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/rcupdate.h>
43 #include <linux/sched.h>
44 #include <linux/seq_file.h>
45 #include <linux/slab.h>
46 #include <linux/smp_lock.h>
47 #include <linux/spinlock.h>
48 #include <linux/stat.h>
49 #include <linux/string.h>
50 #include <linux/time.h>
51 #include <linux/backing-dev.h>
52 #include <linux/sort.h>
54 #include <asm/uaccess.h>
55 #include <asm/atomic.h>
56 #include <asm/semaphore.h>
58 #define CPUSET_SUPER_MAGIC 0x27e0eb
61 * Tracks how many cpusets are currently defined in system.
62 * When there is only one cpuset (the root cpuset) we can
63 * short circuit some hooks.
65 int number_of_cpusets __read_mostly;
67 /* See "Frequency meter" comments, below. */
70 int cnt; /* unprocessed events count */
71 int val; /* most recent output value */
72 time_t time; /* clock (secs) when val computed */
73 spinlock_t lock; /* guards read or write of above */
77 unsigned long flags; /* "unsigned long" so bitops work */
78 cpumask_t cpus_allowed; /* CPUs allowed to tasks in cpuset */
79 nodemask_t mems_allowed; /* Memory Nodes allowed to tasks */
82 * Count is atomic so can incr (fork) or decr (exit) without a lock.
84 atomic_t count; /* count tasks using this cpuset */
87 * We link our 'sibling' struct into our parents 'children'.
88 * Our children link their 'sibling' into our 'children'.
90 struct list_head sibling; /* my parents children */
91 struct list_head children; /* my children */
93 struct cpuset *parent; /* my parent */
94 struct dentry *dentry; /* cpuset fs entry */
97 * Copy of global cpuset_mems_generation as of the most
98 * recent time this cpuset changed its mems_allowed.
102 struct fmeter fmeter; /* memory_pressure filter */
105 /* bits in struct cpuset flags field */
114 /* convenient tests for these bits */
115 static inline int is_cpu_exclusive(const struct cpuset *cs)
117 return !!test_bit(CS_CPU_EXCLUSIVE, &cs->flags);
120 static inline int is_mem_exclusive(const struct cpuset *cs)
122 return !!test_bit(CS_MEM_EXCLUSIVE, &cs->flags);
125 static inline int is_removed(const struct cpuset *cs)
127 return !!test_bit(CS_REMOVED, &cs->flags);
130 static inline int notify_on_release(const struct cpuset *cs)
132 return !!test_bit(CS_NOTIFY_ON_RELEASE, &cs->flags);
135 static inline int is_memory_migrate(const struct cpuset *cs)
137 return !!test_bit(CS_MEMORY_MIGRATE, &cs->flags);
141 * Increment this atomic integer everytime any cpuset changes its
142 * mems_allowed value. Users of cpusets can track this generation
143 * number, and avoid having to lock and reload mems_allowed unless
144 * the cpuset they're using changes generation.
146 * A single, global generation is needed because attach_task() could
147 * reattach a task to a different cpuset, which must not have its
148 * generation numbers aliased with those of that tasks previous cpuset.
150 * Generations are needed for mems_allowed because one task cannot
151 * modify anothers memory placement. So we must enable every task,
152 * on every visit to __alloc_pages(), to efficiently check whether
153 * its current->cpuset->mems_allowed has changed, requiring an update
154 * of its current->mems_allowed.
156 static atomic_t cpuset_mems_generation = ATOMIC_INIT(1);
158 static struct cpuset top_cpuset = {
159 .flags = ((1 << CS_CPU_EXCLUSIVE) | (1 << CS_MEM_EXCLUSIVE)),
160 .cpus_allowed = CPU_MASK_ALL,
161 .mems_allowed = NODE_MASK_ALL,
162 .count = ATOMIC_INIT(0),
163 .sibling = LIST_HEAD_INIT(top_cpuset.sibling),
164 .children = LIST_HEAD_INIT(top_cpuset.children),
167 static struct vfsmount *cpuset_mount;
168 static struct super_block *cpuset_sb;
171 * We have two global cpuset semaphores below. They can nest.
172 * It is ok to first take manage_sem, then nest callback_sem. We also
173 * require taking task_lock() when dereferencing a tasks cpuset pointer.
174 * See "The task_lock() exception", at the end of this comment.
176 * A task must hold both semaphores to modify cpusets. If a task
177 * holds manage_sem, then it blocks others wanting that semaphore,
178 * ensuring that it is the only task able to also acquire callback_sem
179 * and be able to modify cpusets. It can perform various checks on
180 * the cpuset structure first, knowing nothing will change. It can
181 * also allocate memory while just holding manage_sem. While it is
182 * performing these checks, various callback routines can briefly
183 * acquire callback_sem to query cpusets. Once it is ready to make
184 * the changes, it takes callback_sem, blocking everyone else.
186 * Calls to the kernel memory allocator can not be made while holding
187 * callback_sem, as that would risk double tripping on callback_sem
188 * from one of the callbacks into the cpuset code from within
191 * If a task is only holding callback_sem, then it has read-only
194 * The task_struct fields mems_allowed and mems_generation may only
195 * be accessed in the context of that task, so require no locks.
197 * Any task can increment and decrement the count field without lock.
198 * So in general, code holding manage_sem or callback_sem can't rely
199 * on the count field not changing. However, if the count goes to
200 * zero, then only attach_task(), which holds both semaphores, can
201 * increment it again. Because a count of zero means that no tasks
202 * are currently attached, therefore there is no way a task attached
203 * to that cpuset can fork (the other way to increment the count).
204 * So code holding manage_sem or callback_sem can safely assume that
205 * if the count is zero, it will stay zero. Similarly, if a task
206 * holds manage_sem or callback_sem on a cpuset with zero count, it
207 * knows that the cpuset won't be removed, as cpuset_rmdir() needs
208 * both of those semaphores.
210 * A possible optimization to improve parallelism would be to make
211 * callback_sem a R/W semaphore (rwsem), allowing the callback routines
212 * to proceed in parallel, with read access, until the holder of
213 * manage_sem needed to take this rwsem for exclusive write access
214 * and modify some cpusets.
216 * The cpuset_common_file_write handler for operations that modify
217 * the cpuset hierarchy holds manage_sem across the entire operation,
218 * single threading all such cpuset modifications across the system.
220 * The cpuset_common_file_read() handlers only hold callback_sem across
221 * small pieces of code, such as when reading out possibly multi-word
222 * cpumasks and nodemasks.
224 * The fork and exit callbacks cpuset_fork() and cpuset_exit(), don't
225 * (usually) take either semaphore. These are the two most performance
226 * critical pieces of code here. The exception occurs on cpuset_exit(),
227 * when a task in a notify_on_release cpuset exits. Then manage_sem
228 * is taken, and if the cpuset count is zero, a usermode call made
229 * to /sbin/cpuset_release_agent with the name of the cpuset (path
230 * relative to the root of cpuset file system) as the argument.
232 * A cpuset can only be deleted if both its 'count' of using tasks
233 * is zero, and its list of 'children' cpusets is empty. Since all
234 * tasks in the system use _some_ cpuset, and since there is always at
235 * least one task in the system (init, pid == 1), therefore, top_cpuset
236 * always has either children cpusets and/or using tasks. So we don't
237 * need a special hack to ensure that top_cpuset cannot be deleted.
239 * The above "Tale of Two Semaphores" would be complete, but for:
241 * The task_lock() exception
243 * The need for this exception arises from the action of attach_task(),
244 * which overwrites one tasks cpuset pointer with another. It does
245 * so using both semaphores, however there are several performance
246 * critical places that need to reference task->cpuset without the
247 * expense of grabbing a system global semaphore. Therefore except as
248 * noted below, when dereferencing or, as in attach_task(), modifying
249 * a tasks cpuset pointer we use task_lock(), which acts on a spinlock
250 * (task->alloc_lock) already in the task_struct routinely used for
253 * P.S. One more locking exception. RCU is used to guard the
254 * update of a tasks cpuset pointer by attach_task() and the
255 * access of task->cpuset->mems_generation via that pointer in
256 * the routine cpuset_update_task_memory_state().
259 static DECLARE_MUTEX(manage_sem);
260 static DECLARE_MUTEX(callback_sem);
263 * A couple of forward declarations required, due to cyclic reference loop:
264 * cpuset_mkdir -> cpuset_create -> cpuset_populate_dir -> cpuset_add_file
265 * -> cpuset_create_file -> cpuset_dir_inode_operations -> cpuset_mkdir.
268 static int cpuset_mkdir(struct inode *dir, struct dentry *dentry, int mode);
269 static int cpuset_rmdir(struct inode *unused_dir, struct dentry *dentry);
271 static struct backing_dev_info cpuset_backing_dev_info = {
272 .ra_pages = 0, /* No readahead */
273 .capabilities = BDI_CAP_NO_ACCT_DIRTY | BDI_CAP_NO_WRITEBACK,
276 static struct inode *cpuset_new_inode(mode_t mode)
278 struct inode *inode = new_inode(cpuset_sb);
281 inode->i_mode = mode;
282 inode->i_uid = current->fsuid;
283 inode->i_gid = current->fsgid;
284 inode->i_blksize = PAGE_CACHE_SIZE;
286 inode->i_atime = inode->i_mtime = inode->i_ctime = CURRENT_TIME;
287 inode->i_mapping->backing_dev_info = &cpuset_backing_dev_info;
292 static void cpuset_diput(struct dentry *dentry, struct inode *inode)
294 /* is dentry a directory ? if so, kfree() associated cpuset */
295 if (S_ISDIR(inode->i_mode)) {
296 struct cpuset *cs = dentry->d_fsdata;
297 BUG_ON(!(is_removed(cs)));
303 static struct dentry_operations cpuset_dops = {
304 .d_iput = cpuset_diput,
307 static struct dentry *cpuset_get_dentry(struct dentry *parent, const char *name)
309 struct dentry *d = lookup_one_len(name, parent, strlen(name));
311 d->d_op = &cpuset_dops;
315 static void remove_dir(struct dentry *d)
317 struct dentry *parent = dget(d->d_parent);
320 simple_rmdir(parent->d_inode, d);
325 * NOTE : the dentry must have been dget()'ed
327 static void cpuset_d_remove_dir(struct dentry *dentry)
329 struct list_head *node;
331 spin_lock(&dcache_lock);
332 node = dentry->d_subdirs.next;
333 while (node != &dentry->d_subdirs) {
334 struct dentry *d = list_entry(node, struct dentry, d_u.d_child);
338 spin_unlock(&dcache_lock);
340 simple_unlink(dentry->d_inode, d);
342 spin_lock(&dcache_lock);
344 node = dentry->d_subdirs.next;
346 list_del_init(&dentry->d_u.d_child);
347 spin_unlock(&dcache_lock);
351 static struct super_operations cpuset_ops = {
352 .statfs = simple_statfs,
353 .drop_inode = generic_delete_inode,
356 static int cpuset_fill_super(struct super_block *sb, void *unused_data,
362 sb->s_blocksize = PAGE_CACHE_SIZE;
363 sb->s_blocksize_bits = PAGE_CACHE_SHIFT;
364 sb->s_magic = CPUSET_SUPER_MAGIC;
365 sb->s_op = &cpuset_ops;
368 inode = cpuset_new_inode(S_IFDIR | S_IRUGO | S_IXUGO | S_IWUSR);
370 inode->i_op = &simple_dir_inode_operations;
371 inode->i_fop = &simple_dir_operations;
372 /* directories start off with i_nlink == 2 (for "." entry) */
378 root = d_alloc_root(inode);
387 static struct super_block *cpuset_get_sb(struct file_system_type *fs_type,
388 int flags, const char *unused_dev_name,
391 return get_sb_single(fs_type, flags, data, cpuset_fill_super);
394 static struct file_system_type cpuset_fs_type = {
396 .get_sb = cpuset_get_sb,
397 .kill_sb = kill_litter_super,
402 * The files in the cpuset filesystem mostly have a very simple read/write
403 * handling, some common function will take care of it. Nevertheless some cases
404 * (read tasks) are special and therefore I define this structure for every
408 * When reading/writing to a file:
409 * - the cpuset to use in file->f_dentry->d_parent->d_fsdata
410 * - the 'cftype' of the file is file->f_dentry->d_fsdata
416 int (*open) (struct inode *inode, struct file *file);
417 ssize_t (*read) (struct file *file, char __user *buf, size_t nbytes,
419 int (*write) (struct file *file, const char __user *buf, size_t nbytes,
421 int (*release) (struct inode *inode, struct file *file);
424 static inline struct cpuset *__d_cs(struct dentry *dentry)
426 return dentry->d_fsdata;
429 static inline struct cftype *__d_cft(struct dentry *dentry)
431 return dentry->d_fsdata;
435 * Call with manage_sem held. Writes path of cpuset into buf.
436 * Returns 0 on success, -errno on error.
439 static int cpuset_path(const struct cpuset *cs, char *buf, int buflen)
443 start = buf + buflen;
447 int len = cs->dentry->d_name.len;
448 if ((start -= len) < buf)
449 return -ENAMETOOLONG;
450 memcpy(start, cs->dentry->d_name.name, len);
457 return -ENAMETOOLONG;
460 memmove(buf, start, buf + buflen - start);
465 * Notify userspace when a cpuset is released, by running
466 * /sbin/cpuset_release_agent with the name of the cpuset (path
467 * relative to the root of cpuset file system) as the argument.
469 * Most likely, this user command will try to rmdir this cpuset.
471 * This races with the possibility that some other task will be
472 * attached to this cpuset before it is removed, or that some other
473 * user task will 'mkdir' a child cpuset of this cpuset. That's ok.
474 * The presumed 'rmdir' will fail quietly if this cpuset is no longer
475 * unused, and this cpuset will be reprieved from its death sentence,
476 * to continue to serve a useful existence. Next time it's released,
477 * we will get notified again, if it still has 'notify_on_release' set.
479 * The final arg to call_usermodehelper() is 0, which means don't
480 * wait. The separate /sbin/cpuset_release_agent task is forked by
481 * call_usermodehelper(), then control in this thread returns here,
482 * without waiting for the release agent task. We don't bother to
483 * wait because the caller of this routine has no use for the exit
484 * status of the /sbin/cpuset_release_agent task, so no sense holding
485 * our caller up for that.
487 * When we had only one cpuset semaphore, we had to call this
488 * without holding it, to avoid deadlock when call_usermodehelper()
489 * allocated memory. With two locks, we could now call this while
490 * holding manage_sem, but we still don't, so as to minimize
491 * the time manage_sem is held.
494 static void cpuset_release_agent(const char *pathbuf)
496 char *argv[3], *envp[3];
503 argv[i++] = "/sbin/cpuset_release_agent";
504 argv[i++] = (char *)pathbuf;
508 /* minimal command environment */
509 envp[i++] = "HOME=/";
510 envp[i++] = "PATH=/sbin:/bin:/usr/sbin:/usr/bin";
513 call_usermodehelper(argv[0], argv, envp, 0);
518 * Either cs->count of using tasks transitioned to zero, or the
519 * cs->children list of child cpusets just became empty. If this
520 * cs is notify_on_release() and now both the user count is zero and
521 * the list of children is empty, prepare cpuset path in a kmalloc'd
522 * buffer, to be returned via ppathbuf, so that the caller can invoke
523 * cpuset_release_agent() with it later on, once manage_sem is dropped.
524 * Call here with manage_sem held.
526 * This check_for_release() routine is responsible for kmalloc'ing
527 * pathbuf. The above cpuset_release_agent() is responsible for
528 * kfree'ing pathbuf. The caller of these routines is responsible
529 * for providing a pathbuf pointer, initialized to NULL, then
530 * calling check_for_release() with manage_sem held and the address
531 * of the pathbuf pointer, then dropping manage_sem, then calling
532 * cpuset_release_agent() with pathbuf, as set by check_for_release().
535 static void check_for_release(struct cpuset *cs, char **ppathbuf)
537 if (notify_on_release(cs) && atomic_read(&cs->count) == 0 &&
538 list_empty(&cs->children)) {
541 buf = kmalloc(PAGE_SIZE, GFP_KERNEL);
544 if (cpuset_path(cs, buf, PAGE_SIZE) < 0)
552 * Return in *pmask the portion of a cpusets's cpus_allowed that
553 * are online. If none are online, walk up the cpuset hierarchy
554 * until we find one that does have some online cpus. If we get
555 * all the way to the top and still haven't found any online cpus,
556 * return cpu_online_map. Or if passed a NULL cs from an exit'ing
557 * task, return cpu_online_map.
559 * One way or another, we guarantee to return some non-empty subset
562 * Call with callback_sem held.
565 static void guarantee_online_cpus(const struct cpuset *cs, cpumask_t *pmask)
567 while (cs && !cpus_intersects(cs->cpus_allowed, cpu_online_map))
570 cpus_and(*pmask, cs->cpus_allowed, cpu_online_map);
572 *pmask = cpu_online_map;
573 BUG_ON(!cpus_intersects(*pmask, cpu_online_map));
577 * Return in *pmask the portion of a cpusets's mems_allowed that
578 * are online. If none are online, walk up the cpuset hierarchy
579 * until we find one that does have some online mems. If we get
580 * all the way to the top and still haven't found any online mems,
581 * return node_online_map.
583 * One way or another, we guarantee to return some non-empty subset
584 * of node_online_map.
586 * Call with callback_sem held.
589 static void guarantee_online_mems(const struct cpuset *cs, nodemask_t *pmask)
591 while (cs && !nodes_intersects(cs->mems_allowed, node_online_map))
594 nodes_and(*pmask, cs->mems_allowed, node_online_map);
596 *pmask = node_online_map;
597 BUG_ON(!nodes_intersects(*pmask, node_online_map));
601 * cpuset_update_task_memory_state - update task memory placement
603 * If the current tasks cpusets mems_allowed changed behind our
604 * backs, update current->mems_allowed, mems_generation and task NUMA
605 * mempolicy to the new value.
607 * Task mempolicy is updated by rebinding it relative to the
608 * current->cpuset if a task has its memory placement changed.
609 * Do not call this routine if in_interrupt().
611 * Call without callback_sem or task_lock() held. May be called
612 * with or without manage_sem held. Doesn't need task_lock to guard
613 * against another task changing a non-NULL cpuset pointer to NULL,
614 * as that is only done by a task on itself, and if the current task
615 * is here, it is not simultaneously in the exit code NULL'ing its
616 * cpuset pointer. This routine also might acquire callback_sem and
617 * current->mm->mmap_sem during call.
619 * Reading current->cpuset->mems_generation doesn't need task_lock
620 * to guard the current->cpuset derefence, because it is guarded
621 * from concurrent freeing of current->cpuset by attach_task(),
624 * The rcu_dereference() is technically probably not needed,
625 * as I don't actually mind if I see a new cpuset pointer but
626 * an old value of mems_generation. However this really only
627 * matters on alpha systems using cpusets heavily. If I dropped
628 * that rcu_dereference(), it would save them a memory barrier.
629 * For all other arch's, rcu_dereference is a no-op anyway, and for
630 * alpha systems not using cpusets, another planned optimization,
631 * avoiding the rcu critical section for tasks in the root cpuset
632 * which is statically allocated, so can't vanish, will make this
633 * irrelevant. Better to use RCU as intended, than to engage in
634 * some cute trick to save a memory barrier that is impossible to
635 * test, for alpha systems using cpusets heavily, which might not
638 * This routine is needed to update the per-task mems_allowed data,
639 * within the tasks context, when it is trying to allocate memory
640 * (in various mm/mempolicy.c routines) and notices that some other
641 * task has been modifying its cpuset.
644 void cpuset_update_task_memory_state()
646 int my_cpusets_mem_gen;
647 struct task_struct *tsk = current;
650 if (tsk->cpuset == &top_cpuset) {
651 /* Don't need rcu for top_cpuset. It's never freed. */
652 my_cpusets_mem_gen = top_cpuset.mems_generation;
655 cs = rcu_dereference(tsk->cpuset);
656 my_cpusets_mem_gen = cs->mems_generation;
660 if (my_cpusets_mem_gen != tsk->cpuset_mems_generation) {
663 cs = tsk->cpuset; /* Maybe changed when task not locked */
664 guarantee_online_mems(cs, &tsk->mems_allowed);
665 tsk->cpuset_mems_generation = cs->mems_generation;
668 mpol_rebind_task(tsk, &tsk->mems_allowed);
673 * is_cpuset_subset(p, q) - Is cpuset p a subset of cpuset q?
675 * One cpuset is a subset of another if all its allowed CPUs and
676 * Memory Nodes are a subset of the other, and its exclusive flags
677 * are only set if the other's are set. Call holding manage_sem.
680 static int is_cpuset_subset(const struct cpuset *p, const struct cpuset *q)
682 return cpus_subset(p->cpus_allowed, q->cpus_allowed) &&
683 nodes_subset(p->mems_allowed, q->mems_allowed) &&
684 is_cpu_exclusive(p) <= is_cpu_exclusive(q) &&
685 is_mem_exclusive(p) <= is_mem_exclusive(q);
689 * validate_change() - Used to validate that any proposed cpuset change
690 * follows the structural rules for cpusets.
692 * If we replaced the flag and mask values of the current cpuset
693 * (cur) with those values in the trial cpuset (trial), would
694 * our various subset and exclusive rules still be valid? Presumes
697 * 'cur' is the address of an actual, in-use cpuset. Operations
698 * such as list traversal that depend on the actual address of the
699 * cpuset in the list must use cur below, not trial.
701 * 'trial' is the address of bulk structure copy of cur, with
702 * perhaps one or more of the fields cpus_allowed, mems_allowed,
703 * or flags changed to new, trial values.
705 * Return 0 if valid, -errno if not.
708 static int validate_change(const struct cpuset *cur, const struct cpuset *trial)
710 struct cpuset *c, *par;
712 /* Each of our child cpusets must be a subset of us */
713 list_for_each_entry(c, &cur->children, sibling) {
714 if (!is_cpuset_subset(c, trial))
718 /* Remaining checks don't apply to root cpuset */
719 if ((par = cur->parent) == NULL)
722 /* We must be a subset of our parent cpuset */
723 if (!is_cpuset_subset(trial, par))
726 /* If either I or some sibling (!= me) is exclusive, we can't overlap */
727 list_for_each_entry(c, &par->children, sibling) {
728 if ((is_cpu_exclusive(trial) || is_cpu_exclusive(c)) &&
730 cpus_intersects(trial->cpus_allowed, c->cpus_allowed))
732 if ((is_mem_exclusive(trial) || is_mem_exclusive(c)) &&
734 nodes_intersects(trial->mems_allowed, c->mems_allowed))
742 * For a given cpuset cur, partition the system as follows
743 * a. All cpus in the parent cpuset's cpus_allowed that are not part of any
744 * exclusive child cpusets
745 * b. All cpus in the current cpuset's cpus_allowed that are not part of any
746 * exclusive child cpusets
747 * Build these two partitions by calling partition_sched_domains
749 * Call with manage_sem held. May nest a call to the
750 * lock_cpu_hotplug()/unlock_cpu_hotplug() pair.
753 static void update_cpu_domains(struct cpuset *cur)
755 struct cpuset *c, *par = cur->parent;
756 cpumask_t pspan, cspan;
758 if (par == NULL || cpus_empty(cur->cpus_allowed))
762 * Get all cpus from parent's cpus_allowed not part of exclusive
765 pspan = par->cpus_allowed;
766 list_for_each_entry(c, &par->children, sibling) {
767 if (is_cpu_exclusive(c))
768 cpus_andnot(pspan, pspan, c->cpus_allowed);
770 if (is_removed(cur) || !is_cpu_exclusive(cur)) {
771 cpus_or(pspan, pspan, cur->cpus_allowed);
772 if (cpus_equal(pspan, cur->cpus_allowed))
774 cspan = CPU_MASK_NONE;
776 if (cpus_empty(pspan))
778 cspan = cur->cpus_allowed;
780 * Get all cpus from current cpuset's cpus_allowed not part
781 * of exclusive children
783 list_for_each_entry(c, &cur->children, sibling) {
784 if (is_cpu_exclusive(c))
785 cpus_andnot(cspan, cspan, c->cpus_allowed);
790 partition_sched_domains(&pspan, &cspan);
791 unlock_cpu_hotplug();
795 * Call with manage_sem held. May take callback_sem during call.
798 static int update_cpumask(struct cpuset *cs, char *buf)
800 struct cpuset trialcs;
801 int retval, cpus_unchanged;
804 retval = cpulist_parse(buf, trialcs.cpus_allowed);
807 cpus_and(trialcs.cpus_allowed, trialcs.cpus_allowed, cpu_online_map);
808 if (cpus_empty(trialcs.cpus_allowed))
810 retval = validate_change(cs, &trialcs);
813 cpus_unchanged = cpus_equal(cs->cpus_allowed, trialcs.cpus_allowed);
815 cs->cpus_allowed = trialcs.cpus_allowed;
817 if (is_cpu_exclusive(cs) && !cpus_unchanged)
818 update_cpu_domains(cs);
823 * Handle user request to change the 'mems' memory placement
824 * of a cpuset. Needs to validate the request, update the
825 * cpusets mems_allowed and mems_generation, and for each
826 * task in the cpuset, rebind any vma mempolicies and if
827 * the cpuset is marked 'memory_migrate', migrate the tasks
828 * pages to the new memory.
830 * Call with manage_sem held. May take callback_sem during call.
831 * Will take tasklist_lock, scan tasklist for tasks in cpuset cs,
832 * lock each such tasks mm->mmap_sem, scan its vma's and rebind
833 * their mempolicies to the cpusets new mems_allowed.
836 static int update_nodemask(struct cpuset *cs, char *buf)
838 struct cpuset trialcs;
840 struct task_struct *g, *p;
841 struct mm_struct **mmarray;
848 retval = nodelist_parse(buf, trialcs.mems_allowed);
851 nodes_and(trialcs.mems_allowed, trialcs.mems_allowed, node_online_map);
852 oldmem = cs->mems_allowed;
853 if (nodes_equal(oldmem, trialcs.mems_allowed)) {
854 retval = 0; /* Too easy - nothing to do */
857 if (nodes_empty(trialcs.mems_allowed)) {
861 retval = validate_change(cs, &trialcs);
866 cs->mems_allowed = trialcs.mems_allowed;
867 atomic_inc(&cpuset_mems_generation);
868 cs->mems_generation = atomic_read(&cpuset_mems_generation);
871 set_cpuset_being_rebound(cs); /* causes mpol_copy() rebind */
873 fudge = 10; /* spare mmarray[] slots */
874 fudge += cpus_weight(cs->cpus_allowed); /* imagine one fork-bomb/cpu */
878 * Allocate mmarray[] to hold mm reference for each task
879 * in cpuset cs. Can't kmalloc GFP_KERNEL while holding
880 * tasklist_lock. We could use GFP_ATOMIC, but with a
881 * few more lines of code, we can retry until we get a big
882 * enough mmarray[] w/o using GFP_ATOMIC.
885 ntasks = atomic_read(&cs->count); /* guess */
887 mmarray = kmalloc(ntasks * sizeof(*mmarray), GFP_KERNEL);
890 write_lock_irq(&tasklist_lock); /* block fork */
891 if (atomic_read(&cs->count) <= ntasks)
892 break; /* got enough */
893 write_unlock_irq(&tasklist_lock); /* try again */
899 /* Load up mmarray[] with mm reference for each task in cpuset. */
900 do_each_thread(g, p) {
901 struct mm_struct *mm;
905 "Cpuset mempolicy rebind incomplete.\n");
914 } while_each_thread(g, p);
915 write_unlock_irq(&tasklist_lock);
918 * Now that we've dropped the tasklist spinlock, we can
919 * rebind the vma mempolicies of each mm in mmarray[] to their
920 * new cpuset, and release that mm. The mpol_rebind_mm()
921 * call takes mmap_sem, which we couldn't take while holding
922 * tasklist_lock. Forks can happen again now - the mpol_copy()
923 * cpuset_being_rebound check will catch such forks, and rebind
924 * their vma mempolicies too. Because we still hold the global
925 * cpuset manage_sem, we know that no other rebind effort will
926 * be contending for the global variable cpuset_being_rebound.
927 * It's ok if we rebind the same mm twice; mpol_rebind_mm()
928 * is idempotent. Also migrate pages in each mm to new nodes.
930 migrate = is_memory_migrate(cs);
931 for (i = 0; i < n; i++) {
932 struct mm_struct *mm = mmarray[i];
934 mpol_rebind_mm(mm, &cs->mems_allowed);
936 do_migrate_pages(mm, &oldmem, &cs->mems_allowed,
942 /* We're done rebinding vma's to this cpusets new mems_allowed. */
944 set_cpuset_being_rebound(NULL);
951 * Call with manage_sem held.
954 static int update_memory_pressure_enabled(struct cpuset *cs, char *buf)
956 if (simple_strtoul(buf, NULL, 10) != 0)
957 cpuset_memory_pressure_enabled = 1;
959 cpuset_memory_pressure_enabled = 0;
964 * update_flag - read a 0 or a 1 in a file and update associated flag
965 * bit: the bit to update (CS_CPU_EXCLUSIVE, CS_MEM_EXCLUSIVE,
966 * CS_NOTIFY_ON_RELEASE, CS_MEMORY_MIGRATE)
967 * cs: the cpuset to update
968 * buf: the buffer where we read the 0 or 1
970 * Call with manage_sem held.
973 static int update_flag(cpuset_flagbits_t bit, struct cpuset *cs, char *buf)
976 struct cpuset trialcs;
977 int err, cpu_exclusive_changed;
979 turning_on = (simple_strtoul(buf, NULL, 10) != 0);
983 set_bit(bit, &trialcs.flags);
985 clear_bit(bit, &trialcs.flags);
987 err = validate_change(cs, &trialcs);
990 cpu_exclusive_changed =
991 (is_cpu_exclusive(cs) != is_cpu_exclusive(&trialcs));
994 set_bit(bit, &cs->flags);
996 clear_bit(bit, &cs->flags);
999 if (cpu_exclusive_changed)
1000 update_cpu_domains(cs);
1005 * Frequency meter - How fast is some event occuring?
1007 * These routines manage a digitally filtered, constant time based,
1008 * event frequency meter. There are four routines:
1009 * fmeter_init() - initialize a frequency meter.
1010 * fmeter_markevent() - called each time the event happens.
1011 * fmeter_getrate() - returns the recent rate of such events.
1012 * fmeter_update() - internal routine used to update fmeter.
1014 * A common data structure is passed to each of these routines,
1015 * which is used to keep track of the state required to manage the
1016 * frequency meter and its digital filter.
1018 * The filter works on the number of events marked per unit time.
1019 * The filter is single-pole low-pass recursive (IIR). The time unit
1020 * is 1 second. Arithmetic is done using 32-bit integers scaled to
1021 * simulate 3 decimal digits of precision (multiplied by 1000).
1023 * With an FM_COEF of 933, and a time base of 1 second, the filter
1024 * has a half-life of 10 seconds, meaning that if the events quit
1025 * happening, then the rate returned from the fmeter_getrate()
1026 * will be cut in half each 10 seconds, until it converges to zero.
1028 * It is not worth doing a real infinitely recursive filter. If more
1029 * than FM_MAXTICKS ticks have elapsed since the last filter event,
1030 * just compute FM_MAXTICKS ticks worth, by which point the level
1033 * Limit the count of unprocessed events to FM_MAXCNT, so as to avoid
1034 * arithmetic overflow in the fmeter_update() routine.
1036 * Given the simple 32 bit integer arithmetic used, this meter works
1037 * best for reporting rates between one per millisecond (msec) and
1038 * one per 32 (approx) seconds. At constant rates faster than one
1039 * per msec it maxes out at values just under 1,000,000. At constant
1040 * rates between one per msec, and one per second it will stabilize
1041 * to a value N*1000, where N is the rate of events per second.
1042 * At constant rates between one per second and one per 32 seconds,
1043 * it will be choppy, moving up on the seconds that have an event,
1044 * and then decaying until the next event. At rates slower than
1045 * about one in 32 seconds, it decays all the way back to zero between
1049 #define FM_COEF 933 /* coefficient for half-life of 10 secs */
1050 #define FM_MAXTICKS ((time_t)99) /* useless computing more ticks than this */
1051 #define FM_MAXCNT 1000000 /* limit cnt to avoid overflow */
1052 #define FM_SCALE 1000 /* faux fixed point scale */
1054 /* Initialize a frequency meter */
1055 static void fmeter_init(struct fmeter *fmp)
1060 spin_lock_init(&fmp->lock);
1063 /* Internal meter update - process cnt events and update value */
1064 static void fmeter_update(struct fmeter *fmp)
1066 time_t now = get_seconds();
1067 time_t ticks = now - fmp->time;
1072 ticks = min(FM_MAXTICKS, ticks);
1074 fmp->val = (FM_COEF * fmp->val) / FM_SCALE;
1077 fmp->val += ((FM_SCALE - FM_COEF) * fmp->cnt) / FM_SCALE;
1081 /* Process any previous ticks, then bump cnt by one (times scale). */
1082 static void fmeter_markevent(struct fmeter *fmp)
1084 spin_lock(&fmp->lock);
1086 fmp->cnt = min(FM_MAXCNT, fmp->cnt + FM_SCALE);
1087 spin_unlock(&fmp->lock);
1090 /* Process any previous ticks, then return current value. */
1091 static int fmeter_getrate(struct fmeter *fmp)
1095 spin_lock(&fmp->lock);
1098 spin_unlock(&fmp->lock);
1103 * Attack task specified by pid in 'pidbuf' to cpuset 'cs', possibly
1104 * writing the path of the old cpuset in 'ppathbuf' if it needs to be
1105 * notified on release.
1107 * Call holding manage_sem. May take callback_sem and task_lock of
1108 * the task 'pid' during call.
1111 static int attach_task(struct cpuset *cs, char *pidbuf, char **ppathbuf)
1114 struct task_struct *tsk;
1115 struct cpuset *oldcs;
1117 nodemask_t from, to;
1118 struct mm_struct *mm;
1120 if (sscanf(pidbuf, "%d", &pid) != 1)
1122 if (cpus_empty(cs->cpus_allowed) || nodes_empty(cs->mems_allowed))
1126 read_lock(&tasklist_lock);
1128 tsk = find_task_by_pid(pid);
1129 if (!tsk || tsk->flags & PF_EXITING) {
1130 read_unlock(&tasklist_lock);
1134 get_task_struct(tsk);
1135 read_unlock(&tasklist_lock);
1137 if ((current->euid) && (current->euid != tsk->uid)
1138 && (current->euid != tsk->suid)) {
1139 put_task_struct(tsk);
1144 get_task_struct(tsk);
1147 down(&callback_sem);
1150 oldcs = tsk->cpuset;
1154 put_task_struct(tsk);
1157 atomic_inc(&cs->count);
1158 rcu_assign_pointer(tsk->cpuset, cs);
1161 guarantee_online_cpus(cs, &cpus);
1162 set_cpus_allowed(tsk, cpus);
1164 from = oldcs->mems_allowed;
1165 to = cs->mems_allowed;
1169 mm = get_task_mm(tsk);
1171 mpol_rebind_mm(mm, &to);
1175 if (is_memory_migrate(cs))
1176 do_migrate_pages(tsk->mm, &from, &to, MPOL_MF_MOVE_ALL);
1177 put_task_struct(tsk);
1179 if (atomic_dec_and_test(&oldcs->count))
1180 check_for_release(oldcs, ppathbuf);
1184 /* The various types of files and directories in a cpuset file system */
1189 FILE_MEMORY_MIGRATE,
1194 FILE_NOTIFY_ON_RELEASE,
1195 FILE_MEMORY_PRESSURE_ENABLED,
1196 FILE_MEMORY_PRESSURE,
1198 } cpuset_filetype_t;
1200 static ssize_t cpuset_common_file_write(struct file *file, const char __user *userbuf,
1201 size_t nbytes, loff_t *unused_ppos)
1203 struct cpuset *cs = __d_cs(file->f_dentry->d_parent);
1204 struct cftype *cft = __d_cft(file->f_dentry);
1205 cpuset_filetype_t type = cft->private;
1207 char *pathbuf = NULL;
1210 /* Crude upper limit on largest legitimate cpulist user might write. */
1211 if (nbytes > 100 + 6 * NR_CPUS)
1214 /* +1 for nul-terminator */
1215 if ((buffer = kmalloc(nbytes + 1, GFP_KERNEL)) == 0)
1218 if (copy_from_user(buffer, userbuf, nbytes)) {
1222 buffer[nbytes] = 0; /* nul-terminate */
1226 if (is_removed(cs)) {
1233 retval = update_cpumask(cs, buffer);
1236 retval = update_nodemask(cs, buffer);
1238 case FILE_CPU_EXCLUSIVE:
1239 retval = update_flag(CS_CPU_EXCLUSIVE, cs, buffer);
1241 case FILE_MEM_EXCLUSIVE:
1242 retval = update_flag(CS_MEM_EXCLUSIVE, cs, buffer);
1244 case FILE_NOTIFY_ON_RELEASE:
1245 retval = update_flag(CS_NOTIFY_ON_RELEASE, cs, buffer);
1247 case FILE_MEMORY_MIGRATE:
1248 retval = update_flag(CS_MEMORY_MIGRATE, cs, buffer);
1250 case FILE_MEMORY_PRESSURE_ENABLED:
1251 retval = update_memory_pressure_enabled(cs, buffer);
1253 case FILE_MEMORY_PRESSURE:
1257 retval = attach_task(cs, buffer, &pathbuf);
1268 cpuset_release_agent(pathbuf);
1274 static ssize_t cpuset_file_write(struct file *file, const char __user *buf,
1275 size_t nbytes, loff_t *ppos)
1278 struct cftype *cft = __d_cft(file->f_dentry);
1282 /* special function ? */
1284 retval = cft->write(file, buf, nbytes, ppos);
1286 retval = cpuset_common_file_write(file, buf, nbytes, ppos);
1292 * These ascii lists should be read in a single call, by using a user
1293 * buffer large enough to hold the entire map. If read in smaller
1294 * chunks, there is no guarantee of atomicity. Since the display format
1295 * used, list of ranges of sequential numbers, is variable length,
1296 * and since these maps can change value dynamically, one could read
1297 * gibberish by doing partial reads while a list was changing.
1298 * A single large read to a buffer that crosses a page boundary is
1299 * ok, because the result being copied to user land is not recomputed
1300 * across a page fault.
1303 static int cpuset_sprintf_cpulist(char *page, struct cpuset *cs)
1307 down(&callback_sem);
1308 mask = cs->cpus_allowed;
1311 return cpulist_scnprintf(page, PAGE_SIZE, mask);
1314 static int cpuset_sprintf_memlist(char *page, struct cpuset *cs)
1318 down(&callback_sem);
1319 mask = cs->mems_allowed;
1322 return nodelist_scnprintf(page, PAGE_SIZE, mask);
1325 static ssize_t cpuset_common_file_read(struct file *file, char __user *buf,
1326 size_t nbytes, loff_t *ppos)
1328 struct cftype *cft = __d_cft(file->f_dentry);
1329 struct cpuset *cs = __d_cs(file->f_dentry->d_parent);
1330 cpuset_filetype_t type = cft->private;
1335 if (!(page = (char *)__get_free_page(GFP_KERNEL)))
1342 s += cpuset_sprintf_cpulist(s, cs);
1345 s += cpuset_sprintf_memlist(s, cs);
1347 case FILE_CPU_EXCLUSIVE:
1348 *s++ = is_cpu_exclusive(cs) ? '1' : '0';
1350 case FILE_MEM_EXCLUSIVE:
1351 *s++ = is_mem_exclusive(cs) ? '1' : '0';
1353 case FILE_NOTIFY_ON_RELEASE:
1354 *s++ = notify_on_release(cs) ? '1' : '0';
1356 case FILE_MEMORY_MIGRATE:
1357 *s++ = is_memory_migrate(cs) ? '1' : '0';
1359 case FILE_MEMORY_PRESSURE_ENABLED:
1360 *s++ = cpuset_memory_pressure_enabled ? '1' : '0';
1362 case FILE_MEMORY_PRESSURE:
1363 s += sprintf(s, "%d", fmeter_getrate(&cs->fmeter));
1371 retval = simple_read_from_buffer(buf, nbytes, ppos, page, s - page);
1373 free_page((unsigned long)page);
1377 static ssize_t cpuset_file_read(struct file *file, char __user *buf, size_t nbytes,
1381 struct cftype *cft = __d_cft(file->f_dentry);
1385 /* special function ? */
1387 retval = cft->read(file, buf, nbytes, ppos);
1389 retval = cpuset_common_file_read(file, buf, nbytes, ppos);
1394 static int cpuset_file_open(struct inode *inode, struct file *file)
1399 err = generic_file_open(inode, file);
1403 cft = __d_cft(file->f_dentry);
1407 err = cft->open(inode, file);
1414 static int cpuset_file_release(struct inode *inode, struct file *file)
1416 struct cftype *cft = __d_cft(file->f_dentry);
1418 return cft->release(inode, file);
1423 * cpuset_rename - Only allow simple rename of directories in place.
1425 static int cpuset_rename(struct inode *old_dir, struct dentry *old_dentry,
1426 struct inode *new_dir, struct dentry *new_dentry)
1428 if (!S_ISDIR(old_dentry->d_inode->i_mode))
1430 if (new_dentry->d_inode)
1432 if (old_dir != new_dir)
1434 return simple_rename(old_dir, old_dentry, new_dir, new_dentry);
1437 static struct file_operations cpuset_file_operations = {
1438 .read = cpuset_file_read,
1439 .write = cpuset_file_write,
1440 .llseek = generic_file_llseek,
1441 .open = cpuset_file_open,
1442 .release = cpuset_file_release,
1445 static struct inode_operations cpuset_dir_inode_operations = {
1446 .lookup = simple_lookup,
1447 .mkdir = cpuset_mkdir,
1448 .rmdir = cpuset_rmdir,
1449 .rename = cpuset_rename,
1452 static int cpuset_create_file(struct dentry *dentry, int mode)
1454 struct inode *inode;
1458 if (dentry->d_inode)
1461 inode = cpuset_new_inode(mode);
1465 if (S_ISDIR(mode)) {
1466 inode->i_op = &cpuset_dir_inode_operations;
1467 inode->i_fop = &simple_dir_operations;
1469 /* start off with i_nlink == 2 (for "." entry) */
1471 } else if (S_ISREG(mode)) {
1473 inode->i_fop = &cpuset_file_operations;
1476 d_instantiate(dentry, inode);
1477 dget(dentry); /* Extra count - pin the dentry in core */
1482 * cpuset_create_dir - create a directory for an object.
1483 * cs: the cpuset we create the directory for.
1484 * It must have a valid ->parent field
1485 * And we are going to fill its ->dentry field.
1486 * name: The name to give to the cpuset directory. Will be copied.
1487 * mode: mode to set on new directory.
1490 static int cpuset_create_dir(struct cpuset *cs, const char *name, int mode)
1492 struct dentry *dentry = NULL;
1493 struct dentry *parent;
1496 parent = cs->parent->dentry;
1497 dentry = cpuset_get_dentry(parent, name);
1499 return PTR_ERR(dentry);
1500 error = cpuset_create_file(dentry, S_IFDIR | mode);
1502 dentry->d_fsdata = cs;
1503 parent->d_inode->i_nlink++;
1504 cs->dentry = dentry;
1511 static int cpuset_add_file(struct dentry *dir, const struct cftype *cft)
1513 struct dentry *dentry;
1516 mutex_lock(&dir->d_inode->i_mutex);
1517 dentry = cpuset_get_dentry(dir, cft->name);
1518 if (!IS_ERR(dentry)) {
1519 error = cpuset_create_file(dentry, 0644 | S_IFREG);
1521 dentry->d_fsdata = (void *)cft;
1524 error = PTR_ERR(dentry);
1525 mutex_unlock(&dir->d_inode->i_mutex);
1530 * Stuff for reading the 'tasks' file.
1532 * Reading this file can return large amounts of data if a cpuset has
1533 * *lots* of attached tasks. So it may need several calls to read(),
1534 * but we cannot guarantee that the information we produce is correct
1535 * unless we produce it entirely atomically.
1537 * Upon tasks file open(), a struct ctr_struct is allocated, that
1538 * will have a pointer to an array (also allocated here). The struct
1539 * ctr_struct * is stored in file->private_data. Its resources will
1540 * be freed by release() when the file is closed. The array is used
1541 * to sprintf the PIDs and then used by read().
1544 /* cpusets_tasks_read array */
1552 * Load into 'pidarray' up to 'npids' of the tasks using cpuset 'cs'.
1553 * Return actual number of pids loaded. No need to task_lock(p)
1554 * when reading out p->cpuset, as we don't really care if it changes
1555 * on the next cycle, and we are not going to try to dereference it.
1557 static inline int pid_array_load(pid_t *pidarray, int npids, struct cpuset *cs)
1560 struct task_struct *g, *p;
1562 read_lock(&tasklist_lock);
1564 do_each_thread(g, p) {
1565 if (p->cpuset == cs) {
1566 pidarray[n++] = p->pid;
1567 if (unlikely(n == npids))
1570 } while_each_thread(g, p);
1573 read_unlock(&tasklist_lock);
1577 static int cmppid(const void *a, const void *b)
1579 return *(pid_t *)a - *(pid_t *)b;
1583 * Convert array 'a' of 'npids' pid_t's to a string of newline separated
1584 * decimal pids in 'buf'. Don't write more than 'sz' chars, but return
1585 * count 'cnt' of how many chars would be written if buf were large enough.
1587 static int pid_array_to_buf(char *buf, int sz, pid_t *a, int npids)
1592 for (i = 0; i < npids; i++)
1593 cnt += snprintf(buf + cnt, max(sz - cnt, 0), "%d\n", a[i]);
1598 * Handle an open on 'tasks' file. Prepare a buffer listing the
1599 * process id's of tasks currently attached to the cpuset being opened.
1601 * Does not require any specific cpuset semaphores, and does not take any.
1603 static int cpuset_tasks_open(struct inode *unused, struct file *file)
1605 struct cpuset *cs = __d_cs(file->f_dentry->d_parent);
1606 struct ctr_struct *ctr;
1611 if (!(file->f_mode & FMODE_READ))
1614 ctr = kmalloc(sizeof(*ctr), GFP_KERNEL);
1619 * If cpuset gets more users after we read count, we won't have
1620 * enough space - tough. This race is indistinguishable to the
1621 * caller from the case that the additional cpuset users didn't
1622 * show up until sometime later on.
1624 npids = atomic_read(&cs->count);
1625 pidarray = kmalloc(npids * sizeof(pid_t), GFP_KERNEL);
1629 npids = pid_array_load(pidarray, npids, cs);
1630 sort(pidarray, npids, sizeof(pid_t), cmppid, NULL);
1632 /* Call pid_array_to_buf() twice, first just to get bufsz */
1633 ctr->bufsz = pid_array_to_buf(&c, sizeof(c), pidarray, npids) + 1;
1634 ctr->buf = kmalloc(ctr->bufsz, GFP_KERNEL);
1637 ctr->bufsz = pid_array_to_buf(ctr->buf, ctr->bufsz, pidarray, npids);
1640 file->private_data = ctr;
1651 static ssize_t cpuset_tasks_read(struct file *file, char __user *buf,
1652 size_t nbytes, loff_t *ppos)
1654 struct ctr_struct *ctr = file->private_data;
1656 if (*ppos + nbytes > ctr->bufsz)
1657 nbytes = ctr->bufsz - *ppos;
1658 if (copy_to_user(buf, ctr->buf + *ppos, nbytes))
1664 static int cpuset_tasks_release(struct inode *unused_inode, struct file *file)
1666 struct ctr_struct *ctr;
1668 if (file->f_mode & FMODE_READ) {
1669 ctr = file->private_data;
1677 * for the common functions, 'private' gives the type of file
1680 static struct cftype cft_tasks = {
1682 .open = cpuset_tasks_open,
1683 .read = cpuset_tasks_read,
1684 .release = cpuset_tasks_release,
1685 .private = FILE_TASKLIST,
1688 static struct cftype cft_cpus = {
1690 .private = FILE_CPULIST,
1693 static struct cftype cft_mems = {
1695 .private = FILE_MEMLIST,
1698 static struct cftype cft_cpu_exclusive = {
1699 .name = "cpu_exclusive",
1700 .private = FILE_CPU_EXCLUSIVE,
1703 static struct cftype cft_mem_exclusive = {
1704 .name = "mem_exclusive",
1705 .private = FILE_MEM_EXCLUSIVE,
1708 static struct cftype cft_notify_on_release = {
1709 .name = "notify_on_release",
1710 .private = FILE_NOTIFY_ON_RELEASE,
1713 static struct cftype cft_memory_migrate = {
1714 .name = "memory_migrate",
1715 .private = FILE_MEMORY_MIGRATE,
1718 static struct cftype cft_memory_pressure_enabled = {
1719 .name = "memory_pressure_enabled",
1720 .private = FILE_MEMORY_PRESSURE_ENABLED,
1723 static struct cftype cft_memory_pressure = {
1724 .name = "memory_pressure",
1725 .private = FILE_MEMORY_PRESSURE,
1728 static int cpuset_populate_dir(struct dentry *cs_dentry)
1732 if ((err = cpuset_add_file(cs_dentry, &cft_cpus)) < 0)
1734 if ((err = cpuset_add_file(cs_dentry, &cft_mems)) < 0)
1736 if ((err = cpuset_add_file(cs_dentry, &cft_cpu_exclusive)) < 0)
1738 if ((err = cpuset_add_file(cs_dentry, &cft_mem_exclusive)) < 0)
1740 if ((err = cpuset_add_file(cs_dentry, &cft_notify_on_release)) < 0)
1742 if ((err = cpuset_add_file(cs_dentry, &cft_memory_migrate)) < 0)
1744 if ((err = cpuset_add_file(cs_dentry, &cft_memory_pressure)) < 0)
1746 if ((err = cpuset_add_file(cs_dentry, &cft_tasks)) < 0)
1752 * cpuset_create - create a cpuset
1753 * parent: cpuset that will be parent of the new cpuset.
1754 * name: name of the new cpuset. Will be strcpy'ed.
1755 * mode: mode to set on new inode
1757 * Must be called with the semaphore on the parent inode held
1760 static long cpuset_create(struct cpuset *parent, const char *name, int mode)
1765 cs = kmalloc(sizeof(*cs), GFP_KERNEL);
1770 cpuset_update_task_memory_state();
1772 if (notify_on_release(parent))
1773 set_bit(CS_NOTIFY_ON_RELEASE, &cs->flags);
1774 cs->cpus_allowed = CPU_MASK_NONE;
1775 cs->mems_allowed = NODE_MASK_NONE;
1776 atomic_set(&cs->count, 0);
1777 INIT_LIST_HEAD(&cs->sibling);
1778 INIT_LIST_HEAD(&cs->children);
1779 atomic_inc(&cpuset_mems_generation);
1780 cs->mems_generation = atomic_read(&cpuset_mems_generation);
1781 fmeter_init(&cs->fmeter);
1783 cs->parent = parent;
1785 down(&callback_sem);
1786 list_add(&cs->sibling, &cs->parent->children);
1787 number_of_cpusets++;
1790 err = cpuset_create_dir(cs, name, mode);
1795 * Release manage_sem before cpuset_populate_dir() because it
1796 * will down() this new directory's i_mutex and if we race with
1797 * another mkdir, we might deadlock.
1801 err = cpuset_populate_dir(cs->dentry);
1802 /* If err < 0, we have a half-filled directory - oh well ;) */
1805 list_del(&cs->sibling);
1811 static int cpuset_mkdir(struct inode *dir, struct dentry *dentry, int mode)
1813 struct cpuset *c_parent = dentry->d_parent->d_fsdata;
1815 /* the vfs holds inode->i_mutex already */
1816 return cpuset_create(c_parent, dentry->d_name.name, mode | S_IFDIR);
1819 static int cpuset_rmdir(struct inode *unused_dir, struct dentry *dentry)
1821 struct cpuset *cs = dentry->d_fsdata;
1823 struct cpuset *parent;
1824 char *pathbuf = NULL;
1826 /* the vfs holds both inode->i_mutex already */
1829 cpuset_update_task_memory_state();
1830 if (atomic_read(&cs->count) > 0) {
1834 if (!list_empty(&cs->children)) {
1838 parent = cs->parent;
1839 down(&callback_sem);
1840 set_bit(CS_REMOVED, &cs->flags);
1841 if (is_cpu_exclusive(cs))
1842 update_cpu_domains(cs);
1843 list_del(&cs->sibling); /* delete my sibling from parent->children */
1844 spin_lock(&cs->dentry->d_lock);
1845 d = dget(cs->dentry);
1847 spin_unlock(&d->d_lock);
1848 cpuset_d_remove_dir(d);
1850 number_of_cpusets--;
1852 if (list_empty(&parent->children))
1853 check_for_release(parent, &pathbuf);
1855 cpuset_release_agent(pathbuf);
1860 * cpuset_init_early - just enough so that the calls to
1861 * cpuset_update_task_memory_state() in early init code
1865 int __init cpuset_init_early(void)
1867 struct task_struct *tsk = current;
1869 tsk->cpuset = &top_cpuset;
1870 tsk->cpuset->mems_generation = atomic_read(&cpuset_mems_generation);
1875 * cpuset_init - initialize cpusets at system boot
1877 * Description: Initialize top_cpuset and the cpuset internal file system,
1880 int __init cpuset_init(void)
1882 struct dentry *root;
1885 top_cpuset.cpus_allowed = CPU_MASK_ALL;
1886 top_cpuset.mems_allowed = NODE_MASK_ALL;
1888 fmeter_init(&top_cpuset.fmeter);
1889 atomic_inc(&cpuset_mems_generation);
1890 top_cpuset.mems_generation = atomic_read(&cpuset_mems_generation);
1892 init_task.cpuset = &top_cpuset;
1894 err = register_filesystem(&cpuset_fs_type);
1897 cpuset_mount = kern_mount(&cpuset_fs_type);
1898 if (IS_ERR(cpuset_mount)) {
1899 printk(KERN_ERR "cpuset: could not mount!\n");
1900 err = PTR_ERR(cpuset_mount);
1901 cpuset_mount = NULL;
1904 root = cpuset_mount->mnt_sb->s_root;
1905 root->d_fsdata = &top_cpuset;
1906 root->d_inode->i_nlink++;
1907 top_cpuset.dentry = root;
1908 root->d_inode->i_op = &cpuset_dir_inode_operations;
1909 number_of_cpusets = 1;
1910 err = cpuset_populate_dir(root);
1911 /* memory_pressure_enabled is in root cpuset only */
1913 err = cpuset_add_file(root, &cft_memory_pressure_enabled);
1919 * cpuset_init_smp - initialize cpus_allowed
1921 * Description: Finish top cpuset after cpu, node maps are initialized
1924 void __init cpuset_init_smp(void)
1926 top_cpuset.cpus_allowed = cpu_online_map;
1927 top_cpuset.mems_allowed = node_online_map;
1931 * cpuset_fork - attach newly forked task to its parents cpuset.
1932 * @tsk: pointer to task_struct of forking parent process.
1934 * Description: A task inherits its parent's cpuset at fork().
1936 * A pointer to the shared cpuset was automatically copied in fork.c
1937 * by dup_task_struct(). However, we ignore that copy, since it was
1938 * not made under the protection of task_lock(), so might no longer be
1939 * a valid cpuset pointer. attach_task() might have already changed
1940 * current->cpuset, allowing the previously referenced cpuset to
1941 * be removed and freed. Instead, we task_lock(current) and copy
1942 * its present value of current->cpuset for our freshly forked child.
1944 * At the point that cpuset_fork() is called, 'current' is the parent
1945 * task, and the passed argument 'child' points to the child task.
1948 void cpuset_fork(struct task_struct *child)
1951 child->cpuset = current->cpuset;
1952 atomic_inc(&child->cpuset->count);
1953 task_unlock(current);
1957 * cpuset_exit - detach cpuset from exiting task
1958 * @tsk: pointer to task_struct of exiting process
1960 * Description: Detach cpuset from @tsk and release it.
1962 * Note that cpusets marked notify_on_release force every task in
1963 * them to take the global manage_sem semaphore when exiting.
1964 * This could impact scaling on very large systems. Be reluctant to
1965 * use notify_on_release cpusets where very high task exit scaling
1966 * is required on large systems.
1968 * Don't even think about derefencing 'cs' after the cpuset use count
1969 * goes to zero, except inside a critical section guarded by manage_sem
1970 * or callback_sem. Otherwise a zero cpuset use count is a license to
1971 * any other task to nuke the cpuset immediately, via cpuset_rmdir().
1973 * This routine has to take manage_sem, not callback_sem, because
1974 * it is holding that semaphore while calling check_for_release(),
1975 * which calls kmalloc(), so can't be called holding callback__sem().
1977 * We don't need to task_lock() this reference to tsk->cpuset,
1978 * because tsk is already marked PF_EXITING, so attach_task() won't
1979 * mess with it, or task is a failed fork, never visible to attach_task.
1982 void cpuset_exit(struct task_struct *tsk)
1989 if (notify_on_release(cs)) {
1990 char *pathbuf = NULL;
1993 if (atomic_dec_and_test(&cs->count))
1994 check_for_release(cs, &pathbuf);
1996 cpuset_release_agent(pathbuf);
1998 atomic_dec(&cs->count);
2003 * cpuset_cpus_allowed - return cpus_allowed mask from a tasks cpuset.
2004 * @tsk: pointer to task_struct from which to obtain cpuset->cpus_allowed.
2006 * Description: Returns the cpumask_t cpus_allowed of the cpuset
2007 * attached to the specified @tsk. Guaranteed to return some non-empty
2008 * subset of cpu_online_map, even if this means going outside the
2012 cpumask_t cpuset_cpus_allowed(struct task_struct *tsk)
2016 down(&callback_sem);
2018 guarantee_online_cpus(tsk->cpuset, &mask);
2025 void cpuset_init_current_mems_allowed(void)
2027 current->mems_allowed = NODE_MASK_ALL;
2031 * cpuset_mems_allowed - return mems_allowed mask from a tasks cpuset.
2032 * @tsk: pointer to task_struct from which to obtain cpuset->mems_allowed.
2034 * Description: Returns the nodemask_t mems_allowed of the cpuset
2035 * attached to the specified @tsk. Guaranteed to return some non-empty
2036 * subset of node_online_map, even if this means going outside the
2040 nodemask_t cpuset_mems_allowed(struct task_struct *tsk)
2044 down(&callback_sem);
2046 guarantee_online_mems(tsk->cpuset, &mask);
2054 * cpuset_zonelist_valid_mems_allowed - check zonelist vs. curremt mems_allowed
2055 * @zl: the zonelist to be checked
2057 * Are any of the nodes on zonelist zl allowed in current->mems_allowed?
2059 int cpuset_zonelist_valid_mems_allowed(struct zonelist *zl)
2063 for (i = 0; zl->zones[i]; i++) {
2064 int nid = zl->zones[i]->zone_pgdat->node_id;
2066 if (node_isset(nid, current->mems_allowed))
2073 * nearest_exclusive_ancestor() - Returns the nearest mem_exclusive
2074 * ancestor to the specified cpuset. Call holding callback_sem.
2075 * If no ancestor is mem_exclusive (an unusual configuration), then
2076 * returns the root cpuset.
2078 static const struct cpuset *nearest_exclusive_ancestor(const struct cpuset *cs)
2080 while (!is_mem_exclusive(cs) && cs->parent)
2086 * cpuset_zone_allowed - Can we allocate memory on zone z's memory node?
2087 * @z: is this zone on an allowed node?
2088 * @gfp_mask: memory allocation flags (we use __GFP_HARDWALL)
2090 * If we're in interrupt, yes, we can always allocate. If zone
2091 * z's node is in our tasks mems_allowed, yes. If it's not a
2092 * __GFP_HARDWALL request and this zone's nodes is in the nearest
2093 * mem_exclusive cpuset ancestor to this tasks cpuset, yes.
2096 * GFP_USER allocations are marked with the __GFP_HARDWALL bit,
2097 * and do not allow allocations outside the current tasks cpuset.
2098 * GFP_KERNEL allocations are not so marked, so can escape to the
2099 * nearest mem_exclusive ancestor cpuset.
2101 * Scanning up parent cpusets requires callback_sem. The __alloc_pages()
2102 * routine only calls here with __GFP_HARDWALL bit _not_ set if
2103 * it's a GFP_KERNEL allocation, and all nodes in the current tasks
2104 * mems_allowed came up empty on the first pass over the zonelist.
2105 * So only GFP_KERNEL allocations, if all nodes in the cpuset are
2106 * short of memory, might require taking the callback_sem semaphore.
2108 * The first loop over the zonelist in mm/page_alloc.c:__alloc_pages()
2109 * calls here with __GFP_HARDWALL always set in gfp_mask, enforcing
2110 * hardwall cpusets - no allocation on a node outside the cpuset is
2111 * allowed (unless in interrupt, of course).
2113 * The second loop doesn't even call here for GFP_ATOMIC requests
2114 * (if the __alloc_pages() local variable 'wait' is set). That check
2115 * and the checks below have the combined affect in the second loop of
2116 * the __alloc_pages() routine that:
2117 * in_interrupt - any node ok (current task context irrelevant)
2118 * GFP_ATOMIC - any node ok
2119 * GFP_KERNEL - any node in enclosing mem_exclusive cpuset ok
2120 * GFP_USER - only nodes in current tasks mems allowed ok.
2123 int __cpuset_zone_allowed(struct zone *z, gfp_t gfp_mask)
2125 int node; /* node that zone z is on */
2126 const struct cpuset *cs; /* current cpuset ancestors */
2127 int allowed = 1; /* is allocation in zone z allowed? */
2131 node = z->zone_pgdat->node_id;
2132 if (node_isset(node, current->mems_allowed))
2134 if (gfp_mask & __GFP_HARDWALL) /* If hardwall request, stop here */
2137 if (current->flags & PF_EXITING) /* Let dying task have memory */
2140 /* Not hardwall and node outside mems_allowed: scan up cpusets */
2141 down(&callback_sem);
2144 cs = nearest_exclusive_ancestor(current->cpuset);
2145 task_unlock(current);
2147 allowed = node_isset(node, cs->mems_allowed);
2153 * cpuset_excl_nodes_overlap - Do we overlap @p's mem_exclusive ancestors?
2154 * @p: pointer to task_struct of some other task.
2156 * Description: Return true if the nearest mem_exclusive ancestor
2157 * cpusets of tasks @p and current overlap. Used by oom killer to
2158 * determine if task @p's memory usage might impact the memory
2159 * available to the current task.
2161 * Acquires callback_sem - not suitable for calling from a fast path.
2164 int cpuset_excl_nodes_overlap(const struct task_struct *p)
2166 const struct cpuset *cs1, *cs2; /* my and p's cpuset ancestors */
2167 int overlap = 0; /* do cpusets overlap? */
2169 down(&callback_sem);
2172 if (current->flags & PF_EXITING) {
2173 task_unlock(current);
2176 cs1 = nearest_exclusive_ancestor(current->cpuset);
2177 task_unlock(current);
2179 task_lock((struct task_struct *)p);
2180 if (p->flags & PF_EXITING) {
2181 task_unlock((struct task_struct *)p);
2184 cs2 = nearest_exclusive_ancestor(p->cpuset);
2185 task_unlock((struct task_struct *)p);
2187 overlap = nodes_intersects(cs1->mems_allowed, cs2->mems_allowed);
2195 * Collection of memory_pressure is suppressed unless
2196 * this flag is enabled by writing "1" to the special
2197 * cpuset file 'memory_pressure_enabled' in the root cpuset.
2200 int cpuset_memory_pressure_enabled __read_mostly;
2203 * cpuset_memory_pressure_bump - keep stats of per-cpuset reclaims.
2205 * Keep a running average of the rate of synchronous (direct)
2206 * page reclaim efforts initiated by tasks in each cpuset.
2208 * This represents the rate at which some task in the cpuset
2209 * ran low on memory on all nodes it was allowed to use, and
2210 * had to enter the kernels page reclaim code in an effort to
2211 * create more free memory by tossing clean pages or swapping
2212 * or writing dirty pages.
2214 * Display to user space in the per-cpuset read-only file
2215 * "memory_pressure". Value displayed is an integer
2216 * representing the recent rate of entry into the synchronous
2217 * (direct) page reclaim by any task attached to the cpuset.
2220 void __cpuset_memory_pressure_bump(void)
2225 cs = current->cpuset;
2226 fmeter_markevent(&cs->fmeter);
2227 task_unlock(current);
2231 * proc_cpuset_show()
2232 * - Print tasks cpuset path into seq_file.
2233 * - Used for /proc/<pid>/cpuset.
2234 * - No need to task_lock(tsk) on this tsk->cpuset reference, as it
2235 * doesn't really matter if tsk->cpuset changes after we read it,
2236 * and we take manage_sem, keeping attach_task() from changing it
2240 static int proc_cpuset_show(struct seq_file *m, void *v)
2243 struct task_struct *tsk;
2247 buf = kmalloc(PAGE_SIZE, GFP_KERNEL);
2259 retval = cpuset_path(cs, buf, PAGE_SIZE);
2270 static int cpuset_open(struct inode *inode, struct file *file)
2272 struct task_struct *tsk = PROC_I(inode)->task;
2273 return single_open(file, proc_cpuset_show, tsk);
2276 struct file_operations proc_cpuset_operations = {
2277 .open = cpuset_open,
2279 .llseek = seq_lseek,
2280 .release = single_release,
2283 /* Display task cpus_allowed, mems_allowed in /proc/<pid>/status file. */
2284 char *cpuset_task_status_allowed(struct task_struct *task, char *buffer)
2286 buffer += sprintf(buffer, "Cpus_allowed:\t");
2287 buffer += cpumask_scnprintf(buffer, PAGE_SIZE, task->cpus_allowed);
2288 buffer += sprintf(buffer, "\n");
2289 buffer += sprintf(buffer, "Mems_allowed:\t");
2290 buffer += nodemask_scnprintf(buffer, PAGE_SIZE, task->mems_allowed);
2291 buffer += sprintf(buffer, "\n");