4 * Processor and Memory placement constraints for sets of tasks.
6 * Copyright (C) 2003 BULL SA.
7 * Copyright (C) 2004-2006 Silicon Graphics, Inc.
9 * Portions derived from Patrick Mochel's sysfs code.
10 * sysfs is Copyright (c) 2001-3 Patrick Mochel
12 * 2003-10-10 Written by Simon Derr.
13 * 2003-10-22 Updates by Stephen Hemminger.
14 * 2004 May-July Rework by Paul Jackson.
16 * This file is subject to the terms and conditions of the GNU General Public
17 * License. See the file COPYING in the main directory of the Linux
18 * distribution for more details.
21 #include <linux/config.h>
22 #include <linux/cpu.h>
23 #include <linux/cpumask.h>
24 #include <linux/cpuset.h>
25 #include <linux/err.h>
26 #include <linux/errno.h>
27 #include <linux/file.h>
29 #include <linux/init.h>
30 #include <linux/interrupt.h>
31 #include <linux/kernel.h>
32 #include <linux/kmod.h>
33 #include <linux/list.h>
34 #include <linux/mempolicy.h>
36 #include <linux/module.h>
37 #include <linux/mount.h>
38 #include <linux/namei.h>
39 #include <linux/pagemap.h>
40 #include <linux/proc_fs.h>
41 #include <linux/rcupdate.h>
42 #include <linux/sched.h>
43 #include <linux/seq_file.h>
44 #include <linux/security.h>
45 #include <linux/slab.h>
46 #include <linux/smp_lock.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>
53 #include <linux/task_ref.h>
55 #include <asm/uaccess.h>
56 #include <asm/atomic.h>
57 #include <linux/mutex.h>
59 #define CPUSET_SUPER_MAGIC 0x27e0eb
62 * Tracks how many cpusets are currently defined in system.
63 * When there is only one cpuset (the root cpuset) we can
64 * short circuit some hooks.
66 int number_of_cpusets __read_mostly
;
68 /* See "Frequency meter" comments, below. */
71 int cnt
; /* unprocessed events count */
72 int val
; /* most recent output value */
73 time_t time
; /* clock (secs) when val computed */
74 spinlock_t lock
; /* guards read or write of above */
78 unsigned long flags
; /* "unsigned long" so bitops work */
79 cpumask_t cpus_allowed
; /* CPUs allowed to tasks in cpuset */
80 nodemask_t mems_allowed
; /* Memory Nodes allowed to tasks */
83 * Count is atomic so can incr (fork) or decr (exit) without a lock.
85 atomic_t count
; /* count tasks using this cpuset */
88 * We link our 'sibling' struct into our parents 'children'.
89 * Our children link their 'sibling' into our 'children'.
91 struct list_head sibling
; /* my parents children */
92 struct list_head children
; /* my children */
94 struct cpuset
*parent
; /* my parent */
95 struct dentry
*dentry
; /* cpuset fs entry */
98 * Copy of global cpuset_mems_generation as of the most
99 * recent time this cpuset changed its mems_allowed.
103 struct fmeter fmeter
; /* memory_pressure filter */
106 /* bits in struct cpuset flags field */
112 CS_NOTIFY_ON_RELEASE
,
117 /* convenient tests for these bits */
118 static inline int is_cpu_exclusive(const struct cpuset
*cs
)
120 return test_bit(CS_CPU_EXCLUSIVE
, &cs
->flags
);
123 static inline int is_mem_exclusive(const struct cpuset
*cs
)
125 return test_bit(CS_MEM_EXCLUSIVE
, &cs
->flags
);
128 static inline int is_removed(const struct cpuset
*cs
)
130 return test_bit(CS_REMOVED
, &cs
->flags
);
133 static inline int notify_on_release(const struct cpuset
*cs
)
135 return test_bit(CS_NOTIFY_ON_RELEASE
, &cs
->flags
);
138 static inline int is_memory_migrate(const struct cpuset
*cs
)
140 return test_bit(CS_MEMORY_MIGRATE
, &cs
->flags
);
143 static inline int is_spread_page(const struct cpuset
*cs
)
145 return test_bit(CS_SPREAD_PAGE
, &cs
->flags
);
148 static inline int is_spread_slab(const struct cpuset
*cs
)
150 return test_bit(CS_SPREAD_SLAB
, &cs
->flags
);
154 * Increment this integer everytime any cpuset changes its
155 * mems_allowed value. Users of cpusets can track this generation
156 * number, and avoid having to lock and reload mems_allowed unless
157 * the cpuset they're using changes generation.
159 * A single, global generation is needed because attach_task() could
160 * reattach a task to a different cpuset, which must not have its
161 * generation numbers aliased with those of that tasks previous cpuset.
163 * Generations are needed for mems_allowed because one task cannot
164 * modify anothers memory placement. So we must enable every task,
165 * on every visit to __alloc_pages(), to efficiently check whether
166 * its current->cpuset->mems_allowed has changed, requiring an update
167 * of its current->mems_allowed.
169 * Since cpuset_mems_generation is guarded by manage_mutex,
170 * there is no need to mark it atomic.
172 static int cpuset_mems_generation
;
174 static struct cpuset top_cpuset
= {
175 .flags
= ((1 << CS_CPU_EXCLUSIVE
) | (1 << CS_MEM_EXCLUSIVE
)),
176 .cpus_allowed
= CPU_MASK_ALL
,
177 .mems_allowed
= NODE_MASK_ALL
,
178 .count
= ATOMIC_INIT(0),
179 .sibling
= LIST_HEAD_INIT(top_cpuset
.sibling
),
180 .children
= LIST_HEAD_INIT(top_cpuset
.children
),
183 static struct vfsmount
*cpuset_mount
;
184 static struct super_block
*cpuset_sb
;
187 * We have two global cpuset mutexes below. They can nest.
188 * It is ok to first take manage_mutex, then nest callback_mutex. We also
189 * require taking task_lock() when dereferencing a tasks cpuset pointer.
190 * See "The task_lock() exception", at the end of this comment.
192 * A task must hold both mutexes to modify cpusets. If a task
193 * holds manage_mutex, then it blocks others wanting that mutex,
194 * ensuring that it is the only task able to also acquire callback_mutex
195 * and be able to modify cpusets. It can perform various checks on
196 * the cpuset structure first, knowing nothing will change. It can
197 * also allocate memory while just holding manage_mutex. While it is
198 * performing these checks, various callback routines can briefly
199 * acquire callback_mutex to query cpusets. Once it is ready to make
200 * the changes, it takes callback_mutex, blocking everyone else.
202 * Calls to the kernel memory allocator can not be made while holding
203 * callback_mutex, as that would risk double tripping on callback_mutex
204 * from one of the callbacks into the cpuset code from within
207 * If a task is only holding callback_mutex, then it has read-only
210 * The task_struct fields mems_allowed and mems_generation may only
211 * be accessed in the context of that task, so require no locks.
213 * Any task can increment and decrement the count field without lock.
214 * So in general, code holding manage_mutex or callback_mutex can't rely
215 * on the count field not changing. However, if the count goes to
216 * zero, then only attach_task(), which holds both mutexes, can
217 * increment it again. Because a count of zero means that no tasks
218 * are currently attached, therefore there is no way a task attached
219 * to that cpuset can fork (the other way to increment the count).
220 * So code holding manage_mutex or callback_mutex can safely assume that
221 * if the count is zero, it will stay zero. Similarly, if a task
222 * holds manage_mutex or callback_mutex on a cpuset with zero count, it
223 * knows that the cpuset won't be removed, as cpuset_rmdir() needs
224 * both of those mutexes.
226 * The cpuset_common_file_write handler for operations that modify
227 * the cpuset hierarchy holds manage_mutex across the entire operation,
228 * single threading all such cpuset modifications across the system.
230 * The cpuset_common_file_read() handlers only hold callback_mutex across
231 * small pieces of code, such as when reading out possibly multi-word
232 * cpumasks and nodemasks.
234 * The fork and exit callbacks cpuset_fork() and cpuset_exit(), don't
235 * (usually) take either mutex. These are the two most performance
236 * critical pieces of code here. The exception occurs on cpuset_exit(),
237 * when a task in a notify_on_release cpuset exits. Then manage_mutex
238 * is taken, and if the cpuset count is zero, a usermode call made
239 * to /sbin/cpuset_release_agent with the name of the cpuset (path
240 * relative to the root of cpuset file system) as the argument.
242 * A cpuset can only be deleted if both its 'count' of using tasks
243 * is zero, and its list of 'children' cpusets is empty. Since all
244 * tasks in the system use _some_ cpuset, and since there is always at
245 * least one task in the system (init, pid == 1), therefore, top_cpuset
246 * always has either children cpusets and/or using tasks. So we don't
247 * need a special hack to ensure that top_cpuset cannot be deleted.
249 * The above "Tale of Two Semaphores" would be complete, but for:
251 * The task_lock() exception
253 * The need for this exception arises from the action of attach_task(),
254 * which overwrites one tasks cpuset pointer with another. It does
255 * so using both mutexes, however there are several performance
256 * critical places that need to reference task->cpuset without the
257 * expense of grabbing a system global mutex. Therefore except as
258 * noted below, when dereferencing or, as in attach_task(), modifying
259 * a tasks cpuset pointer we use task_lock(), which acts on a spinlock
260 * (task->alloc_lock) already in the task_struct routinely used for
263 * P.S. One more locking exception. RCU is used to guard the
264 * update of a tasks cpuset pointer by attach_task() and the
265 * access of task->cpuset->mems_generation via that pointer in
266 * the routine cpuset_update_task_memory_state().
269 static DEFINE_MUTEX(manage_mutex
);
270 static DEFINE_MUTEX(callback_mutex
);
273 * A couple of forward declarations required, due to cyclic reference loop:
274 * cpuset_mkdir -> cpuset_create -> cpuset_populate_dir -> cpuset_add_file
275 * -> cpuset_create_file -> cpuset_dir_inode_operations -> cpuset_mkdir.
278 static int cpuset_mkdir(struct inode
*dir
, struct dentry
*dentry
, int mode
);
279 static int cpuset_rmdir(struct inode
*unused_dir
, struct dentry
*dentry
);
281 static struct backing_dev_info cpuset_backing_dev_info
= {
282 .ra_pages
= 0, /* No readahead */
283 .capabilities
= BDI_CAP_NO_ACCT_DIRTY
| BDI_CAP_NO_WRITEBACK
,
286 static struct inode
*cpuset_new_inode(mode_t mode
)
288 struct inode
*inode
= new_inode(cpuset_sb
);
291 inode
->i_mode
= mode
;
292 inode
->i_uid
= current
->fsuid
;
293 inode
->i_gid
= current
->fsgid
;
294 inode
->i_blksize
= PAGE_CACHE_SIZE
;
296 inode
->i_atime
= inode
->i_mtime
= inode
->i_ctime
= CURRENT_TIME
;
297 inode
->i_mapping
->backing_dev_info
= &cpuset_backing_dev_info
;
302 static void cpuset_diput(struct dentry
*dentry
, struct inode
*inode
)
304 /* is dentry a directory ? if so, kfree() associated cpuset */
305 if (S_ISDIR(inode
->i_mode
)) {
306 struct cpuset
*cs
= dentry
->d_fsdata
;
307 BUG_ON(!(is_removed(cs
)));
313 static struct dentry_operations cpuset_dops
= {
314 .d_iput
= cpuset_diput
,
317 static struct dentry
*cpuset_get_dentry(struct dentry
*parent
, const char *name
)
319 struct dentry
*d
= lookup_one_len(name
, parent
, strlen(name
));
321 d
->d_op
= &cpuset_dops
;
325 static void remove_dir(struct dentry
*d
)
327 struct dentry
*parent
= dget(d
->d_parent
);
330 simple_rmdir(parent
->d_inode
, d
);
335 * NOTE : the dentry must have been dget()'ed
337 static void cpuset_d_remove_dir(struct dentry
*dentry
)
339 struct list_head
*node
;
341 spin_lock(&dcache_lock
);
342 node
= dentry
->d_subdirs
.next
;
343 while (node
!= &dentry
->d_subdirs
) {
344 struct dentry
*d
= list_entry(node
, struct dentry
, d_u
.d_child
);
348 spin_unlock(&dcache_lock
);
350 simple_unlink(dentry
->d_inode
, d
);
352 spin_lock(&dcache_lock
);
354 node
= dentry
->d_subdirs
.next
;
356 list_del_init(&dentry
->d_u
.d_child
);
357 spin_unlock(&dcache_lock
);
361 static struct super_operations cpuset_ops
= {
362 .statfs
= simple_statfs
,
363 .drop_inode
= generic_delete_inode
,
366 static int cpuset_fill_super(struct super_block
*sb
, void *unused_data
,
372 sb
->s_blocksize
= PAGE_CACHE_SIZE
;
373 sb
->s_blocksize_bits
= PAGE_CACHE_SHIFT
;
374 sb
->s_magic
= CPUSET_SUPER_MAGIC
;
375 sb
->s_op
= &cpuset_ops
;
378 inode
= cpuset_new_inode(S_IFDIR
| S_IRUGO
| S_IXUGO
| S_IWUSR
);
380 inode
->i_op
= &simple_dir_inode_operations
;
381 inode
->i_fop
= &simple_dir_operations
;
382 /* directories start off with i_nlink == 2 (for "." entry) */
388 root
= d_alloc_root(inode
);
397 static int cpuset_get_sb(struct file_system_type
*fs_type
,
398 int flags
, const char *unused_dev_name
,
399 void *data
, struct vfsmount
*mnt
)
401 return get_sb_single(fs_type
, flags
, data
, cpuset_fill_super
, mnt
);
404 static struct file_system_type cpuset_fs_type
= {
406 .get_sb
= cpuset_get_sb
,
407 .kill_sb
= kill_litter_super
,
412 * The files in the cpuset filesystem mostly have a very simple read/write
413 * handling, some common function will take care of it. Nevertheless some cases
414 * (read tasks) are special and therefore I define this structure for every
418 * When reading/writing to a file:
419 * - the cpuset to use in file->f_dentry->d_parent->d_fsdata
420 * - the 'cftype' of the file is file->f_dentry->d_fsdata
426 int (*open
) (struct inode
*inode
, struct file
*file
);
427 ssize_t (*read
) (struct file
*file
, char __user
*buf
, size_t nbytes
,
429 int (*write
) (struct file
*file
, const char __user
*buf
, size_t nbytes
,
431 int (*release
) (struct inode
*inode
, struct file
*file
);
434 static inline struct cpuset
*__d_cs(struct dentry
*dentry
)
436 return dentry
->d_fsdata
;
439 static inline struct cftype
*__d_cft(struct dentry
*dentry
)
441 return dentry
->d_fsdata
;
445 * Call with manage_mutex held. Writes path of cpuset into buf.
446 * Returns 0 on success, -errno on error.
449 static int cpuset_path(const struct cpuset
*cs
, char *buf
, int buflen
)
453 start
= buf
+ buflen
;
457 int len
= cs
->dentry
->d_name
.len
;
458 if ((start
-= len
) < buf
)
459 return -ENAMETOOLONG
;
460 memcpy(start
, cs
->dentry
->d_name
.name
, len
);
467 return -ENAMETOOLONG
;
470 memmove(buf
, start
, buf
+ buflen
- start
);
475 * Notify userspace when a cpuset is released, by running
476 * /sbin/cpuset_release_agent with the name of the cpuset (path
477 * relative to the root of cpuset file system) as the argument.
479 * Most likely, this user command will try to rmdir this cpuset.
481 * This races with the possibility that some other task will be
482 * attached to this cpuset before it is removed, or that some other
483 * user task will 'mkdir' a child cpuset of this cpuset. That's ok.
484 * The presumed 'rmdir' will fail quietly if this cpuset is no longer
485 * unused, and this cpuset will be reprieved from its death sentence,
486 * to continue to serve a useful existence. Next time it's released,
487 * we will get notified again, if it still has 'notify_on_release' set.
489 * The final arg to call_usermodehelper() is 0, which means don't
490 * wait. The separate /sbin/cpuset_release_agent task is forked by
491 * call_usermodehelper(), then control in this thread returns here,
492 * without waiting for the release agent task. We don't bother to
493 * wait because the caller of this routine has no use for the exit
494 * status of the /sbin/cpuset_release_agent task, so no sense holding
495 * our caller up for that.
497 * When we had only one cpuset mutex, we had to call this
498 * without holding it, to avoid deadlock when call_usermodehelper()
499 * allocated memory. With two locks, we could now call this while
500 * holding manage_mutex, but we still don't, so as to minimize
501 * the time manage_mutex is held.
504 static void cpuset_release_agent(const char *pathbuf
)
506 char *argv
[3], *envp
[3];
513 argv
[i
++] = "/sbin/cpuset_release_agent";
514 argv
[i
++] = (char *)pathbuf
;
518 /* minimal command environment */
519 envp
[i
++] = "HOME=/";
520 envp
[i
++] = "PATH=/sbin:/bin:/usr/sbin:/usr/bin";
523 call_usermodehelper(argv
[0], argv
, envp
, 0);
528 * Either cs->count of using tasks transitioned to zero, or the
529 * cs->children list of child cpusets just became empty. If this
530 * cs is notify_on_release() and now both the user count is zero and
531 * the list of children is empty, prepare cpuset path in a kmalloc'd
532 * buffer, to be returned via ppathbuf, so that the caller can invoke
533 * cpuset_release_agent() with it later on, once manage_mutex is dropped.
534 * Call here with manage_mutex held.
536 * This check_for_release() routine is responsible for kmalloc'ing
537 * pathbuf. The above cpuset_release_agent() is responsible for
538 * kfree'ing pathbuf. The caller of these routines is responsible
539 * for providing a pathbuf pointer, initialized to NULL, then
540 * calling check_for_release() with manage_mutex held and the address
541 * of the pathbuf pointer, then dropping manage_mutex, then calling
542 * cpuset_release_agent() with pathbuf, as set by check_for_release().
545 static void check_for_release(struct cpuset
*cs
, char **ppathbuf
)
547 if (notify_on_release(cs
) && atomic_read(&cs
->count
) == 0 &&
548 list_empty(&cs
->children
)) {
551 buf
= kmalloc(PAGE_SIZE
, GFP_KERNEL
);
554 if (cpuset_path(cs
, buf
, PAGE_SIZE
) < 0)
562 * Return in *pmask the portion of a cpusets's cpus_allowed that
563 * are online. If none are online, walk up the cpuset hierarchy
564 * until we find one that does have some online cpus. If we get
565 * all the way to the top and still haven't found any online cpus,
566 * return cpu_online_map. Or if passed a NULL cs from an exit'ing
567 * task, return cpu_online_map.
569 * One way or another, we guarantee to return some non-empty subset
572 * Call with callback_mutex held.
575 static void guarantee_online_cpus(const struct cpuset
*cs
, cpumask_t
*pmask
)
577 while (cs
&& !cpus_intersects(cs
->cpus_allowed
, cpu_online_map
))
580 cpus_and(*pmask
, cs
->cpus_allowed
, cpu_online_map
);
582 *pmask
= cpu_online_map
;
583 BUG_ON(!cpus_intersects(*pmask
, cpu_online_map
));
587 * Return in *pmask the portion of a cpusets's mems_allowed that
588 * are online. If none are online, walk up the cpuset hierarchy
589 * until we find one that does have some online mems. If we get
590 * all the way to the top and still haven't found any online mems,
591 * return node_online_map.
593 * One way or another, we guarantee to return some non-empty subset
594 * of node_online_map.
596 * Call with callback_mutex held.
599 static void guarantee_online_mems(const struct cpuset
*cs
, nodemask_t
*pmask
)
601 while (cs
&& !nodes_intersects(cs
->mems_allowed
, node_online_map
))
604 nodes_and(*pmask
, cs
->mems_allowed
, node_online_map
);
606 *pmask
= node_online_map
;
607 BUG_ON(!nodes_intersects(*pmask
, node_online_map
));
611 * cpuset_update_task_memory_state - update task memory placement
613 * If the current tasks cpusets mems_allowed changed behind our
614 * backs, update current->mems_allowed, mems_generation and task NUMA
615 * mempolicy to the new value.
617 * Task mempolicy is updated by rebinding it relative to the
618 * current->cpuset if a task has its memory placement changed.
619 * Do not call this routine if in_interrupt().
621 * Call without callback_mutex or task_lock() held. May be
622 * called with or without manage_mutex held. Thanks in part to
623 * 'the_top_cpuset_hack', the tasks cpuset pointer will never
624 * be NULL. This routine also might acquire callback_mutex and
625 * current->mm->mmap_sem during call.
627 * Reading current->cpuset->mems_generation doesn't need task_lock
628 * to guard the current->cpuset derefence, because it is guarded
629 * from concurrent freeing of current->cpuset by attach_task(),
632 * The rcu_dereference() is technically probably not needed,
633 * as I don't actually mind if I see a new cpuset pointer but
634 * an old value of mems_generation. However this really only
635 * matters on alpha systems using cpusets heavily. If I dropped
636 * that rcu_dereference(), it would save them a memory barrier.
637 * For all other arch's, rcu_dereference is a no-op anyway, and for
638 * alpha systems not using cpusets, another planned optimization,
639 * avoiding the rcu critical section for tasks in the root cpuset
640 * which is statically allocated, so can't vanish, will make this
641 * irrelevant. Better to use RCU as intended, than to engage in
642 * some cute trick to save a memory barrier that is impossible to
643 * test, for alpha systems using cpusets heavily, which might not
646 * This routine is needed to update the per-task mems_allowed data,
647 * within the tasks context, when it is trying to allocate memory
648 * (in various mm/mempolicy.c routines) and notices that some other
649 * task has been modifying its cpuset.
652 void cpuset_update_task_memory_state(void)
654 int my_cpusets_mem_gen
;
655 struct task_struct
*tsk
= current
;
658 if (tsk
->cpuset
== &top_cpuset
) {
659 /* Don't need rcu for top_cpuset. It's never freed. */
660 my_cpusets_mem_gen
= top_cpuset
.mems_generation
;
663 cs
= rcu_dereference(tsk
->cpuset
);
664 my_cpusets_mem_gen
= cs
->mems_generation
;
668 if (my_cpusets_mem_gen
!= tsk
->cpuset_mems_generation
) {
669 mutex_lock(&callback_mutex
);
671 cs
= tsk
->cpuset
; /* Maybe changed when task not locked */
672 guarantee_online_mems(cs
, &tsk
->mems_allowed
);
673 tsk
->cpuset_mems_generation
= cs
->mems_generation
;
674 if (is_spread_page(cs
))
675 tsk
->flags
|= PF_SPREAD_PAGE
;
677 tsk
->flags
&= ~PF_SPREAD_PAGE
;
678 if (is_spread_slab(cs
))
679 tsk
->flags
|= PF_SPREAD_SLAB
;
681 tsk
->flags
&= ~PF_SPREAD_SLAB
;
683 mutex_unlock(&callback_mutex
);
684 mpol_rebind_task(tsk
, &tsk
->mems_allowed
);
689 * is_cpuset_subset(p, q) - Is cpuset p a subset of cpuset q?
691 * One cpuset is a subset of another if all its allowed CPUs and
692 * Memory Nodes are a subset of the other, and its exclusive flags
693 * are only set if the other's are set. Call holding manage_mutex.
696 static int is_cpuset_subset(const struct cpuset
*p
, const struct cpuset
*q
)
698 return cpus_subset(p
->cpus_allowed
, q
->cpus_allowed
) &&
699 nodes_subset(p
->mems_allowed
, q
->mems_allowed
) &&
700 is_cpu_exclusive(p
) <= is_cpu_exclusive(q
) &&
701 is_mem_exclusive(p
) <= is_mem_exclusive(q
);
705 * validate_change() - Used to validate that any proposed cpuset change
706 * follows the structural rules for cpusets.
708 * If we replaced the flag and mask values of the current cpuset
709 * (cur) with those values in the trial cpuset (trial), would
710 * our various subset and exclusive rules still be valid? Presumes
713 * 'cur' is the address of an actual, in-use cpuset. Operations
714 * such as list traversal that depend on the actual address of the
715 * cpuset in the list must use cur below, not trial.
717 * 'trial' is the address of bulk structure copy of cur, with
718 * perhaps one or more of the fields cpus_allowed, mems_allowed,
719 * or flags changed to new, trial values.
721 * Return 0 if valid, -errno if not.
724 static int validate_change(const struct cpuset
*cur
, const struct cpuset
*trial
)
726 struct cpuset
*c
, *par
;
728 /* Each of our child cpusets must be a subset of us */
729 list_for_each_entry(c
, &cur
->children
, sibling
) {
730 if (!is_cpuset_subset(c
, trial
))
734 /* Remaining checks don't apply to root cpuset */
735 if ((par
= cur
->parent
) == NULL
)
738 /* We must be a subset of our parent cpuset */
739 if (!is_cpuset_subset(trial
, par
))
742 /* If either I or some sibling (!= me) is exclusive, we can't overlap */
743 list_for_each_entry(c
, &par
->children
, sibling
) {
744 if ((is_cpu_exclusive(trial
) || is_cpu_exclusive(c
)) &&
746 cpus_intersects(trial
->cpus_allowed
, c
->cpus_allowed
))
748 if ((is_mem_exclusive(trial
) || is_mem_exclusive(c
)) &&
750 nodes_intersects(trial
->mems_allowed
, c
->mems_allowed
))
758 * For a given cpuset cur, partition the system as follows
759 * a. All cpus in the parent cpuset's cpus_allowed that are not part of any
760 * exclusive child cpusets
761 * b. All cpus in the current cpuset's cpus_allowed that are not part of any
762 * exclusive child cpusets
763 * Build these two partitions by calling partition_sched_domains
765 * Call with manage_mutex held. May nest a call to the
766 * lock_cpu_hotplug()/unlock_cpu_hotplug() pair.
769 static void update_cpu_domains(struct cpuset
*cur
)
771 struct cpuset
*c
, *par
= cur
->parent
;
772 cpumask_t pspan
, cspan
;
774 if (par
== NULL
|| cpus_empty(cur
->cpus_allowed
))
778 * Get all cpus from parent's cpus_allowed not part of exclusive
781 pspan
= par
->cpus_allowed
;
782 list_for_each_entry(c
, &par
->children
, sibling
) {
783 if (is_cpu_exclusive(c
))
784 cpus_andnot(pspan
, pspan
, c
->cpus_allowed
);
786 if (is_removed(cur
) || !is_cpu_exclusive(cur
)) {
787 cpus_or(pspan
, pspan
, cur
->cpus_allowed
);
788 if (cpus_equal(pspan
, cur
->cpus_allowed
))
790 cspan
= CPU_MASK_NONE
;
792 if (cpus_empty(pspan
))
794 cspan
= cur
->cpus_allowed
;
796 * Get all cpus from current cpuset's cpus_allowed not part
797 * of exclusive children
799 list_for_each_entry(c
, &cur
->children
, sibling
) {
800 if (is_cpu_exclusive(c
))
801 cpus_andnot(cspan
, cspan
, c
->cpus_allowed
);
806 partition_sched_domains(&pspan
, &cspan
);
807 unlock_cpu_hotplug();
811 * Call with manage_mutex held. May take callback_mutex during call.
814 static int update_cpumask(struct cpuset
*cs
, char *buf
)
816 struct cpuset trialcs
;
817 int retval
, cpus_unchanged
;
820 retval
= cpulist_parse(buf
, trialcs
.cpus_allowed
);
823 cpus_and(trialcs
.cpus_allowed
, trialcs
.cpus_allowed
, cpu_online_map
);
824 if (cpus_empty(trialcs
.cpus_allowed
))
826 retval
= validate_change(cs
, &trialcs
);
829 cpus_unchanged
= cpus_equal(cs
->cpus_allowed
, trialcs
.cpus_allowed
);
830 mutex_lock(&callback_mutex
);
831 cs
->cpus_allowed
= trialcs
.cpus_allowed
;
832 mutex_unlock(&callback_mutex
);
833 if (is_cpu_exclusive(cs
) && !cpus_unchanged
)
834 update_cpu_domains(cs
);
841 * Migrate memory region from one set of nodes to another.
843 * Temporarilly set tasks mems_allowed to target nodes of migration,
844 * so that the migration code can allocate pages on these nodes.
846 * Call holding manage_mutex, so our current->cpuset won't change
847 * during this call, as manage_mutex holds off any attach_task()
848 * calls. Therefore we don't need to take task_lock around the
849 * call to guarantee_online_mems(), as we know no one is changing
852 * Hold callback_mutex around the two modifications of our tasks
853 * mems_allowed to synchronize with cpuset_mems_allowed().
855 * While the mm_struct we are migrating is typically from some
856 * other task, the task_struct mems_allowed that we are hacking
857 * is for our current task, which must allocate new pages for that
858 * migrating memory region.
860 * We call cpuset_update_task_memory_state() before hacking
861 * our tasks mems_allowed, so that we are assured of being in
862 * sync with our tasks cpuset, and in particular, callbacks to
863 * cpuset_update_task_memory_state() from nested page allocations
864 * won't see any mismatch of our cpuset and task mems_generation
865 * values, so won't overwrite our hacked tasks mems_allowed
869 static void cpuset_migrate_mm(struct mm_struct
*mm
, const nodemask_t
*from
,
870 const nodemask_t
*to
)
872 struct task_struct
*tsk
= current
;
874 cpuset_update_task_memory_state();
876 mutex_lock(&callback_mutex
);
877 tsk
->mems_allowed
= *to
;
878 mutex_unlock(&callback_mutex
);
880 do_migrate_pages(mm
, from
, to
, MPOL_MF_MOVE_ALL
);
882 mutex_lock(&callback_mutex
);
883 guarantee_online_mems(tsk
->cpuset
, &tsk
->mems_allowed
);
884 mutex_unlock(&callback_mutex
);
888 * Handle user request to change the 'mems' memory placement
889 * of a cpuset. Needs to validate the request, update the
890 * cpusets mems_allowed and mems_generation, and for each
891 * task in the cpuset, rebind any vma mempolicies and if
892 * the cpuset is marked 'memory_migrate', migrate the tasks
893 * pages to the new memory.
895 * Call with manage_mutex held. May take callback_mutex during call.
896 * Will take tasklist_lock, scan tasklist for tasks in cpuset cs,
897 * lock each such tasks mm->mmap_sem, scan its vma's and rebind
898 * their mempolicies to the cpusets new mems_allowed.
901 static int update_nodemask(struct cpuset
*cs
, char *buf
)
903 struct cpuset trialcs
;
905 struct task_struct
*g
, *p
;
906 struct mm_struct
**mmarray
;
913 retval
= nodelist_parse(buf
, trialcs
.mems_allowed
);
916 nodes_and(trialcs
.mems_allowed
, trialcs
.mems_allowed
, node_online_map
);
917 oldmem
= cs
->mems_allowed
;
918 if (nodes_equal(oldmem
, trialcs
.mems_allowed
)) {
919 retval
= 0; /* Too easy - nothing to do */
922 if (nodes_empty(trialcs
.mems_allowed
)) {
926 retval
= validate_change(cs
, &trialcs
);
930 mutex_lock(&callback_mutex
);
931 cs
->mems_allowed
= trialcs
.mems_allowed
;
932 cs
->mems_generation
= cpuset_mems_generation
++;
933 mutex_unlock(&callback_mutex
);
935 set_cpuset_being_rebound(cs
); /* causes mpol_copy() rebind */
937 fudge
= 10; /* spare mmarray[] slots */
938 fudge
+= cpus_weight(cs
->cpus_allowed
); /* imagine one fork-bomb/cpu */
942 * Allocate mmarray[] to hold mm reference for each task
943 * in cpuset cs. Can't kmalloc GFP_KERNEL while holding
944 * tasklist_lock. We could use GFP_ATOMIC, but with a
945 * few more lines of code, we can retry until we get a big
946 * enough mmarray[] w/o using GFP_ATOMIC.
949 ntasks
= atomic_read(&cs
->count
); /* guess */
951 mmarray
= kmalloc(ntasks
* sizeof(*mmarray
), GFP_KERNEL
);
954 write_lock_irq(&tasklist_lock
); /* block fork */
955 if (atomic_read(&cs
->count
) <= ntasks
)
956 break; /* got enough */
957 write_unlock_irq(&tasklist_lock
); /* try again */
963 /* Load up mmarray[] with mm reference for each task in cpuset. */
964 do_each_thread(g
, p
) {
965 struct mm_struct
*mm
;
969 "Cpuset mempolicy rebind incomplete.\n");
978 } while_each_thread(g
, p
);
979 write_unlock_irq(&tasklist_lock
);
982 * Now that we've dropped the tasklist spinlock, we can
983 * rebind the vma mempolicies of each mm in mmarray[] to their
984 * new cpuset, and release that mm. The mpol_rebind_mm()
985 * call takes mmap_sem, which we couldn't take while holding
986 * tasklist_lock. Forks can happen again now - the mpol_copy()
987 * cpuset_being_rebound check will catch such forks, and rebind
988 * their vma mempolicies too. Because we still hold the global
989 * cpuset manage_mutex, we know that no other rebind effort will
990 * be contending for the global variable cpuset_being_rebound.
991 * It's ok if we rebind the same mm twice; mpol_rebind_mm()
992 * is idempotent. Also migrate pages in each mm to new nodes.
994 migrate
= is_memory_migrate(cs
);
995 for (i
= 0; i
< n
; i
++) {
996 struct mm_struct
*mm
= mmarray
[i
];
998 mpol_rebind_mm(mm
, &cs
->mems_allowed
);
1000 cpuset_migrate_mm(mm
, &oldmem
, &cs
->mems_allowed
);
1004 /* We're done rebinding vma's to this cpusets new mems_allowed. */
1006 set_cpuset_being_rebound(NULL
);
1013 * Call with manage_mutex held.
1016 static int update_memory_pressure_enabled(struct cpuset
*cs
, char *buf
)
1018 if (simple_strtoul(buf
, NULL
, 10) != 0)
1019 cpuset_memory_pressure_enabled
= 1;
1021 cpuset_memory_pressure_enabled
= 0;
1026 * update_flag - read a 0 or a 1 in a file and update associated flag
1027 * bit: the bit to update (CS_CPU_EXCLUSIVE, CS_MEM_EXCLUSIVE,
1028 * CS_NOTIFY_ON_RELEASE, CS_MEMORY_MIGRATE,
1029 * CS_SPREAD_PAGE, CS_SPREAD_SLAB)
1030 * cs: the cpuset to update
1031 * buf: the buffer where we read the 0 or 1
1033 * Call with manage_mutex held.
1036 static int update_flag(cpuset_flagbits_t bit
, struct cpuset
*cs
, char *buf
)
1039 struct cpuset trialcs
;
1040 int err
, cpu_exclusive_changed
;
1042 turning_on
= (simple_strtoul(buf
, NULL
, 10) != 0);
1046 set_bit(bit
, &trialcs
.flags
);
1048 clear_bit(bit
, &trialcs
.flags
);
1050 err
= validate_change(cs
, &trialcs
);
1053 cpu_exclusive_changed
=
1054 (is_cpu_exclusive(cs
) != is_cpu_exclusive(&trialcs
));
1055 mutex_lock(&callback_mutex
);
1057 set_bit(bit
, &cs
->flags
);
1059 clear_bit(bit
, &cs
->flags
);
1060 mutex_unlock(&callback_mutex
);
1062 if (cpu_exclusive_changed
)
1063 update_cpu_domains(cs
);
1068 * Frequency meter - How fast is some event occuring?
1070 * These routines manage a digitally filtered, constant time based,
1071 * event frequency meter. There are four routines:
1072 * fmeter_init() - initialize a frequency meter.
1073 * fmeter_markevent() - called each time the event happens.
1074 * fmeter_getrate() - returns the recent rate of such events.
1075 * fmeter_update() - internal routine used to update fmeter.
1077 * A common data structure is passed to each of these routines,
1078 * which is used to keep track of the state required to manage the
1079 * frequency meter and its digital filter.
1081 * The filter works on the number of events marked per unit time.
1082 * The filter is single-pole low-pass recursive (IIR). The time unit
1083 * is 1 second. Arithmetic is done using 32-bit integers scaled to
1084 * simulate 3 decimal digits of precision (multiplied by 1000).
1086 * With an FM_COEF of 933, and a time base of 1 second, the filter
1087 * has a half-life of 10 seconds, meaning that if the events quit
1088 * happening, then the rate returned from the fmeter_getrate()
1089 * will be cut in half each 10 seconds, until it converges to zero.
1091 * It is not worth doing a real infinitely recursive filter. If more
1092 * than FM_MAXTICKS ticks have elapsed since the last filter event,
1093 * just compute FM_MAXTICKS ticks worth, by which point the level
1096 * Limit the count of unprocessed events to FM_MAXCNT, so as to avoid
1097 * arithmetic overflow in the fmeter_update() routine.
1099 * Given the simple 32 bit integer arithmetic used, this meter works
1100 * best for reporting rates between one per millisecond (msec) and
1101 * one per 32 (approx) seconds. At constant rates faster than one
1102 * per msec it maxes out at values just under 1,000,000. At constant
1103 * rates between one per msec, and one per second it will stabilize
1104 * to a value N*1000, where N is the rate of events per second.
1105 * At constant rates between one per second and one per 32 seconds,
1106 * it will be choppy, moving up on the seconds that have an event,
1107 * and then decaying until the next event. At rates slower than
1108 * about one in 32 seconds, it decays all the way back to zero between
1112 #define FM_COEF 933 /* coefficient for half-life of 10 secs */
1113 #define FM_MAXTICKS ((time_t)99) /* useless computing more ticks than this */
1114 #define FM_MAXCNT 1000000 /* limit cnt to avoid overflow */
1115 #define FM_SCALE 1000 /* faux fixed point scale */
1117 /* Initialize a frequency meter */
1118 static void fmeter_init(struct fmeter
*fmp
)
1123 spin_lock_init(&fmp
->lock
);
1126 /* Internal meter update - process cnt events and update value */
1127 static void fmeter_update(struct fmeter
*fmp
)
1129 time_t now
= get_seconds();
1130 time_t ticks
= now
- fmp
->time
;
1135 ticks
= min(FM_MAXTICKS
, ticks
);
1137 fmp
->val
= (FM_COEF
* fmp
->val
) / FM_SCALE
;
1140 fmp
->val
+= ((FM_SCALE
- FM_COEF
) * fmp
->cnt
) / FM_SCALE
;
1144 /* Process any previous ticks, then bump cnt by one (times scale). */
1145 static void fmeter_markevent(struct fmeter
*fmp
)
1147 spin_lock(&fmp
->lock
);
1149 fmp
->cnt
= min(FM_MAXCNT
, fmp
->cnt
+ FM_SCALE
);
1150 spin_unlock(&fmp
->lock
);
1153 /* Process any previous ticks, then return current value. */
1154 static int fmeter_getrate(struct fmeter
*fmp
)
1158 spin_lock(&fmp
->lock
);
1161 spin_unlock(&fmp
->lock
);
1166 * Attack task specified by pid in 'pidbuf' to cpuset 'cs', possibly
1167 * writing the path of the old cpuset in 'ppathbuf' if it needs to be
1168 * notified on release.
1170 * Call holding manage_mutex. May take callback_mutex and task_lock of
1171 * the task 'pid' during call.
1174 static int attach_task(struct cpuset
*cs
, char *pidbuf
, char **ppathbuf
)
1177 struct task_struct
*tsk
;
1178 struct cpuset
*oldcs
;
1180 nodemask_t from
, to
;
1181 struct mm_struct
*mm
;
1184 if (sscanf(pidbuf
, "%d", &pid
) != 1)
1186 if (cpus_empty(cs
->cpus_allowed
) || nodes_empty(cs
->mems_allowed
))
1190 read_lock(&tasklist_lock
);
1192 tsk
= find_task_by_pid(pid
);
1193 if (!tsk
|| tsk
->flags
& PF_EXITING
) {
1194 read_unlock(&tasklist_lock
);
1198 get_task_struct(tsk
);
1199 read_unlock(&tasklist_lock
);
1201 if ((current
->euid
) && (current
->euid
!= tsk
->uid
)
1202 && (current
->euid
!= tsk
->suid
)) {
1203 put_task_struct(tsk
);
1208 get_task_struct(tsk
);
1211 retval
= security_task_setscheduler(tsk
, 0, NULL
);
1213 put_task_struct(tsk
);
1217 mutex_lock(&callback_mutex
);
1220 oldcs
= tsk
->cpuset
;
1223 mutex_unlock(&callback_mutex
);
1224 put_task_struct(tsk
);
1227 atomic_inc(&cs
->count
);
1228 rcu_assign_pointer(tsk
->cpuset
, cs
);
1231 guarantee_online_cpus(cs
, &cpus
);
1232 set_cpus_allowed(tsk
, cpus
);
1234 from
= oldcs
->mems_allowed
;
1235 to
= cs
->mems_allowed
;
1237 mutex_unlock(&callback_mutex
);
1239 mm
= get_task_mm(tsk
);
1241 mpol_rebind_mm(mm
, &to
);
1242 if (is_memory_migrate(cs
))
1243 cpuset_migrate_mm(mm
, &from
, &to
);
1247 put_task_struct(tsk
);
1249 if (atomic_dec_and_test(&oldcs
->count
))
1250 check_for_release(oldcs
, ppathbuf
);
1254 /* The various types of files and directories in a cpuset file system */
1259 FILE_MEMORY_MIGRATE
,
1264 FILE_NOTIFY_ON_RELEASE
,
1265 FILE_MEMORY_PRESSURE_ENABLED
,
1266 FILE_MEMORY_PRESSURE
,
1270 } cpuset_filetype_t
;
1272 static ssize_t
cpuset_common_file_write(struct file
*file
, const char __user
*userbuf
,
1273 size_t nbytes
, loff_t
*unused_ppos
)
1275 struct cpuset
*cs
= __d_cs(file
->f_dentry
->d_parent
);
1276 struct cftype
*cft
= __d_cft(file
->f_dentry
);
1277 cpuset_filetype_t type
= cft
->private;
1279 char *pathbuf
= NULL
;
1282 /* Crude upper limit on largest legitimate cpulist user might write. */
1283 if (nbytes
> 100 + 6 * NR_CPUS
)
1286 /* +1 for nul-terminator */
1287 if ((buffer
= kmalloc(nbytes
+ 1, GFP_KERNEL
)) == 0)
1290 if (copy_from_user(buffer
, userbuf
, nbytes
)) {
1294 buffer
[nbytes
] = 0; /* nul-terminate */
1296 mutex_lock(&manage_mutex
);
1298 if (is_removed(cs
)) {
1305 retval
= update_cpumask(cs
, buffer
);
1308 retval
= update_nodemask(cs
, buffer
);
1310 case FILE_CPU_EXCLUSIVE
:
1311 retval
= update_flag(CS_CPU_EXCLUSIVE
, cs
, buffer
);
1313 case FILE_MEM_EXCLUSIVE
:
1314 retval
= update_flag(CS_MEM_EXCLUSIVE
, cs
, buffer
);
1316 case FILE_NOTIFY_ON_RELEASE
:
1317 retval
= update_flag(CS_NOTIFY_ON_RELEASE
, cs
, buffer
);
1319 case FILE_MEMORY_MIGRATE
:
1320 retval
= update_flag(CS_MEMORY_MIGRATE
, cs
, buffer
);
1322 case FILE_MEMORY_PRESSURE_ENABLED
:
1323 retval
= update_memory_pressure_enabled(cs
, buffer
);
1325 case FILE_MEMORY_PRESSURE
:
1328 case FILE_SPREAD_PAGE
:
1329 retval
= update_flag(CS_SPREAD_PAGE
, cs
, buffer
);
1330 cs
->mems_generation
= cpuset_mems_generation
++;
1332 case FILE_SPREAD_SLAB
:
1333 retval
= update_flag(CS_SPREAD_SLAB
, cs
, buffer
);
1334 cs
->mems_generation
= cpuset_mems_generation
++;
1337 retval
= attach_task(cs
, buffer
, &pathbuf
);
1347 mutex_unlock(&manage_mutex
);
1348 cpuset_release_agent(pathbuf
);
1354 static ssize_t
cpuset_file_write(struct file
*file
, const char __user
*buf
,
1355 size_t nbytes
, loff_t
*ppos
)
1358 struct cftype
*cft
= __d_cft(file
->f_dentry
);
1362 /* special function ? */
1364 retval
= cft
->write(file
, buf
, nbytes
, ppos
);
1366 retval
= cpuset_common_file_write(file
, buf
, nbytes
, ppos
);
1372 * These ascii lists should be read in a single call, by using a user
1373 * buffer large enough to hold the entire map. If read in smaller
1374 * chunks, there is no guarantee of atomicity. Since the display format
1375 * used, list of ranges of sequential numbers, is variable length,
1376 * and since these maps can change value dynamically, one could read
1377 * gibberish by doing partial reads while a list was changing.
1378 * A single large read to a buffer that crosses a page boundary is
1379 * ok, because the result being copied to user land is not recomputed
1380 * across a page fault.
1383 static int cpuset_sprintf_cpulist(char *page
, struct cpuset
*cs
)
1387 mutex_lock(&callback_mutex
);
1388 mask
= cs
->cpus_allowed
;
1389 mutex_unlock(&callback_mutex
);
1391 return cpulist_scnprintf(page
, PAGE_SIZE
, mask
);
1394 static int cpuset_sprintf_memlist(char *page
, struct cpuset
*cs
)
1398 mutex_lock(&callback_mutex
);
1399 mask
= cs
->mems_allowed
;
1400 mutex_unlock(&callback_mutex
);
1402 return nodelist_scnprintf(page
, PAGE_SIZE
, mask
);
1405 static ssize_t
cpuset_common_file_read(struct file
*file
, char __user
*buf
,
1406 size_t nbytes
, loff_t
*ppos
)
1408 struct cftype
*cft
= __d_cft(file
->f_dentry
);
1409 struct cpuset
*cs
= __d_cs(file
->f_dentry
->d_parent
);
1410 cpuset_filetype_t type
= cft
->private;
1415 if (!(page
= (char *)__get_free_page(GFP_KERNEL
)))
1422 s
+= cpuset_sprintf_cpulist(s
, cs
);
1425 s
+= cpuset_sprintf_memlist(s
, cs
);
1427 case FILE_CPU_EXCLUSIVE
:
1428 *s
++ = is_cpu_exclusive(cs
) ? '1' : '0';
1430 case FILE_MEM_EXCLUSIVE
:
1431 *s
++ = is_mem_exclusive(cs
) ? '1' : '0';
1433 case FILE_NOTIFY_ON_RELEASE
:
1434 *s
++ = notify_on_release(cs
) ? '1' : '0';
1436 case FILE_MEMORY_MIGRATE
:
1437 *s
++ = is_memory_migrate(cs
) ? '1' : '0';
1439 case FILE_MEMORY_PRESSURE_ENABLED
:
1440 *s
++ = cpuset_memory_pressure_enabled
? '1' : '0';
1442 case FILE_MEMORY_PRESSURE
:
1443 s
+= sprintf(s
, "%d", fmeter_getrate(&cs
->fmeter
));
1445 case FILE_SPREAD_PAGE
:
1446 *s
++ = is_spread_page(cs
) ? '1' : '0';
1448 case FILE_SPREAD_SLAB
:
1449 *s
++ = is_spread_slab(cs
) ? '1' : '0';
1457 retval
= simple_read_from_buffer(buf
, nbytes
, ppos
, page
, s
- page
);
1459 free_page((unsigned long)page
);
1463 static ssize_t
cpuset_file_read(struct file
*file
, char __user
*buf
, size_t nbytes
,
1467 struct cftype
*cft
= __d_cft(file
->f_dentry
);
1471 /* special function ? */
1473 retval
= cft
->read(file
, buf
, nbytes
, ppos
);
1475 retval
= cpuset_common_file_read(file
, buf
, nbytes
, ppos
);
1480 static int cpuset_file_open(struct inode
*inode
, struct file
*file
)
1485 err
= generic_file_open(inode
, file
);
1489 cft
= __d_cft(file
->f_dentry
);
1493 err
= cft
->open(inode
, file
);
1500 static int cpuset_file_release(struct inode
*inode
, struct file
*file
)
1502 struct cftype
*cft
= __d_cft(file
->f_dentry
);
1504 return cft
->release(inode
, file
);
1509 * cpuset_rename - Only allow simple rename of directories in place.
1511 static int cpuset_rename(struct inode
*old_dir
, struct dentry
*old_dentry
,
1512 struct inode
*new_dir
, struct dentry
*new_dentry
)
1514 if (!S_ISDIR(old_dentry
->d_inode
->i_mode
))
1516 if (new_dentry
->d_inode
)
1518 if (old_dir
!= new_dir
)
1520 return simple_rename(old_dir
, old_dentry
, new_dir
, new_dentry
);
1523 static struct file_operations cpuset_file_operations
= {
1524 .read
= cpuset_file_read
,
1525 .write
= cpuset_file_write
,
1526 .llseek
= generic_file_llseek
,
1527 .open
= cpuset_file_open
,
1528 .release
= cpuset_file_release
,
1531 static struct inode_operations cpuset_dir_inode_operations
= {
1532 .lookup
= simple_lookup
,
1533 .mkdir
= cpuset_mkdir
,
1534 .rmdir
= cpuset_rmdir
,
1535 .rename
= cpuset_rename
,
1538 static int cpuset_create_file(struct dentry
*dentry
, int mode
)
1540 struct inode
*inode
;
1544 if (dentry
->d_inode
)
1547 inode
= cpuset_new_inode(mode
);
1551 if (S_ISDIR(mode
)) {
1552 inode
->i_op
= &cpuset_dir_inode_operations
;
1553 inode
->i_fop
= &simple_dir_operations
;
1555 /* start off with i_nlink == 2 (for "." entry) */
1557 } else if (S_ISREG(mode
)) {
1559 inode
->i_fop
= &cpuset_file_operations
;
1562 d_instantiate(dentry
, inode
);
1563 dget(dentry
); /* Extra count - pin the dentry in core */
1568 * cpuset_create_dir - create a directory for an object.
1569 * cs: the cpuset we create the directory for.
1570 * It must have a valid ->parent field
1571 * And we are going to fill its ->dentry field.
1572 * name: The name to give to the cpuset directory. Will be copied.
1573 * mode: mode to set on new directory.
1576 static int cpuset_create_dir(struct cpuset
*cs
, const char *name
, int mode
)
1578 struct dentry
*dentry
= NULL
;
1579 struct dentry
*parent
;
1582 parent
= cs
->parent
->dentry
;
1583 dentry
= cpuset_get_dentry(parent
, name
);
1585 return PTR_ERR(dentry
);
1586 error
= cpuset_create_file(dentry
, S_IFDIR
| mode
);
1588 dentry
->d_fsdata
= cs
;
1589 parent
->d_inode
->i_nlink
++;
1590 cs
->dentry
= dentry
;
1597 static int cpuset_add_file(struct dentry
*dir
, const struct cftype
*cft
)
1599 struct dentry
*dentry
;
1602 mutex_lock(&dir
->d_inode
->i_mutex
);
1603 dentry
= cpuset_get_dentry(dir
, cft
->name
);
1604 if (!IS_ERR(dentry
)) {
1605 error
= cpuset_create_file(dentry
, 0644 | S_IFREG
);
1607 dentry
->d_fsdata
= (void *)cft
;
1610 error
= PTR_ERR(dentry
);
1611 mutex_unlock(&dir
->d_inode
->i_mutex
);
1616 * Stuff for reading the 'tasks' file.
1618 * Reading this file can return large amounts of data if a cpuset has
1619 * *lots* of attached tasks. So it may need several calls to read(),
1620 * but we cannot guarantee that the information we produce is correct
1621 * unless we produce it entirely atomically.
1623 * Upon tasks file open(), a struct ctr_struct is allocated, that
1624 * will have a pointer to an array (also allocated here). The struct
1625 * ctr_struct * is stored in file->private_data. Its resources will
1626 * be freed by release() when the file is closed. The array is used
1627 * to sprintf the PIDs and then used by read().
1630 /* cpusets_tasks_read array */
1638 * Load into 'pidarray' up to 'npids' of the tasks using cpuset 'cs'.
1639 * Return actual number of pids loaded. No need to task_lock(p)
1640 * when reading out p->cpuset, as we don't really care if it changes
1641 * on the next cycle, and we are not going to try to dereference it.
1643 static int pid_array_load(pid_t
*pidarray
, int npids
, struct cpuset
*cs
)
1646 struct task_struct
*g
, *p
;
1648 read_lock(&tasklist_lock
);
1650 do_each_thread(g
, p
) {
1651 if (p
->cpuset
== cs
) {
1652 pidarray
[n
++] = p
->pid
;
1653 if (unlikely(n
== npids
))
1656 } while_each_thread(g
, p
);
1659 read_unlock(&tasklist_lock
);
1663 static int cmppid(const void *a
, const void *b
)
1665 return *(pid_t
*)a
- *(pid_t
*)b
;
1669 * Convert array 'a' of 'npids' pid_t's to a string of newline separated
1670 * decimal pids in 'buf'. Don't write more than 'sz' chars, but return
1671 * count 'cnt' of how many chars would be written if buf were large enough.
1673 static int pid_array_to_buf(char *buf
, int sz
, pid_t
*a
, int npids
)
1678 for (i
= 0; i
< npids
; i
++)
1679 cnt
+= snprintf(buf
+ cnt
, max(sz
- cnt
, 0), "%d\n", a
[i
]);
1684 * Handle an open on 'tasks' file. Prepare a buffer listing the
1685 * process id's of tasks currently attached to the cpuset being opened.
1687 * Does not require any specific cpuset mutexes, and does not take any.
1689 static int cpuset_tasks_open(struct inode
*unused
, struct file
*file
)
1691 struct cpuset
*cs
= __d_cs(file
->f_dentry
->d_parent
);
1692 struct ctr_struct
*ctr
;
1697 if (!(file
->f_mode
& FMODE_READ
))
1700 ctr
= kmalloc(sizeof(*ctr
), GFP_KERNEL
);
1705 * If cpuset gets more users after we read count, we won't have
1706 * enough space - tough. This race is indistinguishable to the
1707 * caller from the case that the additional cpuset users didn't
1708 * show up until sometime later on.
1710 npids
= atomic_read(&cs
->count
);
1711 pidarray
= kmalloc(npids
* sizeof(pid_t
), GFP_KERNEL
);
1715 npids
= pid_array_load(pidarray
, npids
, cs
);
1716 sort(pidarray
, npids
, sizeof(pid_t
), cmppid
, NULL
);
1718 /* Call pid_array_to_buf() twice, first just to get bufsz */
1719 ctr
->bufsz
= pid_array_to_buf(&c
, sizeof(c
), pidarray
, npids
) + 1;
1720 ctr
->buf
= kmalloc(ctr
->bufsz
, GFP_KERNEL
);
1723 ctr
->bufsz
= pid_array_to_buf(ctr
->buf
, ctr
->bufsz
, pidarray
, npids
);
1726 file
->private_data
= ctr
;
1737 static ssize_t
cpuset_tasks_read(struct file
*file
, char __user
*buf
,
1738 size_t nbytes
, loff_t
*ppos
)
1740 struct ctr_struct
*ctr
= file
->private_data
;
1742 if (*ppos
+ nbytes
> ctr
->bufsz
)
1743 nbytes
= ctr
->bufsz
- *ppos
;
1744 if (copy_to_user(buf
, ctr
->buf
+ *ppos
, nbytes
))
1750 static int cpuset_tasks_release(struct inode
*unused_inode
, struct file
*file
)
1752 struct ctr_struct
*ctr
;
1754 if (file
->f_mode
& FMODE_READ
) {
1755 ctr
= file
->private_data
;
1763 * for the common functions, 'private' gives the type of file
1766 static struct cftype cft_tasks
= {
1768 .open
= cpuset_tasks_open
,
1769 .read
= cpuset_tasks_read
,
1770 .release
= cpuset_tasks_release
,
1771 .private = FILE_TASKLIST
,
1774 static struct cftype cft_cpus
= {
1776 .private = FILE_CPULIST
,
1779 static struct cftype cft_mems
= {
1781 .private = FILE_MEMLIST
,
1784 static struct cftype cft_cpu_exclusive
= {
1785 .name
= "cpu_exclusive",
1786 .private = FILE_CPU_EXCLUSIVE
,
1789 static struct cftype cft_mem_exclusive
= {
1790 .name
= "mem_exclusive",
1791 .private = FILE_MEM_EXCLUSIVE
,
1794 static struct cftype cft_notify_on_release
= {
1795 .name
= "notify_on_release",
1796 .private = FILE_NOTIFY_ON_RELEASE
,
1799 static struct cftype cft_memory_migrate
= {
1800 .name
= "memory_migrate",
1801 .private = FILE_MEMORY_MIGRATE
,
1804 static struct cftype cft_memory_pressure_enabled
= {
1805 .name
= "memory_pressure_enabled",
1806 .private = FILE_MEMORY_PRESSURE_ENABLED
,
1809 static struct cftype cft_memory_pressure
= {
1810 .name
= "memory_pressure",
1811 .private = FILE_MEMORY_PRESSURE
,
1814 static struct cftype cft_spread_page
= {
1815 .name
= "memory_spread_page",
1816 .private = FILE_SPREAD_PAGE
,
1819 static struct cftype cft_spread_slab
= {
1820 .name
= "memory_spread_slab",
1821 .private = FILE_SPREAD_SLAB
,
1824 static int cpuset_populate_dir(struct dentry
*cs_dentry
)
1828 if ((err
= cpuset_add_file(cs_dentry
, &cft_cpus
)) < 0)
1830 if ((err
= cpuset_add_file(cs_dentry
, &cft_mems
)) < 0)
1832 if ((err
= cpuset_add_file(cs_dentry
, &cft_cpu_exclusive
)) < 0)
1834 if ((err
= cpuset_add_file(cs_dentry
, &cft_mem_exclusive
)) < 0)
1836 if ((err
= cpuset_add_file(cs_dentry
, &cft_notify_on_release
)) < 0)
1838 if ((err
= cpuset_add_file(cs_dentry
, &cft_memory_migrate
)) < 0)
1840 if ((err
= cpuset_add_file(cs_dentry
, &cft_memory_pressure
)) < 0)
1842 if ((err
= cpuset_add_file(cs_dentry
, &cft_spread_page
)) < 0)
1844 if ((err
= cpuset_add_file(cs_dentry
, &cft_spread_slab
)) < 0)
1846 if ((err
= cpuset_add_file(cs_dentry
, &cft_tasks
)) < 0)
1852 * cpuset_create - create a cpuset
1853 * parent: cpuset that will be parent of the new cpuset.
1854 * name: name of the new cpuset. Will be strcpy'ed.
1855 * mode: mode to set on new inode
1857 * Must be called with the mutex on the parent inode held
1860 static long cpuset_create(struct cpuset
*parent
, const char *name
, int mode
)
1865 cs
= kmalloc(sizeof(*cs
), GFP_KERNEL
);
1869 mutex_lock(&manage_mutex
);
1870 cpuset_update_task_memory_state();
1872 if (notify_on_release(parent
))
1873 set_bit(CS_NOTIFY_ON_RELEASE
, &cs
->flags
);
1874 if (is_spread_page(parent
))
1875 set_bit(CS_SPREAD_PAGE
, &cs
->flags
);
1876 if (is_spread_slab(parent
))
1877 set_bit(CS_SPREAD_SLAB
, &cs
->flags
);
1878 cs
->cpus_allowed
= CPU_MASK_NONE
;
1879 cs
->mems_allowed
= NODE_MASK_NONE
;
1880 atomic_set(&cs
->count
, 0);
1881 INIT_LIST_HEAD(&cs
->sibling
);
1882 INIT_LIST_HEAD(&cs
->children
);
1883 cs
->mems_generation
= cpuset_mems_generation
++;
1884 fmeter_init(&cs
->fmeter
);
1886 cs
->parent
= parent
;
1888 mutex_lock(&callback_mutex
);
1889 list_add(&cs
->sibling
, &cs
->parent
->children
);
1890 number_of_cpusets
++;
1891 mutex_unlock(&callback_mutex
);
1893 err
= cpuset_create_dir(cs
, name
, mode
);
1898 * Release manage_mutex before cpuset_populate_dir() because it
1899 * will down() this new directory's i_mutex and if we race with
1900 * another mkdir, we might deadlock.
1902 mutex_unlock(&manage_mutex
);
1904 err
= cpuset_populate_dir(cs
->dentry
);
1905 /* If err < 0, we have a half-filled directory - oh well ;) */
1908 list_del(&cs
->sibling
);
1909 mutex_unlock(&manage_mutex
);
1914 static int cpuset_mkdir(struct inode
*dir
, struct dentry
*dentry
, int mode
)
1916 struct cpuset
*c_parent
= dentry
->d_parent
->d_fsdata
;
1918 /* the vfs holds inode->i_mutex already */
1919 return cpuset_create(c_parent
, dentry
->d_name
.name
, mode
| S_IFDIR
);
1922 static int cpuset_rmdir(struct inode
*unused_dir
, struct dentry
*dentry
)
1924 struct cpuset
*cs
= dentry
->d_fsdata
;
1926 struct cpuset
*parent
;
1927 char *pathbuf
= NULL
;
1929 /* the vfs holds both inode->i_mutex already */
1931 mutex_lock(&manage_mutex
);
1932 cpuset_update_task_memory_state();
1933 if (atomic_read(&cs
->count
) > 0) {
1934 mutex_unlock(&manage_mutex
);
1937 if (!list_empty(&cs
->children
)) {
1938 mutex_unlock(&manage_mutex
);
1941 parent
= cs
->parent
;
1942 mutex_lock(&callback_mutex
);
1943 set_bit(CS_REMOVED
, &cs
->flags
);
1944 if (is_cpu_exclusive(cs
))
1945 update_cpu_domains(cs
);
1946 list_del(&cs
->sibling
); /* delete my sibling from parent->children */
1947 spin_lock(&cs
->dentry
->d_lock
);
1948 d
= dget(cs
->dentry
);
1950 spin_unlock(&d
->d_lock
);
1951 cpuset_d_remove_dir(d
);
1953 number_of_cpusets
--;
1954 mutex_unlock(&callback_mutex
);
1955 if (list_empty(&parent
->children
))
1956 check_for_release(parent
, &pathbuf
);
1957 mutex_unlock(&manage_mutex
);
1958 cpuset_release_agent(pathbuf
);
1963 * cpuset_init_early - just enough so that the calls to
1964 * cpuset_update_task_memory_state() in early init code
1968 int __init
cpuset_init_early(void)
1970 struct task_struct
*tsk
= current
;
1972 tsk
->cpuset
= &top_cpuset
;
1973 tsk
->cpuset
->mems_generation
= cpuset_mems_generation
++;
1978 * cpuset_init - initialize cpusets at system boot
1980 * Description: Initialize top_cpuset and the cpuset internal file system,
1983 int __init
cpuset_init(void)
1985 struct dentry
*root
;
1988 top_cpuset
.cpus_allowed
= CPU_MASK_ALL
;
1989 top_cpuset
.mems_allowed
= NODE_MASK_ALL
;
1991 fmeter_init(&top_cpuset
.fmeter
);
1992 top_cpuset
.mems_generation
= cpuset_mems_generation
++;
1994 init_task
.cpuset
= &top_cpuset
;
1996 err
= register_filesystem(&cpuset_fs_type
);
1999 cpuset_mount
= kern_mount(&cpuset_fs_type
);
2000 if (IS_ERR(cpuset_mount
)) {
2001 printk(KERN_ERR
"cpuset: could not mount!\n");
2002 err
= PTR_ERR(cpuset_mount
);
2003 cpuset_mount
= NULL
;
2006 root
= cpuset_mount
->mnt_sb
->s_root
;
2007 root
->d_fsdata
= &top_cpuset
;
2008 root
->d_inode
->i_nlink
++;
2009 top_cpuset
.dentry
= root
;
2010 root
->d_inode
->i_op
= &cpuset_dir_inode_operations
;
2011 number_of_cpusets
= 1;
2012 err
= cpuset_populate_dir(root
);
2013 /* memory_pressure_enabled is in root cpuset only */
2015 err
= cpuset_add_file(root
, &cft_memory_pressure_enabled
);
2021 * cpuset_init_smp - initialize cpus_allowed
2023 * Description: Finish top cpuset after cpu, node maps are initialized
2026 void __init
cpuset_init_smp(void)
2028 top_cpuset
.cpus_allowed
= cpu_online_map
;
2029 top_cpuset
.mems_allowed
= node_online_map
;
2033 * cpuset_fork - attach newly forked task to its parents cpuset.
2034 * @tsk: pointer to task_struct of forking parent process.
2036 * Description: A task inherits its parent's cpuset at fork().
2038 * A pointer to the shared cpuset was automatically copied in fork.c
2039 * by dup_task_struct(). However, we ignore that copy, since it was
2040 * not made under the protection of task_lock(), so might no longer be
2041 * a valid cpuset pointer. attach_task() might have already changed
2042 * current->cpuset, allowing the previously referenced cpuset to
2043 * be removed and freed. Instead, we task_lock(current) and copy
2044 * its present value of current->cpuset for our freshly forked child.
2046 * At the point that cpuset_fork() is called, 'current' is the parent
2047 * task, and the passed argument 'child' points to the child task.
2050 void cpuset_fork(struct task_struct
*child
)
2053 child
->cpuset
= current
->cpuset
;
2054 atomic_inc(&child
->cpuset
->count
);
2055 task_unlock(current
);
2059 * cpuset_exit - detach cpuset from exiting task
2060 * @tsk: pointer to task_struct of exiting process
2062 * Description: Detach cpuset from @tsk and release it.
2064 * Note that cpusets marked notify_on_release force every task in
2065 * them to take the global manage_mutex mutex when exiting.
2066 * This could impact scaling on very large systems. Be reluctant to
2067 * use notify_on_release cpusets where very high task exit scaling
2068 * is required on large systems.
2070 * Don't even think about derefencing 'cs' after the cpuset use count
2071 * goes to zero, except inside a critical section guarded by manage_mutex
2072 * or callback_mutex. Otherwise a zero cpuset use count is a license to
2073 * any other task to nuke the cpuset immediately, via cpuset_rmdir().
2075 * This routine has to take manage_mutex, not callback_mutex, because
2076 * it is holding that mutex while calling check_for_release(),
2077 * which calls kmalloc(), so can't be called holding callback_mutex().
2079 * We don't need to task_lock() this reference to tsk->cpuset,
2080 * because tsk is already marked PF_EXITING, so attach_task() won't
2081 * mess with it, or task is a failed fork, never visible to attach_task.
2083 * the_top_cpuset_hack:
2085 * Set the exiting tasks cpuset to the root cpuset (top_cpuset).
2087 * Don't leave a task unable to allocate memory, as that is an
2088 * accident waiting to happen should someone add a callout in
2089 * do_exit() after the cpuset_exit() call that might allocate.
2090 * If a task tries to allocate memory with an invalid cpuset,
2091 * it will oops in cpuset_update_task_memory_state().
2093 * We call cpuset_exit() while the task is still competent to
2094 * handle notify_on_release(), then leave the task attached to
2095 * the root cpuset (top_cpuset) for the remainder of its exit.
2097 * To do this properly, we would increment the reference count on
2098 * top_cpuset, and near the very end of the kernel/exit.c do_exit()
2099 * code we would add a second cpuset function call, to drop that
2100 * reference. This would just create an unnecessary hot spot on
2101 * the top_cpuset reference count, to no avail.
2103 * Normally, holding a reference to a cpuset without bumping its
2104 * count is unsafe. The cpuset could go away, or someone could
2105 * attach us to a different cpuset, decrementing the count on
2106 * the first cpuset that we never incremented. But in this case,
2107 * top_cpuset isn't going away, and either task has PF_EXITING set,
2108 * which wards off any attach_task() attempts, or task is a failed
2109 * fork, never visible to attach_task.
2111 * Another way to do this would be to set the cpuset pointer
2112 * to NULL here, and check in cpuset_update_task_memory_state()
2113 * for a NULL pointer. This hack avoids that NULL check, for no
2114 * cost (other than this way too long comment ;).
2117 void cpuset_exit(struct task_struct
*tsk
)
2122 tsk
->cpuset
= &top_cpuset
; /* the_top_cpuset_hack - see above */
2124 if (notify_on_release(cs
)) {
2125 char *pathbuf
= NULL
;
2127 mutex_lock(&manage_mutex
);
2128 if (atomic_dec_and_test(&cs
->count
))
2129 check_for_release(cs
, &pathbuf
);
2130 mutex_unlock(&manage_mutex
);
2131 cpuset_release_agent(pathbuf
);
2133 atomic_dec(&cs
->count
);
2138 * cpuset_cpus_allowed - return cpus_allowed mask from a tasks cpuset.
2139 * @tsk: pointer to task_struct from which to obtain cpuset->cpus_allowed.
2141 * Description: Returns the cpumask_t cpus_allowed of the cpuset
2142 * attached to the specified @tsk. Guaranteed to return some non-empty
2143 * subset of cpu_online_map, even if this means going outside the
2147 cpumask_t
cpuset_cpus_allowed(struct task_struct
*tsk
)
2151 mutex_lock(&callback_mutex
);
2153 guarantee_online_cpus(tsk
->cpuset
, &mask
);
2155 mutex_unlock(&callback_mutex
);
2160 void cpuset_init_current_mems_allowed(void)
2162 current
->mems_allowed
= NODE_MASK_ALL
;
2166 * cpuset_mems_allowed - return mems_allowed mask from a tasks cpuset.
2167 * @tsk: pointer to task_struct from which to obtain cpuset->mems_allowed.
2169 * Description: Returns the nodemask_t mems_allowed of the cpuset
2170 * attached to the specified @tsk. Guaranteed to return some non-empty
2171 * subset of node_online_map, even if this means going outside the
2175 nodemask_t
cpuset_mems_allowed(struct task_struct
*tsk
)
2179 mutex_lock(&callback_mutex
);
2181 guarantee_online_mems(tsk
->cpuset
, &mask
);
2183 mutex_unlock(&callback_mutex
);
2189 * cpuset_zonelist_valid_mems_allowed - check zonelist vs. curremt mems_allowed
2190 * @zl: the zonelist to be checked
2192 * Are any of the nodes on zonelist zl allowed in current->mems_allowed?
2194 int cpuset_zonelist_valid_mems_allowed(struct zonelist
*zl
)
2198 for (i
= 0; zl
->zones
[i
]; i
++) {
2199 int nid
= zl
->zones
[i
]->zone_pgdat
->node_id
;
2201 if (node_isset(nid
, current
->mems_allowed
))
2208 * nearest_exclusive_ancestor() - Returns the nearest mem_exclusive
2209 * ancestor to the specified cpuset. Call holding callback_mutex.
2210 * If no ancestor is mem_exclusive (an unusual configuration), then
2211 * returns the root cpuset.
2213 static const struct cpuset
*nearest_exclusive_ancestor(const struct cpuset
*cs
)
2215 while (!is_mem_exclusive(cs
) && cs
->parent
)
2221 * cpuset_zone_allowed - Can we allocate memory on zone z's memory node?
2222 * @z: is this zone on an allowed node?
2223 * @gfp_mask: memory allocation flags (we use __GFP_HARDWALL)
2225 * If we're in interrupt, yes, we can always allocate. If zone
2226 * z's node is in our tasks mems_allowed, yes. If it's not a
2227 * __GFP_HARDWALL request and this zone's nodes is in the nearest
2228 * mem_exclusive cpuset ancestor to this tasks cpuset, yes.
2231 * GFP_USER allocations are marked with the __GFP_HARDWALL bit,
2232 * and do not allow allocations outside the current tasks cpuset.
2233 * GFP_KERNEL allocations are not so marked, so can escape to the
2234 * nearest mem_exclusive ancestor cpuset.
2236 * Scanning up parent cpusets requires callback_mutex. The __alloc_pages()
2237 * routine only calls here with __GFP_HARDWALL bit _not_ set if
2238 * it's a GFP_KERNEL allocation, and all nodes in the current tasks
2239 * mems_allowed came up empty on the first pass over the zonelist.
2240 * So only GFP_KERNEL allocations, if all nodes in the cpuset are
2241 * short of memory, might require taking the callback_mutex mutex.
2243 * The first call here from mm/page_alloc:get_page_from_freelist()
2244 * has __GFP_HARDWALL set in gfp_mask, enforcing hardwall cpusets, so
2245 * no allocation on a node outside the cpuset is allowed (unless in
2246 * interrupt, of course).
2248 * The second pass through get_page_from_freelist() doesn't even call
2249 * here for GFP_ATOMIC calls. For those calls, the __alloc_pages()
2250 * variable 'wait' is not set, and the bit ALLOC_CPUSET is not set
2251 * in alloc_flags. That logic and the checks below have the combined
2253 * in_interrupt - any node ok (current task context irrelevant)
2254 * GFP_ATOMIC - any node ok
2255 * GFP_KERNEL - any node in enclosing mem_exclusive cpuset ok
2256 * GFP_USER - only nodes in current tasks mems allowed ok.
2259 * Don't call cpuset_zone_allowed() if you can't sleep, unless you
2260 * pass in the __GFP_HARDWALL flag set in gfp_flag, which disables
2261 * the code that might scan up ancestor cpusets and sleep.
2264 int __cpuset_zone_allowed(struct zone
*z
, gfp_t gfp_mask
)
2266 int node
; /* node that zone z is on */
2267 const struct cpuset
*cs
; /* current cpuset ancestors */
2268 int allowed
; /* is allocation in zone z allowed? */
2272 node
= z
->zone_pgdat
->node_id
;
2273 might_sleep_if(!(gfp_mask
& __GFP_HARDWALL
));
2274 if (node_isset(node
, current
->mems_allowed
))
2276 if (gfp_mask
& __GFP_HARDWALL
) /* If hardwall request, stop here */
2279 if (current
->flags
& PF_EXITING
) /* Let dying task have memory */
2282 /* Not hardwall and node outside mems_allowed: scan up cpusets */
2283 mutex_lock(&callback_mutex
);
2286 cs
= nearest_exclusive_ancestor(current
->cpuset
);
2287 task_unlock(current
);
2289 allowed
= node_isset(node
, cs
->mems_allowed
);
2290 mutex_unlock(&callback_mutex
);
2295 * cpuset_lock - lock out any changes to cpuset structures
2297 * The out of memory (oom) code needs to mutex_lock cpusets
2298 * from being changed while it scans the tasklist looking for a
2299 * task in an overlapping cpuset. Expose callback_mutex via this
2300 * cpuset_lock() routine, so the oom code can lock it, before
2301 * locking the task list. The tasklist_lock is a spinlock, so
2302 * must be taken inside callback_mutex.
2305 void cpuset_lock(void)
2307 mutex_lock(&callback_mutex
);
2311 * cpuset_unlock - release lock on cpuset changes
2313 * Undo the lock taken in a previous cpuset_lock() call.
2316 void cpuset_unlock(void)
2318 mutex_unlock(&callback_mutex
);
2322 * cpuset_mem_spread_node() - On which node to begin search for a page
2324 * If a task is marked PF_SPREAD_PAGE or PF_SPREAD_SLAB (as for
2325 * tasks in a cpuset with is_spread_page or is_spread_slab set),
2326 * and if the memory allocation used cpuset_mem_spread_node()
2327 * to determine on which node to start looking, as it will for
2328 * certain page cache or slab cache pages such as used for file
2329 * system buffers and inode caches, then instead of starting on the
2330 * local node to look for a free page, rather spread the starting
2331 * node around the tasks mems_allowed nodes.
2333 * We don't have to worry about the returned node being offline
2334 * because "it can't happen", and even if it did, it would be ok.
2336 * The routines calling guarantee_online_mems() are careful to
2337 * only set nodes in task->mems_allowed that are online. So it
2338 * should not be possible for the following code to return an
2339 * offline node. But if it did, that would be ok, as this routine
2340 * is not returning the node where the allocation must be, only
2341 * the node where the search should start. The zonelist passed to
2342 * __alloc_pages() will include all nodes. If the slab allocator
2343 * is passed an offline node, it will fall back to the local node.
2344 * See kmem_cache_alloc_node().
2347 int cpuset_mem_spread_node(void)
2351 node
= next_node(current
->cpuset_mem_spread_rotor
, current
->mems_allowed
);
2352 if (node
== MAX_NUMNODES
)
2353 node
= first_node(current
->mems_allowed
);
2354 current
->cpuset_mem_spread_rotor
= node
;
2357 EXPORT_SYMBOL_GPL(cpuset_mem_spread_node
);
2360 * cpuset_excl_nodes_overlap - Do we overlap @p's mem_exclusive ancestors?
2361 * @p: pointer to task_struct of some other task.
2363 * Description: Return true if the nearest mem_exclusive ancestor
2364 * cpusets of tasks @p and current overlap. Used by oom killer to
2365 * determine if task @p's memory usage might impact the memory
2366 * available to the current task.
2368 * Call while holding callback_mutex.
2371 int cpuset_excl_nodes_overlap(const struct task_struct
*p
)
2373 const struct cpuset
*cs1
, *cs2
; /* my and p's cpuset ancestors */
2374 int overlap
= 0; /* do cpusets overlap? */
2377 if (current
->flags
& PF_EXITING
) {
2378 task_unlock(current
);
2381 cs1
= nearest_exclusive_ancestor(current
->cpuset
);
2382 task_unlock(current
);
2384 task_lock((struct task_struct
*)p
);
2385 if (p
->flags
& PF_EXITING
) {
2386 task_unlock((struct task_struct
*)p
);
2389 cs2
= nearest_exclusive_ancestor(p
->cpuset
);
2390 task_unlock((struct task_struct
*)p
);
2392 overlap
= nodes_intersects(cs1
->mems_allowed
, cs2
->mems_allowed
);
2398 * Collection of memory_pressure is suppressed unless
2399 * this flag is enabled by writing "1" to the special
2400 * cpuset file 'memory_pressure_enabled' in the root cpuset.
2403 int cpuset_memory_pressure_enabled __read_mostly
;
2406 * cpuset_memory_pressure_bump - keep stats of per-cpuset reclaims.
2408 * Keep a running average of the rate of synchronous (direct)
2409 * page reclaim efforts initiated by tasks in each cpuset.
2411 * This represents the rate at which some task in the cpuset
2412 * ran low on memory on all nodes it was allowed to use, and
2413 * had to enter the kernels page reclaim code in an effort to
2414 * create more free memory by tossing clean pages or swapping
2415 * or writing dirty pages.
2417 * Display to user space in the per-cpuset read-only file
2418 * "memory_pressure". Value displayed is an integer
2419 * representing the recent rate of entry into the synchronous
2420 * (direct) page reclaim by any task attached to the cpuset.
2423 void __cpuset_memory_pressure_bump(void)
2428 cs
= current
->cpuset
;
2429 fmeter_markevent(&cs
->fmeter
);
2430 task_unlock(current
);
2434 * proc_cpuset_show()
2435 * - Print tasks cpuset path into seq_file.
2436 * - Used for /proc/<pid>/cpuset.
2437 * - No need to task_lock(tsk) on this tsk->cpuset reference, as it
2438 * doesn't really matter if tsk->cpuset changes after we read it,
2439 * and we take manage_mutex, keeping attach_task() from changing it
2440 * anyway. No need to check that tsk->cpuset != NULL, thanks to
2441 * the_top_cpuset_hack in cpuset_exit(), which sets an exiting tasks
2442 * cpuset to top_cpuset.
2444 static int proc_cpuset_show(struct seq_file
*m
, void *v
)
2446 struct task_ref
*tref
;
2447 struct task_struct
*tsk
;
2452 buf
= kmalloc(PAGE_SIZE
, GFP_KERNEL
);
2458 tsk
= get_tref_task(tref
);
2463 mutex_lock(&manage_mutex
);
2465 retval
= cpuset_path(tsk
->cpuset
, buf
, PAGE_SIZE
);
2471 mutex_unlock(&manage_mutex
);
2472 put_task_struct(tsk
);
2479 static int cpuset_open(struct inode
*inode
, struct file
*file
)
2481 struct task_ref
*tref
= PROC_I(inode
)->tref
;
2482 return single_open(file
, proc_cpuset_show
, tref
);
2485 struct file_operations proc_cpuset_operations
= {
2486 .open
= cpuset_open
,
2488 .llseek
= seq_lseek
,
2489 .release
= single_release
,
2492 /* Display task cpus_allowed, mems_allowed in /proc/<pid>/status file. */
2493 char *cpuset_task_status_allowed(struct task_struct
*task
, char *buffer
)
2495 buffer
+= sprintf(buffer
, "Cpus_allowed:\t");
2496 buffer
+= cpumask_scnprintf(buffer
, PAGE_SIZE
, task
->cpus_allowed
);
2497 buffer
+= sprintf(buffer
, "\n");
2498 buffer
+= sprintf(buffer
, "Mems_allowed:\t");
2499 buffer
+= nodemask_scnprintf(buffer
, PAGE_SIZE
, task
->mems_allowed
);
2500 buffer
+= sprintf(buffer
, "\n");