Commit | Line | Data |
---|---|---|
1da177e4 LT |
1 | /* |
2 | * kernel/cpuset.c | |
3 | * | |
4 | * Processor and Memory placement constraints for sets of tasks. | |
5 | * | |
6 | * Copyright (C) 2003 BULL SA. | |
825a46af | 7 | * Copyright (C) 2004-2006 Silicon Graphics, Inc. |
1da177e4 LT |
8 | * |
9 | * Portions derived from Patrick Mochel's sysfs code. | |
10 | * sysfs is Copyright (c) 2001-3 Patrick Mochel | |
1da177e4 | 11 | * |
825a46af | 12 | * 2003-10-10 Written by Simon Derr. |
1da177e4 | 13 | * 2003-10-22 Updates by Stephen Hemminger. |
825a46af | 14 | * 2004 May-July Rework by Paul Jackson. |
1da177e4 LT |
15 | * |
16 | * This file is subject to the terms and conditions of the GNU General Public | |
17 | * License. See the file COPYING in the main directory of the Linux | |
18 | * distribution for more details. | |
19 | */ | |
20 | ||
1da177e4 LT |
21 | #include <linux/cpu.h> |
22 | #include <linux/cpumask.h> | |
23 | #include <linux/cpuset.h> | |
24 | #include <linux/err.h> | |
25 | #include <linux/errno.h> | |
26 | #include <linux/file.h> | |
27 | #include <linux/fs.h> | |
28 | #include <linux/init.h> | |
29 | #include <linux/interrupt.h> | |
30 | #include <linux/kernel.h> | |
31 | #include <linux/kmod.h> | |
32 | #include <linux/list.h> | |
68860ec1 | 33 | #include <linux/mempolicy.h> |
1da177e4 LT |
34 | #include <linux/mm.h> |
35 | #include <linux/module.h> | |
36 | #include <linux/mount.h> | |
37 | #include <linux/namei.h> | |
38 | #include <linux/pagemap.h> | |
39 | #include <linux/proc_fs.h> | |
6b9c2603 | 40 | #include <linux/rcupdate.h> |
1da177e4 LT |
41 | #include <linux/sched.h> |
42 | #include <linux/seq_file.h> | |
22fb52dd | 43 | #include <linux/security.h> |
1da177e4 | 44 | #include <linux/slab.h> |
1da177e4 LT |
45 | #include <linux/spinlock.h> |
46 | #include <linux/stat.h> | |
47 | #include <linux/string.h> | |
48 | #include <linux/time.h> | |
49 | #include <linux/backing-dev.h> | |
50 | #include <linux/sort.h> | |
51 | ||
52 | #include <asm/uaccess.h> | |
53 | #include <asm/atomic.h> | |
3d3f26a7 | 54 | #include <linux/mutex.h> |
1da177e4 | 55 | |
c5b2aff8 | 56 | #define CPUSET_SUPER_MAGIC 0x27e0eb |
1da177e4 | 57 | |
202f72d5 PJ |
58 | /* |
59 | * Tracks how many cpusets are currently defined in system. | |
60 | * When there is only one cpuset (the root cpuset) we can | |
61 | * short circuit some hooks. | |
62 | */ | |
7edc5962 | 63 | int number_of_cpusets __read_mostly; |
202f72d5 | 64 | |
3e0d98b9 PJ |
65 | /* See "Frequency meter" comments, below. */ |
66 | ||
67 | struct fmeter { | |
68 | int cnt; /* unprocessed events count */ | |
69 | int val; /* most recent output value */ | |
70 | time_t time; /* clock (secs) when val computed */ | |
71 | spinlock_t lock; /* guards read or write of above */ | |
72 | }; | |
73 | ||
1da177e4 LT |
74 | struct cpuset { |
75 | unsigned long flags; /* "unsigned long" so bitops work */ | |
76 | cpumask_t cpus_allowed; /* CPUs allowed to tasks in cpuset */ | |
77 | nodemask_t mems_allowed; /* Memory Nodes allowed to tasks */ | |
78 | ||
053199ed PJ |
79 | /* |
80 | * Count is atomic so can incr (fork) or decr (exit) without a lock. | |
81 | */ | |
1da177e4 LT |
82 | atomic_t count; /* count tasks using this cpuset */ |
83 | ||
84 | /* | |
85 | * We link our 'sibling' struct into our parents 'children'. | |
86 | * Our children link their 'sibling' into our 'children'. | |
87 | */ | |
88 | struct list_head sibling; /* my parents children */ | |
89 | struct list_head children; /* my children */ | |
90 | ||
91 | struct cpuset *parent; /* my parent */ | |
92 | struct dentry *dentry; /* cpuset fs entry */ | |
93 | ||
94 | /* | |
95 | * Copy of global cpuset_mems_generation as of the most | |
96 | * recent time this cpuset changed its mems_allowed. | |
97 | */ | |
3e0d98b9 PJ |
98 | int mems_generation; |
99 | ||
100 | struct fmeter fmeter; /* memory_pressure filter */ | |
1da177e4 LT |
101 | }; |
102 | ||
103 | /* bits in struct cpuset flags field */ | |
104 | typedef enum { | |
105 | CS_CPU_EXCLUSIVE, | |
106 | CS_MEM_EXCLUSIVE, | |
45b07ef3 | 107 | CS_MEMORY_MIGRATE, |
1da177e4 | 108 | CS_REMOVED, |
825a46af PJ |
109 | CS_NOTIFY_ON_RELEASE, |
110 | CS_SPREAD_PAGE, | |
111 | CS_SPREAD_SLAB, | |
1da177e4 LT |
112 | } cpuset_flagbits_t; |
113 | ||
114 | /* convenient tests for these bits */ | |
115 | static inline int is_cpu_exclusive(const struct cpuset *cs) | |
116 | { | |
7b5b9ef0 | 117 | return test_bit(CS_CPU_EXCLUSIVE, &cs->flags); |
1da177e4 LT |
118 | } |
119 | ||
120 | static inline int is_mem_exclusive(const struct cpuset *cs) | |
121 | { | |
7b5b9ef0 | 122 | return test_bit(CS_MEM_EXCLUSIVE, &cs->flags); |
1da177e4 LT |
123 | } |
124 | ||
125 | static inline int is_removed(const struct cpuset *cs) | |
126 | { | |
7b5b9ef0 | 127 | return test_bit(CS_REMOVED, &cs->flags); |
1da177e4 LT |
128 | } |
129 | ||
130 | static inline int notify_on_release(const struct cpuset *cs) | |
131 | { | |
7b5b9ef0 | 132 | return test_bit(CS_NOTIFY_ON_RELEASE, &cs->flags); |
1da177e4 LT |
133 | } |
134 | ||
45b07ef3 PJ |
135 | static inline int is_memory_migrate(const struct cpuset *cs) |
136 | { | |
7b5b9ef0 | 137 | return test_bit(CS_MEMORY_MIGRATE, &cs->flags); |
45b07ef3 PJ |
138 | } |
139 | ||
825a46af PJ |
140 | static inline int is_spread_page(const struct cpuset *cs) |
141 | { | |
142 | return test_bit(CS_SPREAD_PAGE, &cs->flags); | |
143 | } | |
144 | ||
145 | static inline int is_spread_slab(const struct cpuset *cs) | |
146 | { | |
147 | return test_bit(CS_SPREAD_SLAB, &cs->flags); | |
148 | } | |
149 | ||
1da177e4 | 150 | /* |
151a4420 | 151 | * Increment this integer everytime any cpuset changes its |
1da177e4 LT |
152 | * mems_allowed value. Users of cpusets can track this generation |
153 | * number, and avoid having to lock and reload mems_allowed unless | |
154 | * the cpuset they're using changes generation. | |
155 | * | |
156 | * A single, global generation is needed because attach_task() could | |
157 | * reattach a task to a different cpuset, which must not have its | |
158 | * generation numbers aliased with those of that tasks previous cpuset. | |
159 | * | |
160 | * Generations are needed for mems_allowed because one task cannot | |
161 | * modify anothers memory placement. So we must enable every task, | |
162 | * on every visit to __alloc_pages(), to efficiently check whether | |
163 | * its current->cpuset->mems_allowed has changed, requiring an update | |
164 | * of its current->mems_allowed. | |
151a4420 PJ |
165 | * |
166 | * Since cpuset_mems_generation is guarded by manage_mutex, | |
167 | * there is no need to mark it atomic. | |
1da177e4 | 168 | */ |
151a4420 | 169 | static int cpuset_mems_generation; |
1da177e4 LT |
170 | |
171 | static struct cpuset top_cpuset = { | |
172 | .flags = ((1 << CS_CPU_EXCLUSIVE) | (1 << CS_MEM_EXCLUSIVE)), | |
173 | .cpus_allowed = CPU_MASK_ALL, | |
174 | .mems_allowed = NODE_MASK_ALL, | |
175 | .count = ATOMIC_INIT(0), | |
176 | .sibling = LIST_HEAD_INIT(top_cpuset.sibling), | |
177 | .children = LIST_HEAD_INIT(top_cpuset.children), | |
1da177e4 LT |
178 | }; |
179 | ||
180 | static struct vfsmount *cpuset_mount; | |
3e0d98b9 | 181 | static struct super_block *cpuset_sb; |
1da177e4 LT |
182 | |
183 | /* | |
3d3f26a7 IM |
184 | * We have two global cpuset mutexes below. They can nest. |
185 | * It is ok to first take manage_mutex, then nest callback_mutex. We also | |
053199ed PJ |
186 | * require taking task_lock() when dereferencing a tasks cpuset pointer. |
187 | * See "The task_lock() exception", at the end of this comment. | |
188 | * | |
3d3f26a7 IM |
189 | * A task must hold both mutexes to modify cpusets. If a task |
190 | * holds manage_mutex, then it blocks others wanting that mutex, | |
191 | * ensuring that it is the only task able to also acquire callback_mutex | |
053199ed PJ |
192 | * and be able to modify cpusets. It can perform various checks on |
193 | * the cpuset structure first, knowing nothing will change. It can | |
3d3f26a7 | 194 | * also allocate memory while just holding manage_mutex. While it is |
053199ed | 195 | * performing these checks, various callback routines can briefly |
3d3f26a7 IM |
196 | * acquire callback_mutex to query cpusets. Once it is ready to make |
197 | * the changes, it takes callback_mutex, blocking everyone else. | |
053199ed PJ |
198 | * |
199 | * Calls to the kernel memory allocator can not be made while holding | |
3d3f26a7 | 200 | * callback_mutex, as that would risk double tripping on callback_mutex |
053199ed PJ |
201 | * from one of the callbacks into the cpuset code from within |
202 | * __alloc_pages(). | |
203 | * | |
3d3f26a7 | 204 | * If a task is only holding callback_mutex, then it has read-only |
053199ed PJ |
205 | * access to cpusets. |
206 | * | |
207 | * The task_struct fields mems_allowed and mems_generation may only | |
208 | * be accessed in the context of that task, so require no locks. | |
209 | * | |
210 | * Any task can increment and decrement the count field without lock. | |
3d3f26a7 | 211 | * So in general, code holding manage_mutex or callback_mutex can't rely |
053199ed | 212 | * on the count field not changing. However, if the count goes to |
3d3f26a7 | 213 | * zero, then only attach_task(), which holds both mutexes, can |
053199ed PJ |
214 | * increment it again. Because a count of zero means that no tasks |
215 | * are currently attached, therefore there is no way a task attached | |
216 | * to that cpuset can fork (the other way to increment the count). | |
3d3f26a7 | 217 | * So code holding manage_mutex or callback_mutex can safely assume that |
053199ed | 218 | * if the count is zero, it will stay zero. Similarly, if a task |
3d3f26a7 | 219 | * holds manage_mutex or callback_mutex on a cpuset with zero count, it |
053199ed | 220 | * knows that the cpuset won't be removed, as cpuset_rmdir() needs |
3d3f26a7 | 221 | * both of those mutexes. |
053199ed PJ |
222 | * |
223 | * The cpuset_common_file_write handler for operations that modify | |
3d3f26a7 | 224 | * the cpuset hierarchy holds manage_mutex across the entire operation, |
053199ed PJ |
225 | * single threading all such cpuset modifications across the system. |
226 | * | |
3d3f26a7 | 227 | * The cpuset_common_file_read() handlers only hold callback_mutex across |
053199ed PJ |
228 | * small pieces of code, such as when reading out possibly multi-word |
229 | * cpumasks and nodemasks. | |
230 | * | |
231 | * The fork and exit callbacks cpuset_fork() and cpuset_exit(), don't | |
3d3f26a7 | 232 | * (usually) take either mutex. These are the two most performance |
053199ed | 233 | * critical pieces of code here. The exception occurs on cpuset_exit(), |
3d3f26a7 | 234 | * when a task in a notify_on_release cpuset exits. Then manage_mutex |
2efe86b8 | 235 | * is taken, and if the cpuset count is zero, a usermode call made |
1da177e4 LT |
236 | * to /sbin/cpuset_release_agent with the name of the cpuset (path |
237 | * relative to the root of cpuset file system) as the argument. | |
238 | * | |
053199ed PJ |
239 | * A cpuset can only be deleted if both its 'count' of using tasks |
240 | * is zero, and its list of 'children' cpusets is empty. Since all | |
241 | * tasks in the system use _some_ cpuset, and since there is always at | |
f400e198 | 242 | * least one task in the system (init), therefore, top_cpuset |
053199ed PJ |
243 | * always has either children cpusets and/or using tasks. So we don't |
244 | * need a special hack to ensure that top_cpuset cannot be deleted. | |
245 | * | |
246 | * The above "Tale of Two Semaphores" would be complete, but for: | |
247 | * | |
248 | * The task_lock() exception | |
249 | * | |
250 | * The need for this exception arises from the action of attach_task(), | |
251 | * which overwrites one tasks cpuset pointer with another. It does | |
3d3f26a7 | 252 | * so using both mutexes, however there are several performance |
053199ed | 253 | * critical places that need to reference task->cpuset without the |
3d3f26a7 | 254 | * expense of grabbing a system global mutex. Therefore except as |
053199ed PJ |
255 | * noted below, when dereferencing or, as in attach_task(), modifying |
256 | * a tasks cpuset pointer we use task_lock(), which acts on a spinlock | |
257 | * (task->alloc_lock) already in the task_struct routinely used for | |
258 | * such matters. | |
6b9c2603 PJ |
259 | * |
260 | * P.S. One more locking exception. RCU is used to guard the | |
261 | * update of a tasks cpuset pointer by attach_task() and the | |
262 | * access of task->cpuset->mems_generation via that pointer in | |
263 | * the routine cpuset_update_task_memory_state(). | |
1da177e4 LT |
264 | */ |
265 | ||
3d3f26a7 IM |
266 | static DEFINE_MUTEX(manage_mutex); |
267 | static DEFINE_MUTEX(callback_mutex); | |
4247bdc6 | 268 | |
1da177e4 LT |
269 | /* |
270 | * A couple of forward declarations required, due to cyclic reference loop: | |
271 | * cpuset_mkdir -> cpuset_create -> cpuset_populate_dir -> cpuset_add_file | |
272 | * -> cpuset_create_file -> cpuset_dir_inode_operations -> cpuset_mkdir. | |
273 | */ | |
274 | ||
275 | static int cpuset_mkdir(struct inode *dir, struct dentry *dentry, int mode); | |
276 | static int cpuset_rmdir(struct inode *unused_dir, struct dentry *dentry); | |
277 | ||
278 | static struct backing_dev_info cpuset_backing_dev_info = { | |
279 | .ra_pages = 0, /* No readahead */ | |
280 | .capabilities = BDI_CAP_NO_ACCT_DIRTY | BDI_CAP_NO_WRITEBACK, | |
281 | }; | |
282 | ||
283 | static struct inode *cpuset_new_inode(mode_t mode) | |
284 | { | |
285 | struct inode *inode = new_inode(cpuset_sb); | |
286 | ||
287 | if (inode) { | |
288 | inode->i_mode = mode; | |
289 | inode->i_uid = current->fsuid; | |
290 | inode->i_gid = current->fsgid; | |
1da177e4 LT |
291 | inode->i_blocks = 0; |
292 | inode->i_atime = inode->i_mtime = inode->i_ctime = CURRENT_TIME; | |
293 | inode->i_mapping->backing_dev_info = &cpuset_backing_dev_info; | |
294 | } | |
295 | return inode; | |
296 | } | |
297 | ||
298 | static void cpuset_diput(struct dentry *dentry, struct inode *inode) | |
299 | { | |
300 | /* is dentry a directory ? if so, kfree() associated cpuset */ | |
301 | if (S_ISDIR(inode->i_mode)) { | |
302 | struct cpuset *cs = dentry->d_fsdata; | |
303 | BUG_ON(!(is_removed(cs))); | |
304 | kfree(cs); | |
305 | } | |
306 | iput(inode); | |
307 | } | |
308 | ||
309 | static struct dentry_operations cpuset_dops = { | |
310 | .d_iput = cpuset_diput, | |
311 | }; | |
312 | ||
313 | static struct dentry *cpuset_get_dentry(struct dentry *parent, const char *name) | |
314 | { | |
5f45f1a7 | 315 | struct dentry *d = lookup_one_len(name, parent, strlen(name)); |
1da177e4 LT |
316 | if (!IS_ERR(d)) |
317 | d->d_op = &cpuset_dops; | |
318 | return d; | |
319 | } | |
320 | ||
321 | static void remove_dir(struct dentry *d) | |
322 | { | |
323 | struct dentry *parent = dget(d->d_parent); | |
324 | ||
325 | d_delete(d); | |
326 | simple_rmdir(parent->d_inode, d); | |
327 | dput(parent); | |
328 | } | |
329 | ||
330 | /* | |
331 | * NOTE : the dentry must have been dget()'ed | |
332 | */ | |
333 | static void cpuset_d_remove_dir(struct dentry *dentry) | |
334 | { | |
335 | struct list_head *node; | |
336 | ||
337 | spin_lock(&dcache_lock); | |
338 | node = dentry->d_subdirs.next; | |
339 | while (node != &dentry->d_subdirs) { | |
5160ee6f | 340 | struct dentry *d = list_entry(node, struct dentry, d_u.d_child); |
1da177e4 LT |
341 | list_del_init(node); |
342 | if (d->d_inode) { | |
343 | d = dget_locked(d); | |
344 | spin_unlock(&dcache_lock); | |
345 | d_delete(d); | |
346 | simple_unlink(dentry->d_inode, d); | |
347 | dput(d); | |
348 | spin_lock(&dcache_lock); | |
349 | } | |
350 | node = dentry->d_subdirs.next; | |
351 | } | |
5160ee6f | 352 | list_del_init(&dentry->d_u.d_child); |
1da177e4 LT |
353 | spin_unlock(&dcache_lock); |
354 | remove_dir(dentry); | |
355 | } | |
356 | ||
357 | static struct super_operations cpuset_ops = { | |
358 | .statfs = simple_statfs, | |
359 | .drop_inode = generic_delete_inode, | |
360 | }; | |
361 | ||
362 | static int cpuset_fill_super(struct super_block *sb, void *unused_data, | |
363 | int unused_silent) | |
364 | { | |
365 | struct inode *inode; | |
366 | struct dentry *root; | |
367 | ||
368 | sb->s_blocksize = PAGE_CACHE_SIZE; | |
369 | sb->s_blocksize_bits = PAGE_CACHE_SHIFT; | |
370 | sb->s_magic = CPUSET_SUPER_MAGIC; | |
371 | sb->s_op = &cpuset_ops; | |
372 | cpuset_sb = sb; | |
373 | ||
374 | inode = cpuset_new_inode(S_IFDIR | S_IRUGO | S_IXUGO | S_IWUSR); | |
375 | if (inode) { | |
376 | inode->i_op = &simple_dir_inode_operations; | |
377 | inode->i_fop = &simple_dir_operations; | |
378 | /* directories start off with i_nlink == 2 (for "." entry) */ | |
d8c76e6f | 379 | inc_nlink(inode); |
1da177e4 LT |
380 | } else { |
381 | return -ENOMEM; | |
382 | } | |
383 | ||
384 | root = d_alloc_root(inode); | |
385 | if (!root) { | |
386 | iput(inode); | |
387 | return -ENOMEM; | |
388 | } | |
389 | sb->s_root = root; | |
390 | return 0; | |
391 | } | |
392 | ||
454e2398 DH |
393 | static int cpuset_get_sb(struct file_system_type *fs_type, |
394 | int flags, const char *unused_dev_name, | |
395 | void *data, struct vfsmount *mnt) | |
1da177e4 | 396 | { |
454e2398 | 397 | return get_sb_single(fs_type, flags, data, cpuset_fill_super, mnt); |
1da177e4 LT |
398 | } |
399 | ||
400 | static struct file_system_type cpuset_fs_type = { | |
401 | .name = "cpuset", | |
402 | .get_sb = cpuset_get_sb, | |
403 | .kill_sb = kill_litter_super, | |
404 | }; | |
405 | ||
406 | /* struct cftype: | |
407 | * | |
408 | * The files in the cpuset filesystem mostly have a very simple read/write | |
409 | * handling, some common function will take care of it. Nevertheless some cases | |
410 | * (read tasks) are special and therefore I define this structure for every | |
411 | * kind of file. | |
412 | * | |
413 | * | |
414 | * When reading/writing to a file: | |
a7a005fd JS |
415 | * - the cpuset to use in file->f_path.dentry->d_parent->d_fsdata |
416 | * - the 'cftype' of the file is file->f_path.dentry->d_fsdata | |
1da177e4 LT |
417 | */ |
418 | ||
419 | struct cftype { | |
420 | char *name; | |
421 | int private; | |
422 | int (*open) (struct inode *inode, struct file *file); | |
423 | ssize_t (*read) (struct file *file, char __user *buf, size_t nbytes, | |
424 | loff_t *ppos); | |
425 | int (*write) (struct file *file, const char __user *buf, size_t nbytes, | |
426 | loff_t *ppos); | |
427 | int (*release) (struct inode *inode, struct file *file); | |
428 | }; | |
429 | ||
430 | static inline struct cpuset *__d_cs(struct dentry *dentry) | |
431 | { | |
432 | return dentry->d_fsdata; | |
433 | } | |
434 | ||
435 | static inline struct cftype *__d_cft(struct dentry *dentry) | |
436 | { | |
437 | return dentry->d_fsdata; | |
438 | } | |
439 | ||
440 | /* | |
3d3f26a7 | 441 | * Call with manage_mutex held. Writes path of cpuset into buf. |
1da177e4 LT |
442 | * Returns 0 on success, -errno on error. |
443 | */ | |
444 | ||
445 | static int cpuset_path(const struct cpuset *cs, char *buf, int buflen) | |
446 | { | |
447 | char *start; | |
448 | ||
449 | start = buf + buflen; | |
450 | ||
451 | *--start = '\0'; | |
452 | for (;;) { | |
453 | int len = cs->dentry->d_name.len; | |
454 | if ((start -= len) < buf) | |
455 | return -ENAMETOOLONG; | |
456 | memcpy(start, cs->dentry->d_name.name, len); | |
457 | cs = cs->parent; | |
458 | if (!cs) | |
459 | break; | |
460 | if (!cs->parent) | |
461 | continue; | |
462 | if (--start < buf) | |
463 | return -ENAMETOOLONG; | |
464 | *start = '/'; | |
465 | } | |
466 | memmove(buf, start, buf + buflen - start); | |
467 | return 0; | |
468 | } | |
469 | ||
470 | /* | |
471 | * Notify userspace when a cpuset is released, by running | |
472 | * /sbin/cpuset_release_agent with the name of the cpuset (path | |
473 | * relative to the root of cpuset file system) as the argument. | |
474 | * | |
475 | * Most likely, this user command will try to rmdir this cpuset. | |
476 | * | |
477 | * This races with the possibility that some other task will be | |
478 | * attached to this cpuset before it is removed, or that some other | |
479 | * user task will 'mkdir' a child cpuset of this cpuset. That's ok. | |
480 | * The presumed 'rmdir' will fail quietly if this cpuset is no longer | |
481 | * unused, and this cpuset will be reprieved from its death sentence, | |
482 | * to continue to serve a useful existence. Next time it's released, | |
483 | * we will get notified again, if it still has 'notify_on_release' set. | |
484 | * | |
3077a260 PJ |
485 | * The final arg to call_usermodehelper() is 0, which means don't |
486 | * wait. The separate /sbin/cpuset_release_agent task is forked by | |
487 | * call_usermodehelper(), then control in this thread returns here, | |
488 | * without waiting for the release agent task. We don't bother to | |
489 | * wait because the caller of this routine has no use for the exit | |
490 | * status of the /sbin/cpuset_release_agent task, so no sense holding | |
491 | * our caller up for that. | |
492 | * | |
3d3f26a7 | 493 | * When we had only one cpuset mutex, we had to call this |
053199ed PJ |
494 | * without holding it, to avoid deadlock when call_usermodehelper() |
495 | * allocated memory. With two locks, we could now call this while | |
3d3f26a7 IM |
496 | * holding manage_mutex, but we still don't, so as to minimize |
497 | * the time manage_mutex is held. | |
1da177e4 LT |
498 | */ |
499 | ||
3077a260 | 500 | static void cpuset_release_agent(const char *pathbuf) |
1da177e4 LT |
501 | { |
502 | char *argv[3], *envp[3]; | |
503 | int i; | |
504 | ||
3077a260 PJ |
505 | if (!pathbuf) |
506 | return; | |
507 | ||
1da177e4 LT |
508 | i = 0; |
509 | argv[i++] = "/sbin/cpuset_release_agent"; | |
3077a260 | 510 | argv[i++] = (char *)pathbuf; |
1da177e4 LT |
511 | argv[i] = NULL; |
512 | ||
513 | i = 0; | |
514 | /* minimal command environment */ | |
515 | envp[i++] = "HOME=/"; | |
516 | envp[i++] = "PATH=/sbin:/bin:/usr/sbin:/usr/bin"; | |
517 | envp[i] = NULL; | |
518 | ||
86313c48 | 519 | call_usermodehelper(argv[0], argv, envp, UMH_WAIT_EXEC); |
3077a260 | 520 | kfree(pathbuf); |
1da177e4 LT |
521 | } |
522 | ||
523 | /* | |
524 | * Either cs->count of using tasks transitioned to zero, or the | |
525 | * cs->children list of child cpusets just became empty. If this | |
526 | * cs is notify_on_release() and now both the user count is zero and | |
3077a260 PJ |
527 | * the list of children is empty, prepare cpuset path in a kmalloc'd |
528 | * buffer, to be returned via ppathbuf, so that the caller can invoke | |
3d3f26a7 IM |
529 | * cpuset_release_agent() with it later on, once manage_mutex is dropped. |
530 | * Call here with manage_mutex held. | |
3077a260 PJ |
531 | * |
532 | * This check_for_release() routine is responsible for kmalloc'ing | |
533 | * pathbuf. The above cpuset_release_agent() is responsible for | |
534 | * kfree'ing pathbuf. The caller of these routines is responsible | |
535 | * for providing a pathbuf pointer, initialized to NULL, then | |
3d3f26a7 IM |
536 | * calling check_for_release() with manage_mutex held and the address |
537 | * of the pathbuf pointer, then dropping manage_mutex, then calling | |
3077a260 | 538 | * cpuset_release_agent() with pathbuf, as set by check_for_release(). |
1da177e4 LT |
539 | */ |
540 | ||
3077a260 | 541 | static void check_for_release(struct cpuset *cs, char **ppathbuf) |
1da177e4 LT |
542 | { |
543 | if (notify_on_release(cs) && atomic_read(&cs->count) == 0 && | |
544 | list_empty(&cs->children)) { | |
545 | char *buf; | |
546 | ||
547 | buf = kmalloc(PAGE_SIZE, GFP_KERNEL); | |
548 | if (!buf) | |
549 | return; | |
550 | if (cpuset_path(cs, buf, PAGE_SIZE) < 0) | |
3077a260 PJ |
551 | kfree(buf); |
552 | else | |
553 | *ppathbuf = buf; | |
1da177e4 LT |
554 | } |
555 | } | |
556 | ||
557 | /* | |
558 | * Return in *pmask the portion of a cpusets's cpus_allowed that | |
559 | * are online. If none are online, walk up the cpuset hierarchy | |
560 | * until we find one that does have some online cpus. If we get | |
561 | * all the way to the top and still haven't found any online cpus, | |
562 | * return cpu_online_map. Or if passed a NULL cs from an exit'ing | |
563 | * task, return cpu_online_map. | |
564 | * | |
565 | * One way or another, we guarantee to return some non-empty subset | |
566 | * of cpu_online_map. | |
567 | * | |
3d3f26a7 | 568 | * Call with callback_mutex held. |
1da177e4 LT |
569 | */ |
570 | ||
571 | static void guarantee_online_cpus(const struct cpuset *cs, cpumask_t *pmask) | |
572 | { | |
573 | while (cs && !cpus_intersects(cs->cpus_allowed, cpu_online_map)) | |
574 | cs = cs->parent; | |
575 | if (cs) | |
576 | cpus_and(*pmask, cs->cpus_allowed, cpu_online_map); | |
577 | else | |
578 | *pmask = cpu_online_map; | |
579 | BUG_ON(!cpus_intersects(*pmask, cpu_online_map)); | |
580 | } | |
581 | ||
582 | /* | |
583 | * Return in *pmask the portion of a cpusets's mems_allowed that | |
0e1e7c7a CL |
584 | * are online, with memory. If none are online with memory, walk |
585 | * up the cpuset hierarchy until we find one that does have some | |
586 | * online mems. If we get all the way to the top and still haven't | |
587 | * found any online mems, return node_states[N_HIGH_MEMORY]. | |
1da177e4 LT |
588 | * |
589 | * One way or another, we guarantee to return some non-empty subset | |
0e1e7c7a | 590 | * of node_states[N_HIGH_MEMORY]. |
1da177e4 | 591 | * |
3d3f26a7 | 592 | * Call with callback_mutex held. |
1da177e4 LT |
593 | */ |
594 | ||
595 | static void guarantee_online_mems(const struct cpuset *cs, nodemask_t *pmask) | |
596 | { | |
0e1e7c7a CL |
597 | while (cs && !nodes_intersects(cs->mems_allowed, |
598 | node_states[N_HIGH_MEMORY])) | |
1da177e4 LT |
599 | cs = cs->parent; |
600 | if (cs) | |
0e1e7c7a CL |
601 | nodes_and(*pmask, cs->mems_allowed, |
602 | node_states[N_HIGH_MEMORY]); | |
1da177e4 | 603 | else |
0e1e7c7a CL |
604 | *pmask = node_states[N_HIGH_MEMORY]; |
605 | BUG_ON(!nodes_intersects(*pmask, node_states[N_HIGH_MEMORY])); | |
1da177e4 LT |
606 | } |
607 | ||
cf2a473c PJ |
608 | /** |
609 | * cpuset_update_task_memory_state - update task memory placement | |
610 | * | |
611 | * If the current tasks cpusets mems_allowed changed behind our | |
612 | * backs, update current->mems_allowed, mems_generation and task NUMA | |
613 | * mempolicy to the new value. | |
053199ed | 614 | * |
cf2a473c PJ |
615 | * Task mempolicy is updated by rebinding it relative to the |
616 | * current->cpuset if a task has its memory placement changed. | |
617 | * Do not call this routine if in_interrupt(). | |
618 | * | |
4a01c8d5 PJ |
619 | * Call without callback_mutex or task_lock() held. May be |
620 | * called with or without manage_mutex held. Thanks in part to | |
621 | * 'the_top_cpuset_hack', the tasks cpuset pointer will never | |
622 | * be NULL. This routine also might acquire callback_mutex and | |
cf2a473c | 623 | * current->mm->mmap_sem during call. |
053199ed | 624 | * |
6b9c2603 PJ |
625 | * Reading current->cpuset->mems_generation doesn't need task_lock |
626 | * to guard the current->cpuset derefence, because it is guarded | |
627 | * from concurrent freeing of current->cpuset by attach_task(), | |
628 | * using RCU. | |
629 | * | |
630 | * The rcu_dereference() is technically probably not needed, | |
631 | * as I don't actually mind if I see a new cpuset pointer but | |
632 | * an old value of mems_generation. However this really only | |
633 | * matters on alpha systems using cpusets heavily. If I dropped | |
634 | * that rcu_dereference(), it would save them a memory barrier. | |
635 | * For all other arch's, rcu_dereference is a no-op anyway, and for | |
636 | * alpha systems not using cpusets, another planned optimization, | |
637 | * avoiding the rcu critical section for tasks in the root cpuset | |
638 | * which is statically allocated, so can't vanish, will make this | |
639 | * irrelevant. Better to use RCU as intended, than to engage in | |
640 | * some cute trick to save a memory barrier that is impossible to | |
641 | * test, for alpha systems using cpusets heavily, which might not | |
642 | * even exist. | |
053199ed PJ |
643 | * |
644 | * This routine is needed to update the per-task mems_allowed data, | |
645 | * within the tasks context, when it is trying to allocate memory | |
646 | * (in various mm/mempolicy.c routines) and notices that some other | |
647 | * task has been modifying its cpuset. | |
1da177e4 LT |
648 | */ |
649 | ||
fe85a998 | 650 | void cpuset_update_task_memory_state(void) |
1da177e4 | 651 | { |
053199ed | 652 | int my_cpusets_mem_gen; |
cf2a473c | 653 | struct task_struct *tsk = current; |
6b9c2603 | 654 | struct cpuset *cs; |
053199ed | 655 | |
03a285f5 PJ |
656 | if (tsk->cpuset == &top_cpuset) { |
657 | /* Don't need rcu for top_cpuset. It's never freed. */ | |
658 | my_cpusets_mem_gen = top_cpuset.mems_generation; | |
659 | } else { | |
660 | rcu_read_lock(); | |
661 | cs = rcu_dereference(tsk->cpuset); | |
662 | my_cpusets_mem_gen = cs->mems_generation; | |
663 | rcu_read_unlock(); | |
664 | } | |
1da177e4 | 665 | |
cf2a473c | 666 | if (my_cpusets_mem_gen != tsk->cpuset_mems_generation) { |
3d3f26a7 | 667 | mutex_lock(&callback_mutex); |
cf2a473c PJ |
668 | task_lock(tsk); |
669 | cs = tsk->cpuset; /* Maybe changed when task not locked */ | |
cf2a473c PJ |
670 | guarantee_online_mems(cs, &tsk->mems_allowed); |
671 | tsk->cpuset_mems_generation = cs->mems_generation; | |
825a46af PJ |
672 | if (is_spread_page(cs)) |
673 | tsk->flags |= PF_SPREAD_PAGE; | |
674 | else | |
675 | tsk->flags &= ~PF_SPREAD_PAGE; | |
676 | if (is_spread_slab(cs)) | |
677 | tsk->flags |= PF_SPREAD_SLAB; | |
678 | else | |
679 | tsk->flags &= ~PF_SPREAD_SLAB; | |
cf2a473c | 680 | task_unlock(tsk); |
3d3f26a7 | 681 | mutex_unlock(&callback_mutex); |
74cb2155 | 682 | mpol_rebind_task(tsk, &tsk->mems_allowed); |
1da177e4 LT |
683 | } |
684 | } | |
685 | ||
686 | /* | |
687 | * is_cpuset_subset(p, q) - Is cpuset p a subset of cpuset q? | |
688 | * | |
689 | * One cpuset is a subset of another if all its allowed CPUs and | |
690 | * Memory Nodes are a subset of the other, and its exclusive flags | |
3d3f26a7 | 691 | * are only set if the other's are set. Call holding manage_mutex. |
1da177e4 LT |
692 | */ |
693 | ||
694 | static int is_cpuset_subset(const struct cpuset *p, const struct cpuset *q) | |
695 | { | |
696 | return cpus_subset(p->cpus_allowed, q->cpus_allowed) && | |
697 | nodes_subset(p->mems_allowed, q->mems_allowed) && | |
698 | is_cpu_exclusive(p) <= is_cpu_exclusive(q) && | |
699 | is_mem_exclusive(p) <= is_mem_exclusive(q); | |
700 | } | |
701 | ||
702 | /* | |
703 | * validate_change() - Used to validate that any proposed cpuset change | |
704 | * follows the structural rules for cpusets. | |
705 | * | |
706 | * If we replaced the flag and mask values of the current cpuset | |
707 | * (cur) with those values in the trial cpuset (trial), would | |
708 | * our various subset and exclusive rules still be valid? Presumes | |
3d3f26a7 | 709 | * manage_mutex held. |
1da177e4 LT |
710 | * |
711 | * 'cur' is the address of an actual, in-use cpuset. Operations | |
712 | * such as list traversal that depend on the actual address of the | |
713 | * cpuset in the list must use cur below, not trial. | |
714 | * | |
715 | * 'trial' is the address of bulk structure copy of cur, with | |
716 | * perhaps one or more of the fields cpus_allowed, mems_allowed, | |
717 | * or flags changed to new, trial values. | |
718 | * | |
719 | * Return 0 if valid, -errno if not. | |
720 | */ | |
721 | ||
722 | static int validate_change(const struct cpuset *cur, const struct cpuset *trial) | |
723 | { | |
724 | struct cpuset *c, *par; | |
725 | ||
726 | /* Each of our child cpusets must be a subset of us */ | |
727 | list_for_each_entry(c, &cur->children, sibling) { | |
728 | if (!is_cpuset_subset(c, trial)) | |
729 | return -EBUSY; | |
730 | } | |
731 | ||
732 | /* Remaining checks don't apply to root cpuset */ | |
69604067 | 733 | if (cur == &top_cpuset) |
1da177e4 LT |
734 | return 0; |
735 | ||
69604067 PJ |
736 | par = cur->parent; |
737 | ||
1da177e4 LT |
738 | /* We must be a subset of our parent cpuset */ |
739 | if (!is_cpuset_subset(trial, par)) | |
740 | return -EACCES; | |
741 | ||
742 | /* If either I or some sibling (!= me) is exclusive, we can't overlap */ | |
743 | list_for_each_entry(c, &par->children, sibling) { | |
744 | if ((is_cpu_exclusive(trial) || is_cpu_exclusive(c)) && | |
745 | c != cur && | |
746 | cpus_intersects(trial->cpus_allowed, c->cpus_allowed)) | |
747 | return -EINVAL; | |
748 | if ((is_mem_exclusive(trial) || is_mem_exclusive(c)) && | |
749 | c != cur && | |
750 | nodes_intersects(trial->mems_allowed, c->mems_allowed)) | |
751 | return -EINVAL; | |
752 | } | |
753 | ||
754 | return 0; | |
755 | } | |
756 | ||
053199ed | 757 | /* |
3d3f26a7 | 758 | * Call with manage_mutex held. May take callback_mutex during call. |
053199ed PJ |
759 | */ |
760 | ||
1da177e4 LT |
761 | static int update_cpumask(struct cpuset *cs, char *buf) |
762 | { | |
763 | struct cpuset trialcs; | |
607717a6 | 764 | int retval; |
1da177e4 | 765 | |
4c4d50f7 PJ |
766 | /* top_cpuset.cpus_allowed tracks cpu_online_map; it's read-only */ |
767 | if (cs == &top_cpuset) | |
768 | return -EACCES; | |
769 | ||
1da177e4 | 770 | trialcs = *cs; |
6f7f02e7 DR |
771 | |
772 | /* | |
773 | * We allow a cpuset's cpus_allowed to be empty; if it has attached | |
774 | * tasks, we'll catch it later when we validate the change and return | |
775 | * -ENOSPC. | |
776 | */ | |
777 | if (!buf[0] || (buf[0] == '\n' && !buf[1])) { | |
778 | cpus_clear(trialcs.cpus_allowed); | |
779 | } else { | |
780 | retval = cpulist_parse(buf, trialcs.cpus_allowed); | |
781 | if (retval < 0) | |
782 | return retval; | |
783 | } | |
1da177e4 | 784 | cpus_and(trialcs.cpus_allowed, trialcs.cpus_allowed, cpu_online_map); |
6f7f02e7 DR |
785 | /* cpus_allowed cannot be empty for a cpuset with attached tasks. */ |
786 | if (atomic_read(&cs->count) && cpus_empty(trialcs.cpus_allowed)) | |
1da177e4 LT |
787 | return -ENOSPC; |
788 | retval = validate_change(cs, &trialcs); | |
85d7b949 DG |
789 | if (retval < 0) |
790 | return retval; | |
3d3f26a7 | 791 | mutex_lock(&callback_mutex); |
85d7b949 | 792 | cs->cpus_allowed = trialcs.cpus_allowed; |
3d3f26a7 | 793 | mutex_unlock(&callback_mutex); |
85d7b949 | 794 | return 0; |
1da177e4 LT |
795 | } |
796 | ||
e4e364e8 PJ |
797 | /* |
798 | * cpuset_migrate_mm | |
799 | * | |
800 | * Migrate memory region from one set of nodes to another. | |
801 | * | |
802 | * Temporarilly set tasks mems_allowed to target nodes of migration, | |
803 | * so that the migration code can allocate pages on these nodes. | |
804 | * | |
805 | * Call holding manage_mutex, so our current->cpuset won't change | |
806 | * during this call, as manage_mutex holds off any attach_task() | |
807 | * calls. Therefore we don't need to take task_lock around the | |
808 | * call to guarantee_online_mems(), as we know no one is changing | |
809 | * our tasks cpuset. | |
810 | * | |
811 | * Hold callback_mutex around the two modifications of our tasks | |
812 | * mems_allowed to synchronize with cpuset_mems_allowed(). | |
813 | * | |
814 | * While the mm_struct we are migrating is typically from some | |
815 | * other task, the task_struct mems_allowed that we are hacking | |
816 | * is for our current task, which must allocate new pages for that | |
817 | * migrating memory region. | |
818 | * | |
819 | * We call cpuset_update_task_memory_state() before hacking | |
820 | * our tasks mems_allowed, so that we are assured of being in | |
821 | * sync with our tasks cpuset, and in particular, callbacks to | |
822 | * cpuset_update_task_memory_state() from nested page allocations | |
823 | * won't see any mismatch of our cpuset and task mems_generation | |
824 | * values, so won't overwrite our hacked tasks mems_allowed | |
825 | * nodemask. | |
826 | */ | |
827 | ||
828 | static void cpuset_migrate_mm(struct mm_struct *mm, const nodemask_t *from, | |
829 | const nodemask_t *to) | |
830 | { | |
831 | struct task_struct *tsk = current; | |
832 | ||
833 | cpuset_update_task_memory_state(); | |
834 | ||
835 | mutex_lock(&callback_mutex); | |
836 | tsk->mems_allowed = *to; | |
837 | mutex_unlock(&callback_mutex); | |
838 | ||
839 | do_migrate_pages(mm, from, to, MPOL_MF_MOVE_ALL); | |
840 | ||
841 | mutex_lock(&callback_mutex); | |
842 | guarantee_online_mems(tsk->cpuset, &tsk->mems_allowed); | |
843 | mutex_unlock(&callback_mutex); | |
844 | } | |
845 | ||
053199ed | 846 | /* |
4225399a PJ |
847 | * Handle user request to change the 'mems' memory placement |
848 | * of a cpuset. Needs to validate the request, update the | |
849 | * cpusets mems_allowed and mems_generation, and for each | |
04c19fa6 PJ |
850 | * task in the cpuset, rebind any vma mempolicies and if |
851 | * the cpuset is marked 'memory_migrate', migrate the tasks | |
852 | * pages to the new memory. | |
4225399a | 853 | * |
3d3f26a7 | 854 | * Call with manage_mutex held. May take callback_mutex during call. |
4225399a PJ |
855 | * Will take tasklist_lock, scan tasklist for tasks in cpuset cs, |
856 | * lock each such tasks mm->mmap_sem, scan its vma's and rebind | |
857 | * their mempolicies to the cpusets new mems_allowed. | |
053199ed PJ |
858 | */ |
859 | ||
1da177e4 LT |
860 | static int update_nodemask(struct cpuset *cs, char *buf) |
861 | { | |
862 | struct cpuset trialcs; | |
04c19fa6 | 863 | nodemask_t oldmem; |
4225399a PJ |
864 | struct task_struct *g, *p; |
865 | struct mm_struct **mmarray; | |
866 | int i, n, ntasks; | |
04c19fa6 | 867 | int migrate; |
4225399a | 868 | int fudge; |
1da177e4 LT |
869 | int retval; |
870 | ||
0e1e7c7a CL |
871 | /* |
872 | * top_cpuset.mems_allowed tracks node_stats[N_HIGH_MEMORY]; | |
873 | * it's read-only | |
874 | */ | |
38837fc7 PJ |
875 | if (cs == &top_cpuset) |
876 | return -EACCES; | |
877 | ||
1da177e4 | 878 | trialcs = *cs; |
6f7f02e7 DR |
879 | |
880 | /* | |
881 | * We allow a cpuset's mems_allowed to be empty; if it has attached | |
882 | * tasks, we'll catch it later when we validate the change and return | |
883 | * -ENOSPC. | |
884 | */ | |
885 | if (!buf[0] || (buf[0] == '\n' && !buf[1])) { | |
886 | nodes_clear(trialcs.mems_allowed); | |
887 | } else { | |
888 | retval = nodelist_parse(buf, trialcs.mems_allowed); | |
889 | if (retval < 0) | |
890 | goto done; | |
0e1e7c7a CL |
891 | if (!nodes_intersects(trialcs.mems_allowed, |
892 | node_states[N_HIGH_MEMORY])) { | |
893 | /* | |
894 | * error if only memoryless nodes specified. | |
895 | */ | |
896 | retval = -ENOSPC; | |
897 | goto done; | |
898 | } | |
6f7f02e7 | 899 | } |
0e1e7c7a CL |
900 | /* |
901 | * Exclude memoryless nodes. We know that trialcs.mems_allowed | |
902 | * contains at least one node with memory. | |
903 | */ | |
904 | nodes_and(trialcs.mems_allowed, trialcs.mems_allowed, | |
905 | node_states[N_HIGH_MEMORY]); | |
04c19fa6 PJ |
906 | oldmem = cs->mems_allowed; |
907 | if (nodes_equal(oldmem, trialcs.mems_allowed)) { | |
908 | retval = 0; /* Too easy - nothing to do */ | |
909 | goto done; | |
910 | } | |
6f7f02e7 DR |
911 | /* mems_allowed cannot be empty for a cpuset with attached tasks. */ |
912 | if (atomic_read(&cs->count) && nodes_empty(trialcs.mems_allowed)) { | |
59dac16f PJ |
913 | retval = -ENOSPC; |
914 | goto done; | |
1da177e4 | 915 | } |
59dac16f PJ |
916 | retval = validate_change(cs, &trialcs); |
917 | if (retval < 0) | |
918 | goto done; | |
919 | ||
3d3f26a7 | 920 | mutex_lock(&callback_mutex); |
59dac16f | 921 | cs->mems_allowed = trialcs.mems_allowed; |
151a4420 | 922 | cs->mems_generation = cpuset_mems_generation++; |
3d3f26a7 | 923 | mutex_unlock(&callback_mutex); |
59dac16f | 924 | |
4225399a PJ |
925 | set_cpuset_being_rebound(cs); /* causes mpol_copy() rebind */ |
926 | ||
927 | fudge = 10; /* spare mmarray[] slots */ | |
928 | fudge += cpus_weight(cs->cpus_allowed); /* imagine one fork-bomb/cpu */ | |
929 | retval = -ENOMEM; | |
930 | ||
931 | /* | |
932 | * Allocate mmarray[] to hold mm reference for each task | |
933 | * in cpuset cs. Can't kmalloc GFP_KERNEL while holding | |
934 | * tasklist_lock. We could use GFP_ATOMIC, but with a | |
935 | * few more lines of code, we can retry until we get a big | |
936 | * enough mmarray[] w/o using GFP_ATOMIC. | |
937 | */ | |
938 | while (1) { | |
939 | ntasks = atomic_read(&cs->count); /* guess */ | |
940 | ntasks += fudge; | |
941 | mmarray = kmalloc(ntasks * sizeof(*mmarray), GFP_KERNEL); | |
942 | if (!mmarray) | |
943 | goto done; | |
c2aef333 | 944 | read_lock(&tasklist_lock); /* block fork */ |
4225399a PJ |
945 | if (atomic_read(&cs->count) <= ntasks) |
946 | break; /* got enough */ | |
c2aef333 | 947 | read_unlock(&tasklist_lock); /* try again */ |
4225399a PJ |
948 | kfree(mmarray); |
949 | } | |
950 | ||
951 | n = 0; | |
952 | ||
953 | /* Load up mmarray[] with mm reference for each task in cpuset. */ | |
954 | do_each_thread(g, p) { | |
955 | struct mm_struct *mm; | |
956 | ||
957 | if (n >= ntasks) { | |
958 | printk(KERN_WARNING | |
959 | "Cpuset mempolicy rebind incomplete.\n"); | |
960 | continue; | |
961 | } | |
962 | if (p->cpuset != cs) | |
963 | continue; | |
964 | mm = get_task_mm(p); | |
965 | if (!mm) | |
966 | continue; | |
967 | mmarray[n++] = mm; | |
968 | } while_each_thread(g, p); | |
c2aef333 | 969 | read_unlock(&tasklist_lock); |
4225399a PJ |
970 | |
971 | /* | |
972 | * Now that we've dropped the tasklist spinlock, we can | |
973 | * rebind the vma mempolicies of each mm in mmarray[] to their | |
974 | * new cpuset, and release that mm. The mpol_rebind_mm() | |
975 | * call takes mmap_sem, which we couldn't take while holding | |
976 | * tasklist_lock. Forks can happen again now - the mpol_copy() | |
977 | * cpuset_being_rebound check will catch such forks, and rebind | |
978 | * their vma mempolicies too. Because we still hold the global | |
3d3f26a7 | 979 | * cpuset manage_mutex, we know that no other rebind effort will |
4225399a PJ |
980 | * be contending for the global variable cpuset_being_rebound. |
981 | * It's ok if we rebind the same mm twice; mpol_rebind_mm() | |
04c19fa6 | 982 | * is idempotent. Also migrate pages in each mm to new nodes. |
4225399a | 983 | */ |
04c19fa6 | 984 | migrate = is_memory_migrate(cs); |
4225399a PJ |
985 | for (i = 0; i < n; i++) { |
986 | struct mm_struct *mm = mmarray[i]; | |
987 | ||
988 | mpol_rebind_mm(mm, &cs->mems_allowed); | |
e4e364e8 PJ |
989 | if (migrate) |
990 | cpuset_migrate_mm(mm, &oldmem, &cs->mems_allowed); | |
4225399a PJ |
991 | mmput(mm); |
992 | } | |
993 | ||
994 | /* We're done rebinding vma's to this cpusets new mems_allowed. */ | |
995 | kfree(mmarray); | |
996 | set_cpuset_being_rebound(NULL); | |
997 | retval = 0; | |
59dac16f | 998 | done: |
1da177e4 LT |
999 | return retval; |
1000 | } | |
1001 | ||
3e0d98b9 | 1002 | /* |
3d3f26a7 | 1003 | * Call with manage_mutex held. |
3e0d98b9 PJ |
1004 | */ |
1005 | ||
1006 | static int update_memory_pressure_enabled(struct cpuset *cs, char *buf) | |
1007 | { | |
1008 | if (simple_strtoul(buf, NULL, 10) != 0) | |
1009 | cpuset_memory_pressure_enabled = 1; | |
1010 | else | |
1011 | cpuset_memory_pressure_enabled = 0; | |
1012 | return 0; | |
1013 | } | |
1014 | ||
1da177e4 LT |
1015 | /* |
1016 | * update_flag - read a 0 or a 1 in a file and update associated flag | |
1017 | * bit: the bit to update (CS_CPU_EXCLUSIVE, CS_MEM_EXCLUSIVE, | |
825a46af PJ |
1018 | * CS_NOTIFY_ON_RELEASE, CS_MEMORY_MIGRATE, |
1019 | * CS_SPREAD_PAGE, CS_SPREAD_SLAB) | |
1da177e4 LT |
1020 | * cs: the cpuset to update |
1021 | * buf: the buffer where we read the 0 or 1 | |
053199ed | 1022 | * |
3d3f26a7 | 1023 | * Call with manage_mutex held. |
1da177e4 LT |
1024 | */ |
1025 | ||
1026 | static int update_flag(cpuset_flagbits_t bit, struct cpuset *cs, char *buf) | |
1027 | { | |
1028 | int turning_on; | |
1029 | struct cpuset trialcs; | |
607717a6 | 1030 | int err; |
1da177e4 LT |
1031 | |
1032 | turning_on = (simple_strtoul(buf, NULL, 10) != 0); | |
1033 | ||
1034 | trialcs = *cs; | |
1035 | if (turning_on) | |
1036 | set_bit(bit, &trialcs.flags); | |
1037 | else | |
1038 | clear_bit(bit, &trialcs.flags); | |
1039 | ||
1040 | err = validate_change(cs, &trialcs); | |
85d7b949 DG |
1041 | if (err < 0) |
1042 | return err; | |
3d3f26a7 | 1043 | mutex_lock(&callback_mutex); |
69604067 | 1044 | cs->flags = trialcs.flags; |
3d3f26a7 | 1045 | mutex_unlock(&callback_mutex); |
85d7b949 | 1046 | |
85d7b949 | 1047 | return 0; |
1da177e4 LT |
1048 | } |
1049 | ||
3e0d98b9 | 1050 | /* |
80f7228b | 1051 | * Frequency meter - How fast is some event occurring? |
3e0d98b9 PJ |
1052 | * |
1053 | * These routines manage a digitally filtered, constant time based, | |
1054 | * event frequency meter. There are four routines: | |
1055 | * fmeter_init() - initialize a frequency meter. | |
1056 | * fmeter_markevent() - called each time the event happens. | |
1057 | * fmeter_getrate() - returns the recent rate of such events. | |
1058 | * fmeter_update() - internal routine used to update fmeter. | |
1059 | * | |
1060 | * A common data structure is passed to each of these routines, | |
1061 | * which is used to keep track of the state required to manage the | |
1062 | * frequency meter and its digital filter. | |
1063 | * | |
1064 | * The filter works on the number of events marked per unit time. | |
1065 | * The filter is single-pole low-pass recursive (IIR). The time unit | |
1066 | * is 1 second. Arithmetic is done using 32-bit integers scaled to | |
1067 | * simulate 3 decimal digits of precision (multiplied by 1000). | |
1068 | * | |
1069 | * With an FM_COEF of 933, and a time base of 1 second, the filter | |
1070 | * has a half-life of 10 seconds, meaning that if the events quit | |
1071 | * happening, then the rate returned from the fmeter_getrate() | |
1072 | * will be cut in half each 10 seconds, until it converges to zero. | |
1073 | * | |
1074 | * It is not worth doing a real infinitely recursive filter. If more | |
1075 | * than FM_MAXTICKS ticks have elapsed since the last filter event, | |
1076 | * just compute FM_MAXTICKS ticks worth, by which point the level | |
1077 | * will be stable. | |
1078 | * | |
1079 | * Limit the count of unprocessed events to FM_MAXCNT, so as to avoid | |
1080 | * arithmetic overflow in the fmeter_update() routine. | |
1081 | * | |
1082 | * Given the simple 32 bit integer arithmetic used, this meter works | |
1083 | * best for reporting rates between one per millisecond (msec) and | |
1084 | * one per 32 (approx) seconds. At constant rates faster than one | |
1085 | * per msec it maxes out at values just under 1,000,000. At constant | |
1086 | * rates between one per msec, and one per second it will stabilize | |
1087 | * to a value N*1000, where N is the rate of events per second. | |
1088 | * At constant rates between one per second and one per 32 seconds, | |
1089 | * it will be choppy, moving up on the seconds that have an event, | |
1090 | * and then decaying until the next event. At rates slower than | |
1091 | * about one in 32 seconds, it decays all the way back to zero between | |
1092 | * each event. | |
1093 | */ | |
1094 | ||
1095 | #define FM_COEF 933 /* coefficient for half-life of 10 secs */ | |
1096 | #define FM_MAXTICKS ((time_t)99) /* useless computing more ticks than this */ | |
1097 | #define FM_MAXCNT 1000000 /* limit cnt to avoid overflow */ | |
1098 | #define FM_SCALE 1000 /* faux fixed point scale */ | |
1099 | ||
1100 | /* Initialize a frequency meter */ | |
1101 | static void fmeter_init(struct fmeter *fmp) | |
1102 | { | |
1103 | fmp->cnt = 0; | |
1104 | fmp->val = 0; | |
1105 | fmp->time = 0; | |
1106 | spin_lock_init(&fmp->lock); | |
1107 | } | |
1108 | ||
1109 | /* Internal meter update - process cnt events and update value */ | |
1110 | static void fmeter_update(struct fmeter *fmp) | |
1111 | { | |
1112 | time_t now = get_seconds(); | |
1113 | time_t ticks = now - fmp->time; | |
1114 | ||
1115 | if (ticks == 0) | |
1116 | return; | |
1117 | ||
1118 | ticks = min(FM_MAXTICKS, ticks); | |
1119 | while (ticks-- > 0) | |
1120 | fmp->val = (FM_COEF * fmp->val) / FM_SCALE; | |
1121 | fmp->time = now; | |
1122 | ||
1123 | fmp->val += ((FM_SCALE - FM_COEF) * fmp->cnt) / FM_SCALE; | |
1124 | fmp->cnt = 0; | |
1125 | } | |
1126 | ||
1127 | /* Process any previous ticks, then bump cnt by one (times scale). */ | |
1128 | static void fmeter_markevent(struct fmeter *fmp) | |
1129 | { | |
1130 | spin_lock(&fmp->lock); | |
1131 | fmeter_update(fmp); | |
1132 | fmp->cnt = min(FM_MAXCNT, fmp->cnt + FM_SCALE); | |
1133 | spin_unlock(&fmp->lock); | |
1134 | } | |
1135 | ||
1136 | /* Process any previous ticks, then return current value. */ | |
1137 | static int fmeter_getrate(struct fmeter *fmp) | |
1138 | { | |
1139 | int val; | |
1140 | ||
1141 | spin_lock(&fmp->lock); | |
1142 | fmeter_update(fmp); | |
1143 | val = fmp->val; | |
1144 | spin_unlock(&fmp->lock); | |
1145 | return val; | |
1146 | } | |
1147 | ||
053199ed PJ |
1148 | /* |
1149 | * Attack task specified by pid in 'pidbuf' to cpuset 'cs', possibly | |
1150 | * writing the path of the old cpuset in 'ppathbuf' if it needs to be | |
1151 | * notified on release. | |
1152 | * | |
3d3f26a7 | 1153 | * Call holding manage_mutex. May take callback_mutex and task_lock of |
053199ed PJ |
1154 | * the task 'pid' during call. |
1155 | */ | |
1156 | ||
3077a260 | 1157 | static int attach_task(struct cpuset *cs, char *pidbuf, char **ppathbuf) |
1da177e4 LT |
1158 | { |
1159 | pid_t pid; | |
1160 | struct task_struct *tsk; | |
1161 | struct cpuset *oldcs; | |
1162 | cpumask_t cpus; | |
45b07ef3 | 1163 | nodemask_t from, to; |
4225399a | 1164 | struct mm_struct *mm; |
22fb52dd | 1165 | int retval; |
1da177e4 | 1166 | |
3077a260 | 1167 | if (sscanf(pidbuf, "%d", &pid) != 1) |
1da177e4 LT |
1168 | return -EIO; |
1169 | if (cpus_empty(cs->cpus_allowed) || nodes_empty(cs->mems_allowed)) | |
1170 | return -ENOSPC; | |
1171 | ||
1172 | if (pid) { | |
1173 | read_lock(&tasklist_lock); | |
1174 | ||
1175 | tsk = find_task_by_pid(pid); | |
053199ed | 1176 | if (!tsk || tsk->flags & PF_EXITING) { |
1da177e4 LT |
1177 | read_unlock(&tasklist_lock); |
1178 | return -ESRCH; | |
1179 | } | |
1180 | ||
1181 | get_task_struct(tsk); | |
1182 | read_unlock(&tasklist_lock); | |
1183 | ||
1184 | if ((current->euid) && (current->euid != tsk->uid) | |
1185 | && (current->euid != tsk->suid)) { | |
1186 | put_task_struct(tsk); | |
1187 | return -EACCES; | |
1188 | } | |
1189 | } else { | |
1190 | tsk = current; | |
1191 | get_task_struct(tsk); | |
1192 | } | |
1193 | ||
22fb52dd DQ |
1194 | retval = security_task_setscheduler(tsk, 0, NULL); |
1195 | if (retval) { | |
1196 | put_task_struct(tsk); | |
1197 | return retval; | |
1198 | } | |
1199 | ||
3d3f26a7 | 1200 | mutex_lock(&callback_mutex); |
053199ed | 1201 | |
1da177e4 LT |
1202 | task_lock(tsk); |
1203 | oldcs = tsk->cpuset; | |
181b6480 PJ |
1204 | /* |
1205 | * After getting 'oldcs' cpuset ptr, be sure still not exiting. | |
1206 | * If 'oldcs' might be the top_cpuset due to the_top_cpuset_hack | |
1207 | * then fail this attach_task(), to avoid breaking top_cpuset.count. | |
1208 | */ | |
1209 | if (tsk->flags & PF_EXITING) { | |
1da177e4 | 1210 | task_unlock(tsk); |
3d3f26a7 | 1211 | mutex_unlock(&callback_mutex); |
1da177e4 LT |
1212 | put_task_struct(tsk); |
1213 | return -ESRCH; | |
1214 | } | |
1215 | atomic_inc(&cs->count); | |
6b9c2603 | 1216 | rcu_assign_pointer(tsk->cpuset, cs); |
1da177e4 LT |
1217 | task_unlock(tsk); |
1218 | ||
1219 | guarantee_online_cpus(cs, &cpus); | |
1220 | set_cpus_allowed(tsk, cpus); | |
1221 | ||
45b07ef3 PJ |
1222 | from = oldcs->mems_allowed; |
1223 | to = cs->mems_allowed; | |
1224 | ||
3d3f26a7 | 1225 | mutex_unlock(&callback_mutex); |
4225399a PJ |
1226 | |
1227 | mm = get_task_mm(tsk); | |
1228 | if (mm) { | |
1229 | mpol_rebind_mm(mm, &to); | |
2741a559 | 1230 | if (is_memory_migrate(cs)) |
e4e364e8 | 1231 | cpuset_migrate_mm(mm, &from, &to); |
4225399a PJ |
1232 | mmput(mm); |
1233 | } | |
1234 | ||
1da177e4 | 1235 | put_task_struct(tsk); |
6b9c2603 | 1236 | synchronize_rcu(); |
1da177e4 | 1237 | if (atomic_dec_and_test(&oldcs->count)) |
3077a260 | 1238 | check_for_release(oldcs, ppathbuf); |
1da177e4 LT |
1239 | return 0; |
1240 | } | |
1241 | ||
1242 | /* The various types of files and directories in a cpuset file system */ | |
1243 | ||
1244 | typedef enum { | |
1245 | FILE_ROOT, | |
1246 | FILE_DIR, | |
45b07ef3 | 1247 | FILE_MEMORY_MIGRATE, |
1da177e4 LT |
1248 | FILE_CPULIST, |
1249 | FILE_MEMLIST, | |
1250 | FILE_CPU_EXCLUSIVE, | |
1251 | FILE_MEM_EXCLUSIVE, | |
1252 | FILE_NOTIFY_ON_RELEASE, | |
3e0d98b9 PJ |
1253 | FILE_MEMORY_PRESSURE_ENABLED, |
1254 | FILE_MEMORY_PRESSURE, | |
825a46af PJ |
1255 | FILE_SPREAD_PAGE, |
1256 | FILE_SPREAD_SLAB, | |
1da177e4 LT |
1257 | FILE_TASKLIST, |
1258 | } cpuset_filetype_t; | |
1259 | ||
d3ed11c3 PM |
1260 | static ssize_t cpuset_common_file_write(struct file *file, |
1261 | const char __user *userbuf, | |
1da177e4 LT |
1262 | size_t nbytes, loff_t *unused_ppos) |
1263 | { | |
a7a005fd JS |
1264 | struct cpuset *cs = __d_cs(file->f_path.dentry->d_parent); |
1265 | struct cftype *cft = __d_cft(file->f_path.dentry); | |
1da177e4 LT |
1266 | cpuset_filetype_t type = cft->private; |
1267 | char *buffer; | |
3077a260 | 1268 | char *pathbuf = NULL; |
1da177e4 LT |
1269 | int retval = 0; |
1270 | ||
1271 | /* Crude upper limit on largest legitimate cpulist user might write. */ | |
d3ed11c3 | 1272 | if (nbytes > 100 + 6 * max(NR_CPUS, MAX_NUMNODES)) |
1da177e4 LT |
1273 | return -E2BIG; |
1274 | ||
1275 | /* +1 for nul-terminator */ | |
1276 | if ((buffer = kmalloc(nbytes + 1, GFP_KERNEL)) == 0) | |
1277 | return -ENOMEM; | |
1278 | ||
1279 | if (copy_from_user(buffer, userbuf, nbytes)) { | |
1280 | retval = -EFAULT; | |
1281 | goto out1; | |
1282 | } | |
1283 | buffer[nbytes] = 0; /* nul-terminate */ | |
1284 | ||
3d3f26a7 | 1285 | mutex_lock(&manage_mutex); |
1da177e4 LT |
1286 | |
1287 | if (is_removed(cs)) { | |
1288 | retval = -ENODEV; | |
1289 | goto out2; | |
1290 | } | |
1291 | ||
1292 | switch (type) { | |
1293 | case FILE_CPULIST: | |
1294 | retval = update_cpumask(cs, buffer); | |
1295 | break; | |
1296 | case FILE_MEMLIST: | |
1297 | retval = update_nodemask(cs, buffer); | |
1298 | break; | |
1299 | case FILE_CPU_EXCLUSIVE: | |
1300 | retval = update_flag(CS_CPU_EXCLUSIVE, cs, buffer); | |
1301 | break; | |
1302 | case FILE_MEM_EXCLUSIVE: | |
1303 | retval = update_flag(CS_MEM_EXCLUSIVE, cs, buffer); | |
1304 | break; | |
1305 | case FILE_NOTIFY_ON_RELEASE: | |
1306 | retval = update_flag(CS_NOTIFY_ON_RELEASE, cs, buffer); | |
1307 | break; | |
45b07ef3 PJ |
1308 | case FILE_MEMORY_MIGRATE: |
1309 | retval = update_flag(CS_MEMORY_MIGRATE, cs, buffer); | |
1310 | break; | |
3e0d98b9 PJ |
1311 | case FILE_MEMORY_PRESSURE_ENABLED: |
1312 | retval = update_memory_pressure_enabled(cs, buffer); | |
1313 | break; | |
1314 | case FILE_MEMORY_PRESSURE: | |
1315 | retval = -EACCES; | |
1316 | break; | |
825a46af PJ |
1317 | case FILE_SPREAD_PAGE: |
1318 | retval = update_flag(CS_SPREAD_PAGE, cs, buffer); | |
151a4420 | 1319 | cs->mems_generation = cpuset_mems_generation++; |
825a46af PJ |
1320 | break; |
1321 | case FILE_SPREAD_SLAB: | |
1322 | retval = update_flag(CS_SPREAD_SLAB, cs, buffer); | |
151a4420 | 1323 | cs->mems_generation = cpuset_mems_generation++; |
825a46af | 1324 | break; |
1da177e4 | 1325 | case FILE_TASKLIST: |
3077a260 | 1326 | retval = attach_task(cs, buffer, &pathbuf); |
1da177e4 LT |
1327 | break; |
1328 | default: | |
1329 | retval = -EINVAL; | |
1330 | goto out2; | |
1331 | } | |
1332 | ||
1333 | if (retval == 0) | |
1334 | retval = nbytes; | |
1335 | out2: | |
3d3f26a7 | 1336 | mutex_unlock(&manage_mutex); |
3077a260 | 1337 | cpuset_release_agent(pathbuf); |
1da177e4 LT |
1338 | out1: |
1339 | kfree(buffer); | |
1340 | return retval; | |
1341 | } | |
1342 | ||
1343 | static ssize_t cpuset_file_write(struct file *file, const char __user *buf, | |
1344 | size_t nbytes, loff_t *ppos) | |
1345 | { | |
1346 | ssize_t retval = 0; | |
a7a005fd | 1347 | struct cftype *cft = __d_cft(file->f_path.dentry); |
1da177e4 LT |
1348 | if (!cft) |
1349 | return -ENODEV; | |
1350 | ||
1351 | /* special function ? */ | |
1352 | if (cft->write) | |
1353 | retval = cft->write(file, buf, nbytes, ppos); | |
1354 | else | |
1355 | retval = cpuset_common_file_write(file, buf, nbytes, ppos); | |
1356 | ||
1357 | return retval; | |
1358 | } | |
1359 | ||
1360 | /* | |
1361 | * These ascii lists should be read in a single call, by using a user | |
1362 | * buffer large enough to hold the entire map. If read in smaller | |
1363 | * chunks, there is no guarantee of atomicity. Since the display format | |
1364 | * used, list of ranges of sequential numbers, is variable length, | |
1365 | * and since these maps can change value dynamically, one could read | |
1366 | * gibberish by doing partial reads while a list was changing. | |
1367 | * A single large read to a buffer that crosses a page boundary is | |
1368 | * ok, because the result being copied to user land is not recomputed | |
1369 | * across a page fault. | |
1370 | */ | |
1371 | ||
1372 | static int cpuset_sprintf_cpulist(char *page, struct cpuset *cs) | |
1373 | { | |
1374 | cpumask_t mask; | |
1375 | ||
3d3f26a7 | 1376 | mutex_lock(&callback_mutex); |
1da177e4 | 1377 | mask = cs->cpus_allowed; |
3d3f26a7 | 1378 | mutex_unlock(&callback_mutex); |
1da177e4 LT |
1379 | |
1380 | return cpulist_scnprintf(page, PAGE_SIZE, mask); | |
1381 | } | |
1382 | ||
1383 | static int cpuset_sprintf_memlist(char *page, struct cpuset *cs) | |
1384 | { | |
1385 | nodemask_t mask; | |
1386 | ||
3d3f26a7 | 1387 | mutex_lock(&callback_mutex); |
1da177e4 | 1388 | mask = cs->mems_allowed; |
3d3f26a7 | 1389 | mutex_unlock(&callback_mutex); |
1da177e4 LT |
1390 | |
1391 | return nodelist_scnprintf(page, PAGE_SIZE, mask); | |
1392 | } | |
1393 | ||
1394 | static ssize_t cpuset_common_file_read(struct file *file, char __user *buf, | |
1395 | size_t nbytes, loff_t *ppos) | |
1396 | { | |
a7a005fd JS |
1397 | struct cftype *cft = __d_cft(file->f_path.dentry); |
1398 | struct cpuset *cs = __d_cs(file->f_path.dentry->d_parent); | |
1da177e4 LT |
1399 | cpuset_filetype_t type = cft->private; |
1400 | char *page; | |
1401 | ssize_t retval = 0; | |
1402 | char *s; | |
1da177e4 | 1403 | |
e12ba74d | 1404 | if (!(page = (char *)__get_free_page(GFP_TEMPORARY))) |
1da177e4 LT |
1405 | return -ENOMEM; |
1406 | ||
1407 | s = page; | |
1408 | ||
1409 | switch (type) { | |
1410 | case FILE_CPULIST: | |
1411 | s += cpuset_sprintf_cpulist(s, cs); | |
1412 | break; | |
1413 | case FILE_MEMLIST: | |
1414 | s += cpuset_sprintf_memlist(s, cs); | |
1415 | break; | |
1416 | case FILE_CPU_EXCLUSIVE: | |
1417 | *s++ = is_cpu_exclusive(cs) ? '1' : '0'; | |
1418 | break; | |
1419 | case FILE_MEM_EXCLUSIVE: | |
1420 | *s++ = is_mem_exclusive(cs) ? '1' : '0'; | |
1421 | break; | |
1422 | case FILE_NOTIFY_ON_RELEASE: | |
1423 | *s++ = notify_on_release(cs) ? '1' : '0'; | |
1424 | break; | |
45b07ef3 PJ |
1425 | case FILE_MEMORY_MIGRATE: |
1426 | *s++ = is_memory_migrate(cs) ? '1' : '0'; | |
1427 | break; | |
3e0d98b9 PJ |
1428 | case FILE_MEMORY_PRESSURE_ENABLED: |
1429 | *s++ = cpuset_memory_pressure_enabled ? '1' : '0'; | |
1430 | break; | |
1431 | case FILE_MEMORY_PRESSURE: | |
1432 | s += sprintf(s, "%d", fmeter_getrate(&cs->fmeter)); | |
1433 | break; | |
825a46af PJ |
1434 | case FILE_SPREAD_PAGE: |
1435 | *s++ = is_spread_page(cs) ? '1' : '0'; | |
1436 | break; | |
1437 | case FILE_SPREAD_SLAB: | |
1438 | *s++ = is_spread_slab(cs) ? '1' : '0'; | |
1439 | break; | |
1da177e4 LT |
1440 | default: |
1441 | retval = -EINVAL; | |
1442 | goto out; | |
1443 | } | |
1444 | *s++ = '\n'; | |
1da177e4 | 1445 | |
eacaa1f5 | 1446 | retval = simple_read_from_buffer(buf, nbytes, ppos, page, s - page); |
1da177e4 LT |
1447 | out: |
1448 | free_page((unsigned long)page); | |
1449 | return retval; | |
1450 | } | |
1451 | ||
1452 | static ssize_t cpuset_file_read(struct file *file, char __user *buf, size_t nbytes, | |
1453 | loff_t *ppos) | |
1454 | { | |
1455 | ssize_t retval = 0; | |
a7a005fd | 1456 | struct cftype *cft = __d_cft(file->f_path.dentry); |
1da177e4 LT |
1457 | if (!cft) |
1458 | return -ENODEV; | |
1459 | ||
1460 | /* special function ? */ | |
1461 | if (cft->read) | |
1462 | retval = cft->read(file, buf, nbytes, ppos); | |
1463 | else | |
1464 | retval = cpuset_common_file_read(file, buf, nbytes, ppos); | |
1465 | ||
1466 | return retval; | |
1467 | } | |
1468 | ||
1469 | static int cpuset_file_open(struct inode *inode, struct file *file) | |
1470 | { | |
1471 | int err; | |
1472 | struct cftype *cft; | |
1473 | ||
1474 | err = generic_file_open(inode, file); | |
1475 | if (err) | |
1476 | return err; | |
1477 | ||
a7a005fd | 1478 | cft = __d_cft(file->f_path.dentry); |
1da177e4 LT |
1479 | if (!cft) |
1480 | return -ENODEV; | |
1481 | if (cft->open) | |
1482 | err = cft->open(inode, file); | |
1483 | else | |
1484 | err = 0; | |
1485 | ||
1486 | return err; | |
1487 | } | |
1488 | ||
1489 | static int cpuset_file_release(struct inode *inode, struct file *file) | |
1490 | { | |
a7a005fd | 1491 | struct cftype *cft = __d_cft(file->f_path.dentry); |
1da177e4 LT |
1492 | if (cft->release) |
1493 | return cft->release(inode, file); | |
1494 | return 0; | |
1495 | } | |
1496 | ||
18a19cb3 PJ |
1497 | /* |
1498 | * cpuset_rename - Only allow simple rename of directories in place. | |
1499 | */ | |
1500 | static int cpuset_rename(struct inode *old_dir, struct dentry *old_dentry, | |
1501 | struct inode *new_dir, struct dentry *new_dentry) | |
1502 | { | |
1503 | if (!S_ISDIR(old_dentry->d_inode->i_mode)) | |
1504 | return -ENOTDIR; | |
1505 | if (new_dentry->d_inode) | |
1506 | return -EEXIST; | |
1507 | if (old_dir != new_dir) | |
1508 | return -EIO; | |
1509 | return simple_rename(old_dir, old_dentry, new_dir, new_dentry); | |
1510 | } | |
1511 | ||
15ad7cdc | 1512 | static const struct file_operations cpuset_file_operations = { |
1da177e4 LT |
1513 | .read = cpuset_file_read, |
1514 | .write = cpuset_file_write, | |
1515 | .llseek = generic_file_llseek, | |
1516 | .open = cpuset_file_open, | |
1517 | .release = cpuset_file_release, | |
1518 | }; | |
1519 | ||
92e1d5be | 1520 | static const struct inode_operations cpuset_dir_inode_operations = { |
1da177e4 LT |
1521 | .lookup = simple_lookup, |
1522 | .mkdir = cpuset_mkdir, | |
1523 | .rmdir = cpuset_rmdir, | |
18a19cb3 | 1524 | .rename = cpuset_rename, |
1da177e4 LT |
1525 | }; |
1526 | ||
1527 | static int cpuset_create_file(struct dentry *dentry, int mode) | |
1528 | { | |
1529 | struct inode *inode; | |
1530 | ||
1531 | if (!dentry) | |
1532 | return -ENOENT; | |
1533 | if (dentry->d_inode) | |
1534 | return -EEXIST; | |
1535 | ||
1536 | inode = cpuset_new_inode(mode); | |
1537 | if (!inode) | |
1538 | return -ENOMEM; | |
1539 | ||
1540 | if (S_ISDIR(mode)) { | |
1541 | inode->i_op = &cpuset_dir_inode_operations; | |
1542 | inode->i_fop = &simple_dir_operations; | |
1543 | ||
1544 | /* start off with i_nlink == 2 (for "." entry) */ | |
d8c76e6f | 1545 | inc_nlink(inode); |
1da177e4 LT |
1546 | } else if (S_ISREG(mode)) { |
1547 | inode->i_size = 0; | |
1548 | inode->i_fop = &cpuset_file_operations; | |
1549 | } | |
1550 | ||
1551 | d_instantiate(dentry, inode); | |
1552 | dget(dentry); /* Extra count - pin the dentry in core */ | |
1553 | return 0; | |
1554 | } | |
1555 | ||
1556 | /* | |
1557 | * cpuset_create_dir - create a directory for an object. | |
c5b2aff8 | 1558 | * cs: the cpuset we create the directory for. |
1da177e4 LT |
1559 | * It must have a valid ->parent field |
1560 | * And we are going to fill its ->dentry field. | |
1561 | * name: The name to give to the cpuset directory. Will be copied. | |
1562 | * mode: mode to set on new directory. | |
1563 | */ | |
1564 | ||
1565 | static int cpuset_create_dir(struct cpuset *cs, const char *name, int mode) | |
1566 | { | |
1567 | struct dentry *dentry = NULL; | |
1568 | struct dentry *parent; | |
1569 | int error = 0; | |
1570 | ||
1571 | parent = cs->parent->dentry; | |
1572 | dentry = cpuset_get_dentry(parent, name); | |
1573 | if (IS_ERR(dentry)) | |
1574 | return PTR_ERR(dentry); | |
1575 | error = cpuset_create_file(dentry, S_IFDIR | mode); | |
1576 | if (!error) { | |
1577 | dentry->d_fsdata = cs; | |
d8c76e6f | 1578 | inc_nlink(parent->d_inode); |
1da177e4 LT |
1579 | cs->dentry = dentry; |
1580 | } | |
1581 | dput(dentry); | |
1582 | ||
1583 | return error; | |
1584 | } | |
1585 | ||
1586 | static int cpuset_add_file(struct dentry *dir, const struct cftype *cft) | |
1587 | { | |
1588 | struct dentry *dentry; | |
1589 | int error; | |
1590 | ||
1b1dcc1b | 1591 | mutex_lock(&dir->d_inode->i_mutex); |
1da177e4 LT |
1592 | dentry = cpuset_get_dentry(dir, cft->name); |
1593 | if (!IS_ERR(dentry)) { | |
1594 | error = cpuset_create_file(dentry, 0644 | S_IFREG); | |
1595 | if (!error) | |
1596 | dentry->d_fsdata = (void *)cft; | |
1597 | dput(dentry); | |
1598 | } else | |
1599 | error = PTR_ERR(dentry); | |
1b1dcc1b | 1600 | mutex_unlock(&dir->d_inode->i_mutex); |
1da177e4 LT |
1601 | return error; |
1602 | } | |
1603 | ||
1604 | /* | |
1605 | * Stuff for reading the 'tasks' file. | |
1606 | * | |
1607 | * Reading this file can return large amounts of data if a cpuset has | |
1608 | * *lots* of attached tasks. So it may need several calls to read(), | |
1609 | * but we cannot guarantee that the information we produce is correct | |
1610 | * unless we produce it entirely atomically. | |
1611 | * | |
1612 | * Upon tasks file open(), a struct ctr_struct is allocated, that | |
1613 | * will have a pointer to an array (also allocated here). The struct | |
1614 | * ctr_struct * is stored in file->private_data. Its resources will | |
1615 | * be freed by release() when the file is closed. The array is used | |
1616 | * to sprintf the PIDs and then used by read(). | |
1617 | */ | |
1618 | ||
1619 | /* cpusets_tasks_read array */ | |
1620 | ||
1621 | struct ctr_struct { | |
1622 | char *buf; | |
1623 | int bufsz; | |
1624 | }; | |
1625 | ||
1626 | /* | |
1627 | * Load into 'pidarray' up to 'npids' of the tasks using cpuset 'cs'. | |
053199ed PJ |
1628 | * Return actual number of pids loaded. No need to task_lock(p) |
1629 | * when reading out p->cpuset, as we don't really care if it changes | |
1630 | * on the next cycle, and we are not going to try to dereference it. | |
1da177e4 | 1631 | */ |
858119e1 | 1632 | static int pid_array_load(pid_t *pidarray, int npids, struct cpuset *cs) |
1da177e4 LT |
1633 | { |
1634 | int n = 0; | |
1635 | struct task_struct *g, *p; | |
1636 | ||
1637 | read_lock(&tasklist_lock); | |
1638 | ||
1639 | do_each_thread(g, p) { | |
1640 | if (p->cpuset == cs) { | |
1da177e4 LT |
1641 | if (unlikely(n == npids)) |
1642 | goto array_full; | |
3e903e7b | 1643 | pidarray[n++] = p->pid; |
1da177e4 LT |
1644 | } |
1645 | } while_each_thread(g, p); | |
1646 | ||
1647 | array_full: | |
1648 | read_unlock(&tasklist_lock); | |
1649 | return n; | |
1650 | } | |
1651 | ||
1652 | static int cmppid(const void *a, const void *b) | |
1653 | { | |
1654 | return *(pid_t *)a - *(pid_t *)b; | |
1655 | } | |
1656 | ||
1657 | /* | |
1658 | * Convert array 'a' of 'npids' pid_t's to a string of newline separated | |
1659 | * decimal pids in 'buf'. Don't write more than 'sz' chars, but return | |
1660 | * count 'cnt' of how many chars would be written if buf were large enough. | |
1661 | */ | |
1662 | static int pid_array_to_buf(char *buf, int sz, pid_t *a, int npids) | |
1663 | { | |
1664 | int cnt = 0; | |
1665 | int i; | |
1666 | ||
1667 | for (i = 0; i < npids; i++) | |
1668 | cnt += snprintf(buf + cnt, max(sz - cnt, 0), "%d\n", a[i]); | |
1669 | return cnt; | |
1670 | } | |
1671 | ||
053199ed PJ |
1672 | /* |
1673 | * Handle an open on 'tasks' file. Prepare a buffer listing the | |
1674 | * process id's of tasks currently attached to the cpuset being opened. | |
1675 | * | |
3d3f26a7 | 1676 | * Does not require any specific cpuset mutexes, and does not take any. |
053199ed | 1677 | */ |
1da177e4 LT |
1678 | static int cpuset_tasks_open(struct inode *unused, struct file *file) |
1679 | { | |
a7a005fd | 1680 | struct cpuset *cs = __d_cs(file->f_path.dentry->d_parent); |
1da177e4 LT |
1681 | struct ctr_struct *ctr; |
1682 | pid_t *pidarray; | |
1683 | int npids; | |
1684 | char c; | |
1685 | ||
1686 | if (!(file->f_mode & FMODE_READ)) | |
1687 | return 0; | |
1688 | ||
1689 | ctr = kmalloc(sizeof(*ctr), GFP_KERNEL); | |
1690 | if (!ctr) | |
1691 | goto err0; | |
1692 | ||
1693 | /* | |
1694 | * If cpuset gets more users after we read count, we won't have | |
1695 | * enough space - tough. This race is indistinguishable to the | |
1696 | * caller from the case that the additional cpuset users didn't | |
1697 | * show up until sometime later on. | |
1698 | */ | |
1699 | npids = atomic_read(&cs->count); | |
1700 | pidarray = kmalloc(npids * sizeof(pid_t), GFP_KERNEL); | |
1701 | if (!pidarray) | |
1702 | goto err1; | |
1703 | ||
1704 | npids = pid_array_load(pidarray, npids, cs); | |
1705 | sort(pidarray, npids, sizeof(pid_t), cmppid, NULL); | |
1706 | ||
1707 | /* Call pid_array_to_buf() twice, first just to get bufsz */ | |
1708 | ctr->bufsz = pid_array_to_buf(&c, sizeof(c), pidarray, npids) + 1; | |
1709 | ctr->buf = kmalloc(ctr->bufsz, GFP_KERNEL); | |
1710 | if (!ctr->buf) | |
1711 | goto err2; | |
1712 | ctr->bufsz = pid_array_to_buf(ctr->buf, ctr->bufsz, pidarray, npids); | |
1713 | ||
1714 | kfree(pidarray); | |
1715 | file->private_data = ctr; | |
1716 | return 0; | |
1717 | ||
1718 | err2: | |
1719 | kfree(pidarray); | |
1720 | err1: | |
1721 | kfree(ctr); | |
1722 | err0: | |
1723 | return -ENOMEM; | |
1724 | } | |
1725 | ||
1726 | static ssize_t cpuset_tasks_read(struct file *file, char __user *buf, | |
1727 | size_t nbytes, loff_t *ppos) | |
1728 | { | |
1729 | struct ctr_struct *ctr = file->private_data; | |
1730 | ||
85badbdf | 1731 | return simple_read_from_buffer(buf, nbytes, ppos, ctr->buf, ctr->bufsz); |
1da177e4 LT |
1732 | } |
1733 | ||
1734 | static int cpuset_tasks_release(struct inode *unused_inode, struct file *file) | |
1735 | { | |
1736 | struct ctr_struct *ctr; | |
1737 | ||
1738 | if (file->f_mode & FMODE_READ) { | |
1739 | ctr = file->private_data; | |
1740 | kfree(ctr->buf); | |
1741 | kfree(ctr); | |
1742 | } | |
1743 | return 0; | |
1744 | } | |
1745 | ||
1746 | /* | |
1747 | * for the common functions, 'private' gives the type of file | |
1748 | */ | |
1749 | ||
1750 | static struct cftype cft_tasks = { | |
1751 | .name = "tasks", | |
1752 | .open = cpuset_tasks_open, | |
1753 | .read = cpuset_tasks_read, | |
1754 | .release = cpuset_tasks_release, | |
1755 | .private = FILE_TASKLIST, | |
1756 | }; | |
1757 | ||
1758 | static struct cftype cft_cpus = { | |
1759 | .name = "cpus", | |
1760 | .private = FILE_CPULIST, | |
1761 | }; | |
1762 | ||
1763 | static struct cftype cft_mems = { | |
1764 | .name = "mems", | |
1765 | .private = FILE_MEMLIST, | |
1766 | }; | |
1767 | ||
1768 | static struct cftype cft_cpu_exclusive = { | |
1769 | .name = "cpu_exclusive", | |
1770 | .private = FILE_CPU_EXCLUSIVE, | |
1771 | }; | |
1772 | ||
1773 | static struct cftype cft_mem_exclusive = { | |
1774 | .name = "mem_exclusive", | |
1775 | .private = FILE_MEM_EXCLUSIVE, | |
1776 | }; | |
1777 | ||
1778 | static struct cftype cft_notify_on_release = { | |
1779 | .name = "notify_on_release", | |
1780 | .private = FILE_NOTIFY_ON_RELEASE, | |
1781 | }; | |
1782 | ||
45b07ef3 PJ |
1783 | static struct cftype cft_memory_migrate = { |
1784 | .name = "memory_migrate", | |
1785 | .private = FILE_MEMORY_MIGRATE, | |
1786 | }; | |
1787 | ||
3e0d98b9 PJ |
1788 | static struct cftype cft_memory_pressure_enabled = { |
1789 | .name = "memory_pressure_enabled", | |
1790 | .private = FILE_MEMORY_PRESSURE_ENABLED, | |
1791 | }; | |
1792 | ||
1793 | static struct cftype cft_memory_pressure = { | |
1794 | .name = "memory_pressure", | |
1795 | .private = FILE_MEMORY_PRESSURE, | |
1796 | }; | |
1797 | ||
825a46af PJ |
1798 | static struct cftype cft_spread_page = { |
1799 | .name = "memory_spread_page", | |
1800 | .private = FILE_SPREAD_PAGE, | |
1801 | }; | |
1802 | ||
1803 | static struct cftype cft_spread_slab = { | |
1804 | .name = "memory_spread_slab", | |
1805 | .private = FILE_SPREAD_SLAB, | |
1806 | }; | |
1807 | ||
1da177e4 LT |
1808 | static int cpuset_populate_dir(struct dentry *cs_dentry) |
1809 | { | |
1810 | int err; | |
1811 | ||
1812 | if ((err = cpuset_add_file(cs_dentry, &cft_cpus)) < 0) | |
1813 | return err; | |
1814 | if ((err = cpuset_add_file(cs_dentry, &cft_mems)) < 0) | |
1815 | return err; | |
1816 | if ((err = cpuset_add_file(cs_dentry, &cft_cpu_exclusive)) < 0) | |
1817 | return err; | |
1818 | if ((err = cpuset_add_file(cs_dentry, &cft_mem_exclusive)) < 0) | |
1819 | return err; | |
1820 | if ((err = cpuset_add_file(cs_dentry, &cft_notify_on_release)) < 0) | |
1821 | return err; | |
45b07ef3 PJ |
1822 | if ((err = cpuset_add_file(cs_dentry, &cft_memory_migrate)) < 0) |
1823 | return err; | |
3e0d98b9 PJ |
1824 | if ((err = cpuset_add_file(cs_dentry, &cft_memory_pressure)) < 0) |
1825 | return err; | |
825a46af PJ |
1826 | if ((err = cpuset_add_file(cs_dentry, &cft_spread_page)) < 0) |
1827 | return err; | |
1828 | if ((err = cpuset_add_file(cs_dentry, &cft_spread_slab)) < 0) | |
1829 | return err; | |
1da177e4 LT |
1830 | if ((err = cpuset_add_file(cs_dentry, &cft_tasks)) < 0) |
1831 | return err; | |
1832 | return 0; | |
1833 | } | |
1834 | ||
1835 | /* | |
1836 | * cpuset_create - create a cpuset | |
1837 | * parent: cpuset that will be parent of the new cpuset. | |
1838 | * name: name of the new cpuset. Will be strcpy'ed. | |
1839 | * mode: mode to set on new inode | |
1840 | * | |
3d3f26a7 | 1841 | * Must be called with the mutex on the parent inode held |
1da177e4 LT |
1842 | */ |
1843 | ||
1844 | static long cpuset_create(struct cpuset *parent, const char *name, int mode) | |
1845 | { | |
1846 | struct cpuset *cs; | |
1847 | int err; | |
1848 | ||
1849 | cs = kmalloc(sizeof(*cs), GFP_KERNEL); | |
1850 | if (!cs) | |
1851 | return -ENOMEM; | |
1852 | ||
3d3f26a7 | 1853 | mutex_lock(&manage_mutex); |
cf2a473c | 1854 | cpuset_update_task_memory_state(); |
1da177e4 LT |
1855 | cs->flags = 0; |
1856 | if (notify_on_release(parent)) | |
1857 | set_bit(CS_NOTIFY_ON_RELEASE, &cs->flags); | |
825a46af PJ |
1858 | if (is_spread_page(parent)) |
1859 | set_bit(CS_SPREAD_PAGE, &cs->flags); | |
1860 | if (is_spread_slab(parent)) | |
1861 | set_bit(CS_SPREAD_SLAB, &cs->flags); | |
1da177e4 LT |
1862 | cs->cpus_allowed = CPU_MASK_NONE; |
1863 | cs->mems_allowed = NODE_MASK_NONE; | |
1864 | atomic_set(&cs->count, 0); | |
1865 | INIT_LIST_HEAD(&cs->sibling); | |
1866 | INIT_LIST_HEAD(&cs->children); | |
151a4420 | 1867 | cs->mems_generation = cpuset_mems_generation++; |
3e0d98b9 | 1868 | fmeter_init(&cs->fmeter); |
1da177e4 LT |
1869 | |
1870 | cs->parent = parent; | |
1871 | ||
3d3f26a7 | 1872 | mutex_lock(&callback_mutex); |
1da177e4 | 1873 | list_add(&cs->sibling, &cs->parent->children); |
202f72d5 | 1874 | number_of_cpusets++; |
3d3f26a7 | 1875 | mutex_unlock(&callback_mutex); |
1da177e4 LT |
1876 | |
1877 | err = cpuset_create_dir(cs, name, mode); | |
1878 | if (err < 0) | |
1879 | goto err; | |
1880 | ||
1881 | /* | |
3d3f26a7 | 1882 | * Release manage_mutex before cpuset_populate_dir() because it |
1b1dcc1b | 1883 | * will down() this new directory's i_mutex and if we race with |
1da177e4 LT |
1884 | * another mkdir, we might deadlock. |
1885 | */ | |
3d3f26a7 | 1886 | mutex_unlock(&manage_mutex); |
1da177e4 LT |
1887 | |
1888 | err = cpuset_populate_dir(cs->dentry); | |
1889 | /* If err < 0, we have a half-filled directory - oh well ;) */ | |
1890 | return 0; | |
1891 | err: | |
1892 | list_del(&cs->sibling); | |
3d3f26a7 | 1893 | mutex_unlock(&manage_mutex); |
1da177e4 LT |
1894 | kfree(cs); |
1895 | return err; | |
1896 | } | |
1897 | ||
1898 | static int cpuset_mkdir(struct inode *dir, struct dentry *dentry, int mode) | |
1899 | { | |
1900 | struct cpuset *c_parent = dentry->d_parent->d_fsdata; | |
1901 | ||
1b1dcc1b | 1902 | /* the vfs holds inode->i_mutex already */ |
1da177e4 LT |
1903 | return cpuset_create(c_parent, dentry->d_name.name, mode | S_IFDIR); |
1904 | } | |
1905 | ||
1906 | static int cpuset_rmdir(struct inode *unused_dir, struct dentry *dentry) | |
1907 | { | |
1908 | struct cpuset *cs = dentry->d_fsdata; | |
1909 | struct dentry *d; | |
1910 | struct cpuset *parent; | |
3077a260 | 1911 | char *pathbuf = NULL; |
1da177e4 | 1912 | |
1b1dcc1b | 1913 | /* the vfs holds both inode->i_mutex already */ |
1da177e4 | 1914 | |
3d3f26a7 | 1915 | mutex_lock(&manage_mutex); |
cf2a473c | 1916 | cpuset_update_task_memory_state(); |
1da177e4 | 1917 | if (atomic_read(&cs->count) > 0) { |
3d3f26a7 | 1918 | mutex_unlock(&manage_mutex); |
1da177e4 LT |
1919 | return -EBUSY; |
1920 | } | |
1921 | if (!list_empty(&cs->children)) { | |
3d3f26a7 | 1922 | mutex_unlock(&manage_mutex); |
1da177e4 LT |
1923 | return -EBUSY; |
1924 | } | |
1da177e4 | 1925 | parent = cs->parent; |
3d3f26a7 | 1926 | mutex_lock(&callback_mutex); |
1da177e4 LT |
1927 | set_bit(CS_REMOVED, &cs->flags); |
1928 | list_del(&cs->sibling); /* delete my sibling from parent->children */ | |
85d7b949 | 1929 | spin_lock(&cs->dentry->d_lock); |
1da177e4 LT |
1930 | d = dget(cs->dentry); |
1931 | cs->dentry = NULL; | |
1932 | spin_unlock(&d->d_lock); | |
1933 | cpuset_d_remove_dir(d); | |
1934 | dput(d); | |
202f72d5 | 1935 | number_of_cpusets--; |
3d3f26a7 | 1936 | mutex_unlock(&callback_mutex); |
053199ed PJ |
1937 | if (list_empty(&parent->children)) |
1938 | check_for_release(parent, &pathbuf); | |
3d3f26a7 | 1939 | mutex_unlock(&manage_mutex); |
3077a260 | 1940 | cpuset_release_agent(pathbuf); |
1da177e4 LT |
1941 | return 0; |
1942 | } | |
1943 | ||
c417f024 PJ |
1944 | /* |
1945 | * cpuset_init_early - just enough so that the calls to | |
1946 | * cpuset_update_task_memory_state() in early init code | |
1947 | * are harmless. | |
1948 | */ | |
1949 | ||
1950 | int __init cpuset_init_early(void) | |
1951 | { | |
1952 | struct task_struct *tsk = current; | |
1953 | ||
1954 | tsk->cpuset = &top_cpuset; | |
151a4420 | 1955 | tsk->cpuset->mems_generation = cpuset_mems_generation++; |
c417f024 PJ |
1956 | return 0; |
1957 | } | |
1958 | ||
1da177e4 LT |
1959 | /** |
1960 | * cpuset_init - initialize cpusets at system boot | |
1961 | * | |
1962 | * Description: Initialize top_cpuset and the cpuset internal file system, | |
1963 | **/ | |
1964 | ||
1965 | int __init cpuset_init(void) | |
1966 | { | |
1967 | struct dentry *root; | |
1968 | int err; | |
1969 | ||
1970 | top_cpuset.cpus_allowed = CPU_MASK_ALL; | |
1971 | top_cpuset.mems_allowed = NODE_MASK_ALL; | |
1972 | ||
3e0d98b9 | 1973 | fmeter_init(&top_cpuset.fmeter); |
151a4420 | 1974 | top_cpuset.mems_generation = cpuset_mems_generation++; |
1da177e4 LT |
1975 | |
1976 | init_task.cpuset = &top_cpuset; | |
1977 | ||
1978 | err = register_filesystem(&cpuset_fs_type); | |
1979 | if (err < 0) | |
1980 | goto out; | |
1981 | cpuset_mount = kern_mount(&cpuset_fs_type); | |
1982 | if (IS_ERR(cpuset_mount)) { | |
1983 | printk(KERN_ERR "cpuset: could not mount!\n"); | |
1984 | err = PTR_ERR(cpuset_mount); | |
1985 | cpuset_mount = NULL; | |
1986 | goto out; | |
1987 | } | |
1988 | root = cpuset_mount->mnt_sb->s_root; | |
1989 | root->d_fsdata = &top_cpuset; | |
d8c76e6f | 1990 | inc_nlink(root->d_inode); |
1da177e4 LT |
1991 | top_cpuset.dentry = root; |
1992 | root->d_inode->i_op = &cpuset_dir_inode_operations; | |
202f72d5 | 1993 | number_of_cpusets = 1; |
1da177e4 | 1994 | err = cpuset_populate_dir(root); |
3e0d98b9 PJ |
1995 | /* memory_pressure_enabled is in root cpuset only */ |
1996 | if (err == 0) | |
1997 | err = cpuset_add_file(root, &cft_memory_pressure_enabled); | |
1da177e4 LT |
1998 | out: |
1999 | return err; | |
2000 | } | |
2001 | ||
b1aac8bb PJ |
2002 | /* |
2003 | * If common_cpu_mem_hotplug_unplug(), below, unplugs any CPUs | |
2004 | * or memory nodes, we need to walk over the cpuset hierarchy, | |
2005 | * removing that CPU or node from all cpusets. If this removes the | |
2006 | * last CPU or node from a cpuset, then the guarantee_online_cpus() | |
2007 | * or guarantee_online_mems() code will use that emptied cpusets | |
2008 | * parent online CPUs or nodes. Cpusets that were already empty of | |
2009 | * CPUs or nodes are left empty. | |
2010 | * | |
2011 | * This routine is intentionally inefficient in a couple of regards. | |
2012 | * It will check all cpusets in a subtree even if the top cpuset of | |
2013 | * the subtree has no offline CPUs or nodes. It checks both CPUs and | |
2014 | * nodes, even though the caller could have been coded to know that | |
2015 | * only one of CPUs or nodes needed to be checked on a given call. | |
2016 | * This was done to minimize text size rather than cpu cycles. | |
2017 | * | |
2018 | * Call with both manage_mutex and callback_mutex held. | |
2019 | * | |
2020 | * Recursive, on depth of cpuset subtree. | |
2021 | */ | |
2022 | ||
2023 | static void guarantee_online_cpus_mems_in_subtree(const struct cpuset *cur) | |
2024 | { | |
2025 | struct cpuset *c; | |
2026 | ||
2027 | /* Each of our child cpusets mems must be online */ | |
2028 | list_for_each_entry(c, &cur->children, sibling) { | |
2029 | guarantee_online_cpus_mems_in_subtree(c); | |
2030 | if (!cpus_empty(c->cpus_allowed)) | |
2031 | guarantee_online_cpus(c, &c->cpus_allowed); | |
2032 | if (!nodes_empty(c->mems_allowed)) | |
2033 | guarantee_online_mems(c, &c->mems_allowed); | |
2034 | } | |
2035 | } | |
2036 | ||
2037 | /* | |
2038 | * The cpus_allowed and mems_allowed nodemasks in the top_cpuset track | |
0e1e7c7a CL |
2039 | * cpu_online_map and node_states[N_HIGH_MEMORY]. Force the top cpuset to |
2040 | * track what's online after any CPU or memory node hotplug or unplug | |
2041 | * event. | |
b1aac8bb PJ |
2042 | * |
2043 | * To ensure that we don't remove a CPU or node from the top cpuset | |
2044 | * that is currently in use by a child cpuset (which would violate | |
2045 | * the rule that cpusets must be subsets of their parent), we first | |
2046 | * call the recursive routine guarantee_online_cpus_mems_in_subtree(). | |
2047 | * | |
2048 | * Since there are two callers of this routine, one for CPU hotplug | |
2049 | * events and one for memory node hotplug events, we could have coded | |
2050 | * two separate routines here. We code it as a single common routine | |
2051 | * in order to minimize text size. | |
2052 | */ | |
2053 | ||
2054 | static void common_cpu_mem_hotplug_unplug(void) | |
2055 | { | |
2056 | mutex_lock(&manage_mutex); | |
2057 | mutex_lock(&callback_mutex); | |
2058 | ||
2059 | guarantee_online_cpus_mems_in_subtree(&top_cpuset); | |
2060 | top_cpuset.cpus_allowed = cpu_online_map; | |
0e1e7c7a | 2061 | top_cpuset.mems_allowed = node_states[N_HIGH_MEMORY]; |
b1aac8bb PJ |
2062 | |
2063 | mutex_unlock(&callback_mutex); | |
2064 | mutex_unlock(&manage_mutex); | |
2065 | } | |
b1aac8bb | 2066 | |
4c4d50f7 PJ |
2067 | /* |
2068 | * The top_cpuset tracks what CPUs and Memory Nodes are online, | |
2069 | * period. This is necessary in order to make cpusets transparent | |
2070 | * (of no affect) on systems that are actively using CPU hotplug | |
2071 | * but making no active use of cpusets. | |
2072 | * | |
38837fc7 PJ |
2073 | * This routine ensures that top_cpuset.cpus_allowed tracks |
2074 | * cpu_online_map on each CPU hotplug (cpuhp) event. | |
4c4d50f7 PJ |
2075 | */ |
2076 | ||
4c4d50f7 PJ |
2077 | static int cpuset_handle_cpuhp(struct notifier_block *nb, |
2078 | unsigned long phase, void *cpu) | |
2079 | { | |
ac076758 AK |
2080 | if (phase == CPU_DYING || phase == CPU_DYING_FROZEN) |
2081 | return NOTIFY_DONE; | |
2082 | ||
b1aac8bb | 2083 | common_cpu_mem_hotplug_unplug(); |
4c4d50f7 PJ |
2084 | return 0; |
2085 | } | |
4c4d50f7 | 2086 | |
b1aac8bb | 2087 | #ifdef CONFIG_MEMORY_HOTPLUG |
38837fc7 | 2088 | /* |
0e1e7c7a CL |
2089 | * Keep top_cpuset.mems_allowed tracking node_states[N_HIGH_MEMORY]. |
2090 | * Call this routine anytime after you change | |
2091 | * node_states[N_HIGH_MEMORY]. | |
38837fc7 PJ |
2092 | * See also the previous routine cpuset_handle_cpuhp(). |
2093 | */ | |
2094 | ||
1af98928 | 2095 | void cpuset_track_online_nodes(void) |
38837fc7 | 2096 | { |
b1aac8bb | 2097 | common_cpu_mem_hotplug_unplug(); |
38837fc7 PJ |
2098 | } |
2099 | #endif | |
2100 | ||
1da177e4 LT |
2101 | /** |
2102 | * cpuset_init_smp - initialize cpus_allowed | |
2103 | * | |
2104 | * Description: Finish top cpuset after cpu, node maps are initialized | |
2105 | **/ | |
2106 | ||
2107 | void __init cpuset_init_smp(void) | |
2108 | { | |
2109 | top_cpuset.cpus_allowed = cpu_online_map; | |
0e1e7c7a | 2110 | top_cpuset.mems_allowed = node_states[N_HIGH_MEMORY]; |
4c4d50f7 PJ |
2111 | |
2112 | hotcpu_notifier(cpuset_handle_cpuhp, 0); | |
1da177e4 LT |
2113 | } |
2114 | ||
2115 | /** | |
2116 | * cpuset_fork - attach newly forked task to its parents cpuset. | |
d9fd8a6d | 2117 | * @tsk: pointer to task_struct of forking parent process. |
1da177e4 | 2118 | * |
053199ed PJ |
2119 | * Description: A task inherits its parent's cpuset at fork(). |
2120 | * | |
2121 | * A pointer to the shared cpuset was automatically copied in fork.c | |
2122 | * by dup_task_struct(). However, we ignore that copy, since it was | |
2123 | * not made under the protection of task_lock(), so might no longer be | |
2124 | * a valid cpuset pointer. attach_task() might have already changed | |
2125 | * current->cpuset, allowing the previously referenced cpuset to | |
2126 | * be removed and freed. Instead, we task_lock(current) and copy | |
2127 | * its present value of current->cpuset for our freshly forked child. | |
2128 | * | |
2129 | * At the point that cpuset_fork() is called, 'current' is the parent | |
2130 | * task, and the passed argument 'child' points to the child task. | |
1da177e4 LT |
2131 | **/ |
2132 | ||
053199ed | 2133 | void cpuset_fork(struct task_struct *child) |
1da177e4 | 2134 | { |
053199ed PJ |
2135 | task_lock(current); |
2136 | child->cpuset = current->cpuset; | |
2137 | atomic_inc(&child->cpuset->count); | |
2138 | task_unlock(current); | |
1da177e4 LT |
2139 | } |
2140 | ||
2141 | /** | |
2142 | * cpuset_exit - detach cpuset from exiting task | |
2143 | * @tsk: pointer to task_struct of exiting process | |
2144 | * | |
2145 | * Description: Detach cpuset from @tsk and release it. | |
2146 | * | |
053199ed | 2147 | * Note that cpusets marked notify_on_release force every task in |
3d3f26a7 | 2148 | * them to take the global manage_mutex mutex when exiting. |
053199ed PJ |
2149 | * This could impact scaling on very large systems. Be reluctant to |
2150 | * use notify_on_release cpusets where very high task exit scaling | |
2151 | * is required on large systems. | |
2152 | * | |
2153 | * Don't even think about derefencing 'cs' after the cpuset use count | |
3d3f26a7 IM |
2154 | * goes to zero, except inside a critical section guarded by manage_mutex |
2155 | * or callback_mutex. Otherwise a zero cpuset use count is a license to | |
053199ed PJ |
2156 | * any other task to nuke the cpuset immediately, via cpuset_rmdir(). |
2157 | * | |
3d3f26a7 IM |
2158 | * This routine has to take manage_mutex, not callback_mutex, because |
2159 | * it is holding that mutex while calling check_for_release(), | |
2160 | * which calls kmalloc(), so can't be called holding callback_mutex(). | |
053199ed | 2161 | * |
8488bc35 | 2162 | * the_top_cpuset_hack: |
06fed338 PJ |
2163 | * |
2164 | * Set the exiting tasks cpuset to the root cpuset (top_cpuset). | |
2165 | * | |
2166 | * Don't leave a task unable to allocate memory, as that is an | |
2167 | * accident waiting to happen should someone add a callout in | |
2168 | * do_exit() after the cpuset_exit() call that might allocate. | |
2169 | * If a task tries to allocate memory with an invalid cpuset, | |
2170 | * it will oops in cpuset_update_task_memory_state(). | |
2171 | * | |
2172 | * We call cpuset_exit() while the task is still competent to | |
2173 | * handle notify_on_release(), then leave the task attached to | |
2174 | * the root cpuset (top_cpuset) for the remainder of its exit. | |
2175 | * | |
2176 | * To do this properly, we would increment the reference count on | |
2177 | * top_cpuset, and near the very end of the kernel/exit.c do_exit() | |
2178 | * code we would add a second cpuset function call, to drop that | |
2179 | * reference. This would just create an unnecessary hot spot on | |
2180 | * the top_cpuset reference count, to no avail. | |
2181 | * | |
2182 | * Normally, holding a reference to a cpuset without bumping its | |
2183 | * count is unsafe. The cpuset could go away, or someone could | |
2184 | * attach us to a different cpuset, decrementing the count on | |
2185 | * the first cpuset that we never incremented. But in this case, | |
2186 | * top_cpuset isn't going away, and either task has PF_EXITING set, | |
2187 | * which wards off any attach_task() attempts, or task is a failed | |
2188 | * fork, never visible to attach_task. | |
2189 | * | |
2190 | * Another way to do this would be to set the cpuset pointer | |
2191 | * to NULL here, and check in cpuset_update_task_memory_state() | |
2192 | * for a NULL pointer. This hack avoids that NULL check, for no | |
2193 | * cost (other than this way too long comment ;). | |
1da177e4 LT |
2194 | **/ |
2195 | ||
2196 | void cpuset_exit(struct task_struct *tsk) | |
2197 | { | |
2198 | struct cpuset *cs; | |
2199 | ||
dd9037a2 | 2200 | task_lock(current); |
1da177e4 | 2201 | cs = tsk->cpuset; |
8488bc35 | 2202 | tsk->cpuset = &top_cpuset; /* the_top_cpuset_hack - see above */ |
dd9037a2 | 2203 | task_unlock(current); |
1da177e4 | 2204 | |
2efe86b8 | 2205 | if (notify_on_release(cs)) { |
3077a260 PJ |
2206 | char *pathbuf = NULL; |
2207 | ||
3d3f26a7 | 2208 | mutex_lock(&manage_mutex); |
2efe86b8 | 2209 | if (atomic_dec_and_test(&cs->count)) |
3077a260 | 2210 | check_for_release(cs, &pathbuf); |
3d3f26a7 | 2211 | mutex_unlock(&manage_mutex); |
3077a260 | 2212 | cpuset_release_agent(pathbuf); |
2efe86b8 PJ |
2213 | } else { |
2214 | atomic_dec(&cs->count); | |
1da177e4 LT |
2215 | } |
2216 | } | |
2217 | ||
2218 | /** | |
2219 | * cpuset_cpus_allowed - return cpus_allowed mask from a tasks cpuset. | |
2220 | * @tsk: pointer to task_struct from which to obtain cpuset->cpus_allowed. | |
2221 | * | |
2222 | * Description: Returns the cpumask_t cpus_allowed of the cpuset | |
2223 | * attached to the specified @tsk. Guaranteed to return some non-empty | |
2224 | * subset of cpu_online_map, even if this means going outside the | |
2225 | * tasks cpuset. | |
2226 | **/ | |
2227 | ||
909d75a3 | 2228 | cpumask_t cpuset_cpus_allowed(struct task_struct *tsk) |
1da177e4 LT |
2229 | { |
2230 | cpumask_t mask; | |
2231 | ||
3d3f26a7 | 2232 | mutex_lock(&callback_mutex); |
909d75a3 | 2233 | task_lock(tsk); |
1da177e4 | 2234 | guarantee_online_cpus(tsk->cpuset, &mask); |
909d75a3 | 2235 | task_unlock(tsk); |
3d3f26a7 | 2236 | mutex_unlock(&callback_mutex); |
1da177e4 LT |
2237 | |
2238 | return mask; | |
2239 | } | |
2240 | ||
2241 | void cpuset_init_current_mems_allowed(void) | |
2242 | { | |
2243 | current->mems_allowed = NODE_MASK_ALL; | |
2244 | } | |
2245 | ||
909d75a3 PJ |
2246 | /** |
2247 | * cpuset_mems_allowed - return mems_allowed mask from a tasks cpuset. | |
2248 | * @tsk: pointer to task_struct from which to obtain cpuset->mems_allowed. | |
2249 | * | |
2250 | * Description: Returns the nodemask_t mems_allowed of the cpuset | |
2251 | * attached to the specified @tsk. Guaranteed to return some non-empty | |
0e1e7c7a | 2252 | * subset of node_states[N_HIGH_MEMORY], even if this means going outside the |
909d75a3 PJ |
2253 | * tasks cpuset. |
2254 | **/ | |
2255 | ||
2256 | nodemask_t cpuset_mems_allowed(struct task_struct *tsk) | |
2257 | { | |
2258 | nodemask_t mask; | |
2259 | ||
3d3f26a7 | 2260 | mutex_lock(&callback_mutex); |
909d75a3 PJ |
2261 | task_lock(tsk); |
2262 | guarantee_online_mems(tsk->cpuset, &mask); | |
2263 | task_unlock(tsk); | |
3d3f26a7 | 2264 | mutex_unlock(&callback_mutex); |
909d75a3 PJ |
2265 | |
2266 | return mask; | |
2267 | } | |
2268 | ||
d9fd8a6d RD |
2269 | /** |
2270 | * cpuset_zonelist_valid_mems_allowed - check zonelist vs. curremt mems_allowed | |
2271 | * @zl: the zonelist to be checked | |
2272 | * | |
1da177e4 LT |
2273 | * Are any of the nodes on zonelist zl allowed in current->mems_allowed? |
2274 | */ | |
2275 | int cpuset_zonelist_valid_mems_allowed(struct zonelist *zl) | |
2276 | { | |
2277 | int i; | |
2278 | ||
2279 | for (i = 0; zl->zones[i]; i++) { | |
89fa3024 | 2280 | int nid = zone_to_nid(zl->zones[i]); |
1da177e4 LT |
2281 | |
2282 | if (node_isset(nid, current->mems_allowed)) | |
2283 | return 1; | |
2284 | } | |
2285 | return 0; | |
2286 | } | |
2287 | ||
9bf2229f PJ |
2288 | /* |
2289 | * nearest_exclusive_ancestor() - Returns the nearest mem_exclusive | |
3d3f26a7 | 2290 | * ancestor to the specified cpuset. Call holding callback_mutex. |
9bf2229f PJ |
2291 | * If no ancestor is mem_exclusive (an unusual configuration), then |
2292 | * returns the root cpuset. | |
2293 | */ | |
2294 | static const struct cpuset *nearest_exclusive_ancestor(const struct cpuset *cs) | |
2295 | { | |
2296 | while (!is_mem_exclusive(cs) && cs->parent) | |
2297 | cs = cs->parent; | |
2298 | return cs; | |
2299 | } | |
2300 | ||
d9fd8a6d | 2301 | /** |
02a0e53d | 2302 | * cpuset_zone_allowed_softwall - Can we allocate on zone z's memory node? |
9bf2229f | 2303 | * @z: is this zone on an allowed node? |
02a0e53d | 2304 | * @gfp_mask: memory allocation flags |
d9fd8a6d | 2305 | * |
02a0e53d PJ |
2306 | * If we're in interrupt, yes, we can always allocate. If |
2307 | * __GFP_THISNODE is set, yes, we can always allocate. If zone | |
9bf2229f PJ |
2308 | * z's node is in our tasks mems_allowed, yes. If it's not a |
2309 | * __GFP_HARDWALL request and this zone's nodes is in the nearest | |
2310 | * mem_exclusive cpuset ancestor to this tasks cpuset, yes. | |
c596d9f3 DR |
2311 | * If the task has been OOM killed and has access to memory reserves |
2312 | * as specified by the TIF_MEMDIE flag, yes. | |
9bf2229f PJ |
2313 | * Otherwise, no. |
2314 | * | |
02a0e53d PJ |
2315 | * If __GFP_HARDWALL is set, cpuset_zone_allowed_softwall() |
2316 | * reduces to cpuset_zone_allowed_hardwall(). Otherwise, | |
2317 | * cpuset_zone_allowed_softwall() might sleep, and might allow a zone | |
2318 | * from an enclosing cpuset. | |
2319 | * | |
2320 | * cpuset_zone_allowed_hardwall() only handles the simpler case of | |
2321 | * hardwall cpusets, and never sleeps. | |
2322 | * | |
2323 | * The __GFP_THISNODE placement logic is really handled elsewhere, | |
2324 | * by forcibly using a zonelist starting at a specified node, and by | |
2325 | * (in get_page_from_freelist()) refusing to consider the zones for | |
2326 | * any node on the zonelist except the first. By the time any such | |
2327 | * calls get to this routine, we should just shut up and say 'yes'. | |
2328 | * | |
9bf2229f | 2329 | * GFP_USER allocations are marked with the __GFP_HARDWALL bit, |
c596d9f3 DR |
2330 | * and do not allow allocations outside the current tasks cpuset |
2331 | * unless the task has been OOM killed as is marked TIF_MEMDIE. | |
9bf2229f | 2332 | * GFP_KERNEL allocations are not so marked, so can escape to the |
02a0e53d | 2333 | * nearest enclosing mem_exclusive ancestor cpuset. |
9bf2229f | 2334 | * |
02a0e53d PJ |
2335 | * Scanning up parent cpusets requires callback_mutex. The |
2336 | * __alloc_pages() routine only calls here with __GFP_HARDWALL bit | |
2337 | * _not_ set if it's a GFP_KERNEL allocation, and all nodes in the | |
2338 | * current tasks mems_allowed came up empty on the first pass over | |
2339 | * the zonelist. So only GFP_KERNEL allocations, if all nodes in the | |
2340 | * cpuset are short of memory, might require taking the callback_mutex | |
2341 | * mutex. | |
9bf2229f | 2342 | * |
36be57ff | 2343 | * The first call here from mm/page_alloc:get_page_from_freelist() |
02a0e53d PJ |
2344 | * has __GFP_HARDWALL set in gfp_mask, enforcing hardwall cpusets, |
2345 | * so no allocation on a node outside the cpuset is allowed (unless | |
2346 | * in interrupt, of course). | |
36be57ff PJ |
2347 | * |
2348 | * The second pass through get_page_from_freelist() doesn't even call | |
2349 | * here for GFP_ATOMIC calls. For those calls, the __alloc_pages() | |
2350 | * variable 'wait' is not set, and the bit ALLOC_CPUSET is not set | |
2351 | * in alloc_flags. That logic and the checks below have the combined | |
2352 | * affect that: | |
9bf2229f PJ |
2353 | * in_interrupt - any node ok (current task context irrelevant) |
2354 | * GFP_ATOMIC - any node ok | |
c596d9f3 | 2355 | * TIF_MEMDIE - any node ok |
9bf2229f PJ |
2356 | * GFP_KERNEL - any node in enclosing mem_exclusive cpuset ok |
2357 | * GFP_USER - only nodes in current tasks mems allowed ok. | |
36be57ff PJ |
2358 | * |
2359 | * Rule: | |
02a0e53d | 2360 | * Don't call cpuset_zone_allowed_softwall if you can't sleep, unless you |
36be57ff PJ |
2361 | * pass in the __GFP_HARDWALL flag set in gfp_flag, which disables |
2362 | * the code that might scan up ancestor cpusets and sleep. | |
02a0e53d | 2363 | */ |
9bf2229f | 2364 | |
02a0e53d | 2365 | int __cpuset_zone_allowed_softwall(struct zone *z, gfp_t gfp_mask) |
1da177e4 | 2366 | { |
9bf2229f PJ |
2367 | int node; /* node that zone z is on */ |
2368 | const struct cpuset *cs; /* current cpuset ancestors */ | |
29afd49b | 2369 | int allowed; /* is allocation in zone z allowed? */ |
9bf2229f | 2370 | |
9b819d20 | 2371 | if (in_interrupt() || (gfp_mask & __GFP_THISNODE)) |
9bf2229f | 2372 | return 1; |
89fa3024 | 2373 | node = zone_to_nid(z); |
92d1dbd2 | 2374 | might_sleep_if(!(gfp_mask & __GFP_HARDWALL)); |
9bf2229f PJ |
2375 | if (node_isset(node, current->mems_allowed)) |
2376 | return 1; | |
c596d9f3 DR |
2377 | /* |
2378 | * Allow tasks that have access to memory reserves because they have | |
2379 | * been OOM killed to get memory anywhere. | |
2380 | */ | |
2381 | if (unlikely(test_thread_flag(TIF_MEMDIE))) | |
2382 | return 1; | |
9bf2229f PJ |
2383 | if (gfp_mask & __GFP_HARDWALL) /* If hardwall request, stop here */ |
2384 | return 0; | |
2385 | ||
5563e770 BP |
2386 | if (current->flags & PF_EXITING) /* Let dying task have memory */ |
2387 | return 1; | |
2388 | ||
9bf2229f | 2389 | /* Not hardwall and node outside mems_allowed: scan up cpusets */ |
3d3f26a7 | 2390 | mutex_lock(&callback_mutex); |
053199ed | 2391 | |
053199ed PJ |
2392 | task_lock(current); |
2393 | cs = nearest_exclusive_ancestor(current->cpuset); | |
2394 | task_unlock(current); | |
2395 | ||
9bf2229f | 2396 | allowed = node_isset(node, cs->mems_allowed); |
3d3f26a7 | 2397 | mutex_unlock(&callback_mutex); |
9bf2229f | 2398 | return allowed; |
1da177e4 LT |
2399 | } |
2400 | ||
02a0e53d PJ |
2401 | /* |
2402 | * cpuset_zone_allowed_hardwall - Can we allocate on zone z's memory node? | |
2403 | * @z: is this zone on an allowed node? | |
2404 | * @gfp_mask: memory allocation flags | |
2405 | * | |
2406 | * If we're in interrupt, yes, we can always allocate. | |
2407 | * If __GFP_THISNODE is set, yes, we can always allocate. If zone | |
c596d9f3 DR |
2408 | * z's node is in our tasks mems_allowed, yes. If the task has been |
2409 | * OOM killed and has access to memory reserves as specified by the | |
2410 | * TIF_MEMDIE flag, yes. Otherwise, no. | |
02a0e53d PJ |
2411 | * |
2412 | * The __GFP_THISNODE placement logic is really handled elsewhere, | |
2413 | * by forcibly using a zonelist starting at a specified node, and by | |
2414 | * (in get_page_from_freelist()) refusing to consider the zones for | |
2415 | * any node on the zonelist except the first. By the time any such | |
2416 | * calls get to this routine, we should just shut up and say 'yes'. | |
2417 | * | |
2418 | * Unlike the cpuset_zone_allowed_softwall() variant, above, | |
2419 | * this variant requires that the zone be in the current tasks | |
2420 | * mems_allowed or that we're in interrupt. It does not scan up the | |
2421 | * cpuset hierarchy for the nearest enclosing mem_exclusive cpuset. | |
2422 | * It never sleeps. | |
2423 | */ | |
2424 | ||
2425 | int __cpuset_zone_allowed_hardwall(struct zone *z, gfp_t gfp_mask) | |
2426 | { | |
2427 | int node; /* node that zone z is on */ | |
2428 | ||
2429 | if (in_interrupt() || (gfp_mask & __GFP_THISNODE)) | |
2430 | return 1; | |
2431 | node = zone_to_nid(z); | |
2432 | if (node_isset(node, current->mems_allowed)) | |
2433 | return 1; | |
c596d9f3 DR |
2434 | /* |
2435 | * Allow tasks that have access to memory reserves because they have | |
2436 | * been OOM killed to get memory anywhere. | |
2437 | */ | |
2438 | if (unlikely(test_thread_flag(TIF_MEMDIE))) | |
2439 | return 1; | |
02a0e53d PJ |
2440 | return 0; |
2441 | } | |
2442 | ||
505970b9 PJ |
2443 | /** |
2444 | * cpuset_lock - lock out any changes to cpuset structures | |
2445 | * | |
3d3f26a7 | 2446 | * The out of memory (oom) code needs to mutex_lock cpusets |
505970b9 | 2447 | * from being changed while it scans the tasklist looking for a |
3d3f26a7 | 2448 | * task in an overlapping cpuset. Expose callback_mutex via this |
505970b9 PJ |
2449 | * cpuset_lock() routine, so the oom code can lock it, before |
2450 | * locking the task list. The tasklist_lock is a spinlock, so | |
3d3f26a7 | 2451 | * must be taken inside callback_mutex. |
505970b9 PJ |
2452 | */ |
2453 | ||
2454 | void cpuset_lock(void) | |
2455 | { | |
3d3f26a7 | 2456 | mutex_lock(&callback_mutex); |
505970b9 PJ |
2457 | } |
2458 | ||
2459 | /** | |
2460 | * cpuset_unlock - release lock on cpuset changes | |
2461 | * | |
2462 | * Undo the lock taken in a previous cpuset_lock() call. | |
2463 | */ | |
2464 | ||
2465 | void cpuset_unlock(void) | |
2466 | { | |
3d3f26a7 | 2467 | mutex_unlock(&callback_mutex); |
505970b9 PJ |
2468 | } |
2469 | ||
825a46af PJ |
2470 | /** |
2471 | * cpuset_mem_spread_node() - On which node to begin search for a page | |
2472 | * | |
2473 | * If a task is marked PF_SPREAD_PAGE or PF_SPREAD_SLAB (as for | |
2474 | * tasks in a cpuset with is_spread_page or is_spread_slab set), | |
2475 | * and if the memory allocation used cpuset_mem_spread_node() | |
2476 | * to determine on which node to start looking, as it will for | |
2477 | * certain page cache or slab cache pages such as used for file | |
2478 | * system buffers and inode caches, then instead of starting on the | |
2479 | * local node to look for a free page, rather spread the starting | |
2480 | * node around the tasks mems_allowed nodes. | |
2481 | * | |
2482 | * We don't have to worry about the returned node being offline | |
2483 | * because "it can't happen", and even if it did, it would be ok. | |
2484 | * | |
2485 | * The routines calling guarantee_online_mems() are careful to | |
2486 | * only set nodes in task->mems_allowed that are online. So it | |
2487 | * should not be possible for the following code to return an | |
2488 | * offline node. But if it did, that would be ok, as this routine | |
2489 | * is not returning the node where the allocation must be, only | |
2490 | * the node where the search should start. The zonelist passed to | |
2491 | * __alloc_pages() will include all nodes. If the slab allocator | |
2492 | * is passed an offline node, it will fall back to the local node. | |
2493 | * See kmem_cache_alloc_node(). | |
2494 | */ | |
2495 | ||
2496 | int cpuset_mem_spread_node(void) | |
2497 | { | |
2498 | int node; | |
2499 | ||
2500 | node = next_node(current->cpuset_mem_spread_rotor, current->mems_allowed); | |
2501 | if (node == MAX_NUMNODES) | |
2502 | node = first_node(current->mems_allowed); | |
2503 | current->cpuset_mem_spread_rotor = node; | |
2504 | return node; | |
2505 | } | |
2506 | EXPORT_SYMBOL_GPL(cpuset_mem_spread_node); | |
2507 | ||
ef08e3b4 | 2508 | /** |
bbe373f2 DR |
2509 | * cpuset_mems_allowed_intersects - Does @tsk1's mems_allowed intersect @tsk2's? |
2510 | * @tsk1: pointer to task_struct of some task. | |
2511 | * @tsk2: pointer to task_struct of some other task. | |
2512 | * | |
2513 | * Description: Return true if @tsk1's mems_allowed intersects the | |
2514 | * mems_allowed of @tsk2. Used by the OOM killer to determine if | |
2515 | * one of the task's memory usage might impact the memory available | |
2516 | * to the other. | |
ef08e3b4 PJ |
2517 | **/ |
2518 | ||
bbe373f2 DR |
2519 | int cpuset_mems_allowed_intersects(const struct task_struct *tsk1, |
2520 | const struct task_struct *tsk2) | |
ef08e3b4 | 2521 | { |
bbe373f2 | 2522 | return nodes_intersects(tsk1->mems_allowed, tsk2->mems_allowed); |
ef08e3b4 PJ |
2523 | } |
2524 | ||
3e0d98b9 PJ |
2525 | /* |
2526 | * Collection of memory_pressure is suppressed unless | |
2527 | * this flag is enabled by writing "1" to the special | |
2528 | * cpuset file 'memory_pressure_enabled' in the root cpuset. | |
2529 | */ | |
2530 | ||
c5b2aff8 | 2531 | int cpuset_memory_pressure_enabled __read_mostly; |
3e0d98b9 PJ |
2532 | |
2533 | /** | |
2534 | * cpuset_memory_pressure_bump - keep stats of per-cpuset reclaims. | |
2535 | * | |
2536 | * Keep a running average of the rate of synchronous (direct) | |
2537 | * page reclaim efforts initiated by tasks in each cpuset. | |
2538 | * | |
2539 | * This represents the rate at which some task in the cpuset | |
2540 | * ran low on memory on all nodes it was allowed to use, and | |
2541 | * had to enter the kernels page reclaim code in an effort to | |
2542 | * create more free memory by tossing clean pages or swapping | |
2543 | * or writing dirty pages. | |
2544 | * | |
2545 | * Display to user space in the per-cpuset read-only file | |
2546 | * "memory_pressure". Value displayed is an integer | |
2547 | * representing the recent rate of entry into the synchronous | |
2548 | * (direct) page reclaim by any task attached to the cpuset. | |
2549 | **/ | |
2550 | ||
2551 | void __cpuset_memory_pressure_bump(void) | |
2552 | { | |
2553 | struct cpuset *cs; | |
2554 | ||
2555 | task_lock(current); | |
2556 | cs = current->cpuset; | |
2557 | fmeter_markevent(&cs->fmeter); | |
2558 | task_unlock(current); | |
2559 | } | |
2560 | ||
1da177e4 LT |
2561 | /* |
2562 | * proc_cpuset_show() | |
2563 | * - Print tasks cpuset path into seq_file. | |
2564 | * - Used for /proc/<pid>/cpuset. | |
053199ed PJ |
2565 | * - No need to task_lock(tsk) on this tsk->cpuset reference, as it |
2566 | * doesn't really matter if tsk->cpuset changes after we read it, | |
3d3f26a7 | 2567 | * and we take manage_mutex, keeping attach_task() from changing it |
8488bc35 PJ |
2568 | * anyway. No need to check that tsk->cpuset != NULL, thanks to |
2569 | * the_top_cpuset_hack in cpuset_exit(), which sets an exiting tasks | |
2570 | * cpuset to top_cpuset. | |
1da177e4 | 2571 | */ |
1da177e4 LT |
2572 | static int proc_cpuset_show(struct seq_file *m, void *v) |
2573 | { | |
13b41b09 | 2574 | struct pid *pid; |
1da177e4 LT |
2575 | struct task_struct *tsk; |
2576 | char *buf; | |
99f89551 | 2577 | int retval; |
1da177e4 | 2578 | |
99f89551 | 2579 | retval = -ENOMEM; |
1da177e4 LT |
2580 | buf = kmalloc(PAGE_SIZE, GFP_KERNEL); |
2581 | if (!buf) | |
99f89551 EB |
2582 | goto out; |
2583 | ||
2584 | retval = -ESRCH; | |
13b41b09 EB |
2585 | pid = m->private; |
2586 | tsk = get_pid_task(pid, PIDTYPE_PID); | |
99f89551 EB |
2587 | if (!tsk) |
2588 | goto out_free; | |
1da177e4 | 2589 | |
99f89551 | 2590 | retval = -EINVAL; |
3d3f26a7 | 2591 | mutex_lock(&manage_mutex); |
99f89551 | 2592 | |
8488bc35 | 2593 | retval = cpuset_path(tsk->cpuset, buf, PAGE_SIZE); |
1da177e4 | 2594 | if (retval < 0) |
99f89551 | 2595 | goto out_unlock; |
1da177e4 LT |
2596 | seq_puts(m, buf); |
2597 | seq_putc(m, '\n'); | |
99f89551 | 2598 | out_unlock: |
3d3f26a7 | 2599 | mutex_unlock(&manage_mutex); |
99f89551 EB |
2600 | put_task_struct(tsk); |
2601 | out_free: | |
1da177e4 | 2602 | kfree(buf); |
99f89551 | 2603 | out: |
1da177e4 LT |
2604 | return retval; |
2605 | } | |
2606 | ||
2607 | static int cpuset_open(struct inode *inode, struct file *file) | |
2608 | { | |
13b41b09 EB |
2609 | struct pid *pid = PROC_I(inode)->pid; |
2610 | return single_open(file, proc_cpuset_show, pid); | |
1da177e4 LT |
2611 | } |
2612 | ||
9a32144e | 2613 | const struct file_operations proc_cpuset_operations = { |
1da177e4 LT |
2614 | .open = cpuset_open, |
2615 | .read = seq_read, | |
2616 | .llseek = seq_lseek, | |
2617 | .release = single_release, | |
2618 | }; | |
2619 | ||
2620 | /* Display task cpus_allowed, mems_allowed in /proc/<pid>/status file. */ | |
2621 | char *cpuset_task_status_allowed(struct task_struct *task, char *buffer) | |
2622 | { | |
2623 | buffer += sprintf(buffer, "Cpus_allowed:\t"); | |
2624 | buffer += cpumask_scnprintf(buffer, PAGE_SIZE, task->cpus_allowed); | |
2625 | buffer += sprintf(buffer, "\n"); | |
2626 | buffer += sprintf(buffer, "Mems_allowed:\t"); | |
2627 | buffer += nodemask_scnprintf(buffer, PAGE_SIZE, task->mems_allowed); | |
2628 | buffer += sprintf(buffer, "\n"); | |
2629 | return buffer; | |
2630 | } |