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