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 /* See "Frequency meter" comments, below. */
74 int cnt
; /* unprocessed events count */
75 int val
; /* most recent output value */
76 time_t time
; /* clock (secs) when val computed */
77 spinlock_t lock
; /* guards read or write of above */
81 struct cgroup_subsys_state css
;
83 unsigned long flags
; /* "unsigned long" so bitops work */
84 cpumask_var_t cpus_allowed
; /* CPUs allowed to tasks in cpuset */
85 nodemask_t mems_allowed
; /* Memory Nodes allowed to tasks */
88 * This is old Memory Nodes tasks took on.
90 * - top_cpuset.old_mems_allowed is initialized to mems_allowed.
91 * - A new cpuset's old_mems_allowed is initialized when some
92 * task is moved into it.
93 * - old_mems_allowed is used in cpuset_migrate_mm() when we change
94 * cpuset.mems_allowed and have tasks' nodemask updated, and
95 * then old_mems_allowed is updated to mems_allowed.
97 nodemask_t old_mems_allowed
;
99 struct fmeter fmeter
; /* memory_pressure filter */
102 * Tasks are being attached to this cpuset. Used to prevent
103 * zeroing cpus/mems_allowed between ->can_attach() and ->attach().
105 int attach_in_progress
;
107 /* partition number for rebuild_sched_domains() */
110 /* for custom sched domain */
111 int relax_domain_level
;
114 static inline struct cpuset
*css_cs(struct cgroup_subsys_state
*css
)
116 return css
? container_of(css
, struct cpuset
, css
) : NULL
;
119 /* Retrieve the cpuset for a task */
120 static inline struct cpuset
*task_cs(struct task_struct
*task
)
122 return css_cs(task_css(task
, cpuset_cgrp_id
));
125 static inline struct cpuset
*parent_cs(struct cpuset
*cs
)
127 return css_cs(css_parent(&cs
->css
));
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_css: 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_css, parent_cs) \
211 css_for_each_child((pos_css), &(parent_cs)->css) \
212 if (is_cpuset_online(((child_cs) = css_cs((pos_css)))))
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_css: 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_css by calling
222 * css_rightmost_descendant() to skip subtree. @root_cs is included in the
223 * iteration and the first node to be visited.
225 #define cpuset_for_each_descendant_pre(des_cs, pos_css, root_cs) \
226 css_for_each_descendant_pre((pos_css), &(root_cs)->css) \
227 if (is_cpuset_online(((des_cs) = css_cs((pos_css)))))
230 * There are two global mutexes guarding cpuset structures - cpuset_mutex
231 * and callback_mutex. The latter may nest inside the former. We also
232 * require taking task_lock() when dereferencing a task's cpuset pointer.
233 * See "The task_lock() exception", at the end of this comment.
235 * A task must hold both mutexes to modify cpusets. If a task holds
236 * cpuset_mutex, then it blocks others wanting that mutex, ensuring that it
237 * is the only task able to also acquire callback_mutex and be able to
238 * modify cpusets. It can perform various checks on the cpuset structure
239 * first, knowing nothing will change. It can also allocate memory while
240 * just holding cpuset_mutex. While it is performing these checks, various
241 * callback routines can briefly acquire callback_mutex to query cpusets.
242 * Once it is ready to make the changes, it takes callback_mutex, blocking
245 * Calls to the kernel memory allocator can not be made while holding
246 * callback_mutex, as that would risk double tripping on callback_mutex
247 * from one of the callbacks into the cpuset code from within
250 * If a task is only holding callback_mutex, then it has read-only
253 * Now, the task_struct fields mems_allowed and mempolicy may be changed
254 * by other task, we use alloc_lock in the task_struct fields to protect
257 * The cpuset_common_file_read() handlers only hold callback_mutex across
258 * small pieces of code, such as when reading out possibly multi-word
259 * cpumasks and nodemasks.
261 * Accessing a task's cpuset should be done in accordance with the
262 * guidelines for accessing subsystem state in kernel/cgroup.c
265 static DEFINE_MUTEX(cpuset_mutex
);
266 static DEFINE_MUTEX(callback_mutex
);
269 * CPU / memory hotplug is handled asynchronously.
271 static void cpuset_hotplug_workfn(struct work_struct
*work
);
272 static DECLARE_WORK(cpuset_hotplug_work
, cpuset_hotplug_workfn
);
274 static DECLARE_WAIT_QUEUE_HEAD(cpuset_attach_wq
);
277 * This is ugly, but preserves the userspace API for existing cpuset
278 * users. If someone tries to mount the "cpuset" filesystem, we
279 * silently switch it to mount "cgroup" instead
281 static struct dentry
*cpuset_mount(struct file_system_type
*fs_type
,
282 int flags
, const char *unused_dev_name
, void *data
)
284 struct file_system_type
*cgroup_fs
= get_fs_type("cgroup");
285 struct dentry
*ret
= ERR_PTR(-ENODEV
);
289 "release_agent=/sbin/cpuset_release_agent";
290 ret
= cgroup_fs
->mount(cgroup_fs
, flags
,
291 unused_dev_name
, mountopts
);
292 put_filesystem(cgroup_fs
);
297 static struct file_system_type cpuset_fs_type
= {
299 .mount
= cpuset_mount
,
303 * Return in pmask the portion of a cpusets's cpus_allowed that
304 * are online. If none are online, walk up the cpuset hierarchy
305 * until we find one that does have some online cpus. The top
306 * cpuset always has some cpus online.
308 * One way or another, we guarantee to return some non-empty subset
309 * of cpu_online_mask.
311 * Call with callback_mutex held.
313 static void guarantee_online_cpus(struct cpuset
*cs
, struct cpumask
*pmask
)
315 while (!cpumask_intersects(cs
->cpus_allowed
, cpu_online_mask
))
317 cpumask_and(pmask
, cs
->cpus_allowed
, cpu_online_mask
);
321 * Return in *pmask the portion of a cpusets's mems_allowed that
322 * are online, with memory. If none are online with memory, walk
323 * up the cpuset hierarchy until we find one that does have some
324 * online mems. The top cpuset always has some mems online.
326 * One way or another, we guarantee to return some non-empty subset
327 * of node_states[N_MEMORY].
329 * Call with callback_mutex held.
331 static void guarantee_online_mems(struct cpuset
*cs
, nodemask_t
*pmask
)
333 while (!nodes_intersects(cs
->mems_allowed
, node_states
[N_MEMORY
]))
335 nodes_and(*pmask
, cs
->mems_allowed
, node_states
[N_MEMORY
]);
339 * update task's spread flag if cpuset's page/slab spread flag is set
341 * Called with callback_mutex/cpuset_mutex held
343 static void cpuset_update_task_spread_flag(struct cpuset
*cs
,
344 struct task_struct
*tsk
)
346 if (is_spread_page(cs
))
347 tsk
->flags
|= PF_SPREAD_PAGE
;
349 tsk
->flags
&= ~PF_SPREAD_PAGE
;
350 if (is_spread_slab(cs
))
351 tsk
->flags
|= PF_SPREAD_SLAB
;
353 tsk
->flags
&= ~PF_SPREAD_SLAB
;
357 * is_cpuset_subset(p, q) - Is cpuset p a subset of cpuset q?
359 * One cpuset is a subset of another if all its allowed CPUs and
360 * Memory Nodes are a subset of the other, and its exclusive flags
361 * are only set if the other's are set. Call holding cpuset_mutex.
364 static int is_cpuset_subset(const struct cpuset
*p
, const struct cpuset
*q
)
366 return cpumask_subset(p
->cpus_allowed
, q
->cpus_allowed
) &&
367 nodes_subset(p
->mems_allowed
, q
->mems_allowed
) &&
368 is_cpu_exclusive(p
) <= is_cpu_exclusive(q
) &&
369 is_mem_exclusive(p
) <= is_mem_exclusive(q
);
373 * alloc_trial_cpuset - allocate a trial cpuset
374 * @cs: the cpuset that the trial cpuset duplicates
376 static struct cpuset
*alloc_trial_cpuset(struct cpuset
*cs
)
378 struct cpuset
*trial
;
380 trial
= kmemdup(cs
, sizeof(*cs
), GFP_KERNEL
);
384 if (!alloc_cpumask_var(&trial
->cpus_allowed
, GFP_KERNEL
)) {
388 cpumask_copy(trial
->cpus_allowed
, cs
->cpus_allowed
);
394 * free_trial_cpuset - free the trial cpuset
395 * @trial: the trial cpuset to be freed
397 static void free_trial_cpuset(struct cpuset
*trial
)
399 free_cpumask_var(trial
->cpus_allowed
);
404 * validate_change() - Used to validate that any proposed cpuset change
405 * follows the structural rules for cpusets.
407 * If we replaced the flag and mask values of the current cpuset
408 * (cur) with those values in the trial cpuset (trial), would
409 * our various subset and exclusive rules still be valid? Presumes
412 * 'cur' is the address of an actual, in-use cpuset. Operations
413 * such as list traversal that depend on the actual address of the
414 * cpuset in the list must use cur below, not trial.
416 * 'trial' is the address of bulk structure copy of cur, with
417 * perhaps one or more of the fields cpus_allowed, mems_allowed,
418 * or flags changed to new, trial values.
420 * Return 0 if valid, -errno if not.
423 static int validate_change(struct cpuset
*cur
, struct cpuset
*trial
)
425 struct cgroup_subsys_state
*css
;
426 struct cpuset
*c
, *par
;
431 /* Each of our child cpusets must be a subset of us */
433 cpuset_for_each_child(c
, css
, cur
)
434 if (!is_cpuset_subset(c
, trial
))
437 /* Remaining checks don't apply to root cpuset */
439 if (cur
== &top_cpuset
)
442 par
= parent_cs(cur
);
444 /* We must be a subset of our parent cpuset */
446 if (!is_cpuset_subset(trial
, par
))
450 * If either I or some sibling (!= me) is exclusive, we can't
454 cpuset_for_each_child(c
, css
, par
) {
455 if ((is_cpu_exclusive(trial
) || is_cpu_exclusive(c
)) &&
457 cpumask_intersects(trial
->cpus_allowed
, c
->cpus_allowed
))
459 if ((is_mem_exclusive(trial
) || is_mem_exclusive(c
)) &&
461 nodes_intersects(trial
->mems_allowed
, c
->mems_allowed
))
466 * Cpusets with tasks - existing or newly being attached - can't
467 * be changed to have empty cpus_allowed or mems_allowed.
470 if ((cgroup_has_tasks(cur
->css
.cgroup
) || cur
->attach_in_progress
)) {
471 if (!cpumask_empty(cur
->cpus_allowed
) &&
472 cpumask_empty(trial
->cpus_allowed
))
474 if (!nodes_empty(cur
->mems_allowed
) &&
475 nodes_empty(trial
->mems_allowed
))
487 * Helper routine for generate_sched_domains().
488 * Do cpusets a, b have overlapping cpus_allowed masks?
490 static int cpusets_overlap(struct cpuset
*a
, struct cpuset
*b
)
492 return cpumask_intersects(a
->cpus_allowed
, b
->cpus_allowed
);
496 update_domain_attr(struct sched_domain_attr
*dattr
, struct cpuset
*c
)
498 if (dattr
->relax_domain_level
< c
->relax_domain_level
)
499 dattr
->relax_domain_level
= c
->relax_domain_level
;
503 static void update_domain_attr_tree(struct sched_domain_attr
*dattr
,
504 struct cpuset
*root_cs
)
507 struct cgroup_subsys_state
*pos_css
;
510 cpuset_for_each_descendant_pre(cp
, pos_css
, root_cs
) {
514 /* skip the whole subtree if @cp doesn't have any CPU */
515 if (cpumask_empty(cp
->cpus_allowed
)) {
516 pos_css
= css_rightmost_descendant(pos_css
);
520 if (is_sched_load_balance(cp
))
521 update_domain_attr(dattr
, cp
);
527 * generate_sched_domains()
529 * This function builds a partial partition of the systems CPUs
530 * A 'partial partition' is a set of non-overlapping subsets whose
531 * union is a subset of that set.
532 * The output of this function needs to be passed to kernel/sched/core.c
533 * partition_sched_domains() routine, which will rebuild the scheduler's
534 * load balancing domains (sched domains) as specified by that partial
537 * See "What is sched_load_balance" in Documentation/cgroups/cpusets.txt
538 * for a background explanation of this.
540 * Does not return errors, on the theory that the callers of this
541 * routine would rather not worry about failures to rebuild sched
542 * domains when operating in the severe memory shortage situations
543 * that could cause allocation failures below.
545 * Must be called with cpuset_mutex held.
547 * The three key local variables below are:
548 * q - a linked-list queue of cpuset pointers, used to implement a
549 * top-down scan of all cpusets. This scan loads a pointer
550 * to each cpuset marked is_sched_load_balance into the
551 * array 'csa'. For our purposes, rebuilding the schedulers
552 * sched domains, we can ignore !is_sched_load_balance cpusets.
553 * csa - (for CpuSet Array) Array of pointers to all the cpusets
554 * that need to be load balanced, for convenient iterative
555 * access by the subsequent code that finds the best partition,
556 * i.e the set of domains (subsets) of CPUs such that the
557 * cpus_allowed of every cpuset marked is_sched_load_balance
558 * is a subset of one of these domains, while there are as
559 * many such domains as possible, each as small as possible.
560 * doms - Conversion of 'csa' to an array of cpumasks, for passing to
561 * the kernel/sched/core.c routine partition_sched_domains() in a
562 * convenient format, that can be easily compared to the prior
563 * value to determine what partition elements (sched domains)
564 * were changed (added or removed.)
566 * Finding the best partition (set of domains):
567 * The triple nested loops below over i, j, k scan over the
568 * load balanced cpusets (using the array of cpuset pointers in
569 * csa[]) looking for pairs of cpusets that have overlapping
570 * cpus_allowed, but which don't have the same 'pn' partition
571 * number and gives them in the same partition number. It keeps
572 * looping on the 'restart' label until it can no longer find
575 * The union of the cpus_allowed masks from the set of
576 * all cpusets having the same 'pn' value then form the one
577 * element of the partition (one sched domain) to be passed to
578 * partition_sched_domains().
580 static int generate_sched_domains(cpumask_var_t
**domains
,
581 struct sched_domain_attr
**attributes
)
583 struct cpuset
*cp
; /* scans q */
584 struct cpuset
**csa
; /* array of all cpuset ptrs */
585 int csn
; /* how many cpuset ptrs in csa so far */
586 int i
, j
, k
; /* indices for partition finding loops */
587 cpumask_var_t
*doms
; /* resulting partition; i.e. sched domains */
588 struct sched_domain_attr
*dattr
; /* attributes for custom domains */
589 int ndoms
= 0; /* number of sched domains in result */
590 int nslot
; /* next empty doms[] struct cpumask slot */
591 struct cgroup_subsys_state
*pos_css
;
597 /* Special case for the 99% of systems with one, full, sched domain */
598 if (is_sched_load_balance(&top_cpuset
)) {
600 doms
= alloc_sched_domains(ndoms
);
604 dattr
= kmalloc(sizeof(struct sched_domain_attr
), GFP_KERNEL
);
606 *dattr
= SD_ATTR_INIT
;
607 update_domain_attr_tree(dattr
, &top_cpuset
);
609 cpumask_copy(doms
[0], top_cpuset
.cpus_allowed
);
614 csa
= kmalloc(number_of_cpusets
* sizeof(cp
), GFP_KERNEL
);
620 cpuset_for_each_descendant_pre(cp
, pos_css
, &top_cpuset
) {
621 if (cp
== &top_cpuset
)
624 * Continue traversing beyond @cp iff @cp has some CPUs and
625 * isn't load balancing. The former is obvious. The
626 * latter: All child cpusets contain a subset of the
627 * parent's cpus, so just skip them, and then we call
628 * update_domain_attr_tree() to calc relax_domain_level of
629 * the corresponding sched domain.
631 if (!cpumask_empty(cp
->cpus_allowed
) &&
632 !is_sched_load_balance(cp
))
635 if (is_sched_load_balance(cp
))
638 /* skip @cp's subtree */
639 pos_css
= css_rightmost_descendant(pos_css
);
643 for (i
= 0; i
< csn
; i
++)
648 /* Find the best partition (set of sched domains) */
649 for (i
= 0; i
< csn
; i
++) {
650 struct cpuset
*a
= csa
[i
];
653 for (j
= 0; j
< csn
; j
++) {
654 struct cpuset
*b
= csa
[j
];
657 if (apn
!= bpn
&& cpusets_overlap(a
, b
)) {
658 for (k
= 0; k
< csn
; k
++) {
659 struct cpuset
*c
= csa
[k
];
664 ndoms
--; /* one less element */
671 * Now we know how many domains to create.
672 * Convert <csn, csa> to <ndoms, doms> and populate cpu masks.
674 doms
= alloc_sched_domains(ndoms
);
679 * The rest of the code, including the scheduler, can deal with
680 * dattr==NULL case. No need to abort if alloc fails.
682 dattr
= kmalloc(ndoms
* sizeof(struct sched_domain_attr
), GFP_KERNEL
);
684 for (nslot
= 0, i
= 0; i
< csn
; i
++) {
685 struct cpuset
*a
= csa
[i
];
690 /* Skip completed partitions */
696 if (nslot
== ndoms
) {
697 static int warnings
= 10;
700 "rebuild_sched_domains confused:"
701 " nslot %d, ndoms %d, csn %d, i %d,"
703 nslot
, ndoms
, csn
, i
, apn
);
711 *(dattr
+ nslot
) = SD_ATTR_INIT
;
712 for (j
= i
; j
< csn
; j
++) {
713 struct cpuset
*b
= csa
[j
];
716 cpumask_or(dp
, dp
, b
->cpus_allowed
);
718 update_domain_attr_tree(dattr
+ nslot
, b
);
720 /* Done with this partition */
726 BUG_ON(nslot
!= ndoms
);
732 * Fallback to the default domain if kmalloc() failed.
733 * See comments in partition_sched_domains().
744 * Rebuild scheduler domains.
746 * If the flag 'sched_load_balance' of any cpuset with non-empty
747 * 'cpus' changes, or if the 'cpus' allowed changes in any cpuset
748 * which has that flag enabled, or if any cpuset with a non-empty
749 * 'cpus' is removed, then call this routine to rebuild the
750 * scheduler's dynamic sched domains.
752 * Call with cpuset_mutex held. Takes get_online_cpus().
754 static void rebuild_sched_domains_locked(void)
756 struct sched_domain_attr
*attr
;
760 lockdep_assert_held(&cpuset_mutex
);
764 * We have raced with CPU hotplug. Don't do anything to avoid
765 * passing doms with offlined cpu to partition_sched_domains().
766 * Anyways, hotplug work item will rebuild sched domains.
768 if (!cpumask_equal(top_cpuset
.cpus_allowed
, cpu_active_mask
))
771 /* Generate domain masks and attrs */
772 ndoms
= generate_sched_domains(&doms
, &attr
);
774 /* Have scheduler rebuild the domains */
775 partition_sched_domains(ndoms
, doms
, attr
);
779 #else /* !CONFIG_SMP */
780 static void rebuild_sched_domains_locked(void)
783 #endif /* CONFIG_SMP */
785 void rebuild_sched_domains(void)
787 mutex_lock(&cpuset_mutex
);
788 rebuild_sched_domains_locked();
789 mutex_unlock(&cpuset_mutex
);
793 * effective_cpumask_cpuset - return nearest ancestor with non-empty cpus
794 * @cs: the cpuset in interest
796 * A cpuset's effective cpumask is the cpumask of the nearest ancestor
797 * with non-empty cpus. We use effective cpumask whenever:
798 * - we update tasks' cpus_allowed. (they take on the ancestor's cpumask
799 * if the cpuset they reside in has no cpus)
800 * - we want to retrieve task_cs(tsk)'s cpus_allowed.
802 * Called with cpuset_mutex held. cpuset_cpus_allowed_fallback() is an
803 * exception. See comments there.
805 static struct cpuset
*effective_cpumask_cpuset(struct cpuset
*cs
)
807 while (cpumask_empty(cs
->cpus_allowed
))
813 * effective_nodemask_cpuset - return nearest ancestor with non-empty mems
814 * @cs: the cpuset in interest
816 * A cpuset's effective nodemask is the nodemask of the nearest ancestor
817 * with non-empty memss. We use effective nodemask whenever:
818 * - we update tasks' mems_allowed. (they take on the ancestor's nodemask
819 * if the cpuset they reside in has no mems)
820 * - we want to retrieve task_cs(tsk)'s mems_allowed.
822 * Called with cpuset_mutex held.
824 static struct cpuset
*effective_nodemask_cpuset(struct cpuset
*cs
)
826 while (nodes_empty(cs
->mems_allowed
))
832 * update_tasks_cpumask - Update the cpumasks of tasks in the cpuset.
833 * @cs: the cpuset in which each task's cpus_allowed mask needs to be changed
835 * Iterate through each task of @cs updating its cpus_allowed to the
836 * effective cpuset's. As this function is called with cpuset_mutex held,
837 * cpuset membership stays stable.
839 static void update_tasks_cpumask(struct cpuset
*cs
)
841 struct cpuset
*cpus_cs
= effective_cpumask_cpuset(cs
);
842 struct css_task_iter it
;
843 struct task_struct
*task
;
845 css_task_iter_start(&cs
->css
, &it
);
846 while ((task
= css_task_iter_next(&it
)))
847 set_cpus_allowed_ptr(task
, cpus_cs
->cpus_allowed
);
848 css_task_iter_end(&it
);
852 * update_tasks_cpumask_hier - Update the cpumasks of tasks in the hierarchy.
853 * @root_cs: the root cpuset of the hierarchy
854 * @update_root: update root cpuset or not?
856 * This will update cpumasks of tasks in @root_cs and all other empty cpusets
857 * which take on cpumask of @root_cs.
859 * Called with cpuset_mutex held
861 static void update_tasks_cpumask_hier(struct cpuset
*root_cs
, bool update_root
)
864 struct cgroup_subsys_state
*pos_css
;
867 cpuset_for_each_descendant_pre(cp
, pos_css
, root_cs
) {
872 /* skip the whole subtree if @cp have some CPU */
873 if (!cpumask_empty(cp
->cpus_allowed
)) {
874 pos_css
= css_rightmost_descendant(pos_css
);
878 if (!css_tryget(&cp
->css
))
882 update_tasks_cpumask(cp
);
891 * update_cpumask - update the cpus_allowed mask of a cpuset and all tasks in it
892 * @cs: the cpuset to consider
893 * @buf: buffer of cpu numbers written to this cpuset
895 static int update_cpumask(struct cpuset
*cs
, struct cpuset
*trialcs
,
899 int is_load_balanced
;
901 /* top_cpuset.cpus_allowed tracks cpu_online_mask; it's read-only */
902 if (cs
== &top_cpuset
)
906 * An empty cpus_allowed is ok only if the cpuset has no tasks.
907 * Since cpulist_parse() fails on an empty mask, we special case
908 * that parsing. The validate_change() call ensures that cpusets
909 * with tasks have cpus.
912 cpumask_clear(trialcs
->cpus_allowed
);
914 retval
= cpulist_parse(buf
, trialcs
->cpus_allowed
);
918 if (!cpumask_subset(trialcs
->cpus_allowed
, cpu_active_mask
))
922 /* Nothing to do if the cpus didn't change */
923 if (cpumask_equal(cs
->cpus_allowed
, trialcs
->cpus_allowed
))
926 retval
= validate_change(cs
, trialcs
);
930 is_load_balanced
= is_sched_load_balance(trialcs
);
932 mutex_lock(&callback_mutex
);
933 cpumask_copy(cs
->cpus_allowed
, trialcs
->cpus_allowed
);
934 mutex_unlock(&callback_mutex
);
936 update_tasks_cpumask_hier(cs
, true);
938 if (is_load_balanced
)
939 rebuild_sched_domains_locked();
946 * Migrate memory region from one set of nodes to another.
948 * Temporarilly set tasks mems_allowed to target nodes of migration,
949 * so that the migration code can allocate pages on these nodes.
951 * Call holding cpuset_mutex, so current's cpuset won't change
952 * during this call, as manage_mutex holds off any cpuset_attach()
953 * calls. Therefore we don't need to take task_lock around the
954 * call to guarantee_online_mems(), as we know no one is changing
957 * While the mm_struct we are migrating is typically from some
958 * other task, the task_struct mems_allowed that we are hacking
959 * is for our current task, which must allocate new pages for that
960 * migrating memory region.
963 static void cpuset_migrate_mm(struct mm_struct
*mm
, const nodemask_t
*from
,
964 const nodemask_t
*to
)
966 struct task_struct
*tsk
= current
;
967 struct cpuset
*mems_cs
;
969 tsk
->mems_allowed
= *to
;
971 do_migrate_pages(mm
, from
, to
, MPOL_MF_MOVE_ALL
);
973 mems_cs
= effective_nodemask_cpuset(task_cs(tsk
));
974 guarantee_online_mems(mems_cs
, &tsk
->mems_allowed
);
978 * cpuset_change_task_nodemask - change task's mems_allowed and mempolicy
979 * @tsk: the task to change
980 * @newmems: new nodes that the task will be set
982 * In order to avoid seeing no nodes if the old and new nodes are disjoint,
983 * we structure updates as setting all new allowed nodes, then clearing newly
986 static void cpuset_change_task_nodemask(struct task_struct
*tsk
,
992 * Allow tasks that have access to memory reserves because they have
993 * been OOM killed to get memory anywhere.
995 if (unlikely(test_thread_flag(TIF_MEMDIE
)))
997 if (current
->flags
& PF_EXITING
) /* Let dying task have memory */
1002 * Determine if a loop is necessary if another thread is doing
1003 * get_mems_allowed(). If at least one node remains unchanged and
1004 * tsk does not have a mempolicy, then an empty nodemask will not be
1005 * possible when mems_allowed is larger than a word.
1007 need_loop
= task_has_mempolicy(tsk
) ||
1008 !nodes_intersects(*newmems
, tsk
->mems_allowed
);
1011 local_irq_disable();
1012 write_seqcount_begin(&tsk
->mems_allowed_seq
);
1015 nodes_or(tsk
->mems_allowed
, tsk
->mems_allowed
, *newmems
);
1016 mpol_rebind_task(tsk
, newmems
, MPOL_REBIND_STEP1
);
1018 mpol_rebind_task(tsk
, newmems
, MPOL_REBIND_STEP2
);
1019 tsk
->mems_allowed
= *newmems
;
1022 write_seqcount_end(&tsk
->mems_allowed_seq
);
1029 static void *cpuset_being_rebound
;
1032 * update_tasks_nodemask - Update the nodemasks of tasks in the cpuset.
1033 * @cs: the cpuset in which each task's mems_allowed mask needs to be changed
1035 * Iterate through each task of @cs updating its mems_allowed to the
1036 * effective cpuset's. As this function is called with cpuset_mutex held,
1037 * cpuset membership stays stable.
1039 static void update_tasks_nodemask(struct cpuset
*cs
)
1041 static nodemask_t newmems
; /* protected by cpuset_mutex */
1042 struct cpuset
*mems_cs
= effective_nodemask_cpuset(cs
);
1043 struct css_task_iter it
;
1044 struct task_struct
*task
;
1046 cpuset_being_rebound
= cs
; /* causes mpol_dup() rebind */
1048 guarantee_online_mems(mems_cs
, &newmems
);
1051 * The mpol_rebind_mm() call takes mmap_sem, which we couldn't
1052 * take while holding tasklist_lock. Forks can happen - the
1053 * mpol_dup() cpuset_being_rebound check will catch such forks,
1054 * and rebind their vma mempolicies too. Because we still hold
1055 * the global cpuset_mutex, we know that no other rebind effort
1056 * will be contending for the global variable cpuset_being_rebound.
1057 * It's ok if we rebind the same mm twice; mpol_rebind_mm()
1058 * is idempotent. Also migrate pages in each mm to new nodes.
1060 css_task_iter_start(&cs
->css
, &it
);
1061 while ((task
= css_task_iter_next(&it
))) {
1062 struct mm_struct
*mm
;
1065 cpuset_change_task_nodemask(task
, &newmems
);
1067 mm
= get_task_mm(task
);
1071 migrate
= is_memory_migrate(cs
);
1073 mpol_rebind_mm(mm
, &cs
->mems_allowed
);
1075 cpuset_migrate_mm(mm
, &cs
->old_mems_allowed
, &newmems
);
1078 css_task_iter_end(&it
);
1081 * All the tasks' nodemasks have been updated, update
1082 * cs->old_mems_allowed.
1084 cs
->old_mems_allowed
= newmems
;
1086 /* We're done rebinding vmas to this cpuset's new mems_allowed. */
1087 cpuset_being_rebound
= NULL
;
1091 * update_tasks_nodemask_hier - Update the nodemasks of tasks in the hierarchy.
1092 * @cs: the root cpuset of the hierarchy
1093 * @update_root: update the root cpuset or not?
1095 * This will update nodemasks of tasks in @root_cs and all other empty cpusets
1096 * which take on nodemask of @root_cs.
1098 * Called with cpuset_mutex held
1100 static void update_tasks_nodemask_hier(struct cpuset
*root_cs
, bool update_root
)
1103 struct cgroup_subsys_state
*pos_css
;
1106 cpuset_for_each_descendant_pre(cp
, pos_css
, root_cs
) {
1107 if (cp
== root_cs
) {
1111 /* skip the whole subtree if @cp have some CPU */
1112 if (!nodes_empty(cp
->mems_allowed
)) {
1113 pos_css
= css_rightmost_descendant(pos_css
);
1117 if (!css_tryget(&cp
->css
))
1121 update_tasks_nodemask(cp
);
1130 * Handle user request to change the 'mems' memory placement
1131 * of a cpuset. Needs to validate the request, update the
1132 * cpusets mems_allowed, and for each task in the cpuset,
1133 * update mems_allowed and rebind task's mempolicy and any vma
1134 * mempolicies and if the cpuset is marked 'memory_migrate',
1135 * migrate the tasks pages to the new memory.
1137 * Call with cpuset_mutex held. May take callback_mutex during call.
1138 * Will take tasklist_lock, scan tasklist for tasks in cpuset cs,
1139 * lock each such tasks mm->mmap_sem, scan its vma's and rebind
1140 * their mempolicies to the cpusets new mems_allowed.
1142 static int update_nodemask(struct cpuset
*cs
, struct cpuset
*trialcs
,
1148 * top_cpuset.mems_allowed tracks node_stats[N_MEMORY];
1151 if (cs
== &top_cpuset
) {
1157 * An empty mems_allowed is ok iff there are no tasks in the cpuset.
1158 * Since nodelist_parse() fails on an empty mask, we special case
1159 * that parsing. The validate_change() call ensures that cpusets
1160 * with tasks have memory.
1163 nodes_clear(trialcs
->mems_allowed
);
1165 retval
= nodelist_parse(buf
, trialcs
->mems_allowed
);
1169 if (!nodes_subset(trialcs
->mems_allowed
,
1170 node_states
[N_MEMORY
])) {
1176 if (nodes_equal(cs
->mems_allowed
, trialcs
->mems_allowed
)) {
1177 retval
= 0; /* Too easy - nothing to do */
1180 retval
= validate_change(cs
, trialcs
);
1184 mutex_lock(&callback_mutex
);
1185 cs
->mems_allowed
= trialcs
->mems_allowed
;
1186 mutex_unlock(&callback_mutex
);
1188 update_tasks_nodemask_hier(cs
, true);
1193 int current_cpuset_is_being_rebound(void)
1195 return task_cs(current
) == cpuset_being_rebound
;
1198 static int update_relax_domain_level(struct cpuset
*cs
, s64 val
)
1201 if (val
< -1 || val
>= sched_domain_level_max
)
1205 if (val
!= cs
->relax_domain_level
) {
1206 cs
->relax_domain_level
= val
;
1207 if (!cpumask_empty(cs
->cpus_allowed
) &&
1208 is_sched_load_balance(cs
))
1209 rebuild_sched_domains_locked();
1216 * update_tasks_flags - update the spread flags of tasks in the cpuset.
1217 * @cs: the cpuset in which each task's spread flags needs to be changed
1219 * Iterate through each task of @cs updating its spread flags. As this
1220 * function is called with cpuset_mutex held, cpuset membership stays
1223 static void update_tasks_flags(struct cpuset
*cs
)
1225 struct css_task_iter it
;
1226 struct task_struct
*task
;
1228 css_task_iter_start(&cs
->css
, &it
);
1229 while ((task
= css_task_iter_next(&it
)))
1230 cpuset_update_task_spread_flag(cs
, task
);
1231 css_task_iter_end(&it
);
1235 * update_flag - read a 0 or a 1 in a file and update associated flag
1236 * bit: the bit to update (see cpuset_flagbits_t)
1237 * cs: the cpuset to update
1238 * turning_on: whether the flag is being set or cleared
1240 * Call with cpuset_mutex held.
1243 static int update_flag(cpuset_flagbits_t bit
, struct cpuset
*cs
,
1246 struct cpuset
*trialcs
;
1247 int balance_flag_changed
;
1248 int spread_flag_changed
;
1251 trialcs
= alloc_trial_cpuset(cs
);
1256 set_bit(bit
, &trialcs
->flags
);
1258 clear_bit(bit
, &trialcs
->flags
);
1260 err
= validate_change(cs
, trialcs
);
1264 balance_flag_changed
= (is_sched_load_balance(cs
) !=
1265 is_sched_load_balance(trialcs
));
1267 spread_flag_changed
= ((is_spread_slab(cs
) != is_spread_slab(trialcs
))
1268 || (is_spread_page(cs
) != is_spread_page(trialcs
)));
1270 mutex_lock(&callback_mutex
);
1271 cs
->flags
= trialcs
->flags
;
1272 mutex_unlock(&callback_mutex
);
1274 if (!cpumask_empty(trialcs
->cpus_allowed
) && balance_flag_changed
)
1275 rebuild_sched_domains_locked();
1277 if (spread_flag_changed
)
1278 update_tasks_flags(cs
);
1280 free_trial_cpuset(trialcs
);
1285 * Frequency meter - How fast is some event occurring?
1287 * These routines manage a digitally filtered, constant time based,
1288 * event frequency meter. There are four routines:
1289 * fmeter_init() - initialize a frequency meter.
1290 * fmeter_markevent() - called each time the event happens.
1291 * fmeter_getrate() - returns the recent rate of such events.
1292 * fmeter_update() - internal routine used to update fmeter.
1294 * A common data structure is passed to each of these routines,
1295 * which is used to keep track of the state required to manage the
1296 * frequency meter and its digital filter.
1298 * The filter works on the number of events marked per unit time.
1299 * The filter is single-pole low-pass recursive (IIR). The time unit
1300 * is 1 second. Arithmetic is done using 32-bit integers scaled to
1301 * simulate 3 decimal digits of precision (multiplied by 1000).
1303 * With an FM_COEF of 933, and a time base of 1 second, the filter
1304 * has a half-life of 10 seconds, meaning that if the events quit
1305 * happening, then the rate returned from the fmeter_getrate()
1306 * will be cut in half each 10 seconds, until it converges to zero.
1308 * It is not worth doing a real infinitely recursive filter. If more
1309 * than FM_MAXTICKS ticks have elapsed since the last filter event,
1310 * just compute FM_MAXTICKS ticks worth, by which point the level
1313 * Limit the count of unprocessed events to FM_MAXCNT, so as to avoid
1314 * arithmetic overflow in the fmeter_update() routine.
1316 * Given the simple 32 bit integer arithmetic used, this meter works
1317 * best for reporting rates between one per millisecond (msec) and
1318 * one per 32 (approx) seconds. At constant rates faster than one
1319 * per msec it maxes out at values just under 1,000,000. At constant
1320 * rates between one per msec, and one per second it will stabilize
1321 * to a value N*1000, where N is the rate of events per second.
1322 * At constant rates between one per second and one per 32 seconds,
1323 * it will be choppy, moving up on the seconds that have an event,
1324 * and then decaying until the next event. At rates slower than
1325 * about one in 32 seconds, it decays all the way back to zero between
1329 #define FM_COEF 933 /* coefficient for half-life of 10 secs */
1330 #define FM_MAXTICKS ((time_t)99) /* useless computing more ticks than this */
1331 #define FM_MAXCNT 1000000 /* limit cnt to avoid overflow */
1332 #define FM_SCALE 1000 /* faux fixed point scale */
1334 /* Initialize a frequency meter */
1335 static void fmeter_init(struct fmeter
*fmp
)
1340 spin_lock_init(&fmp
->lock
);
1343 /* Internal meter update - process cnt events and update value */
1344 static void fmeter_update(struct fmeter
*fmp
)
1346 time_t now
= get_seconds();
1347 time_t ticks
= now
- fmp
->time
;
1352 ticks
= min(FM_MAXTICKS
, ticks
);
1354 fmp
->val
= (FM_COEF
* fmp
->val
) / FM_SCALE
;
1357 fmp
->val
+= ((FM_SCALE
- FM_COEF
) * fmp
->cnt
) / FM_SCALE
;
1361 /* Process any previous ticks, then bump cnt by one (times scale). */
1362 static void fmeter_markevent(struct fmeter
*fmp
)
1364 spin_lock(&fmp
->lock
);
1366 fmp
->cnt
= min(FM_MAXCNT
, fmp
->cnt
+ FM_SCALE
);
1367 spin_unlock(&fmp
->lock
);
1370 /* Process any previous ticks, then return current value. */
1371 static int fmeter_getrate(struct fmeter
*fmp
)
1375 spin_lock(&fmp
->lock
);
1378 spin_unlock(&fmp
->lock
);
1382 /* Called by cgroups to determine if a cpuset is usable; cpuset_mutex held */
1383 static int cpuset_can_attach(struct cgroup_subsys_state
*css
,
1384 struct cgroup_taskset
*tset
)
1386 struct cpuset
*cs
= css_cs(css
);
1387 struct task_struct
*task
;
1390 mutex_lock(&cpuset_mutex
);
1393 * We allow to move tasks into an empty cpuset if sane_behavior
1397 if (!cgroup_sane_behavior(css
->cgroup
) &&
1398 (cpumask_empty(cs
->cpus_allowed
) || nodes_empty(cs
->mems_allowed
)))
1401 cgroup_taskset_for_each(task
, tset
) {
1403 * Kthreads which disallow setaffinity shouldn't be moved
1404 * to a new cpuset; we don't want to change their cpu
1405 * affinity and isolating such threads by their set of
1406 * allowed nodes is unnecessary. Thus, cpusets are not
1407 * applicable for such threads. This prevents checking for
1408 * success of set_cpus_allowed_ptr() on all attached tasks
1409 * before cpus_allowed may be changed.
1412 if (task
->flags
& PF_NO_SETAFFINITY
)
1414 ret
= security_task_setscheduler(task
);
1420 * Mark attach is in progress. This makes validate_change() fail
1421 * changes which zero cpus/mems_allowed.
1423 cs
->attach_in_progress
++;
1426 mutex_unlock(&cpuset_mutex
);
1430 static void cpuset_cancel_attach(struct cgroup_subsys_state
*css
,
1431 struct cgroup_taskset
*tset
)
1433 mutex_lock(&cpuset_mutex
);
1434 css_cs(css
)->attach_in_progress
--;
1435 mutex_unlock(&cpuset_mutex
);
1439 * Protected by cpuset_mutex. cpus_attach is used only by cpuset_attach()
1440 * but we can't allocate it dynamically there. Define it global and
1441 * allocate from cpuset_init().
1443 static cpumask_var_t cpus_attach
;
1445 static void cpuset_attach(struct cgroup_subsys_state
*css
,
1446 struct cgroup_taskset
*tset
)
1448 /* static buf protected by cpuset_mutex */
1449 static nodemask_t cpuset_attach_nodemask_to
;
1450 struct mm_struct
*mm
;
1451 struct task_struct
*task
;
1452 struct task_struct
*leader
= cgroup_taskset_first(tset
);
1453 struct cgroup_subsys_state
*oldcss
= cgroup_taskset_cur_css(tset
,
1455 struct cpuset
*cs
= css_cs(css
);
1456 struct cpuset
*oldcs
= css_cs(oldcss
);
1457 struct cpuset
*cpus_cs
= effective_cpumask_cpuset(cs
);
1458 struct cpuset
*mems_cs
= effective_nodemask_cpuset(cs
);
1460 mutex_lock(&cpuset_mutex
);
1462 /* prepare for attach */
1463 if (cs
== &top_cpuset
)
1464 cpumask_copy(cpus_attach
, cpu_possible_mask
);
1466 guarantee_online_cpus(cpus_cs
, cpus_attach
);
1468 guarantee_online_mems(mems_cs
, &cpuset_attach_nodemask_to
);
1470 cgroup_taskset_for_each(task
, tset
) {
1472 * can_attach beforehand should guarantee that this doesn't
1473 * fail. TODO: have a better way to handle failure here
1475 WARN_ON_ONCE(set_cpus_allowed_ptr(task
, cpus_attach
));
1477 cpuset_change_task_nodemask(task
, &cpuset_attach_nodemask_to
);
1478 cpuset_update_task_spread_flag(cs
, task
);
1482 * Change mm, possibly for multiple threads in a threadgroup. This is
1483 * expensive and may sleep.
1485 cpuset_attach_nodemask_to
= cs
->mems_allowed
;
1486 mm
= get_task_mm(leader
);
1488 struct cpuset
*mems_oldcs
= effective_nodemask_cpuset(oldcs
);
1490 mpol_rebind_mm(mm
, &cpuset_attach_nodemask_to
);
1493 * old_mems_allowed is the same with mems_allowed here, except
1494 * if this task is being moved automatically due to hotplug.
1495 * In that case @mems_allowed has been updated and is empty,
1496 * so @old_mems_allowed is the right nodesets that we migrate
1499 if (is_memory_migrate(cs
)) {
1500 cpuset_migrate_mm(mm
, &mems_oldcs
->old_mems_allowed
,
1501 &cpuset_attach_nodemask_to
);
1506 cs
->old_mems_allowed
= cpuset_attach_nodemask_to
;
1508 cs
->attach_in_progress
--;
1509 if (!cs
->attach_in_progress
)
1510 wake_up(&cpuset_attach_wq
);
1512 mutex_unlock(&cpuset_mutex
);
1515 /* The various types of files and directories in a cpuset file system */
1518 FILE_MEMORY_MIGRATE
,
1524 FILE_SCHED_LOAD_BALANCE
,
1525 FILE_SCHED_RELAX_DOMAIN_LEVEL
,
1526 FILE_MEMORY_PRESSURE_ENABLED
,
1527 FILE_MEMORY_PRESSURE
,
1530 } cpuset_filetype_t
;
1532 static int cpuset_write_u64(struct cgroup_subsys_state
*css
, struct cftype
*cft
,
1535 struct cpuset
*cs
= css_cs(css
);
1536 cpuset_filetype_t type
= cft
->private;
1539 mutex_lock(&cpuset_mutex
);
1540 if (!is_cpuset_online(cs
)) {
1546 case FILE_CPU_EXCLUSIVE
:
1547 retval
= update_flag(CS_CPU_EXCLUSIVE
, cs
, val
);
1549 case FILE_MEM_EXCLUSIVE
:
1550 retval
= update_flag(CS_MEM_EXCLUSIVE
, cs
, val
);
1552 case FILE_MEM_HARDWALL
:
1553 retval
= update_flag(CS_MEM_HARDWALL
, cs
, val
);
1555 case FILE_SCHED_LOAD_BALANCE
:
1556 retval
= update_flag(CS_SCHED_LOAD_BALANCE
, cs
, val
);
1558 case FILE_MEMORY_MIGRATE
:
1559 retval
= update_flag(CS_MEMORY_MIGRATE
, cs
, val
);
1561 case FILE_MEMORY_PRESSURE_ENABLED
:
1562 cpuset_memory_pressure_enabled
= !!val
;
1564 case FILE_MEMORY_PRESSURE
:
1567 case FILE_SPREAD_PAGE
:
1568 retval
= update_flag(CS_SPREAD_PAGE
, cs
, val
);
1570 case FILE_SPREAD_SLAB
:
1571 retval
= update_flag(CS_SPREAD_SLAB
, cs
, val
);
1578 mutex_unlock(&cpuset_mutex
);
1582 static int cpuset_write_s64(struct cgroup_subsys_state
*css
, struct cftype
*cft
,
1585 struct cpuset
*cs
= css_cs(css
);
1586 cpuset_filetype_t type
= cft
->private;
1587 int retval
= -ENODEV
;
1589 mutex_lock(&cpuset_mutex
);
1590 if (!is_cpuset_online(cs
))
1594 case FILE_SCHED_RELAX_DOMAIN_LEVEL
:
1595 retval
= update_relax_domain_level(cs
, val
);
1602 mutex_unlock(&cpuset_mutex
);
1607 * Common handling for a write to a "cpus" or "mems" file.
1609 static int cpuset_write_resmask(struct cgroup_subsys_state
*css
,
1610 struct cftype
*cft
, const char *buf
)
1612 struct cpuset
*cs
= css_cs(css
);
1613 struct cpuset
*trialcs
;
1614 int retval
= -ENODEV
;
1617 * CPU or memory hotunplug may leave @cs w/o any execution
1618 * resources, in which case the hotplug code asynchronously updates
1619 * configuration and transfers all tasks to the nearest ancestor
1620 * which can execute.
1622 * As writes to "cpus" or "mems" may restore @cs's execution
1623 * resources, wait for the previously scheduled operations before
1624 * proceeding, so that we don't end up keep removing tasks added
1625 * after execution capability is restored.
1627 flush_work(&cpuset_hotplug_work
);
1629 mutex_lock(&cpuset_mutex
);
1630 if (!is_cpuset_online(cs
))
1633 trialcs
= alloc_trial_cpuset(cs
);
1639 switch (cft
->private) {
1641 retval
= update_cpumask(cs
, trialcs
, buf
);
1644 retval
= update_nodemask(cs
, trialcs
, buf
);
1651 free_trial_cpuset(trialcs
);
1653 mutex_unlock(&cpuset_mutex
);
1658 * These ascii lists should be read in a single call, by using a user
1659 * buffer large enough to hold the entire map. If read in smaller
1660 * chunks, there is no guarantee of atomicity. Since the display format
1661 * used, list of ranges of sequential numbers, is variable length,
1662 * and since these maps can change value dynamically, one could read
1663 * gibberish by doing partial reads while a list was changing.
1665 static int cpuset_common_seq_show(struct seq_file
*sf
, void *v
)
1667 struct cpuset
*cs
= css_cs(seq_css(sf
));
1668 cpuset_filetype_t type
= seq_cft(sf
)->private;
1673 count
= seq_get_buf(sf
, &buf
);
1676 mutex_lock(&callback_mutex
);
1680 s
+= cpulist_scnprintf(s
, count
, cs
->cpus_allowed
);
1683 s
+= nodelist_scnprintf(s
, count
, cs
->mems_allowed
);
1690 if (s
< buf
+ count
- 1) {
1692 seq_commit(sf
, s
- buf
);
1697 mutex_unlock(&callback_mutex
);
1701 static u64
cpuset_read_u64(struct cgroup_subsys_state
*css
, struct cftype
*cft
)
1703 struct cpuset
*cs
= css_cs(css
);
1704 cpuset_filetype_t type
= cft
->private;
1706 case FILE_CPU_EXCLUSIVE
:
1707 return is_cpu_exclusive(cs
);
1708 case FILE_MEM_EXCLUSIVE
:
1709 return is_mem_exclusive(cs
);
1710 case FILE_MEM_HARDWALL
:
1711 return is_mem_hardwall(cs
);
1712 case FILE_SCHED_LOAD_BALANCE
:
1713 return is_sched_load_balance(cs
);
1714 case FILE_MEMORY_MIGRATE
:
1715 return is_memory_migrate(cs
);
1716 case FILE_MEMORY_PRESSURE_ENABLED
:
1717 return cpuset_memory_pressure_enabled
;
1718 case FILE_MEMORY_PRESSURE
:
1719 return fmeter_getrate(&cs
->fmeter
);
1720 case FILE_SPREAD_PAGE
:
1721 return is_spread_page(cs
);
1722 case FILE_SPREAD_SLAB
:
1723 return is_spread_slab(cs
);
1728 /* Unreachable but makes gcc happy */
1732 static s64
cpuset_read_s64(struct cgroup_subsys_state
*css
, struct cftype
*cft
)
1734 struct cpuset
*cs
= css_cs(css
);
1735 cpuset_filetype_t type
= cft
->private;
1737 case FILE_SCHED_RELAX_DOMAIN_LEVEL
:
1738 return cs
->relax_domain_level
;
1743 /* Unrechable but makes gcc happy */
1749 * for the common functions, 'private' gives the type of file
1752 static struct cftype files
[] = {
1755 .seq_show
= cpuset_common_seq_show
,
1756 .write_string
= cpuset_write_resmask
,
1757 .max_write_len
= (100U + 6 * NR_CPUS
),
1758 .private = FILE_CPULIST
,
1763 .seq_show
= cpuset_common_seq_show
,
1764 .write_string
= cpuset_write_resmask
,
1765 .max_write_len
= (100U + 6 * MAX_NUMNODES
),
1766 .private = FILE_MEMLIST
,
1770 .name
= "cpu_exclusive",
1771 .read_u64
= cpuset_read_u64
,
1772 .write_u64
= cpuset_write_u64
,
1773 .private = FILE_CPU_EXCLUSIVE
,
1777 .name
= "mem_exclusive",
1778 .read_u64
= cpuset_read_u64
,
1779 .write_u64
= cpuset_write_u64
,
1780 .private = FILE_MEM_EXCLUSIVE
,
1784 .name
= "mem_hardwall",
1785 .read_u64
= cpuset_read_u64
,
1786 .write_u64
= cpuset_write_u64
,
1787 .private = FILE_MEM_HARDWALL
,
1791 .name
= "sched_load_balance",
1792 .read_u64
= cpuset_read_u64
,
1793 .write_u64
= cpuset_write_u64
,
1794 .private = FILE_SCHED_LOAD_BALANCE
,
1798 .name
= "sched_relax_domain_level",
1799 .read_s64
= cpuset_read_s64
,
1800 .write_s64
= cpuset_write_s64
,
1801 .private = FILE_SCHED_RELAX_DOMAIN_LEVEL
,
1805 .name
= "memory_migrate",
1806 .read_u64
= cpuset_read_u64
,
1807 .write_u64
= cpuset_write_u64
,
1808 .private = FILE_MEMORY_MIGRATE
,
1812 .name
= "memory_pressure",
1813 .read_u64
= cpuset_read_u64
,
1814 .write_u64
= cpuset_write_u64
,
1815 .private = FILE_MEMORY_PRESSURE
,
1820 .name
= "memory_spread_page",
1821 .read_u64
= cpuset_read_u64
,
1822 .write_u64
= cpuset_write_u64
,
1823 .private = FILE_SPREAD_PAGE
,
1827 .name
= "memory_spread_slab",
1828 .read_u64
= cpuset_read_u64
,
1829 .write_u64
= cpuset_write_u64
,
1830 .private = FILE_SPREAD_SLAB
,
1834 .name
= "memory_pressure_enabled",
1835 .flags
= CFTYPE_ONLY_ON_ROOT
,
1836 .read_u64
= cpuset_read_u64
,
1837 .write_u64
= cpuset_write_u64
,
1838 .private = FILE_MEMORY_PRESSURE_ENABLED
,
1845 * cpuset_css_alloc - allocate a cpuset css
1846 * cgrp: control group that the new cpuset will be part of
1849 static struct cgroup_subsys_state
*
1850 cpuset_css_alloc(struct cgroup_subsys_state
*parent_css
)
1855 return &top_cpuset
.css
;
1857 cs
= kzalloc(sizeof(*cs
), GFP_KERNEL
);
1859 return ERR_PTR(-ENOMEM
);
1860 if (!alloc_cpumask_var(&cs
->cpus_allowed
, GFP_KERNEL
)) {
1862 return ERR_PTR(-ENOMEM
);
1865 set_bit(CS_SCHED_LOAD_BALANCE
, &cs
->flags
);
1866 cpumask_clear(cs
->cpus_allowed
);
1867 nodes_clear(cs
->mems_allowed
);
1868 fmeter_init(&cs
->fmeter
);
1869 cs
->relax_domain_level
= -1;
1874 static int cpuset_css_online(struct cgroup_subsys_state
*css
)
1876 struct cpuset
*cs
= css_cs(css
);
1877 struct cpuset
*parent
= parent_cs(cs
);
1878 struct cpuset
*tmp_cs
;
1879 struct cgroup_subsys_state
*pos_css
;
1884 mutex_lock(&cpuset_mutex
);
1886 set_bit(CS_ONLINE
, &cs
->flags
);
1887 if (is_spread_page(parent
))
1888 set_bit(CS_SPREAD_PAGE
, &cs
->flags
);
1889 if (is_spread_slab(parent
))
1890 set_bit(CS_SPREAD_SLAB
, &cs
->flags
);
1892 number_of_cpusets
++;
1894 if (!test_bit(CGRP_CPUSET_CLONE_CHILDREN
, &css
->cgroup
->flags
))
1898 * Clone @parent's configuration if CGRP_CPUSET_CLONE_CHILDREN is
1899 * set. This flag handling is implemented in cgroup core for
1900 * histrical reasons - the flag may be specified during mount.
1902 * Currently, if any sibling cpusets have exclusive cpus or mem, we
1903 * refuse to clone the configuration - thereby refusing the task to
1904 * be entered, and as a result refusing the sys_unshare() or
1905 * clone() which initiated it. If this becomes a problem for some
1906 * users who wish to allow that scenario, then this could be
1907 * changed to grant parent->cpus_allowed-sibling_cpus_exclusive
1908 * (and likewise for mems) to the new cgroup.
1911 cpuset_for_each_child(tmp_cs
, pos_css
, parent
) {
1912 if (is_mem_exclusive(tmp_cs
) || is_cpu_exclusive(tmp_cs
)) {
1919 mutex_lock(&callback_mutex
);
1920 cs
->mems_allowed
= parent
->mems_allowed
;
1921 cpumask_copy(cs
->cpus_allowed
, parent
->cpus_allowed
);
1922 mutex_unlock(&callback_mutex
);
1924 mutex_unlock(&cpuset_mutex
);
1929 * If the cpuset being removed has its flag 'sched_load_balance'
1930 * enabled, then simulate turning sched_load_balance off, which
1931 * will call rebuild_sched_domains_locked().
1934 static void cpuset_css_offline(struct cgroup_subsys_state
*css
)
1936 struct cpuset
*cs
= css_cs(css
);
1938 mutex_lock(&cpuset_mutex
);
1940 if (is_sched_load_balance(cs
))
1941 update_flag(CS_SCHED_LOAD_BALANCE
, cs
, 0);
1943 number_of_cpusets
--;
1944 clear_bit(CS_ONLINE
, &cs
->flags
);
1946 mutex_unlock(&cpuset_mutex
);
1949 static void cpuset_css_free(struct cgroup_subsys_state
*css
)
1951 struct cpuset
*cs
= css_cs(css
);
1953 free_cpumask_var(cs
->cpus_allowed
);
1957 struct cgroup_subsys cpuset_cgrp_subsys
= {
1958 .css_alloc
= cpuset_css_alloc
,
1959 .css_online
= cpuset_css_online
,
1960 .css_offline
= cpuset_css_offline
,
1961 .css_free
= cpuset_css_free
,
1962 .can_attach
= cpuset_can_attach
,
1963 .cancel_attach
= cpuset_cancel_attach
,
1964 .attach
= cpuset_attach
,
1965 .base_cftypes
= files
,
1970 * cpuset_init - initialize cpusets at system boot
1972 * Description: Initialize top_cpuset and the cpuset internal file system,
1975 int __init
cpuset_init(void)
1979 if (!alloc_cpumask_var(&top_cpuset
.cpus_allowed
, GFP_KERNEL
))
1982 cpumask_setall(top_cpuset
.cpus_allowed
);
1983 nodes_setall(top_cpuset
.mems_allowed
);
1985 fmeter_init(&top_cpuset
.fmeter
);
1986 set_bit(CS_SCHED_LOAD_BALANCE
, &top_cpuset
.flags
);
1987 top_cpuset
.relax_domain_level
= -1;
1989 err
= register_filesystem(&cpuset_fs_type
);
1993 if (!alloc_cpumask_var(&cpus_attach
, GFP_KERNEL
))
1996 number_of_cpusets
= 1;
2001 * If CPU and/or memory hotplug handlers, below, unplug any CPUs
2002 * or memory nodes, we need to walk over the cpuset hierarchy,
2003 * removing that CPU or node from all cpusets. If this removes the
2004 * last CPU or node from a cpuset, then move the tasks in the empty
2005 * cpuset to its next-highest non-empty parent.
2007 static void remove_tasks_in_empty_cpuset(struct cpuset
*cs
)
2009 struct cpuset
*parent
;
2012 * Find its next-highest non-empty parent, (top cpuset
2013 * has online cpus, so can't be empty).
2015 parent
= parent_cs(cs
);
2016 while (cpumask_empty(parent
->cpus_allowed
) ||
2017 nodes_empty(parent
->mems_allowed
))
2018 parent
= parent_cs(parent
);
2020 if (cgroup_transfer_tasks(parent
->css
.cgroup
, cs
->css
.cgroup
)) {
2021 printk(KERN_ERR
"cpuset: failed to transfer tasks out of empty cpuset ");
2022 pr_cont_cgroup_name(cs
->css
.cgroup
);
2028 * cpuset_hotplug_update_tasks - update tasks in a cpuset for hotunplug
2029 * @cs: cpuset in interest
2031 * Compare @cs's cpu and mem masks against top_cpuset and if some have gone
2032 * offline, update @cs accordingly. If @cs ends up with no CPU or memory,
2033 * all its tasks are moved to the nearest ancestor with both resources.
2035 static void cpuset_hotplug_update_tasks(struct cpuset
*cs
)
2037 static cpumask_t off_cpus
;
2038 static nodemask_t off_mems
;
2040 bool sane
= cgroup_sane_behavior(cs
->css
.cgroup
);
2043 wait_event(cpuset_attach_wq
, cs
->attach_in_progress
== 0);
2045 mutex_lock(&cpuset_mutex
);
2048 * We have raced with task attaching. We wait until attaching
2049 * is finished, so we won't attach a task to an empty cpuset.
2051 if (cs
->attach_in_progress
) {
2052 mutex_unlock(&cpuset_mutex
);
2056 cpumask_andnot(&off_cpus
, cs
->cpus_allowed
, top_cpuset
.cpus_allowed
);
2057 nodes_andnot(off_mems
, cs
->mems_allowed
, top_cpuset
.mems_allowed
);
2059 mutex_lock(&callback_mutex
);
2060 cpumask_andnot(cs
->cpus_allowed
, cs
->cpus_allowed
, &off_cpus
);
2061 mutex_unlock(&callback_mutex
);
2064 * If sane_behavior flag is set, we need to update tasks' cpumask
2065 * for empty cpuset to take on ancestor's cpumask. Otherwise, don't
2066 * call update_tasks_cpumask() if the cpuset becomes empty, as
2067 * the tasks in it will be migrated to an ancestor.
2069 if ((sane
&& cpumask_empty(cs
->cpus_allowed
)) ||
2070 (!cpumask_empty(&off_cpus
) && !cpumask_empty(cs
->cpus_allowed
)))
2071 update_tasks_cpumask(cs
);
2073 mutex_lock(&callback_mutex
);
2074 nodes_andnot(cs
->mems_allowed
, cs
->mems_allowed
, off_mems
);
2075 mutex_unlock(&callback_mutex
);
2078 * If sane_behavior flag is set, we need to update tasks' nodemask
2079 * for empty cpuset to take on ancestor's nodemask. Otherwise, don't
2080 * call update_tasks_nodemask() if the cpuset becomes empty, as
2081 * the tasks in it will be migratd to an ancestor.
2083 if ((sane
&& nodes_empty(cs
->mems_allowed
)) ||
2084 (!nodes_empty(off_mems
) && !nodes_empty(cs
->mems_allowed
)))
2085 update_tasks_nodemask(cs
);
2087 is_empty
= cpumask_empty(cs
->cpus_allowed
) ||
2088 nodes_empty(cs
->mems_allowed
);
2090 mutex_unlock(&cpuset_mutex
);
2093 * If sane_behavior flag is set, we'll keep tasks in empty cpusets.
2095 * Otherwise move tasks to the nearest ancestor with execution
2096 * resources. This is full cgroup operation which will
2097 * also call back into cpuset. Should be done outside any lock.
2099 if (!sane
&& is_empty
)
2100 remove_tasks_in_empty_cpuset(cs
);
2104 * cpuset_hotplug_workfn - handle CPU/memory hotunplug for a cpuset
2106 * This function is called after either CPU or memory configuration has
2107 * changed and updates cpuset accordingly. The top_cpuset is always
2108 * synchronized to cpu_active_mask and N_MEMORY, which is necessary in
2109 * order to make cpusets transparent (of no affect) on systems that are
2110 * actively using CPU hotplug but making no active use of cpusets.
2112 * Non-root cpusets are only affected by offlining. If any CPUs or memory
2113 * nodes have been taken down, cpuset_hotplug_update_tasks() is invoked on
2116 * Note that CPU offlining during suspend is ignored. We don't modify
2117 * cpusets across suspend/resume cycles at all.
2119 static void cpuset_hotplug_workfn(struct work_struct
*work
)
2121 static cpumask_t new_cpus
;
2122 static nodemask_t new_mems
;
2123 bool cpus_updated
, mems_updated
;
2125 mutex_lock(&cpuset_mutex
);
2127 /* fetch the available cpus/mems and find out which changed how */
2128 cpumask_copy(&new_cpus
, cpu_active_mask
);
2129 new_mems
= node_states
[N_MEMORY
];
2131 cpus_updated
= !cpumask_equal(top_cpuset
.cpus_allowed
, &new_cpus
);
2132 mems_updated
= !nodes_equal(top_cpuset
.mems_allowed
, new_mems
);
2134 /* synchronize cpus_allowed to cpu_active_mask */
2136 mutex_lock(&callback_mutex
);
2137 cpumask_copy(top_cpuset
.cpus_allowed
, &new_cpus
);
2138 mutex_unlock(&callback_mutex
);
2139 /* we don't mess with cpumasks of tasks in top_cpuset */
2142 /* synchronize mems_allowed to N_MEMORY */
2144 mutex_lock(&callback_mutex
);
2145 top_cpuset
.mems_allowed
= new_mems
;
2146 mutex_unlock(&callback_mutex
);
2147 update_tasks_nodemask(&top_cpuset
);
2150 mutex_unlock(&cpuset_mutex
);
2152 /* if cpus or mems changed, we need to propagate to descendants */
2153 if (cpus_updated
|| mems_updated
) {
2155 struct cgroup_subsys_state
*pos_css
;
2158 cpuset_for_each_descendant_pre(cs
, pos_css
, &top_cpuset
) {
2159 if (cs
== &top_cpuset
|| !css_tryget(&cs
->css
))
2163 cpuset_hotplug_update_tasks(cs
);
2171 /* rebuild sched domains if cpus_allowed has changed */
2173 rebuild_sched_domains();
2176 void cpuset_update_active_cpus(bool cpu_online
)
2179 * We're inside cpu hotplug critical region which usually nests
2180 * inside cgroup synchronization. Bounce actual hotplug processing
2181 * to a work item to avoid reverse locking order.
2183 * We still need to do partition_sched_domains() synchronously;
2184 * otherwise, the scheduler will get confused and put tasks to the
2185 * dead CPU. Fall back to the default single domain.
2186 * cpuset_hotplug_workfn() will rebuild it as necessary.
2188 partition_sched_domains(1, NULL
, NULL
);
2189 schedule_work(&cpuset_hotplug_work
);
2193 * Keep top_cpuset.mems_allowed tracking node_states[N_MEMORY].
2194 * Call this routine anytime after node_states[N_MEMORY] changes.
2195 * See cpuset_update_active_cpus() for CPU hotplug handling.
2197 static int cpuset_track_online_nodes(struct notifier_block
*self
,
2198 unsigned long action
, void *arg
)
2200 schedule_work(&cpuset_hotplug_work
);
2204 static struct notifier_block cpuset_track_online_nodes_nb
= {
2205 .notifier_call
= cpuset_track_online_nodes
,
2206 .priority
= 10, /* ??! */
2210 * cpuset_init_smp - initialize cpus_allowed
2212 * Description: Finish top cpuset after cpu, node maps are initialized
2214 void __init
cpuset_init_smp(void)
2216 cpumask_copy(top_cpuset
.cpus_allowed
, cpu_active_mask
);
2217 top_cpuset
.mems_allowed
= node_states
[N_MEMORY
];
2218 top_cpuset
.old_mems_allowed
= top_cpuset
.mems_allowed
;
2220 register_hotmemory_notifier(&cpuset_track_online_nodes_nb
);
2224 * cpuset_cpus_allowed - return cpus_allowed mask from a tasks cpuset.
2225 * @tsk: pointer to task_struct from which to obtain cpuset->cpus_allowed.
2226 * @pmask: pointer to struct cpumask variable to receive cpus_allowed set.
2228 * Description: Returns the cpumask_var_t cpus_allowed of the cpuset
2229 * attached to the specified @tsk. Guaranteed to return some non-empty
2230 * subset of cpu_online_mask, even if this means going outside the
2234 void cpuset_cpus_allowed(struct task_struct
*tsk
, struct cpumask
*pmask
)
2236 struct cpuset
*cpus_cs
;
2238 mutex_lock(&callback_mutex
);
2240 cpus_cs
= effective_cpumask_cpuset(task_cs(tsk
));
2241 guarantee_online_cpus(cpus_cs
, pmask
);
2243 mutex_unlock(&callback_mutex
);
2246 void cpuset_cpus_allowed_fallback(struct task_struct
*tsk
)
2248 struct cpuset
*cpus_cs
;
2251 cpus_cs
= effective_cpumask_cpuset(task_cs(tsk
));
2252 do_set_cpus_allowed(tsk
, cpus_cs
->cpus_allowed
);
2256 * We own tsk->cpus_allowed, nobody can change it under us.
2258 * But we used cs && cs->cpus_allowed lockless and thus can
2259 * race with cgroup_attach_task() or update_cpumask() and get
2260 * the wrong tsk->cpus_allowed. However, both cases imply the
2261 * subsequent cpuset_change_cpumask()->set_cpus_allowed_ptr()
2262 * which takes task_rq_lock().
2264 * If we are called after it dropped the lock we must see all
2265 * changes in tsk_cs()->cpus_allowed. Otherwise we can temporary
2266 * set any mask even if it is not right from task_cs() pov,
2267 * the pending set_cpus_allowed_ptr() will fix things.
2269 * select_fallback_rq() will fix things ups and set cpu_possible_mask
2274 void cpuset_init_current_mems_allowed(void)
2276 nodes_setall(current
->mems_allowed
);
2280 * cpuset_mems_allowed - return mems_allowed mask from a tasks cpuset.
2281 * @tsk: pointer to task_struct from which to obtain cpuset->mems_allowed.
2283 * Description: Returns the nodemask_t mems_allowed of the cpuset
2284 * attached to the specified @tsk. Guaranteed to return some non-empty
2285 * subset of node_states[N_MEMORY], even if this means going outside the
2289 nodemask_t
cpuset_mems_allowed(struct task_struct
*tsk
)
2291 struct cpuset
*mems_cs
;
2294 mutex_lock(&callback_mutex
);
2296 mems_cs
= effective_nodemask_cpuset(task_cs(tsk
));
2297 guarantee_online_mems(mems_cs
, &mask
);
2299 mutex_unlock(&callback_mutex
);
2305 * cpuset_nodemask_valid_mems_allowed - check nodemask vs. curremt mems_allowed
2306 * @nodemask: the nodemask to be checked
2308 * Are any of the nodes in the nodemask allowed in current->mems_allowed?
2310 int cpuset_nodemask_valid_mems_allowed(nodemask_t
*nodemask
)
2312 return nodes_intersects(*nodemask
, current
->mems_allowed
);
2316 * nearest_hardwall_ancestor() - Returns the nearest mem_exclusive or
2317 * mem_hardwall ancestor to the specified cpuset. Call holding
2318 * callback_mutex. If no ancestor is mem_exclusive or mem_hardwall
2319 * (an unusual configuration), then returns the root cpuset.
2321 static struct cpuset
*nearest_hardwall_ancestor(struct cpuset
*cs
)
2323 while (!(is_mem_exclusive(cs
) || is_mem_hardwall(cs
)) && parent_cs(cs
))
2329 * cpuset_node_allowed_softwall - Can we allocate on a memory node?
2330 * @node: is this an allowed node?
2331 * @gfp_mask: memory allocation flags
2333 * If we're in interrupt, yes, we can always allocate. If __GFP_THISNODE is
2334 * set, yes, we can always allocate. If node is in our task's mems_allowed,
2335 * yes. If it's not a __GFP_HARDWALL request and this node is in the nearest
2336 * hardwalled cpuset ancestor to this task's cpuset, yes. If the task has been
2337 * OOM killed and has access to memory reserves as specified by the TIF_MEMDIE
2341 * If __GFP_HARDWALL is set, cpuset_node_allowed_softwall() reduces to
2342 * cpuset_node_allowed_hardwall(). Otherwise, cpuset_node_allowed_softwall()
2343 * might sleep, and might allow a node from an enclosing cpuset.
2345 * cpuset_node_allowed_hardwall() only handles the simpler case of hardwall
2346 * cpusets, and never sleeps.
2348 * The __GFP_THISNODE placement logic is really handled elsewhere,
2349 * by forcibly using a zonelist starting at a specified node, and by
2350 * (in get_page_from_freelist()) refusing to consider the zones for
2351 * any node on the zonelist except the first. By the time any such
2352 * calls get to this routine, we should just shut up and say 'yes'.
2354 * GFP_USER allocations are marked with the __GFP_HARDWALL bit,
2355 * and do not allow allocations outside the current tasks cpuset
2356 * unless the task has been OOM killed as is marked TIF_MEMDIE.
2357 * GFP_KERNEL allocations are not so marked, so can escape to the
2358 * nearest enclosing hardwalled ancestor cpuset.
2360 * Scanning up parent cpusets requires callback_mutex. The
2361 * __alloc_pages() routine only calls here with __GFP_HARDWALL bit
2362 * _not_ set if it's a GFP_KERNEL allocation, and all nodes in the
2363 * current tasks mems_allowed came up empty on the first pass over
2364 * the zonelist. So only GFP_KERNEL allocations, if all nodes in the
2365 * cpuset are short of memory, might require taking the callback_mutex
2368 * The first call here from mm/page_alloc:get_page_from_freelist()
2369 * has __GFP_HARDWALL set in gfp_mask, enforcing hardwall cpusets,
2370 * so no allocation on a node outside the cpuset is allowed (unless
2371 * in interrupt, of course).
2373 * The second pass through get_page_from_freelist() doesn't even call
2374 * here for GFP_ATOMIC calls. For those calls, the __alloc_pages()
2375 * variable 'wait' is not set, and the bit ALLOC_CPUSET is not set
2376 * in alloc_flags. That logic and the checks below have the combined
2378 * in_interrupt - any node ok (current task context irrelevant)
2379 * GFP_ATOMIC - any node ok
2380 * TIF_MEMDIE - any node ok
2381 * GFP_KERNEL - any node in enclosing hardwalled cpuset ok
2382 * GFP_USER - only nodes in current tasks mems allowed ok.
2385 * Don't call cpuset_node_allowed_softwall if you can't sleep, unless you
2386 * pass in the __GFP_HARDWALL flag set in gfp_flag, which disables
2387 * the code that might scan up ancestor cpusets and sleep.
2389 int __cpuset_node_allowed_softwall(int node
, gfp_t gfp_mask
)
2391 struct cpuset
*cs
; /* current cpuset ancestors */
2392 int allowed
; /* is allocation in zone z allowed? */
2394 if (in_interrupt() || (gfp_mask
& __GFP_THISNODE
))
2396 might_sleep_if(!(gfp_mask
& __GFP_HARDWALL
));
2397 if (node_isset(node
, current
->mems_allowed
))
2400 * Allow tasks that have access to memory reserves because they have
2401 * been OOM killed to get memory anywhere.
2403 if (unlikely(test_thread_flag(TIF_MEMDIE
)))
2405 if (gfp_mask
& __GFP_HARDWALL
) /* If hardwall request, stop here */
2408 if (current
->flags
& PF_EXITING
) /* Let dying task have memory */
2411 /* Not hardwall and node outside mems_allowed: scan up cpusets */
2412 mutex_lock(&callback_mutex
);
2415 cs
= nearest_hardwall_ancestor(task_cs(current
));
2416 task_unlock(current
);
2418 allowed
= node_isset(node
, cs
->mems_allowed
);
2419 mutex_unlock(&callback_mutex
);
2424 * cpuset_node_allowed_hardwall - Can we allocate on a memory node?
2425 * @node: is this an allowed node?
2426 * @gfp_mask: memory allocation flags
2428 * If we're in interrupt, yes, we can always allocate. If __GFP_THISNODE is
2429 * set, yes, we can always allocate. If node is in our task's mems_allowed,
2430 * yes. If the task has been OOM killed and has access to memory reserves as
2431 * specified by the TIF_MEMDIE flag, yes.
2434 * The __GFP_THISNODE placement logic is really handled elsewhere,
2435 * by forcibly using a zonelist starting at a specified node, and by
2436 * (in get_page_from_freelist()) refusing to consider the zones for
2437 * any node on the zonelist except the first. By the time any such
2438 * calls get to this routine, we should just shut up and say 'yes'.
2440 * Unlike the cpuset_node_allowed_softwall() variant, above,
2441 * this variant requires that the node be in the current task's
2442 * mems_allowed or that we're in interrupt. It does not scan up the
2443 * cpuset hierarchy for the nearest enclosing mem_exclusive cpuset.
2446 int __cpuset_node_allowed_hardwall(int node
, gfp_t gfp_mask
)
2448 if (in_interrupt() || (gfp_mask
& __GFP_THISNODE
))
2450 if (node_isset(node
, current
->mems_allowed
))
2453 * Allow tasks that have access to memory reserves because they have
2454 * been OOM killed to get memory anywhere.
2456 if (unlikely(test_thread_flag(TIF_MEMDIE
)))
2462 * cpuset_mem_spread_node() - On which node to begin search for a file page
2463 * cpuset_slab_spread_node() - On which node to begin search for a slab page
2465 * If a task is marked PF_SPREAD_PAGE or PF_SPREAD_SLAB (as for
2466 * tasks in a cpuset with is_spread_page or is_spread_slab set),
2467 * and if the memory allocation used cpuset_mem_spread_node()
2468 * to determine on which node to start looking, as it will for
2469 * certain page cache or slab cache pages such as used for file
2470 * system buffers and inode caches, then instead of starting on the
2471 * local node to look for a free page, rather spread the starting
2472 * node around the tasks mems_allowed nodes.
2474 * We don't have to worry about the returned node being offline
2475 * because "it can't happen", and even if it did, it would be ok.
2477 * The routines calling guarantee_online_mems() are careful to
2478 * only set nodes in task->mems_allowed that are online. So it
2479 * should not be possible for the following code to return an
2480 * offline node. But if it did, that would be ok, as this routine
2481 * is not returning the node where the allocation must be, only
2482 * the node where the search should start. The zonelist passed to
2483 * __alloc_pages() will include all nodes. If the slab allocator
2484 * is passed an offline node, it will fall back to the local node.
2485 * See kmem_cache_alloc_node().
2488 static int cpuset_spread_node(int *rotor
)
2492 node
= next_node(*rotor
, current
->mems_allowed
);
2493 if (node
== MAX_NUMNODES
)
2494 node
= first_node(current
->mems_allowed
);
2499 int cpuset_mem_spread_node(void)
2501 if (current
->cpuset_mem_spread_rotor
== NUMA_NO_NODE
)
2502 current
->cpuset_mem_spread_rotor
=
2503 node_random(¤t
->mems_allowed
);
2505 return cpuset_spread_node(¤t
->cpuset_mem_spread_rotor
);
2508 int cpuset_slab_spread_node(void)
2510 if (current
->cpuset_slab_spread_rotor
== NUMA_NO_NODE
)
2511 current
->cpuset_slab_spread_rotor
=
2512 node_random(¤t
->mems_allowed
);
2514 return cpuset_spread_node(¤t
->cpuset_slab_spread_rotor
);
2517 EXPORT_SYMBOL_GPL(cpuset_mem_spread_node
);
2520 * cpuset_mems_allowed_intersects - Does @tsk1's mems_allowed intersect @tsk2's?
2521 * @tsk1: pointer to task_struct of some task.
2522 * @tsk2: pointer to task_struct of some other task.
2524 * Description: Return true if @tsk1's mems_allowed intersects the
2525 * mems_allowed of @tsk2. Used by the OOM killer to determine if
2526 * one of the task's memory usage might impact the memory available
2530 int cpuset_mems_allowed_intersects(const struct task_struct
*tsk1
,
2531 const struct task_struct
*tsk2
)
2533 return nodes_intersects(tsk1
->mems_allowed
, tsk2
->mems_allowed
);
2536 #define CPUSET_NODELIST_LEN (256)
2539 * cpuset_print_task_mems_allowed - prints task's cpuset and mems_allowed
2540 * @task: pointer to task_struct of some task.
2542 * Description: Prints @task's name, cpuset name, and cached copy of its
2543 * mems_allowed to the kernel log. Must hold task_lock(task) to allow
2544 * dereferencing task_cs(task).
2546 void cpuset_print_task_mems_allowed(struct task_struct
*tsk
)
2548 /* Statically allocated to prevent using excess stack. */
2549 static char cpuset_nodelist
[CPUSET_NODELIST_LEN
];
2550 static DEFINE_SPINLOCK(cpuset_buffer_lock
);
2551 struct cgroup
*cgrp
= task_cs(tsk
)->css
.cgroup
;
2553 spin_lock(&cpuset_buffer_lock
);
2555 nodelist_scnprintf(cpuset_nodelist
, CPUSET_NODELIST_LEN
,
2557 printk(KERN_INFO
"%s cpuset=", tsk
->comm
);
2558 pr_cont_cgroup_name(cgrp
);
2559 pr_cont(" mems_allowed=%s\n", cpuset_nodelist
);
2561 spin_unlock(&cpuset_buffer_lock
);
2565 * Collection of memory_pressure is suppressed unless
2566 * this flag is enabled by writing "1" to the special
2567 * cpuset file 'memory_pressure_enabled' in the root cpuset.
2570 int cpuset_memory_pressure_enabled __read_mostly
;
2573 * cpuset_memory_pressure_bump - keep stats of per-cpuset reclaims.
2575 * Keep a running average of the rate of synchronous (direct)
2576 * page reclaim efforts initiated by tasks in each cpuset.
2578 * This represents the rate at which some task in the cpuset
2579 * ran low on memory on all nodes it was allowed to use, and
2580 * had to enter the kernels page reclaim code in an effort to
2581 * create more free memory by tossing clean pages or swapping
2582 * or writing dirty pages.
2584 * Display to user space in the per-cpuset read-only file
2585 * "memory_pressure". Value displayed is an integer
2586 * representing the recent rate of entry into the synchronous
2587 * (direct) page reclaim by any task attached to the cpuset.
2590 void __cpuset_memory_pressure_bump(void)
2593 fmeter_markevent(&task_cs(current
)->fmeter
);
2594 task_unlock(current
);
2597 #ifdef CONFIG_PROC_PID_CPUSET
2599 * proc_cpuset_show()
2600 * - Print tasks cpuset path into seq_file.
2601 * - Used for /proc/<pid>/cpuset.
2602 * - No need to task_lock(tsk) on this tsk->cpuset reference, as it
2603 * doesn't really matter if tsk->cpuset changes after we read it,
2604 * and we take cpuset_mutex, keeping cpuset_attach() from changing it
2607 int proc_cpuset_show(struct seq_file
*m
, void *unused_v
)
2610 struct task_struct
*tsk
;
2612 struct cgroup_subsys_state
*css
;
2616 buf
= kmalloc(PATH_MAX
, GFP_KERNEL
);
2622 tsk
= get_pid_task(pid
, PIDTYPE_PID
);
2626 retval
= -ENAMETOOLONG
;
2628 css
= task_css(tsk
, cpuset_cgrp_id
);
2629 p
= cgroup_path(css
->cgroup
, buf
, PATH_MAX
);
2637 put_task_struct(tsk
);
2643 #endif /* CONFIG_PROC_PID_CPUSET */
2645 /* Display task mems_allowed in /proc/<pid>/status file. */
2646 void cpuset_task_status_allowed(struct seq_file
*m
, struct task_struct
*task
)
2648 seq_printf(m
, "Mems_allowed:\t");
2649 seq_nodemask(m
, &task
->mems_allowed
);
2650 seq_printf(m
, "\n");
2651 seq_printf(m
, "Mems_allowed_list:\t");
2652 seq_nodemask_list(m
, &task
->mems_allowed
);
2653 seq_printf(m
, "\n");