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>
62 #include <linux/wait.h>
65 * Tracks how many cpusets are currently defined in system.
66 * When there is only one cpuset (the root cpuset) we can
67 * short circuit some hooks.
69 int number_of_cpusets __read_mostly
;
71 /* Forward declare cgroup structures */
72 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 */
91 * This is old Memory Nodes tasks took on.
93 * - top_cpuset.old_mems_allowed is initialized to mems_allowed.
94 * - A new cpuset's old_mems_allowed is initialized when some
95 * task is moved into it.
96 * - old_mems_allowed is used in cpuset_migrate_mm() when we change
97 * cpuset.mems_allowed and have tasks' nodemask updated, and
98 * then old_mems_allowed is updated to mems_allowed.
100 nodemask_t old_mems_allowed
;
102 struct fmeter fmeter
; /* memory_pressure filter */
105 * Tasks are being attached to this cpuset. Used to prevent
106 * zeroing cpus/mems_allowed between ->can_attach() and ->attach().
108 int attach_in_progress
;
110 /* partition number for rebuild_sched_domains() */
113 /* for custom sched domain */
114 int relax_domain_level
;
117 static inline struct cpuset
*css_cs(struct cgroup_subsys_state
*css
)
119 return css
? container_of(css
, struct cpuset
, css
) : NULL
;
122 /* Retrieve the cpuset for a task */
123 static inline struct cpuset
*task_cs(struct task_struct
*task
)
125 return css_cs(task_css(task
, cpuset_subsys_id
));
128 static inline struct cpuset
*parent_cs(struct cpuset
*cs
)
130 return css_cs(css_parent(&cs
->css
));
134 static inline bool task_has_mempolicy(struct task_struct
*task
)
136 return task
->mempolicy
;
139 static inline bool task_has_mempolicy(struct task_struct
*task
)
146 /* bits in struct cpuset flags field */
153 CS_SCHED_LOAD_BALANCE
,
158 /* convenient tests for these bits */
159 static inline bool is_cpuset_online(const struct cpuset
*cs
)
161 return test_bit(CS_ONLINE
, &cs
->flags
);
164 static inline int is_cpu_exclusive(const struct cpuset
*cs
)
166 return test_bit(CS_CPU_EXCLUSIVE
, &cs
->flags
);
169 static inline int is_mem_exclusive(const struct cpuset
*cs
)
171 return test_bit(CS_MEM_EXCLUSIVE
, &cs
->flags
);
174 static inline int is_mem_hardwall(const struct cpuset
*cs
)
176 return test_bit(CS_MEM_HARDWALL
, &cs
->flags
);
179 static inline int is_sched_load_balance(const struct cpuset
*cs
)
181 return test_bit(CS_SCHED_LOAD_BALANCE
, &cs
->flags
);
184 static inline int is_memory_migrate(const struct cpuset
*cs
)
186 return test_bit(CS_MEMORY_MIGRATE
, &cs
->flags
);
189 static inline int is_spread_page(const struct cpuset
*cs
)
191 return test_bit(CS_SPREAD_PAGE
, &cs
->flags
);
194 static inline int is_spread_slab(const struct cpuset
*cs
)
196 return test_bit(CS_SPREAD_SLAB
, &cs
->flags
);
199 static struct cpuset top_cpuset
= {
200 .flags
= ((1 << CS_ONLINE
) | (1 << CS_CPU_EXCLUSIVE
) |
201 (1 << CS_MEM_EXCLUSIVE
)),
205 * cpuset_for_each_child - traverse online children of a cpuset
206 * @child_cs: loop cursor pointing to the current child
207 * @pos_css: used for iteration
208 * @parent_cs: target cpuset to walk children of
210 * Walk @child_cs through the online children of @parent_cs. Must be used
211 * with RCU read locked.
213 #define cpuset_for_each_child(child_cs, pos_css, parent_cs) \
214 css_for_each_child((pos_css), &(parent_cs)->css) \
215 if (is_cpuset_online(((child_cs) = css_cs((pos_css)))))
218 * cpuset_for_each_descendant_pre - pre-order walk of a cpuset's descendants
219 * @des_cs: loop cursor pointing to the current descendant
220 * @pos_css: used for iteration
221 * @root_cs: target cpuset to walk ancestor of
223 * Walk @des_cs through the online descendants of @root_cs. Must be used
224 * with RCU read locked. The caller may modify @pos_css by calling
225 * css_rightmost_descendant() to skip subtree.
227 #define cpuset_for_each_descendant_pre(des_cs, pos_css, root_cs) \
228 css_for_each_descendant_pre((pos_css), &(root_cs)->css) \
229 if (is_cpuset_online(((des_cs) = css_cs((pos_css)))))
232 * There are two global mutexes guarding cpuset structures - cpuset_mutex
233 * and callback_mutex. The latter may nest inside the former. We also
234 * require taking task_lock() when dereferencing a task's cpuset pointer.
235 * See "The task_lock() exception", at the end of this comment.
237 * A task must hold both mutexes to modify cpusets. If a task holds
238 * cpuset_mutex, then it blocks others wanting that mutex, ensuring that it
239 * is the only task able to also acquire callback_mutex and be able to
240 * modify cpusets. It can perform various checks on the cpuset structure
241 * first, knowing nothing will change. It can also allocate memory while
242 * just holding cpuset_mutex. While it is performing these checks, various
243 * callback routines can briefly acquire callback_mutex to query cpusets.
244 * Once it is ready to make the changes, it takes callback_mutex, blocking
247 * Calls to the kernel memory allocator can not be made while holding
248 * callback_mutex, as that would risk double tripping on callback_mutex
249 * from one of the callbacks into the cpuset code from within
252 * If a task is only holding callback_mutex, then it has read-only
255 * Now, the task_struct fields mems_allowed and mempolicy may be changed
256 * by other task, we use alloc_lock in the task_struct fields to protect
259 * The cpuset_common_file_read() handlers only hold callback_mutex across
260 * small pieces of code, such as when reading out possibly multi-word
261 * cpumasks and nodemasks.
263 * Accessing a task's cpuset should be done in accordance with the
264 * guidelines for accessing subsystem state in kernel/cgroup.c
267 static DEFINE_MUTEX(cpuset_mutex
);
268 static DEFINE_MUTEX(callback_mutex
);
271 * CPU / memory hotplug is handled asynchronously.
273 static void cpuset_hotplug_workfn(struct work_struct
*work
);
274 static DECLARE_WORK(cpuset_hotplug_work
, cpuset_hotplug_workfn
);
276 static DECLARE_WAIT_QUEUE_HEAD(cpuset_attach_wq
);
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. The top
308 * cpuset always has some cpus online.
310 * One way or another, we guarantee to return some non-empty subset
311 * of cpu_online_mask.
313 * Call with callback_mutex held.
315 static void guarantee_online_cpus(struct cpuset
*cs
, struct cpumask
*pmask
)
317 while (!cpumask_intersects(cs
->cpus_allowed
, cpu_online_mask
))
319 cpumask_and(pmask
, cs
->cpus_allowed
, cpu_online_mask
);
323 * Return in *pmask the portion of a cpusets's mems_allowed that
324 * are online, with memory. If none are online with memory, walk
325 * up the cpuset hierarchy until we find one that does have some
326 * online mems. The top cpuset always has some mems online.
328 * One way or another, we guarantee to return some non-empty subset
329 * of node_states[N_MEMORY].
331 * Call with callback_mutex held.
333 static void guarantee_online_mems(struct cpuset
*cs
, nodemask_t
*pmask
)
335 while (!nodes_intersects(cs
->mems_allowed
, node_states
[N_MEMORY
]))
337 nodes_and(*pmask
, cs
->mems_allowed
, node_states
[N_MEMORY
]);
341 * update task's spread flag if cpuset's page/slab spread flag is set
343 * Called with callback_mutex/cpuset_mutex held
345 static void cpuset_update_task_spread_flag(struct cpuset
*cs
,
346 struct task_struct
*tsk
)
348 if (is_spread_page(cs
))
349 tsk
->flags
|= PF_SPREAD_PAGE
;
351 tsk
->flags
&= ~PF_SPREAD_PAGE
;
352 if (is_spread_slab(cs
))
353 tsk
->flags
|= PF_SPREAD_SLAB
;
355 tsk
->flags
&= ~PF_SPREAD_SLAB
;
359 * is_cpuset_subset(p, q) - Is cpuset p a subset of cpuset q?
361 * One cpuset is a subset of another if all its allowed CPUs and
362 * Memory Nodes are a subset of the other, and its exclusive flags
363 * are only set if the other's are set. Call holding cpuset_mutex.
366 static int is_cpuset_subset(const struct cpuset
*p
, const struct cpuset
*q
)
368 return cpumask_subset(p
->cpus_allowed
, q
->cpus_allowed
) &&
369 nodes_subset(p
->mems_allowed
, q
->mems_allowed
) &&
370 is_cpu_exclusive(p
) <= is_cpu_exclusive(q
) &&
371 is_mem_exclusive(p
) <= is_mem_exclusive(q
);
375 * alloc_trial_cpuset - allocate a trial cpuset
376 * @cs: the cpuset that the trial cpuset duplicates
378 static struct cpuset
*alloc_trial_cpuset(struct cpuset
*cs
)
380 struct cpuset
*trial
;
382 trial
= kmemdup(cs
, sizeof(*cs
), GFP_KERNEL
);
386 if (!alloc_cpumask_var(&trial
->cpus_allowed
, GFP_KERNEL
)) {
390 cpumask_copy(trial
->cpus_allowed
, cs
->cpus_allowed
);
396 * free_trial_cpuset - free the trial cpuset
397 * @trial: the trial cpuset to be freed
399 static void free_trial_cpuset(struct cpuset
*trial
)
401 free_cpumask_var(trial
->cpus_allowed
);
406 * validate_change() - Used to validate that any proposed cpuset change
407 * follows the structural rules for cpusets.
409 * If we replaced the flag and mask values of the current cpuset
410 * (cur) with those values in the trial cpuset (trial), would
411 * our various subset and exclusive rules still be valid? Presumes
414 * 'cur' is the address of an actual, in-use cpuset. Operations
415 * such as list traversal that depend on the actual address of the
416 * cpuset in the list must use cur below, not trial.
418 * 'trial' is the address of bulk structure copy of cur, with
419 * perhaps one or more of the fields cpus_allowed, mems_allowed,
420 * or flags changed to new, trial values.
422 * Return 0 if valid, -errno if not.
425 static int validate_change(struct cpuset
*cur
, struct cpuset
*trial
)
427 struct cgroup_subsys_state
*css
;
428 struct cpuset
*c
, *par
;
433 /* Each of our child cpusets must be a subset of us */
435 cpuset_for_each_child(c
, css
, cur
)
436 if (!is_cpuset_subset(c
, trial
))
439 /* Remaining checks don't apply to root cpuset */
441 if (cur
== &top_cpuset
)
444 par
= parent_cs(cur
);
446 /* We must be a subset of our parent cpuset */
448 if (!is_cpuset_subset(trial
, par
))
452 * If either I or some sibling (!= me) is exclusive, we can't
456 cpuset_for_each_child(c
, css
, par
) {
457 if ((is_cpu_exclusive(trial
) || is_cpu_exclusive(c
)) &&
459 cpumask_intersects(trial
->cpus_allowed
, c
->cpus_allowed
))
461 if ((is_mem_exclusive(trial
) || is_mem_exclusive(c
)) &&
463 nodes_intersects(trial
->mems_allowed
, c
->mems_allowed
))
468 * Cpusets with tasks - existing or newly being attached - can't
469 * have empty cpus_allowed or mems_allowed.
472 if ((cgroup_task_count(cur
->css
.cgroup
) || cur
->attach_in_progress
) &&
473 (cpumask_empty(trial
->cpus_allowed
) &&
474 nodes_empty(trial
->mems_allowed
)))
485 * Helper routine for generate_sched_domains().
486 * Do cpusets a, b have overlapping cpus_allowed masks?
488 static int cpusets_overlap(struct cpuset
*a
, struct cpuset
*b
)
490 return cpumask_intersects(a
->cpus_allowed
, b
->cpus_allowed
);
494 update_domain_attr(struct sched_domain_attr
*dattr
, struct cpuset
*c
)
496 if (dattr
->relax_domain_level
< c
->relax_domain_level
)
497 dattr
->relax_domain_level
= c
->relax_domain_level
;
501 static void update_domain_attr_tree(struct sched_domain_attr
*dattr
,
502 struct cpuset
*root_cs
)
505 struct cgroup_subsys_state
*pos_css
;
508 cpuset_for_each_descendant_pre(cp
, pos_css
, root_cs
) {
509 /* skip the whole subtree if @cp doesn't have any CPU */
510 if (cpumask_empty(cp
->cpus_allowed
)) {
511 pos_css
= css_rightmost_descendant(pos_css
);
515 if (is_sched_load_balance(cp
))
516 update_domain_attr(dattr
, cp
);
522 * generate_sched_domains()
524 * This function builds a partial partition of the systems CPUs
525 * A 'partial partition' is a set of non-overlapping subsets whose
526 * union is a subset of that set.
527 * The output of this function needs to be passed to kernel/sched/core.c
528 * partition_sched_domains() routine, which will rebuild the scheduler's
529 * load balancing domains (sched domains) as specified by that partial
532 * See "What is sched_load_balance" in Documentation/cgroups/cpusets.txt
533 * for a background explanation of this.
535 * Does not return errors, on the theory that the callers of this
536 * routine would rather not worry about failures to rebuild sched
537 * domains when operating in the severe memory shortage situations
538 * that could cause allocation failures below.
540 * Must be called with cpuset_mutex held.
542 * The three key local variables below are:
543 * q - a linked-list queue of cpuset pointers, used to implement a
544 * top-down scan of all cpusets. This scan loads a pointer
545 * to each cpuset marked is_sched_load_balance into the
546 * array 'csa'. For our purposes, rebuilding the schedulers
547 * sched domains, we can ignore !is_sched_load_balance cpusets.
548 * csa - (for CpuSet Array) Array of pointers to all the cpusets
549 * that need to be load balanced, for convenient iterative
550 * access by the subsequent code that finds the best partition,
551 * i.e the set of domains (subsets) of CPUs such that the
552 * cpus_allowed of every cpuset marked is_sched_load_balance
553 * is a subset of one of these domains, while there are as
554 * many such domains as possible, each as small as possible.
555 * doms - Conversion of 'csa' to an array of cpumasks, for passing to
556 * the kernel/sched/core.c routine partition_sched_domains() in a
557 * convenient format, that can be easily compared to the prior
558 * value to determine what partition elements (sched domains)
559 * were changed (added or removed.)
561 * Finding the best partition (set of domains):
562 * The triple nested loops below over i, j, k scan over the
563 * load balanced cpusets (using the array of cpuset pointers in
564 * csa[]) looking for pairs of cpusets that have overlapping
565 * cpus_allowed, but which don't have the same 'pn' partition
566 * number and gives them in the same partition number. It keeps
567 * looping on the 'restart' label until it can no longer find
570 * The union of the cpus_allowed masks from the set of
571 * all cpusets having the same 'pn' value then form the one
572 * element of the partition (one sched domain) to be passed to
573 * partition_sched_domains().
575 static int generate_sched_domains(cpumask_var_t
**domains
,
576 struct sched_domain_attr
**attributes
)
578 struct cpuset
*cp
; /* scans q */
579 struct cpuset
**csa
; /* array of all cpuset ptrs */
580 int csn
; /* how many cpuset ptrs in csa so far */
581 int i
, j
, k
; /* indices for partition finding loops */
582 cpumask_var_t
*doms
; /* resulting partition; i.e. sched domains */
583 struct sched_domain_attr
*dattr
; /* attributes for custom domains */
584 int ndoms
= 0; /* number of sched domains in result */
585 int nslot
; /* next empty doms[] struct cpumask slot */
586 struct cgroup_subsys_state
*pos_css
;
592 /* Special case for the 99% of systems with one, full, sched domain */
593 if (is_sched_load_balance(&top_cpuset
)) {
595 doms
= alloc_sched_domains(ndoms
);
599 dattr
= kmalloc(sizeof(struct sched_domain_attr
), GFP_KERNEL
);
601 *dattr
= SD_ATTR_INIT
;
602 update_domain_attr_tree(dattr
, &top_cpuset
);
604 cpumask_copy(doms
[0], top_cpuset
.cpus_allowed
);
609 csa
= kmalloc(number_of_cpusets
* sizeof(cp
), GFP_KERNEL
);
615 cpuset_for_each_descendant_pre(cp
, pos_css
, &top_cpuset
) {
617 * Continue traversing beyond @cp iff @cp has some CPUs and
618 * isn't load balancing. The former is obvious. The
619 * latter: All child cpusets contain a subset of the
620 * parent's cpus, so just skip them, and then we call
621 * update_domain_attr_tree() to calc relax_domain_level of
622 * the corresponding sched domain.
624 if (!cpumask_empty(cp
->cpus_allowed
) &&
625 !is_sched_load_balance(cp
))
628 if (is_sched_load_balance(cp
))
631 /* skip @cp's subtree */
632 pos_css
= css_rightmost_descendant(pos_css
);
636 for (i
= 0; i
< csn
; i
++)
641 /* Find the best partition (set of sched domains) */
642 for (i
= 0; i
< csn
; i
++) {
643 struct cpuset
*a
= csa
[i
];
646 for (j
= 0; j
< csn
; j
++) {
647 struct cpuset
*b
= csa
[j
];
650 if (apn
!= bpn
&& cpusets_overlap(a
, b
)) {
651 for (k
= 0; k
< csn
; k
++) {
652 struct cpuset
*c
= csa
[k
];
657 ndoms
--; /* one less element */
664 * Now we know how many domains to create.
665 * Convert <csn, csa> to <ndoms, doms> and populate cpu masks.
667 doms
= alloc_sched_domains(ndoms
);
672 * The rest of the code, including the scheduler, can deal with
673 * dattr==NULL case. No need to abort if alloc fails.
675 dattr
= kmalloc(ndoms
* sizeof(struct sched_domain_attr
), GFP_KERNEL
);
677 for (nslot
= 0, i
= 0; i
< csn
; i
++) {
678 struct cpuset
*a
= csa
[i
];
683 /* Skip completed partitions */
689 if (nslot
== ndoms
) {
690 static int warnings
= 10;
693 "rebuild_sched_domains confused:"
694 " nslot %d, ndoms %d, csn %d, i %d,"
696 nslot
, ndoms
, csn
, i
, apn
);
704 *(dattr
+ nslot
) = SD_ATTR_INIT
;
705 for (j
= i
; j
< csn
; j
++) {
706 struct cpuset
*b
= csa
[j
];
709 cpumask_or(dp
, dp
, b
->cpus_allowed
);
711 update_domain_attr_tree(dattr
+ nslot
, b
);
713 /* Done with this partition */
719 BUG_ON(nslot
!= ndoms
);
725 * Fallback to the default domain if kmalloc() failed.
726 * See comments in partition_sched_domains().
737 * Rebuild scheduler domains.
739 * If the flag 'sched_load_balance' of any cpuset with non-empty
740 * 'cpus' changes, or if the 'cpus' allowed changes in any cpuset
741 * which has that flag enabled, or if any cpuset with a non-empty
742 * 'cpus' is removed, then call this routine to rebuild the
743 * scheduler's dynamic sched domains.
745 * Call with cpuset_mutex held. Takes get_online_cpus().
747 static void rebuild_sched_domains_locked(void)
749 struct sched_domain_attr
*attr
;
753 lockdep_assert_held(&cpuset_mutex
);
757 * We have raced with CPU hotplug. Don't do anything to avoid
758 * passing doms with offlined cpu to partition_sched_domains().
759 * Anyways, hotplug work item will rebuild sched domains.
761 if (!cpumask_equal(top_cpuset
.cpus_allowed
, cpu_active_mask
))
764 /* Generate domain masks and attrs */
765 ndoms
= generate_sched_domains(&doms
, &attr
);
767 /* Have scheduler rebuild the domains */
768 partition_sched_domains(ndoms
, doms
, attr
);
772 #else /* !CONFIG_SMP */
773 static void rebuild_sched_domains_locked(void)
776 #endif /* CONFIG_SMP */
778 void rebuild_sched_domains(void)
780 mutex_lock(&cpuset_mutex
);
781 rebuild_sched_domains_locked();
782 mutex_unlock(&cpuset_mutex
);
786 * effective_cpumask_cpuset - return nearest ancestor with non-empty cpus
787 * @cs: the cpuset in interest
789 * A cpuset's effective cpumask is the cpumask of the nearest ancestor
790 * with non-empty cpus. We use effective cpumask whenever:
791 * - we update tasks' cpus_allowed. (they take on the ancestor's cpumask
792 * if the cpuset they reside in has no cpus)
793 * - we want to retrieve task_cs(tsk)'s cpus_allowed.
795 * Called with cpuset_mutex held. cpuset_cpus_allowed_fallback() is an
796 * exception. See comments there.
798 static struct cpuset
*effective_cpumask_cpuset(struct cpuset
*cs
)
800 while (cpumask_empty(cs
->cpus_allowed
))
806 * effective_nodemask_cpuset - return nearest ancestor with non-empty mems
807 * @cs: the cpuset in interest
809 * A cpuset's effective nodemask is the nodemask of the nearest ancestor
810 * with non-empty memss. We use effective nodemask whenever:
811 * - we update tasks' mems_allowed. (they take on the ancestor's nodemask
812 * if the cpuset they reside in has no mems)
813 * - we want to retrieve task_cs(tsk)'s mems_allowed.
815 * Called with cpuset_mutex held.
817 static struct cpuset
*effective_nodemask_cpuset(struct cpuset
*cs
)
819 while (nodes_empty(cs
->mems_allowed
))
825 * cpuset_change_cpumask - make a task's cpus_allowed the same as its cpuset's
827 * @data: cpuset to @tsk belongs to
829 * Called by css_scan_tasks() for each task in a cgroup whose cpus_allowed
830 * mask needs to be changed.
832 * We don't need to re-check for the cgroup/cpuset membership, since we're
833 * holding cpuset_mutex at this point.
835 static void cpuset_change_cpumask(struct task_struct
*tsk
, void *data
)
837 struct cpuset
*cs
= data
;
838 struct cpuset
*cpus_cs
= effective_cpumask_cpuset(cs
);
840 set_cpus_allowed_ptr(tsk
, cpus_cs
->cpus_allowed
);
844 * update_tasks_cpumask - Update the cpumasks of tasks in the cpuset.
845 * @cs: the cpuset in which each task's cpus_allowed mask needs to be changed
846 * @heap: if NULL, defer allocating heap memory to css_scan_tasks()
848 * Called with cpuset_mutex held
850 * The css_scan_tasks() function will scan all the tasks in a cgroup,
851 * calling callback functions for each.
853 * No return value. It's guaranteed that css_scan_tasks() always returns 0
856 static void update_tasks_cpumask(struct cpuset
*cs
, struct ptr_heap
*heap
)
858 css_scan_tasks(&cs
->css
, NULL
, cpuset_change_cpumask
, cs
, heap
);
862 * update_tasks_cpumask_hier - Update the cpumasks of tasks in the hierarchy.
863 * @root_cs: the root cpuset of the hierarchy
864 * @update_root: update root cpuset or not?
865 * @heap: the heap used by css_scan_tasks()
867 * This will update cpumasks of tasks in @root_cs and all other empty cpusets
868 * which take on cpumask of @root_cs.
870 * Called with cpuset_mutex held
872 static void update_tasks_cpumask_hier(struct cpuset
*root_cs
,
873 bool update_root
, struct ptr_heap
*heap
)
876 struct cgroup_subsys_state
*pos_css
;
879 update_tasks_cpumask(root_cs
, heap
);
882 cpuset_for_each_descendant_pre(cp
, pos_css
, root_cs
) {
883 /* skip the whole subtree if @cp have some CPU */
884 if (!cpumask_empty(cp
->cpus_allowed
)) {
885 pos_css
= css_rightmost_descendant(pos_css
);
888 if (!css_tryget(&cp
->css
))
892 update_tasks_cpumask(cp
, heap
);
901 * update_cpumask - update the cpus_allowed mask of a cpuset and all tasks in it
902 * @cs: the cpuset to consider
903 * @buf: buffer of cpu numbers written to this cpuset
905 static int update_cpumask(struct cpuset
*cs
, struct cpuset
*trialcs
,
908 struct ptr_heap heap
;
910 int is_load_balanced
;
912 /* top_cpuset.cpus_allowed tracks cpu_online_mask; it's read-only */
913 if (cs
== &top_cpuset
)
917 * An empty cpus_allowed is ok only if the cpuset has no tasks.
918 * Since cpulist_parse() fails on an empty mask, we special case
919 * that parsing. The validate_change() call ensures that cpusets
920 * with tasks have cpus.
923 cpumask_clear(trialcs
->cpus_allowed
);
925 retval
= cpulist_parse(buf
, trialcs
->cpus_allowed
);
929 if (!cpumask_subset(trialcs
->cpus_allowed
, cpu_active_mask
))
933 /* Nothing to do if the cpus didn't change */
934 if (cpumask_equal(cs
->cpus_allowed
, trialcs
->cpus_allowed
))
937 retval
= validate_change(cs
, trialcs
);
941 retval
= heap_init(&heap
, PAGE_SIZE
, GFP_KERNEL
, NULL
);
945 is_load_balanced
= is_sched_load_balance(trialcs
);
947 mutex_lock(&callback_mutex
);
948 cpumask_copy(cs
->cpus_allowed
, trialcs
->cpus_allowed
);
949 mutex_unlock(&callback_mutex
);
951 update_tasks_cpumask_hier(cs
, true, &heap
);
955 if (is_load_balanced
)
956 rebuild_sched_domains_locked();
963 * Migrate memory region from one set of nodes to another.
965 * Temporarilly set tasks mems_allowed to target nodes of migration,
966 * so that the migration code can allocate pages on these nodes.
968 * Call holding cpuset_mutex, so current's cpuset won't change
969 * during this call, as manage_mutex holds off any cpuset_attach()
970 * calls. Therefore we don't need to take task_lock around the
971 * call to guarantee_online_mems(), as we know no one is changing
974 * While the mm_struct we are migrating is typically from some
975 * other task, the task_struct mems_allowed that we are hacking
976 * is for our current task, which must allocate new pages for that
977 * migrating memory region.
980 static void cpuset_migrate_mm(struct mm_struct
*mm
, const nodemask_t
*from
,
981 const nodemask_t
*to
)
983 struct task_struct
*tsk
= current
;
984 struct cpuset
*mems_cs
;
986 tsk
->mems_allowed
= *to
;
988 do_migrate_pages(mm
, from
, to
, MPOL_MF_MOVE_ALL
);
990 mems_cs
= effective_nodemask_cpuset(task_cs(tsk
));
991 guarantee_online_mems(mems_cs
, &tsk
->mems_allowed
);
995 * cpuset_change_task_nodemask - change task's mems_allowed and mempolicy
996 * @tsk: the task to change
997 * @newmems: new nodes that the task will be set
999 * In order to avoid seeing no nodes if the old and new nodes are disjoint,
1000 * we structure updates as setting all new allowed nodes, then clearing newly
1003 static void cpuset_change_task_nodemask(struct task_struct
*tsk
,
1004 nodemask_t
*newmems
)
1009 * Allow tasks that have access to memory reserves because they have
1010 * been OOM killed to get memory anywhere.
1012 if (unlikely(test_thread_flag(TIF_MEMDIE
)))
1014 if (current
->flags
& PF_EXITING
) /* Let dying task have memory */
1019 * Determine if a loop is necessary if another thread is doing
1020 * get_mems_allowed(). If at least one node remains unchanged and
1021 * tsk does not have a mempolicy, then an empty nodemask will not be
1022 * possible when mems_allowed is larger than a word.
1024 need_loop
= task_has_mempolicy(tsk
) ||
1025 !nodes_intersects(*newmems
, tsk
->mems_allowed
);
1028 write_seqcount_begin(&tsk
->mems_allowed_seq
);
1030 nodes_or(tsk
->mems_allowed
, tsk
->mems_allowed
, *newmems
);
1031 mpol_rebind_task(tsk
, newmems
, MPOL_REBIND_STEP1
);
1033 mpol_rebind_task(tsk
, newmems
, MPOL_REBIND_STEP2
);
1034 tsk
->mems_allowed
= *newmems
;
1037 write_seqcount_end(&tsk
->mems_allowed_seq
);
1042 struct cpuset_change_nodemask_arg
{
1044 nodemask_t
*newmems
;
1048 * Update task's mems_allowed and rebind its mempolicy and vmas' mempolicy
1049 * of it to cpuset's new mems_allowed, and migrate pages to new nodes if
1050 * memory_migrate flag is set. Called with cpuset_mutex held.
1052 static void cpuset_change_nodemask(struct task_struct
*p
, void *data
)
1054 struct cpuset_change_nodemask_arg
*arg
= data
;
1055 struct cpuset
*cs
= arg
->cs
;
1056 struct mm_struct
*mm
;
1059 cpuset_change_task_nodemask(p
, arg
->newmems
);
1061 mm
= get_task_mm(p
);
1065 migrate
= is_memory_migrate(cs
);
1067 mpol_rebind_mm(mm
, &cs
->mems_allowed
);
1069 cpuset_migrate_mm(mm
, &cs
->old_mems_allowed
, arg
->newmems
);
1073 static void *cpuset_being_rebound
;
1076 * update_tasks_nodemask - Update the nodemasks of tasks in the cpuset.
1077 * @cs: the cpuset in which each task's mems_allowed mask needs to be changed
1078 * @heap: if NULL, defer allocating heap memory to css_scan_tasks()
1080 * Called with cpuset_mutex held. No return value. It's guaranteed that
1081 * css_scan_tasks() always returns 0 if @heap != NULL.
1083 static void update_tasks_nodemask(struct cpuset
*cs
, struct ptr_heap
*heap
)
1085 static nodemask_t newmems
; /* protected by cpuset_mutex */
1086 struct cpuset
*mems_cs
= effective_nodemask_cpuset(cs
);
1087 struct cpuset_change_nodemask_arg arg
= { .cs
= cs
,
1088 .newmems
= &newmems
};
1090 cpuset_being_rebound
= cs
; /* causes mpol_dup() rebind */
1092 guarantee_online_mems(mems_cs
, &newmems
);
1095 * The mpol_rebind_mm() call takes mmap_sem, which we couldn't
1096 * take while holding tasklist_lock. Forks can happen - the
1097 * mpol_dup() cpuset_being_rebound check will catch such forks,
1098 * and rebind their vma mempolicies too. Because we still hold
1099 * the global cpuset_mutex, we know that no other rebind effort
1100 * will be contending for the global variable cpuset_being_rebound.
1101 * It's ok if we rebind the same mm twice; mpol_rebind_mm()
1102 * is idempotent. Also migrate pages in each mm to new nodes.
1104 css_scan_tasks(&cs
->css
, NULL
, cpuset_change_nodemask
, &arg
, heap
);
1107 * All the tasks' nodemasks have been updated, update
1108 * cs->old_mems_allowed.
1110 cs
->old_mems_allowed
= newmems
;
1112 /* We're done rebinding vmas to this cpuset's new mems_allowed. */
1113 cpuset_being_rebound
= NULL
;
1117 * update_tasks_nodemask_hier - Update the nodemasks of tasks in the hierarchy.
1118 * @cs: the root cpuset of the hierarchy
1119 * @update_root: update the root cpuset or not?
1120 * @heap: the heap used by css_scan_tasks()
1122 * This will update nodemasks of tasks in @root_cs and all other empty cpusets
1123 * which take on nodemask of @root_cs.
1125 * Called with cpuset_mutex held
1127 static void update_tasks_nodemask_hier(struct cpuset
*root_cs
,
1128 bool update_root
, struct ptr_heap
*heap
)
1131 struct cgroup_subsys_state
*pos_css
;
1134 update_tasks_nodemask(root_cs
, heap
);
1137 cpuset_for_each_descendant_pre(cp
, pos_css
, root_cs
) {
1138 /* skip the whole subtree if @cp have some CPU */
1139 if (!nodes_empty(cp
->mems_allowed
)) {
1140 pos_css
= css_rightmost_descendant(pos_css
);
1143 if (!css_tryget(&cp
->css
))
1147 update_tasks_nodemask(cp
, heap
);
1156 * Handle user request to change the 'mems' memory placement
1157 * of a cpuset. Needs to validate the request, update the
1158 * cpusets mems_allowed, and for each task in the cpuset,
1159 * update mems_allowed and rebind task's mempolicy and any vma
1160 * mempolicies and if the cpuset is marked 'memory_migrate',
1161 * migrate the tasks pages to the new memory.
1163 * Call with cpuset_mutex held. May take callback_mutex during call.
1164 * Will take tasklist_lock, scan tasklist for tasks in cpuset cs,
1165 * lock each such tasks mm->mmap_sem, scan its vma's and rebind
1166 * their mempolicies to the cpusets new mems_allowed.
1168 static int update_nodemask(struct cpuset
*cs
, struct cpuset
*trialcs
,
1172 struct ptr_heap heap
;
1175 * top_cpuset.mems_allowed tracks node_stats[N_MEMORY];
1178 if (cs
== &top_cpuset
) {
1184 * An empty mems_allowed is ok iff there are no tasks in the cpuset.
1185 * Since nodelist_parse() fails on an empty mask, we special case
1186 * that parsing. The validate_change() call ensures that cpusets
1187 * with tasks have memory.
1190 nodes_clear(trialcs
->mems_allowed
);
1192 retval
= nodelist_parse(buf
, trialcs
->mems_allowed
);
1196 if (!nodes_subset(trialcs
->mems_allowed
,
1197 node_states
[N_MEMORY
])) {
1203 if (nodes_equal(cs
->mems_allowed
, trialcs
->mems_allowed
)) {
1204 retval
= 0; /* Too easy - nothing to do */
1207 retval
= validate_change(cs
, trialcs
);
1211 retval
= heap_init(&heap
, PAGE_SIZE
, GFP_KERNEL
, NULL
);
1215 mutex_lock(&callback_mutex
);
1216 cs
->mems_allowed
= trialcs
->mems_allowed
;
1217 mutex_unlock(&callback_mutex
);
1219 update_tasks_nodemask_hier(cs
, true, &heap
);
1226 int current_cpuset_is_being_rebound(void)
1228 return task_cs(current
) == cpuset_being_rebound
;
1231 static int update_relax_domain_level(struct cpuset
*cs
, s64 val
)
1234 if (val
< -1 || val
>= sched_domain_level_max
)
1238 if (val
!= cs
->relax_domain_level
) {
1239 cs
->relax_domain_level
= val
;
1240 if (!cpumask_empty(cs
->cpus_allowed
) &&
1241 is_sched_load_balance(cs
))
1242 rebuild_sched_domains_locked();
1249 * cpuset_change_flag - make a task's spread flags the same as its cpuset's
1250 * @tsk: task to be updated
1251 * @data: cpuset to @tsk belongs to
1253 * Called by css_scan_tasks() for each task in a cgroup.
1255 * We don't need to re-check for the cgroup/cpuset membership, since we're
1256 * holding cpuset_mutex at this point.
1258 static void cpuset_change_flag(struct task_struct
*tsk
, void *data
)
1260 struct cpuset
*cs
= data
;
1262 cpuset_update_task_spread_flag(cs
, tsk
);
1266 * update_tasks_flags - update the spread flags of tasks in the cpuset.
1267 * @cs: the cpuset in which each task's spread flags needs to be changed
1268 * @heap: if NULL, defer allocating heap memory to css_scan_tasks()
1270 * Called with cpuset_mutex held
1272 * The css_scan_tasks() function will scan all the tasks in a cgroup,
1273 * calling callback functions for each.
1275 * No return value. It's guaranteed that css_scan_tasks() always returns 0
1278 static void update_tasks_flags(struct cpuset
*cs
, struct ptr_heap
*heap
)
1280 css_scan_tasks(&cs
->css
, NULL
, cpuset_change_flag
, cs
, heap
);
1284 * update_flag - read a 0 or a 1 in a file and update associated flag
1285 * bit: the bit to update (see cpuset_flagbits_t)
1286 * cs: the cpuset to update
1287 * turning_on: whether the flag is being set or cleared
1289 * Call with cpuset_mutex held.
1292 static int update_flag(cpuset_flagbits_t bit
, struct cpuset
*cs
,
1295 struct cpuset
*trialcs
;
1296 int balance_flag_changed
;
1297 int spread_flag_changed
;
1298 struct ptr_heap heap
;
1301 trialcs
= alloc_trial_cpuset(cs
);
1306 set_bit(bit
, &trialcs
->flags
);
1308 clear_bit(bit
, &trialcs
->flags
);
1310 err
= validate_change(cs
, trialcs
);
1314 err
= heap_init(&heap
, PAGE_SIZE
, GFP_KERNEL
, NULL
);
1318 balance_flag_changed
= (is_sched_load_balance(cs
) !=
1319 is_sched_load_balance(trialcs
));
1321 spread_flag_changed
= ((is_spread_slab(cs
) != is_spread_slab(trialcs
))
1322 || (is_spread_page(cs
) != is_spread_page(trialcs
)));
1324 mutex_lock(&callback_mutex
);
1325 cs
->flags
= trialcs
->flags
;
1326 mutex_unlock(&callback_mutex
);
1328 if (!cpumask_empty(trialcs
->cpus_allowed
) && balance_flag_changed
)
1329 rebuild_sched_domains_locked();
1331 if (spread_flag_changed
)
1332 update_tasks_flags(cs
, &heap
);
1335 free_trial_cpuset(trialcs
);
1340 * Frequency meter - How fast is some event occurring?
1342 * These routines manage a digitally filtered, constant time based,
1343 * event frequency meter. There are four routines:
1344 * fmeter_init() - initialize a frequency meter.
1345 * fmeter_markevent() - called each time the event happens.
1346 * fmeter_getrate() - returns the recent rate of such events.
1347 * fmeter_update() - internal routine used to update fmeter.
1349 * A common data structure is passed to each of these routines,
1350 * which is used to keep track of the state required to manage the
1351 * frequency meter and its digital filter.
1353 * The filter works on the number of events marked per unit time.
1354 * The filter is single-pole low-pass recursive (IIR). The time unit
1355 * is 1 second. Arithmetic is done using 32-bit integers scaled to
1356 * simulate 3 decimal digits of precision (multiplied by 1000).
1358 * With an FM_COEF of 933, and a time base of 1 second, the filter
1359 * has a half-life of 10 seconds, meaning that if the events quit
1360 * happening, then the rate returned from the fmeter_getrate()
1361 * will be cut in half each 10 seconds, until it converges to zero.
1363 * It is not worth doing a real infinitely recursive filter. If more
1364 * than FM_MAXTICKS ticks have elapsed since the last filter event,
1365 * just compute FM_MAXTICKS ticks worth, by which point the level
1368 * Limit the count of unprocessed events to FM_MAXCNT, so as to avoid
1369 * arithmetic overflow in the fmeter_update() routine.
1371 * Given the simple 32 bit integer arithmetic used, this meter works
1372 * best for reporting rates between one per millisecond (msec) and
1373 * one per 32 (approx) seconds. At constant rates faster than one
1374 * per msec it maxes out at values just under 1,000,000. At constant
1375 * rates between one per msec, and one per second it will stabilize
1376 * to a value N*1000, where N is the rate of events per second.
1377 * At constant rates between one per second and one per 32 seconds,
1378 * it will be choppy, moving up on the seconds that have an event,
1379 * and then decaying until the next event. At rates slower than
1380 * about one in 32 seconds, it decays all the way back to zero between
1384 #define FM_COEF 933 /* coefficient for half-life of 10 secs */
1385 #define FM_MAXTICKS ((time_t)99) /* useless computing more ticks than this */
1386 #define FM_MAXCNT 1000000 /* limit cnt to avoid overflow */
1387 #define FM_SCALE 1000 /* faux fixed point scale */
1389 /* Initialize a frequency meter */
1390 static void fmeter_init(struct fmeter
*fmp
)
1395 spin_lock_init(&fmp
->lock
);
1398 /* Internal meter update - process cnt events and update value */
1399 static void fmeter_update(struct fmeter
*fmp
)
1401 time_t now
= get_seconds();
1402 time_t ticks
= now
- fmp
->time
;
1407 ticks
= min(FM_MAXTICKS
, ticks
);
1409 fmp
->val
= (FM_COEF
* fmp
->val
) / FM_SCALE
;
1412 fmp
->val
+= ((FM_SCALE
- FM_COEF
) * fmp
->cnt
) / FM_SCALE
;
1416 /* Process any previous ticks, then bump cnt by one (times scale). */
1417 static void fmeter_markevent(struct fmeter
*fmp
)
1419 spin_lock(&fmp
->lock
);
1421 fmp
->cnt
= min(FM_MAXCNT
, fmp
->cnt
+ FM_SCALE
);
1422 spin_unlock(&fmp
->lock
);
1425 /* Process any previous ticks, then return current value. */
1426 static int fmeter_getrate(struct fmeter
*fmp
)
1430 spin_lock(&fmp
->lock
);
1433 spin_unlock(&fmp
->lock
);
1437 /* Called by cgroups to determine if a cpuset is usable; cpuset_mutex held */
1438 static int cpuset_can_attach(struct cgroup_subsys_state
*css
,
1439 struct cgroup_taskset
*tset
)
1441 struct cpuset
*cs
= css_cs(css
);
1442 struct task_struct
*task
;
1445 mutex_lock(&cpuset_mutex
);
1448 * We allow to move tasks into an empty cpuset if sane_behavior
1452 if (!cgroup_sane_behavior(css
->cgroup
) &&
1453 (cpumask_empty(cs
->cpus_allowed
) || nodes_empty(cs
->mems_allowed
)))
1456 cgroup_taskset_for_each(task
, css
, tset
) {
1458 * Kthreads which disallow setaffinity shouldn't be moved
1459 * to a new cpuset; we don't want to change their cpu
1460 * affinity and isolating such threads by their set of
1461 * allowed nodes is unnecessary. Thus, cpusets are not
1462 * applicable for such threads. This prevents checking for
1463 * success of set_cpus_allowed_ptr() on all attached tasks
1464 * before cpus_allowed may be changed.
1467 if (task
->flags
& PF_NO_SETAFFINITY
)
1469 ret
= security_task_setscheduler(task
);
1475 * Mark attach is in progress. This makes validate_change() fail
1476 * changes which zero cpus/mems_allowed.
1478 cs
->attach_in_progress
++;
1481 mutex_unlock(&cpuset_mutex
);
1485 static void cpuset_cancel_attach(struct cgroup_subsys_state
*css
,
1486 struct cgroup_taskset
*tset
)
1488 mutex_lock(&cpuset_mutex
);
1489 css_cs(css
)->attach_in_progress
--;
1490 mutex_unlock(&cpuset_mutex
);
1494 * Protected by cpuset_mutex. cpus_attach is used only by cpuset_attach()
1495 * but we can't allocate it dynamically there. Define it global and
1496 * allocate from cpuset_init().
1498 static cpumask_var_t cpus_attach
;
1500 static void cpuset_attach(struct cgroup_subsys_state
*css
,
1501 struct cgroup_taskset
*tset
)
1503 /* static buf protected by cpuset_mutex */
1504 static nodemask_t cpuset_attach_nodemask_to
;
1505 struct mm_struct
*mm
;
1506 struct task_struct
*task
;
1507 struct task_struct
*leader
= cgroup_taskset_first(tset
);
1508 struct cgroup_subsys_state
*oldcss
= cgroup_taskset_cur_css(tset
,
1510 struct cpuset
*cs
= css_cs(css
);
1511 struct cpuset
*oldcs
= css_cs(oldcss
);
1512 struct cpuset
*cpus_cs
= effective_cpumask_cpuset(cs
);
1513 struct cpuset
*mems_cs
= effective_nodemask_cpuset(cs
);
1515 mutex_lock(&cpuset_mutex
);
1517 /* prepare for attach */
1518 if (cs
== &top_cpuset
)
1519 cpumask_copy(cpus_attach
, cpu_possible_mask
);
1521 guarantee_online_cpus(cpus_cs
, cpus_attach
);
1523 guarantee_online_mems(mems_cs
, &cpuset_attach_nodemask_to
);
1525 cgroup_taskset_for_each(task
, css
, tset
) {
1527 * can_attach beforehand should guarantee that this doesn't
1528 * fail. TODO: have a better way to handle failure here
1530 WARN_ON_ONCE(set_cpus_allowed_ptr(task
, cpus_attach
));
1532 cpuset_change_task_nodemask(task
, &cpuset_attach_nodemask_to
);
1533 cpuset_update_task_spread_flag(cs
, task
);
1537 * Change mm, possibly for multiple threads in a threadgroup. This is
1538 * expensive and may sleep.
1540 cpuset_attach_nodemask_to
= cs
->mems_allowed
;
1541 mm
= get_task_mm(leader
);
1543 struct cpuset
*mems_oldcs
= effective_nodemask_cpuset(oldcs
);
1545 mpol_rebind_mm(mm
, &cpuset_attach_nodemask_to
);
1548 * old_mems_allowed is the same with mems_allowed here, except
1549 * if this task is being moved automatically due to hotplug.
1550 * In that case @mems_allowed has been updated and is empty,
1551 * so @old_mems_allowed is the right nodesets that we migrate
1554 if (is_memory_migrate(cs
)) {
1555 cpuset_migrate_mm(mm
, &mems_oldcs
->old_mems_allowed
,
1556 &cpuset_attach_nodemask_to
);
1561 cs
->old_mems_allowed
= cpuset_attach_nodemask_to
;
1563 cs
->attach_in_progress
--;
1564 if (!cs
->attach_in_progress
)
1565 wake_up(&cpuset_attach_wq
);
1567 mutex_unlock(&cpuset_mutex
);
1570 /* The various types of files and directories in a cpuset file system */
1573 FILE_MEMORY_MIGRATE
,
1579 FILE_SCHED_LOAD_BALANCE
,
1580 FILE_SCHED_RELAX_DOMAIN_LEVEL
,
1581 FILE_MEMORY_PRESSURE_ENABLED
,
1582 FILE_MEMORY_PRESSURE
,
1585 } cpuset_filetype_t
;
1587 static int cpuset_write_u64(struct cgroup_subsys_state
*css
, struct cftype
*cft
,
1590 struct cpuset
*cs
= css_cs(css
);
1591 cpuset_filetype_t type
= cft
->private;
1592 int retval
= -ENODEV
;
1594 mutex_lock(&cpuset_mutex
);
1595 if (!is_cpuset_online(cs
))
1599 case FILE_CPU_EXCLUSIVE
:
1600 retval
= update_flag(CS_CPU_EXCLUSIVE
, cs
, val
);
1602 case FILE_MEM_EXCLUSIVE
:
1603 retval
= update_flag(CS_MEM_EXCLUSIVE
, cs
, val
);
1605 case FILE_MEM_HARDWALL
:
1606 retval
= update_flag(CS_MEM_HARDWALL
, cs
, val
);
1608 case FILE_SCHED_LOAD_BALANCE
:
1609 retval
= update_flag(CS_SCHED_LOAD_BALANCE
, cs
, val
);
1611 case FILE_MEMORY_MIGRATE
:
1612 retval
= update_flag(CS_MEMORY_MIGRATE
, cs
, val
);
1614 case FILE_MEMORY_PRESSURE_ENABLED
:
1615 cpuset_memory_pressure_enabled
= !!val
;
1617 case FILE_MEMORY_PRESSURE
:
1620 case FILE_SPREAD_PAGE
:
1621 retval
= update_flag(CS_SPREAD_PAGE
, cs
, val
);
1623 case FILE_SPREAD_SLAB
:
1624 retval
= update_flag(CS_SPREAD_SLAB
, cs
, val
);
1631 mutex_unlock(&cpuset_mutex
);
1635 static int cpuset_write_s64(struct cgroup_subsys_state
*css
, struct cftype
*cft
,
1638 struct cpuset
*cs
= css_cs(css
);
1639 cpuset_filetype_t type
= cft
->private;
1640 int retval
= -ENODEV
;
1642 mutex_lock(&cpuset_mutex
);
1643 if (!is_cpuset_online(cs
))
1647 case FILE_SCHED_RELAX_DOMAIN_LEVEL
:
1648 retval
= update_relax_domain_level(cs
, val
);
1655 mutex_unlock(&cpuset_mutex
);
1660 * Common handling for a write to a "cpus" or "mems" file.
1662 static int cpuset_write_resmask(struct cgroup_subsys_state
*css
,
1663 struct cftype
*cft
, const char *buf
)
1665 struct cpuset
*cs
= css_cs(css
);
1666 struct cpuset
*trialcs
;
1667 int retval
= -ENODEV
;
1670 * CPU or memory hotunplug may leave @cs w/o any execution
1671 * resources, in which case the hotplug code asynchronously updates
1672 * configuration and transfers all tasks to the nearest ancestor
1673 * which can execute.
1675 * As writes to "cpus" or "mems" may restore @cs's execution
1676 * resources, wait for the previously scheduled operations before
1677 * proceeding, so that we don't end up keep removing tasks added
1678 * after execution capability is restored.
1680 flush_work(&cpuset_hotplug_work
);
1682 mutex_lock(&cpuset_mutex
);
1683 if (!is_cpuset_online(cs
))
1686 trialcs
= alloc_trial_cpuset(cs
);
1692 switch (cft
->private) {
1694 retval
= update_cpumask(cs
, trialcs
, buf
);
1697 retval
= update_nodemask(cs
, trialcs
, buf
);
1704 free_trial_cpuset(trialcs
);
1706 mutex_unlock(&cpuset_mutex
);
1711 * These ascii lists should be read in a single call, by using a user
1712 * buffer large enough to hold the entire map. If read in smaller
1713 * chunks, there is no guarantee of atomicity. Since the display format
1714 * used, list of ranges of sequential numbers, is variable length,
1715 * and since these maps can change value dynamically, one could read
1716 * gibberish by doing partial reads while a list was changing.
1717 * A single large read to a buffer that crosses a page boundary is
1718 * ok, because the result being copied to user land is not recomputed
1719 * across a page fault.
1722 static size_t cpuset_sprintf_cpulist(char *page
, struct cpuset
*cs
)
1726 mutex_lock(&callback_mutex
);
1727 count
= cpulist_scnprintf(page
, PAGE_SIZE
, cs
->cpus_allowed
);
1728 mutex_unlock(&callback_mutex
);
1733 static size_t cpuset_sprintf_memlist(char *page
, struct cpuset
*cs
)
1737 mutex_lock(&callback_mutex
);
1738 count
= nodelist_scnprintf(page
, PAGE_SIZE
, cs
->mems_allowed
);
1739 mutex_unlock(&callback_mutex
);
1744 static ssize_t
cpuset_common_file_read(struct cgroup_subsys_state
*css
,
1745 struct cftype
*cft
, struct file
*file
,
1746 char __user
*buf
, size_t nbytes
,
1749 struct cpuset
*cs
= css_cs(css
);
1750 cpuset_filetype_t type
= cft
->private;
1755 if (!(page
= (char *)__get_free_page(GFP_TEMPORARY
)))
1762 s
+= cpuset_sprintf_cpulist(s
, cs
);
1765 s
+= cpuset_sprintf_memlist(s
, cs
);
1773 retval
= simple_read_from_buffer(buf
, nbytes
, ppos
, page
, s
- page
);
1775 free_page((unsigned long)page
);
1779 static u64
cpuset_read_u64(struct cgroup_subsys_state
*css
, struct cftype
*cft
)
1781 struct cpuset
*cs
= css_cs(css
);
1782 cpuset_filetype_t type
= cft
->private;
1784 case FILE_CPU_EXCLUSIVE
:
1785 return is_cpu_exclusive(cs
);
1786 case FILE_MEM_EXCLUSIVE
:
1787 return is_mem_exclusive(cs
);
1788 case FILE_MEM_HARDWALL
:
1789 return is_mem_hardwall(cs
);
1790 case FILE_SCHED_LOAD_BALANCE
:
1791 return is_sched_load_balance(cs
);
1792 case FILE_MEMORY_MIGRATE
:
1793 return is_memory_migrate(cs
);
1794 case FILE_MEMORY_PRESSURE_ENABLED
:
1795 return cpuset_memory_pressure_enabled
;
1796 case FILE_MEMORY_PRESSURE
:
1797 return fmeter_getrate(&cs
->fmeter
);
1798 case FILE_SPREAD_PAGE
:
1799 return is_spread_page(cs
);
1800 case FILE_SPREAD_SLAB
:
1801 return is_spread_slab(cs
);
1806 /* Unreachable but makes gcc happy */
1810 static s64
cpuset_read_s64(struct cgroup_subsys_state
*css
, struct cftype
*cft
)
1812 struct cpuset
*cs
= css_cs(css
);
1813 cpuset_filetype_t type
= cft
->private;
1815 case FILE_SCHED_RELAX_DOMAIN_LEVEL
:
1816 return cs
->relax_domain_level
;
1821 /* Unrechable but makes gcc happy */
1827 * for the common functions, 'private' gives the type of file
1830 static struct cftype files
[] = {
1833 .read
= cpuset_common_file_read
,
1834 .write_string
= cpuset_write_resmask
,
1835 .max_write_len
= (100U + 6 * NR_CPUS
),
1836 .private = FILE_CPULIST
,
1841 .read
= cpuset_common_file_read
,
1842 .write_string
= cpuset_write_resmask
,
1843 .max_write_len
= (100U + 6 * MAX_NUMNODES
),
1844 .private = FILE_MEMLIST
,
1848 .name
= "cpu_exclusive",
1849 .read_u64
= cpuset_read_u64
,
1850 .write_u64
= cpuset_write_u64
,
1851 .private = FILE_CPU_EXCLUSIVE
,
1855 .name
= "mem_exclusive",
1856 .read_u64
= cpuset_read_u64
,
1857 .write_u64
= cpuset_write_u64
,
1858 .private = FILE_MEM_EXCLUSIVE
,
1862 .name
= "mem_hardwall",
1863 .read_u64
= cpuset_read_u64
,
1864 .write_u64
= cpuset_write_u64
,
1865 .private = FILE_MEM_HARDWALL
,
1869 .name
= "sched_load_balance",
1870 .read_u64
= cpuset_read_u64
,
1871 .write_u64
= cpuset_write_u64
,
1872 .private = FILE_SCHED_LOAD_BALANCE
,
1876 .name
= "sched_relax_domain_level",
1877 .read_s64
= cpuset_read_s64
,
1878 .write_s64
= cpuset_write_s64
,
1879 .private = FILE_SCHED_RELAX_DOMAIN_LEVEL
,
1883 .name
= "memory_migrate",
1884 .read_u64
= cpuset_read_u64
,
1885 .write_u64
= cpuset_write_u64
,
1886 .private = FILE_MEMORY_MIGRATE
,
1890 .name
= "memory_pressure",
1891 .read_u64
= cpuset_read_u64
,
1892 .write_u64
= cpuset_write_u64
,
1893 .private = FILE_MEMORY_PRESSURE
,
1898 .name
= "memory_spread_page",
1899 .read_u64
= cpuset_read_u64
,
1900 .write_u64
= cpuset_write_u64
,
1901 .private = FILE_SPREAD_PAGE
,
1905 .name
= "memory_spread_slab",
1906 .read_u64
= cpuset_read_u64
,
1907 .write_u64
= cpuset_write_u64
,
1908 .private = FILE_SPREAD_SLAB
,
1912 .name
= "memory_pressure_enabled",
1913 .flags
= CFTYPE_ONLY_ON_ROOT
,
1914 .read_u64
= cpuset_read_u64
,
1915 .write_u64
= cpuset_write_u64
,
1916 .private = FILE_MEMORY_PRESSURE_ENABLED
,
1923 * cpuset_css_alloc - allocate a cpuset css
1924 * cgrp: control group that the new cpuset will be part of
1927 static struct cgroup_subsys_state
*
1928 cpuset_css_alloc(struct cgroup_subsys_state
*parent_css
)
1933 return &top_cpuset
.css
;
1935 cs
= kzalloc(sizeof(*cs
), GFP_KERNEL
);
1937 return ERR_PTR(-ENOMEM
);
1938 if (!alloc_cpumask_var(&cs
->cpus_allowed
, GFP_KERNEL
)) {
1940 return ERR_PTR(-ENOMEM
);
1943 set_bit(CS_SCHED_LOAD_BALANCE
, &cs
->flags
);
1944 cpumask_clear(cs
->cpus_allowed
);
1945 nodes_clear(cs
->mems_allowed
);
1946 fmeter_init(&cs
->fmeter
);
1947 cs
->relax_domain_level
= -1;
1952 static int cpuset_css_online(struct cgroup_subsys_state
*css
)
1954 struct cpuset
*cs
= css_cs(css
);
1955 struct cpuset
*parent
= parent_cs(cs
);
1956 struct cpuset
*tmp_cs
;
1957 struct cgroup_subsys_state
*pos_css
;
1962 mutex_lock(&cpuset_mutex
);
1964 set_bit(CS_ONLINE
, &cs
->flags
);
1965 if (is_spread_page(parent
))
1966 set_bit(CS_SPREAD_PAGE
, &cs
->flags
);
1967 if (is_spread_slab(parent
))
1968 set_bit(CS_SPREAD_SLAB
, &cs
->flags
);
1970 number_of_cpusets
++;
1972 if (!test_bit(CGRP_CPUSET_CLONE_CHILDREN
, &css
->cgroup
->flags
))
1976 * Clone @parent's configuration if CGRP_CPUSET_CLONE_CHILDREN is
1977 * set. This flag handling is implemented in cgroup core for
1978 * histrical reasons - the flag may be specified during mount.
1980 * Currently, if any sibling cpusets have exclusive cpus or mem, we
1981 * refuse to clone the configuration - thereby refusing the task to
1982 * be entered, and as a result refusing the sys_unshare() or
1983 * clone() which initiated it. If this becomes a problem for some
1984 * users who wish to allow that scenario, then this could be
1985 * changed to grant parent->cpus_allowed-sibling_cpus_exclusive
1986 * (and likewise for mems) to the new cgroup.
1989 cpuset_for_each_child(tmp_cs
, pos_css
, parent
) {
1990 if (is_mem_exclusive(tmp_cs
) || is_cpu_exclusive(tmp_cs
)) {
1997 mutex_lock(&callback_mutex
);
1998 cs
->mems_allowed
= parent
->mems_allowed
;
1999 cpumask_copy(cs
->cpus_allowed
, parent
->cpus_allowed
);
2000 mutex_unlock(&callback_mutex
);
2002 mutex_unlock(&cpuset_mutex
);
2007 * If the cpuset being removed has its flag 'sched_load_balance'
2008 * enabled, then simulate turning sched_load_balance off, which
2009 * will call rebuild_sched_domains_locked().
2012 static void cpuset_css_offline(struct cgroup_subsys_state
*css
)
2014 struct cpuset
*cs
= css_cs(css
);
2016 mutex_lock(&cpuset_mutex
);
2018 if (is_sched_load_balance(cs
))
2019 update_flag(CS_SCHED_LOAD_BALANCE
, cs
, 0);
2021 number_of_cpusets
--;
2022 clear_bit(CS_ONLINE
, &cs
->flags
);
2024 mutex_unlock(&cpuset_mutex
);
2027 static void cpuset_css_free(struct cgroup_subsys_state
*css
)
2029 struct cpuset
*cs
= css_cs(css
);
2031 free_cpumask_var(cs
->cpus_allowed
);
2035 struct cgroup_subsys cpuset_subsys
= {
2037 .css_alloc
= cpuset_css_alloc
,
2038 .css_online
= cpuset_css_online
,
2039 .css_offline
= cpuset_css_offline
,
2040 .css_free
= cpuset_css_free
,
2041 .can_attach
= cpuset_can_attach
,
2042 .cancel_attach
= cpuset_cancel_attach
,
2043 .attach
= cpuset_attach
,
2044 .subsys_id
= cpuset_subsys_id
,
2045 .base_cftypes
= files
,
2050 * cpuset_init - initialize cpusets at system boot
2052 * Description: Initialize top_cpuset and the cpuset internal file system,
2055 int __init
cpuset_init(void)
2059 if (!alloc_cpumask_var(&top_cpuset
.cpus_allowed
, GFP_KERNEL
))
2062 cpumask_setall(top_cpuset
.cpus_allowed
);
2063 nodes_setall(top_cpuset
.mems_allowed
);
2065 fmeter_init(&top_cpuset
.fmeter
);
2066 set_bit(CS_SCHED_LOAD_BALANCE
, &top_cpuset
.flags
);
2067 top_cpuset
.relax_domain_level
= -1;
2069 err
= register_filesystem(&cpuset_fs_type
);
2073 if (!alloc_cpumask_var(&cpus_attach
, GFP_KERNEL
))
2076 number_of_cpusets
= 1;
2081 * If CPU and/or memory hotplug handlers, below, unplug any CPUs
2082 * or memory nodes, we need to walk over the cpuset hierarchy,
2083 * removing that CPU or node from all cpusets. If this removes the
2084 * last CPU or node from a cpuset, then move the tasks in the empty
2085 * cpuset to its next-highest non-empty parent.
2087 static void remove_tasks_in_empty_cpuset(struct cpuset
*cs
)
2089 struct cpuset
*parent
;
2092 * Find its next-highest non-empty parent, (top cpuset
2093 * has online cpus, so can't be empty).
2095 parent
= parent_cs(cs
);
2096 while (cpumask_empty(parent
->cpus_allowed
) ||
2097 nodes_empty(parent
->mems_allowed
))
2098 parent
= parent_cs(parent
);
2100 if (cgroup_transfer_tasks(parent
->css
.cgroup
, cs
->css
.cgroup
)) {
2102 printk(KERN_ERR
"cpuset: failed to transfer tasks out of empty cpuset %s\n",
2103 cgroup_name(cs
->css
.cgroup
));
2109 * cpuset_hotplug_update_tasks - update tasks in a cpuset for hotunplug
2110 * @cs: cpuset in interest
2112 * Compare @cs's cpu and mem masks against top_cpuset and if some have gone
2113 * offline, update @cs accordingly. If @cs ends up with no CPU or memory,
2114 * all its tasks are moved to the nearest ancestor with both resources.
2116 static void cpuset_hotplug_update_tasks(struct cpuset
*cs
)
2118 static cpumask_t off_cpus
;
2119 static nodemask_t off_mems
;
2121 bool sane
= cgroup_sane_behavior(cs
->css
.cgroup
);
2124 wait_event(cpuset_attach_wq
, cs
->attach_in_progress
== 0);
2126 mutex_lock(&cpuset_mutex
);
2129 * We have raced with task attaching. We wait until attaching
2130 * is finished, so we won't attach a task to an empty cpuset.
2132 if (cs
->attach_in_progress
) {
2133 mutex_unlock(&cpuset_mutex
);
2137 cpumask_andnot(&off_cpus
, cs
->cpus_allowed
, top_cpuset
.cpus_allowed
);
2138 nodes_andnot(off_mems
, cs
->mems_allowed
, top_cpuset
.mems_allowed
);
2140 mutex_lock(&callback_mutex
);
2141 cpumask_andnot(cs
->cpus_allowed
, cs
->cpus_allowed
, &off_cpus
);
2142 mutex_unlock(&callback_mutex
);
2145 * If sane_behavior flag is set, we need to update tasks' cpumask
2146 * for empty cpuset to take on ancestor's cpumask. Otherwise, don't
2147 * call update_tasks_cpumask() if the cpuset becomes empty, as
2148 * the tasks in it will be migrated to an ancestor.
2150 if ((sane
&& cpumask_empty(cs
->cpus_allowed
)) ||
2151 (!cpumask_empty(&off_cpus
) && !cpumask_empty(cs
->cpus_allowed
)))
2152 update_tasks_cpumask(cs
, NULL
);
2154 mutex_lock(&callback_mutex
);
2155 nodes_andnot(cs
->mems_allowed
, cs
->mems_allowed
, off_mems
);
2156 mutex_unlock(&callback_mutex
);
2159 * If sane_behavior flag is set, we need to update tasks' nodemask
2160 * for empty cpuset to take on ancestor's nodemask. Otherwise, don't
2161 * call update_tasks_nodemask() if the cpuset becomes empty, as
2162 * the tasks in it will be migratd to an ancestor.
2164 if ((sane
&& nodes_empty(cs
->mems_allowed
)) ||
2165 (!nodes_empty(off_mems
) && !nodes_empty(cs
->mems_allowed
)))
2166 update_tasks_nodemask(cs
, NULL
);
2168 is_empty
= cpumask_empty(cs
->cpus_allowed
) ||
2169 nodes_empty(cs
->mems_allowed
);
2171 mutex_unlock(&cpuset_mutex
);
2174 * If sane_behavior flag is set, we'll keep tasks in empty cpusets.
2176 * Otherwise move tasks to the nearest ancestor with execution
2177 * resources. This is full cgroup operation which will
2178 * also call back into cpuset. Should be done outside any lock.
2180 if (!sane
&& is_empty
)
2181 remove_tasks_in_empty_cpuset(cs
);
2185 * cpuset_hotplug_workfn - handle CPU/memory hotunplug for a cpuset
2187 * This function is called after either CPU or memory configuration has
2188 * changed and updates cpuset accordingly. The top_cpuset is always
2189 * synchronized to cpu_active_mask and N_MEMORY, which is necessary in
2190 * order to make cpusets transparent (of no affect) on systems that are
2191 * actively using CPU hotplug but making no active use of cpusets.
2193 * Non-root cpusets are only affected by offlining. If any CPUs or memory
2194 * nodes have been taken down, cpuset_hotplug_update_tasks() is invoked on
2197 * Note that CPU offlining during suspend is ignored. We don't modify
2198 * cpusets across suspend/resume cycles at all.
2200 static void cpuset_hotplug_workfn(struct work_struct
*work
)
2202 static cpumask_t new_cpus
;
2203 static nodemask_t new_mems
;
2204 bool cpus_updated
, mems_updated
;
2206 mutex_lock(&cpuset_mutex
);
2208 /* fetch the available cpus/mems and find out which changed how */
2209 cpumask_copy(&new_cpus
, cpu_active_mask
);
2210 new_mems
= node_states
[N_MEMORY
];
2212 cpus_updated
= !cpumask_equal(top_cpuset
.cpus_allowed
, &new_cpus
);
2213 mems_updated
= !nodes_equal(top_cpuset
.mems_allowed
, new_mems
);
2215 /* synchronize cpus_allowed to cpu_active_mask */
2217 mutex_lock(&callback_mutex
);
2218 cpumask_copy(top_cpuset
.cpus_allowed
, &new_cpus
);
2219 mutex_unlock(&callback_mutex
);
2220 /* we don't mess with cpumasks of tasks in top_cpuset */
2223 /* synchronize mems_allowed to N_MEMORY */
2225 mutex_lock(&callback_mutex
);
2226 top_cpuset
.mems_allowed
= new_mems
;
2227 mutex_unlock(&callback_mutex
);
2228 update_tasks_nodemask(&top_cpuset
, NULL
);
2231 mutex_unlock(&cpuset_mutex
);
2233 /* if cpus or mems changed, we need to propagate to descendants */
2234 if (cpus_updated
|| mems_updated
) {
2236 struct cgroup_subsys_state
*pos_css
;
2239 cpuset_for_each_descendant_pre(cs
, pos_css
, &top_cpuset
) {
2240 if (!css_tryget(&cs
->css
))
2244 cpuset_hotplug_update_tasks(cs
);
2252 /* rebuild sched domains if cpus_allowed has changed */
2254 rebuild_sched_domains();
2257 void cpuset_update_active_cpus(bool cpu_online
)
2260 * We're inside cpu hotplug critical region which usually nests
2261 * inside cgroup synchronization. Bounce actual hotplug processing
2262 * to a work item to avoid reverse locking order.
2264 * We still need to do partition_sched_domains() synchronously;
2265 * otherwise, the scheduler will get confused and put tasks to the
2266 * dead CPU. Fall back to the default single domain.
2267 * cpuset_hotplug_workfn() will rebuild it as necessary.
2269 partition_sched_domains(1, NULL
, NULL
);
2270 schedule_work(&cpuset_hotplug_work
);
2274 * Keep top_cpuset.mems_allowed tracking node_states[N_MEMORY].
2275 * Call this routine anytime after node_states[N_MEMORY] changes.
2276 * See cpuset_update_active_cpus() for CPU hotplug handling.
2278 static int cpuset_track_online_nodes(struct notifier_block
*self
,
2279 unsigned long action
, void *arg
)
2281 schedule_work(&cpuset_hotplug_work
);
2285 static struct notifier_block cpuset_track_online_nodes_nb
= {
2286 .notifier_call
= cpuset_track_online_nodes
,
2287 .priority
= 10, /* ??! */
2291 * cpuset_init_smp - initialize cpus_allowed
2293 * Description: Finish top cpuset after cpu, node maps are initialized
2295 void __init
cpuset_init_smp(void)
2297 cpumask_copy(top_cpuset
.cpus_allowed
, cpu_active_mask
);
2298 top_cpuset
.mems_allowed
= node_states
[N_MEMORY
];
2299 top_cpuset
.old_mems_allowed
= top_cpuset
.mems_allowed
;
2301 register_hotmemory_notifier(&cpuset_track_online_nodes_nb
);
2305 * cpuset_cpus_allowed - return cpus_allowed mask from a tasks cpuset.
2306 * @tsk: pointer to task_struct from which to obtain cpuset->cpus_allowed.
2307 * @pmask: pointer to struct cpumask variable to receive cpus_allowed set.
2309 * Description: Returns the cpumask_var_t cpus_allowed of the cpuset
2310 * attached to the specified @tsk. Guaranteed to return some non-empty
2311 * subset of cpu_online_mask, even if this means going outside the
2315 void cpuset_cpus_allowed(struct task_struct
*tsk
, struct cpumask
*pmask
)
2317 struct cpuset
*cpus_cs
;
2319 mutex_lock(&callback_mutex
);
2321 cpus_cs
= effective_cpumask_cpuset(task_cs(tsk
));
2322 guarantee_online_cpus(cpus_cs
, pmask
);
2324 mutex_unlock(&callback_mutex
);
2327 void cpuset_cpus_allowed_fallback(struct task_struct
*tsk
)
2329 struct cpuset
*cpus_cs
;
2332 cpus_cs
= effective_cpumask_cpuset(task_cs(tsk
));
2333 do_set_cpus_allowed(tsk
, cpus_cs
->cpus_allowed
);
2337 * We own tsk->cpus_allowed, nobody can change it under us.
2339 * But we used cs && cs->cpus_allowed lockless and thus can
2340 * race with cgroup_attach_task() or update_cpumask() and get
2341 * the wrong tsk->cpus_allowed. However, both cases imply the
2342 * subsequent cpuset_change_cpumask()->set_cpus_allowed_ptr()
2343 * which takes task_rq_lock().
2345 * If we are called after it dropped the lock we must see all
2346 * changes in tsk_cs()->cpus_allowed. Otherwise we can temporary
2347 * set any mask even if it is not right from task_cs() pov,
2348 * the pending set_cpus_allowed_ptr() will fix things.
2350 * select_fallback_rq() will fix things ups and set cpu_possible_mask
2355 void cpuset_init_current_mems_allowed(void)
2357 nodes_setall(current
->mems_allowed
);
2361 * cpuset_mems_allowed - return mems_allowed mask from a tasks cpuset.
2362 * @tsk: pointer to task_struct from which to obtain cpuset->mems_allowed.
2364 * Description: Returns the nodemask_t mems_allowed of the cpuset
2365 * attached to the specified @tsk. Guaranteed to return some non-empty
2366 * subset of node_states[N_MEMORY], even if this means going outside the
2370 nodemask_t
cpuset_mems_allowed(struct task_struct
*tsk
)
2372 struct cpuset
*mems_cs
;
2375 mutex_lock(&callback_mutex
);
2377 mems_cs
= effective_nodemask_cpuset(task_cs(tsk
));
2378 guarantee_online_mems(mems_cs
, &mask
);
2380 mutex_unlock(&callback_mutex
);
2386 * cpuset_nodemask_valid_mems_allowed - check nodemask vs. curremt mems_allowed
2387 * @nodemask: the nodemask to be checked
2389 * Are any of the nodes in the nodemask allowed in current->mems_allowed?
2391 int cpuset_nodemask_valid_mems_allowed(nodemask_t
*nodemask
)
2393 return nodes_intersects(*nodemask
, current
->mems_allowed
);
2397 * nearest_hardwall_ancestor() - Returns the nearest mem_exclusive or
2398 * mem_hardwall ancestor to the specified cpuset. Call holding
2399 * callback_mutex. If no ancestor is mem_exclusive or mem_hardwall
2400 * (an unusual configuration), then returns the root cpuset.
2402 static struct cpuset
*nearest_hardwall_ancestor(struct cpuset
*cs
)
2404 while (!(is_mem_exclusive(cs
) || is_mem_hardwall(cs
)) && parent_cs(cs
))
2410 * cpuset_node_allowed_softwall - Can we allocate on a memory node?
2411 * @node: is this an allowed node?
2412 * @gfp_mask: memory allocation flags
2414 * If we're in interrupt, yes, we can always allocate. If __GFP_THISNODE is
2415 * set, yes, we can always allocate. If node is in our task's mems_allowed,
2416 * yes. If it's not a __GFP_HARDWALL request and this node is in the nearest
2417 * hardwalled cpuset ancestor to this task's cpuset, yes. If the task has been
2418 * OOM killed and has access to memory reserves as specified by the TIF_MEMDIE
2422 * If __GFP_HARDWALL is set, cpuset_node_allowed_softwall() reduces to
2423 * cpuset_node_allowed_hardwall(). Otherwise, cpuset_node_allowed_softwall()
2424 * might sleep, and might allow a node from an enclosing cpuset.
2426 * cpuset_node_allowed_hardwall() only handles the simpler case of hardwall
2427 * cpusets, and never sleeps.
2429 * The __GFP_THISNODE placement logic is really handled elsewhere,
2430 * by forcibly using a zonelist starting at a specified node, and by
2431 * (in get_page_from_freelist()) refusing to consider the zones for
2432 * any node on the zonelist except the first. By the time any such
2433 * calls get to this routine, we should just shut up and say 'yes'.
2435 * GFP_USER allocations are marked with the __GFP_HARDWALL bit,
2436 * and do not allow allocations outside the current tasks cpuset
2437 * unless the task has been OOM killed as is marked TIF_MEMDIE.
2438 * GFP_KERNEL allocations are not so marked, so can escape to the
2439 * nearest enclosing hardwalled ancestor cpuset.
2441 * Scanning up parent cpusets requires callback_mutex. The
2442 * __alloc_pages() routine only calls here with __GFP_HARDWALL bit
2443 * _not_ set if it's a GFP_KERNEL allocation, and all nodes in the
2444 * current tasks mems_allowed came up empty on the first pass over
2445 * the zonelist. So only GFP_KERNEL allocations, if all nodes in the
2446 * cpuset are short of memory, might require taking the callback_mutex
2449 * The first call here from mm/page_alloc:get_page_from_freelist()
2450 * has __GFP_HARDWALL set in gfp_mask, enforcing hardwall cpusets,
2451 * so no allocation on a node outside the cpuset is allowed (unless
2452 * in interrupt, of course).
2454 * The second pass through get_page_from_freelist() doesn't even call
2455 * here for GFP_ATOMIC calls. For those calls, the __alloc_pages()
2456 * variable 'wait' is not set, and the bit ALLOC_CPUSET is not set
2457 * in alloc_flags. That logic and the checks below have the combined
2459 * in_interrupt - any node ok (current task context irrelevant)
2460 * GFP_ATOMIC - any node ok
2461 * TIF_MEMDIE - any node ok
2462 * GFP_KERNEL - any node in enclosing hardwalled cpuset ok
2463 * GFP_USER - only nodes in current tasks mems allowed ok.
2466 * Don't call cpuset_node_allowed_softwall if you can't sleep, unless you
2467 * pass in the __GFP_HARDWALL flag set in gfp_flag, which disables
2468 * the code that might scan up ancestor cpusets and sleep.
2470 int __cpuset_node_allowed_softwall(int node
, gfp_t gfp_mask
)
2472 struct cpuset
*cs
; /* current cpuset ancestors */
2473 int allowed
; /* is allocation in zone z allowed? */
2475 if (in_interrupt() || (gfp_mask
& __GFP_THISNODE
))
2477 might_sleep_if(!(gfp_mask
& __GFP_HARDWALL
));
2478 if (node_isset(node
, current
->mems_allowed
))
2481 * Allow tasks that have access to memory reserves because they have
2482 * been OOM killed to get memory anywhere.
2484 if (unlikely(test_thread_flag(TIF_MEMDIE
)))
2486 if (gfp_mask
& __GFP_HARDWALL
) /* If hardwall request, stop here */
2489 if (current
->flags
& PF_EXITING
) /* Let dying task have memory */
2492 /* Not hardwall and node outside mems_allowed: scan up cpusets */
2493 mutex_lock(&callback_mutex
);
2496 cs
= nearest_hardwall_ancestor(task_cs(current
));
2497 task_unlock(current
);
2499 allowed
= node_isset(node
, cs
->mems_allowed
);
2500 mutex_unlock(&callback_mutex
);
2505 * cpuset_node_allowed_hardwall - Can we allocate on a memory node?
2506 * @node: is this an allowed node?
2507 * @gfp_mask: memory allocation flags
2509 * If we're in interrupt, yes, we can always allocate. If __GFP_THISNODE is
2510 * set, yes, we can always allocate. If node is in our task's mems_allowed,
2511 * yes. If the task has been OOM killed and has access to memory reserves as
2512 * specified by the TIF_MEMDIE flag, yes.
2515 * The __GFP_THISNODE placement logic is really handled elsewhere,
2516 * by forcibly using a zonelist starting at a specified node, and by
2517 * (in get_page_from_freelist()) refusing to consider the zones for
2518 * any node on the zonelist except the first. By the time any such
2519 * calls get to this routine, we should just shut up and say 'yes'.
2521 * Unlike the cpuset_node_allowed_softwall() variant, above,
2522 * this variant requires that the node be in the current task's
2523 * mems_allowed or that we're in interrupt. It does not scan up the
2524 * cpuset hierarchy for the nearest enclosing mem_exclusive cpuset.
2527 int __cpuset_node_allowed_hardwall(int node
, gfp_t gfp_mask
)
2529 if (in_interrupt() || (gfp_mask
& __GFP_THISNODE
))
2531 if (node_isset(node
, current
->mems_allowed
))
2534 * Allow tasks that have access to memory reserves because they have
2535 * been OOM killed to get memory anywhere.
2537 if (unlikely(test_thread_flag(TIF_MEMDIE
)))
2543 * cpuset_mem_spread_node() - On which node to begin search for a file page
2544 * cpuset_slab_spread_node() - On which node to begin search for a slab page
2546 * If a task is marked PF_SPREAD_PAGE or PF_SPREAD_SLAB (as for
2547 * tasks in a cpuset with is_spread_page or is_spread_slab set),
2548 * and if the memory allocation used cpuset_mem_spread_node()
2549 * to determine on which node to start looking, as it will for
2550 * certain page cache or slab cache pages such as used for file
2551 * system buffers and inode caches, then instead of starting on the
2552 * local node to look for a free page, rather spread the starting
2553 * node around the tasks mems_allowed nodes.
2555 * We don't have to worry about the returned node being offline
2556 * because "it can't happen", and even if it did, it would be ok.
2558 * The routines calling guarantee_online_mems() are careful to
2559 * only set nodes in task->mems_allowed that are online. So it
2560 * should not be possible for the following code to return an
2561 * offline node. But if it did, that would be ok, as this routine
2562 * is not returning the node where the allocation must be, only
2563 * the node where the search should start. The zonelist passed to
2564 * __alloc_pages() will include all nodes. If the slab allocator
2565 * is passed an offline node, it will fall back to the local node.
2566 * See kmem_cache_alloc_node().
2569 static int cpuset_spread_node(int *rotor
)
2573 node
= next_node(*rotor
, current
->mems_allowed
);
2574 if (node
== MAX_NUMNODES
)
2575 node
= first_node(current
->mems_allowed
);
2580 int cpuset_mem_spread_node(void)
2582 if (current
->cpuset_mem_spread_rotor
== NUMA_NO_NODE
)
2583 current
->cpuset_mem_spread_rotor
=
2584 node_random(¤t
->mems_allowed
);
2586 return cpuset_spread_node(¤t
->cpuset_mem_spread_rotor
);
2589 int cpuset_slab_spread_node(void)
2591 if (current
->cpuset_slab_spread_rotor
== NUMA_NO_NODE
)
2592 current
->cpuset_slab_spread_rotor
=
2593 node_random(¤t
->mems_allowed
);
2595 return cpuset_spread_node(¤t
->cpuset_slab_spread_rotor
);
2598 EXPORT_SYMBOL_GPL(cpuset_mem_spread_node
);
2601 * cpuset_mems_allowed_intersects - Does @tsk1's mems_allowed intersect @tsk2's?
2602 * @tsk1: pointer to task_struct of some task.
2603 * @tsk2: pointer to task_struct of some other task.
2605 * Description: Return true if @tsk1's mems_allowed intersects the
2606 * mems_allowed of @tsk2. Used by the OOM killer to determine if
2607 * one of the task's memory usage might impact the memory available
2611 int cpuset_mems_allowed_intersects(const struct task_struct
*tsk1
,
2612 const struct task_struct
*tsk2
)
2614 return nodes_intersects(tsk1
->mems_allowed
, tsk2
->mems_allowed
);
2617 #define CPUSET_NODELIST_LEN (256)
2620 * cpuset_print_task_mems_allowed - prints task's cpuset and mems_allowed
2621 * @task: pointer to task_struct of some task.
2623 * Description: Prints @task's name, cpuset name, and cached copy of its
2624 * mems_allowed to the kernel log. Must hold task_lock(task) to allow
2625 * dereferencing task_cs(task).
2627 void cpuset_print_task_mems_allowed(struct task_struct
*tsk
)
2629 /* Statically allocated to prevent using excess stack. */
2630 static char cpuset_nodelist
[CPUSET_NODELIST_LEN
];
2631 static DEFINE_SPINLOCK(cpuset_buffer_lock
);
2633 struct cgroup
*cgrp
= task_cs(tsk
)->css
.cgroup
;
2636 spin_lock(&cpuset_buffer_lock
);
2638 nodelist_scnprintf(cpuset_nodelist
, CPUSET_NODELIST_LEN
,
2640 printk(KERN_INFO
"%s cpuset=%s mems_allowed=%s\n",
2641 tsk
->comm
, cgroup_name(cgrp
), cpuset_nodelist
);
2643 spin_unlock(&cpuset_buffer_lock
);
2648 * Collection of memory_pressure is suppressed unless
2649 * this flag is enabled by writing "1" to the special
2650 * cpuset file 'memory_pressure_enabled' in the root cpuset.
2653 int cpuset_memory_pressure_enabled __read_mostly
;
2656 * cpuset_memory_pressure_bump - keep stats of per-cpuset reclaims.
2658 * Keep a running average of the rate of synchronous (direct)
2659 * page reclaim efforts initiated by tasks in each cpuset.
2661 * This represents the rate at which some task in the cpuset
2662 * ran low on memory on all nodes it was allowed to use, and
2663 * had to enter the kernels page reclaim code in an effort to
2664 * create more free memory by tossing clean pages or swapping
2665 * or writing dirty pages.
2667 * Display to user space in the per-cpuset read-only file
2668 * "memory_pressure". Value displayed is an integer
2669 * representing the recent rate of entry into the synchronous
2670 * (direct) page reclaim by any task attached to the cpuset.
2673 void __cpuset_memory_pressure_bump(void)
2676 fmeter_markevent(&task_cs(current
)->fmeter
);
2677 task_unlock(current
);
2680 #ifdef CONFIG_PROC_PID_CPUSET
2682 * proc_cpuset_show()
2683 * - Print tasks cpuset path into seq_file.
2684 * - Used for /proc/<pid>/cpuset.
2685 * - No need to task_lock(tsk) on this tsk->cpuset reference, as it
2686 * doesn't really matter if tsk->cpuset changes after we read it,
2687 * and we take cpuset_mutex, keeping cpuset_attach() from changing it
2690 int proc_cpuset_show(struct seq_file
*m
, void *unused_v
)
2693 struct task_struct
*tsk
;
2695 struct cgroup_subsys_state
*css
;
2699 buf
= kmalloc(PAGE_SIZE
, GFP_KERNEL
);
2705 tsk
= get_pid_task(pid
, PIDTYPE_PID
);
2710 css
= task_css(tsk
, cpuset_subsys_id
);
2711 retval
= cgroup_path(css
->cgroup
, buf
, PAGE_SIZE
);
2718 put_task_struct(tsk
);
2724 #endif /* CONFIG_PROC_PID_CPUSET */
2726 /* Display task mems_allowed in /proc/<pid>/status file. */
2727 void cpuset_task_status_allowed(struct seq_file
*m
, struct task_struct
*task
)
2729 seq_printf(m
, "Mems_allowed:\t");
2730 seq_nodemask(m
, &task
->mems_allowed
);
2731 seq_printf(m
, "\n");
2732 seq_printf(m
, "Mems_allowed_list:\t");
2733 seq_nodemask_list(m
, &task
->mems_allowed
);
2734 seq_printf(m
, "\n");