4 * Processor and Memory placement constraints for sets of tasks.
6 * Copyright (C) 2003 BULL SA.
7 * Copyright (C) 2004-2007 Silicon Graphics, Inc.
8 * Copyright (C) 2006 Google, Inc
10 * Portions derived from Patrick Mochel's sysfs code.
11 * sysfs is Copyright (c) 2001-3 Patrick Mochel
13 * 2003-10-10 Written by Simon Derr.
14 * 2003-10-22 Updates by Stephen Hemminger.
15 * 2004 May-July Rework by Paul Jackson.
16 * 2006 Rework by Paul Menage to use generic cgroups
17 * 2008 Rework of the scheduler domains and CPU hotplug handling
20 * This file is subject to the terms and conditions of the GNU General Public
21 * License. See the file COPYING in the main directory of the Linux
22 * distribution for more details.
25 #include <linux/cpu.h>
26 #include <linux/cpumask.h>
27 #include <linux/cpuset.h>
28 #include <linux/err.h>
29 #include <linux/errno.h>
30 #include <linux/file.h>
32 #include <linux/init.h>
33 #include <linux/interrupt.h>
34 #include <linux/kernel.h>
35 #include <linux/kmod.h>
36 #include <linux/list.h>
37 #include <linux/mempolicy.h>
39 #include <linux/memory.h>
40 #include <linux/export.h>
41 #include <linux/mount.h>
42 #include <linux/namei.h>
43 #include <linux/pagemap.h>
44 #include <linux/proc_fs.h>
45 #include <linux/rcupdate.h>
46 #include <linux/sched.h>
47 #include <linux/seq_file.h>
48 #include <linux/security.h>
49 #include <linux/slab.h>
50 #include <linux/spinlock.h>
51 #include <linux/stat.h>
52 #include <linux/string.h>
53 #include <linux/time.h>
54 #include <linux/backing-dev.h>
55 #include <linux/sort.h>
57 #include <asm/uaccess.h>
58 #include <linux/atomic.h>
59 #include <linux/mutex.h>
60 #include <linux/workqueue.h>
61 #include <linux/cgroup.h>
64 * Tracks how many cpusets are currently defined in system.
65 * When there is only one cpuset (the root cpuset) we can
66 * short circuit some hooks.
68 int number_of_cpusets __read_mostly
;
70 /* Forward declare cgroup structures */
71 struct cgroup_subsys cpuset_subsys
;
74 /* See "Frequency meter" comments, below. */
77 int cnt
; /* unprocessed events count */
78 int val
; /* most recent output value */
79 time_t time
; /* clock (secs) when val computed */
80 spinlock_t lock
; /* guards read or write of above */
84 struct cgroup_subsys_state css
;
86 unsigned long flags
; /* "unsigned long" so bitops work */
87 cpumask_var_t cpus_allowed
; /* CPUs allowed to tasks in cpuset */
88 nodemask_t mems_allowed
; /* Memory Nodes allowed to tasks */
90 struct cpuset
*parent
; /* my parent */
92 struct fmeter fmeter
; /* memory_pressure filter */
95 * Tasks are being attached to this cpuset. Used to prevent
96 * zeroing cpus/mems_allowed between ->can_attach() and ->attach().
98 int attach_in_progress
;
100 /* partition number for rebuild_sched_domains() */
103 /* for custom sched domain */
104 int relax_domain_level
;
106 /* used for walking a cpuset hierarchy */
107 struct list_head stack_list
;
109 struct work_struct hotplug_work
;
112 /* Retrieve the cpuset for a cgroup */
113 static inline struct cpuset
*cgroup_cs(struct cgroup
*cont
)
115 return container_of(cgroup_subsys_state(cont
, cpuset_subsys_id
),
119 /* Retrieve the cpuset for a task */
120 static inline struct cpuset
*task_cs(struct task_struct
*task
)
122 return container_of(task_subsys_state(task
, cpuset_subsys_id
),
127 static inline bool task_has_mempolicy(struct task_struct
*task
)
129 return task
->mempolicy
;
132 static inline bool task_has_mempolicy(struct task_struct
*task
)
139 /* bits in struct cpuset flags field */
146 CS_SCHED_LOAD_BALANCE
,
151 /* convenient tests for these bits */
152 static inline bool is_cpuset_online(const struct cpuset
*cs
)
154 return test_bit(CS_ONLINE
, &cs
->flags
);
157 static inline int is_cpu_exclusive(const struct cpuset
*cs
)
159 return test_bit(CS_CPU_EXCLUSIVE
, &cs
->flags
);
162 static inline int is_mem_exclusive(const struct cpuset
*cs
)
164 return test_bit(CS_MEM_EXCLUSIVE
, &cs
->flags
);
167 static inline int is_mem_hardwall(const struct cpuset
*cs
)
169 return test_bit(CS_MEM_HARDWALL
, &cs
->flags
);
172 static inline int is_sched_load_balance(const struct cpuset
*cs
)
174 return test_bit(CS_SCHED_LOAD_BALANCE
, &cs
->flags
);
177 static inline int is_memory_migrate(const struct cpuset
*cs
)
179 return test_bit(CS_MEMORY_MIGRATE
, &cs
->flags
);
182 static inline int is_spread_page(const struct cpuset
*cs
)
184 return test_bit(CS_SPREAD_PAGE
, &cs
->flags
);
187 static inline int is_spread_slab(const struct cpuset
*cs
)
189 return test_bit(CS_SPREAD_SLAB
, &cs
->flags
);
192 static struct cpuset top_cpuset
= {
193 .flags
= ((1 << CS_ONLINE
) | (1 << CS_CPU_EXCLUSIVE
) |
194 (1 << CS_MEM_EXCLUSIVE
)),
198 * cpuset_for_each_child - traverse online children of a cpuset
199 * @child_cs: loop cursor pointing to the current child
200 * @pos_cgrp: used for iteration
201 * @parent_cs: target cpuset to walk children of
203 * Walk @child_cs through the online children of @parent_cs. Must be used
204 * with RCU read locked.
206 #define cpuset_for_each_child(child_cs, pos_cgrp, parent_cs) \
207 cgroup_for_each_child((pos_cgrp), (parent_cs)->css.cgroup) \
208 if (is_cpuset_online(((child_cs) = cgroup_cs((pos_cgrp)))))
211 * There are two global mutexes guarding cpuset structures. The first
212 * is the main control groups cgroup_mutex, accessed via
213 * cgroup_lock()/cgroup_unlock(). The second is the cpuset-specific
214 * callback_mutex, below. They can nest. It is ok to first take
215 * cgroup_mutex, then nest callback_mutex. We also require taking
216 * task_lock() when dereferencing a task's cpuset pointer. See "The
217 * task_lock() exception", at the end of this comment.
219 * A task must hold both mutexes to modify cpusets. If a task
220 * holds cgroup_mutex, then it blocks others wanting that mutex,
221 * ensuring that it is the only task able to also acquire callback_mutex
222 * and be able to modify cpusets. It can perform various checks on
223 * the cpuset structure first, knowing nothing will change. It can
224 * also allocate memory while just holding cgroup_mutex. While it is
225 * performing these checks, various callback routines can briefly
226 * acquire callback_mutex to query cpusets. Once it is ready to make
227 * the changes, it takes callback_mutex, blocking everyone else.
229 * Calls to the kernel memory allocator can not be made while holding
230 * callback_mutex, as that would risk double tripping on callback_mutex
231 * from one of the callbacks into the cpuset code from within
234 * If a task is only holding callback_mutex, then it has read-only
237 * Now, the task_struct fields mems_allowed and mempolicy may be changed
238 * by other task, we use alloc_lock in the task_struct fields to protect
241 * The cpuset_common_file_read() handlers only hold callback_mutex across
242 * small pieces of code, such as when reading out possibly multi-word
243 * cpumasks and nodemasks.
245 * Accessing a task's cpuset should be done in accordance with the
246 * guidelines for accessing subsystem state in kernel/cgroup.c
249 static DEFINE_MUTEX(callback_mutex
);
252 * cpuset_buffer_lock protects both the cpuset_name and cpuset_nodelist
253 * buffers. They are statically allocated to prevent using excess stack
254 * when calling cpuset_print_task_mems_allowed().
256 #define CPUSET_NAME_LEN (128)
257 #define CPUSET_NODELIST_LEN (256)
258 static char cpuset_name
[CPUSET_NAME_LEN
];
259 static char cpuset_nodelist
[CPUSET_NODELIST_LEN
];
260 static DEFINE_SPINLOCK(cpuset_buffer_lock
);
263 * CPU / memory hotplug is handled asynchronously.
265 static struct workqueue_struct
*cpuset_propagate_hotplug_wq
;
267 static void cpuset_hotplug_workfn(struct work_struct
*work
);
268 static void cpuset_propagate_hotplug_workfn(struct work_struct
*work
);
270 static DECLARE_WORK(cpuset_hotplug_work
, cpuset_hotplug_workfn
);
273 * This is ugly, but preserves the userspace API for existing cpuset
274 * users. If someone tries to mount the "cpuset" filesystem, we
275 * silently switch it to mount "cgroup" instead
277 static struct dentry
*cpuset_mount(struct file_system_type
*fs_type
,
278 int flags
, const char *unused_dev_name
, void *data
)
280 struct file_system_type
*cgroup_fs
= get_fs_type("cgroup");
281 struct dentry
*ret
= ERR_PTR(-ENODEV
);
285 "release_agent=/sbin/cpuset_release_agent";
286 ret
= cgroup_fs
->mount(cgroup_fs
, flags
,
287 unused_dev_name
, mountopts
);
288 put_filesystem(cgroup_fs
);
293 static struct file_system_type cpuset_fs_type
= {
295 .mount
= cpuset_mount
,
299 * Return in pmask the portion of a cpusets's cpus_allowed that
300 * are online. If none are online, walk up the cpuset hierarchy
301 * until we find one that does have some online cpus. If we get
302 * all the way to the top and still haven't found any online cpus,
303 * return cpu_online_mask. Or if passed a NULL cs from an exit'ing
304 * task, return cpu_online_mask.
306 * One way or another, we guarantee to return some non-empty subset
307 * of cpu_online_mask.
309 * Call with callback_mutex held.
312 static void guarantee_online_cpus(const struct cpuset
*cs
,
313 struct cpumask
*pmask
)
315 while (cs
&& !cpumask_intersects(cs
->cpus_allowed
, cpu_online_mask
))
318 cpumask_and(pmask
, cs
->cpus_allowed
, cpu_online_mask
);
320 cpumask_copy(pmask
, cpu_online_mask
);
321 BUG_ON(!cpumask_intersects(pmask
, cpu_online_mask
));
325 * Return in *pmask the portion of a cpusets's mems_allowed that
326 * are online, with memory. If none are online with memory, walk
327 * up the cpuset hierarchy until we find one that does have some
328 * online mems. If we get all the way to the top and still haven't
329 * found any online mems, return node_states[N_MEMORY].
331 * One way or another, we guarantee to return some non-empty subset
332 * of node_states[N_MEMORY].
334 * Call with callback_mutex held.
337 static void guarantee_online_mems(const struct cpuset
*cs
, nodemask_t
*pmask
)
339 while (cs
&& !nodes_intersects(cs
->mems_allowed
,
340 node_states
[N_MEMORY
]))
343 nodes_and(*pmask
, cs
->mems_allowed
,
344 node_states
[N_MEMORY
]);
346 *pmask
= node_states
[N_MEMORY
];
347 BUG_ON(!nodes_intersects(*pmask
, node_states
[N_MEMORY
]));
351 * update task's spread flag if cpuset's page/slab spread flag is set
353 * Called with callback_mutex/cgroup_mutex held
355 static void cpuset_update_task_spread_flag(struct cpuset
*cs
,
356 struct task_struct
*tsk
)
358 if (is_spread_page(cs
))
359 tsk
->flags
|= PF_SPREAD_PAGE
;
361 tsk
->flags
&= ~PF_SPREAD_PAGE
;
362 if (is_spread_slab(cs
))
363 tsk
->flags
|= PF_SPREAD_SLAB
;
365 tsk
->flags
&= ~PF_SPREAD_SLAB
;
369 * is_cpuset_subset(p, q) - Is cpuset p a subset of cpuset q?
371 * One cpuset is a subset of another if all its allowed CPUs and
372 * Memory Nodes are a subset of the other, and its exclusive flags
373 * are only set if the other's are set. Call holding cgroup_mutex.
376 static int is_cpuset_subset(const struct cpuset
*p
, const struct cpuset
*q
)
378 return cpumask_subset(p
->cpus_allowed
, q
->cpus_allowed
) &&
379 nodes_subset(p
->mems_allowed
, q
->mems_allowed
) &&
380 is_cpu_exclusive(p
) <= is_cpu_exclusive(q
) &&
381 is_mem_exclusive(p
) <= is_mem_exclusive(q
);
385 * alloc_trial_cpuset - allocate a trial cpuset
386 * @cs: the cpuset that the trial cpuset duplicates
388 static struct cpuset
*alloc_trial_cpuset(const struct cpuset
*cs
)
390 struct cpuset
*trial
;
392 trial
= kmemdup(cs
, sizeof(*cs
), GFP_KERNEL
);
396 if (!alloc_cpumask_var(&trial
->cpus_allowed
, GFP_KERNEL
)) {
400 cpumask_copy(trial
->cpus_allowed
, cs
->cpus_allowed
);
406 * free_trial_cpuset - free the trial cpuset
407 * @trial: the trial cpuset to be freed
409 static void free_trial_cpuset(struct cpuset
*trial
)
411 free_cpumask_var(trial
->cpus_allowed
);
416 * validate_change() - Used to validate that any proposed cpuset change
417 * follows the structural rules for cpusets.
419 * If we replaced the flag and mask values of the current cpuset
420 * (cur) with those values in the trial cpuset (trial), would
421 * our various subset and exclusive rules still be valid? Presumes
424 * 'cur' is the address of an actual, in-use cpuset. Operations
425 * such as list traversal that depend on the actual address of the
426 * cpuset in the list must use cur below, not trial.
428 * 'trial' is the address of bulk structure copy of cur, with
429 * perhaps one or more of the fields cpus_allowed, mems_allowed,
430 * or flags changed to new, trial values.
432 * Return 0 if valid, -errno if not.
435 static int validate_change(const struct cpuset
*cur
, const struct cpuset
*trial
)
438 struct cpuset
*c
, *par
;
443 /* Each of our child cpusets must be a subset of us */
445 cpuset_for_each_child(c
, cont
, cur
)
446 if (!is_cpuset_subset(c
, trial
))
449 /* Remaining checks don't apply to root cpuset */
451 if (cur
== &top_cpuset
)
456 /* We must be a subset of our parent cpuset */
458 if (!is_cpuset_subset(trial
, par
))
462 * If either I or some sibling (!= me) is exclusive, we can't
466 cpuset_for_each_child(c
, cont
, par
) {
467 if ((is_cpu_exclusive(trial
) || is_cpu_exclusive(c
)) &&
469 cpumask_intersects(trial
->cpus_allowed
, c
->cpus_allowed
))
471 if ((is_mem_exclusive(trial
) || is_mem_exclusive(c
)) &&
473 nodes_intersects(trial
->mems_allowed
, c
->mems_allowed
))
478 * Cpusets with tasks - existing or newly being attached - can't
479 * have empty cpus_allowed or mems_allowed.
482 if ((cgroup_task_count(cur
->css
.cgroup
) || cur
->attach_in_progress
) &&
483 (cpumask_empty(trial
->cpus_allowed
) ||
484 nodes_empty(trial
->mems_allowed
)))
495 * Helper routine for generate_sched_domains().
496 * Do cpusets a, b have overlapping cpus_allowed masks?
498 static int cpusets_overlap(struct cpuset
*a
, struct cpuset
*b
)
500 return cpumask_intersects(a
->cpus_allowed
, b
->cpus_allowed
);
504 update_domain_attr(struct sched_domain_attr
*dattr
, struct cpuset
*c
)
506 if (dattr
->relax_domain_level
< c
->relax_domain_level
)
507 dattr
->relax_domain_level
= c
->relax_domain_level
;
512 update_domain_attr_tree(struct sched_domain_attr
*dattr
, struct cpuset
*c
)
516 list_add(&c
->stack_list
, &q
);
517 while (!list_empty(&q
)) {
520 struct cpuset
*child
;
522 cp
= list_first_entry(&q
, struct cpuset
, stack_list
);
525 if (cpumask_empty(cp
->cpus_allowed
))
528 if (is_sched_load_balance(cp
))
529 update_domain_attr(dattr
, cp
);
532 cpuset_for_each_child(child
, cont
, cp
)
533 list_add_tail(&child
->stack_list
, &q
);
539 * generate_sched_domains()
541 * This function builds a partial partition of the systems CPUs
542 * A 'partial partition' is a set of non-overlapping subsets whose
543 * union is a subset of that set.
544 * The output of this function needs to be passed to kernel/sched.c
545 * partition_sched_domains() routine, which will rebuild the scheduler's
546 * load balancing domains (sched domains) as specified by that partial
549 * See "What is sched_load_balance" in Documentation/cgroups/cpusets.txt
550 * for a background explanation of this.
552 * Does not return errors, on the theory that the callers of this
553 * routine would rather not worry about failures to rebuild sched
554 * domains when operating in the severe memory shortage situations
555 * that could cause allocation failures below.
557 * Must be called with cgroup_lock held.
559 * The three key local variables below are:
560 * q - a linked-list queue of cpuset pointers, used to implement a
561 * top-down scan of all cpusets. This scan loads a pointer
562 * to each cpuset marked is_sched_load_balance into the
563 * array 'csa'. For our purposes, rebuilding the schedulers
564 * sched domains, we can ignore !is_sched_load_balance cpusets.
565 * csa - (for CpuSet Array) Array of pointers to all the cpusets
566 * that need to be load balanced, for convenient iterative
567 * access by the subsequent code that finds the best partition,
568 * i.e the set of domains (subsets) of CPUs such that the
569 * cpus_allowed of every cpuset marked is_sched_load_balance
570 * is a subset of one of these domains, while there are as
571 * many such domains as possible, each as small as possible.
572 * doms - Conversion of 'csa' to an array of cpumasks, for passing to
573 * the kernel/sched.c routine partition_sched_domains() in a
574 * convenient format, that can be easily compared to the prior
575 * value to determine what partition elements (sched domains)
576 * were changed (added or removed.)
578 * Finding the best partition (set of domains):
579 * The triple nested loops below over i, j, k scan over the
580 * load balanced cpusets (using the array of cpuset pointers in
581 * csa[]) looking for pairs of cpusets that have overlapping
582 * cpus_allowed, but which don't have the same 'pn' partition
583 * number and gives them in the same partition number. It keeps
584 * looping on the 'restart' label until it can no longer find
587 * The union of the cpus_allowed masks from the set of
588 * all cpusets having the same 'pn' value then form the one
589 * element of the partition (one sched domain) to be passed to
590 * partition_sched_domains().
592 static int generate_sched_domains(cpumask_var_t
**domains
,
593 struct sched_domain_attr
**attributes
)
595 LIST_HEAD(q
); /* queue of cpusets to be scanned */
596 struct cpuset
*cp
; /* scans q */
597 struct cpuset
**csa
; /* array of all cpuset ptrs */
598 int csn
; /* how many cpuset ptrs in csa so far */
599 int i
, j
, k
; /* indices for partition finding loops */
600 cpumask_var_t
*doms
; /* resulting partition; i.e. sched domains */
601 struct sched_domain_attr
*dattr
; /* attributes for custom domains */
602 int ndoms
= 0; /* number of sched domains in result */
603 int nslot
; /* next empty doms[] struct cpumask slot */
609 /* Special case for the 99% of systems with one, full, sched domain */
610 if (is_sched_load_balance(&top_cpuset
)) {
612 doms
= alloc_sched_domains(ndoms
);
616 dattr
= kmalloc(sizeof(struct sched_domain_attr
), GFP_KERNEL
);
618 *dattr
= SD_ATTR_INIT
;
619 update_domain_attr_tree(dattr
, &top_cpuset
);
621 cpumask_copy(doms
[0], top_cpuset
.cpus_allowed
);
626 csa
= kmalloc(number_of_cpusets
* sizeof(cp
), GFP_KERNEL
);
631 list_add(&top_cpuset
.stack_list
, &q
);
632 while (!list_empty(&q
)) {
634 struct cpuset
*child
; /* scans child cpusets of cp */
636 cp
= list_first_entry(&q
, struct cpuset
, stack_list
);
639 if (cpumask_empty(cp
->cpus_allowed
))
643 * All child cpusets contain a subset of the parent's cpus, so
644 * just skip them, and then we call update_domain_attr_tree()
645 * to calc relax_domain_level of the corresponding sched
648 if (is_sched_load_balance(cp
)) {
654 cpuset_for_each_child(child
, cont
, cp
)
655 list_add_tail(&child
->stack_list
, &q
);
659 for (i
= 0; i
< csn
; i
++)
664 /* Find the best partition (set of sched domains) */
665 for (i
= 0; i
< csn
; i
++) {
666 struct cpuset
*a
= csa
[i
];
669 for (j
= 0; j
< csn
; j
++) {
670 struct cpuset
*b
= csa
[j
];
673 if (apn
!= bpn
&& cpusets_overlap(a
, b
)) {
674 for (k
= 0; k
< csn
; k
++) {
675 struct cpuset
*c
= csa
[k
];
680 ndoms
--; /* one less element */
687 * Now we know how many domains to create.
688 * Convert <csn, csa> to <ndoms, doms> and populate cpu masks.
690 doms
= alloc_sched_domains(ndoms
);
695 * The rest of the code, including the scheduler, can deal with
696 * dattr==NULL case. No need to abort if alloc fails.
698 dattr
= kmalloc(ndoms
* sizeof(struct sched_domain_attr
), GFP_KERNEL
);
700 for (nslot
= 0, i
= 0; i
< csn
; i
++) {
701 struct cpuset
*a
= csa
[i
];
706 /* Skip completed partitions */
712 if (nslot
== ndoms
) {
713 static int warnings
= 10;
716 "rebuild_sched_domains confused:"
717 " nslot %d, ndoms %d, csn %d, i %d,"
719 nslot
, ndoms
, csn
, i
, apn
);
727 *(dattr
+ nslot
) = SD_ATTR_INIT
;
728 for (j
= i
; j
< csn
; j
++) {
729 struct cpuset
*b
= csa
[j
];
732 cpumask_or(dp
, dp
, b
->cpus_allowed
);
734 update_domain_attr_tree(dattr
+ nslot
, b
);
736 /* Done with this partition */
742 BUG_ON(nslot
!= ndoms
);
748 * Fallback to the default domain if kmalloc() failed.
749 * See comments in partition_sched_domains().
760 * Rebuild scheduler domains.
762 * If the flag 'sched_load_balance' of any cpuset with non-empty
763 * 'cpus' changes, or if the 'cpus' allowed changes in any cpuset
764 * which has that flag enabled, or if any cpuset with a non-empty
765 * 'cpus' is removed, then call this routine to rebuild the
766 * scheduler's dynamic sched domains.
768 * Call with cgroup_mutex held. Takes get_online_cpus().
770 static void rebuild_sched_domains_locked(void)
772 struct sched_domain_attr
*attr
;
776 WARN_ON_ONCE(!cgroup_lock_is_held());
779 /* Generate domain masks and attrs */
780 ndoms
= generate_sched_domains(&doms
, &attr
);
782 /* Have scheduler rebuild the domains */
783 partition_sched_domains(ndoms
, doms
, attr
);
787 #else /* !CONFIG_SMP */
788 static void rebuild_sched_domains_locked(void)
792 static int generate_sched_domains(cpumask_var_t
**domains
,
793 struct sched_domain_attr
**attributes
)
798 #endif /* CONFIG_SMP */
800 void rebuild_sched_domains(void)
803 rebuild_sched_domains_locked();
808 * cpuset_test_cpumask - test a task's cpus_allowed versus its cpuset's
810 * @scan: struct cgroup_scanner contained in its struct cpuset_hotplug_scanner
812 * Call with cgroup_mutex held. May take callback_mutex during call.
813 * Called for each task in a cgroup by cgroup_scan_tasks().
814 * Return nonzero if this tasks's cpus_allowed mask should be changed (in other
815 * words, if its mask is not equal to its cpuset's mask).
817 static int cpuset_test_cpumask(struct task_struct
*tsk
,
818 struct cgroup_scanner
*scan
)
820 return !cpumask_equal(&tsk
->cpus_allowed
,
821 (cgroup_cs(scan
->cg
))->cpus_allowed
);
825 * cpuset_change_cpumask - make a task's cpus_allowed the same as its cpuset's
827 * @scan: struct cgroup_scanner containing the cgroup of the task
829 * Called by cgroup_scan_tasks() for each task in a cgroup whose
830 * cpus_allowed mask needs to be changed.
832 * We don't need to re-check for the cgroup/cpuset membership, since we're
833 * holding cgroup_lock() at this point.
835 static void cpuset_change_cpumask(struct task_struct
*tsk
,
836 struct cgroup_scanner
*scan
)
838 set_cpus_allowed_ptr(tsk
, ((cgroup_cs(scan
->cg
))->cpus_allowed
));
842 * update_tasks_cpumask - Update the cpumasks of tasks in the cpuset.
843 * @cs: the cpuset in which each task's cpus_allowed mask needs to be changed
844 * @heap: if NULL, defer allocating heap memory to cgroup_scan_tasks()
846 * Called with cgroup_mutex held
848 * The cgroup_scan_tasks() function will scan all the tasks in a cgroup,
849 * calling callback functions for each.
851 * No return value. It's guaranteed that cgroup_scan_tasks() always returns 0
854 static void update_tasks_cpumask(struct cpuset
*cs
, struct ptr_heap
*heap
)
856 struct cgroup_scanner scan
;
858 scan
.cg
= cs
->css
.cgroup
;
859 scan
.test_task
= cpuset_test_cpumask
;
860 scan
.process_task
= cpuset_change_cpumask
;
862 cgroup_scan_tasks(&scan
);
866 * update_cpumask - update the cpus_allowed mask of a cpuset and all tasks in it
867 * @cs: the cpuset to consider
868 * @buf: buffer of cpu numbers written to this cpuset
870 static int update_cpumask(struct cpuset
*cs
, struct cpuset
*trialcs
,
873 struct ptr_heap heap
;
875 int is_load_balanced
;
877 /* top_cpuset.cpus_allowed tracks cpu_online_mask; it's read-only */
878 if (cs
== &top_cpuset
)
882 * An empty cpus_allowed is ok only if the cpuset has no tasks.
883 * Since cpulist_parse() fails on an empty mask, we special case
884 * that parsing. The validate_change() call ensures that cpusets
885 * with tasks have cpus.
888 cpumask_clear(trialcs
->cpus_allowed
);
890 retval
= cpulist_parse(buf
, trialcs
->cpus_allowed
);
894 if (!cpumask_subset(trialcs
->cpus_allowed
, cpu_active_mask
))
897 retval
= validate_change(cs
, trialcs
);
901 /* Nothing to do if the cpus didn't change */
902 if (cpumask_equal(cs
->cpus_allowed
, trialcs
->cpus_allowed
))
905 retval
= heap_init(&heap
, PAGE_SIZE
, GFP_KERNEL
, NULL
);
909 is_load_balanced
= is_sched_load_balance(trialcs
);
911 mutex_lock(&callback_mutex
);
912 cpumask_copy(cs
->cpus_allowed
, trialcs
->cpus_allowed
);
913 mutex_unlock(&callback_mutex
);
916 * Scan tasks in the cpuset, and update the cpumasks of any
917 * that need an update.
919 update_tasks_cpumask(cs
, &heap
);
923 if (is_load_balanced
)
924 rebuild_sched_domains_locked();
931 * Migrate memory region from one set of nodes to another.
933 * Temporarilly set tasks mems_allowed to target nodes of migration,
934 * so that the migration code can allocate pages on these nodes.
936 * Call holding cgroup_mutex, so current's cpuset won't change
937 * during this call, as manage_mutex holds off any cpuset_attach()
938 * calls. Therefore we don't need to take task_lock around the
939 * call to guarantee_online_mems(), as we know no one is changing
942 * While the mm_struct we are migrating is typically from some
943 * other task, the task_struct mems_allowed that we are hacking
944 * is for our current task, which must allocate new pages for that
945 * migrating memory region.
948 static void cpuset_migrate_mm(struct mm_struct
*mm
, const nodemask_t
*from
,
949 const nodemask_t
*to
)
951 struct task_struct
*tsk
= current
;
953 tsk
->mems_allowed
= *to
;
955 do_migrate_pages(mm
, from
, to
, MPOL_MF_MOVE_ALL
);
957 guarantee_online_mems(task_cs(tsk
),&tsk
->mems_allowed
);
961 * cpuset_change_task_nodemask - change task's mems_allowed and mempolicy
962 * @tsk: the task to change
963 * @newmems: new nodes that the task will be set
965 * In order to avoid seeing no nodes if the old and new nodes are disjoint,
966 * we structure updates as setting all new allowed nodes, then clearing newly
969 static void cpuset_change_task_nodemask(struct task_struct
*tsk
,
975 * Allow tasks that have access to memory reserves because they have
976 * been OOM killed to get memory anywhere.
978 if (unlikely(test_thread_flag(TIF_MEMDIE
)))
980 if (current
->flags
& PF_EXITING
) /* Let dying task have memory */
985 * Determine if a loop is necessary if another thread is doing
986 * get_mems_allowed(). If at least one node remains unchanged and
987 * tsk does not have a mempolicy, then an empty nodemask will not be
988 * possible when mems_allowed is larger than a word.
990 need_loop
= task_has_mempolicy(tsk
) ||
991 !nodes_intersects(*newmems
, tsk
->mems_allowed
);
994 write_seqcount_begin(&tsk
->mems_allowed_seq
);
996 nodes_or(tsk
->mems_allowed
, tsk
->mems_allowed
, *newmems
);
997 mpol_rebind_task(tsk
, newmems
, MPOL_REBIND_STEP1
);
999 mpol_rebind_task(tsk
, newmems
, MPOL_REBIND_STEP2
);
1000 tsk
->mems_allowed
= *newmems
;
1003 write_seqcount_end(&tsk
->mems_allowed_seq
);
1009 * Update task's mems_allowed and rebind its mempolicy and vmas' mempolicy
1010 * of it to cpuset's new mems_allowed, and migrate pages to new nodes if
1011 * memory_migrate flag is set. Called with cgroup_mutex held.
1013 static void cpuset_change_nodemask(struct task_struct
*p
,
1014 struct cgroup_scanner
*scan
)
1016 struct mm_struct
*mm
;
1019 const nodemask_t
*oldmem
= scan
->data
;
1020 static nodemask_t newmems
; /* protected by cgroup_mutex */
1022 cs
= cgroup_cs(scan
->cg
);
1023 guarantee_online_mems(cs
, &newmems
);
1025 cpuset_change_task_nodemask(p
, &newmems
);
1027 mm
= get_task_mm(p
);
1031 migrate
= is_memory_migrate(cs
);
1033 mpol_rebind_mm(mm
, &cs
->mems_allowed
);
1035 cpuset_migrate_mm(mm
, oldmem
, &cs
->mems_allowed
);
1039 static void *cpuset_being_rebound
;
1042 * update_tasks_nodemask - Update the nodemasks of tasks in the cpuset.
1043 * @cs: the cpuset in which each task's mems_allowed mask needs to be changed
1044 * @oldmem: old mems_allowed of cpuset cs
1045 * @heap: if NULL, defer allocating heap memory to cgroup_scan_tasks()
1047 * Called with cgroup_mutex held
1048 * No return value. It's guaranteed that cgroup_scan_tasks() always returns 0
1051 static void update_tasks_nodemask(struct cpuset
*cs
, const nodemask_t
*oldmem
,
1052 struct ptr_heap
*heap
)
1054 struct cgroup_scanner scan
;
1056 cpuset_being_rebound
= cs
; /* causes mpol_dup() rebind */
1058 scan
.cg
= cs
->css
.cgroup
;
1059 scan
.test_task
= NULL
;
1060 scan
.process_task
= cpuset_change_nodemask
;
1062 scan
.data
= (nodemask_t
*)oldmem
;
1065 * The mpol_rebind_mm() call takes mmap_sem, which we couldn't
1066 * take while holding tasklist_lock. Forks can happen - the
1067 * mpol_dup() cpuset_being_rebound check will catch such forks,
1068 * and rebind their vma mempolicies too. Because we still hold
1069 * the global cgroup_mutex, we know that no other rebind effort
1070 * will be contending for the global variable cpuset_being_rebound.
1071 * It's ok if we rebind the same mm twice; mpol_rebind_mm()
1072 * is idempotent. Also migrate pages in each mm to new nodes.
1074 cgroup_scan_tasks(&scan
);
1076 /* We're done rebinding vmas to this cpuset's new mems_allowed. */
1077 cpuset_being_rebound
= NULL
;
1081 * Handle user request to change the 'mems' memory placement
1082 * of a cpuset. Needs to validate the request, update the
1083 * cpusets mems_allowed, and for each task in the cpuset,
1084 * update mems_allowed and rebind task's mempolicy and any vma
1085 * mempolicies and if the cpuset is marked 'memory_migrate',
1086 * migrate the tasks pages to the new memory.
1088 * Call with cgroup_mutex held. May take callback_mutex during call.
1089 * Will take tasklist_lock, scan tasklist for tasks in cpuset cs,
1090 * lock each such tasks mm->mmap_sem, scan its vma's and rebind
1091 * their mempolicies to the cpusets new mems_allowed.
1093 static int update_nodemask(struct cpuset
*cs
, struct cpuset
*trialcs
,
1096 NODEMASK_ALLOC(nodemask_t
, oldmem
, GFP_KERNEL
);
1098 struct ptr_heap heap
;
1104 * top_cpuset.mems_allowed tracks node_stats[N_MEMORY];
1107 if (cs
== &top_cpuset
) {
1113 * An empty mems_allowed is ok iff there are no tasks in the cpuset.
1114 * Since nodelist_parse() fails on an empty mask, we special case
1115 * that parsing. The validate_change() call ensures that cpusets
1116 * with tasks have memory.
1119 nodes_clear(trialcs
->mems_allowed
);
1121 retval
= nodelist_parse(buf
, trialcs
->mems_allowed
);
1125 if (!nodes_subset(trialcs
->mems_allowed
,
1126 node_states
[N_MEMORY
])) {
1131 *oldmem
= cs
->mems_allowed
;
1132 if (nodes_equal(*oldmem
, trialcs
->mems_allowed
)) {
1133 retval
= 0; /* Too easy - nothing to do */
1136 retval
= validate_change(cs
, trialcs
);
1140 retval
= heap_init(&heap
, PAGE_SIZE
, GFP_KERNEL
, NULL
);
1144 mutex_lock(&callback_mutex
);
1145 cs
->mems_allowed
= trialcs
->mems_allowed
;
1146 mutex_unlock(&callback_mutex
);
1148 update_tasks_nodemask(cs
, oldmem
, &heap
);
1152 NODEMASK_FREE(oldmem
);
1156 int current_cpuset_is_being_rebound(void)
1158 return task_cs(current
) == cpuset_being_rebound
;
1161 static int update_relax_domain_level(struct cpuset
*cs
, s64 val
)
1164 if (val
< -1 || val
>= sched_domain_level_max
)
1168 if (val
!= cs
->relax_domain_level
) {
1169 cs
->relax_domain_level
= val
;
1170 if (!cpumask_empty(cs
->cpus_allowed
) &&
1171 is_sched_load_balance(cs
))
1172 rebuild_sched_domains_locked();
1179 * cpuset_change_flag - make a task's spread flags the same as its cpuset's
1180 * @tsk: task to be updated
1181 * @scan: struct cgroup_scanner containing the cgroup of the task
1183 * Called by cgroup_scan_tasks() for each task in a cgroup.
1185 * We don't need to re-check for the cgroup/cpuset membership, since we're
1186 * holding cgroup_lock() at this point.
1188 static void cpuset_change_flag(struct task_struct
*tsk
,
1189 struct cgroup_scanner
*scan
)
1191 cpuset_update_task_spread_flag(cgroup_cs(scan
->cg
), tsk
);
1195 * update_tasks_flags - update the spread flags of tasks in the cpuset.
1196 * @cs: the cpuset in which each task's spread flags needs to be changed
1197 * @heap: if NULL, defer allocating heap memory to cgroup_scan_tasks()
1199 * Called with cgroup_mutex held
1201 * The cgroup_scan_tasks() function will scan all the tasks in a cgroup,
1202 * calling callback functions for each.
1204 * No return value. It's guaranteed that cgroup_scan_tasks() always returns 0
1207 static void update_tasks_flags(struct cpuset
*cs
, struct ptr_heap
*heap
)
1209 struct cgroup_scanner scan
;
1211 scan
.cg
= cs
->css
.cgroup
;
1212 scan
.test_task
= NULL
;
1213 scan
.process_task
= cpuset_change_flag
;
1215 cgroup_scan_tasks(&scan
);
1219 * update_flag - read a 0 or a 1 in a file and update associated flag
1220 * bit: the bit to update (see cpuset_flagbits_t)
1221 * cs: the cpuset to update
1222 * turning_on: whether the flag is being set or cleared
1224 * Call with cgroup_mutex held.
1227 static int update_flag(cpuset_flagbits_t bit
, struct cpuset
*cs
,
1230 struct cpuset
*trialcs
;
1231 int balance_flag_changed
;
1232 int spread_flag_changed
;
1233 struct ptr_heap heap
;
1236 trialcs
= alloc_trial_cpuset(cs
);
1241 set_bit(bit
, &trialcs
->flags
);
1243 clear_bit(bit
, &trialcs
->flags
);
1245 err
= validate_change(cs
, trialcs
);
1249 err
= heap_init(&heap
, PAGE_SIZE
, GFP_KERNEL
, NULL
);
1253 balance_flag_changed
= (is_sched_load_balance(cs
) !=
1254 is_sched_load_balance(trialcs
));
1256 spread_flag_changed
= ((is_spread_slab(cs
) != is_spread_slab(trialcs
))
1257 || (is_spread_page(cs
) != is_spread_page(trialcs
)));
1259 mutex_lock(&callback_mutex
);
1260 cs
->flags
= trialcs
->flags
;
1261 mutex_unlock(&callback_mutex
);
1263 if (!cpumask_empty(trialcs
->cpus_allowed
) && balance_flag_changed
)
1264 rebuild_sched_domains_locked();
1266 if (spread_flag_changed
)
1267 update_tasks_flags(cs
, &heap
);
1270 free_trial_cpuset(trialcs
);
1275 * Frequency meter - How fast is some event occurring?
1277 * These routines manage a digitally filtered, constant time based,
1278 * event frequency meter. There are four routines:
1279 * fmeter_init() - initialize a frequency meter.
1280 * fmeter_markevent() - called each time the event happens.
1281 * fmeter_getrate() - returns the recent rate of such events.
1282 * fmeter_update() - internal routine used to update fmeter.
1284 * A common data structure is passed to each of these routines,
1285 * which is used to keep track of the state required to manage the
1286 * frequency meter and its digital filter.
1288 * The filter works on the number of events marked per unit time.
1289 * The filter is single-pole low-pass recursive (IIR). The time unit
1290 * is 1 second. Arithmetic is done using 32-bit integers scaled to
1291 * simulate 3 decimal digits of precision (multiplied by 1000).
1293 * With an FM_COEF of 933, and a time base of 1 second, the filter
1294 * has a half-life of 10 seconds, meaning that if the events quit
1295 * happening, then the rate returned from the fmeter_getrate()
1296 * will be cut in half each 10 seconds, until it converges to zero.
1298 * It is not worth doing a real infinitely recursive filter. If more
1299 * than FM_MAXTICKS ticks have elapsed since the last filter event,
1300 * just compute FM_MAXTICKS ticks worth, by which point the level
1303 * Limit the count of unprocessed events to FM_MAXCNT, so as to avoid
1304 * arithmetic overflow in the fmeter_update() routine.
1306 * Given the simple 32 bit integer arithmetic used, this meter works
1307 * best for reporting rates between one per millisecond (msec) and
1308 * one per 32 (approx) seconds. At constant rates faster than one
1309 * per msec it maxes out at values just under 1,000,000. At constant
1310 * rates between one per msec, and one per second it will stabilize
1311 * to a value N*1000, where N is the rate of events per second.
1312 * At constant rates between one per second and one per 32 seconds,
1313 * it will be choppy, moving up on the seconds that have an event,
1314 * and then decaying until the next event. At rates slower than
1315 * about one in 32 seconds, it decays all the way back to zero between
1319 #define FM_COEF 933 /* coefficient for half-life of 10 secs */
1320 #define FM_MAXTICKS ((time_t)99) /* useless computing more ticks than this */
1321 #define FM_MAXCNT 1000000 /* limit cnt to avoid overflow */
1322 #define FM_SCALE 1000 /* faux fixed point scale */
1324 /* Initialize a frequency meter */
1325 static void fmeter_init(struct fmeter
*fmp
)
1330 spin_lock_init(&fmp
->lock
);
1333 /* Internal meter update - process cnt events and update value */
1334 static void fmeter_update(struct fmeter
*fmp
)
1336 time_t now
= get_seconds();
1337 time_t ticks
= now
- fmp
->time
;
1342 ticks
= min(FM_MAXTICKS
, ticks
);
1344 fmp
->val
= (FM_COEF
* fmp
->val
) / FM_SCALE
;
1347 fmp
->val
+= ((FM_SCALE
- FM_COEF
) * fmp
->cnt
) / FM_SCALE
;
1351 /* Process any previous ticks, then bump cnt by one (times scale). */
1352 static void fmeter_markevent(struct fmeter
*fmp
)
1354 spin_lock(&fmp
->lock
);
1356 fmp
->cnt
= min(FM_MAXCNT
, fmp
->cnt
+ FM_SCALE
);
1357 spin_unlock(&fmp
->lock
);
1360 /* Process any previous ticks, then return current value. */
1361 static int fmeter_getrate(struct fmeter
*fmp
)
1365 spin_lock(&fmp
->lock
);
1368 spin_unlock(&fmp
->lock
);
1372 /* Called by cgroups to determine if a cpuset is usable; cgroup_mutex held */
1373 static int cpuset_can_attach(struct cgroup
*cgrp
, struct cgroup_taskset
*tset
)
1375 struct cpuset
*cs
= cgroup_cs(cgrp
);
1376 struct task_struct
*task
;
1379 if (cpumask_empty(cs
->cpus_allowed
) || nodes_empty(cs
->mems_allowed
))
1382 cgroup_taskset_for_each(task
, cgrp
, tset
) {
1384 * Kthreads bound to specific cpus cannot be moved to a new
1385 * cpuset; we cannot change their cpu affinity and
1386 * isolating such threads by their set of allowed nodes is
1387 * unnecessary. Thus, cpusets are not applicable for such
1388 * threads. This prevents checking for success of
1389 * set_cpus_allowed_ptr() on all attached tasks before
1390 * cpus_allowed may be changed.
1392 if (task
->flags
& PF_THREAD_BOUND
)
1394 if ((ret
= security_task_setscheduler(task
)))
1399 * Mark attach is in progress. This makes validate_change() fail
1400 * changes which zero cpus/mems_allowed.
1402 cs
->attach_in_progress
++;
1407 static void cpuset_cancel_attach(struct cgroup
*cgrp
,
1408 struct cgroup_taskset
*tset
)
1410 cgroup_cs(cgrp
)->attach_in_progress
--;
1414 * Protected by cgroup_mutex. cpus_attach is used only by cpuset_attach()
1415 * but we can't allocate it dynamically there. Define it global and
1416 * allocate from cpuset_init().
1418 static cpumask_var_t cpus_attach
;
1420 static void cpuset_attach(struct cgroup
*cgrp
, struct cgroup_taskset
*tset
)
1422 /* static bufs protected by cgroup_mutex */
1423 static nodemask_t cpuset_attach_nodemask_from
;
1424 static nodemask_t cpuset_attach_nodemask_to
;
1425 struct mm_struct
*mm
;
1426 struct task_struct
*task
;
1427 struct task_struct
*leader
= cgroup_taskset_first(tset
);
1428 struct cgroup
*oldcgrp
= cgroup_taskset_cur_cgroup(tset
);
1429 struct cpuset
*cs
= cgroup_cs(cgrp
);
1430 struct cpuset
*oldcs
= cgroup_cs(oldcgrp
);
1432 /* prepare for attach */
1433 if (cs
== &top_cpuset
)
1434 cpumask_copy(cpus_attach
, cpu_possible_mask
);
1436 guarantee_online_cpus(cs
, cpus_attach
);
1438 guarantee_online_mems(cs
, &cpuset_attach_nodemask_to
);
1440 cgroup_taskset_for_each(task
, cgrp
, tset
) {
1442 * can_attach beforehand should guarantee that this doesn't
1443 * fail. TODO: have a better way to handle failure here
1445 WARN_ON_ONCE(set_cpus_allowed_ptr(task
, cpus_attach
));
1447 cpuset_change_task_nodemask(task
, &cpuset_attach_nodemask_to
);
1448 cpuset_update_task_spread_flag(cs
, task
);
1452 * Change mm, possibly for multiple threads in a threadgroup. This is
1453 * expensive and may sleep.
1455 cpuset_attach_nodemask_from
= oldcs
->mems_allowed
;
1456 cpuset_attach_nodemask_to
= cs
->mems_allowed
;
1457 mm
= get_task_mm(leader
);
1459 mpol_rebind_mm(mm
, &cpuset_attach_nodemask_to
);
1460 if (is_memory_migrate(cs
))
1461 cpuset_migrate_mm(mm
, &cpuset_attach_nodemask_from
,
1462 &cpuset_attach_nodemask_to
);
1466 cs
->attach_in_progress
--;
1469 /* The various types of files and directories in a cpuset file system */
1472 FILE_MEMORY_MIGRATE
,
1478 FILE_SCHED_LOAD_BALANCE
,
1479 FILE_SCHED_RELAX_DOMAIN_LEVEL
,
1480 FILE_MEMORY_PRESSURE_ENABLED
,
1481 FILE_MEMORY_PRESSURE
,
1484 } cpuset_filetype_t
;
1486 static int cpuset_write_u64(struct cgroup
*cgrp
, struct cftype
*cft
, u64 val
)
1489 struct cpuset
*cs
= cgroup_cs(cgrp
);
1490 cpuset_filetype_t type
= cft
->private;
1492 if (!cgroup_lock_live_group(cgrp
))
1496 case FILE_CPU_EXCLUSIVE
:
1497 retval
= update_flag(CS_CPU_EXCLUSIVE
, cs
, val
);
1499 case FILE_MEM_EXCLUSIVE
:
1500 retval
= update_flag(CS_MEM_EXCLUSIVE
, cs
, val
);
1502 case FILE_MEM_HARDWALL
:
1503 retval
= update_flag(CS_MEM_HARDWALL
, cs
, val
);
1505 case FILE_SCHED_LOAD_BALANCE
:
1506 retval
= update_flag(CS_SCHED_LOAD_BALANCE
, cs
, val
);
1508 case FILE_MEMORY_MIGRATE
:
1509 retval
= update_flag(CS_MEMORY_MIGRATE
, cs
, val
);
1511 case FILE_MEMORY_PRESSURE_ENABLED
:
1512 cpuset_memory_pressure_enabled
= !!val
;
1514 case FILE_MEMORY_PRESSURE
:
1517 case FILE_SPREAD_PAGE
:
1518 retval
= update_flag(CS_SPREAD_PAGE
, cs
, val
);
1520 case FILE_SPREAD_SLAB
:
1521 retval
= update_flag(CS_SPREAD_SLAB
, cs
, val
);
1531 static int cpuset_write_s64(struct cgroup
*cgrp
, struct cftype
*cft
, s64 val
)
1534 struct cpuset
*cs
= cgroup_cs(cgrp
);
1535 cpuset_filetype_t type
= cft
->private;
1537 if (!cgroup_lock_live_group(cgrp
))
1541 case FILE_SCHED_RELAX_DOMAIN_LEVEL
:
1542 retval
= update_relax_domain_level(cs
, val
);
1553 * Common handling for a write to a "cpus" or "mems" file.
1555 static int cpuset_write_resmask(struct cgroup
*cgrp
, struct cftype
*cft
,
1559 struct cpuset
*cs
= cgroup_cs(cgrp
);
1560 struct cpuset
*trialcs
;
1563 * CPU or memory hotunplug may leave @cs w/o any execution
1564 * resources, in which case the hotplug code asynchronously updates
1565 * configuration and transfers all tasks to the nearest ancestor
1566 * which can execute.
1568 * As writes to "cpus" or "mems" may restore @cs's execution
1569 * resources, wait for the previously scheduled operations before
1570 * proceeding, so that we don't end up keep removing tasks added
1571 * after execution capability is restored.
1573 flush_work(&cpuset_hotplug_work
);
1575 if (!cgroup_lock_live_group(cgrp
))
1578 trialcs
= alloc_trial_cpuset(cs
);
1584 switch (cft
->private) {
1586 retval
= update_cpumask(cs
, trialcs
, buf
);
1589 retval
= update_nodemask(cs
, trialcs
, buf
);
1596 free_trial_cpuset(trialcs
);
1603 * These ascii lists should be read in a single call, by using a user
1604 * buffer large enough to hold the entire map. If read in smaller
1605 * chunks, there is no guarantee of atomicity. Since the display format
1606 * used, list of ranges of sequential numbers, is variable length,
1607 * and since these maps can change value dynamically, one could read
1608 * gibberish by doing partial reads while a list was changing.
1609 * A single large read to a buffer that crosses a page boundary is
1610 * ok, because the result being copied to user land is not recomputed
1611 * across a page fault.
1614 static size_t cpuset_sprintf_cpulist(char *page
, struct cpuset
*cs
)
1618 mutex_lock(&callback_mutex
);
1619 count
= cpulist_scnprintf(page
, PAGE_SIZE
, cs
->cpus_allowed
);
1620 mutex_unlock(&callback_mutex
);
1625 static size_t cpuset_sprintf_memlist(char *page
, struct cpuset
*cs
)
1629 mutex_lock(&callback_mutex
);
1630 count
= nodelist_scnprintf(page
, PAGE_SIZE
, cs
->mems_allowed
);
1631 mutex_unlock(&callback_mutex
);
1636 static ssize_t
cpuset_common_file_read(struct cgroup
*cont
,
1640 size_t nbytes
, loff_t
*ppos
)
1642 struct cpuset
*cs
= cgroup_cs(cont
);
1643 cpuset_filetype_t type
= cft
->private;
1648 if (!(page
= (char *)__get_free_page(GFP_TEMPORARY
)))
1655 s
+= cpuset_sprintf_cpulist(s
, cs
);
1658 s
+= cpuset_sprintf_memlist(s
, cs
);
1666 retval
= simple_read_from_buffer(buf
, nbytes
, ppos
, page
, s
- page
);
1668 free_page((unsigned long)page
);
1672 static u64
cpuset_read_u64(struct cgroup
*cont
, struct cftype
*cft
)
1674 struct cpuset
*cs
= cgroup_cs(cont
);
1675 cpuset_filetype_t type
= cft
->private;
1677 case FILE_CPU_EXCLUSIVE
:
1678 return is_cpu_exclusive(cs
);
1679 case FILE_MEM_EXCLUSIVE
:
1680 return is_mem_exclusive(cs
);
1681 case FILE_MEM_HARDWALL
:
1682 return is_mem_hardwall(cs
);
1683 case FILE_SCHED_LOAD_BALANCE
:
1684 return is_sched_load_balance(cs
);
1685 case FILE_MEMORY_MIGRATE
:
1686 return is_memory_migrate(cs
);
1687 case FILE_MEMORY_PRESSURE_ENABLED
:
1688 return cpuset_memory_pressure_enabled
;
1689 case FILE_MEMORY_PRESSURE
:
1690 return fmeter_getrate(&cs
->fmeter
);
1691 case FILE_SPREAD_PAGE
:
1692 return is_spread_page(cs
);
1693 case FILE_SPREAD_SLAB
:
1694 return is_spread_slab(cs
);
1699 /* Unreachable but makes gcc happy */
1703 static s64
cpuset_read_s64(struct cgroup
*cont
, struct cftype
*cft
)
1705 struct cpuset
*cs
= cgroup_cs(cont
);
1706 cpuset_filetype_t type
= cft
->private;
1708 case FILE_SCHED_RELAX_DOMAIN_LEVEL
:
1709 return cs
->relax_domain_level
;
1714 /* Unrechable but makes gcc happy */
1720 * for the common functions, 'private' gives the type of file
1723 static struct cftype files
[] = {
1726 .read
= cpuset_common_file_read
,
1727 .write_string
= cpuset_write_resmask
,
1728 .max_write_len
= (100U + 6 * NR_CPUS
),
1729 .private = FILE_CPULIST
,
1734 .read
= cpuset_common_file_read
,
1735 .write_string
= cpuset_write_resmask
,
1736 .max_write_len
= (100U + 6 * MAX_NUMNODES
),
1737 .private = FILE_MEMLIST
,
1741 .name
= "cpu_exclusive",
1742 .read_u64
= cpuset_read_u64
,
1743 .write_u64
= cpuset_write_u64
,
1744 .private = FILE_CPU_EXCLUSIVE
,
1748 .name
= "mem_exclusive",
1749 .read_u64
= cpuset_read_u64
,
1750 .write_u64
= cpuset_write_u64
,
1751 .private = FILE_MEM_EXCLUSIVE
,
1755 .name
= "mem_hardwall",
1756 .read_u64
= cpuset_read_u64
,
1757 .write_u64
= cpuset_write_u64
,
1758 .private = FILE_MEM_HARDWALL
,
1762 .name
= "sched_load_balance",
1763 .read_u64
= cpuset_read_u64
,
1764 .write_u64
= cpuset_write_u64
,
1765 .private = FILE_SCHED_LOAD_BALANCE
,
1769 .name
= "sched_relax_domain_level",
1770 .read_s64
= cpuset_read_s64
,
1771 .write_s64
= cpuset_write_s64
,
1772 .private = FILE_SCHED_RELAX_DOMAIN_LEVEL
,
1776 .name
= "memory_migrate",
1777 .read_u64
= cpuset_read_u64
,
1778 .write_u64
= cpuset_write_u64
,
1779 .private = FILE_MEMORY_MIGRATE
,
1783 .name
= "memory_pressure",
1784 .read_u64
= cpuset_read_u64
,
1785 .write_u64
= cpuset_write_u64
,
1786 .private = FILE_MEMORY_PRESSURE
,
1791 .name
= "memory_spread_page",
1792 .read_u64
= cpuset_read_u64
,
1793 .write_u64
= cpuset_write_u64
,
1794 .private = FILE_SPREAD_PAGE
,
1798 .name
= "memory_spread_slab",
1799 .read_u64
= cpuset_read_u64
,
1800 .write_u64
= cpuset_write_u64
,
1801 .private = FILE_SPREAD_SLAB
,
1805 .name
= "memory_pressure_enabled",
1806 .flags
= CFTYPE_ONLY_ON_ROOT
,
1807 .read_u64
= cpuset_read_u64
,
1808 .write_u64
= cpuset_write_u64
,
1809 .private = FILE_MEMORY_PRESSURE_ENABLED
,
1816 * cpuset_css_alloc - allocate a cpuset css
1817 * cont: control group that the new cpuset will be part of
1820 static struct cgroup_subsys_state
*cpuset_css_alloc(struct cgroup
*cont
)
1825 return &top_cpuset
.css
;
1827 cs
= kzalloc(sizeof(*cs
), GFP_KERNEL
);
1829 return ERR_PTR(-ENOMEM
);
1830 if (!alloc_cpumask_var(&cs
->cpus_allowed
, GFP_KERNEL
)) {
1832 return ERR_PTR(-ENOMEM
);
1835 set_bit(CS_SCHED_LOAD_BALANCE
, &cs
->flags
);
1836 cpumask_clear(cs
->cpus_allowed
);
1837 nodes_clear(cs
->mems_allowed
);
1838 fmeter_init(&cs
->fmeter
);
1839 INIT_WORK(&cs
->hotplug_work
, cpuset_propagate_hotplug_workfn
);
1840 cs
->relax_domain_level
= -1;
1841 cs
->parent
= cgroup_cs(cont
->parent
);
1846 static int cpuset_css_online(struct cgroup
*cgrp
)
1848 struct cpuset
*cs
= cgroup_cs(cgrp
);
1849 struct cpuset
*parent
= cs
->parent
;
1850 struct cpuset
*tmp_cs
;
1851 struct cgroup
*pos_cg
;
1856 set_bit(CS_ONLINE
, &cs
->flags
);
1857 if (is_spread_page(parent
))
1858 set_bit(CS_SPREAD_PAGE
, &cs
->flags
);
1859 if (is_spread_slab(parent
))
1860 set_bit(CS_SPREAD_SLAB
, &cs
->flags
);
1862 number_of_cpusets
++;
1864 if (!test_bit(CGRP_CPUSET_CLONE_CHILDREN
, &cgrp
->flags
))
1868 * Clone @parent's configuration if CGRP_CPUSET_CLONE_CHILDREN is
1869 * set. This flag handling is implemented in cgroup core for
1870 * histrical reasons - the flag may be specified during mount.
1872 * Currently, if any sibling cpusets have exclusive cpus or mem, we
1873 * refuse to clone the configuration - thereby refusing the task to
1874 * be entered, and as a result refusing the sys_unshare() or
1875 * clone() which initiated it. If this becomes a problem for some
1876 * users who wish to allow that scenario, then this could be
1877 * changed to grant parent->cpus_allowed-sibling_cpus_exclusive
1878 * (and likewise for mems) to the new cgroup.
1881 cpuset_for_each_child(tmp_cs
, pos_cg
, parent
) {
1882 if (is_mem_exclusive(tmp_cs
) || is_cpu_exclusive(tmp_cs
)) {
1889 mutex_lock(&callback_mutex
);
1890 cs
->mems_allowed
= parent
->mems_allowed
;
1891 cpumask_copy(cs
->cpus_allowed
, parent
->cpus_allowed
);
1892 mutex_unlock(&callback_mutex
);
1897 static void cpuset_css_offline(struct cgroup
*cgrp
)
1899 struct cpuset
*cs
= cgroup_cs(cgrp
);
1901 /* css_offline is called w/o cgroup_mutex, grab it */
1904 if (is_sched_load_balance(cs
))
1905 update_flag(CS_SCHED_LOAD_BALANCE
, cs
, 0);
1907 number_of_cpusets
--;
1908 clear_bit(CS_ONLINE
, &cs
->flags
);
1914 * If the cpuset being removed has its flag 'sched_load_balance'
1915 * enabled, then simulate turning sched_load_balance off, which
1916 * will call rebuild_sched_domains_locked().
1919 static void cpuset_css_free(struct cgroup
*cont
)
1921 struct cpuset
*cs
= cgroup_cs(cont
);
1923 free_cpumask_var(cs
->cpus_allowed
);
1927 struct cgroup_subsys cpuset_subsys
= {
1929 .css_alloc
= cpuset_css_alloc
,
1930 .css_online
= cpuset_css_online
,
1931 .css_offline
= cpuset_css_offline
,
1932 .css_free
= cpuset_css_free
,
1933 .can_attach
= cpuset_can_attach
,
1934 .cancel_attach
= cpuset_cancel_attach
,
1935 .attach
= cpuset_attach
,
1936 .subsys_id
= cpuset_subsys_id
,
1937 .base_cftypes
= files
,
1942 * cpuset_init - initialize cpusets at system boot
1944 * Description: Initialize top_cpuset and the cpuset internal file system,
1947 int __init
cpuset_init(void)
1951 if (!alloc_cpumask_var(&top_cpuset
.cpus_allowed
, GFP_KERNEL
))
1954 cpumask_setall(top_cpuset
.cpus_allowed
);
1955 nodes_setall(top_cpuset
.mems_allowed
);
1957 fmeter_init(&top_cpuset
.fmeter
);
1958 set_bit(CS_SCHED_LOAD_BALANCE
, &top_cpuset
.flags
);
1959 top_cpuset
.relax_domain_level
= -1;
1961 err
= register_filesystem(&cpuset_fs_type
);
1965 if (!alloc_cpumask_var(&cpus_attach
, GFP_KERNEL
))
1968 number_of_cpusets
= 1;
1973 * cpuset_do_move_task - move a given task to another cpuset
1974 * @tsk: pointer to task_struct the task to move
1975 * @scan: struct cgroup_scanner contained in its struct cpuset_hotplug_scanner
1977 * Called by cgroup_scan_tasks() for each task in a cgroup.
1978 * Return nonzero to stop the walk through the tasks.
1980 static void cpuset_do_move_task(struct task_struct
*tsk
,
1981 struct cgroup_scanner
*scan
)
1983 struct cgroup
*new_cgroup
= scan
->data
;
1985 cgroup_attach_task(new_cgroup
, tsk
);
1989 * move_member_tasks_to_cpuset - move tasks from one cpuset to another
1990 * @from: cpuset in which the tasks currently reside
1991 * @to: cpuset to which the tasks will be moved
1993 * Called with cgroup_mutex held
1994 * callback_mutex must not be held, as cpuset_attach() will take it.
1996 * The cgroup_scan_tasks() function will scan all the tasks in a cgroup,
1997 * calling callback functions for each.
1999 static void move_member_tasks_to_cpuset(struct cpuset
*from
, struct cpuset
*to
)
2001 struct cgroup_scanner scan
;
2003 scan
.cg
= from
->css
.cgroup
;
2004 scan
.test_task
= NULL
; /* select all tasks in cgroup */
2005 scan
.process_task
= cpuset_do_move_task
;
2007 scan
.data
= to
->css
.cgroup
;
2009 if (cgroup_scan_tasks(&scan
))
2010 printk(KERN_ERR
"move_member_tasks_to_cpuset: "
2011 "cgroup_scan_tasks failed\n");
2015 * If CPU and/or memory hotplug handlers, below, unplug any CPUs
2016 * or memory nodes, we need to walk over the cpuset hierarchy,
2017 * removing that CPU or node from all cpusets. If this removes the
2018 * last CPU or node from a cpuset, then move the tasks in the empty
2019 * cpuset to its next-highest non-empty parent.
2021 * Called with cgroup_mutex held
2022 * callback_mutex must not be held, as cpuset_attach() will take it.
2024 static void remove_tasks_in_empty_cpuset(struct cpuset
*cs
)
2026 struct cpuset
*parent
;
2029 * Find its next-highest non-empty parent, (top cpuset
2030 * has online cpus, so can't be empty).
2032 parent
= cs
->parent
;
2033 while (cpumask_empty(parent
->cpus_allowed
) ||
2034 nodes_empty(parent
->mems_allowed
))
2035 parent
= parent
->parent
;
2037 move_member_tasks_to_cpuset(cs
, parent
);
2041 * Helper function to traverse cpusets.
2042 * It can be used to walk the cpuset tree from top to bottom, completing
2043 * one layer before dropping down to the next (thus always processing a
2044 * node before any of its children).
2046 static struct cpuset
*cpuset_next(struct list_head
*queue
)
2049 struct cpuset
*child
; /* scans child cpusets of cp */
2050 struct cgroup
*cont
;
2052 if (list_empty(queue
))
2055 cp
= list_first_entry(queue
, struct cpuset
, stack_list
);
2056 list_del(queue
->next
);
2058 cpuset_for_each_child(child
, cont
, cp
)
2059 list_add_tail(&child
->stack_list
, queue
);
2066 * cpuset_propagate_hotplug_workfn - propagate CPU/memory hotplug to a cpuset
2067 * @cs: cpuset in interest
2069 * Compare @cs's cpu and mem masks against top_cpuset and if some have gone
2070 * offline, update @cs accordingly. If @cs ends up with no CPU or memory,
2071 * all its tasks are moved to the nearest ancestor with both resources.
2073 static void cpuset_propagate_hotplug_workfn(struct work_struct
*work
)
2075 static cpumask_t off_cpus
;
2076 static nodemask_t off_mems
, tmp_mems
;
2077 struct cpuset
*cs
= container_of(work
, struct cpuset
, hotplug_work
);
2081 cpumask_andnot(&off_cpus
, cs
->cpus_allowed
, top_cpuset
.cpus_allowed
);
2082 nodes_andnot(off_mems
, cs
->mems_allowed
, top_cpuset
.mems_allowed
);
2084 /* remove offline cpus from @cs */
2085 if (!cpumask_empty(&off_cpus
)) {
2086 mutex_lock(&callback_mutex
);
2087 cpumask_andnot(cs
->cpus_allowed
, cs
->cpus_allowed
, &off_cpus
);
2088 mutex_unlock(&callback_mutex
);
2089 update_tasks_cpumask(cs
, NULL
);
2092 /* remove offline mems from @cs */
2093 if (!nodes_empty(off_mems
)) {
2094 tmp_mems
= cs
->mems_allowed
;
2095 mutex_lock(&callback_mutex
);
2096 nodes_andnot(cs
->mems_allowed
, cs
->mems_allowed
, off_mems
);
2097 mutex_unlock(&callback_mutex
);
2098 update_tasks_nodemask(cs
, &tmp_mems
, NULL
);
2101 if (cpumask_empty(cs
->cpus_allowed
) || nodes_empty(cs
->mems_allowed
))
2102 remove_tasks_in_empty_cpuset(cs
);
2106 /* the following may free @cs, should be the last operation */
2111 * schedule_cpuset_propagate_hotplug - schedule hotplug propagation to a cpuset
2112 * @cs: cpuset of interest
2114 * Schedule cpuset_propagate_hotplug_workfn() which will update CPU and
2115 * memory masks according to top_cpuset.
2117 static void schedule_cpuset_propagate_hotplug(struct cpuset
*cs
)
2120 * Pin @cs. The refcnt will be released when the work item
2121 * finishes executing.
2123 if (!css_tryget(&cs
->css
))
2127 * Queue @cs->hotplug_work. If already pending, lose the css ref.
2128 * cpuset_propagate_hotplug_wq is ordered and propagation will
2129 * happen in the order this function is called.
2131 if (!queue_work(cpuset_propagate_hotplug_wq
, &cs
->hotplug_work
))
2136 * cpuset_hotplug_workfn - handle CPU/memory hotunplug for a cpuset
2138 * This function is called after either CPU or memory configuration has
2139 * changed and updates cpuset accordingly. The top_cpuset is always
2140 * synchronized to cpu_active_mask and N_MEMORY, which is necessary in
2141 * order to make cpusets transparent (of no affect) on systems that are
2142 * actively using CPU hotplug but making no active use of cpusets.
2144 * Non-root cpusets are only affected by offlining. If any CPUs or memory
2145 * nodes have been taken down, cpuset_propagate_hotplug() is invoked on all
2148 * Note that CPU offlining during suspend is ignored. We don't modify
2149 * cpusets across suspend/resume cycles at all.
2151 static void cpuset_hotplug_workfn(struct work_struct
*work
)
2153 static cpumask_t new_cpus
, tmp_cpus
;
2154 static nodemask_t new_mems
, tmp_mems
;
2155 bool cpus_updated
, mems_updated
;
2156 bool cpus_offlined
, mems_offlined
;
2160 /* fetch the available cpus/mems and find out which changed how */
2161 cpumask_copy(&new_cpus
, cpu_active_mask
);
2162 new_mems
= node_states
[N_MEMORY
];
2164 cpus_updated
= !cpumask_equal(top_cpuset
.cpus_allowed
, &new_cpus
);
2165 cpus_offlined
= cpumask_andnot(&tmp_cpus
, top_cpuset
.cpus_allowed
,
2168 mems_updated
= !nodes_equal(top_cpuset
.mems_allowed
, new_mems
);
2169 nodes_andnot(tmp_mems
, top_cpuset
.mems_allowed
, new_mems
);
2170 mems_offlined
= !nodes_empty(tmp_mems
);
2172 /* synchronize cpus_allowed to cpu_active_mask */
2174 mutex_lock(&callback_mutex
);
2175 cpumask_copy(top_cpuset
.cpus_allowed
, &new_cpus
);
2176 mutex_unlock(&callback_mutex
);
2177 /* we don't mess with cpumasks of tasks in top_cpuset */
2180 /* synchronize mems_allowed to N_MEMORY */
2182 tmp_mems
= top_cpuset
.mems_allowed
;
2183 mutex_lock(&callback_mutex
);
2184 top_cpuset
.mems_allowed
= new_mems
;
2185 mutex_unlock(&callback_mutex
);
2186 update_tasks_nodemask(&top_cpuset
, &tmp_mems
, NULL
);
2189 /* if cpus or mems went down, we need to propagate to descendants */
2190 if (cpus_offlined
|| mems_offlined
) {
2194 list_add_tail(&top_cpuset
.stack_list
, &queue
);
2195 while ((cs
= cpuset_next(&queue
)))
2196 if (cs
!= &top_cpuset
)
2197 schedule_cpuset_propagate_hotplug(cs
);
2202 /* wait for propagations to finish */
2203 flush_workqueue(cpuset_propagate_hotplug_wq
);
2205 /* rebuild sched domains if cpus_allowed has changed */
2207 struct sched_domain_attr
*attr
;
2208 cpumask_var_t
*doms
;
2212 ndoms
= generate_sched_domains(&doms
, &attr
);
2215 partition_sched_domains(ndoms
, doms
, attr
);
2219 void cpuset_update_active_cpus(bool cpu_online
)
2222 * We're inside cpu hotplug critical region which usually nests
2223 * inside cgroup synchronization. Bounce actual hotplug processing
2224 * to a work item to avoid reverse locking order.
2226 * We still need to do partition_sched_domains() synchronously;
2227 * otherwise, the scheduler will get confused and put tasks to the
2228 * dead CPU. Fall back to the default single domain.
2229 * cpuset_hotplug_workfn() will rebuild it as necessary.
2231 partition_sched_domains(1, NULL
, NULL
);
2232 schedule_work(&cpuset_hotplug_work
);
2235 #ifdef CONFIG_MEMORY_HOTPLUG
2237 * Keep top_cpuset.mems_allowed tracking node_states[N_MEMORY].
2238 * Call this routine anytime after node_states[N_MEMORY] changes.
2239 * See cpuset_update_active_cpus() for CPU hotplug handling.
2241 static int cpuset_track_online_nodes(struct notifier_block
*self
,
2242 unsigned long action
, void *arg
)
2244 schedule_work(&cpuset_hotplug_work
);
2250 * cpuset_init_smp - initialize cpus_allowed
2252 * Description: Finish top cpuset after cpu, node maps are initialized
2255 void __init
cpuset_init_smp(void)
2257 cpumask_copy(top_cpuset
.cpus_allowed
, cpu_active_mask
);
2258 top_cpuset
.mems_allowed
= node_states
[N_MEMORY
];
2260 hotplug_memory_notifier(cpuset_track_online_nodes
, 10);
2262 cpuset_propagate_hotplug_wq
=
2263 alloc_ordered_workqueue("cpuset_hotplug", 0);
2264 BUG_ON(!cpuset_propagate_hotplug_wq
);
2268 * cpuset_cpus_allowed - return cpus_allowed mask from a tasks cpuset.
2269 * @tsk: pointer to task_struct from which to obtain cpuset->cpus_allowed.
2270 * @pmask: pointer to struct cpumask variable to receive cpus_allowed set.
2272 * Description: Returns the cpumask_var_t cpus_allowed of the cpuset
2273 * attached to the specified @tsk. Guaranteed to return some non-empty
2274 * subset of cpu_online_mask, even if this means going outside the
2278 void cpuset_cpus_allowed(struct task_struct
*tsk
, struct cpumask
*pmask
)
2280 mutex_lock(&callback_mutex
);
2282 guarantee_online_cpus(task_cs(tsk
), pmask
);
2284 mutex_unlock(&callback_mutex
);
2287 void cpuset_cpus_allowed_fallback(struct task_struct
*tsk
)
2289 const struct cpuset
*cs
;
2294 do_set_cpus_allowed(tsk
, cs
->cpus_allowed
);
2298 * We own tsk->cpus_allowed, nobody can change it under us.
2300 * But we used cs && cs->cpus_allowed lockless and thus can
2301 * race with cgroup_attach_task() or update_cpumask() and get
2302 * the wrong tsk->cpus_allowed. However, both cases imply the
2303 * subsequent cpuset_change_cpumask()->set_cpus_allowed_ptr()
2304 * which takes task_rq_lock().
2306 * If we are called after it dropped the lock we must see all
2307 * changes in tsk_cs()->cpus_allowed. Otherwise we can temporary
2308 * set any mask even if it is not right from task_cs() pov,
2309 * the pending set_cpus_allowed_ptr() will fix things.
2311 * select_fallback_rq() will fix things ups and set cpu_possible_mask
2316 void cpuset_init_current_mems_allowed(void)
2318 nodes_setall(current
->mems_allowed
);
2322 * cpuset_mems_allowed - return mems_allowed mask from a tasks cpuset.
2323 * @tsk: pointer to task_struct from which to obtain cpuset->mems_allowed.
2325 * Description: Returns the nodemask_t mems_allowed of the cpuset
2326 * attached to the specified @tsk. Guaranteed to return some non-empty
2327 * subset of node_states[N_MEMORY], even if this means going outside the
2331 nodemask_t
cpuset_mems_allowed(struct task_struct
*tsk
)
2335 mutex_lock(&callback_mutex
);
2337 guarantee_online_mems(task_cs(tsk
), &mask
);
2339 mutex_unlock(&callback_mutex
);
2345 * cpuset_nodemask_valid_mems_allowed - check nodemask vs. curremt mems_allowed
2346 * @nodemask: the nodemask to be checked
2348 * Are any of the nodes in the nodemask allowed in current->mems_allowed?
2350 int cpuset_nodemask_valid_mems_allowed(nodemask_t
*nodemask
)
2352 return nodes_intersects(*nodemask
, current
->mems_allowed
);
2356 * nearest_hardwall_ancestor() - Returns the nearest mem_exclusive or
2357 * mem_hardwall ancestor to the specified cpuset. Call holding
2358 * callback_mutex. If no ancestor is mem_exclusive or mem_hardwall
2359 * (an unusual configuration), then returns the root cpuset.
2361 static const struct cpuset
*nearest_hardwall_ancestor(const struct cpuset
*cs
)
2363 while (!(is_mem_exclusive(cs
) || is_mem_hardwall(cs
)) && cs
->parent
)
2369 * cpuset_node_allowed_softwall - Can we allocate on a memory node?
2370 * @node: is this an allowed node?
2371 * @gfp_mask: memory allocation flags
2373 * If we're in interrupt, yes, we can always allocate. If __GFP_THISNODE is
2374 * set, yes, we can always allocate. If node is in our task's mems_allowed,
2375 * yes. If it's not a __GFP_HARDWALL request and this node is in the nearest
2376 * hardwalled cpuset ancestor to this task's cpuset, yes. If the task has been
2377 * OOM killed and has access to memory reserves as specified by the TIF_MEMDIE
2381 * If __GFP_HARDWALL is set, cpuset_node_allowed_softwall() reduces to
2382 * cpuset_node_allowed_hardwall(). Otherwise, cpuset_node_allowed_softwall()
2383 * might sleep, and might allow a node from an enclosing cpuset.
2385 * cpuset_node_allowed_hardwall() only handles the simpler case of hardwall
2386 * cpusets, and never sleeps.
2388 * The __GFP_THISNODE placement logic is really handled elsewhere,
2389 * by forcibly using a zonelist starting at a specified node, and by
2390 * (in get_page_from_freelist()) refusing to consider the zones for
2391 * any node on the zonelist except the first. By the time any such
2392 * calls get to this routine, we should just shut up and say 'yes'.
2394 * GFP_USER allocations are marked with the __GFP_HARDWALL bit,
2395 * and do not allow allocations outside the current tasks cpuset
2396 * unless the task has been OOM killed as is marked TIF_MEMDIE.
2397 * GFP_KERNEL allocations are not so marked, so can escape to the
2398 * nearest enclosing hardwalled ancestor cpuset.
2400 * Scanning up parent cpusets requires callback_mutex. The
2401 * __alloc_pages() routine only calls here with __GFP_HARDWALL bit
2402 * _not_ set if it's a GFP_KERNEL allocation, and all nodes in the
2403 * current tasks mems_allowed came up empty on the first pass over
2404 * the zonelist. So only GFP_KERNEL allocations, if all nodes in the
2405 * cpuset are short of memory, might require taking the callback_mutex
2408 * The first call here from mm/page_alloc:get_page_from_freelist()
2409 * has __GFP_HARDWALL set in gfp_mask, enforcing hardwall cpusets,
2410 * so no allocation on a node outside the cpuset is allowed (unless
2411 * in interrupt, of course).
2413 * The second pass through get_page_from_freelist() doesn't even call
2414 * here for GFP_ATOMIC calls. For those calls, the __alloc_pages()
2415 * variable 'wait' is not set, and the bit ALLOC_CPUSET is not set
2416 * in alloc_flags. That logic and the checks below have the combined
2418 * in_interrupt - any node ok (current task context irrelevant)
2419 * GFP_ATOMIC - any node ok
2420 * TIF_MEMDIE - any node ok
2421 * GFP_KERNEL - any node in enclosing hardwalled cpuset ok
2422 * GFP_USER - only nodes in current tasks mems allowed ok.
2425 * Don't call cpuset_node_allowed_softwall if you can't sleep, unless you
2426 * pass in the __GFP_HARDWALL flag set in gfp_flag, which disables
2427 * the code that might scan up ancestor cpusets and sleep.
2429 int __cpuset_node_allowed_softwall(int node
, gfp_t gfp_mask
)
2431 const struct cpuset
*cs
; /* current cpuset ancestors */
2432 int allowed
; /* is allocation in zone z allowed? */
2434 if (in_interrupt() || (gfp_mask
& __GFP_THISNODE
))
2436 might_sleep_if(!(gfp_mask
& __GFP_HARDWALL
));
2437 if (node_isset(node
, current
->mems_allowed
))
2440 * Allow tasks that have access to memory reserves because they have
2441 * been OOM killed to get memory anywhere.
2443 if (unlikely(test_thread_flag(TIF_MEMDIE
)))
2445 if (gfp_mask
& __GFP_HARDWALL
) /* If hardwall request, stop here */
2448 if (current
->flags
& PF_EXITING
) /* Let dying task have memory */
2451 /* Not hardwall and node outside mems_allowed: scan up cpusets */
2452 mutex_lock(&callback_mutex
);
2455 cs
= nearest_hardwall_ancestor(task_cs(current
));
2456 task_unlock(current
);
2458 allowed
= node_isset(node
, cs
->mems_allowed
);
2459 mutex_unlock(&callback_mutex
);
2464 * cpuset_node_allowed_hardwall - Can we allocate on a memory node?
2465 * @node: is this an allowed node?
2466 * @gfp_mask: memory allocation flags
2468 * If we're in interrupt, yes, we can always allocate. If __GFP_THISNODE is
2469 * set, yes, we can always allocate. If node is in our task's mems_allowed,
2470 * yes. If the task has been OOM killed and has access to memory reserves as
2471 * specified by the TIF_MEMDIE flag, yes.
2474 * The __GFP_THISNODE placement logic is really handled elsewhere,
2475 * by forcibly using a zonelist starting at a specified node, and by
2476 * (in get_page_from_freelist()) refusing to consider the zones for
2477 * any node on the zonelist except the first. By the time any such
2478 * calls get to this routine, we should just shut up and say 'yes'.
2480 * Unlike the cpuset_node_allowed_softwall() variant, above,
2481 * this variant requires that the node be in the current task's
2482 * mems_allowed or that we're in interrupt. It does not scan up the
2483 * cpuset hierarchy for the nearest enclosing mem_exclusive cpuset.
2486 int __cpuset_node_allowed_hardwall(int node
, gfp_t gfp_mask
)
2488 if (in_interrupt() || (gfp_mask
& __GFP_THISNODE
))
2490 if (node_isset(node
, current
->mems_allowed
))
2493 * Allow tasks that have access to memory reserves because they have
2494 * been OOM killed to get memory anywhere.
2496 if (unlikely(test_thread_flag(TIF_MEMDIE
)))
2502 * cpuset_mem_spread_node() - On which node to begin search for a file page
2503 * cpuset_slab_spread_node() - On which node to begin search for a slab page
2505 * If a task is marked PF_SPREAD_PAGE or PF_SPREAD_SLAB (as for
2506 * tasks in a cpuset with is_spread_page or is_spread_slab set),
2507 * and if the memory allocation used cpuset_mem_spread_node()
2508 * to determine on which node to start looking, as it will for
2509 * certain page cache or slab cache pages such as used for file
2510 * system buffers and inode caches, then instead of starting on the
2511 * local node to look for a free page, rather spread the starting
2512 * node around the tasks mems_allowed nodes.
2514 * We don't have to worry about the returned node being offline
2515 * because "it can't happen", and even if it did, it would be ok.
2517 * The routines calling guarantee_online_mems() are careful to
2518 * only set nodes in task->mems_allowed that are online. So it
2519 * should not be possible for the following code to return an
2520 * offline node. But if it did, that would be ok, as this routine
2521 * is not returning the node where the allocation must be, only
2522 * the node where the search should start. The zonelist passed to
2523 * __alloc_pages() will include all nodes. If the slab allocator
2524 * is passed an offline node, it will fall back to the local node.
2525 * See kmem_cache_alloc_node().
2528 static int cpuset_spread_node(int *rotor
)
2532 node
= next_node(*rotor
, current
->mems_allowed
);
2533 if (node
== MAX_NUMNODES
)
2534 node
= first_node(current
->mems_allowed
);
2539 int cpuset_mem_spread_node(void)
2541 if (current
->cpuset_mem_spread_rotor
== NUMA_NO_NODE
)
2542 current
->cpuset_mem_spread_rotor
=
2543 node_random(¤t
->mems_allowed
);
2545 return cpuset_spread_node(¤t
->cpuset_mem_spread_rotor
);
2548 int cpuset_slab_spread_node(void)
2550 if (current
->cpuset_slab_spread_rotor
== NUMA_NO_NODE
)
2551 current
->cpuset_slab_spread_rotor
=
2552 node_random(¤t
->mems_allowed
);
2554 return cpuset_spread_node(¤t
->cpuset_slab_spread_rotor
);
2557 EXPORT_SYMBOL_GPL(cpuset_mem_spread_node
);
2560 * cpuset_mems_allowed_intersects - Does @tsk1's mems_allowed intersect @tsk2's?
2561 * @tsk1: pointer to task_struct of some task.
2562 * @tsk2: pointer to task_struct of some other task.
2564 * Description: Return true if @tsk1's mems_allowed intersects the
2565 * mems_allowed of @tsk2. Used by the OOM killer to determine if
2566 * one of the task's memory usage might impact the memory available
2570 int cpuset_mems_allowed_intersects(const struct task_struct
*tsk1
,
2571 const struct task_struct
*tsk2
)
2573 return nodes_intersects(tsk1
->mems_allowed
, tsk2
->mems_allowed
);
2577 * cpuset_print_task_mems_allowed - prints task's cpuset and mems_allowed
2578 * @task: pointer to task_struct of some task.
2580 * Description: Prints @task's name, cpuset name, and cached copy of its
2581 * mems_allowed to the kernel log. Must hold task_lock(task) to allow
2582 * dereferencing task_cs(task).
2584 void cpuset_print_task_mems_allowed(struct task_struct
*tsk
)
2586 struct dentry
*dentry
;
2588 dentry
= task_cs(tsk
)->css
.cgroup
->dentry
;
2589 spin_lock(&cpuset_buffer_lock
);
2590 snprintf(cpuset_name
, CPUSET_NAME_LEN
,
2591 dentry
? (const char *)dentry
->d_name
.name
: "/");
2592 nodelist_scnprintf(cpuset_nodelist
, CPUSET_NODELIST_LEN
,
2594 printk(KERN_INFO
"%s cpuset=%s mems_allowed=%s\n",
2595 tsk
->comm
, cpuset_name
, cpuset_nodelist
);
2596 spin_unlock(&cpuset_buffer_lock
);
2600 * Collection of memory_pressure is suppressed unless
2601 * this flag is enabled by writing "1" to the special
2602 * cpuset file 'memory_pressure_enabled' in the root cpuset.
2605 int cpuset_memory_pressure_enabled __read_mostly
;
2608 * cpuset_memory_pressure_bump - keep stats of per-cpuset reclaims.
2610 * Keep a running average of the rate of synchronous (direct)
2611 * page reclaim efforts initiated by tasks in each cpuset.
2613 * This represents the rate at which some task in the cpuset
2614 * ran low on memory on all nodes it was allowed to use, and
2615 * had to enter the kernels page reclaim code in an effort to
2616 * create more free memory by tossing clean pages or swapping
2617 * or writing dirty pages.
2619 * Display to user space in the per-cpuset read-only file
2620 * "memory_pressure". Value displayed is an integer
2621 * representing the recent rate of entry into the synchronous
2622 * (direct) page reclaim by any task attached to the cpuset.
2625 void __cpuset_memory_pressure_bump(void)
2628 fmeter_markevent(&task_cs(current
)->fmeter
);
2629 task_unlock(current
);
2632 #ifdef CONFIG_PROC_PID_CPUSET
2634 * proc_cpuset_show()
2635 * - Print tasks cpuset path into seq_file.
2636 * - Used for /proc/<pid>/cpuset.
2637 * - No need to task_lock(tsk) on this tsk->cpuset reference, as it
2638 * doesn't really matter if tsk->cpuset changes after we read it,
2639 * and we take cgroup_mutex, keeping cpuset_attach() from changing it
2642 static int proc_cpuset_show(struct seq_file
*m
, void *unused_v
)
2645 struct task_struct
*tsk
;
2647 struct cgroup_subsys_state
*css
;
2651 buf
= kmalloc(PAGE_SIZE
, GFP_KERNEL
);
2657 tsk
= get_pid_task(pid
, PIDTYPE_PID
);
2663 css
= task_subsys_state(tsk
, cpuset_subsys_id
);
2664 retval
= cgroup_path(css
->cgroup
, buf
, PAGE_SIZE
);
2671 put_task_struct(tsk
);
2678 static int cpuset_open(struct inode
*inode
, struct file
*file
)
2680 struct pid
*pid
= PROC_I(inode
)->pid
;
2681 return single_open(file
, proc_cpuset_show
, pid
);
2684 const struct file_operations proc_cpuset_operations
= {
2685 .open
= cpuset_open
,
2687 .llseek
= seq_lseek
,
2688 .release
= single_release
,
2690 #endif /* CONFIG_PROC_PID_CPUSET */
2692 /* Display task mems_allowed in /proc/<pid>/status file. */
2693 void cpuset_task_status_allowed(struct seq_file
*m
, struct task_struct
*task
)
2695 seq_printf(m
, "Mems_allowed:\t");
2696 seq_nodemask(m
, &task
->mems_allowed
);
2697 seq_printf(m
, "\n");
2698 seq_printf(m
, "Mems_allowed_list:\t");
2699 seq_nodemask_list(m
, &task
->mems_allowed
);
2700 seq_printf(m
, "\n");