ia64/linux-2.6.18-xen.hg

view kernel/cpuset.c @ 912:dd42cdb0ab89

[IA64] Build blktap2 driver by default in x86 builds.

add CONFIG_XEN_BLKDEV_TAP2=y to buildconfigs/linux-defconfig_xen_ia64.

Signed-off-by: Isaku Yamahata <yamahata@valinux.co.jp>
author Isaku Yamahata <yamahata@valinux.co.jp>
date Mon Jun 29 12:09:16 2009 +0900 (2009-06-29)
parents 831230e53067
children
line source
1 /*
2 * kernel/cpuset.c
3 *
4 * Processor and Memory placement constraints for sets of tasks.
5 *
6 * Copyright (C) 2003 BULL SA.
7 * Copyright (C) 2004-2006 Silicon Graphics, Inc.
8 *
9 * Portions derived from Patrick Mochel's sysfs code.
10 * sysfs is Copyright (c) 2001-3 Patrick Mochel
11 *
12 * 2003-10-10 Written by Simon Derr.
13 * 2003-10-22 Updates by Stephen Hemminger.
14 * 2004 May-July Rework by Paul Jackson.
15 *
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.
19 */
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
59 /*
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.
63 */
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 */
73 };
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 */
80 /*
81 * Count is atomic so can incr (fork) or decr (exit) without a lock.
82 */
83 atomic_t count; /* count tasks using this cpuset */
85 /*
86 * We link our 'sibling' struct into our parents 'children'.
87 * Our children link their 'sibling' into our 'children'.
88 */
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 */
95 /*
96 * Copy of global cpuset_mems_generation as of the most
97 * recent time this cpuset changed its mems_allowed.
98 */
99 int mems_generation;
101 struct fmeter fmeter; /* memory_pressure filter */
102 };
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)
117 {
118 return test_bit(CS_CPU_EXCLUSIVE, &cs->flags);
119 }
121 static inline int is_mem_exclusive(const struct cpuset *cs)
122 {
123 return test_bit(CS_MEM_EXCLUSIVE, &cs->flags);
124 }
126 static inline int is_removed(const struct cpuset *cs)
127 {
128 return test_bit(CS_REMOVED, &cs->flags);
129 }
131 static inline int notify_on_release(const struct cpuset *cs)
132 {
133 return test_bit(CS_NOTIFY_ON_RELEASE, &cs->flags);
134 }
136 static inline int is_memory_migrate(const struct cpuset *cs)
137 {
138 return test_bit(CS_MEMORY_MIGRATE, &cs->flags);
139 }
141 static inline int is_spread_page(const struct cpuset *cs)
142 {
143 return test_bit(CS_SPREAD_PAGE, &cs->flags);
144 }
146 static inline int is_spread_slab(const struct cpuset *cs)
147 {
148 return test_bit(CS_SPREAD_SLAB, &cs->flags);
149 }
151 /*
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.
156 *
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.
160 *
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.
166 *
167 * Since cpuset_mems_generation is guarded by manage_mutex,
168 * there is no need to mark it atomic.
169 */
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),
179 };
181 static struct vfsmount *cpuset_mount;
182 static struct super_block *cpuset_sb;
184 /*
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.
189 *
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.
199 *
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().
204 *
205 * If a task is only holding callback_mutex, then it has read-only
206 * access to cpusets.
207 *
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.
210 *
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.
223 *
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.
227 *
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.
231 *
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.
239 *
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, pid == 1), 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.
246 *
247 * The above "Tale of Two Semaphores" would be complete, but for:
248 *
249 * The task_lock() exception
250 *
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.
260 *
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().
265 */
267 static DEFINE_MUTEX(manage_mutex);
268 static DEFINE_MUTEX(callback_mutex);
270 /*
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.
274 */
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,
282 };
284 static struct inode *cpuset_new_inode(mode_t mode)
285 {
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_blksize = PAGE_CACHE_SIZE;
293 inode->i_blocks = 0;
294 inode->i_atime = inode->i_mtime = inode->i_ctime = CURRENT_TIME;
295 inode->i_mapping->backing_dev_info = &cpuset_backing_dev_info;
296 }
297 return inode;
298 }
300 static void cpuset_diput(struct dentry *dentry, struct inode *inode)
301 {
302 /* is dentry a directory ? if so, kfree() associated cpuset */
303 if (S_ISDIR(inode->i_mode)) {
304 struct cpuset *cs = dentry->d_fsdata;
305 BUG_ON(!(is_removed(cs)));
306 kfree(cs);
307 }
308 iput(inode);
309 }
311 static struct dentry_operations cpuset_dops = {
312 .d_iput = cpuset_diput,
313 };
315 static struct dentry *cpuset_get_dentry(struct dentry *parent, const char *name)
316 {
317 struct dentry *d = lookup_one_len(name, parent, strlen(name));
318 if (!IS_ERR(d))
319 d->d_op = &cpuset_dops;
320 return d;
321 }
323 static void remove_dir(struct dentry *d)
324 {
325 struct dentry *parent = dget(d->d_parent);
327 d_delete(d);
328 simple_rmdir(parent->d_inode, d);
329 dput(parent);
330 }
332 /*
333 * NOTE : the dentry must have been dget()'ed
334 */
335 static void cpuset_d_remove_dir(struct dentry *dentry)
336 {
337 struct list_head *node;
339 spin_lock(&dcache_lock);
340 node = dentry->d_subdirs.next;
341 while (node != &dentry->d_subdirs) {
342 struct dentry *d = list_entry(node, struct dentry, d_u.d_child);
343 list_del_init(node);
344 if (d->d_inode) {
345 d = dget_locked(d);
346 spin_unlock(&dcache_lock);
347 d_delete(d);
348 simple_unlink(dentry->d_inode, d);
349 dput(d);
350 spin_lock(&dcache_lock);
351 }
352 node = dentry->d_subdirs.next;
353 }
354 list_del_init(&dentry->d_u.d_child);
355 spin_unlock(&dcache_lock);
356 remove_dir(dentry);
357 }
359 static struct super_operations cpuset_ops = {
360 .statfs = simple_statfs,
361 .drop_inode = generic_delete_inode,
362 };
364 static int cpuset_fill_super(struct super_block *sb, void *unused_data,
365 int unused_silent)
366 {
367 struct inode *inode;
368 struct dentry *root;
370 sb->s_blocksize = PAGE_CACHE_SIZE;
371 sb->s_blocksize_bits = PAGE_CACHE_SHIFT;
372 sb->s_magic = CPUSET_SUPER_MAGIC;
373 sb->s_op = &cpuset_ops;
374 cpuset_sb = sb;
376 inode = cpuset_new_inode(S_IFDIR | S_IRUGO | S_IXUGO | S_IWUSR);
377 if (inode) {
378 inode->i_op = &simple_dir_inode_operations;
379 inode->i_fop = &simple_dir_operations;
380 /* directories start off with i_nlink == 2 (for "." entry) */
381 inode->i_nlink++;
382 } else {
383 return -ENOMEM;
384 }
386 root = d_alloc_root(inode);
387 if (!root) {
388 iput(inode);
389 return -ENOMEM;
390 }
391 sb->s_root = root;
392 return 0;
393 }
395 static int cpuset_get_sb(struct file_system_type *fs_type,
396 int flags, const char *unused_dev_name,
397 void *data, struct vfsmount *mnt)
398 {
399 return get_sb_single(fs_type, flags, data, cpuset_fill_super, mnt);
400 }
402 static struct file_system_type cpuset_fs_type = {
403 .name = "cpuset",
404 .get_sb = cpuset_get_sb,
405 .kill_sb = kill_litter_super,
406 };
408 /* struct cftype:
409 *
410 * The files in the cpuset filesystem mostly have a very simple read/write
411 * handling, some common function will take care of it. Nevertheless some cases
412 * (read tasks) are special and therefore I define this structure for every
413 * kind of file.
414 *
415 *
416 * When reading/writing to a file:
417 * - the cpuset to use in file->f_dentry->d_parent->d_fsdata
418 * - the 'cftype' of the file is file->f_dentry->d_fsdata
419 */
421 struct cftype {
422 char *name;
423 int private;
424 int (*open) (struct inode *inode, struct file *file);
425 ssize_t (*read) (struct file *file, char __user *buf, size_t nbytes,
426 loff_t *ppos);
427 int (*write) (struct file *file, const char __user *buf, size_t nbytes,
428 loff_t *ppos);
429 int (*release) (struct inode *inode, struct file *file);
430 };
432 static inline struct cpuset *__d_cs(struct dentry *dentry)
433 {
434 return dentry->d_fsdata;
435 }
437 static inline struct cftype *__d_cft(struct dentry *dentry)
438 {
439 return dentry->d_fsdata;
440 }
442 /*
443 * Call with manage_mutex held. Writes path of cpuset into buf.
444 * Returns 0 on success, -errno on error.
445 */
447 static int cpuset_path(const struct cpuset *cs, char *buf, int buflen)
448 {
449 char *start;
451 start = buf + buflen;
453 *--start = '\0';
454 for (;;) {
455 int len = cs->dentry->d_name.len;
456 if ((start -= len) < buf)
457 return -ENAMETOOLONG;
458 memcpy(start, cs->dentry->d_name.name, len);
459 cs = cs->parent;
460 if (!cs)
461 break;
462 if (!cs->parent)
463 continue;
464 if (--start < buf)
465 return -ENAMETOOLONG;
466 *start = '/';
467 }
468 memmove(buf, start, buf + buflen - start);
469 return 0;
470 }
472 /*
473 * Notify userspace when a cpuset is released, by running
474 * /sbin/cpuset_release_agent with the name of the cpuset (path
475 * relative to the root of cpuset file system) as the argument.
476 *
477 * Most likely, this user command will try to rmdir this cpuset.
478 *
479 * This races with the possibility that some other task will be
480 * attached to this cpuset before it is removed, or that some other
481 * user task will 'mkdir' a child cpuset of this cpuset. That's ok.
482 * The presumed 'rmdir' will fail quietly if this cpuset is no longer
483 * unused, and this cpuset will be reprieved from its death sentence,
484 * to continue to serve a useful existence. Next time it's released,
485 * we will get notified again, if it still has 'notify_on_release' set.
486 *
487 * The final arg to call_usermodehelper() is 0, which means don't
488 * wait. The separate /sbin/cpuset_release_agent task is forked by
489 * call_usermodehelper(), then control in this thread returns here,
490 * without waiting for the release agent task. We don't bother to
491 * wait because the caller of this routine has no use for the exit
492 * status of the /sbin/cpuset_release_agent task, so no sense holding
493 * our caller up for that.
494 *
495 * When we had only one cpuset mutex, we had to call this
496 * without holding it, to avoid deadlock when call_usermodehelper()
497 * allocated memory. With two locks, we could now call this while
498 * holding manage_mutex, but we still don't, so as to minimize
499 * the time manage_mutex is held.
500 */
502 static void cpuset_release_agent(const char *pathbuf)
503 {
504 char *argv[3], *envp[3];
505 int i;
507 if (!pathbuf)
508 return;
510 i = 0;
511 argv[i++] = "/sbin/cpuset_release_agent";
512 argv[i++] = (char *)pathbuf;
513 argv[i] = NULL;
515 i = 0;
516 /* minimal command environment */
517 envp[i++] = "HOME=/";
518 envp[i++] = "PATH=/sbin:/bin:/usr/sbin:/usr/bin";
519 envp[i] = NULL;
521 call_usermodehelper(argv[0], argv, envp, 0);
522 kfree(pathbuf);
523 }
525 /*
526 * Either cs->count of using tasks transitioned to zero, or the
527 * cs->children list of child cpusets just became empty. If this
528 * cs is notify_on_release() and now both the user count is zero and
529 * the list of children is empty, prepare cpuset path in a kmalloc'd
530 * buffer, to be returned via ppathbuf, so that the caller can invoke
531 * cpuset_release_agent() with it later on, once manage_mutex is dropped.
532 * Call here with manage_mutex held.
533 *
534 * This check_for_release() routine is responsible for kmalloc'ing
535 * pathbuf. The above cpuset_release_agent() is responsible for
536 * kfree'ing pathbuf. The caller of these routines is responsible
537 * for providing a pathbuf pointer, initialized to NULL, then
538 * calling check_for_release() with manage_mutex held and the address
539 * of the pathbuf pointer, then dropping manage_mutex, then calling
540 * cpuset_release_agent() with pathbuf, as set by check_for_release().
541 */
543 static void check_for_release(struct cpuset *cs, char **ppathbuf)
544 {
545 if (notify_on_release(cs) && atomic_read(&cs->count) == 0 &&
546 list_empty(&cs->children)) {
547 char *buf;
549 buf = kmalloc(PAGE_SIZE, GFP_KERNEL);
550 if (!buf)
551 return;
552 if (cpuset_path(cs, buf, PAGE_SIZE) < 0)
553 kfree(buf);
554 else
555 *ppathbuf = buf;
556 }
557 }
559 /*
560 * Return in *pmask the portion of a cpusets's cpus_allowed that
561 * are online. If none are online, walk up the cpuset hierarchy
562 * until we find one that does have some online cpus. If we get
563 * all the way to the top and still haven't found any online cpus,
564 * return cpu_online_map. Or if passed a NULL cs from an exit'ing
565 * task, return cpu_online_map.
566 *
567 * One way or another, we guarantee to return some non-empty subset
568 * of cpu_online_map.
569 *
570 * Call with callback_mutex held.
571 */
573 static void guarantee_online_cpus(const struct cpuset *cs, cpumask_t *pmask)
574 {
575 while (cs && !cpus_intersects(cs->cpus_allowed, cpu_online_map))
576 cs = cs->parent;
577 if (cs)
578 cpus_and(*pmask, cs->cpus_allowed, cpu_online_map);
579 else
580 *pmask = cpu_online_map;
581 BUG_ON(!cpus_intersects(*pmask, cpu_online_map));
582 }
584 /*
585 * Return in *pmask the portion of a cpusets's mems_allowed that
586 * are online. If none are online, walk up the cpuset hierarchy
587 * until we find one that does have some online mems. If we get
588 * all the way to the top and still haven't found any online mems,
589 * return node_online_map.
590 *
591 * One way or another, we guarantee to return some non-empty subset
592 * of node_online_map.
593 *
594 * Call with callback_mutex held.
595 */
597 static void guarantee_online_mems(const struct cpuset *cs, nodemask_t *pmask)
598 {
599 while (cs && !nodes_intersects(cs->mems_allowed, node_online_map))
600 cs = cs->parent;
601 if (cs)
602 nodes_and(*pmask, cs->mems_allowed, node_online_map);
603 else
604 *pmask = node_online_map;
605 BUG_ON(!nodes_intersects(*pmask, node_online_map));
606 }
608 /**
609 * cpuset_update_task_memory_state - update task memory placement
610 *
611 * If the current tasks cpusets mems_allowed changed behind our
612 * backs, update current->mems_allowed, mems_generation and task NUMA
613 * mempolicy to the new value.
614 *
615 * Task mempolicy is updated by rebinding it relative to the
616 * current->cpuset if a task has its memory placement changed.
617 * Do not call this routine if in_interrupt().
618 *
619 * Call without callback_mutex or task_lock() held. May be
620 * called with or without manage_mutex held. Thanks in part to
621 * 'the_top_cpuset_hack', the tasks cpuset pointer will never
622 * be NULL. This routine also might acquire callback_mutex and
623 * current->mm->mmap_sem during call.
624 *
625 * Reading current->cpuset->mems_generation doesn't need task_lock
626 * to guard the current->cpuset derefence, because it is guarded
627 * from concurrent freeing of current->cpuset by attach_task(),
628 * using RCU.
629 *
630 * The rcu_dereference() is technically probably not needed,
631 * as I don't actually mind if I see a new cpuset pointer but
632 * an old value of mems_generation. However this really only
633 * matters on alpha systems using cpusets heavily. If I dropped
634 * that rcu_dereference(), it would save them a memory barrier.
635 * For all other arch's, rcu_dereference is a no-op anyway, and for
636 * alpha systems not using cpusets, another planned optimization,
637 * avoiding the rcu critical section for tasks in the root cpuset
638 * which is statically allocated, so can't vanish, will make this
639 * irrelevant. Better to use RCU as intended, than to engage in
640 * some cute trick to save a memory barrier that is impossible to
641 * test, for alpha systems using cpusets heavily, which might not
642 * even exist.
643 *
644 * This routine is needed to update the per-task mems_allowed data,
645 * within the tasks context, when it is trying to allocate memory
646 * (in various mm/mempolicy.c routines) and notices that some other
647 * task has been modifying its cpuset.
648 */
650 void cpuset_update_task_memory_state(void)
651 {
652 int my_cpusets_mem_gen;
653 struct task_struct *tsk = current;
654 struct cpuset *cs;
656 if (tsk->cpuset == &top_cpuset) {
657 /* Don't need rcu for top_cpuset. It's never freed. */
658 my_cpusets_mem_gen = top_cpuset.mems_generation;
659 } else {
660 rcu_read_lock();
661 cs = rcu_dereference(tsk->cpuset);
662 my_cpusets_mem_gen = cs->mems_generation;
663 rcu_read_unlock();
664 }
666 if (my_cpusets_mem_gen != tsk->cpuset_mems_generation) {
667 mutex_lock(&callback_mutex);
668 task_lock(tsk);
669 cs = tsk->cpuset; /* Maybe changed when task not locked */
670 guarantee_online_mems(cs, &tsk->mems_allowed);
671 tsk->cpuset_mems_generation = cs->mems_generation;
672 if (is_spread_page(cs))
673 tsk->flags |= PF_SPREAD_PAGE;
674 else
675 tsk->flags &= ~PF_SPREAD_PAGE;
676 if (is_spread_slab(cs))
677 tsk->flags |= PF_SPREAD_SLAB;
678 else
679 tsk->flags &= ~PF_SPREAD_SLAB;
680 task_unlock(tsk);
681 mutex_unlock(&callback_mutex);
682 mpol_rebind_task(tsk, &tsk->mems_allowed);
683 }
684 }
686 /*
687 * is_cpuset_subset(p, q) - Is cpuset p a subset of cpuset q?
688 *
689 * One cpuset is a subset of another if all its allowed CPUs and
690 * Memory Nodes are a subset of the other, and its exclusive flags
691 * are only set if the other's are set. Call holding manage_mutex.
692 */
694 static int is_cpuset_subset(const struct cpuset *p, const struct cpuset *q)
695 {
696 return cpus_subset(p->cpus_allowed, q->cpus_allowed) &&
697 nodes_subset(p->mems_allowed, q->mems_allowed) &&
698 is_cpu_exclusive(p) <= is_cpu_exclusive(q) &&
699 is_mem_exclusive(p) <= is_mem_exclusive(q);
700 }
702 /*
703 * validate_change() - Used to validate that any proposed cpuset change
704 * follows the structural rules for cpusets.
705 *
706 * If we replaced the flag and mask values of the current cpuset
707 * (cur) with those values in the trial cpuset (trial), would
708 * our various subset and exclusive rules still be valid? Presumes
709 * manage_mutex held.
710 *
711 * 'cur' is the address of an actual, in-use cpuset. Operations
712 * such as list traversal that depend on the actual address of the
713 * cpuset in the list must use cur below, not trial.
714 *
715 * 'trial' is the address of bulk structure copy of cur, with
716 * perhaps one or more of the fields cpus_allowed, mems_allowed,
717 * or flags changed to new, trial values.
718 *
719 * Return 0 if valid, -errno if not.
720 */
722 static int validate_change(const struct cpuset *cur, const struct cpuset *trial)
723 {
724 struct cpuset *c, *par;
726 /* Each of our child cpusets must be a subset of us */
727 list_for_each_entry(c, &cur->children, sibling) {
728 if (!is_cpuset_subset(c, trial))
729 return -EBUSY;
730 }
732 /* Remaining checks don't apply to root cpuset */
733 if ((par = cur->parent) == NULL)
734 return 0;
736 /* We must be a subset of our parent cpuset */
737 if (!is_cpuset_subset(trial, par))
738 return -EACCES;
740 /* If either I or some sibling (!= me) is exclusive, we can't overlap */
741 list_for_each_entry(c, &par->children, sibling) {
742 if ((is_cpu_exclusive(trial) || is_cpu_exclusive(c)) &&
743 c != cur &&
744 cpus_intersects(trial->cpus_allowed, c->cpus_allowed))
745 return -EINVAL;
746 if ((is_mem_exclusive(trial) || is_mem_exclusive(c)) &&
747 c != cur &&
748 nodes_intersects(trial->mems_allowed, c->mems_allowed))
749 return -EINVAL;
750 }
752 return 0;
753 }
755 /*
756 * For a given cpuset cur, partition the system as follows
757 * a. All cpus in the parent cpuset's cpus_allowed that are not part of any
758 * exclusive child cpusets
759 * b. All cpus in the current cpuset's cpus_allowed that are not part of any
760 * exclusive child cpusets
761 * Build these two partitions by calling partition_sched_domains
762 *
763 * Call with manage_mutex held. May nest a call to the
764 * lock_cpu_hotplug()/unlock_cpu_hotplug() pair.
765 * Must not be called holding callback_mutex, because we must
766 * not call lock_cpu_hotplug() while holding callback_mutex.
767 */
769 static void update_cpu_domains(struct cpuset *cur)
770 {
771 struct cpuset *c, *par = cur->parent;
772 cpumask_t pspan, cspan;
774 if (par == NULL || cpus_empty(cur->cpus_allowed))
775 return;
777 /*
778 * Get all cpus from parent's cpus_allowed not part of exclusive
779 * children
780 */
781 pspan = par->cpus_allowed;
782 list_for_each_entry(c, &par->children, sibling) {
783 if (is_cpu_exclusive(c))
784 cpus_andnot(pspan, pspan, c->cpus_allowed);
785 }
786 if (!is_cpu_exclusive(cur)) {
787 cpus_or(pspan, pspan, cur->cpus_allowed);
788 if (cpus_equal(pspan, cur->cpus_allowed))
789 return;
790 cspan = CPU_MASK_NONE;
791 } else {
792 if (cpus_empty(pspan))
793 return;
794 cspan = cur->cpus_allowed;
795 /*
796 * Get all cpus from current cpuset's cpus_allowed not part
797 * of exclusive children
798 */
799 list_for_each_entry(c, &cur->children, sibling) {
800 if (is_cpu_exclusive(c))
801 cpus_andnot(cspan, cspan, c->cpus_allowed);
802 }
803 }
805 lock_cpu_hotplug();
806 partition_sched_domains(&pspan, &cspan);
807 unlock_cpu_hotplug();
808 }
810 /*
811 * Call with manage_mutex held. May take callback_mutex during call.
812 */
814 static int update_cpumask(struct cpuset *cs, char *buf)
815 {
816 struct cpuset trialcs;
817 int retval, cpus_unchanged;
819 /* top_cpuset.cpus_allowed tracks cpu_online_map; it's read-only */
820 if (cs == &top_cpuset)
821 return -EACCES;
823 trialcs = *cs;
824 retval = cpulist_parse(buf, trialcs.cpus_allowed);
825 if (retval < 0)
826 return retval;
827 cpus_and(trialcs.cpus_allowed, trialcs.cpus_allowed, cpu_online_map);
828 if (cpus_empty(trialcs.cpus_allowed))
829 return -ENOSPC;
830 retval = validate_change(cs, &trialcs);
831 if (retval < 0)
832 return retval;
833 cpus_unchanged = cpus_equal(cs->cpus_allowed, trialcs.cpus_allowed);
834 mutex_lock(&callback_mutex);
835 cs->cpus_allowed = trialcs.cpus_allowed;
836 mutex_unlock(&callback_mutex);
837 if (is_cpu_exclusive(cs) && !cpus_unchanged)
838 update_cpu_domains(cs);
839 return 0;
840 }
842 /*
843 * cpuset_migrate_mm
844 *
845 * Migrate memory region from one set of nodes to another.
846 *
847 * Temporarilly set tasks mems_allowed to target nodes of migration,
848 * so that the migration code can allocate pages on these nodes.
849 *
850 * Call holding manage_mutex, so our current->cpuset won't change
851 * during this call, as manage_mutex holds off any attach_task()
852 * calls. Therefore we don't need to take task_lock around the
853 * call to guarantee_online_mems(), as we know no one is changing
854 * our tasks cpuset.
855 *
856 * Hold callback_mutex around the two modifications of our tasks
857 * mems_allowed to synchronize with cpuset_mems_allowed().
858 *
859 * While the mm_struct we are migrating is typically from some
860 * other task, the task_struct mems_allowed that we are hacking
861 * is for our current task, which must allocate new pages for that
862 * migrating memory region.
863 *
864 * We call cpuset_update_task_memory_state() before hacking
865 * our tasks mems_allowed, so that we are assured of being in
866 * sync with our tasks cpuset, and in particular, callbacks to
867 * cpuset_update_task_memory_state() from nested page allocations
868 * won't see any mismatch of our cpuset and task mems_generation
869 * values, so won't overwrite our hacked tasks mems_allowed
870 * nodemask.
871 */
873 static void cpuset_migrate_mm(struct mm_struct *mm, const nodemask_t *from,
874 const nodemask_t *to)
875 {
876 struct task_struct *tsk = current;
878 cpuset_update_task_memory_state();
880 mutex_lock(&callback_mutex);
881 tsk->mems_allowed = *to;
882 mutex_unlock(&callback_mutex);
884 do_migrate_pages(mm, from, to, MPOL_MF_MOVE_ALL);
886 mutex_lock(&callback_mutex);
887 guarantee_online_mems(tsk->cpuset, &tsk->mems_allowed);
888 mutex_unlock(&callback_mutex);
889 }
891 /*
892 * Handle user request to change the 'mems' memory placement
893 * of a cpuset. Needs to validate the request, update the
894 * cpusets mems_allowed and mems_generation, and for each
895 * task in the cpuset, rebind any vma mempolicies and if
896 * the cpuset is marked 'memory_migrate', migrate the tasks
897 * pages to the new memory.
898 *
899 * Call with manage_mutex held. May take callback_mutex during call.
900 * Will take tasklist_lock, scan tasklist for tasks in cpuset cs,
901 * lock each such tasks mm->mmap_sem, scan its vma's and rebind
902 * their mempolicies to the cpusets new mems_allowed.
903 */
905 static int update_nodemask(struct cpuset *cs, char *buf)
906 {
907 struct cpuset trialcs;
908 nodemask_t oldmem;
909 struct task_struct *g, *p;
910 struct mm_struct **mmarray;
911 int i, n, ntasks;
912 int migrate;
913 int fudge;
914 int retval;
916 trialcs = *cs;
917 retval = nodelist_parse(buf, trialcs.mems_allowed);
918 if (retval < 0)
919 goto done;
920 nodes_and(trialcs.mems_allowed, trialcs.mems_allowed, node_online_map);
921 oldmem = cs->mems_allowed;
922 if (nodes_equal(oldmem, trialcs.mems_allowed)) {
923 retval = 0; /* Too easy - nothing to do */
924 goto done;
925 }
926 if (nodes_empty(trialcs.mems_allowed)) {
927 retval = -ENOSPC;
928 goto done;
929 }
930 retval = validate_change(cs, &trialcs);
931 if (retval < 0)
932 goto done;
934 mutex_lock(&callback_mutex);
935 cs->mems_allowed = trialcs.mems_allowed;
936 cs->mems_generation = cpuset_mems_generation++;
937 mutex_unlock(&callback_mutex);
939 set_cpuset_being_rebound(cs); /* causes mpol_copy() rebind */
941 fudge = 10; /* spare mmarray[] slots */
942 fudge += cpus_weight(cs->cpus_allowed); /* imagine one fork-bomb/cpu */
943 retval = -ENOMEM;
945 /*
946 * Allocate mmarray[] to hold mm reference for each task
947 * in cpuset cs. Can't kmalloc GFP_KERNEL while holding
948 * tasklist_lock. We could use GFP_ATOMIC, but with a
949 * few more lines of code, we can retry until we get a big
950 * enough mmarray[] w/o using GFP_ATOMIC.
951 */
952 while (1) {
953 ntasks = atomic_read(&cs->count); /* guess */
954 ntasks += fudge;
955 mmarray = kmalloc(ntasks * sizeof(*mmarray), GFP_KERNEL);
956 if (!mmarray)
957 goto done;
958 write_lock_irq(&tasklist_lock); /* block fork */
959 if (atomic_read(&cs->count) <= ntasks)
960 break; /* got enough */
961 write_unlock_irq(&tasklist_lock); /* try again */
962 kfree(mmarray);
963 }
965 n = 0;
967 /* Load up mmarray[] with mm reference for each task in cpuset. */
968 do_each_thread(g, p) {
969 struct mm_struct *mm;
971 if (n >= ntasks) {
972 printk(KERN_WARNING
973 "Cpuset mempolicy rebind incomplete.\n");
974 continue;
975 }
976 if (p->cpuset != cs)
977 continue;
978 mm = get_task_mm(p);
979 if (!mm)
980 continue;
981 mmarray[n++] = mm;
982 } while_each_thread(g, p);
983 write_unlock_irq(&tasklist_lock);
985 /*
986 * Now that we've dropped the tasklist spinlock, we can
987 * rebind the vma mempolicies of each mm in mmarray[] to their
988 * new cpuset, and release that mm. The mpol_rebind_mm()
989 * call takes mmap_sem, which we couldn't take while holding
990 * tasklist_lock. Forks can happen again now - the mpol_copy()
991 * cpuset_being_rebound check will catch such forks, and rebind
992 * their vma mempolicies too. Because we still hold the global
993 * cpuset manage_mutex, we know that no other rebind effort will
994 * be contending for the global variable cpuset_being_rebound.
995 * It's ok if we rebind the same mm twice; mpol_rebind_mm()
996 * is idempotent. Also migrate pages in each mm to new nodes.
997 */
998 migrate = is_memory_migrate(cs);
999 for (i = 0; i < n; i++) {
1000 struct mm_struct *mm = mmarray[i];
1002 mpol_rebind_mm(mm, &cs->mems_allowed);
1003 if (migrate)
1004 cpuset_migrate_mm(mm, &oldmem, &cs->mems_allowed);
1005 mmput(mm);
1008 /* We're done rebinding vma's to this cpusets new mems_allowed. */
1009 kfree(mmarray);
1010 set_cpuset_being_rebound(NULL);
1011 retval = 0;
1012 done:
1013 return retval;
1016 /*
1017 * Call with manage_mutex held.
1018 */
1020 static int update_memory_pressure_enabled(struct cpuset *cs, char *buf)
1022 if (simple_strtoul(buf, NULL, 10) != 0)
1023 cpuset_memory_pressure_enabled = 1;
1024 else
1025 cpuset_memory_pressure_enabled = 0;
1026 return 0;
1029 /*
1030 * update_flag - read a 0 or a 1 in a file and update associated flag
1031 * bit: the bit to update (CS_CPU_EXCLUSIVE, CS_MEM_EXCLUSIVE,
1032 * CS_NOTIFY_ON_RELEASE, CS_MEMORY_MIGRATE,
1033 * CS_SPREAD_PAGE, CS_SPREAD_SLAB)
1034 * cs: the cpuset to update
1035 * buf: the buffer where we read the 0 or 1
1037 * Call with manage_mutex held.
1038 */
1040 static int update_flag(cpuset_flagbits_t bit, struct cpuset *cs, char *buf)
1042 int turning_on;
1043 struct cpuset trialcs;
1044 int err, cpu_exclusive_changed;
1046 turning_on = (simple_strtoul(buf, NULL, 10) != 0);
1048 trialcs = *cs;
1049 if (turning_on)
1050 set_bit(bit, &trialcs.flags);
1051 else
1052 clear_bit(bit, &trialcs.flags);
1054 err = validate_change(cs, &trialcs);
1055 if (err < 0)
1056 return err;
1057 cpu_exclusive_changed =
1058 (is_cpu_exclusive(cs) != is_cpu_exclusive(&trialcs));
1059 mutex_lock(&callback_mutex);
1060 if (turning_on)
1061 set_bit(bit, &cs->flags);
1062 else
1063 clear_bit(bit, &cs->flags);
1064 mutex_unlock(&callback_mutex);
1066 if (cpu_exclusive_changed)
1067 update_cpu_domains(cs);
1068 return 0;
1071 /*
1072 * Frequency meter - How fast is some event occurring?
1074 * These routines manage a digitally filtered, constant time based,
1075 * event frequency meter. There are four routines:
1076 * fmeter_init() - initialize a frequency meter.
1077 * fmeter_markevent() - called each time the event happens.
1078 * fmeter_getrate() - returns the recent rate of such events.
1079 * fmeter_update() - internal routine used to update fmeter.
1081 * A common data structure is passed to each of these routines,
1082 * which is used to keep track of the state required to manage the
1083 * frequency meter and its digital filter.
1085 * The filter works on the number of events marked per unit time.
1086 * The filter is single-pole low-pass recursive (IIR). The time unit
1087 * is 1 second. Arithmetic is done using 32-bit integers scaled to
1088 * simulate 3 decimal digits of precision (multiplied by 1000).
1090 * With an FM_COEF of 933, and a time base of 1 second, the filter
1091 * has a half-life of 10 seconds, meaning that if the events quit
1092 * happening, then the rate returned from the fmeter_getrate()
1093 * will be cut in half each 10 seconds, until it converges to zero.
1095 * It is not worth doing a real infinitely recursive filter. If more
1096 * than FM_MAXTICKS ticks have elapsed since the last filter event,
1097 * just compute FM_MAXTICKS ticks worth, by which point the level
1098 * will be stable.
1100 * Limit the count of unprocessed events to FM_MAXCNT, so as to avoid
1101 * arithmetic overflow in the fmeter_update() routine.
1103 * Given the simple 32 bit integer arithmetic used, this meter works
1104 * best for reporting rates between one per millisecond (msec) and
1105 * one per 32 (approx) seconds. At constant rates faster than one
1106 * per msec it maxes out at values just under 1,000,000. At constant
1107 * rates between one per msec, and one per second it will stabilize
1108 * to a value N*1000, where N is the rate of events per second.
1109 * At constant rates between one per second and one per 32 seconds,
1110 * it will be choppy, moving up on the seconds that have an event,
1111 * and then decaying until the next event. At rates slower than
1112 * about one in 32 seconds, it decays all the way back to zero between
1113 * each event.
1114 */
1116 #define FM_COEF 933 /* coefficient for half-life of 10 secs */
1117 #define FM_MAXTICKS ((time_t)99) /* useless computing more ticks than this */
1118 #define FM_MAXCNT 1000000 /* limit cnt to avoid overflow */
1119 #define FM_SCALE 1000 /* faux fixed point scale */
1121 /* Initialize a frequency meter */
1122 static void fmeter_init(struct fmeter *fmp)
1124 fmp->cnt = 0;
1125 fmp->val = 0;
1126 fmp->time = 0;
1127 spin_lock_init(&fmp->lock);
1130 /* Internal meter update - process cnt events and update value */
1131 static void fmeter_update(struct fmeter *fmp)
1133 time_t now = get_seconds();
1134 time_t ticks = now - fmp->time;
1136 if (ticks == 0)
1137 return;
1139 ticks = min(FM_MAXTICKS, ticks);
1140 while (ticks-- > 0)
1141 fmp->val = (FM_COEF * fmp->val) / FM_SCALE;
1142 fmp->time = now;
1144 fmp->val += ((FM_SCALE - FM_COEF) * fmp->cnt) / FM_SCALE;
1145 fmp->cnt = 0;
1148 /* Process any previous ticks, then bump cnt by one (times scale). */
1149 static void fmeter_markevent(struct fmeter *fmp)
1151 spin_lock(&fmp->lock);
1152 fmeter_update(fmp);
1153 fmp->cnt = min(FM_MAXCNT, fmp->cnt + FM_SCALE);
1154 spin_unlock(&fmp->lock);
1157 /* Process any previous ticks, then return current value. */
1158 static int fmeter_getrate(struct fmeter *fmp)
1160 int val;
1162 spin_lock(&fmp->lock);
1163 fmeter_update(fmp);
1164 val = fmp->val;
1165 spin_unlock(&fmp->lock);
1166 return val;
1169 /*
1170 * Attack task specified by pid in 'pidbuf' to cpuset 'cs', possibly
1171 * writing the path of the old cpuset in 'ppathbuf' if it needs to be
1172 * notified on release.
1174 * Call holding manage_mutex. May take callback_mutex and task_lock of
1175 * the task 'pid' during call.
1176 */
1178 static int attach_task(struct cpuset *cs, char *pidbuf, char **ppathbuf)
1180 pid_t pid;
1181 struct task_struct *tsk;
1182 struct cpuset *oldcs;
1183 cpumask_t cpus;
1184 nodemask_t from, to;
1185 struct mm_struct *mm;
1186 int retval;
1188 if (sscanf(pidbuf, "%d", &pid) != 1)
1189 return -EIO;
1190 if (cpus_empty(cs->cpus_allowed) || nodes_empty(cs->mems_allowed))
1191 return -ENOSPC;
1193 if (pid) {
1194 read_lock(&tasklist_lock);
1196 tsk = find_task_by_pid(pid);
1197 if (!tsk || tsk->flags & PF_EXITING) {
1198 read_unlock(&tasklist_lock);
1199 return -ESRCH;
1202 get_task_struct(tsk);
1203 read_unlock(&tasklist_lock);
1205 if ((current->euid) && (current->euid != tsk->uid)
1206 && (current->euid != tsk->suid)) {
1207 put_task_struct(tsk);
1208 return -EACCES;
1210 } else {
1211 tsk = current;
1212 get_task_struct(tsk);
1215 retval = security_task_setscheduler(tsk, 0, NULL);
1216 if (retval) {
1217 put_task_struct(tsk);
1218 return retval;
1221 mutex_lock(&callback_mutex);
1223 task_lock(tsk);
1224 oldcs = tsk->cpuset;
1225 if (!oldcs) {
1226 task_unlock(tsk);
1227 mutex_unlock(&callback_mutex);
1228 put_task_struct(tsk);
1229 return -ESRCH;
1231 atomic_inc(&cs->count);
1232 rcu_assign_pointer(tsk->cpuset, cs);
1233 task_unlock(tsk);
1235 guarantee_online_cpus(cs, &cpus);
1236 set_cpus_allowed(tsk, cpus);
1238 from = oldcs->mems_allowed;
1239 to = cs->mems_allowed;
1241 mutex_unlock(&callback_mutex);
1243 mm = get_task_mm(tsk);
1244 if (mm) {
1245 mpol_rebind_mm(mm, &to);
1246 if (is_memory_migrate(cs))
1247 cpuset_migrate_mm(mm, &from, &to);
1248 mmput(mm);
1251 put_task_struct(tsk);
1252 synchronize_rcu();
1253 if (atomic_dec_and_test(&oldcs->count))
1254 check_for_release(oldcs, ppathbuf);
1255 return 0;
1258 /* The various types of files and directories in a cpuset file system */
1260 typedef enum {
1261 FILE_ROOT,
1262 FILE_DIR,
1263 FILE_MEMORY_MIGRATE,
1264 FILE_CPULIST,
1265 FILE_MEMLIST,
1266 FILE_CPU_EXCLUSIVE,
1267 FILE_MEM_EXCLUSIVE,
1268 FILE_NOTIFY_ON_RELEASE,
1269 FILE_MEMORY_PRESSURE_ENABLED,
1270 FILE_MEMORY_PRESSURE,
1271 FILE_SPREAD_PAGE,
1272 FILE_SPREAD_SLAB,
1273 FILE_TASKLIST,
1274 } cpuset_filetype_t;
1276 static ssize_t cpuset_common_file_write(struct file *file, const char __user *userbuf,
1277 size_t nbytes, loff_t *unused_ppos)
1279 struct cpuset *cs = __d_cs(file->f_dentry->d_parent);
1280 struct cftype *cft = __d_cft(file->f_dentry);
1281 cpuset_filetype_t type = cft->private;
1282 char *buffer;
1283 char *pathbuf = NULL;
1284 int retval = 0;
1286 /* Crude upper limit on largest legitimate cpulist user might write. */
1287 if (nbytes > 100 + 6 * NR_CPUS)
1288 return -E2BIG;
1290 /* +1 for nul-terminator */
1291 if ((buffer = kmalloc(nbytes + 1, GFP_KERNEL)) == 0)
1292 return -ENOMEM;
1294 if (copy_from_user(buffer, userbuf, nbytes)) {
1295 retval = -EFAULT;
1296 goto out1;
1298 buffer[nbytes] = 0; /* nul-terminate */
1300 mutex_lock(&manage_mutex);
1302 if (is_removed(cs)) {
1303 retval = -ENODEV;
1304 goto out2;
1307 switch (type) {
1308 case FILE_CPULIST:
1309 retval = update_cpumask(cs, buffer);
1310 break;
1311 case FILE_MEMLIST:
1312 retval = update_nodemask(cs, buffer);
1313 break;
1314 case FILE_CPU_EXCLUSIVE:
1315 retval = update_flag(CS_CPU_EXCLUSIVE, cs, buffer);
1316 break;
1317 case FILE_MEM_EXCLUSIVE:
1318 retval = update_flag(CS_MEM_EXCLUSIVE, cs, buffer);
1319 break;
1320 case FILE_NOTIFY_ON_RELEASE:
1321 retval = update_flag(CS_NOTIFY_ON_RELEASE, cs, buffer);
1322 break;
1323 case FILE_MEMORY_MIGRATE:
1324 retval = update_flag(CS_MEMORY_MIGRATE, cs, buffer);
1325 break;
1326 case FILE_MEMORY_PRESSURE_ENABLED:
1327 retval = update_memory_pressure_enabled(cs, buffer);
1328 break;
1329 case FILE_MEMORY_PRESSURE:
1330 retval = -EACCES;
1331 break;
1332 case FILE_SPREAD_PAGE:
1333 retval = update_flag(CS_SPREAD_PAGE, cs, buffer);
1334 cs->mems_generation = cpuset_mems_generation++;
1335 break;
1336 case FILE_SPREAD_SLAB:
1337 retval = update_flag(CS_SPREAD_SLAB, cs, buffer);
1338 cs->mems_generation = cpuset_mems_generation++;
1339 break;
1340 case FILE_TASKLIST:
1341 retval = attach_task(cs, buffer, &pathbuf);
1342 break;
1343 default:
1344 retval = -EINVAL;
1345 goto out2;
1348 if (retval == 0)
1349 retval = nbytes;
1350 out2:
1351 mutex_unlock(&manage_mutex);
1352 cpuset_release_agent(pathbuf);
1353 out1:
1354 kfree(buffer);
1355 return retval;
1358 static ssize_t cpuset_file_write(struct file *file, const char __user *buf,
1359 size_t nbytes, loff_t *ppos)
1361 ssize_t retval = 0;
1362 struct cftype *cft = __d_cft(file->f_dentry);
1363 if (!cft)
1364 return -ENODEV;
1366 /* special function ? */
1367 if (cft->write)
1368 retval = cft->write(file, buf, nbytes, ppos);
1369 else
1370 retval = cpuset_common_file_write(file, buf, nbytes, ppos);
1372 return retval;
1375 /*
1376 * These ascii lists should be read in a single call, by using a user
1377 * buffer large enough to hold the entire map. If read in smaller
1378 * chunks, there is no guarantee of atomicity. Since the display format
1379 * used, list of ranges of sequential numbers, is variable length,
1380 * and since these maps can change value dynamically, one could read
1381 * gibberish by doing partial reads while a list was changing.
1382 * A single large read to a buffer that crosses a page boundary is
1383 * ok, because the result being copied to user land is not recomputed
1384 * across a page fault.
1385 */
1387 static int cpuset_sprintf_cpulist(char *page, struct cpuset *cs)
1389 cpumask_t mask;
1391 mutex_lock(&callback_mutex);
1392 mask = cs->cpus_allowed;
1393 mutex_unlock(&callback_mutex);
1395 return cpulist_scnprintf(page, PAGE_SIZE, mask);
1398 static int cpuset_sprintf_memlist(char *page, struct cpuset *cs)
1400 nodemask_t mask;
1402 mutex_lock(&callback_mutex);
1403 mask = cs->mems_allowed;
1404 mutex_unlock(&callback_mutex);
1406 return nodelist_scnprintf(page, PAGE_SIZE, mask);
1409 static ssize_t cpuset_common_file_read(struct file *file, char __user *buf,
1410 size_t nbytes, loff_t *ppos)
1412 struct cftype *cft = __d_cft(file->f_dentry);
1413 struct cpuset *cs = __d_cs(file->f_dentry->d_parent);
1414 cpuset_filetype_t type = cft->private;
1415 char *page;
1416 ssize_t retval = 0;
1417 char *s;
1419 if (!(page = (char *)__get_free_page(GFP_KERNEL)))
1420 return -ENOMEM;
1422 s = page;
1424 switch (type) {
1425 case FILE_CPULIST:
1426 s += cpuset_sprintf_cpulist(s, cs);
1427 break;
1428 case FILE_MEMLIST:
1429 s += cpuset_sprintf_memlist(s, cs);
1430 break;
1431 case FILE_CPU_EXCLUSIVE:
1432 *s++ = is_cpu_exclusive(cs) ? '1' : '0';
1433 break;
1434 case FILE_MEM_EXCLUSIVE:
1435 *s++ = is_mem_exclusive(cs) ? '1' : '0';
1436 break;
1437 case FILE_NOTIFY_ON_RELEASE:
1438 *s++ = notify_on_release(cs) ? '1' : '0';
1439 break;
1440 case FILE_MEMORY_MIGRATE:
1441 *s++ = is_memory_migrate(cs) ? '1' : '0';
1442 break;
1443 case FILE_MEMORY_PRESSURE_ENABLED:
1444 *s++ = cpuset_memory_pressure_enabled ? '1' : '0';
1445 break;
1446 case FILE_MEMORY_PRESSURE:
1447 s += sprintf(s, "%d", fmeter_getrate(&cs->fmeter));
1448 break;
1449 case FILE_SPREAD_PAGE:
1450 *s++ = is_spread_page(cs) ? '1' : '0';
1451 break;
1452 case FILE_SPREAD_SLAB:
1453 *s++ = is_spread_slab(cs) ? '1' : '0';
1454 break;
1455 default:
1456 retval = -EINVAL;
1457 goto out;
1459 *s++ = '\n';
1461 retval = simple_read_from_buffer(buf, nbytes, ppos, page, s - page);
1462 out:
1463 free_page((unsigned long)page);
1464 return retval;
1467 static ssize_t cpuset_file_read(struct file *file, char __user *buf, size_t nbytes,
1468 loff_t *ppos)
1470 ssize_t retval = 0;
1471 struct cftype *cft = __d_cft(file->f_dentry);
1472 if (!cft)
1473 return -ENODEV;
1475 /* special function ? */
1476 if (cft->read)
1477 retval = cft->read(file, buf, nbytes, ppos);
1478 else
1479 retval = cpuset_common_file_read(file, buf, nbytes, ppos);
1481 return retval;
1484 static int cpuset_file_open(struct inode *inode, struct file *file)
1486 int err;
1487 struct cftype *cft;
1489 err = generic_file_open(inode, file);
1490 if (err)
1491 return err;
1493 cft = __d_cft(file->f_dentry);
1494 if (!cft)
1495 return -ENODEV;
1496 if (cft->open)
1497 err = cft->open(inode, file);
1498 else
1499 err = 0;
1501 return err;
1504 static int cpuset_file_release(struct inode *inode, struct file *file)
1506 struct cftype *cft = __d_cft(file->f_dentry);
1507 if (cft->release)
1508 return cft->release(inode, file);
1509 return 0;
1512 /*
1513 * cpuset_rename - Only allow simple rename of directories in place.
1514 */
1515 static int cpuset_rename(struct inode *old_dir, struct dentry *old_dentry,
1516 struct inode *new_dir, struct dentry *new_dentry)
1518 if (!S_ISDIR(old_dentry->d_inode->i_mode))
1519 return -ENOTDIR;
1520 if (new_dentry->d_inode)
1521 return -EEXIST;
1522 if (old_dir != new_dir)
1523 return -EIO;
1524 return simple_rename(old_dir, old_dentry, new_dir, new_dentry);
1527 static struct file_operations cpuset_file_operations = {
1528 .read = cpuset_file_read,
1529 .write = cpuset_file_write,
1530 .llseek = generic_file_llseek,
1531 .open = cpuset_file_open,
1532 .release = cpuset_file_release,
1533 };
1535 static struct inode_operations cpuset_dir_inode_operations = {
1536 .lookup = simple_lookup,
1537 .mkdir = cpuset_mkdir,
1538 .rmdir = cpuset_rmdir,
1539 .rename = cpuset_rename,
1540 };
1542 static int cpuset_create_file(struct dentry *dentry, int mode)
1544 struct inode *inode;
1546 if (!dentry)
1547 return -ENOENT;
1548 if (dentry->d_inode)
1549 return -EEXIST;
1551 inode = cpuset_new_inode(mode);
1552 if (!inode)
1553 return -ENOMEM;
1555 if (S_ISDIR(mode)) {
1556 inode->i_op = &cpuset_dir_inode_operations;
1557 inode->i_fop = &simple_dir_operations;
1559 /* start off with i_nlink == 2 (for "." entry) */
1560 inode->i_nlink++;
1561 } else if (S_ISREG(mode)) {
1562 inode->i_size = 0;
1563 inode->i_fop = &cpuset_file_operations;
1566 d_instantiate(dentry, inode);
1567 dget(dentry); /* Extra count - pin the dentry in core */
1568 return 0;
1571 /*
1572 * cpuset_create_dir - create a directory for an object.
1573 * cs: the cpuset we create the directory for.
1574 * It must have a valid ->parent field
1575 * And we are going to fill its ->dentry field.
1576 * name: The name to give to the cpuset directory. Will be copied.
1577 * mode: mode to set on new directory.
1578 */
1580 static int cpuset_create_dir(struct cpuset *cs, const char *name, int mode)
1582 struct dentry *dentry = NULL;
1583 struct dentry *parent;
1584 int error = 0;
1586 parent = cs->parent->dentry;
1587 dentry = cpuset_get_dentry(parent, name);
1588 if (IS_ERR(dentry))
1589 return PTR_ERR(dentry);
1590 error = cpuset_create_file(dentry, S_IFDIR | mode);
1591 if (!error) {
1592 dentry->d_fsdata = cs;
1593 parent->d_inode->i_nlink++;
1594 cs->dentry = dentry;
1596 dput(dentry);
1598 return error;
1601 static int cpuset_add_file(struct dentry *dir, const struct cftype *cft)
1603 struct dentry *dentry;
1604 int error;
1606 mutex_lock(&dir->d_inode->i_mutex);
1607 dentry = cpuset_get_dentry(dir, cft->name);
1608 if (!IS_ERR(dentry)) {
1609 error = cpuset_create_file(dentry, 0644 | S_IFREG);
1610 if (!error)
1611 dentry->d_fsdata = (void *)cft;
1612 dput(dentry);
1613 } else
1614 error = PTR_ERR(dentry);
1615 mutex_unlock(&dir->d_inode->i_mutex);
1616 return error;
1619 /*
1620 * Stuff for reading the 'tasks' file.
1622 * Reading this file can return large amounts of data if a cpuset has
1623 * *lots* of attached tasks. So it may need several calls to read(),
1624 * but we cannot guarantee that the information we produce is correct
1625 * unless we produce it entirely atomically.
1627 * Upon tasks file open(), a struct ctr_struct is allocated, that
1628 * will have a pointer to an array (also allocated here). The struct
1629 * ctr_struct * is stored in file->private_data. Its resources will
1630 * be freed by release() when the file is closed. The array is used
1631 * to sprintf the PIDs and then used by read().
1632 */
1634 /* cpusets_tasks_read array */
1636 struct ctr_struct {
1637 char *buf;
1638 int bufsz;
1639 };
1641 /*
1642 * Load into 'pidarray' up to 'npids' of the tasks using cpuset 'cs'.
1643 * Return actual number of pids loaded. No need to task_lock(p)
1644 * when reading out p->cpuset, as we don't really care if it changes
1645 * on the next cycle, and we are not going to try to dereference it.
1646 */
1647 static int pid_array_load(pid_t *pidarray, int npids, struct cpuset *cs)
1649 int n = 0;
1650 struct task_struct *g, *p;
1652 read_lock(&tasklist_lock);
1654 do_each_thread(g, p) {
1655 if (p->cpuset == cs) {
1656 pidarray[n++] = p->pid;
1657 if (unlikely(n == npids))
1658 goto array_full;
1660 } while_each_thread(g, p);
1662 array_full:
1663 read_unlock(&tasklist_lock);
1664 return n;
1667 static int cmppid(const void *a, const void *b)
1669 return *(pid_t *)a - *(pid_t *)b;
1672 /*
1673 * Convert array 'a' of 'npids' pid_t's to a string of newline separated
1674 * decimal pids in 'buf'. Don't write more than 'sz' chars, but return
1675 * count 'cnt' of how many chars would be written if buf were large enough.
1676 */
1677 static int pid_array_to_buf(char *buf, int sz, pid_t *a, int npids)
1679 int cnt = 0;
1680 int i;
1682 for (i = 0; i < npids; i++)
1683 cnt += snprintf(buf + cnt, max(sz - cnt, 0), "%d\n", a[i]);
1684 return cnt;
1687 /*
1688 * Handle an open on 'tasks' file. Prepare a buffer listing the
1689 * process id's of tasks currently attached to the cpuset being opened.
1691 * Does not require any specific cpuset mutexes, and does not take any.
1692 */
1693 static int cpuset_tasks_open(struct inode *unused, struct file *file)
1695 struct cpuset *cs = __d_cs(file->f_dentry->d_parent);
1696 struct ctr_struct *ctr;
1697 pid_t *pidarray;
1698 int npids;
1699 char c;
1701 if (!(file->f_mode & FMODE_READ))
1702 return 0;
1704 ctr = kmalloc(sizeof(*ctr), GFP_KERNEL);
1705 if (!ctr)
1706 goto err0;
1708 /*
1709 * If cpuset gets more users after we read count, we won't have
1710 * enough space - tough. This race is indistinguishable to the
1711 * caller from the case that the additional cpuset users didn't
1712 * show up until sometime later on.
1713 */
1714 npids = atomic_read(&cs->count);
1715 pidarray = kmalloc(npids * sizeof(pid_t), GFP_KERNEL);
1716 if (!pidarray)
1717 goto err1;
1719 npids = pid_array_load(pidarray, npids, cs);
1720 sort(pidarray, npids, sizeof(pid_t), cmppid, NULL);
1722 /* Call pid_array_to_buf() twice, first just to get bufsz */
1723 ctr->bufsz = pid_array_to_buf(&c, sizeof(c), pidarray, npids) + 1;
1724 ctr->buf = kmalloc(ctr->bufsz, GFP_KERNEL);
1725 if (!ctr->buf)
1726 goto err2;
1727 ctr->bufsz = pid_array_to_buf(ctr->buf, ctr->bufsz, pidarray, npids);
1729 kfree(pidarray);
1730 file->private_data = ctr;
1731 return 0;
1733 err2:
1734 kfree(pidarray);
1735 err1:
1736 kfree(ctr);
1737 err0:
1738 return -ENOMEM;
1741 static ssize_t cpuset_tasks_read(struct file *file, char __user *buf,
1742 size_t nbytes, loff_t *ppos)
1744 struct ctr_struct *ctr = file->private_data;
1746 if (*ppos + nbytes > ctr->bufsz)
1747 nbytes = ctr->bufsz - *ppos;
1748 if (copy_to_user(buf, ctr->buf + *ppos, nbytes))
1749 return -EFAULT;
1750 *ppos += nbytes;
1751 return nbytes;
1754 static int cpuset_tasks_release(struct inode *unused_inode, struct file *file)
1756 struct ctr_struct *ctr;
1758 if (file->f_mode & FMODE_READ) {
1759 ctr = file->private_data;
1760 kfree(ctr->buf);
1761 kfree(ctr);
1763 return 0;
1766 /*
1767 * for the common functions, 'private' gives the type of file
1768 */
1770 static struct cftype cft_tasks = {
1771 .name = "tasks",
1772 .open = cpuset_tasks_open,
1773 .read = cpuset_tasks_read,
1774 .release = cpuset_tasks_release,
1775 .private = FILE_TASKLIST,
1776 };
1778 static struct cftype cft_cpus = {
1779 .name = "cpus",
1780 .private = FILE_CPULIST,
1781 };
1783 static struct cftype cft_mems = {
1784 .name = "mems",
1785 .private = FILE_MEMLIST,
1786 };
1788 static struct cftype cft_cpu_exclusive = {
1789 .name = "cpu_exclusive",
1790 .private = FILE_CPU_EXCLUSIVE,
1791 };
1793 static struct cftype cft_mem_exclusive = {
1794 .name = "mem_exclusive",
1795 .private = FILE_MEM_EXCLUSIVE,
1796 };
1798 static struct cftype cft_notify_on_release = {
1799 .name = "notify_on_release",
1800 .private = FILE_NOTIFY_ON_RELEASE,
1801 };
1803 static struct cftype cft_memory_migrate = {
1804 .name = "memory_migrate",
1805 .private = FILE_MEMORY_MIGRATE,
1806 };
1808 static struct cftype cft_memory_pressure_enabled = {
1809 .name = "memory_pressure_enabled",
1810 .private = FILE_MEMORY_PRESSURE_ENABLED,
1811 };
1813 static struct cftype cft_memory_pressure = {
1814 .name = "memory_pressure",
1815 .private = FILE_MEMORY_PRESSURE,
1816 };
1818 static struct cftype cft_spread_page = {
1819 .name = "memory_spread_page",
1820 .private = FILE_SPREAD_PAGE,
1821 };
1823 static struct cftype cft_spread_slab = {
1824 .name = "memory_spread_slab",
1825 .private = FILE_SPREAD_SLAB,
1826 };
1828 static int cpuset_populate_dir(struct dentry *cs_dentry)
1830 int err;
1832 if ((err = cpuset_add_file(cs_dentry, &cft_cpus)) < 0)
1833 return err;
1834 if ((err = cpuset_add_file(cs_dentry, &cft_mems)) < 0)
1835 return err;
1836 if ((err = cpuset_add_file(cs_dentry, &cft_cpu_exclusive)) < 0)
1837 return err;
1838 if ((err = cpuset_add_file(cs_dentry, &cft_mem_exclusive)) < 0)
1839 return err;
1840 if ((err = cpuset_add_file(cs_dentry, &cft_notify_on_release)) < 0)
1841 return err;
1842 if ((err = cpuset_add_file(cs_dentry, &cft_memory_migrate)) < 0)
1843 return err;
1844 if ((err = cpuset_add_file(cs_dentry, &cft_memory_pressure)) < 0)
1845 return err;
1846 if ((err = cpuset_add_file(cs_dentry, &cft_spread_page)) < 0)
1847 return err;
1848 if ((err = cpuset_add_file(cs_dentry, &cft_spread_slab)) < 0)
1849 return err;
1850 if ((err = cpuset_add_file(cs_dentry, &cft_tasks)) < 0)
1851 return err;
1852 return 0;
1855 /*
1856 * cpuset_create - create a cpuset
1857 * parent: cpuset that will be parent of the new cpuset.
1858 * name: name of the new cpuset. Will be strcpy'ed.
1859 * mode: mode to set on new inode
1861 * Must be called with the mutex on the parent inode held
1862 */
1864 static long cpuset_create(struct cpuset *parent, const char *name, int mode)
1866 struct cpuset *cs;
1867 int err;
1869 cs = kmalloc(sizeof(*cs), GFP_KERNEL);
1870 if (!cs)
1871 return -ENOMEM;
1873 mutex_lock(&manage_mutex);
1874 cpuset_update_task_memory_state();
1875 cs->flags = 0;
1876 if (notify_on_release(parent))
1877 set_bit(CS_NOTIFY_ON_RELEASE, &cs->flags);
1878 if (is_spread_page(parent))
1879 set_bit(CS_SPREAD_PAGE, &cs->flags);
1880 if (is_spread_slab(parent))
1881 set_bit(CS_SPREAD_SLAB, &cs->flags);
1882 cs->cpus_allowed = CPU_MASK_NONE;
1883 cs->mems_allowed = NODE_MASK_NONE;
1884 atomic_set(&cs->count, 0);
1885 INIT_LIST_HEAD(&cs->sibling);
1886 INIT_LIST_HEAD(&cs->children);
1887 cs->mems_generation = cpuset_mems_generation++;
1888 fmeter_init(&cs->fmeter);
1890 cs->parent = parent;
1892 mutex_lock(&callback_mutex);
1893 list_add(&cs->sibling, &cs->parent->children);
1894 number_of_cpusets++;
1895 mutex_unlock(&callback_mutex);
1897 err = cpuset_create_dir(cs, name, mode);
1898 if (err < 0)
1899 goto err;
1901 /*
1902 * Release manage_mutex before cpuset_populate_dir() because it
1903 * will down() this new directory's i_mutex and if we race with
1904 * another mkdir, we might deadlock.
1905 */
1906 mutex_unlock(&manage_mutex);
1908 err = cpuset_populate_dir(cs->dentry);
1909 /* If err < 0, we have a half-filled directory - oh well ;) */
1910 return 0;
1911 err:
1912 list_del(&cs->sibling);
1913 mutex_unlock(&manage_mutex);
1914 kfree(cs);
1915 return err;
1918 static int cpuset_mkdir(struct inode *dir, struct dentry *dentry, int mode)
1920 struct cpuset *c_parent = dentry->d_parent->d_fsdata;
1922 /* the vfs holds inode->i_mutex already */
1923 return cpuset_create(c_parent, dentry->d_name.name, mode | S_IFDIR);
1926 /*
1927 * Locking note on the strange update_flag() call below:
1929 * If the cpuset being removed is marked cpu_exclusive, then simulate
1930 * turning cpu_exclusive off, which will call update_cpu_domains().
1931 * The lock_cpu_hotplug() call in update_cpu_domains() must not be
1932 * made while holding callback_mutex. Elsewhere the kernel nests
1933 * callback_mutex inside lock_cpu_hotplug() calls. So the reverse
1934 * nesting would risk an ABBA deadlock.
1935 */
1937 static int cpuset_rmdir(struct inode *unused_dir, struct dentry *dentry)
1939 struct cpuset *cs = dentry->d_fsdata;
1940 struct dentry *d;
1941 struct cpuset *parent;
1942 char *pathbuf = NULL;
1944 /* the vfs holds both inode->i_mutex already */
1946 mutex_lock(&manage_mutex);
1947 cpuset_update_task_memory_state();
1948 if (atomic_read(&cs->count) > 0) {
1949 mutex_unlock(&manage_mutex);
1950 return -EBUSY;
1952 if (!list_empty(&cs->children)) {
1953 mutex_unlock(&manage_mutex);
1954 return -EBUSY;
1956 if (is_cpu_exclusive(cs)) {
1957 int retval = update_flag(CS_CPU_EXCLUSIVE, cs, "0");
1958 if (retval < 0) {
1959 mutex_unlock(&manage_mutex);
1960 return retval;
1963 parent = cs->parent;
1964 mutex_lock(&callback_mutex);
1965 set_bit(CS_REMOVED, &cs->flags);
1966 list_del(&cs->sibling); /* delete my sibling from parent->children */
1967 spin_lock(&cs->dentry->d_lock);
1968 d = dget(cs->dentry);
1969 cs->dentry = NULL;
1970 spin_unlock(&d->d_lock);
1971 cpuset_d_remove_dir(d);
1972 dput(d);
1973 number_of_cpusets--;
1974 mutex_unlock(&callback_mutex);
1975 if (list_empty(&parent->children))
1976 check_for_release(parent, &pathbuf);
1977 mutex_unlock(&manage_mutex);
1978 cpuset_release_agent(pathbuf);
1979 return 0;
1982 /*
1983 * cpuset_init_early - just enough so that the calls to
1984 * cpuset_update_task_memory_state() in early init code
1985 * are harmless.
1986 */
1988 int __init cpuset_init_early(void)
1990 struct task_struct *tsk = current;
1992 tsk->cpuset = &top_cpuset;
1993 tsk->cpuset->mems_generation = cpuset_mems_generation++;
1994 return 0;
1997 /**
1998 * cpuset_init - initialize cpusets at system boot
2000 * Description: Initialize top_cpuset and the cpuset internal file system,
2001 **/
2003 int __init cpuset_init(void)
2005 struct dentry *root;
2006 int err;
2008 top_cpuset.cpus_allowed = CPU_MASK_ALL;
2009 top_cpuset.mems_allowed = NODE_MASK_ALL;
2011 fmeter_init(&top_cpuset.fmeter);
2012 top_cpuset.mems_generation = cpuset_mems_generation++;
2014 init_task.cpuset = &top_cpuset;
2016 err = register_filesystem(&cpuset_fs_type);
2017 if (err < 0)
2018 goto out;
2019 cpuset_mount = kern_mount(&cpuset_fs_type);
2020 if (IS_ERR(cpuset_mount)) {
2021 printk(KERN_ERR "cpuset: could not mount!\n");
2022 err = PTR_ERR(cpuset_mount);
2023 cpuset_mount = NULL;
2024 goto out;
2026 root = cpuset_mount->mnt_sb->s_root;
2027 root->d_fsdata = &top_cpuset;
2028 root->d_inode->i_nlink++;
2029 top_cpuset.dentry = root;
2030 root->d_inode->i_op = &cpuset_dir_inode_operations;
2031 number_of_cpusets = 1;
2032 err = cpuset_populate_dir(root);
2033 /* memory_pressure_enabled is in root cpuset only */
2034 if (err == 0)
2035 err = cpuset_add_file(root, &cft_memory_pressure_enabled);
2036 out:
2037 return err;
2040 /*
2041 * The top_cpuset tracks what CPUs and Memory Nodes are online,
2042 * period. This is necessary in order to make cpusets transparent
2043 * (of no affect) on systems that are actively using CPU hotplug
2044 * but making no active use of cpusets.
2046 * This handles CPU hotplug (cpuhp) events. If someday Memory
2047 * Nodes can be hotplugged (dynamically changing node_online_map)
2048 * then we should handle that too, perhaps in a similar way.
2049 */
2051 #ifdef CONFIG_HOTPLUG_CPU
2052 static int cpuset_handle_cpuhp(struct notifier_block *nb,
2053 unsigned long phase, void *cpu)
2055 mutex_lock(&manage_mutex);
2056 mutex_lock(&callback_mutex);
2058 top_cpuset.cpus_allowed = cpu_online_map;
2060 mutex_unlock(&callback_mutex);
2061 mutex_unlock(&manage_mutex);
2063 return 0;
2065 #endif
2067 /**
2068 * cpuset_init_smp - initialize cpus_allowed
2070 * Description: Finish top cpuset after cpu, node maps are initialized
2071 **/
2073 void __init cpuset_init_smp(void)
2075 top_cpuset.cpus_allowed = cpu_online_map;
2076 top_cpuset.mems_allowed = node_online_map;
2078 hotcpu_notifier(cpuset_handle_cpuhp, 0);
2081 /**
2082 * cpuset_fork - attach newly forked task to its parents cpuset.
2083 * @tsk: pointer to task_struct of forking parent process.
2085 * Description: A task inherits its parent's cpuset at fork().
2087 * A pointer to the shared cpuset was automatically copied in fork.c
2088 * by dup_task_struct(). However, we ignore that copy, since it was
2089 * not made under the protection of task_lock(), so might no longer be
2090 * a valid cpuset pointer. attach_task() might have already changed
2091 * current->cpuset, allowing the previously referenced cpuset to
2092 * be removed and freed. Instead, we task_lock(current) and copy
2093 * its present value of current->cpuset for our freshly forked child.
2095 * At the point that cpuset_fork() is called, 'current' is the parent
2096 * task, and the passed argument 'child' points to the child task.
2097 **/
2099 void cpuset_fork(struct task_struct *child)
2101 task_lock(current);
2102 child->cpuset = current->cpuset;
2103 atomic_inc(&child->cpuset->count);
2104 task_unlock(current);
2107 /**
2108 * cpuset_exit - detach cpuset from exiting task
2109 * @tsk: pointer to task_struct of exiting process
2111 * Description: Detach cpuset from @tsk and release it.
2113 * Note that cpusets marked notify_on_release force every task in
2114 * them to take the global manage_mutex mutex when exiting.
2115 * This could impact scaling on very large systems. Be reluctant to
2116 * use notify_on_release cpusets where very high task exit scaling
2117 * is required on large systems.
2119 * Don't even think about derefencing 'cs' after the cpuset use count
2120 * goes to zero, except inside a critical section guarded by manage_mutex
2121 * or callback_mutex. Otherwise a zero cpuset use count is a license to
2122 * any other task to nuke the cpuset immediately, via cpuset_rmdir().
2124 * This routine has to take manage_mutex, not callback_mutex, because
2125 * it is holding that mutex while calling check_for_release(),
2126 * which calls kmalloc(), so can't be called holding callback_mutex().
2128 * We don't need to task_lock() this reference to tsk->cpuset,
2129 * because tsk is already marked PF_EXITING, so attach_task() won't
2130 * mess with it, or task is a failed fork, never visible to attach_task.
2132 * the_top_cpuset_hack:
2134 * Set the exiting tasks cpuset to the root cpuset (top_cpuset).
2136 * Don't leave a task unable to allocate memory, as that is an
2137 * accident waiting to happen should someone add a callout in
2138 * do_exit() after the cpuset_exit() call that might allocate.
2139 * If a task tries to allocate memory with an invalid cpuset,
2140 * it will oops in cpuset_update_task_memory_state().
2142 * We call cpuset_exit() while the task is still competent to
2143 * handle notify_on_release(), then leave the task attached to
2144 * the root cpuset (top_cpuset) for the remainder of its exit.
2146 * To do this properly, we would increment the reference count on
2147 * top_cpuset, and near the very end of the kernel/exit.c do_exit()
2148 * code we would add a second cpuset function call, to drop that
2149 * reference. This would just create an unnecessary hot spot on
2150 * the top_cpuset reference count, to no avail.
2152 * Normally, holding a reference to a cpuset without bumping its
2153 * count is unsafe. The cpuset could go away, or someone could
2154 * attach us to a different cpuset, decrementing the count on
2155 * the first cpuset that we never incremented. But in this case,
2156 * top_cpuset isn't going away, and either task has PF_EXITING set,
2157 * which wards off any attach_task() attempts, or task is a failed
2158 * fork, never visible to attach_task.
2160 * Another way to do this would be to set the cpuset pointer
2161 * to NULL here, and check in cpuset_update_task_memory_state()
2162 * for a NULL pointer. This hack avoids that NULL check, for no
2163 * cost (other than this way too long comment ;).
2164 **/
2166 void cpuset_exit(struct task_struct *tsk)
2168 struct cpuset *cs;
2170 cs = tsk->cpuset;
2171 tsk->cpuset = &top_cpuset; /* the_top_cpuset_hack - see above */
2173 if (notify_on_release(cs)) {
2174 char *pathbuf = NULL;
2176 mutex_lock(&manage_mutex);
2177 if (atomic_dec_and_test(&cs->count))
2178 check_for_release(cs, &pathbuf);
2179 mutex_unlock(&manage_mutex);
2180 cpuset_release_agent(pathbuf);
2181 } else {
2182 atomic_dec(&cs->count);
2186 /**
2187 * cpuset_cpus_allowed - return cpus_allowed mask from a tasks cpuset.
2188 * @tsk: pointer to task_struct from which to obtain cpuset->cpus_allowed.
2190 * Description: Returns the cpumask_t cpus_allowed of the cpuset
2191 * attached to the specified @tsk. Guaranteed to return some non-empty
2192 * subset of cpu_online_map, even if this means going outside the
2193 * tasks cpuset.
2194 **/
2196 cpumask_t cpuset_cpus_allowed(struct task_struct *tsk)
2198 cpumask_t mask;
2200 mutex_lock(&callback_mutex);
2201 task_lock(tsk);
2202 guarantee_online_cpus(tsk->cpuset, &mask);
2203 task_unlock(tsk);
2204 mutex_unlock(&callback_mutex);
2206 return mask;
2209 void cpuset_init_current_mems_allowed(void)
2211 current->mems_allowed = NODE_MASK_ALL;
2214 /**
2215 * cpuset_mems_allowed - return mems_allowed mask from a tasks cpuset.
2216 * @tsk: pointer to task_struct from which to obtain cpuset->mems_allowed.
2218 * Description: Returns the nodemask_t mems_allowed of the cpuset
2219 * attached to the specified @tsk. Guaranteed to return some non-empty
2220 * subset of node_online_map, even if this means going outside the
2221 * tasks cpuset.
2222 **/
2224 nodemask_t cpuset_mems_allowed(struct task_struct *tsk)
2226 nodemask_t mask;
2228 mutex_lock(&callback_mutex);
2229 task_lock(tsk);
2230 guarantee_online_mems(tsk->cpuset, &mask);
2231 task_unlock(tsk);
2232 mutex_unlock(&callback_mutex);
2234 return mask;
2237 /**
2238 * cpuset_zonelist_valid_mems_allowed - check zonelist vs. curremt mems_allowed
2239 * @zl: the zonelist to be checked
2241 * Are any of the nodes on zonelist zl allowed in current->mems_allowed?
2242 */
2243 int cpuset_zonelist_valid_mems_allowed(struct zonelist *zl)
2245 int i;
2247 for (i = 0; zl->zones[i]; i++) {
2248 int nid = zl->zones[i]->zone_pgdat->node_id;
2250 if (node_isset(nid, current->mems_allowed))
2251 return 1;
2253 return 0;
2256 /*
2257 * nearest_exclusive_ancestor() - Returns the nearest mem_exclusive
2258 * ancestor to the specified cpuset. Call holding callback_mutex.
2259 * If no ancestor is mem_exclusive (an unusual configuration), then
2260 * returns the root cpuset.
2261 */
2262 static const struct cpuset *nearest_exclusive_ancestor(const struct cpuset *cs)
2264 while (!is_mem_exclusive(cs) && cs->parent)
2265 cs = cs->parent;
2266 return cs;
2269 /**
2270 * cpuset_zone_allowed - Can we allocate memory on zone z's memory node?
2271 * @z: is this zone on an allowed node?
2272 * @gfp_mask: memory allocation flags (we use __GFP_HARDWALL)
2274 * If we're in interrupt, yes, we can always allocate. If zone
2275 * z's node is in our tasks mems_allowed, yes. If it's not a
2276 * __GFP_HARDWALL request and this zone's nodes is in the nearest
2277 * mem_exclusive cpuset ancestor to this tasks cpuset, yes.
2278 * Otherwise, no.
2280 * GFP_USER allocations are marked with the __GFP_HARDWALL bit,
2281 * and do not allow allocations outside the current tasks cpuset.
2282 * GFP_KERNEL allocations are not so marked, so can escape to the
2283 * nearest mem_exclusive ancestor cpuset.
2285 * Scanning up parent cpusets requires callback_mutex. The __alloc_pages()
2286 * routine only calls here with __GFP_HARDWALL bit _not_ set if
2287 * it's a GFP_KERNEL allocation, and all nodes in the current tasks
2288 * mems_allowed came up empty on the first pass over the zonelist.
2289 * So only GFP_KERNEL allocations, if all nodes in the cpuset are
2290 * short of memory, might require taking the callback_mutex mutex.
2292 * The first call here from mm/page_alloc:get_page_from_freelist()
2293 * has __GFP_HARDWALL set in gfp_mask, enforcing hardwall cpusets, so
2294 * no allocation on a node outside the cpuset is allowed (unless in
2295 * interrupt, of course).
2297 * The second pass through get_page_from_freelist() doesn't even call
2298 * here for GFP_ATOMIC calls. For those calls, the __alloc_pages()
2299 * variable 'wait' is not set, and the bit ALLOC_CPUSET is not set
2300 * in alloc_flags. That logic and the checks below have the combined
2301 * affect that:
2302 * in_interrupt - any node ok (current task context irrelevant)
2303 * GFP_ATOMIC - any node ok
2304 * GFP_KERNEL - any node in enclosing mem_exclusive cpuset ok
2305 * GFP_USER - only nodes in current tasks mems allowed ok.
2307 * Rule:
2308 * Don't call cpuset_zone_allowed() if you can't sleep, unless you
2309 * pass in the __GFP_HARDWALL flag set in gfp_flag, which disables
2310 * the code that might scan up ancestor cpusets and sleep.
2311 **/
2313 int __cpuset_zone_allowed(struct zone *z, gfp_t gfp_mask)
2315 int node; /* node that zone z is on */
2316 const struct cpuset *cs; /* current cpuset ancestors */
2317 int allowed; /* is allocation in zone z allowed? */
2319 if (in_interrupt())
2320 return 1;
2321 node = z->zone_pgdat->node_id;
2322 might_sleep_if(!(gfp_mask & __GFP_HARDWALL));
2323 if (node_isset(node, current->mems_allowed))
2324 return 1;
2325 if (gfp_mask & __GFP_HARDWALL) /* If hardwall request, stop here */
2326 return 0;
2328 if (current->flags & PF_EXITING) /* Let dying task have memory */
2329 return 1;
2331 /* Not hardwall and node outside mems_allowed: scan up cpusets */
2332 mutex_lock(&callback_mutex);
2334 task_lock(current);
2335 cs = nearest_exclusive_ancestor(current->cpuset);
2336 task_unlock(current);
2338 allowed = node_isset(node, cs->mems_allowed);
2339 mutex_unlock(&callback_mutex);
2340 return allowed;
2343 /**
2344 * cpuset_lock - lock out any changes to cpuset structures
2346 * The out of memory (oom) code needs to mutex_lock cpusets
2347 * from being changed while it scans the tasklist looking for a
2348 * task in an overlapping cpuset. Expose callback_mutex via this
2349 * cpuset_lock() routine, so the oom code can lock it, before
2350 * locking the task list. The tasklist_lock is a spinlock, so
2351 * must be taken inside callback_mutex.
2352 */
2354 void cpuset_lock(void)
2356 mutex_lock(&callback_mutex);
2359 /**
2360 * cpuset_unlock - release lock on cpuset changes
2362 * Undo the lock taken in a previous cpuset_lock() call.
2363 */
2365 void cpuset_unlock(void)
2367 mutex_unlock(&callback_mutex);
2370 /**
2371 * cpuset_mem_spread_node() - On which node to begin search for a page
2373 * If a task is marked PF_SPREAD_PAGE or PF_SPREAD_SLAB (as for
2374 * tasks in a cpuset with is_spread_page or is_spread_slab set),
2375 * and if the memory allocation used cpuset_mem_spread_node()
2376 * to determine on which node to start looking, as it will for
2377 * certain page cache or slab cache pages such as used for file
2378 * system buffers and inode caches, then instead of starting on the
2379 * local node to look for a free page, rather spread the starting
2380 * node around the tasks mems_allowed nodes.
2382 * We don't have to worry about the returned node being offline
2383 * because "it can't happen", and even if it did, it would be ok.
2385 * The routines calling guarantee_online_mems() are careful to
2386 * only set nodes in task->mems_allowed that are online. So it
2387 * should not be possible for the following code to return an
2388 * offline node. But if it did, that would be ok, as this routine
2389 * is not returning the node where the allocation must be, only
2390 * the node where the search should start. The zonelist passed to
2391 * __alloc_pages() will include all nodes. If the slab allocator
2392 * is passed an offline node, it will fall back to the local node.
2393 * See kmem_cache_alloc_node().
2394 */
2396 int cpuset_mem_spread_node(void)
2398 int node;
2400 node = next_node(current->cpuset_mem_spread_rotor, current->mems_allowed);
2401 if (node == MAX_NUMNODES)
2402 node = first_node(current->mems_allowed);
2403 current->cpuset_mem_spread_rotor = node;
2404 return node;
2406 EXPORT_SYMBOL_GPL(cpuset_mem_spread_node);
2408 /**
2409 * cpuset_excl_nodes_overlap - Do we overlap @p's mem_exclusive ancestors?
2410 * @p: pointer to task_struct of some other task.
2412 * Description: Return true if the nearest mem_exclusive ancestor
2413 * cpusets of tasks @p and current overlap. Used by oom killer to
2414 * determine if task @p's memory usage might impact the memory
2415 * available to the current task.
2417 * Call while holding callback_mutex.
2418 **/
2420 int cpuset_excl_nodes_overlap(const struct task_struct *p)
2422 const struct cpuset *cs1, *cs2; /* my and p's cpuset ancestors */
2423 int overlap = 1; /* do cpusets overlap? */
2425 task_lock(current);
2426 if (current->flags & PF_EXITING) {
2427 task_unlock(current);
2428 goto done;
2430 cs1 = nearest_exclusive_ancestor(current->cpuset);
2431 task_unlock(current);
2433 task_lock((struct task_struct *)p);
2434 if (p->flags & PF_EXITING) {
2435 task_unlock((struct task_struct *)p);
2436 goto done;
2438 cs2 = nearest_exclusive_ancestor(p->cpuset);
2439 task_unlock((struct task_struct *)p);
2441 overlap = nodes_intersects(cs1->mems_allowed, cs2->mems_allowed);
2442 done:
2443 return overlap;
2446 /*
2447 * Collection of memory_pressure is suppressed unless
2448 * this flag is enabled by writing "1" to the special
2449 * cpuset file 'memory_pressure_enabled' in the root cpuset.
2450 */
2452 int cpuset_memory_pressure_enabled __read_mostly;
2454 /**
2455 * cpuset_memory_pressure_bump - keep stats of per-cpuset reclaims.
2457 * Keep a running average of the rate of synchronous (direct)
2458 * page reclaim efforts initiated by tasks in each cpuset.
2460 * This represents the rate at which some task in the cpuset
2461 * ran low on memory on all nodes it was allowed to use, and
2462 * had to enter the kernels page reclaim code in an effort to
2463 * create more free memory by tossing clean pages or swapping
2464 * or writing dirty pages.
2466 * Display to user space in the per-cpuset read-only file
2467 * "memory_pressure". Value displayed is an integer
2468 * representing the recent rate of entry into the synchronous
2469 * (direct) page reclaim by any task attached to the cpuset.
2470 **/
2472 void __cpuset_memory_pressure_bump(void)
2474 struct cpuset *cs;
2476 task_lock(current);
2477 cs = current->cpuset;
2478 fmeter_markevent(&cs->fmeter);
2479 task_unlock(current);
2482 /*
2483 * proc_cpuset_show()
2484 * - Print tasks cpuset path into seq_file.
2485 * - Used for /proc/<pid>/cpuset.
2486 * - No need to task_lock(tsk) on this tsk->cpuset reference, as it
2487 * doesn't really matter if tsk->cpuset changes after we read it,
2488 * and we take manage_mutex, keeping attach_task() from changing it
2489 * anyway. No need to check that tsk->cpuset != NULL, thanks to
2490 * the_top_cpuset_hack in cpuset_exit(), which sets an exiting tasks
2491 * cpuset to top_cpuset.
2492 */
2493 static int proc_cpuset_show(struct seq_file *m, void *v)
2495 struct pid *pid;
2496 struct task_struct *tsk;
2497 char *buf;
2498 int retval;
2500 retval = -ENOMEM;
2501 buf = kmalloc(PAGE_SIZE, GFP_KERNEL);
2502 if (!buf)
2503 goto out;
2505 retval = -ESRCH;
2506 pid = m->private;
2507 tsk = get_pid_task(pid, PIDTYPE_PID);
2508 if (!tsk)
2509 goto out_free;
2511 retval = -EINVAL;
2512 mutex_lock(&manage_mutex);
2514 retval = cpuset_path(tsk->cpuset, buf, PAGE_SIZE);
2515 if (retval < 0)
2516 goto out_unlock;
2517 seq_puts(m, buf);
2518 seq_putc(m, '\n');
2519 out_unlock:
2520 mutex_unlock(&manage_mutex);
2521 put_task_struct(tsk);
2522 out_free:
2523 kfree(buf);
2524 out:
2525 return retval;
2528 static int cpuset_open(struct inode *inode, struct file *file)
2530 struct pid *pid = PROC_I(inode)->pid;
2531 return single_open(file, proc_cpuset_show, pid);
2534 struct file_operations proc_cpuset_operations = {
2535 .open = cpuset_open,
2536 .read = seq_read,
2537 .llseek = seq_lseek,
2538 .release = single_release,
2539 };
2541 /* Display task cpus_allowed, mems_allowed in /proc/<pid>/status file. */
2542 char *cpuset_task_status_allowed(struct task_struct *task, char *buffer)
2544 buffer += sprintf(buffer, "Cpus_allowed:\t");
2545 buffer += cpumask_scnprintf(buffer, PAGE_SIZE, task->cpus_allowed);
2546 buffer += sprintf(buffer, "\n");
2547 buffer += sprintf(buffer, "Mems_allowed:\t");
2548 buffer += nodemask_scnprintf(buffer, PAGE_SIZE, task->mems_allowed);
2549 buffer += sprintf(buffer, "\n");
2550 return buffer;