cpusets: allow TIF_MEMDIE threads to allocate anywhere
[linux-2.6/mini2440.git] / kernel / cpuset.c
blobd240349cbf0f282d56e12205b68d19ecce215b40
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/cpu.h>
22 #include <linux/cpumask.h>
23 #include <linux/cpuset.h>
24 #include <linux/err.h>
25 #include <linux/errno.h>
26 #include <linux/file.h>
27 #include <linux/fs.h>
28 #include <linux/init.h>
29 #include <linux/interrupt.h>
30 #include <linux/kernel.h>
31 #include <linux/kmod.h>
32 #include <linux/list.h>
33 #include <linux/mempolicy.h>
34 #include <linux/mm.h>
35 #include <linux/module.h>
36 #include <linux/mount.h>
37 #include <linux/namei.h>
38 #include <linux/pagemap.h>
39 #include <linux/proc_fs.h>
40 #include <linux/rcupdate.h>
41 #include <linux/sched.h>
42 #include <linux/seq_file.h>
43 #include <linux/security.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 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 * Since cpuset_mems_generation is guarded by manage_mutex,
168 * there is no need to mark it atomic.
170 static int cpuset_mems_generation;
172 static struct cpuset top_cpuset = {
173 .flags = ((1 << CS_CPU_EXCLUSIVE) | (1 << CS_MEM_EXCLUSIVE)),
174 .cpus_allowed = CPU_MASK_ALL,
175 .mems_allowed = NODE_MASK_ALL,
176 .count = ATOMIC_INIT(0),
177 .sibling = LIST_HEAD_INIT(top_cpuset.sibling),
178 .children = LIST_HEAD_INIT(top_cpuset.children),
181 static struct vfsmount *cpuset_mount;
182 static struct super_block *cpuset_sb;
185 * We have two global cpuset mutexes below. They can nest.
186 * It is ok to first take manage_mutex, then nest callback_mutex. We also
187 * require taking task_lock() when dereferencing a tasks cpuset pointer.
188 * See "The task_lock() exception", at the end of this comment.
190 * A task must hold both mutexes to modify cpusets. If a task
191 * holds manage_mutex, then it blocks others wanting that mutex,
192 * ensuring that it is the only task able to also acquire callback_mutex
193 * and be able to modify cpusets. It can perform various checks on
194 * the cpuset structure first, knowing nothing will change. It can
195 * also allocate memory while just holding manage_mutex. While it is
196 * performing these checks, various callback routines can briefly
197 * acquire callback_mutex to query cpusets. Once it is ready to make
198 * the changes, it takes callback_mutex, blocking everyone else.
200 * Calls to the kernel memory allocator can not be made while holding
201 * callback_mutex, as that would risk double tripping on callback_mutex
202 * from one of the callbacks into the cpuset code from within
203 * __alloc_pages().
205 * If a task is only holding callback_mutex, then it has read-only
206 * access to cpusets.
208 * The task_struct fields mems_allowed and mems_generation may only
209 * be accessed in the context of that task, so require no locks.
211 * Any task can increment and decrement the count field without lock.
212 * So in general, code holding manage_mutex or callback_mutex can't rely
213 * on the count field not changing. However, if the count goes to
214 * zero, then only attach_task(), which holds both mutexes, can
215 * increment it again. Because a count of zero means that no tasks
216 * are currently attached, therefore there is no way a task attached
217 * to that cpuset can fork (the other way to increment the count).
218 * So code holding manage_mutex or callback_mutex can safely assume that
219 * if the count is zero, it will stay zero. Similarly, if a task
220 * holds manage_mutex or callback_mutex on a cpuset with zero count, it
221 * knows that the cpuset won't be removed, as cpuset_rmdir() needs
222 * both of those mutexes.
224 * The cpuset_common_file_write handler for operations that modify
225 * the cpuset hierarchy holds manage_mutex across the entire operation,
226 * single threading all such cpuset modifications across the system.
228 * The cpuset_common_file_read() handlers only hold callback_mutex across
229 * small pieces of code, such as when reading out possibly multi-word
230 * cpumasks and nodemasks.
232 * The fork and exit callbacks cpuset_fork() and cpuset_exit(), don't
233 * (usually) take either mutex. These are the two most performance
234 * critical pieces of code here. The exception occurs on cpuset_exit(),
235 * when a task in a notify_on_release cpuset exits. Then manage_mutex
236 * is taken, and if the cpuset count is zero, a usermode call made
237 * to /sbin/cpuset_release_agent with the name of the cpuset (path
238 * relative to the root of cpuset file system) as the argument.
240 * A cpuset can only be deleted if both its 'count' of using tasks
241 * is zero, and its list of 'children' cpusets is empty. Since all
242 * tasks in the system use _some_ cpuset, and since there is always at
243 * least one task in the system (init), therefore, top_cpuset
244 * always has either children cpusets and/or using tasks. So we don't
245 * need a special hack to ensure that top_cpuset cannot be deleted.
247 * The above "Tale of Two Semaphores" would be complete, but for:
249 * The task_lock() exception
251 * The need for this exception arises from the action of attach_task(),
252 * which overwrites one tasks cpuset pointer with another. It does
253 * so using both mutexes, however there are several performance
254 * critical places that need to reference task->cpuset without the
255 * expense of grabbing a system global mutex. Therefore except as
256 * noted below, when dereferencing or, as in attach_task(), modifying
257 * a tasks cpuset pointer we use task_lock(), which acts on a spinlock
258 * (task->alloc_lock) already in the task_struct routinely used for
259 * such matters.
261 * P.S. One more locking exception. RCU is used to guard the
262 * update of a tasks cpuset pointer by attach_task() and the
263 * access of task->cpuset->mems_generation via that pointer in
264 * the routine cpuset_update_task_memory_state().
267 static DEFINE_MUTEX(manage_mutex);
268 static DEFINE_MUTEX(callback_mutex);
271 * A couple of forward declarations required, due to cyclic reference loop:
272 * cpuset_mkdir -> cpuset_create -> cpuset_populate_dir -> cpuset_add_file
273 * -> cpuset_create_file -> cpuset_dir_inode_operations -> cpuset_mkdir.
276 static int cpuset_mkdir(struct inode *dir, struct dentry *dentry, int mode);
277 static int cpuset_rmdir(struct inode *unused_dir, struct dentry *dentry);
279 static struct backing_dev_info cpuset_backing_dev_info = {
280 .ra_pages = 0, /* No readahead */
281 .capabilities = BDI_CAP_NO_ACCT_DIRTY | BDI_CAP_NO_WRITEBACK,
284 static struct inode *cpuset_new_inode(mode_t mode)
286 struct inode *inode = new_inode(cpuset_sb);
288 if (inode) {
289 inode->i_mode = mode;
290 inode->i_uid = current->fsuid;
291 inode->i_gid = current->fsgid;
292 inode->i_blocks = 0;
293 inode->i_atime = inode->i_mtime = inode->i_ctime = CURRENT_TIME;
294 inode->i_mapping->backing_dev_info = &cpuset_backing_dev_info;
296 return inode;
299 static void cpuset_diput(struct dentry *dentry, struct inode *inode)
301 /* is dentry a directory ? if so, kfree() associated cpuset */
302 if (S_ISDIR(inode->i_mode)) {
303 struct cpuset *cs = dentry->d_fsdata;
304 BUG_ON(!(is_removed(cs)));
305 kfree(cs);
307 iput(inode);
310 static struct dentry_operations cpuset_dops = {
311 .d_iput = cpuset_diput,
314 static struct dentry *cpuset_get_dentry(struct dentry *parent, const char *name)
316 struct dentry *d = lookup_one_len(name, parent, strlen(name));
317 if (!IS_ERR(d))
318 d->d_op = &cpuset_dops;
319 return d;
322 static void remove_dir(struct dentry *d)
324 struct dentry *parent = dget(d->d_parent);
326 d_delete(d);
327 simple_rmdir(parent->d_inode, d);
328 dput(parent);
332 * NOTE : the dentry must have been dget()'ed
334 static void cpuset_d_remove_dir(struct dentry *dentry)
336 struct list_head *node;
338 spin_lock(&dcache_lock);
339 node = dentry->d_subdirs.next;
340 while (node != &dentry->d_subdirs) {
341 struct dentry *d = list_entry(node, struct dentry, d_u.d_child);
342 list_del_init(node);
343 if (d->d_inode) {
344 d = dget_locked(d);
345 spin_unlock(&dcache_lock);
346 d_delete(d);
347 simple_unlink(dentry->d_inode, d);
348 dput(d);
349 spin_lock(&dcache_lock);
351 node = dentry->d_subdirs.next;
353 list_del_init(&dentry->d_u.d_child);
354 spin_unlock(&dcache_lock);
355 remove_dir(dentry);
358 static struct super_operations cpuset_ops = {
359 .statfs = simple_statfs,
360 .drop_inode = generic_delete_inode,
363 static int cpuset_fill_super(struct super_block *sb, void *unused_data,
364 int unused_silent)
366 struct inode *inode;
367 struct dentry *root;
369 sb->s_blocksize = PAGE_CACHE_SIZE;
370 sb->s_blocksize_bits = PAGE_CACHE_SHIFT;
371 sb->s_magic = CPUSET_SUPER_MAGIC;
372 sb->s_op = &cpuset_ops;
373 cpuset_sb = sb;
375 inode = cpuset_new_inode(S_IFDIR | S_IRUGO | S_IXUGO | S_IWUSR);
376 if (inode) {
377 inode->i_op = &simple_dir_inode_operations;
378 inode->i_fop = &simple_dir_operations;
379 /* directories start off with i_nlink == 2 (for "." entry) */
380 inc_nlink(inode);
381 } else {
382 return -ENOMEM;
385 root = d_alloc_root(inode);
386 if (!root) {
387 iput(inode);
388 return -ENOMEM;
390 sb->s_root = root;
391 return 0;
394 static int cpuset_get_sb(struct file_system_type *fs_type,
395 int flags, const char *unused_dev_name,
396 void *data, struct vfsmount *mnt)
398 return get_sb_single(fs_type, flags, data, cpuset_fill_super, mnt);
401 static struct file_system_type cpuset_fs_type = {
402 .name = "cpuset",
403 .get_sb = cpuset_get_sb,
404 .kill_sb = kill_litter_super,
407 /* struct cftype:
409 * The files in the cpuset filesystem mostly have a very simple read/write
410 * handling, some common function will take care of it. Nevertheless some cases
411 * (read tasks) are special and therefore I define this structure for every
412 * kind of file.
415 * When reading/writing to a file:
416 * - the cpuset to use in file->f_path.dentry->d_parent->d_fsdata
417 * - the 'cftype' of the file is file->f_path.dentry->d_fsdata
420 struct cftype {
421 char *name;
422 int private;
423 int (*open) (struct inode *inode, struct file *file);
424 ssize_t (*read) (struct file *file, char __user *buf, size_t nbytes,
425 loff_t *ppos);
426 int (*write) (struct file *file, const char __user *buf, size_t nbytes,
427 loff_t *ppos);
428 int (*release) (struct inode *inode, struct file *file);
431 static inline struct cpuset *__d_cs(struct dentry *dentry)
433 return dentry->d_fsdata;
436 static inline struct cftype *__d_cft(struct dentry *dentry)
438 return dentry->d_fsdata;
442 * Call with manage_mutex held. Writes path of cpuset into buf.
443 * Returns 0 on success, -errno on error.
446 static int cpuset_path(const struct cpuset *cs, char *buf, int buflen)
448 char *start;
450 start = buf + buflen;
452 *--start = '\0';
453 for (;;) {
454 int len = cs->dentry->d_name.len;
455 if ((start -= len) < buf)
456 return -ENAMETOOLONG;
457 memcpy(start, cs->dentry->d_name.name, len);
458 cs = cs->parent;
459 if (!cs)
460 break;
461 if (!cs->parent)
462 continue;
463 if (--start < buf)
464 return -ENAMETOOLONG;
465 *start = '/';
467 memmove(buf, start, buf + buflen - start);
468 return 0;
472 * Notify userspace when a cpuset is released, by running
473 * /sbin/cpuset_release_agent with the name of the cpuset (path
474 * relative to the root of cpuset file system) as the argument.
476 * Most likely, this user command will try to rmdir this cpuset.
478 * This races with the possibility that some other task will be
479 * attached to this cpuset before it is removed, or that some other
480 * user task will 'mkdir' a child cpuset of this cpuset. That's ok.
481 * The presumed 'rmdir' will fail quietly if this cpuset is no longer
482 * unused, and this cpuset will be reprieved from its death sentence,
483 * to continue to serve a useful existence. Next time it's released,
484 * we will get notified again, if it still has 'notify_on_release' set.
486 * The final arg to call_usermodehelper() is 0, which means don't
487 * wait. The separate /sbin/cpuset_release_agent task is forked by
488 * call_usermodehelper(), then control in this thread returns here,
489 * without waiting for the release agent task. We don't bother to
490 * wait because the caller of this routine has no use for the exit
491 * status of the /sbin/cpuset_release_agent task, so no sense holding
492 * our caller up for that.
494 * When we had only one cpuset mutex, we had to call this
495 * without holding it, to avoid deadlock when call_usermodehelper()
496 * allocated memory. With two locks, we could now call this while
497 * holding manage_mutex, but we still don't, so as to minimize
498 * the time manage_mutex is held.
501 static void cpuset_release_agent(const char *pathbuf)
503 char *argv[3], *envp[3];
504 int i;
506 if (!pathbuf)
507 return;
509 i = 0;
510 argv[i++] = "/sbin/cpuset_release_agent";
511 argv[i++] = (char *)pathbuf;
512 argv[i] = NULL;
514 i = 0;
515 /* minimal command environment */
516 envp[i++] = "HOME=/";
517 envp[i++] = "PATH=/sbin:/bin:/usr/sbin:/usr/bin";
518 envp[i] = NULL;
520 call_usermodehelper(argv[0], argv, envp, 0);
521 kfree(pathbuf);
525 * Either cs->count of using tasks transitioned to zero, or the
526 * cs->children list of child cpusets just became empty. If this
527 * cs is notify_on_release() and now both the user count is zero and
528 * the list of children is empty, prepare cpuset path in a kmalloc'd
529 * buffer, to be returned via ppathbuf, so that the caller can invoke
530 * cpuset_release_agent() with it later on, once manage_mutex is dropped.
531 * Call here with manage_mutex held.
533 * This check_for_release() routine is responsible for kmalloc'ing
534 * pathbuf. The above cpuset_release_agent() is responsible for
535 * kfree'ing pathbuf. The caller of these routines is responsible
536 * for providing a pathbuf pointer, initialized to NULL, then
537 * calling check_for_release() with manage_mutex held and the address
538 * of the pathbuf pointer, then dropping manage_mutex, then calling
539 * cpuset_release_agent() with pathbuf, as set by check_for_release().
542 static void check_for_release(struct cpuset *cs, char **ppathbuf)
544 if (notify_on_release(cs) && atomic_read(&cs->count) == 0 &&
545 list_empty(&cs->children)) {
546 char *buf;
548 buf = kmalloc(PAGE_SIZE, GFP_KERNEL);
549 if (!buf)
550 return;
551 if (cpuset_path(cs, buf, PAGE_SIZE) < 0)
552 kfree(buf);
553 else
554 *ppathbuf = buf;
559 * Return in *pmask the portion of a cpusets's cpus_allowed that
560 * are online. If none are online, walk up the cpuset hierarchy
561 * until we find one that does have some online cpus. If we get
562 * all the way to the top and still haven't found any online cpus,
563 * return cpu_online_map. Or if passed a NULL cs from an exit'ing
564 * task, return cpu_online_map.
566 * One way or another, we guarantee to return some non-empty subset
567 * of cpu_online_map.
569 * Call with callback_mutex held.
572 static void guarantee_online_cpus(const struct cpuset *cs, cpumask_t *pmask)
574 while (cs && !cpus_intersects(cs->cpus_allowed, cpu_online_map))
575 cs = cs->parent;
576 if (cs)
577 cpus_and(*pmask, cs->cpus_allowed, cpu_online_map);
578 else
579 *pmask = cpu_online_map;
580 BUG_ON(!cpus_intersects(*pmask, cpu_online_map));
584 * Return in *pmask the portion of a cpusets's mems_allowed that
585 * are online. If none are online, walk up the cpuset hierarchy
586 * until we find one that does have some online mems. If we get
587 * all the way to the top and still haven't found any online mems,
588 * return node_online_map.
590 * One way or another, we guarantee to return some non-empty subset
591 * of node_online_map.
593 * Call with callback_mutex held.
596 static void guarantee_online_mems(const struct cpuset *cs, nodemask_t *pmask)
598 while (cs && !nodes_intersects(cs->mems_allowed, node_online_map))
599 cs = cs->parent;
600 if (cs)
601 nodes_and(*pmask, cs->mems_allowed, node_online_map);
602 else
603 *pmask = node_online_map;
604 BUG_ON(!nodes_intersects(*pmask, node_online_map));
608 * cpuset_update_task_memory_state - update task memory placement
610 * If the current tasks cpusets mems_allowed changed behind our
611 * backs, update current->mems_allowed, mems_generation and task NUMA
612 * mempolicy to the new value.
614 * Task mempolicy is updated by rebinding it relative to the
615 * current->cpuset if a task has its memory placement changed.
616 * Do not call this routine if in_interrupt().
618 * Call without callback_mutex or task_lock() held. May be
619 * called with or without manage_mutex held. Thanks in part to
620 * 'the_top_cpuset_hack', the tasks cpuset pointer will never
621 * be NULL. 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 (cur == &top_cpuset)
733 return 0;
735 par = cur->parent;
737 /* We must be a subset of our parent cpuset */
738 if (!is_cpuset_subset(trial, par))
739 return -EACCES;
741 /* If either I or some sibling (!= me) is exclusive, we can't overlap */
742 list_for_each_entry(c, &par->children, sibling) {
743 if ((is_cpu_exclusive(trial) || is_cpu_exclusive(c)) &&
744 c != cur &&
745 cpus_intersects(trial->cpus_allowed, c->cpus_allowed))
746 return -EINVAL;
747 if ((is_mem_exclusive(trial) || is_mem_exclusive(c)) &&
748 c != cur &&
749 nodes_intersects(trial->mems_allowed, c->mems_allowed))
750 return -EINVAL;
753 return 0;
757 * For a given cpuset cur, partition the system as follows
758 * a. All cpus in the parent cpuset's cpus_allowed that are not part of any
759 * exclusive child cpusets
760 * b. All cpus in the current cpuset's cpus_allowed that are not part of any
761 * exclusive child cpusets
762 * Build these two partitions by calling partition_sched_domains
764 * Call with manage_mutex held. May nest a call to the
765 * lock_cpu_hotplug()/unlock_cpu_hotplug() pair.
766 * Must not be called holding callback_mutex, because we must
767 * not call lock_cpu_hotplug() while holding callback_mutex.
770 static void update_cpu_domains(struct cpuset *cur)
772 struct cpuset *c, *par = cur->parent;
773 cpumask_t pspan, cspan;
775 if (par == NULL || cpus_empty(cur->cpus_allowed))
776 return;
779 * Get all cpus from parent's cpus_allowed not part of exclusive
780 * children
782 pspan = par->cpus_allowed;
783 list_for_each_entry(c, &par->children, sibling) {
784 if (is_cpu_exclusive(c))
785 cpus_andnot(pspan, pspan, c->cpus_allowed);
787 if (!is_cpu_exclusive(cur)) {
788 cpus_or(pspan, pspan, cur->cpus_allowed);
789 if (cpus_equal(pspan, cur->cpus_allowed))
790 return;
791 cspan = CPU_MASK_NONE;
792 } else {
793 if (cpus_empty(pspan))
794 return;
795 cspan = cur->cpus_allowed;
797 * Get all cpus from current cpuset's cpus_allowed not part
798 * of exclusive children
800 list_for_each_entry(c, &cur->children, sibling) {
801 if (is_cpu_exclusive(c))
802 cpus_andnot(cspan, cspan, c->cpus_allowed);
806 lock_cpu_hotplug();
807 partition_sched_domains(&pspan, &cspan);
808 unlock_cpu_hotplug();
812 * Call with manage_mutex held. May take callback_mutex during call.
815 static int update_cpumask(struct cpuset *cs, char *buf)
817 struct cpuset trialcs;
818 int retval, cpus_unchanged;
820 /* top_cpuset.cpus_allowed tracks cpu_online_map; it's read-only */
821 if (cs == &top_cpuset)
822 return -EACCES;
824 trialcs = *cs;
825 retval = cpulist_parse(buf, trialcs.cpus_allowed);
826 if (retval < 0)
827 return retval;
828 cpus_and(trialcs.cpus_allowed, trialcs.cpus_allowed, cpu_online_map);
829 if (cpus_empty(trialcs.cpus_allowed))
830 return -ENOSPC;
831 retval = validate_change(cs, &trialcs);
832 if (retval < 0)
833 return retval;
834 cpus_unchanged = cpus_equal(cs->cpus_allowed, trialcs.cpus_allowed);
835 mutex_lock(&callback_mutex);
836 cs->cpus_allowed = trialcs.cpus_allowed;
837 mutex_unlock(&callback_mutex);
838 if (is_cpu_exclusive(cs) && !cpus_unchanged)
839 update_cpu_domains(cs);
840 return 0;
844 * cpuset_migrate_mm
846 * Migrate memory region from one set of nodes to another.
848 * Temporarilly set tasks mems_allowed to target nodes of migration,
849 * so that the migration code can allocate pages on these nodes.
851 * Call holding manage_mutex, so our current->cpuset won't change
852 * during this call, as manage_mutex holds off any attach_task()
853 * calls. Therefore we don't need to take task_lock around the
854 * call to guarantee_online_mems(), as we know no one is changing
855 * our tasks cpuset.
857 * Hold callback_mutex around the two modifications of our tasks
858 * mems_allowed to synchronize with cpuset_mems_allowed().
860 * While the mm_struct we are migrating is typically from some
861 * other task, the task_struct mems_allowed that we are hacking
862 * is for our current task, which must allocate new pages for that
863 * migrating memory region.
865 * We call cpuset_update_task_memory_state() before hacking
866 * our tasks mems_allowed, so that we are assured of being in
867 * sync with our tasks cpuset, and in particular, callbacks to
868 * cpuset_update_task_memory_state() from nested page allocations
869 * won't see any mismatch of our cpuset and task mems_generation
870 * values, so won't overwrite our hacked tasks mems_allowed
871 * nodemask.
874 static void cpuset_migrate_mm(struct mm_struct *mm, const nodemask_t *from,
875 const nodemask_t *to)
877 struct task_struct *tsk = current;
879 cpuset_update_task_memory_state();
881 mutex_lock(&callback_mutex);
882 tsk->mems_allowed = *to;
883 mutex_unlock(&callback_mutex);
885 do_migrate_pages(mm, from, to, MPOL_MF_MOVE_ALL);
887 mutex_lock(&callback_mutex);
888 guarantee_online_mems(tsk->cpuset, &tsk->mems_allowed);
889 mutex_unlock(&callback_mutex);
893 * Handle user request to change the 'mems' memory placement
894 * of a cpuset. Needs to validate the request, update the
895 * cpusets mems_allowed and mems_generation, and for each
896 * task in the cpuset, rebind any vma mempolicies and if
897 * the cpuset is marked 'memory_migrate', migrate the tasks
898 * pages to the new memory.
900 * Call with manage_mutex held. May take callback_mutex during call.
901 * Will take tasklist_lock, scan tasklist for tasks in cpuset cs,
902 * lock each such tasks mm->mmap_sem, scan its vma's and rebind
903 * their mempolicies to the cpusets new mems_allowed.
906 static int update_nodemask(struct cpuset *cs, char *buf)
908 struct cpuset trialcs;
909 nodemask_t oldmem;
910 struct task_struct *g, *p;
911 struct mm_struct **mmarray;
912 int i, n, ntasks;
913 int migrate;
914 int fudge;
915 int retval;
917 /* top_cpuset.mems_allowed tracks node_online_map; it's read-only */
918 if (cs == &top_cpuset)
919 return -EACCES;
921 trialcs = *cs;
922 retval = nodelist_parse(buf, trialcs.mems_allowed);
923 if (retval < 0)
924 goto done;
925 nodes_and(trialcs.mems_allowed, trialcs.mems_allowed, node_online_map);
926 oldmem = cs->mems_allowed;
927 if (nodes_equal(oldmem, trialcs.mems_allowed)) {
928 retval = 0; /* Too easy - nothing to do */
929 goto done;
931 if (nodes_empty(trialcs.mems_allowed)) {
932 retval = -ENOSPC;
933 goto done;
935 retval = validate_change(cs, &trialcs);
936 if (retval < 0)
937 goto done;
939 mutex_lock(&callback_mutex);
940 cs->mems_allowed = trialcs.mems_allowed;
941 cs->mems_generation = cpuset_mems_generation++;
942 mutex_unlock(&callback_mutex);
944 set_cpuset_being_rebound(cs); /* causes mpol_copy() rebind */
946 fudge = 10; /* spare mmarray[] slots */
947 fudge += cpus_weight(cs->cpus_allowed); /* imagine one fork-bomb/cpu */
948 retval = -ENOMEM;
951 * Allocate mmarray[] to hold mm reference for each task
952 * in cpuset cs. Can't kmalloc GFP_KERNEL while holding
953 * tasklist_lock. We could use GFP_ATOMIC, but with a
954 * few more lines of code, we can retry until we get a big
955 * enough mmarray[] w/o using GFP_ATOMIC.
957 while (1) {
958 ntasks = atomic_read(&cs->count); /* guess */
959 ntasks += fudge;
960 mmarray = kmalloc(ntasks * sizeof(*mmarray), GFP_KERNEL);
961 if (!mmarray)
962 goto done;
963 write_lock_irq(&tasklist_lock); /* block fork */
964 if (atomic_read(&cs->count) <= ntasks)
965 break; /* got enough */
966 write_unlock_irq(&tasklist_lock); /* try again */
967 kfree(mmarray);
970 n = 0;
972 /* Load up mmarray[] with mm reference for each task in cpuset. */
973 do_each_thread(g, p) {
974 struct mm_struct *mm;
976 if (n >= ntasks) {
977 printk(KERN_WARNING
978 "Cpuset mempolicy rebind incomplete.\n");
979 continue;
981 if (p->cpuset != cs)
982 continue;
983 mm = get_task_mm(p);
984 if (!mm)
985 continue;
986 mmarray[n++] = mm;
987 } while_each_thread(g, p);
988 write_unlock_irq(&tasklist_lock);
991 * Now that we've dropped the tasklist spinlock, we can
992 * rebind the vma mempolicies of each mm in mmarray[] to their
993 * new cpuset, and release that mm. The mpol_rebind_mm()
994 * call takes mmap_sem, which we couldn't take while holding
995 * tasklist_lock. Forks can happen again now - the mpol_copy()
996 * cpuset_being_rebound check will catch such forks, and rebind
997 * their vma mempolicies too. Because we still hold the global
998 * cpuset manage_mutex, we know that no other rebind effort will
999 * be contending for the global variable cpuset_being_rebound.
1000 * It's ok if we rebind the same mm twice; mpol_rebind_mm()
1001 * is idempotent. Also migrate pages in each mm to new nodes.
1003 migrate = is_memory_migrate(cs);
1004 for (i = 0; i < n; i++) {
1005 struct mm_struct *mm = mmarray[i];
1007 mpol_rebind_mm(mm, &cs->mems_allowed);
1008 if (migrate)
1009 cpuset_migrate_mm(mm, &oldmem, &cs->mems_allowed);
1010 mmput(mm);
1013 /* We're done rebinding vma's to this cpusets new mems_allowed. */
1014 kfree(mmarray);
1015 set_cpuset_being_rebound(NULL);
1016 retval = 0;
1017 done:
1018 return retval;
1022 * Call with manage_mutex held.
1025 static int update_memory_pressure_enabled(struct cpuset *cs, char *buf)
1027 if (simple_strtoul(buf, NULL, 10) != 0)
1028 cpuset_memory_pressure_enabled = 1;
1029 else
1030 cpuset_memory_pressure_enabled = 0;
1031 return 0;
1035 * update_flag - read a 0 or a 1 in a file and update associated flag
1036 * bit: the bit to update (CS_CPU_EXCLUSIVE, CS_MEM_EXCLUSIVE,
1037 * CS_NOTIFY_ON_RELEASE, CS_MEMORY_MIGRATE,
1038 * CS_SPREAD_PAGE, CS_SPREAD_SLAB)
1039 * cs: the cpuset to update
1040 * buf: the buffer where we read the 0 or 1
1042 * Call with manage_mutex held.
1045 static int update_flag(cpuset_flagbits_t bit, struct cpuset *cs, char *buf)
1047 int turning_on;
1048 struct cpuset trialcs;
1049 int err, cpu_exclusive_changed;
1051 turning_on = (simple_strtoul(buf, NULL, 10) != 0);
1053 trialcs = *cs;
1054 if (turning_on)
1055 set_bit(bit, &trialcs.flags);
1056 else
1057 clear_bit(bit, &trialcs.flags);
1059 err = validate_change(cs, &trialcs);
1060 if (err < 0)
1061 return err;
1062 cpu_exclusive_changed =
1063 (is_cpu_exclusive(cs) != is_cpu_exclusive(&trialcs));
1064 mutex_lock(&callback_mutex);
1065 cs->flags = trialcs.flags;
1066 mutex_unlock(&callback_mutex);
1068 if (cpu_exclusive_changed)
1069 update_cpu_domains(cs);
1070 return 0;
1074 * Frequency meter - How fast is some event occurring?
1076 * These routines manage a digitally filtered, constant time based,
1077 * event frequency meter. There are four routines:
1078 * fmeter_init() - initialize a frequency meter.
1079 * fmeter_markevent() - called each time the event happens.
1080 * fmeter_getrate() - returns the recent rate of such events.
1081 * fmeter_update() - internal routine used to update fmeter.
1083 * A common data structure is passed to each of these routines,
1084 * which is used to keep track of the state required to manage the
1085 * frequency meter and its digital filter.
1087 * The filter works on the number of events marked per unit time.
1088 * The filter is single-pole low-pass recursive (IIR). The time unit
1089 * is 1 second. Arithmetic is done using 32-bit integers scaled to
1090 * simulate 3 decimal digits of precision (multiplied by 1000).
1092 * With an FM_COEF of 933, and a time base of 1 second, the filter
1093 * has a half-life of 10 seconds, meaning that if the events quit
1094 * happening, then the rate returned from the fmeter_getrate()
1095 * will be cut in half each 10 seconds, until it converges to zero.
1097 * It is not worth doing a real infinitely recursive filter. If more
1098 * than FM_MAXTICKS ticks have elapsed since the last filter event,
1099 * just compute FM_MAXTICKS ticks worth, by which point the level
1100 * will be stable.
1102 * Limit the count of unprocessed events to FM_MAXCNT, so as to avoid
1103 * arithmetic overflow in the fmeter_update() routine.
1105 * Given the simple 32 bit integer arithmetic used, this meter works
1106 * best for reporting rates between one per millisecond (msec) and
1107 * one per 32 (approx) seconds. At constant rates faster than one
1108 * per msec it maxes out at values just under 1,000,000. At constant
1109 * rates between one per msec, and one per second it will stabilize
1110 * to a value N*1000, where N is the rate of events per second.
1111 * At constant rates between one per second and one per 32 seconds,
1112 * it will be choppy, moving up on the seconds that have an event,
1113 * and then decaying until the next event. At rates slower than
1114 * about one in 32 seconds, it decays all the way back to zero between
1115 * each event.
1118 #define FM_COEF 933 /* coefficient for half-life of 10 secs */
1119 #define FM_MAXTICKS ((time_t)99) /* useless computing more ticks than this */
1120 #define FM_MAXCNT 1000000 /* limit cnt to avoid overflow */
1121 #define FM_SCALE 1000 /* faux fixed point scale */
1123 /* Initialize a frequency meter */
1124 static void fmeter_init(struct fmeter *fmp)
1126 fmp->cnt = 0;
1127 fmp->val = 0;
1128 fmp->time = 0;
1129 spin_lock_init(&fmp->lock);
1132 /* Internal meter update - process cnt events and update value */
1133 static void fmeter_update(struct fmeter *fmp)
1135 time_t now = get_seconds();
1136 time_t ticks = now - fmp->time;
1138 if (ticks == 0)
1139 return;
1141 ticks = min(FM_MAXTICKS, ticks);
1142 while (ticks-- > 0)
1143 fmp->val = (FM_COEF * fmp->val) / FM_SCALE;
1144 fmp->time = now;
1146 fmp->val += ((FM_SCALE - FM_COEF) * fmp->cnt) / FM_SCALE;
1147 fmp->cnt = 0;
1150 /* Process any previous ticks, then bump cnt by one (times scale). */
1151 static void fmeter_markevent(struct fmeter *fmp)
1153 spin_lock(&fmp->lock);
1154 fmeter_update(fmp);
1155 fmp->cnt = min(FM_MAXCNT, fmp->cnt + FM_SCALE);
1156 spin_unlock(&fmp->lock);
1159 /* Process any previous ticks, then return current value. */
1160 static int fmeter_getrate(struct fmeter *fmp)
1162 int val;
1164 spin_lock(&fmp->lock);
1165 fmeter_update(fmp);
1166 val = fmp->val;
1167 spin_unlock(&fmp->lock);
1168 return val;
1172 * Attack task specified by pid in 'pidbuf' to cpuset 'cs', possibly
1173 * writing the path of the old cpuset in 'ppathbuf' if it needs to be
1174 * notified on release.
1176 * Call holding manage_mutex. May take callback_mutex and task_lock of
1177 * the task 'pid' during call.
1180 static int attach_task(struct cpuset *cs, char *pidbuf, char **ppathbuf)
1182 pid_t pid;
1183 struct task_struct *tsk;
1184 struct cpuset *oldcs;
1185 cpumask_t cpus;
1186 nodemask_t from, to;
1187 struct mm_struct *mm;
1188 int retval;
1190 if (sscanf(pidbuf, "%d", &pid) != 1)
1191 return -EIO;
1192 if (cpus_empty(cs->cpus_allowed) || nodes_empty(cs->mems_allowed))
1193 return -ENOSPC;
1195 if (pid) {
1196 read_lock(&tasklist_lock);
1198 tsk = find_task_by_pid(pid);
1199 if (!tsk || tsk->flags & PF_EXITING) {
1200 read_unlock(&tasklist_lock);
1201 return -ESRCH;
1204 get_task_struct(tsk);
1205 read_unlock(&tasklist_lock);
1207 if ((current->euid) && (current->euid != tsk->uid)
1208 && (current->euid != tsk->suid)) {
1209 put_task_struct(tsk);
1210 return -EACCES;
1212 } else {
1213 tsk = current;
1214 get_task_struct(tsk);
1217 retval = security_task_setscheduler(tsk, 0, NULL);
1218 if (retval) {
1219 put_task_struct(tsk);
1220 return retval;
1223 mutex_lock(&callback_mutex);
1225 task_lock(tsk);
1226 oldcs = tsk->cpuset;
1228 * After getting 'oldcs' cpuset ptr, be sure still not exiting.
1229 * If 'oldcs' might be the top_cpuset due to the_top_cpuset_hack
1230 * then fail this attach_task(), to avoid breaking top_cpuset.count.
1232 if (tsk->flags & PF_EXITING) {
1233 task_unlock(tsk);
1234 mutex_unlock(&callback_mutex);
1235 put_task_struct(tsk);
1236 return -ESRCH;
1238 atomic_inc(&cs->count);
1239 rcu_assign_pointer(tsk->cpuset, cs);
1240 task_unlock(tsk);
1242 guarantee_online_cpus(cs, &cpus);
1243 set_cpus_allowed(tsk, cpus);
1245 from = oldcs->mems_allowed;
1246 to = cs->mems_allowed;
1248 mutex_unlock(&callback_mutex);
1250 mm = get_task_mm(tsk);
1251 if (mm) {
1252 mpol_rebind_mm(mm, &to);
1253 if (is_memory_migrate(cs))
1254 cpuset_migrate_mm(mm, &from, &to);
1255 mmput(mm);
1258 put_task_struct(tsk);
1259 synchronize_rcu();
1260 if (atomic_dec_and_test(&oldcs->count))
1261 check_for_release(oldcs, ppathbuf);
1262 return 0;
1265 /* The various types of files and directories in a cpuset file system */
1267 typedef enum {
1268 FILE_ROOT,
1269 FILE_DIR,
1270 FILE_MEMORY_MIGRATE,
1271 FILE_CPULIST,
1272 FILE_MEMLIST,
1273 FILE_CPU_EXCLUSIVE,
1274 FILE_MEM_EXCLUSIVE,
1275 FILE_NOTIFY_ON_RELEASE,
1276 FILE_MEMORY_PRESSURE_ENABLED,
1277 FILE_MEMORY_PRESSURE,
1278 FILE_SPREAD_PAGE,
1279 FILE_SPREAD_SLAB,
1280 FILE_TASKLIST,
1281 } cpuset_filetype_t;
1283 static ssize_t cpuset_common_file_write(struct file *file,
1284 const char __user *userbuf,
1285 size_t nbytes, loff_t *unused_ppos)
1287 struct cpuset *cs = __d_cs(file->f_path.dentry->d_parent);
1288 struct cftype *cft = __d_cft(file->f_path.dentry);
1289 cpuset_filetype_t type = cft->private;
1290 char *buffer;
1291 char *pathbuf = NULL;
1292 int retval = 0;
1294 /* Crude upper limit on largest legitimate cpulist user might write. */
1295 if (nbytes > 100 + 6 * max(NR_CPUS, MAX_NUMNODES))
1296 return -E2BIG;
1298 /* +1 for nul-terminator */
1299 if ((buffer = kmalloc(nbytes + 1, GFP_KERNEL)) == 0)
1300 return -ENOMEM;
1302 if (copy_from_user(buffer, userbuf, nbytes)) {
1303 retval = -EFAULT;
1304 goto out1;
1306 buffer[nbytes] = 0; /* nul-terminate */
1308 mutex_lock(&manage_mutex);
1310 if (is_removed(cs)) {
1311 retval = -ENODEV;
1312 goto out2;
1315 switch (type) {
1316 case FILE_CPULIST:
1317 retval = update_cpumask(cs, buffer);
1318 break;
1319 case FILE_MEMLIST:
1320 retval = update_nodemask(cs, buffer);
1321 break;
1322 case FILE_CPU_EXCLUSIVE:
1323 retval = update_flag(CS_CPU_EXCLUSIVE, cs, buffer);
1324 break;
1325 case FILE_MEM_EXCLUSIVE:
1326 retval = update_flag(CS_MEM_EXCLUSIVE, cs, buffer);
1327 break;
1328 case FILE_NOTIFY_ON_RELEASE:
1329 retval = update_flag(CS_NOTIFY_ON_RELEASE, cs, buffer);
1330 break;
1331 case FILE_MEMORY_MIGRATE:
1332 retval = update_flag(CS_MEMORY_MIGRATE, cs, buffer);
1333 break;
1334 case FILE_MEMORY_PRESSURE_ENABLED:
1335 retval = update_memory_pressure_enabled(cs, buffer);
1336 break;
1337 case FILE_MEMORY_PRESSURE:
1338 retval = -EACCES;
1339 break;
1340 case FILE_SPREAD_PAGE:
1341 retval = update_flag(CS_SPREAD_PAGE, cs, buffer);
1342 cs->mems_generation = cpuset_mems_generation++;
1343 break;
1344 case FILE_SPREAD_SLAB:
1345 retval = update_flag(CS_SPREAD_SLAB, cs, buffer);
1346 cs->mems_generation = cpuset_mems_generation++;
1347 break;
1348 case FILE_TASKLIST:
1349 retval = attach_task(cs, buffer, &pathbuf);
1350 break;
1351 default:
1352 retval = -EINVAL;
1353 goto out2;
1356 if (retval == 0)
1357 retval = nbytes;
1358 out2:
1359 mutex_unlock(&manage_mutex);
1360 cpuset_release_agent(pathbuf);
1361 out1:
1362 kfree(buffer);
1363 return retval;
1366 static ssize_t cpuset_file_write(struct file *file, const char __user *buf,
1367 size_t nbytes, loff_t *ppos)
1369 ssize_t retval = 0;
1370 struct cftype *cft = __d_cft(file->f_path.dentry);
1371 if (!cft)
1372 return -ENODEV;
1374 /* special function ? */
1375 if (cft->write)
1376 retval = cft->write(file, buf, nbytes, ppos);
1377 else
1378 retval = cpuset_common_file_write(file, buf, nbytes, ppos);
1380 return retval;
1384 * These ascii lists should be read in a single call, by using a user
1385 * buffer large enough to hold the entire map. If read in smaller
1386 * chunks, there is no guarantee of atomicity. Since the display format
1387 * used, list of ranges of sequential numbers, is variable length,
1388 * and since these maps can change value dynamically, one could read
1389 * gibberish by doing partial reads while a list was changing.
1390 * A single large read to a buffer that crosses a page boundary is
1391 * ok, because the result being copied to user land is not recomputed
1392 * across a page fault.
1395 static int cpuset_sprintf_cpulist(char *page, struct cpuset *cs)
1397 cpumask_t mask;
1399 mutex_lock(&callback_mutex);
1400 mask = cs->cpus_allowed;
1401 mutex_unlock(&callback_mutex);
1403 return cpulist_scnprintf(page, PAGE_SIZE, mask);
1406 static int cpuset_sprintf_memlist(char *page, struct cpuset *cs)
1408 nodemask_t mask;
1410 mutex_lock(&callback_mutex);
1411 mask = cs->mems_allowed;
1412 mutex_unlock(&callback_mutex);
1414 return nodelist_scnprintf(page, PAGE_SIZE, mask);
1417 static ssize_t cpuset_common_file_read(struct file *file, char __user *buf,
1418 size_t nbytes, loff_t *ppos)
1420 struct cftype *cft = __d_cft(file->f_path.dentry);
1421 struct cpuset *cs = __d_cs(file->f_path.dentry->d_parent);
1422 cpuset_filetype_t type = cft->private;
1423 char *page;
1424 ssize_t retval = 0;
1425 char *s;
1427 if (!(page = (char *)__get_free_page(GFP_KERNEL)))
1428 return -ENOMEM;
1430 s = page;
1432 switch (type) {
1433 case FILE_CPULIST:
1434 s += cpuset_sprintf_cpulist(s, cs);
1435 break;
1436 case FILE_MEMLIST:
1437 s += cpuset_sprintf_memlist(s, cs);
1438 break;
1439 case FILE_CPU_EXCLUSIVE:
1440 *s++ = is_cpu_exclusive(cs) ? '1' : '0';
1441 break;
1442 case FILE_MEM_EXCLUSIVE:
1443 *s++ = is_mem_exclusive(cs) ? '1' : '0';
1444 break;
1445 case FILE_NOTIFY_ON_RELEASE:
1446 *s++ = notify_on_release(cs) ? '1' : '0';
1447 break;
1448 case FILE_MEMORY_MIGRATE:
1449 *s++ = is_memory_migrate(cs) ? '1' : '0';
1450 break;
1451 case FILE_MEMORY_PRESSURE_ENABLED:
1452 *s++ = cpuset_memory_pressure_enabled ? '1' : '0';
1453 break;
1454 case FILE_MEMORY_PRESSURE:
1455 s += sprintf(s, "%d", fmeter_getrate(&cs->fmeter));
1456 break;
1457 case FILE_SPREAD_PAGE:
1458 *s++ = is_spread_page(cs) ? '1' : '0';
1459 break;
1460 case FILE_SPREAD_SLAB:
1461 *s++ = is_spread_slab(cs) ? '1' : '0';
1462 break;
1463 default:
1464 retval = -EINVAL;
1465 goto out;
1467 *s++ = '\n';
1469 retval = simple_read_from_buffer(buf, nbytes, ppos, page, s - page);
1470 out:
1471 free_page((unsigned long)page);
1472 return retval;
1475 static ssize_t cpuset_file_read(struct file *file, char __user *buf, size_t nbytes,
1476 loff_t *ppos)
1478 ssize_t retval = 0;
1479 struct cftype *cft = __d_cft(file->f_path.dentry);
1480 if (!cft)
1481 return -ENODEV;
1483 /* special function ? */
1484 if (cft->read)
1485 retval = cft->read(file, buf, nbytes, ppos);
1486 else
1487 retval = cpuset_common_file_read(file, buf, nbytes, ppos);
1489 return retval;
1492 static int cpuset_file_open(struct inode *inode, struct file *file)
1494 int err;
1495 struct cftype *cft;
1497 err = generic_file_open(inode, file);
1498 if (err)
1499 return err;
1501 cft = __d_cft(file->f_path.dentry);
1502 if (!cft)
1503 return -ENODEV;
1504 if (cft->open)
1505 err = cft->open(inode, file);
1506 else
1507 err = 0;
1509 return err;
1512 static int cpuset_file_release(struct inode *inode, struct file *file)
1514 struct cftype *cft = __d_cft(file->f_path.dentry);
1515 if (cft->release)
1516 return cft->release(inode, file);
1517 return 0;
1521 * cpuset_rename - Only allow simple rename of directories in place.
1523 static int cpuset_rename(struct inode *old_dir, struct dentry *old_dentry,
1524 struct inode *new_dir, struct dentry *new_dentry)
1526 if (!S_ISDIR(old_dentry->d_inode->i_mode))
1527 return -ENOTDIR;
1528 if (new_dentry->d_inode)
1529 return -EEXIST;
1530 if (old_dir != new_dir)
1531 return -EIO;
1532 return simple_rename(old_dir, old_dentry, new_dir, new_dentry);
1535 static const struct file_operations cpuset_file_operations = {
1536 .read = cpuset_file_read,
1537 .write = cpuset_file_write,
1538 .llseek = generic_file_llseek,
1539 .open = cpuset_file_open,
1540 .release = cpuset_file_release,
1543 static const struct inode_operations cpuset_dir_inode_operations = {
1544 .lookup = simple_lookup,
1545 .mkdir = cpuset_mkdir,
1546 .rmdir = cpuset_rmdir,
1547 .rename = cpuset_rename,
1550 static int cpuset_create_file(struct dentry *dentry, int mode)
1552 struct inode *inode;
1554 if (!dentry)
1555 return -ENOENT;
1556 if (dentry->d_inode)
1557 return -EEXIST;
1559 inode = cpuset_new_inode(mode);
1560 if (!inode)
1561 return -ENOMEM;
1563 if (S_ISDIR(mode)) {
1564 inode->i_op = &cpuset_dir_inode_operations;
1565 inode->i_fop = &simple_dir_operations;
1567 /* start off with i_nlink == 2 (for "." entry) */
1568 inc_nlink(inode);
1569 } else if (S_ISREG(mode)) {
1570 inode->i_size = 0;
1571 inode->i_fop = &cpuset_file_operations;
1574 d_instantiate(dentry, inode);
1575 dget(dentry); /* Extra count - pin the dentry in core */
1576 return 0;
1580 * cpuset_create_dir - create a directory for an object.
1581 * cs: the cpuset we create the directory for.
1582 * It must have a valid ->parent field
1583 * And we are going to fill its ->dentry field.
1584 * name: The name to give to the cpuset directory. Will be copied.
1585 * mode: mode to set on new directory.
1588 static int cpuset_create_dir(struct cpuset *cs, const char *name, int mode)
1590 struct dentry *dentry = NULL;
1591 struct dentry *parent;
1592 int error = 0;
1594 parent = cs->parent->dentry;
1595 dentry = cpuset_get_dentry(parent, name);
1596 if (IS_ERR(dentry))
1597 return PTR_ERR(dentry);
1598 error = cpuset_create_file(dentry, S_IFDIR | mode);
1599 if (!error) {
1600 dentry->d_fsdata = cs;
1601 inc_nlink(parent->d_inode);
1602 cs->dentry = dentry;
1604 dput(dentry);
1606 return error;
1609 static int cpuset_add_file(struct dentry *dir, const struct cftype *cft)
1611 struct dentry *dentry;
1612 int error;
1614 mutex_lock(&dir->d_inode->i_mutex);
1615 dentry = cpuset_get_dentry(dir, cft->name);
1616 if (!IS_ERR(dentry)) {
1617 error = cpuset_create_file(dentry, 0644 | S_IFREG);
1618 if (!error)
1619 dentry->d_fsdata = (void *)cft;
1620 dput(dentry);
1621 } else
1622 error = PTR_ERR(dentry);
1623 mutex_unlock(&dir->d_inode->i_mutex);
1624 return error;
1628 * Stuff for reading the 'tasks' file.
1630 * Reading this file can return large amounts of data if a cpuset has
1631 * *lots* of attached tasks. So it may need several calls to read(),
1632 * but we cannot guarantee that the information we produce is correct
1633 * unless we produce it entirely atomically.
1635 * Upon tasks file open(), a struct ctr_struct is allocated, that
1636 * will have a pointer to an array (also allocated here). The struct
1637 * ctr_struct * is stored in file->private_data. Its resources will
1638 * be freed by release() when the file is closed. The array is used
1639 * to sprintf the PIDs and then used by read().
1642 /* cpusets_tasks_read array */
1644 struct ctr_struct {
1645 char *buf;
1646 int bufsz;
1650 * Load into 'pidarray' up to 'npids' of the tasks using cpuset 'cs'.
1651 * Return actual number of pids loaded. No need to task_lock(p)
1652 * when reading out p->cpuset, as we don't really care if it changes
1653 * on the next cycle, and we are not going to try to dereference it.
1655 static int pid_array_load(pid_t *pidarray, int npids, struct cpuset *cs)
1657 int n = 0;
1658 struct task_struct *g, *p;
1660 read_lock(&tasklist_lock);
1662 do_each_thread(g, p) {
1663 if (p->cpuset == cs) {
1664 pidarray[n++] = p->pid;
1665 if (unlikely(n == npids))
1666 goto array_full;
1668 } while_each_thread(g, p);
1670 array_full:
1671 read_unlock(&tasklist_lock);
1672 return n;
1675 static int cmppid(const void *a, const void *b)
1677 return *(pid_t *)a - *(pid_t *)b;
1681 * Convert array 'a' of 'npids' pid_t's to a string of newline separated
1682 * decimal pids in 'buf'. Don't write more than 'sz' chars, but return
1683 * count 'cnt' of how many chars would be written if buf were large enough.
1685 static int pid_array_to_buf(char *buf, int sz, pid_t *a, int npids)
1687 int cnt = 0;
1688 int i;
1690 for (i = 0; i < npids; i++)
1691 cnt += snprintf(buf + cnt, max(sz - cnt, 0), "%d\n", a[i]);
1692 return cnt;
1696 * Handle an open on 'tasks' file. Prepare a buffer listing the
1697 * process id's of tasks currently attached to the cpuset being opened.
1699 * Does not require any specific cpuset mutexes, and does not take any.
1701 static int cpuset_tasks_open(struct inode *unused, struct file *file)
1703 struct cpuset *cs = __d_cs(file->f_path.dentry->d_parent);
1704 struct ctr_struct *ctr;
1705 pid_t *pidarray;
1706 int npids;
1707 char c;
1709 if (!(file->f_mode & FMODE_READ))
1710 return 0;
1712 ctr = kmalloc(sizeof(*ctr), GFP_KERNEL);
1713 if (!ctr)
1714 goto err0;
1717 * If cpuset gets more users after we read count, we won't have
1718 * enough space - tough. This race is indistinguishable to the
1719 * caller from the case that the additional cpuset users didn't
1720 * show up until sometime later on.
1722 npids = atomic_read(&cs->count);
1723 pidarray = kmalloc(npids * sizeof(pid_t), GFP_KERNEL);
1724 if (!pidarray)
1725 goto err1;
1727 npids = pid_array_load(pidarray, npids, cs);
1728 sort(pidarray, npids, sizeof(pid_t), cmppid, NULL);
1730 /* Call pid_array_to_buf() twice, first just to get bufsz */
1731 ctr->bufsz = pid_array_to_buf(&c, sizeof(c), pidarray, npids) + 1;
1732 ctr->buf = kmalloc(ctr->bufsz, GFP_KERNEL);
1733 if (!ctr->buf)
1734 goto err2;
1735 ctr->bufsz = pid_array_to_buf(ctr->buf, ctr->bufsz, pidarray, npids);
1737 kfree(pidarray);
1738 file->private_data = ctr;
1739 return 0;
1741 err2:
1742 kfree(pidarray);
1743 err1:
1744 kfree(ctr);
1745 err0:
1746 return -ENOMEM;
1749 static ssize_t cpuset_tasks_read(struct file *file, char __user *buf,
1750 size_t nbytes, loff_t *ppos)
1752 struct ctr_struct *ctr = file->private_data;
1754 if (*ppos + nbytes > ctr->bufsz)
1755 nbytes = ctr->bufsz - *ppos;
1756 if (copy_to_user(buf, ctr->buf + *ppos, nbytes))
1757 return -EFAULT;
1758 *ppos += nbytes;
1759 return nbytes;
1762 static int cpuset_tasks_release(struct inode *unused_inode, struct file *file)
1764 struct ctr_struct *ctr;
1766 if (file->f_mode & FMODE_READ) {
1767 ctr = file->private_data;
1768 kfree(ctr->buf);
1769 kfree(ctr);
1771 return 0;
1775 * for the common functions, 'private' gives the type of file
1778 static struct cftype cft_tasks = {
1779 .name = "tasks",
1780 .open = cpuset_tasks_open,
1781 .read = cpuset_tasks_read,
1782 .release = cpuset_tasks_release,
1783 .private = FILE_TASKLIST,
1786 static struct cftype cft_cpus = {
1787 .name = "cpus",
1788 .private = FILE_CPULIST,
1791 static struct cftype cft_mems = {
1792 .name = "mems",
1793 .private = FILE_MEMLIST,
1796 static struct cftype cft_cpu_exclusive = {
1797 .name = "cpu_exclusive",
1798 .private = FILE_CPU_EXCLUSIVE,
1801 static struct cftype cft_mem_exclusive = {
1802 .name = "mem_exclusive",
1803 .private = FILE_MEM_EXCLUSIVE,
1806 static struct cftype cft_notify_on_release = {
1807 .name = "notify_on_release",
1808 .private = FILE_NOTIFY_ON_RELEASE,
1811 static struct cftype cft_memory_migrate = {
1812 .name = "memory_migrate",
1813 .private = FILE_MEMORY_MIGRATE,
1816 static struct cftype cft_memory_pressure_enabled = {
1817 .name = "memory_pressure_enabled",
1818 .private = FILE_MEMORY_PRESSURE_ENABLED,
1821 static struct cftype cft_memory_pressure = {
1822 .name = "memory_pressure",
1823 .private = FILE_MEMORY_PRESSURE,
1826 static struct cftype cft_spread_page = {
1827 .name = "memory_spread_page",
1828 .private = FILE_SPREAD_PAGE,
1831 static struct cftype cft_spread_slab = {
1832 .name = "memory_spread_slab",
1833 .private = FILE_SPREAD_SLAB,
1836 static int cpuset_populate_dir(struct dentry *cs_dentry)
1838 int err;
1840 if ((err = cpuset_add_file(cs_dentry, &cft_cpus)) < 0)
1841 return err;
1842 if ((err = cpuset_add_file(cs_dentry, &cft_mems)) < 0)
1843 return err;
1844 if ((err = cpuset_add_file(cs_dentry, &cft_cpu_exclusive)) < 0)
1845 return err;
1846 if ((err = cpuset_add_file(cs_dentry, &cft_mem_exclusive)) < 0)
1847 return err;
1848 if ((err = cpuset_add_file(cs_dentry, &cft_notify_on_release)) < 0)
1849 return err;
1850 if ((err = cpuset_add_file(cs_dentry, &cft_memory_migrate)) < 0)
1851 return err;
1852 if ((err = cpuset_add_file(cs_dentry, &cft_memory_pressure)) < 0)
1853 return err;
1854 if ((err = cpuset_add_file(cs_dentry, &cft_spread_page)) < 0)
1855 return err;
1856 if ((err = cpuset_add_file(cs_dentry, &cft_spread_slab)) < 0)
1857 return err;
1858 if ((err = cpuset_add_file(cs_dentry, &cft_tasks)) < 0)
1859 return err;
1860 return 0;
1864 * cpuset_create - create a cpuset
1865 * parent: cpuset that will be parent of the new cpuset.
1866 * name: name of the new cpuset. Will be strcpy'ed.
1867 * mode: mode to set on new inode
1869 * Must be called with the mutex on the parent inode held
1872 static long cpuset_create(struct cpuset *parent, const char *name, int mode)
1874 struct cpuset *cs;
1875 int err;
1877 cs = kmalloc(sizeof(*cs), GFP_KERNEL);
1878 if (!cs)
1879 return -ENOMEM;
1881 mutex_lock(&manage_mutex);
1882 cpuset_update_task_memory_state();
1883 cs->flags = 0;
1884 if (notify_on_release(parent))
1885 set_bit(CS_NOTIFY_ON_RELEASE, &cs->flags);
1886 if (is_spread_page(parent))
1887 set_bit(CS_SPREAD_PAGE, &cs->flags);
1888 if (is_spread_slab(parent))
1889 set_bit(CS_SPREAD_SLAB, &cs->flags);
1890 cs->cpus_allowed = CPU_MASK_NONE;
1891 cs->mems_allowed = NODE_MASK_NONE;
1892 atomic_set(&cs->count, 0);
1893 INIT_LIST_HEAD(&cs->sibling);
1894 INIT_LIST_HEAD(&cs->children);
1895 cs->mems_generation = cpuset_mems_generation++;
1896 fmeter_init(&cs->fmeter);
1898 cs->parent = parent;
1900 mutex_lock(&callback_mutex);
1901 list_add(&cs->sibling, &cs->parent->children);
1902 number_of_cpusets++;
1903 mutex_unlock(&callback_mutex);
1905 err = cpuset_create_dir(cs, name, mode);
1906 if (err < 0)
1907 goto err;
1910 * Release manage_mutex before cpuset_populate_dir() because it
1911 * will down() this new directory's i_mutex and if we race with
1912 * another mkdir, we might deadlock.
1914 mutex_unlock(&manage_mutex);
1916 err = cpuset_populate_dir(cs->dentry);
1917 /* If err < 0, we have a half-filled directory - oh well ;) */
1918 return 0;
1919 err:
1920 list_del(&cs->sibling);
1921 mutex_unlock(&manage_mutex);
1922 kfree(cs);
1923 return err;
1926 static int cpuset_mkdir(struct inode *dir, struct dentry *dentry, int mode)
1928 struct cpuset *c_parent = dentry->d_parent->d_fsdata;
1930 /* the vfs holds inode->i_mutex already */
1931 return cpuset_create(c_parent, dentry->d_name.name, mode | S_IFDIR);
1935 * Locking note on the strange update_flag() call below:
1937 * If the cpuset being removed is marked cpu_exclusive, then simulate
1938 * turning cpu_exclusive off, which will call update_cpu_domains().
1939 * The lock_cpu_hotplug() call in update_cpu_domains() must not be
1940 * made while holding callback_mutex. Elsewhere the kernel nests
1941 * callback_mutex inside lock_cpu_hotplug() calls. So the reverse
1942 * nesting would risk an ABBA deadlock.
1945 static int cpuset_rmdir(struct inode *unused_dir, struct dentry *dentry)
1947 struct cpuset *cs = dentry->d_fsdata;
1948 struct dentry *d;
1949 struct cpuset *parent;
1950 char *pathbuf = NULL;
1952 /* the vfs holds both inode->i_mutex already */
1954 mutex_lock(&manage_mutex);
1955 cpuset_update_task_memory_state();
1956 if (atomic_read(&cs->count) > 0) {
1957 mutex_unlock(&manage_mutex);
1958 return -EBUSY;
1960 if (!list_empty(&cs->children)) {
1961 mutex_unlock(&manage_mutex);
1962 return -EBUSY;
1964 if (is_cpu_exclusive(cs)) {
1965 int retval = update_flag(CS_CPU_EXCLUSIVE, cs, "0");
1966 if (retval < 0) {
1967 mutex_unlock(&manage_mutex);
1968 return retval;
1971 parent = cs->parent;
1972 mutex_lock(&callback_mutex);
1973 set_bit(CS_REMOVED, &cs->flags);
1974 list_del(&cs->sibling); /* delete my sibling from parent->children */
1975 spin_lock(&cs->dentry->d_lock);
1976 d = dget(cs->dentry);
1977 cs->dentry = NULL;
1978 spin_unlock(&d->d_lock);
1979 cpuset_d_remove_dir(d);
1980 dput(d);
1981 number_of_cpusets--;
1982 mutex_unlock(&callback_mutex);
1983 if (list_empty(&parent->children))
1984 check_for_release(parent, &pathbuf);
1985 mutex_unlock(&manage_mutex);
1986 cpuset_release_agent(pathbuf);
1987 return 0;
1991 * cpuset_init_early - just enough so that the calls to
1992 * cpuset_update_task_memory_state() in early init code
1993 * are harmless.
1996 int __init cpuset_init_early(void)
1998 struct task_struct *tsk = current;
2000 tsk->cpuset = &top_cpuset;
2001 tsk->cpuset->mems_generation = cpuset_mems_generation++;
2002 return 0;
2006 * cpuset_init - initialize cpusets at system boot
2008 * Description: Initialize top_cpuset and the cpuset internal file system,
2011 int __init cpuset_init(void)
2013 struct dentry *root;
2014 int err;
2016 top_cpuset.cpus_allowed = CPU_MASK_ALL;
2017 top_cpuset.mems_allowed = NODE_MASK_ALL;
2019 fmeter_init(&top_cpuset.fmeter);
2020 top_cpuset.mems_generation = cpuset_mems_generation++;
2022 init_task.cpuset = &top_cpuset;
2024 err = register_filesystem(&cpuset_fs_type);
2025 if (err < 0)
2026 goto out;
2027 cpuset_mount = kern_mount(&cpuset_fs_type);
2028 if (IS_ERR(cpuset_mount)) {
2029 printk(KERN_ERR "cpuset: could not mount!\n");
2030 err = PTR_ERR(cpuset_mount);
2031 cpuset_mount = NULL;
2032 goto out;
2034 root = cpuset_mount->mnt_sb->s_root;
2035 root->d_fsdata = &top_cpuset;
2036 inc_nlink(root->d_inode);
2037 top_cpuset.dentry = root;
2038 root->d_inode->i_op = &cpuset_dir_inode_operations;
2039 number_of_cpusets = 1;
2040 err = cpuset_populate_dir(root);
2041 /* memory_pressure_enabled is in root cpuset only */
2042 if (err == 0)
2043 err = cpuset_add_file(root, &cft_memory_pressure_enabled);
2044 out:
2045 return err;
2049 * If common_cpu_mem_hotplug_unplug(), below, unplugs any CPUs
2050 * or memory nodes, we need to walk over the cpuset hierarchy,
2051 * removing that CPU or node from all cpusets. If this removes the
2052 * last CPU or node from a cpuset, then the guarantee_online_cpus()
2053 * or guarantee_online_mems() code will use that emptied cpusets
2054 * parent online CPUs or nodes. Cpusets that were already empty of
2055 * CPUs or nodes are left empty.
2057 * This routine is intentionally inefficient in a couple of regards.
2058 * It will check all cpusets in a subtree even if the top cpuset of
2059 * the subtree has no offline CPUs or nodes. It checks both CPUs and
2060 * nodes, even though the caller could have been coded to know that
2061 * only one of CPUs or nodes needed to be checked on a given call.
2062 * This was done to minimize text size rather than cpu cycles.
2064 * Call with both manage_mutex and callback_mutex held.
2066 * Recursive, on depth of cpuset subtree.
2069 static void guarantee_online_cpus_mems_in_subtree(const struct cpuset *cur)
2071 struct cpuset *c;
2073 /* Each of our child cpusets mems must be online */
2074 list_for_each_entry(c, &cur->children, sibling) {
2075 guarantee_online_cpus_mems_in_subtree(c);
2076 if (!cpus_empty(c->cpus_allowed))
2077 guarantee_online_cpus(c, &c->cpus_allowed);
2078 if (!nodes_empty(c->mems_allowed))
2079 guarantee_online_mems(c, &c->mems_allowed);
2084 * The cpus_allowed and mems_allowed nodemasks in the top_cpuset track
2085 * cpu_online_map and node_online_map. Force the top cpuset to track
2086 * whats online after any CPU or memory node hotplug or unplug event.
2088 * To ensure that we don't remove a CPU or node from the top cpuset
2089 * that is currently in use by a child cpuset (which would violate
2090 * the rule that cpusets must be subsets of their parent), we first
2091 * call the recursive routine guarantee_online_cpus_mems_in_subtree().
2093 * Since there are two callers of this routine, one for CPU hotplug
2094 * events and one for memory node hotplug events, we could have coded
2095 * two separate routines here. We code it as a single common routine
2096 * in order to minimize text size.
2099 static void common_cpu_mem_hotplug_unplug(void)
2101 mutex_lock(&manage_mutex);
2102 mutex_lock(&callback_mutex);
2104 guarantee_online_cpus_mems_in_subtree(&top_cpuset);
2105 top_cpuset.cpus_allowed = cpu_online_map;
2106 top_cpuset.mems_allowed = node_online_map;
2108 mutex_unlock(&callback_mutex);
2109 mutex_unlock(&manage_mutex);
2113 * The top_cpuset tracks what CPUs and Memory Nodes are online,
2114 * period. This is necessary in order to make cpusets transparent
2115 * (of no affect) on systems that are actively using CPU hotplug
2116 * but making no active use of cpusets.
2118 * This routine ensures that top_cpuset.cpus_allowed tracks
2119 * cpu_online_map on each CPU hotplug (cpuhp) event.
2122 static int cpuset_handle_cpuhp(struct notifier_block *nb,
2123 unsigned long phase, void *cpu)
2125 common_cpu_mem_hotplug_unplug();
2126 return 0;
2129 #ifdef CONFIG_MEMORY_HOTPLUG
2131 * Keep top_cpuset.mems_allowed tracking node_online_map.
2132 * Call this routine anytime after you change node_online_map.
2133 * See also the previous routine cpuset_handle_cpuhp().
2136 void cpuset_track_online_nodes(void)
2138 common_cpu_mem_hotplug_unplug();
2140 #endif
2143 * cpuset_init_smp - initialize cpus_allowed
2145 * Description: Finish top cpuset after cpu, node maps are initialized
2148 void __init cpuset_init_smp(void)
2150 top_cpuset.cpus_allowed = cpu_online_map;
2151 top_cpuset.mems_allowed = node_online_map;
2153 hotcpu_notifier(cpuset_handle_cpuhp, 0);
2157 * cpuset_fork - attach newly forked task to its parents cpuset.
2158 * @tsk: pointer to task_struct of forking parent process.
2160 * Description: A task inherits its parent's cpuset at fork().
2162 * A pointer to the shared cpuset was automatically copied in fork.c
2163 * by dup_task_struct(). However, we ignore that copy, since it was
2164 * not made under the protection of task_lock(), so might no longer be
2165 * a valid cpuset pointer. attach_task() might have already changed
2166 * current->cpuset, allowing the previously referenced cpuset to
2167 * be removed and freed. Instead, we task_lock(current) and copy
2168 * its present value of current->cpuset for our freshly forked child.
2170 * At the point that cpuset_fork() is called, 'current' is the parent
2171 * task, and the passed argument 'child' points to the child task.
2174 void cpuset_fork(struct task_struct *child)
2176 task_lock(current);
2177 child->cpuset = current->cpuset;
2178 atomic_inc(&child->cpuset->count);
2179 task_unlock(current);
2183 * cpuset_exit - detach cpuset from exiting task
2184 * @tsk: pointer to task_struct of exiting process
2186 * Description: Detach cpuset from @tsk and release it.
2188 * Note that cpusets marked notify_on_release force every task in
2189 * them to take the global manage_mutex mutex when exiting.
2190 * This could impact scaling on very large systems. Be reluctant to
2191 * use notify_on_release cpusets where very high task exit scaling
2192 * is required on large systems.
2194 * Don't even think about derefencing 'cs' after the cpuset use count
2195 * goes to zero, except inside a critical section guarded by manage_mutex
2196 * or callback_mutex. Otherwise a zero cpuset use count is a license to
2197 * any other task to nuke the cpuset immediately, via cpuset_rmdir().
2199 * This routine has to take manage_mutex, not callback_mutex, because
2200 * it is holding that mutex while calling check_for_release(),
2201 * which calls kmalloc(), so can't be called holding callback_mutex().
2203 * We don't need to task_lock() this reference to tsk->cpuset,
2204 * because tsk is already marked PF_EXITING, so attach_task() won't
2205 * mess with it, or task is a failed fork, never visible to attach_task.
2207 * the_top_cpuset_hack:
2209 * Set the exiting tasks cpuset to the root cpuset (top_cpuset).
2211 * Don't leave a task unable to allocate memory, as that is an
2212 * accident waiting to happen should someone add a callout in
2213 * do_exit() after the cpuset_exit() call that might allocate.
2214 * If a task tries to allocate memory with an invalid cpuset,
2215 * it will oops in cpuset_update_task_memory_state().
2217 * We call cpuset_exit() while the task is still competent to
2218 * handle notify_on_release(), then leave the task attached to
2219 * the root cpuset (top_cpuset) for the remainder of its exit.
2221 * To do this properly, we would increment the reference count on
2222 * top_cpuset, and near the very end of the kernel/exit.c do_exit()
2223 * code we would add a second cpuset function call, to drop that
2224 * reference. This would just create an unnecessary hot spot on
2225 * the top_cpuset reference count, to no avail.
2227 * Normally, holding a reference to a cpuset without bumping its
2228 * count is unsafe. The cpuset could go away, or someone could
2229 * attach us to a different cpuset, decrementing the count on
2230 * the first cpuset that we never incremented. But in this case,
2231 * top_cpuset isn't going away, and either task has PF_EXITING set,
2232 * which wards off any attach_task() attempts, or task is a failed
2233 * fork, never visible to attach_task.
2235 * Another way to do this would be to set the cpuset pointer
2236 * to NULL here, and check in cpuset_update_task_memory_state()
2237 * for a NULL pointer. This hack avoids that NULL check, for no
2238 * cost (other than this way too long comment ;).
2241 void cpuset_exit(struct task_struct *tsk)
2243 struct cpuset *cs;
2245 cs = tsk->cpuset;
2246 tsk->cpuset = &top_cpuset; /* the_top_cpuset_hack - see above */
2248 if (notify_on_release(cs)) {
2249 char *pathbuf = NULL;
2251 mutex_lock(&manage_mutex);
2252 if (atomic_dec_and_test(&cs->count))
2253 check_for_release(cs, &pathbuf);
2254 mutex_unlock(&manage_mutex);
2255 cpuset_release_agent(pathbuf);
2256 } else {
2257 atomic_dec(&cs->count);
2262 * cpuset_cpus_allowed - return cpus_allowed mask from a tasks cpuset.
2263 * @tsk: pointer to task_struct from which to obtain cpuset->cpus_allowed.
2265 * Description: Returns the cpumask_t cpus_allowed of the cpuset
2266 * attached to the specified @tsk. Guaranteed to return some non-empty
2267 * subset of cpu_online_map, even if this means going outside the
2268 * tasks cpuset.
2271 cpumask_t cpuset_cpus_allowed(struct task_struct *tsk)
2273 cpumask_t mask;
2275 mutex_lock(&callback_mutex);
2276 task_lock(tsk);
2277 guarantee_online_cpus(tsk->cpuset, &mask);
2278 task_unlock(tsk);
2279 mutex_unlock(&callback_mutex);
2281 return mask;
2284 void cpuset_init_current_mems_allowed(void)
2286 current->mems_allowed = NODE_MASK_ALL;
2290 * cpuset_mems_allowed - return mems_allowed mask from a tasks cpuset.
2291 * @tsk: pointer to task_struct from which to obtain cpuset->mems_allowed.
2293 * Description: Returns the nodemask_t mems_allowed of the cpuset
2294 * attached to the specified @tsk. Guaranteed to return some non-empty
2295 * subset of node_online_map, even if this means going outside the
2296 * tasks cpuset.
2299 nodemask_t cpuset_mems_allowed(struct task_struct *tsk)
2301 nodemask_t mask;
2303 mutex_lock(&callback_mutex);
2304 task_lock(tsk);
2305 guarantee_online_mems(tsk->cpuset, &mask);
2306 task_unlock(tsk);
2307 mutex_unlock(&callback_mutex);
2309 return mask;
2313 * cpuset_zonelist_valid_mems_allowed - check zonelist vs. curremt mems_allowed
2314 * @zl: the zonelist to be checked
2316 * Are any of the nodes on zonelist zl allowed in current->mems_allowed?
2318 int cpuset_zonelist_valid_mems_allowed(struct zonelist *zl)
2320 int i;
2322 for (i = 0; zl->zones[i]; i++) {
2323 int nid = zone_to_nid(zl->zones[i]);
2325 if (node_isset(nid, current->mems_allowed))
2326 return 1;
2328 return 0;
2332 * nearest_exclusive_ancestor() - Returns the nearest mem_exclusive
2333 * ancestor to the specified cpuset. Call holding callback_mutex.
2334 * If no ancestor is mem_exclusive (an unusual configuration), then
2335 * returns the root cpuset.
2337 static const struct cpuset *nearest_exclusive_ancestor(const struct cpuset *cs)
2339 while (!is_mem_exclusive(cs) && cs->parent)
2340 cs = cs->parent;
2341 return cs;
2345 * cpuset_zone_allowed_softwall - Can we allocate on zone z's memory node?
2346 * @z: is this zone on an allowed node?
2347 * @gfp_mask: memory allocation flags
2349 * If we're in interrupt, yes, we can always allocate. If
2350 * __GFP_THISNODE is set, yes, we can always allocate. If zone
2351 * z's node is in our tasks mems_allowed, yes. If it's not a
2352 * __GFP_HARDWALL request and this zone's nodes is in the nearest
2353 * mem_exclusive cpuset ancestor to this tasks cpuset, yes.
2354 * If the task has been OOM killed and has access to memory reserves
2355 * as specified by the TIF_MEMDIE flag, yes.
2356 * Otherwise, no.
2358 * If __GFP_HARDWALL is set, cpuset_zone_allowed_softwall()
2359 * reduces to cpuset_zone_allowed_hardwall(). Otherwise,
2360 * cpuset_zone_allowed_softwall() might sleep, and might allow a zone
2361 * from an enclosing cpuset.
2363 * cpuset_zone_allowed_hardwall() only handles the simpler case of
2364 * hardwall cpusets, and never sleeps.
2366 * The __GFP_THISNODE placement logic is really handled elsewhere,
2367 * by forcibly using a zonelist starting at a specified node, and by
2368 * (in get_page_from_freelist()) refusing to consider the zones for
2369 * any node on the zonelist except the first. By the time any such
2370 * calls get to this routine, we should just shut up and say 'yes'.
2372 * GFP_USER allocations are marked with the __GFP_HARDWALL bit,
2373 * and do not allow allocations outside the current tasks cpuset
2374 * unless the task has been OOM killed as is marked TIF_MEMDIE.
2375 * GFP_KERNEL allocations are not so marked, so can escape to the
2376 * nearest enclosing mem_exclusive ancestor cpuset.
2378 * Scanning up parent cpusets requires callback_mutex. The
2379 * __alloc_pages() routine only calls here with __GFP_HARDWALL bit
2380 * _not_ set if it's a GFP_KERNEL allocation, and all nodes in the
2381 * current tasks mems_allowed came up empty on the first pass over
2382 * the zonelist. So only GFP_KERNEL allocations, if all nodes in the
2383 * cpuset are short of memory, might require taking the callback_mutex
2384 * mutex.
2386 * The first call here from mm/page_alloc:get_page_from_freelist()
2387 * has __GFP_HARDWALL set in gfp_mask, enforcing hardwall cpusets,
2388 * so no allocation on a node outside the cpuset is allowed (unless
2389 * in interrupt, of course).
2391 * The second pass through get_page_from_freelist() doesn't even call
2392 * here for GFP_ATOMIC calls. For those calls, the __alloc_pages()
2393 * variable 'wait' is not set, and the bit ALLOC_CPUSET is not set
2394 * in alloc_flags. That logic and the checks below have the combined
2395 * affect that:
2396 * in_interrupt - any node ok (current task context irrelevant)
2397 * GFP_ATOMIC - any node ok
2398 * TIF_MEMDIE - any node ok
2399 * GFP_KERNEL - any node in enclosing mem_exclusive cpuset ok
2400 * GFP_USER - only nodes in current tasks mems allowed ok.
2402 * Rule:
2403 * Don't call cpuset_zone_allowed_softwall if you can't sleep, unless you
2404 * pass in the __GFP_HARDWALL flag set in gfp_flag, which disables
2405 * the code that might scan up ancestor cpusets and sleep.
2408 int __cpuset_zone_allowed_softwall(struct zone *z, gfp_t gfp_mask)
2410 int node; /* node that zone z is on */
2411 const struct cpuset *cs; /* current cpuset ancestors */
2412 int allowed; /* is allocation in zone z allowed? */
2414 if (in_interrupt() || (gfp_mask & __GFP_THISNODE))
2415 return 1;
2416 node = zone_to_nid(z);
2417 might_sleep_if(!(gfp_mask & __GFP_HARDWALL));
2418 if (node_isset(node, current->mems_allowed))
2419 return 1;
2421 * Allow tasks that have access to memory reserves because they have
2422 * been OOM killed to get memory anywhere.
2424 if (unlikely(test_thread_flag(TIF_MEMDIE)))
2425 return 1;
2426 if (gfp_mask & __GFP_HARDWALL) /* If hardwall request, stop here */
2427 return 0;
2429 if (current->flags & PF_EXITING) /* Let dying task have memory */
2430 return 1;
2432 /* Not hardwall and node outside mems_allowed: scan up cpusets */
2433 mutex_lock(&callback_mutex);
2435 task_lock(current);
2436 cs = nearest_exclusive_ancestor(current->cpuset);
2437 task_unlock(current);
2439 allowed = node_isset(node, cs->mems_allowed);
2440 mutex_unlock(&callback_mutex);
2441 return allowed;
2445 * cpuset_zone_allowed_hardwall - Can we allocate on zone z's memory node?
2446 * @z: is this zone on an allowed node?
2447 * @gfp_mask: memory allocation flags
2449 * If we're in interrupt, yes, we can always allocate.
2450 * If __GFP_THISNODE is set, yes, we can always allocate. If zone
2451 * z's node is in our tasks mems_allowed, yes. If the task has been
2452 * OOM killed and has access to memory reserves as specified by the
2453 * TIF_MEMDIE flag, yes. Otherwise, no.
2455 * The __GFP_THISNODE placement logic is really handled elsewhere,
2456 * by forcibly using a zonelist starting at a specified node, and by
2457 * (in get_page_from_freelist()) refusing to consider the zones for
2458 * any node on the zonelist except the first. By the time any such
2459 * calls get to this routine, we should just shut up and say 'yes'.
2461 * Unlike the cpuset_zone_allowed_softwall() variant, above,
2462 * this variant requires that the zone be in the current tasks
2463 * mems_allowed or that we're in interrupt. It does not scan up the
2464 * cpuset hierarchy for the nearest enclosing mem_exclusive cpuset.
2465 * It never sleeps.
2468 int __cpuset_zone_allowed_hardwall(struct zone *z, gfp_t gfp_mask)
2470 int node; /* node that zone z is on */
2472 if (in_interrupt() || (gfp_mask & __GFP_THISNODE))
2473 return 1;
2474 node = zone_to_nid(z);
2475 if (node_isset(node, current->mems_allowed))
2476 return 1;
2478 * Allow tasks that have access to memory reserves because they have
2479 * been OOM killed to get memory anywhere.
2481 if (unlikely(test_thread_flag(TIF_MEMDIE)))
2482 return 1;
2483 return 0;
2487 * cpuset_lock - lock out any changes to cpuset structures
2489 * The out of memory (oom) code needs to mutex_lock cpusets
2490 * from being changed while it scans the tasklist looking for a
2491 * task in an overlapping cpuset. Expose callback_mutex via this
2492 * cpuset_lock() routine, so the oom code can lock it, before
2493 * locking the task list. The tasklist_lock is a spinlock, so
2494 * must be taken inside callback_mutex.
2497 void cpuset_lock(void)
2499 mutex_lock(&callback_mutex);
2503 * cpuset_unlock - release lock on cpuset changes
2505 * Undo the lock taken in a previous cpuset_lock() call.
2508 void cpuset_unlock(void)
2510 mutex_unlock(&callback_mutex);
2514 * cpuset_mem_spread_node() - On which node to begin search for a page
2516 * If a task is marked PF_SPREAD_PAGE or PF_SPREAD_SLAB (as for
2517 * tasks in a cpuset with is_spread_page or is_spread_slab set),
2518 * and if the memory allocation used cpuset_mem_spread_node()
2519 * to determine on which node to start looking, as it will for
2520 * certain page cache or slab cache pages such as used for file
2521 * system buffers and inode caches, then instead of starting on the
2522 * local node to look for a free page, rather spread the starting
2523 * node around the tasks mems_allowed nodes.
2525 * We don't have to worry about the returned node being offline
2526 * because "it can't happen", and even if it did, it would be ok.
2528 * The routines calling guarantee_online_mems() are careful to
2529 * only set nodes in task->mems_allowed that are online. So it
2530 * should not be possible for the following code to return an
2531 * offline node. But if it did, that would be ok, as this routine
2532 * is not returning the node where the allocation must be, only
2533 * the node where the search should start. The zonelist passed to
2534 * __alloc_pages() will include all nodes. If the slab allocator
2535 * is passed an offline node, it will fall back to the local node.
2536 * See kmem_cache_alloc_node().
2539 int cpuset_mem_spread_node(void)
2541 int node;
2543 node = next_node(current->cpuset_mem_spread_rotor, current->mems_allowed);
2544 if (node == MAX_NUMNODES)
2545 node = first_node(current->mems_allowed);
2546 current->cpuset_mem_spread_rotor = node;
2547 return node;
2549 EXPORT_SYMBOL_GPL(cpuset_mem_spread_node);
2552 * cpuset_excl_nodes_overlap - Do we overlap @p's mem_exclusive ancestors?
2553 * @p: pointer to task_struct of some other task.
2555 * Description: Return true if the nearest mem_exclusive ancestor
2556 * cpusets of tasks @p and current overlap. Used by oom killer to
2557 * determine if task @p's memory usage might impact the memory
2558 * available to the current task.
2560 * Call while holding callback_mutex.
2563 int cpuset_excl_nodes_overlap(const struct task_struct *p)
2565 const struct cpuset *cs1, *cs2; /* my and p's cpuset ancestors */
2566 int overlap = 1; /* do cpusets overlap? */
2568 task_lock(current);
2569 if (current->flags & PF_EXITING) {
2570 task_unlock(current);
2571 goto done;
2573 cs1 = nearest_exclusive_ancestor(current->cpuset);
2574 task_unlock(current);
2576 task_lock((struct task_struct *)p);
2577 if (p->flags & PF_EXITING) {
2578 task_unlock((struct task_struct *)p);
2579 goto done;
2581 cs2 = nearest_exclusive_ancestor(p->cpuset);
2582 task_unlock((struct task_struct *)p);
2584 overlap = nodes_intersects(cs1->mems_allowed, cs2->mems_allowed);
2585 done:
2586 return overlap;
2590 * Collection of memory_pressure is suppressed unless
2591 * this flag is enabled by writing "1" to the special
2592 * cpuset file 'memory_pressure_enabled' in the root cpuset.
2595 int cpuset_memory_pressure_enabled __read_mostly;
2598 * cpuset_memory_pressure_bump - keep stats of per-cpuset reclaims.
2600 * Keep a running average of the rate of synchronous (direct)
2601 * page reclaim efforts initiated by tasks in each cpuset.
2603 * This represents the rate at which some task in the cpuset
2604 * ran low on memory on all nodes it was allowed to use, and
2605 * had to enter the kernels page reclaim code in an effort to
2606 * create more free memory by tossing clean pages or swapping
2607 * or writing dirty pages.
2609 * Display to user space in the per-cpuset read-only file
2610 * "memory_pressure". Value displayed is an integer
2611 * representing the recent rate of entry into the synchronous
2612 * (direct) page reclaim by any task attached to the cpuset.
2615 void __cpuset_memory_pressure_bump(void)
2617 struct cpuset *cs;
2619 task_lock(current);
2620 cs = current->cpuset;
2621 fmeter_markevent(&cs->fmeter);
2622 task_unlock(current);
2626 * proc_cpuset_show()
2627 * - Print tasks cpuset path into seq_file.
2628 * - Used for /proc/<pid>/cpuset.
2629 * - No need to task_lock(tsk) on this tsk->cpuset reference, as it
2630 * doesn't really matter if tsk->cpuset changes after we read it,
2631 * and we take manage_mutex, keeping attach_task() from changing it
2632 * anyway. No need to check that tsk->cpuset != NULL, thanks to
2633 * the_top_cpuset_hack in cpuset_exit(), which sets an exiting tasks
2634 * cpuset to top_cpuset.
2636 static int proc_cpuset_show(struct seq_file *m, void *v)
2638 struct pid *pid;
2639 struct task_struct *tsk;
2640 char *buf;
2641 int retval;
2643 retval = -ENOMEM;
2644 buf = kmalloc(PAGE_SIZE, GFP_KERNEL);
2645 if (!buf)
2646 goto out;
2648 retval = -ESRCH;
2649 pid = m->private;
2650 tsk = get_pid_task(pid, PIDTYPE_PID);
2651 if (!tsk)
2652 goto out_free;
2654 retval = -EINVAL;
2655 mutex_lock(&manage_mutex);
2657 retval = cpuset_path(tsk->cpuset, buf, PAGE_SIZE);
2658 if (retval < 0)
2659 goto out_unlock;
2660 seq_puts(m, buf);
2661 seq_putc(m, '\n');
2662 out_unlock:
2663 mutex_unlock(&manage_mutex);
2664 put_task_struct(tsk);
2665 out_free:
2666 kfree(buf);
2667 out:
2668 return retval;
2671 static int cpuset_open(struct inode *inode, struct file *file)
2673 struct pid *pid = PROC_I(inode)->pid;
2674 return single_open(file, proc_cpuset_show, pid);
2677 const struct file_operations proc_cpuset_operations = {
2678 .open = cpuset_open,
2679 .read = seq_read,
2680 .llseek = seq_lseek,
2681 .release = single_release,
2684 /* Display task cpus_allowed, mems_allowed in /proc/<pid>/status file. */
2685 char *cpuset_task_status_allowed(struct task_struct *task, char *buffer)
2687 buffer += sprintf(buffer, "Cpus_allowed:\t");
2688 buffer += cpumask_scnprintf(buffer, PAGE_SIZE, task->cpus_allowed);
2689 buffer += sprintf(buffer, "\n");
2690 buffer += sprintf(buffer, "Mems_allowed:\t");
2691 buffer += nodemask_scnprintf(buffer, PAGE_SIZE, task->mems_allowed);
2692 buffer += sprintf(buffer, "\n");
2693 return buffer;