Commit | Line | Data |
---|---|---|
1da177e4 LT |
1 | /* |
2 | * kernel/cpuset.c | |
3 | * | |
4 | * Processor and Memory placement constraints for sets of tasks. | |
5 | * | |
6 | * Copyright (C) 2003 BULL SA. | |
029190c5 | 7 | * Copyright (C) 2004-2007 Silicon Graphics, Inc. |
8793d854 | 8 | * Copyright (C) 2006 Google, Inc |
1da177e4 LT |
9 | * |
10 | * Portions derived from Patrick Mochel's sysfs code. | |
11 | * sysfs is Copyright (c) 2001-3 Patrick Mochel | |
1da177e4 | 12 | * |
825a46af | 13 | * 2003-10-10 Written by Simon Derr. |
1da177e4 | 14 | * 2003-10-22 Updates by Stephen Hemminger. |
825a46af | 15 | * 2004 May-July Rework by Paul Jackson. |
8793d854 | 16 | * 2006 Rework by Paul Menage to use generic cgroups |
cf417141 MK |
17 | * 2008 Rework of the scheduler domains and CPU hotplug handling |
18 | * by Max Krasnyansky | |
1da177e4 LT |
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 | ||
1da177e4 LT |
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> | |
68860ec1 | 37 | #include <linux/mempolicy.h> |
1da177e4 | 38 | #include <linux/mm.h> |
f481891f | 39 | #include <linux/memory.h> |
9984de1a | 40 | #include <linux/export.h> |
1da177e4 LT |
41 | #include <linux/mount.h> |
42 | #include <linux/namei.h> | |
43 | #include <linux/pagemap.h> | |
44 | #include <linux/proc_fs.h> | |
6b9c2603 | 45 | #include <linux/rcupdate.h> |
1da177e4 LT |
46 | #include <linux/sched.h> |
47 | #include <linux/seq_file.h> | |
22fb52dd | 48 | #include <linux/security.h> |
1da177e4 | 49 | #include <linux/slab.h> |
1da177e4 LT |
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> | |
60063497 | 58 | #include <linux/atomic.h> |
3d3f26a7 | 59 | #include <linux/mutex.h> |
956db3ca CW |
60 | #include <linux/workqueue.h> |
61 | #include <linux/cgroup.h> | |
1da177e4 | 62 | |
f90d4118 MX |
63 | /* |
64 | * Workqueue for cpuset related tasks. | |
65 | * | |
66 | * Using kevent workqueue may cause deadlock when memory_migrate | |
67 | * is set. So we create a separate workqueue thread for cpuset. | |
68 | */ | |
69 | static struct workqueue_struct *cpuset_wq; | |
70 | ||
202f72d5 PJ |
71 | /* |
72 | * Tracks how many cpusets are currently defined in system. | |
73 | * When there is only one cpuset (the root cpuset) we can | |
74 | * short circuit some hooks. | |
75 | */ | |
7edc5962 | 76 | int number_of_cpusets __read_mostly; |
202f72d5 | 77 | |
2df167a3 | 78 | /* Forward declare cgroup structures */ |
8793d854 PM |
79 | struct cgroup_subsys cpuset_subsys; |
80 | struct cpuset; | |
81 | ||
3e0d98b9 PJ |
82 | /* See "Frequency meter" comments, below. */ |
83 | ||
84 | struct fmeter { | |
85 | int cnt; /* unprocessed events count */ | |
86 | int val; /* most recent output value */ | |
87 | time_t time; /* clock (secs) when val computed */ | |
88 | spinlock_t lock; /* guards read or write of above */ | |
89 | }; | |
90 | ||
1da177e4 | 91 | struct cpuset { |
8793d854 PM |
92 | struct cgroup_subsys_state css; |
93 | ||
1da177e4 | 94 | unsigned long flags; /* "unsigned long" so bitops work */ |
300ed6cb | 95 | cpumask_var_t cpus_allowed; /* CPUs allowed to tasks in cpuset */ |
1da177e4 LT |
96 | nodemask_t mems_allowed; /* Memory Nodes allowed to tasks */ |
97 | ||
1da177e4 | 98 | struct cpuset *parent; /* my parent */ |
1da177e4 | 99 | |
3e0d98b9 | 100 | struct fmeter fmeter; /* memory_pressure filter */ |
029190c5 PJ |
101 | |
102 | /* partition number for rebuild_sched_domains() */ | |
103 | int pn; | |
956db3ca | 104 | |
1d3504fc HS |
105 | /* for custom sched domain */ |
106 | int relax_domain_level; | |
107 | ||
732bee7a | 108 | /* used for walking a cpuset hierarchy */ |
956db3ca | 109 | struct list_head stack_list; |
1da177e4 LT |
110 | }; |
111 | ||
8793d854 PM |
112 | /* Retrieve the cpuset for a cgroup */ |
113 | static inline struct cpuset *cgroup_cs(struct cgroup *cont) | |
114 | { | |
115 | return container_of(cgroup_subsys_state(cont, cpuset_subsys_id), | |
116 | struct cpuset, css); | |
117 | } | |
118 | ||
119 | /* Retrieve the cpuset for a task */ | |
120 | static inline struct cpuset *task_cs(struct task_struct *task) | |
121 | { | |
122 | return container_of(task_subsys_state(task, cpuset_subsys_id), | |
123 | struct cpuset, css); | |
124 | } | |
8793d854 | 125 | |
b246272e DR |
126 | #ifdef CONFIG_NUMA |
127 | static inline bool task_has_mempolicy(struct task_struct *task) | |
128 | { | |
129 | return task->mempolicy; | |
130 | } | |
131 | #else | |
132 | static inline bool task_has_mempolicy(struct task_struct *task) | |
133 | { | |
134 | return false; | |
135 | } | |
136 | #endif | |
137 | ||
138 | ||
1da177e4 LT |
139 | /* bits in struct cpuset flags field */ |
140 | typedef enum { | |
efeb77b2 | 141 | CS_ONLINE, |
1da177e4 LT |
142 | CS_CPU_EXCLUSIVE, |
143 | CS_MEM_EXCLUSIVE, | |
78608366 | 144 | CS_MEM_HARDWALL, |
45b07ef3 | 145 | CS_MEMORY_MIGRATE, |
029190c5 | 146 | CS_SCHED_LOAD_BALANCE, |
825a46af PJ |
147 | CS_SPREAD_PAGE, |
148 | CS_SPREAD_SLAB, | |
1da177e4 LT |
149 | } cpuset_flagbits_t; |
150 | ||
151 | /* convenient tests for these bits */ | |
efeb77b2 TH |
152 | static inline bool is_cpuset_online(const struct cpuset *cs) |
153 | { | |
154 | return test_bit(CS_ONLINE, &cs->flags); | |
155 | } | |
156 | ||
1da177e4 LT |
157 | static inline int is_cpu_exclusive(const struct cpuset *cs) |
158 | { | |
7b5b9ef0 | 159 | return test_bit(CS_CPU_EXCLUSIVE, &cs->flags); |
1da177e4 LT |
160 | } |
161 | ||
162 | static inline int is_mem_exclusive(const struct cpuset *cs) | |
163 | { | |
7b5b9ef0 | 164 | return test_bit(CS_MEM_EXCLUSIVE, &cs->flags); |
1da177e4 LT |
165 | } |
166 | ||
78608366 PM |
167 | static inline int is_mem_hardwall(const struct cpuset *cs) |
168 | { | |
169 | return test_bit(CS_MEM_HARDWALL, &cs->flags); | |
170 | } | |
171 | ||
029190c5 PJ |
172 | static inline int is_sched_load_balance(const struct cpuset *cs) |
173 | { | |
174 | return test_bit(CS_SCHED_LOAD_BALANCE, &cs->flags); | |
175 | } | |
176 | ||
45b07ef3 PJ |
177 | static inline int is_memory_migrate(const struct cpuset *cs) |
178 | { | |
7b5b9ef0 | 179 | return test_bit(CS_MEMORY_MIGRATE, &cs->flags); |
45b07ef3 PJ |
180 | } |
181 | ||
825a46af PJ |
182 | static inline int is_spread_page(const struct cpuset *cs) |
183 | { | |
184 | return test_bit(CS_SPREAD_PAGE, &cs->flags); | |
185 | } | |
186 | ||
187 | static inline int is_spread_slab(const struct cpuset *cs) | |
188 | { | |
189 | return test_bit(CS_SPREAD_SLAB, &cs->flags); | |
190 | } | |
191 | ||
1da177e4 | 192 | static struct cpuset top_cpuset = { |
efeb77b2 TH |
193 | .flags = ((1 << CS_ONLINE) | (1 << CS_CPU_EXCLUSIVE) | |
194 | (1 << CS_MEM_EXCLUSIVE)), | |
1da177e4 LT |
195 | }; |
196 | ||
ae8086ce TH |
197 | /** |
198 | * cpuset_for_each_child - traverse online children of a cpuset | |
199 | * @child_cs: loop cursor pointing to the current child | |
200 | * @pos_cgrp: used for iteration | |
201 | * @parent_cs: target cpuset to walk children of | |
202 | * | |
203 | * Walk @child_cs through the online children of @parent_cs. Must be used | |
204 | * with RCU read locked. | |
205 | */ | |
206 | #define cpuset_for_each_child(child_cs, pos_cgrp, parent_cs) \ | |
207 | cgroup_for_each_child((pos_cgrp), (parent_cs)->css.cgroup) \ | |
208 | if (is_cpuset_online(((child_cs) = cgroup_cs((pos_cgrp))))) | |
209 | ||
1da177e4 | 210 | /* |
2df167a3 PM |
211 | * There are two global mutexes guarding cpuset structures. The first |
212 | * is the main control groups cgroup_mutex, accessed via | |
213 | * cgroup_lock()/cgroup_unlock(). The second is the cpuset-specific | |
214 | * callback_mutex, below. They can nest. It is ok to first take | |
215 | * cgroup_mutex, then nest callback_mutex. We also require taking | |
216 | * task_lock() when dereferencing a task's cpuset pointer. See "The | |
217 | * task_lock() exception", at the end of this comment. | |
053199ed | 218 | * |
3d3f26a7 | 219 | * A task must hold both mutexes to modify cpusets. If a task |
2df167a3 | 220 | * holds cgroup_mutex, then it blocks others wanting that mutex, |
3d3f26a7 | 221 | * ensuring that it is the only task able to also acquire callback_mutex |
053199ed PJ |
222 | * and be able to modify cpusets. It can perform various checks on |
223 | * the cpuset structure first, knowing nothing will change. It can | |
2df167a3 | 224 | * also allocate memory while just holding cgroup_mutex. While it is |
053199ed | 225 | * performing these checks, various callback routines can briefly |
3d3f26a7 IM |
226 | * acquire callback_mutex to query cpusets. Once it is ready to make |
227 | * the changes, it takes callback_mutex, blocking everyone else. | |
053199ed PJ |
228 | * |
229 | * Calls to the kernel memory allocator can not be made while holding | |
3d3f26a7 | 230 | * callback_mutex, as that would risk double tripping on callback_mutex |
053199ed PJ |
231 | * from one of the callbacks into the cpuset code from within |
232 | * __alloc_pages(). | |
233 | * | |
3d3f26a7 | 234 | * If a task is only holding callback_mutex, then it has read-only |
053199ed PJ |
235 | * access to cpusets. |
236 | * | |
58568d2a MX |
237 | * Now, the task_struct fields mems_allowed and mempolicy may be changed |
238 | * by other task, we use alloc_lock in the task_struct fields to protect | |
239 | * them. | |
053199ed | 240 | * |
3d3f26a7 | 241 | * The cpuset_common_file_read() handlers only hold callback_mutex across |
053199ed PJ |
242 | * small pieces of code, such as when reading out possibly multi-word |
243 | * cpumasks and nodemasks. | |
244 | * | |
2df167a3 PM |
245 | * Accessing a task's cpuset should be done in accordance with the |
246 | * guidelines for accessing subsystem state in kernel/cgroup.c | |
1da177e4 LT |
247 | */ |
248 | ||
3d3f26a7 | 249 | static DEFINE_MUTEX(callback_mutex); |
4247bdc6 | 250 | |
75aa1994 DR |
251 | /* |
252 | * cpuset_buffer_lock protects both the cpuset_name and cpuset_nodelist | |
253 | * buffers. They are statically allocated to prevent using excess stack | |
254 | * when calling cpuset_print_task_mems_allowed(). | |
255 | */ | |
256 | #define CPUSET_NAME_LEN (128) | |
257 | #define CPUSET_NODELIST_LEN (256) | |
258 | static char cpuset_name[CPUSET_NAME_LEN]; | |
259 | static char cpuset_nodelist[CPUSET_NODELIST_LEN]; | |
260 | static DEFINE_SPINLOCK(cpuset_buffer_lock); | |
261 | ||
cf417141 MK |
262 | /* |
263 | * This is ugly, but preserves the userspace API for existing cpuset | |
8793d854 | 264 | * users. If someone tries to mount the "cpuset" filesystem, we |
cf417141 MK |
265 | * silently switch it to mount "cgroup" instead |
266 | */ | |
f7e83571 AV |
267 | static struct dentry *cpuset_mount(struct file_system_type *fs_type, |
268 | int flags, const char *unused_dev_name, void *data) | |
1da177e4 | 269 | { |
8793d854 | 270 | struct file_system_type *cgroup_fs = get_fs_type("cgroup"); |
f7e83571 | 271 | struct dentry *ret = ERR_PTR(-ENODEV); |
8793d854 PM |
272 | if (cgroup_fs) { |
273 | char mountopts[] = | |
274 | "cpuset,noprefix," | |
275 | "release_agent=/sbin/cpuset_release_agent"; | |
f7e83571 AV |
276 | ret = cgroup_fs->mount(cgroup_fs, flags, |
277 | unused_dev_name, mountopts); | |
8793d854 PM |
278 | put_filesystem(cgroup_fs); |
279 | } | |
280 | return ret; | |
1da177e4 LT |
281 | } |
282 | ||
283 | static struct file_system_type cpuset_fs_type = { | |
284 | .name = "cpuset", | |
f7e83571 | 285 | .mount = cpuset_mount, |
1da177e4 LT |
286 | }; |
287 | ||
1da177e4 | 288 | /* |
300ed6cb | 289 | * Return in pmask the portion of a cpusets's cpus_allowed that |
1da177e4 LT |
290 | * are online. If none are online, walk up the cpuset hierarchy |
291 | * until we find one that does have some online cpus. If we get | |
292 | * all the way to the top and still haven't found any online cpus, | |
5f054e31 RR |
293 | * return cpu_online_mask. Or if passed a NULL cs from an exit'ing |
294 | * task, return cpu_online_mask. | |
1da177e4 LT |
295 | * |
296 | * One way or another, we guarantee to return some non-empty subset | |
5f054e31 | 297 | * of cpu_online_mask. |
1da177e4 | 298 | * |
3d3f26a7 | 299 | * Call with callback_mutex held. |
1da177e4 LT |
300 | */ |
301 | ||
6af866af LZ |
302 | static void guarantee_online_cpus(const struct cpuset *cs, |
303 | struct cpumask *pmask) | |
1da177e4 | 304 | { |
300ed6cb | 305 | while (cs && !cpumask_intersects(cs->cpus_allowed, cpu_online_mask)) |
1da177e4 LT |
306 | cs = cs->parent; |
307 | if (cs) | |
300ed6cb | 308 | cpumask_and(pmask, cs->cpus_allowed, cpu_online_mask); |
1da177e4 | 309 | else |
300ed6cb LZ |
310 | cpumask_copy(pmask, cpu_online_mask); |
311 | BUG_ON(!cpumask_intersects(pmask, cpu_online_mask)); | |
1da177e4 LT |
312 | } |
313 | ||
314 | /* | |
315 | * Return in *pmask the portion of a cpusets's mems_allowed that | |
0e1e7c7a CL |
316 | * are online, with memory. If none are online with memory, walk |
317 | * up the cpuset hierarchy until we find one that does have some | |
318 | * online mems. If we get all the way to the top and still haven't | |
38d7bee9 | 319 | * found any online mems, return node_states[N_MEMORY]. |
1da177e4 LT |
320 | * |
321 | * One way or another, we guarantee to return some non-empty subset | |
38d7bee9 | 322 | * of node_states[N_MEMORY]. |
1da177e4 | 323 | * |
3d3f26a7 | 324 | * Call with callback_mutex held. |
1da177e4 LT |
325 | */ |
326 | ||
327 | static void guarantee_online_mems(const struct cpuset *cs, nodemask_t *pmask) | |
328 | { | |
0e1e7c7a | 329 | while (cs && !nodes_intersects(cs->mems_allowed, |
38d7bee9 | 330 | node_states[N_MEMORY])) |
1da177e4 LT |
331 | cs = cs->parent; |
332 | if (cs) | |
0e1e7c7a | 333 | nodes_and(*pmask, cs->mems_allowed, |
38d7bee9 | 334 | node_states[N_MEMORY]); |
1da177e4 | 335 | else |
38d7bee9 LJ |
336 | *pmask = node_states[N_MEMORY]; |
337 | BUG_ON(!nodes_intersects(*pmask, node_states[N_MEMORY])); | |
1da177e4 LT |
338 | } |
339 | ||
f3b39d47 MX |
340 | /* |
341 | * update task's spread flag if cpuset's page/slab spread flag is set | |
342 | * | |
343 | * Called with callback_mutex/cgroup_mutex held | |
344 | */ | |
345 | static void cpuset_update_task_spread_flag(struct cpuset *cs, | |
346 | struct task_struct *tsk) | |
347 | { | |
348 | if (is_spread_page(cs)) | |
349 | tsk->flags |= PF_SPREAD_PAGE; | |
350 | else | |
351 | tsk->flags &= ~PF_SPREAD_PAGE; | |
352 | if (is_spread_slab(cs)) | |
353 | tsk->flags |= PF_SPREAD_SLAB; | |
354 | else | |
355 | tsk->flags &= ~PF_SPREAD_SLAB; | |
356 | } | |
357 | ||
1da177e4 LT |
358 | /* |
359 | * is_cpuset_subset(p, q) - Is cpuset p a subset of cpuset q? | |
360 | * | |
361 | * One cpuset is a subset of another if all its allowed CPUs and | |
362 | * Memory Nodes are a subset of the other, and its exclusive flags | |
2df167a3 | 363 | * are only set if the other's are set. Call holding cgroup_mutex. |
1da177e4 LT |
364 | */ |
365 | ||
366 | static int is_cpuset_subset(const struct cpuset *p, const struct cpuset *q) | |
367 | { | |
300ed6cb | 368 | return cpumask_subset(p->cpus_allowed, q->cpus_allowed) && |
1da177e4 LT |
369 | nodes_subset(p->mems_allowed, q->mems_allowed) && |
370 | is_cpu_exclusive(p) <= is_cpu_exclusive(q) && | |
371 | is_mem_exclusive(p) <= is_mem_exclusive(q); | |
372 | } | |
373 | ||
645fcc9d LZ |
374 | /** |
375 | * alloc_trial_cpuset - allocate a trial cpuset | |
376 | * @cs: the cpuset that the trial cpuset duplicates | |
377 | */ | |
378 | static struct cpuset *alloc_trial_cpuset(const struct cpuset *cs) | |
379 | { | |
300ed6cb LZ |
380 | struct cpuset *trial; |
381 | ||
382 | trial = kmemdup(cs, sizeof(*cs), GFP_KERNEL); | |
383 | if (!trial) | |
384 | return NULL; | |
385 | ||
386 | if (!alloc_cpumask_var(&trial->cpus_allowed, GFP_KERNEL)) { | |
387 | kfree(trial); | |
388 | return NULL; | |
389 | } | |
390 | cpumask_copy(trial->cpus_allowed, cs->cpus_allowed); | |
391 | ||
392 | return trial; | |
645fcc9d LZ |
393 | } |
394 | ||
395 | /** | |
396 | * free_trial_cpuset - free the trial cpuset | |
397 | * @trial: the trial cpuset to be freed | |
398 | */ | |
399 | static void free_trial_cpuset(struct cpuset *trial) | |
400 | { | |
300ed6cb | 401 | free_cpumask_var(trial->cpus_allowed); |
645fcc9d LZ |
402 | kfree(trial); |
403 | } | |
404 | ||
1da177e4 LT |
405 | /* |
406 | * validate_change() - Used to validate that any proposed cpuset change | |
407 | * follows the structural rules for cpusets. | |
408 | * | |
409 | * If we replaced the flag and mask values of the current cpuset | |
410 | * (cur) with those values in the trial cpuset (trial), would | |
411 | * our various subset and exclusive rules still be valid? Presumes | |
2df167a3 | 412 | * cgroup_mutex held. |
1da177e4 LT |
413 | * |
414 | * 'cur' is the address of an actual, in-use cpuset. Operations | |
415 | * such as list traversal that depend on the actual address of the | |
416 | * cpuset in the list must use cur below, not trial. | |
417 | * | |
418 | * 'trial' is the address of bulk structure copy of cur, with | |
419 | * perhaps one or more of the fields cpus_allowed, mems_allowed, | |
420 | * or flags changed to new, trial values. | |
421 | * | |
422 | * Return 0 if valid, -errno if not. | |
423 | */ | |
424 | ||
425 | static int validate_change(const struct cpuset *cur, const struct cpuset *trial) | |
426 | { | |
8793d854 | 427 | struct cgroup *cont; |
1da177e4 | 428 | struct cpuset *c, *par; |
ae8086ce TH |
429 | int ret; |
430 | ||
431 | rcu_read_lock(); | |
1da177e4 LT |
432 | |
433 | /* Each of our child cpusets must be a subset of us */ | |
ae8086ce TH |
434 | ret = -EBUSY; |
435 | cpuset_for_each_child(c, cont, cur) | |
436 | if (!is_cpuset_subset(c, trial)) | |
437 | goto out; | |
1da177e4 LT |
438 | |
439 | /* Remaining checks don't apply to root cpuset */ | |
ae8086ce | 440 | ret = 0; |
69604067 | 441 | if (cur == &top_cpuset) |
ae8086ce | 442 | goto out; |
1da177e4 | 443 | |
69604067 PJ |
444 | par = cur->parent; |
445 | ||
1da177e4 | 446 | /* We must be a subset of our parent cpuset */ |
ae8086ce | 447 | ret = -EACCES; |
1da177e4 | 448 | if (!is_cpuset_subset(trial, par)) |
ae8086ce | 449 | goto out; |
1da177e4 | 450 | |
2df167a3 PM |
451 | /* |
452 | * If either I or some sibling (!= me) is exclusive, we can't | |
453 | * overlap | |
454 | */ | |
ae8086ce TH |
455 | ret = -EINVAL; |
456 | cpuset_for_each_child(c, cont, par) { | |
1da177e4 LT |
457 | if ((is_cpu_exclusive(trial) || is_cpu_exclusive(c)) && |
458 | c != cur && | |
300ed6cb | 459 | cpumask_intersects(trial->cpus_allowed, c->cpus_allowed)) |
ae8086ce | 460 | goto out; |
1da177e4 LT |
461 | if ((is_mem_exclusive(trial) || is_mem_exclusive(c)) && |
462 | c != cur && | |
463 | nodes_intersects(trial->mems_allowed, c->mems_allowed)) | |
ae8086ce | 464 | goto out; |
1da177e4 LT |
465 | } |
466 | ||
020958b6 | 467 | /* Cpusets with tasks can't have empty cpus_allowed or mems_allowed */ |
ae8086ce TH |
468 | ret = -ENOSPC; |
469 | if (cgroup_task_count(cur->css.cgroup) && | |
470 | (cpumask_empty(trial->cpus_allowed) || | |
471 | nodes_empty(trial->mems_allowed))) | |
472 | goto out; | |
020958b6 | 473 | |
ae8086ce TH |
474 | ret = 0; |
475 | out: | |
476 | rcu_read_unlock(); | |
477 | return ret; | |
1da177e4 LT |
478 | } |
479 | ||
db7f47cf | 480 | #ifdef CONFIG_SMP |
029190c5 | 481 | /* |
cf417141 | 482 | * Helper routine for generate_sched_domains(). |
029190c5 PJ |
483 | * Do cpusets a, b have overlapping cpus_allowed masks? |
484 | */ | |
029190c5 PJ |
485 | static int cpusets_overlap(struct cpuset *a, struct cpuset *b) |
486 | { | |
300ed6cb | 487 | return cpumask_intersects(a->cpus_allowed, b->cpus_allowed); |
029190c5 PJ |
488 | } |
489 | ||
1d3504fc HS |
490 | static void |
491 | update_domain_attr(struct sched_domain_attr *dattr, struct cpuset *c) | |
492 | { | |
1d3504fc HS |
493 | if (dattr->relax_domain_level < c->relax_domain_level) |
494 | dattr->relax_domain_level = c->relax_domain_level; | |
495 | return; | |
496 | } | |
497 | ||
f5393693 LJ |
498 | static void |
499 | update_domain_attr_tree(struct sched_domain_attr *dattr, struct cpuset *c) | |
500 | { | |
501 | LIST_HEAD(q); | |
502 | ||
503 | list_add(&c->stack_list, &q); | |
504 | while (!list_empty(&q)) { | |
505 | struct cpuset *cp; | |
506 | struct cgroup *cont; | |
507 | struct cpuset *child; | |
508 | ||
509 | cp = list_first_entry(&q, struct cpuset, stack_list); | |
510 | list_del(q.next); | |
511 | ||
300ed6cb | 512 | if (cpumask_empty(cp->cpus_allowed)) |
f5393693 LJ |
513 | continue; |
514 | ||
515 | if (is_sched_load_balance(cp)) | |
516 | update_domain_attr(dattr, cp); | |
517 | ||
ae8086ce TH |
518 | rcu_read_lock(); |
519 | cpuset_for_each_child(child, cont, cp) | |
f5393693 | 520 | list_add_tail(&child->stack_list, &q); |
ae8086ce | 521 | rcu_read_unlock(); |
f5393693 LJ |
522 | } |
523 | } | |
524 | ||
029190c5 | 525 | /* |
cf417141 MK |
526 | * generate_sched_domains() |
527 | * | |
528 | * This function builds a partial partition of the systems CPUs | |
529 | * A 'partial partition' is a set of non-overlapping subsets whose | |
530 | * union is a subset of that set. | |
531 | * The output of this function needs to be passed to kernel/sched.c | |
532 | * partition_sched_domains() routine, which will rebuild the scheduler's | |
533 | * load balancing domains (sched domains) as specified by that partial | |
534 | * partition. | |
029190c5 | 535 | * |
45ce80fb | 536 | * See "What is sched_load_balance" in Documentation/cgroups/cpusets.txt |
029190c5 PJ |
537 | * for a background explanation of this. |
538 | * | |
539 | * Does not return errors, on the theory that the callers of this | |
540 | * routine would rather not worry about failures to rebuild sched | |
541 | * domains when operating in the severe memory shortage situations | |
542 | * that could cause allocation failures below. | |
543 | * | |
cf417141 | 544 | * Must be called with cgroup_lock held. |
029190c5 PJ |
545 | * |
546 | * The three key local variables below are: | |
aeed6824 | 547 | * q - a linked-list queue of cpuset pointers, used to implement a |
029190c5 PJ |
548 | * top-down scan of all cpusets. This scan loads a pointer |
549 | * to each cpuset marked is_sched_load_balance into the | |
550 | * array 'csa'. For our purposes, rebuilding the schedulers | |
551 | * sched domains, we can ignore !is_sched_load_balance cpusets. | |
552 | * csa - (for CpuSet Array) Array of pointers to all the cpusets | |
553 | * that need to be load balanced, for convenient iterative | |
554 | * access by the subsequent code that finds the best partition, | |
555 | * i.e the set of domains (subsets) of CPUs such that the | |
556 | * cpus_allowed of every cpuset marked is_sched_load_balance | |
557 | * is a subset of one of these domains, while there are as | |
558 | * many such domains as possible, each as small as possible. | |
559 | * doms - Conversion of 'csa' to an array of cpumasks, for passing to | |
560 | * the kernel/sched.c routine partition_sched_domains() in a | |
561 | * convenient format, that can be easily compared to the prior | |
562 | * value to determine what partition elements (sched domains) | |
563 | * were changed (added or removed.) | |
564 | * | |
565 | * Finding the best partition (set of domains): | |
566 | * The triple nested loops below over i, j, k scan over the | |
567 | * load balanced cpusets (using the array of cpuset pointers in | |
568 | * csa[]) looking for pairs of cpusets that have overlapping | |
569 | * cpus_allowed, but which don't have the same 'pn' partition | |
570 | * number and gives them in the same partition number. It keeps | |
571 | * looping on the 'restart' label until it can no longer find | |
572 | * any such pairs. | |
573 | * | |
574 | * The union of the cpus_allowed masks from the set of | |
575 | * all cpusets having the same 'pn' value then form the one | |
576 | * element of the partition (one sched domain) to be passed to | |
577 | * partition_sched_domains(). | |
578 | */ | |
acc3f5d7 | 579 | static int generate_sched_domains(cpumask_var_t **domains, |
cf417141 | 580 | struct sched_domain_attr **attributes) |
029190c5 | 581 | { |
cf417141 | 582 | LIST_HEAD(q); /* queue of cpusets to be scanned */ |
029190c5 PJ |
583 | struct cpuset *cp; /* scans q */ |
584 | struct cpuset **csa; /* array of all cpuset ptrs */ | |
585 | int csn; /* how many cpuset ptrs in csa so far */ | |
586 | int i, j, k; /* indices for partition finding loops */ | |
acc3f5d7 | 587 | cpumask_var_t *doms; /* resulting partition; i.e. sched domains */ |
1d3504fc | 588 | struct sched_domain_attr *dattr; /* attributes for custom domains */ |
1583715d | 589 | int ndoms = 0; /* number of sched domains in result */ |
6af866af | 590 | int nslot; /* next empty doms[] struct cpumask slot */ |
029190c5 | 591 | |
029190c5 | 592 | doms = NULL; |
1d3504fc | 593 | dattr = NULL; |
cf417141 | 594 | csa = NULL; |
029190c5 PJ |
595 | |
596 | /* Special case for the 99% of systems with one, full, sched domain */ | |
597 | if (is_sched_load_balance(&top_cpuset)) { | |
acc3f5d7 RR |
598 | ndoms = 1; |
599 | doms = alloc_sched_domains(ndoms); | |
029190c5 | 600 | if (!doms) |
cf417141 MK |
601 | goto done; |
602 | ||
1d3504fc HS |
603 | dattr = kmalloc(sizeof(struct sched_domain_attr), GFP_KERNEL); |
604 | if (dattr) { | |
605 | *dattr = SD_ATTR_INIT; | |
93a65575 | 606 | update_domain_attr_tree(dattr, &top_cpuset); |
1d3504fc | 607 | } |
acc3f5d7 | 608 | cpumask_copy(doms[0], top_cpuset.cpus_allowed); |
cf417141 | 609 | |
cf417141 | 610 | goto done; |
029190c5 PJ |
611 | } |
612 | ||
029190c5 PJ |
613 | csa = kmalloc(number_of_cpusets * sizeof(cp), GFP_KERNEL); |
614 | if (!csa) | |
615 | goto done; | |
616 | csn = 0; | |
617 | ||
aeed6824 LZ |
618 | list_add(&top_cpuset.stack_list, &q); |
619 | while (!list_empty(&q)) { | |
029190c5 PJ |
620 | struct cgroup *cont; |
621 | struct cpuset *child; /* scans child cpusets of cp */ | |
489a5393 | 622 | |
aeed6824 LZ |
623 | cp = list_first_entry(&q, struct cpuset, stack_list); |
624 | list_del(q.next); | |
625 | ||
300ed6cb | 626 | if (cpumask_empty(cp->cpus_allowed)) |
489a5393 LJ |
627 | continue; |
628 | ||
f5393693 LJ |
629 | /* |
630 | * All child cpusets contain a subset of the parent's cpus, so | |
631 | * just skip them, and then we call update_domain_attr_tree() | |
632 | * to calc relax_domain_level of the corresponding sched | |
633 | * domain. | |
634 | */ | |
635 | if (is_sched_load_balance(cp)) { | |
029190c5 | 636 | csa[csn++] = cp; |
f5393693 LJ |
637 | continue; |
638 | } | |
489a5393 | 639 | |
ae8086ce TH |
640 | rcu_read_lock(); |
641 | cpuset_for_each_child(child, cont, cp) | |
aeed6824 | 642 | list_add_tail(&child->stack_list, &q); |
ae8086ce | 643 | rcu_read_unlock(); |
029190c5 PJ |
644 | } |
645 | ||
646 | for (i = 0; i < csn; i++) | |
647 | csa[i]->pn = i; | |
648 | ndoms = csn; | |
649 | ||
650 | restart: | |
651 | /* Find the best partition (set of sched domains) */ | |
652 | for (i = 0; i < csn; i++) { | |
653 | struct cpuset *a = csa[i]; | |
654 | int apn = a->pn; | |
655 | ||
656 | for (j = 0; j < csn; j++) { | |
657 | struct cpuset *b = csa[j]; | |
658 | int bpn = b->pn; | |
659 | ||
660 | if (apn != bpn && cpusets_overlap(a, b)) { | |
661 | for (k = 0; k < csn; k++) { | |
662 | struct cpuset *c = csa[k]; | |
663 | ||
664 | if (c->pn == bpn) | |
665 | c->pn = apn; | |
666 | } | |
667 | ndoms--; /* one less element */ | |
668 | goto restart; | |
669 | } | |
670 | } | |
671 | } | |
672 | ||
cf417141 MK |
673 | /* |
674 | * Now we know how many domains to create. | |
675 | * Convert <csn, csa> to <ndoms, doms> and populate cpu masks. | |
676 | */ | |
acc3f5d7 | 677 | doms = alloc_sched_domains(ndoms); |
700018e0 | 678 | if (!doms) |
cf417141 | 679 | goto done; |
cf417141 MK |
680 | |
681 | /* | |
682 | * The rest of the code, including the scheduler, can deal with | |
683 | * dattr==NULL case. No need to abort if alloc fails. | |
684 | */ | |
1d3504fc | 685 | dattr = kmalloc(ndoms * sizeof(struct sched_domain_attr), GFP_KERNEL); |
029190c5 PJ |
686 | |
687 | for (nslot = 0, i = 0; i < csn; i++) { | |
688 | struct cpuset *a = csa[i]; | |
6af866af | 689 | struct cpumask *dp; |
029190c5 PJ |
690 | int apn = a->pn; |
691 | ||
cf417141 MK |
692 | if (apn < 0) { |
693 | /* Skip completed partitions */ | |
694 | continue; | |
695 | } | |
696 | ||
acc3f5d7 | 697 | dp = doms[nslot]; |
cf417141 MK |
698 | |
699 | if (nslot == ndoms) { | |
700 | static int warnings = 10; | |
701 | if (warnings) { | |
702 | printk(KERN_WARNING | |
703 | "rebuild_sched_domains confused:" | |
704 | " nslot %d, ndoms %d, csn %d, i %d," | |
705 | " apn %d\n", | |
706 | nslot, ndoms, csn, i, apn); | |
707 | warnings--; | |
029190c5 | 708 | } |
cf417141 MK |
709 | continue; |
710 | } | |
029190c5 | 711 | |
6af866af | 712 | cpumask_clear(dp); |
cf417141 MK |
713 | if (dattr) |
714 | *(dattr + nslot) = SD_ATTR_INIT; | |
715 | for (j = i; j < csn; j++) { | |
716 | struct cpuset *b = csa[j]; | |
717 | ||
718 | if (apn == b->pn) { | |
300ed6cb | 719 | cpumask_or(dp, dp, b->cpus_allowed); |
cf417141 MK |
720 | if (dattr) |
721 | update_domain_attr_tree(dattr + nslot, b); | |
722 | ||
723 | /* Done with this partition */ | |
724 | b->pn = -1; | |
029190c5 | 725 | } |
029190c5 | 726 | } |
cf417141 | 727 | nslot++; |
029190c5 PJ |
728 | } |
729 | BUG_ON(nslot != ndoms); | |
730 | ||
cf417141 MK |
731 | done: |
732 | kfree(csa); | |
733 | ||
700018e0 LZ |
734 | /* |
735 | * Fallback to the default domain if kmalloc() failed. | |
736 | * See comments in partition_sched_domains(). | |
737 | */ | |
738 | if (doms == NULL) | |
739 | ndoms = 1; | |
740 | ||
cf417141 MK |
741 | *domains = doms; |
742 | *attributes = dattr; | |
743 | return ndoms; | |
744 | } | |
745 | ||
746 | /* | |
747 | * Rebuild scheduler domains. | |
748 | * | |
749 | * Call with neither cgroup_mutex held nor within get_online_cpus(). | |
750 | * Takes both cgroup_mutex and get_online_cpus(). | |
751 | * | |
752 | * Cannot be directly called from cpuset code handling changes | |
753 | * to the cpuset pseudo-filesystem, because it cannot be called | |
754 | * from code that already holds cgroup_mutex. | |
755 | */ | |
756 | static void do_rebuild_sched_domains(struct work_struct *unused) | |
757 | { | |
758 | struct sched_domain_attr *attr; | |
acc3f5d7 | 759 | cpumask_var_t *doms; |
cf417141 MK |
760 | int ndoms; |
761 | ||
86ef5c9a | 762 | get_online_cpus(); |
cf417141 MK |
763 | |
764 | /* Generate domain masks and attrs */ | |
765 | cgroup_lock(); | |
766 | ndoms = generate_sched_domains(&doms, &attr); | |
767 | cgroup_unlock(); | |
768 | ||
769 | /* Have scheduler rebuild the domains */ | |
770 | partition_sched_domains(ndoms, doms, attr); | |
771 | ||
86ef5c9a | 772 | put_online_cpus(); |
cf417141 | 773 | } |
db7f47cf PM |
774 | #else /* !CONFIG_SMP */ |
775 | static void do_rebuild_sched_domains(struct work_struct *unused) | |
776 | { | |
777 | } | |
778 | ||
e1b8090b | 779 | static int generate_sched_domains(cpumask_var_t **domains, |
db7f47cf PM |
780 | struct sched_domain_attr **attributes) |
781 | { | |
782 | *domains = NULL; | |
783 | return 1; | |
784 | } | |
785 | #endif /* CONFIG_SMP */ | |
029190c5 | 786 | |
cf417141 MK |
787 | static DECLARE_WORK(rebuild_sched_domains_work, do_rebuild_sched_domains); |
788 | ||
789 | /* | |
790 | * Rebuild scheduler domains, asynchronously via workqueue. | |
791 | * | |
792 | * If the flag 'sched_load_balance' of any cpuset with non-empty | |
793 | * 'cpus' changes, or if the 'cpus' allowed changes in any cpuset | |
794 | * which has that flag enabled, or if any cpuset with a non-empty | |
795 | * 'cpus' is removed, then call this routine to rebuild the | |
796 | * scheduler's dynamic sched domains. | |
797 | * | |
798 | * The rebuild_sched_domains() and partition_sched_domains() | |
799 | * routines must nest cgroup_lock() inside get_online_cpus(), | |
800 | * but such cpuset changes as these must nest that locking the | |
801 | * other way, holding cgroup_lock() for much of the code. | |
802 | * | |
803 | * So in order to avoid an ABBA deadlock, the cpuset code handling | |
804 | * these user changes delegates the actual sched domain rebuilding | |
805 | * to a separate workqueue thread, which ends up processing the | |
806 | * above do_rebuild_sched_domains() function. | |
807 | */ | |
808 | static void async_rebuild_sched_domains(void) | |
809 | { | |
f90d4118 | 810 | queue_work(cpuset_wq, &rebuild_sched_domains_work); |
cf417141 MK |
811 | } |
812 | ||
813 | /* | |
814 | * Accomplishes the same scheduler domain rebuild as the above | |
815 | * async_rebuild_sched_domains(), however it directly calls the | |
816 | * rebuild routine synchronously rather than calling it via an | |
817 | * asynchronous work thread. | |
818 | * | |
819 | * This can only be called from code that is not holding | |
820 | * cgroup_mutex (not nested in a cgroup_lock() call.) | |
821 | */ | |
822 | void rebuild_sched_domains(void) | |
823 | { | |
824 | do_rebuild_sched_domains(NULL); | |
029190c5 PJ |
825 | } |
826 | ||
58f4790b CW |
827 | /** |
828 | * cpuset_test_cpumask - test a task's cpus_allowed versus its cpuset's | |
829 | * @tsk: task to test | |
830 | * @scan: struct cgroup_scanner contained in its struct cpuset_hotplug_scanner | |
831 | * | |
2df167a3 | 832 | * Call with cgroup_mutex held. May take callback_mutex during call. |
58f4790b CW |
833 | * Called for each task in a cgroup by cgroup_scan_tasks(). |
834 | * Return nonzero if this tasks's cpus_allowed mask should be changed (in other | |
835 | * words, if its mask is not equal to its cpuset's mask). | |
053199ed | 836 | */ |
9e0c914c AB |
837 | static int cpuset_test_cpumask(struct task_struct *tsk, |
838 | struct cgroup_scanner *scan) | |
58f4790b | 839 | { |
300ed6cb | 840 | return !cpumask_equal(&tsk->cpus_allowed, |
58f4790b CW |
841 | (cgroup_cs(scan->cg))->cpus_allowed); |
842 | } | |
053199ed | 843 | |
58f4790b CW |
844 | /** |
845 | * cpuset_change_cpumask - make a task's cpus_allowed the same as its cpuset's | |
846 | * @tsk: task to test | |
847 | * @scan: struct cgroup_scanner containing the cgroup of the task | |
848 | * | |
849 | * Called by cgroup_scan_tasks() for each task in a cgroup whose | |
850 | * cpus_allowed mask needs to be changed. | |
851 | * | |
852 | * We don't need to re-check for the cgroup/cpuset membership, since we're | |
853 | * holding cgroup_lock() at this point. | |
854 | */ | |
9e0c914c AB |
855 | static void cpuset_change_cpumask(struct task_struct *tsk, |
856 | struct cgroup_scanner *scan) | |
58f4790b | 857 | { |
300ed6cb | 858 | set_cpus_allowed_ptr(tsk, ((cgroup_cs(scan->cg))->cpus_allowed)); |
58f4790b CW |
859 | } |
860 | ||
0b2f630a MX |
861 | /** |
862 | * update_tasks_cpumask - Update the cpumasks of tasks in the cpuset. | |
863 | * @cs: the cpuset in which each task's cpus_allowed mask needs to be changed | |
4e74339a | 864 | * @heap: if NULL, defer allocating heap memory to cgroup_scan_tasks() |
0b2f630a MX |
865 | * |
866 | * Called with cgroup_mutex held | |
867 | * | |
868 | * The cgroup_scan_tasks() function will scan all the tasks in a cgroup, | |
869 | * calling callback functions for each. | |
870 | * | |
4e74339a LZ |
871 | * No return value. It's guaranteed that cgroup_scan_tasks() always returns 0 |
872 | * if @heap != NULL. | |
0b2f630a | 873 | */ |
4e74339a | 874 | static void update_tasks_cpumask(struct cpuset *cs, struct ptr_heap *heap) |
0b2f630a MX |
875 | { |
876 | struct cgroup_scanner scan; | |
0b2f630a MX |
877 | |
878 | scan.cg = cs->css.cgroup; | |
879 | scan.test_task = cpuset_test_cpumask; | |
880 | scan.process_task = cpuset_change_cpumask; | |
4e74339a LZ |
881 | scan.heap = heap; |
882 | cgroup_scan_tasks(&scan); | |
0b2f630a MX |
883 | } |
884 | ||
58f4790b CW |
885 | /** |
886 | * update_cpumask - update the cpus_allowed mask of a cpuset and all tasks in it | |
887 | * @cs: the cpuset to consider | |
888 | * @buf: buffer of cpu numbers written to this cpuset | |
889 | */ | |
645fcc9d LZ |
890 | static int update_cpumask(struct cpuset *cs, struct cpuset *trialcs, |
891 | const char *buf) | |
1da177e4 | 892 | { |
4e74339a | 893 | struct ptr_heap heap; |
58f4790b CW |
894 | int retval; |
895 | int is_load_balanced; | |
1da177e4 | 896 | |
5f054e31 | 897 | /* top_cpuset.cpus_allowed tracks cpu_online_mask; it's read-only */ |
4c4d50f7 PJ |
898 | if (cs == &top_cpuset) |
899 | return -EACCES; | |
900 | ||
6f7f02e7 | 901 | /* |
c8d9c90c | 902 | * An empty cpus_allowed is ok only if the cpuset has no tasks. |
020958b6 PJ |
903 | * Since cpulist_parse() fails on an empty mask, we special case |
904 | * that parsing. The validate_change() call ensures that cpusets | |
905 | * with tasks have cpus. | |
6f7f02e7 | 906 | */ |
020958b6 | 907 | if (!*buf) { |
300ed6cb | 908 | cpumask_clear(trialcs->cpus_allowed); |
6f7f02e7 | 909 | } else { |
300ed6cb | 910 | retval = cpulist_parse(buf, trialcs->cpus_allowed); |
6f7f02e7 DR |
911 | if (retval < 0) |
912 | return retval; | |
37340746 | 913 | |
6ad4c188 | 914 | if (!cpumask_subset(trialcs->cpus_allowed, cpu_active_mask)) |
37340746 | 915 | return -EINVAL; |
6f7f02e7 | 916 | } |
645fcc9d | 917 | retval = validate_change(cs, trialcs); |
85d7b949 DG |
918 | if (retval < 0) |
919 | return retval; | |
029190c5 | 920 | |
8707d8b8 | 921 | /* Nothing to do if the cpus didn't change */ |
300ed6cb | 922 | if (cpumask_equal(cs->cpus_allowed, trialcs->cpus_allowed)) |
8707d8b8 | 923 | return 0; |
58f4790b | 924 | |
4e74339a LZ |
925 | retval = heap_init(&heap, PAGE_SIZE, GFP_KERNEL, NULL); |
926 | if (retval) | |
927 | return retval; | |
928 | ||
645fcc9d | 929 | is_load_balanced = is_sched_load_balance(trialcs); |
029190c5 | 930 | |
3d3f26a7 | 931 | mutex_lock(&callback_mutex); |
300ed6cb | 932 | cpumask_copy(cs->cpus_allowed, trialcs->cpus_allowed); |
3d3f26a7 | 933 | mutex_unlock(&callback_mutex); |
029190c5 | 934 | |
8707d8b8 PM |
935 | /* |
936 | * Scan tasks in the cpuset, and update the cpumasks of any | |
58f4790b | 937 | * that need an update. |
8707d8b8 | 938 | */ |
4e74339a LZ |
939 | update_tasks_cpumask(cs, &heap); |
940 | ||
941 | heap_free(&heap); | |
58f4790b | 942 | |
8707d8b8 | 943 | if (is_load_balanced) |
cf417141 | 944 | async_rebuild_sched_domains(); |
85d7b949 | 945 | return 0; |
1da177e4 LT |
946 | } |
947 | ||
e4e364e8 PJ |
948 | /* |
949 | * cpuset_migrate_mm | |
950 | * | |
951 | * Migrate memory region from one set of nodes to another. | |
952 | * | |
953 | * Temporarilly set tasks mems_allowed to target nodes of migration, | |
954 | * so that the migration code can allocate pages on these nodes. | |
955 | * | |
2df167a3 | 956 | * Call holding cgroup_mutex, so current's cpuset won't change |
c8d9c90c | 957 | * during this call, as manage_mutex holds off any cpuset_attach() |
e4e364e8 PJ |
958 | * calls. Therefore we don't need to take task_lock around the |
959 | * call to guarantee_online_mems(), as we know no one is changing | |
2df167a3 | 960 | * our task's cpuset. |
e4e364e8 | 961 | * |
e4e364e8 PJ |
962 | * While the mm_struct we are migrating is typically from some |
963 | * other task, the task_struct mems_allowed that we are hacking | |
964 | * is for our current task, which must allocate new pages for that | |
965 | * migrating memory region. | |
e4e364e8 PJ |
966 | */ |
967 | ||
968 | static void cpuset_migrate_mm(struct mm_struct *mm, const nodemask_t *from, | |
969 | const nodemask_t *to) | |
970 | { | |
971 | struct task_struct *tsk = current; | |
972 | ||
e4e364e8 | 973 | tsk->mems_allowed = *to; |
e4e364e8 PJ |
974 | |
975 | do_migrate_pages(mm, from, to, MPOL_MF_MOVE_ALL); | |
976 | ||
8793d854 | 977 | guarantee_online_mems(task_cs(tsk),&tsk->mems_allowed); |
e4e364e8 PJ |
978 | } |
979 | ||
3b6766fe | 980 | /* |
58568d2a MX |
981 | * cpuset_change_task_nodemask - change task's mems_allowed and mempolicy |
982 | * @tsk: the task to change | |
983 | * @newmems: new nodes that the task will be set | |
984 | * | |
985 | * In order to avoid seeing no nodes if the old and new nodes are disjoint, | |
986 | * we structure updates as setting all new allowed nodes, then clearing newly | |
987 | * disallowed ones. | |
58568d2a MX |
988 | */ |
989 | static void cpuset_change_task_nodemask(struct task_struct *tsk, | |
990 | nodemask_t *newmems) | |
991 | { | |
b246272e | 992 | bool need_loop; |
89e8a244 | 993 | |
c0ff7453 MX |
994 | /* |
995 | * Allow tasks that have access to memory reserves because they have | |
996 | * been OOM killed to get memory anywhere. | |
997 | */ | |
998 | if (unlikely(test_thread_flag(TIF_MEMDIE))) | |
999 | return; | |
1000 | if (current->flags & PF_EXITING) /* Let dying task have memory */ | |
1001 | return; | |
1002 | ||
1003 | task_lock(tsk); | |
b246272e DR |
1004 | /* |
1005 | * Determine if a loop is necessary if another thread is doing | |
1006 | * get_mems_allowed(). If at least one node remains unchanged and | |
1007 | * tsk does not have a mempolicy, then an empty nodemask will not be | |
1008 | * possible when mems_allowed is larger than a word. | |
1009 | */ | |
1010 | need_loop = task_has_mempolicy(tsk) || | |
1011 | !nodes_intersects(*newmems, tsk->mems_allowed); | |
c0ff7453 | 1012 | |
cc9a6c87 MG |
1013 | if (need_loop) |
1014 | write_seqcount_begin(&tsk->mems_allowed_seq); | |
c0ff7453 | 1015 | |
cc9a6c87 MG |
1016 | nodes_or(tsk->mems_allowed, tsk->mems_allowed, *newmems); |
1017 | mpol_rebind_task(tsk, newmems, MPOL_REBIND_STEP1); | |
c0ff7453 MX |
1018 | |
1019 | mpol_rebind_task(tsk, newmems, MPOL_REBIND_STEP2); | |
58568d2a | 1020 | tsk->mems_allowed = *newmems; |
cc9a6c87 MG |
1021 | |
1022 | if (need_loop) | |
1023 | write_seqcount_end(&tsk->mems_allowed_seq); | |
1024 | ||
c0ff7453 | 1025 | task_unlock(tsk); |
58568d2a MX |
1026 | } |
1027 | ||
1028 | /* | |
1029 | * Update task's mems_allowed and rebind its mempolicy and vmas' mempolicy | |
1030 | * of it to cpuset's new mems_allowed, and migrate pages to new nodes if | |
1031 | * memory_migrate flag is set. Called with cgroup_mutex held. | |
3b6766fe LZ |
1032 | */ |
1033 | static void cpuset_change_nodemask(struct task_struct *p, | |
1034 | struct cgroup_scanner *scan) | |
1035 | { | |
1036 | struct mm_struct *mm; | |
1037 | struct cpuset *cs; | |
1038 | int migrate; | |
1039 | const nodemask_t *oldmem = scan->data; | |
ee24d379 | 1040 | static nodemask_t newmems; /* protected by cgroup_mutex */ |
58568d2a MX |
1041 | |
1042 | cs = cgroup_cs(scan->cg); | |
ee24d379 | 1043 | guarantee_online_mems(cs, &newmems); |
58568d2a | 1044 | |
ee24d379 | 1045 | cpuset_change_task_nodemask(p, &newmems); |
53feb297 | 1046 | |
3b6766fe LZ |
1047 | mm = get_task_mm(p); |
1048 | if (!mm) | |
1049 | return; | |
1050 | ||
3b6766fe LZ |
1051 | migrate = is_memory_migrate(cs); |
1052 | ||
1053 | mpol_rebind_mm(mm, &cs->mems_allowed); | |
1054 | if (migrate) | |
1055 | cpuset_migrate_mm(mm, oldmem, &cs->mems_allowed); | |
1056 | mmput(mm); | |
1057 | } | |
1058 | ||
8793d854 PM |
1059 | static void *cpuset_being_rebound; |
1060 | ||
0b2f630a MX |
1061 | /** |
1062 | * update_tasks_nodemask - Update the nodemasks of tasks in the cpuset. | |
1063 | * @cs: the cpuset in which each task's mems_allowed mask needs to be changed | |
1064 | * @oldmem: old mems_allowed of cpuset cs | |
010cfac4 | 1065 | * @heap: if NULL, defer allocating heap memory to cgroup_scan_tasks() |
0b2f630a MX |
1066 | * |
1067 | * Called with cgroup_mutex held | |
010cfac4 LZ |
1068 | * No return value. It's guaranteed that cgroup_scan_tasks() always returns 0 |
1069 | * if @heap != NULL. | |
0b2f630a | 1070 | */ |
010cfac4 LZ |
1071 | static void update_tasks_nodemask(struct cpuset *cs, const nodemask_t *oldmem, |
1072 | struct ptr_heap *heap) | |
1da177e4 | 1073 | { |
3b6766fe | 1074 | struct cgroup_scanner scan; |
59dac16f | 1075 | |
846a16bf | 1076 | cpuset_being_rebound = cs; /* causes mpol_dup() rebind */ |
4225399a | 1077 | |
3b6766fe LZ |
1078 | scan.cg = cs->css.cgroup; |
1079 | scan.test_task = NULL; | |
1080 | scan.process_task = cpuset_change_nodemask; | |
010cfac4 | 1081 | scan.heap = heap; |
3b6766fe | 1082 | scan.data = (nodemask_t *)oldmem; |
4225399a PJ |
1083 | |
1084 | /* | |
3b6766fe LZ |
1085 | * The mpol_rebind_mm() call takes mmap_sem, which we couldn't |
1086 | * take while holding tasklist_lock. Forks can happen - the | |
1087 | * mpol_dup() cpuset_being_rebound check will catch such forks, | |
1088 | * and rebind their vma mempolicies too. Because we still hold | |
1089 | * the global cgroup_mutex, we know that no other rebind effort | |
1090 | * will be contending for the global variable cpuset_being_rebound. | |
4225399a | 1091 | * It's ok if we rebind the same mm twice; mpol_rebind_mm() |
04c19fa6 | 1092 | * is idempotent. Also migrate pages in each mm to new nodes. |
4225399a | 1093 | */ |
010cfac4 | 1094 | cgroup_scan_tasks(&scan); |
4225399a | 1095 | |
2df167a3 | 1096 | /* We're done rebinding vmas to this cpuset's new mems_allowed. */ |
8793d854 | 1097 | cpuset_being_rebound = NULL; |
1da177e4 LT |
1098 | } |
1099 | ||
0b2f630a MX |
1100 | /* |
1101 | * Handle user request to change the 'mems' memory placement | |
1102 | * of a cpuset. Needs to validate the request, update the | |
58568d2a MX |
1103 | * cpusets mems_allowed, and for each task in the cpuset, |
1104 | * update mems_allowed and rebind task's mempolicy and any vma | |
1105 | * mempolicies and if the cpuset is marked 'memory_migrate', | |
1106 | * migrate the tasks pages to the new memory. | |
0b2f630a MX |
1107 | * |
1108 | * Call with cgroup_mutex held. May take callback_mutex during call. | |
1109 | * Will take tasklist_lock, scan tasklist for tasks in cpuset cs, | |
1110 | * lock each such tasks mm->mmap_sem, scan its vma's and rebind | |
1111 | * their mempolicies to the cpusets new mems_allowed. | |
1112 | */ | |
645fcc9d LZ |
1113 | static int update_nodemask(struct cpuset *cs, struct cpuset *trialcs, |
1114 | const char *buf) | |
0b2f630a | 1115 | { |
53feb297 | 1116 | NODEMASK_ALLOC(nodemask_t, oldmem, GFP_KERNEL); |
0b2f630a | 1117 | int retval; |
010cfac4 | 1118 | struct ptr_heap heap; |
0b2f630a | 1119 | |
53feb297 MX |
1120 | if (!oldmem) |
1121 | return -ENOMEM; | |
1122 | ||
0b2f630a | 1123 | /* |
38d7bee9 | 1124 | * top_cpuset.mems_allowed tracks node_stats[N_MEMORY]; |
0b2f630a MX |
1125 | * it's read-only |
1126 | */ | |
53feb297 MX |
1127 | if (cs == &top_cpuset) { |
1128 | retval = -EACCES; | |
1129 | goto done; | |
1130 | } | |
0b2f630a | 1131 | |
0b2f630a MX |
1132 | /* |
1133 | * An empty mems_allowed is ok iff there are no tasks in the cpuset. | |
1134 | * Since nodelist_parse() fails on an empty mask, we special case | |
1135 | * that parsing. The validate_change() call ensures that cpusets | |
1136 | * with tasks have memory. | |
1137 | */ | |
1138 | if (!*buf) { | |
645fcc9d | 1139 | nodes_clear(trialcs->mems_allowed); |
0b2f630a | 1140 | } else { |
645fcc9d | 1141 | retval = nodelist_parse(buf, trialcs->mems_allowed); |
0b2f630a MX |
1142 | if (retval < 0) |
1143 | goto done; | |
1144 | ||
645fcc9d | 1145 | if (!nodes_subset(trialcs->mems_allowed, |
38d7bee9 | 1146 | node_states[N_MEMORY])) { |
53feb297 MX |
1147 | retval = -EINVAL; |
1148 | goto done; | |
1149 | } | |
0b2f630a | 1150 | } |
53feb297 MX |
1151 | *oldmem = cs->mems_allowed; |
1152 | if (nodes_equal(*oldmem, trialcs->mems_allowed)) { | |
0b2f630a MX |
1153 | retval = 0; /* Too easy - nothing to do */ |
1154 | goto done; | |
1155 | } | |
645fcc9d | 1156 | retval = validate_change(cs, trialcs); |
0b2f630a MX |
1157 | if (retval < 0) |
1158 | goto done; | |
1159 | ||
010cfac4 LZ |
1160 | retval = heap_init(&heap, PAGE_SIZE, GFP_KERNEL, NULL); |
1161 | if (retval < 0) | |
1162 | goto done; | |
1163 | ||
0b2f630a | 1164 | mutex_lock(&callback_mutex); |
645fcc9d | 1165 | cs->mems_allowed = trialcs->mems_allowed; |
0b2f630a MX |
1166 | mutex_unlock(&callback_mutex); |
1167 | ||
53feb297 | 1168 | update_tasks_nodemask(cs, oldmem, &heap); |
010cfac4 LZ |
1169 | |
1170 | heap_free(&heap); | |
0b2f630a | 1171 | done: |
53feb297 | 1172 | NODEMASK_FREE(oldmem); |
0b2f630a MX |
1173 | return retval; |
1174 | } | |
1175 | ||
8793d854 PM |
1176 | int current_cpuset_is_being_rebound(void) |
1177 | { | |
1178 | return task_cs(current) == cpuset_being_rebound; | |
1179 | } | |
1180 | ||
5be7a479 | 1181 | static int update_relax_domain_level(struct cpuset *cs, s64 val) |
1d3504fc | 1182 | { |
db7f47cf | 1183 | #ifdef CONFIG_SMP |
60495e77 | 1184 | if (val < -1 || val >= sched_domain_level_max) |
30e0e178 | 1185 | return -EINVAL; |
db7f47cf | 1186 | #endif |
1d3504fc HS |
1187 | |
1188 | if (val != cs->relax_domain_level) { | |
1189 | cs->relax_domain_level = val; | |
300ed6cb LZ |
1190 | if (!cpumask_empty(cs->cpus_allowed) && |
1191 | is_sched_load_balance(cs)) | |
cf417141 | 1192 | async_rebuild_sched_domains(); |
1d3504fc HS |
1193 | } |
1194 | ||
1195 | return 0; | |
1196 | } | |
1197 | ||
950592f7 MX |
1198 | /* |
1199 | * cpuset_change_flag - make a task's spread flags the same as its cpuset's | |
1200 | * @tsk: task to be updated | |
1201 | * @scan: struct cgroup_scanner containing the cgroup of the task | |
1202 | * | |
1203 | * Called by cgroup_scan_tasks() for each task in a cgroup. | |
1204 | * | |
1205 | * We don't need to re-check for the cgroup/cpuset membership, since we're | |
1206 | * holding cgroup_lock() at this point. | |
1207 | */ | |
1208 | static void cpuset_change_flag(struct task_struct *tsk, | |
1209 | struct cgroup_scanner *scan) | |
1210 | { | |
1211 | cpuset_update_task_spread_flag(cgroup_cs(scan->cg), tsk); | |
1212 | } | |
1213 | ||
1214 | /* | |
1215 | * update_tasks_flags - update the spread flags of tasks in the cpuset. | |
1216 | * @cs: the cpuset in which each task's spread flags needs to be changed | |
1217 | * @heap: if NULL, defer allocating heap memory to cgroup_scan_tasks() | |
1218 | * | |
1219 | * Called with cgroup_mutex held | |
1220 | * | |
1221 | * The cgroup_scan_tasks() function will scan all the tasks in a cgroup, | |
1222 | * calling callback functions for each. | |
1223 | * | |
1224 | * No return value. It's guaranteed that cgroup_scan_tasks() always returns 0 | |
1225 | * if @heap != NULL. | |
1226 | */ | |
1227 | static void update_tasks_flags(struct cpuset *cs, struct ptr_heap *heap) | |
1228 | { | |
1229 | struct cgroup_scanner scan; | |
1230 | ||
1231 | scan.cg = cs->css.cgroup; | |
1232 | scan.test_task = NULL; | |
1233 | scan.process_task = cpuset_change_flag; | |
1234 | scan.heap = heap; | |
1235 | cgroup_scan_tasks(&scan); | |
1236 | } | |
1237 | ||
1da177e4 LT |
1238 | /* |
1239 | * update_flag - read a 0 or a 1 in a file and update associated flag | |
78608366 PM |
1240 | * bit: the bit to update (see cpuset_flagbits_t) |
1241 | * cs: the cpuset to update | |
1242 | * turning_on: whether the flag is being set or cleared | |
053199ed | 1243 | * |
2df167a3 | 1244 | * Call with cgroup_mutex held. |
1da177e4 LT |
1245 | */ |
1246 | ||
700fe1ab PM |
1247 | static int update_flag(cpuset_flagbits_t bit, struct cpuset *cs, |
1248 | int turning_on) | |
1da177e4 | 1249 | { |
645fcc9d | 1250 | struct cpuset *trialcs; |
40b6a762 | 1251 | int balance_flag_changed; |
950592f7 MX |
1252 | int spread_flag_changed; |
1253 | struct ptr_heap heap; | |
1254 | int err; | |
1da177e4 | 1255 | |
645fcc9d LZ |
1256 | trialcs = alloc_trial_cpuset(cs); |
1257 | if (!trialcs) | |
1258 | return -ENOMEM; | |
1259 | ||
1da177e4 | 1260 | if (turning_on) |
645fcc9d | 1261 | set_bit(bit, &trialcs->flags); |
1da177e4 | 1262 | else |
645fcc9d | 1263 | clear_bit(bit, &trialcs->flags); |
1da177e4 | 1264 | |
645fcc9d | 1265 | err = validate_change(cs, trialcs); |
85d7b949 | 1266 | if (err < 0) |
645fcc9d | 1267 | goto out; |
029190c5 | 1268 | |
950592f7 MX |
1269 | err = heap_init(&heap, PAGE_SIZE, GFP_KERNEL, NULL); |
1270 | if (err < 0) | |
1271 | goto out; | |
1272 | ||
029190c5 | 1273 | balance_flag_changed = (is_sched_load_balance(cs) != |
645fcc9d | 1274 | is_sched_load_balance(trialcs)); |
029190c5 | 1275 | |
950592f7 MX |
1276 | spread_flag_changed = ((is_spread_slab(cs) != is_spread_slab(trialcs)) |
1277 | || (is_spread_page(cs) != is_spread_page(trialcs))); | |
1278 | ||
3d3f26a7 | 1279 | mutex_lock(&callback_mutex); |
645fcc9d | 1280 | cs->flags = trialcs->flags; |
3d3f26a7 | 1281 | mutex_unlock(&callback_mutex); |
85d7b949 | 1282 | |
300ed6cb | 1283 | if (!cpumask_empty(trialcs->cpus_allowed) && balance_flag_changed) |
cf417141 | 1284 | async_rebuild_sched_domains(); |
029190c5 | 1285 | |
950592f7 MX |
1286 | if (spread_flag_changed) |
1287 | update_tasks_flags(cs, &heap); | |
1288 | heap_free(&heap); | |
645fcc9d LZ |
1289 | out: |
1290 | free_trial_cpuset(trialcs); | |
1291 | return err; | |
1da177e4 LT |
1292 | } |
1293 | ||
3e0d98b9 | 1294 | /* |
80f7228b | 1295 | * Frequency meter - How fast is some event occurring? |
3e0d98b9 PJ |
1296 | * |
1297 | * These routines manage a digitally filtered, constant time based, | |
1298 | * event frequency meter. There are four routines: | |
1299 | * fmeter_init() - initialize a frequency meter. | |
1300 | * fmeter_markevent() - called each time the event happens. | |
1301 | * fmeter_getrate() - returns the recent rate of such events. | |
1302 | * fmeter_update() - internal routine used to update fmeter. | |
1303 | * | |
1304 | * A common data structure is passed to each of these routines, | |
1305 | * which is used to keep track of the state required to manage the | |
1306 | * frequency meter and its digital filter. | |
1307 | * | |
1308 | * The filter works on the number of events marked per unit time. | |
1309 | * The filter is single-pole low-pass recursive (IIR). The time unit | |
1310 | * is 1 second. Arithmetic is done using 32-bit integers scaled to | |
1311 | * simulate 3 decimal digits of precision (multiplied by 1000). | |
1312 | * | |
1313 | * With an FM_COEF of 933, and a time base of 1 second, the filter | |
1314 | * has a half-life of 10 seconds, meaning that if the events quit | |
1315 | * happening, then the rate returned from the fmeter_getrate() | |
1316 | * will be cut in half each 10 seconds, until it converges to zero. | |
1317 | * | |
1318 | * It is not worth doing a real infinitely recursive filter. If more | |
1319 | * than FM_MAXTICKS ticks have elapsed since the last filter event, | |
1320 | * just compute FM_MAXTICKS ticks worth, by which point the level | |
1321 | * will be stable. | |
1322 | * | |
1323 | * Limit the count of unprocessed events to FM_MAXCNT, so as to avoid | |
1324 | * arithmetic overflow in the fmeter_update() routine. | |
1325 | * | |
1326 | * Given the simple 32 bit integer arithmetic used, this meter works | |
1327 | * best for reporting rates between one per millisecond (msec) and | |
1328 | * one per 32 (approx) seconds. At constant rates faster than one | |
1329 | * per msec it maxes out at values just under 1,000,000. At constant | |
1330 | * rates between one per msec, and one per second it will stabilize | |
1331 | * to a value N*1000, where N is the rate of events per second. | |
1332 | * At constant rates between one per second and one per 32 seconds, | |
1333 | * it will be choppy, moving up on the seconds that have an event, | |
1334 | * and then decaying until the next event. At rates slower than | |
1335 | * about one in 32 seconds, it decays all the way back to zero between | |
1336 | * each event. | |
1337 | */ | |
1338 | ||
1339 | #define FM_COEF 933 /* coefficient for half-life of 10 secs */ | |
1340 | #define FM_MAXTICKS ((time_t)99) /* useless computing more ticks than this */ | |
1341 | #define FM_MAXCNT 1000000 /* limit cnt to avoid overflow */ | |
1342 | #define FM_SCALE 1000 /* faux fixed point scale */ | |
1343 | ||
1344 | /* Initialize a frequency meter */ | |
1345 | static void fmeter_init(struct fmeter *fmp) | |
1346 | { | |
1347 | fmp->cnt = 0; | |
1348 | fmp->val = 0; | |
1349 | fmp->time = 0; | |
1350 | spin_lock_init(&fmp->lock); | |
1351 | } | |
1352 | ||
1353 | /* Internal meter update - process cnt events and update value */ | |
1354 | static void fmeter_update(struct fmeter *fmp) | |
1355 | { | |
1356 | time_t now = get_seconds(); | |
1357 | time_t ticks = now - fmp->time; | |
1358 | ||
1359 | if (ticks == 0) | |
1360 | return; | |
1361 | ||
1362 | ticks = min(FM_MAXTICKS, ticks); | |
1363 | while (ticks-- > 0) | |
1364 | fmp->val = (FM_COEF * fmp->val) / FM_SCALE; | |
1365 | fmp->time = now; | |
1366 | ||
1367 | fmp->val += ((FM_SCALE - FM_COEF) * fmp->cnt) / FM_SCALE; | |
1368 | fmp->cnt = 0; | |
1369 | } | |
1370 | ||
1371 | /* Process any previous ticks, then bump cnt by one (times scale). */ | |
1372 | static void fmeter_markevent(struct fmeter *fmp) | |
1373 | { | |
1374 | spin_lock(&fmp->lock); | |
1375 | fmeter_update(fmp); | |
1376 | fmp->cnt = min(FM_MAXCNT, fmp->cnt + FM_SCALE); | |
1377 | spin_unlock(&fmp->lock); | |
1378 | } | |
1379 | ||
1380 | /* Process any previous ticks, then return current value. */ | |
1381 | static int fmeter_getrate(struct fmeter *fmp) | |
1382 | { | |
1383 | int val; | |
1384 | ||
1385 | spin_lock(&fmp->lock); | |
1386 | fmeter_update(fmp); | |
1387 | val = fmp->val; | |
1388 | spin_unlock(&fmp->lock); | |
1389 | return val; | |
1390 | } | |
1391 | ||
2df167a3 | 1392 | /* Called by cgroups to determine if a cpuset is usable; cgroup_mutex held */ |
761b3ef5 | 1393 | static int cpuset_can_attach(struct cgroup *cgrp, struct cgroup_taskset *tset) |
f780bdb7 | 1394 | { |
2f7ee569 | 1395 | struct cpuset *cs = cgroup_cs(cgrp); |
bb9d97b6 TH |
1396 | struct task_struct *task; |
1397 | int ret; | |
1da177e4 | 1398 | |
300ed6cb | 1399 | if (cpumask_empty(cs->cpus_allowed) || nodes_empty(cs->mems_allowed)) |
1da177e4 | 1400 | return -ENOSPC; |
9985b0ba | 1401 | |
bb9d97b6 TH |
1402 | cgroup_taskset_for_each(task, cgrp, tset) { |
1403 | /* | |
1404 | * Kthreads bound to specific cpus cannot be moved to a new | |
1405 | * cpuset; we cannot change their cpu affinity and | |
1406 | * isolating such threads by their set of allowed nodes is | |
1407 | * unnecessary. Thus, cpusets are not applicable for such | |
1408 | * threads. This prevents checking for success of | |
1409 | * set_cpus_allowed_ptr() on all attached tasks before | |
1410 | * cpus_allowed may be changed. | |
1411 | */ | |
1412 | if (task->flags & PF_THREAD_BOUND) | |
1413 | return -EINVAL; | |
1414 | if ((ret = security_task_setscheduler(task))) | |
1415 | return ret; | |
1416 | } | |
f780bdb7 | 1417 | |
94196f51 | 1418 | return 0; |
8793d854 | 1419 | } |
1da177e4 | 1420 | |
4e4c9a14 TH |
1421 | /* |
1422 | * Protected by cgroup_mutex. cpus_attach is used only by cpuset_attach() | |
1423 | * but we can't allocate it dynamically there. Define it global and | |
1424 | * allocate from cpuset_init(). | |
1425 | */ | |
1426 | static cpumask_var_t cpus_attach; | |
1427 | ||
761b3ef5 | 1428 | static void cpuset_attach(struct cgroup *cgrp, struct cgroup_taskset *tset) |
8793d854 | 1429 | { |
4e4c9a14 TH |
1430 | /* static bufs protected by cgroup_mutex */ |
1431 | static nodemask_t cpuset_attach_nodemask_from; | |
1432 | static nodemask_t cpuset_attach_nodemask_to; | |
8793d854 | 1433 | struct mm_struct *mm; |
bb9d97b6 TH |
1434 | struct task_struct *task; |
1435 | struct task_struct *leader = cgroup_taskset_first(tset); | |
2f7ee569 TH |
1436 | struct cgroup *oldcgrp = cgroup_taskset_cur_cgroup(tset); |
1437 | struct cpuset *cs = cgroup_cs(cgrp); | |
1438 | struct cpuset *oldcs = cgroup_cs(oldcgrp); | |
22fb52dd | 1439 | |
4e4c9a14 TH |
1440 | /* prepare for attach */ |
1441 | if (cs == &top_cpuset) | |
1442 | cpumask_copy(cpus_attach, cpu_possible_mask); | |
1443 | else | |
1444 | guarantee_online_cpus(cs, cpus_attach); | |
1445 | ||
1446 | guarantee_online_mems(cs, &cpuset_attach_nodemask_to); | |
1447 | ||
bb9d97b6 TH |
1448 | cgroup_taskset_for_each(task, cgrp, tset) { |
1449 | /* | |
1450 | * can_attach beforehand should guarantee that this doesn't | |
1451 | * fail. TODO: have a better way to handle failure here | |
1452 | */ | |
1453 | WARN_ON_ONCE(set_cpus_allowed_ptr(task, cpus_attach)); | |
1454 | ||
1455 | cpuset_change_task_nodemask(task, &cpuset_attach_nodemask_to); | |
1456 | cpuset_update_task_spread_flag(cs, task); | |
1457 | } | |
22fb52dd | 1458 | |
f780bdb7 BB |
1459 | /* |
1460 | * Change mm, possibly for multiple threads in a threadgroup. This is | |
1461 | * expensive and may sleep. | |
1462 | */ | |
1463 | cpuset_attach_nodemask_from = oldcs->mems_allowed; | |
1464 | cpuset_attach_nodemask_to = cs->mems_allowed; | |
bb9d97b6 | 1465 | mm = get_task_mm(leader); |
4225399a | 1466 | if (mm) { |
f780bdb7 | 1467 | mpol_rebind_mm(mm, &cpuset_attach_nodemask_to); |
2741a559 | 1468 | if (is_memory_migrate(cs)) |
f780bdb7 BB |
1469 | cpuset_migrate_mm(mm, &cpuset_attach_nodemask_from, |
1470 | &cpuset_attach_nodemask_to); | |
4225399a PJ |
1471 | mmput(mm); |
1472 | } | |
1da177e4 LT |
1473 | } |
1474 | ||
1475 | /* The various types of files and directories in a cpuset file system */ | |
1476 | ||
1477 | typedef enum { | |
45b07ef3 | 1478 | FILE_MEMORY_MIGRATE, |
1da177e4 LT |
1479 | FILE_CPULIST, |
1480 | FILE_MEMLIST, | |
1481 | FILE_CPU_EXCLUSIVE, | |
1482 | FILE_MEM_EXCLUSIVE, | |
78608366 | 1483 | FILE_MEM_HARDWALL, |
029190c5 | 1484 | FILE_SCHED_LOAD_BALANCE, |
1d3504fc | 1485 | FILE_SCHED_RELAX_DOMAIN_LEVEL, |
3e0d98b9 PJ |
1486 | FILE_MEMORY_PRESSURE_ENABLED, |
1487 | FILE_MEMORY_PRESSURE, | |
825a46af PJ |
1488 | FILE_SPREAD_PAGE, |
1489 | FILE_SPREAD_SLAB, | |
1da177e4 LT |
1490 | } cpuset_filetype_t; |
1491 | ||
700fe1ab PM |
1492 | static int cpuset_write_u64(struct cgroup *cgrp, struct cftype *cft, u64 val) |
1493 | { | |
1494 | int retval = 0; | |
1495 | struct cpuset *cs = cgroup_cs(cgrp); | |
1496 | cpuset_filetype_t type = cft->private; | |
1497 | ||
e3712395 | 1498 | if (!cgroup_lock_live_group(cgrp)) |
700fe1ab | 1499 | return -ENODEV; |
700fe1ab PM |
1500 | |
1501 | switch (type) { | |
1da177e4 | 1502 | case FILE_CPU_EXCLUSIVE: |
700fe1ab | 1503 | retval = update_flag(CS_CPU_EXCLUSIVE, cs, val); |
1da177e4 LT |
1504 | break; |
1505 | case FILE_MEM_EXCLUSIVE: | |
700fe1ab | 1506 | retval = update_flag(CS_MEM_EXCLUSIVE, cs, val); |
1da177e4 | 1507 | break; |
78608366 PM |
1508 | case FILE_MEM_HARDWALL: |
1509 | retval = update_flag(CS_MEM_HARDWALL, cs, val); | |
1510 | break; | |
029190c5 | 1511 | case FILE_SCHED_LOAD_BALANCE: |
700fe1ab | 1512 | retval = update_flag(CS_SCHED_LOAD_BALANCE, cs, val); |
1d3504fc | 1513 | break; |
45b07ef3 | 1514 | case FILE_MEMORY_MIGRATE: |
700fe1ab | 1515 | retval = update_flag(CS_MEMORY_MIGRATE, cs, val); |
45b07ef3 | 1516 | break; |
3e0d98b9 | 1517 | case FILE_MEMORY_PRESSURE_ENABLED: |
700fe1ab | 1518 | cpuset_memory_pressure_enabled = !!val; |
3e0d98b9 PJ |
1519 | break; |
1520 | case FILE_MEMORY_PRESSURE: | |
1521 | retval = -EACCES; | |
1522 | break; | |
825a46af | 1523 | case FILE_SPREAD_PAGE: |
700fe1ab | 1524 | retval = update_flag(CS_SPREAD_PAGE, cs, val); |
825a46af PJ |
1525 | break; |
1526 | case FILE_SPREAD_SLAB: | |
700fe1ab | 1527 | retval = update_flag(CS_SPREAD_SLAB, cs, val); |
825a46af | 1528 | break; |
1da177e4 LT |
1529 | default: |
1530 | retval = -EINVAL; | |
700fe1ab | 1531 | break; |
1da177e4 | 1532 | } |
8793d854 | 1533 | cgroup_unlock(); |
1da177e4 LT |
1534 | return retval; |
1535 | } | |
1536 | ||
5be7a479 PM |
1537 | static int cpuset_write_s64(struct cgroup *cgrp, struct cftype *cft, s64 val) |
1538 | { | |
1539 | int retval = 0; | |
1540 | struct cpuset *cs = cgroup_cs(cgrp); | |
1541 | cpuset_filetype_t type = cft->private; | |
1542 | ||
e3712395 | 1543 | if (!cgroup_lock_live_group(cgrp)) |
5be7a479 | 1544 | return -ENODEV; |
e3712395 | 1545 | |
5be7a479 PM |
1546 | switch (type) { |
1547 | case FILE_SCHED_RELAX_DOMAIN_LEVEL: | |
1548 | retval = update_relax_domain_level(cs, val); | |
1549 | break; | |
1550 | default: | |
1551 | retval = -EINVAL; | |
1552 | break; | |
1553 | } | |
1554 | cgroup_unlock(); | |
1555 | return retval; | |
1556 | } | |
1557 | ||
e3712395 PM |
1558 | /* |
1559 | * Common handling for a write to a "cpus" or "mems" file. | |
1560 | */ | |
1561 | static int cpuset_write_resmask(struct cgroup *cgrp, struct cftype *cft, | |
1562 | const char *buf) | |
1563 | { | |
1564 | int retval = 0; | |
645fcc9d LZ |
1565 | struct cpuset *cs = cgroup_cs(cgrp); |
1566 | struct cpuset *trialcs; | |
e3712395 PM |
1567 | |
1568 | if (!cgroup_lock_live_group(cgrp)) | |
1569 | return -ENODEV; | |
1570 | ||
645fcc9d | 1571 | trialcs = alloc_trial_cpuset(cs); |
b75f38d6 LZ |
1572 | if (!trialcs) { |
1573 | retval = -ENOMEM; | |
1574 | goto out; | |
1575 | } | |
645fcc9d | 1576 | |
e3712395 PM |
1577 | switch (cft->private) { |
1578 | case FILE_CPULIST: | |
645fcc9d | 1579 | retval = update_cpumask(cs, trialcs, buf); |
e3712395 PM |
1580 | break; |
1581 | case FILE_MEMLIST: | |
645fcc9d | 1582 | retval = update_nodemask(cs, trialcs, buf); |
e3712395 PM |
1583 | break; |
1584 | default: | |
1585 | retval = -EINVAL; | |
1586 | break; | |
1587 | } | |
645fcc9d LZ |
1588 | |
1589 | free_trial_cpuset(trialcs); | |
b75f38d6 | 1590 | out: |
e3712395 PM |
1591 | cgroup_unlock(); |
1592 | return retval; | |
1593 | } | |
1594 | ||
1da177e4 LT |
1595 | /* |
1596 | * These ascii lists should be read in a single call, by using a user | |
1597 | * buffer large enough to hold the entire map. If read in smaller | |
1598 | * chunks, there is no guarantee of atomicity. Since the display format | |
1599 | * used, list of ranges of sequential numbers, is variable length, | |
1600 | * and since these maps can change value dynamically, one could read | |
1601 | * gibberish by doing partial reads while a list was changing. | |
1602 | * A single large read to a buffer that crosses a page boundary is | |
1603 | * ok, because the result being copied to user land is not recomputed | |
1604 | * across a page fault. | |
1605 | */ | |
1606 | ||
9303e0c4 | 1607 | static size_t cpuset_sprintf_cpulist(char *page, struct cpuset *cs) |
1da177e4 | 1608 | { |
9303e0c4 | 1609 | size_t count; |
1da177e4 | 1610 | |
3d3f26a7 | 1611 | mutex_lock(&callback_mutex); |
9303e0c4 | 1612 | count = cpulist_scnprintf(page, PAGE_SIZE, cs->cpus_allowed); |
3d3f26a7 | 1613 | mutex_unlock(&callback_mutex); |
1da177e4 | 1614 | |
9303e0c4 | 1615 | return count; |
1da177e4 LT |
1616 | } |
1617 | ||
9303e0c4 | 1618 | static size_t cpuset_sprintf_memlist(char *page, struct cpuset *cs) |
1da177e4 | 1619 | { |
9303e0c4 | 1620 | size_t count; |
1da177e4 | 1621 | |
3d3f26a7 | 1622 | mutex_lock(&callback_mutex); |
9303e0c4 | 1623 | count = nodelist_scnprintf(page, PAGE_SIZE, cs->mems_allowed); |
3d3f26a7 | 1624 | mutex_unlock(&callback_mutex); |
1da177e4 | 1625 | |
9303e0c4 | 1626 | return count; |
1da177e4 LT |
1627 | } |
1628 | ||
8793d854 PM |
1629 | static ssize_t cpuset_common_file_read(struct cgroup *cont, |
1630 | struct cftype *cft, | |
1631 | struct file *file, | |
1632 | char __user *buf, | |
1633 | size_t nbytes, loff_t *ppos) | |
1da177e4 | 1634 | { |
8793d854 | 1635 | struct cpuset *cs = cgroup_cs(cont); |
1da177e4 LT |
1636 | cpuset_filetype_t type = cft->private; |
1637 | char *page; | |
1638 | ssize_t retval = 0; | |
1639 | char *s; | |
1da177e4 | 1640 | |
e12ba74d | 1641 | if (!(page = (char *)__get_free_page(GFP_TEMPORARY))) |
1da177e4 LT |
1642 | return -ENOMEM; |
1643 | ||
1644 | s = page; | |
1645 | ||
1646 | switch (type) { | |
1647 | case FILE_CPULIST: | |
1648 | s += cpuset_sprintf_cpulist(s, cs); | |
1649 | break; | |
1650 | case FILE_MEMLIST: | |
1651 | s += cpuset_sprintf_memlist(s, cs); | |
1652 | break; | |
1da177e4 LT |
1653 | default: |
1654 | retval = -EINVAL; | |
1655 | goto out; | |
1656 | } | |
1657 | *s++ = '\n'; | |
1da177e4 | 1658 | |
eacaa1f5 | 1659 | retval = simple_read_from_buffer(buf, nbytes, ppos, page, s - page); |
1da177e4 LT |
1660 | out: |
1661 | free_page((unsigned long)page); | |
1662 | return retval; | |
1663 | } | |
1664 | ||
700fe1ab PM |
1665 | static u64 cpuset_read_u64(struct cgroup *cont, struct cftype *cft) |
1666 | { | |
1667 | struct cpuset *cs = cgroup_cs(cont); | |
1668 | cpuset_filetype_t type = cft->private; | |
1669 | switch (type) { | |
1670 | case FILE_CPU_EXCLUSIVE: | |
1671 | return is_cpu_exclusive(cs); | |
1672 | case FILE_MEM_EXCLUSIVE: | |
1673 | return is_mem_exclusive(cs); | |
78608366 PM |
1674 | case FILE_MEM_HARDWALL: |
1675 | return is_mem_hardwall(cs); | |
700fe1ab PM |
1676 | case FILE_SCHED_LOAD_BALANCE: |
1677 | return is_sched_load_balance(cs); | |
1678 | case FILE_MEMORY_MIGRATE: | |
1679 | return is_memory_migrate(cs); | |
1680 | case FILE_MEMORY_PRESSURE_ENABLED: | |
1681 | return cpuset_memory_pressure_enabled; | |
1682 | case FILE_MEMORY_PRESSURE: | |
1683 | return fmeter_getrate(&cs->fmeter); | |
1684 | case FILE_SPREAD_PAGE: | |
1685 | return is_spread_page(cs); | |
1686 | case FILE_SPREAD_SLAB: | |
1687 | return is_spread_slab(cs); | |
1688 | default: | |
1689 | BUG(); | |
1690 | } | |
cf417141 MK |
1691 | |
1692 | /* Unreachable but makes gcc happy */ | |
1693 | return 0; | |
700fe1ab | 1694 | } |
1da177e4 | 1695 | |
5be7a479 PM |
1696 | static s64 cpuset_read_s64(struct cgroup *cont, struct cftype *cft) |
1697 | { | |
1698 | struct cpuset *cs = cgroup_cs(cont); | |
1699 | cpuset_filetype_t type = cft->private; | |
1700 | switch (type) { | |
1701 | case FILE_SCHED_RELAX_DOMAIN_LEVEL: | |
1702 | return cs->relax_domain_level; | |
1703 | default: | |
1704 | BUG(); | |
1705 | } | |
cf417141 MK |
1706 | |
1707 | /* Unrechable but makes gcc happy */ | |
1708 | return 0; | |
5be7a479 PM |
1709 | } |
1710 | ||
1da177e4 LT |
1711 | |
1712 | /* | |
1713 | * for the common functions, 'private' gives the type of file | |
1714 | */ | |
1715 | ||
addf2c73 PM |
1716 | static struct cftype files[] = { |
1717 | { | |
1718 | .name = "cpus", | |
1719 | .read = cpuset_common_file_read, | |
e3712395 PM |
1720 | .write_string = cpuset_write_resmask, |
1721 | .max_write_len = (100U + 6 * NR_CPUS), | |
addf2c73 PM |
1722 | .private = FILE_CPULIST, |
1723 | }, | |
1724 | ||
1725 | { | |
1726 | .name = "mems", | |
1727 | .read = cpuset_common_file_read, | |
e3712395 PM |
1728 | .write_string = cpuset_write_resmask, |
1729 | .max_write_len = (100U + 6 * MAX_NUMNODES), | |
addf2c73 PM |
1730 | .private = FILE_MEMLIST, |
1731 | }, | |
1732 | ||
1733 | { | |
1734 | .name = "cpu_exclusive", | |
1735 | .read_u64 = cpuset_read_u64, | |
1736 | .write_u64 = cpuset_write_u64, | |
1737 | .private = FILE_CPU_EXCLUSIVE, | |
1738 | }, | |
1739 | ||
1740 | { | |
1741 | .name = "mem_exclusive", | |
1742 | .read_u64 = cpuset_read_u64, | |
1743 | .write_u64 = cpuset_write_u64, | |
1744 | .private = FILE_MEM_EXCLUSIVE, | |
1745 | }, | |
1746 | ||
78608366 PM |
1747 | { |
1748 | .name = "mem_hardwall", | |
1749 | .read_u64 = cpuset_read_u64, | |
1750 | .write_u64 = cpuset_write_u64, | |
1751 | .private = FILE_MEM_HARDWALL, | |
1752 | }, | |
1753 | ||
addf2c73 PM |
1754 | { |
1755 | .name = "sched_load_balance", | |
1756 | .read_u64 = cpuset_read_u64, | |
1757 | .write_u64 = cpuset_write_u64, | |
1758 | .private = FILE_SCHED_LOAD_BALANCE, | |
1759 | }, | |
1760 | ||
1761 | { | |
1762 | .name = "sched_relax_domain_level", | |
5be7a479 PM |
1763 | .read_s64 = cpuset_read_s64, |
1764 | .write_s64 = cpuset_write_s64, | |
addf2c73 PM |
1765 | .private = FILE_SCHED_RELAX_DOMAIN_LEVEL, |
1766 | }, | |
1767 | ||
1768 | { | |
1769 | .name = "memory_migrate", | |
1770 | .read_u64 = cpuset_read_u64, | |
1771 | .write_u64 = cpuset_write_u64, | |
1772 | .private = FILE_MEMORY_MIGRATE, | |
1773 | }, | |
1774 | ||
1775 | { | |
1776 | .name = "memory_pressure", | |
1777 | .read_u64 = cpuset_read_u64, | |
1778 | .write_u64 = cpuset_write_u64, | |
1779 | .private = FILE_MEMORY_PRESSURE, | |
099fca32 | 1780 | .mode = S_IRUGO, |
addf2c73 PM |
1781 | }, |
1782 | ||
1783 | { | |
1784 | .name = "memory_spread_page", | |
1785 | .read_u64 = cpuset_read_u64, | |
1786 | .write_u64 = cpuset_write_u64, | |
1787 | .private = FILE_SPREAD_PAGE, | |
1788 | }, | |
1789 | ||
1790 | { | |
1791 | .name = "memory_spread_slab", | |
1792 | .read_u64 = cpuset_read_u64, | |
1793 | .write_u64 = cpuset_write_u64, | |
1794 | .private = FILE_SPREAD_SLAB, | |
1795 | }, | |
3e0d98b9 | 1796 | |
4baf6e33 TH |
1797 | { |
1798 | .name = "memory_pressure_enabled", | |
1799 | .flags = CFTYPE_ONLY_ON_ROOT, | |
1800 | .read_u64 = cpuset_read_u64, | |
1801 | .write_u64 = cpuset_write_u64, | |
1802 | .private = FILE_MEMORY_PRESSURE_ENABLED, | |
1803 | }, | |
1da177e4 | 1804 | |
4baf6e33 TH |
1805 | { } /* terminate */ |
1806 | }; | |
1da177e4 LT |
1807 | |
1808 | /* | |
92fb9748 | 1809 | * cpuset_css_alloc - allocate a cpuset css |
2df167a3 | 1810 | * cont: control group that the new cpuset will be part of |
1da177e4 LT |
1811 | */ |
1812 | ||
92fb9748 | 1813 | static struct cgroup_subsys_state *cpuset_css_alloc(struct cgroup *cont) |
1da177e4 | 1814 | { |
c8f699bb | 1815 | struct cpuset *cs; |
1da177e4 | 1816 | |
c8f699bb | 1817 | if (!cont->parent) |
8793d854 | 1818 | return &top_cpuset.css; |
033fa1c5 | 1819 | |
c8f699bb | 1820 | cs = kzalloc(sizeof(*cs), GFP_KERNEL); |
1da177e4 | 1821 | if (!cs) |
8793d854 | 1822 | return ERR_PTR(-ENOMEM); |
300ed6cb LZ |
1823 | if (!alloc_cpumask_var(&cs->cpus_allowed, GFP_KERNEL)) { |
1824 | kfree(cs); | |
1825 | return ERR_PTR(-ENOMEM); | |
1826 | } | |
1da177e4 | 1827 | |
029190c5 | 1828 | set_bit(CS_SCHED_LOAD_BALANCE, &cs->flags); |
300ed6cb | 1829 | cpumask_clear(cs->cpus_allowed); |
f9a86fcb | 1830 | nodes_clear(cs->mems_allowed); |
3e0d98b9 | 1831 | fmeter_init(&cs->fmeter); |
1d3504fc | 1832 | cs->relax_domain_level = -1; |
c8f699bb TH |
1833 | cs->parent = cgroup_cs(cont->parent); |
1834 | ||
1835 | return &cs->css; | |
1836 | } | |
1837 | ||
1838 | static int cpuset_css_online(struct cgroup *cgrp) | |
1839 | { | |
1840 | struct cpuset *cs = cgroup_cs(cgrp); | |
1841 | struct cpuset *parent = cs->parent; | |
ae8086ce TH |
1842 | struct cpuset *tmp_cs; |
1843 | struct cgroup *pos_cg; | |
c8f699bb TH |
1844 | |
1845 | if (!parent) | |
1846 | return 0; | |
1847 | ||
efeb77b2 | 1848 | set_bit(CS_ONLINE, &cs->flags); |
c8f699bb TH |
1849 | if (is_spread_page(parent)) |
1850 | set_bit(CS_SPREAD_PAGE, &cs->flags); | |
1851 | if (is_spread_slab(parent)) | |
1852 | set_bit(CS_SPREAD_SLAB, &cs->flags); | |
1da177e4 | 1853 | |
202f72d5 | 1854 | number_of_cpusets++; |
033fa1c5 | 1855 | |
c8f699bb TH |
1856 | if (!test_bit(CGRP_CPUSET_CLONE_CHILDREN, &cgrp->flags)) |
1857 | return 0; | |
033fa1c5 TH |
1858 | |
1859 | /* | |
1860 | * Clone @parent's configuration if CGRP_CPUSET_CLONE_CHILDREN is | |
1861 | * set. This flag handling is implemented in cgroup core for | |
1862 | * histrical reasons - the flag may be specified during mount. | |
1863 | * | |
1864 | * Currently, if any sibling cpusets have exclusive cpus or mem, we | |
1865 | * refuse to clone the configuration - thereby refusing the task to | |
1866 | * be entered, and as a result refusing the sys_unshare() or | |
1867 | * clone() which initiated it. If this becomes a problem for some | |
1868 | * users who wish to allow that scenario, then this could be | |
1869 | * changed to grant parent->cpus_allowed-sibling_cpus_exclusive | |
1870 | * (and likewise for mems) to the new cgroup. | |
1871 | */ | |
ae8086ce TH |
1872 | rcu_read_lock(); |
1873 | cpuset_for_each_child(tmp_cs, pos_cg, parent) { | |
1874 | if (is_mem_exclusive(tmp_cs) || is_cpu_exclusive(tmp_cs)) { | |
1875 | rcu_read_unlock(); | |
c8f699bb | 1876 | return 0; |
ae8086ce | 1877 | } |
033fa1c5 | 1878 | } |
ae8086ce | 1879 | rcu_read_unlock(); |
033fa1c5 TH |
1880 | |
1881 | mutex_lock(&callback_mutex); | |
1882 | cs->mems_allowed = parent->mems_allowed; | |
1883 | cpumask_copy(cs->cpus_allowed, parent->cpus_allowed); | |
1884 | mutex_unlock(&callback_mutex); | |
c8f699bb TH |
1885 | |
1886 | return 0; | |
1887 | } | |
1888 | ||
1889 | static void cpuset_css_offline(struct cgroup *cgrp) | |
1890 | { | |
1891 | struct cpuset *cs = cgroup_cs(cgrp); | |
1892 | ||
1893 | /* css_offline is called w/o cgroup_mutex, grab it */ | |
1894 | cgroup_lock(); | |
1895 | ||
1896 | if (is_sched_load_balance(cs)) | |
1897 | update_flag(CS_SCHED_LOAD_BALANCE, cs, 0); | |
1898 | ||
1899 | number_of_cpusets--; | |
efeb77b2 | 1900 | clear_bit(CS_ONLINE, &cs->flags); |
c8f699bb TH |
1901 | |
1902 | cgroup_unlock(); | |
1da177e4 LT |
1903 | } |
1904 | ||
029190c5 | 1905 | /* |
029190c5 PJ |
1906 | * If the cpuset being removed has its flag 'sched_load_balance' |
1907 | * enabled, then simulate turning sched_load_balance off, which | |
cf417141 | 1908 | * will call async_rebuild_sched_domains(). |
029190c5 PJ |
1909 | */ |
1910 | ||
92fb9748 | 1911 | static void cpuset_css_free(struct cgroup *cont) |
1da177e4 | 1912 | { |
8793d854 | 1913 | struct cpuset *cs = cgroup_cs(cont); |
1da177e4 | 1914 | |
300ed6cb | 1915 | free_cpumask_var(cs->cpus_allowed); |
8793d854 | 1916 | kfree(cs); |
1da177e4 LT |
1917 | } |
1918 | ||
8793d854 PM |
1919 | struct cgroup_subsys cpuset_subsys = { |
1920 | .name = "cpuset", | |
92fb9748 | 1921 | .css_alloc = cpuset_css_alloc, |
c8f699bb TH |
1922 | .css_online = cpuset_css_online, |
1923 | .css_offline = cpuset_css_offline, | |
92fb9748 | 1924 | .css_free = cpuset_css_free, |
8793d854 PM |
1925 | .can_attach = cpuset_can_attach, |
1926 | .attach = cpuset_attach, | |
8793d854 | 1927 | .subsys_id = cpuset_subsys_id, |
4baf6e33 | 1928 | .base_cftypes = files, |
8793d854 PM |
1929 | .early_init = 1, |
1930 | }; | |
1931 | ||
1da177e4 LT |
1932 | /** |
1933 | * cpuset_init - initialize cpusets at system boot | |
1934 | * | |
1935 | * Description: Initialize top_cpuset and the cpuset internal file system, | |
1936 | **/ | |
1937 | ||
1938 | int __init cpuset_init(void) | |
1939 | { | |
8793d854 | 1940 | int err = 0; |
1da177e4 | 1941 | |
58568d2a MX |
1942 | if (!alloc_cpumask_var(&top_cpuset.cpus_allowed, GFP_KERNEL)) |
1943 | BUG(); | |
1944 | ||
300ed6cb | 1945 | cpumask_setall(top_cpuset.cpus_allowed); |
f9a86fcb | 1946 | nodes_setall(top_cpuset.mems_allowed); |
1da177e4 | 1947 | |
3e0d98b9 | 1948 | fmeter_init(&top_cpuset.fmeter); |
029190c5 | 1949 | set_bit(CS_SCHED_LOAD_BALANCE, &top_cpuset.flags); |
1d3504fc | 1950 | top_cpuset.relax_domain_level = -1; |
1da177e4 | 1951 | |
1da177e4 LT |
1952 | err = register_filesystem(&cpuset_fs_type); |
1953 | if (err < 0) | |
8793d854 PM |
1954 | return err; |
1955 | ||
2341d1b6 LZ |
1956 | if (!alloc_cpumask_var(&cpus_attach, GFP_KERNEL)) |
1957 | BUG(); | |
1958 | ||
202f72d5 | 1959 | number_of_cpusets = 1; |
8793d854 | 1960 | return 0; |
1da177e4 LT |
1961 | } |
1962 | ||
956db3ca CW |
1963 | /** |
1964 | * cpuset_do_move_task - move a given task to another cpuset | |
1965 | * @tsk: pointer to task_struct the task to move | |
1966 | * @scan: struct cgroup_scanner contained in its struct cpuset_hotplug_scanner | |
1967 | * | |
1968 | * Called by cgroup_scan_tasks() for each task in a cgroup. | |
1969 | * Return nonzero to stop the walk through the tasks. | |
1970 | */ | |
9e0c914c AB |
1971 | static void cpuset_do_move_task(struct task_struct *tsk, |
1972 | struct cgroup_scanner *scan) | |
956db3ca | 1973 | { |
7f81b1ae | 1974 | struct cgroup *new_cgroup = scan->data; |
956db3ca | 1975 | |
7f81b1ae | 1976 | cgroup_attach_task(new_cgroup, tsk); |
956db3ca CW |
1977 | } |
1978 | ||
1979 | /** | |
1980 | * move_member_tasks_to_cpuset - move tasks from one cpuset to another | |
1981 | * @from: cpuset in which the tasks currently reside | |
1982 | * @to: cpuset to which the tasks will be moved | |
1983 | * | |
c8d9c90c PJ |
1984 | * Called with cgroup_mutex held |
1985 | * callback_mutex must not be held, as cpuset_attach() will take it. | |
956db3ca CW |
1986 | * |
1987 | * The cgroup_scan_tasks() function will scan all the tasks in a cgroup, | |
1988 | * calling callback functions for each. | |
1989 | */ | |
1990 | static void move_member_tasks_to_cpuset(struct cpuset *from, struct cpuset *to) | |
1991 | { | |
7f81b1ae | 1992 | struct cgroup_scanner scan; |
956db3ca | 1993 | |
7f81b1ae LZ |
1994 | scan.cg = from->css.cgroup; |
1995 | scan.test_task = NULL; /* select all tasks in cgroup */ | |
1996 | scan.process_task = cpuset_do_move_task; | |
1997 | scan.heap = NULL; | |
1998 | scan.data = to->css.cgroup; | |
956db3ca | 1999 | |
7f81b1ae | 2000 | if (cgroup_scan_tasks(&scan)) |
956db3ca CW |
2001 | printk(KERN_ERR "move_member_tasks_to_cpuset: " |
2002 | "cgroup_scan_tasks failed\n"); | |
2003 | } | |
2004 | ||
b1aac8bb | 2005 | /* |
cf417141 | 2006 | * If CPU and/or memory hotplug handlers, below, unplug any CPUs |
b1aac8bb PJ |
2007 | * or memory nodes, we need to walk over the cpuset hierarchy, |
2008 | * removing that CPU or node from all cpusets. If this removes the | |
956db3ca CW |
2009 | * last CPU or node from a cpuset, then move the tasks in the empty |
2010 | * cpuset to its next-highest non-empty parent. | |
b1aac8bb | 2011 | * |
c8d9c90c PJ |
2012 | * Called with cgroup_mutex held |
2013 | * callback_mutex must not be held, as cpuset_attach() will take it. | |
b1aac8bb | 2014 | */ |
956db3ca CW |
2015 | static void remove_tasks_in_empty_cpuset(struct cpuset *cs) |
2016 | { | |
2017 | struct cpuset *parent; | |
2018 | ||
956db3ca CW |
2019 | /* |
2020 | * Find its next-highest non-empty parent, (top cpuset | |
2021 | * has online cpus, so can't be empty). | |
2022 | */ | |
2023 | parent = cs->parent; | |
300ed6cb | 2024 | while (cpumask_empty(parent->cpus_allowed) || |
b4501295 | 2025 | nodes_empty(parent->mems_allowed)) |
956db3ca | 2026 | parent = parent->parent; |
956db3ca CW |
2027 | |
2028 | move_member_tasks_to_cpuset(cs, parent); | |
2029 | } | |
2030 | ||
80d1fa64 SB |
2031 | /* |
2032 | * Helper function to traverse cpusets. | |
2033 | * It can be used to walk the cpuset tree from top to bottom, completing | |
2034 | * one layer before dropping down to the next (thus always processing a | |
2035 | * node before any of its children). | |
2036 | */ | |
2037 | static struct cpuset *cpuset_next(struct list_head *queue) | |
2038 | { | |
2039 | struct cpuset *cp; | |
2040 | struct cpuset *child; /* scans child cpusets of cp */ | |
2041 | struct cgroup *cont; | |
2042 | ||
2043 | if (list_empty(queue)) | |
2044 | return NULL; | |
2045 | ||
2046 | cp = list_first_entry(queue, struct cpuset, stack_list); | |
2047 | list_del(queue->next); | |
ae8086ce TH |
2048 | rcu_read_lock(); |
2049 | cpuset_for_each_child(child, cont, cp) | |
80d1fa64 | 2050 | list_add_tail(&child->stack_list, queue); |
ae8086ce | 2051 | rcu_read_unlock(); |
80d1fa64 SB |
2052 | |
2053 | return cp; | |
2054 | } | |
2055 | ||
deb7aa30 TH |
2056 | /** |
2057 | * cpuset_propagate_hotplug - propagate CPU/memory hotplug to a cpuset | |
2058 | * @cs: cpuset in interest | |
956db3ca | 2059 | * |
deb7aa30 TH |
2060 | * Compare @cs's cpu and mem masks against top_cpuset and if some have gone |
2061 | * offline, update @cs accordingly. If @cs ends up with no CPU or memory, | |
2062 | * all its tasks are moved to the nearest ancestor with both resources. | |
956db3ca | 2063 | * |
deb7aa30 | 2064 | * Should be called with cgroup_mutex held. |
956db3ca | 2065 | */ |
deb7aa30 | 2066 | static void cpuset_propagate_hotplug(struct cpuset *cs) |
b1aac8bb | 2067 | { |
deb7aa30 TH |
2068 | static cpumask_t off_cpus; |
2069 | static nodemask_t off_mems, tmp_mems; | |
7ddf96b0 | 2070 | |
deb7aa30 | 2071 | WARN_ON_ONCE(!cgroup_lock_is_held()); |
7ddf96b0 | 2072 | |
deb7aa30 TH |
2073 | cpumask_andnot(&off_cpus, cs->cpus_allowed, top_cpuset.cpus_allowed); |
2074 | nodes_andnot(off_mems, cs->mems_allowed, top_cpuset.mems_allowed); | |
7ddf96b0 | 2075 | |
deb7aa30 TH |
2076 | /* remove offline cpus from @cs */ |
2077 | if (!cpumask_empty(&off_cpus)) { | |
2078 | mutex_lock(&callback_mutex); | |
2079 | cpumask_andnot(cs->cpus_allowed, cs->cpus_allowed, &off_cpus); | |
2080 | mutex_unlock(&callback_mutex); | |
2081 | update_tasks_cpumask(cs, NULL); | |
2082 | } | |
b4501295 | 2083 | |
deb7aa30 TH |
2084 | /* remove offline mems from @cs */ |
2085 | if (!nodes_empty(off_mems)) { | |
2086 | tmp_mems = cs->mems_allowed; | |
2087 | mutex_lock(&callback_mutex); | |
2088 | nodes_andnot(cs->mems_allowed, cs->mems_allowed, off_mems); | |
2089 | mutex_unlock(&callback_mutex); | |
2090 | update_tasks_nodemask(cs, &tmp_mems, NULL); | |
b1aac8bb | 2091 | } |
deb7aa30 TH |
2092 | |
2093 | if (cpumask_empty(cs->cpus_allowed) || nodes_empty(cs->mems_allowed)) | |
2094 | remove_tasks_in_empty_cpuset(cs); | |
b1aac8bb PJ |
2095 | } |
2096 | ||
deb7aa30 TH |
2097 | /** |
2098 | * cpuset_handle_hotplug - handle CPU/memory hot[un]plug | |
d35be8ba | 2099 | * |
deb7aa30 TH |
2100 | * This function is called after either CPU or memory configuration has |
2101 | * changed and updates cpuset accordingly. The top_cpuset is always | |
2102 | * synchronized to cpu_active_mask and N_MEMORY, which is necessary in | |
2103 | * order to make cpusets transparent (of no affect) on systems that are | |
2104 | * actively using CPU hotplug but making no active use of cpusets. | |
cf417141 | 2105 | * |
deb7aa30 TH |
2106 | * Non-root cpusets are only affected by offlining. If any CPUs or memory |
2107 | * nodes have been taken down, cpuset_propagate_hotplug() is invoked on all | |
2108 | * descendants. | |
7ddf96b0 | 2109 | * |
deb7aa30 TH |
2110 | * Note that CPU offlining during suspend is ignored. We don't modify |
2111 | * cpusets across suspend/resume cycles at all. | |
4c4d50f7 | 2112 | */ |
deb7aa30 | 2113 | static void cpuset_handle_hotplug(void) |
4c4d50f7 | 2114 | { |
deb7aa30 TH |
2115 | static cpumask_t new_cpus, tmp_cpus; |
2116 | static nodemask_t new_mems, tmp_mems; | |
2117 | bool cpus_updated, mems_updated; | |
2118 | bool cpus_offlined, mems_offlined; | |
cf417141 | 2119 | |
cf417141 | 2120 | cgroup_lock(); |
7ddf96b0 | 2121 | |
deb7aa30 TH |
2122 | /* fetch the available cpus/mems and find out which changed how */ |
2123 | cpumask_copy(&new_cpus, cpu_active_mask); | |
2124 | new_mems = node_states[N_MEMORY]; | |
2125 | ||
2126 | cpus_updated = !cpumask_equal(top_cpuset.cpus_allowed, &new_cpus); | |
2127 | cpus_offlined = cpumask_andnot(&tmp_cpus, top_cpuset.cpus_allowed, | |
2128 | &new_cpus); | |
2129 | ||
2130 | mems_updated = !nodes_equal(top_cpuset.mems_allowed, new_mems); | |
2131 | nodes_andnot(tmp_mems, top_cpuset.mems_allowed, new_mems); | |
2132 | mems_offlined = !nodes_empty(tmp_mems); | |
2133 | ||
2134 | /* synchronize cpus_allowed to cpu_active_mask */ | |
2135 | if (cpus_updated) { | |
2136 | mutex_lock(&callback_mutex); | |
2137 | cpumask_copy(top_cpuset.cpus_allowed, &new_cpus); | |
2138 | mutex_unlock(&callback_mutex); | |
2139 | /* we don't mess with cpumasks of tasks in top_cpuset */ | |
2140 | } | |
2141 | ||
2142 | /* synchronize mems_allowed to N_MEMORY */ | |
2143 | if (mems_updated) { | |
2144 | tmp_mems = top_cpuset.mems_allowed; | |
2145 | mutex_lock(&callback_mutex); | |
2146 | top_cpuset.mems_allowed = new_mems; | |
2147 | mutex_unlock(&callback_mutex); | |
2148 | update_tasks_nodemask(&top_cpuset, &tmp_mems, NULL); | |
2149 | } | |
2150 | ||
2151 | /* if cpus or mems went down, we need to propagate to descendants */ | |
2152 | if (cpus_offlined || mems_offlined) { | |
2153 | struct cpuset *cs; | |
2154 | LIST_HEAD(queue); | |
2155 | ||
2156 | list_add_tail(&top_cpuset.stack_list, &queue); | |
2157 | while ((cs = cpuset_next(&queue))) | |
2158 | if (cs != &top_cpuset) | |
2159 | cpuset_propagate_hotplug(cs); | |
2160 | } | |
7ddf96b0 | 2161 | |
cf417141 MK |
2162 | cgroup_unlock(); |
2163 | ||
deb7aa30 TH |
2164 | /* rebuild sched domains if cpus_allowed has changed */ |
2165 | if (cpus_updated) { | |
2166 | struct sched_domain_attr *attr; | |
2167 | cpumask_var_t *doms; | |
2168 | int ndoms; | |
2169 | ||
2170 | cgroup_lock(); | |
2171 | ndoms = generate_sched_domains(&doms, &attr); | |
2172 | cgroup_unlock(); | |
2173 | ||
2174 | partition_sched_domains(ndoms, doms, attr); | |
2175 | } | |
2176 | } | |
2177 | ||
2178 | void cpuset_update_active_cpus(bool cpu_online) | |
2179 | { | |
2180 | cpuset_handle_hotplug(); | |
4c4d50f7 | 2181 | } |
4c4d50f7 | 2182 | |
b1aac8bb | 2183 | #ifdef CONFIG_MEMORY_HOTPLUG |
38837fc7 | 2184 | /* |
38d7bee9 LJ |
2185 | * Keep top_cpuset.mems_allowed tracking node_states[N_MEMORY]. |
2186 | * Call this routine anytime after node_states[N_MEMORY] changes. | |
a1cd2b13 | 2187 | * See cpuset_update_active_cpus() for CPU hotplug handling. |
38837fc7 | 2188 | */ |
f481891f MX |
2189 | static int cpuset_track_online_nodes(struct notifier_block *self, |
2190 | unsigned long action, void *arg) | |
38837fc7 | 2191 | { |
deb7aa30 | 2192 | cpuset_handle_hotplug(); |
f481891f | 2193 | return NOTIFY_OK; |
38837fc7 PJ |
2194 | } |
2195 | #endif | |
2196 | ||
1da177e4 LT |
2197 | /** |
2198 | * cpuset_init_smp - initialize cpus_allowed | |
2199 | * | |
2200 | * Description: Finish top cpuset after cpu, node maps are initialized | |
2201 | **/ | |
2202 | ||
2203 | void __init cpuset_init_smp(void) | |
2204 | { | |
6ad4c188 | 2205 | cpumask_copy(top_cpuset.cpus_allowed, cpu_active_mask); |
38d7bee9 | 2206 | top_cpuset.mems_allowed = node_states[N_MEMORY]; |
4c4d50f7 | 2207 | |
f481891f | 2208 | hotplug_memory_notifier(cpuset_track_online_nodes, 10); |
f90d4118 MX |
2209 | |
2210 | cpuset_wq = create_singlethread_workqueue("cpuset"); | |
2211 | BUG_ON(!cpuset_wq); | |
1da177e4 LT |
2212 | } |
2213 | ||
2214 | /** | |
1da177e4 LT |
2215 | * cpuset_cpus_allowed - return cpus_allowed mask from a tasks cpuset. |
2216 | * @tsk: pointer to task_struct from which to obtain cpuset->cpus_allowed. | |
6af866af | 2217 | * @pmask: pointer to struct cpumask variable to receive cpus_allowed set. |
1da177e4 | 2218 | * |
300ed6cb | 2219 | * Description: Returns the cpumask_var_t cpus_allowed of the cpuset |
1da177e4 | 2220 | * attached to the specified @tsk. Guaranteed to return some non-empty |
5f054e31 | 2221 | * subset of cpu_online_mask, even if this means going outside the |
1da177e4 LT |
2222 | * tasks cpuset. |
2223 | **/ | |
2224 | ||
6af866af | 2225 | void cpuset_cpus_allowed(struct task_struct *tsk, struct cpumask *pmask) |
1da177e4 | 2226 | { |
3d3f26a7 | 2227 | mutex_lock(&callback_mutex); |
909d75a3 | 2228 | task_lock(tsk); |
f9a86fcb | 2229 | guarantee_online_cpus(task_cs(tsk), pmask); |
909d75a3 | 2230 | task_unlock(tsk); |
897f0b3c | 2231 | mutex_unlock(&callback_mutex); |
1da177e4 LT |
2232 | } |
2233 | ||
2baab4e9 | 2234 | void cpuset_cpus_allowed_fallback(struct task_struct *tsk) |
9084bb82 ON |
2235 | { |
2236 | const struct cpuset *cs; | |
9084bb82 ON |
2237 | |
2238 | rcu_read_lock(); | |
2239 | cs = task_cs(tsk); | |
2240 | if (cs) | |
1e1b6c51 | 2241 | do_set_cpus_allowed(tsk, cs->cpus_allowed); |
9084bb82 ON |
2242 | rcu_read_unlock(); |
2243 | ||
2244 | /* | |
2245 | * We own tsk->cpus_allowed, nobody can change it under us. | |
2246 | * | |
2247 | * But we used cs && cs->cpus_allowed lockless and thus can | |
2248 | * race with cgroup_attach_task() or update_cpumask() and get | |
2249 | * the wrong tsk->cpus_allowed. However, both cases imply the | |
2250 | * subsequent cpuset_change_cpumask()->set_cpus_allowed_ptr() | |
2251 | * which takes task_rq_lock(). | |
2252 | * | |
2253 | * If we are called after it dropped the lock we must see all | |
2254 | * changes in tsk_cs()->cpus_allowed. Otherwise we can temporary | |
2255 | * set any mask even if it is not right from task_cs() pov, | |
2256 | * the pending set_cpus_allowed_ptr() will fix things. | |
2baab4e9 PZ |
2257 | * |
2258 | * select_fallback_rq() will fix things ups and set cpu_possible_mask | |
2259 | * if required. | |
9084bb82 | 2260 | */ |
9084bb82 ON |
2261 | } |
2262 | ||
1da177e4 LT |
2263 | void cpuset_init_current_mems_allowed(void) |
2264 | { | |
f9a86fcb | 2265 | nodes_setall(current->mems_allowed); |
1da177e4 LT |
2266 | } |
2267 | ||
909d75a3 PJ |
2268 | /** |
2269 | * cpuset_mems_allowed - return mems_allowed mask from a tasks cpuset. | |
2270 | * @tsk: pointer to task_struct from which to obtain cpuset->mems_allowed. | |
2271 | * | |
2272 | * Description: Returns the nodemask_t mems_allowed of the cpuset | |
2273 | * attached to the specified @tsk. Guaranteed to return some non-empty | |
38d7bee9 | 2274 | * subset of node_states[N_MEMORY], even if this means going outside the |
909d75a3 PJ |
2275 | * tasks cpuset. |
2276 | **/ | |
2277 | ||
2278 | nodemask_t cpuset_mems_allowed(struct task_struct *tsk) | |
2279 | { | |
2280 | nodemask_t mask; | |
2281 | ||
3d3f26a7 | 2282 | mutex_lock(&callback_mutex); |
909d75a3 | 2283 | task_lock(tsk); |
8793d854 | 2284 | guarantee_online_mems(task_cs(tsk), &mask); |
909d75a3 | 2285 | task_unlock(tsk); |
3d3f26a7 | 2286 | mutex_unlock(&callback_mutex); |
909d75a3 PJ |
2287 | |
2288 | return mask; | |
2289 | } | |
2290 | ||
d9fd8a6d | 2291 | /** |
19770b32 MG |
2292 | * cpuset_nodemask_valid_mems_allowed - check nodemask vs. curremt mems_allowed |
2293 | * @nodemask: the nodemask to be checked | |
d9fd8a6d | 2294 | * |
19770b32 | 2295 | * Are any of the nodes in the nodemask allowed in current->mems_allowed? |
1da177e4 | 2296 | */ |
19770b32 | 2297 | int cpuset_nodemask_valid_mems_allowed(nodemask_t *nodemask) |
1da177e4 | 2298 | { |
19770b32 | 2299 | return nodes_intersects(*nodemask, current->mems_allowed); |
1da177e4 LT |
2300 | } |
2301 | ||
9bf2229f | 2302 | /* |
78608366 PM |
2303 | * nearest_hardwall_ancestor() - Returns the nearest mem_exclusive or |
2304 | * mem_hardwall ancestor to the specified cpuset. Call holding | |
2305 | * callback_mutex. If no ancestor is mem_exclusive or mem_hardwall | |
2306 | * (an unusual configuration), then returns the root cpuset. | |
9bf2229f | 2307 | */ |
78608366 | 2308 | static const struct cpuset *nearest_hardwall_ancestor(const struct cpuset *cs) |
9bf2229f | 2309 | { |
78608366 | 2310 | while (!(is_mem_exclusive(cs) || is_mem_hardwall(cs)) && cs->parent) |
9bf2229f PJ |
2311 | cs = cs->parent; |
2312 | return cs; | |
2313 | } | |
2314 | ||
d9fd8a6d | 2315 | /** |
a1bc5a4e DR |
2316 | * cpuset_node_allowed_softwall - Can we allocate on a memory node? |
2317 | * @node: is this an allowed node? | |
02a0e53d | 2318 | * @gfp_mask: memory allocation flags |
d9fd8a6d | 2319 | * |
a1bc5a4e DR |
2320 | * If we're in interrupt, yes, we can always allocate. If __GFP_THISNODE is |
2321 | * set, yes, we can always allocate. If node is in our task's mems_allowed, | |
2322 | * yes. If it's not a __GFP_HARDWALL request and this node is in the nearest | |
2323 | * hardwalled cpuset ancestor to this task's cpuset, yes. If the task has been | |
2324 | * OOM killed and has access to memory reserves as specified by the TIF_MEMDIE | |
2325 | * flag, yes. | |
9bf2229f PJ |
2326 | * Otherwise, no. |
2327 | * | |
a1bc5a4e DR |
2328 | * If __GFP_HARDWALL is set, cpuset_node_allowed_softwall() reduces to |
2329 | * cpuset_node_allowed_hardwall(). Otherwise, cpuset_node_allowed_softwall() | |
2330 | * might sleep, and might allow a node from an enclosing cpuset. | |
02a0e53d | 2331 | * |
a1bc5a4e DR |
2332 | * cpuset_node_allowed_hardwall() only handles the simpler case of hardwall |
2333 | * cpusets, and never sleeps. | |
02a0e53d PJ |
2334 | * |
2335 | * The __GFP_THISNODE placement logic is really handled elsewhere, | |
2336 | * by forcibly using a zonelist starting at a specified node, and by | |
2337 | * (in get_page_from_freelist()) refusing to consider the zones for | |
2338 | * any node on the zonelist except the first. By the time any such | |
2339 | * calls get to this routine, we should just shut up and say 'yes'. | |
2340 | * | |
9bf2229f | 2341 | * GFP_USER allocations are marked with the __GFP_HARDWALL bit, |
c596d9f3 DR |
2342 | * and do not allow allocations outside the current tasks cpuset |
2343 | * unless the task has been OOM killed as is marked TIF_MEMDIE. | |
9bf2229f | 2344 | * GFP_KERNEL allocations are not so marked, so can escape to the |
78608366 | 2345 | * nearest enclosing hardwalled ancestor cpuset. |
9bf2229f | 2346 | * |
02a0e53d PJ |
2347 | * Scanning up parent cpusets requires callback_mutex. The |
2348 | * __alloc_pages() routine only calls here with __GFP_HARDWALL bit | |
2349 | * _not_ set if it's a GFP_KERNEL allocation, and all nodes in the | |
2350 | * current tasks mems_allowed came up empty on the first pass over | |
2351 | * the zonelist. So only GFP_KERNEL allocations, if all nodes in the | |
2352 | * cpuset are short of memory, might require taking the callback_mutex | |
2353 | * mutex. | |
9bf2229f | 2354 | * |
36be57ff | 2355 | * The first call here from mm/page_alloc:get_page_from_freelist() |
02a0e53d PJ |
2356 | * has __GFP_HARDWALL set in gfp_mask, enforcing hardwall cpusets, |
2357 | * so no allocation on a node outside the cpuset is allowed (unless | |
2358 | * in interrupt, of course). | |
36be57ff PJ |
2359 | * |
2360 | * The second pass through get_page_from_freelist() doesn't even call | |
2361 | * here for GFP_ATOMIC calls. For those calls, the __alloc_pages() | |
2362 | * variable 'wait' is not set, and the bit ALLOC_CPUSET is not set | |
2363 | * in alloc_flags. That logic and the checks below have the combined | |
2364 | * affect that: | |
9bf2229f PJ |
2365 | * in_interrupt - any node ok (current task context irrelevant) |
2366 | * GFP_ATOMIC - any node ok | |
c596d9f3 | 2367 | * TIF_MEMDIE - any node ok |
78608366 | 2368 | * GFP_KERNEL - any node in enclosing hardwalled cpuset ok |
9bf2229f | 2369 | * GFP_USER - only nodes in current tasks mems allowed ok. |
36be57ff PJ |
2370 | * |
2371 | * Rule: | |
a1bc5a4e | 2372 | * Don't call cpuset_node_allowed_softwall if you can't sleep, unless you |
36be57ff PJ |
2373 | * pass in the __GFP_HARDWALL flag set in gfp_flag, which disables |
2374 | * the code that might scan up ancestor cpusets and sleep. | |
02a0e53d | 2375 | */ |
a1bc5a4e | 2376 | int __cpuset_node_allowed_softwall(int node, gfp_t gfp_mask) |
1da177e4 | 2377 | { |
9bf2229f | 2378 | const struct cpuset *cs; /* current cpuset ancestors */ |
29afd49b | 2379 | int allowed; /* is allocation in zone z allowed? */ |
9bf2229f | 2380 | |
9b819d20 | 2381 | if (in_interrupt() || (gfp_mask & __GFP_THISNODE)) |
9bf2229f | 2382 | return 1; |
92d1dbd2 | 2383 | might_sleep_if(!(gfp_mask & __GFP_HARDWALL)); |
9bf2229f PJ |
2384 | if (node_isset(node, current->mems_allowed)) |
2385 | return 1; | |
c596d9f3 DR |
2386 | /* |
2387 | * Allow tasks that have access to memory reserves because they have | |
2388 | * been OOM killed to get memory anywhere. | |
2389 | */ | |
2390 | if (unlikely(test_thread_flag(TIF_MEMDIE))) | |
2391 | return 1; | |
9bf2229f PJ |
2392 | if (gfp_mask & __GFP_HARDWALL) /* If hardwall request, stop here */ |
2393 | return 0; | |
2394 | ||
5563e770 BP |
2395 | if (current->flags & PF_EXITING) /* Let dying task have memory */ |
2396 | return 1; | |
2397 | ||
9bf2229f | 2398 | /* Not hardwall and node outside mems_allowed: scan up cpusets */ |
3d3f26a7 | 2399 | mutex_lock(&callback_mutex); |
053199ed | 2400 | |
053199ed | 2401 | task_lock(current); |
78608366 | 2402 | cs = nearest_hardwall_ancestor(task_cs(current)); |
053199ed PJ |
2403 | task_unlock(current); |
2404 | ||
9bf2229f | 2405 | allowed = node_isset(node, cs->mems_allowed); |
3d3f26a7 | 2406 | mutex_unlock(&callback_mutex); |
9bf2229f | 2407 | return allowed; |
1da177e4 LT |
2408 | } |
2409 | ||
02a0e53d | 2410 | /* |
a1bc5a4e DR |
2411 | * cpuset_node_allowed_hardwall - Can we allocate on a memory node? |
2412 | * @node: is this an allowed node? | |
02a0e53d PJ |
2413 | * @gfp_mask: memory allocation flags |
2414 | * | |
a1bc5a4e DR |
2415 | * If we're in interrupt, yes, we can always allocate. If __GFP_THISNODE is |
2416 | * set, yes, we can always allocate. If node is in our task's mems_allowed, | |
2417 | * yes. If the task has been OOM killed and has access to memory reserves as | |
2418 | * specified by the TIF_MEMDIE flag, yes. | |
2419 | * Otherwise, no. | |
02a0e53d PJ |
2420 | * |
2421 | * The __GFP_THISNODE placement logic is really handled elsewhere, | |
2422 | * by forcibly using a zonelist starting at a specified node, and by | |
2423 | * (in get_page_from_freelist()) refusing to consider the zones for | |
2424 | * any node on the zonelist except the first. By the time any such | |
2425 | * calls get to this routine, we should just shut up and say 'yes'. | |
2426 | * | |
a1bc5a4e DR |
2427 | * Unlike the cpuset_node_allowed_softwall() variant, above, |
2428 | * this variant requires that the node be in the current task's | |
02a0e53d PJ |
2429 | * mems_allowed or that we're in interrupt. It does not scan up the |
2430 | * cpuset hierarchy for the nearest enclosing mem_exclusive cpuset. | |
2431 | * It never sleeps. | |
2432 | */ | |
a1bc5a4e | 2433 | int __cpuset_node_allowed_hardwall(int node, gfp_t gfp_mask) |
02a0e53d | 2434 | { |
02a0e53d PJ |
2435 | if (in_interrupt() || (gfp_mask & __GFP_THISNODE)) |
2436 | return 1; | |
02a0e53d PJ |
2437 | if (node_isset(node, current->mems_allowed)) |
2438 | return 1; | |
dedf8b79 DW |
2439 | /* |
2440 | * Allow tasks that have access to memory reserves because they have | |
2441 | * been OOM killed to get memory anywhere. | |
2442 | */ | |
2443 | if (unlikely(test_thread_flag(TIF_MEMDIE))) | |
2444 | return 1; | |
02a0e53d PJ |
2445 | return 0; |
2446 | } | |
2447 | ||
825a46af | 2448 | /** |
6adef3eb JS |
2449 | * cpuset_mem_spread_node() - On which node to begin search for a file page |
2450 | * cpuset_slab_spread_node() - On which node to begin search for a slab page | |
825a46af PJ |
2451 | * |
2452 | * If a task is marked PF_SPREAD_PAGE or PF_SPREAD_SLAB (as for | |
2453 | * tasks in a cpuset with is_spread_page or is_spread_slab set), | |
2454 | * and if the memory allocation used cpuset_mem_spread_node() | |
2455 | * to determine on which node to start looking, as it will for | |
2456 | * certain page cache or slab cache pages such as used for file | |
2457 | * system buffers and inode caches, then instead of starting on the | |
2458 | * local node to look for a free page, rather spread the starting | |
2459 | * node around the tasks mems_allowed nodes. | |
2460 | * | |
2461 | * We don't have to worry about the returned node being offline | |
2462 | * because "it can't happen", and even if it did, it would be ok. | |
2463 | * | |
2464 | * The routines calling guarantee_online_mems() are careful to | |
2465 | * only set nodes in task->mems_allowed that are online. So it | |
2466 | * should not be possible for the following code to return an | |
2467 | * offline node. But if it did, that would be ok, as this routine | |
2468 | * is not returning the node where the allocation must be, only | |
2469 | * the node where the search should start. The zonelist passed to | |
2470 | * __alloc_pages() will include all nodes. If the slab allocator | |
2471 | * is passed an offline node, it will fall back to the local node. | |
2472 | * See kmem_cache_alloc_node(). | |
2473 | */ | |
2474 | ||
6adef3eb | 2475 | static int cpuset_spread_node(int *rotor) |
825a46af PJ |
2476 | { |
2477 | int node; | |
2478 | ||
6adef3eb | 2479 | node = next_node(*rotor, current->mems_allowed); |
825a46af PJ |
2480 | if (node == MAX_NUMNODES) |
2481 | node = first_node(current->mems_allowed); | |
6adef3eb | 2482 | *rotor = node; |
825a46af PJ |
2483 | return node; |
2484 | } | |
6adef3eb JS |
2485 | |
2486 | int cpuset_mem_spread_node(void) | |
2487 | { | |
778d3b0f MH |
2488 | if (current->cpuset_mem_spread_rotor == NUMA_NO_NODE) |
2489 | current->cpuset_mem_spread_rotor = | |
2490 | node_random(¤t->mems_allowed); | |
2491 | ||
6adef3eb JS |
2492 | return cpuset_spread_node(¤t->cpuset_mem_spread_rotor); |
2493 | } | |
2494 | ||
2495 | int cpuset_slab_spread_node(void) | |
2496 | { | |
778d3b0f MH |
2497 | if (current->cpuset_slab_spread_rotor == NUMA_NO_NODE) |
2498 | current->cpuset_slab_spread_rotor = | |
2499 | node_random(¤t->mems_allowed); | |
2500 | ||
6adef3eb JS |
2501 | return cpuset_spread_node(¤t->cpuset_slab_spread_rotor); |
2502 | } | |
2503 | ||
825a46af PJ |
2504 | EXPORT_SYMBOL_GPL(cpuset_mem_spread_node); |
2505 | ||
ef08e3b4 | 2506 | /** |
bbe373f2 DR |
2507 | * cpuset_mems_allowed_intersects - Does @tsk1's mems_allowed intersect @tsk2's? |
2508 | * @tsk1: pointer to task_struct of some task. | |
2509 | * @tsk2: pointer to task_struct of some other task. | |
2510 | * | |
2511 | * Description: Return true if @tsk1's mems_allowed intersects the | |
2512 | * mems_allowed of @tsk2. Used by the OOM killer to determine if | |
2513 | * one of the task's memory usage might impact the memory available | |
2514 | * to the other. | |
ef08e3b4 PJ |
2515 | **/ |
2516 | ||
bbe373f2 DR |
2517 | int cpuset_mems_allowed_intersects(const struct task_struct *tsk1, |
2518 | const struct task_struct *tsk2) | |
ef08e3b4 | 2519 | { |
bbe373f2 | 2520 | return nodes_intersects(tsk1->mems_allowed, tsk2->mems_allowed); |
ef08e3b4 PJ |
2521 | } |
2522 | ||
75aa1994 DR |
2523 | /** |
2524 | * cpuset_print_task_mems_allowed - prints task's cpuset and mems_allowed | |
2525 | * @task: pointer to task_struct of some task. | |
2526 | * | |
2527 | * Description: Prints @task's name, cpuset name, and cached copy of its | |
2528 | * mems_allowed to the kernel log. Must hold task_lock(task) to allow | |
2529 | * dereferencing task_cs(task). | |
2530 | */ | |
2531 | void cpuset_print_task_mems_allowed(struct task_struct *tsk) | |
2532 | { | |
2533 | struct dentry *dentry; | |
2534 | ||
2535 | dentry = task_cs(tsk)->css.cgroup->dentry; | |
2536 | spin_lock(&cpuset_buffer_lock); | |
2537 | snprintf(cpuset_name, CPUSET_NAME_LEN, | |
2538 | dentry ? (const char *)dentry->d_name.name : "/"); | |
2539 | nodelist_scnprintf(cpuset_nodelist, CPUSET_NODELIST_LEN, | |
2540 | tsk->mems_allowed); | |
2541 | printk(KERN_INFO "%s cpuset=%s mems_allowed=%s\n", | |
2542 | tsk->comm, cpuset_name, cpuset_nodelist); | |
2543 | spin_unlock(&cpuset_buffer_lock); | |
2544 | } | |
2545 | ||
3e0d98b9 PJ |
2546 | /* |
2547 | * Collection of memory_pressure is suppressed unless | |
2548 | * this flag is enabled by writing "1" to the special | |
2549 | * cpuset file 'memory_pressure_enabled' in the root cpuset. | |
2550 | */ | |
2551 | ||
c5b2aff8 | 2552 | int cpuset_memory_pressure_enabled __read_mostly; |
3e0d98b9 PJ |
2553 | |
2554 | /** | |
2555 | * cpuset_memory_pressure_bump - keep stats of per-cpuset reclaims. | |
2556 | * | |
2557 | * Keep a running average of the rate of synchronous (direct) | |
2558 | * page reclaim efforts initiated by tasks in each cpuset. | |
2559 | * | |
2560 | * This represents the rate at which some task in the cpuset | |
2561 | * ran low on memory on all nodes it was allowed to use, and | |
2562 | * had to enter the kernels page reclaim code in an effort to | |
2563 | * create more free memory by tossing clean pages or swapping | |
2564 | * or writing dirty pages. | |
2565 | * | |
2566 | * Display to user space in the per-cpuset read-only file | |
2567 | * "memory_pressure". Value displayed is an integer | |
2568 | * representing the recent rate of entry into the synchronous | |
2569 | * (direct) page reclaim by any task attached to the cpuset. | |
2570 | **/ | |
2571 | ||
2572 | void __cpuset_memory_pressure_bump(void) | |
2573 | { | |
3e0d98b9 | 2574 | task_lock(current); |
8793d854 | 2575 | fmeter_markevent(&task_cs(current)->fmeter); |
3e0d98b9 PJ |
2576 | task_unlock(current); |
2577 | } | |
2578 | ||
8793d854 | 2579 | #ifdef CONFIG_PROC_PID_CPUSET |
1da177e4 LT |
2580 | /* |
2581 | * proc_cpuset_show() | |
2582 | * - Print tasks cpuset path into seq_file. | |
2583 | * - Used for /proc/<pid>/cpuset. | |
053199ed PJ |
2584 | * - No need to task_lock(tsk) on this tsk->cpuset reference, as it |
2585 | * doesn't really matter if tsk->cpuset changes after we read it, | |
c8d9c90c | 2586 | * and we take cgroup_mutex, keeping cpuset_attach() from changing it |
2df167a3 | 2587 | * anyway. |
1da177e4 | 2588 | */ |
029190c5 | 2589 | static int proc_cpuset_show(struct seq_file *m, void *unused_v) |
1da177e4 | 2590 | { |
13b41b09 | 2591 | struct pid *pid; |
1da177e4 LT |
2592 | struct task_struct *tsk; |
2593 | char *buf; | |
8793d854 | 2594 | struct cgroup_subsys_state *css; |
99f89551 | 2595 | int retval; |
1da177e4 | 2596 | |
99f89551 | 2597 | retval = -ENOMEM; |
1da177e4 LT |
2598 | buf = kmalloc(PAGE_SIZE, GFP_KERNEL); |
2599 | if (!buf) | |
99f89551 EB |
2600 | goto out; |
2601 | ||
2602 | retval = -ESRCH; | |
13b41b09 EB |
2603 | pid = m->private; |
2604 | tsk = get_pid_task(pid, PIDTYPE_PID); | |
99f89551 EB |
2605 | if (!tsk) |
2606 | goto out_free; | |
1da177e4 | 2607 | |
99f89551 | 2608 | retval = -EINVAL; |
8793d854 PM |
2609 | cgroup_lock(); |
2610 | css = task_subsys_state(tsk, cpuset_subsys_id); | |
2611 | retval = cgroup_path(css->cgroup, buf, PAGE_SIZE); | |
1da177e4 | 2612 | if (retval < 0) |
99f89551 | 2613 | goto out_unlock; |
1da177e4 LT |
2614 | seq_puts(m, buf); |
2615 | seq_putc(m, '\n'); | |
99f89551 | 2616 | out_unlock: |
8793d854 | 2617 | cgroup_unlock(); |
99f89551 EB |
2618 | put_task_struct(tsk); |
2619 | out_free: | |
1da177e4 | 2620 | kfree(buf); |
99f89551 | 2621 | out: |
1da177e4 LT |
2622 | return retval; |
2623 | } | |
2624 | ||
2625 | static int cpuset_open(struct inode *inode, struct file *file) | |
2626 | { | |
13b41b09 EB |
2627 | struct pid *pid = PROC_I(inode)->pid; |
2628 | return single_open(file, proc_cpuset_show, pid); | |
1da177e4 LT |
2629 | } |
2630 | ||
9a32144e | 2631 | const struct file_operations proc_cpuset_operations = { |
1da177e4 LT |
2632 | .open = cpuset_open, |
2633 | .read = seq_read, | |
2634 | .llseek = seq_lseek, | |
2635 | .release = single_release, | |
2636 | }; | |
8793d854 | 2637 | #endif /* CONFIG_PROC_PID_CPUSET */ |
1da177e4 | 2638 | |
d01d4827 | 2639 | /* Display task mems_allowed in /proc/<pid>/status file. */ |
df5f8314 EB |
2640 | void cpuset_task_status_allowed(struct seq_file *m, struct task_struct *task) |
2641 | { | |
df5f8314 | 2642 | seq_printf(m, "Mems_allowed:\t"); |
30e8e136 | 2643 | seq_nodemask(m, &task->mems_allowed); |
df5f8314 | 2644 | seq_printf(m, "\n"); |
39106dcf | 2645 | seq_printf(m, "Mems_allowed_list:\t"); |
30e8e136 | 2646 | seq_nodemask_list(m, &task->mems_allowed); |
39106dcf | 2647 | seq_printf(m, "\n"); |
1da177e4 | 2648 | } |