4 * Processor and Memory placement constraints for sets of tasks.
6 * Copyright (C) 2003 BULL SA.
7 * Copyright (C) 2004 Silicon Graphics, Inc.
9 * Portions derived from Patrick Mochel's sysfs code.
10 * sysfs is Copyright (c) 2001-3 Patrick Mochel
11 * Portions Copyright (c) 2004 Silicon Graphics, Inc.
13 * 2003-10-10 Written by Simon Derr <simon.derr@bull.net>
14 * 2003-10-22 Updates by Stephen Hemminger.
15 * 2004 May-July Rework by Paul Jackson <pj@sgi.com>
17 * This file is subject to the terms and conditions of the GNU General Public
18 * License. See the file COPYING in the main directory of the Linux
19 * distribution for more details.
22 #include <linux/config.h>
23 #include <linux/cpu.h>
24 #include <linux/cpumask.h>
25 #include <linux/cpuset.h>
26 #include <linux/err.h>
27 #include <linux/errno.h>
28 #include <linux/file.h>
30 #include <linux/init.h>
31 #include <linux/interrupt.h>
32 #include <linux/kernel.h>
33 #include <linux/kmod.h>
34 #include <linux/list.h>
35 #include <linux/mempolicy.h>
37 #include <linux/module.h>
38 #include <linux/mount.h>
39 #include <linux/namei.h>
40 #include <linux/pagemap.h>
41 #include <linux/proc_fs.h>
42 #include <linux/sched.h>
43 #include <linux/seq_file.h>
44 #include <linux/slab.h>
45 #include <linux/smp_lock.h>
46 #include <linux/spinlock.h>
47 #include <linux/stat.h>
48 #include <linux/string.h>
49 #include <linux/time.h>
50 #include <linux/backing-dev.h>
51 #include <linux/sort.h>
53 #include <asm/uaccess.h>
54 #include <asm/atomic.h>
55 #include <asm/semaphore.h>
57 #define CPUSET_SUPER_MAGIC 0x27e0eb
60 * Tracks how many cpusets are currently defined in system.
61 * When there is only one cpuset (the root cpuset) we can
62 * short circuit some hooks.
64 int number_of_cpusets;
66 /* See "Frequency meter" comments, below. */
69 int cnt; /* unprocessed events count */
70 int val; /* most recent output value */
71 time_t time; /* clock (secs) when val computed */
72 spinlock_t lock; /* guards read or write of above */
76 unsigned long flags; /* "unsigned long" so bitops work */
77 cpumask_t cpus_allowed; /* CPUs allowed to tasks in cpuset */
78 nodemask_t mems_allowed; /* Memory Nodes allowed to tasks */
81 * Count is atomic so can incr (fork) or decr (exit) without a lock.
83 atomic_t count; /* count tasks using this cpuset */
86 * We link our 'sibling' struct into our parents 'children'.
87 * Our children link their 'sibling' into our 'children'.
89 struct list_head sibling; /* my parents children */
90 struct list_head children; /* my children */
92 struct cpuset *parent; /* my parent */
93 struct dentry *dentry; /* cpuset fs entry */
96 * Copy of global cpuset_mems_generation as of the most
97 * recent time this cpuset changed its mems_allowed.
101 struct fmeter fmeter; /* memory_pressure filter */
104 /* bits in struct cpuset flags field */
113 /* convenient tests for these bits */
114 static inline int is_cpu_exclusive(const struct cpuset *cs)
116 return !!test_bit(CS_CPU_EXCLUSIVE, &cs->flags);
119 static inline int is_mem_exclusive(const struct cpuset *cs)
121 return !!test_bit(CS_MEM_EXCLUSIVE, &cs->flags);
124 static inline int is_removed(const struct cpuset *cs)
126 return !!test_bit(CS_REMOVED, &cs->flags);
129 static inline int notify_on_release(const struct cpuset *cs)
131 return !!test_bit(CS_NOTIFY_ON_RELEASE, &cs->flags);
134 static inline int is_memory_migrate(const struct cpuset *cs)
136 return !!test_bit(CS_MEMORY_MIGRATE, &cs->flags);
140 * Increment this atomic integer everytime any cpuset changes its
141 * mems_allowed value. Users of cpusets can track this generation
142 * number, and avoid having to lock and reload mems_allowed unless
143 * the cpuset they're using changes generation.
145 * A single, global generation is needed because attach_task() could
146 * reattach a task to a different cpuset, which must not have its
147 * generation numbers aliased with those of that tasks previous cpuset.
149 * Generations are needed for mems_allowed because one task cannot
150 * modify anothers memory placement. So we must enable every task,
151 * on every visit to __alloc_pages(), to efficiently check whether
152 * its current->cpuset->mems_allowed has changed, requiring an update
153 * of its current->mems_allowed.
155 static atomic_t cpuset_mems_generation = ATOMIC_INIT(1);
157 static struct cpuset top_cpuset = {
158 .flags = ((1 << CS_CPU_EXCLUSIVE) | (1 << CS_MEM_EXCLUSIVE)),
159 .cpus_allowed = CPU_MASK_ALL,
160 .mems_allowed = NODE_MASK_ALL,
161 .count = ATOMIC_INIT(0),
162 .sibling = LIST_HEAD_INIT(top_cpuset.sibling),
163 .children = LIST_HEAD_INIT(top_cpuset.children),
166 static struct vfsmount *cpuset_mount;
167 static struct super_block *cpuset_sb;
170 * We have two global cpuset semaphores below. They can nest.
171 * It is ok to first take manage_sem, then nest callback_sem. We also
172 * require taking task_lock() when dereferencing a tasks cpuset pointer.
173 * See "The task_lock() exception", at the end of this comment.
175 * A task must hold both semaphores to modify cpusets. If a task
176 * holds manage_sem, then it blocks others wanting that semaphore,
177 * ensuring that it is the only task able to also acquire callback_sem
178 * and be able to modify cpusets. It can perform various checks on
179 * the cpuset structure first, knowing nothing will change. It can
180 * also allocate memory while just holding manage_sem. While it is
181 * performing these checks, various callback routines can briefly
182 * acquire callback_sem to query cpusets. Once it is ready to make
183 * the changes, it takes callback_sem, blocking everyone else.
185 * Calls to the kernel memory allocator can not be made while holding
186 * callback_sem, as that would risk double tripping on callback_sem
187 * from one of the callbacks into the cpuset code from within
190 * If a task is only holding callback_sem, then it has read-only
193 * The task_struct fields mems_allowed and mems_generation may only
194 * be accessed in the context of that task, so require no locks.
196 * Any task can increment and decrement the count field without lock.
197 * So in general, code holding manage_sem or callback_sem can't rely
198 * on the count field not changing. However, if the count goes to
199 * zero, then only attach_task(), which holds both semaphores, can
200 * increment it again. Because a count of zero means that no tasks
201 * are currently attached, therefore there is no way a task attached
202 * to that cpuset can fork (the other way to increment the count).
203 * So code holding manage_sem or callback_sem can safely assume that
204 * if the count is zero, it will stay zero. Similarly, if a task
205 * holds manage_sem or callback_sem on a cpuset with zero count, it
206 * knows that the cpuset won't be removed, as cpuset_rmdir() needs
207 * both of those semaphores.
209 * A possible optimization to improve parallelism would be to make
210 * callback_sem a R/W semaphore (rwsem), allowing the callback routines
211 * to proceed in parallel, with read access, until the holder of
212 * manage_sem needed to take this rwsem for exclusive write access
213 * and modify some cpusets.
215 * The cpuset_common_file_write handler for operations that modify
216 * the cpuset hierarchy holds manage_sem across the entire operation,
217 * single threading all such cpuset modifications across the system.
219 * The cpuset_common_file_read() handlers only hold callback_sem across
220 * small pieces of code, such as when reading out possibly multi-word
221 * cpumasks and nodemasks.
223 * The fork and exit callbacks cpuset_fork() and cpuset_exit(), don't
224 * (usually) take either semaphore. These are the two most performance
225 * critical pieces of code here. The exception occurs on cpuset_exit(),
226 * when a task in a notify_on_release cpuset exits. Then manage_sem
227 * is taken, and if the cpuset count is zero, a usermode call made
228 * to /sbin/cpuset_release_agent with the name of the cpuset (path
229 * relative to the root of cpuset file system) as the argument.
231 * A cpuset can only be deleted if both its 'count' of using tasks
232 * is zero, and its list of 'children' cpusets is empty. Since all
233 * tasks in the system use _some_ cpuset, and since there is always at
234 * least one task in the system (init, pid == 1), therefore, top_cpuset
235 * always has either children cpusets and/or using tasks. So we don't
236 * need a special hack to ensure that top_cpuset cannot be deleted.
238 * The above "Tale of Two Semaphores" would be complete, but for:
240 * The task_lock() exception
242 * The need for this exception arises from the action of attach_task(),
243 * which overwrites one tasks cpuset pointer with another. It does
244 * so using both semaphores, however there are several performance
245 * critical places that need to reference task->cpuset without the
246 * expense of grabbing a system global semaphore. Therefore except as
247 * noted below, when dereferencing or, as in attach_task(), modifying
248 * a tasks cpuset pointer we use task_lock(), which acts on a spinlock
249 * (task->alloc_lock) already in the task_struct routinely used for
253 static DECLARE_MUTEX(manage_sem);
254 static DECLARE_MUTEX(callback_sem);
257 * A couple of forward declarations required, due to cyclic reference loop:
258 * cpuset_mkdir -> cpuset_create -> cpuset_populate_dir -> cpuset_add_file
259 * -> cpuset_create_file -> cpuset_dir_inode_operations -> cpuset_mkdir.
262 static int cpuset_mkdir(struct inode *dir, struct dentry *dentry, int mode);
263 static int cpuset_rmdir(struct inode *unused_dir, struct dentry *dentry);
265 static struct backing_dev_info cpuset_backing_dev_info = {
266 .ra_pages = 0, /* No readahead */
267 .capabilities = BDI_CAP_NO_ACCT_DIRTY | BDI_CAP_NO_WRITEBACK,
270 static struct inode *cpuset_new_inode(mode_t mode)
272 struct inode *inode = new_inode(cpuset_sb);
275 inode->i_mode = mode;
276 inode->i_uid = current->fsuid;
277 inode->i_gid = current->fsgid;
278 inode->i_blksize = PAGE_CACHE_SIZE;
280 inode->i_atime = inode->i_mtime = inode->i_ctime = CURRENT_TIME;
281 inode->i_mapping->backing_dev_info = &cpuset_backing_dev_info;
286 static void cpuset_diput(struct dentry *dentry, struct inode *inode)
288 /* is dentry a directory ? if so, kfree() associated cpuset */
289 if (S_ISDIR(inode->i_mode)) {
290 struct cpuset *cs = dentry->d_fsdata;
291 BUG_ON(!(is_removed(cs)));
297 static struct dentry_operations cpuset_dops = {
298 .d_iput = cpuset_diput,
301 static struct dentry *cpuset_get_dentry(struct dentry *parent, const char *name)
303 struct dentry *d = lookup_one_len(name, parent, strlen(name));
305 d->d_op = &cpuset_dops;
309 static void remove_dir(struct dentry *d)
311 struct dentry *parent = dget(d->d_parent);
314 simple_rmdir(parent->d_inode, d);
319 * NOTE : the dentry must have been dget()'ed
321 static void cpuset_d_remove_dir(struct dentry *dentry)
323 struct list_head *node;
325 spin_lock(&dcache_lock);
326 node = dentry->d_subdirs.next;
327 while (node != &dentry->d_subdirs) {
328 struct dentry *d = list_entry(node, struct dentry, d_child);
332 spin_unlock(&dcache_lock);
334 simple_unlink(dentry->d_inode, d);
336 spin_lock(&dcache_lock);
338 node = dentry->d_subdirs.next;
340 list_del_init(&dentry->d_child);
341 spin_unlock(&dcache_lock);
345 static struct super_operations cpuset_ops = {
346 .statfs = simple_statfs,
347 .drop_inode = generic_delete_inode,
350 static int cpuset_fill_super(struct super_block *sb, void *unused_data,
356 sb->s_blocksize = PAGE_CACHE_SIZE;
357 sb->s_blocksize_bits = PAGE_CACHE_SHIFT;
358 sb->s_magic = CPUSET_SUPER_MAGIC;
359 sb->s_op = &cpuset_ops;
362 inode = cpuset_new_inode(S_IFDIR | S_IRUGO | S_IXUGO | S_IWUSR);
364 inode->i_op = &simple_dir_inode_operations;
365 inode->i_fop = &simple_dir_operations;
366 /* directories start off with i_nlink == 2 (for "." entry) */
372 root = d_alloc_root(inode);
381 static struct super_block *cpuset_get_sb(struct file_system_type *fs_type,
382 int flags, const char *unused_dev_name,
385 return get_sb_single(fs_type, flags, data, cpuset_fill_super);
388 static struct file_system_type cpuset_fs_type = {
390 .get_sb = cpuset_get_sb,
391 .kill_sb = kill_litter_super,
396 * The files in the cpuset filesystem mostly have a very simple read/write
397 * handling, some common function will take care of it. Nevertheless some cases
398 * (read tasks) are special and therefore I define this structure for every
402 * When reading/writing to a file:
403 * - the cpuset to use in file->f_dentry->d_parent->d_fsdata
404 * - the 'cftype' of the file is file->f_dentry->d_fsdata
410 int (*open) (struct inode *inode, struct file *file);
411 ssize_t (*read) (struct file *file, char __user *buf, size_t nbytes,
413 int (*write) (struct file *file, const char __user *buf, size_t nbytes,
415 int (*release) (struct inode *inode, struct file *file);
418 static inline struct cpuset *__d_cs(struct dentry *dentry)
420 return dentry->d_fsdata;
423 static inline struct cftype *__d_cft(struct dentry *dentry)
425 return dentry->d_fsdata;
429 * Call with manage_sem held. Writes path of cpuset into buf.
430 * Returns 0 on success, -errno on error.
433 static int cpuset_path(const struct cpuset *cs, char *buf, int buflen)
437 start = buf + buflen;
441 int len = cs->dentry->d_name.len;
442 if ((start -= len) < buf)
443 return -ENAMETOOLONG;
444 memcpy(start, cs->dentry->d_name.name, len);
451 return -ENAMETOOLONG;
454 memmove(buf, start, buf + buflen - start);
459 * Notify userspace when a cpuset is released, by running
460 * /sbin/cpuset_release_agent with the name of the cpuset (path
461 * relative to the root of cpuset file system) as the argument.
463 * Most likely, this user command will try to rmdir this cpuset.
465 * This races with the possibility that some other task will be
466 * attached to this cpuset before it is removed, or that some other
467 * user task will 'mkdir' a child cpuset of this cpuset. That's ok.
468 * The presumed 'rmdir' will fail quietly if this cpuset is no longer
469 * unused, and this cpuset will be reprieved from its death sentence,
470 * to continue to serve a useful existence. Next time it's released,
471 * we will get notified again, if it still has 'notify_on_release' set.
473 * The final arg to call_usermodehelper() is 0, which means don't
474 * wait. The separate /sbin/cpuset_release_agent task is forked by
475 * call_usermodehelper(), then control in this thread returns here,
476 * without waiting for the release agent task. We don't bother to
477 * wait because the caller of this routine has no use for the exit
478 * status of the /sbin/cpuset_release_agent task, so no sense holding
479 * our caller up for that.
481 * When we had only one cpuset semaphore, we had to call this
482 * without holding it, to avoid deadlock when call_usermodehelper()
483 * allocated memory. With two locks, we could now call this while
484 * holding manage_sem, but we still don't, so as to minimize
485 * the time manage_sem is held.
488 static void cpuset_release_agent(const char *pathbuf)
490 char *argv[3], *envp[3];
497 argv[i++] = "/sbin/cpuset_release_agent";
498 argv[i++] = (char *)pathbuf;
502 /* minimal command environment */
503 envp[i++] = "HOME=/";
504 envp[i++] = "PATH=/sbin:/bin:/usr/sbin:/usr/bin";
507 call_usermodehelper(argv[0], argv, envp, 0);
512 * Either cs->count of using tasks transitioned to zero, or the
513 * cs->children list of child cpusets just became empty. If this
514 * cs is notify_on_release() and now both the user count is zero and
515 * the list of children is empty, prepare cpuset path in a kmalloc'd
516 * buffer, to be returned via ppathbuf, so that the caller can invoke
517 * cpuset_release_agent() with it later on, once manage_sem is dropped.
518 * Call here with manage_sem held.
520 * This check_for_release() routine is responsible for kmalloc'ing
521 * pathbuf. The above cpuset_release_agent() is responsible for
522 * kfree'ing pathbuf. The caller of these routines is responsible
523 * for providing a pathbuf pointer, initialized to NULL, then
524 * calling check_for_release() with manage_sem held and the address
525 * of the pathbuf pointer, then dropping manage_sem, then calling
526 * cpuset_release_agent() with pathbuf, as set by check_for_release().
529 static void check_for_release(struct cpuset *cs, char **ppathbuf)
531 if (notify_on_release(cs) && atomic_read(&cs->count) == 0 &&
532 list_empty(&cs->children)) {
535 buf = kmalloc(PAGE_SIZE, GFP_KERNEL);
538 if (cpuset_path(cs, buf, PAGE_SIZE) < 0)
546 * Return in *pmask the portion of a cpusets's cpus_allowed that
547 * are online. If none are online, walk up the cpuset hierarchy
548 * until we find one that does have some online cpus. If we get
549 * all the way to the top and still haven't found any online cpus,
550 * return cpu_online_map. Or if passed a NULL cs from an exit'ing
551 * task, return cpu_online_map.
553 * One way or another, we guarantee to return some non-empty subset
556 * Call with callback_sem held.
559 static void guarantee_online_cpus(const struct cpuset *cs, cpumask_t *pmask)
561 while (cs && !cpus_intersects(cs->cpus_allowed, cpu_online_map))
564 cpus_and(*pmask, cs->cpus_allowed, cpu_online_map);
566 *pmask = cpu_online_map;
567 BUG_ON(!cpus_intersects(*pmask, cpu_online_map));
571 * Return in *pmask the portion of a cpusets's mems_allowed that
572 * are online. If none are online, walk up the cpuset hierarchy
573 * until we find one that does have some online mems. If we get
574 * all the way to the top and still haven't found any online mems,
575 * return node_online_map.
577 * One way or another, we guarantee to return some non-empty subset
578 * of node_online_map.
580 * Call with callback_sem held.
583 static void guarantee_online_mems(const struct cpuset *cs, nodemask_t *pmask)
585 while (cs && !nodes_intersects(cs->mems_allowed, node_online_map))
588 nodes_and(*pmask, cs->mems_allowed, node_online_map);
590 *pmask = node_online_map;
591 BUG_ON(!nodes_intersects(*pmask, node_online_map));
595 * cpuset_update_task_memory_state - update task memory placement
597 * If the current tasks cpusets mems_allowed changed behind our
598 * backs, update current->mems_allowed, mems_generation and task NUMA
599 * mempolicy to the new value.
601 * Task mempolicy is updated by rebinding it relative to the
602 * current->cpuset if a task has its memory placement changed.
603 * Do not call this routine if in_interrupt().
605 * Call without callback_sem or task_lock() held. May be called
606 * with or without manage_sem held. Except in early boot or
607 * an exiting task, when tsk->cpuset is NULL, this routine will
608 * acquire task_lock(). We don't need to use task_lock to guard
609 * against another task changing a non-NULL cpuset pointer to NULL,
610 * as that is only done by a task on itself, and if the current task
611 * is here, it is not simultaneously in the exit code NULL'ing its
612 * cpuset pointer. This routine also might acquire callback_sem and
613 * current->mm->mmap_sem during call.
615 * The task_lock() is required to dereference current->cpuset safely.
616 * Without it, we could pick up the pointer value of current->cpuset
617 * in one instruction, and then attach_task could give us a different
618 * cpuset, and then the cpuset we had could be removed and freed,
619 * and then on our next instruction, we could dereference a no longer
620 * valid cpuset pointer to get its mems_generation field.
622 * This routine is needed to update the per-task mems_allowed data,
623 * within the tasks context, when it is trying to allocate memory
624 * (in various mm/mempolicy.c routines) and notices that some other
625 * task has been modifying its cpuset.
628 void cpuset_update_task_memory_state()
630 int my_cpusets_mem_gen;
631 struct task_struct *tsk = current;
632 struct cpuset *cs = tsk->cpuset;
638 my_cpusets_mem_gen = cs->mems_generation;
641 if (my_cpusets_mem_gen != tsk->cpuset_mems_generation) {
642 nodemask_t oldmem = tsk->mems_allowed;
647 cs = tsk->cpuset; /* Maybe changed when task not locked */
648 migrate = is_memory_migrate(cs);
649 guarantee_online_mems(cs, &tsk->mems_allowed);
650 tsk->cpuset_mems_generation = cs->mems_generation;
653 mpol_rebind_task(tsk, &tsk->mems_allowed);
654 if (!nodes_equal(oldmem, tsk->mems_allowed)) {
656 do_migrate_pages(tsk->mm, &oldmem,
665 * is_cpuset_subset(p, q) - Is cpuset p a subset of cpuset q?
667 * One cpuset is a subset of another if all its allowed CPUs and
668 * Memory Nodes are a subset of the other, and its exclusive flags
669 * are only set if the other's are set. Call holding manage_sem.
672 static int is_cpuset_subset(const struct cpuset *p, const struct cpuset *q)
674 return cpus_subset(p->cpus_allowed, q->cpus_allowed) &&
675 nodes_subset(p->mems_allowed, q->mems_allowed) &&
676 is_cpu_exclusive(p) <= is_cpu_exclusive(q) &&
677 is_mem_exclusive(p) <= is_mem_exclusive(q);
681 * validate_change() - Used to validate that any proposed cpuset change
682 * follows the structural rules for cpusets.
684 * If we replaced the flag and mask values of the current cpuset
685 * (cur) with those values in the trial cpuset (trial), would
686 * our various subset and exclusive rules still be valid? Presumes
689 * 'cur' is the address of an actual, in-use cpuset. Operations
690 * such as list traversal that depend on the actual address of the
691 * cpuset in the list must use cur below, not trial.
693 * 'trial' is the address of bulk structure copy of cur, with
694 * perhaps one or more of the fields cpus_allowed, mems_allowed,
695 * or flags changed to new, trial values.
697 * Return 0 if valid, -errno if not.
700 static int validate_change(const struct cpuset *cur, const struct cpuset *trial)
702 struct cpuset *c, *par;
704 /* Each of our child cpusets must be a subset of us */
705 list_for_each_entry(c, &cur->children, sibling) {
706 if (!is_cpuset_subset(c, trial))
710 /* Remaining checks don't apply to root cpuset */
711 if ((par = cur->parent) == NULL)
714 /* We must be a subset of our parent cpuset */
715 if (!is_cpuset_subset(trial, par))
718 /* If either I or some sibling (!= me) is exclusive, we can't overlap */
719 list_for_each_entry(c, &par->children, sibling) {
720 if ((is_cpu_exclusive(trial) || is_cpu_exclusive(c)) &&
722 cpus_intersects(trial->cpus_allowed, c->cpus_allowed))
724 if ((is_mem_exclusive(trial) || is_mem_exclusive(c)) &&
726 nodes_intersects(trial->mems_allowed, c->mems_allowed))
734 * For a given cpuset cur, partition the system as follows
735 * a. All cpus in the parent cpuset's cpus_allowed that are not part of any
736 * exclusive child cpusets
737 * b. All cpus in the current cpuset's cpus_allowed that are not part of any
738 * exclusive child cpusets
739 * Build these two partitions by calling partition_sched_domains
741 * Call with manage_sem held. May nest a call to the
742 * lock_cpu_hotplug()/unlock_cpu_hotplug() pair.
745 static void update_cpu_domains(struct cpuset *cur)
747 struct cpuset *c, *par = cur->parent;
748 cpumask_t pspan, cspan;
750 if (par == NULL || cpus_empty(cur->cpus_allowed))
754 * Get all cpus from parent's cpus_allowed not part of exclusive
757 pspan = par->cpus_allowed;
758 list_for_each_entry(c, &par->children, sibling) {
759 if (is_cpu_exclusive(c))
760 cpus_andnot(pspan, pspan, c->cpus_allowed);
762 if (is_removed(cur) || !is_cpu_exclusive(cur)) {
763 cpus_or(pspan, pspan, cur->cpus_allowed);
764 if (cpus_equal(pspan, cur->cpus_allowed))
766 cspan = CPU_MASK_NONE;
768 if (cpus_empty(pspan))
770 cspan = cur->cpus_allowed;
772 * Get all cpus from current cpuset's cpus_allowed not part
773 * of exclusive children
775 list_for_each_entry(c, &cur->children, sibling) {
776 if (is_cpu_exclusive(c))
777 cpus_andnot(cspan, cspan, c->cpus_allowed);
782 partition_sched_domains(&pspan, &cspan);
783 unlock_cpu_hotplug();
787 * Call with manage_sem held. May take callback_sem during call.
790 static int update_cpumask(struct cpuset *cs, char *buf)
792 struct cpuset trialcs;
793 int retval, cpus_unchanged;
796 retval = cpulist_parse(buf, trialcs.cpus_allowed);
799 cpus_and(trialcs.cpus_allowed, trialcs.cpus_allowed, cpu_online_map);
800 if (cpus_empty(trialcs.cpus_allowed))
802 retval = validate_change(cs, &trialcs);
805 cpus_unchanged = cpus_equal(cs->cpus_allowed, trialcs.cpus_allowed);
807 cs->cpus_allowed = trialcs.cpus_allowed;
809 if (is_cpu_exclusive(cs) && !cpus_unchanged)
810 update_cpu_domains(cs);
815 * Handle user request to change the 'mems' memory placement
816 * of a cpuset. Needs to validate the request, update the
817 * cpusets mems_allowed and mems_generation, and for each
818 * task in the cpuset, rebind any vma mempolicies.
820 * Call with manage_sem held. May take callback_sem during call.
821 * Will take tasklist_lock, scan tasklist for tasks in cpuset cs,
822 * lock each such tasks mm->mmap_sem, scan its vma's and rebind
823 * their mempolicies to the cpusets new mems_allowed.
826 static int update_nodemask(struct cpuset *cs, char *buf)
828 struct cpuset trialcs;
829 struct task_struct *g, *p;
830 struct mm_struct **mmarray;
836 retval = nodelist_parse(buf, trialcs.mems_allowed);
839 nodes_and(trialcs.mems_allowed, trialcs.mems_allowed, node_online_map);
840 if (nodes_empty(trialcs.mems_allowed)) {
844 retval = validate_change(cs, &trialcs);
849 cs->mems_allowed = trialcs.mems_allowed;
850 atomic_inc(&cpuset_mems_generation);
851 cs->mems_generation = atomic_read(&cpuset_mems_generation);
854 set_cpuset_being_rebound(cs); /* causes mpol_copy() rebind */
856 fudge = 10; /* spare mmarray[] slots */
857 fudge += cpus_weight(cs->cpus_allowed); /* imagine one fork-bomb/cpu */
861 * Allocate mmarray[] to hold mm reference for each task
862 * in cpuset cs. Can't kmalloc GFP_KERNEL while holding
863 * tasklist_lock. We could use GFP_ATOMIC, but with a
864 * few more lines of code, we can retry until we get a big
865 * enough mmarray[] w/o using GFP_ATOMIC.
868 ntasks = atomic_read(&cs->count); /* guess */
870 mmarray = kmalloc(ntasks * sizeof(*mmarray), GFP_KERNEL);
873 write_lock_irq(&tasklist_lock); /* block fork */
874 if (atomic_read(&cs->count) <= ntasks)
875 break; /* got enough */
876 write_unlock_irq(&tasklist_lock); /* try again */
882 /* Load up mmarray[] with mm reference for each task in cpuset. */
883 do_each_thread(g, p) {
884 struct mm_struct *mm;
888 "Cpuset mempolicy rebind incomplete.\n");
897 } while_each_thread(g, p);
898 write_unlock_irq(&tasklist_lock);
901 * Now that we've dropped the tasklist spinlock, we can
902 * rebind the vma mempolicies of each mm in mmarray[] to their
903 * new cpuset, and release that mm. The mpol_rebind_mm()
904 * call takes mmap_sem, which we couldn't take while holding
905 * tasklist_lock. Forks can happen again now - the mpol_copy()
906 * cpuset_being_rebound check will catch such forks, and rebind
907 * their vma mempolicies too. Because we still hold the global
908 * cpuset manage_sem, we know that no other rebind effort will
909 * be contending for the global variable cpuset_being_rebound.
910 * It's ok if we rebind the same mm twice; mpol_rebind_mm()
913 for (i = 0; i < n; i++) {
914 struct mm_struct *mm = mmarray[i];
916 mpol_rebind_mm(mm, &cs->mems_allowed);
920 /* We're done rebinding vma's to this cpusets new mems_allowed. */
922 set_cpuset_being_rebound(NULL);
929 * Call with manage_sem held.
932 static int update_memory_pressure_enabled(struct cpuset *cs, char *buf)
934 if (simple_strtoul(buf, NULL, 10) != 0)
935 cpuset_memory_pressure_enabled = 1;
937 cpuset_memory_pressure_enabled = 0;
942 * update_flag - read a 0 or a 1 in a file and update associated flag
943 * bit: the bit to update (CS_CPU_EXCLUSIVE, CS_MEM_EXCLUSIVE,
944 * CS_NOTIFY_ON_RELEASE, CS_MEMORY_MIGRATE)
945 * cs: the cpuset to update
946 * buf: the buffer where we read the 0 or 1
948 * Call with manage_sem held.
951 static int update_flag(cpuset_flagbits_t bit, struct cpuset *cs, char *buf)
954 struct cpuset trialcs;
955 int err, cpu_exclusive_changed;
957 turning_on = (simple_strtoul(buf, NULL, 10) != 0);
961 set_bit(bit, &trialcs.flags);
963 clear_bit(bit, &trialcs.flags);
965 err = validate_change(cs, &trialcs);
968 cpu_exclusive_changed =
969 (is_cpu_exclusive(cs) != is_cpu_exclusive(&trialcs));
972 set_bit(bit, &cs->flags);
974 clear_bit(bit, &cs->flags);
977 if (cpu_exclusive_changed)
978 update_cpu_domains(cs);
983 * Frequency meter - How fast is some event occuring?
985 * These routines manage a digitally filtered, constant time based,
986 * event frequency meter. There are four routines:
987 * fmeter_init() - initialize a frequency meter.
988 * fmeter_markevent() - called each time the event happens.
989 * fmeter_getrate() - returns the recent rate of such events.
990 * fmeter_update() - internal routine used to update fmeter.
992 * A common data structure is passed to each of these routines,
993 * which is used to keep track of the state required to manage the
994 * frequency meter and its digital filter.
996 * The filter works on the number of events marked per unit time.
997 * The filter is single-pole low-pass recursive (IIR). The time unit
998 * is 1 second. Arithmetic is done using 32-bit integers scaled to
999 * simulate 3 decimal digits of precision (multiplied by 1000).
1001 * With an FM_COEF of 933, and a time base of 1 second, the filter
1002 * has a half-life of 10 seconds, meaning that if the events quit
1003 * happening, then the rate returned from the fmeter_getrate()
1004 * will be cut in half each 10 seconds, until it converges to zero.
1006 * It is not worth doing a real infinitely recursive filter. If more
1007 * than FM_MAXTICKS ticks have elapsed since the last filter event,
1008 * just compute FM_MAXTICKS ticks worth, by which point the level
1011 * Limit the count of unprocessed events to FM_MAXCNT, so as to avoid
1012 * arithmetic overflow in the fmeter_update() routine.
1014 * Given the simple 32 bit integer arithmetic used, this meter works
1015 * best for reporting rates between one per millisecond (msec) and
1016 * one per 32 (approx) seconds. At constant rates faster than one
1017 * per msec it maxes out at values just under 1,000,000. At constant
1018 * rates between one per msec, and one per second it will stabilize
1019 * to a value N*1000, where N is the rate of events per second.
1020 * At constant rates between one per second and one per 32 seconds,
1021 * it will be choppy, moving up on the seconds that have an event,
1022 * and then decaying until the next event. At rates slower than
1023 * about one in 32 seconds, it decays all the way back to zero between
1027 #define FM_COEF 933 /* coefficient for half-life of 10 secs */
1028 #define FM_MAXTICKS ((time_t)99) /* useless computing more ticks than this */
1029 #define FM_MAXCNT 1000000 /* limit cnt to avoid overflow */
1030 #define FM_SCALE 1000 /* faux fixed point scale */
1032 /* Initialize a frequency meter */
1033 static void fmeter_init(struct fmeter *fmp)
1038 spin_lock_init(&fmp->lock);
1041 /* Internal meter update - process cnt events and update value */
1042 static void fmeter_update(struct fmeter *fmp)
1044 time_t now = get_seconds();
1045 time_t ticks = now - fmp->time;
1050 ticks = min(FM_MAXTICKS, ticks);
1052 fmp->val = (FM_COEF * fmp->val) / FM_SCALE;
1055 fmp->val += ((FM_SCALE - FM_COEF) * fmp->cnt) / FM_SCALE;
1059 /* Process any previous ticks, then bump cnt by one (times scale). */
1060 static void fmeter_markevent(struct fmeter *fmp)
1062 spin_lock(&fmp->lock);
1064 fmp->cnt = min(FM_MAXCNT, fmp->cnt + FM_SCALE);
1065 spin_unlock(&fmp->lock);
1068 /* Process any previous ticks, then return current value. */
1069 static int fmeter_getrate(struct fmeter *fmp)
1073 spin_lock(&fmp->lock);
1076 spin_unlock(&fmp->lock);
1081 * Attack task specified by pid in 'pidbuf' to cpuset 'cs', possibly
1082 * writing the path of the old cpuset in 'ppathbuf' if it needs to be
1083 * notified on release.
1085 * Call holding manage_sem. May take callback_sem and task_lock of
1086 * the task 'pid' during call.
1089 static int attach_task(struct cpuset *cs, char *pidbuf, char **ppathbuf)
1092 struct task_struct *tsk;
1093 struct cpuset *oldcs;
1095 nodemask_t from, to;
1096 struct mm_struct *mm;
1098 if (sscanf(pidbuf, "%d", &pid) != 1)
1100 if (cpus_empty(cs->cpus_allowed) || nodes_empty(cs->mems_allowed))
1104 read_lock(&tasklist_lock);
1106 tsk = find_task_by_pid(pid);
1107 if (!tsk || tsk->flags & PF_EXITING) {
1108 read_unlock(&tasklist_lock);
1112 get_task_struct(tsk);
1113 read_unlock(&tasklist_lock);
1115 if ((current->euid) && (current->euid != tsk->uid)
1116 && (current->euid != tsk->suid)) {
1117 put_task_struct(tsk);
1122 get_task_struct(tsk);
1125 down(&callback_sem);
1128 oldcs = tsk->cpuset;
1132 put_task_struct(tsk);
1135 atomic_inc(&cs->count);
1139 guarantee_online_cpus(cs, &cpus);
1140 set_cpus_allowed(tsk, cpus);
1142 from = oldcs->mems_allowed;
1143 to = cs->mems_allowed;
1147 mm = get_task_mm(tsk);
1149 mpol_rebind_mm(mm, &to);
1153 if (is_memory_migrate(cs))
1154 do_migrate_pages(tsk->mm, &from, &to, MPOL_MF_MOVE_ALL);
1155 put_task_struct(tsk);
1156 if (atomic_dec_and_test(&oldcs->count))
1157 check_for_release(oldcs, ppathbuf);
1161 /* The various types of files and directories in a cpuset file system */
1166 FILE_MEMORY_MIGRATE,
1171 FILE_NOTIFY_ON_RELEASE,
1172 FILE_MEMORY_PRESSURE_ENABLED,
1173 FILE_MEMORY_PRESSURE,
1175 } cpuset_filetype_t;
1177 static ssize_t cpuset_common_file_write(struct file *file, const char __user *userbuf,
1178 size_t nbytes, loff_t *unused_ppos)
1180 struct cpuset *cs = __d_cs(file->f_dentry->d_parent);
1181 struct cftype *cft = __d_cft(file->f_dentry);
1182 cpuset_filetype_t type = cft->private;
1184 char *pathbuf = NULL;
1187 /* Crude upper limit on largest legitimate cpulist user might write. */
1188 if (nbytes > 100 + 6 * NR_CPUS)
1191 /* +1 for nul-terminator */
1192 if ((buffer = kmalloc(nbytes + 1, GFP_KERNEL)) == 0)
1195 if (copy_from_user(buffer, userbuf, nbytes)) {
1199 buffer[nbytes] = 0; /* nul-terminate */
1203 if (is_removed(cs)) {
1210 retval = update_cpumask(cs, buffer);
1213 retval = update_nodemask(cs, buffer);
1215 case FILE_CPU_EXCLUSIVE:
1216 retval = update_flag(CS_CPU_EXCLUSIVE, cs, buffer);
1218 case FILE_MEM_EXCLUSIVE:
1219 retval = update_flag(CS_MEM_EXCLUSIVE, cs, buffer);
1221 case FILE_NOTIFY_ON_RELEASE:
1222 retval = update_flag(CS_NOTIFY_ON_RELEASE, cs, buffer);
1224 case FILE_MEMORY_MIGRATE:
1225 retval = update_flag(CS_MEMORY_MIGRATE, cs, buffer);
1227 case FILE_MEMORY_PRESSURE_ENABLED:
1228 retval = update_memory_pressure_enabled(cs, buffer);
1230 case FILE_MEMORY_PRESSURE:
1234 retval = attach_task(cs, buffer, &pathbuf);
1245 cpuset_release_agent(pathbuf);
1251 static ssize_t cpuset_file_write(struct file *file, const char __user *buf,
1252 size_t nbytes, loff_t *ppos)
1255 struct cftype *cft = __d_cft(file->f_dentry);
1259 /* special function ? */
1261 retval = cft->write(file, buf, nbytes, ppos);
1263 retval = cpuset_common_file_write(file, buf, nbytes, ppos);
1269 * These ascii lists should be read in a single call, by using a user
1270 * buffer large enough to hold the entire map. If read in smaller
1271 * chunks, there is no guarantee of atomicity. Since the display format
1272 * used, list of ranges of sequential numbers, is variable length,
1273 * and since these maps can change value dynamically, one could read
1274 * gibberish by doing partial reads while a list was changing.
1275 * A single large read to a buffer that crosses a page boundary is
1276 * ok, because the result being copied to user land is not recomputed
1277 * across a page fault.
1280 static int cpuset_sprintf_cpulist(char *page, struct cpuset *cs)
1284 down(&callback_sem);
1285 mask = cs->cpus_allowed;
1288 return cpulist_scnprintf(page, PAGE_SIZE, mask);
1291 static int cpuset_sprintf_memlist(char *page, struct cpuset *cs)
1295 down(&callback_sem);
1296 mask = cs->mems_allowed;
1299 return nodelist_scnprintf(page, PAGE_SIZE, mask);
1302 static ssize_t cpuset_common_file_read(struct file *file, char __user *buf,
1303 size_t nbytes, loff_t *ppos)
1305 struct cftype *cft = __d_cft(file->f_dentry);
1306 struct cpuset *cs = __d_cs(file->f_dentry->d_parent);
1307 cpuset_filetype_t type = cft->private;
1312 if (!(page = (char *)__get_free_page(GFP_KERNEL)))
1319 s += cpuset_sprintf_cpulist(s, cs);
1322 s += cpuset_sprintf_memlist(s, cs);
1324 case FILE_CPU_EXCLUSIVE:
1325 *s++ = is_cpu_exclusive(cs) ? '1' : '0';
1327 case FILE_MEM_EXCLUSIVE:
1328 *s++ = is_mem_exclusive(cs) ? '1' : '0';
1330 case FILE_NOTIFY_ON_RELEASE:
1331 *s++ = notify_on_release(cs) ? '1' : '0';
1333 case FILE_MEMORY_MIGRATE:
1334 *s++ = is_memory_migrate(cs) ? '1' : '0';
1336 case FILE_MEMORY_PRESSURE_ENABLED:
1337 *s++ = cpuset_memory_pressure_enabled ? '1' : '0';
1339 case FILE_MEMORY_PRESSURE:
1340 s += sprintf(s, "%d", fmeter_getrate(&cs->fmeter));
1348 retval = simple_read_from_buffer(buf, nbytes, ppos, page, s - page);
1350 free_page((unsigned long)page);
1354 static ssize_t cpuset_file_read(struct file *file, char __user *buf, size_t nbytes,
1358 struct cftype *cft = __d_cft(file->f_dentry);
1362 /* special function ? */
1364 retval = cft->read(file, buf, nbytes, ppos);
1366 retval = cpuset_common_file_read(file, buf, nbytes, ppos);
1371 static int cpuset_file_open(struct inode *inode, struct file *file)
1376 err = generic_file_open(inode, file);
1380 cft = __d_cft(file->f_dentry);
1384 err = cft->open(inode, file);
1391 static int cpuset_file_release(struct inode *inode, struct file *file)
1393 struct cftype *cft = __d_cft(file->f_dentry);
1395 return cft->release(inode, file);
1400 * cpuset_rename - Only allow simple rename of directories in place.
1402 static int cpuset_rename(struct inode *old_dir, struct dentry *old_dentry,
1403 struct inode *new_dir, struct dentry *new_dentry)
1405 if (!S_ISDIR(old_dentry->d_inode->i_mode))
1407 if (new_dentry->d_inode)
1409 if (old_dir != new_dir)
1411 return simple_rename(old_dir, old_dentry, new_dir, new_dentry);
1414 static struct file_operations cpuset_file_operations = {
1415 .read = cpuset_file_read,
1416 .write = cpuset_file_write,
1417 .llseek = generic_file_llseek,
1418 .open = cpuset_file_open,
1419 .release = cpuset_file_release,
1422 static struct inode_operations cpuset_dir_inode_operations = {
1423 .lookup = simple_lookup,
1424 .mkdir = cpuset_mkdir,
1425 .rmdir = cpuset_rmdir,
1426 .rename = cpuset_rename,
1429 static int cpuset_create_file(struct dentry *dentry, int mode)
1431 struct inode *inode;
1435 if (dentry->d_inode)
1438 inode = cpuset_new_inode(mode);
1442 if (S_ISDIR(mode)) {
1443 inode->i_op = &cpuset_dir_inode_operations;
1444 inode->i_fop = &simple_dir_operations;
1446 /* start off with i_nlink == 2 (for "." entry) */
1448 } else if (S_ISREG(mode)) {
1450 inode->i_fop = &cpuset_file_operations;
1453 d_instantiate(dentry, inode);
1454 dget(dentry); /* Extra count - pin the dentry in core */
1459 * cpuset_create_dir - create a directory for an object.
1460 * cs: the cpuset we create the directory for.
1461 * It must have a valid ->parent field
1462 * And we are going to fill its ->dentry field.
1463 * name: The name to give to the cpuset directory. Will be copied.
1464 * mode: mode to set on new directory.
1467 static int cpuset_create_dir(struct cpuset *cs, const char *name, int mode)
1469 struct dentry *dentry = NULL;
1470 struct dentry *parent;
1473 parent = cs->parent->dentry;
1474 dentry = cpuset_get_dentry(parent, name);
1476 return PTR_ERR(dentry);
1477 error = cpuset_create_file(dentry, S_IFDIR | mode);
1479 dentry->d_fsdata = cs;
1480 parent->d_inode->i_nlink++;
1481 cs->dentry = dentry;
1488 static int cpuset_add_file(struct dentry *dir, const struct cftype *cft)
1490 struct dentry *dentry;
1493 down(&dir->d_inode->i_sem);
1494 dentry = cpuset_get_dentry(dir, cft->name);
1495 if (!IS_ERR(dentry)) {
1496 error = cpuset_create_file(dentry, 0644 | S_IFREG);
1498 dentry->d_fsdata = (void *)cft;
1501 error = PTR_ERR(dentry);
1502 up(&dir->d_inode->i_sem);
1507 * Stuff for reading the 'tasks' file.
1509 * Reading this file can return large amounts of data if a cpuset has
1510 * *lots* of attached tasks. So it may need several calls to read(),
1511 * but we cannot guarantee that the information we produce is correct
1512 * unless we produce it entirely atomically.
1514 * Upon tasks file open(), a struct ctr_struct is allocated, that
1515 * will have a pointer to an array (also allocated here). The struct
1516 * ctr_struct * is stored in file->private_data. Its resources will
1517 * be freed by release() when the file is closed. The array is used
1518 * to sprintf the PIDs and then used by read().
1521 /* cpusets_tasks_read array */
1529 * Load into 'pidarray' up to 'npids' of the tasks using cpuset 'cs'.
1530 * Return actual number of pids loaded. No need to task_lock(p)
1531 * when reading out p->cpuset, as we don't really care if it changes
1532 * on the next cycle, and we are not going to try to dereference it.
1534 static inline int pid_array_load(pid_t *pidarray, int npids, struct cpuset *cs)
1537 struct task_struct *g, *p;
1539 read_lock(&tasklist_lock);
1541 do_each_thread(g, p) {
1542 if (p->cpuset == cs) {
1543 pidarray[n++] = p->pid;
1544 if (unlikely(n == npids))
1547 } while_each_thread(g, p);
1550 read_unlock(&tasklist_lock);
1554 static int cmppid(const void *a, const void *b)
1556 return *(pid_t *)a - *(pid_t *)b;
1560 * Convert array 'a' of 'npids' pid_t's to a string of newline separated
1561 * decimal pids in 'buf'. Don't write more than 'sz' chars, but return
1562 * count 'cnt' of how many chars would be written if buf were large enough.
1564 static int pid_array_to_buf(char *buf, int sz, pid_t *a, int npids)
1569 for (i = 0; i < npids; i++)
1570 cnt += snprintf(buf + cnt, max(sz - cnt, 0), "%d\n", a[i]);
1575 * Handle an open on 'tasks' file. Prepare a buffer listing the
1576 * process id's of tasks currently attached to the cpuset being opened.
1578 * Does not require any specific cpuset semaphores, and does not take any.
1580 static int cpuset_tasks_open(struct inode *unused, struct file *file)
1582 struct cpuset *cs = __d_cs(file->f_dentry->d_parent);
1583 struct ctr_struct *ctr;
1588 if (!(file->f_mode & FMODE_READ))
1591 ctr = kmalloc(sizeof(*ctr), GFP_KERNEL);
1596 * If cpuset gets more users after we read count, we won't have
1597 * enough space - tough. This race is indistinguishable to the
1598 * caller from the case that the additional cpuset users didn't
1599 * show up until sometime later on.
1601 npids = atomic_read(&cs->count);
1602 pidarray = kmalloc(npids * sizeof(pid_t), GFP_KERNEL);
1606 npids = pid_array_load(pidarray, npids, cs);
1607 sort(pidarray, npids, sizeof(pid_t), cmppid, NULL);
1609 /* Call pid_array_to_buf() twice, first just to get bufsz */
1610 ctr->bufsz = pid_array_to_buf(&c, sizeof(c), pidarray, npids) + 1;
1611 ctr->buf = kmalloc(ctr->bufsz, GFP_KERNEL);
1614 ctr->bufsz = pid_array_to_buf(ctr->buf, ctr->bufsz, pidarray, npids);
1617 file->private_data = ctr;
1628 static ssize_t cpuset_tasks_read(struct file *file, char __user *buf,
1629 size_t nbytes, loff_t *ppos)
1631 struct ctr_struct *ctr = file->private_data;
1633 if (*ppos + nbytes > ctr->bufsz)
1634 nbytes = ctr->bufsz - *ppos;
1635 if (copy_to_user(buf, ctr->buf + *ppos, nbytes))
1641 static int cpuset_tasks_release(struct inode *unused_inode, struct file *file)
1643 struct ctr_struct *ctr;
1645 if (file->f_mode & FMODE_READ) {
1646 ctr = file->private_data;
1654 * for the common functions, 'private' gives the type of file
1657 static struct cftype cft_tasks = {
1659 .open = cpuset_tasks_open,
1660 .read = cpuset_tasks_read,
1661 .release = cpuset_tasks_release,
1662 .private = FILE_TASKLIST,
1665 static struct cftype cft_cpus = {
1667 .private = FILE_CPULIST,
1670 static struct cftype cft_mems = {
1672 .private = FILE_MEMLIST,
1675 static struct cftype cft_cpu_exclusive = {
1676 .name = "cpu_exclusive",
1677 .private = FILE_CPU_EXCLUSIVE,
1680 static struct cftype cft_mem_exclusive = {
1681 .name = "mem_exclusive",
1682 .private = FILE_MEM_EXCLUSIVE,
1685 static struct cftype cft_notify_on_release = {
1686 .name = "notify_on_release",
1687 .private = FILE_NOTIFY_ON_RELEASE,
1690 static struct cftype cft_memory_migrate = {
1691 .name = "memory_migrate",
1692 .private = FILE_MEMORY_MIGRATE,
1695 static struct cftype cft_memory_pressure_enabled = {
1696 .name = "memory_pressure_enabled",
1697 .private = FILE_MEMORY_PRESSURE_ENABLED,
1700 static struct cftype cft_memory_pressure = {
1701 .name = "memory_pressure",
1702 .private = FILE_MEMORY_PRESSURE,
1705 static int cpuset_populate_dir(struct dentry *cs_dentry)
1709 if ((err = cpuset_add_file(cs_dentry, &cft_cpus)) < 0)
1711 if ((err = cpuset_add_file(cs_dentry, &cft_mems)) < 0)
1713 if ((err = cpuset_add_file(cs_dentry, &cft_cpu_exclusive)) < 0)
1715 if ((err = cpuset_add_file(cs_dentry, &cft_mem_exclusive)) < 0)
1717 if ((err = cpuset_add_file(cs_dentry, &cft_notify_on_release)) < 0)
1719 if ((err = cpuset_add_file(cs_dentry, &cft_memory_migrate)) < 0)
1721 if ((err = cpuset_add_file(cs_dentry, &cft_memory_pressure)) < 0)
1723 if ((err = cpuset_add_file(cs_dentry, &cft_tasks)) < 0)
1729 * cpuset_create - create a cpuset
1730 * parent: cpuset that will be parent of the new cpuset.
1731 * name: name of the new cpuset. Will be strcpy'ed.
1732 * mode: mode to set on new inode
1734 * Must be called with the semaphore on the parent inode held
1737 static long cpuset_create(struct cpuset *parent, const char *name, int mode)
1742 cs = kmalloc(sizeof(*cs), GFP_KERNEL);
1747 cpuset_update_task_memory_state();
1749 if (notify_on_release(parent))
1750 set_bit(CS_NOTIFY_ON_RELEASE, &cs->flags);
1751 cs->cpus_allowed = CPU_MASK_NONE;
1752 cs->mems_allowed = NODE_MASK_NONE;
1753 atomic_set(&cs->count, 0);
1754 INIT_LIST_HEAD(&cs->sibling);
1755 INIT_LIST_HEAD(&cs->children);
1756 atomic_inc(&cpuset_mems_generation);
1757 cs->mems_generation = atomic_read(&cpuset_mems_generation);
1758 fmeter_init(&cs->fmeter);
1760 cs->parent = parent;
1762 down(&callback_sem);
1763 list_add(&cs->sibling, &cs->parent->children);
1764 number_of_cpusets++;
1767 err = cpuset_create_dir(cs, name, mode);
1772 * Release manage_sem before cpuset_populate_dir() because it
1773 * will down() this new directory's i_sem and if we race with
1774 * another mkdir, we might deadlock.
1778 err = cpuset_populate_dir(cs->dentry);
1779 /* If err < 0, we have a half-filled directory - oh well ;) */
1782 list_del(&cs->sibling);
1788 static int cpuset_mkdir(struct inode *dir, struct dentry *dentry, int mode)
1790 struct cpuset *c_parent = dentry->d_parent->d_fsdata;
1792 /* the vfs holds inode->i_sem already */
1793 return cpuset_create(c_parent, dentry->d_name.name, mode | S_IFDIR);
1796 static int cpuset_rmdir(struct inode *unused_dir, struct dentry *dentry)
1798 struct cpuset *cs = dentry->d_fsdata;
1800 struct cpuset *parent;
1801 char *pathbuf = NULL;
1803 /* the vfs holds both inode->i_sem already */
1806 cpuset_update_task_memory_state();
1807 if (atomic_read(&cs->count) > 0) {
1811 if (!list_empty(&cs->children)) {
1815 parent = cs->parent;
1816 down(&callback_sem);
1817 set_bit(CS_REMOVED, &cs->flags);
1818 if (is_cpu_exclusive(cs))
1819 update_cpu_domains(cs);
1820 list_del(&cs->sibling); /* delete my sibling from parent->children */
1821 spin_lock(&cs->dentry->d_lock);
1822 d = dget(cs->dentry);
1824 spin_unlock(&d->d_lock);
1825 cpuset_d_remove_dir(d);
1827 number_of_cpusets--;
1829 if (list_empty(&parent->children))
1830 check_for_release(parent, &pathbuf);
1832 cpuset_release_agent(pathbuf);
1837 * cpuset_init - initialize cpusets at system boot
1839 * Description: Initialize top_cpuset and the cpuset internal file system,
1842 int __init cpuset_init(void)
1844 struct dentry *root;
1847 top_cpuset.cpus_allowed = CPU_MASK_ALL;
1848 top_cpuset.mems_allowed = NODE_MASK_ALL;
1850 fmeter_init(&top_cpuset.fmeter);
1851 atomic_inc(&cpuset_mems_generation);
1852 top_cpuset.mems_generation = atomic_read(&cpuset_mems_generation);
1854 init_task.cpuset = &top_cpuset;
1856 err = register_filesystem(&cpuset_fs_type);
1859 cpuset_mount = kern_mount(&cpuset_fs_type);
1860 if (IS_ERR(cpuset_mount)) {
1861 printk(KERN_ERR "cpuset: could not mount!\n");
1862 err = PTR_ERR(cpuset_mount);
1863 cpuset_mount = NULL;
1866 root = cpuset_mount->mnt_sb->s_root;
1867 root->d_fsdata = &top_cpuset;
1868 root->d_inode->i_nlink++;
1869 top_cpuset.dentry = root;
1870 root->d_inode->i_op = &cpuset_dir_inode_operations;
1871 number_of_cpusets = 1;
1872 err = cpuset_populate_dir(root);
1873 /* memory_pressure_enabled is in root cpuset only */
1875 err = cpuset_add_file(root, &cft_memory_pressure_enabled);
1881 * cpuset_init_smp - initialize cpus_allowed
1883 * Description: Finish top cpuset after cpu, node maps are initialized
1886 void __init cpuset_init_smp(void)
1888 top_cpuset.cpus_allowed = cpu_online_map;
1889 top_cpuset.mems_allowed = node_online_map;
1893 * cpuset_fork - attach newly forked task to its parents cpuset.
1894 * @tsk: pointer to task_struct of forking parent process.
1896 * Description: A task inherits its parent's cpuset at fork().
1898 * A pointer to the shared cpuset was automatically copied in fork.c
1899 * by dup_task_struct(). However, we ignore that copy, since it was
1900 * not made under the protection of task_lock(), so might no longer be
1901 * a valid cpuset pointer. attach_task() might have already changed
1902 * current->cpuset, allowing the previously referenced cpuset to
1903 * be removed and freed. Instead, we task_lock(current) and copy
1904 * its present value of current->cpuset for our freshly forked child.
1906 * At the point that cpuset_fork() is called, 'current' is the parent
1907 * task, and the passed argument 'child' points to the child task.
1910 void cpuset_fork(struct task_struct *child)
1913 child->cpuset = current->cpuset;
1914 atomic_inc(&child->cpuset->count);
1915 task_unlock(current);
1919 * cpuset_exit - detach cpuset from exiting task
1920 * @tsk: pointer to task_struct of exiting process
1922 * Description: Detach cpuset from @tsk and release it.
1924 * Note that cpusets marked notify_on_release force every task in
1925 * them to take the global manage_sem semaphore when exiting.
1926 * This could impact scaling on very large systems. Be reluctant to
1927 * use notify_on_release cpusets where very high task exit scaling
1928 * is required on large systems.
1930 * Don't even think about derefencing 'cs' after the cpuset use count
1931 * goes to zero, except inside a critical section guarded by manage_sem
1932 * or callback_sem. Otherwise a zero cpuset use count is a license to
1933 * any other task to nuke the cpuset immediately, via cpuset_rmdir().
1935 * This routine has to take manage_sem, not callback_sem, because
1936 * it is holding that semaphore while calling check_for_release(),
1937 * which calls kmalloc(), so can't be called holding callback__sem().
1939 * We don't need to task_lock() this reference to tsk->cpuset,
1940 * because tsk is already marked PF_EXITING, so attach_task() won't
1941 * mess with it, or task is a failed fork, never visible to attach_task.
1944 void cpuset_exit(struct task_struct *tsk)
1951 if (notify_on_release(cs)) {
1952 char *pathbuf = NULL;
1955 if (atomic_dec_and_test(&cs->count))
1956 check_for_release(cs, &pathbuf);
1958 cpuset_release_agent(pathbuf);
1960 atomic_dec(&cs->count);
1965 * cpuset_cpus_allowed - return cpus_allowed mask from a tasks cpuset.
1966 * @tsk: pointer to task_struct from which to obtain cpuset->cpus_allowed.
1968 * Description: Returns the cpumask_t cpus_allowed of the cpuset
1969 * attached to the specified @tsk. Guaranteed to return some non-empty
1970 * subset of cpu_online_map, even if this means going outside the
1974 cpumask_t cpuset_cpus_allowed(struct task_struct *tsk)
1978 down(&callback_sem);
1980 guarantee_online_cpus(tsk->cpuset, &mask);
1987 void cpuset_init_current_mems_allowed(void)
1989 current->mems_allowed = NODE_MASK_ALL;
1993 * cpuset_mems_allowed - return mems_allowed mask from a tasks cpuset.
1994 * @tsk: pointer to task_struct from which to obtain cpuset->mems_allowed.
1996 * Description: Returns the nodemask_t mems_allowed of the cpuset
1997 * attached to the specified @tsk. Guaranteed to return some non-empty
1998 * subset of node_online_map, even if this means going outside the
2002 nodemask_t cpuset_mems_allowed(struct task_struct *tsk)
2006 down(&callback_sem);
2008 guarantee_online_mems(tsk->cpuset, &mask);
2016 * cpuset_zonelist_valid_mems_allowed - check zonelist vs. curremt mems_allowed
2017 * @zl: the zonelist to be checked
2019 * Are any of the nodes on zonelist zl allowed in current->mems_allowed?
2021 int cpuset_zonelist_valid_mems_allowed(struct zonelist *zl)
2025 for (i = 0; zl->zones[i]; i++) {
2026 int nid = zl->zones[i]->zone_pgdat->node_id;
2028 if (node_isset(nid, current->mems_allowed))
2035 * nearest_exclusive_ancestor() - Returns the nearest mem_exclusive
2036 * ancestor to the specified cpuset. Call holding callback_sem.
2037 * If no ancestor is mem_exclusive (an unusual configuration), then
2038 * returns the root cpuset.
2040 static const struct cpuset *nearest_exclusive_ancestor(const struct cpuset *cs)
2042 while (!is_mem_exclusive(cs) && cs->parent)
2048 * cpuset_zone_allowed - Can we allocate memory on zone z's memory node?
2049 * @z: is this zone on an allowed node?
2050 * @gfp_mask: memory allocation flags (we use __GFP_HARDWALL)
2052 * If we're in interrupt, yes, we can always allocate. If zone
2053 * z's node is in our tasks mems_allowed, yes. If it's not a
2054 * __GFP_HARDWALL request and this zone's nodes is in the nearest
2055 * mem_exclusive cpuset ancestor to this tasks cpuset, yes.
2058 * GFP_USER allocations are marked with the __GFP_HARDWALL bit,
2059 * and do not allow allocations outside the current tasks cpuset.
2060 * GFP_KERNEL allocations are not so marked, so can escape to the
2061 * nearest mem_exclusive ancestor cpuset.
2063 * Scanning up parent cpusets requires callback_sem. The __alloc_pages()
2064 * routine only calls here with __GFP_HARDWALL bit _not_ set if
2065 * it's a GFP_KERNEL allocation, and all nodes in the current tasks
2066 * mems_allowed came up empty on the first pass over the zonelist.
2067 * So only GFP_KERNEL allocations, if all nodes in the cpuset are
2068 * short of memory, might require taking the callback_sem semaphore.
2070 * The first loop over the zonelist in mm/page_alloc.c:__alloc_pages()
2071 * calls here with __GFP_HARDWALL always set in gfp_mask, enforcing
2072 * hardwall cpusets - no allocation on a node outside the cpuset is
2073 * allowed (unless in interrupt, of course).
2075 * The second loop doesn't even call here for GFP_ATOMIC requests
2076 * (if the __alloc_pages() local variable 'wait' is set). That check
2077 * and the checks below have the combined affect in the second loop of
2078 * the __alloc_pages() routine that:
2079 * in_interrupt - any node ok (current task context irrelevant)
2080 * GFP_ATOMIC - any node ok
2081 * GFP_KERNEL - any node in enclosing mem_exclusive cpuset ok
2082 * GFP_USER - only nodes in current tasks mems allowed ok.
2085 int __cpuset_zone_allowed(struct zone *z, gfp_t gfp_mask)
2087 int node; /* node that zone z is on */
2088 const struct cpuset *cs; /* current cpuset ancestors */
2089 int allowed = 1; /* is allocation in zone z allowed? */
2093 node = z->zone_pgdat->node_id;
2094 if (node_isset(node, current->mems_allowed))
2096 if (gfp_mask & __GFP_HARDWALL) /* If hardwall request, stop here */
2099 if (current->flags & PF_EXITING) /* Let dying task have memory */
2102 /* Not hardwall and node outside mems_allowed: scan up cpusets */
2103 down(&callback_sem);
2106 cs = nearest_exclusive_ancestor(current->cpuset);
2107 task_unlock(current);
2109 allowed = node_isset(node, cs->mems_allowed);
2115 * cpuset_excl_nodes_overlap - Do we overlap @p's mem_exclusive ancestors?
2116 * @p: pointer to task_struct of some other task.
2118 * Description: Return true if the nearest mem_exclusive ancestor
2119 * cpusets of tasks @p and current overlap. Used by oom killer to
2120 * determine if task @p's memory usage might impact the memory
2121 * available to the current task.
2123 * Acquires callback_sem - not suitable for calling from a fast path.
2126 int cpuset_excl_nodes_overlap(const struct task_struct *p)
2128 const struct cpuset *cs1, *cs2; /* my and p's cpuset ancestors */
2129 int overlap = 0; /* do cpusets overlap? */
2131 down(&callback_sem);
2134 if (current->flags & PF_EXITING) {
2135 task_unlock(current);
2138 cs1 = nearest_exclusive_ancestor(current->cpuset);
2139 task_unlock(current);
2141 task_lock((struct task_struct *)p);
2142 if (p->flags & PF_EXITING) {
2143 task_unlock((struct task_struct *)p);
2146 cs2 = nearest_exclusive_ancestor(p->cpuset);
2147 task_unlock((struct task_struct *)p);
2149 overlap = nodes_intersects(cs1->mems_allowed, cs2->mems_allowed);
2157 * Collection of memory_pressure is suppressed unless
2158 * this flag is enabled by writing "1" to the special
2159 * cpuset file 'memory_pressure_enabled' in the root cpuset.
2162 int cpuset_memory_pressure_enabled __read_mostly;
2165 * cpuset_memory_pressure_bump - keep stats of per-cpuset reclaims.
2167 * Keep a running average of the rate of synchronous (direct)
2168 * page reclaim efforts initiated by tasks in each cpuset.
2170 * This represents the rate at which some task in the cpuset
2171 * ran low on memory on all nodes it was allowed to use, and
2172 * had to enter the kernels page reclaim code in an effort to
2173 * create more free memory by tossing clean pages or swapping
2174 * or writing dirty pages.
2176 * Display to user space in the per-cpuset read-only file
2177 * "memory_pressure". Value displayed is an integer
2178 * representing the recent rate of entry into the synchronous
2179 * (direct) page reclaim by any task attached to the cpuset.
2182 void __cpuset_memory_pressure_bump(void)
2187 cs = current->cpuset;
2188 fmeter_markevent(&cs->fmeter);
2189 task_unlock(current);
2193 * proc_cpuset_show()
2194 * - Print tasks cpuset path into seq_file.
2195 * - Used for /proc/<pid>/cpuset.
2196 * - No need to task_lock(tsk) on this tsk->cpuset reference, as it
2197 * doesn't really matter if tsk->cpuset changes after we read it,
2198 * and we take manage_sem, keeping attach_task() from changing it
2202 static int proc_cpuset_show(struct seq_file *m, void *v)
2205 struct task_struct *tsk;
2209 buf = kmalloc(PAGE_SIZE, GFP_KERNEL);
2221 retval = cpuset_path(cs, buf, PAGE_SIZE);
2232 static int cpuset_open(struct inode *inode, struct file *file)
2234 struct task_struct *tsk = PROC_I(inode)->task;
2235 return single_open(file, proc_cpuset_show, tsk);
2238 struct file_operations proc_cpuset_operations = {
2239 .open = cpuset_open,
2241 .llseek = seq_lseek,
2242 .release = single_release,
2245 /* Display task cpus_allowed, mems_allowed in /proc/<pid>/status file. */
2246 char *cpuset_task_status_allowed(struct task_struct *task, char *buffer)
2248 buffer += sprintf(buffer, "Cpus_allowed:\t");
2249 buffer += cpumask_scnprintf(buffer, PAGE_SIZE, task->cpus_allowed);
2250 buffer += sprintf(buffer, "\n");
2251 buffer += sprintf(buffer, "Mems_allowed:\t");
2252 buffer += nodemask_scnprintf(buffer, PAGE_SIZE, task->mems_allowed);
2253 buffer += sprintf(buffer, "\n");