4 * Processor and Memory placement constraints for sets of tasks.
6 * Copyright (C) 2003 BULL SA.
7 * Copyright (C) 2004 Silicon Graphics, Inc.
9 * Portions derived from Patrick Mochel's sysfs code.
10 * sysfs is Copyright (c) 2001-3 Patrick Mochel
11 * Portions Copyright (c) 2004 Silicon Graphics, Inc.
13 * 2003-10-10 Written by Simon Derr <simon.derr@bull.net>
14 * 2003-10-22 Updates by Stephen Hemminger.
15 * 2004 May-July Rework by Paul Jackson <pj@sgi.com>
17 * This file is subject to the terms and conditions of the GNU General Public
18 * License. See the file COPYING in the main directory of the Linux
19 * distribution for more details.
22 #include <linux/config.h>
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/sched.h>
43 #include <linux/seq_file.h>
44 #include <linux/slab.h>
45 #include <linux/smp_lock.h>
46 #include <linux/spinlock.h>
47 #include <linux/stat.h>
48 #include <linux/string.h>
49 #include <linux/time.h>
50 #include <linux/backing-dev.h>
51 #include <linux/sort.h>
53 #include <asm/uaccess.h>
54 #include <asm/atomic.h>
55 #include <asm/semaphore.h>
57 #define CPUSET_SUPER_MAGIC 0x27e0eb
60 * Tracks how many cpusets are currently defined in system.
61 * When there is only one cpuset (the root cpuset) we can
62 * short circuit some hooks.
64 int number_of_cpusets
;
66 /* See "Frequency meter" comments, below. */
69 int cnt
; /* unprocessed events count */
70 int val
; /* most recent output value */
71 time_t time
; /* clock (secs) when val computed */
72 spinlock_t lock
; /* guards read or write of above */
76 unsigned long flags
; /* "unsigned long" so bitops work */
77 cpumask_t cpus_allowed
; /* CPUs allowed to tasks in cpuset */
78 nodemask_t mems_allowed
; /* Memory Nodes allowed to tasks */
81 * Count is atomic so can incr (fork) or decr (exit) without a lock.
83 atomic_t count
; /* count tasks using this cpuset */
86 * We link our 'sibling' struct into our parents 'children'.
87 * Our children link their 'sibling' into our 'children'.
89 struct list_head sibling
; /* my parents children */
90 struct list_head children
; /* my children */
92 struct cpuset
*parent
; /* my parent */
93 struct dentry
*dentry
; /* cpuset fs entry */
96 * Copy of global cpuset_mems_generation as of the most
97 * recent time this cpuset changed its mems_allowed.
101 struct fmeter fmeter
; /* memory_pressure filter */
104 /* bits in struct cpuset flags field */
113 /* convenient tests for these bits */
114 static inline int is_cpu_exclusive(const struct cpuset
*cs
)
116 return !!test_bit(CS_CPU_EXCLUSIVE
, &cs
->flags
);
119 static inline int is_mem_exclusive(const struct cpuset
*cs
)
121 return !!test_bit(CS_MEM_EXCLUSIVE
, &cs
->flags
);
124 static inline int is_removed(const struct cpuset
*cs
)
126 return !!test_bit(CS_REMOVED
, &cs
->flags
);
129 static inline int notify_on_release(const struct cpuset
*cs
)
131 return !!test_bit(CS_NOTIFY_ON_RELEASE
, &cs
->flags
);
134 static inline int is_memory_migrate(const struct cpuset
*cs
)
136 return !!test_bit(CS_MEMORY_MIGRATE
, &cs
->flags
);
140 * Increment this atomic integer everytime any cpuset changes its
141 * mems_allowed value. Users of cpusets can track this generation
142 * number, and avoid having to lock and reload mems_allowed unless
143 * the cpuset they're using changes generation.
145 * A single, global generation is needed because attach_task() could
146 * reattach a task to a different cpuset, which must not have its
147 * generation numbers aliased with those of that tasks previous cpuset.
149 * Generations are needed for mems_allowed because one task cannot
150 * modify anothers memory placement. So we must enable every task,
151 * on every visit to __alloc_pages(), to efficiently check whether
152 * its current->cpuset->mems_allowed has changed, requiring an update
153 * of its current->mems_allowed.
155 static atomic_t cpuset_mems_generation
= ATOMIC_INIT(1);
157 static struct cpuset top_cpuset
= {
158 .flags
= ((1 << CS_CPU_EXCLUSIVE
) | (1 << CS_MEM_EXCLUSIVE
)),
159 .cpus_allowed
= CPU_MASK_ALL
,
160 .mems_allowed
= NODE_MASK_ALL
,
161 .count
= ATOMIC_INIT(0),
162 .sibling
= LIST_HEAD_INIT(top_cpuset
.sibling
),
163 .children
= LIST_HEAD_INIT(top_cpuset
.children
),
166 static struct vfsmount
*cpuset_mount
;
167 static struct super_block
*cpuset_sb
;
170 * We have two global cpuset semaphores below. They can nest.
171 * It is ok to first take manage_sem, then nest callback_sem. We also
172 * require taking task_lock() when dereferencing a tasks cpuset pointer.
173 * See "The task_lock() exception", at the end of this comment.
175 * A task must hold both semaphores to modify cpusets. If a task
176 * holds manage_sem, then it blocks others wanting that semaphore,
177 * ensuring that it is the only task able to also acquire callback_sem
178 * and be able to modify cpusets. It can perform various checks on
179 * the cpuset structure first, knowing nothing will change. It can
180 * also allocate memory while just holding manage_sem. While it is
181 * performing these checks, various callback routines can briefly
182 * acquire callback_sem to query cpusets. Once it is ready to make
183 * the changes, it takes callback_sem, blocking everyone else.
185 * Calls to the kernel memory allocator can not be made while holding
186 * callback_sem, as that would risk double tripping on callback_sem
187 * from one of the callbacks into the cpuset code from within
190 * If a task is only holding callback_sem, then it has read-only
193 * The task_struct fields mems_allowed and mems_generation may only
194 * be accessed in the context of that task, so require no locks.
196 * Any task can increment and decrement the count field without lock.
197 * So in general, code holding manage_sem or callback_sem can't rely
198 * on the count field not changing. However, if the count goes to
199 * zero, then only attach_task(), which holds both semaphores, can
200 * increment it again. Because a count of zero means that no tasks
201 * are currently attached, therefore there is no way a task attached
202 * to that cpuset can fork (the other way to increment the count).
203 * So code holding manage_sem or callback_sem can safely assume that
204 * if the count is zero, it will stay zero. Similarly, if a task
205 * holds manage_sem or callback_sem on a cpuset with zero count, it
206 * knows that the cpuset won't be removed, as cpuset_rmdir() needs
207 * both of those semaphores.
209 * A possible optimization to improve parallelism would be to make
210 * callback_sem a R/W semaphore (rwsem), allowing the callback routines
211 * to proceed in parallel, with read access, until the holder of
212 * manage_sem needed to take this rwsem for exclusive write access
213 * and modify some cpusets.
215 * The cpuset_common_file_write handler for operations that modify
216 * the cpuset hierarchy holds manage_sem across the entire operation,
217 * single threading all such cpuset modifications across the system.
219 * The cpuset_common_file_read() handlers only hold callback_sem across
220 * small pieces of code, such as when reading out possibly multi-word
221 * cpumasks and nodemasks.
223 * The fork and exit callbacks cpuset_fork() and cpuset_exit(), don't
224 * (usually) take either semaphore. These are the two most performance
225 * critical pieces of code here. The exception occurs on cpuset_exit(),
226 * when a task in a notify_on_release cpuset exits. Then manage_sem
227 * is taken, and if the cpuset count is zero, a usermode call made
228 * to /sbin/cpuset_release_agent with the name of the cpuset (path
229 * relative to the root of cpuset file system) as the argument.
231 * A cpuset can only be deleted if both its 'count' of using tasks
232 * is zero, and its list of 'children' cpusets is empty. Since all
233 * tasks in the system use _some_ cpuset, and since there is always at
234 * least one task in the system (init, pid == 1), therefore, top_cpuset
235 * always has either children cpusets and/or using tasks. So we don't
236 * need a special hack to ensure that top_cpuset cannot be deleted.
238 * The above "Tale of Two Semaphores" would be complete, but for:
240 * The task_lock() exception
242 * The need for this exception arises from the action of attach_task(),
243 * which overwrites one tasks cpuset pointer with another. It does
244 * so using both semaphores, however there are several performance
245 * critical places that need to reference task->cpuset without the
246 * expense of grabbing a system global semaphore. Therefore except as
247 * noted below, when dereferencing or, as in attach_task(), modifying
248 * a tasks cpuset pointer we use task_lock(), which acts on a spinlock
249 * (task->alloc_lock) already in the task_struct routinely used for
253 static DECLARE_MUTEX(manage_sem
);
254 static DECLARE_MUTEX(callback_sem
);
257 * A couple of forward declarations required, due to cyclic reference loop:
258 * cpuset_mkdir -> cpuset_create -> cpuset_populate_dir -> cpuset_add_file
259 * -> cpuset_create_file -> cpuset_dir_inode_operations -> cpuset_mkdir.
262 static int cpuset_mkdir(struct inode
*dir
, struct dentry
*dentry
, int mode
);
263 static int cpuset_rmdir(struct inode
*unused_dir
, struct dentry
*dentry
);
265 static struct backing_dev_info cpuset_backing_dev_info
= {
266 .ra_pages
= 0, /* No readahead */
267 .capabilities
= BDI_CAP_NO_ACCT_DIRTY
| BDI_CAP_NO_WRITEBACK
,
270 static struct inode
*cpuset_new_inode(mode_t mode
)
272 struct inode
*inode
= new_inode(cpuset_sb
);
275 inode
->i_mode
= mode
;
276 inode
->i_uid
= current
->fsuid
;
277 inode
->i_gid
= current
->fsgid
;
278 inode
->i_blksize
= PAGE_CACHE_SIZE
;
280 inode
->i_atime
= inode
->i_mtime
= inode
->i_ctime
= CURRENT_TIME
;
281 inode
->i_mapping
->backing_dev_info
= &cpuset_backing_dev_info
;
286 static void cpuset_diput(struct dentry
*dentry
, struct inode
*inode
)
288 /* is dentry a directory ? if so, kfree() associated cpuset */
289 if (S_ISDIR(inode
->i_mode
)) {
290 struct cpuset
*cs
= dentry
->d_fsdata
;
291 BUG_ON(!(is_removed(cs
)));
297 static struct dentry_operations cpuset_dops
= {
298 .d_iput
= cpuset_diput
,
301 static struct dentry
*cpuset_get_dentry(struct dentry
*parent
, const char *name
)
303 struct dentry
*d
= lookup_one_len(name
, parent
, strlen(name
));
305 d
->d_op
= &cpuset_dops
;
309 static void remove_dir(struct dentry
*d
)
311 struct dentry
*parent
= dget(d
->d_parent
);
314 simple_rmdir(parent
->d_inode
, d
);
319 * NOTE : the dentry must have been dget()'ed
321 static void cpuset_d_remove_dir(struct dentry
*dentry
)
323 struct list_head
*node
;
325 spin_lock(&dcache_lock
);
326 node
= dentry
->d_subdirs
.next
;
327 while (node
!= &dentry
->d_subdirs
) {
328 struct dentry
*d
= list_entry(node
, struct dentry
, d_child
);
332 spin_unlock(&dcache_lock
);
334 simple_unlink(dentry
->d_inode
, d
);
336 spin_lock(&dcache_lock
);
338 node
= dentry
->d_subdirs
.next
;
340 list_del_init(&dentry
->d_child
);
341 spin_unlock(&dcache_lock
);
345 static struct super_operations cpuset_ops
= {
346 .statfs
= simple_statfs
,
347 .drop_inode
= generic_delete_inode
,
350 static int cpuset_fill_super(struct super_block
*sb
, void *unused_data
,
356 sb
->s_blocksize
= PAGE_CACHE_SIZE
;
357 sb
->s_blocksize_bits
= PAGE_CACHE_SHIFT
;
358 sb
->s_magic
= CPUSET_SUPER_MAGIC
;
359 sb
->s_op
= &cpuset_ops
;
362 inode
= cpuset_new_inode(S_IFDIR
| S_IRUGO
| S_IXUGO
| S_IWUSR
);
364 inode
->i_op
= &simple_dir_inode_operations
;
365 inode
->i_fop
= &simple_dir_operations
;
366 /* directories start off with i_nlink == 2 (for "." entry) */
372 root
= d_alloc_root(inode
);
381 static struct super_block
*cpuset_get_sb(struct file_system_type
*fs_type
,
382 int flags
, const char *unused_dev_name
,
385 return get_sb_single(fs_type
, flags
, data
, cpuset_fill_super
);
388 static struct file_system_type cpuset_fs_type
= {
390 .get_sb
= cpuset_get_sb
,
391 .kill_sb
= kill_litter_super
,
396 * The files in the cpuset filesystem mostly have a very simple read/write
397 * handling, some common function will take care of it. Nevertheless some cases
398 * (read tasks) are special and therefore I define this structure for every
402 * When reading/writing to a file:
403 * - the cpuset to use in file->f_dentry->d_parent->d_fsdata
404 * - the 'cftype' of the file is file->f_dentry->d_fsdata
410 int (*open
) (struct inode
*inode
, struct file
*file
);
411 ssize_t (*read
) (struct file
*file
, char __user
*buf
, size_t nbytes
,
413 int (*write
) (struct file
*file
, const char __user
*buf
, size_t nbytes
,
415 int (*release
) (struct inode
*inode
, struct file
*file
);
418 static inline struct cpuset
*__d_cs(struct dentry
*dentry
)
420 return dentry
->d_fsdata
;
423 static inline struct cftype
*__d_cft(struct dentry
*dentry
)
425 return dentry
->d_fsdata
;
429 * Call with manage_sem held. Writes path of cpuset into buf.
430 * Returns 0 on success, -errno on error.
433 static int cpuset_path(const struct cpuset
*cs
, char *buf
, int buflen
)
437 start
= buf
+ buflen
;
441 int len
= cs
->dentry
->d_name
.len
;
442 if ((start
-= len
) < buf
)
443 return -ENAMETOOLONG
;
444 memcpy(start
, cs
->dentry
->d_name
.name
, len
);
451 return -ENAMETOOLONG
;
454 memmove(buf
, start
, buf
+ buflen
- start
);
459 * Notify userspace when a cpuset is released, by running
460 * /sbin/cpuset_release_agent with the name of the cpuset (path
461 * relative to the root of cpuset file system) as the argument.
463 * Most likely, this user command will try to rmdir this cpuset.
465 * This races with the possibility that some other task will be
466 * attached to this cpuset before it is removed, or that some other
467 * user task will 'mkdir' a child cpuset of this cpuset. That's ok.
468 * The presumed 'rmdir' will fail quietly if this cpuset is no longer
469 * unused, and this cpuset will be reprieved from its death sentence,
470 * to continue to serve a useful existence. Next time it's released,
471 * we will get notified again, if it still has 'notify_on_release' set.
473 * The final arg to call_usermodehelper() is 0, which means don't
474 * wait. The separate /sbin/cpuset_release_agent task is forked by
475 * call_usermodehelper(), then control in this thread returns here,
476 * without waiting for the release agent task. We don't bother to
477 * wait because the caller of this routine has no use for the exit
478 * status of the /sbin/cpuset_release_agent task, so no sense holding
479 * our caller up for that.
481 * When we had only one cpuset semaphore, we had to call this
482 * without holding it, to avoid deadlock when call_usermodehelper()
483 * allocated memory. With two locks, we could now call this while
484 * holding manage_sem, but we still don't, so as to minimize
485 * the time manage_sem is held.
488 static void cpuset_release_agent(const char *pathbuf
)
490 char *argv
[3], *envp
[3];
497 argv
[i
++] = "/sbin/cpuset_release_agent";
498 argv
[i
++] = (char *)pathbuf
;
502 /* minimal command environment */
503 envp
[i
++] = "HOME=/";
504 envp
[i
++] = "PATH=/sbin:/bin:/usr/sbin:/usr/bin";
507 call_usermodehelper(argv
[0], argv
, envp
, 0);
512 * Either cs->count of using tasks transitioned to zero, or the
513 * cs->children list of child cpusets just became empty. If this
514 * cs is notify_on_release() and now both the user count is zero and
515 * the list of children is empty, prepare cpuset path in a kmalloc'd
516 * buffer, to be returned via ppathbuf, so that the caller can invoke
517 * cpuset_release_agent() with it later on, once manage_sem is dropped.
518 * Call here with manage_sem held.
520 * This check_for_release() routine is responsible for kmalloc'ing
521 * pathbuf. The above cpuset_release_agent() is responsible for
522 * kfree'ing pathbuf. The caller of these routines is responsible
523 * for providing a pathbuf pointer, initialized to NULL, then
524 * calling check_for_release() with manage_sem held and the address
525 * of the pathbuf pointer, then dropping manage_sem, then calling
526 * cpuset_release_agent() with pathbuf, as set by check_for_release().
529 static void check_for_release(struct cpuset
*cs
, char **ppathbuf
)
531 if (notify_on_release(cs
) && atomic_read(&cs
->count
) == 0 &&
532 list_empty(&cs
->children
)) {
535 buf
= kmalloc(PAGE_SIZE
, GFP_KERNEL
);
538 if (cpuset_path(cs
, buf
, PAGE_SIZE
) < 0)
546 * Return in *pmask the portion of a cpusets's cpus_allowed that
547 * are online. If none are online, walk up the cpuset hierarchy
548 * until we find one that does have some online cpus. If we get
549 * all the way to the top and still haven't found any online cpus,
550 * return cpu_online_map. Or if passed a NULL cs from an exit'ing
551 * task, return cpu_online_map.
553 * One way or another, we guarantee to return some non-empty subset
556 * Call with callback_sem held.
559 static void guarantee_online_cpus(const struct cpuset
*cs
, cpumask_t
*pmask
)
561 while (cs
&& !cpus_intersects(cs
->cpus_allowed
, cpu_online_map
))
564 cpus_and(*pmask
, cs
->cpus_allowed
, cpu_online_map
);
566 *pmask
= cpu_online_map
;
567 BUG_ON(!cpus_intersects(*pmask
, cpu_online_map
));
571 * Return in *pmask the portion of a cpusets's mems_allowed that
572 * are online. If none are online, walk up the cpuset hierarchy
573 * until we find one that does have some online mems. If we get
574 * all the way to the top and still haven't found any online mems,
575 * return node_online_map.
577 * One way or another, we guarantee to return some non-empty subset
578 * of node_online_map.
580 * Call with callback_sem held.
583 static void guarantee_online_mems(const struct cpuset
*cs
, nodemask_t
*pmask
)
585 while (cs
&& !nodes_intersects(cs
->mems_allowed
, node_online_map
))
588 nodes_and(*pmask
, cs
->mems_allowed
, node_online_map
);
590 *pmask
= node_online_map
;
591 BUG_ON(!nodes_intersects(*pmask
, node_online_map
));
595 * cpuset_update_task_memory_state - update task memory placement
597 * If the current tasks cpusets mems_allowed changed behind our
598 * backs, update current->mems_allowed, mems_generation and task NUMA
599 * mempolicy to the new value.
601 * Task mempolicy is updated by rebinding it relative to the
602 * current->cpuset if a task has its memory placement changed.
603 * Do not call this routine if in_interrupt().
605 * Call without callback_sem or task_lock() held. May be called
606 * with or without manage_sem held. Except in early boot or
607 * an exiting task, when tsk->cpuset is NULL, this routine will
608 * acquire task_lock(). We don't need to use task_lock to guard
609 * against another task changing a non-NULL cpuset pointer to NULL,
610 * as that is only done by a task on itself, and if the current task
611 * is here, it is not simultaneously in the exit code NULL'ing its
612 * cpuset pointer. This routine also might acquire callback_sem and
613 * current->mm->mmap_sem during call.
615 * The task_lock() is required to dereference current->cpuset safely.
616 * Without it, we could pick up the pointer value of current->cpuset
617 * in one instruction, and then attach_task could give us a different
618 * cpuset, and then the cpuset we had could be removed and freed,
619 * and then on our next instruction, we could dereference a no longer
620 * valid cpuset pointer to get its mems_generation field.
622 * This routine is needed to update the per-task mems_allowed data,
623 * within the tasks context, when it is trying to allocate memory
624 * (in various mm/mempolicy.c routines) and notices that some other
625 * task has been modifying its cpuset.
628 void cpuset_update_task_memory_state()
630 int my_cpusets_mem_gen
;
631 struct task_struct
*tsk
= current
;
632 struct cpuset
*cs
= tsk
->cpuset
;
638 my_cpusets_mem_gen
= cs
->mems_generation
;
641 if (my_cpusets_mem_gen
!= tsk
->cpuset_mems_generation
) {
642 nodemask_t oldmem
= tsk
->mems_allowed
;
647 cs
= tsk
->cpuset
; /* Maybe changed when task not locked */
648 migrate
= is_memory_migrate(cs
);
649 guarantee_online_mems(cs
, &tsk
->mems_allowed
);
650 tsk
->cpuset_mems_generation
= cs
->mems_generation
;
653 mpol_rebind_task(tsk
, &tsk
->mems_allowed
);
654 if (!nodes_equal(oldmem
, tsk
->mems_allowed
)) {
656 do_migrate_pages(tsk
->mm
, &oldmem
,
665 * is_cpuset_subset(p, q) - Is cpuset p a subset of cpuset q?
667 * One cpuset is a subset of another if all its allowed CPUs and
668 * Memory Nodes are a subset of the other, and its exclusive flags
669 * are only set if the other's are set. Call holding manage_sem.
672 static int is_cpuset_subset(const struct cpuset
*p
, const struct cpuset
*q
)
674 return cpus_subset(p
->cpus_allowed
, q
->cpus_allowed
) &&
675 nodes_subset(p
->mems_allowed
, q
->mems_allowed
) &&
676 is_cpu_exclusive(p
) <= is_cpu_exclusive(q
) &&
677 is_mem_exclusive(p
) <= is_mem_exclusive(q
);
681 * validate_change() - Used to validate that any proposed cpuset change
682 * follows the structural rules for cpusets.
684 * If we replaced the flag and mask values of the current cpuset
685 * (cur) with those values in the trial cpuset (trial), would
686 * our various subset and exclusive rules still be valid? Presumes
689 * 'cur' is the address of an actual, in-use cpuset. Operations
690 * such as list traversal that depend on the actual address of the
691 * cpuset in the list must use cur below, not trial.
693 * 'trial' is the address of bulk structure copy of cur, with
694 * perhaps one or more of the fields cpus_allowed, mems_allowed,
695 * or flags changed to new, trial values.
697 * Return 0 if valid, -errno if not.
700 static int validate_change(const struct cpuset
*cur
, const struct cpuset
*trial
)
702 struct cpuset
*c
, *par
;
704 /* Each of our child cpusets must be a subset of us */
705 list_for_each_entry(c
, &cur
->children
, sibling
) {
706 if (!is_cpuset_subset(c
, trial
))
710 /* Remaining checks don't apply to root cpuset */
711 if ((par
= cur
->parent
) == NULL
)
714 /* We must be a subset of our parent cpuset */
715 if (!is_cpuset_subset(trial
, par
))
718 /* If either I or some sibling (!= me) is exclusive, we can't overlap */
719 list_for_each_entry(c
, &par
->children
, sibling
) {
720 if ((is_cpu_exclusive(trial
) || is_cpu_exclusive(c
)) &&
722 cpus_intersects(trial
->cpus_allowed
, c
->cpus_allowed
))
724 if ((is_mem_exclusive(trial
) || is_mem_exclusive(c
)) &&
726 nodes_intersects(trial
->mems_allowed
, c
->mems_allowed
))
734 * For a given cpuset cur, partition the system as follows
735 * a. All cpus in the parent cpuset's cpus_allowed that are not part of any
736 * exclusive child cpusets
737 * b. All cpus in the current cpuset's cpus_allowed that are not part of any
738 * exclusive child cpusets
739 * Build these two partitions by calling partition_sched_domains
741 * Call with manage_sem held. May nest a call to the
742 * lock_cpu_hotplug()/unlock_cpu_hotplug() pair.
745 static void update_cpu_domains(struct cpuset
*cur
)
747 struct cpuset
*c
, *par
= cur
->parent
;
748 cpumask_t pspan
, cspan
;
750 if (par
== NULL
|| cpus_empty(cur
->cpus_allowed
))
754 * Get all cpus from parent's cpus_allowed not part of exclusive
757 pspan
= par
->cpus_allowed
;
758 list_for_each_entry(c
, &par
->children
, sibling
) {
759 if (is_cpu_exclusive(c
))
760 cpus_andnot(pspan
, pspan
, c
->cpus_allowed
);
762 if (is_removed(cur
) || !is_cpu_exclusive(cur
)) {
763 cpus_or(pspan
, pspan
, cur
->cpus_allowed
);
764 if (cpus_equal(pspan
, cur
->cpus_allowed
))
766 cspan
= CPU_MASK_NONE
;
768 if (cpus_empty(pspan
))
770 cspan
= cur
->cpus_allowed
;
772 * Get all cpus from current cpuset's cpus_allowed not part
773 * of exclusive children
775 list_for_each_entry(c
, &cur
->children
, sibling
) {
776 if (is_cpu_exclusive(c
))
777 cpus_andnot(cspan
, cspan
, c
->cpus_allowed
);
782 partition_sched_domains(&pspan
, &cspan
);
783 unlock_cpu_hotplug();
787 * Call with manage_sem held. May take callback_sem during call.
790 static int update_cpumask(struct cpuset
*cs
, char *buf
)
792 struct cpuset trialcs
;
793 int retval
, cpus_unchanged
;
796 retval
= cpulist_parse(buf
, trialcs
.cpus_allowed
);
799 cpus_and(trialcs
.cpus_allowed
, trialcs
.cpus_allowed
, cpu_online_map
);
800 if (cpus_empty(trialcs
.cpus_allowed
))
802 retval
= validate_change(cs
, &trialcs
);
805 cpus_unchanged
= cpus_equal(cs
->cpus_allowed
, trialcs
.cpus_allowed
);
807 cs
->cpus_allowed
= trialcs
.cpus_allowed
;
809 if (is_cpu_exclusive(cs
) && !cpus_unchanged
)
810 update_cpu_domains(cs
);
815 * Call with manage_sem held. May take callback_sem during call.
818 static int update_nodemask(struct cpuset
*cs
, char *buf
)
820 struct cpuset trialcs
;
824 retval
= nodelist_parse(buf
, trialcs
.mems_allowed
);
827 nodes_and(trialcs
.mems_allowed
, trialcs
.mems_allowed
, node_online_map
);
828 if (nodes_empty(trialcs
.mems_allowed
)) {
832 retval
= validate_change(cs
, &trialcs
);
837 cs
->mems_allowed
= trialcs
.mems_allowed
;
838 atomic_inc(&cpuset_mems_generation
);
839 cs
->mems_generation
= atomic_read(&cpuset_mems_generation
);
847 * Call with manage_sem held.
850 static int update_memory_pressure_enabled(struct cpuset
*cs
, char *buf
)
852 if (simple_strtoul(buf
, NULL
, 10) != 0)
853 cpuset_memory_pressure_enabled
= 1;
855 cpuset_memory_pressure_enabled
= 0;
860 * update_flag - read a 0 or a 1 in a file and update associated flag
861 * bit: the bit to update (CS_CPU_EXCLUSIVE, CS_MEM_EXCLUSIVE,
862 * CS_NOTIFY_ON_RELEASE, CS_MEMORY_MIGRATE)
863 * cs: the cpuset to update
864 * buf: the buffer where we read the 0 or 1
866 * Call with manage_sem held.
869 static int update_flag(cpuset_flagbits_t bit
, struct cpuset
*cs
, char *buf
)
872 struct cpuset trialcs
;
873 int err
, cpu_exclusive_changed
;
875 turning_on
= (simple_strtoul(buf
, NULL
, 10) != 0);
879 set_bit(bit
, &trialcs
.flags
);
881 clear_bit(bit
, &trialcs
.flags
);
883 err
= validate_change(cs
, &trialcs
);
886 cpu_exclusive_changed
=
887 (is_cpu_exclusive(cs
) != is_cpu_exclusive(&trialcs
));
890 set_bit(bit
, &cs
->flags
);
892 clear_bit(bit
, &cs
->flags
);
895 if (cpu_exclusive_changed
)
896 update_cpu_domains(cs
);
901 * Frequency meter - How fast is some event occuring?
903 * These routines manage a digitally filtered, constant time based,
904 * event frequency meter. There are four routines:
905 * fmeter_init() - initialize a frequency meter.
906 * fmeter_markevent() - called each time the event happens.
907 * fmeter_getrate() - returns the recent rate of such events.
908 * fmeter_update() - internal routine used to update fmeter.
910 * A common data structure is passed to each of these routines,
911 * which is used to keep track of the state required to manage the
912 * frequency meter and its digital filter.
914 * The filter works on the number of events marked per unit time.
915 * The filter is single-pole low-pass recursive (IIR). The time unit
916 * is 1 second. Arithmetic is done using 32-bit integers scaled to
917 * simulate 3 decimal digits of precision (multiplied by 1000).
919 * With an FM_COEF of 933, and a time base of 1 second, the filter
920 * has a half-life of 10 seconds, meaning that if the events quit
921 * happening, then the rate returned from the fmeter_getrate()
922 * will be cut in half each 10 seconds, until it converges to zero.
924 * It is not worth doing a real infinitely recursive filter. If more
925 * than FM_MAXTICKS ticks have elapsed since the last filter event,
926 * just compute FM_MAXTICKS ticks worth, by which point the level
929 * Limit the count of unprocessed events to FM_MAXCNT, so as to avoid
930 * arithmetic overflow in the fmeter_update() routine.
932 * Given the simple 32 bit integer arithmetic used, this meter works
933 * best for reporting rates between one per millisecond (msec) and
934 * one per 32 (approx) seconds. At constant rates faster than one
935 * per msec it maxes out at values just under 1,000,000. At constant
936 * rates between one per msec, and one per second it will stabilize
937 * to a value N*1000, where N is the rate of events per second.
938 * At constant rates between one per second and one per 32 seconds,
939 * it will be choppy, moving up on the seconds that have an event,
940 * and then decaying until the next event. At rates slower than
941 * about one in 32 seconds, it decays all the way back to zero between
945 #define FM_COEF 933 /* coefficient for half-life of 10 secs */
946 #define FM_MAXTICKS ((time_t)99) /* useless computing more ticks than this */
947 #define FM_MAXCNT 1000000 /* limit cnt to avoid overflow */
948 #define FM_SCALE 1000 /* faux fixed point scale */
950 /* Initialize a frequency meter */
951 static void fmeter_init(struct fmeter
*fmp
)
956 spin_lock_init(&fmp
->lock
);
959 /* Internal meter update - process cnt events and update value */
960 static void fmeter_update(struct fmeter
*fmp
)
962 time_t now
= get_seconds();
963 time_t ticks
= now
- fmp
->time
;
968 ticks
= min(FM_MAXTICKS
, ticks
);
970 fmp
->val
= (FM_COEF
* fmp
->val
) / FM_SCALE
;
973 fmp
->val
+= ((FM_SCALE
- FM_COEF
) * fmp
->cnt
) / FM_SCALE
;
977 /* Process any previous ticks, then bump cnt by one (times scale). */
978 static void fmeter_markevent(struct fmeter
*fmp
)
980 spin_lock(&fmp
->lock
);
982 fmp
->cnt
= min(FM_MAXCNT
, fmp
->cnt
+ FM_SCALE
);
983 spin_unlock(&fmp
->lock
);
986 /* Process any previous ticks, then return current value. */
987 static int fmeter_getrate(struct fmeter
*fmp
)
991 spin_lock(&fmp
->lock
);
994 spin_unlock(&fmp
->lock
);
999 * Attack task specified by pid in 'pidbuf' to cpuset 'cs', possibly
1000 * writing the path of the old cpuset in 'ppathbuf' if it needs to be
1001 * notified on release.
1003 * Call holding manage_sem. May take callback_sem and task_lock of
1004 * the task 'pid' during call.
1007 static int attach_task(struct cpuset
*cs
, char *pidbuf
, char **ppathbuf
)
1010 struct task_struct
*tsk
;
1011 struct cpuset
*oldcs
;
1013 nodemask_t from
, to
;
1015 if (sscanf(pidbuf
, "%d", &pid
) != 1)
1017 if (cpus_empty(cs
->cpus_allowed
) || nodes_empty(cs
->mems_allowed
))
1021 read_lock(&tasklist_lock
);
1023 tsk
= find_task_by_pid(pid
);
1024 if (!tsk
|| tsk
->flags
& PF_EXITING
) {
1025 read_unlock(&tasklist_lock
);
1029 get_task_struct(tsk
);
1030 read_unlock(&tasklist_lock
);
1032 if ((current
->euid
) && (current
->euid
!= tsk
->uid
)
1033 && (current
->euid
!= tsk
->suid
)) {
1034 put_task_struct(tsk
);
1039 get_task_struct(tsk
);
1042 down(&callback_sem
);
1045 oldcs
= tsk
->cpuset
;
1049 put_task_struct(tsk
);
1052 atomic_inc(&cs
->count
);
1056 guarantee_online_cpus(cs
, &cpus
);
1057 set_cpus_allowed(tsk
, cpus
);
1059 from
= oldcs
->mems_allowed
;
1060 to
= cs
->mems_allowed
;
1063 if (is_memory_migrate(cs
))
1064 do_migrate_pages(tsk
->mm
, &from
, &to
, MPOL_MF_MOVE_ALL
);
1065 put_task_struct(tsk
);
1066 if (atomic_dec_and_test(&oldcs
->count
))
1067 check_for_release(oldcs
, ppathbuf
);
1071 /* The various types of files and directories in a cpuset file system */
1076 FILE_MEMORY_MIGRATE
,
1081 FILE_NOTIFY_ON_RELEASE
,
1082 FILE_MEMORY_PRESSURE_ENABLED
,
1083 FILE_MEMORY_PRESSURE
,
1085 } cpuset_filetype_t
;
1087 static ssize_t
cpuset_common_file_write(struct file
*file
, const char __user
*userbuf
,
1088 size_t nbytes
, loff_t
*unused_ppos
)
1090 struct cpuset
*cs
= __d_cs(file
->f_dentry
->d_parent
);
1091 struct cftype
*cft
= __d_cft(file
->f_dentry
);
1092 cpuset_filetype_t type
= cft
->private;
1094 char *pathbuf
= NULL
;
1097 /* Crude upper limit on largest legitimate cpulist user might write. */
1098 if (nbytes
> 100 + 6 * NR_CPUS
)
1101 /* +1 for nul-terminator */
1102 if ((buffer
= kmalloc(nbytes
+ 1, GFP_KERNEL
)) == 0)
1105 if (copy_from_user(buffer
, userbuf
, nbytes
)) {
1109 buffer
[nbytes
] = 0; /* nul-terminate */
1113 if (is_removed(cs
)) {
1120 retval
= update_cpumask(cs
, buffer
);
1123 retval
= update_nodemask(cs
, buffer
);
1125 case FILE_CPU_EXCLUSIVE
:
1126 retval
= update_flag(CS_CPU_EXCLUSIVE
, cs
, buffer
);
1128 case FILE_MEM_EXCLUSIVE
:
1129 retval
= update_flag(CS_MEM_EXCLUSIVE
, cs
, buffer
);
1131 case FILE_NOTIFY_ON_RELEASE
:
1132 retval
= update_flag(CS_NOTIFY_ON_RELEASE
, cs
, buffer
);
1134 case FILE_MEMORY_MIGRATE
:
1135 retval
= update_flag(CS_MEMORY_MIGRATE
, cs
, buffer
);
1137 case FILE_MEMORY_PRESSURE_ENABLED
:
1138 retval
= update_memory_pressure_enabled(cs
, buffer
);
1140 case FILE_MEMORY_PRESSURE
:
1144 retval
= attach_task(cs
, buffer
, &pathbuf
);
1155 cpuset_release_agent(pathbuf
);
1161 static ssize_t
cpuset_file_write(struct file
*file
, const char __user
*buf
,
1162 size_t nbytes
, loff_t
*ppos
)
1165 struct cftype
*cft
= __d_cft(file
->f_dentry
);
1169 /* special function ? */
1171 retval
= cft
->write(file
, buf
, nbytes
, ppos
);
1173 retval
= cpuset_common_file_write(file
, buf
, nbytes
, ppos
);
1179 * These ascii lists should be read in a single call, by using a user
1180 * buffer large enough to hold the entire map. If read in smaller
1181 * chunks, there is no guarantee of atomicity. Since the display format
1182 * used, list of ranges of sequential numbers, is variable length,
1183 * and since these maps can change value dynamically, one could read
1184 * gibberish by doing partial reads while a list was changing.
1185 * A single large read to a buffer that crosses a page boundary is
1186 * ok, because the result being copied to user land is not recomputed
1187 * across a page fault.
1190 static int cpuset_sprintf_cpulist(char *page
, struct cpuset
*cs
)
1194 down(&callback_sem
);
1195 mask
= cs
->cpus_allowed
;
1198 return cpulist_scnprintf(page
, PAGE_SIZE
, mask
);
1201 static int cpuset_sprintf_memlist(char *page
, struct cpuset
*cs
)
1205 down(&callback_sem
);
1206 mask
= cs
->mems_allowed
;
1209 return nodelist_scnprintf(page
, PAGE_SIZE
, mask
);
1212 static ssize_t
cpuset_common_file_read(struct file
*file
, char __user
*buf
,
1213 size_t nbytes
, loff_t
*ppos
)
1215 struct cftype
*cft
= __d_cft(file
->f_dentry
);
1216 struct cpuset
*cs
= __d_cs(file
->f_dentry
->d_parent
);
1217 cpuset_filetype_t type
= cft
->private;
1222 if (!(page
= (char *)__get_free_page(GFP_KERNEL
)))
1229 s
+= cpuset_sprintf_cpulist(s
, cs
);
1232 s
+= cpuset_sprintf_memlist(s
, cs
);
1234 case FILE_CPU_EXCLUSIVE
:
1235 *s
++ = is_cpu_exclusive(cs
) ? '1' : '0';
1237 case FILE_MEM_EXCLUSIVE
:
1238 *s
++ = is_mem_exclusive(cs
) ? '1' : '0';
1240 case FILE_NOTIFY_ON_RELEASE
:
1241 *s
++ = notify_on_release(cs
) ? '1' : '0';
1243 case FILE_MEMORY_MIGRATE
:
1244 *s
++ = is_memory_migrate(cs
) ? '1' : '0';
1246 case FILE_MEMORY_PRESSURE_ENABLED
:
1247 *s
++ = cpuset_memory_pressure_enabled
? '1' : '0';
1249 case FILE_MEMORY_PRESSURE
:
1250 s
+= sprintf(s
, "%d", fmeter_getrate(&cs
->fmeter
));
1258 retval
= simple_read_from_buffer(buf
, nbytes
, ppos
, page
, s
- page
);
1260 free_page((unsigned long)page
);
1264 static ssize_t
cpuset_file_read(struct file
*file
, char __user
*buf
, size_t nbytes
,
1268 struct cftype
*cft
= __d_cft(file
->f_dentry
);
1272 /* special function ? */
1274 retval
= cft
->read(file
, buf
, nbytes
, ppos
);
1276 retval
= cpuset_common_file_read(file
, buf
, nbytes
, ppos
);
1281 static int cpuset_file_open(struct inode
*inode
, struct file
*file
)
1286 err
= generic_file_open(inode
, file
);
1290 cft
= __d_cft(file
->f_dentry
);
1294 err
= cft
->open(inode
, file
);
1301 static int cpuset_file_release(struct inode
*inode
, struct file
*file
)
1303 struct cftype
*cft
= __d_cft(file
->f_dentry
);
1305 return cft
->release(inode
, file
);
1310 * cpuset_rename - Only allow simple rename of directories in place.
1312 static int cpuset_rename(struct inode
*old_dir
, struct dentry
*old_dentry
,
1313 struct inode
*new_dir
, struct dentry
*new_dentry
)
1315 if (!S_ISDIR(old_dentry
->d_inode
->i_mode
))
1317 if (new_dentry
->d_inode
)
1319 if (old_dir
!= new_dir
)
1321 return simple_rename(old_dir
, old_dentry
, new_dir
, new_dentry
);
1324 static struct file_operations cpuset_file_operations
= {
1325 .read
= cpuset_file_read
,
1326 .write
= cpuset_file_write
,
1327 .llseek
= generic_file_llseek
,
1328 .open
= cpuset_file_open
,
1329 .release
= cpuset_file_release
,
1332 static struct inode_operations cpuset_dir_inode_operations
= {
1333 .lookup
= simple_lookup
,
1334 .mkdir
= cpuset_mkdir
,
1335 .rmdir
= cpuset_rmdir
,
1336 .rename
= cpuset_rename
,
1339 static int cpuset_create_file(struct dentry
*dentry
, int mode
)
1341 struct inode
*inode
;
1345 if (dentry
->d_inode
)
1348 inode
= cpuset_new_inode(mode
);
1352 if (S_ISDIR(mode
)) {
1353 inode
->i_op
= &cpuset_dir_inode_operations
;
1354 inode
->i_fop
= &simple_dir_operations
;
1356 /* start off with i_nlink == 2 (for "." entry) */
1358 } else if (S_ISREG(mode
)) {
1360 inode
->i_fop
= &cpuset_file_operations
;
1363 d_instantiate(dentry
, inode
);
1364 dget(dentry
); /* Extra count - pin the dentry in core */
1369 * cpuset_create_dir - create a directory for an object.
1370 * cs: the cpuset we create the directory for.
1371 * It must have a valid ->parent field
1372 * And we are going to fill its ->dentry field.
1373 * name: The name to give to the cpuset directory. Will be copied.
1374 * mode: mode to set on new directory.
1377 static int cpuset_create_dir(struct cpuset
*cs
, const char *name
, int mode
)
1379 struct dentry
*dentry
= NULL
;
1380 struct dentry
*parent
;
1383 parent
= cs
->parent
->dentry
;
1384 dentry
= cpuset_get_dentry(parent
, name
);
1386 return PTR_ERR(dentry
);
1387 error
= cpuset_create_file(dentry
, S_IFDIR
| mode
);
1389 dentry
->d_fsdata
= cs
;
1390 parent
->d_inode
->i_nlink
++;
1391 cs
->dentry
= dentry
;
1398 static int cpuset_add_file(struct dentry
*dir
, const struct cftype
*cft
)
1400 struct dentry
*dentry
;
1403 down(&dir
->d_inode
->i_sem
);
1404 dentry
= cpuset_get_dentry(dir
, cft
->name
);
1405 if (!IS_ERR(dentry
)) {
1406 error
= cpuset_create_file(dentry
, 0644 | S_IFREG
);
1408 dentry
->d_fsdata
= (void *)cft
;
1411 error
= PTR_ERR(dentry
);
1412 up(&dir
->d_inode
->i_sem
);
1417 * Stuff for reading the 'tasks' file.
1419 * Reading this file can return large amounts of data if a cpuset has
1420 * *lots* of attached tasks. So it may need several calls to read(),
1421 * but we cannot guarantee that the information we produce is correct
1422 * unless we produce it entirely atomically.
1424 * Upon tasks file open(), a struct ctr_struct is allocated, that
1425 * will have a pointer to an array (also allocated here). The struct
1426 * ctr_struct * is stored in file->private_data. Its resources will
1427 * be freed by release() when the file is closed. The array is used
1428 * to sprintf the PIDs and then used by read().
1431 /* cpusets_tasks_read array */
1439 * Load into 'pidarray' up to 'npids' of the tasks using cpuset 'cs'.
1440 * Return actual number of pids loaded. No need to task_lock(p)
1441 * when reading out p->cpuset, as we don't really care if it changes
1442 * on the next cycle, and we are not going to try to dereference it.
1444 static inline int pid_array_load(pid_t
*pidarray
, int npids
, struct cpuset
*cs
)
1447 struct task_struct
*g
, *p
;
1449 read_lock(&tasklist_lock
);
1451 do_each_thread(g
, p
) {
1452 if (p
->cpuset
== cs
) {
1453 pidarray
[n
++] = p
->pid
;
1454 if (unlikely(n
== npids
))
1457 } while_each_thread(g
, p
);
1460 read_unlock(&tasklist_lock
);
1464 static int cmppid(const void *a
, const void *b
)
1466 return *(pid_t
*)a
- *(pid_t
*)b
;
1470 * Convert array 'a' of 'npids' pid_t's to a string of newline separated
1471 * decimal pids in 'buf'. Don't write more than 'sz' chars, but return
1472 * count 'cnt' of how many chars would be written if buf were large enough.
1474 static int pid_array_to_buf(char *buf
, int sz
, pid_t
*a
, int npids
)
1479 for (i
= 0; i
< npids
; i
++)
1480 cnt
+= snprintf(buf
+ cnt
, max(sz
- cnt
, 0), "%d\n", a
[i
]);
1485 * Handle an open on 'tasks' file. Prepare a buffer listing the
1486 * process id's of tasks currently attached to the cpuset being opened.
1488 * Does not require any specific cpuset semaphores, and does not take any.
1490 static int cpuset_tasks_open(struct inode
*unused
, struct file
*file
)
1492 struct cpuset
*cs
= __d_cs(file
->f_dentry
->d_parent
);
1493 struct ctr_struct
*ctr
;
1498 if (!(file
->f_mode
& FMODE_READ
))
1501 ctr
= kmalloc(sizeof(*ctr
), GFP_KERNEL
);
1506 * If cpuset gets more users after we read count, we won't have
1507 * enough space - tough. This race is indistinguishable to the
1508 * caller from the case that the additional cpuset users didn't
1509 * show up until sometime later on.
1511 npids
= atomic_read(&cs
->count
);
1512 pidarray
= kmalloc(npids
* sizeof(pid_t
), GFP_KERNEL
);
1516 npids
= pid_array_load(pidarray
, npids
, cs
);
1517 sort(pidarray
, npids
, sizeof(pid_t
), cmppid
, NULL
);
1519 /* Call pid_array_to_buf() twice, first just to get bufsz */
1520 ctr
->bufsz
= pid_array_to_buf(&c
, sizeof(c
), pidarray
, npids
) + 1;
1521 ctr
->buf
= kmalloc(ctr
->bufsz
, GFP_KERNEL
);
1524 ctr
->bufsz
= pid_array_to_buf(ctr
->buf
, ctr
->bufsz
, pidarray
, npids
);
1527 file
->private_data
= ctr
;
1538 static ssize_t
cpuset_tasks_read(struct file
*file
, char __user
*buf
,
1539 size_t nbytes
, loff_t
*ppos
)
1541 struct ctr_struct
*ctr
= file
->private_data
;
1543 if (*ppos
+ nbytes
> ctr
->bufsz
)
1544 nbytes
= ctr
->bufsz
- *ppos
;
1545 if (copy_to_user(buf
, ctr
->buf
+ *ppos
, nbytes
))
1551 static int cpuset_tasks_release(struct inode
*unused_inode
, struct file
*file
)
1553 struct ctr_struct
*ctr
;
1555 if (file
->f_mode
& FMODE_READ
) {
1556 ctr
= file
->private_data
;
1564 * for the common functions, 'private' gives the type of file
1567 static struct cftype cft_tasks
= {
1569 .open
= cpuset_tasks_open
,
1570 .read
= cpuset_tasks_read
,
1571 .release
= cpuset_tasks_release
,
1572 .private = FILE_TASKLIST
,
1575 static struct cftype cft_cpus
= {
1577 .private = FILE_CPULIST
,
1580 static struct cftype cft_mems
= {
1582 .private = FILE_MEMLIST
,
1585 static struct cftype cft_cpu_exclusive
= {
1586 .name
= "cpu_exclusive",
1587 .private = FILE_CPU_EXCLUSIVE
,
1590 static struct cftype cft_mem_exclusive
= {
1591 .name
= "mem_exclusive",
1592 .private = FILE_MEM_EXCLUSIVE
,
1595 static struct cftype cft_notify_on_release
= {
1596 .name
= "notify_on_release",
1597 .private = FILE_NOTIFY_ON_RELEASE
,
1600 static struct cftype cft_memory_migrate
= {
1601 .name
= "memory_migrate",
1602 .private = FILE_MEMORY_MIGRATE
,
1605 static struct cftype cft_memory_pressure_enabled
= {
1606 .name
= "memory_pressure_enabled",
1607 .private = FILE_MEMORY_PRESSURE_ENABLED
,
1610 static struct cftype cft_memory_pressure
= {
1611 .name
= "memory_pressure",
1612 .private = FILE_MEMORY_PRESSURE
,
1615 static int cpuset_populate_dir(struct dentry
*cs_dentry
)
1619 if ((err
= cpuset_add_file(cs_dentry
, &cft_cpus
)) < 0)
1621 if ((err
= cpuset_add_file(cs_dentry
, &cft_mems
)) < 0)
1623 if ((err
= cpuset_add_file(cs_dentry
, &cft_cpu_exclusive
)) < 0)
1625 if ((err
= cpuset_add_file(cs_dentry
, &cft_mem_exclusive
)) < 0)
1627 if ((err
= cpuset_add_file(cs_dentry
, &cft_notify_on_release
)) < 0)
1629 if ((err
= cpuset_add_file(cs_dentry
, &cft_memory_migrate
)) < 0)
1631 if ((err
= cpuset_add_file(cs_dentry
, &cft_memory_pressure
)) < 0)
1633 if ((err
= cpuset_add_file(cs_dentry
, &cft_tasks
)) < 0)
1639 * cpuset_create - create a cpuset
1640 * parent: cpuset that will be parent of the new cpuset.
1641 * name: name of the new cpuset. Will be strcpy'ed.
1642 * mode: mode to set on new inode
1644 * Must be called with the semaphore on the parent inode held
1647 static long cpuset_create(struct cpuset
*parent
, const char *name
, int mode
)
1652 cs
= kmalloc(sizeof(*cs
), GFP_KERNEL
);
1657 cpuset_update_task_memory_state();
1659 if (notify_on_release(parent
))
1660 set_bit(CS_NOTIFY_ON_RELEASE
, &cs
->flags
);
1661 cs
->cpus_allowed
= CPU_MASK_NONE
;
1662 cs
->mems_allowed
= NODE_MASK_NONE
;
1663 atomic_set(&cs
->count
, 0);
1664 INIT_LIST_HEAD(&cs
->sibling
);
1665 INIT_LIST_HEAD(&cs
->children
);
1666 atomic_inc(&cpuset_mems_generation
);
1667 cs
->mems_generation
= atomic_read(&cpuset_mems_generation
);
1668 fmeter_init(&cs
->fmeter
);
1670 cs
->parent
= parent
;
1672 down(&callback_sem
);
1673 list_add(&cs
->sibling
, &cs
->parent
->children
);
1674 number_of_cpusets
++;
1677 err
= cpuset_create_dir(cs
, name
, mode
);
1682 * Release manage_sem before cpuset_populate_dir() because it
1683 * will down() this new directory's i_sem and if we race with
1684 * another mkdir, we might deadlock.
1688 err
= cpuset_populate_dir(cs
->dentry
);
1689 /* If err < 0, we have a half-filled directory - oh well ;) */
1692 list_del(&cs
->sibling
);
1698 static int cpuset_mkdir(struct inode
*dir
, struct dentry
*dentry
, int mode
)
1700 struct cpuset
*c_parent
= dentry
->d_parent
->d_fsdata
;
1702 /* the vfs holds inode->i_sem already */
1703 return cpuset_create(c_parent
, dentry
->d_name
.name
, mode
| S_IFDIR
);
1706 static int cpuset_rmdir(struct inode
*unused_dir
, struct dentry
*dentry
)
1708 struct cpuset
*cs
= dentry
->d_fsdata
;
1710 struct cpuset
*parent
;
1711 char *pathbuf
= NULL
;
1713 /* the vfs holds both inode->i_sem already */
1716 cpuset_update_task_memory_state();
1717 if (atomic_read(&cs
->count
) > 0) {
1721 if (!list_empty(&cs
->children
)) {
1725 parent
= cs
->parent
;
1726 down(&callback_sem
);
1727 set_bit(CS_REMOVED
, &cs
->flags
);
1728 if (is_cpu_exclusive(cs
))
1729 update_cpu_domains(cs
);
1730 list_del(&cs
->sibling
); /* delete my sibling from parent->children */
1731 spin_lock(&cs
->dentry
->d_lock
);
1732 d
= dget(cs
->dentry
);
1734 spin_unlock(&d
->d_lock
);
1735 cpuset_d_remove_dir(d
);
1737 number_of_cpusets
--;
1739 if (list_empty(&parent
->children
))
1740 check_for_release(parent
, &pathbuf
);
1742 cpuset_release_agent(pathbuf
);
1747 * cpuset_init - initialize cpusets at system boot
1749 * Description: Initialize top_cpuset and the cpuset internal file system,
1752 int __init
cpuset_init(void)
1754 struct dentry
*root
;
1757 top_cpuset
.cpus_allowed
= CPU_MASK_ALL
;
1758 top_cpuset
.mems_allowed
= NODE_MASK_ALL
;
1760 fmeter_init(&top_cpuset
.fmeter
);
1761 atomic_inc(&cpuset_mems_generation
);
1762 top_cpuset
.mems_generation
= atomic_read(&cpuset_mems_generation
);
1764 init_task
.cpuset
= &top_cpuset
;
1766 err
= register_filesystem(&cpuset_fs_type
);
1769 cpuset_mount
= kern_mount(&cpuset_fs_type
);
1770 if (IS_ERR(cpuset_mount
)) {
1771 printk(KERN_ERR
"cpuset: could not mount!\n");
1772 err
= PTR_ERR(cpuset_mount
);
1773 cpuset_mount
= NULL
;
1776 root
= cpuset_mount
->mnt_sb
->s_root
;
1777 root
->d_fsdata
= &top_cpuset
;
1778 root
->d_inode
->i_nlink
++;
1779 top_cpuset
.dentry
= root
;
1780 root
->d_inode
->i_op
= &cpuset_dir_inode_operations
;
1781 number_of_cpusets
= 1;
1782 err
= cpuset_populate_dir(root
);
1783 /* memory_pressure_enabled is in root cpuset only */
1785 err
= cpuset_add_file(root
, &cft_memory_pressure_enabled
);
1791 * cpuset_init_smp - initialize cpus_allowed
1793 * Description: Finish top cpuset after cpu, node maps are initialized
1796 void __init
cpuset_init_smp(void)
1798 top_cpuset
.cpus_allowed
= cpu_online_map
;
1799 top_cpuset
.mems_allowed
= node_online_map
;
1803 * cpuset_fork - attach newly forked task to its parents cpuset.
1804 * @tsk: pointer to task_struct of forking parent process.
1806 * Description: A task inherits its parent's cpuset at fork().
1808 * A pointer to the shared cpuset was automatically copied in fork.c
1809 * by dup_task_struct(). However, we ignore that copy, since it was
1810 * not made under the protection of task_lock(), so might no longer be
1811 * a valid cpuset pointer. attach_task() might have already changed
1812 * current->cpuset, allowing the previously referenced cpuset to
1813 * be removed and freed. Instead, we task_lock(current) and copy
1814 * its present value of current->cpuset for our freshly forked child.
1816 * At the point that cpuset_fork() is called, 'current' is the parent
1817 * task, and the passed argument 'child' points to the child task.
1820 void cpuset_fork(struct task_struct
*child
)
1823 child
->cpuset
= current
->cpuset
;
1824 atomic_inc(&child
->cpuset
->count
);
1825 task_unlock(current
);
1829 * cpuset_exit - detach cpuset from exiting task
1830 * @tsk: pointer to task_struct of exiting process
1832 * Description: Detach cpuset from @tsk and release it.
1834 * Note that cpusets marked notify_on_release force every task in
1835 * them to take the global manage_sem semaphore when exiting.
1836 * This could impact scaling on very large systems. Be reluctant to
1837 * use notify_on_release cpusets where very high task exit scaling
1838 * is required on large systems.
1840 * Don't even think about derefencing 'cs' after the cpuset use count
1841 * goes to zero, except inside a critical section guarded by manage_sem
1842 * or callback_sem. Otherwise a zero cpuset use count is a license to
1843 * any other task to nuke the cpuset immediately, via cpuset_rmdir().
1845 * This routine has to take manage_sem, not callback_sem, because
1846 * it is holding that semaphore while calling check_for_release(),
1847 * which calls kmalloc(), so can't be called holding callback__sem().
1849 * We don't need to task_lock() this reference to tsk->cpuset,
1850 * because tsk is already marked PF_EXITING, so attach_task() won't
1851 * mess with it, or task is a failed fork, never visible to attach_task.
1854 void cpuset_exit(struct task_struct
*tsk
)
1861 if (notify_on_release(cs
)) {
1862 char *pathbuf
= NULL
;
1865 if (atomic_dec_and_test(&cs
->count
))
1866 check_for_release(cs
, &pathbuf
);
1868 cpuset_release_agent(pathbuf
);
1870 atomic_dec(&cs
->count
);
1875 * cpuset_cpus_allowed - return cpus_allowed mask from a tasks cpuset.
1876 * @tsk: pointer to task_struct from which to obtain cpuset->cpus_allowed.
1878 * Description: Returns the cpumask_t cpus_allowed of the cpuset
1879 * attached to the specified @tsk. Guaranteed to return some non-empty
1880 * subset of cpu_online_map, even if this means going outside the
1884 cpumask_t
cpuset_cpus_allowed(struct task_struct
*tsk
)
1888 down(&callback_sem
);
1890 guarantee_online_cpus(tsk
->cpuset
, &mask
);
1897 void cpuset_init_current_mems_allowed(void)
1899 current
->mems_allowed
= NODE_MASK_ALL
;
1903 * cpuset_mems_allowed - return mems_allowed mask from a tasks cpuset.
1904 * @tsk: pointer to task_struct from which to obtain cpuset->mems_allowed.
1906 * Description: Returns the nodemask_t mems_allowed of the cpuset
1907 * attached to the specified @tsk. Guaranteed to return some non-empty
1908 * subset of node_online_map, even if this means going outside the
1912 nodemask_t
cpuset_mems_allowed(struct task_struct
*tsk
)
1916 down(&callback_sem
);
1918 guarantee_online_mems(tsk
->cpuset
, &mask
);
1926 * cpuset_zonelist_valid_mems_allowed - check zonelist vs. curremt mems_allowed
1927 * @zl: the zonelist to be checked
1929 * Are any of the nodes on zonelist zl allowed in current->mems_allowed?
1931 int cpuset_zonelist_valid_mems_allowed(struct zonelist
*zl
)
1935 for (i
= 0; zl
->zones
[i
]; i
++) {
1936 int nid
= zl
->zones
[i
]->zone_pgdat
->node_id
;
1938 if (node_isset(nid
, current
->mems_allowed
))
1945 * nearest_exclusive_ancestor() - Returns the nearest mem_exclusive
1946 * ancestor to the specified cpuset. Call holding callback_sem.
1947 * If no ancestor is mem_exclusive (an unusual configuration), then
1948 * returns the root cpuset.
1950 static const struct cpuset
*nearest_exclusive_ancestor(const struct cpuset
*cs
)
1952 while (!is_mem_exclusive(cs
) && cs
->parent
)
1958 * cpuset_zone_allowed - Can we allocate memory on zone z's memory node?
1959 * @z: is this zone on an allowed node?
1960 * @gfp_mask: memory allocation flags (we use __GFP_HARDWALL)
1962 * If we're in interrupt, yes, we can always allocate. If zone
1963 * z's node is in our tasks mems_allowed, yes. If it's not a
1964 * __GFP_HARDWALL request and this zone's nodes is in the nearest
1965 * mem_exclusive cpuset ancestor to this tasks cpuset, yes.
1968 * GFP_USER allocations are marked with the __GFP_HARDWALL bit,
1969 * and do not allow allocations outside the current tasks cpuset.
1970 * GFP_KERNEL allocations are not so marked, so can escape to the
1971 * nearest mem_exclusive ancestor cpuset.
1973 * Scanning up parent cpusets requires callback_sem. The __alloc_pages()
1974 * routine only calls here with __GFP_HARDWALL bit _not_ set if
1975 * it's a GFP_KERNEL allocation, and all nodes in the current tasks
1976 * mems_allowed came up empty on the first pass over the zonelist.
1977 * So only GFP_KERNEL allocations, if all nodes in the cpuset are
1978 * short of memory, might require taking the callback_sem semaphore.
1980 * The first loop over the zonelist in mm/page_alloc.c:__alloc_pages()
1981 * calls here with __GFP_HARDWALL always set in gfp_mask, enforcing
1982 * hardwall cpusets - no allocation on a node outside the cpuset is
1983 * allowed (unless in interrupt, of course).
1985 * The second loop doesn't even call here for GFP_ATOMIC requests
1986 * (if the __alloc_pages() local variable 'wait' is set). That check
1987 * and the checks below have the combined affect in the second loop of
1988 * the __alloc_pages() routine that:
1989 * in_interrupt - any node ok (current task context irrelevant)
1990 * GFP_ATOMIC - any node ok
1991 * GFP_KERNEL - any node in enclosing mem_exclusive cpuset ok
1992 * GFP_USER - only nodes in current tasks mems allowed ok.
1995 int __cpuset_zone_allowed(struct zone
*z
, gfp_t gfp_mask
)
1997 int node
; /* node that zone z is on */
1998 const struct cpuset
*cs
; /* current cpuset ancestors */
1999 int allowed
= 1; /* is allocation in zone z allowed? */
2003 node
= z
->zone_pgdat
->node_id
;
2004 if (node_isset(node
, current
->mems_allowed
))
2006 if (gfp_mask
& __GFP_HARDWALL
) /* If hardwall request, stop here */
2009 if (current
->flags
& PF_EXITING
) /* Let dying task have memory */
2012 /* Not hardwall and node outside mems_allowed: scan up cpusets */
2013 down(&callback_sem
);
2016 cs
= nearest_exclusive_ancestor(current
->cpuset
);
2017 task_unlock(current
);
2019 allowed
= node_isset(node
, cs
->mems_allowed
);
2025 * cpuset_excl_nodes_overlap - Do we overlap @p's mem_exclusive ancestors?
2026 * @p: pointer to task_struct of some other task.
2028 * Description: Return true if the nearest mem_exclusive ancestor
2029 * cpusets of tasks @p and current overlap. Used by oom killer to
2030 * determine if task @p's memory usage might impact the memory
2031 * available to the current task.
2033 * Acquires callback_sem - not suitable for calling from a fast path.
2036 int cpuset_excl_nodes_overlap(const struct task_struct
*p
)
2038 const struct cpuset
*cs1
, *cs2
; /* my and p's cpuset ancestors */
2039 int overlap
= 0; /* do cpusets overlap? */
2041 down(&callback_sem
);
2044 if (current
->flags
& PF_EXITING
) {
2045 task_unlock(current
);
2048 cs1
= nearest_exclusive_ancestor(current
->cpuset
);
2049 task_unlock(current
);
2051 task_lock((struct task_struct
*)p
);
2052 if (p
->flags
& PF_EXITING
) {
2053 task_unlock((struct task_struct
*)p
);
2056 cs2
= nearest_exclusive_ancestor(p
->cpuset
);
2057 task_unlock((struct task_struct
*)p
);
2059 overlap
= nodes_intersects(cs1
->mems_allowed
, cs2
->mems_allowed
);
2067 * Collection of memory_pressure is suppressed unless
2068 * this flag is enabled by writing "1" to the special
2069 * cpuset file 'memory_pressure_enabled' in the root cpuset.
2072 int cpuset_memory_pressure_enabled __read_mostly
;
2075 * cpuset_memory_pressure_bump - keep stats of per-cpuset reclaims.
2077 * Keep a running average of the rate of synchronous (direct)
2078 * page reclaim efforts initiated by tasks in each cpuset.
2080 * This represents the rate at which some task in the cpuset
2081 * ran low on memory on all nodes it was allowed to use, and
2082 * had to enter the kernels page reclaim code in an effort to
2083 * create more free memory by tossing clean pages or swapping
2084 * or writing dirty pages.
2086 * Display to user space in the per-cpuset read-only file
2087 * "memory_pressure". Value displayed is an integer
2088 * representing the recent rate of entry into the synchronous
2089 * (direct) page reclaim by any task attached to the cpuset.
2092 void __cpuset_memory_pressure_bump(void)
2097 cs
= current
->cpuset
;
2098 fmeter_markevent(&cs
->fmeter
);
2099 task_unlock(current
);
2103 * proc_cpuset_show()
2104 * - Print tasks cpuset path into seq_file.
2105 * - Used for /proc/<pid>/cpuset.
2106 * - No need to task_lock(tsk) on this tsk->cpuset reference, as it
2107 * doesn't really matter if tsk->cpuset changes after we read it,
2108 * and we take manage_sem, keeping attach_task() from changing it
2112 static int proc_cpuset_show(struct seq_file
*m
, void *v
)
2115 struct task_struct
*tsk
;
2119 buf
= kmalloc(PAGE_SIZE
, GFP_KERNEL
);
2131 retval
= cpuset_path(cs
, buf
, PAGE_SIZE
);
2142 static int cpuset_open(struct inode
*inode
, struct file
*file
)
2144 struct task_struct
*tsk
= PROC_I(inode
)->task
;
2145 return single_open(file
, proc_cpuset_show
, tsk
);
2148 struct file_operations proc_cpuset_operations
= {
2149 .open
= cpuset_open
,
2151 .llseek
= seq_lseek
,
2152 .release
= single_release
,
2155 /* Display task cpus_allowed, mems_allowed in /proc/<pid>/status file. */
2156 char *cpuset_task_status_allowed(struct task_struct
*task
, char *buffer
)
2158 buffer
+= sprintf(buffer
, "Cpus_allowed:\t");
2159 buffer
+= cpumask_scnprintf(buffer
, PAGE_SIZE
, task
->cpus_allowed
);
2160 buffer
+= sprintf(buffer
, "\n");
2161 buffer
+= sprintf(buffer
, "Mems_allowed:\t");
2162 buffer
+= nodemask_scnprintf(buffer
, PAGE_SIZE
, task
->mems_allowed
);
2163 buffer
+= sprintf(buffer
, "\n");