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. | |
029190c5 | 7 | * Copyright (C) 2004-2007 Silicon Graphics, Inc. |
8793d854 | 8 | * Copyright (C) 2006 Google, Inc |
1da177e4 LT |
9 | * |
10 | * Portions derived from Patrick Mochel's sysfs code. | |
11 | * sysfs is Copyright (c) 2001-3 Patrick Mochel | |
1da177e4 | 12 | * |
825a46af | 13 | * 2003-10-10 Written by Simon Derr. |
1da177e4 | 14 | * 2003-10-22 Updates by Stephen Hemminger. |
825a46af | 15 | * 2004 May-July Rework by Paul Jackson. |
8793d854 | 16 | * 2006 Rework by Paul Menage to use generic cgroups |
1da177e4 LT |
17 | * |
18 | * This file is subject to the terms and conditions of the GNU General Public | |
19 | * License. See the file COPYING in the main directory of the Linux | |
20 | * distribution for more details. | |
21 | */ | |
22 | ||
1da177e4 LT |
23 | #include <linux/cpu.h> |
24 | #include <linux/cpumask.h> | |
25 | #include <linux/cpuset.h> | |
26 | #include <linux/err.h> | |
27 | #include <linux/errno.h> | |
28 | #include <linux/file.h> | |
29 | #include <linux/fs.h> | |
30 | #include <linux/init.h> | |
31 | #include <linux/interrupt.h> | |
32 | #include <linux/kernel.h> | |
33 | #include <linux/kmod.h> | |
34 | #include <linux/list.h> | |
68860ec1 | 35 | #include <linux/mempolicy.h> |
1da177e4 LT |
36 | #include <linux/mm.h> |
37 | #include <linux/module.h> | |
38 | #include <linux/mount.h> | |
39 | #include <linux/namei.h> | |
40 | #include <linux/pagemap.h> | |
41 | #include <linux/proc_fs.h> | |
6b9c2603 | 42 | #include <linux/rcupdate.h> |
1da177e4 LT |
43 | #include <linux/sched.h> |
44 | #include <linux/seq_file.h> | |
22fb52dd | 45 | #include <linux/security.h> |
1da177e4 | 46 | #include <linux/slab.h> |
1da177e4 LT |
47 | #include <linux/spinlock.h> |
48 | #include <linux/stat.h> | |
49 | #include <linux/string.h> | |
50 | #include <linux/time.h> | |
51 | #include <linux/backing-dev.h> | |
52 | #include <linux/sort.h> | |
53 | ||
54 | #include <asm/uaccess.h> | |
55 | #include <asm/atomic.h> | |
3d3f26a7 | 56 | #include <linux/mutex.h> |
029190c5 | 57 | #include <linux/kfifo.h> |
956db3ca CW |
58 | #include <linux/workqueue.h> |
59 | #include <linux/cgroup.h> | |
1da177e4 | 60 | |
202f72d5 PJ |
61 | /* |
62 | * Tracks how many cpusets are currently defined in system. | |
63 | * When there is only one cpuset (the root cpuset) we can | |
64 | * short circuit some hooks. | |
65 | */ | |
7edc5962 | 66 | int number_of_cpusets __read_mostly; |
202f72d5 | 67 | |
2df167a3 | 68 | /* Forward declare cgroup structures */ |
8793d854 PM |
69 | struct cgroup_subsys cpuset_subsys; |
70 | struct cpuset; | |
71 | ||
3e0d98b9 PJ |
72 | /* See "Frequency meter" comments, below. */ |
73 | ||
74 | struct fmeter { | |
75 | int cnt; /* unprocessed events count */ | |
76 | int val; /* most recent output value */ | |
77 | time_t time; /* clock (secs) when val computed */ | |
78 | spinlock_t lock; /* guards read or write of above */ | |
79 | }; | |
80 | ||
1da177e4 | 81 | struct cpuset { |
8793d854 PM |
82 | struct cgroup_subsys_state css; |
83 | ||
1da177e4 LT |
84 | unsigned long flags; /* "unsigned long" so bitops work */ |
85 | cpumask_t cpus_allowed; /* CPUs allowed to tasks in cpuset */ | |
86 | nodemask_t mems_allowed; /* Memory Nodes allowed to tasks */ | |
87 | ||
1da177e4 | 88 | struct cpuset *parent; /* my parent */ |
1da177e4 LT |
89 | |
90 | /* | |
91 | * Copy of global cpuset_mems_generation as of the most | |
92 | * recent time this cpuset changed its mems_allowed. | |
93 | */ | |
3e0d98b9 PJ |
94 | int mems_generation; |
95 | ||
96 | struct fmeter fmeter; /* memory_pressure filter */ | |
029190c5 PJ |
97 | |
98 | /* partition number for rebuild_sched_domains() */ | |
99 | int pn; | |
956db3ca | 100 | |
1d3504fc HS |
101 | /* for custom sched domain */ |
102 | int relax_domain_level; | |
103 | ||
956db3ca CW |
104 | /* used for walking a cpuset heirarchy */ |
105 | struct list_head stack_list; | |
1da177e4 LT |
106 | }; |
107 | ||
8793d854 PM |
108 | /* Retrieve the cpuset for a cgroup */ |
109 | static inline struct cpuset *cgroup_cs(struct cgroup *cont) | |
110 | { | |
111 | return container_of(cgroup_subsys_state(cont, cpuset_subsys_id), | |
112 | struct cpuset, css); | |
113 | } | |
114 | ||
115 | /* Retrieve the cpuset for a task */ | |
116 | static inline struct cpuset *task_cs(struct task_struct *task) | |
117 | { | |
118 | return container_of(task_subsys_state(task, cpuset_subsys_id), | |
119 | struct cpuset, css); | |
120 | } | |
956db3ca CW |
121 | struct cpuset_hotplug_scanner { |
122 | struct cgroup_scanner scan; | |
123 | struct cgroup *to; | |
124 | }; | |
8793d854 | 125 | |
1da177e4 LT |
126 | /* bits in struct cpuset flags field */ |
127 | typedef enum { | |
128 | CS_CPU_EXCLUSIVE, | |
129 | CS_MEM_EXCLUSIVE, | |
45b07ef3 | 130 | CS_MEMORY_MIGRATE, |
029190c5 | 131 | CS_SCHED_LOAD_BALANCE, |
825a46af PJ |
132 | CS_SPREAD_PAGE, |
133 | CS_SPREAD_SLAB, | |
1da177e4 LT |
134 | } cpuset_flagbits_t; |
135 | ||
136 | /* convenient tests for these bits */ | |
137 | static inline int is_cpu_exclusive(const struct cpuset *cs) | |
138 | { | |
7b5b9ef0 | 139 | return test_bit(CS_CPU_EXCLUSIVE, &cs->flags); |
1da177e4 LT |
140 | } |
141 | ||
142 | static inline int is_mem_exclusive(const struct cpuset *cs) | |
143 | { | |
7b5b9ef0 | 144 | return test_bit(CS_MEM_EXCLUSIVE, &cs->flags); |
1da177e4 LT |
145 | } |
146 | ||
029190c5 PJ |
147 | static inline int is_sched_load_balance(const struct cpuset *cs) |
148 | { | |
149 | return test_bit(CS_SCHED_LOAD_BALANCE, &cs->flags); | |
150 | } | |
151 | ||
45b07ef3 PJ |
152 | static inline int is_memory_migrate(const struct cpuset *cs) |
153 | { | |
7b5b9ef0 | 154 | return test_bit(CS_MEMORY_MIGRATE, &cs->flags); |
45b07ef3 PJ |
155 | } |
156 | ||
825a46af PJ |
157 | static inline int is_spread_page(const struct cpuset *cs) |
158 | { | |
159 | return test_bit(CS_SPREAD_PAGE, &cs->flags); | |
160 | } | |
161 | ||
162 | static inline int is_spread_slab(const struct cpuset *cs) | |
163 | { | |
164 | return test_bit(CS_SPREAD_SLAB, &cs->flags); | |
165 | } | |
166 | ||
1da177e4 | 167 | /* |
151a4420 | 168 | * Increment this integer everytime any cpuset changes its |
1da177e4 LT |
169 | * mems_allowed value. Users of cpusets can track this generation |
170 | * number, and avoid having to lock and reload mems_allowed unless | |
171 | * the cpuset they're using changes generation. | |
172 | * | |
2df167a3 | 173 | * A single, global generation is needed because cpuset_attach_task() could |
1da177e4 LT |
174 | * reattach a task to a different cpuset, which must not have its |
175 | * generation numbers aliased with those of that tasks previous cpuset. | |
176 | * | |
177 | * Generations are needed for mems_allowed because one task cannot | |
2df167a3 | 178 | * modify another's memory placement. So we must enable every task, |
1da177e4 LT |
179 | * on every visit to __alloc_pages(), to efficiently check whether |
180 | * its current->cpuset->mems_allowed has changed, requiring an update | |
181 | * of its current->mems_allowed. | |
151a4420 | 182 | * |
2df167a3 | 183 | * Since writes to cpuset_mems_generation are guarded by the cgroup lock |
151a4420 | 184 | * there is no need to mark it atomic. |
1da177e4 | 185 | */ |
151a4420 | 186 | static int cpuset_mems_generation; |
1da177e4 LT |
187 | |
188 | static struct cpuset top_cpuset = { | |
189 | .flags = ((1 << CS_CPU_EXCLUSIVE) | (1 << CS_MEM_EXCLUSIVE)), | |
190 | .cpus_allowed = CPU_MASK_ALL, | |
191 | .mems_allowed = NODE_MASK_ALL, | |
1da177e4 LT |
192 | }; |
193 | ||
1da177e4 | 194 | /* |
2df167a3 PM |
195 | * There are two global mutexes guarding cpuset structures. The first |
196 | * is the main control groups cgroup_mutex, accessed via | |
197 | * cgroup_lock()/cgroup_unlock(). The second is the cpuset-specific | |
198 | * callback_mutex, below. They can nest. It is ok to first take | |
199 | * cgroup_mutex, then nest callback_mutex. We also require taking | |
200 | * task_lock() when dereferencing a task's cpuset pointer. See "The | |
201 | * task_lock() exception", at the end of this comment. | |
053199ed | 202 | * |
3d3f26a7 | 203 | * A task must hold both mutexes to modify cpusets. If a task |
2df167a3 | 204 | * holds cgroup_mutex, then it blocks others wanting that mutex, |
3d3f26a7 | 205 | * ensuring that it is the only task able to also acquire callback_mutex |
053199ed PJ |
206 | * and be able to modify cpusets. It can perform various checks on |
207 | * the cpuset structure first, knowing nothing will change. It can | |
2df167a3 | 208 | * also allocate memory while just holding cgroup_mutex. While it is |
053199ed | 209 | * performing these checks, various callback routines can briefly |
3d3f26a7 IM |
210 | * acquire callback_mutex to query cpusets. Once it is ready to make |
211 | * the changes, it takes callback_mutex, blocking everyone else. | |
053199ed PJ |
212 | * |
213 | * Calls to the kernel memory allocator can not be made while holding | |
3d3f26a7 | 214 | * callback_mutex, as that would risk double tripping on callback_mutex |
053199ed PJ |
215 | * from one of the callbacks into the cpuset code from within |
216 | * __alloc_pages(). | |
217 | * | |
3d3f26a7 | 218 | * If a task is only holding callback_mutex, then it has read-only |
053199ed PJ |
219 | * access to cpusets. |
220 | * | |
221 | * The task_struct fields mems_allowed and mems_generation may only | |
222 | * be accessed in the context of that task, so require no locks. | |
223 | * | |
053199ed | 224 | * The cpuset_common_file_write handler for operations that modify |
2df167a3 | 225 | * the cpuset hierarchy holds cgroup_mutex across the entire operation, |
053199ed PJ |
226 | * single threading all such cpuset modifications across the system. |
227 | * | |
3d3f26a7 | 228 | * The cpuset_common_file_read() handlers only hold callback_mutex across |
053199ed PJ |
229 | * small pieces of code, such as when reading out possibly multi-word |
230 | * cpumasks and nodemasks. | |
231 | * | |
2df167a3 PM |
232 | * Accessing a task's cpuset should be done in accordance with the |
233 | * guidelines for accessing subsystem state in kernel/cgroup.c | |
1da177e4 LT |
234 | */ |
235 | ||
3d3f26a7 | 236 | static DEFINE_MUTEX(callback_mutex); |
4247bdc6 | 237 | |
8793d854 PM |
238 | /* This is ugly, but preserves the userspace API for existing cpuset |
239 | * users. If someone tries to mount the "cpuset" filesystem, we | |
240 | * silently switch it to mount "cgroup" instead */ | |
454e2398 DH |
241 | static int cpuset_get_sb(struct file_system_type *fs_type, |
242 | int flags, const char *unused_dev_name, | |
243 | void *data, struct vfsmount *mnt) | |
1da177e4 | 244 | { |
8793d854 PM |
245 | struct file_system_type *cgroup_fs = get_fs_type("cgroup"); |
246 | int ret = -ENODEV; | |
247 | if (cgroup_fs) { | |
248 | char mountopts[] = | |
249 | "cpuset,noprefix," | |
250 | "release_agent=/sbin/cpuset_release_agent"; | |
251 | ret = cgroup_fs->get_sb(cgroup_fs, flags, | |
252 | unused_dev_name, mountopts, mnt); | |
253 | put_filesystem(cgroup_fs); | |
254 | } | |
255 | return ret; | |
1da177e4 LT |
256 | } |
257 | ||
258 | static struct file_system_type cpuset_fs_type = { | |
259 | .name = "cpuset", | |
260 | .get_sb = cpuset_get_sb, | |
1da177e4 LT |
261 | }; |
262 | ||
1da177e4 LT |
263 | /* |
264 | * Return in *pmask the portion of a cpusets's cpus_allowed that | |
265 | * are online. If none are online, walk up the cpuset hierarchy | |
266 | * until we find one that does have some online cpus. If we get | |
267 | * all the way to the top and still haven't found any online cpus, | |
268 | * return cpu_online_map. Or if passed a NULL cs from an exit'ing | |
269 | * task, return cpu_online_map. | |
270 | * | |
271 | * One way or another, we guarantee to return some non-empty subset | |
272 | * of cpu_online_map. | |
273 | * | |
3d3f26a7 | 274 | * Call with callback_mutex held. |
1da177e4 LT |
275 | */ |
276 | ||
277 | static void guarantee_online_cpus(const struct cpuset *cs, cpumask_t *pmask) | |
278 | { | |
279 | while (cs && !cpus_intersects(cs->cpus_allowed, cpu_online_map)) | |
280 | cs = cs->parent; | |
281 | if (cs) | |
282 | cpus_and(*pmask, cs->cpus_allowed, cpu_online_map); | |
283 | else | |
284 | *pmask = cpu_online_map; | |
285 | BUG_ON(!cpus_intersects(*pmask, cpu_online_map)); | |
286 | } | |
287 | ||
288 | /* | |
289 | * Return in *pmask the portion of a cpusets's mems_allowed that | |
0e1e7c7a CL |
290 | * are online, with memory. If none are online with memory, walk |
291 | * up the cpuset hierarchy until we find one that does have some | |
292 | * online mems. If we get all the way to the top and still haven't | |
293 | * found any online mems, return node_states[N_HIGH_MEMORY]. | |
1da177e4 LT |
294 | * |
295 | * One way or another, we guarantee to return some non-empty subset | |
0e1e7c7a | 296 | * of node_states[N_HIGH_MEMORY]. |
1da177e4 | 297 | * |
3d3f26a7 | 298 | * Call with callback_mutex held. |
1da177e4 LT |
299 | */ |
300 | ||
301 | static void guarantee_online_mems(const struct cpuset *cs, nodemask_t *pmask) | |
302 | { | |
0e1e7c7a CL |
303 | while (cs && !nodes_intersects(cs->mems_allowed, |
304 | node_states[N_HIGH_MEMORY])) | |
1da177e4 LT |
305 | cs = cs->parent; |
306 | if (cs) | |
0e1e7c7a CL |
307 | nodes_and(*pmask, cs->mems_allowed, |
308 | node_states[N_HIGH_MEMORY]); | |
1da177e4 | 309 | else |
0e1e7c7a CL |
310 | *pmask = node_states[N_HIGH_MEMORY]; |
311 | BUG_ON(!nodes_intersects(*pmask, node_states[N_HIGH_MEMORY])); | |
1da177e4 LT |
312 | } |
313 | ||
cf2a473c PJ |
314 | /** |
315 | * cpuset_update_task_memory_state - update task memory placement | |
316 | * | |
317 | * If the current tasks cpusets mems_allowed changed behind our | |
318 | * backs, update current->mems_allowed, mems_generation and task NUMA | |
319 | * mempolicy to the new value. | |
053199ed | 320 | * |
cf2a473c PJ |
321 | * Task mempolicy is updated by rebinding it relative to the |
322 | * current->cpuset if a task has its memory placement changed. | |
323 | * Do not call this routine if in_interrupt(). | |
324 | * | |
4a01c8d5 | 325 | * Call without callback_mutex or task_lock() held. May be |
2df167a3 PM |
326 | * called with or without cgroup_mutex held. Thanks in part to |
327 | * 'the_top_cpuset_hack', the task's cpuset pointer will never | |
41f7f60d DR |
328 | * be NULL. This routine also might acquire callback_mutex during |
329 | * call. | |
053199ed | 330 | * |
6b9c2603 PJ |
331 | * Reading current->cpuset->mems_generation doesn't need task_lock |
332 | * to guard the current->cpuset derefence, because it is guarded | |
2df167a3 | 333 | * from concurrent freeing of current->cpuset using RCU. |
6b9c2603 PJ |
334 | * |
335 | * The rcu_dereference() is technically probably not needed, | |
336 | * as I don't actually mind if I see a new cpuset pointer but | |
337 | * an old value of mems_generation. However this really only | |
338 | * matters on alpha systems using cpusets heavily. If I dropped | |
339 | * that rcu_dereference(), it would save them a memory barrier. | |
340 | * For all other arch's, rcu_dereference is a no-op anyway, and for | |
341 | * alpha systems not using cpusets, another planned optimization, | |
342 | * avoiding the rcu critical section for tasks in the root cpuset | |
343 | * which is statically allocated, so can't vanish, will make this | |
344 | * irrelevant. Better to use RCU as intended, than to engage in | |
345 | * some cute trick to save a memory barrier that is impossible to | |
346 | * test, for alpha systems using cpusets heavily, which might not | |
347 | * even exist. | |
053199ed PJ |
348 | * |
349 | * This routine is needed to update the per-task mems_allowed data, | |
350 | * within the tasks context, when it is trying to allocate memory | |
351 | * (in various mm/mempolicy.c routines) and notices that some other | |
352 | * task has been modifying its cpuset. | |
1da177e4 LT |
353 | */ |
354 | ||
fe85a998 | 355 | void cpuset_update_task_memory_state(void) |
1da177e4 | 356 | { |
053199ed | 357 | int my_cpusets_mem_gen; |
cf2a473c | 358 | struct task_struct *tsk = current; |
6b9c2603 | 359 | struct cpuset *cs; |
053199ed | 360 | |
8793d854 | 361 | if (task_cs(tsk) == &top_cpuset) { |
03a285f5 PJ |
362 | /* Don't need rcu for top_cpuset. It's never freed. */ |
363 | my_cpusets_mem_gen = top_cpuset.mems_generation; | |
364 | } else { | |
365 | rcu_read_lock(); | |
8793d854 | 366 | my_cpusets_mem_gen = task_cs(current)->mems_generation; |
03a285f5 PJ |
367 | rcu_read_unlock(); |
368 | } | |
1da177e4 | 369 | |
cf2a473c | 370 | if (my_cpusets_mem_gen != tsk->cpuset_mems_generation) { |
3d3f26a7 | 371 | mutex_lock(&callback_mutex); |
cf2a473c | 372 | task_lock(tsk); |
8793d854 | 373 | cs = task_cs(tsk); /* Maybe changed when task not locked */ |
cf2a473c PJ |
374 | guarantee_online_mems(cs, &tsk->mems_allowed); |
375 | tsk->cpuset_mems_generation = cs->mems_generation; | |
825a46af PJ |
376 | if (is_spread_page(cs)) |
377 | tsk->flags |= PF_SPREAD_PAGE; | |
378 | else | |
379 | tsk->flags &= ~PF_SPREAD_PAGE; | |
380 | if (is_spread_slab(cs)) | |
381 | tsk->flags |= PF_SPREAD_SLAB; | |
382 | else | |
383 | tsk->flags &= ~PF_SPREAD_SLAB; | |
cf2a473c | 384 | task_unlock(tsk); |
3d3f26a7 | 385 | mutex_unlock(&callback_mutex); |
74cb2155 | 386 | mpol_rebind_task(tsk, &tsk->mems_allowed); |
1da177e4 LT |
387 | } |
388 | } | |
389 | ||
390 | /* | |
391 | * is_cpuset_subset(p, q) - Is cpuset p a subset of cpuset q? | |
392 | * | |
393 | * One cpuset is a subset of another if all its allowed CPUs and | |
394 | * Memory Nodes are a subset of the other, and its exclusive flags | |
2df167a3 | 395 | * are only set if the other's are set. Call holding cgroup_mutex. |
1da177e4 LT |
396 | */ |
397 | ||
398 | static int is_cpuset_subset(const struct cpuset *p, const struct cpuset *q) | |
399 | { | |
400 | return cpus_subset(p->cpus_allowed, q->cpus_allowed) && | |
401 | nodes_subset(p->mems_allowed, q->mems_allowed) && | |
402 | is_cpu_exclusive(p) <= is_cpu_exclusive(q) && | |
403 | is_mem_exclusive(p) <= is_mem_exclusive(q); | |
404 | } | |
405 | ||
406 | /* | |
407 | * validate_change() - Used to validate that any proposed cpuset change | |
408 | * follows the structural rules for cpusets. | |
409 | * | |
410 | * If we replaced the flag and mask values of the current cpuset | |
411 | * (cur) with those values in the trial cpuset (trial), would | |
412 | * our various subset and exclusive rules still be valid? Presumes | |
2df167a3 | 413 | * cgroup_mutex held. |
1da177e4 LT |
414 | * |
415 | * 'cur' is the address of an actual, in-use cpuset. Operations | |
416 | * such as list traversal that depend on the actual address of the | |
417 | * cpuset in the list must use cur below, not trial. | |
418 | * | |
419 | * 'trial' is the address of bulk structure copy of cur, with | |
420 | * perhaps one or more of the fields cpus_allowed, mems_allowed, | |
421 | * or flags changed to new, trial values. | |
422 | * | |
423 | * Return 0 if valid, -errno if not. | |
424 | */ | |
425 | ||
426 | static int validate_change(const struct cpuset *cur, const struct cpuset *trial) | |
427 | { | |
8793d854 | 428 | struct cgroup *cont; |
1da177e4 LT |
429 | struct cpuset *c, *par; |
430 | ||
431 | /* Each of our child cpusets must be a subset of us */ | |
8793d854 PM |
432 | list_for_each_entry(cont, &cur->css.cgroup->children, sibling) { |
433 | if (!is_cpuset_subset(cgroup_cs(cont), trial)) | |
1da177e4 LT |
434 | return -EBUSY; |
435 | } | |
436 | ||
437 | /* Remaining checks don't apply to root cpuset */ | |
69604067 | 438 | if (cur == &top_cpuset) |
1da177e4 LT |
439 | return 0; |
440 | ||
69604067 PJ |
441 | par = cur->parent; |
442 | ||
1da177e4 LT |
443 | /* We must be a subset of our parent cpuset */ |
444 | if (!is_cpuset_subset(trial, par)) | |
445 | return -EACCES; | |
446 | ||
2df167a3 PM |
447 | /* |
448 | * If either I or some sibling (!= me) is exclusive, we can't | |
449 | * overlap | |
450 | */ | |
8793d854 PM |
451 | list_for_each_entry(cont, &par->css.cgroup->children, sibling) { |
452 | c = cgroup_cs(cont); | |
1da177e4 LT |
453 | if ((is_cpu_exclusive(trial) || is_cpu_exclusive(c)) && |
454 | c != cur && | |
455 | cpus_intersects(trial->cpus_allowed, c->cpus_allowed)) | |
456 | return -EINVAL; | |
457 | if ((is_mem_exclusive(trial) || is_mem_exclusive(c)) && | |
458 | c != cur && | |
459 | nodes_intersects(trial->mems_allowed, c->mems_allowed)) | |
460 | return -EINVAL; | |
461 | } | |
462 | ||
020958b6 PJ |
463 | /* Cpusets with tasks can't have empty cpus_allowed or mems_allowed */ |
464 | if (cgroup_task_count(cur->css.cgroup)) { | |
465 | if (cpus_empty(trial->cpus_allowed) || | |
466 | nodes_empty(trial->mems_allowed)) { | |
467 | return -ENOSPC; | |
468 | } | |
469 | } | |
470 | ||
1da177e4 LT |
471 | return 0; |
472 | } | |
473 | ||
029190c5 PJ |
474 | /* |
475 | * Helper routine for rebuild_sched_domains(). | |
476 | * Do cpusets a, b have overlapping cpus_allowed masks? | |
477 | */ | |
478 | ||
479 | static int cpusets_overlap(struct cpuset *a, struct cpuset *b) | |
480 | { | |
481 | return cpus_intersects(a->cpus_allowed, b->cpus_allowed); | |
482 | } | |
483 | ||
1d3504fc HS |
484 | static void |
485 | update_domain_attr(struct sched_domain_attr *dattr, struct cpuset *c) | |
486 | { | |
487 | if (!dattr) | |
488 | return; | |
489 | if (dattr->relax_domain_level < c->relax_domain_level) | |
490 | dattr->relax_domain_level = c->relax_domain_level; | |
491 | return; | |
492 | } | |
493 | ||
029190c5 PJ |
494 | /* |
495 | * rebuild_sched_domains() | |
496 | * | |
497 | * If the flag 'sched_load_balance' of any cpuset with non-empty | |
498 | * 'cpus' changes, or if the 'cpus' allowed changes in any cpuset | |
499 | * which has that flag enabled, or if any cpuset with a non-empty | |
500 | * 'cpus' is removed, then call this routine to rebuild the | |
501 | * scheduler's dynamic sched domains. | |
502 | * | |
503 | * This routine builds a partial partition of the systems CPUs | |
504 | * (the set of non-overlappping cpumask_t's in the array 'part' | |
505 | * below), and passes that partial partition to the kernel/sched.c | |
506 | * partition_sched_domains() routine, which will rebuild the | |
507 | * schedulers load balancing domains (sched domains) as specified | |
508 | * by that partial partition. A 'partial partition' is a set of | |
509 | * non-overlapping subsets whose union is a subset of that set. | |
510 | * | |
511 | * See "What is sched_load_balance" in Documentation/cpusets.txt | |
512 | * for a background explanation of this. | |
513 | * | |
514 | * Does not return errors, on the theory that the callers of this | |
515 | * routine would rather not worry about failures to rebuild sched | |
516 | * domains when operating in the severe memory shortage situations | |
517 | * that could cause allocation failures below. | |
518 | * | |
519 | * Call with cgroup_mutex held. May take callback_mutex during | |
520 | * call due to the kfifo_alloc() and kmalloc() calls. May nest | |
86ef5c9a | 521 | * a call to the get_online_cpus()/put_online_cpus() pair. |
029190c5 | 522 | * Must not be called holding callback_mutex, because we must not |
86ef5c9a GS |
523 | * call get_online_cpus() while holding callback_mutex. Elsewhere |
524 | * the kernel nests callback_mutex inside get_online_cpus() calls. | |
029190c5 PJ |
525 | * So the reverse nesting would risk an ABBA deadlock. |
526 | * | |
527 | * The three key local variables below are: | |
528 | * q - a kfifo queue of cpuset pointers, used to implement a | |
529 | * top-down scan of all cpusets. This scan loads a pointer | |
530 | * to each cpuset marked is_sched_load_balance into the | |
531 | * array 'csa'. For our purposes, rebuilding the schedulers | |
532 | * sched domains, we can ignore !is_sched_load_balance cpusets. | |
533 | * csa - (for CpuSet Array) Array of pointers to all the cpusets | |
534 | * that need to be load balanced, for convenient iterative | |
535 | * access by the subsequent code that finds the best partition, | |
536 | * i.e the set of domains (subsets) of CPUs such that the | |
537 | * cpus_allowed of every cpuset marked is_sched_load_balance | |
538 | * is a subset of one of these domains, while there are as | |
539 | * many such domains as possible, each as small as possible. | |
540 | * doms - Conversion of 'csa' to an array of cpumasks, for passing to | |
541 | * the kernel/sched.c routine partition_sched_domains() in a | |
542 | * convenient format, that can be easily compared to the prior | |
543 | * value to determine what partition elements (sched domains) | |
544 | * were changed (added or removed.) | |
545 | * | |
546 | * Finding the best partition (set of domains): | |
547 | * The triple nested loops below over i, j, k scan over the | |
548 | * load balanced cpusets (using the array of cpuset pointers in | |
549 | * csa[]) looking for pairs of cpusets that have overlapping | |
550 | * cpus_allowed, but which don't have the same 'pn' partition | |
551 | * number and gives them in the same partition number. It keeps | |
552 | * looping on the 'restart' label until it can no longer find | |
553 | * any such pairs. | |
554 | * | |
555 | * The union of the cpus_allowed masks from the set of | |
556 | * all cpusets having the same 'pn' value then form the one | |
557 | * element of the partition (one sched domain) to be passed to | |
558 | * partition_sched_domains(). | |
559 | */ | |
560 | ||
561 | static void rebuild_sched_domains(void) | |
562 | { | |
563 | struct kfifo *q; /* queue of cpusets to be scanned */ | |
564 | struct cpuset *cp; /* scans q */ | |
565 | struct cpuset **csa; /* array of all cpuset ptrs */ | |
566 | int csn; /* how many cpuset ptrs in csa so far */ | |
567 | int i, j, k; /* indices for partition finding loops */ | |
568 | cpumask_t *doms; /* resulting partition; i.e. sched domains */ | |
1d3504fc | 569 | struct sched_domain_attr *dattr; /* attributes for custom domains */ |
029190c5 PJ |
570 | int ndoms; /* number of sched domains in result */ |
571 | int nslot; /* next empty doms[] cpumask_t slot */ | |
572 | ||
573 | q = NULL; | |
574 | csa = NULL; | |
575 | doms = NULL; | |
1d3504fc | 576 | dattr = NULL; |
029190c5 PJ |
577 | |
578 | /* Special case for the 99% of systems with one, full, sched domain */ | |
579 | if (is_sched_load_balance(&top_cpuset)) { | |
580 | ndoms = 1; | |
581 | doms = kmalloc(sizeof(cpumask_t), GFP_KERNEL); | |
582 | if (!doms) | |
583 | goto rebuild; | |
1d3504fc HS |
584 | dattr = kmalloc(sizeof(struct sched_domain_attr), GFP_KERNEL); |
585 | if (dattr) { | |
586 | *dattr = SD_ATTR_INIT; | |
587 | update_domain_attr(dattr, &top_cpuset); | |
588 | } | |
029190c5 PJ |
589 | *doms = top_cpuset.cpus_allowed; |
590 | goto rebuild; | |
591 | } | |
592 | ||
593 | q = kfifo_alloc(number_of_cpusets * sizeof(cp), GFP_KERNEL, NULL); | |
594 | if (IS_ERR(q)) | |
595 | goto done; | |
596 | csa = kmalloc(number_of_cpusets * sizeof(cp), GFP_KERNEL); | |
597 | if (!csa) | |
598 | goto done; | |
599 | csn = 0; | |
600 | ||
601 | cp = &top_cpuset; | |
602 | __kfifo_put(q, (void *)&cp, sizeof(cp)); | |
603 | while (__kfifo_get(q, (void *)&cp, sizeof(cp))) { | |
604 | struct cgroup *cont; | |
605 | struct cpuset *child; /* scans child cpusets of cp */ | |
606 | if (is_sched_load_balance(cp)) | |
607 | csa[csn++] = cp; | |
608 | list_for_each_entry(cont, &cp->css.cgroup->children, sibling) { | |
609 | child = cgroup_cs(cont); | |
610 | __kfifo_put(q, (void *)&child, sizeof(cp)); | |
611 | } | |
612 | } | |
613 | ||
614 | for (i = 0; i < csn; i++) | |
615 | csa[i]->pn = i; | |
616 | ndoms = csn; | |
617 | ||
618 | restart: | |
619 | /* Find the best partition (set of sched domains) */ | |
620 | for (i = 0; i < csn; i++) { | |
621 | struct cpuset *a = csa[i]; | |
622 | int apn = a->pn; | |
623 | ||
624 | for (j = 0; j < csn; j++) { | |
625 | struct cpuset *b = csa[j]; | |
626 | int bpn = b->pn; | |
627 | ||
628 | if (apn != bpn && cpusets_overlap(a, b)) { | |
629 | for (k = 0; k < csn; k++) { | |
630 | struct cpuset *c = csa[k]; | |
631 | ||
632 | if (c->pn == bpn) | |
633 | c->pn = apn; | |
634 | } | |
635 | ndoms--; /* one less element */ | |
636 | goto restart; | |
637 | } | |
638 | } | |
639 | } | |
640 | ||
641 | /* Convert <csn, csa> to <ndoms, doms> */ | |
642 | doms = kmalloc(ndoms * sizeof(cpumask_t), GFP_KERNEL); | |
643 | if (!doms) | |
644 | goto rebuild; | |
1d3504fc | 645 | dattr = kmalloc(ndoms * sizeof(struct sched_domain_attr), GFP_KERNEL); |
029190c5 PJ |
646 | |
647 | for (nslot = 0, i = 0; i < csn; i++) { | |
648 | struct cpuset *a = csa[i]; | |
649 | int apn = a->pn; | |
650 | ||
651 | if (apn >= 0) { | |
652 | cpumask_t *dp = doms + nslot; | |
653 | ||
654 | if (nslot == ndoms) { | |
655 | static int warnings = 10; | |
656 | if (warnings) { | |
657 | printk(KERN_WARNING | |
658 | "rebuild_sched_domains confused:" | |
659 | " nslot %d, ndoms %d, csn %d, i %d," | |
660 | " apn %d\n", | |
661 | nslot, ndoms, csn, i, apn); | |
662 | warnings--; | |
663 | } | |
664 | continue; | |
665 | } | |
666 | ||
667 | cpus_clear(*dp); | |
1d3504fc HS |
668 | if (dattr) |
669 | *(dattr + nslot) = SD_ATTR_INIT; | |
029190c5 PJ |
670 | for (j = i; j < csn; j++) { |
671 | struct cpuset *b = csa[j]; | |
672 | ||
673 | if (apn == b->pn) { | |
674 | cpus_or(*dp, *dp, b->cpus_allowed); | |
675 | b->pn = -1; | |
1d3504fc | 676 | update_domain_attr(dattr, b); |
029190c5 PJ |
677 | } |
678 | } | |
679 | nslot++; | |
680 | } | |
681 | } | |
682 | BUG_ON(nslot != ndoms); | |
683 | ||
684 | rebuild: | |
685 | /* Have scheduler rebuild sched domains */ | |
86ef5c9a | 686 | get_online_cpus(); |
1d3504fc | 687 | partition_sched_domains(ndoms, doms, dattr); |
86ef5c9a | 688 | put_online_cpus(); |
029190c5 PJ |
689 | |
690 | done: | |
691 | if (q && !IS_ERR(q)) | |
692 | kfifo_free(q); | |
693 | kfree(csa); | |
694 | /* Don't kfree(doms) -- partition_sched_domains() does that. */ | |
1d3504fc | 695 | /* Don't kfree(dattr) -- partition_sched_domains() does that. */ |
029190c5 PJ |
696 | } |
697 | ||
8707d8b8 PM |
698 | static inline int started_after_time(struct task_struct *t1, |
699 | struct timespec *time, | |
700 | struct task_struct *t2) | |
701 | { | |
702 | int start_diff = timespec_compare(&t1->start_time, time); | |
703 | if (start_diff > 0) { | |
704 | return 1; | |
705 | } else if (start_diff < 0) { | |
706 | return 0; | |
707 | } else { | |
708 | /* | |
709 | * Arbitrarily, if two processes started at the same | |
710 | * time, we'll say that the lower pointer value | |
711 | * started first. Note that t2 may have exited by now | |
712 | * so this may not be a valid pointer any longer, but | |
713 | * that's fine - it still serves to distinguish | |
714 | * between two tasks started (effectively) | |
715 | * simultaneously. | |
716 | */ | |
717 | return t1 > t2; | |
718 | } | |
719 | } | |
720 | ||
721 | static inline int started_after(void *p1, void *p2) | |
722 | { | |
723 | struct task_struct *t1 = p1; | |
724 | struct task_struct *t2 = p2; | |
725 | return started_after_time(t1, &t2->start_time, t2); | |
726 | } | |
727 | ||
58f4790b CW |
728 | /** |
729 | * cpuset_test_cpumask - test a task's cpus_allowed versus its cpuset's | |
730 | * @tsk: task to test | |
731 | * @scan: struct cgroup_scanner contained in its struct cpuset_hotplug_scanner | |
732 | * | |
2df167a3 | 733 | * Call with cgroup_mutex held. May take callback_mutex during call. |
58f4790b CW |
734 | * Called for each task in a cgroup by cgroup_scan_tasks(). |
735 | * Return nonzero if this tasks's cpus_allowed mask should be changed (in other | |
736 | * words, if its mask is not equal to its cpuset's mask). | |
053199ed | 737 | */ |
58f4790b CW |
738 | int cpuset_test_cpumask(struct task_struct *tsk, struct cgroup_scanner *scan) |
739 | { | |
740 | return !cpus_equal(tsk->cpus_allowed, | |
741 | (cgroup_cs(scan->cg))->cpus_allowed); | |
742 | } | |
053199ed | 743 | |
58f4790b CW |
744 | /** |
745 | * cpuset_change_cpumask - make a task's cpus_allowed the same as its cpuset's | |
746 | * @tsk: task to test | |
747 | * @scan: struct cgroup_scanner containing the cgroup of the task | |
748 | * | |
749 | * Called by cgroup_scan_tasks() for each task in a cgroup whose | |
750 | * cpus_allowed mask needs to be changed. | |
751 | * | |
752 | * We don't need to re-check for the cgroup/cpuset membership, since we're | |
753 | * holding cgroup_lock() at this point. | |
754 | */ | |
755 | void cpuset_change_cpumask(struct task_struct *tsk, struct cgroup_scanner *scan) | |
756 | { | |
f9a86fcb | 757 | set_cpus_allowed_ptr(tsk, &((cgroup_cs(scan->cg))->cpus_allowed)); |
58f4790b CW |
758 | } |
759 | ||
760 | /** | |
761 | * update_cpumask - update the cpus_allowed mask of a cpuset and all tasks in it | |
762 | * @cs: the cpuset to consider | |
763 | * @buf: buffer of cpu numbers written to this cpuset | |
764 | */ | |
1da177e4 LT |
765 | static int update_cpumask(struct cpuset *cs, char *buf) |
766 | { | |
767 | struct cpuset trialcs; | |
58f4790b | 768 | struct cgroup_scanner scan; |
8707d8b8 | 769 | struct ptr_heap heap; |
58f4790b CW |
770 | int retval; |
771 | int is_load_balanced; | |
1da177e4 | 772 | |
4c4d50f7 PJ |
773 | /* top_cpuset.cpus_allowed tracks cpu_online_map; it's read-only */ |
774 | if (cs == &top_cpuset) | |
775 | return -EACCES; | |
776 | ||
1da177e4 | 777 | trialcs = *cs; |
6f7f02e7 DR |
778 | |
779 | /* | |
c8d9c90c | 780 | * An empty cpus_allowed is ok only if the cpuset has no tasks. |
020958b6 PJ |
781 | * Since cpulist_parse() fails on an empty mask, we special case |
782 | * that parsing. The validate_change() call ensures that cpusets | |
783 | * with tasks have cpus. | |
6f7f02e7 | 784 | */ |
020958b6 PJ |
785 | buf = strstrip(buf); |
786 | if (!*buf) { | |
6f7f02e7 DR |
787 | cpus_clear(trialcs.cpus_allowed); |
788 | } else { | |
789 | retval = cpulist_parse(buf, trialcs.cpus_allowed); | |
790 | if (retval < 0) | |
791 | return retval; | |
792 | } | |
1da177e4 | 793 | cpus_and(trialcs.cpus_allowed, trialcs.cpus_allowed, cpu_online_map); |
1da177e4 | 794 | retval = validate_change(cs, &trialcs); |
85d7b949 DG |
795 | if (retval < 0) |
796 | return retval; | |
029190c5 | 797 | |
8707d8b8 PM |
798 | /* Nothing to do if the cpus didn't change */ |
799 | if (cpus_equal(cs->cpus_allowed, trialcs.cpus_allowed)) | |
800 | return 0; | |
58f4790b | 801 | |
8707d8b8 PM |
802 | retval = heap_init(&heap, PAGE_SIZE, GFP_KERNEL, &started_after); |
803 | if (retval) | |
804 | return retval; | |
805 | ||
029190c5 PJ |
806 | is_load_balanced = is_sched_load_balance(&trialcs); |
807 | ||
3d3f26a7 | 808 | mutex_lock(&callback_mutex); |
85d7b949 | 809 | cs->cpus_allowed = trialcs.cpus_allowed; |
3d3f26a7 | 810 | mutex_unlock(&callback_mutex); |
029190c5 | 811 | |
8707d8b8 PM |
812 | /* |
813 | * Scan tasks in the cpuset, and update the cpumasks of any | |
58f4790b | 814 | * that need an update. |
8707d8b8 | 815 | */ |
58f4790b CW |
816 | scan.cg = cs->css.cgroup; |
817 | scan.test_task = cpuset_test_cpumask; | |
818 | scan.process_task = cpuset_change_cpumask; | |
819 | scan.heap = &heap; | |
820 | cgroup_scan_tasks(&scan); | |
8707d8b8 | 821 | heap_free(&heap); |
58f4790b | 822 | |
8707d8b8 | 823 | if (is_load_balanced) |
029190c5 | 824 | rebuild_sched_domains(); |
85d7b949 | 825 | return 0; |
1da177e4 LT |
826 | } |
827 | ||
e4e364e8 PJ |
828 | /* |
829 | * cpuset_migrate_mm | |
830 | * | |
831 | * Migrate memory region from one set of nodes to another. | |
832 | * | |
833 | * Temporarilly set tasks mems_allowed to target nodes of migration, | |
834 | * so that the migration code can allocate pages on these nodes. | |
835 | * | |
2df167a3 | 836 | * Call holding cgroup_mutex, so current's cpuset won't change |
c8d9c90c | 837 | * during this call, as manage_mutex holds off any cpuset_attach() |
e4e364e8 PJ |
838 | * calls. Therefore we don't need to take task_lock around the |
839 | * call to guarantee_online_mems(), as we know no one is changing | |
2df167a3 | 840 | * our task's cpuset. |
e4e364e8 PJ |
841 | * |
842 | * Hold callback_mutex around the two modifications of our tasks | |
843 | * mems_allowed to synchronize with cpuset_mems_allowed(). | |
844 | * | |
845 | * While the mm_struct we are migrating is typically from some | |
846 | * other task, the task_struct mems_allowed that we are hacking | |
847 | * is for our current task, which must allocate new pages for that | |
848 | * migrating memory region. | |
849 | * | |
850 | * We call cpuset_update_task_memory_state() before hacking | |
851 | * our tasks mems_allowed, so that we are assured of being in | |
852 | * sync with our tasks cpuset, and in particular, callbacks to | |
853 | * cpuset_update_task_memory_state() from nested page allocations | |
854 | * won't see any mismatch of our cpuset and task mems_generation | |
855 | * values, so won't overwrite our hacked tasks mems_allowed | |
856 | * nodemask. | |
857 | */ | |
858 | ||
859 | static void cpuset_migrate_mm(struct mm_struct *mm, const nodemask_t *from, | |
860 | const nodemask_t *to) | |
861 | { | |
862 | struct task_struct *tsk = current; | |
863 | ||
864 | cpuset_update_task_memory_state(); | |
865 | ||
866 | mutex_lock(&callback_mutex); | |
867 | tsk->mems_allowed = *to; | |
868 | mutex_unlock(&callback_mutex); | |
869 | ||
870 | do_migrate_pages(mm, from, to, MPOL_MF_MOVE_ALL); | |
871 | ||
872 | mutex_lock(&callback_mutex); | |
8793d854 | 873 | guarantee_online_mems(task_cs(tsk),&tsk->mems_allowed); |
e4e364e8 PJ |
874 | mutex_unlock(&callback_mutex); |
875 | } | |
876 | ||
053199ed | 877 | /* |
4225399a PJ |
878 | * Handle user request to change the 'mems' memory placement |
879 | * of a cpuset. Needs to validate the request, update the | |
880 | * cpusets mems_allowed and mems_generation, and for each | |
04c19fa6 PJ |
881 | * task in the cpuset, rebind any vma mempolicies and if |
882 | * the cpuset is marked 'memory_migrate', migrate the tasks | |
883 | * pages to the new memory. | |
4225399a | 884 | * |
2df167a3 | 885 | * Call with cgroup_mutex held. May take callback_mutex during call. |
4225399a PJ |
886 | * Will take tasklist_lock, scan tasklist for tasks in cpuset cs, |
887 | * lock each such tasks mm->mmap_sem, scan its vma's and rebind | |
888 | * their mempolicies to the cpusets new mems_allowed. | |
053199ed PJ |
889 | */ |
890 | ||
8793d854 PM |
891 | static void *cpuset_being_rebound; |
892 | ||
1da177e4 LT |
893 | static int update_nodemask(struct cpuset *cs, char *buf) |
894 | { | |
895 | struct cpuset trialcs; | |
04c19fa6 | 896 | nodemask_t oldmem; |
8793d854 | 897 | struct task_struct *p; |
4225399a PJ |
898 | struct mm_struct **mmarray; |
899 | int i, n, ntasks; | |
04c19fa6 | 900 | int migrate; |
4225399a | 901 | int fudge; |
1da177e4 | 902 | int retval; |
8793d854 | 903 | struct cgroup_iter it; |
1da177e4 | 904 | |
0e1e7c7a CL |
905 | /* |
906 | * top_cpuset.mems_allowed tracks node_stats[N_HIGH_MEMORY]; | |
907 | * it's read-only | |
908 | */ | |
38837fc7 PJ |
909 | if (cs == &top_cpuset) |
910 | return -EACCES; | |
911 | ||
1da177e4 | 912 | trialcs = *cs; |
6f7f02e7 DR |
913 | |
914 | /* | |
020958b6 PJ |
915 | * An empty mems_allowed is ok iff there are no tasks in the cpuset. |
916 | * Since nodelist_parse() fails on an empty mask, we special case | |
917 | * that parsing. The validate_change() call ensures that cpusets | |
918 | * with tasks have memory. | |
6f7f02e7 | 919 | */ |
020958b6 PJ |
920 | buf = strstrip(buf); |
921 | if (!*buf) { | |
6f7f02e7 DR |
922 | nodes_clear(trialcs.mems_allowed); |
923 | } else { | |
924 | retval = nodelist_parse(buf, trialcs.mems_allowed); | |
925 | if (retval < 0) | |
926 | goto done; | |
927 | } | |
0e1e7c7a CL |
928 | nodes_and(trialcs.mems_allowed, trialcs.mems_allowed, |
929 | node_states[N_HIGH_MEMORY]); | |
04c19fa6 PJ |
930 | oldmem = cs->mems_allowed; |
931 | if (nodes_equal(oldmem, trialcs.mems_allowed)) { | |
932 | retval = 0; /* Too easy - nothing to do */ | |
933 | goto done; | |
934 | } | |
59dac16f PJ |
935 | retval = validate_change(cs, &trialcs); |
936 | if (retval < 0) | |
937 | goto done; | |
938 | ||
3d3f26a7 | 939 | mutex_lock(&callback_mutex); |
59dac16f | 940 | cs->mems_allowed = trialcs.mems_allowed; |
151a4420 | 941 | cs->mems_generation = cpuset_mems_generation++; |
3d3f26a7 | 942 | mutex_unlock(&callback_mutex); |
59dac16f | 943 | |
8793d854 | 944 | cpuset_being_rebound = cs; /* causes mpol_copy() rebind */ |
4225399a PJ |
945 | |
946 | fudge = 10; /* spare mmarray[] slots */ | |
947 | fudge += cpus_weight(cs->cpus_allowed); /* imagine one fork-bomb/cpu */ | |
948 | retval = -ENOMEM; | |
949 | ||
950 | /* | |
951 | * Allocate mmarray[] to hold mm reference for each task | |
952 | * in cpuset cs. Can't kmalloc GFP_KERNEL while holding | |
953 | * tasklist_lock. We could use GFP_ATOMIC, but with a | |
954 | * few more lines of code, we can retry until we get a big | |
955 | * enough mmarray[] w/o using GFP_ATOMIC. | |
956 | */ | |
957 | while (1) { | |
8793d854 | 958 | ntasks = cgroup_task_count(cs->css.cgroup); /* guess */ |
4225399a PJ |
959 | ntasks += fudge; |
960 | mmarray = kmalloc(ntasks * sizeof(*mmarray), GFP_KERNEL); | |
961 | if (!mmarray) | |
962 | goto done; | |
c2aef333 | 963 | read_lock(&tasklist_lock); /* block fork */ |
8793d854 | 964 | if (cgroup_task_count(cs->css.cgroup) <= ntasks) |
4225399a | 965 | break; /* got enough */ |
c2aef333 | 966 | read_unlock(&tasklist_lock); /* try again */ |
4225399a PJ |
967 | kfree(mmarray); |
968 | } | |
969 | ||
970 | n = 0; | |
971 | ||
972 | /* Load up mmarray[] with mm reference for each task in cpuset. */ | |
8793d854 PM |
973 | cgroup_iter_start(cs->css.cgroup, &it); |
974 | while ((p = cgroup_iter_next(cs->css.cgroup, &it))) { | |
4225399a PJ |
975 | struct mm_struct *mm; |
976 | ||
977 | if (n >= ntasks) { | |
978 | printk(KERN_WARNING | |
979 | "Cpuset mempolicy rebind incomplete.\n"); | |
8793d854 | 980 | break; |
4225399a | 981 | } |
4225399a PJ |
982 | mm = get_task_mm(p); |
983 | if (!mm) | |
984 | continue; | |
985 | mmarray[n++] = mm; | |
8793d854 PM |
986 | } |
987 | cgroup_iter_end(cs->css.cgroup, &it); | |
c2aef333 | 988 | read_unlock(&tasklist_lock); |
4225399a PJ |
989 | |
990 | /* | |
991 | * Now that we've dropped the tasklist spinlock, we can | |
992 | * rebind the vma mempolicies of each mm in mmarray[] to their | |
993 | * new cpuset, and release that mm. The mpol_rebind_mm() | |
994 | * call takes mmap_sem, which we couldn't take while holding | |
995 | * tasklist_lock. Forks can happen again now - the mpol_copy() | |
996 | * cpuset_being_rebound check will catch such forks, and rebind | |
997 | * their vma mempolicies too. Because we still hold the global | |
2df167a3 | 998 | * cgroup_mutex, we know that no other rebind effort will |
4225399a PJ |
999 | * be contending for the global variable cpuset_being_rebound. |
1000 | * It's ok if we rebind the same mm twice; mpol_rebind_mm() | |
04c19fa6 | 1001 | * is idempotent. Also migrate pages in each mm to new nodes. |
4225399a | 1002 | */ |
04c19fa6 | 1003 | migrate = is_memory_migrate(cs); |
4225399a PJ |
1004 | for (i = 0; i < n; i++) { |
1005 | struct mm_struct *mm = mmarray[i]; | |
1006 | ||
1007 | mpol_rebind_mm(mm, &cs->mems_allowed); | |
e4e364e8 PJ |
1008 | if (migrate) |
1009 | cpuset_migrate_mm(mm, &oldmem, &cs->mems_allowed); | |
4225399a PJ |
1010 | mmput(mm); |
1011 | } | |
1012 | ||
2df167a3 | 1013 | /* We're done rebinding vmas to this cpuset's new mems_allowed. */ |
4225399a | 1014 | kfree(mmarray); |
8793d854 | 1015 | cpuset_being_rebound = NULL; |
4225399a | 1016 | retval = 0; |
59dac16f | 1017 | done: |
1da177e4 LT |
1018 | return retval; |
1019 | } | |
1020 | ||
8793d854 PM |
1021 | int current_cpuset_is_being_rebound(void) |
1022 | { | |
1023 | return task_cs(current) == cpuset_being_rebound; | |
1024 | } | |
1025 | ||
3e0d98b9 | 1026 | /* |
2df167a3 | 1027 | * Call with cgroup_mutex held. |
3e0d98b9 PJ |
1028 | */ |
1029 | ||
1030 | static int update_memory_pressure_enabled(struct cpuset *cs, char *buf) | |
1031 | { | |
1032 | if (simple_strtoul(buf, NULL, 10) != 0) | |
1033 | cpuset_memory_pressure_enabled = 1; | |
1034 | else | |
1035 | cpuset_memory_pressure_enabled = 0; | |
1036 | return 0; | |
1037 | } | |
1038 | ||
1d3504fc HS |
1039 | static int update_relax_domain_level(struct cpuset *cs, char *buf) |
1040 | { | |
1041 | int val = simple_strtol(buf, NULL, 10); | |
1042 | ||
1043 | if (val < 0) | |
1044 | val = -1; | |
1045 | ||
1046 | if (val != cs->relax_domain_level) { | |
1047 | cs->relax_domain_level = val; | |
1048 | rebuild_sched_domains(); | |
1049 | } | |
1050 | ||
1051 | return 0; | |
1052 | } | |
1053 | ||
1da177e4 LT |
1054 | /* |
1055 | * update_flag - read a 0 or a 1 in a file and update associated flag | |
1056 | * bit: the bit to update (CS_CPU_EXCLUSIVE, CS_MEM_EXCLUSIVE, | |
029190c5 | 1057 | * CS_SCHED_LOAD_BALANCE, |
825a46af PJ |
1058 | * CS_NOTIFY_ON_RELEASE, CS_MEMORY_MIGRATE, |
1059 | * CS_SPREAD_PAGE, CS_SPREAD_SLAB) | |
1da177e4 LT |
1060 | * cs: the cpuset to update |
1061 | * buf: the buffer where we read the 0 or 1 | |
053199ed | 1062 | * |
2df167a3 | 1063 | * Call with cgroup_mutex held. |
1da177e4 LT |
1064 | */ |
1065 | ||
1066 | static int update_flag(cpuset_flagbits_t bit, struct cpuset *cs, char *buf) | |
1067 | { | |
1068 | int turning_on; | |
1069 | struct cpuset trialcs; | |
607717a6 | 1070 | int err; |
029190c5 | 1071 | int cpus_nonempty, balance_flag_changed; |
1da177e4 LT |
1072 | |
1073 | turning_on = (simple_strtoul(buf, NULL, 10) != 0); | |
1074 | ||
1075 | trialcs = *cs; | |
1076 | if (turning_on) | |
1077 | set_bit(bit, &trialcs.flags); | |
1078 | else | |
1079 | clear_bit(bit, &trialcs.flags); | |
1080 | ||
1081 | err = validate_change(cs, &trialcs); | |
85d7b949 DG |
1082 | if (err < 0) |
1083 | return err; | |
029190c5 PJ |
1084 | |
1085 | cpus_nonempty = !cpus_empty(trialcs.cpus_allowed); | |
1086 | balance_flag_changed = (is_sched_load_balance(cs) != | |
1087 | is_sched_load_balance(&trialcs)); | |
1088 | ||
3d3f26a7 | 1089 | mutex_lock(&callback_mutex); |
69604067 | 1090 | cs->flags = trialcs.flags; |
3d3f26a7 | 1091 | mutex_unlock(&callback_mutex); |
85d7b949 | 1092 | |
029190c5 PJ |
1093 | if (cpus_nonempty && balance_flag_changed) |
1094 | rebuild_sched_domains(); | |
1095 | ||
85d7b949 | 1096 | return 0; |
1da177e4 LT |
1097 | } |
1098 | ||
3e0d98b9 | 1099 | /* |
80f7228b | 1100 | * Frequency meter - How fast is some event occurring? |
3e0d98b9 PJ |
1101 | * |
1102 | * These routines manage a digitally filtered, constant time based, | |
1103 | * event frequency meter. There are four routines: | |
1104 | * fmeter_init() - initialize a frequency meter. | |
1105 | * fmeter_markevent() - called each time the event happens. | |
1106 | * fmeter_getrate() - returns the recent rate of such events. | |
1107 | * fmeter_update() - internal routine used to update fmeter. | |
1108 | * | |
1109 | * A common data structure is passed to each of these routines, | |
1110 | * which is used to keep track of the state required to manage the | |
1111 | * frequency meter and its digital filter. | |
1112 | * | |
1113 | * The filter works on the number of events marked per unit time. | |
1114 | * The filter is single-pole low-pass recursive (IIR). The time unit | |
1115 | * is 1 second. Arithmetic is done using 32-bit integers scaled to | |
1116 | * simulate 3 decimal digits of precision (multiplied by 1000). | |
1117 | * | |
1118 | * With an FM_COEF of 933, and a time base of 1 second, the filter | |
1119 | * has a half-life of 10 seconds, meaning that if the events quit | |
1120 | * happening, then the rate returned from the fmeter_getrate() | |
1121 | * will be cut in half each 10 seconds, until it converges to zero. | |
1122 | * | |
1123 | * It is not worth doing a real infinitely recursive filter. If more | |
1124 | * than FM_MAXTICKS ticks have elapsed since the last filter event, | |
1125 | * just compute FM_MAXTICKS ticks worth, by which point the level | |
1126 | * will be stable. | |
1127 | * | |
1128 | * Limit the count of unprocessed events to FM_MAXCNT, so as to avoid | |
1129 | * arithmetic overflow in the fmeter_update() routine. | |
1130 | * | |
1131 | * Given the simple 32 bit integer arithmetic used, this meter works | |
1132 | * best for reporting rates between one per millisecond (msec) and | |
1133 | * one per 32 (approx) seconds. At constant rates faster than one | |
1134 | * per msec it maxes out at values just under 1,000,000. At constant | |
1135 | * rates between one per msec, and one per second it will stabilize | |
1136 | * to a value N*1000, where N is the rate of events per second. | |
1137 | * At constant rates between one per second and one per 32 seconds, | |
1138 | * it will be choppy, moving up on the seconds that have an event, | |
1139 | * and then decaying until the next event. At rates slower than | |
1140 | * about one in 32 seconds, it decays all the way back to zero between | |
1141 | * each event. | |
1142 | */ | |
1143 | ||
1144 | #define FM_COEF 933 /* coefficient for half-life of 10 secs */ | |
1145 | #define FM_MAXTICKS ((time_t)99) /* useless computing more ticks than this */ | |
1146 | #define FM_MAXCNT 1000000 /* limit cnt to avoid overflow */ | |
1147 | #define FM_SCALE 1000 /* faux fixed point scale */ | |
1148 | ||
1149 | /* Initialize a frequency meter */ | |
1150 | static void fmeter_init(struct fmeter *fmp) | |
1151 | { | |
1152 | fmp->cnt = 0; | |
1153 | fmp->val = 0; | |
1154 | fmp->time = 0; | |
1155 | spin_lock_init(&fmp->lock); | |
1156 | } | |
1157 | ||
1158 | /* Internal meter update - process cnt events and update value */ | |
1159 | static void fmeter_update(struct fmeter *fmp) | |
1160 | { | |
1161 | time_t now = get_seconds(); | |
1162 | time_t ticks = now - fmp->time; | |
1163 | ||
1164 | if (ticks == 0) | |
1165 | return; | |
1166 | ||
1167 | ticks = min(FM_MAXTICKS, ticks); | |
1168 | while (ticks-- > 0) | |
1169 | fmp->val = (FM_COEF * fmp->val) / FM_SCALE; | |
1170 | fmp->time = now; | |
1171 | ||
1172 | fmp->val += ((FM_SCALE - FM_COEF) * fmp->cnt) / FM_SCALE; | |
1173 | fmp->cnt = 0; | |
1174 | } | |
1175 | ||
1176 | /* Process any previous ticks, then bump cnt by one (times scale). */ | |
1177 | static void fmeter_markevent(struct fmeter *fmp) | |
1178 | { | |
1179 | spin_lock(&fmp->lock); | |
1180 | fmeter_update(fmp); | |
1181 | fmp->cnt = min(FM_MAXCNT, fmp->cnt + FM_SCALE); | |
1182 | spin_unlock(&fmp->lock); | |
1183 | } | |
1184 | ||
1185 | /* Process any previous ticks, then return current value. */ | |
1186 | static int fmeter_getrate(struct fmeter *fmp) | |
1187 | { | |
1188 | int val; | |
1189 | ||
1190 | spin_lock(&fmp->lock); | |
1191 | fmeter_update(fmp); | |
1192 | val = fmp->val; | |
1193 | spin_unlock(&fmp->lock); | |
1194 | return val; | |
1195 | } | |
1196 | ||
2df167a3 | 1197 | /* Called by cgroups to determine if a cpuset is usable; cgroup_mutex held */ |
8793d854 PM |
1198 | static int cpuset_can_attach(struct cgroup_subsys *ss, |
1199 | struct cgroup *cont, struct task_struct *tsk) | |
1da177e4 | 1200 | { |
8793d854 | 1201 | struct cpuset *cs = cgroup_cs(cont); |
1da177e4 | 1202 | |
1da177e4 LT |
1203 | if (cpus_empty(cs->cpus_allowed) || nodes_empty(cs->mems_allowed)) |
1204 | return -ENOSPC; | |
1205 | ||
8793d854 PM |
1206 | return security_task_setscheduler(tsk, 0, NULL); |
1207 | } | |
1da177e4 | 1208 | |
8793d854 PM |
1209 | static void cpuset_attach(struct cgroup_subsys *ss, |
1210 | struct cgroup *cont, struct cgroup *oldcont, | |
1211 | struct task_struct *tsk) | |
1212 | { | |
1213 | cpumask_t cpus; | |
1214 | nodemask_t from, to; | |
1215 | struct mm_struct *mm; | |
1216 | struct cpuset *cs = cgroup_cs(cont); | |
1217 | struct cpuset *oldcs = cgroup_cs(oldcont); | |
22fb52dd | 1218 | |
3d3f26a7 | 1219 | mutex_lock(&callback_mutex); |
1da177e4 | 1220 | guarantee_online_cpus(cs, &cpus); |
f9a86fcb | 1221 | set_cpus_allowed_ptr(tsk, &cpus); |
8793d854 | 1222 | mutex_unlock(&callback_mutex); |
1da177e4 | 1223 | |
45b07ef3 PJ |
1224 | from = oldcs->mems_allowed; |
1225 | to = cs->mems_allowed; | |
4225399a PJ |
1226 | mm = get_task_mm(tsk); |
1227 | if (mm) { | |
1228 | mpol_rebind_mm(mm, &to); | |
2741a559 | 1229 | if (is_memory_migrate(cs)) |
e4e364e8 | 1230 | cpuset_migrate_mm(mm, &from, &to); |
4225399a PJ |
1231 | mmput(mm); |
1232 | } | |
1233 | ||
1da177e4 LT |
1234 | } |
1235 | ||
1236 | /* The various types of files and directories in a cpuset file system */ | |
1237 | ||
1238 | typedef enum { | |
45b07ef3 | 1239 | FILE_MEMORY_MIGRATE, |
1da177e4 LT |
1240 | FILE_CPULIST, |
1241 | FILE_MEMLIST, | |
1242 | FILE_CPU_EXCLUSIVE, | |
1243 | FILE_MEM_EXCLUSIVE, | |
029190c5 | 1244 | FILE_SCHED_LOAD_BALANCE, |
1d3504fc | 1245 | FILE_SCHED_RELAX_DOMAIN_LEVEL, |
3e0d98b9 PJ |
1246 | FILE_MEMORY_PRESSURE_ENABLED, |
1247 | FILE_MEMORY_PRESSURE, | |
825a46af PJ |
1248 | FILE_SPREAD_PAGE, |
1249 | FILE_SPREAD_SLAB, | |
1da177e4 LT |
1250 | } cpuset_filetype_t; |
1251 | ||
8793d854 PM |
1252 | static ssize_t cpuset_common_file_write(struct cgroup *cont, |
1253 | struct cftype *cft, | |
1254 | struct file *file, | |
d3ed11c3 | 1255 | const char __user *userbuf, |
1da177e4 LT |
1256 | size_t nbytes, loff_t *unused_ppos) |
1257 | { | |
8793d854 | 1258 | struct cpuset *cs = cgroup_cs(cont); |
1da177e4 LT |
1259 | cpuset_filetype_t type = cft->private; |
1260 | char *buffer; | |
1261 | int retval = 0; | |
1262 | ||
1263 | /* Crude upper limit on largest legitimate cpulist user might write. */ | |
029190c5 | 1264 | if (nbytes > 100U + 6 * max(NR_CPUS, MAX_NUMNODES)) |
1da177e4 LT |
1265 | return -E2BIG; |
1266 | ||
1267 | /* +1 for nul-terminator */ | |
1268 | if ((buffer = kmalloc(nbytes + 1, GFP_KERNEL)) == 0) | |
1269 | return -ENOMEM; | |
1270 | ||
1271 | if (copy_from_user(buffer, userbuf, nbytes)) { | |
1272 | retval = -EFAULT; | |
1273 | goto out1; | |
1274 | } | |
1275 | buffer[nbytes] = 0; /* nul-terminate */ | |
1276 | ||
8793d854 | 1277 | cgroup_lock(); |
1da177e4 | 1278 | |
8793d854 | 1279 | if (cgroup_is_removed(cont)) { |
1da177e4 LT |
1280 | retval = -ENODEV; |
1281 | goto out2; | |
1282 | } | |
1283 | ||
1284 | switch (type) { | |
1285 | case FILE_CPULIST: | |
1286 | retval = update_cpumask(cs, buffer); | |
1287 | break; | |
1288 | case FILE_MEMLIST: | |
1289 | retval = update_nodemask(cs, buffer); | |
1290 | break; | |
1291 | case FILE_CPU_EXCLUSIVE: | |
1292 | retval = update_flag(CS_CPU_EXCLUSIVE, cs, buffer); | |
1293 | break; | |
1294 | case FILE_MEM_EXCLUSIVE: | |
1295 | retval = update_flag(CS_MEM_EXCLUSIVE, cs, buffer); | |
1296 | break; | |
029190c5 PJ |
1297 | case FILE_SCHED_LOAD_BALANCE: |
1298 | retval = update_flag(CS_SCHED_LOAD_BALANCE, cs, buffer); | |
1299 | break; | |
1d3504fc HS |
1300 | case FILE_SCHED_RELAX_DOMAIN_LEVEL: |
1301 | retval = update_relax_domain_level(cs, buffer); | |
1302 | break; | |
45b07ef3 PJ |
1303 | case FILE_MEMORY_MIGRATE: |
1304 | retval = update_flag(CS_MEMORY_MIGRATE, cs, buffer); | |
1305 | break; | |
3e0d98b9 PJ |
1306 | case FILE_MEMORY_PRESSURE_ENABLED: |
1307 | retval = update_memory_pressure_enabled(cs, buffer); | |
1308 | break; | |
1309 | case FILE_MEMORY_PRESSURE: | |
1310 | retval = -EACCES; | |
1311 | break; | |
825a46af PJ |
1312 | case FILE_SPREAD_PAGE: |
1313 | retval = update_flag(CS_SPREAD_PAGE, cs, buffer); | |
151a4420 | 1314 | cs->mems_generation = cpuset_mems_generation++; |
825a46af PJ |
1315 | break; |
1316 | case FILE_SPREAD_SLAB: | |
1317 | retval = update_flag(CS_SPREAD_SLAB, cs, buffer); | |
151a4420 | 1318 | cs->mems_generation = cpuset_mems_generation++; |
825a46af | 1319 | break; |
1da177e4 LT |
1320 | default: |
1321 | retval = -EINVAL; | |
1322 | goto out2; | |
1323 | } | |
1324 | ||
1325 | if (retval == 0) | |
1326 | retval = nbytes; | |
1327 | out2: | |
8793d854 | 1328 | cgroup_unlock(); |
1da177e4 LT |
1329 | out1: |
1330 | kfree(buffer); | |
1331 | return retval; | |
1332 | } | |
1333 | ||
1da177e4 LT |
1334 | /* |
1335 | * These ascii lists should be read in a single call, by using a user | |
1336 | * buffer large enough to hold the entire map. If read in smaller | |
1337 | * chunks, there is no guarantee of atomicity. Since the display format | |
1338 | * used, list of ranges of sequential numbers, is variable length, | |
1339 | * and since these maps can change value dynamically, one could read | |
1340 | * gibberish by doing partial reads while a list was changing. | |
1341 | * A single large read to a buffer that crosses a page boundary is | |
1342 | * ok, because the result being copied to user land is not recomputed | |
1343 | * across a page fault. | |
1344 | */ | |
1345 | ||
1346 | static int cpuset_sprintf_cpulist(char *page, struct cpuset *cs) | |
1347 | { | |
1348 | cpumask_t mask; | |
1349 | ||
3d3f26a7 | 1350 | mutex_lock(&callback_mutex); |
1da177e4 | 1351 | mask = cs->cpus_allowed; |
3d3f26a7 | 1352 | mutex_unlock(&callback_mutex); |
1da177e4 LT |
1353 | |
1354 | return cpulist_scnprintf(page, PAGE_SIZE, mask); | |
1355 | } | |
1356 | ||
1357 | static int cpuset_sprintf_memlist(char *page, struct cpuset *cs) | |
1358 | { | |
1359 | nodemask_t mask; | |
1360 | ||
3d3f26a7 | 1361 | mutex_lock(&callback_mutex); |
1da177e4 | 1362 | mask = cs->mems_allowed; |
3d3f26a7 | 1363 | mutex_unlock(&callback_mutex); |
1da177e4 LT |
1364 | |
1365 | return nodelist_scnprintf(page, PAGE_SIZE, mask); | |
1366 | } | |
1367 | ||
8793d854 PM |
1368 | static ssize_t cpuset_common_file_read(struct cgroup *cont, |
1369 | struct cftype *cft, | |
1370 | struct file *file, | |
1371 | char __user *buf, | |
1372 | size_t nbytes, loff_t *ppos) | |
1da177e4 | 1373 | { |
8793d854 | 1374 | struct cpuset *cs = cgroup_cs(cont); |
1da177e4 LT |
1375 | cpuset_filetype_t type = cft->private; |
1376 | char *page; | |
1377 | ssize_t retval = 0; | |
1378 | char *s; | |
1da177e4 | 1379 | |
e12ba74d | 1380 | if (!(page = (char *)__get_free_page(GFP_TEMPORARY))) |
1da177e4 LT |
1381 | return -ENOMEM; |
1382 | ||
1383 | s = page; | |
1384 | ||
1385 | switch (type) { | |
1386 | case FILE_CPULIST: | |
1387 | s += cpuset_sprintf_cpulist(s, cs); | |
1388 | break; | |
1389 | case FILE_MEMLIST: | |
1390 | s += cpuset_sprintf_memlist(s, cs); | |
1391 | break; | |
1392 | case FILE_CPU_EXCLUSIVE: | |
1393 | *s++ = is_cpu_exclusive(cs) ? '1' : '0'; | |
1394 | break; | |
1395 | case FILE_MEM_EXCLUSIVE: | |
1396 | *s++ = is_mem_exclusive(cs) ? '1' : '0'; | |
1397 | break; | |
029190c5 PJ |
1398 | case FILE_SCHED_LOAD_BALANCE: |
1399 | *s++ = is_sched_load_balance(cs) ? '1' : '0'; | |
1400 | break; | |
1d3504fc HS |
1401 | case FILE_SCHED_RELAX_DOMAIN_LEVEL: |
1402 | s += sprintf(s, "%d", cs->relax_domain_level); | |
1403 | break; | |
45b07ef3 PJ |
1404 | case FILE_MEMORY_MIGRATE: |
1405 | *s++ = is_memory_migrate(cs) ? '1' : '0'; | |
1406 | break; | |
3e0d98b9 PJ |
1407 | case FILE_MEMORY_PRESSURE_ENABLED: |
1408 | *s++ = cpuset_memory_pressure_enabled ? '1' : '0'; | |
1409 | break; | |
1410 | case FILE_MEMORY_PRESSURE: | |
1411 | s += sprintf(s, "%d", fmeter_getrate(&cs->fmeter)); | |
1412 | break; | |
825a46af PJ |
1413 | case FILE_SPREAD_PAGE: |
1414 | *s++ = is_spread_page(cs) ? '1' : '0'; | |
1415 | break; | |
1416 | case FILE_SPREAD_SLAB: | |
1417 | *s++ = is_spread_slab(cs) ? '1' : '0'; | |
1418 | break; | |
1da177e4 LT |
1419 | default: |
1420 | retval = -EINVAL; | |
1421 | goto out; | |
1422 | } | |
1423 | *s++ = '\n'; | |
1da177e4 | 1424 | |
eacaa1f5 | 1425 | retval = simple_read_from_buffer(buf, nbytes, ppos, page, s - page); |
1da177e4 LT |
1426 | out: |
1427 | free_page((unsigned long)page); | |
1428 | return retval; | |
1429 | } | |
1430 | ||
1da177e4 | 1431 | |
1da177e4 | 1432 | |
1da177e4 | 1433 | |
1da177e4 LT |
1434 | |
1435 | /* | |
1436 | * for the common functions, 'private' gives the type of file | |
1437 | */ | |
1438 | ||
1da177e4 LT |
1439 | static struct cftype cft_cpus = { |
1440 | .name = "cpus", | |
8793d854 PM |
1441 | .read = cpuset_common_file_read, |
1442 | .write = cpuset_common_file_write, | |
1da177e4 LT |
1443 | .private = FILE_CPULIST, |
1444 | }; | |
1445 | ||
1446 | static struct cftype cft_mems = { | |
1447 | .name = "mems", | |
8793d854 PM |
1448 | .read = cpuset_common_file_read, |
1449 | .write = cpuset_common_file_write, | |
1da177e4 LT |
1450 | .private = FILE_MEMLIST, |
1451 | }; | |
1452 | ||
1453 | static struct cftype cft_cpu_exclusive = { | |
1454 | .name = "cpu_exclusive", | |
8793d854 PM |
1455 | .read = cpuset_common_file_read, |
1456 | .write = cpuset_common_file_write, | |
1da177e4 LT |
1457 | .private = FILE_CPU_EXCLUSIVE, |
1458 | }; | |
1459 | ||
1460 | static struct cftype cft_mem_exclusive = { | |
1461 | .name = "mem_exclusive", | |
8793d854 PM |
1462 | .read = cpuset_common_file_read, |
1463 | .write = cpuset_common_file_write, | |
1da177e4 LT |
1464 | .private = FILE_MEM_EXCLUSIVE, |
1465 | }; | |
1466 | ||
029190c5 PJ |
1467 | static struct cftype cft_sched_load_balance = { |
1468 | .name = "sched_load_balance", | |
1469 | .read = cpuset_common_file_read, | |
1470 | .write = cpuset_common_file_write, | |
1471 | .private = FILE_SCHED_LOAD_BALANCE, | |
1472 | }; | |
1473 | ||
1d3504fc HS |
1474 | static struct cftype cft_sched_relax_domain_level = { |
1475 | .name = "sched_relax_domain_level", | |
1476 | .read = cpuset_common_file_read, | |
1477 | .write = cpuset_common_file_write, | |
1478 | .private = FILE_SCHED_RELAX_DOMAIN_LEVEL, | |
1479 | }; | |
1480 | ||
45b07ef3 PJ |
1481 | static struct cftype cft_memory_migrate = { |
1482 | .name = "memory_migrate", | |
8793d854 PM |
1483 | .read = cpuset_common_file_read, |
1484 | .write = cpuset_common_file_write, | |
45b07ef3 PJ |
1485 | .private = FILE_MEMORY_MIGRATE, |
1486 | }; | |
1487 | ||
3e0d98b9 PJ |
1488 | static struct cftype cft_memory_pressure_enabled = { |
1489 | .name = "memory_pressure_enabled", | |
8793d854 PM |
1490 | .read = cpuset_common_file_read, |
1491 | .write = cpuset_common_file_write, | |
3e0d98b9 PJ |
1492 | .private = FILE_MEMORY_PRESSURE_ENABLED, |
1493 | }; | |
1494 | ||
1495 | static struct cftype cft_memory_pressure = { | |
1496 | .name = "memory_pressure", | |
8793d854 PM |
1497 | .read = cpuset_common_file_read, |
1498 | .write = cpuset_common_file_write, | |
3e0d98b9 PJ |
1499 | .private = FILE_MEMORY_PRESSURE, |
1500 | }; | |
1501 | ||
825a46af PJ |
1502 | static struct cftype cft_spread_page = { |
1503 | .name = "memory_spread_page", | |
8793d854 PM |
1504 | .read = cpuset_common_file_read, |
1505 | .write = cpuset_common_file_write, | |
825a46af PJ |
1506 | .private = FILE_SPREAD_PAGE, |
1507 | }; | |
1508 | ||
1509 | static struct cftype cft_spread_slab = { | |
1510 | .name = "memory_spread_slab", | |
8793d854 PM |
1511 | .read = cpuset_common_file_read, |
1512 | .write = cpuset_common_file_write, | |
825a46af PJ |
1513 | .private = FILE_SPREAD_SLAB, |
1514 | }; | |
1515 | ||
8793d854 | 1516 | static int cpuset_populate(struct cgroup_subsys *ss, struct cgroup *cont) |
1da177e4 LT |
1517 | { |
1518 | int err; | |
1519 | ||
8793d854 | 1520 | if ((err = cgroup_add_file(cont, ss, &cft_cpus)) < 0) |
1da177e4 | 1521 | return err; |
8793d854 | 1522 | if ((err = cgroup_add_file(cont, ss, &cft_mems)) < 0) |
1da177e4 | 1523 | return err; |
8793d854 | 1524 | if ((err = cgroup_add_file(cont, ss, &cft_cpu_exclusive)) < 0) |
1da177e4 | 1525 | return err; |
8793d854 | 1526 | if ((err = cgroup_add_file(cont, ss, &cft_mem_exclusive)) < 0) |
1da177e4 | 1527 | return err; |
8793d854 | 1528 | if ((err = cgroup_add_file(cont, ss, &cft_memory_migrate)) < 0) |
1da177e4 | 1529 | return err; |
029190c5 PJ |
1530 | if ((err = cgroup_add_file(cont, ss, &cft_sched_load_balance)) < 0) |
1531 | return err; | |
1d3504fc HS |
1532 | if ((err = cgroup_add_file(cont, ss, |
1533 | &cft_sched_relax_domain_level)) < 0) | |
1534 | return err; | |
8793d854 | 1535 | if ((err = cgroup_add_file(cont, ss, &cft_memory_pressure)) < 0) |
45b07ef3 | 1536 | return err; |
8793d854 | 1537 | if ((err = cgroup_add_file(cont, ss, &cft_spread_page)) < 0) |
3e0d98b9 | 1538 | return err; |
8793d854 | 1539 | if ((err = cgroup_add_file(cont, ss, &cft_spread_slab)) < 0) |
1da177e4 | 1540 | return err; |
8793d854 PM |
1541 | /* memory_pressure_enabled is in root cpuset only */ |
1542 | if (err == 0 && !cont->parent) | |
1543 | err = cgroup_add_file(cont, ss, | |
1544 | &cft_memory_pressure_enabled); | |
1da177e4 LT |
1545 | return 0; |
1546 | } | |
1547 | ||
8793d854 PM |
1548 | /* |
1549 | * post_clone() is called at the end of cgroup_clone(). | |
1550 | * 'cgroup' was just created automatically as a result of | |
1551 | * a cgroup_clone(), and the current task is about to | |
1552 | * be moved into 'cgroup'. | |
1553 | * | |
1554 | * Currently we refuse to set up the cgroup - thereby | |
1555 | * refusing the task to be entered, and as a result refusing | |
1556 | * the sys_unshare() or clone() which initiated it - if any | |
1557 | * sibling cpusets have exclusive cpus or mem. | |
1558 | * | |
1559 | * If this becomes a problem for some users who wish to | |
1560 | * allow that scenario, then cpuset_post_clone() could be | |
1561 | * changed to grant parent->cpus_allowed-sibling_cpus_exclusive | |
2df167a3 PM |
1562 | * (and likewise for mems) to the new cgroup. Called with cgroup_mutex |
1563 | * held. | |
8793d854 PM |
1564 | */ |
1565 | static void cpuset_post_clone(struct cgroup_subsys *ss, | |
1566 | struct cgroup *cgroup) | |
1567 | { | |
1568 | struct cgroup *parent, *child; | |
1569 | struct cpuset *cs, *parent_cs; | |
1570 | ||
1571 | parent = cgroup->parent; | |
1572 | list_for_each_entry(child, &parent->children, sibling) { | |
1573 | cs = cgroup_cs(child); | |
1574 | if (is_mem_exclusive(cs) || is_cpu_exclusive(cs)) | |
1575 | return; | |
1576 | } | |
1577 | cs = cgroup_cs(cgroup); | |
1578 | parent_cs = cgroup_cs(parent); | |
1579 | ||
1580 | cs->mems_allowed = parent_cs->mems_allowed; | |
1581 | cs->cpus_allowed = parent_cs->cpus_allowed; | |
1582 | return; | |
1583 | } | |
1584 | ||
1da177e4 LT |
1585 | /* |
1586 | * cpuset_create - create a cpuset | |
2df167a3 PM |
1587 | * ss: cpuset cgroup subsystem |
1588 | * cont: control group that the new cpuset will be part of | |
1da177e4 LT |
1589 | */ |
1590 | ||
8793d854 PM |
1591 | static struct cgroup_subsys_state *cpuset_create( |
1592 | struct cgroup_subsys *ss, | |
1593 | struct cgroup *cont) | |
1da177e4 LT |
1594 | { |
1595 | struct cpuset *cs; | |
8793d854 | 1596 | struct cpuset *parent; |
1da177e4 | 1597 | |
8793d854 PM |
1598 | if (!cont->parent) { |
1599 | /* This is early initialization for the top cgroup */ | |
1600 | top_cpuset.mems_generation = cpuset_mems_generation++; | |
1601 | return &top_cpuset.css; | |
1602 | } | |
1603 | parent = cgroup_cs(cont->parent); | |
1da177e4 LT |
1604 | cs = kmalloc(sizeof(*cs), GFP_KERNEL); |
1605 | if (!cs) | |
8793d854 | 1606 | return ERR_PTR(-ENOMEM); |
1da177e4 | 1607 | |
cf2a473c | 1608 | cpuset_update_task_memory_state(); |
1da177e4 | 1609 | cs->flags = 0; |
825a46af PJ |
1610 | if (is_spread_page(parent)) |
1611 | set_bit(CS_SPREAD_PAGE, &cs->flags); | |
1612 | if (is_spread_slab(parent)) | |
1613 | set_bit(CS_SPREAD_SLAB, &cs->flags); | |
029190c5 | 1614 | set_bit(CS_SCHED_LOAD_BALANCE, &cs->flags); |
f9a86fcb MT |
1615 | cpus_clear(cs->cpus_allowed); |
1616 | nodes_clear(cs->mems_allowed); | |
151a4420 | 1617 | cs->mems_generation = cpuset_mems_generation++; |
3e0d98b9 | 1618 | fmeter_init(&cs->fmeter); |
1d3504fc | 1619 | cs->relax_domain_level = -1; |
1da177e4 LT |
1620 | |
1621 | cs->parent = parent; | |
202f72d5 | 1622 | number_of_cpusets++; |
8793d854 | 1623 | return &cs->css ; |
1da177e4 LT |
1624 | } |
1625 | ||
029190c5 PJ |
1626 | /* |
1627 | * Locking note on the strange update_flag() call below: | |
1628 | * | |
1629 | * If the cpuset being removed has its flag 'sched_load_balance' | |
1630 | * enabled, then simulate turning sched_load_balance off, which | |
86ef5c9a | 1631 | * will call rebuild_sched_domains(). The get_online_cpus() |
029190c5 PJ |
1632 | * call in rebuild_sched_domains() must not be made while holding |
1633 | * callback_mutex. Elsewhere the kernel nests callback_mutex inside | |
86ef5c9a | 1634 | * get_online_cpus() calls. So the reverse nesting would risk an |
029190c5 PJ |
1635 | * ABBA deadlock. |
1636 | */ | |
1637 | ||
8793d854 | 1638 | static void cpuset_destroy(struct cgroup_subsys *ss, struct cgroup *cont) |
1da177e4 | 1639 | { |
8793d854 | 1640 | struct cpuset *cs = cgroup_cs(cont); |
1da177e4 | 1641 | |
cf2a473c | 1642 | cpuset_update_task_memory_state(); |
029190c5 PJ |
1643 | |
1644 | if (is_sched_load_balance(cs)) | |
1645 | update_flag(CS_SCHED_LOAD_BALANCE, cs, "0"); | |
1646 | ||
202f72d5 | 1647 | number_of_cpusets--; |
8793d854 | 1648 | kfree(cs); |
1da177e4 LT |
1649 | } |
1650 | ||
8793d854 PM |
1651 | struct cgroup_subsys cpuset_subsys = { |
1652 | .name = "cpuset", | |
1653 | .create = cpuset_create, | |
1654 | .destroy = cpuset_destroy, | |
1655 | .can_attach = cpuset_can_attach, | |
1656 | .attach = cpuset_attach, | |
1657 | .populate = cpuset_populate, | |
1658 | .post_clone = cpuset_post_clone, | |
1659 | .subsys_id = cpuset_subsys_id, | |
1660 | .early_init = 1, | |
1661 | }; | |
1662 | ||
c417f024 PJ |
1663 | /* |
1664 | * cpuset_init_early - just enough so that the calls to | |
1665 | * cpuset_update_task_memory_state() in early init code | |
1666 | * are harmless. | |
1667 | */ | |
1668 | ||
1669 | int __init cpuset_init_early(void) | |
1670 | { | |
8793d854 | 1671 | top_cpuset.mems_generation = cpuset_mems_generation++; |
c417f024 PJ |
1672 | return 0; |
1673 | } | |
1674 | ||
8793d854 | 1675 | |
1da177e4 LT |
1676 | /** |
1677 | * cpuset_init - initialize cpusets at system boot | |
1678 | * | |
1679 | * Description: Initialize top_cpuset and the cpuset internal file system, | |
1680 | **/ | |
1681 | ||
1682 | int __init cpuset_init(void) | |
1683 | { | |
8793d854 | 1684 | int err = 0; |
1da177e4 | 1685 | |
f9a86fcb MT |
1686 | cpus_setall(top_cpuset.cpus_allowed); |
1687 | nodes_setall(top_cpuset.mems_allowed); | |
1da177e4 | 1688 | |
3e0d98b9 | 1689 | fmeter_init(&top_cpuset.fmeter); |
151a4420 | 1690 | top_cpuset.mems_generation = cpuset_mems_generation++; |
029190c5 | 1691 | set_bit(CS_SCHED_LOAD_BALANCE, &top_cpuset.flags); |
1d3504fc | 1692 | top_cpuset.relax_domain_level = -1; |
1da177e4 | 1693 | |
1da177e4 LT |
1694 | err = register_filesystem(&cpuset_fs_type); |
1695 | if (err < 0) | |
8793d854 PM |
1696 | return err; |
1697 | ||
202f72d5 | 1698 | number_of_cpusets = 1; |
8793d854 | 1699 | return 0; |
1da177e4 LT |
1700 | } |
1701 | ||
956db3ca CW |
1702 | /** |
1703 | * cpuset_do_move_task - move a given task to another cpuset | |
1704 | * @tsk: pointer to task_struct the task to move | |
1705 | * @scan: struct cgroup_scanner contained in its struct cpuset_hotplug_scanner | |
1706 | * | |
1707 | * Called by cgroup_scan_tasks() for each task in a cgroup. | |
1708 | * Return nonzero to stop the walk through the tasks. | |
1709 | */ | |
1710 | void cpuset_do_move_task(struct task_struct *tsk, struct cgroup_scanner *scan) | |
1711 | { | |
1712 | struct cpuset_hotplug_scanner *chsp; | |
1713 | ||
1714 | chsp = container_of(scan, struct cpuset_hotplug_scanner, scan); | |
1715 | cgroup_attach_task(chsp->to, tsk); | |
1716 | } | |
1717 | ||
1718 | /** | |
1719 | * move_member_tasks_to_cpuset - move tasks from one cpuset to another | |
1720 | * @from: cpuset in which the tasks currently reside | |
1721 | * @to: cpuset to which the tasks will be moved | |
1722 | * | |
c8d9c90c PJ |
1723 | * Called with cgroup_mutex held |
1724 | * callback_mutex must not be held, as cpuset_attach() will take it. | |
956db3ca CW |
1725 | * |
1726 | * The cgroup_scan_tasks() function will scan all the tasks in a cgroup, | |
1727 | * calling callback functions for each. | |
1728 | */ | |
1729 | static void move_member_tasks_to_cpuset(struct cpuset *from, struct cpuset *to) | |
1730 | { | |
1731 | struct cpuset_hotplug_scanner scan; | |
1732 | ||
1733 | scan.scan.cg = from->css.cgroup; | |
1734 | scan.scan.test_task = NULL; /* select all tasks in cgroup */ | |
1735 | scan.scan.process_task = cpuset_do_move_task; | |
1736 | scan.scan.heap = NULL; | |
1737 | scan.to = to->css.cgroup; | |
1738 | ||
1739 | if (cgroup_scan_tasks((struct cgroup_scanner *)&scan)) | |
1740 | printk(KERN_ERR "move_member_tasks_to_cpuset: " | |
1741 | "cgroup_scan_tasks failed\n"); | |
1742 | } | |
1743 | ||
b1aac8bb PJ |
1744 | /* |
1745 | * If common_cpu_mem_hotplug_unplug(), below, unplugs any CPUs | |
1746 | * or memory nodes, we need to walk over the cpuset hierarchy, | |
1747 | * removing that CPU or node from all cpusets. If this removes the | |
956db3ca CW |
1748 | * last CPU or node from a cpuset, then move the tasks in the empty |
1749 | * cpuset to its next-highest non-empty parent. | |
b1aac8bb | 1750 | * |
c8d9c90c PJ |
1751 | * Called with cgroup_mutex held |
1752 | * callback_mutex must not be held, as cpuset_attach() will take it. | |
b1aac8bb | 1753 | */ |
956db3ca CW |
1754 | static void remove_tasks_in_empty_cpuset(struct cpuset *cs) |
1755 | { | |
1756 | struct cpuset *parent; | |
1757 | ||
c8d9c90c PJ |
1758 | /* |
1759 | * The cgroup's css_sets list is in use if there are tasks | |
1760 | * in the cpuset; the list is empty if there are none; | |
1761 | * the cs->css.refcnt seems always 0. | |
1762 | */ | |
956db3ca CW |
1763 | if (list_empty(&cs->css.cgroup->css_sets)) |
1764 | return; | |
b1aac8bb | 1765 | |
956db3ca CW |
1766 | /* |
1767 | * Find its next-highest non-empty parent, (top cpuset | |
1768 | * has online cpus, so can't be empty). | |
1769 | */ | |
1770 | parent = cs->parent; | |
b4501295 PJ |
1771 | while (cpus_empty(parent->cpus_allowed) || |
1772 | nodes_empty(parent->mems_allowed)) | |
956db3ca | 1773 | parent = parent->parent; |
956db3ca CW |
1774 | |
1775 | move_member_tasks_to_cpuset(cs, parent); | |
1776 | } | |
1777 | ||
1778 | /* | |
1779 | * Walk the specified cpuset subtree and look for empty cpusets. | |
1780 | * The tasks of such cpuset must be moved to a parent cpuset. | |
1781 | * | |
2df167a3 | 1782 | * Called with cgroup_mutex held. We take callback_mutex to modify |
956db3ca CW |
1783 | * cpus_allowed and mems_allowed. |
1784 | * | |
1785 | * This walk processes the tree from top to bottom, completing one layer | |
1786 | * before dropping down to the next. It always processes a node before | |
1787 | * any of its children. | |
1788 | * | |
1789 | * For now, since we lack memory hot unplug, we'll never see a cpuset | |
1790 | * that has tasks along with an empty 'mems'. But if we did see such | |
1791 | * a cpuset, we'd handle it just like we do if its 'cpus' was empty. | |
1792 | */ | |
1793 | static void scan_for_empty_cpusets(const struct cpuset *root) | |
b1aac8bb | 1794 | { |
956db3ca CW |
1795 | struct cpuset *cp; /* scans cpusets being updated */ |
1796 | struct cpuset *child; /* scans child cpusets of cp */ | |
1797 | struct list_head queue; | |
8793d854 | 1798 | struct cgroup *cont; |
b1aac8bb | 1799 | |
956db3ca CW |
1800 | INIT_LIST_HEAD(&queue); |
1801 | ||
1802 | list_add_tail((struct list_head *)&root->stack_list, &queue); | |
1803 | ||
956db3ca CW |
1804 | while (!list_empty(&queue)) { |
1805 | cp = container_of(queue.next, struct cpuset, stack_list); | |
1806 | list_del(queue.next); | |
1807 | list_for_each_entry(cont, &cp->css.cgroup->children, sibling) { | |
1808 | child = cgroup_cs(cont); | |
1809 | list_add_tail(&child->stack_list, &queue); | |
1810 | } | |
1811 | cont = cp->css.cgroup; | |
b4501295 PJ |
1812 | |
1813 | /* Continue past cpusets with all cpus, mems online */ | |
1814 | if (cpus_subset(cp->cpus_allowed, cpu_online_map) && | |
1815 | nodes_subset(cp->mems_allowed, node_states[N_HIGH_MEMORY])) | |
1816 | continue; | |
1817 | ||
956db3ca | 1818 | /* Remove offline cpus and mems from this cpuset. */ |
b4501295 | 1819 | mutex_lock(&callback_mutex); |
956db3ca CW |
1820 | cpus_and(cp->cpus_allowed, cp->cpus_allowed, cpu_online_map); |
1821 | nodes_and(cp->mems_allowed, cp->mems_allowed, | |
1822 | node_states[N_HIGH_MEMORY]); | |
b4501295 PJ |
1823 | mutex_unlock(&callback_mutex); |
1824 | ||
1825 | /* Move tasks from the empty cpuset to a parent */ | |
c8d9c90c | 1826 | if (cpus_empty(cp->cpus_allowed) || |
b4501295 | 1827 | nodes_empty(cp->mems_allowed)) |
956db3ca | 1828 | remove_tasks_in_empty_cpuset(cp); |
b1aac8bb PJ |
1829 | } |
1830 | } | |
1831 | ||
1832 | /* | |
1833 | * The cpus_allowed and mems_allowed nodemasks in the top_cpuset track | |
0e1e7c7a | 1834 | * cpu_online_map and node_states[N_HIGH_MEMORY]. Force the top cpuset to |
956db3ca | 1835 | * track what's online after any CPU or memory node hotplug or unplug event. |
b1aac8bb PJ |
1836 | * |
1837 | * Since there are two callers of this routine, one for CPU hotplug | |
1838 | * events and one for memory node hotplug events, we could have coded | |
1839 | * two separate routines here. We code it as a single common routine | |
1840 | * in order to minimize text size. | |
1841 | */ | |
1842 | ||
1843 | static void common_cpu_mem_hotplug_unplug(void) | |
1844 | { | |
8793d854 | 1845 | cgroup_lock(); |
b1aac8bb | 1846 | |
b1aac8bb | 1847 | top_cpuset.cpus_allowed = cpu_online_map; |
0e1e7c7a | 1848 | top_cpuset.mems_allowed = node_states[N_HIGH_MEMORY]; |
956db3ca | 1849 | scan_for_empty_cpusets(&top_cpuset); |
b1aac8bb | 1850 | |
8793d854 | 1851 | cgroup_unlock(); |
b1aac8bb | 1852 | } |
b1aac8bb | 1853 | |
4c4d50f7 PJ |
1854 | /* |
1855 | * The top_cpuset tracks what CPUs and Memory Nodes are online, | |
1856 | * period. This is necessary in order to make cpusets transparent | |
1857 | * (of no affect) on systems that are actively using CPU hotplug | |
1858 | * but making no active use of cpusets. | |
1859 | * | |
38837fc7 PJ |
1860 | * This routine ensures that top_cpuset.cpus_allowed tracks |
1861 | * cpu_online_map on each CPU hotplug (cpuhp) event. | |
4c4d50f7 PJ |
1862 | */ |
1863 | ||
029190c5 PJ |
1864 | static int cpuset_handle_cpuhp(struct notifier_block *unused_nb, |
1865 | unsigned long phase, void *unused_cpu) | |
4c4d50f7 | 1866 | { |
ac076758 AK |
1867 | if (phase == CPU_DYING || phase == CPU_DYING_FROZEN) |
1868 | return NOTIFY_DONE; | |
1869 | ||
b1aac8bb | 1870 | common_cpu_mem_hotplug_unplug(); |
4c4d50f7 PJ |
1871 | return 0; |
1872 | } | |
4c4d50f7 | 1873 | |
b1aac8bb | 1874 | #ifdef CONFIG_MEMORY_HOTPLUG |
38837fc7 | 1875 | /* |
0e1e7c7a CL |
1876 | * Keep top_cpuset.mems_allowed tracking node_states[N_HIGH_MEMORY]. |
1877 | * Call this routine anytime after you change | |
1878 | * node_states[N_HIGH_MEMORY]. | |
38837fc7 PJ |
1879 | * See also the previous routine cpuset_handle_cpuhp(). |
1880 | */ | |
1881 | ||
1af98928 | 1882 | void cpuset_track_online_nodes(void) |
38837fc7 | 1883 | { |
b1aac8bb | 1884 | common_cpu_mem_hotplug_unplug(); |
38837fc7 PJ |
1885 | } |
1886 | #endif | |
1887 | ||
1da177e4 LT |
1888 | /** |
1889 | * cpuset_init_smp - initialize cpus_allowed | |
1890 | * | |
1891 | * Description: Finish top cpuset after cpu, node maps are initialized | |
1892 | **/ | |
1893 | ||
1894 | void __init cpuset_init_smp(void) | |
1895 | { | |
1896 | top_cpuset.cpus_allowed = cpu_online_map; | |
0e1e7c7a | 1897 | top_cpuset.mems_allowed = node_states[N_HIGH_MEMORY]; |
4c4d50f7 PJ |
1898 | |
1899 | hotcpu_notifier(cpuset_handle_cpuhp, 0); | |
1da177e4 LT |
1900 | } |
1901 | ||
1902 | /** | |
3077a260 | 1903 | |
1da177e4 LT |
1904 | * cpuset_cpus_allowed - return cpus_allowed mask from a tasks cpuset. |
1905 | * @tsk: pointer to task_struct from which to obtain cpuset->cpus_allowed. | |
f9a86fcb | 1906 | * @pmask: pointer to cpumask_t variable to receive cpus_allowed set. |
1da177e4 LT |
1907 | * |
1908 | * Description: Returns the cpumask_t cpus_allowed of the cpuset | |
1909 | * attached to the specified @tsk. Guaranteed to return some non-empty | |
1910 | * subset of cpu_online_map, even if this means going outside the | |
1911 | * tasks cpuset. | |
1912 | **/ | |
1913 | ||
f9a86fcb | 1914 | void cpuset_cpus_allowed(struct task_struct *tsk, cpumask_t *pmask) |
1da177e4 | 1915 | { |
3d3f26a7 | 1916 | mutex_lock(&callback_mutex); |
f9a86fcb | 1917 | cpuset_cpus_allowed_locked(tsk, pmask); |
470fd646 | 1918 | mutex_unlock(&callback_mutex); |
470fd646 CW |
1919 | } |
1920 | ||
1921 | /** | |
1922 | * cpuset_cpus_allowed_locked - return cpus_allowed mask from a tasks cpuset. | |
2df167a3 | 1923 | * Must be called with callback_mutex held. |
470fd646 | 1924 | **/ |
f9a86fcb | 1925 | void cpuset_cpus_allowed_locked(struct task_struct *tsk, cpumask_t *pmask) |
470fd646 | 1926 | { |
909d75a3 | 1927 | task_lock(tsk); |
f9a86fcb | 1928 | guarantee_online_cpus(task_cs(tsk), pmask); |
909d75a3 | 1929 | task_unlock(tsk); |
1da177e4 LT |
1930 | } |
1931 | ||
1932 | void cpuset_init_current_mems_allowed(void) | |
1933 | { | |
f9a86fcb | 1934 | nodes_setall(current->mems_allowed); |
1da177e4 LT |
1935 | } |
1936 | ||
909d75a3 PJ |
1937 | /** |
1938 | * cpuset_mems_allowed - return mems_allowed mask from a tasks cpuset. | |
1939 | * @tsk: pointer to task_struct from which to obtain cpuset->mems_allowed. | |
1940 | * | |
1941 | * Description: Returns the nodemask_t mems_allowed of the cpuset | |
1942 | * attached to the specified @tsk. Guaranteed to return some non-empty | |
0e1e7c7a | 1943 | * subset of node_states[N_HIGH_MEMORY], even if this means going outside the |
909d75a3 PJ |
1944 | * tasks cpuset. |
1945 | **/ | |
1946 | ||
1947 | nodemask_t cpuset_mems_allowed(struct task_struct *tsk) | |
1948 | { | |
1949 | nodemask_t mask; | |
1950 | ||
3d3f26a7 | 1951 | mutex_lock(&callback_mutex); |
909d75a3 | 1952 | task_lock(tsk); |
8793d854 | 1953 | guarantee_online_mems(task_cs(tsk), &mask); |
909d75a3 | 1954 | task_unlock(tsk); |
3d3f26a7 | 1955 | mutex_unlock(&callback_mutex); |
909d75a3 PJ |
1956 | |
1957 | return mask; | |
1958 | } | |
1959 | ||
d9fd8a6d RD |
1960 | /** |
1961 | * cpuset_zonelist_valid_mems_allowed - check zonelist vs. curremt mems_allowed | |
1962 | * @zl: the zonelist to be checked | |
1963 | * | |
1da177e4 LT |
1964 | * Are any of the nodes on zonelist zl allowed in current->mems_allowed? |
1965 | */ | |
1966 | int cpuset_zonelist_valid_mems_allowed(struct zonelist *zl) | |
1967 | { | |
1968 | int i; | |
1969 | ||
1970 | for (i = 0; zl->zones[i]; i++) { | |
89fa3024 | 1971 | int nid = zone_to_nid(zl->zones[i]); |
1da177e4 LT |
1972 | |
1973 | if (node_isset(nid, current->mems_allowed)) | |
1974 | return 1; | |
1975 | } | |
1976 | return 0; | |
1977 | } | |
1978 | ||
9bf2229f PJ |
1979 | /* |
1980 | * nearest_exclusive_ancestor() - Returns the nearest mem_exclusive | |
3d3f26a7 | 1981 | * ancestor to the specified cpuset. Call holding callback_mutex. |
9bf2229f PJ |
1982 | * If no ancestor is mem_exclusive (an unusual configuration), then |
1983 | * returns the root cpuset. | |
1984 | */ | |
1985 | static const struct cpuset *nearest_exclusive_ancestor(const struct cpuset *cs) | |
1986 | { | |
1987 | while (!is_mem_exclusive(cs) && cs->parent) | |
1988 | cs = cs->parent; | |
1989 | return cs; | |
1990 | } | |
1991 | ||
d9fd8a6d | 1992 | /** |
02a0e53d | 1993 | * cpuset_zone_allowed_softwall - Can we allocate on zone z's memory node? |
9bf2229f | 1994 | * @z: is this zone on an allowed node? |
02a0e53d | 1995 | * @gfp_mask: memory allocation flags |
d9fd8a6d | 1996 | * |
02a0e53d PJ |
1997 | * If we're in interrupt, yes, we can always allocate. If |
1998 | * __GFP_THISNODE is set, yes, we can always allocate. If zone | |
9bf2229f PJ |
1999 | * z's node is in our tasks mems_allowed, yes. If it's not a |
2000 | * __GFP_HARDWALL request and this zone's nodes is in the nearest | |
2001 | * mem_exclusive cpuset ancestor to this tasks cpuset, yes. | |
c596d9f3 DR |
2002 | * If the task has been OOM killed and has access to memory reserves |
2003 | * as specified by the TIF_MEMDIE flag, yes. | |
9bf2229f PJ |
2004 | * Otherwise, no. |
2005 | * | |
02a0e53d PJ |
2006 | * If __GFP_HARDWALL is set, cpuset_zone_allowed_softwall() |
2007 | * reduces to cpuset_zone_allowed_hardwall(). Otherwise, | |
2008 | * cpuset_zone_allowed_softwall() might sleep, and might allow a zone | |
2009 | * from an enclosing cpuset. | |
2010 | * | |
2011 | * cpuset_zone_allowed_hardwall() only handles the simpler case of | |
2012 | * hardwall cpusets, and never sleeps. | |
2013 | * | |
2014 | * The __GFP_THISNODE placement logic is really handled elsewhere, | |
2015 | * by forcibly using a zonelist starting at a specified node, and by | |
2016 | * (in get_page_from_freelist()) refusing to consider the zones for | |
2017 | * any node on the zonelist except the first. By the time any such | |
2018 | * calls get to this routine, we should just shut up and say 'yes'. | |
2019 | * | |
9bf2229f | 2020 | * GFP_USER allocations are marked with the __GFP_HARDWALL bit, |
c596d9f3 DR |
2021 | * and do not allow allocations outside the current tasks cpuset |
2022 | * unless the task has been OOM killed as is marked TIF_MEMDIE. | |
9bf2229f | 2023 | * GFP_KERNEL allocations are not so marked, so can escape to the |
02a0e53d | 2024 | * nearest enclosing mem_exclusive ancestor cpuset. |
9bf2229f | 2025 | * |
02a0e53d PJ |
2026 | * Scanning up parent cpusets requires callback_mutex. The |
2027 | * __alloc_pages() routine only calls here with __GFP_HARDWALL bit | |
2028 | * _not_ set if it's a GFP_KERNEL allocation, and all nodes in the | |
2029 | * current tasks mems_allowed came up empty on the first pass over | |
2030 | * the zonelist. So only GFP_KERNEL allocations, if all nodes in the | |
2031 | * cpuset are short of memory, might require taking the callback_mutex | |
2032 | * mutex. | |
9bf2229f | 2033 | * |
36be57ff | 2034 | * The first call here from mm/page_alloc:get_page_from_freelist() |
02a0e53d PJ |
2035 | * has __GFP_HARDWALL set in gfp_mask, enforcing hardwall cpusets, |
2036 | * so no allocation on a node outside the cpuset is allowed (unless | |
2037 | * in interrupt, of course). | |
36be57ff PJ |
2038 | * |
2039 | * The second pass through get_page_from_freelist() doesn't even call | |
2040 | * here for GFP_ATOMIC calls. For those calls, the __alloc_pages() | |
2041 | * variable 'wait' is not set, and the bit ALLOC_CPUSET is not set | |
2042 | * in alloc_flags. That logic and the checks below have the combined | |
2043 | * affect that: | |
9bf2229f PJ |
2044 | * in_interrupt - any node ok (current task context irrelevant) |
2045 | * GFP_ATOMIC - any node ok | |
c596d9f3 | 2046 | * TIF_MEMDIE - any node ok |
9bf2229f PJ |
2047 | * GFP_KERNEL - any node in enclosing mem_exclusive cpuset ok |
2048 | * GFP_USER - only nodes in current tasks mems allowed ok. | |
36be57ff PJ |
2049 | * |
2050 | * Rule: | |
02a0e53d | 2051 | * Don't call cpuset_zone_allowed_softwall if you can't sleep, unless you |
36be57ff PJ |
2052 | * pass in the __GFP_HARDWALL flag set in gfp_flag, which disables |
2053 | * the code that might scan up ancestor cpusets and sleep. | |
02a0e53d | 2054 | */ |
9bf2229f | 2055 | |
02a0e53d | 2056 | int __cpuset_zone_allowed_softwall(struct zone *z, gfp_t gfp_mask) |
1da177e4 | 2057 | { |
9bf2229f PJ |
2058 | int node; /* node that zone z is on */ |
2059 | const struct cpuset *cs; /* current cpuset ancestors */ | |
29afd49b | 2060 | int allowed; /* is allocation in zone z allowed? */ |
9bf2229f | 2061 | |
9b819d20 | 2062 | if (in_interrupt() || (gfp_mask & __GFP_THISNODE)) |
9bf2229f | 2063 | return 1; |
89fa3024 | 2064 | node = zone_to_nid(z); |
92d1dbd2 | 2065 | might_sleep_if(!(gfp_mask & __GFP_HARDWALL)); |
9bf2229f PJ |
2066 | if (node_isset(node, current->mems_allowed)) |
2067 | return 1; | |
c596d9f3 DR |
2068 | /* |
2069 | * Allow tasks that have access to memory reserves because they have | |
2070 | * been OOM killed to get memory anywhere. | |
2071 | */ | |
2072 | if (unlikely(test_thread_flag(TIF_MEMDIE))) | |
2073 | return 1; | |
9bf2229f PJ |
2074 | if (gfp_mask & __GFP_HARDWALL) /* If hardwall request, stop here */ |
2075 | return 0; | |
2076 | ||
5563e770 BP |
2077 | if (current->flags & PF_EXITING) /* Let dying task have memory */ |
2078 | return 1; | |
2079 | ||
9bf2229f | 2080 | /* Not hardwall and node outside mems_allowed: scan up cpusets */ |
3d3f26a7 | 2081 | mutex_lock(&callback_mutex); |
053199ed | 2082 | |
053199ed | 2083 | task_lock(current); |
8793d854 | 2084 | cs = nearest_exclusive_ancestor(task_cs(current)); |
053199ed PJ |
2085 | task_unlock(current); |
2086 | ||
9bf2229f | 2087 | allowed = node_isset(node, cs->mems_allowed); |
3d3f26a7 | 2088 | mutex_unlock(&callback_mutex); |
9bf2229f | 2089 | return allowed; |
1da177e4 LT |
2090 | } |
2091 | ||
02a0e53d PJ |
2092 | /* |
2093 | * cpuset_zone_allowed_hardwall - Can we allocate on zone z's memory node? | |
2094 | * @z: is this zone on an allowed node? | |
2095 | * @gfp_mask: memory allocation flags | |
2096 | * | |
2097 | * If we're in interrupt, yes, we can always allocate. | |
2098 | * If __GFP_THISNODE is set, yes, we can always allocate. If zone | |
c596d9f3 DR |
2099 | * z's node is in our tasks mems_allowed, yes. If the task has been |
2100 | * OOM killed and has access to memory reserves as specified by the | |
2101 | * TIF_MEMDIE flag, yes. Otherwise, no. | |
02a0e53d PJ |
2102 | * |
2103 | * The __GFP_THISNODE placement logic is really handled elsewhere, | |
2104 | * by forcibly using a zonelist starting at a specified node, and by | |
2105 | * (in get_page_from_freelist()) refusing to consider the zones for | |
2106 | * any node on the zonelist except the first. By the time any such | |
2107 | * calls get to this routine, we should just shut up and say 'yes'. | |
2108 | * | |
2109 | * Unlike the cpuset_zone_allowed_softwall() variant, above, | |
2110 | * this variant requires that the zone be in the current tasks | |
2111 | * mems_allowed or that we're in interrupt. It does not scan up the | |
2112 | * cpuset hierarchy for the nearest enclosing mem_exclusive cpuset. | |
2113 | * It never sleeps. | |
2114 | */ | |
2115 | ||
2116 | int __cpuset_zone_allowed_hardwall(struct zone *z, gfp_t gfp_mask) | |
2117 | { | |
2118 | int node; /* node that zone z is on */ | |
2119 | ||
2120 | if (in_interrupt() || (gfp_mask & __GFP_THISNODE)) | |
2121 | return 1; | |
2122 | node = zone_to_nid(z); | |
2123 | if (node_isset(node, current->mems_allowed)) | |
2124 | return 1; | |
dedf8b79 DW |
2125 | /* |
2126 | * Allow tasks that have access to memory reserves because they have | |
2127 | * been OOM killed to get memory anywhere. | |
2128 | */ | |
2129 | if (unlikely(test_thread_flag(TIF_MEMDIE))) | |
2130 | return 1; | |
02a0e53d PJ |
2131 | return 0; |
2132 | } | |
2133 | ||
505970b9 PJ |
2134 | /** |
2135 | * cpuset_lock - lock out any changes to cpuset structures | |
2136 | * | |
3d3f26a7 | 2137 | * The out of memory (oom) code needs to mutex_lock cpusets |
505970b9 | 2138 | * from being changed while it scans the tasklist looking for a |
3d3f26a7 | 2139 | * task in an overlapping cpuset. Expose callback_mutex via this |
505970b9 PJ |
2140 | * cpuset_lock() routine, so the oom code can lock it, before |
2141 | * locking the task list. The tasklist_lock is a spinlock, so | |
3d3f26a7 | 2142 | * must be taken inside callback_mutex. |
505970b9 PJ |
2143 | */ |
2144 | ||
2145 | void cpuset_lock(void) | |
2146 | { | |
3d3f26a7 | 2147 | mutex_lock(&callback_mutex); |
505970b9 PJ |
2148 | } |
2149 | ||
2150 | /** | |
2151 | * cpuset_unlock - release lock on cpuset changes | |
2152 | * | |
2153 | * Undo the lock taken in a previous cpuset_lock() call. | |
2154 | */ | |
2155 | ||
2156 | void cpuset_unlock(void) | |
2157 | { | |
3d3f26a7 | 2158 | mutex_unlock(&callback_mutex); |
505970b9 PJ |
2159 | } |
2160 | ||
825a46af PJ |
2161 | /** |
2162 | * cpuset_mem_spread_node() - On which node to begin search for a page | |
2163 | * | |
2164 | * If a task is marked PF_SPREAD_PAGE or PF_SPREAD_SLAB (as for | |
2165 | * tasks in a cpuset with is_spread_page or is_spread_slab set), | |
2166 | * and if the memory allocation used cpuset_mem_spread_node() | |
2167 | * to determine on which node to start looking, as it will for | |
2168 | * certain page cache or slab cache pages such as used for file | |
2169 | * system buffers and inode caches, then instead of starting on the | |
2170 | * local node to look for a free page, rather spread the starting | |
2171 | * node around the tasks mems_allowed nodes. | |
2172 | * | |
2173 | * We don't have to worry about the returned node being offline | |
2174 | * because "it can't happen", and even if it did, it would be ok. | |
2175 | * | |
2176 | * The routines calling guarantee_online_mems() are careful to | |
2177 | * only set nodes in task->mems_allowed that are online. So it | |
2178 | * should not be possible for the following code to return an | |
2179 | * offline node. But if it did, that would be ok, as this routine | |
2180 | * is not returning the node where the allocation must be, only | |
2181 | * the node where the search should start. The zonelist passed to | |
2182 | * __alloc_pages() will include all nodes. If the slab allocator | |
2183 | * is passed an offline node, it will fall back to the local node. | |
2184 | * See kmem_cache_alloc_node(). | |
2185 | */ | |
2186 | ||
2187 | int cpuset_mem_spread_node(void) | |
2188 | { | |
2189 | int node; | |
2190 | ||
2191 | node = next_node(current->cpuset_mem_spread_rotor, current->mems_allowed); | |
2192 | if (node == MAX_NUMNODES) | |
2193 | node = first_node(current->mems_allowed); | |
2194 | current->cpuset_mem_spread_rotor = node; | |
2195 | return node; | |
2196 | } | |
2197 | EXPORT_SYMBOL_GPL(cpuset_mem_spread_node); | |
2198 | ||
ef08e3b4 | 2199 | /** |
bbe373f2 DR |
2200 | * cpuset_mems_allowed_intersects - Does @tsk1's mems_allowed intersect @tsk2's? |
2201 | * @tsk1: pointer to task_struct of some task. | |
2202 | * @tsk2: pointer to task_struct of some other task. | |
2203 | * | |
2204 | * Description: Return true if @tsk1's mems_allowed intersects the | |
2205 | * mems_allowed of @tsk2. Used by the OOM killer to determine if | |
2206 | * one of the task's memory usage might impact the memory available | |
2207 | * to the other. | |
ef08e3b4 PJ |
2208 | **/ |
2209 | ||
bbe373f2 DR |
2210 | int cpuset_mems_allowed_intersects(const struct task_struct *tsk1, |
2211 | const struct task_struct *tsk2) | |
ef08e3b4 | 2212 | { |
bbe373f2 | 2213 | return nodes_intersects(tsk1->mems_allowed, tsk2->mems_allowed); |
ef08e3b4 PJ |
2214 | } |
2215 | ||
3e0d98b9 PJ |
2216 | /* |
2217 | * Collection of memory_pressure is suppressed unless | |
2218 | * this flag is enabled by writing "1" to the special | |
2219 | * cpuset file 'memory_pressure_enabled' in the root cpuset. | |
2220 | */ | |
2221 | ||
c5b2aff8 | 2222 | int cpuset_memory_pressure_enabled __read_mostly; |
3e0d98b9 PJ |
2223 | |
2224 | /** | |
2225 | * cpuset_memory_pressure_bump - keep stats of per-cpuset reclaims. | |
2226 | * | |
2227 | * Keep a running average of the rate of synchronous (direct) | |
2228 | * page reclaim efforts initiated by tasks in each cpuset. | |
2229 | * | |
2230 | * This represents the rate at which some task in the cpuset | |
2231 | * ran low on memory on all nodes it was allowed to use, and | |
2232 | * had to enter the kernels page reclaim code in an effort to | |
2233 | * create more free memory by tossing clean pages or swapping | |
2234 | * or writing dirty pages. | |
2235 | * | |
2236 | * Display to user space in the per-cpuset read-only file | |
2237 | * "memory_pressure". Value displayed is an integer | |
2238 | * representing the recent rate of entry into the synchronous | |
2239 | * (direct) page reclaim by any task attached to the cpuset. | |
2240 | **/ | |
2241 | ||
2242 | void __cpuset_memory_pressure_bump(void) | |
2243 | { | |
3e0d98b9 | 2244 | task_lock(current); |
8793d854 | 2245 | fmeter_markevent(&task_cs(current)->fmeter); |
3e0d98b9 PJ |
2246 | task_unlock(current); |
2247 | } | |
2248 | ||
8793d854 | 2249 | #ifdef CONFIG_PROC_PID_CPUSET |
1da177e4 LT |
2250 | /* |
2251 | * proc_cpuset_show() | |
2252 | * - Print tasks cpuset path into seq_file. | |
2253 | * - Used for /proc/<pid>/cpuset. | |
053199ed PJ |
2254 | * - No need to task_lock(tsk) on this tsk->cpuset reference, as it |
2255 | * doesn't really matter if tsk->cpuset changes after we read it, | |
c8d9c90c | 2256 | * and we take cgroup_mutex, keeping cpuset_attach() from changing it |
2df167a3 | 2257 | * anyway. |
1da177e4 | 2258 | */ |
029190c5 | 2259 | static int proc_cpuset_show(struct seq_file *m, void *unused_v) |
1da177e4 | 2260 | { |
13b41b09 | 2261 | struct pid *pid; |
1da177e4 LT |
2262 | struct task_struct *tsk; |
2263 | char *buf; | |
8793d854 | 2264 | struct cgroup_subsys_state *css; |
99f89551 | 2265 | int retval; |
1da177e4 | 2266 | |
99f89551 | 2267 | retval = -ENOMEM; |
1da177e4 LT |
2268 | buf = kmalloc(PAGE_SIZE, GFP_KERNEL); |
2269 | if (!buf) | |
99f89551 EB |
2270 | goto out; |
2271 | ||
2272 | retval = -ESRCH; | |
13b41b09 EB |
2273 | pid = m->private; |
2274 | tsk = get_pid_task(pid, PIDTYPE_PID); | |
99f89551 EB |
2275 | if (!tsk) |
2276 | goto out_free; | |
1da177e4 | 2277 | |
99f89551 | 2278 | retval = -EINVAL; |
8793d854 PM |
2279 | cgroup_lock(); |
2280 | css = task_subsys_state(tsk, cpuset_subsys_id); | |
2281 | retval = cgroup_path(css->cgroup, buf, PAGE_SIZE); | |
1da177e4 | 2282 | if (retval < 0) |
99f89551 | 2283 | goto out_unlock; |
1da177e4 LT |
2284 | seq_puts(m, buf); |
2285 | seq_putc(m, '\n'); | |
99f89551 | 2286 | out_unlock: |
8793d854 | 2287 | cgroup_unlock(); |
99f89551 EB |
2288 | put_task_struct(tsk); |
2289 | out_free: | |
1da177e4 | 2290 | kfree(buf); |
99f89551 | 2291 | out: |
1da177e4 LT |
2292 | return retval; |
2293 | } | |
2294 | ||
2295 | static int cpuset_open(struct inode *inode, struct file *file) | |
2296 | { | |
13b41b09 EB |
2297 | struct pid *pid = PROC_I(inode)->pid; |
2298 | return single_open(file, proc_cpuset_show, pid); | |
1da177e4 LT |
2299 | } |
2300 | ||
9a32144e | 2301 | const struct file_operations proc_cpuset_operations = { |
1da177e4 LT |
2302 | .open = cpuset_open, |
2303 | .read = seq_read, | |
2304 | .llseek = seq_lseek, | |
2305 | .release = single_release, | |
2306 | }; | |
8793d854 | 2307 | #endif /* CONFIG_PROC_PID_CPUSET */ |
1da177e4 LT |
2308 | |
2309 | /* Display task cpus_allowed, mems_allowed in /proc/<pid>/status file. */ | |
df5f8314 EB |
2310 | void cpuset_task_status_allowed(struct seq_file *m, struct task_struct *task) |
2311 | { | |
2312 | seq_printf(m, "Cpus_allowed:\t"); | |
2313 | m->count += cpumask_scnprintf(m->buf + m->count, m->size - m->count, | |
2314 | task->cpus_allowed); | |
2315 | seq_printf(m, "\n"); | |
39106dcf MT |
2316 | seq_printf(m, "Cpus_allowed_list:\t"); |
2317 | m->count += cpulist_scnprintf(m->buf + m->count, m->size - m->count, | |
2318 | task->cpus_allowed); | |
2319 | seq_printf(m, "\n"); | |
df5f8314 EB |
2320 | seq_printf(m, "Mems_allowed:\t"); |
2321 | m->count += nodemask_scnprintf(m->buf + m->count, m->size - m->count, | |
2322 | task->mems_allowed); | |
2323 | seq_printf(m, "\n"); | |
39106dcf MT |
2324 | seq_printf(m, "Mems_allowed_list:\t"); |
2325 | m->count += nodelist_scnprintf(m->buf + m->count, m->size - m->count, | |
2326 | task->mems_allowed); | |
2327 | seq_printf(m, "\n"); | |
1da177e4 | 2328 | } |