4 * Processor and Memory placement constraints for sets of tasks.
6 * Copyright (C) 2003 BULL SA.
7 * Copyright (C) 2004-2006 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
18 * This file is subject to the terms and conditions of the GNU General Public
19 * License. See the file COPYING in the main directory of the Linux
20 * distribution for more details.
23 #include <linux/cpu.h>
24 #include <linux/cpumask.h>
25 #include <linux/cpuset.h>
26 #include <linux/err.h>
27 #include <linux/errno.h>
28 #include <linux/file.h>
30 #include <linux/init.h>
31 #include <linux/interrupt.h>
32 #include <linux/kernel.h>
33 #include <linux/kmod.h>
34 #include <linux/list.h>
35 #include <linux/mempolicy.h>
37 #include <linux/module.h>
38 #include <linux/mount.h>
39 #include <linux/namei.h>
40 #include <linux/pagemap.h>
41 #include <linux/proc_fs.h>
42 #include <linux/rcupdate.h>
43 #include <linux/sched.h>
44 #include <linux/seq_file.h>
45 #include <linux/security.h>
46 #include <linux/slab.h>
47 #include <linux/spinlock.h>
48 #include <linux/stat.h>
49 #include <linux/string.h>
50 #include <linux/time.h>
51 #include <linux/backing-dev.h>
52 #include <linux/sort.h>
54 #include <asm/uaccess.h>
55 #include <asm/atomic.h>
56 #include <linux/mutex.h>
59 * Tracks how many cpusets are currently defined in system.
60 * When there is only one cpuset (the root cpuset) we can
61 * short circuit some hooks.
63 int number_of_cpusets __read_mostly
;
65 /* Retrieve the cpuset from a cgroup */
66 struct cgroup_subsys cpuset_subsys
;
69 /* See "Frequency meter" comments, below. */
72 int cnt
; /* unprocessed events count */
73 int val
; /* most recent output value */
74 time_t time
; /* clock (secs) when val computed */
75 spinlock_t lock
; /* guards read or write of above */
79 struct cgroup_subsys_state css
;
81 unsigned long flags
; /* "unsigned long" so bitops work */
82 cpumask_t cpus_allowed
; /* CPUs allowed to tasks in cpuset */
83 nodemask_t mems_allowed
; /* Memory Nodes allowed to tasks */
85 struct cpuset
*parent
; /* my parent */
88 * Copy of global cpuset_mems_generation as of the most
89 * recent time this cpuset changed its mems_allowed.
93 struct fmeter fmeter
; /* memory_pressure filter */
96 /* Retrieve the cpuset for a cgroup */
97 static inline struct cpuset
*cgroup_cs(struct cgroup
*cont
)
99 return container_of(cgroup_subsys_state(cont
, cpuset_subsys_id
),
103 /* Retrieve the cpuset for a task */
104 static inline struct cpuset
*task_cs(struct task_struct
*task
)
106 return container_of(task_subsys_state(task
, cpuset_subsys_id
),
111 /* bits in struct cpuset flags field */
120 /* convenient tests for these bits */
121 static inline int is_cpu_exclusive(const struct cpuset
*cs
)
123 return test_bit(CS_CPU_EXCLUSIVE
, &cs
->flags
);
126 static inline int is_mem_exclusive(const struct cpuset
*cs
)
128 return test_bit(CS_MEM_EXCLUSIVE
, &cs
->flags
);
131 static inline int is_memory_migrate(const struct cpuset
*cs
)
133 return test_bit(CS_MEMORY_MIGRATE
, &cs
->flags
);
136 static inline int is_spread_page(const struct cpuset
*cs
)
138 return test_bit(CS_SPREAD_PAGE
, &cs
->flags
);
141 static inline int is_spread_slab(const struct cpuset
*cs
)
143 return test_bit(CS_SPREAD_SLAB
, &cs
->flags
);
147 * Increment this integer everytime any cpuset changes its
148 * mems_allowed value. Users of cpusets can track this generation
149 * number, and avoid having to lock and reload mems_allowed unless
150 * the cpuset they're using changes generation.
152 * A single, global generation is needed because attach_task() could
153 * reattach a task to a different cpuset, which must not have its
154 * generation numbers aliased with those of that tasks previous cpuset.
156 * Generations are needed for mems_allowed because one task cannot
157 * modify anothers memory placement. So we must enable every task,
158 * on every visit to __alloc_pages(), to efficiently check whether
159 * its current->cpuset->mems_allowed has changed, requiring an update
160 * of its current->mems_allowed.
162 * Since cpuset_mems_generation is guarded by manage_mutex,
163 * there is no need to mark it atomic.
165 static int cpuset_mems_generation
;
167 static struct cpuset top_cpuset
= {
168 .flags
= ((1 << CS_CPU_EXCLUSIVE
) | (1 << CS_MEM_EXCLUSIVE
)),
169 .cpus_allowed
= CPU_MASK_ALL
,
170 .mems_allowed
= NODE_MASK_ALL
,
174 * We have two global cpuset mutexes below. They can nest.
175 * It is ok to first take manage_mutex, then nest callback_mutex. We also
176 * require taking task_lock() when dereferencing a tasks cpuset pointer.
177 * See "The task_lock() exception", at the end of this comment.
179 * A task must hold both mutexes to modify cpusets. If a task
180 * holds manage_mutex, then it blocks others wanting that mutex,
181 * ensuring that it is the only task able to also acquire callback_mutex
182 * and be able to modify cpusets. It can perform various checks on
183 * the cpuset structure first, knowing nothing will change. It can
184 * also allocate memory while just holding manage_mutex. While it is
185 * performing these checks, various callback routines can briefly
186 * acquire callback_mutex to query cpusets. Once it is ready to make
187 * the changes, it takes callback_mutex, blocking everyone else.
189 * Calls to the kernel memory allocator can not be made while holding
190 * callback_mutex, as that would risk double tripping on callback_mutex
191 * from one of the callbacks into the cpuset code from within
194 * If a task is only holding callback_mutex, then it has read-only
197 * The task_struct fields mems_allowed and mems_generation may only
198 * be accessed in the context of that task, so require no locks.
200 * Any task can increment and decrement the count field without lock.
201 * So in general, code holding manage_mutex or callback_mutex can't rely
202 * on the count field not changing. However, if the count goes to
203 * zero, then only attach_task(), which holds both mutexes, can
204 * increment it again. Because a count of zero means that no tasks
205 * are currently attached, therefore there is no way a task attached
206 * to that cpuset can fork (the other way to increment the count).
207 * So code holding manage_mutex or callback_mutex can safely assume that
208 * if the count is zero, it will stay zero. Similarly, if a task
209 * holds manage_mutex or callback_mutex on a cpuset with zero count, it
210 * knows that the cpuset won't be removed, as cpuset_rmdir() needs
211 * both of those mutexes.
213 * The cpuset_common_file_write handler for operations that modify
214 * the cpuset hierarchy holds manage_mutex across the entire operation,
215 * single threading all such cpuset modifications across the system.
217 * The cpuset_common_file_read() handlers only hold callback_mutex across
218 * small pieces of code, such as when reading out possibly multi-word
219 * cpumasks and nodemasks.
221 * The fork and exit callbacks cpuset_fork() and cpuset_exit(), don't
222 * (usually) take either mutex. These are the two most performance
223 * critical pieces of code here. The exception occurs on cpuset_exit(),
224 * when a task in a notify_on_release cpuset exits. Then manage_mutex
225 * is taken, and if the cpuset count is zero, a usermode call made
226 * to /sbin/cpuset_release_agent with the name of the cpuset (path
227 * relative to the root of cpuset file system) as the argument.
229 * A cpuset can only be deleted if both its 'count' of using tasks
230 * is zero, and its list of 'children' cpusets is empty. Since all
231 * tasks in the system use _some_ cpuset, and since there is always at
232 * least one task in the system (init), therefore, top_cpuset
233 * always has either children cpusets and/or using tasks. So we don't
234 * need a special hack to ensure that top_cpuset cannot be deleted.
236 * The above "Tale of Two Semaphores" would be complete, but for:
238 * The task_lock() exception
240 * The need for this exception arises from the action of attach_task(),
241 * which overwrites one tasks cpuset pointer with another. It does
242 * so using both mutexes, however there are several performance
243 * critical places that need to reference task->cpuset without the
244 * expense of grabbing a system global mutex. Therefore except as
245 * noted below, when dereferencing or, as in attach_task(), modifying
246 * a tasks cpuset pointer we use task_lock(), which acts on a spinlock
247 * (task->alloc_lock) already in the task_struct routinely used for
250 * P.S. One more locking exception. RCU is used to guard the
251 * update of a tasks cpuset pointer by attach_task() and the
252 * access of task->cpuset->mems_generation via that pointer in
253 * the routine cpuset_update_task_memory_state().
256 static DEFINE_MUTEX(callback_mutex
);
258 /* This is ugly, but preserves the userspace API for existing cpuset
259 * users. If someone tries to mount the "cpuset" filesystem, we
260 * silently switch it to mount "cgroup" instead */
261 static int cpuset_get_sb(struct file_system_type
*fs_type
,
262 int flags
, const char *unused_dev_name
,
263 void *data
, struct vfsmount
*mnt
)
265 struct file_system_type
*cgroup_fs
= get_fs_type("cgroup");
270 "release_agent=/sbin/cpuset_release_agent";
271 ret
= cgroup_fs
->get_sb(cgroup_fs
, flags
,
272 unused_dev_name
, mountopts
, mnt
);
273 put_filesystem(cgroup_fs
);
278 static struct file_system_type cpuset_fs_type
= {
280 .get_sb
= cpuset_get_sb
,
284 * Return in *pmask the portion of a cpusets's cpus_allowed that
285 * are online. If none are online, walk up the cpuset hierarchy
286 * until we find one that does have some online cpus. If we get
287 * all the way to the top and still haven't found any online cpus,
288 * return cpu_online_map. Or if passed a NULL cs from an exit'ing
289 * task, return cpu_online_map.
291 * One way or another, we guarantee to return some non-empty subset
294 * Call with callback_mutex held.
297 static void guarantee_online_cpus(const struct cpuset
*cs
, cpumask_t
*pmask
)
299 while (cs
&& !cpus_intersects(cs
->cpus_allowed
, cpu_online_map
))
302 cpus_and(*pmask
, cs
->cpus_allowed
, cpu_online_map
);
304 *pmask
= cpu_online_map
;
305 BUG_ON(!cpus_intersects(*pmask
, cpu_online_map
));
309 * Return in *pmask the portion of a cpusets's mems_allowed that
310 * are online, with memory. If none are online with memory, walk
311 * up the cpuset hierarchy until we find one that does have some
312 * online mems. If we get all the way to the top and still haven't
313 * found any online mems, return node_states[N_HIGH_MEMORY].
315 * One way or another, we guarantee to return some non-empty subset
316 * of node_states[N_HIGH_MEMORY].
318 * Call with callback_mutex held.
321 static void guarantee_online_mems(const struct cpuset
*cs
, nodemask_t
*pmask
)
323 while (cs
&& !nodes_intersects(cs
->mems_allowed
,
324 node_states
[N_HIGH_MEMORY
]))
327 nodes_and(*pmask
, cs
->mems_allowed
,
328 node_states
[N_HIGH_MEMORY
]);
330 *pmask
= node_states
[N_HIGH_MEMORY
];
331 BUG_ON(!nodes_intersects(*pmask
, node_states
[N_HIGH_MEMORY
]));
335 * cpuset_update_task_memory_state - update task memory placement
337 * If the current tasks cpusets mems_allowed changed behind our
338 * backs, update current->mems_allowed, mems_generation and task NUMA
339 * mempolicy to the new value.
341 * Task mempolicy is updated by rebinding it relative to the
342 * current->cpuset if a task has its memory placement changed.
343 * Do not call this routine if in_interrupt().
345 * Call without callback_mutex or task_lock() held. May be
346 * called with or without manage_mutex held. Thanks in part to
347 * 'the_top_cpuset_hack', the tasks cpuset pointer will never
348 * be NULL. This routine also might acquire callback_mutex and
349 * current->mm->mmap_sem during call.
351 * Reading current->cpuset->mems_generation doesn't need task_lock
352 * to guard the current->cpuset derefence, because it is guarded
353 * from concurrent freeing of current->cpuset by attach_task(),
356 * The rcu_dereference() is technically probably not needed,
357 * as I don't actually mind if I see a new cpuset pointer but
358 * an old value of mems_generation. However this really only
359 * matters on alpha systems using cpusets heavily. If I dropped
360 * that rcu_dereference(), it would save them a memory barrier.
361 * For all other arch's, rcu_dereference is a no-op anyway, and for
362 * alpha systems not using cpusets, another planned optimization,
363 * avoiding the rcu critical section for tasks in the root cpuset
364 * which is statically allocated, so can't vanish, will make this
365 * irrelevant. Better to use RCU as intended, than to engage in
366 * some cute trick to save a memory barrier that is impossible to
367 * test, for alpha systems using cpusets heavily, which might not
370 * This routine is needed to update the per-task mems_allowed data,
371 * within the tasks context, when it is trying to allocate memory
372 * (in various mm/mempolicy.c routines) and notices that some other
373 * task has been modifying its cpuset.
376 void cpuset_update_task_memory_state(void)
378 int my_cpusets_mem_gen
;
379 struct task_struct
*tsk
= current
;
382 if (task_cs(tsk
) == &top_cpuset
) {
383 /* Don't need rcu for top_cpuset. It's never freed. */
384 my_cpusets_mem_gen
= top_cpuset
.mems_generation
;
387 my_cpusets_mem_gen
= task_cs(current
)->mems_generation
;
391 if (my_cpusets_mem_gen
!= tsk
->cpuset_mems_generation
) {
392 mutex_lock(&callback_mutex
);
394 cs
= task_cs(tsk
); /* Maybe changed when task not locked */
395 guarantee_online_mems(cs
, &tsk
->mems_allowed
);
396 tsk
->cpuset_mems_generation
= cs
->mems_generation
;
397 if (is_spread_page(cs
))
398 tsk
->flags
|= PF_SPREAD_PAGE
;
400 tsk
->flags
&= ~PF_SPREAD_PAGE
;
401 if (is_spread_slab(cs
))
402 tsk
->flags
|= PF_SPREAD_SLAB
;
404 tsk
->flags
&= ~PF_SPREAD_SLAB
;
406 mutex_unlock(&callback_mutex
);
407 mpol_rebind_task(tsk
, &tsk
->mems_allowed
);
412 * is_cpuset_subset(p, q) - Is cpuset p a subset of cpuset q?
414 * One cpuset is a subset of another if all its allowed CPUs and
415 * Memory Nodes are a subset of the other, and its exclusive flags
416 * are only set if the other's are set. Call holding manage_mutex.
419 static int is_cpuset_subset(const struct cpuset
*p
, const struct cpuset
*q
)
421 return cpus_subset(p
->cpus_allowed
, q
->cpus_allowed
) &&
422 nodes_subset(p
->mems_allowed
, q
->mems_allowed
) &&
423 is_cpu_exclusive(p
) <= is_cpu_exclusive(q
) &&
424 is_mem_exclusive(p
) <= is_mem_exclusive(q
);
428 * validate_change() - Used to validate that any proposed cpuset change
429 * follows the structural rules for cpusets.
431 * If we replaced the flag and mask values of the current cpuset
432 * (cur) with those values in the trial cpuset (trial), would
433 * our various subset and exclusive rules still be valid? Presumes
436 * 'cur' is the address of an actual, in-use cpuset. Operations
437 * such as list traversal that depend on the actual address of the
438 * cpuset in the list must use cur below, not trial.
440 * 'trial' is the address of bulk structure copy of cur, with
441 * perhaps one or more of the fields cpus_allowed, mems_allowed,
442 * or flags changed to new, trial values.
444 * Return 0 if valid, -errno if not.
447 static int validate_change(const struct cpuset
*cur
, const struct cpuset
*trial
)
450 struct cpuset
*c
, *par
;
452 /* Each of our child cpusets must be a subset of us */
453 list_for_each_entry(cont
, &cur
->css
.cgroup
->children
, sibling
) {
454 if (!is_cpuset_subset(cgroup_cs(cont
), trial
))
458 /* Remaining checks don't apply to root cpuset */
459 if (cur
== &top_cpuset
)
464 /* We must be a subset of our parent cpuset */
465 if (!is_cpuset_subset(trial
, par
))
468 /* If either I or some sibling (!= me) is exclusive, we can't overlap */
469 list_for_each_entry(cont
, &par
->css
.cgroup
->children
, sibling
) {
471 if ((is_cpu_exclusive(trial
) || is_cpu_exclusive(c
)) &&
473 cpus_intersects(trial
->cpus_allowed
, c
->cpus_allowed
))
475 if ((is_mem_exclusive(trial
) || is_mem_exclusive(c
)) &&
477 nodes_intersects(trial
->mems_allowed
, c
->mems_allowed
))
485 * Call with manage_mutex held. May take callback_mutex during call.
488 static int update_cpumask(struct cpuset
*cs
, char *buf
)
490 struct cpuset trialcs
;
493 /* top_cpuset.cpus_allowed tracks cpu_online_map; it's read-only */
494 if (cs
== &top_cpuset
)
500 * We allow a cpuset's cpus_allowed to be empty; if it has attached
501 * tasks, we'll catch it later when we validate the change and return
504 if (!buf
[0] || (buf
[0] == '\n' && !buf
[1])) {
505 cpus_clear(trialcs
.cpus_allowed
);
507 retval
= cpulist_parse(buf
, trialcs
.cpus_allowed
);
511 cpus_and(trialcs
.cpus_allowed
, trialcs
.cpus_allowed
, cpu_online_map
);
512 /* cpus_allowed cannot be empty for a cpuset with attached tasks. */
513 if (cgroup_task_count(cs
->css
.cgroup
) &&
514 cpus_empty(trialcs
.cpus_allowed
))
516 retval
= validate_change(cs
, &trialcs
);
519 mutex_lock(&callback_mutex
);
520 cs
->cpus_allowed
= trialcs
.cpus_allowed
;
521 mutex_unlock(&callback_mutex
);
528 * Migrate memory region from one set of nodes to another.
530 * Temporarilly set tasks mems_allowed to target nodes of migration,
531 * so that the migration code can allocate pages on these nodes.
533 * Call holding manage_mutex, so our current->cpuset won't change
534 * during this call, as manage_mutex holds off any attach_task()
535 * calls. Therefore we don't need to take task_lock around the
536 * call to guarantee_online_mems(), as we know no one is changing
539 * Hold callback_mutex around the two modifications of our tasks
540 * mems_allowed to synchronize with cpuset_mems_allowed().
542 * While the mm_struct we are migrating is typically from some
543 * other task, the task_struct mems_allowed that we are hacking
544 * is for our current task, which must allocate new pages for that
545 * migrating memory region.
547 * We call cpuset_update_task_memory_state() before hacking
548 * our tasks mems_allowed, so that we are assured of being in
549 * sync with our tasks cpuset, and in particular, callbacks to
550 * cpuset_update_task_memory_state() from nested page allocations
551 * won't see any mismatch of our cpuset and task mems_generation
552 * values, so won't overwrite our hacked tasks mems_allowed
556 static void cpuset_migrate_mm(struct mm_struct
*mm
, const nodemask_t
*from
,
557 const nodemask_t
*to
)
559 struct task_struct
*tsk
= current
;
561 cpuset_update_task_memory_state();
563 mutex_lock(&callback_mutex
);
564 tsk
->mems_allowed
= *to
;
565 mutex_unlock(&callback_mutex
);
567 do_migrate_pages(mm
, from
, to
, MPOL_MF_MOVE_ALL
);
569 mutex_lock(&callback_mutex
);
570 guarantee_online_mems(task_cs(tsk
),&tsk
->mems_allowed
);
571 mutex_unlock(&callback_mutex
);
575 * Handle user request to change the 'mems' memory placement
576 * of a cpuset. Needs to validate the request, update the
577 * cpusets mems_allowed and mems_generation, and for each
578 * task in the cpuset, rebind any vma mempolicies and if
579 * the cpuset is marked 'memory_migrate', migrate the tasks
580 * pages to the new memory.
582 * Call with manage_mutex held. May take callback_mutex during call.
583 * Will take tasklist_lock, scan tasklist for tasks in cpuset cs,
584 * lock each such tasks mm->mmap_sem, scan its vma's and rebind
585 * their mempolicies to the cpusets new mems_allowed.
588 static void *cpuset_being_rebound
;
590 static int update_nodemask(struct cpuset
*cs
, char *buf
)
592 struct cpuset trialcs
;
594 struct task_struct
*p
;
595 struct mm_struct
**mmarray
;
600 struct cgroup_iter it
;
603 * top_cpuset.mems_allowed tracks node_stats[N_HIGH_MEMORY];
606 if (cs
== &top_cpuset
)
612 * We allow a cpuset's mems_allowed to be empty; if it has attached
613 * tasks, we'll catch it later when we validate the change and return
616 if (!buf
[0] || (buf
[0] == '\n' && !buf
[1])) {
617 nodes_clear(trialcs
.mems_allowed
);
619 retval
= nodelist_parse(buf
, trialcs
.mems_allowed
);
622 if (!nodes_intersects(trialcs
.mems_allowed
,
623 node_states
[N_HIGH_MEMORY
])) {
625 * error if only memoryless nodes specified.
632 * Exclude memoryless nodes. We know that trialcs.mems_allowed
633 * contains at least one node with memory.
635 nodes_and(trialcs
.mems_allowed
, trialcs
.mems_allowed
,
636 node_states
[N_HIGH_MEMORY
]);
637 oldmem
= cs
->mems_allowed
;
638 if (nodes_equal(oldmem
, trialcs
.mems_allowed
)) {
639 retval
= 0; /* Too easy - nothing to do */
642 /* mems_allowed cannot be empty for a cpuset with attached tasks. */
643 if (cgroup_task_count(cs
->css
.cgroup
) &&
644 nodes_empty(trialcs
.mems_allowed
)) {
648 retval
= validate_change(cs
, &trialcs
);
652 mutex_lock(&callback_mutex
);
653 cs
->mems_allowed
= trialcs
.mems_allowed
;
654 cs
->mems_generation
= cpuset_mems_generation
++;
655 mutex_unlock(&callback_mutex
);
657 cpuset_being_rebound
= cs
; /* causes mpol_copy() rebind */
659 fudge
= 10; /* spare mmarray[] slots */
660 fudge
+= cpus_weight(cs
->cpus_allowed
); /* imagine one fork-bomb/cpu */
664 * Allocate mmarray[] to hold mm reference for each task
665 * in cpuset cs. Can't kmalloc GFP_KERNEL while holding
666 * tasklist_lock. We could use GFP_ATOMIC, but with a
667 * few more lines of code, we can retry until we get a big
668 * enough mmarray[] w/o using GFP_ATOMIC.
671 ntasks
= cgroup_task_count(cs
->css
.cgroup
); /* guess */
673 mmarray
= kmalloc(ntasks
* sizeof(*mmarray
), GFP_KERNEL
);
676 read_lock(&tasklist_lock
); /* block fork */
677 if (cgroup_task_count(cs
->css
.cgroup
) <= ntasks
)
678 break; /* got enough */
679 read_unlock(&tasklist_lock
); /* try again */
685 /* Load up mmarray[] with mm reference for each task in cpuset. */
686 cgroup_iter_start(cs
->css
.cgroup
, &it
);
687 while ((p
= cgroup_iter_next(cs
->css
.cgroup
, &it
))) {
688 struct mm_struct
*mm
;
692 "Cpuset mempolicy rebind incomplete.\n");
700 cgroup_iter_end(cs
->css
.cgroup
, &it
);
701 read_unlock(&tasklist_lock
);
704 * Now that we've dropped the tasklist spinlock, we can
705 * rebind the vma mempolicies of each mm in mmarray[] to their
706 * new cpuset, and release that mm. The mpol_rebind_mm()
707 * call takes mmap_sem, which we couldn't take while holding
708 * tasklist_lock. Forks can happen again now - the mpol_copy()
709 * cpuset_being_rebound check will catch such forks, and rebind
710 * their vma mempolicies too. Because we still hold the global
711 * cpuset manage_mutex, we know that no other rebind effort will
712 * be contending for the global variable cpuset_being_rebound.
713 * It's ok if we rebind the same mm twice; mpol_rebind_mm()
714 * is idempotent. Also migrate pages in each mm to new nodes.
716 migrate
= is_memory_migrate(cs
);
717 for (i
= 0; i
< n
; i
++) {
718 struct mm_struct
*mm
= mmarray
[i
];
720 mpol_rebind_mm(mm
, &cs
->mems_allowed
);
722 cpuset_migrate_mm(mm
, &oldmem
, &cs
->mems_allowed
);
726 /* We're done rebinding vma's to this cpusets new mems_allowed. */
728 cpuset_being_rebound
= NULL
;
734 int current_cpuset_is_being_rebound(void)
736 return task_cs(current
) == cpuset_being_rebound
;
740 * Call with manage_mutex held.
743 static int update_memory_pressure_enabled(struct cpuset
*cs
, char *buf
)
745 if (simple_strtoul(buf
, NULL
, 10) != 0)
746 cpuset_memory_pressure_enabled
= 1;
748 cpuset_memory_pressure_enabled
= 0;
753 * update_flag - read a 0 or a 1 in a file and update associated flag
754 * bit: the bit to update (CS_CPU_EXCLUSIVE, CS_MEM_EXCLUSIVE,
755 * CS_NOTIFY_ON_RELEASE, CS_MEMORY_MIGRATE,
756 * CS_SPREAD_PAGE, CS_SPREAD_SLAB)
757 * cs: the cpuset to update
758 * buf: the buffer where we read the 0 or 1
760 * Call with manage_mutex held.
763 static int update_flag(cpuset_flagbits_t bit
, struct cpuset
*cs
, char *buf
)
766 struct cpuset trialcs
;
769 turning_on
= (simple_strtoul(buf
, NULL
, 10) != 0);
773 set_bit(bit
, &trialcs
.flags
);
775 clear_bit(bit
, &trialcs
.flags
);
777 err
= validate_change(cs
, &trialcs
);
780 mutex_lock(&callback_mutex
);
781 cs
->flags
= trialcs
.flags
;
782 mutex_unlock(&callback_mutex
);
788 * Frequency meter - How fast is some event occurring?
790 * These routines manage a digitally filtered, constant time based,
791 * event frequency meter. There are four routines:
792 * fmeter_init() - initialize a frequency meter.
793 * fmeter_markevent() - called each time the event happens.
794 * fmeter_getrate() - returns the recent rate of such events.
795 * fmeter_update() - internal routine used to update fmeter.
797 * A common data structure is passed to each of these routines,
798 * which is used to keep track of the state required to manage the
799 * frequency meter and its digital filter.
801 * The filter works on the number of events marked per unit time.
802 * The filter is single-pole low-pass recursive (IIR). The time unit
803 * is 1 second. Arithmetic is done using 32-bit integers scaled to
804 * simulate 3 decimal digits of precision (multiplied by 1000).
806 * With an FM_COEF of 933, and a time base of 1 second, the filter
807 * has a half-life of 10 seconds, meaning that if the events quit
808 * happening, then the rate returned from the fmeter_getrate()
809 * will be cut in half each 10 seconds, until it converges to zero.
811 * It is not worth doing a real infinitely recursive filter. If more
812 * than FM_MAXTICKS ticks have elapsed since the last filter event,
813 * just compute FM_MAXTICKS ticks worth, by which point the level
816 * Limit the count of unprocessed events to FM_MAXCNT, so as to avoid
817 * arithmetic overflow in the fmeter_update() routine.
819 * Given the simple 32 bit integer arithmetic used, this meter works
820 * best for reporting rates between one per millisecond (msec) and
821 * one per 32 (approx) seconds. At constant rates faster than one
822 * per msec it maxes out at values just under 1,000,000. At constant
823 * rates between one per msec, and one per second it will stabilize
824 * to a value N*1000, where N is the rate of events per second.
825 * At constant rates between one per second and one per 32 seconds,
826 * it will be choppy, moving up on the seconds that have an event,
827 * and then decaying until the next event. At rates slower than
828 * about one in 32 seconds, it decays all the way back to zero between
832 #define FM_COEF 933 /* coefficient for half-life of 10 secs */
833 #define FM_MAXTICKS ((time_t)99) /* useless computing more ticks than this */
834 #define FM_MAXCNT 1000000 /* limit cnt to avoid overflow */
835 #define FM_SCALE 1000 /* faux fixed point scale */
837 /* Initialize a frequency meter */
838 static void fmeter_init(struct fmeter
*fmp
)
843 spin_lock_init(&fmp
->lock
);
846 /* Internal meter update - process cnt events and update value */
847 static void fmeter_update(struct fmeter
*fmp
)
849 time_t now
= get_seconds();
850 time_t ticks
= now
- fmp
->time
;
855 ticks
= min(FM_MAXTICKS
, ticks
);
857 fmp
->val
= (FM_COEF
* fmp
->val
) / FM_SCALE
;
860 fmp
->val
+= ((FM_SCALE
- FM_COEF
) * fmp
->cnt
) / FM_SCALE
;
864 /* Process any previous ticks, then bump cnt by one (times scale). */
865 static void fmeter_markevent(struct fmeter
*fmp
)
867 spin_lock(&fmp
->lock
);
869 fmp
->cnt
= min(FM_MAXCNT
, fmp
->cnt
+ FM_SCALE
);
870 spin_unlock(&fmp
->lock
);
873 /* Process any previous ticks, then return current value. */
874 static int fmeter_getrate(struct fmeter
*fmp
)
878 spin_lock(&fmp
->lock
);
881 spin_unlock(&fmp
->lock
);
885 static int cpuset_can_attach(struct cgroup_subsys
*ss
,
886 struct cgroup
*cont
, struct task_struct
*tsk
)
888 struct cpuset
*cs
= cgroup_cs(cont
);
890 if (cpus_empty(cs
->cpus_allowed
) || nodes_empty(cs
->mems_allowed
))
893 return security_task_setscheduler(tsk
, 0, NULL
);
896 static void cpuset_attach(struct cgroup_subsys
*ss
,
897 struct cgroup
*cont
, struct cgroup
*oldcont
,
898 struct task_struct
*tsk
)
902 struct mm_struct
*mm
;
903 struct cpuset
*cs
= cgroup_cs(cont
);
904 struct cpuset
*oldcs
= cgroup_cs(oldcont
);
906 mutex_lock(&callback_mutex
);
907 guarantee_online_cpus(cs
, &cpus
);
908 set_cpus_allowed(tsk
, cpus
);
909 mutex_unlock(&callback_mutex
);
911 from
= oldcs
->mems_allowed
;
912 to
= cs
->mems_allowed
;
913 mm
= get_task_mm(tsk
);
915 mpol_rebind_mm(mm
, &to
);
916 if (is_memory_migrate(cs
))
917 cpuset_migrate_mm(mm
, &from
, &to
);
923 /* The various types of files and directories in a cpuset file system */
931 FILE_MEMORY_PRESSURE_ENABLED
,
932 FILE_MEMORY_PRESSURE
,
937 static ssize_t
cpuset_common_file_write(struct cgroup
*cont
,
940 const char __user
*userbuf
,
941 size_t nbytes
, loff_t
*unused_ppos
)
943 struct cpuset
*cs
= cgroup_cs(cont
);
944 cpuset_filetype_t type
= cft
->private;
948 /* Crude upper limit on largest legitimate cpulist user might write. */
949 if (nbytes
> 100 + 6 * max(NR_CPUS
, MAX_NUMNODES
))
952 /* +1 for nul-terminator */
953 if ((buffer
= kmalloc(nbytes
+ 1, GFP_KERNEL
)) == 0)
956 if (copy_from_user(buffer
, userbuf
, nbytes
)) {
960 buffer
[nbytes
] = 0; /* nul-terminate */
964 if (cgroup_is_removed(cont
)) {
971 retval
= update_cpumask(cs
, buffer
);
974 retval
= update_nodemask(cs
, buffer
);
976 case FILE_CPU_EXCLUSIVE
:
977 retval
= update_flag(CS_CPU_EXCLUSIVE
, cs
, buffer
);
979 case FILE_MEM_EXCLUSIVE
:
980 retval
= update_flag(CS_MEM_EXCLUSIVE
, cs
, buffer
);
982 case FILE_MEMORY_MIGRATE
:
983 retval
= update_flag(CS_MEMORY_MIGRATE
, cs
, buffer
);
985 case FILE_MEMORY_PRESSURE_ENABLED
:
986 retval
= update_memory_pressure_enabled(cs
, buffer
);
988 case FILE_MEMORY_PRESSURE
:
991 case FILE_SPREAD_PAGE
:
992 retval
= update_flag(CS_SPREAD_PAGE
, cs
, buffer
);
993 cs
->mems_generation
= cpuset_mems_generation
++;
995 case FILE_SPREAD_SLAB
:
996 retval
= update_flag(CS_SPREAD_SLAB
, cs
, buffer
);
997 cs
->mems_generation
= cpuset_mems_generation
++;
1014 * These ascii lists should be read in a single call, by using a user
1015 * buffer large enough to hold the entire map. If read in smaller
1016 * chunks, there is no guarantee of atomicity. Since the display format
1017 * used, list of ranges of sequential numbers, is variable length,
1018 * and since these maps can change value dynamically, one could read
1019 * gibberish by doing partial reads while a list was changing.
1020 * A single large read to a buffer that crosses a page boundary is
1021 * ok, because the result being copied to user land is not recomputed
1022 * across a page fault.
1025 static int cpuset_sprintf_cpulist(char *page
, struct cpuset
*cs
)
1029 mutex_lock(&callback_mutex
);
1030 mask
= cs
->cpus_allowed
;
1031 mutex_unlock(&callback_mutex
);
1033 return cpulist_scnprintf(page
, PAGE_SIZE
, mask
);
1036 static int cpuset_sprintf_memlist(char *page
, struct cpuset
*cs
)
1040 mutex_lock(&callback_mutex
);
1041 mask
= cs
->mems_allowed
;
1042 mutex_unlock(&callback_mutex
);
1044 return nodelist_scnprintf(page
, PAGE_SIZE
, mask
);
1047 static ssize_t
cpuset_common_file_read(struct cgroup
*cont
,
1051 size_t nbytes
, loff_t
*ppos
)
1053 struct cpuset
*cs
= cgroup_cs(cont
);
1054 cpuset_filetype_t type
= cft
->private;
1059 if (!(page
= (char *)__get_free_page(GFP_TEMPORARY
)))
1066 s
+= cpuset_sprintf_cpulist(s
, cs
);
1069 s
+= cpuset_sprintf_memlist(s
, cs
);
1071 case FILE_CPU_EXCLUSIVE
:
1072 *s
++ = is_cpu_exclusive(cs
) ? '1' : '0';
1074 case FILE_MEM_EXCLUSIVE
:
1075 *s
++ = is_mem_exclusive(cs
) ? '1' : '0';
1077 case FILE_MEMORY_MIGRATE
:
1078 *s
++ = is_memory_migrate(cs
) ? '1' : '0';
1080 case FILE_MEMORY_PRESSURE_ENABLED
:
1081 *s
++ = cpuset_memory_pressure_enabled
? '1' : '0';
1083 case FILE_MEMORY_PRESSURE
:
1084 s
+= sprintf(s
, "%d", fmeter_getrate(&cs
->fmeter
));
1086 case FILE_SPREAD_PAGE
:
1087 *s
++ = is_spread_page(cs
) ? '1' : '0';
1089 case FILE_SPREAD_SLAB
:
1090 *s
++ = is_spread_slab(cs
) ? '1' : '0';
1098 retval
= simple_read_from_buffer(buf
, nbytes
, ppos
, page
, s
- page
);
1100 free_page((unsigned long)page
);
1109 * for the common functions, 'private' gives the type of file
1112 static struct cftype cft_cpus
= {
1114 .read
= cpuset_common_file_read
,
1115 .write
= cpuset_common_file_write
,
1116 .private = FILE_CPULIST
,
1119 static struct cftype cft_mems
= {
1121 .read
= cpuset_common_file_read
,
1122 .write
= cpuset_common_file_write
,
1123 .private = FILE_MEMLIST
,
1126 static struct cftype cft_cpu_exclusive
= {
1127 .name
= "cpu_exclusive",
1128 .read
= cpuset_common_file_read
,
1129 .write
= cpuset_common_file_write
,
1130 .private = FILE_CPU_EXCLUSIVE
,
1133 static struct cftype cft_mem_exclusive
= {
1134 .name
= "mem_exclusive",
1135 .read
= cpuset_common_file_read
,
1136 .write
= cpuset_common_file_write
,
1137 .private = FILE_MEM_EXCLUSIVE
,
1140 static struct cftype cft_memory_migrate
= {
1141 .name
= "memory_migrate",
1142 .read
= cpuset_common_file_read
,
1143 .write
= cpuset_common_file_write
,
1144 .private = FILE_MEMORY_MIGRATE
,
1147 static struct cftype cft_memory_pressure_enabled
= {
1148 .name
= "memory_pressure_enabled",
1149 .read
= cpuset_common_file_read
,
1150 .write
= cpuset_common_file_write
,
1151 .private = FILE_MEMORY_PRESSURE_ENABLED
,
1154 static struct cftype cft_memory_pressure
= {
1155 .name
= "memory_pressure",
1156 .read
= cpuset_common_file_read
,
1157 .write
= cpuset_common_file_write
,
1158 .private = FILE_MEMORY_PRESSURE
,
1161 static struct cftype cft_spread_page
= {
1162 .name
= "memory_spread_page",
1163 .read
= cpuset_common_file_read
,
1164 .write
= cpuset_common_file_write
,
1165 .private = FILE_SPREAD_PAGE
,
1168 static struct cftype cft_spread_slab
= {
1169 .name
= "memory_spread_slab",
1170 .read
= cpuset_common_file_read
,
1171 .write
= cpuset_common_file_write
,
1172 .private = FILE_SPREAD_SLAB
,
1175 static int cpuset_populate(struct cgroup_subsys
*ss
, struct cgroup
*cont
)
1179 if ((err
= cgroup_add_file(cont
, ss
, &cft_cpus
)) < 0)
1181 if ((err
= cgroup_add_file(cont
, ss
, &cft_mems
)) < 0)
1183 if ((err
= cgroup_add_file(cont
, ss
, &cft_cpu_exclusive
)) < 0)
1185 if ((err
= cgroup_add_file(cont
, ss
, &cft_mem_exclusive
)) < 0)
1187 if ((err
= cgroup_add_file(cont
, ss
, &cft_memory_migrate
)) < 0)
1189 if ((err
= cgroup_add_file(cont
, ss
, &cft_memory_pressure
)) < 0)
1191 if ((err
= cgroup_add_file(cont
, ss
, &cft_spread_page
)) < 0)
1193 if ((err
= cgroup_add_file(cont
, ss
, &cft_spread_slab
)) < 0)
1195 /* memory_pressure_enabled is in root cpuset only */
1196 if (err
== 0 && !cont
->parent
)
1197 err
= cgroup_add_file(cont
, ss
,
1198 &cft_memory_pressure_enabled
);
1203 * post_clone() is called at the end of cgroup_clone().
1204 * 'cgroup' was just created automatically as a result of
1205 * a cgroup_clone(), and the current task is about to
1206 * be moved into 'cgroup'.
1208 * Currently we refuse to set up the cgroup - thereby
1209 * refusing the task to be entered, and as a result refusing
1210 * the sys_unshare() or clone() which initiated it - if any
1211 * sibling cpusets have exclusive cpus or mem.
1213 * If this becomes a problem for some users who wish to
1214 * allow that scenario, then cpuset_post_clone() could be
1215 * changed to grant parent->cpus_allowed-sibling_cpus_exclusive
1216 * (and likewise for mems) to the new cgroup.
1218 static void cpuset_post_clone(struct cgroup_subsys
*ss
,
1219 struct cgroup
*cgroup
)
1221 struct cgroup
*parent
, *child
;
1222 struct cpuset
*cs
, *parent_cs
;
1224 parent
= cgroup
->parent
;
1225 list_for_each_entry(child
, &parent
->children
, sibling
) {
1226 cs
= cgroup_cs(child
);
1227 if (is_mem_exclusive(cs
) || is_cpu_exclusive(cs
))
1230 cs
= cgroup_cs(cgroup
);
1231 parent_cs
= cgroup_cs(parent
);
1233 cs
->mems_allowed
= parent_cs
->mems_allowed
;
1234 cs
->cpus_allowed
= parent_cs
->cpus_allowed
;
1239 * cpuset_create - create a cpuset
1240 * parent: cpuset that will be parent of the new cpuset.
1241 * name: name of the new cpuset. Will be strcpy'ed.
1242 * mode: mode to set on new inode
1244 * Must be called with the mutex on the parent inode held
1247 static struct cgroup_subsys_state
*cpuset_create(
1248 struct cgroup_subsys
*ss
,
1249 struct cgroup
*cont
)
1252 struct cpuset
*parent
;
1254 if (!cont
->parent
) {
1255 /* This is early initialization for the top cgroup */
1256 top_cpuset
.mems_generation
= cpuset_mems_generation
++;
1257 return &top_cpuset
.css
;
1259 parent
= cgroup_cs(cont
->parent
);
1260 cs
= kmalloc(sizeof(*cs
), GFP_KERNEL
);
1262 return ERR_PTR(-ENOMEM
);
1264 cpuset_update_task_memory_state();
1266 if (is_spread_page(parent
))
1267 set_bit(CS_SPREAD_PAGE
, &cs
->flags
);
1268 if (is_spread_slab(parent
))
1269 set_bit(CS_SPREAD_SLAB
, &cs
->flags
);
1270 cs
->cpus_allowed
= CPU_MASK_NONE
;
1271 cs
->mems_allowed
= NODE_MASK_NONE
;
1272 cs
->mems_generation
= cpuset_mems_generation
++;
1273 fmeter_init(&cs
->fmeter
);
1275 cs
->parent
= parent
;
1276 number_of_cpusets
++;
1280 static void cpuset_destroy(struct cgroup_subsys
*ss
, struct cgroup
*cont
)
1282 struct cpuset
*cs
= cgroup_cs(cont
);
1284 cpuset_update_task_memory_state();
1285 number_of_cpusets
--;
1289 struct cgroup_subsys cpuset_subsys
= {
1291 .create
= cpuset_create
,
1292 .destroy
= cpuset_destroy
,
1293 .can_attach
= cpuset_can_attach
,
1294 .attach
= cpuset_attach
,
1295 .populate
= cpuset_populate
,
1296 .post_clone
= cpuset_post_clone
,
1297 .subsys_id
= cpuset_subsys_id
,
1302 * cpuset_init_early - just enough so that the calls to
1303 * cpuset_update_task_memory_state() in early init code
1307 int __init
cpuset_init_early(void)
1309 top_cpuset
.mems_generation
= cpuset_mems_generation
++;
1315 * cpuset_init - initialize cpusets at system boot
1317 * Description: Initialize top_cpuset and the cpuset internal file system,
1320 int __init
cpuset_init(void)
1324 top_cpuset
.cpus_allowed
= CPU_MASK_ALL
;
1325 top_cpuset
.mems_allowed
= NODE_MASK_ALL
;
1327 fmeter_init(&top_cpuset
.fmeter
);
1328 top_cpuset
.mems_generation
= cpuset_mems_generation
++;
1330 err
= register_filesystem(&cpuset_fs_type
);
1334 number_of_cpusets
= 1;
1339 * If common_cpu_mem_hotplug_unplug(), below, unplugs any CPUs
1340 * or memory nodes, we need to walk over the cpuset hierarchy,
1341 * removing that CPU or node from all cpusets. If this removes the
1342 * last CPU or node from a cpuset, then the guarantee_online_cpus()
1343 * or guarantee_online_mems() code will use that emptied cpusets
1344 * parent online CPUs or nodes. Cpusets that were already empty of
1345 * CPUs or nodes are left empty.
1347 * This routine is intentionally inefficient in a couple of regards.
1348 * It will check all cpusets in a subtree even if the top cpuset of
1349 * the subtree has no offline CPUs or nodes. It checks both CPUs and
1350 * nodes, even though the caller could have been coded to know that
1351 * only one of CPUs or nodes needed to be checked on a given call.
1352 * This was done to minimize text size rather than cpu cycles.
1354 * Call with both manage_mutex and callback_mutex held.
1356 * Recursive, on depth of cpuset subtree.
1359 static void guarantee_online_cpus_mems_in_subtree(const struct cpuset
*cur
)
1361 struct cgroup
*cont
;
1364 /* Each of our child cpusets mems must be online */
1365 list_for_each_entry(cont
, &cur
->css
.cgroup
->children
, sibling
) {
1366 c
= cgroup_cs(cont
);
1367 guarantee_online_cpus_mems_in_subtree(c
);
1368 if (!cpus_empty(c
->cpus_allowed
))
1369 guarantee_online_cpus(c
, &c
->cpus_allowed
);
1370 if (!nodes_empty(c
->mems_allowed
))
1371 guarantee_online_mems(c
, &c
->mems_allowed
);
1376 * The cpus_allowed and mems_allowed nodemasks in the top_cpuset track
1377 * cpu_online_map and node_states[N_HIGH_MEMORY]. Force the top cpuset to
1378 * track what's online after any CPU or memory node hotplug or unplug
1381 * To ensure that we don't remove a CPU or node from the top cpuset
1382 * that is currently in use by a child cpuset (which would violate
1383 * the rule that cpusets must be subsets of their parent), we first
1384 * call the recursive routine guarantee_online_cpus_mems_in_subtree().
1386 * Since there are two callers of this routine, one for CPU hotplug
1387 * events and one for memory node hotplug events, we could have coded
1388 * two separate routines here. We code it as a single common routine
1389 * in order to minimize text size.
1392 static void common_cpu_mem_hotplug_unplug(void)
1395 mutex_lock(&callback_mutex
);
1397 guarantee_online_cpus_mems_in_subtree(&top_cpuset
);
1398 top_cpuset
.cpus_allowed
= cpu_online_map
;
1399 top_cpuset
.mems_allowed
= node_states
[N_HIGH_MEMORY
];
1401 mutex_unlock(&callback_mutex
);
1406 * The top_cpuset tracks what CPUs and Memory Nodes are online,
1407 * period. This is necessary in order to make cpusets transparent
1408 * (of no affect) on systems that are actively using CPU hotplug
1409 * but making no active use of cpusets.
1411 * This routine ensures that top_cpuset.cpus_allowed tracks
1412 * cpu_online_map on each CPU hotplug (cpuhp) event.
1415 static int cpuset_handle_cpuhp(struct notifier_block
*nb
,
1416 unsigned long phase
, void *cpu
)
1418 if (phase
== CPU_DYING
|| phase
== CPU_DYING_FROZEN
)
1421 common_cpu_mem_hotplug_unplug();
1425 #ifdef CONFIG_MEMORY_HOTPLUG
1427 * Keep top_cpuset.mems_allowed tracking node_states[N_HIGH_MEMORY].
1428 * Call this routine anytime after you change
1429 * node_states[N_HIGH_MEMORY].
1430 * See also the previous routine cpuset_handle_cpuhp().
1433 void cpuset_track_online_nodes(void)
1435 common_cpu_mem_hotplug_unplug();
1440 * cpuset_init_smp - initialize cpus_allowed
1442 * Description: Finish top cpuset after cpu, node maps are initialized
1445 void __init
cpuset_init_smp(void)
1447 top_cpuset
.cpus_allowed
= cpu_online_map
;
1448 top_cpuset
.mems_allowed
= node_states
[N_HIGH_MEMORY
];
1450 hotcpu_notifier(cpuset_handle_cpuhp
, 0);
1455 * cpuset_cpus_allowed - return cpus_allowed mask from a tasks cpuset.
1456 * @tsk: pointer to task_struct from which to obtain cpuset->cpus_allowed.
1458 * Description: Returns the cpumask_t cpus_allowed of the cpuset
1459 * attached to the specified @tsk. Guaranteed to return some non-empty
1460 * subset of cpu_online_map, even if this means going outside the
1464 cpumask_t
cpuset_cpus_allowed(struct task_struct
*tsk
)
1468 mutex_lock(&callback_mutex
);
1470 guarantee_online_cpus(task_cs(tsk
), &mask
);
1472 mutex_unlock(&callback_mutex
);
1477 void cpuset_init_current_mems_allowed(void)
1479 current
->mems_allowed
= NODE_MASK_ALL
;
1483 * cpuset_mems_allowed - return mems_allowed mask from a tasks cpuset.
1484 * @tsk: pointer to task_struct from which to obtain cpuset->mems_allowed.
1486 * Description: Returns the nodemask_t mems_allowed of the cpuset
1487 * attached to the specified @tsk. Guaranteed to return some non-empty
1488 * subset of node_states[N_HIGH_MEMORY], even if this means going outside the
1492 nodemask_t
cpuset_mems_allowed(struct task_struct
*tsk
)
1496 mutex_lock(&callback_mutex
);
1498 guarantee_online_mems(task_cs(tsk
), &mask
);
1500 mutex_unlock(&callback_mutex
);
1506 * cpuset_zonelist_valid_mems_allowed - check zonelist vs. curremt mems_allowed
1507 * @zl: the zonelist to be checked
1509 * Are any of the nodes on zonelist zl allowed in current->mems_allowed?
1511 int cpuset_zonelist_valid_mems_allowed(struct zonelist
*zl
)
1515 for (i
= 0; zl
->zones
[i
]; i
++) {
1516 int nid
= zone_to_nid(zl
->zones
[i
]);
1518 if (node_isset(nid
, current
->mems_allowed
))
1525 * nearest_exclusive_ancestor() - Returns the nearest mem_exclusive
1526 * ancestor to the specified cpuset. Call holding callback_mutex.
1527 * If no ancestor is mem_exclusive (an unusual configuration), then
1528 * returns the root cpuset.
1530 static const struct cpuset
*nearest_exclusive_ancestor(const struct cpuset
*cs
)
1532 while (!is_mem_exclusive(cs
) && cs
->parent
)
1538 * cpuset_zone_allowed_softwall - Can we allocate on zone z's memory node?
1539 * @z: is this zone on an allowed node?
1540 * @gfp_mask: memory allocation flags
1542 * If we're in interrupt, yes, we can always allocate. If
1543 * __GFP_THISNODE is set, yes, we can always allocate. If zone
1544 * z's node is in our tasks mems_allowed, yes. If it's not a
1545 * __GFP_HARDWALL request and this zone's nodes is in the nearest
1546 * mem_exclusive cpuset ancestor to this tasks cpuset, yes.
1547 * If the task has been OOM killed and has access to memory reserves
1548 * as specified by the TIF_MEMDIE flag, yes.
1551 * If __GFP_HARDWALL is set, cpuset_zone_allowed_softwall()
1552 * reduces to cpuset_zone_allowed_hardwall(). Otherwise,
1553 * cpuset_zone_allowed_softwall() might sleep, and might allow a zone
1554 * from an enclosing cpuset.
1556 * cpuset_zone_allowed_hardwall() only handles the simpler case of
1557 * hardwall cpusets, and never sleeps.
1559 * The __GFP_THISNODE placement logic is really handled elsewhere,
1560 * by forcibly using a zonelist starting at a specified node, and by
1561 * (in get_page_from_freelist()) refusing to consider the zones for
1562 * any node on the zonelist except the first. By the time any such
1563 * calls get to this routine, we should just shut up and say 'yes'.
1565 * GFP_USER allocations are marked with the __GFP_HARDWALL bit,
1566 * and do not allow allocations outside the current tasks cpuset
1567 * unless the task has been OOM killed as is marked TIF_MEMDIE.
1568 * GFP_KERNEL allocations are not so marked, so can escape to the
1569 * nearest enclosing mem_exclusive ancestor cpuset.
1571 * Scanning up parent cpusets requires callback_mutex. The
1572 * __alloc_pages() routine only calls here with __GFP_HARDWALL bit
1573 * _not_ set if it's a GFP_KERNEL allocation, and all nodes in the
1574 * current tasks mems_allowed came up empty on the first pass over
1575 * the zonelist. So only GFP_KERNEL allocations, if all nodes in the
1576 * cpuset are short of memory, might require taking the callback_mutex
1579 * The first call here from mm/page_alloc:get_page_from_freelist()
1580 * has __GFP_HARDWALL set in gfp_mask, enforcing hardwall cpusets,
1581 * so no allocation on a node outside the cpuset is allowed (unless
1582 * in interrupt, of course).
1584 * The second pass through get_page_from_freelist() doesn't even call
1585 * here for GFP_ATOMIC calls. For those calls, the __alloc_pages()
1586 * variable 'wait' is not set, and the bit ALLOC_CPUSET is not set
1587 * in alloc_flags. That logic and the checks below have the combined
1589 * in_interrupt - any node ok (current task context irrelevant)
1590 * GFP_ATOMIC - any node ok
1591 * TIF_MEMDIE - any node ok
1592 * GFP_KERNEL - any node in enclosing mem_exclusive cpuset ok
1593 * GFP_USER - only nodes in current tasks mems allowed ok.
1596 * Don't call cpuset_zone_allowed_softwall if you can't sleep, unless you
1597 * pass in the __GFP_HARDWALL flag set in gfp_flag, which disables
1598 * the code that might scan up ancestor cpusets and sleep.
1601 int __cpuset_zone_allowed_softwall(struct zone
*z
, gfp_t gfp_mask
)
1603 int node
; /* node that zone z is on */
1604 const struct cpuset
*cs
; /* current cpuset ancestors */
1605 int allowed
; /* is allocation in zone z allowed? */
1607 if (in_interrupt() || (gfp_mask
& __GFP_THISNODE
))
1609 node
= zone_to_nid(z
);
1610 might_sleep_if(!(gfp_mask
& __GFP_HARDWALL
));
1611 if (node_isset(node
, current
->mems_allowed
))
1614 * Allow tasks that have access to memory reserves because they have
1615 * been OOM killed to get memory anywhere.
1617 if (unlikely(test_thread_flag(TIF_MEMDIE
)))
1619 if (gfp_mask
& __GFP_HARDWALL
) /* If hardwall request, stop here */
1622 if (current
->flags
& PF_EXITING
) /* Let dying task have memory */
1625 /* Not hardwall and node outside mems_allowed: scan up cpusets */
1626 mutex_lock(&callback_mutex
);
1629 cs
= nearest_exclusive_ancestor(task_cs(current
));
1630 task_unlock(current
);
1632 allowed
= node_isset(node
, cs
->mems_allowed
);
1633 mutex_unlock(&callback_mutex
);
1638 * cpuset_zone_allowed_hardwall - Can we allocate on zone z's memory node?
1639 * @z: is this zone on an allowed node?
1640 * @gfp_mask: memory allocation flags
1642 * If we're in interrupt, yes, we can always allocate.
1643 * If __GFP_THISNODE is set, yes, we can always allocate. If zone
1644 * z's node is in our tasks mems_allowed, yes. If the task has been
1645 * OOM killed and has access to memory reserves as specified by the
1646 * TIF_MEMDIE flag, yes. Otherwise, no.
1648 * The __GFP_THISNODE placement logic is really handled elsewhere,
1649 * by forcibly using a zonelist starting at a specified node, and by
1650 * (in get_page_from_freelist()) refusing to consider the zones for
1651 * any node on the zonelist except the first. By the time any such
1652 * calls get to this routine, we should just shut up and say 'yes'.
1654 * Unlike the cpuset_zone_allowed_softwall() variant, above,
1655 * this variant requires that the zone be in the current tasks
1656 * mems_allowed or that we're in interrupt. It does not scan up the
1657 * cpuset hierarchy for the nearest enclosing mem_exclusive cpuset.
1661 int __cpuset_zone_allowed_hardwall(struct zone
*z
, gfp_t gfp_mask
)
1663 int node
; /* node that zone z is on */
1665 if (in_interrupt() || (gfp_mask
& __GFP_THISNODE
))
1667 node
= zone_to_nid(z
);
1668 if (node_isset(node
, current
->mems_allowed
))
1671 * Allow tasks that have access to memory reserves because they have
1672 * been OOM killed to get memory anywhere.
1674 if (unlikely(test_thread_flag(TIF_MEMDIE
)))
1680 * cpuset_lock - lock out any changes to cpuset structures
1682 * The out of memory (oom) code needs to mutex_lock cpusets
1683 * from being changed while it scans the tasklist looking for a
1684 * task in an overlapping cpuset. Expose callback_mutex via this
1685 * cpuset_lock() routine, so the oom code can lock it, before
1686 * locking the task list. The tasklist_lock is a spinlock, so
1687 * must be taken inside callback_mutex.
1690 void cpuset_lock(void)
1692 mutex_lock(&callback_mutex
);
1696 * cpuset_unlock - release lock on cpuset changes
1698 * Undo the lock taken in a previous cpuset_lock() call.
1701 void cpuset_unlock(void)
1703 mutex_unlock(&callback_mutex
);
1707 * cpuset_mem_spread_node() - On which node to begin search for a page
1709 * If a task is marked PF_SPREAD_PAGE or PF_SPREAD_SLAB (as for
1710 * tasks in a cpuset with is_spread_page or is_spread_slab set),
1711 * and if the memory allocation used cpuset_mem_spread_node()
1712 * to determine on which node to start looking, as it will for
1713 * certain page cache or slab cache pages such as used for file
1714 * system buffers and inode caches, then instead of starting on the
1715 * local node to look for a free page, rather spread the starting
1716 * node around the tasks mems_allowed nodes.
1718 * We don't have to worry about the returned node being offline
1719 * because "it can't happen", and even if it did, it would be ok.
1721 * The routines calling guarantee_online_mems() are careful to
1722 * only set nodes in task->mems_allowed that are online. So it
1723 * should not be possible for the following code to return an
1724 * offline node. But if it did, that would be ok, as this routine
1725 * is not returning the node where the allocation must be, only
1726 * the node where the search should start. The zonelist passed to
1727 * __alloc_pages() will include all nodes. If the slab allocator
1728 * is passed an offline node, it will fall back to the local node.
1729 * See kmem_cache_alloc_node().
1732 int cpuset_mem_spread_node(void)
1736 node
= next_node(current
->cpuset_mem_spread_rotor
, current
->mems_allowed
);
1737 if (node
== MAX_NUMNODES
)
1738 node
= first_node(current
->mems_allowed
);
1739 current
->cpuset_mem_spread_rotor
= node
;
1742 EXPORT_SYMBOL_GPL(cpuset_mem_spread_node
);
1745 * cpuset_mems_allowed_intersects - Does @tsk1's mems_allowed intersect @tsk2's?
1746 * @tsk1: pointer to task_struct of some task.
1747 * @tsk2: pointer to task_struct of some other task.
1749 * Description: Return true if @tsk1's mems_allowed intersects the
1750 * mems_allowed of @tsk2. Used by the OOM killer to determine if
1751 * one of the task's memory usage might impact the memory available
1755 int cpuset_mems_allowed_intersects(const struct task_struct
*tsk1
,
1756 const struct task_struct
*tsk2
)
1758 return nodes_intersects(tsk1
->mems_allowed
, tsk2
->mems_allowed
);
1762 * Collection of memory_pressure is suppressed unless
1763 * this flag is enabled by writing "1" to the special
1764 * cpuset file 'memory_pressure_enabled' in the root cpuset.
1767 int cpuset_memory_pressure_enabled __read_mostly
;
1770 * cpuset_memory_pressure_bump - keep stats of per-cpuset reclaims.
1772 * Keep a running average of the rate of synchronous (direct)
1773 * page reclaim efforts initiated by tasks in each cpuset.
1775 * This represents the rate at which some task in the cpuset
1776 * ran low on memory on all nodes it was allowed to use, and
1777 * had to enter the kernels page reclaim code in an effort to
1778 * create more free memory by tossing clean pages or swapping
1779 * or writing dirty pages.
1781 * Display to user space in the per-cpuset read-only file
1782 * "memory_pressure". Value displayed is an integer
1783 * representing the recent rate of entry into the synchronous
1784 * (direct) page reclaim by any task attached to the cpuset.
1787 void __cpuset_memory_pressure_bump(void)
1790 fmeter_markevent(&task_cs(current
)->fmeter
);
1791 task_unlock(current
);
1794 #ifdef CONFIG_PROC_PID_CPUSET
1796 * proc_cpuset_show()
1797 * - Print tasks cpuset path into seq_file.
1798 * - Used for /proc/<pid>/cpuset.
1799 * - No need to task_lock(tsk) on this tsk->cpuset reference, as it
1800 * doesn't really matter if tsk->cpuset changes after we read it,
1801 * and we take manage_mutex, keeping attach_task() from changing it
1802 * anyway. No need to check that tsk->cpuset != NULL, thanks to
1803 * the_top_cpuset_hack in cpuset_exit(), which sets an exiting tasks
1804 * cpuset to top_cpuset.
1806 static int proc_cpuset_show(struct seq_file
*m
, void *v
)
1809 struct task_struct
*tsk
;
1811 struct cgroup_subsys_state
*css
;
1815 buf
= kmalloc(PAGE_SIZE
, GFP_KERNEL
);
1821 tsk
= get_pid_task(pid
, PIDTYPE_PID
);
1827 css
= task_subsys_state(tsk
, cpuset_subsys_id
);
1828 retval
= cgroup_path(css
->cgroup
, buf
, PAGE_SIZE
);
1835 put_task_struct(tsk
);
1842 static int cpuset_open(struct inode
*inode
, struct file
*file
)
1844 struct pid
*pid
= PROC_I(inode
)->pid
;
1845 return single_open(file
, proc_cpuset_show
, pid
);
1848 const struct file_operations proc_cpuset_operations
= {
1849 .open
= cpuset_open
,
1851 .llseek
= seq_lseek
,
1852 .release
= single_release
,
1854 #endif /* CONFIG_PROC_PID_CPUSET */
1856 /* Display task cpus_allowed, mems_allowed in /proc/<pid>/status file. */
1857 char *cpuset_task_status_allowed(struct task_struct
*task
, char *buffer
)
1859 buffer
+= sprintf(buffer
, "Cpus_allowed:\t");
1860 buffer
+= cpumask_scnprintf(buffer
, PAGE_SIZE
, task
->cpus_allowed
);
1861 buffer
+= sprintf(buffer
, "\n");
1862 buffer
+= sprintf(buffer
, "Mems_allowed:\t");
1863 buffer
+= nodemask_scnprintf(buffer
, PAGE_SIZE
, task
->mems_allowed
);
1864 buffer
+= sprintf(buffer
, "\n");