed
))
1386 cgroup_taskset_for_each(task
, cgrp
, tset
) {
1388 * Kthreads bound to specific cpus cannot be moved to a new
1389 * cpuset; we cannot change their cpu affinity and
1390 * isolating such threads by their set of allowed nodes is
1391 * unnecessary. Thus, cpusets are not applicable for such
1392 * threads. This prevents checking for success of
1393 * set_cpus_allowed_ptr() on all attached tasks before
1394 * cpus_allowed may be changed.
1397 if (task
->flags
& PF_THREAD_BOUND
)
1399 ret
= security_task_setscheduler(task
);
1405 * Mark attach is in progress. This makes validate_change() fail
1406 * changes which zero cpus/mems_allowed.
1408 cs
->attach_in_progress
++;
1411 mutex_unlock(&cpuset_mutex
);
1415 static void cpuset_cancel_attach(struct cgroup
*cgrp
,
1416 struct cgroup_taskset
*tset
)
1418 mutex_lock(&cpuset_mutex
);
1419 cgroup_cs(cgrp
)->attach_in_progress
--;
1420 mutex_unlock(&cpuset_mutex
);
1424 * Protected by cpuset_mutex. cpus_attach is used only by cpuset_attach()
1425 * but we can't allocate it dynamically there. Define it global and
1426 * allocate from cpuset_init().
1428 static cpumask_var_t cpus_attach
;
1430 static void cpuset_attach(struct cgroup
*cgrp
, struct cgroup_taskset
*tset
)
1432 /* static bufs protected by cpuset_mutex */
1433 static nodemask_t cpuset_attach_nodemask_from
;
1434 static nodemask_t cpuset_attach_nodemask_to
;
1435 struct mm_struct
*mm
;
1436 struct task_struct
*task
;
1437 struct task_struct
*leader
= cgroup_taskset_first(tset
);
1438 struct cgroup
*oldcgrp
= cgroup_taskset_cur_cgroup(tset
);
1439 struct cpuset
*cs
= cgroup_cs(cgrp
);
1440 struct cpuset
*oldcs
= cgroup_cs(oldcgrp
);
1442 mutex_lock(&cpuset_mutex
);
1444 /* prepare for attach */
1445 if (cs
== &top_cpuset
)
1446 cpumask_copy(cpus_attach
, cpu_possible_mask
);
1448 guarantee_online_cpus(cs
, cpus_attach
);
1450 guarantee_online_mems(cs
, &cpuset_attach_nodemask_to
);
1452 cgroup_taskset_for_each(task
, cgrp
, tset
) {
1454 * can_attach beforehand should guarantee that this doesn't
1455 * fail. TODO: have a better way to handle failure here
1457 WARN_ON_ONCE(set_cpus_allowed_ptr(task
, cpus_attach
));
1459 cpuset_change_task_nodemask(task
, &cpuset_attach_nodemask_to
);
1460 cpuset_update_task_spread_flag(cs
, task
);
1464 * Change mm, possibly for multiple threads in a threadgroup. This is
1465 * expensive and may sleep.
1467 cpuset_attach_nodemask_from
= oldcs
->mems_allowed
;
1468 cpuset_attach_nodemask_to
= cs
->mems_allowed
;
1469 mm
= get_task_mm(leader
);
1471 mpol_rebind_mm(mm
, &cpuset_attach_nodemask_to
);
1472 if (is_memory_migrate(cs
))
1473 cpuset_migrate_mm(mm
, &cpuset_attach_nodemask_from
,
1474 &cpuset_attach_nodemask_to
);
1478 cs
->attach_in_progress
--;
1481 * We may have raced with CPU/memory hotunplug. Trigger hotplug
1482 * propagation if @cs doesn't have any CPU or memory. It will move
1483 * the newly added tasks to the nearest parent which can execute.
1485 if (cpumask_empty(cs
->cpus_allowed
) || nodes_empty(cs
->mems_allowed
))
1486 schedule_cpuset_propagate_hotplug(cs
);
1488 mutex_unlock(&cpuset_mutex
);
1491 /* The various types of files and directories in a cpuset file system */
1494 FILE_MEMORY_MIGRATE
,
1500 FILE_SCHED_LOAD_BALANCE
,
1501 FILE_SCHED_RELAX_DOMAIN_LEVEL
,
1502 FILE_MEMORY_PRESSURE_ENABLED
,
1503 FILE_MEMORY_PRESSURE
,
1506 } cpuset_filetype_t
;
1508 static int cpuset_write_u64(struct cgroup
*cgrp
, struct cftype
*cft
, u64 val
)
1510 struct cpuset
*cs
= cgroup_cs(cgrp
);
1511 cpuset_filetype_t type
= cft
->private;
1512 int retval
= -ENODEV
;
1514 mutex_lock(&cpuset_mutex
);
1515 if (!is_cpuset_online(cs
))
1519 case FILE_CPU_EXCLUSIVE
:
1520 retval
= update_flag(CS_CPU_EXCLUSIVE
, cs
, val
);
1522 case FILE_MEM_EXCLUSIVE
:
1523 retval
= update_flag(CS_MEM_EXCLUSIVE
, cs
, val
);
1525 case FILE_MEM_HARDWALL
:
1526 retval
= update_flag(CS_MEM_HARDWALL
, cs
, val
);
1528 case FILE_SCHED_LOAD_BALANCE
:
1529 retval
= update_flag(CS_SCHED_LOAD_BALANCE
, cs
, val
);
1531 case FILE_MEMORY_MIGRATE
:
1532 retval
= update_flag(CS_MEMORY_MIGRATE
, cs
, val
);
1534 case FILE_MEMORY_PRESSURE_ENABLED
:
1535 cpuset_memory_pressure_enabled
= !!val
;
1537 case FILE_MEMORY_PRESSURE
:
1540 case FILE_SPREAD_PAGE
:
1541 retval
= update_flag(CS_SPREAD_PAGE
, cs
, val
);
1543 case FILE_SPREAD_SLAB
:
1544 retval
= update_flag(CS_SPREAD_SLAB
, cs
, val
);
1551 mutex_unlock(&cpuset_mutex
);
1555 static int cpuset_write_s64(struct cgroup
*cgrp
, struct cftype
*cft
, s64 val
)
1557 struct cpuset
*cs
= cgroup_cs(cgrp
);
1558 cpuset_filetype_t type
= cft
->private;
1559 int retval
= -ENODEV
;
1561 mutex_lock(&cpuset_mutex
);
1562 if (!is_cpuset_online(cs
))
1566 case FILE_SCHED_RELAX_DOMAIN_LEVEL
:
1567 retval
= update_relax_domain_level(cs
, val
);
1574 mutex_unlock(&cpuset_mutex
);
1579 * Common handling for a write to a "cpus" or "mems" file.
1581 static int cpuset_write_resmask(struct cgroup
*cgrp
, struct cftype
*cft
,
1584 struct cpuset
*cs
= cgroup_cs(cgrp
);
1585 struct cpuset
*trialcs
;
1586 int retval
= -ENODEV
;
1589 * CPU or memory hotunplug may leave @cs w/o any execution
1590 * resources, in which case the hotplug code asynchronously updates
1591 * configuration and transfers all tasks to the nearest ancestor
1592 * which can execute.
1594 * As writes to "cpus" or "mems" may restore @cs's execution
1595 * resources, wait for the previously scheduled operations before
1596 * proceeding, so that we don't end up keep removing tasks added
1597 * after execution capability is restored.
1599 * Flushing cpuset_hotplug_work is enough to synchronize against
1600 * hotplug hanlding; however, cpuset_attach() may schedule
1601 * propagation work directly. Flush the workqueue too.
1603 flush_work(&cpuset_hotplug_work
);
1604 flush_workqueue(cpuset_propagate_hotplug_wq
);
1606 mutex_lock(&cpuset_mutex
);
1607 if (!is_cpuset_online(cs
))
1610 trialcs
= alloc_trial_cpuset(cs
);
1616 switch (cft
->private) {
1618 retval
= update_cpumask(cs
, trialcs
, buf
);
1621 retval
= update_nodemask(cs
, trialcs
, buf
);
1628 free_trial_cpuset(trialcs
);
1630 mutex_unlock(&cpuset_mutex
);
1635 * These ascii lists should be read in a single call, by using a user
1636 * buffer large enough to hold the entire map. If read in smaller
1637 * chunks, there is no guarantee of atomicity. Since the display format
1638 * used, list of ranges of sequential numbers, is variable length,
1639 * and since these maps can change value dynamically, one could read
1640 * gibberish by doing partial reads while a list was changing.
1641 * A single large read to a buffer that crosses a page boundary is
1642 * ok, because the result being copied to user land is not recomputed
1643 * across a page fault.
1646 static size_t cpuset_sprintf_cpulist(char *page
, struct cpuset
*cs
)
1650 mutex_lock(&callback_mutex
);
1651 count
= cpulist_scnprintf(page
, PAGE_SIZE
, cs
->cpus_allowed
);
1652 mutex_unlock(&callback_mutex
);
1657 static size_t cpuset_sprintf_memlist(char *page
, struct cpuset
*cs
)
1661 mutex_lock(&callback_mutex
);
1662 count
= nodelist_scnprintf(page
, PAGE_SIZE
, cs
->mems_allowed
);
1663 mutex_unlock(&callback_mutex
);
1668 static ssize_t
cpuset_common_file_read(struct cgroup
*cont
,
1672 size_t nbytes
, loff_t
*ppos
)
1674 struct cpuset
*cs
= cgroup_cs(cont
);
1675 cpuset_filetype_t type
= cft
->private;
1680 if (!(page
= (char *)__get_free_page(GFP_TEMPORARY
)))
1687 s
+= cpuset_sprintf_cpulist(s
, cs
);
1690 s
+= cpuset_sprintf_memlist(s
, cs
);
1698 retval
= simple_read_from_buffer(buf
, nbytes
, ppos
, page
, s
- page
);
1700 free_page((unsigned long)page
);
1704 static u64
cpuset_read_u64(struct cgroup
*cont
, struct cftype
*cft
)
1706 struct cpuset
*cs
= cgroup_cs(cont
);
1707 cpuset_filetype_t type
= cft
->private;
1709 case FILE_CPU_EXCLUSIVE
:
1710 return is_cpu_exclusive(cs
);
1711 case FILE_MEM_EXCLUSIVE
:
1712 return is_mem_exclusive(cs
);
1713 case FILE_MEM_HARDWALL
:
1714 return is_mem_hardwall(cs
);
1715 case FILE_SCHED_LOAD_BALANCE
:
1716 return is_sched_load_balance(cs
);
1717 case FILE_MEMORY_MIGRATE
:
1718 return is_memory_migrate(cs
);
1719 case FILE_MEMORY_PRESSURE_ENABLED
:
1720 return cpuset_memory_pressure_enabled
;
1721 case FILE_MEMORY_PRESSURE
:
1722 return fmeter_getrate(&cs
->fmeter
);
1723 case FILE_SPREAD_PAGE
:
1724 return is_spread_page(cs
);
1725 case FILE_SPREAD_SLAB
:
1726 return is_spread_slab(cs
);
1731 /* Unreachable but makes gcc happy */
1735 static s64
cpuset_read_s64(struct cgroup
*cont
, struct cftype
*cft
)
1737 struct cpuset
*cs
= cgroup_cs(cont
);
1738 cpuset_filetype_t type
= cft
->private;
1740 case FILE_SCHED_RELAX_DOMAIN_LEVEL
:
1741 return cs
->relax_domain_level
;
1746 /* Unrechable but makes gcc happy */
1752 * for the common functions, 'private' gives the type of file
1755 static struct cftype files
[] = {
1758 .read
= cpuset_common_file_read
,
1759 .write_string
= cpuset_write_resmask
,
1760 .max_write_len
= (100U + 6 * NR_CPUS
),
1761 .private = FILE_CPULIST
,
1766 .read
= cpuset_common_file_read
,
1767 .write_string
= cpuset_write_resmask
,
1768 .max_write_len
= (100U + 6 * MAX_NUMNODES
),
1769 .private = FILE_MEMLIST
,
1773 .name
= "cpu_exclusive",
1774 .read_u64
= cpuset_read_u64
,
1775 .write_u64
= cpuset_write_u64
,
1776 .private = FILE_CPU_EXCLUSIVE
,
1780 .name
= "mem_exclusive",
1781 .read_u64
= cpuset_read_u64
,
1782 .write_u64
= cpuset_write_u64
,
1783 .private = FILE_MEM_EXCLUSIVE
,
1787 .name
= "mem_hardwall",
1788 .read_u64
= cpuset_read_u64
,
1789 .write_u64
= cpuset_write_u64
,
1790 .private = FILE_MEM_HARDWALL
,
1794 .name
= "sched_load_balance",
1795 .read_u64
= cpuset_read_u64
,
1796 .write_u64
= cpuset_write_u64
,
1797 .private = FILE_SCHED_LOAD_BALANCE
,
1801 .name
= "sched_relax_domain_level",
1802 .read_s64
= cpuset_read_s64
,
1803 .write_s64
= cpuset_write_s64
,
1804 .private = FILE_SCHED_RELAX_DOMAIN_LEVEL
,
1808 .name
= "memory_migrate",
1809 .read_u64
= cpuset_read_u64
,
1810 .write_u64
= cpuset_write_u64
,
1811 .private = FILE_MEMORY_MIGRATE
,
1815 .name
= "memory_pressure",
1816 .read_u64
= cpuset_read_u64
,
1817 .write_u64
= cpuset_write_u64
,
1818 .private = FILE_MEMORY_PRESSURE
,
1823 .name
= "memory_spread_page",
1824 .read_u64
= cpuset_read_u64
,
1825 .write_u64
= cpuset_write_u64
,
1826 .private = FILE_SPREAD_PAGE
,
1830 .name
= "memory_spread_slab",
1831 .read_u64
= cpuset_read_u64
,
1832 .write_u64
= cpuset_write_u64
,
1833 .private = FILE_SPREAD_SLAB
,
1837 .name
= "memory_pressure_enabled",
1838 .flags
= CFTYPE_ONLY_ON_ROOT
,
1839 .read_u64
= cpuset_read_u64
,
1840 .write_u64
= cpuset_write_u64
,
1841 .private = FILE_MEMORY_PRESSURE_ENABLED
,
1848 * cpuset_css_alloc - allocate a cpuset css
1849 * cont: control group that the new cpuset will be part of
1852 static struct cgroup_subsys_state
*cpuset_css_alloc(struct cgroup
*cont
)
1857 return &top_cpuset
.css
;
1859 cs
= kzalloc(sizeof(*cs
), GFP_KERNEL
);
1861 return ERR_PTR(-ENOMEM
);
1862 if (!alloc_cpumask_var(&cs
->cpus_allowed
, GFP_KERNEL
)) {
1864 return ERR_PTR(-ENOMEM
);
1867 set_bit(CS_SCHED_LOAD_BALANCE
, &cs
->flags
);
1868 cpumask_clear(cs
->cpus_allowed
);
1869 nodes_clear(cs
->mems_allowed
);
1870 fmeter_init(&cs
->fmeter
);
1871 INIT_WORK(&cs
->hotplug_work
, cpuset_propagate_hotplug_workfn
);
1872 cs
->relax_domain_level
= -1;
1877 static int cpuset_css_online(struct cgroup
*cgrp
)
1879 struct cpuset
*cs
= cgroup_cs(cgrp
);
1880 struct cpuset
*parent
= parent_cs(cs
);
1881 struct cpuset
*tmp_cs
;
1882 struct cgroup
*pos_cg
;
1887 mutex_lock(&cpuset_mutex
);
1889 set_bit(CS_ONLINE
, &cs
->flags
);
1890 if (is_spread_page(parent
))
1891 set_bit(CS_SPREAD_PAGE
, &cs
->flags
);
1892 if (is_spread_slab(parent
))
1893 set_bit(CS_SPREAD_SLAB
, &cs
->flags
);
1895 number_of_cpusets
++;
1897 if (!test_bit(CGRP_CPUSET_CLONE_CHILDREN
, &cgrp
->flags
))
1901 * Clone @parent's configuration if CGRP_CPUSET_CLONE_CHILDREN is
1902 * set. This flag handling is implemented in cgroup core for
1903 * histrical reasons - the flag may be specified during mount.
1905 * Currently, if any sibling cpusets have exclusive cpus or mem, we
1906 * refuse to clone the configuration - thereby refusing the task to
1907 * be entered, and as a result refusing the sys_unshare() or
1908 * clone() which initiated it. If this becomes a problem for some
1909 * users who wish to allow that scenario, then this could be
1910 * changed to grant parent->cpus_allowed-sibling_cpus_exclusive
1911 * (and likewise for mems) to the new cgroup.
1914 cpuset_for_each_child(tmp_cs
, pos_cg
, parent
) {
1915 if (is_mem_exclusive(tmp_cs
) || is_cpu_exclusive(tmp_cs
)) {
1922 mutex_lock(&callback_mutex
);
1923 cs
->mems_allowed
= parent
->mems_allowed
;
1924 cpumask_copy(cs
->cpus_allowed
, parent
->cpus_allowed
);
1925 mutex_unlock(&callback_mutex
);
1927 mutex_unlock(&cpuset_mutex
);
1931 static void cpuset_css_offline(struct cgroup
*cgrp
)
1933 struct cpuset
*cs
= cgroup_cs(cgrp
);
1935 mutex_lock(&cpuset_mutex
);
1937 if (is_sched_load_balance(cs
))
1938 update_flag(CS_SCHED_LOAD_BALANCE
, cs
, 0);
1940 number_of_cpusets
--;
1941 clear_bit(CS_ONLINE
, &cs
->flags
);
1943 mutex_unlock(&cpuset_mutex
);
1947 * If the cpuset being removed has its flag 'sched_load_balance'
1948 * enabled, then simulate turning sched_load_balance off, which
1949 * will call rebuild_sched_domains_locked().
1952 static void cpuset_css_free(struct cgroup
*cont
)
1954 struct cpuset
*cs
= cgroup_cs(cont
);
1956 free_cpumask_var(cs
->cpus_allowed
);
1960 struct cgroup_subsys cpuset_subsys
= {
1962 .css_alloc
= cpuset_css_alloc
,
1963 .css_online
= cpuset_css_online
,
1964 .css_offline
= cpuset_css_offline
,
1965 .css_free
= cpuset_css_free
,
1966 .can_attach
= cpuset_can_attach
,
1967 .cancel_attach
= cpuset_cancel_attach
,
1968 .attach
= cpuset_attach
,
1969 .subsys_id
= cpuset_subsys_id
,
1970 .base_cftypes
= files
,
1975 * cpuset_init - initialize cpusets at system boot
1977 * Description: Initialize top_cpuset and the cpuset internal file system,
1980 int __init
cpuset_init(void)
1984 if (!alloc_cpumask_var(&top_cpuset
.cpus_allowed
, GFP_KERNEL
))
1987 cpumask_setall(top_cpuset
.cpus_allowed
);
1988 nodes_setall(top_cpuset
.mems_allowed
);
1990 fmeter_init(&top_cpuset
.fmeter
);
1991 set_bit(CS_SCHED_LOAD_BALANCE
, &top_cpuset
.flags
);
1992 top_cpuset
.relax_domain_level
= -1;
1994 err
= register_filesystem(&cpuset_fs_type
);
1998 if (!alloc_cpumask_var(&cpus_attach
, GFP_KERNEL
))
2001 number_of_cpusets
= 1;
2006 * If CPU and/or memory hotplug handlers, below, unplug any CPUs
2007 * or memory nodes, we need to walk over the cpuset hierarchy,
2008 * removing that CPU or node from all cpusets. If this removes the
2009 * last CPU or node from a cpuset, then move the tasks in the empty
2010 * cpuset to its next-highest non-empty parent.
2012 static void remove_tasks_in_empty_cpuset(struct cpuset
*cs
)
2014 struct cpuset
*parent
;
2017 * Find its next-highest non-empty parent, (top cpuset
2018 * has online cpus, so can't be empty).
2020 parent
= parent_cs(cs
);
2021 while (cpumask_empty(parent
->cpus_allowed
) ||
2022 nodes_empty(parent
->mems_allowed
))
2023 parent
= parent_cs(parent
);
2025 if (cgroup_transfer_tasks(parent
->css
.cgroup
, cs
->css
.cgroup
)) {
2027 printk(KERN_ERR
"cpuset: failed to transfer tasks out of empty cpuset %s\n",
2028 cgroup_name(cs
->css
.cgroup
));
2034 * cpuset_propagate_hotplug_workfn - propagate CPU/memory hotplug to a cpuset
2035 * @cs: cpuset in interest
2037 * Compare @cs's cpu and mem masks against top_cpuset and if some have gone
2038 * offline, update @cs accordingly. If @cs ends up with no CPU or memory,
2039 * all its tasks are moved to the nearest ancestor with both resources.
2041 static void cpuset_propagate_hotplug_workfn(struct work_struct
*work
)
2043 static cpumask_t off_cpus
;
2044 static nodemask_t off_mems
, tmp_mems
;
2045 struct cpuset
*cs
= container_of(work
, struct cpuset
, hotplug_work
);
2048 mutex_lock(&cpuset_mutex
);
2050 cpumask_andnot(&off_cpus
, cs
->cpus_allowed
, top_cpuset
.cpus_allowed
);
2051 nodes_andnot(off_mems
, cs
->mems_allowed
, top_cpuset
.mems_allowed
);
2053 /* remove offline cpus from @cs */
2054 if (!cpumask_empty(&off_cpus
)) {
2055 mutex_lock(&callback_mutex
);
2056 cpumask_andnot(cs
->cpus_allowed
, cs
->cpus_allowed
, &off_cpus
);
2057 mutex_unlock(&callback_mutex
);
2058 update_tasks_cpumask(cs
, NULL
);
2061 /* remove offline mems from @cs */
2062 if (!nodes_empty(off_mems
)) {
2063 tmp_mems
= cs
->mems_allowed
;
2064 mutex_lock(&callback_mutex
);
2065 nodes_andnot(cs
->mems_allowed
, cs
->mems_allowed
, off_mems
);
2066 mutex_unlock(&callback_mutex
);
2067 update_tasks_nodemask(cs
, &tmp_mems
, NULL
);
2070 is_empty
= cpumask_empty(cs
->cpus_allowed
) ||
2071 nodes_empty(cs
->mems_allowed
);
2073 mutex_unlock(&cpuset_mutex
);
2076 * If @cs became empty, move tasks to the nearest ancestor with
2077 * execution resources. This is full cgroup operation which will
2078 * also call back into cpuset. Should be done outside any lock.
2081 remove_tasks_in_empty_cpuset(cs
);
2083 /* the following may free @cs, should be the last operation */
2088 * schedule_cpuset_propagate_hotplug - schedule hotplug propagation to a cpuset
2089 * @cs: cpuset of interest
2091 * Schedule cpuset_propagate_hotplug_workfn() which will update CPU and
2092 * memory masks according to top_cpuset.
2094 static void schedule_cpuset_propagate_hotplug(struct cpuset
*cs
)
2097 * Pin @cs. The refcnt will be released when the work item
2098 * finishes executing.
2100 if (!css_tryget(&cs
->css
))
2104 * Queue @cs->hotplug_work. If already pending, lose the css ref.
2105 * cpuset_propagate_hotplug_wq is ordered and propagation will
2106 * happen in the order this function is called.
2108 if (!queue_work(cpuset_propagate_hotplug_wq
, &cs
->hotplug_work
))
2113 * cpuset_hotplug_workfn - handle CPU/memory hotunplug for a cpuset
2115 * This function is called after either CPU or memory configuration has
2116 * changed and updates cpuset accordingly. The top_cpuset is always
2117 * synchronized to cpu_active_mask and N_MEMORY, which is necessary in
2118 * order to make cpusets transparent (of no affect) on systems that are
2119 * actively using CPU hotplug but making no active use of cpusets.
2121 * Non-root cpusets are only affected by offlining. If any CPUs or memory
2122 * nodes have been taken down, cpuset_propagate_hotplug() is invoked on all
2125 * Note that CPU offlining during suspend is ignored. We don't modify
2126 * cpusets across suspend/resume cycles at all.
2128 static void cpuset_hotplug_workfn(struct work_struct
*work
)
2130 static cpumask_t new_cpus
, tmp_cpus
;
2131 static nodemask_t new_mems
, tmp_mems
;
2132 bool cpus_updated
, mems_updated
;
2133 bool cpus_offlined
, mems_offlined
;
2135 mutex_lock(&cpuset_mutex
);
2137 /* fetch the available cpus/mems and find out which changed how */
2138 cpumask_copy(&new_cpus
, cpu_active_mask
);
2139 new_mems
= node_states
[N_MEMORY
];
2141 cpus_updated
= !cpumask_equal(top_cpuset
.cpus_allowed
, &new_cpus
);
2142 cpus_offlined
= cpumask_andnot(&tmp_cpus
, top_cpuset
.cpus_allowed
,
2145 mems_updated
= !nodes_equal(top_cpuset
.mems_allowed
, new_mems
);
2146 nodes_andnot(tmp_mems
, top_cpuset
.mems_allowed
, new_mems
);
2147 mems_offlined
= !nodes_empty(tmp_mems
);
2149 /* synchronize cpus_allowed to cpu_active_mask */
2151 mutex_lock(&callback_mutex
);
2152 cpumask_copy(top_cpuset
.cpus_allowed
, &new_cpus
);
2153 mutex_unlock(&callback_mutex
);
2154 /* we don't mess with cpumasks of tasks in top_cpuset */
2157 /* synchronize mems_allowed to N_MEMORY */
2159 tmp_mems
= top_cpuset
.mems_allowed
;
2160 mutex_lock(&callback_mutex
);
2161 top_cpuset
.mems_allowed
= new_mems
;
2162 mutex_unlock(&callback_mutex
);
2163 update_tasks_nodemask(&top_cpuset
, &tmp_mems
, NULL
);
2166 /* if cpus or mems went down, we need to propagate to descendants */
2167 if (cpus_offlined
|| mems_offlined
) {
2169 struct cgroup
*pos_cgrp
;
2172 cpuset_for_each_descendant_pre(cs
, pos_cgrp
, &top_cpuset
)
2173 schedule_cpuset_propagate_hotplug(cs
);
2177 mutex_unlock(&cpuset_mutex
);
2179 /* wait for propagations to finish */
2180 flush_workqueue(cpuset_propagate_hotplug_wq
);
2182 /* rebuild sched domains if cpus_allowed has changed */
2184 rebuild_sched_domains();
2187 void cpuset_update_active_cpus(bool cpu_online
)
2190 * We're inside cpu hotplug critical region which usually nests
2191 * inside cgroup synchronization. Bounce actual hotplug processing
2192 * to a work item to avoid reverse locking order.
2194 * We still need to do partition_sched_domains() synchronously;
2195 * otherwise, the scheduler will get confused and put tasks to the
2196 * dead CPU. Fall back to the default single domain.
2197 * cpuset_hotplug_workfn() will rebuild it as necessary.
2199 partition_sched_domains(1, NULL
, NULL
);
2200 schedule_work(&cpuset_hotplug_work
);
2203 #ifdef CONFIG_MEMORY_HOTPLUG
2205 * Keep top_cpuset.mems_allowed tracking node_states[N_MEMORY].
2206 * Call this routine anytime after node_states[N_MEMORY] changes.
2207 * See cpuset_update_active_cpus() for CPU hotplug handling.
2209 static int cpuset_track_online_nodes(struct notifier_block
*self
,
2210 unsigned long action
, void *arg
)
2212 schedule_work(&cpuset_hotplug_work
);
2218 * cpuset_init_smp - initialize cpus_allowed
2220 * Description: Finish top cpuset after cpu, node maps are initialized
2223 void __init
cpuset_init_smp(void)
2225 cpumask_copy(top_cpuset
.cpus_allowed
, cpu_active_mask
);
2226 top_cpuset
.mems_allowed
= node_states
[N_MEMORY
];
2228 hotplug_memory_notifier(cpuset_track_online_nodes
, 10);
2230 cpuset_propagate_hotplug_wq
=
2231 alloc_ordered_workqueue("cpuset_hotplug", 0);
2232 BUG_ON(!cpuset_propagate_hotplug_wq
);
2236 * cpuset_cpus_allowed - return cpus_allowed mask from a tasks cpuset.
2237 * @tsk: pointer to task_struct from which to obtain cpuset->cpus_allowed.
2238 * @pmask: pointer to struct cpumask variable to receive cpus_allowed set.
2240 * Description: Returns the cpumask_var_t cpus_allowed of the cpuset
2241 * attached to the specified @tsk. Guaranteed to return some non-empty
2242 * subset of cpu_online_mask, even if this means going outside the
2246 void cpuset_cpus_allowed(struct task_struct
*tsk
, struct cpumask
*pmask
)
2248 mutex_lock(&callback_mutex
);
2250 guarantee_online_cpus(task_cs(tsk
), pmask
);
2252 mutex_unlock(&callback_mutex
);
2255 void cpuset_cpus_allowed_fallback(struct task_struct
*tsk
)
2257 const struct cpuset
*cs
;
2262 do_set_cpus_allowed(tsk
, cs
->cpus_allowed
);
2266 * We own tsk->cpus_allowed, nobody can change it under us.
2268 * But we used cs && cs->cpus_allowed lockless and thus can
2269 * race with cgroup_attach_task() or update_cpumask() and get
2270 * the wrong tsk->cpus_allowed. However, both cases imply the
2271 * subsequent cpuset_change_cpumask()->set_cpus_allowed_ptr()
2272 * which takes task_rq_lock().
2274 * If we are called after it dropped the lock we must see all
2275 * changes in tsk_cs()->cpus_allowed. Otherwise we can temporary
2276 * set any mask even if it is not right from task_cs() pov,
2277 * the pending set_cpus_allowed_ptr() will fix things.
2279 * select_fallback_rq() will fix things ups and set cpu_possible_mask
2284 void cpuset_init_current_mems_allowed(void)
2286 nodes_setall(current
->mems_allowed
);
2290 * cpuset_mems_allowed - return mems_allowed mask from a tasks cpuset.
2291 * @tsk: pointer to task_struct from which to obtain cpuset->mems_allowed.
2293 * Description: Returns the nodemask_t mems_allowed of the cpuset
2294 * attached to the specified @tsk. Guaranteed to return some non-empty
2295 * subset of node_states[N_MEMORY], even if this means going outside the
2299 nodemask_t
cpuset_mems_allowed(struct task_struct
*tsk
)
2303 mutex_lock(&callback_mutex
);
2305 guarantee_online_mems(task_cs(tsk
), &mask
);
2307 mutex_unlock(&callback_mutex
);
2313 * cpuset_nodemask_valid_mems_allowed - check nodemask vs. curremt mems_allowed
2314 * @nodemask: the nodemask to be checked
2316 * Are any of the nodes in the nodemask allowed in current->mems_allowed?
2318 int cpuset_nodemask_valid_mems_allowed(nodemask_t
*nodemask
)
2320 return nodes_intersects(*nodemask
, current
->mems_allowed
);
2324 * nearest_hardwall_ancestor() - Returns the nearest mem_exclusive or
2325 * mem_hardwall ancestor to the specified cpuset. Call holding
2326 * callback_mutex. If no ancestor is mem_exclusive or mem_hardwall
2327 * (an unusual configuration), then returns the root cpuset.
2329 static const struct cpuset
*nearest_hardwall_ancestor(const struct cpuset
*cs
)
2331 while (!(is_mem_exclusive(cs
) || is_mem_hardwall(cs
)) && parent_cs(cs
))
2337 * cpuset_node_allowed_softwall - Can we allocate on a memory node?
2338 * @node: is this an allowed node?
2339 * @gfp_mask: memory allocation flags
2341 * If we're in interrupt, yes, we can always allocate. If __GFP_THISNODE is
2342 * set, yes, we can always allocate. If node is in our task's mems_allowed,
2343 * yes. If it's not a __GFP_HARDWALL request and this node is in the nearest
2344 * hardwalled cpuset ancestor to this task's cpuset, yes. If the task has been
2345 * OOM killed and has access to memory reserves as specified by the TIF_MEMDIE
2349 * If __GFP_HARDWALL is set, cpuset_node_allowed_softwall() reduces to
2350 * cpuset_node_allowed_hardwall(). Otherwise, cpuset_node_allowed_softwall()
2351 * might sleep, and might allow a node from an enclosing cpuset.
2353 * cpuset_node_allowed_hardwall() only handles the simpler case of hardwall
2354 * cpusets, and never sleeps.
2356 * The __GFP_THISNODE placement logic is really handled elsewhere,
2357 * by forcibly using a zonelist starting at a specified node, and by
2358 * (in get_page_from_freelist()) refusing to consider the zones for
2359 * any node on the zonelist except the first. By the time any such
2360 * calls get to this routine, we should just shut up and say 'yes'.
2362 * GFP_USER allocations are marked with the __GFP_HARDWALL bit,
2363 * and do not allow allocations outside the current tasks cpuset
2364 * unless the task has been OOM killed as is marked TIF_MEMDIE.
2365 * GFP_KERNEL allocations are not so marked, so can escape to the
2366 * nearest enclosing hardwalled ancestor cpuset.
2368 * Scanning up parent cpusets requires callback_mutex. The
2369 * __alloc_pages() routine only calls here with __GFP_HARDWALL bit
2370 * _not_ set if it's a GFP_KERNEL allocation, and all nodes in the
2371 * current tasks mems_allowed came up empty on the first pass over
2372 * the zonelist. So only GFP_KERNEL allocations, if all nodes in the
2373 * cpuset are short of memory, might require taking the callback_mutex
2376 * The first call here from mm/page_alloc:get_page_from_freelist()
2377 * has __GFP_HARDWALL set in gfp_mask, enforcing hardwall cpusets,
2378 * so no allocation on a node outside the cpuset is allowed (unless
2379 * in interrupt, of course).
2381 * The second pass through get_page_from_freelist() doesn't even call
2382 * here for GFP_ATOMIC calls. For those calls, the __alloc_pages()
2383 * variable 'wait' is not set, and the bit ALLOC_CPUSET is not set
2384 * in alloc_flags. That logic and the checks below have the combined
2386 * in_interrupt - any node ok (current task context irrelevant)
2387 * GFP_ATOMIC - any node ok
2388 * TIF_MEMDIE - any node ok
2389 * GFP_KERNEL - any node in enclosing hardwalled cpuset ok
2390 * GFP_USER - only nodes in current tasks mems allowed ok.
2393 * Don't call cpuset_node_allowed_softwall if you can't sleep, unless you
2394 * pass in the __GFP_HARDWALL flag set in gfp_flag, which disables
2395 * the code that might scan up ancestor cpusets and sleep.
2397 int __cpuset_node_allowed_softwall(int node
, gfp_t gfp_mask
)
2399 const struct cpuset
*cs
; /* current cpuset ancestors */
2400 int allowed
; /* is allocation in zone z allowed? */
2402 if (in_interrupt() || (gfp_mask
& __GFP_THISNODE
))
2404 might_sleep_if(!(gfp_mask
& __GFP_HARDWALL
));
2405 if (node_isset(node
, current
->mems_allowed
))
2408 * Allow tasks that have access to memory reserves because they have
2409 * been OOM killed to get memory anywhere.
2411 if (unlikely(test_thread_flag(TIF_MEMDIE
)))
2413 if (gfp_mask
& __GFP_HARDWALL
) /* If hardwall request, stop here */
2416 if (current
->flags
& PF_EXITING
) /* Let dying task have memory */
2419 /* Not hardwall and node outside mems_allowed: scan up cpusets */
2420 mutex_lock(&callback_mutex
);
2423 cs
= nearest_hardwall_ancestor(task_cs(current
));
2424 task_unlock(current
);
2426 allowed
= node_isset(node
, cs
->mems_allowed
);
2427 mutex_unlock(&callback_mutex
);
2432 * cpuset_node_allowed_hardwall - Can we allocate on a memory node?
2433 * @node: is this an allowed node?
2434 * @gfp_mask: memory allocation flags
2436 * If we're in interrupt, yes, we can always allocate. If __GFP_THISNODE is
2437 * set, yes, we can always allocate. If node is in our task's mems_allowed,
2438 * yes. If the task has been OOM killed and has access to memory reserves as
2439 * specified by the TIF_MEMDIE flag, yes.
2442 * The __GFP_THISNODE placement logic is really handled elsewhere,
2443 * by forcibly using a zonelist starting at a specified node, and by
2444 * (in get_page_from_freelist()) refusing to consider the zones for
2445 * any node on the zonelist except the first. By the time any such
2446 * calls get to this routine, we should just shut up and say 'yes'.
2448 * Unlike the cpuset_node_allowed_softwall() variant, above,
2449 * this variant requires that the node be in the current task's
2450 * mems_allowed or that we're in interrupt. It does not scan up the
2451 * cpuset hierarchy for the nearest enclosing mem_exclusive cpuset.
2454 int __cpuset_node_allowed_hardwall(int node
, gfp_t gfp_mask
)
2456 if (in_interrupt() || (gfp_mask
& __GFP_THISNODE
))
2458 if (node_isset(node
, current
->mems_allowed
))
2461 * Allow tasks that have access to memory reserves because they have
2462 * been OOM killed to get memory anywhere.
2464 if (unlikely(test_thread_flag(TIF_MEMDIE
)))
2470 * cpuset_mem_spread_node() - On which node to begin search for a file page
2471 * cpuset_slab_spread_node() - On which node to begin search for a slab page
2473 * If a task is marked PF_SPREAD_PAGE or PF_SPREAD_SLAB (as for
2474 * tasks in a cpuset with is_spread_page or is_spread_slab set),
2475 * and if the memory allocation used cpuset_mem_spread_node()
2476 * to determine on which node to start looking, as it will for
2477 * certain page cache or slab cache pages such as used for file
2478 * system buffers and inode caches, then instead of starting on the
2479 * local node to look for a free page, rather spread the starting
2480 * node around the tasks mems_allowed nodes.
2482 * We don't have to worry about the returned node being offline
2483 * because "it can't happen", and even if it did, it would be ok.
2485 * The routines calling guarantee_online_mems() are careful to
2486 * only set nodes in task->mems_allowed that are online. So it
2487 * should not be possible for the following code to return an
2488 * offline node. But if it did, that would be ok, as this routine
2489 * is not returning the node where the allocation must be, only
2490 * the node where the search should start. The zonelist passed to
2491 * __alloc_pages() will include all nodes. If the slab allocator
2492 * is passed an offline node, it will fall back to the local node.
2493 * See kmem_cache_alloc_node().
2496 static int cpuset_spread_node(int *rotor
)
2500 node
= next_node(*rotor
, current
->mems_allowed
);
2501 if (node
== MAX_NUMNODES
)
2502 node
= first_node(current
->mems_allowed
);
2507 int cpuset_mem_spread_node(void)
2509 if (current
->cpuset_mem_spread_rotor
== NUMA_NO_NODE
)
2510 current
->cpuset_mem_spread_rotor
=
2511 node_random(¤t
->mems_allowed
);
2513 return cpuset_spread_node(¤t
->cpuset_mem_spread_rotor
);
2516 int cpuset_slab_spread_node(void)
2518 if (current
->cpuset_slab_spread_rotor
== NUMA_NO_NODE
)
2519 current
->cpuset_slab_spread_rotor
=
2520 node_random(¤t
->mems_allowed
);
2522 return cpuset_spread_node(¤t
->cpuset_slab_spread_rotor
);
2525 EXPORT_SYMBOL_GPL(cpuset_mem_spread_node
);
2528 * cpuset_mems_allowed_intersects - Does @tsk1's mems_allowed intersect @tsk2's?
2529 * @tsk1: pointer to task_struct of some task.
2530 * @tsk2: pointer to task_struct of some other task.
2532 * Description: Return true if @tsk1's mems_allowed intersects the
2533 * mems_allowed of @tsk2. Used by the OOM killer to determine if
2534 * one of the task's memory usage might impact the memory available
2538 int cpuset_mems_allowed_intersects(const struct task_struct
*tsk1
,
2539 const struct task_struct
*tsk2
)
2541 return nodes_intersects(tsk1
->mems_allowed
, tsk2
->mems_allowed
);
2544 #define CPUSET_NODELIST_LEN (256)
2547 * cpuset_print_task_mems_allowed - prints task's cpuset and mems_allowed
2548 * @task: pointer to task_struct of some task.
2550 * Description: Prints @task's name, cpuset name, and cached copy of its
2551 * mems_allowed to the kernel log. Must hold task_lock(task) to allow
2552 * dereferencing task_cs(task).
2554 void cpuset_print_task_mems_allowed(struct task_struct
*tsk
)
2556 /* Statically allocated to prevent using excess stack. */
2557 static char cpuset_nodelist
[CPUSET_NODELIST_LEN
];
2558 static DEFINE_SPINLOCK(cpuset_buffer_lock
);
2560 struct cgroup
*cgrp
= task_cs(tsk
)->css
.cgroup
;
2563 spin_lock(&cpuset_buffer_lock
);
2565 nodelist_scnprintf(cpuset_nodelist
, CPUSET_NODELIST_LEN
,
2567 printk(KERN_INFO
"%s cpuset=%s mems_allowed=%s\n",
2568 tsk
->comm
, cgroup_name(cgrp
), cpuset_nodelist
);
2570 spin_unlock(&cpuset_buffer_lock
);
2575 * Collection of memory_pressure is suppressed unless
2576 * this flag is enabled by writing "1" to the special
2577 * cpuset file 'memory_pressure_enabled' in the root cpuset.
2580 int cpuset_memory_pressure_enabled __read_mostly
;
2583 * cpuset_memory_pressure_bump - keep stats of per-cpuset reclaims.
2585 * Keep a running average of the rate of synchronous (direct)
2586 * page reclaim efforts initiated by tasks in each cpuset.
2588 * This represents the rate at which some task in the cpuset
2589 * ran low on memory on all nodes it was allowed to use, and
2590 * had to enter the kernels page reclaim code in an effort to
2591 * create more free memory by tossing clean pages or swapping
2592 * or writing dirty pages.
2594 * Display to user space in the per-cpuset read-only file
2595 * "memory_pressure". Value displayed is an integer
2596 * representing the recent rate of entry into the synchronous
2597 * (direct) page reclaim by any task attached to the cpuset.
2600 void __cpuset_memory_pressure_bump(void)
2603 fmeter_markevent(&task_cs(current
)->fmeter
);
2604 task_unlock(current
);
2607 #ifdef CONFIG_PROC_PID_CPUSET
2609 * proc_cpuset_show()
2610 * - Print tasks cpuset path into seq_file.
2611 * - Used for /proc/<pid>/cpuset.
2612 * - No need to task_lock(tsk) on this tsk->cpuset reference, as it
2613 * doesn't really matter if tsk->cpuset changes after we read it,
2614 * and we take cpuset_mutex, keeping cpuset_attach() from changing it
2617 static int proc_cpuset_show(struct seq_file
*m
, void *unused_v
)
2620 struct task_struct
*tsk
;
2622 struct cgroup_subsys_state
*css
;
2626 buf
= kmalloc(PAGE_SIZE
, GFP_KERNEL
);
2632 tsk
= get_pid_task(pid
, PIDTYPE_PID
);
2637 css
= task_subsys_state(tsk
, cpuset_subsys_id
);
2638 retval
= cgroup_path(css
->cgroup
, buf
, PAGE_SIZE
);
2645 put_task_struct(tsk
);
2652 static int cpuset_open(struct inode
*inode
, struct file
*file
)
2654 struct pid
*pid
= PROC_I(inode
)->pid
;
2655 return single_open(file
, proc_cpuset_show
, pid
);
2658 const struct file_operations proc_cpuset_operations
= {
2659 .open
= cpuset_open
,
2661 .llseek
= seq_lseek
,
2662 .release
= single_release
,
2664 #endif /* CONFIG_PROC_PID_CPUSET */
2666 /* Display task mems_allowed in /proc/<pid>/status file. */
2667 void cpuset_task_status_allowed(struct seq_file
*m
, struct task_struct
*task
)
2669 seq_printf(m
, "Mems_allowed:\t");
2670 seq_nodemask(m
, &task
->mems_allowed
);
2671 seq_printf(m
, "\n");
2672 seq_printf(m
, "Mems_allowed_list:\t");
2673 seq_nodemask_list(m
, &task
->mems_allowed
);
2674 seq_printf(m
, "\n");