]>
Commit | Line | Data |
---|---|---|
1 | /* | |
2 | * kernel/cpuset.c | |
3 | * | |
4 | * Processor and Memory placement constraints for sets of tasks. | |
5 | * | |
6 | * Copyright (C) 2003 BULL SA. | |
7 | * Copyright (C) 2004-2007 Silicon Graphics, Inc. | |
8 | * Copyright (C) 2006 Google, Inc | |
9 | * | |
10 | * Portions derived from Patrick Mochel's sysfs code. | |
11 | * sysfs is Copyright (c) 2001-3 Patrick Mochel | |
12 | * | |
13 | * 2003-10-10 Written by Simon Derr. | |
14 | * 2003-10-22 Updates by Stephen Hemminger. | |
15 | * 2004 May-July Rework by Paul Jackson. | |
16 | * 2006 Rework by Paul Menage to use generic cgroups | |
17 | * 2008 Rework of the scheduler domains and CPU hotplug handling | |
18 | * by Max Krasnyansky | |
19 | * | |
20 | * This file is subject to the terms and conditions of the GNU General Public | |
21 | * License. See the file COPYING in the main directory of the Linux | |
22 | * distribution for more details. | |
23 | */ | |
24 | ||
25 | #include <linux/cpu.h> | |
26 | #include <linux/cpumask.h> | |
27 | #include <linux/cpuset.h> | |
28 | #include <linux/err.h> | |
29 | #include <linux/errno.h> | |
30 | #include <linux/file.h> | |
31 | #include <linux/fs.h> | |
32 | #include <linux/init.h> | |
33 | #include <linux/interrupt.h> | |
34 | #include <linux/kernel.h> | |
35 | #include <linux/kmod.h> | |
36 | #include <linux/list.h> | |
37 | #include <linux/mempolicy.h> | |
38 | #include <linux/mm.h> | |
39 | #include <linux/memory.h> | |
40 | #include <linux/module.h> | |
41 | #include <linux/mount.h> | |
42 | #include <linux/namei.h> | |
43 | #include <linux/pagemap.h> | |
44 | #include <linux/proc_fs.h> | |
45 | #include <linux/rcupdate.h> | |
46 | #include <linux/sched.h> | |
47 | #include <linux/seq_file.h> | |
48 | #include <linux/security.h> | |
49 | #include <linux/slab.h> | |
50 | #include <linux/spinlock.h> | |
51 | #include <linux/stat.h> | |
52 | #include <linux/string.h> | |
53 | #include <linux/time.h> | |
54 | #include <linux/backing-dev.h> | |
55 | #include <linux/sort.h> | |
56 | ||
57 | #include <asm/uaccess.h> | |
58 | #include <asm/atomic.h> | |
59 | #include <linux/mutex.h> | |
60 | #include <linux/workqueue.h> | |
61 | #include <linux/cgroup.h> | |
62 | ||
63 | /* | |
64 | * Tracks how many cpusets are currently defined in system. | |
65 | * When there is only one cpuset (the root cpuset) we can | |
66 | * short circuit some hooks. | |
67 | */ | |
68 | int number_of_cpusets __read_mostly; | |
69 | ||
70 | /* Forward declare cgroup structures */ | |
71 | struct cgroup_subsys cpuset_subsys; | |
72 | struct cpuset; | |
73 | ||
74 | /* See "Frequency meter" comments, below. */ | |
75 | ||
76 | struct fmeter { | |
77 | int cnt; /* unprocessed events count */ | |
78 | int val; /* most recent output value */ | |
79 | time_t time; /* clock (secs) when val computed */ | |
80 | spinlock_t lock; /* guards read or write of above */ | |
81 | }; | |
82 | ||
83 | struct cpuset { | |
84 | struct cgroup_subsys_state css; | |
85 | ||
86 | unsigned long flags; /* "unsigned long" so bitops work */ | |
87 | cpumask_t cpus_allowed; /* CPUs allowed to tasks in cpuset */ | |
88 | nodemask_t mems_allowed; /* Memory Nodes allowed to tasks */ | |
89 | ||
90 | struct cpuset *parent; /* my parent */ | |
91 | ||
92 | /* | |
93 | * Copy of global cpuset_mems_generation as of the most | |
94 | * recent time this cpuset changed its mems_allowed. | |
95 | */ | |
96 | int mems_generation; | |
97 | ||
98 | struct fmeter fmeter; /* memory_pressure filter */ | |
99 | ||
100 | /* partition number for rebuild_sched_domains() */ | |
101 | int pn; | |
102 | ||
103 | /* for custom sched domain */ | |
104 | int relax_domain_level; | |
105 | ||
106 | /* used for walking a cpuset heirarchy */ | |
107 | struct list_head stack_list; | |
108 | }; | |
109 | ||
110 | /* Retrieve the cpuset for a cgroup */ | |
111 | static inline struct cpuset *cgroup_cs(struct cgroup *cont) | |
112 | { | |
113 | return container_of(cgroup_subsys_state(cont, cpuset_subsys_id), | |
114 | struct cpuset, css); | |
115 | } | |
116 | ||
117 | /* Retrieve the cpuset for a task */ | |
118 | static inline struct cpuset *task_cs(struct task_struct *task) | |
119 | { | |
120 | return container_of(task_subsys_state(task, cpuset_subsys_id), | |
121 | struct cpuset, css); | |
122 | } | |
123 | struct cpuset_hotplug_scanner { | |
124 | struct cgroup_scanner scan; | |
125 | struct cgroup *to; | |
126 | }; | |
127 | ||
128 | /* bits in struct cpuset flags field */ | |
129 | typedef enum { | |
130 | CS_CPU_EXCLUSIVE, | |
131 | CS_MEM_EXCLUSIVE, | |
132 | CS_MEM_HARDWALL, | |
133 | CS_MEMORY_MIGRATE, | |
134 | CS_SCHED_LOAD_BALANCE, | |
135 | CS_SPREAD_PAGE, | |
136 | CS_SPREAD_SLAB, | |
137 | } cpuset_flagbits_t; | |
138 | ||
139 | /* convenient tests for these bits */ | |
140 | static inline int is_cpu_exclusive(const struct cpuset *cs) | |
141 | { | |
142 | return test_bit(CS_CPU_EXCLUSIVE, &cs->flags); | |
143 | } | |
144 | ||
145 | static inline int is_mem_exclusive(const struct cpuset *cs) | |
146 | { | |
147 | return test_bit(CS_MEM_EXCLUSIVE, &cs->flags); | |
148 | } | |
149 | ||
150 | static inline int is_mem_hardwall(const struct cpuset *cs) | |
151 | { | |
152 | return test_bit(CS_MEM_HARDWALL, &cs->flags); | |
153 | } | |
154 | ||
155 | static inline int is_sched_load_balance(const struct cpuset *cs) | |
156 | { | |
157 | return test_bit(CS_SCHED_LOAD_BALANCE, &cs->flags); | |
158 | } | |
159 | ||
160 | static inline int is_memory_migrate(const struct cpuset *cs) | |
161 | { | |
162 | return test_bit(CS_MEMORY_MIGRATE, &cs->flags); | |
163 | } | |
164 | ||
165 | static inline int is_spread_page(const struct cpuset *cs) | |
166 | { | |
167 | return test_bit(CS_SPREAD_PAGE, &cs->flags); | |
168 | } | |
169 | ||
170 | static inline int is_spread_slab(const struct cpuset *cs) | |
171 | { | |
172 | return test_bit(CS_SPREAD_SLAB, &cs->flags); | |
173 | } | |
174 | ||
175 | /* | |
176 | * Increment this integer everytime any cpuset changes its | |
177 | * mems_allowed value. Users of cpusets can track this generation | |
178 | * number, and avoid having to lock and reload mems_allowed unless | |
179 | * the cpuset they're using changes generation. | |
180 | * | |
181 | * A single, global generation is needed because cpuset_attach_task() could | |
182 | * reattach a task to a different cpuset, which must not have its | |
183 | * generation numbers aliased with those of that tasks previous cpuset. | |
184 | * | |
185 | * Generations are needed for mems_allowed because one task cannot | |
186 | * modify another's memory placement. So we must enable every task, | |
187 | * on every visit to __alloc_pages(), to efficiently check whether | |
188 | * its current->cpuset->mems_allowed has changed, requiring an update | |
189 | * of its current->mems_allowed. | |
190 | * | |
191 | * Since writes to cpuset_mems_generation are guarded by the cgroup lock | |
192 | * there is no need to mark it atomic. | |
193 | */ | |
194 | static int cpuset_mems_generation; | |
195 | ||
196 | static struct cpuset top_cpuset = { | |
197 | .flags = ((1 << CS_CPU_EXCLUSIVE) | (1 << CS_MEM_EXCLUSIVE)), | |
198 | .cpus_allowed = CPU_MASK_ALL, | |
199 | .mems_allowed = NODE_MASK_ALL, | |
200 | }; | |
201 | ||
202 | /* | |
203 | * There are two global mutexes guarding cpuset structures. The first | |
204 | * is the main control groups cgroup_mutex, accessed via | |
205 | * cgroup_lock()/cgroup_unlock(). The second is the cpuset-specific | |
206 | * callback_mutex, below. They can nest. It is ok to first take | |
207 | * cgroup_mutex, then nest callback_mutex. We also require taking | |
208 | * task_lock() when dereferencing a task's cpuset pointer. See "The | |
209 | * task_lock() exception", at the end of this comment. | |
210 | * | |
211 | * A task must hold both mutexes to modify cpusets. If a task | |
212 | * holds cgroup_mutex, then it blocks others wanting that mutex, | |
213 | * ensuring that it is the only task able to also acquire callback_mutex | |
214 | * and be able to modify cpusets. It can perform various checks on | |
215 | * the cpuset structure first, knowing nothing will change. It can | |
216 | * also allocate memory while just holding cgroup_mutex. While it is | |
217 | * performing these checks, various callback routines can briefly | |
218 | * acquire callback_mutex to query cpusets. Once it is ready to make | |
219 | * the changes, it takes callback_mutex, blocking everyone else. | |
220 | * | |
221 | * Calls to the kernel memory allocator can not be made while holding | |
222 | * callback_mutex, as that would risk double tripping on callback_mutex | |
223 | * from one of the callbacks into the cpuset code from within | |
224 | * __alloc_pages(). | |
225 | * | |
226 | * If a task is only holding callback_mutex, then it has read-only | |
227 | * access to cpusets. | |
228 | * | |
229 | * The task_struct fields mems_allowed and mems_generation may only | |
230 | * be accessed in the context of that task, so require no locks. | |
231 | * | |
232 | * The cpuset_common_file_read() handlers only hold callback_mutex across | |
233 | * small pieces of code, such as when reading out possibly multi-word | |
234 | * cpumasks and nodemasks. | |
235 | * | |
236 | * Accessing a task's cpuset should be done in accordance with the | |
237 | * guidelines for accessing subsystem state in kernel/cgroup.c | |
238 | */ | |
239 | ||
240 | static DEFINE_MUTEX(callback_mutex); | |
241 | ||
242 | /* | |
243 | * cpuset_buffer_lock protects both the cpuset_name and cpuset_nodelist | |
244 | * buffers. They are statically allocated to prevent using excess stack | |
245 | * when calling cpuset_print_task_mems_allowed(). | |
246 | */ | |
247 | #define CPUSET_NAME_LEN (128) | |
248 | #define CPUSET_NODELIST_LEN (256) | |
249 | static char cpuset_name[CPUSET_NAME_LEN]; | |
250 | static char cpuset_nodelist[CPUSET_NODELIST_LEN]; | |
251 | static DEFINE_SPINLOCK(cpuset_buffer_lock); | |
252 | ||
253 | /* | |
254 | * This is ugly, but preserves the userspace API for existing cpuset | |
255 | * users. If someone tries to mount the "cpuset" filesystem, we | |
256 | * silently switch it to mount "cgroup" instead | |
257 | */ | |
258 | static int cpuset_get_sb(struct file_system_type *fs_type, | |
259 | int flags, const char *unused_dev_name, | |
260 | void *data, struct vfsmount *mnt) | |
261 | { | |
262 | struct file_system_type *cgroup_fs = get_fs_type("cgroup"); | |
263 | int ret = -ENODEV; | |
264 | if (cgroup_fs) { | |
265 | char mountopts[] = | |
266 | "cpuset,noprefix," | |
267 | "release_agent=/sbin/cpuset_release_agent"; | |
268 | ret = cgroup_fs->get_sb(cgroup_fs, flags, | |
269 | unused_dev_name, mountopts, mnt); | |
270 | put_filesystem(cgroup_fs); | |
271 | } | |
272 | return ret; | |
273 | } | |
274 | ||
275 | static struct file_system_type cpuset_fs_type = { | |
276 | .name = "cpuset", | |
277 | .get_sb = cpuset_get_sb, | |
278 | }; | |
279 | ||
280 | /* | |
281 | * Return in *pmask the portion of a cpusets's cpus_allowed that | |
282 | * are online. If none are online, walk up the cpuset hierarchy | |
283 | * until we find one that does have some online cpus. If we get | |
284 | * all the way to the top and still haven't found any online cpus, | |
285 | * return cpu_online_map. Or if passed a NULL cs from an exit'ing | |
286 | * task, return cpu_online_map. | |
287 | * | |
288 | * One way or another, we guarantee to return some non-empty subset | |
289 | * of cpu_online_map. | |
290 | * | |
291 | * Call with callback_mutex held. | |
292 | */ | |
293 | ||
294 | static void guarantee_online_cpus(const struct cpuset *cs, cpumask_t *pmask) | |
295 | { | |
296 | while (cs && !cpus_intersects(cs->cpus_allowed, cpu_online_map)) | |
297 | cs = cs->parent; | |
298 | if (cs) | |
299 | cpus_and(*pmask, cs->cpus_allowed, cpu_online_map); | |
300 | else | |
301 | *pmask = cpu_online_map; | |
302 | BUG_ON(!cpus_intersects(*pmask, cpu_online_map)); | |
303 | } | |
304 | ||
305 | /* | |
306 | * Return in *pmask the portion of a cpusets's mems_allowed that | |
307 | * are online, with memory. If none are online with memory, walk | |
308 | * up the cpuset hierarchy until we find one that does have some | |
309 | * online mems. If we get all the way to the top and still haven't | |
310 | * found any online mems, return node_states[N_HIGH_MEMORY]. | |
311 | * | |
312 | * One way or another, we guarantee to return some non-empty subset | |
313 | * of node_states[N_HIGH_MEMORY]. | |
314 | * | |
315 | * Call with callback_mutex held. | |
316 | */ | |
317 | ||
318 | static void guarantee_online_mems(const struct cpuset *cs, nodemask_t *pmask) | |
319 | { | |
320 | while (cs && !nodes_intersects(cs->mems_allowed, | |
321 | node_states[N_HIGH_MEMORY])) | |
322 | cs = cs->parent; | |
323 | if (cs) | |
324 | nodes_and(*pmask, cs->mems_allowed, | |
325 | node_states[N_HIGH_MEMORY]); | |
326 | else | |
327 | *pmask = node_states[N_HIGH_MEMORY]; | |
328 | BUG_ON(!nodes_intersects(*pmask, node_states[N_HIGH_MEMORY])); | |
329 | } | |
330 | ||
331 | /** | |
332 | * cpuset_update_task_memory_state - update task memory placement | |
333 | * | |
334 | * If the current tasks cpusets mems_allowed changed behind our | |
335 | * backs, update current->mems_allowed, mems_generation and task NUMA | |
336 | * mempolicy to the new value. | |
337 | * | |
338 | * Task mempolicy is updated by rebinding it relative to the | |
339 | * current->cpuset if a task has its memory placement changed. | |
340 | * Do not call this routine if in_interrupt(). | |
341 | * | |
342 | * Call without callback_mutex or task_lock() held. May be | |
343 | * called with or without cgroup_mutex held. Thanks in part to | |
344 | * 'the_top_cpuset_hack', the task's cpuset pointer will never | |
345 | * be NULL. This routine also might acquire callback_mutex during | |
346 | * call. | |
347 | * | |
348 | * Reading current->cpuset->mems_generation doesn't need task_lock | |
349 | * to guard the current->cpuset derefence, because it is guarded | |
350 | * from concurrent freeing of current->cpuset using RCU. | |
351 | * | |
352 | * The rcu_dereference() is technically probably not needed, | |
353 | * as I don't actually mind if I see a new cpuset pointer but | |
354 | * an old value of mems_generation. However this really only | |
355 | * matters on alpha systems using cpusets heavily. If I dropped | |
356 | * that rcu_dereference(), it would save them a memory barrier. | |
357 | * For all other arch's, rcu_dereference is a no-op anyway, and for | |
358 | * alpha systems not using cpusets, another planned optimization, | |
359 | * avoiding the rcu critical section for tasks in the root cpuset | |
360 | * which is statically allocated, so can't vanish, will make this | |
361 | * irrelevant. Better to use RCU as intended, than to engage in | |
362 | * some cute trick to save a memory barrier that is impossible to | |
363 | * test, for alpha systems using cpusets heavily, which might not | |
364 | * even exist. | |
365 | * | |
366 | * This routine is needed to update the per-task mems_allowed data, | |
367 | * within the tasks context, when it is trying to allocate memory | |
368 | * (in various mm/mempolicy.c routines) and notices that some other | |
369 | * task has been modifying its cpuset. | |
370 | */ | |
371 | ||
372 | void cpuset_update_task_memory_state(void) | |
373 | { | |
374 | int my_cpusets_mem_gen; | |
375 | struct task_struct *tsk = current; | |
376 | struct cpuset *cs; | |
377 | ||
378 | rcu_read_lock(); | |
379 | my_cpusets_mem_gen = task_cs(tsk)->mems_generation; | |
380 | rcu_read_unlock(); | |
381 | ||
382 | if (my_cpusets_mem_gen != tsk->cpuset_mems_generation) { | |
383 | mutex_lock(&callback_mutex); | |
384 | task_lock(tsk); | |
385 | cs = task_cs(tsk); /* Maybe changed when task not locked */ | |
386 | guarantee_online_mems(cs, &tsk->mems_allowed); | |
387 | tsk->cpuset_mems_generation = cs->mems_generation; | |
388 | if (is_spread_page(cs)) | |
389 | tsk->flags |= PF_SPREAD_PAGE; | |
390 | else | |
391 | tsk->flags &= ~PF_SPREAD_PAGE; | |
392 | if (is_spread_slab(cs)) | |
393 | tsk->flags |= PF_SPREAD_SLAB; | |
394 | else | |
395 | tsk->flags &= ~PF_SPREAD_SLAB; | |
396 | task_unlock(tsk); | |
397 | mutex_unlock(&callback_mutex); | |
398 | mpol_rebind_task(tsk, &tsk->mems_allowed); | |
399 | } | |
400 | } | |
401 | ||
402 | /* | |
403 | * is_cpuset_subset(p, q) - Is cpuset p a subset of cpuset q? | |
404 | * | |
405 | * One cpuset is a subset of another if all its allowed CPUs and | |
406 | * Memory Nodes are a subset of the other, and its exclusive flags | |
407 | * are only set if the other's are set. Call holding cgroup_mutex. | |
408 | */ | |
409 | ||
410 | static int is_cpuset_subset(const struct cpuset *p, const struct cpuset *q) | |
411 | { | |
412 | return cpus_subset(p->cpus_allowed, q->cpus_allowed) && | |
413 | nodes_subset(p->mems_allowed, q->mems_allowed) && | |
414 | is_cpu_exclusive(p) <= is_cpu_exclusive(q) && | |
415 | is_mem_exclusive(p) <= is_mem_exclusive(q); | |
416 | } | |
417 | ||
418 | /* | |
419 | * validate_change() - Used to validate that any proposed cpuset change | |
420 | * follows the structural rules for cpusets. | |
421 | * | |
422 | * If we replaced the flag and mask values of the current cpuset | |
423 | * (cur) with those values in the trial cpuset (trial), would | |
424 | * our various subset and exclusive rules still be valid? Presumes | |
425 | * cgroup_mutex held. | |
426 | * | |
427 | * 'cur' is the address of an actual, in-use cpuset. Operations | |
428 | * such as list traversal that depend on the actual address of the | |
429 | * cpuset in the list must use cur below, not trial. | |
430 | * | |
431 | * 'trial' is the address of bulk structure copy of cur, with | |
432 | * perhaps one or more of the fields cpus_allowed, mems_allowed, | |
433 | * or flags changed to new, trial values. | |
434 | * | |
435 | * Return 0 if valid, -errno if not. | |
436 | */ | |
437 | ||
438 | static int validate_change(const struct cpuset *cur, const struct cpuset *trial) | |
439 | { | |
440 | struct cgroup *cont; | |
441 | struct cpuset *c, *par; | |
442 | ||
443 | /* Each of our child cpusets must be a subset of us */ | |
444 | list_for_each_entry(cont, &cur->css.cgroup->children, sibling) { | |
445 | if (!is_cpuset_subset(cgroup_cs(cont), trial)) | |
446 | return -EBUSY; | |
447 | } | |
448 | ||
449 | /* Remaining checks don't apply to root cpuset */ | |
450 | if (cur == &top_cpuset) | |
451 | return 0; | |
452 | ||
453 | par = cur->parent; | |
454 | ||
455 | /* We must be a subset of our parent cpuset */ | |
456 | if (!is_cpuset_subset(trial, par)) | |
457 | return -EACCES; | |
458 | ||
459 | /* | |
460 | * If either I or some sibling (!= me) is exclusive, we can't | |
461 | * overlap | |
462 | */ | |
463 | list_for_each_entry(cont, &par->css.cgroup->children, sibling) { | |
464 | c = cgroup_cs(cont); | |
465 | if ((is_cpu_exclusive(trial) || is_cpu_exclusive(c)) && | |
466 | c != cur && | |
467 | cpus_intersects(trial->cpus_allowed, c->cpus_allowed)) | |
468 | return -EINVAL; | |
469 | if ((is_mem_exclusive(trial) || is_mem_exclusive(c)) && | |
470 | c != cur && | |
471 | nodes_intersects(trial->mems_allowed, c->mems_allowed)) | |
472 | return -EINVAL; | |
473 | } | |
474 | ||
475 | /* Cpusets with tasks can't have empty cpus_allowed or mems_allowed */ | |
476 | if (cgroup_task_count(cur->css.cgroup)) { | |
477 | if (cpus_empty(trial->cpus_allowed) || | |
478 | nodes_empty(trial->mems_allowed)) { | |
479 | return -ENOSPC; | |
480 | } | |
481 | } | |
482 | ||
483 | return 0; | |
484 | } | |
485 | ||
486 | /* | |
487 | * Helper routine for generate_sched_domains(). | |
488 | * Do cpusets a, b have overlapping cpus_allowed masks? | |
489 | */ | |
490 | static int cpusets_overlap(struct cpuset *a, struct cpuset *b) | |
491 | { | |
492 | return cpus_intersects(a->cpus_allowed, b->cpus_allowed); | |
493 | } | |
494 | ||
495 | static void | |
496 | update_domain_attr(struct sched_domain_attr *dattr, struct cpuset *c) | |
497 | { | |
498 | if (dattr->relax_domain_level < c->relax_domain_level) | |
499 | dattr->relax_domain_level = c->relax_domain_level; | |
500 | return; | |
501 | } | |
502 | ||
503 | static void | |
504 | update_domain_attr_tree(struct sched_domain_attr *dattr, struct cpuset *c) | |
505 | { | |
506 | LIST_HEAD(q); | |
507 | ||
508 | list_add(&c->stack_list, &q); | |
509 | while (!list_empty(&q)) { | |
510 | struct cpuset *cp; | |
511 | struct cgroup *cont; | |
512 | struct cpuset *child; | |
513 | ||
514 | cp = list_first_entry(&q, struct cpuset, stack_list); | |
515 | list_del(q.next); | |
516 | ||
517 | if (cpus_empty(cp->cpus_allowed)) | |
518 | continue; | |
519 | ||
520 | if (is_sched_load_balance(cp)) | |
521 | update_domain_attr(dattr, cp); | |
522 | ||
523 | list_for_each_entry(cont, &cp->css.cgroup->children, sibling) { | |
524 | child = cgroup_cs(cont); | |
525 | list_add_tail(&child->stack_list, &q); | |
526 | } | |
527 | } | |
528 | } | |
529 | ||
530 | /* | |
531 | * generate_sched_domains() | |
532 | * | |
533 | * This function builds a partial partition of the systems CPUs | |
534 | * A 'partial partition' is a set of non-overlapping subsets whose | |
535 | * union is a subset of that set. | |
536 | * The output of this function needs to be passed to kernel/sched.c | |
537 | * partition_sched_domains() routine, which will rebuild the scheduler's | |
538 | * load balancing domains (sched domains) as specified by that partial | |
539 | * partition. | |
540 | * | |
541 | * See "What is sched_load_balance" in Documentation/cpusets.txt | |
542 | * for a background explanation of this. | |
543 | * | |
544 | * Does not return errors, on the theory that the callers of this | |
545 | * routine would rather not worry about failures to rebuild sched | |
546 | * domains when operating in the severe memory shortage situations | |
547 | * that could cause allocation failures below. | |
548 | * | |
549 | * Must be called with cgroup_lock held. | |
550 | * | |
551 | * The three key local variables below are: | |
552 | * q - a linked-list queue of cpuset pointers, used to implement a | |
553 | * top-down scan of all cpusets. This scan loads a pointer | |
554 | * to each cpuset marked is_sched_load_balance into the | |
555 | * array 'csa'. For our purposes, rebuilding the schedulers | |
556 | * sched domains, we can ignore !is_sched_load_balance cpusets. | |
557 | * csa - (for CpuSet Array) Array of pointers to all the cpusets | |
558 | * that need to be load balanced, for convenient iterative | |
559 | * access by the subsequent code that finds the best partition, | |
560 | * i.e the set of domains (subsets) of CPUs such that the | |
561 | * cpus_allowed of every cpuset marked is_sched_load_balance | |
562 | * is a subset of one of these domains, while there are as | |
563 | * many such domains as possible, each as small as possible. | |
564 | * doms - Conversion of 'csa' to an array of cpumasks, for passing to | |
565 | * the kernel/sched.c routine partition_sched_domains() in a | |
566 | * convenient format, that can be easily compared to the prior | |
567 | * value to determine what partition elements (sched domains) | |
568 | * were changed (added or removed.) | |
569 | * | |
570 | * Finding the best partition (set of domains): | |
571 | * The triple nested loops below over i, j, k scan over the | |
572 | * load balanced cpusets (using the array of cpuset pointers in | |
573 | * csa[]) looking for pairs of cpusets that have overlapping | |
574 | * cpus_allowed, but which don't have the same 'pn' partition | |
575 | * number and gives them in the same partition number. It keeps | |
576 | * looping on the 'restart' label until it can no longer find | |
577 | * any such pairs. | |
578 | * | |
579 | * The union of the cpus_allowed masks from the set of | |
580 | * all cpusets having the same 'pn' value then form the one | |
581 | * element of the partition (one sched domain) to be passed to | |
582 | * partition_sched_domains(). | |
583 | */ | |
584 | static int generate_sched_domains(cpumask_t **domains, | |
585 | struct sched_domain_attr **attributes) | |
586 | { | |
587 | LIST_HEAD(q); /* queue of cpusets to be scanned */ | |
588 | struct cpuset *cp; /* scans q */ | |
589 | struct cpuset **csa; /* array of all cpuset ptrs */ | |
590 | int csn; /* how many cpuset ptrs in csa so far */ | |
591 | int i, j, k; /* indices for partition finding loops */ | |
592 | cpumask_t *doms; /* resulting partition; i.e. sched domains */ | |
593 | struct sched_domain_attr *dattr; /* attributes for custom domains */ | |
594 | int ndoms = 0; /* number of sched domains in result */ | |
595 | int nslot; /* next empty doms[] cpumask_t slot */ | |
596 | ||
597 | doms = NULL; | |
598 | dattr = NULL; | |
599 | csa = NULL; | |
600 | ||
601 | /* Special case for the 99% of systems with one, full, sched domain */ | |
602 | if (is_sched_load_balance(&top_cpuset)) { | |
603 | doms = kmalloc(sizeof(cpumask_t), GFP_KERNEL); | |
604 | if (!doms) | |
605 | goto done; | |
606 | ||
607 | dattr = kmalloc(sizeof(struct sched_domain_attr), GFP_KERNEL); | |
608 | if (dattr) { | |
609 | *dattr = SD_ATTR_INIT; | |
610 | update_domain_attr_tree(dattr, &top_cpuset); | |
611 | } | |
612 | *doms = top_cpuset.cpus_allowed; | |
613 | ||
614 | ndoms = 1; | |
615 | goto done; | |
616 | } | |
617 | ||
618 | csa = kmalloc(number_of_cpusets * sizeof(cp), GFP_KERNEL); | |
619 | if (!csa) | |
620 | goto done; | |
621 | csn = 0; | |
622 | ||
623 | list_add(&top_cpuset.stack_list, &q); | |
624 | while (!list_empty(&q)) { | |
625 | struct cgroup *cont; | |
626 | struct cpuset *child; /* scans child cpusets of cp */ | |
627 | ||
628 | cp = list_first_entry(&q, struct cpuset, stack_list); | |
629 | list_del(q.next); | |
630 | ||
631 | if (cpus_empty(cp->cpus_allowed)) | |
632 | continue; | |
633 | ||
634 | /* | |
635 | * All child cpusets contain a subset of the parent's cpus, so | |
636 | * just skip them, and then we call update_domain_attr_tree() | |
637 | * to calc relax_domain_level of the corresponding sched | |
638 | * domain. | |
639 | */ | |
640 | if (is_sched_load_balance(cp)) { | |
641 | csa[csn++] = cp; | |
642 | continue; | |
643 | } | |
644 | ||
645 | list_for_each_entry(cont, &cp->css.cgroup->children, sibling) { | |
646 | child = cgroup_cs(cont); | |
647 | list_add_tail(&child->stack_list, &q); | |
648 | } | |
649 | } | |
650 | ||
651 | for (i = 0; i < csn; i++) | |
652 | csa[i]->pn = i; | |
653 | ndoms = csn; | |
654 | ||
655 | restart: | |
656 | /* Find the best partition (set of sched domains) */ | |
657 | for (i = 0; i < csn; i++) { | |
658 | struct cpuset *a = csa[i]; | |
659 | int apn = a->pn; | |
660 | ||
661 | for (j = 0; j < csn; j++) { | |
662 | struct cpuset *b = csa[j]; | |
663 | int bpn = b->pn; | |
664 | ||
665 | if (apn != bpn && cpusets_overlap(a, b)) { | |
666 | for (k = 0; k < csn; k++) { | |
667 | struct cpuset *c = csa[k]; | |
668 | ||
669 | if (c->pn == bpn) | |
670 | c->pn = apn; | |
671 | } | |
672 | ndoms--; /* one less element */ | |
673 | goto restart; | |
674 | } | |
675 | } | |
676 | } | |
677 | ||
678 | /* | |
679 | * Now we know how many domains to create. | |
680 | * Convert <csn, csa> to <ndoms, doms> and populate cpu masks. | |
681 | */ | |
682 | doms = kmalloc(ndoms * sizeof(cpumask_t), GFP_KERNEL); | |
683 | if (!doms) | |
684 | goto done; | |
685 | ||
686 | /* | |
687 | * The rest of the code, including the scheduler, can deal with | |
688 | * dattr==NULL case. No need to abort if alloc fails. | |
689 | */ | |
690 | dattr = kmalloc(ndoms * sizeof(struct sched_domain_attr), GFP_KERNEL); | |
691 | ||
692 | for (nslot = 0, i = 0; i < csn; i++) { | |
693 | struct cpuset *a = csa[i]; | |
694 | cpumask_t *dp; | |
695 | int apn = a->pn; | |
696 | ||
697 | if (apn < 0) { | |
698 | /* Skip completed partitions */ | |
699 | continue; | |
700 | } | |
701 | ||
702 | dp = doms + nslot; | |
703 | ||
704 | if (nslot == ndoms) { | |
705 | static int warnings = 10; | |
706 | if (warnings) { | |
707 | printk(KERN_WARNING | |
708 | "rebuild_sched_domains confused:" | |
709 | " nslot %d, ndoms %d, csn %d, i %d," | |
710 | " apn %d\n", | |
711 | nslot, ndoms, csn, i, apn); | |
712 | warnings--; | |
713 | } | |
714 | continue; | |
715 | } | |
716 | ||
717 | cpus_clear(*dp); | |
718 | if (dattr) | |
719 | *(dattr + nslot) = SD_ATTR_INIT; | |
720 | for (j = i; j < csn; j++) { | |
721 | struct cpuset *b = csa[j]; | |
722 | ||
723 | if (apn == b->pn) { | |
724 | cpus_or(*dp, *dp, b->cpus_allowed); | |
725 | if (dattr) | |
726 | update_domain_attr_tree(dattr + nslot, b); | |
727 | ||
728 | /* Done with this partition */ | |
729 | b->pn = -1; | |
730 | } | |
731 | } | |
732 | nslot++; | |
733 | } | |
734 | BUG_ON(nslot != ndoms); | |
735 | ||
736 | done: | |
737 | kfree(csa); | |
738 | ||
739 | /* | |
740 | * Fallback to the default domain if kmalloc() failed. | |
741 | * See comments in partition_sched_domains(). | |
742 | */ | |
743 | if (doms == NULL) | |
744 | ndoms = 1; | |
745 | ||
746 | *domains = doms; | |
747 | *attributes = dattr; | |
748 | return ndoms; | |
749 | } | |
750 | ||
751 | /* | |
752 | * Rebuild scheduler domains. | |
753 | * | |
754 | * Call with neither cgroup_mutex held nor within get_online_cpus(). | |
755 | * Takes both cgroup_mutex and get_online_cpus(). | |
756 | * | |
757 | * Cannot be directly called from cpuset code handling changes | |
758 | * to the cpuset pseudo-filesystem, because it cannot be called | |
759 | * from code that already holds cgroup_mutex. | |
760 | */ | |
761 | static void do_rebuild_sched_domains(struct work_struct *unused) | |
762 | { | |
763 | struct sched_domain_attr *attr; | |
764 | cpumask_t *doms; | |
765 | int ndoms; | |
766 | ||
767 | get_online_cpus(); | |
768 | ||
769 | /* Generate domain masks and attrs */ | |
770 | cgroup_lock(); | |
771 | ndoms = generate_sched_domains(&doms, &attr); | |
772 | cgroup_unlock(); | |
773 | ||
774 | /* Have scheduler rebuild the domains */ | |
775 | partition_sched_domains(ndoms, doms, attr); | |
776 | ||
777 | put_online_cpus(); | |
778 | } | |
779 | ||
780 | static DECLARE_WORK(rebuild_sched_domains_work, do_rebuild_sched_domains); | |
781 | ||
782 | /* | |
783 | * Rebuild scheduler domains, asynchronously via workqueue. | |
784 | * | |
785 | * If the flag 'sched_load_balance' of any cpuset with non-empty | |
786 | * 'cpus' changes, or if the 'cpus' allowed changes in any cpuset | |
787 | * which has that flag enabled, or if any cpuset with a non-empty | |
788 | * 'cpus' is removed, then call this routine to rebuild the | |
789 | * scheduler's dynamic sched domains. | |
790 | * | |
791 | * The rebuild_sched_domains() and partition_sched_domains() | |
792 | * routines must nest cgroup_lock() inside get_online_cpus(), | |
793 | * but such cpuset changes as these must nest that locking the | |
794 | * other way, holding cgroup_lock() for much of the code. | |
795 | * | |
796 | * So in order to avoid an ABBA deadlock, the cpuset code handling | |
797 | * these user changes delegates the actual sched domain rebuilding | |
798 | * to a separate workqueue thread, which ends up processing the | |
799 | * above do_rebuild_sched_domains() function. | |
800 | */ | |
801 | static void async_rebuild_sched_domains(void) | |
802 | { | |
803 | schedule_work(&rebuild_sched_domains_work); | |
804 | } | |
805 | ||
806 | /* | |
807 | * Accomplishes the same scheduler domain rebuild as the above | |
808 | * async_rebuild_sched_domains(), however it directly calls the | |
809 | * rebuild routine synchronously rather than calling it via an | |
810 | * asynchronous work thread. | |
811 | * | |
812 | * This can only be called from code that is not holding | |
813 | * cgroup_mutex (not nested in a cgroup_lock() call.) | |
814 | */ | |
815 | void rebuild_sched_domains(void) | |
816 | { | |
817 | do_rebuild_sched_domains(NULL); | |
818 | } | |
819 | ||
820 | /** | |
821 | * cpuset_test_cpumask - test a task's cpus_allowed versus its cpuset's | |
822 | * @tsk: task to test | |
823 | * @scan: struct cgroup_scanner contained in its struct cpuset_hotplug_scanner | |
824 | * | |
825 | * Call with cgroup_mutex held. May take callback_mutex during call. | |
826 | * Called for each task in a cgroup by cgroup_scan_tasks(). | |
827 | * Return nonzero if this tasks's cpus_allowed mask should be changed (in other | |
828 | * words, if its mask is not equal to its cpuset's mask). | |
829 | */ | |
830 | static int cpuset_test_cpumask(struct task_struct *tsk, | |
831 | struct cgroup_scanner *scan) | |
832 | { | |
833 | return !cpus_equal(tsk->cpus_allowed, | |
834 | (cgroup_cs(scan->cg))->cpus_allowed); | |
835 | } | |
836 | ||
837 | /** | |
838 | * cpuset_change_cpumask - make a task's cpus_allowed the same as its cpuset's | |
839 | * @tsk: task to test | |
840 | * @scan: struct cgroup_scanner containing the cgroup of the task | |
841 | * | |
842 | * Called by cgroup_scan_tasks() for each task in a cgroup whose | |
843 | * cpus_allowed mask needs to be changed. | |
844 | * | |
845 | * We don't need to re-check for the cgroup/cpuset membership, since we're | |
846 | * holding cgroup_lock() at this point. | |
847 | */ | |
848 | static void cpuset_change_cpumask(struct task_struct *tsk, | |
849 | struct cgroup_scanner *scan) | |
850 | { | |
851 | set_cpus_allowed_ptr(tsk, &((cgroup_cs(scan->cg))->cpus_allowed)); | |
852 | } | |
853 | ||
854 | /** | |
855 | * update_tasks_cpumask - Update the cpumasks of tasks in the cpuset. | |
856 | * @cs: the cpuset in which each task's cpus_allowed mask needs to be changed | |
857 | * @heap: if NULL, defer allocating heap memory to cgroup_scan_tasks() | |
858 | * | |
859 | * Called with cgroup_mutex held | |
860 | * | |
861 | * The cgroup_scan_tasks() function will scan all the tasks in a cgroup, | |
862 | * calling callback functions for each. | |
863 | * | |
864 | * No return value. It's guaranteed that cgroup_scan_tasks() always returns 0 | |
865 | * if @heap != NULL. | |
866 | */ | |
867 | static void update_tasks_cpumask(struct cpuset *cs, struct ptr_heap *heap) | |
868 | { | |
869 | struct cgroup_scanner scan; | |
870 | ||
871 | scan.cg = cs->css.cgroup; | |
872 | scan.test_task = cpuset_test_cpumask; | |
873 | scan.process_task = cpuset_change_cpumask; | |
874 | scan.heap = heap; | |
875 | cgroup_scan_tasks(&scan); | |
876 | } | |
877 | ||
878 | /** | |
879 | * update_cpumask - update the cpus_allowed mask of a cpuset and all tasks in it | |
880 | * @cs: the cpuset to consider | |
881 | * @buf: buffer of cpu numbers written to this cpuset | |
882 | */ | |
883 | static int update_cpumask(struct cpuset *cs, const char *buf) | |
884 | { | |
885 | struct ptr_heap heap; | |
886 | struct cpuset trialcs; | |
887 | int retval; | |
888 | int is_load_balanced; | |
889 | ||
890 | /* top_cpuset.cpus_allowed tracks cpu_online_map; it's read-only */ | |
891 | if (cs == &top_cpuset) | |
892 | return -EACCES; | |
893 | ||
894 | trialcs = *cs; | |
895 | ||
896 | /* | |
897 | * An empty cpus_allowed is ok only if the cpuset has no tasks. | |
898 | * Since cpulist_parse() fails on an empty mask, we special case | |
899 | * that parsing. The validate_change() call ensures that cpusets | |
900 | * with tasks have cpus. | |
901 | */ | |
902 | if (!*buf) { | |
903 | cpus_clear(trialcs.cpus_allowed); | |
904 | } else { | |
905 | retval = cpulist_parse(buf, &trialcs.cpus_allowed); | |
906 | if (retval < 0) | |
907 | return retval; | |
908 | ||
909 | if (!cpus_subset(trialcs.cpus_allowed, cpu_online_map)) | |
910 | return -EINVAL; | |
911 | } | |
912 | retval = validate_change(cs, &trialcs); | |
913 | if (retval < 0) | |
914 | return retval; | |
915 | ||
916 | /* Nothing to do if the cpus didn't change */ | |
917 | if (cpus_equal(cs->cpus_allowed, trialcs.cpus_allowed)) | |
918 | return 0; | |
919 | ||
920 | retval = heap_init(&heap, PAGE_SIZE, GFP_KERNEL, NULL); | |
921 | if (retval) | |
922 | return retval; | |
923 | ||
924 | is_load_balanced = is_sched_load_balance(&trialcs); | |
925 | ||
926 | mutex_lock(&callback_mutex); | |
927 | cs->cpus_allowed = trialcs.cpus_allowed; | |
928 | mutex_unlock(&callback_mutex); | |
929 | ||
930 | /* | |
931 | * Scan tasks in the cpuset, and update the cpumasks of any | |
932 | * that need an update. | |
933 | */ | |
934 | update_tasks_cpumask(cs, &heap); | |
935 | ||
936 | heap_free(&heap); | |
937 | ||
938 | if (is_load_balanced) | |
939 | async_rebuild_sched_domains(); | |
940 | return 0; | |
941 | } | |
942 | ||
943 | /* | |
944 | * cpuset_migrate_mm | |
945 | * | |
946 | * Migrate memory region from one set of nodes to another. | |
947 | * | |
948 | * Temporarilly set tasks mems_allowed to target nodes of migration, | |
949 | * so that the migration code can allocate pages on these nodes. | |
950 | * | |
951 | * Call holding cgroup_mutex, so current's cpuset won't change | |
952 | * during this call, as manage_mutex holds off any cpuset_attach() | |
953 | * calls. Therefore we don't need to take task_lock around the | |
954 | * call to guarantee_online_mems(), as we know no one is changing | |
955 | * our task's cpuset. | |
956 | * | |
957 | * Hold callback_mutex around the two modifications of our tasks | |
958 | * mems_allowed to synchronize with cpuset_mems_allowed(). | |
959 | * | |
960 | * While the mm_struct we are migrating is typically from some | |
961 | * other task, the task_struct mems_allowed that we are hacking | |
962 | * is for our current task, which must allocate new pages for that | |
963 | * migrating memory region. | |
964 | * | |
965 | * We call cpuset_update_task_memory_state() before hacking | |
966 | * our tasks mems_allowed, so that we are assured of being in | |
967 | * sync with our tasks cpuset, and in particular, callbacks to | |
968 | * cpuset_update_task_memory_state() from nested page allocations | |
969 | * won't see any mismatch of our cpuset and task mems_generation | |
970 | * values, so won't overwrite our hacked tasks mems_allowed | |
971 | * nodemask. | |
972 | */ | |
973 | ||
974 | static void cpuset_migrate_mm(struct mm_struct *mm, const nodemask_t *from, | |
975 | const nodemask_t *to) | |
976 | { | |
977 | struct task_struct *tsk = current; | |
978 | ||
979 | cpuset_update_task_memory_state(); | |
980 | ||
981 | mutex_lock(&callback_mutex); | |
982 | tsk->mems_allowed = *to; | |
983 | mutex_unlock(&callback_mutex); | |
984 | ||
985 | do_migrate_pages(mm, from, to, MPOL_MF_MOVE_ALL); | |
986 | ||
987 | mutex_lock(&callback_mutex); | |
988 | guarantee_online_mems(task_cs(tsk),&tsk->mems_allowed); | |
989 | mutex_unlock(&callback_mutex); | |
990 | } | |
991 | ||
992 | static void *cpuset_being_rebound; | |
993 | ||
994 | /** | |
995 | * update_tasks_nodemask - Update the nodemasks of tasks in the cpuset. | |
996 | * @cs: the cpuset in which each task's mems_allowed mask needs to be changed | |
997 | * @oldmem: old mems_allowed of cpuset cs | |
998 | * | |
999 | * Called with cgroup_mutex held | |
1000 | * Return 0 if successful, -errno if not. | |
1001 | */ | |
1002 | static int update_tasks_nodemask(struct cpuset *cs, const nodemask_t *oldmem) | |
1003 | { | |
1004 | struct task_struct *p; | |
1005 | struct mm_struct **mmarray; | |
1006 | int i, n, ntasks; | |
1007 | int migrate; | |
1008 | int fudge; | |
1009 | struct cgroup_iter it; | |
1010 | int retval; | |
1011 | ||
1012 | cpuset_being_rebound = cs; /* causes mpol_dup() rebind */ | |
1013 | ||
1014 | fudge = 10; /* spare mmarray[] slots */ | |
1015 | fudge += cpus_weight(cs->cpus_allowed); /* imagine one fork-bomb/cpu */ | |
1016 | retval = -ENOMEM; | |
1017 | ||
1018 | /* | |
1019 | * Allocate mmarray[] to hold mm reference for each task | |
1020 | * in cpuset cs. Can't kmalloc GFP_KERNEL while holding | |
1021 | * tasklist_lock. We could use GFP_ATOMIC, but with a | |
1022 | * few more lines of code, we can retry until we get a big | |
1023 | * enough mmarray[] w/o using GFP_ATOMIC. | |
1024 | */ | |
1025 | while (1) { | |
1026 | ntasks = cgroup_task_count(cs->css.cgroup); /* guess */ | |
1027 | ntasks += fudge; | |
1028 | mmarray = kmalloc(ntasks * sizeof(*mmarray), GFP_KERNEL); | |
1029 | if (!mmarray) | |
1030 | goto done; | |
1031 | read_lock(&tasklist_lock); /* block fork */ | |
1032 | if (cgroup_task_count(cs->css.cgroup) <= ntasks) | |
1033 | break; /* got enough */ | |
1034 | read_unlock(&tasklist_lock); /* try again */ | |
1035 | kfree(mmarray); | |
1036 | } | |
1037 | ||
1038 | n = 0; | |
1039 | ||
1040 | /* Load up mmarray[] with mm reference for each task in cpuset. */ | |
1041 | cgroup_iter_start(cs->css.cgroup, &it); | |
1042 | while ((p = cgroup_iter_next(cs->css.cgroup, &it))) { | |
1043 | struct mm_struct *mm; | |
1044 | ||
1045 | if (n >= ntasks) { | |
1046 | printk(KERN_WARNING | |
1047 | "Cpuset mempolicy rebind incomplete.\n"); | |
1048 | break; | |
1049 | } | |
1050 | mm = get_task_mm(p); | |
1051 | if (!mm) | |
1052 | continue; | |
1053 | mmarray[n++] = mm; | |
1054 | } | |
1055 | cgroup_iter_end(cs->css.cgroup, &it); | |
1056 | read_unlock(&tasklist_lock); | |
1057 | ||
1058 | /* | |
1059 | * Now that we've dropped the tasklist spinlock, we can | |
1060 | * rebind the vma mempolicies of each mm in mmarray[] to their | |
1061 | * new cpuset, and release that mm. The mpol_rebind_mm() | |
1062 | * call takes mmap_sem, which we couldn't take while holding | |
1063 | * tasklist_lock. Forks can happen again now - the mpol_dup() | |
1064 | * cpuset_being_rebound check will catch such forks, and rebind | |
1065 | * their vma mempolicies too. Because we still hold the global | |
1066 | * cgroup_mutex, we know that no other rebind effort will | |
1067 | * be contending for the global variable cpuset_being_rebound. | |
1068 | * It's ok if we rebind the same mm twice; mpol_rebind_mm() | |
1069 | * is idempotent. Also migrate pages in each mm to new nodes. | |
1070 | */ | |
1071 | migrate = is_memory_migrate(cs); | |
1072 | for (i = 0; i < n; i++) { | |
1073 | struct mm_struct *mm = mmarray[i]; | |
1074 | ||
1075 | mpol_rebind_mm(mm, &cs->mems_allowed); | |
1076 | if (migrate) | |
1077 | cpuset_migrate_mm(mm, oldmem, &cs->mems_allowed); | |
1078 | mmput(mm); | |
1079 | } | |
1080 | ||
1081 | /* We're done rebinding vmas to this cpuset's new mems_allowed. */ | |
1082 | kfree(mmarray); | |
1083 | cpuset_being_rebound = NULL; | |
1084 | retval = 0; | |
1085 | done: | |
1086 | return retval; | |
1087 | } | |
1088 | ||
1089 | /* | |
1090 | * Handle user request to change the 'mems' memory placement | |
1091 | * of a cpuset. Needs to validate the request, update the | |
1092 | * cpusets mems_allowed and mems_generation, and for each | |
1093 | * task in the cpuset, rebind any vma mempolicies and if | |
1094 | * the cpuset is marked 'memory_migrate', migrate the tasks | |
1095 | * pages to the new memory. | |
1096 | * | |
1097 | * Call with cgroup_mutex held. May take callback_mutex during call. | |
1098 | * Will take tasklist_lock, scan tasklist for tasks in cpuset cs, | |
1099 | * lock each such tasks mm->mmap_sem, scan its vma's and rebind | |
1100 | * their mempolicies to the cpusets new mems_allowed. | |
1101 | */ | |
1102 | static int update_nodemask(struct cpuset *cs, const char *buf) | |
1103 | { | |
1104 | struct cpuset trialcs; | |
1105 | nodemask_t oldmem; | |
1106 | int retval; | |
1107 | ||
1108 | /* | |
1109 | * top_cpuset.mems_allowed tracks node_stats[N_HIGH_MEMORY]; | |
1110 | * it's read-only | |
1111 | */ | |
1112 | if (cs == &top_cpuset) | |
1113 | return -EACCES; | |
1114 | ||
1115 | trialcs = *cs; | |
1116 | ||
1117 | /* | |
1118 | * An empty mems_allowed is ok iff there are no tasks in the cpuset. | |
1119 | * Since nodelist_parse() fails on an empty mask, we special case | |
1120 | * that parsing. The validate_change() call ensures that cpusets | |
1121 | * with tasks have memory. | |
1122 | */ | |
1123 | if (!*buf) { | |
1124 | nodes_clear(trialcs.mems_allowed); | |
1125 | } else { | |
1126 | retval = nodelist_parse(buf, trialcs.mems_allowed); | |
1127 | if (retval < 0) | |
1128 | goto done; | |
1129 | ||
1130 | if (!nodes_subset(trialcs.mems_allowed, | |
1131 | node_states[N_HIGH_MEMORY])) | |
1132 | return -EINVAL; | |
1133 | } | |
1134 | oldmem = cs->mems_allowed; | |
1135 | if (nodes_equal(oldmem, trialcs.mems_allowed)) { | |
1136 | retval = 0; /* Too easy - nothing to do */ | |
1137 | goto done; | |
1138 | } | |
1139 | retval = validate_change(cs, &trialcs); | |
1140 | if (retval < 0) | |
1141 | goto done; | |
1142 | ||
1143 | mutex_lock(&callback_mutex); | |
1144 | cs->mems_allowed = trialcs.mems_allowed; | |
1145 | cs->mems_generation = cpuset_mems_generation++; | |
1146 | mutex_unlock(&callback_mutex); | |
1147 | ||
1148 | retval = update_tasks_nodemask(cs, &oldmem); | |
1149 | done: | |
1150 | return retval; | |
1151 | } | |
1152 | ||
1153 | int current_cpuset_is_being_rebound(void) | |
1154 | { | |
1155 | return task_cs(current) == cpuset_being_rebound; | |
1156 | } | |
1157 | ||
1158 | static int update_relax_domain_level(struct cpuset *cs, s64 val) | |
1159 | { | |
1160 | if (val < -1 || val >= SD_LV_MAX) | |
1161 | return -EINVAL; | |
1162 | ||
1163 | if (val != cs->relax_domain_level) { | |
1164 | cs->relax_domain_level = val; | |
1165 | if (!cpus_empty(cs->cpus_allowed) && is_sched_load_balance(cs)) | |
1166 | async_rebuild_sched_domains(); | |
1167 | } | |
1168 | ||
1169 | return 0; | |
1170 | } | |
1171 | ||
1172 | /* | |
1173 | * update_flag - read a 0 or a 1 in a file and update associated flag | |
1174 | * bit: the bit to update (see cpuset_flagbits_t) | |
1175 | * cs: the cpuset to update | |
1176 | * turning_on: whether the flag is being set or cleared | |
1177 | * | |
1178 | * Call with cgroup_mutex held. | |
1179 | */ | |
1180 | ||
1181 | static int update_flag(cpuset_flagbits_t bit, struct cpuset *cs, | |
1182 | int turning_on) | |
1183 | { | |
1184 | struct cpuset trialcs; | |
1185 | int err; | |
1186 | int balance_flag_changed; | |
1187 | ||
1188 | trialcs = *cs; | |
1189 | if (turning_on) | |
1190 | set_bit(bit, &trialcs.flags); | |
1191 | else | |
1192 | clear_bit(bit, &trialcs.flags); | |
1193 | ||
1194 | err = validate_change(cs, &trialcs); | |
1195 | if (err < 0) | |
1196 | return err; | |
1197 | ||
1198 | balance_flag_changed = (is_sched_load_balance(cs) != | |
1199 | is_sched_load_balance(&trialcs)); | |
1200 | ||
1201 | mutex_lock(&callback_mutex); | |
1202 | cs->flags = trialcs.flags; | |
1203 | mutex_unlock(&callback_mutex); | |
1204 | ||
1205 | if (!cpus_empty(trialcs.cpus_allowed) && balance_flag_changed) | |
1206 | async_rebuild_sched_domains(); | |
1207 | ||
1208 | return 0; | |
1209 | } | |
1210 | ||
1211 | /* | |
1212 | * Frequency meter - How fast is some event occurring? | |
1213 | * | |
1214 | * These routines manage a digitally filtered, constant time based, | |
1215 | * event frequency meter. There are four routines: | |
1216 | * fmeter_init() - initialize a frequency meter. | |
1217 | * fmeter_markevent() - called each time the event happens. | |
1218 | * fmeter_getrate() - returns the recent rate of such events. | |
1219 | * fmeter_update() - internal routine used to update fmeter. | |
1220 | * | |
1221 | * A common data structure is passed to each of these routines, | |
1222 | * which is used to keep track of the state required to manage the | |
1223 | * frequency meter and its digital filter. | |
1224 | * | |
1225 | * The filter works on the number of events marked per unit time. | |
1226 | * The filter is single-pole low-pass recursive (IIR). The time unit | |
1227 | * is 1 second. Arithmetic is done using 32-bit integers scaled to | |
1228 | * simulate 3 decimal digits of precision (multiplied by 1000). | |
1229 | * | |
1230 | * With an FM_COEF of 933, and a time base of 1 second, the filter | |
1231 | * has a half-life of 10 seconds, meaning that if the events quit | |
1232 | * happening, then the rate returned from the fmeter_getrate() | |
1233 | * will be cut in half each 10 seconds, until it converges to zero. | |
1234 | * | |
1235 | * It is not worth doing a real infinitely recursive filter. If more | |
1236 | * than FM_MAXTICKS ticks have elapsed since the last filter event, | |
1237 | * just compute FM_MAXTICKS ticks worth, by which point the level | |
1238 | * will be stable. | |
1239 | * | |
1240 | * Limit the count of unprocessed events to FM_MAXCNT, so as to avoid | |
1241 | * arithmetic overflow in the fmeter_update() routine. | |
1242 | * | |
1243 | * Given the simple 32 bit integer arithmetic used, this meter works | |
1244 | * best for reporting rates between one per millisecond (msec) and | |
1245 | * one per 32 (approx) seconds. At constant rates faster than one | |
1246 | * per msec it maxes out at values just under 1,000,000. At constant | |
1247 | * rates between one per msec, and one per second it will stabilize | |
1248 | * to a value N*1000, where N is the rate of events per second. | |
1249 | * At constant rates between one per second and one per 32 seconds, | |
1250 | * it will be choppy, moving up on the seconds that have an event, | |
1251 | * and then decaying until the next event. At rates slower than | |
1252 | * about one in 32 seconds, it decays all the way back to zero between | |
1253 | * each event. | |
1254 | */ | |
1255 | ||
1256 | #define FM_COEF 933 /* coefficient for half-life of 10 secs */ | |
1257 | #define FM_MAXTICKS ((time_t)99) /* useless computing more ticks than this */ | |
1258 | #define FM_MAXCNT 1000000 /* limit cnt to avoid overflow */ | |
1259 | #define FM_SCALE 1000 /* faux fixed point scale */ | |
1260 | ||
1261 | /* Initialize a frequency meter */ | |
1262 | static void fmeter_init(struct fmeter *fmp) | |
1263 | { | |
1264 | fmp->cnt = 0; | |
1265 | fmp->val = 0; | |
1266 | fmp->time = 0; | |
1267 | spin_lock_init(&fmp->lock); | |
1268 | } | |
1269 | ||
1270 | /* Internal meter update - process cnt events and update value */ | |
1271 | static void fmeter_update(struct fmeter *fmp) | |
1272 | { | |
1273 | time_t now = get_seconds(); | |
1274 | time_t ticks = now - fmp->time; | |
1275 | ||
1276 | if (ticks == 0) | |
1277 | return; | |
1278 | ||
1279 | ticks = min(FM_MAXTICKS, ticks); | |
1280 | while (ticks-- > 0) | |
1281 | fmp->val = (FM_COEF * fmp->val) / FM_SCALE; | |
1282 | fmp->time = now; | |
1283 | ||
1284 | fmp->val += ((FM_SCALE - FM_COEF) * fmp->cnt) / FM_SCALE; | |
1285 | fmp->cnt = 0; | |
1286 | } | |
1287 | ||
1288 | /* Process any previous ticks, then bump cnt by one (times scale). */ | |
1289 | static void fmeter_markevent(struct fmeter *fmp) | |
1290 | { | |
1291 | spin_lock(&fmp->lock); | |
1292 | fmeter_update(fmp); | |
1293 | fmp->cnt = min(FM_MAXCNT, fmp->cnt + FM_SCALE); | |
1294 | spin_unlock(&fmp->lock); | |
1295 | } | |
1296 | ||
1297 | /* Process any previous ticks, then return current value. */ | |
1298 | static int fmeter_getrate(struct fmeter *fmp) | |
1299 | { | |
1300 | int val; | |
1301 | ||
1302 | spin_lock(&fmp->lock); | |
1303 | fmeter_update(fmp); | |
1304 | val = fmp->val; | |
1305 | spin_unlock(&fmp->lock); | |
1306 | return val; | |
1307 | } | |
1308 | ||
1309 | /* Protected by cgroup_lock */ | |
1310 | static cpumask_var_t cpus_attach; | |
1311 | ||
1312 | /* Called by cgroups to determine if a cpuset is usable; cgroup_mutex held */ | |
1313 | static int cpuset_can_attach(struct cgroup_subsys *ss, | |
1314 | struct cgroup *cont, struct task_struct *tsk) | |
1315 | { | |
1316 | struct cpuset *cs = cgroup_cs(cont); | |
1317 | int ret = 0; | |
1318 | ||
1319 | if (cpus_empty(cs->cpus_allowed) || nodes_empty(cs->mems_allowed)) | |
1320 | return -ENOSPC; | |
1321 | ||
1322 | if (tsk->flags & PF_THREAD_BOUND) { | |
1323 | mutex_lock(&callback_mutex); | |
1324 | if (!cpus_equal(tsk->cpus_allowed, cs->cpus_allowed)) | |
1325 | ret = -EINVAL; | |
1326 | mutex_unlock(&callback_mutex); | |
1327 | } | |
1328 | ||
1329 | return ret < 0 ? ret : security_task_setscheduler(tsk, 0, NULL); | |
1330 | } | |
1331 | ||
1332 | static void cpuset_attach(struct cgroup_subsys *ss, | |
1333 | struct cgroup *cont, struct cgroup *oldcont, | |
1334 | struct task_struct *tsk) | |
1335 | { | |
1336 | nodemask_t from, to; | |
1337 | struct mm_struct *mm; | |
1338 | struct cpuset *cs = cgroup_cs(cont); | |
1339 | struct cpuset *oldcs = cgroup_cs(oldcont); | |
1340 | int err; | |
1341 | ||
1342 | if (cs == &top_cpuset) { | |
1343 | cpumask_copy(cpus_attach, cpu_possible_mask); | |
1344 | } else { | |
1345 | mutex_lock(&callback_mutex); | |
1346 | guarantee_online_cpus(cs, cpus_attach); | |
1347 | mutex_unlock(&callback_mutex); | |
1348 | } | |
1349 | err = set_cpus_allowed_ptr(tsk, cpus_attach); | |
1350 | if (err) | |
1351 | return; | |
1352 | ||
1353 | from = oldcs->mems_allowed; | |
1354 | to = cs->mems_allowed; | |
1355 | mm = get_task_mm(tsk); | |
1356 | if (mm) { | |
1357 | mpol_rebind_mm(mm, &to); | |
1358 | if (is_memory_migrate(cs)) | |
1359 | cpuset_migrate_mm(mm, &from, &to); | |
1360 | mmput(mm); | |
1361 | } | |
1362 | } | |
1363 | ||
1364 | /* The various types of files and directories in a cpuset file system */ | |
1365 | ||
1366 | typedef enum { | |
1367 | FILE_MEMORY_MIGRATE, | |
1368 | FILE_CPULIST, | |
1369 | FILE_MEMLIST, | |
1370 | FILE_CPU_EXCLUSIVE, | |
1371 | FILE_MEM_EXCLUSIVE, | |
1372 | FILE_MEM_HARDWALL, | |
1373 | FILE_SCHED_LOAD_BALANCE, | |
1374 | FILE_SCHED_RELAX_DOMAIN_LEVEL, | |
1375 | FILE_MEMORY_PRESSURE_ENABLED, | |
1376 | FILE_MEMORY_PRESSURE, | |
1377 | FILE_SPREAD_PAGE, | |
1378 | FILE_SPREAD_SLAB, | |
1379 | } cpuset_filetype_t; | |
1380 | ||
1381 | static int cpuset_write_u64(struct cgroup *cgrp, struct cftype *cft, u64 val) | |
1382 | { | |
1383 | int retval = 0; | |
1384 | struct cpuset *cs = cgroup_cs(cgrp); | |
1385 | cpuset_filetype_t type = cft->private; | |
1386 | ||
1387 | if (!cgroup_lock_live_group(cgrp)) | |
1388 | return -ENODEV; | |
1389 | ||
1390 | switch (type) { | |
1391 | case FILE_CPU_EXCLUSIVE: | |
1392 | retval = update_flag(CS_CPU_EXCLUSIVE, cs, val); | |
1393 | break; | |
1394 | case FILE_MEM_EXCLUSIVE: | |
1395 | retval = update_flag(CS_MEM_EXCLUSIVE, cs, val); | |
1396 | break; | |
1397 | case FILE_MEM_HARDWALL: | |
1398 | retval = update_flag(CS_MEM_HARDWALL, cs, val); | |
1399 | break; | |
1400 | case FILE_SCHED_LOAD_BALANCE: | |
1401 | retval = update_flag(CS_SCHED_LOAD_BALANCE, cs, val); | |
1402 | break; | |
1403 | case FILE_MEMORY_MIGRATE: | |
1404 | retval = update_flag(CS_MEMORY_MIGRATE, cs, val); | |
1405 | break; | |
1406 | case FILE_MEMORY_PRESSURE_ENABLED: | |
1407 | cpuset_memory_pressure_enabled = !!val; | |
1408 | break; | |
1409 | case FILE_MEMORY_PRESSURE: | |
1410 | retval = -EACCES; | |
1411 | break; | |
1412 | case FILE_SPREAD_PAGE: | |
1413 | retval = update_flag(CS_SPREAD_PAGE, cs, val); | |
1414 | cs->mems_generation = cpuset_mems_generation++; | |
1415 | break; | |
1416 | case FILE_SPREAD_SLAB: | |
1417 | retval = update_flag(CS_SPREAD_SLAB, cs, val); | |
1418 | cs->mems_generation = cpuset_mems_generation++; | |
1419 | break; | |
1420 | default: | |
1421 | retval = -EINVAL; | |
1422 | break; | |
1423 | } | |
1424 | cgroup_unlock(); | |
1425 | return retval; | |
1426 | } | |
1427 | ||
1428 | static int cpuset_write_s64(struct cgroup *cgrp, struct cftype *cft, s64 val) | |
1429 | { | |
1430 | int retval = 0; | |
1431 | struct cpuset *cs = cgroup_cs(cgrp); | |
1432 | cpuset_filetype_t type = cft->private; | |
1433 | ||
1434 | if (!cgroup_lock_live_group(cgrp)) | |
1435 | return -ENODEV; | |
1436 | ||
1437 | switch (type) { | |
1438 | case FILE_SCHED_RELAX_DOMAIN_LEVEL: | |
1439 | retval = update_relax_domain_level(cs, val); | |
1440 | break; | |
1441 | default: | |
1442 | retval = -EINVAL; | |
1443 | break; | |
1444 | } | |
1445 | cgroup_unlock(); | |
1446 | return retval; | |
1447 | } | |
1448 | ||
1449 | /* | |
1450 | * Common handling for a write to a "cpus" or "mems" file. | |
1451 | */ | |
1452 | static int cpuset_write_resmask(struct cgroup *cgrp, struct cftype *cft, | |
1453 | const char *buf) | |
1454 | { | |
1455 | int retval = 0; | |
1456 | ||
1457 | if (!cgroup_lock_live_group(cgrp)) | |
1458 | return -ENODEV; | |
1459 | ||
1460 | switch (cft->private) { | |
1461 | case FILE_CPULIST: | |
1462 | retval = update_cpumask(cgroup_cs(cgrp), buf); | |
1463 | break; | |
1464 | case FILE_MEMLIST: | |
1465 | retval = update_nodemask(cgroup_cs(cgrp), buf); | |
1466 | break; | |
1467 | default: | |
1468 | retval = -EINVAL; | |
1469 | break; | |
1470 | } | |
1471 | cgroup_unlock(); | |
1472 | return retval; | |
1473 | } | |
1474 | ||
1475 | /* | |
1476 | * These ascii lists should be read in a single call, by using a user | |
1477 | * buffer large enough to hold the entire map. If read in smaller | |
1478 | * chunks, there is no guarantee of atomicity. Since the display format | |
1479 | * used, list of ranges of sequential numbers, is variable length, | |
1480 | * and since these maps can change value dynamically, one could read | |
1481 | * gibberish by doing partial reads while a list was changing. | |
1482 | * A single large read to a buffer that crosses a page boundary is | |
1483 | * ok, because the result being copied to user land is not recomputed | |
1484 | * across a page fault. | |
1485 | */ | |
1486 | ||
1487 | static int cpuset_sprintf_cpulist(char *page, struct cpuset *cs) | |
1488 | { | |
1489 | int ret; | |
1490 | ||
1491 | mutex_lock(&callback_mutex); | |
1492 | ret = cpulist_scnprintf(page, PAGE_SIZE, &cs->cpus_allowed); | |
1493 | mutex_unlock(&callback_mutex); | |
1494 | ||
1495 | return ret; | |
1496 | } | |
1497 | ||
1498 | static int cpuset_sprintf_memlist(char *page, struct cpuset *cs) | |
1499 | { | |
1500 | nodemask_t mask; | |
1501 | ||
1502 | mutex_lock(&callback_mutex); | |
1503 | mask = cs->mems_allowed; | |
1504 | mutex_unlock(&callback_mutex); | |
1505 | ||
1506 | return nodelist_scnprintf(page, PAGE_SIZE, mask); | |
1507 | } | |
1508 | ||
1509 | static ssize_t cpuset_common_file_read(struct cgroup *cont, | |
1510 | struct cftype *cft, | |
1511 | struct file *file, | |
1512 | char __user *buf, | |
1513 | size_t nbytes, loff_t *ppos) | |
1514 | { | |
1515 | struct cpuset *cs = cgroup_cs(cont); | |
1516 | cpuset_filetype_t type = cft->private; | |
1517 | char *page; | |
1518 | ssize_t retval = 0; | |
1519 | char *s; | |
1520 | ||
1521 | if (!(page = (char *)__get_free_page(GFP_TEMPORARY))) | |
1522 | return -ENOMEM; | |
1523 | ||
1524 | s = page; | |
1525 | ||
1526 | switch (type) { | |
1527 | case FILE_CPULIST: | |
1528 | s += cpuset_sprintf_cpulist(s, cs); | |
1529 | break; | |
1530 | case FILE_MEMLIST: | |
1531 | s += cpuset_sprintf_memlist(s, cs); | |
1532 | break; | |
1533 | default: | |
1534 | retval = -EINVAL; | |
1535 | goto out; | |
1536 | } | |
1537 | *s++ = '\n'; | |
1538 | ||
1539 | retval = simple_read_from_buffer(buf, nbytes, ppos, page, s - page); | |
1540 | out: | |
1541 | free_page((unsigned long)page); | |
1542 | return retval; | |
1543 | } | |
1544 | ||
1545 | static u64 cpuset_read_u64(struct cgroup *cont, struct cftype *cft) | |
1546 | { | |
1547 | struct cpuset *cs = cgroup_cs(cont); | |
1548 | cpuset_filetype_t type = cft->private; | |
1549 | switch (type) { | |
1550 | case FILE_CPU_EXCLUSIVE: | |
1551 | return is_cpu_exclusive(cs); | |
1552 | case FILE_MEM_EXCLUSIVE: | |
1553 | return is_mem_exclusive(cs); | |
1554 | case FILE_MEM_HARDWALL: | |
1555 | return is_mem_hardwall(cs); | |
1556 | case FILE_SCHED_LOAD_BALANCE: | |
1557 | return is_sched_load_balance(cs); | |
1558 | case FILE_MEMORY_MIGRATE: | |
1559 | return is_memory_migrate(cs); | |
1560 | case FILE_MEMORY_PRESSURE_ENABLED: | |
1561 | return cpuset_memory_pressure_enabled; | |
1562 | case FILE_MEMORY_PRESSURE: | |
1563 | return fmeter_getrate(&cs->fmeter); | |
1564 | case FILE_SPREAD_PAGE: | |
1565 | return is_spread_page(cs); | |
1566 | case FILE_SPREAD_SLAB: | |
1567 | return is_spread_slab(cs); | |
1568 | default: | |
1569 | BUG(); | |
1570 | } | |
1571 | ||
1572 | /* Unreachable but makes gcc happy */ | |
1573 | return 0; | |
1574 | } | |
1575 | ||
1576 | static s64 cpuset_read_s64(struct cgroup *cont, struct cftype *cft) | |
1577 | { | |
1578 | struct cpuset *cs = cgroup_cs(cont); | |
1579 | cpuset_filetype_t type = cft->private; | |
1580 | switch (type) { | |
1581 | case FILE_SCHED_RELAX_DOMAIN_LEVEL: | |
1582 | return cs->relax_domain_level; | |
1583 | default: | |
1584 | BUG(); | |
1585 | } | |
1586 | ||
1587 | /* Unrechable but makes gcc happy */ | |
1588 | return 0; | |
1589 | } | |
1590 | ||
1591 | ||
1592 | /* | |
1593 | * for the common functions, 'private' gives the type of file | |
1594 | */ | |
1595 | ||
1596 | static struct cftype files[] = { | |
1597 | { | |
1598 | .name = "cpus", | |
1599 | .read = cpuset_common_file_read, | |
1600 | .write_string = cpuset_write_resmask, | |
1601 | .max_write_len = (100U + 6 * NR_CPUS), | |
1602 | .private = FILE_CPULIST, | |
1603 | }, | |
1604 | ||
1605 | { | |
1606 | .name = "mems", | |
1607 | .read = cpuset_common_file_read, | |
1608 | .write_string = cpuset_write_resmask, | |
1609 | .max_write_len = (100U + 6 * MAX_NUMNODES), | |
1610 | .private = FILE_MEMLIST, | |
1611 | }, | |
1612 | ||
1613 | { | |
1614 | .name = "cpu_exclusive", | |
1615 | .read_u64 = cpuset_read_u64, | |
1616 | .write_u64 = cpuset_write_u64, | |
1617 | .private = FILE_CPU_EXCLUSIVE, | |
1618 | }, | |
1619 | ||
1620 | { | |
1621 | .name = "mem_exclusive", | |
1622 | .read_u64 = cpuset_read_u64, | |
1623 | .write_u64 = cpuset_write_u64, | |
1624 | .private = FILE_MEM_EXCLUSIVE, | |
1625 | }, | |
1626 | ||
1627 | { | |
1628 | .name = "mem_hardwall", | |
1629 | .read_u64 = cpuset_read_u64, | |
1630 | .write_u64 = cpuset_write_u64, | |
1631 | .private = FILE_MEM_HARDWALL, | |
1632 | }, | |
1633 | ||
1634 | { | |
1635 | .name = "sched_load_balance", | |
1636 | .read_u64 = cpuset_read_u64, | |
1637 | .write_u64 = cpuset_write_u64, | |
1638 | .private = FILE_SCHED_LOAD_BALANCE, | |
1639 | }, | |
1640 | ||
1641 | { | |
1642 | .name = "sched_relax_domain_level", | |
1643 | .read_s64 = cpuset_read_s64, | |
1644 | .write_s64 = cpuset_write_s64, | |
1645 | .private = FILE_SCHED_RELAX_DOMAIN_LEVEL, | |
1646 | }, | |
1647 | ||
1648 | { | |
1649 | .name = "memory_migrate", | |
1650 | .read_u64 = cpuset_read_u64, | |
1651 | .write_u64 = cpuset_write_u64, | |
1652 | .private = FILE_MEMORY_MIGRATE, | |
1653 | }, | |
1654 | ||
1655 | { | |
1656 | .name = "memory_pressure", | |
1657 | .read_u64 = cpuset_read_u64, | |
1658 | .write_u64 = cpuset_write_u64, | |
1659 | .private = FILE_MEMORY_PRESSURE, | |
1660 | }, | |
1661 | ||
1662 | { | |
1663 | .name = "memory_spread_page", | |
1664 | .read_u64 = cpuset_read_u64, | |
1665 | .write_u64 = cpuset_write_u64, | |
1666 | .private = FILE_SPREAD_PAGE, | |
1667 | }, | |
1668 | ||
1669 | { | |
1670 | .name = "memory_spread_slab", | |
1671 | .read_u64 = cpuset_read_u64, | |
1672 | .write_u64 = cpuset_write_u64, | |
1673 | .private = FILE_SPREAD_SLAB, | |
1674 | }, | |
1675 | }; | |
1676 | ||
1677 | static struct cftype cft_memory_pressure_enabled = { | |
1678 | .name = "memory_pressure_enabled", | |
1679 | .read_u64 = cpuset_read_u64, | |
1680 | .write_u64 = cpuset_write_u64, | |
1681 | .private = FILE_MEMORY_PRESSURE_ENABLED, | |
1682 | }; | |
1683 | ||
1684 | static int cpuset_populate(struct cgroup_subsys *ss, struct cgroup *cont) | |
1685 | { | |
1686 | int err; | |
1687 | ||
1688 | err = cgroup_add_files(cont, ss, files, ARRAY_SIZE(files)); | |
1689 | if (err) | |
1690 | return err; | |
1691 | /* memory_pressure_enabled is in root cpuset only */ | |
1692 | if (!cont->parent) | |
1693 | err = cgroup_add_file(cont, ss, | |
1694 | &cft_memory_pressure_enabled); | |
1695 | return err; | |
1696 | } | |
1697 | ||
1698 | /* | |
1699 | * post_clone() is called at the end of cgroup_clone(). | |
1700 | * 'cgroup' was just created automatically as a result of | |
1701 | * a cgroup_clone(), and the current task is about to | |
1702 | * be moved into 'cgroup'. | |
1703 | * | |
1704 | * Currently we refuse to set up the cgroup - thereby | |
1705 | * refusing the task to be entered, and as a result refusing | |
1706 | * the sys_unshare() or clone() which initiated it - if any | |
1707 | * sibling cpusets have exclusive cpus or mem. | |
1708 | * | |
1709 | * If this becomes a problem for some users who wish to | |
1710 | * allow that scenario, then cpuset_post_clone() could be | |
1711 | * changed to grant parent->cpus_allowed-sibling_cpus_exclusive | |
1712 | * (and likewise for mems) to the new cgroup. Called with cgroup_mutex | |
1713 | * held. | |
1714 | */ | |
1715 | static void cpuset_post_clone(struct cgroup_subsys *ss, | |
1716 | struct cgroup *cgroup) | |
1717 | { | |
1718 | struct cgroup *parent, *child; | |
1719 | struct cpuset *cs, *parent_cs; | |
1720 | ||
1721 | parent = cgroup->parent; | |
1722 | list_for_each_entry(child, &parent->children, sibling) { | |
1723 | cs = cgroup_cs(child); | |
1724 | if (is_mem_exclusive(cs) || is_cpu_exclusive(cs)) | |
1725 | return; | |
1726 | } | |
1727 | cs = cgroup_cs(cgroup); | |
1728 | parent_cs = cgroup_cs(parent); | |
1729 | ||
1730 | cs->mems_allowed = parent_cs->mems_allowed; | |
1731 | cs->cpus_allowed = parent_cs->cpus_allowed; | |
1732 | return; | |
1733 | } | |
1734 | ||
1735 | /* | |
1736 | * cpuset_create - create a cpuset | |
1737 | * ss: cpuset cgroup subsystem | |
1738 | * cont: control group that the new cpuset will be part of | |
1739 | */ | |
1740 | ||
1741 | static struct cgroup_subsys_state *cpuset_create( | |
1742 | struct cgroup_subsys *ss, | |
1743 | struct cgroup *cont) | |
1744 | { | |
1745 | struct cpuset *cs; | |
1746 | struct cpuset *parent; | |
1747 | ||
1748 | if (!cont->parent) { | |
1749 | /* This is early initialization for the top cgroup */ | |
1750 | top_cpuset.mems_generation = cpuset_mems_generation++; | |
1751 | return &top_cpuset.css; | |
1752 | } | |
1753 | parent = cgroup_cs(cont->parent); | |
1754 | cs = kmalloc(sizeof(*cs), GFP_KERNEL); | |
1755 | if (!cs) | |
1756 | return ERR_PTR(-ENOMEM); | |
1757 | ||
1758 | cpuset_update_task_memory_state(); | |
1759 | cs->flags = 0; | |
1760 | if (is_spread_page(parent)) | |
1761 | set_bit(CS_SPREAD_PAGE, &cs->flags); | |
1762 | if (is_spread_slab(parent)) | |
1763 | set_bit(CS_SPREAD_SLAB, &cs->flags); | |
1764 | set_bit(CS_SCHED_LOAD_BALANCE, &cs->flags); | |
1765 | cpus_clear(cs->cpus_allowed); | |
1766 | nodes_clear(cs->mems_allowed); | |
1767 | cs->mems_generation = cpuset_mems_generation++; | |
1768 | fmeter_init(&cs->fmeter); | |
1769 | cs->relax_domain_level = -1; | |
1770 | ||
1771 | cs->parent = parent; | |
1772 | number_of_cpusets++; | |
1773 | return &cs->css ; | |
1774 | } | |
1775 | ||
1776 | /* | |
1777 | * If the cpuset being removed has its flag 'sched_load_balance' | |
1778 | * enabled, then simulate turning sched_load_balance off, which | |
1779 | * will call async_rebuild_sched_domains(). | |
1780 | */ | |
1781 | ||
1782 | static void cpuset_destroy(struct cgroup_subsys *ss, struct cgroup *cont) | |
1783 | { | |
1784 | struct cpuset *cs = cgroup_cs(cont); | |
1785 | ||
1786 | cpuset_update_task_memory_state(); | |
1787 | ||
1788 | if (is_sched_load_balance(cs)) | |
1789 | update_flag(CS_SCHED_LOAD_BALANCE, cs, 0); | |
1790 | ||
1791 | number_of_cpusets--; | |
1792 | kfree(cs); | |
1793 | } | |
1794 | ||
1795 | struct cgroup_subsys cpuset_subsys = { | |
1796 | .name = "cpuset", | |
1797 | .create = cpuset_create, | |
1798 | .destroy = cpuset_destroy, | |
1799 | .can_attach = cpuset_can_attach, | |
1800 | .attach = cpuset_attach, | |
1801 | .populate = cpuset_populate, | |
1802 | .post_clone = cpuset_post_clone, | |
1803 | .subsys_id = cpuset_subsys_id, | |
1804 | .early_init = 1, | |
1805 | }; | |
1806 | ||
1807 | /* | |
1808 | * cpuset_init_early - just enough so that the calls to | |
1809 | * cpuset_update_task_memory_state() in early init code | |
1810 | * are harmless. | |
1811 | */ | |
1812 | ||
1813 | int __init cpuset_init_early(void) | |
1814 | { | |
1815 | top_cpuset.mems_generation = cpuset_mems_generation++; | |
1816 | return 0; | |
1817 | } | |
1818 | ||
1819 | ||
1820 | /** | |
1821 | * cpuset_init - initialize cpusets at system boot | |
1822 | * | |
1823 | * Description: Initialize top_cpuset and the cpuset internal file system, | |
1824 | **/ | |
1825 | ||
1826 | int __init cpuset_init(void) | |
1827 | { | |
1828 | int err = 0; | |
1829 | ||
1830 | cpus_setall(top_cpuset.cpus_allowed); | |
1831 | nodes_setall(top_cpuset.mems_allowed); | |
1832 | ||
1833 | fmeter_init(&top_cpuset.fmeter); | |
1834 | top_cpuset.mems_generation = cpuset_mems_generation++; | |
1835 | set_bit(CS_SCHED_LOAD_BALANCE, &top_cpuset.flags); | |
1836 | top_cpuset.relax_domain_level = -1; | |
1837 | ||
1838 | err = register_filesystem(&cpuset_fs_type); | |
1839 | if (err < 0) | |
1840 | return err; | |
1841 | ||
1842 | if (!alloc_cpumask_var(&cpus_attach, GFP_KERNEL)) | |
1843 | BUG(); | |
1844 | ||
1845 | number_of_cpusets = 1; | |
1846 | return 0; | |
1847 | } | |
1848 | ||
1849 | /** | |
1850 | * cpuset_do_move_task - move a given task to another cpuset | |
1851 | * @tsk: pointer to task_struct the task to move | |
1852 | * @scan: struct cgroup_scanner contained in its struct cpuset_hotplug_scanner | |
1853 | * | |
1854 | * Called by cgroup_scan_tasks() for each task in a cgroup. | |
1855 | * Return nonzero to stop the walk through the tasks. | |
1856 | */ | |
1857 | static void cpuset_do_move_task(struct task_struct *tsk, | |
1858 | struct cgroup_scanner *scan) | |
1859 | { | |
1860 | struct cpuset_hotplug_scanner *chsp; | |
1861 | ||
1862 | chsp = container_of(scan, struct cpuset_hotplug_scanner, scan); | |
1863 | cgroup_attach_task(chsp->to, tsk); | |
1864 | } | |
1865 | ||
1866 | /** | |
1867 | * move_member_tasks_to_cpuset - move tasks from one cpuset to another | |
1868 | * @from: cpuset in which the tasks currently reside | |
1869 | * @to: cpuset to which the tasks will be moved | |
1870 | * | |
1871 | * Called with cgroup_mutex held | |
1872 | * callback_mutex must not be held, as cpuset_attach() will take it. | |
1873 | * | |
1874 | * The cgroup_scan_tasks() function will scan all the tasks in a cgroup, | |
1875 | * calling callback functions for each. | |
1876 | */ | |
1877 | static void move_member_tasks_to_cpuset(struct cpuset *from, struct cpuset *to) | |
1878 | { | |
1879 | struct cpuset_hotplug_scanner scan; | |
1880 | ||
1881 | scan.scan.cg = from->css.cgroup; | |
1882 | scan.scan.test_task = NULL; /* select all tasks in cgroup */ | |
1883 | scan.scan.process_task = cpuset_do_move_task; | |
1884 | scan.scan.heap = NULL; | |
1885 | scan.to = to->css.cgroup; | |
1886 | ||
1887 | if (cgroup_scan_tasks(&scan.scan)) | |
1888 | printk(KERN_ERR "move_member_tasks_to_cpuset: " | |
1889 | "cgroup_scan_tasks failed\n"); | |
1890 | } | |
1891 | ||
1892 | /* | |
1893 | * If CPU and/or memory hotplug handlers, below, unplug any CPUs | |
1894 | * or memory nodes, we need to walk over the cpuset hierarchy, | |
1895 | * removing that CPU or node from all cpusets. If this removes the | |
1896 | * last CPU or node from a cpuset, then move the tasks in the empty | |
1897 | * cpuset to its next-highest non-empty parent. | |
1898 | * | |
1899 | * Called with cgroup_mutex held | |
1900 | * callback_mutex must not be held, as cpuset_attach() will take it. | |
1901 | */ | |
1902 | static void remove_tasks_in_empty_cpuset(struct cpuset *cs) | |
1903 | { | |
1904 | struct cpuset *parent; | |
1905 | ||
1906 | /* | |
1907 | * The cgroup's css_sets list is in use if there are tasks | |
1908 | * in the cpuset; the list is empty if there are none; | |
1909 | * the cs->css.refcnt seems always 0. | |
1910 | */ | |
1911 | if (list_empty(&cs->css.cgroup->css_sets)) | |
1912 | return; | |
1913 | ||
1914 | /* | |
1915 | * Find its next-highest non-empty parent, (top cpuset | |
1916 | * has online cpus, so can't be empty). | |
1917 | */ | |
1918 | parent = cs->parent; | |
1919 | while (cpus_empty(parent->cpus_allowed) || | |
1920 | nodes_empty(parent->mems_allowed)) | |
1921 | parent = parent->parent; | |
1922 | ||
1923 | move_member_tasks_to_cpuset(cs, parent); | |
1924 | } | |
1925 | ||
1926 | /* | |
1927 | * Walk the specified cpuset subtree and look for empty cpusets. | |
1928 | * The tasks of such cpuset must be moved to a parent cpuset. | |
1929 | * | |
1930 | * Called with cgroup_mutex held. We take callback_mutex to modify | |
1931 | * cpus_allowed and mems_allowed. | |
1932 | * | |
1933 | * This walk processes the tree from top to bottom, completing one layer | |
1934 | * before dropping down to the next. It always processes a node before | |
1935 | * any of its children. | |
1936 | * | |
1937 | * For now, since we lack memory hot unplug, we'll never see a cpuset | |
1938 | * that has tasks along with an empty 'mems'. But if we did see such | |
1939 | * a cpuset, we'd handle it just like we do if its 'cpus' was empty. | |
1940 | */ | |
1941 | static void scan_for_empty_cpusets(struct cpuset *root) | |
1942 | { | |
1943 | LIST_HEAD(queue); | |
1944 | struct cpuset *cp; /* scans cpusets being updated */ | |
1945 | struct cpuset *child; /* scans child cpusets of cp */ | |
1946 | struct cgroup *cont; | |
1947 | nodemask_t oldmems; | |
1948 | ||
1949 | list_add_tail((struct list_head *)&root->stack_list, &queue); | |
1950 | ||
1951 | while (!list_empty(&queue)) { | |
1952 | cp = list_first_entry(&queue, struct cpuset, stack_list); | |
1953 | list_del(queue.next); | |
1954 | list_for_each_entry(cont, &cp->css.cgroup->children, sibling) { | |
1955 | child = cgroup_cs(cont); | |
1956 | list_add_tail(&child->stack_list, &queue); | |
1957 | } | |
1958 | ||
1959 | /* Continue past cpusets with all cpus, mems online */ | |
1960 | if (cpus_subset(cp->cpus_allowed, cpu_online_map) && | |
1961 | nodes_subset(cp->mems_allowed, node_states[N_HIGH_MEMORY])) | |
1962 | continue; | |
1963 | ||
1964 | oldmems = cp->mems_allowed; | |
1965 | ||
1966 | /* Remove offline cpus and mems from this cpuset. */ | |
1967 | mutex_lock(&callback_mutex); | |
1968 | cpus_and(cp->cpus_allowed, cp->cpus_allowed, cpu_online_map); | |
1969 | nodes_and(cp->mems_allowed, cp->mems_allowed, | |
1970 | node_states[N_HIGH_MEMORY]); | |
1971 | mutex_unlock(&callback_mutex); | |
1972 | ||
1973 | /* Move tasks from the empty cpuset to a parent */ | |
1974 | if (cpus_empty(cp->cpus_allowed) || | |
1975 | nodes_empty(cp->mems_allowed)) | |
1976 | remove_tasks_in_empty_cpuset(cp); | |
1977 | else { | |
1978 | update_tasks_cpumask(cp, NULL); | |
1979 | update_tasks_nodemask(cp, &oldmems); | |
1980 | } | |
1981 | } | |
1982 | } | |
1983 | ||
1984 | /* | |
1985 | * The top_cpuset tracks what CPUs and Memory Nodes are online, | |
1986 | * period. This is necessary in order to make cpusets transparent | |
1987 | * (of no affect) on systems that are actively using CPU hotplug | |
1988 | * but making no active use of cpusets. | |
1989 | * | |
1990 | * This routine ensures that top_cpuset.cpus_allowed tracks | |
1991 | * cpu_online_map on each CPU hotplug (cpuhp) event. | |
1992 | * | |
1993 | * Called within get_online_cpus(). Needs to call cgroup_lock() | |
1994 | * before calling generate_sched_domains(). | |
1995 | */ | |
1996 | static int cpuset_track_online_cpus(struct notifier_block *unused_nb, | |
1997 | unsigned long phase, void *unused_cpu) | |
1998 | { | |
1999 | struct sched_domain_attr *attr; | |
2000 | cpumask_t *doms; | |
2001 | int ndoms; | |
2002 | ||
2003 | switch (phase) { | |
2004 | case CPU_ONLINE: | |
2005 | case CPU_ONLINE_FROZEN: | |
2006 | case CPU_DEAD: | |
2007 | case CPU_DEAD_FROZEN: | |
2008 | break; | |
2009 | ||
2010 | default: | |
2011 | return NOTIFY_DONE; | |
2012 | } | |
2013 | ||
2014 | cgroup_lock(); | |
2015 | top_cpuset.cpus_allowed = cpu_online_map; | |
2016 | scan_for_empty_cpusets(&top_cpuset); | |
2017 | ndoms = generate_sched_domains(&doms, &attr); | |
2018 | cgroup_unlock(); | |
2019 | ||
2020 | /* Have scheduler rebuild the domains */ | |
2021 | partition_sched_domains(ndoms, doms, attr); | |
2022 | ||
2023 | return NOTIFY_OK; | |
2024 | } | |
2025 | ||
2026 | #ifdef CONFIG_MEMORY_HOTPLUG | |
2027 | /* | |
2028 | * Keep top_cpuset.mems_allowed tracking node_states[N_HIGH_MEMORY]. | |
2029 | * Call this routine anytime after node_states[N_HIGH_MEMORY] changes. | |
2030 | * See also the previous routine cpuset_track_online_cpus(). | |
2031 | */ | |
2032 | static int cpuset_track_online_nodes(struct notifier_block *self, | |
2033 | unsigned long action, void *arg) | |
2034 | { | |
2035 | cgroup_lock(); | |
2036 | switch (action) { | |
2037 | case MEM_ONLINE: | |
2038 | top_cpuset.mems_allowed = node_states[N_HIGH_MEMORY]; | |
2039 | break; | |
2040 | case MEM_OFFLINE: | |
2041 | top_cpuset.mems_allowed = node_states[N_HIGH_MEMORY]; | |
2042 | scan_for_empty_cpusets(&top_cpuset); | |
2043 | break; | |
2044 | default: | |
2045 | break; | |
2046 | } | |
2047 | cgroup_unlock(); | |
2048 | return NOTIFY_OK; | |
2049 | } | |
2050 | #endif | |
2051 | ||
2052 | /** | |
2053 | * cpuset_init_smp - initialize cpus_allowed | |
2054 | * | |
2055 | * Description: Finish top cpuset after cpu, node maps are initialized | |
2056 | **/ | |
2057 | ||
2058 | void __init cpuset_init_smp(void) | |
2059 | { | |
2060 | top_cpuset.cpus_allowed = cpu_online_map; | |
2061 | top_cpuset.mems_allowed = node_states[N_HIGH_MEMORY]; | |
2062 | ||
2063 | hotcpu_notifier(cpuset_track_online_cpus, 0); | |
2064 | hotplug_memory_notifier(cpuset_track_online_nodes, 10); | |
2065 | } | |
2066 | ||
2067 | /** | |
2068 | * cpuset_cpus_allowed - return cpus_allowed mask from a tasks cpuset. | |
2069 | * @tsk: pointer to task_struct from which to obtain cpuset->cpus_allowed. | |
2070 | * @pmask: pointer to cpumask_t variable to receive cpus_allowed set. | |
2071 | * | |
2072 | * Description: Returns the cpumask_t cpus_allowed of the cpuset | |
2073 | * attached to the specified @tsk. Guaranteed to return some non-empty | |
2074 | * subset of cpu_online_map, even if this means going outside the | |
2075 | * tasks cpuset. | |
2076 | **/ | |
2077 | ||
2078 | void cpuset_cpus_allowed(struct task_struct *tsk, cpumask_t *pmask) | |
2079 | { | |
2080 | mutex_lock(&callback_mutex); | |
2081 | cpuset_cpus_allowed_locked(tsk, pmask); | |
2082 | mutex_unlock(&callback_mutex); | |
2083 | } | |
2084 | ||
2085 | /** | |
2086 | * cpuset_cpus_allowed_locked - return cpus_allowed mask from a tasks cpuset. | |
2087 | * Must be called with callback_mutex held. | |
2088 | **/ | |
2089 | void cpuset_cpus_allowed_locked(struct task_struct *tsk, cpumask_t *pmask) | |
2090 | { | |
2091 | task_lock(tsk); | |
2092 | guarantee_online_cpus(task_cs(tsk), pmask); | |
2093 | task_unlock(tsk); | |
2094 | } | |
2095 | ||
2096 | void cpuset_init_current_mems_allowed(void) | |
2097 | { | |
2098 | nodes_setall(current->mems_allowed); | |
2099 | } | |
2100 | ||
2101 | /** | |
2102 | * cpuset_mems_allowed - return mems_allowed mask from a tasks cpuset. | |
2103 | * @tsk: pointer to task_struct from which to obtain cpuset->mems_allowed. | |
2104 | * | |
2105 | * Description: Returns the nodemask_t mems_allowed of the cpuset | |
2106 | * attached to the specified @tsk. Guaranteed to return some non-empty | |
2107 | * subset of node_states[N_HIGH_MEMORY], even if this means going outside the | |
2108 | * tasks cpuset. | |
2109 | **/ | |
2110 | ||
2111 | nodemask_t cpuset_mems_allowed(struct task_struct *tsk) | |
2112 | { | |
2113 | nodemask_t mask; | |
2114 | ||
2115 | mutex_lock(&callback_mutex); | |
2116 | task_lock(tsk); | |
2117 | guarantee_online_mems(task_cs(tsk), &mask); | |
2118 | task_unlock(tsk); | |
2119 | mutex_unlock(&callback_mutex); | |
2120 | ||
2121 | return mask; | |
2122 | } | |
2123 | ||
2124 | /** | |
2125 | * cpuset_nodemask_valid_mems_allowed - check nodemask vs. curremt mems_allowed | |
2126 | * @nodemask: the nodemask to be checked | |
2127 | * | |
2128 | * Are any of the nodes in the nodemask allowed in current->mems_allowed? | |
2129 | */ | |
2130 | int cpuset_nodemask_valid_mems_allowed(nodemask_t *nodemask) | |
2131 | { | |
2132 | return nodes_intersects(*nodemask, current->mems_allowed); | |
2133 | } | |
2134 | ||
2135 | /* | |
2136 | * nearest_hardwall_ancestor() - Returns the nearest mem_exclusive or | |
2137 | * mem_hardwall ancestor to the specified cpuset. Call holding | |
2138 | * callback_mutex. If no ancestor is mem_exclusive or mem_hardwall | |
2139 | * (an unusual configuration), then returns the root cpuset. | |
2140 | */ | |
2141 | static const struct cpuset *nearest_hardwall_ancestor(const struct cpuset *cs) | |
2142 | { | |
2143 | while (!(is_mem_exclusive(cs) || is_mem_hardwall(cs)) && cs->parent) | |
2144 | cs = cs->parent; | |
2145 | return cs; | |
2146 | } | |
2147 | ||
2148 | /** | |
2149 | * cpuset_zone_allowed_softwall - Can we allocate on zone z's memory node? | |
2150 | * @z: is this zone on an allowed node? | |
2151 | * @gfp_mask: memory allocation flags | |
2152 | * | |
2153 | * If we're in interrupt, yes, we can always allocate. If | |
2154 | * __GFP_THISNODE is set, yes, we can always allocate. If zone | |
2155 | * z's node is in our tasks mems_allowed, yes. If it's not a | |
2156 | * __GFP_HARDWALL request and this zone's nodes is in the nearest | |
2157 | * hardwalled cpuset ancestor to this tasks cpuset, yes. | |
2158 | * If the task has been OOM killed and has access to memory reserves | |
2159 | * as specified by the TIF_MEMDIE flag, yes. | |
2160 | * Otherwise, no. | |
2161 | * | |
2162 | * If __GFP_HARDWALL is set, cpuset_zone_allowed_softwall() | |
2163 | * reduces to cpuset_zone_allowed_hardwall(). Otherwise, | |
2164 | * cpuset_zone_allowed_softwall() might sleep, and might allow a zone | |
2165 | * from an enclosing cpuset. | |
2166 | * | |
2167 | * cpuset_zone_allowed_hardwall() only handles the simpler case of | |
2168 | * hardwall cpusets, and never sleeps. | |
2169 | * | |
2170 | * The __GFP_THISNODE placement logic is really handled elsewhere, | |
2171 | * by forcibly using a zonelist starting at a specified node, and by | |
2172 | * (in get_page_from_freelist()) refusing to consider the zones for | |
2173 | * any node on the zonelist except the first. By the time any such | |
2174 | * calls get to this routine, we should just shut up and say 'yes'. | |
2175 | * | |
2176 | * GFP_USER allocations are marked with the __GFP_HARDWALL bit, | |
2177 | * and do not allow allocations outside the current tasks cpuset | |
2178 | * unless the task has been OOM killed as is marked TIF_MEMDIE. | |
2179 | * GFP_KERNEL allocations are not so marked, so can escape to the | |
2180 | * nearest enclosing hardwalled ancestor cpuset. | |
2181 | * | |
2182 | * Scanning up parent cpusets requires callback_mutex. The | |
2183 | * __alloc_pages() routine only calls here with __GFP_HARDWALL bit | |
2184 | * _not_ set if it's a GFP_KERNEL allocation, and all nodes in the | |
2185 | * current tasks mems_allowed came up empty on the first pass over | |
2186 | * the zonelist. So only GFP_KERNEL allocations, if all nodes in the | |
2187 | * cpuset are short of memory, might require taking the callback_mutex | |
2188 | * mutex. | |
2189 | * | |
2190 | * The first call here from mm/page_alloc:get_page_from_freelist() | |
2191 | * has __GFP_HARDWALL set in gfp_mask, enforcing hardwall cpusets, | |
2192 | * so no allocation on a node outside the cpuset is allowed (unless | |
2193 | * in interrupt, of course). | |
2194 | * | |
2195 | * The second pass through get_page_from_freelist() doesn't even call | |
2196 | * here for GFP_ATOMIC calls. For those calls, the __alloc_pages() | |
2197 | * variable 'wait' is not set, and the bit ALLOC_CPUSET is not set | |
2198 | * in alloc_flags. That logic and the checks below have the combined | |
2199 | * affect that: | |
2200 | * in_interrupt - any node ok (current task context irrelevant) | |
2201 | * GFP_ATOMIC - any node ok | |
2202 | * TIF_MEMDIE - any node ok | |
2203 | * GFP_KERNEL - any node in enclosing hardwalled cpuset ok | |
2204 | * GFP_USER - only nodes in current tasks mems allowed ok. | |
2205 | * | |
2206 | * Rule: | |
2207 | * Don't call cpuset_zone_allowed_softwall if you can't sleep, unless you | |
2208 | * pass in the __GFP_HARDWALL flag set in gfp_flag, which disables | |
2209 | * the code that might scan up ancestor cpusets and sleep. | |
2210 | */ | |
2211 | ||
2212 | int __cpuset_zone_allowed_softwall(struct zone *z, gfp_t gfp_mask) | |
2213 | { | |
2214 | int node; /* node that zone z is on */ | |
2215 | const struct cpuset *cs; /* current cpuset ancestors */ | |
2216 | int allowed; /* is allocation in zone z allowed? */ | |
2217 | ||
2218 | if (in_interrupt() || (gfp_mask & __GFP_THISNODE)) | |
2219 | return 1; | |
2220 | node = zone_to_nid(z); | |
2221 | might_sleep_if(!(gfp_mask & __GFP_HARDWALL)); | |
2222 | if (node_isset(node, current->mems_allowed)) | |
2223 | return 1; | |
2224 | /* | |
2225 | * Allow tasks that have access to memory reserves because they have | |
2226 | * been OOM killed to get memory anywhere. | |
2227 | */ | |
2228 | if (unlikely(test_thread_flag(TIF_MEMDIE))) | |
2229 | return 1; | |
2230 | if (gfp_mask & __GFP_HARDWALL) /* If hardwall request, stop here */ | |
2231 | return 0; | |
2232 | ||
2233 | if (current->flags & PF_EXITING) /* Let dying task have memory */ | |
2234 | return 1; | |
2235 | ||
2236 | /* Not hardwall and node outside mems_allowed: scan up cpusets */ | |
2237 | mutex_lock(&callback_mutex); | |
2238 | ||
2239 | task_lock(current); | |
2240 | cs = nearest_hardwall_ancestor(task_cs(current)); | |
2241 | task_unlock(current); | |
2242 | ||
2243 | allowed = node_isset(node, cs->mems_allowed); | |
2244 | mutex_unlock(&callback_mutex); | |
2245 | return allowed; | |
2246 | } | |
2247 | ||
2248 | /* | |
2249 | * cpuset_zone_allowed_hardwall - Can we allocate on zone z's memory node? | |
2250 | * @z: is this zone on an allowed node? | |
2251 | * @gfp_mask: memory allocation flags | |
2252 | * | |
2253 | * If we're in interrupt, yes, we can always allocate. | |
2254 | * If __GFP_THISNODE is set, yes, we can always allocate. If zone | |
2255 | * z's node is in our tasks mems_allowed, yes. If the task has been | |
2256 | * OOM killed and has access to memory reserves as specified by the | |
2257 | * TIF_MEMDIE flag, yes. Otherwise, no. | |
2258 | * | |
2259 | * The __GFP_THISNODE placement logic is really handled elsewhere, | |
2260 | * by forcibly using a zonelist starting at a specified node, and by | |
2261 | * (in get_page_from_freelist()) refusing to consider the zones for | |
2262 | * any node on the zonelist except the first. By the time any such | |
2263 | * calls get to this routine, we should just shut up and say 'yes'. | |
2264 | * | |
2265 | * Unlike the cpuset_zone_allowed_softwall() variant, above, | |
2266 | * this variant requires that the zone be in the current tasks | |
2267 | * mems_allowed or that we're in interrupt. It does not scan up the | |
2268 | * cpuset hierarchy for the nearest enclosing mem_exclusive cpuset. | |
2269 | * It never sleeps. | |
2270 | */ | |
2271 | ||
2272 | int __cpuset_zone_allowed_hardwall(struct zone *z, gfp_t gfp_mask) | |
2273 | { | |
2274 | int node; /* node that zone z is on */ | |
2275 | ||
2276 | if (in_interrupt() || (gfp_mask & __GFP_THISNODE)) | |
2277 | return 1; | |
2278 | node = zone_to_nid(z); | |
2279 | if (node_isset(node, current->mems_allowed)) | |
2280 | return 1; | |
2281 | /* | |
2282 | * Allow tasks that have access to memory reserves because they have | |
2283 | * been OOM killed to get memory anywhere. | |
2284 | */ | |
2285 | if (unlikely(test_thread_flag(TIF_MEMDIE))) | |
2286 | return 1; | |
2287 | return 0; | |
2288 | } | |
2289 | ||
2290 | /** | |
2291 | * cpuset_lock - lock out any changes to cpuset structures | |
2292 | * | |
2293 | * The out of memory (oom) code needs to mutex_lock cpusets | |
2294 | * from being changed while it scans the tasklist looking for a | |
2295 | * task in an overlapping cpuset. Expose callback_mutex via this | |
2296 | * cpuset_lock() routine, so the oom code can lock it, before | |
2297 | * locking the task list. The tasklist_lock is a spinlock, so | |
2298 | * must be taken inside callback_mutex. | |
2299 | */ | |
2300 | ||
2301 | void cpuset_lock(void) | |
2302 | { | |
2303 | mutex_lock(&callback_mutex); | |
2304 | } | |
2305 | ||
2306 | /** | |
2307 | * cpuset_unlock - release lock on cpuset changes | |
2308 | * | |
2309 | * Undo the lock taken in a previous cpuset_lock() call. | |
2310 | */ | |
2311 | ||
2312 | void cpuset_unlock(void) | |
2313 | { | |
2314 | mutex_unlock(&callback_mutex); | |
2315 | } | |
2316 | ||
2317 | /** | |
2318 | * cpuset_mem_spread_node() - On which node to begin search for a page | |
2319 | * | |
2320 | * If a task is marked PF_SPREAD_PAGE or PF_SPREAD_SLAB (as for | |
2321 | * tasks in a cpuset with is_spread_page or is_spread_slab set), | |
2322 | * and if the memory allocation used cpuset_mem_spread_node() | |
2323 | * to determine on which node to start looking, as it will for | |
2324 | * certain page cache or slab cache pages such as used for file | |
2325 | * system buffers and inode caches, then instead of starting on the | |
2326 | * local node to look for a free page, rather spread the starting | |
2327 | * node around the tasks mems_allowed nodes. | |
2328 | * | |
2329 | * We don't have to worry about the returned node being offline | |
2330 | * because "it can't happen", and even if it did, it would be ok. | |
2331 | * | |
2332 | * The routines calling guarantee_online_mems() are careful to | |
2333 | * only set nodes in task->mems_allowed that are online. So it | |
2334 | * should not be possible for the following code to return an | |
2335 | * offline node. But if it did, that would be ok, as this routine | |
2336 | * is not returning the node where the allocation must be, only | |
2337 | * the node where the search should start. The zonelist passed to | |
2338 | * __alloc_pages() will include all nodes. If the slab allocator | |
2339 | * is passed an offline node, it will fall back to the local node. | |
2340 | * See kmem_cache_alloc_node(). | |
2341 | */ | |
2342 | ||
2343 | int cpuset_mem_spread_node(void) | |
2344 | { | |
2345 | int node; | |
2346 | ||
2347 | node = next_node(current->cpuset_mem_spread_rotor, current->mems_allowed); | |
2348 | if (node == MAX_NUMNODES) | |
2349 | node = first_node(current->mems_allowed); | |
2350 | current->cpuset_mem_spread_rotor = node; | |
2351 | return node; | |
2352 | } | |
2353 | EXPORT_SYMBOL_GPL(cpuset_mem_spread_node); | |
2354 | ||
2355 | /** | |
2356 | * cpuset_mems_allowed_intersects - Does @tsk1's mems_allowed intersect @tsk2's? | |
2357 | * @tsk1: pointer to task_struct of some task. | |
2358 | * @tsk2: pointer to task_struct of some other task. | |
2359 | * | |
2360 | * Description: Return true if @tsk1's mems_allowed intersects the | |
2361 | * mems_allowed of @tsk2. Used by the OOM killer to determine if | |
2362 | * one of the task's memory usage might impact the memory available | |
2363 | * to the other. | |
2364 | **/ | |
2365 | ||
2366 | int cpuset_mems_allowed_intersects(const struct task_struct *tsk1, | |
2367 | const struct task_struct *tsk2) | |
2368 | { | |
2369 | return nodes_intersects(tsk1->mems_allowed, tsk2->mems_allowed); | |
2370 | } | |
2371 | ||
2372 | /** | |
2373 | * cpuset_print_task_mems_allowed - prints task's cpuset and mems_allowed | |
2374 | * @task: pointer to task_struct of some task. | |
2375 | * | |
2376 | * Description: Prints @task's name, cpuset name, and cached copy of its | |
2377 | * mems_allowed to the kernel log. Must hold task_lock(task) to allow | |
2378 | * dereferencing task_cs(task). | |
2379 | */ | |
2380 | void cpuset_print_task_mems_allowed(struct task_struct *tsk) | |
2381 | { | |
2382 | struct dentry *dentry; | |
2383 | ||
2384 | dentry = task_cs(tsk)->css.cgroup->dentry; | |
2385 | spin_lock(&cpuset_buffer_lock); | |
2386 | snprintf(cpuset_name, CPUSET_NAME_LEN, | |
2387 | dentry ? (const char *)dentry->d_name.name : "/"); | |
2388 | nodelist_scnprintf(cpuset_nodelist, CPUSET_NODELIST_LEN, | |
2389 | tsk->mems_allowed); | |
2390 | printk(KERN_INFO "%s cpuset=%s mems_allowed=%s\n", | |
2391 | tsk->comm, cpuset_name, cpuset_nodelist); | |
2392 | spin_unlock(&cpuset_buffer_lock); | |
2393 | } | |
2394 | ||
2395 | /* | |
2396 | * Collection of memory_pressure is suppressed unless | |
2397 | * this flag is enabled by writing "1" to the special | |
2398 | * cpuset file 'memory_pressure_enabled' in the root cpuset. | |
2399 | */ | |
2400 | ||
2401 | int cpuset_memory_pressure_enabled __read_mostly; | |
2402 | ||
2403 | /** | |
2404 | * cpuset_memory_pressure_bump - keep stats of per-cpuset reclaims. | |
2405 | * | |
2406 | * Keep a running average of the rate of synchronous (direct) | |
2407 | * page reclaim efforts initiated by tasks in each cpuset. | |
2408 | * | |
2409 | * This represents the rate at which some task in the cpuset | |
2410 | * ran low on memory on all nodes it was allowed to use, and | |
2411 | * had to enter the kernels page reclaim code in an effort to | |
2412 | * create more free memory by tossing clean pages or swapping | |
2413 | * or writing dirty pages. | |
2414 | * | |
2415 | * Display to user space in the per-cpuset read-only file | |
2416 | * "memory_pressure". Value displayed is an integer | |
2417 | * representing the recent rate of entry into the synchronous | |
2418 | * (direct) page reclaim by any task attached to the cpuset. | |
2419 | **/ | |
2420 | ||
2421 | void __cpuset_memory_pressure_bump(void) | |
2422 | { | |
2423 | task_lock(current); | |
2424 | fmeter_markevent(&task_cs(current)->fmeter); | |
2425 | task_unlock(current); | |
2426 | } | |
2427 | ||
2428 | #ifdef CONFIG_PROC_PID_CPUSET | |
2429 | /* | |
2430 | * proc_cpuset_show() | |
2431 | * - Print tasks cpuset path into seq_file. | |
2432 | * - Used for /proc/<pid>/cpuset. | |
2433 | * - No need to task_lock(tsk) on this tsk->cpuset reference, as it | |
2434 | * doesn't really matter if tsk->cpuset changes after we read it, | |
2435 | * and we take cgroup_mutex, keeping cpuset_attach() from changing it | |
2436 | * anyway. | |
2437 | */ | |
2438 | static int proc_cpuset_show(struct seq_file *m, void *unused_v) | |
2439 | { | |
2440 | struct pid *pid; | |
2441 | struct task_struct *tsk; | |
2442 | char *buf; | |
2443 | struct cgroup_subsys_state *css; | |
2444 | int retval; | |
2445 | ||
2446 | retval = -ENOMEM; | |
2447 | buf = kmalloc(PAGE_SIZE, GFP_KERNEL); | |
2448 | if (!buf) | |
2449 | goto out; | |
2450 | ||
2451 | retval = -ESRCH; | |
2452 | pid = m->private; | |
2453 | tsk = get_pid_task(pid, PIDTYPE_PID); | |
2454 | if (!tsk) | |
2455 | goto out_free; | |
2456 | ||
2457 | retval = -EINVAL; | |
2458 | cgroup_lock(); | |
2459 | css = task_subsys_state(tsk, cpuset_subsys_id); | |
2460 | retval = cgroup_path(css->cgroup, buf, PAGE_SIZE); | |
2461 | if (retval < 0) | |
2462 | goto out_unlock; | |
2463 | seq_puts(m, buf); | |
2464 | seq_putc(m, '\n'); | |
2465 | out_unlock: | |
2466 | cgroup_unlock(); | |
2467 | put_task_struct(tsk); | |
2468 | out_free: | |
2469 | kfree(buf); | |
2470 | out: | |
2471 | return retval; | |
2472 | } | |
2473 | ||
2474 | static int cpuset_open(struct inode *inode, struct file *file) | |
2475 | { | |
2476 | struct pid *pid = PROC_I(inode)->pid; | |
2477 | return single_open(file, proc_cpuset_show, pid); | |
2478 | } | |
2479 | ||
2480 | const struct file_operations proc_cpuset_operations = { | |
2481 | .open = cpuset_open, | |
2482 | .read = seq_read, | |
2483 | .llseek = seq_lseek, | |
2484 | .release = single_release, | |
2485 | }; | |
2486 | #endif /* CONFIG_PROC_PID_CPUSET */ | |
2487 | ||
2488 | /* Display task cpus_allowed, mems_allowed in /proc/<pid>/status file. */ | |
2489 | void cpuset_task_status_allowed(struct seq_file *m, struct task_struct *task) | |
2490 | { | |
2491 | seq_printf(m, "Cpus_allowed:\t"); | |
2492 | seq_cpumask(m, &task->cpus_allowed); | |
2493 | seq_printf(m, "\n"); | |
2494 | seq_printf(m, "Cpus_allowed_list:\t"); | |
2495 | seq_cpumask_list(m, &task->cpus_allowed); | |
2496 | seq_printf(m, "\n"); | |
2497 | seq_printf(m, "Mems_allowed:\t"); | |
2498 | seq_nodemask(m, &task->mems_allowed); | |
2499 | seq_printf(m, "\n"); | |
2500 | seq_printf(m, "Mems_allowed_list:\t"); | |
2501 | seq_nodemask_list(m, &task->mems_allowed); | |
2502 | seq_printf(m, "\n"); | |
2503 | } |