[PATCH] cpuset memory spread basic implementation
[linux-2.6/pdupreez.git] / kernel / cpuset.c
blob38f18b33de6c9a3a47d54a7980c7f5b051595fd4
1 /*
2 * kernel/cpuset.c
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>
28 #include <linux/fs.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>
35 #include <linux/mm.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/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 <linux/mutex.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 __read_mostly;
66 /* See "Frequency meter" comments, below. */
68 struct fmeter {
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 */
75 struct cpuset {
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.
99 int mems_generation;
101 struct fmeter fmeter; /* memory_pressure filter */
104 /* bits in struct cpuset flags field */
105 typedef enum {
106 CS_CPU_EXCLUSIVE,
107 CS_MEM_EXCLUSIVE,
108 CS_MEMORY_MIGRATE,
109 CS_REMOVED,
110 CS_NOTIFY_ON_RELEASE,
111 CS_SPREAD_PAGE,
112 CS_SPREAD_SLAB,
113 } cpuset_flagbits_t;
115 /* convenient tests for these bits */
116 static inline int is_cpu_exclusive(const struct cpuset *cs)
118 return test_bit(CS_CPU_EXCLUSIVE, &cs->flags);
121 static inline int is_mem_exclusive(const struct cpuset *cs)
123 return test_bit(CS_MEM_EXCLUSIVE, &cs->flags);
126 static inline int is_removed(const struct cpuset *cs)
128 return test_bit(CS_REMOVED, &cs->flags);
131 static inline int notify_on_release(const struct cpuset *cs)
133 return test_bit(CS_NOTIFY_ON_RELEASE, &cs->flags);
136 static inline int is_memory_migrate(const struct cpuset *cs)
138 return test_bit(CS_MEMORY_MIGRATE, &cs->flags);
141 static inline int is_spread_page(const struct cpuset *cs)
143 return test_bit(CS_SPREAD_PAGE, &cs->flags);
146 static inline int is_spread_slab(const struct cpuset *cs)
148 return test_bit(CS_SPREAD_SLAB, &cs->flags);
152 * Increment this atomic integer everytime any cpuset changes its
153 * mems_allowed value. Users of cpusets can track this generation
154 * number, and avoid having to lock and reload mems_allowed unless
155 * the cpuset they're using changes generation.
157 * A single, global generation is needed because attach_task() could
158 * reattach a task to a different cpuset, which must not have its
159 * generation numbers aliased with those of that tasks previous cpuset.
161 * Generations are needed for mems_allowed because one task cannot
162 * modify anothers memory placement. So we must enable every task,
163 * on every visit to __alloc_pages(), to efficiently check whether
164 * its current->cpuset->mems_allowed has changed, requiring an update
165 * of its current->mems_allowed.
167 static atomic_t cpuset_mems_generation = ATOMIC_INIT(1);
169 static struct cpuset top_cpuset = {
170 .flags = ((1 << CS_CPU_EXCLUSIVE) | (1 << CS_MEM_EXCLUSIVE)),
171 .cpus_allowed = CPU_MASK_ALL,
172 .mems_allowed = NODE_MASK_ALL,
173 .count = ATOMIC_INIT(0),
174 .sibling = LIST_HEAD_INIT(top_cpuset.sibling),
175 .children = LIST_HEAD_INIT(top_cpuset.children),
178 static struct vfsmount *cpuset_mount;
179 static struct super_block *cpuset_sb;
182 * We have two global cpuset mutexes below. They can nest.
183 * It is ok to first take manage_mutex, then nest callback_mutex. We also
184 * require taking task_lock() when dereferencing a tasks cpuset pointer.
185 * See "The task_lock() exception", at the end of this comment.
187 * A task must hold both mutexes to modify cpusets. If a task
188 * holds manage_mutex, then it blocks others wanting that mutex,
189 * ensuring that it is the only task able to also acquire callback_mutex
190 * and be able to modify cpusets. It can perform various checks on
191 * the cpuset structure first, knowing nothing will change. It can
192 * also allocate memory while just holding manage_mutex. While it is
193 * performing these checks, various callback routines can briefly
194 * acquire callback_mutex to query cpusets. Once it is ready to make
195 * the changes, it takes callback_mutex, blocking everyone else.
197 * Calls to the kernel memory allocator can not be made while holding
198 * callback_mutex, as that would risk double tripping on callback_mutex
199 * from one of the callbacks into the cpuset code from within
200 * __alloc_pages().
202 * If a task is only holding callback_mutex, then it has read-only
203 * access to cpusets.
205 * The task_struct fields mems_allowed and mems_generation may only
206 * be accessed in the context of that task, so require no locks.
208 * Any task can increment and decrement the count field without lock.
209 * So in general, code holding manage_mutex or callback_mutex can't rely
210 * on the count field not changing. However, if the count goes to
211 * zero, then only attach_task(), which holds both mutexes, can
212 * increment it again. Because a count of zero means that no tasks
213 * are currently attached, therefore there is no way a task attached
214 * to that cpuset can fork (the other way to increment the count).
215 * So code holding manage_mutex or callback_mutex can safely assume that
216 * if the count is zero, it will stay zero. Similarly, if a task
217 * holds manage_mutex or callback_mutex on a cpuset with zero count, it
218 * knows that the cpuset won't be removed, as cpuset_rmdir() needs
219 * both of those mutexes.
221 * The cpuset_common_file_write handler for operations that modify
222 * the cpuset hierarchy holds manage_mutex across the entire operation,
223 * single threading all such cpuset modifications across the system.
225 * The cpuset_common_file_read() handlers only hold callback_mutex across
226 * small pieces of code, such as when reading out possibly multi-word
227 * cpumasks and nodemasks.
229 * The fork and exit callbacks cpuset_fork() and cpuset_exit(), don't
230 * (usually) take either mutex. These are the two most performance
231 * critical pieces of code here. The exception occurs on cpuset_exit(),
232 * when a task in a notify_on_release cpuset exits. Then manage_mutex
233 * is taken, and if the cpuset count is zero, a usermode call made
234 * to /sbin/cpuset_release_agent with the name of the cpuset (path
235 * relative to the root of cpuset file system) as the argument.
237 * A cpuset can only be deleted if both its 'count' of using tasks
238 * is zero, and its list of 'children' cpusets is empty. Since all
239 * tasks in the system use _some_ cpuset, and since there is always at
240 * least one task in the system (init, pid == 1), therefore, top_cpuset
241 * always has either children cpusets and/or using tasks. So we don't
242 * need a special hack to ensure that top_cpuset cannot be deleted.
244 * The above "Tale of Two Semaphores" would be complete, but for:
246 * The task_lock() exception
248 * The need for this exception arises from the action of attach_task(),
249 * which overwrites one tasks cpuset pointer with another. It does
250 * so using both mutexes, however there are several performance
251 * critical places that need to reference task->cpuset without the
252 * expense of grabbing a system global mutex. Therefore except as
253 * noted below, when dereferencing or, as in attach_task(), modifying
254 * a tasks cpuset pointer we use task_lock(), which acts on a spinlock
255 * (task->alloc_lock) already in the task_struct routinely used for
256 * such matters.
258 * P.S. One more locking exception. RCU is used to guard the
259 * update of a tasks cpuset pointer by attach_task() and the
260 * access of task->cpuset->mems_generation via that pointer in
261 * the routine cpuset_update_task_memory_state().
264 static DEFINE_MUTEX(manage_mutex);
265 static DEFINE_MUTEX(callback_mutex);
268 * A couple of forward declarations required, due to cyclic reference loop:
269 * cpuset_mkdir -> cpuset_create -> cpuset_populate_dir -> cpuset_add_file
270 * -> cpuset_create_file -> cpuset_dir_inode_operations -> cpuset_mkdir.
273 static int cpuset_mkdir(struct inode *dir, struct dentry *dentry, int mode);
274 static int cpuset_rmdir(struct inode *unused_dir, struct dentry *dentry);
276 static struct backing_dev_info cpuset_backing_dev_info = {
277 .ra_pages = 0, /* No readahead */
278 .capabilities = BDI_CAP_NO_ACCT_DIRTY | BDI_CAP_NO_WRITEBACK,
281 static struct inode *cpuset_new_inode(mode_t mode)
283 struct inode *inode = new_inode(cpuset_sb);
285 if (inode) {
286 inode->i_mode = mode;
287 inode->i_uid = current->fsuid;
288 inode->i_gid = current->fsgid;
289 inode->i_blksize = PAGE_CACHE_SIZE;
290 inode->i_blocks = 0;
291 inode->i_atime = inode->i_mtime = inode->i_ctime = CURRENT_TIME;
292 inode->i_mapping->backing_dev_info = &cpuset_backing_dev_info;
294 return inode;
297 static void cpuset_diput(struct dentry *dentry, struct inode *inode)
299 /* is dentry a directory ? if so, kfree() associated cpuset */
300 if (S_ISDIR(inode->i_mode)) {
301 struct cpuset *cs = dentry->d_fsdata;
302 BUG_ON(!(is_removed(cs)));
303 kfree(cs);
305 iput(inode);
308 static struct dentry_operations cpuset_dops = {
309 .d_iput = cpuset_diput,
312 static struct dentry *cpuset_get_dentry(struct dentry *parent, const char *name)
314 struct dentry *d = lookup_one_len(name, parent, strlen(name));
315 if (!IS_ERR(d))
316 d->d_op = &cpuset_dops;
317 return d;
320 static void remove_dir(struct dentry *d)
322 struct dentry *parent = dget(d->d_parent);
324 d_delete(d);
325 simple_rmdir(parent->d_inode, d);
326 dput(parent);
330 * NOTE : the dentry must have been dget()'ed
332 static void cpuset_d_remove_dir(struct dentry *dentry)
334 struct list_head *node;
336 spin_lock(&dcache_lock);
337 node = dentry->d_subdirs.next;
338 while (node != &dentry->d_subdirs) {
339 struct dentry *d = list_entry(node, struct dentry, d_u.d_child);
340 list_del_init(node);
341 if (d->d_inode) {
342 d = dget_locked(d);
343 spin_unlock(&dcache_lock);
344 d_delete(d);
345 simple_unlink(dentry->d_inode, d);
346 dput(d);
347 spin_lock(&dcache_lock);
349 node = dentry->d_subdirs.next;
351 list_del_init(&dentry->d_u.d_child);
352 spin_unlock(&dcache_lock);
353 remove_dir(dentry);
356 static struct super_operations cpuset_ops = {
357 .statfs = simple_statfs,
358 .drop_inode = generic_delete_inode,
361 static int cpuset_fill_super(struct super_block *sb, void *unused_data,
362 int unused_silent)
364 struct inode *inode;
365 struct dentry *root;
367 sb->s_blocksize = PAGE_CACHE_SIZE;
368 sb->s_blocksize_bits = PAGE_CACHE_SHIFT;
369 sb->s_magic = CPUSET_SUPER_MAGIC;
370 sb->s_op = &cpuset_ops;
371 cpuset_sb = sb;
373 inode = cpuset_new_inode(S_IFDIR | S_IRUGO | S_IXUGO | S_IWUSR);
374 if (inode) {
375 inode->i_op = &simple_dir_inode_operations;
376 inode->i_fop = &simple_dir_operations;
377 /* directories start off with i_nlink == 2 (for "." entry) */
378 inode->i_nlink++;
379 } else {
380 return -ENOMEM;
383 root = d_alloc_root(inode);
384 if (!root) {
385 iput(inode);
386 return -ENOMEM;
388 sb->s_root = root;
389 return 0;
392 static struct super_block *cpuset_get_sb(struct file_system_type *fs_type,
393 int flags, const char *unused_dev_name,
394 void *data)
396 return get_sb_single(fs_type, flags, data, cpuset_fill_super);
399 static struct file_system_type cpuset_fs_type = {
400 .name = "cpuset",
401 .get_sb = cpuset_get_sb,
402 .kill_sb = kill_litter_super,
405 /* struct cftype:
407 * The files in the cpuset filesystem mostly have a very simple read/write
408 * handling, some common function will take care of it. Nevertheless some cases
409 * (read tasks) are special and therefore I define this structure for every
410 * kind of file.
413 * When reading/writing to a file:
414 * - the cpuset to use in file->f_dentry->d_parent->d_fsdata
415 * - the 'cftype' of the file is file->f_dentry->d_fsdata
418 struct cftype {
419 char *name;
420 int private;
421 int (*open) (struct inode *inode, struct file *file);
422 ssize_t (*read) (struct file *file, char __user *buf, size_t nbytes,
423 loff_t *ppos);
424 int (*write) (struct file *file, const char __user *buf, size_t nbytes,
425 loff_t *ppos);
426 int (*release) (struct inode *inode, struct file *file);
429 static inline struct cpuset *__d_cs(struct dentry *dentry)
431 return dentry->d_fsdata;
434 static inline struct cftype *__d_cft(struct dentry *dentry)
436 return dentry->d_fsdata;
440 * Call with manage_mutex held. Writes path of cpuset into buf.
441 * Returns 0 on success, -errno on error.
444 static int cpuset_path(const struct cpuset *cs, char *buf, int buflen)
446 char *start;
448 start = buf + buflen;
450 *--start = '\0';
451 for (;;) {
452 int len = cs->dentry->d_name.len;
453 if ((start -= len) < buf)
454 return -ENAMETOOLONG;
455 memcpy(start, cs->dentry->d_name.name, len);
456 cs = cs->parent;
457 if (!cs)
458 break;
459 if (!cs->parent)
460 continue;
461 if (--start < buf)
462 return -ENAMETOOLONG;
463 *start = '/';
465 memmove(buf, start, buf + buflen - start);
466 return 0;
470 * Notify userspace when a cpuset is released, by running
471 * /sbin/cpuset_release_agent with the name of the cpuset (path
472 * relative to the root of cpuset file system) as the argument.
474 * Most likely, this user command will try to rmdir this cpuset.
476 * This races with the possibility that some other task will be
477 * attached to this cpuset before it is removed, or that some other
478 * user task will 'mkdir' a child cpuset of this cpuset. That's ok.
479 * The presumed 'rmdir' will fail quietly if this cpuset is no longer
480 * unused, and this cpuset will be reprieved from its death sentence,
481 * to continue to serve a useful existence. Next time it's released,
482 * we will get notified again, if it still has 'notify_on_release' set.
484 * The final arg to call_usermodehelper() is 0, which means don't
485 * wait. The separate /sbin/cpuset_release_agent task is forked by
486 * call_usermodehelper(), then control in this thread returns here,
487 * without waiting for the release agent task. We don't bother to
488 * wait because the caller of this routine has no use for the exit
489 * status of the /sbin/cpuset_release_agent task, so no sense holding
490 * our caller up for that.
492 * When we had only one cpuset mutex, we had to call this
493 * without holding it, to avoid deadlock when call_usermodehelper()
494 * allocated memory. With two locks, we could now call this while
495 * holding manage_mutex, but we still don't, so as to minimize
496 * the time manage_mutex is held.
499 static void cpuset_release_agent(const char *pathbuf)
501 char *argv[3], *envp[3];
502 int i;
504 if (!pathbuf)
505 return;
507 i = 0;
508 argv[i++] = "/sbin/cpuset_release_agent";
509 argv[i++] = (char *)pathbuf;
510 argv[i] = NULL;
512 i = 0;
513 /* minimal command environment */
514 envp[i++] = "HOME=/";
515 envp[i++] = "PATH=/sbin:/bin:/usr/sbin:/usr/bin";
516 envp[i] = NULL;
518 call_usermodehelper(argv[0], argv, envp, 0);
519 kfree(pathbuf);
523 * Either cs->count of using tasks transitioned to zero, or the
524 * cs->children list of child cpusets just became empty. If this
525 * cs is notify_on_release() and now both the user count is zero and
526 * the list of children is empty, prepare cpuset path in a kmalloc'd
527 * buffer, to be returned via ppathbuf, so that the caller can invoke
528 * cpuset_release_agent() with it later on, once manage_mutex is dropped.
529 * Call here with manage_mutex held.
531 * This check_for_release() routine is responsible for kmalloc'ing
532 * pathbuf. The above cpuset_release_agent() is responsible for
533 * kfree'ing pathbuf. The caller of these routines is responsible
534 * for providing a pathbuf pointer, initialized to NULL, then
535 * calling check_for_release() with manage_mutex held and the address
536 * of the pathbuf pointer, then dropping manage_mutex, then calling
537 * cpuset_release_agent() with pathbuf, as set by check_for_release().
540 static void check_for_release(struct cpuset *cs, char **ppathbuf)
542 if (notify_on_release(cs) && atomic_read(&cs->count) == 0 &&
543 list_empty(&cs->children)) {
544 char *buf;
546 buf = kmalloc(PAGE_SIZE, GFP_KERNEL);
547 if (!buf)
548 return;
549 if (cpuset_path(cs, buf, PAGE_SIZE) < 0)
550 kfree(buf);
551 else
552 *ppathbuf = buf;
557 * Return in *pmask the portion of a cpusets's cpus_allowed that
558 * are online. If none are online, walk up the cpuset hierarchy
559 * until we find one that does have some online cpus. If we get
560 * all the way to the top and still haven't found any online cpus,
561 * return cpu_online_map. Or if passed a NULL cs from an exit'ing
562 * task, return cpu_online_map.
564 * One way or another, we guarantee to return some non-empty subset
565 * of cpu_online_map.
567 * Call with callback_mutex held.
570 static void guarantee_online_cpus(const struct cpuset *cs, cpumask_t *pmask)
572 while (cs && !cpus_intersects(cs->cpus_allowed, cpu_online_map))
573 cs = cs->parent;
574 if (cs)
575 cpus_and(*pmask, cs->cpus_allowed, cpu_online_map);
576 else
577 *pmask = cpu_online_map;
578 BUG_ON(!cpus_intersects(*pmask, cpu_online_map));
582 * Return in *pmask the portion of a cpusets's mems_allowed that
583 * are online. If none are online, walk up the cpuset hierarchy
584 * until we find one that does have some online mems. If we get
585 * all the way to the top and still haven't found any online mems,
586 * return node_online_map.
588 * One way or another, we guarantee to return some non-empty subset
589 * of node_online_map.
591 * Call with callback_mutex held.
594 static void guarantee_online_mems(const struct cpuset *cs, nodemask_t *pmask)
596 while (cs && !nodes_intersects(cs->mems_allowed, node_online_map))
597 cs = cs->parent;
598 if (cs)
599 nodes_and(*pmask, cs->mems_allowed, node_online_map);
600 else
601 *pmask = node_online_map;
602 BUG_ON(!nodes_intersects(*pmask, node_online_map));
606 * cpuset_update_task_memory_state - update task memory placement
608 * If the current tasks cpusets mems_allowed changed behind our
609 * backs, update current->mems_allowed, mems_generation and task NUMA
610 * mempolicy to the new value.
612 * Task mempolicy is updated by rebinding it relative to the
613 * current->cpuset if a task has its memory placement changed.
614 * Do not call this routine if in_interrupt().
616 * Call without callback_mutex or task_lock() held. May be called
617 * with or without manage_mutex held. Doesn't need task_lock to guard
618 * against another task changing a non-NULL cpuset pointer to NULL,
619 * as that is only done by a task on itself, and if the current task
620 * is here, it is not simultaneously in the exit code NULL'ing its
621 * cpuset pointer. This routine also might acquire callback_mutex and
622 * current->mm->mmap_sem during call.
624 * Reading current->cpuset->mems_generation doesn't need task_lock
625 * to guard the current->cpuset derefence, because it is guarded
626 * from concurrent freeing of current->cpuset by attach_task(),
627 * using RCU.
629 * The rcu_dereference() is technically probably not needed,
630 * as I don't actually mind if I see a new cpuset pointer but
631 * an old value of mems_generation. However this really only
632 * matters on alpha systems using cpusets heavily. If I dropped
633 * that rcu_dereference(), it would save them a memory barrier.
634 * For all other arch's, rcu_dereference is a no-op anyway, and for
635 * alpha systems not using cpusets, another planned optimization,
636 * avoiding the rcu critical section for tasks in the root cpuset
637 * which is statically allocated, so can't vanish, will make this
638 * irrelevant. Better to use RCU as intended, than to engage in
639 * some cute trick to save a memory barrier that is impossible to
640 * test, for alpha systems using cpusets heavily, which might not
641 * even exist.
643 * This routine is needed to update the per-task mems_allowed data,
644 * within the tasks context, when it is trying to allocate memory
645 * (in various mm/mempolicy.c routines) and notices that some other
646 * task has been modifying its cpuset.
649 void cpuset_update_task_memory_state(void)
651 int my_cpusets_mem_gen;
652 struct task_struct *tsk = current;
653 struct cpuset *cs;
655 if (tsk->cpuset == &top_cpuset) {
656 /* Don't need rcu for top_cpuset. It's never freed. */
657 my_cpusets_mem_gen = top_cpuset.mems_generation;
658 } else {
659 rcu_read_lock();
660 cs = rcu_dereference(tsk->cpuset);
661 my_cpusets_mem_gen = cs->mems_generation;
662 rcu_read_unlock();
665 if (my_cpusets_mem_gen != tsk->cpuset_mems_generation) {
666 mutex_lock(&callback_mutex);
667 task_lock(tsk);
668 cs = tsk->cpuset; /* Maybe changed when task not locked */
669 guarantee_online_mems(cs, &tsk->mems_allowed);
670 tsk->cpuset_mems_generation = cs->mems_generation;
671 if (is_spread_page(cs))
672 tsk->flags |= PF_SPREAD_PAGE;
673 else
674 tsk->flags &= ~PF_SPREAD_PAGE;
675 if (is_spread_slab(cs))
676 tsk->flags |= PF_SPREAD_SLAB;
677 else
678 tsk->flags &= ~PF_SPREAD_SLAB;
679 task_unlock(tsk);
680 mutex_unlock(&callback_mutex);
681 mpol_rebind_task(tsk, &tsk->mems_allowed);
686 * is_cpuset_subset(p, q) - Is cpuset p a subset of cpuset q?
688 * One cpuset is a subset of another if all its allowed CPUs and
689 * Memory Nodes are a subset of the other, and its exclusive flags
690 * are only set if the other's are set. Call holding manage_mutex.
693 static int is_cpuset_subset(const struct cpuset *p, const struct cpuset *q)
695 return cpus_subset(p->cpus_allowed, q->cpus_allowed) &&
696 nodes_subset(p->mems_allowed, q->mems_allowed) &&
697 is_cpu_exclusive(p) <= is_cpu_exclusive(q) &&
698 is_mem_exclusive(p) <= is_mem_exclusive(q);
702 * validate_change() - Used to validate that any proposed cpuset change
703 * follows the structural rules for cpusets.
705 * If we replaced the flag and mask values of the current cpuset
706 * (cur) with those values in the trial cpuset (trial), would
707 * our various subset and exclusive rules still be valid? Presumes
708 * manage_mutex held.
710 * 'cur' is the address of an actual, in-use cpuset. Operations
711 * such as list traversal that depend on the actual address of the
712 * cpuset in the list must use cur below, not trial.
714 * 'trial' is the address of bulk structure copy of cur, with
715 * perhaps one or more of the fields cpus_allowed, mems_allowed,
716 * or flags changed to new, trial values.
718 * Return 0 if valid, -errno if not.
721 static int validate_change(const struct cpuset *cur, const struct cpuset *trial)
723 struct cpuset *c, *par;
725 /* Each of our child cpusets must be a subset of us */
726 list_for_each_entry(c, &cur->children, sibling) {
727 if (!is_cpuset_subset(c, trial))
728 return -EBUSY;
731 /* Remaining checks don't apply to root cpuset */
732 if ((par = cur->parent) == NULL)
733 return 0;
735 /* We must be a subset of our parent cpuset */
736 if (!is_cpuset_subset(trial, par))
737 return -EACCES;
739 /* If either I or some sibling (!= me) is exclusive, we can't overlap */
740 list_for_each_entry(c, &par->children, sibling) {
741 if ((is_cpu_exclusive(trial) || is_cpu_exclusive(c)) &&
742 c != cur &&
743 cpus_intersects(trial->cpus_allowed, c->cpus_allowed))
744 return -EINVAL;
745 if ((is_mem_exclusive(trial) || is_mem_exclusive(c)) &&
746 c != cur &&
747 nodes_intersects(trial->mems_allowed, c->mems_allowed))
748 return -EINVAL;
751 return 0;
755 * For a given cpuset cur, partition the system as follows
756 * a. All cpus in the parent cpuset's cpus_allowed that are not part of any
757 * exclusive child cpusets
758 * b. All cpus in the current cpuset's cpus_allowed that are not part of any
759 * exclusive child cpusets
760 * Build these two partitions by calling partition_sched_domains
762 * Call with manage_mutex held. May nest a call to the
763 * lock_cpu_hotplug()/unlock_cpu_hotplug() pair.
766 static void update_cpu_domains(struct cpuset *cur)
768 struct cpuset *c, *par = cur->parent;
769 cpumask_t pspan, cspan;
771 if (par == NULL || cpus_empty(cur->cpus_allowed))
772 return;
775 * Get all cpus from parent's cpus_allowed not part of exclusive
776 * children
778 pspan = par->cpus_allowed;
779 list_for_each_entry(c, &par->children, sibling) {
780 if (is_cpu_exclusive(c))
781 cpus_andnot(pspan, pspan, c->cpus_allowed);
783 if (is_removed(cur) || !is_cpu_exclusive(cur)) {
784 cpus_or(pspan, pspan, cur->cpus_allowed);
785 if (cpus_equal(pspan, cur->cpus_allowed))
786 return;
787 cspan = CPU_MASK_NONE;
788 } else {
789 if (cpus_empty(pspan))
790 return;
791 cspan = cur->cpus_allowed;
793 * Get all cpus from current cpuset's cpus_allowed not part
794 * of exclusive children
796 list_for_each_entry(c, &cur->children, sibling) {
797 if (is_cpu_exclusive(c))
798 cpus_andnot(cspan, cspan, c->cpus_allowed);
802 lock_cpu_hotplug();
803 partition_sched_domains(&pspan, &cspan);
804 unlock_cpu_hotplug();
808 * Call with manage_mutex held. May take callback_mutex during call.
811 static int update_cpumask(struct cpuset *cs, char *buf)
813 struct cpuset trialcs;
814 int retval, cpus_unchanged;
816 trialcs = *cs;
817 retval = cpulist_parse(buf, trialcs.cpus_allowed);
818 if (retval < 0)
819 return retval;
820 cpus_and(trialcs.cpus_allowed, trialcs.cpus_allowed, cpu_online_map);
821 if (cpus_empty(trialcs.cpus_allowed))
822 return -ENOSPC;
823 retval = validate_change(cs, &trialcs);
824 if (retval < 0)
825 return retval;
826 cpus_unchanged = cpus_equal(cs->cpus_allowed, trialcs.cpus_allowed);
827 mutex_lock(&callback_mutex);
828 cs->cpus_allowed = trialcs.cpus_allowed;
829 mutex_unlock(&callback_mutex);
830 if (is_cpu_exclusive(cs) && !cpus_unchanged)
831 update_cpu_domains(cs);
832 return 0;
836 * Handle user request to change the 'mems' memory placement
837 * of a cpuset. Needs to validate the request, update the
838 * cpusets mems_allowed and mems_generation, and for each
839 * task in the cpuset, rebind any vma mempolicies and if
840 * the cpuset is marked 'memory_migrate', migrate the tasks
841 * pages to the new memory.
843 * Call with manage_mutex held. May take callback_mutex during call.
844 * Will take tasklist_lock, scan tasklist for tasks in cpuset cs,
845 * lock each such tasks mm->mmap_sem, scan its vma's and rebind
846 * their mempolicies to the cpusets new mems_allowed.
849 static int update_nodemask(struct cpuset *cs, char *buf)
851 struct cpuset trialcs;
852 nodemask_t oldmem;
853 struct task_struct *g, *p;
854 struct mm_struct **mmarray;
855 int i, n, ntasks;
856 int migrate;
857 int fudge;
858 int retval;
860 trialcs = *cs;
861 retval = nodelist_parse(buf, trialcs.mems_allowed);
862 if (retval < 0)
863 goto done;
864 nodes_and(trialcs.mems_allowed, trialcs.mems_allowed, node_online_map);
865 oldmem = cs->mems_allowed;
866 if (nodes_equal(oldmem, trialcs.mems_allowed)) {
867 retval = 0; /* Too easy - nothing to do */
868 goto done;
870 if (nodes_empty(trialcs.mems_allowed)) {
871 retval = -ENOSPC;
872 goto done;
874 retval = validate_change(cs, &trialcs);
875 if (retval < 0)
876 goto done;
878 mutex_lock(&callback_mutex);
879 cs->mems_allowed = trialcs.mems_allowed;
880 cs->mems_generation = atomic_inc_return(&cpuset_mems_generation);
881 mutex_unlock(&callback_mutex);
883 set_cpuset_being_rebound(cs); /* causes mpol_copy() rebind */
885 fudge = 10; /* spare mmarray[] slots */
886 fudge += cpus_weight(cs->cpus_allowed); /* imagine one fork-bomb/cpu */
887 retval = -ENOMEM;
890 * Allocate mmarray[] to hold mm reference for each task
891 * in cpuset cs. Can't kmalloc GFP_KERNEL while holding
892 * tasklist_lock. We could use GFP_ATOMIC, but with a
893 * few more lines of code, we can retry until we get a big
894 * enough mmarray[] w/o using GFP_ATOMIC.
896 while (1) {
897 ntasks = atomic_read(&cs->count); /* guess */
898 ntasks += fudge;
899 mmarray = kmalloc(ntasks * sizeof(*mmarray), GFP_KERNEL);
900 if (!mmarray)
901 goto done;
902 write_lock_irq(&tasklist_lock); /* block fork */
903 if (atomic_read(&cs->count) <= ntasks)
904 break; /* got enough */
905 write_unlock_irq(&tasklist_lock); /* try again */
906 kfree(mmarray);
909 n = 0;
911 /* Load up mmarray[] with mm reference for each task in cpuset. */
912 do_each_thread(g, p) {
913 struct mm_struct *mm;
915 if (n >= ntasks) {
916 printk(KERN_WARNING
917 "Cpuset mempolicy rebind incomplete.\n");
918 continue;
920 if (p->cpuset != cs)
921 continue;
922 mm = get_task_mm(p);
923 if (!mm)
924 continue;
925 mmarray[n++] = mm;
926 } while_each_thread(g, p);
927 write_unlock_irq(&tasklist_lock);
930 * Now that we've dropped the tasklist spinlock, we can
931 * rebind the vma mempolicies of each mm in mmarray[] to their
932 * new cpuset, and release that mm. The mpol_rebind_mm()
933 * call takes mmap_sem, which we couldn't take while holding
934 * tasklist_lock. Forks can happen again now - the mpol_copy()
935 * cpuset_being_rebound check will catch such forks, and rebind
936 * their vma mempolicies too. Because we still hold the global
937 * cpuset manage_mutex, we know that no other rebind effort will
938 * be contending for the global variable cpuset_being_rebound.
939 * It's ok if we rebind the same mm twice; mpol_rebind_mm()
940 * is idempotent. Also migrate pages in each mm to new nodes.
942 migrate = is_memory_migrate(cs);
943 for (i = 0; i < n; i++) {
944 struct mm_struct *mm = mmarray[i];
946 mpol_rebind_mm(mm, &cs->mems_allowed);
947 if (migrate) {
948 do_migrate_pages(mm, &oldmem, &cs->mems_allowed,
949 MPOL_MF_MOVE_ALL);
951 mmput(mm);
954 /* We're done rebinding vma's to this cpusets new mems_allowed. */
955 kfree(mmarray);
956 set_cpuset_being_rebound(NULL);
957 retval = 0;
958 done:
959 return retval;
963 * Call with manage_mutex held.
966 static int update_memory_pressure_enabled(struct cpuset *cs, char *buf)
968 if (simple_strtoul(buf, NULL, 10) != 0)
969 cpuset_memory_pressure_enabled = 1;
970 else
971 cpuset_memory_pressure_enabled = 0;
972 return 0;
976 * update_flag - read a 0 or a 1 in a file and update associated flag
977 * bit: the bit to update (CS_CPU_EXCLUSIVE, CS_MEM_EXCLUSIVE,
978 * CS_NOTIFY_ON_RELEASE, CS_MEMORY_MIGRATE,
979 * CS_SPREAD_PAGE, CS_SPREAD_SLAB)
980 * cs: the cpuset to update
981 * buf: the buffer where we read the 0 or 1
983 * Call with manage_mutex held.
986 static int update_flag(cpuset_flagbits_t bit, struct cpuset *cs, char *buf)
988 int turning_on;
989 struct cpuset trialcs;
990 int err, cpu_exclusive_changed;
992 turning_on = (simple_strtoul(buf, NULL, 10) != 0);
994 trialcs = *cs;
995 if (turning_on)
996 set_bit(bit, &trialcs.flags);
997 else
998 clear_bit(bit, &trialcs.flags);
1000 err = validate_change(cs, &trialcs);
1001 if (err < 0)
1002 return err;
1003 cpu_exclusive_changed =
1004 (is_cpu_exclusive(cs) != is_cpu_exclusive(&trialcs));
1005 mutex_lock(&callback_mutex);
1006 if (turning_on)
1007 set_bit(bit, &cs->flags);
1008 else
1009 clear_bit(bit, &cs->flags);
1010 mutex_unlock(&callback_mutex);
1012 if (cpu_exclusive_changed)
1013 update_cpu_domains(cs);
1014 return 0;
1018 * Frequency meter - How fast is some event occuring?
1020 * These routines manage a digitally filtered, constant time based,
1021 * event frequency meter. There are four routines:
1022 * fmeter_init() - initialize a frequency meter.
1023 * fmeter_markevent() - called each time the event happens.
1024 * fmeter_getrate() - returns the recent rate of such events.
1025 * fmeter_update() - internal routine used to update fmeter.
1027 * A common data structure is passed to each of these routines,
1028 * which is used to keep track of the state required to manage the
1029 * frequency meter and its digital filter.
1031 * The filter works on the number of events marked per unit time.
1032 * The filter is single-pole low-pass recursive (IIR). The time unit
1033 * is 1 second. Arithmetic is done using 32-bit integers scaled to
1034 * simulate 3 decimal digits of precision (multiplied by 1000).
1036 * With an FM_COEF of 933, and a time base of 1 second, the filter
1037 * has a half-life of 10 seconds, meaning that if the events quit
1038 * happening, then the rate returned from the fmeter_getrate()
1039 * will be cut in half each 10 seconds, until it converges to zero.
1041 * It is not worth doing a real infinitely recursive filter. If more
1042 * than FM_MAXTICKS ticks have elapsed since the last filter event,
1043 * just compute FM_MAXTICKS ticks worth, by which point the level
1044 * will be stable.
1046 * Limit the count of unprocessed events to FM_MAXCNT, so as to avoid
1047 * arithmetic overflow in the fmeter_update() routine.
1049 * Given the simple 32 bit integer arithmetic used, this meter works
1050 * best for reporting rates between one per millisecond (msec) and
1051 * one per 32 (approx) seconds. At constant rates faster than one
1052 * per msec it maxes out at values just under 1,000,000. At constant
1053 * rates between one per msec, and one per second it will stabilize
1054 * to a value N*1000, where N is the rate of events per second.
1055 * At constant rates between one per second and one per 32 seconds,
1056 * it will be choppy, moving up on the seconds that have an event,
1057 * and then decaying until the next event. At rates slower than
1058 * about one in 32 seconds, it decays all the way back to zero between
1059 * each event.
1062 #define FM_COEF 933 /* coefficient for half-life of 10 secs */
1063 #define FM_MAXTICKS ((time_t)99) /* useless computing more ticks than this */
1064 #define FM_MAXCNT 1000000 /* limit cnt to avoid overflow */
1065 #define FM_SCALE 1000 /* faux fixed point scale */
1067 /* Initialize a frequency meter */
1068 static void fmeter_init(struct fmeter *fmp)
1070 fmp->cnt = 0;
1071 fmp->val = 0;
1072 fmp->time = 0;
1073 spin_lock_init(&fmp->lock);
1076 /* Internal meter update - process cnt events and update value */
1077 static void fmeter_update(struct fmeter *fmp)
1079 time_t now = get_seconds();
1080 time_t ticks = now - fmp->time;
1082 if (ticks == 0)
1083 return;
1085 ticks = min(FM_MAXTICKS, ticks);
1086 while (ticks-- > 0)
1087 fmp->val = (FM_COEF * fmp->val) / FM_SCALE;
1088 fmp->time = now;
1090 fmp->val += ((FM_SCALE - FM_COEF) * fmp->cnt) / FM_SCALE;
1091 fmp->cnt = 0;
1094 /* Process any previous ticks, then bump cnt by one (times scale). */
1095 static void fmeter_markevent(struct fmeter *fmp)
1097 spin_lock(&fmp->lock);
1098 fmeter_update(fmp);
1099 fmp->cnt = min(FM_MAXCNT, fmp->cnt + FM_SCALE);
1100 spin_unlock(&fmp->lock);
1103 /* Process any previous ticks, then return current value. */
1104 static int fmeter_getrate(struct fmeter *fmp)
1106 int val;
1108 spin_lock(&fmp->lock);
1109 fmeter_update(fmp);
1110 val = fmp->val;
1111 spin_unlock(&fmp->lock);
1112 return val;
1116 * Attack task specified by pid in 'pidbuf' to cpuset 'cs', possibly
1117 * writing the path of the old cpuset in 'ppathbuf' if it needs to be
1118 * notified on release.
1120 * Call holding manage_mutex. May take callback_mutex and task_lock of
1121 * the task 'pid' during call.
1124 static int attach_task(struct cpuset *cs, char *pidbuf, char **ppathbuf)
1126 pid_t pid;
1127 struct task_struct *tsk;
1128 struct cpuset *oldcs;
1129 cpumask_t cpus;
1130 nodemask_t from, to;
1131 struct mm_struct *mm;
1133 if (sscanf(pidbuf, "%d", &pid) != 1)
1134 return -EIO;
1135 if (cpus_empty(cs->cpus_allowed) || nodes_empty(cs->mems_allowed))
1136 return -ENOSPC;
1138 if (pid) {
1139 read_lock(&tasklist_lock);
1141 tsk = find_task_by_pid(pid);
1142 if (!tsk || tsk->flags & PF_EXITING) {
1143 read_unlock(&tasklist_lock);
1144 return -ESRCH;
1147 get_task_struct(tsk);
1148 read_unlock(&tasklist_lock);
1150 if ((current->euid) && (current->euid != tsk->uid)
1151 && (current->euid != tsk->suid)) {
1152 put_task_struct(tsk);
1153 return -EACCES;
1155 } else {
1156 tsk = current;
1157 get_task_struct(tsk);
1160 mutex_lock(&callback_mutex);
1162 task_lock(tsk);
1163 oldcs = tsk->cpuset;
1164 if (!oldcs) {
1165 task_unlock(tsk);
1166 mutex_unlock(&callback_mutex);
1167 put_task_struct(tsk);
1168 return -ESRCH;
1170 atomic_inc(&cs->count);
1171 rcu_assign_pointer(tsk->cpuset, cs);
1172 task_unlock(tsk);
1174 guarantee_online_cpus(cs, &cpus);
1175 set_cpus_allowed(tsk, cpus);
1177 from = oldcs->mems_allowed;
1178 to = cs->mems_allowed;
1180 mutex_unlock(&callback_mutex);
1182 mm = get_task_mm(tsk);
1183 if (mm) {
1184 mpol_rebind_mm(mm, &to);
1185 mmput(mm);
1188 if (is_memory_migrate(cs))
1189 do_migrate_pages(tsk->mm, &from, &to, MPOL_MF_MOVE_ALL);
1190 put_task_struct(tsk);
1191 synchronize_rcu();
1192 if (atomic_dec_and_test(&oldcs->count))
1193 check_for_release(oldcs, ppathbuf);
1194 return 0;
1197 /* The various types of files and directories in a cpuset file system */
1199 typedef enum {
1200 FILE_ROOT,
1201 FILE_DIR,
1202 FILE_MEMORY_MIGRATE,
1203 FILE_CPULIST,
1204 FILE_MEMLIST,
1205 FILE_CPU_EXCLUSIVE,
1206 FILE_MEM_EXCLUSIVE,
1207 FILE_NOTIFY_ON_RELEASE,
1208 FILE_MEMORY_PRESSURE_ENABLED,
1209 FILE_MEMORY_PRESSURE,
1210 FILE_SPREAD_PAGE,
1211 FILE_SPREAD_SLAB,
1212 FILE_TASKLIST,
1213 } cpuset_filetype_t;
1215 static ssize_t cpuset_common_file_write(struct file *file, const char __user *userbuf,
1216 size_t nbytes, loff_t *unused_ppos)
1218 struct cpuset *cs = __d_cs(file->f_dentry->d_parent);
1219 struct cftype *cft = __d_cft(file->f_dentry);
1220 cpuset_filetype_t type = cft->private;
1221 char *buffer;
1222 char *pathbuf = NULL;
1223 int retval = 0;
1225 /* Crude upper limit on largest legitimate cpulist user might write. */
1226 if (nbytes > 100 + 6 * NR_CPUS)
1227 return -E2BIG;
1229 /* +1 for nul-terminator */
1230 if ((buffer = kmalloc(nbytes + 1, GFP_KERNEL)) == 0)
1231 return -ENOMEM;
1233 if (copy_from_user(buffer, userbuf, nbytes)) {
1234 retval = -EFAULT;
1235 goto out1;
1237 buffer[nbytes] = 0; /* nul-terminate */
1239 mutex_lock(&manage_mutex);
1241 if (is_removed(cs)) {
1242 retval = -ENODEV;
1243 goto out2;
1246 switch (type) {
1247 case FILE_CPULIST:
1248 retval = update_cpumask(cs, buffer);
1249 break;
1250 case FILE_MEMLIST:
1251 retval = update_nodemask(cs, buffer);
1252 break;
1253 case FILE_CPU_EXCLUSIVE:
1254 retval = update_flag(CS_CPU_EXCLUSIVE, cs, buffer);
1255 break;
1256 case FILE_MEM_EXCLUSIVE:
1257 retval = update_flag(CS_MEM_EXCLUSIVE, cs, buffer);
1258 break;
1259 case FILE_NOTIFY_ON_RELEASE:
1260 retval = update_flag(CS_NOTIFY_ON_RELEASE, cs, buffer);
1261 break;
1262 case FILE_MEMORY_MIGRATE:
1263 retval = update_flag(CS_MEMORY_MIGRATE, cs, buffer);
1264 break;
1265 case FILE_MEMORY_PRESSURE_ENABLED:
1266 retval = update_memory_pressure_enabled(cs, buffer);
1267 break;
1268 case FILE_MEMORY_PRESSURE:
1269 retval = -EACCES;
1270 break;
1271 case FILE_SPREAD_PAGE:
1272 retval = update_flag(CS_SPREAD_PAGE, cs, buffer);
1273 cs->mems_generation = atomic_inc_return(&cpuset_mems_generation);
1274 break;
1275 case FILE_SPREAD_SLAB:
1276 retval = update_flag(CS_SPREAD_SLAB, cs, buffer);
1277 cs->mems_generation = atomic_inc_return(&cpuset_mems_generation);
1278 break;
1279 case FILE_TASKLIST:
1280 retval = attach_task(cs, buffer, &pathbuf);
1281 break;
1282 default:
1283 retval = -EINVAL;
1284 goto out2;
1287 if (retval == 0)
1288 retval = nbytes;
1289 out2:
1290 mutex_unlock(&manage_mutex);
1291 cpuset_release_agent(pathbuf);
1292 out1:
1293 kfree(buffer);
1294 return retval;
1297 static ssize_t cpuset_file_write(struct file *file, const char __user *buf,
1298 size_t nbytes, loff_t *ppos)
1300 ssize_t retval = 0;
1301 struct cftype *cft = __d_cft(file->f_dentry);
1302 if (!cft)
1303 return -ENODEV;
1305 /* special function ? */
1306 if (cft->write)
1307 retval = cft->write(file, buf, nbytes, ppos);
1308 else
1309 retval = cpuset_common_file_write(file, buf, nbytes, ppos);
1311 return retval;
1315 * These ascii lists should be read in a single call, by using a user
1316 * buffer large enough to hold the entire map. If read in smaller
1317 * chunks, there is no guarantee of atomicity. Since the display format
1318 * used, list of ranges of sequential numbers, is variable length,
1319 * and since these maps can change value dynamically, one could read
1320 * gibberish by doing partial reads while a list was changing.
1321 * A single large read to a buffer that crosses a page boundary is
1322 * ok, because the result being copied to user land is not recomputed
1323 * across a page fault.
1326 static int cpuset_sprintf_cpulist(char *page, struct cpuset *cs)
1328 cpumask_t mask;
1330 mutex_lock(&callback_mutex);
1331 mask = cs->cpus_allowed;
1332 mutex_unlock(&callback_mutex);
1334 return cpulist_scnprintf(page, PAGE_SIZE, mask);
1337 static int cpuset_sprintf_memlist(char *page, struct cpuset *cs)
1339 nodemask_t mask;
1341 mutex_lock(&callback_mutex);
1342 mask = cs->mems_allowed;
1343 mutex_unlock(&callback_mutex);
1345 return nodelist_scnprintf(page, PAGE_SIZE, mask);
1348 static ssize_t cpuset_common_file_read(struct file *file, char __user *buf,
1349 size_t nbytes, loff_t *ppos)
1351 struct cftype *cft = __d_cft(file->f_dentry);
1352 struct cpuset *cs = __d_cs(file->f_dentry->d_parent);
1353 cpuset_filetype_t type = cft->private;
1354 char *page;
1355 ssize_t retval = 0;
1356 char *s;
1358 if (!(page = (char *)__get_free_page(GFP_KERNEL)))
1359 return -ENOMEM;
1361 s = page;
1363 switch (type) {
1364 case FILE_CPULIST:
1365 s += cpuset_sprintf_cpulist(s, cs);
1366 break;
1367 case FILE_MEMLIST:
1368 s += cpuset_sprintf_memlist(s, cs);
1369 break;
1370 case FILE_CPU_EXCLUSIVE:
1371 *s++ = is_cpu_exclusive(cs) ? '1' : '0';
1372 break;
1373 case FILE_MEM_EXCLUSIVE:
1374 *s++ = is_mem_exclusive(cs) ? '1' : '0';
1375 break;
1376 case FILE_NOTIFY_ON_RELEASE:
1377 *s++ = notify_on_release(cs) ? '1' : '0';
1378 break;
1379 case FILE_MEMORY_MIGRATE:
1380 *s++ = is_memory_migrate(cs) ? '1' : '0';
1381 break;
1382 case FILE_MEMORY_PRESSURE_ENABLED:
1383 *s++ = cpuset_memory_pressure_enabled ? '1' : '0';
1384 break;
1385 case FILE_MEMORY_PRESSURE:
1386 s += sprintf(s, "%d", fmeter_getrate(&cs->fmeter));
1387 break;
1388 case FILE_SPREAD_PAGE:
1389 *s++ = is_spread_page(cs) ? '1' : '0';
1390 break;
1391 case FILE_SPREAD_SLAB:
1392 *s++ = is_spread_slab(cs) ? '1' : '0';
1393 break;
1394 default:
1395 retval = -EINVAL;
1396 goto out;
1398 *s++ = '\n';
1400 retval = simple_read_from_buffer(buf, nbytes, ppos, page, s - page);
1401 out:
1402 free_page((unsigned long)page);
1403 return retval;
1406 static ssize_t cpuset_file_read(struct file *file, char __user *buf, size_t nbytes,
1407 loff_t *ppos)
1409 ssize_t retval = 0;
1410 struct cftype *cft = __d_cft(file->f_dentry);
1411 if (!cft)
1412 return -ENODEV;
1414 /* special function ? */
1415 if (cft->read)
1416 retval = cft->read(file, buf, nbytes, ppos);
1417 else
1418 retval = cpuset_common_file_read(file, buf, nbytes, ppos);
1420 return retval;
1423 static int cpuset_file_open(struct inode *inode, struct file *file)
1425 int err;
1426 struct cftype *cft;
1428 err = generic_file_open(inode, file);
1429 if (err)
1430 return err;
1432 cft = __d_cft(file->f_dentry);
1433 if (!cft)
1434 return -ENODEV;
1435 if (cft->open)
1436 err = cft->open(inode, file);
1437 else
1438 err = 0;
1440 return err;
1443 static int cpuset_file_release(struct inode *inode, struct file *file)
1445 struct cftype *cft = __d_cft(file->f_dentry);
1446 if (cft->release)
1447 return cft->release(inode, file);
1448 return 0;
1452 * cpuset_rename - Only allow simple rename of directories in place.
1454 static int cpuset_rename(struct inode *old_dir, struct dentry *old_dentry,
1455 struct inode *new_dir, struct dentry *new_dentry)
1457 if (!S_ISDIR(old_dentry->d_inode->i_mode))
1458 return -ENOTDIR;
1459 if (new_dentry->d_inode)
1460 return -EEXIST;
1461 if (old_dir != new_dir)
1462 return -EIO;
1463 return simple_rename(old_dir, old_dentry, new_dir, new_dentry);
1466 static struct file_operations cpuset_file_operations = {
1467 .read = cpuset_file_read,
1468 .write = cpuset_file_write,
1469 .llseek = generic_file_llseek,
1470 .open = cpuset_file_open,
1471 .release = cpuset_file_release,
1474 static struct inode_operations cpuset_dir_inode_operations = {
1475 .lookup = simple_lookup,
1476 .mkdir = cpuset_mkdir,
1477 .rmdir = cpuset_rmdir,
1478 .rename = cpuset_rename,
1481 static int cpuset_create_file(struct dentry *dentry, int mode)
1483 struct inode *inode;
1485 if (!dentry)
1486 return -ENOENT;
1487 if (dentry->d_inode)
1488 return -EEXIST;
1490 inode = cpuset_new_inode(mode);
1491 if (!inode)
1492 return -ENOMEM;
1494 if (S_ISDIR(mode)) {
1495 inode->i_op = &cpuset_dir_inode_operations;
1496 inode->i_fop = &simple_dir_operations;
1498 /* start off with i_nlink == 2 (for "." entry) */
1499 inode->i_nlink++;
1500 } else if (S_ISREG(mode)) {
1501 inode->i_size = 0;
1502 inode->i_fop = &cpuset_file_operations;
1505 d_instantiate(dentry, inode);
1506 dget(dentry); /* Extra count - pin the dentry in core */
1507 return 0;
1511 * cpuset_create_dir - create a directory for an object.
1512 * cs: the cpuset we create the directory for.
1513 * It must have a valid ->parent field
1514 * And we are going to fill its ->dentry field.
1515 * name: The name to give to the cpuset directory. Will be copied.
1516 * mode: mode to set on new directory.
1519 static int cpuset_create_dir(struct cpuset *cs, const char *name, int mode)
1521 struct dentry *dentry = NULL;
1522 struct dentry *parent;
1523 int error = 0;
1525 parent = cs->parent->dentry;
1526 dentry = cpuset_get_dentry(parent, name);
1527 if (IS_ERR(dentry))
1528 return PTR_ERR(dentry);
1529 error = cpuset_create_file(dentry, S_IFDIR | mode);
1530 if (!error) {
1531 dentry->d_fsdata = cs;
1532 parent->d_inode->i_nlink++;
1533 cs->dentry = dentry;
1535 dput(dentry);
1537 return error;
1540 static int cpuset_add_file(struct dentry *dir, const struct cftype *cft)
1542 struct dentry *dentry;
1543 int error;
1545 mutex_lock(&dir->d_inode->i_mutex);
1546 dentry = cpuset_get_dentry(dir, cft->name);
1547 if (!IS_ERR(dentry)) {
1548 error = cpuset_create_file(dentry, 0644 | S_IFREG);
1549 if (!error)
1550 dentry->d_fsdata = (void *)cft;
1551 dput(dentry);
1552 } else
1553 error = PTR_ERR(dentry);
1554 mutex_unlock(&dir->d_inode->i_mutex);
1555 return error;
1559 * Stuff for reading the 'tasks' file.
1561 * Reading this file can return large amounts of data if a cpuset has
1562 * *lots* of attached tasks. So it may need several calls to read(),
1563 * but we cannot guarantee that the information we produce is correct
1564 * unless we produce it entirely atomically.
1566 * Upon tasks file open(), a struct ctr_struct is allocated, that
1567 * will have a pointer to an array (also allocated here). The struct
1568 * ctr_struct * is stored in file->private_data. Its resources will
1569 * be freed by release() when the file is closed. The array is used
1570 * to sprintf the PIDs and then used by read().
1573 /* cpusets_tasks_read array */
1575 struct ctr_struct {
1576 char *buf;
1577 int bufsz;
1581 * Load into 'pidarray' up to 'npids' of the tasks using cpuset 'cs'.
1582 * Return actual number of pids loaded. No need to task_lock(p)
1583 * when reading out p->cpuset, as we don't really care if it changes
1584 * on the next cycle, and we are not going to try to dereference it.
1586 static int pid_array_load(pid_t *pidarray, int npids, struct cpuset *cs)
1588 int n = 0;
1589 struct task_struct *g, *p;
1591 read_lock(&tasklist_lock);
1593 do_each_thread(g, p) {
1594 if (p->cpuset == cs) {
1595 pidarray[n++] = p->pid;
1596 if (unlikely(n == npids))
1597 goto array_full;
1599 } while_each_thread(g, p);
1601 array_full:
1602 read_unlock(&tasklist_lock);
1603 return n;
1606 static int cmppid(const void *a, const void *b)
1608 return *(pid_t *)a - *(pid_t *)b;
1612 * Convert array 'a' of 'npids' pid_t's to a string of newline separated
1613 * decimal pids in 'buf'. Don't write more than 'sz' chars, but return
1614 * count 'cnt' of how many chars would be written if buf were large enough.
1616 static int pid_array_to_buf(char *buf, int sz, pid_t *a, int npids)
1618 int cnt = 0;
1619 int i;
1621 for (i = 0; i < npids; i++)
1622 cnt += snprintf(buf + cnt, max(sz - cnt, 0), "%d\n", a[i]);
1623 return cnt;
1627 * Handle an open on 'tasks' file. Prepare a buffer listing the
1628 * process id's of tasks currently attached to the cpuset being opened.
1630 * Does not require any specific cpuset mutexes, and does not take any.
1632 static int cpuset_tasks_open(struct inode *unused, struct file *file)
1634 struct cpuset *cs = __d_cs(file->f_dentry->d_parent);
1635 struct ctr_struct *ctr;
1636 pid_t *pidarray;
1637 int npids;
1638 char c;
1640 if (!(file->f_mode & FMODE_READ))
1641 return 0;
1643 ctr = kmalloc(sizeof(*ctr), GFP_KERNEL);
1644 if (!ctr)
1645 goto err0;
1648 * If cpuset gets more users after we read count, we won't have
1649 * enough space - tough. This race is indistinguishable to the
1650 * caller from the case that the additional cpuset users didn't
1651 * show up until sometime later on.
1653 npids = atomic_read(&cs->count);
1654 pidarray = kmalloc(npids * sizeof(pid_t), GFP_KERNEL);
1655 if (!pidarray)
1656 goto err1;
1658 npids = pid_array_load(pidarray, npids, cs);
1659 sort(pidarray, npids, sizeof(pid_t), cmppid, NULL);
1661 /* Call pid_array_to_buf() twice, first just to get bufsz */
1662 ctr->bufsz = pid_array_to_buf(&c, sizeof(c), pidarray, npids) + 1;
1663 ctr->buf = kmalloc(ctr->bufsz, GFP_KERNEL);
1664 if (!ctr->buf)
1665 goto err2;
1666 ctr->bufsz = pid_array_to_buf(ctr->buf, ctr->bufsz, pidarray, npids);
1668 kfree(pidarray);
1669 file->private_data = ctr;
1670 return 0;
1672 err2:
1673 kfree(pidarray);
1674 err1:
1675 kfree(ctr);
1676 err0:
1677 return -ENOMEM;
1680 static ssize_t cpuset_tasks_read(struct file *file, char __user *buf,
1681 size_t nbytes, loff_t *ppos)
1683 struct ctr_struct *ctr = file->private_data;
1685 if (*ppos + nbytes > ctr->bufsz)
1686 nbytes = ctr->bufsz - *ppos;
1687 if (copy_to_user(buf, ctr->buf + *ppos, nbytes))
1688 return -EFAULT;
1689 *ppos += nbytes;
1690 return nbytes;
1693 static int cpuset_tasks_release(struct inode *unused_inode, struct file *file)
1695 struct ctr_struct *ctr;
1697 if (file->f_mode & FMODE_READ) {
1698 ctr = file->private_data;
1699 kfree(ctr->buf);
1700 kfree(ctr);
1702 return 0;
1706 * for the common functions, 'private' gives the type of file
1709 static struct cftype cft_tasks = {
1710 .name = "tasks",
1711 .open = cpuset_tasks_open,
1712 .read = cpuset_tasks_read,
1713 .release = cpuset_tasks_release,
1714 .private = FILE_TASKLIST,
1717 static struct cftype cft_cpus = {
1718 .name = "cpus",
1719 .private = FILE_CPULIST,
1722 static struct cftype cft_mems = {
1723 .name = "mems",
1724 .private = FILE_MEMLIST,
1727 static struct cftype cft_cpu_exclusive = {
1728 .name = "cpu_exclusive",
1729 .private = FILE_CPU_EXCLUSIVE,
1732 static struct cftype cft_mem_exclusive = {
1733 .name = "mem_exclusive",
1734 .private = FILE_MEM_EXCLUSIVE,
1737 static struct cftype cft_notify_on_release = {
1738 .name = "notify_on_release",
1739 .private = FILE_NOTIFY_ON_RELEASE,
1742 static struct cftype cft_memory_migrate = {
1743 .name = "memory_migrate",
1744 .private = FILE_MEMORY_MIGRATE,
1747 static struct cftype cft_memory_pressure_enabled = {
1748 .name = "memory_pressure_enabled",
1749 .private = FILE_MEMORY_PRESSURE_ENABLED,
1752 static struct cftype cft_memory_pressure = {
1753 .name = "memory_pressure",
1754 .private = FILE_MEMORY_PRESSURE,
1757 static struct cftype cft_spread_page = {
1758 .name = "memory_spread_page",
1759 .private = FILE_SPREAD_PAGE,
1762 static struct cftype cft_spread_slab = {
1763 .name = "memory_spread_slab",
1764 .private = FILE_SPREAD_SLAB,
1767 static int cpuset_populate_dir(struct dentry *cs_dentry)
1769 int err;
1771 if ((err = cpuset_add_file(cs_dentry, &cft_cpus)) < 0)
1772 return err;
1773 if ((err = cpuset_add_file(cs_dentry, &cft_mems)) < 0)
1774 return err;
1775 if ((err = cpuset_add_file(cs_dentry, &cft_cpu_exclusive)) < 0)
1776 return err;
1777 if ((err = cpuset_add_file(cs_dentry, &cft_mem_exclusive)) < 0)
1778 return err;
1779 if ((err = cpuset_add_file(cs_dentry, &cft_notify_on_release)) < 0)
1780 return err;
1781 if ((err = cpuset_add_file(cs_dentry, &cft_memory_migrate)) < 0)
1782 return err;
1783 if ((err = cpuset_add_file(cs_dentry, &cft_memory_pressure)) < 0)
1784 return err;
1785 if ((err = cpuset_add_file(cs_dentry, &cft_spread_page)) < 0)
1786 return err;
1787 if ((err = cpuset_add_file(cs_dentry, &cft_spread_slab)) < 0)
1788 return err;
1789 if ((err = cpuset_add_file(cs_dentry, &cft_tasks)) < 0)
1790 return err;
1791 return 0;
1795 * cpuset_create - create a cpuset
1796 * parent: cpuset that will be parent of the new cpuset.
1797 * name: name of the new cpuset. Will be strcpy'ed.
1798 * mode: mode to set on new inode
1800 * Must be called with the mutex on the parent inode held
1803 static long cpuset_create(struct cpuset *parent, const char *name, int mode)
1805 struct cpuset *cs;
1806 int err;
1808 cs = kmalloc(sizeof(*cs), GFP_KERNEL);
1809 if (!cs)
1810 return -ENOMEM;
1812 mutex_lock(&manage_mutex);
1813 cpuset_update_task_memory_state();
1814 cs->flags = 0;
1815 if (notify_on_release(parent))
1816 set_bit(CS_NOTIFY_ON_RELEASE, &cs->flags);
1817 if (is_spread_page(parent))
1818 set_bit(CS_SPREAD_PAGE, &cs->flags);
1819 if (is_spread_slab(parent))
1820 set_bit(CS_SPREAD_SLAB, &cs->flags);
1821 cs->cpus_allowed = CPU_MASK_NONE;
1822 cs->mems_allowed = NODE_MASK_NONE;
1823 atomic_set(&cs->count, 0);
1824 INIT_LIST_HEAD(&cs->sibling);
1825 INIT_LIST_HEAD(&cs->children);
1826 cs->mems_generation = atomic_inc_return(&cpuset_mems_generation);
1827 fmeter_init(&cs->fmeter);
1829 cs->parent = parent;
1831 mutex_lock(&callback_mutex);
1832 list_add(&cs->sibling, &cs->parent->children);
1833 number_of_cpusets++;
1834 mutex_unlock(&callback_mutex);
1836 err = cpuset_create_dir(cs, name, mode);
1837 if (err < 0)
1838 goto err;
1841 * Release manage_mutex before cpuset_populate_dir() because it
1842 * will down() this new directory's i_mutex and if we race with
1843 * another mkdir, we might deadlock.
1845 mutex_unlock(&manage_mutex);
1847 err = cpuset_populate_dir(cs->dentry);
1848 /* If err < 0, we have a half-filled directory - oh well ;) */
1849 return 0;
1850 err:
1851 list_del(&cs->sibling);
1852 mutex_unlock(&manage_mutex);
1853 kfree(cs);
1854 return err;
1857 static int cpuset_mkdir(struct inode *dir, struct dentry *dentry, int mode)
1859 struct cpuset *c_parent = dentry->d_parent->d_fsdata;
1861 /* the vfs holds inode->i_mutex already */
1862 return cpuset_create(c_parent, dentry->d_name.name, mode | S_IFDIR);
1865 static int cpuset_rmdir(struct inode *unused_dir, struct dentry *dentry)
1867 struct cpuset *cs = dentry->d_fsdata;
1868 struct dentry *d;
1869 struct cpuset *parent;
1870 char *pathbuf = NULL;
1872 /* the vfs holds both inode->i_mutex already */
1874 mutex_lock(&manage_mutex);
1875 cpuset_update_task_memory_state();
1876 if (atomic_read(&cs->count) > 0) {
1877 mutex_unlock(&manage_mutex);
1878 return -EBUSY;
1880 if (!list_empty(&cs->children)) {
1881 mutex_unlock(&manage_mutex);
1882 return -EBUSY;
1884 parent = cs->parent;
1885 mutex_lock(&callback_mutex);
1886 set_bit(CS_REMOVED, &cs->flags);
1887 if (is_cpu_exclusive(cs))
1888 update_cpu_domains(cs);
1889 list_del(&cs->sibling); /* delete my sibling from parent->children */
1890 spin_lock(&cs->dentry->d_lock);
1891 d = dget(cs->dentry);
1892 cs->dentry = NULL;
1893 spin_unlock(&d->d_lock);
1894 cpuset_d_remove_dir(d);
1895 dput(d);
1896 number_of_cpusets--;
1897 mutex_unlock(&callback_mutex);
1898 if (list_empty(&parent->children))
1899 check_for_release(parent, &pathbuf);
1900 mutex_unlock(&manage_mutex);
1901 cpuset_release_agent(pathbuf);
1902 return 0;
1906 * cpuset_init_early - just enough so that the calls to
1907 * cpuset_update_task_memory_state() in early init code
1908 * are harmless.
1911 int __init cpuset_init_early(void)
1913 struct task_struct *tsk = current;
1915 tsk->cpuset = &top_cpuset;
1916 tsk->cpuset->mems_generation = atomic_inc_return(&cpuset_mems_generation);
1917 return 0;
1921 * cpuset_init - initialize cpusets at system boot
1923 * Description: Initialize top_cpuset and the cpuset internal file system,
1926 int __init cpuset_init(void)
1928 struct dentry *root;
1929 int err;
1931 top_cpuset.cpus_allowed = CPU_MASK_ALL;
1932 top_cpuset.mems_allowed = NODE_MASK_ALL;
1934 fmeter_init(&top_cpuset.fmeter);
1935 top_cpuset.mems_generation = atomic_inc_return(&cpuset_mems_generation);
1937 init_task.cpuset = &top_cpuset;
1939 err = register_filesystem(&cpuset_fs_type);
1940 if (err < 0)
1941 goto out;
1942 cpuset_mount = kern_mount(&cpuset_fs_type);
1943 if (IS_ERR(cpuset_mount)) {
1944 printk(KERN_ERR "cpuset: could not mount!\n");
1945 err = PTR_ERR(cpuset_mount);
1946 cpuset_mount = NULL;
1947 goto out;
1949 root = cpuset_mount->mnt_sb->s_root;
1950 root->d_fsdata = &top_cpuset;
1951 root->d_inode->i_nlink++;
1952 top_cpuset.dentry = root;
1953 root->d_inode->i_op = &cpuset_dir_inode_operations;
1954 number_of_cpusets = 1;
1955 err = cpuset_populate_dir(root);
1956 /* memory_pressure_enabled is in root cpuset only */
1957 if (err == 0)
1958 err = cpuset_add_file(root, &cft_memory_pressure_enabled);
1959 out:
1960 return err;
1964 * cpuset_init_smp - initialize cpus_allowed
1966 * Description: Finish top cpuset after cpu, node maps are initialized
1969 void __init cpuset_init_smp(void)
1971 top_cpuset.cpus_allowed = cpu_online_map;
1972 top_cpuset.mems_allowed = node_online_map;
1976 * cpuset_fork - attach newly forked task to its parents cpuset.
1977 * @tsk: pointer to task_struct of forking parent process.
1979 * Description: A task inherits its parent's cpuset at fork().
1981 * A pointer to the shared cpuset was automatically copied in fork.c
1982 * by dup_task_struct(). However, we ignore that copy, since it was
1983 * not made under the protection of task_lock(), so might no longer be
1984 * a valid cpuset pointer. attach_task() might have already changed
1985 * current->cpuset, allowing the previously referenced cpuset to
1986 * be removed and freed. Instead, we task_lock(current) and copy
1987 * its present value of current->cpuset for our freshly forked child.
1989 * At the point that cpuset_fork() is called, 'current' is the parent
1990 * task, and the passed argument 'child' points to the child task.
1993 void cpuset_fork(struct task_struct *child)
1995 task_lock(current);
1996 child->cpuset = current->cpuset;
1997 atomic_inc(&child->cpuset->count);
1998 task_unlock(current);
2002 * cpuset_exit - detach cpuset from exiting task
2003 * @tsk: pointer to task_struct of exiting process
2005 * Description: Detach cpuset from @tsk and release it.
2007 * Note that cpusets marked notify_on_release force every task in
2008 * them to take the global manage_mutex mutex when exiting.
2009 * This could impact scaling on very large systems. Be reluctant to
2010 * use notify_on_release cpusets where very high task exit scaling
2011 * is required on large systems.
2013 * Don't even think about derefencing 'cs' after the cpuset use count
2014 * goes to zero, except inside a critical section guarded by manage_mutex
2015 * or callback_mutex. Otherwise a zero cpuset use count is a license to
2016 * any other task to nuke the cpuset immediately, via cpuset_rmdir().
2018 * This routine has to take manage_mutex, not callback_mutex, because
2019 * it is holding that mutex while calling check_for_release(),
2020 * which calls kmalloc(), so can't be called holding callback_mutex().
2022 * We don't need to task_lock() this reference to tsk->cpuset,
2023 * because tsk is already marked PF_EXITING, so attach_task() won't
2024 * mess with it, or task is a failed fork, never visible to attach_task.
2026 * Hack:
2028 * Set the exiting tasks cpuset to the root cpuset (top_cpuset).
2030 * Don't leave a task unable to allocate memory, as that is an
2031 * accident waiting to happen should someone add a callout in
2032 * do_exit() after the cpuset_exit() call that might allocate.
2033 * If a task tries to allocate memory with an invalid cpuset,
2034 * it will oops in cpuset_update_task_memory_state().
2036 * We call cpuset_exit() while the task is still competent to
2037 * handle notify_on_release(), then leave the task attached to
2038 * the root cpuset (top_cpuset) for the remainder of its exit.
2040 * To do this properly, we would increment the reference count on
2041 * top_cpuset, and near the very end of the kernel/exit.c do_exit()
2042 * code we would add a second cpuset function call, to drop that
2043 * reference. This would just create an unnecessary hot spot on
2044 * the top_cpuset reference count, to no avail.
2046 * Normally, holding a reference to a cpuset without bumping its
2047 * count is unsafe. The cpuset could go away, or someone could
2048 * attach us to a different cpuset, decrementing the count on
2049 * the first cpuset that we never incremented. But in this case,
2050 * top_cpuset isn't going away, and either task has PF_EXITING set,
2051 * which wards off any attach_task() attempts, or task is a failed
2052 * fork, never visible to attach_task.
2054 * Another way to do this would be to set the cpuset pointer
2055 * to NULL here, and check in cpuset_update_task_memory_state()
2056 * for a NULL pointer. This hack avoids that NULL check, for no
2057 * cost (other than this way too long comment ;).
2060 void cpuset_exit(struct task_struct *tsk)
2062 struct cpuset *cs;
2064 cs = tsk->cpuset;
2065 tsk->cpuset = &top_cpuset; /* Hack - see comment above */
2067 if (notify_on_release(cs)) {
2068 char *pathbuf = NULL;
2070 mutex_lock(&manage_mutex);
2071 if (atomic_dec_and_test(&cs->count))
2072 check_for_release(cs, &pathbuf);
2073 mutex_unlock(&manage_mutex);
2074 cpuset_release_agent(pathbuf);
2075 } else {
2076 atomic_dec(&cs->count);
2081 * cpuset_cpus_allowed - return cpus_allowed mask from a tasks cpuset.
2082 * @tsk: pointer to task_struct from which to obtain cpuset->cpus_allowed.
2084 * Description: Returns the cpumask_t cpus_allowed of the cpuset
2085 * attached to the specified @tsk. Guaranteed to return some non-empty
2086 * subset of cpu_online_map, even if this means going outside the
2087 * tasks cpuset.
2090 cpumask_t cpuset_cpus_allowed(struct task_struct *tsk)
2092 cpumask_t mask;
2094 mutex_lock(&callback_mutex);
2095 task_lock(tsk);
2096 guarantee_online_cpus(tsk->cpuset, &mask);
2097 task_unlock(tsk);
2098 mutex_unlock(&callback_mutex);
2100 return mask;
2103 void cpuset_init_current_mems_allowed(void)
2105 current->mems_allowed = NODE_MASK_ALL;
2109 * cpuset_mems_allowed - return mems_allowed mask from a tasks cpuset.
2110 * @tsk: pointer to task_struct from which to obtain cpuset->mems_allowed.
2112 * Description: Returns the nodemask_t mems_allowed of the cpuset
2113 * attached to the specified @tsk. Guaranteed to return some non-empty
2114 * subset of node_online_map, even if this means going outside the
2115 * tasks cpuset.
2118 nodemask_t cpuset_mems_allowed(struct task_struct *tsk)
2120 nodemask_t mask;
2122 mutex_lock(&callback_mutex);
2123 task_lock(tsk);
2124 guarantee_online_mems(tsk->cpuset, &mask);
2125 task_unlock(tsk);
2126 mutex_unlock(&callback_mutex);
2128 return mask;
2132 * cpuset_zonelist_valid_mems_allowed - check zonelist vs. curremt mems_allowed
2133 * @zl: the zonelist to be checked
2135 * Are any of the nodes on zonelist zl allowed in current->mems_allowed?
2137 int cpuset_zonelist_valid_mems_allowed(struct zonelist *zl)
2139 int i;
2141 for (i = 0; zl->zones[i]; i++) {
2142 int nid = zl->zones[i]->zone_pgdat->node_id;
2144 if (node_isset(nid, current->mems_allowed))
2145 return 1;
2147 return 0;
2151 * nearest_exclusive_ancestor() - Returns the nearest mem_exclusive
2152 * ancestor to the specified cpuset. Call holding callback_mutex.
2153 * If no ancestor is mem_exclusive (an unusual configuration), then
2154 * returns the root cpuset.
2156 static const struct cpuset *nearest_exclusive_ancestor(const struct cpuset *cs)
2158 while (!is_mem_exclusive(cs) && cs->parent)
2159 cs = cs->parent;
2160 return cs;
2164 * cpuset_zone_allowed - Can we allocate memory on zone z's memory node?
2165 * @z: is this zone on an allowed node?
2166 * @gfp_mask: memory allocation flags (we use __GFP_HARDWALL)
2168 * If we're in interrupt, yes, we can always allocate. If zone
2169 * z's node is in our tasks mems_allowed, yes. If it's not a
2170 * __GFP_HARDWALL request and this zone's nodes is in the nearest
2171 * mem_exclusive cpuset ancestor to this tasks cpuset, yes.
2172 * Otherwise, no.
2174 * GFP_USER allocations are marked with the __GFP_HARDWALL bit,
2175 * and do not allow allocations outside the current tasks cpuset.
2176 * GFP_KERNEL allocations are not so marked, so can escape to the
2177 * nearest mem_exclusive ancestor cpuset.
2179 * Scanning up parent cpusets requires callback_mutex. The __alloc_pages()
2180 * routine only calls here with __GFP_HARDWALL bit _not_ set if
2181 * it's a GFP_KERNEL allocation, and all nodes in the current tasks
2182 * mems_allowed came up empty on the first pass over the zonelist.
2183 * So only GFP_KERNEL allocations, if all nodes in the cpuset are
2184 * short of memory, might require taking the callback_mutex mutex.
2186 * The first loop over the zonelist in mm/page_alloc.c:__alloc_pages()
2187 * calls here with __GFP_HARDWALL always set in gfp_mask, enforcing
2188 * hardwall cpusets - no allocation on a node outside the cpuset is
2189 * allowed (unless in interrupt, of course).
2191 * The second loop doesn't even call here for GFP_ATOMIC requests
2192 * (if the __alloc_pages() local variable 'wait' is set). That check
2193 * and the checks below have the combined affect in the second loop of
2194 * the __alloc_pages() routine that:
2195 * in_interrupt - any node ok (current task context irrelevant)
2196 * GFP_ATOMIC - any node ok
2197 * GFP_KERNEL - any node in enclosing mem_exclusive cpuset ok
2198 * GFP_USER - only nodes in current tasks mems allowed ok.
2201 int __cpuset_zone_allowed(struct zone *z, gfp_t gfp_mask)
2203 int node; /* node that zone z is on */
2204 const struct cpuset *cs; /* current cpuset ancestors */
2205 int allowed = 1; /* is allocation in zone z allowed? */
2207 if (in_interrupt())
2208 return 1;
2209 node = z->zone_pgdat->node_id;
2210 if (node_isset(node, current->mems_allowed))
2211 return 1;
2212 if (gfp_mask & __GFP_HARDWALL) /* If hardwall request, stop here */
2213 return 0;
2215 if (current->flags & PF_EXITING) /* Let dying task have memory */
2216 return 1;
2218 /* Not hardwall and node outside mems_allowed: scan up cpusets */
2219 mutex_lock(&callback_mutex);
2221 task_lock(current);
2222 cs = nearest_exclusive_ancestor(current->cpuset);
2223 task_unlock(current);
2225 allowed = node_isset(node, cs->mems_allowed);
2226 mutex_unlock(&callback_mutex);
2227 return allowed;
2231 * cpuset_lock - lock out any changes to cpuset structures
2233 * The out of memory (oom) code needs to mutex_lock cpusets
2234 * from being changed while it scans the tasklist looking for a
2235 * task in an overlapping cpuset. Expose callback_mutex via this
2236 * cpuset_lock() routine, so the oom code can lock it, before
2237 * locking the task list. The tasklist_lock is a spinlock, so
2238 * must be taken inside callback_mutex.
2241 void cpuset_lock(void)
2243 mutex_lock(&callback_mutex);
2247 * cpuset_unlock - release lock on cpuset changes
2249 * Undo the lock taken in a previous cpuset_lock() call.
2252 void cpuset_unlock(void)
2254 mutex_unlock(&callback_mutex);
2258 * cpuset_mem_spread_node() - On which node to begin search for a page
2260 * If a task is marked PF_SPREAD_PAGE or PF_SPREAD_SLAB (as for
2261 * tasks in a cpuset with is_spread_page or is_spread_slab set),
2262 * and if the memory allocation used cpuset_mem_spread_node()
2263 * to determine on which node to start looking, as it will for
2264 * certain page cache or slab cache pages such as used for file
2265 * system buffers and inode caches, then instead of starting on the
2266 * local node to look for a free page, rather spread the starting
2267 * node around the tasks mems_allowed nodes.
2269 * We don't have to worry about the returned node being offline
2270 * because "it can't happen", and even if it did, it would be ok.
2272 * The routines calling guarantee_online_mems() are careful to
2273 * only set nodes in task->mems_allowed that are online. So it
2274 * should not be possible for the following code to return an
2275 * offline node. But if it did, that would be ok, as this routine
2276 * is not returning the node where the allocation must be, only
2277 * the node where the search should start. The zonelist passed to
2278 * __alloc_pages() will include all nodes. If the slab allocator
2279 * is passed an offline node, it will fall back to the local node.
2280 * See kmem_cache_alloc_node().
2283 int cpuset_mem_spread_node(void)
2285 int node;
2287 node = next_node(current->cpuset_mem_spread_rotor, current->mems_allowed);
2288 if (node == MAX_NUMNODES)
2289 node = first_node(current->mems_allowed);
2290 current->cpuset_mem_spread_rotor = node;
2291 return node;
2293 EXPORT_SYMBOL_GPL(cpuset_mem_spread_node);
2296 * cpuset_excl_nodes_overlap - Do we overlap @p's mem_exclusive ancestors?
2297 * @p: pointer to task_struct of some other task.
2299 * Description: Return true if the nearest mem_exclusive ancestor
2300 * cpusets of tasks @p and current overlap. Used by oom killer to
2301 * determine if task @p's memory usage might impact the memory
2302 * available to the current task.
2304 * Call while holding callback_mutex.
2307 int cpuset_excl_nodes_overlap(const struct task_struct *p)
2309 const struct cpuset *cs1, *cs2; /* my and p's cpuset ancestors */
2310 int overlap = 0; /* do cpusets overlap? */
2312 task_lock(current);
2313 if (current->flags & PF_EXITING) {
2314 task_unlock(current);
2315 goto done;
2317 cs1 = nearest_exclusive_ancestor(current->cpuset);
2318 task_unlock(current);
2320 task_lock((struct task_struct *)p);
2321 if (p->flags & PF_EXITING) {
2322 task_unlock((struct task_struct *)p);
2323 goto done;
2325 cs2 = nearest_exclusive_ancestor(p->cpuset);
2326 task_unlock((struct task_struct *)p);
2328 overlap = nodes_intersects(cs1->mems_allowed, cs2->mems_allowed);
2329 done:
2330 return overlap;
2334 * Collection of memory_pressure is suppressed unless
2335 * this flag is enabled by writing "1" to the special
2336 * cpuset file 'memory_pressure_enabled' in the root cpuset.
2339 int cpuset_memory_pressure_enabled __read_mostly;
2342 * cpuset_memory_pressure_bump - keep stats of per-cpuset reclaims.
2344 * Keep a running average of the rate of synchronous (direct)
2345 * page reclaim efforts initiated by tasks in each cpuset.
2347 * This represents the rate at which some task in the cpuset
2348 * ran low on memory on all nodes it was allowed to use, and
2349 * had to enter the kernels page reclaim code in an effort to
2350 * create more free memory by tossing clean pages or swapping
2351 * or writing dirty pages.
2353 * Display to user space in the per-cpuset read-only file
2354 * "memory_pressure". Value displayed is an integer
2355 * representing the recent rate of entry into the synchronous
2356 * (direct) page reclaim by any task attached to the cpuset.
2359 void __cpuset_memory_pressure_bump(void)
2361 struct cpuset *cs;
2363 task_lock(current);
2364 cs = current->cpuset;
2365 fmeter_markevent(&cs->fmeter);
2366 task_unlock(current);
2370 * proc_cpuset_show()
2371 * - Print tasks cpuset path into seq_file.
2372 * - Used for /proc/<pid>/cpuset.
2373 * - No need to task_lock(tsk) on this tsk->cpuset reference, as it
2374 * doesn't really matter if tsk->cpuset changes after we read it,
2375 * and we take manage_mutex, keeping attach_task() from changing it
2376 * anyway.
2379 static int proc_cpuset_show(struct seq_file *m, void *v)
2381 struct cpuset *cs;
2382 struct task_struct *tsk;
2383 char *buf;
2384 int retval = 0;
2386 buf = kmalloc(PAGE_SIZE, GFP_KERNEL);
2387 if (!buf)
2388 return -ENOMEM;
2390 tsk = m->private;
2391 mutex_lock(&manage_mutex);
2392 cs = tsk->cpuset;
2393 if (!cs) {
2394 retval = -EINVAL;
2395 goto out;
2398 retval = cpuset_path(cs, buf, PAGE_SIZE);
2399 if (retval < 0)
2400 goto out;
2401 seq_puts(m, buf);
2402 seq_putc(m, '\n');
2403 out:
2404 mutex_unlock(&manage_mutex);
2405 kfree(buf);
2406 return retval;
2409 static int cpuset_open(struct inode *inode, struct file *file)
2411 struct task_struct *tsk = PROC_I(inode)->task;
2412 return single_open(file, proc_cpuset_show, tsk);
2415 struct file_operations proc_cpuset_operations = {
2416 .open = cpuset_open,
2417 .read = seq_read,
2418 .llseek = seq_lseek,
2419 .release = single_release,
2422 /* Display task cpus_allowed, mems_allowed in /proc/<pid>/status file. */
2423 char *cpuset_task_status_allowed(struct task_struct *task, char *buffer)
2425 buffer += sprintf(buffer, "Cpus_allowed:\t");
2426 buffer += cpumask_scnprintf(buffer, PAGE_SIZE, task->cpus_allowed);
2427 buffer += sprintf(buffer, "\n");
2428 buffer += sprintf(buffer, "Mems_allowed:\t");
2429 buffer += nodemask_scnprintf(buffer, PAGE_SIZE, task->mems_allowed);
2430 buffer += sprintf(buffer, "\n");
2431 return buffer;