1 /*
2 * kernel/cpuset.c
3 *
4 * Processor and Memory placement constraints for sets of tasks.
5 *
6 * Copyright (C) 2003 BULL SA.
7 * Copyright (C) 2004-2007 Silicon Graphics, Inc.
8 * Copyright (C) 2006 Google, Inc
9 *
10 * Portions derived from Patrick Mochel's sysfs code.
11 * sysfs is Copyright (c) 2001-3 Patrick Mochel
12 *
13 * 2003-10-10 Written by Simon Derr.
14 * 2003-10-22 Updates by Stephen Hemminger.
15 * 2004 May-July Rework by Paul Jackson.
16 * 2006 Rework by Paul Menage to use generic cgroups
17 * 2008 Rework of the scheduler domains and CPU hotplug handling
18 * by Max Krasnyansky
19 *
20 * This file is subject to the terms and conditions of the GNU General Public
21 * License. See the file COPYING in the main directory of the Linux
22 * distribution for more details.
23 */
24
25 #include <linux/cpu.h>
26 #include <linux/cpumask.h>
27 #include <linux/cpuset.h>
28 #include <linux/err.h>
29 #include <linux/errno.h>
30 #include <linux/file.h>
31 #include <linux/fs.h>
32 #include <linux/init.h>
33 #include <linux/interrupt.h>
34 #include <linux/kernel.h>
35 #include <linux/kmod.h>
36 #include <linux/list.h>
37 #include <linux/mempolicy.h>
38 #include <linux/mm.h>
39 #include <linux/memory.h>
40 #include <linux/export.h>
41 #include <linux/mount.h>
42 #include <linux/namei.h>
43 #include <linux/pagemap.h>
44 #include <linux/proc_fs.h>
45 #include <linux/rcupdate.h>
46 #include <linux/sched.h>
47 #include <linux/seq_file.h>
48 #include <linux/security.h>
49 #include <linux/slab.h>
50 #include <linux/spinlock.h>
51 #include <linux/stat.h>
52 #include <linux/string.h>
53 #include <linux/time.h>
54 #include <linux/backing-dev.h>
55 #include <linux/sort.h>
56
57 #include <asm/uaccess.h>
58 #include <linux/atomic.h>
59 #include <linux/mutex.h>
60 #include <linux/workqueue.h>
61 #include <linux/cgroup.h>
62
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
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 */
76 int number_of_cpusets __read_mostly;
77
78 /* Forward declare cgroup structures */
79 struct cgroup_subsys cpuset_subsys;
80 struct cpuset;
81
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
91 struct cpuset {
92 struct cgroup_subsys_state css;
93
94 unsigned long flags; /* "unsigned long" so bitops work */
95 cpumask_var_t cpus_allowed; /* CPUs allowed to tasks in cpuset */
96 nodemask_t mems_allowed; /* Memory Nodes allowed to tasks */
97
98 struct cpuset *parent; /* my parent */
99
100 struct fmeter fmeter; /* memory_pressure filter */
101
102 /* partition number for rebuild_sched_domains() */
103 int pn;
104
105 /* for custom sched domain */
106 int relax_domain_level;
107
108 /* used for walking a cpuset hierarchy */
109 struct list_head stack_list;
110 };
111
112 /* Retrieve the cpuset for a cgroup */
cgroup_cs(struct cgroup * cont)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 */
task_cs(struct task_struct * 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 }
125
126 #ifdef CONFIG_NUMA
task_has_mempolicy(struct task_struct * task)127 static inline bool task_has_mempolicy(struct task_struct *task)
128 {
129 return task->mempolicy;
130 }
131 #else
task_has_mempolicy(struct task_struct * task)132 static inline bool task_has_mempolicy(struct task_struct *task)
133 {
134 return false;
135 }
136 #endif
137
138
139 /* bits in struct cpuset flags field */
140 typedef enum {
141 CS_CPU_EXCLUSIVE,
142 CS_MEM_EXCLUSIVE,
143 CS_MEM_HARDWALL,
144 CS_MEMORY_MIGRATE,
145 CS_SCHED_LOAD_BALANCE,
146 CS_SPREAD_PAGE,
147 CS_SPREAD_SLAB,
148 } cpuset_flagbits_t;
149
150 /* convenient tests for these bits */
is_cpu_exclusive(const struct cpuset * cs)151 static inline int is_cpu_exclusive(const struct cpuset *cs)
152 {
153 return test_bit(CS_CPU_EXCLUSIVE, &cs->flags);
154 }
155
is_mem_exclusive(const struct cpuset * cs)156 static inline int is_mem_exclusive(const struct cpuset *cs)
157 {
158 return test_bit(CS_MEM_EXCLUSIVE, &cs->flags);
159 }
160
is_mem_hardwall(const struct cpuset * cs)161 static inline int is_mem_hardwall(const struct cpuset *cs)
162 {
163 return test_bit(CS_MEM_HARDWALL, &cs->flags);
164 }
165
is_sched_load_balance(const struct cpuset * cs)166 static inline int is_sched_load_balance(const struct cpuset *cs)
167 {
168 return test_bit(CS_SCHED_LOAD_BALANCE, &cs->flags);
169 }
170
is_memory_migrate(const struct cpuset * cs)171 static inline int is_memory_migrate(const struct cpuset *cs)
172 {
173 return test_bit(CS_MEMORY_MIGRATE, &cs->flags);
174 }
175
is_spread_page(const struct cpuset * cs)176 static inline int is_spread_page(const struct cpuset *cs)
177 {
178 return test_bit(CS_SPREAD_PAGE, &cs->flags);
179 }
180
is_spread_slab(const struct cpuset * cs)181 static inline int is_spread_slab(const struct cpuset *cs)
182 {
183 return test_bit(CS_SPREAD_SLAB, &cs->flags);
184 }
185
186 static struct cpuset top_cpuset = {
187 .flags = ((1 << CS_CPU_EXCLUSIVE) | (1 << CS_MEM_EXCLUSIVE)),
188 };
189
190 /*
191 * There are two global mutexes guarding cpuset structures. The first
192 * is the main control groups cgroup_mutex, accessed via
193 * cgroup_lock()/cgroup_unlock(). The second is the cpuset-specific
194 * callback_mutex, below. They can nest. It is ok to first take
195 * cgroup_mutex, then nest callback_mutex. We also require taking
196 * task_lock() when dereferencing a task's cpuset pointer. See "The
197 * task_lock() exception", at the end of this comment.
198 *
199 * A task must hold both mutexes to modify cpusets. If a task
200 * holds cgroup_mutex, then it blocks others wanting that mutex,
201 * ensuring that it is the only task able to also acquire callback_mutex
202 * and be able to modify cpusets. It can perform various checks on
203 * the cpuset structure first, knowing nothing will change. It can
204 * also allocate memory while just holding cgroup_mutex. While it is
205 * performing these checks, various callback routines can briefly
206 * acquire callback_mutex to query cpusets. Once it is ready to make
207 * the changes, it takes callback_mutex, blocking everyone else.
208 *
209 * Calls to the kernel memory allocator can not be made while holding
210 * callback_mutex, as that would risk double tripping on callback_mutex
211 * from one of the callbacks into the cpuset code from within
212 * __alloc_pages().
213 *
214 * If a task is only holding callback_mutex, then it has read-only
215 * access to cpusets.
216 *
217 * Now, the task_struct fields mems_allowed and mempolicy may be changed
218 * by other task, we use alloc_lock in the task_struct fields to protect
219 * them.
220 *
221 * The cpuset_common_file_read() handlers only hold callback_mutex across
222 * small pieces of code, such as when reading out possibly multi-word
223 * cpumasks and nodemasks.
224 *
225 * Accessing a task's cpuset should be done in accordance with the
226 * guidelines for accessing subsystem state in kernel/cgroup.c
227 */
228
229 static DEFINE_MUTEX(callback_mutex);
230
231 /*
232 * cpuset_buffer_lock protects both the cpuset_name and cpuset_nodelist
233 * buffers. They are statically allocated to prevent using excess stack
234 * when calling cpuset_print_task_mems_allowed().
235 */
236 #define CPUSET_NAME_LEN (128)
237 #define CPUSET_NODELIST_LEN (256)
238 static char cpuset_name[CPUSET_NAME_LEN];
239 static char cpuset_nodelist[CPUSET_NODELIST_LEN];
240 static DEFINE_SPINLOCK(cpuset_buffer_lock);
241
242 /*
243 * This is ugly, but preserves the userspace API for existing cpuset
244 * users. If someone tries to mount the "cpuset" filesystem, we
245 * silently switch it to mount "cgroup" instead
246 */
cpuset_mount(struct file_system_type * fs_type,int flags,const char * unused_dev_name,void * data)247 static struct dentry *cpuset_mount(struct file_system_type *fs_type,
248 int flags, const char *unused_dev_name, void *data)
249 {
250 struct file_system_type *cgroup_fs = get_fs_type("cgroup");
251 struct dentry *ret = ERR_PTR(-ENODEV);
252 if (cgroup_fs) {
253 char mountopts[] =
254 "cpuset,noprefix,"
255 "release_agent=/sbin/cpuset_release_agent";
256 ret = cgroup_fs->mount(cgroup_fs, flags,
257 unused_dev_name, mountopts);
258 put_filesystem(cgroup_fs);
259 }
260 return ret;
261 }
262
263 static struct file_system_type cpuset_fs_type = {
264 .name = "cpuset",
265 .mount = cpuset_mount,
266 };
267
268 /*
269 * Return in pmask the portion of a cpusets's cpus_allowed that
270 * are online. If none are online, walk up the cpuset hierarchy
271 * until we find one that does have some online cpus. If we get
272 * all the way to the top and still haven't found any online cpus,
273 * return cpu_online_mask. Or if passed a NULL cs from an exit'ing
274 * task, return cpu_online_mask.
275 *
276 * One way or another, we guarantee to return some non-empty subset
277 * of cpu_online_mask.
278 *
279 * Call with callback_mutex held.
280 */
281
guarantee_online_cpus(const struct cpuset * cs,struct cpumask * pmask)282 static void guarantee_online_cpus(const struct cpuset *cs,
283 struct cpumask *pmask)
284 {
285 while (cs && !cpumask_intersects(cs->cpus_allowed, cpu_online_mask))
286 cs = cs->parent;
287 if (cs)
288 cpumask_and(pmask, cs->cpus_allowed, cpu_online_mask);
289 else
290 cpumask_copy(pmask, cpu_online_mask);
291 BUG_ON(!cpumask_intersects(pmask, cpu_online_mask));
292 }
293
294 /*
295 * Return in *pmask the portion of a cpusets's mems_allowed that
296 * are online, with memory. If none are online with memory, walk
297 * up the cpuset hierarchy until we find one that does have some
298 * online mems. If we get all the way to the top and still haven't
299 * found any online mems, return node_states[N_HIGH_MEMORY].
300 *
301 * One way or another, we guarantee to return some non-empty subset
302 * of node_states[N_HIGH_MEMORY].
303 *
304 * Call with callback_mutex held.
305 */
306
guarantee_online_mems(const struct cpuset * cs,nodemask_t * pmask)307 static void guarantee_online_mems(const struct cpuset *cs, nodemask_t *pmask)
308 {
309 while (cs && !nodes_intersects(cs->mems_allowed,
310 node_states[N_HIGH_MEMORY]))
311 cs = cs->parent;
312 if (cs)
313 nodes_and(*pmask, cs->mems_allowed,
314 node_states[N_HIGH_MEMORY]);
315 else
316 *pmask = node_states[N_HIGH_MEMORY];
317 BUG_ON(!nodes_intersects(*pmask, node_states[N_HIGH_MEMORY]));
318 }
319
320 /*
321 * update task's spread flag if cpuset's page/slab spread flag is set
322 *
323 * Called with callback_mutex/cgroup_mutex held
324 */
cpuset_update_task_spread_flag(struct cpuset * cs,struct task_struct * tsk)325 static void cpuset_update_task_spread_flag(struct cpuset *cs,
326 struct task_struct *tsk)
327 {
328 if (is_spread_page(cs))
329 tsk->flags |= PF_SPREAD_PAGE;
330 else
331 tsk->flags &= ~PF_SPREAD_PAGE;
332 if (is_spread_slab(cs))
333 tsk->flags |= PF_SPREAD_SLAB;
334 else
335 tsk->flags &= ~PF_SPREAD_SLAB;
336 }
337
338 /*
339 * is_cpuset_subset(p, q) - Is cpuset p a subset of cpuset q?
340 *
341 * One cpuset is a subset of another if all its allowed CPUs and
342 * Memory Nodes are a subset of the other, and its exclusive flags
343 * are only set if the other's are set. Call holding cgroup_mutex.
344 */
345
is_cpuset_subset(const struct cpuset * p,const struct cpuset * q)346 static int is_cpuset_subset(const struct cpuset *p, const struct cpuset *q)
347 {
348 return cpumask_subset(p->cpus_allowed, q->cpus_allowed) &&
349 nodes_subset(p->mems_allowed, q->mems_allowed) &&
350 is_cpu_exclusive(p) <= is_cpu_exclusive(q) &&
351 is_mem_exclusive(p) <= is_mem_exclusive(q);
352 }
353
354 /**
355 * alloc_trial_cpuset - allocate a trial cpuset
356 * @cs: the cpuset that the trial cpuset duplicates
357 */
alloc_trial_cpuset(const struct cpuset * cs)358 static struct cpuset *alloc_trial_cpuset(const struct cpuset *cs)
359 {
360 struct cpuset *trial;
361
362 trial = kmemdup(cs, sizeof(*cs), GFP_KERNEL);
363 if (!trial)
364 return NULL;
365
366 if (!alloc_cpumask_var(&trial->cpus_allowed, GFP_KERNEL)) {
367 kfree(trial);
368 return NULL;
369 }
370 cpumask_copy(trial->cpus_allowed, cs->cpus_allowed);
371
372 return trial;
373 }
374
375 /**
376 * free_trial_cpuset - free the trial cpuset
377 * @trial: the trial cpuset to be freed
378 */
free_trial_cpuset(struct cpuset * trial)379 static void free_trial_cpuset(struct cpuset *trial)
380 {
381 free_cpumask_var(trial->cpus_allowed);
382 kfree(trial);
383 }
384
385 /*
386 * validate_change() - Used to validate that any proposed cpuset change
387 * follows the structural rules for cpusets.
388 *
389 * If we replaced the flag and mask values of the current cpuset
390 * (cur) with those values in the trial cpuset (trial), would
391 * our various subset and exclusive rules still be valid? Presumes
392 * cgroup_mutex held.
393 *
394 * 'cur' is the address of an actual, in-use cpuset. Operations
395 * such as list traversal that depend on the actual address of the
396 * cpuset in the list must use cur below, not trial.
397 *
398 * 'trial' is the address of bulk structure copy of cur, with
399 * perhaps one or more of the fields cpus_allowed, mems_allowed,
400 * or flags changed to new, trial values.
401 *
402 * Return 0 if valid, -errno if not.
403 */
404
validate_change(const struct cpuset * cur,const struct cpuset * trial)405 static int validate_change(const struct cpuset *cur, const struct cpuset *trial)
406 {
407 struct cgroup *cont;
408 struct cpuset *c, *par;
409
410 /* Each of our child cpusets must be a subset of us */
411 list_for_each_entry(cont, &cur->css.cgroup->children, sibling) {
412 if (!is_cpuset_subset(cgroup_cs(cont), trial))
413 return -EBUSY;
414 }
415
416 /* Remaining checks don't apply to root cpuset */
417 if (cur == &top_cpuset)
418 return 0;
419
420 par = cur->parent;
421
422 /* We must be a subset of our parent cpuset */
423 if (!is_cpuset_subset(trial, par))
424 return -EACCES;
425
426 /*
427 * If either I or some sibling (!= me) is exclusive, we can't
428 * overlap
429 */
430 list_for_each_entry(cont, &par->css.cgroup->children, sibling) {
431 c = cgroup_cs(cont);
432 if ((is_cpu_exclusive(trial) || is_cpu_exclusive(c)) &&
433 c != cur &&
434 cpumask_intersects(trial->cpus_allowed, c->cpus_allowed))
435 return -EINVAL;
436 if ((is_mem_exclusive(trial) || is_mem_exclusive(c)) &&
437 c != cur &&
438 nodes_intersects(trial->mems_allowed, c->mems_allowed))
439 return -EINVAL;
440 }
441
442 /* Cpusets with tasks can't have empty cpus_allowed or mems_allowed */
443 if (cgroup_task_count(cur->css.cgroup)) {
444 if (cpumask_empty(trial->cpus_allowed) ||
445 nodes_empty(trial->mems_allowed)) {
446 return -ENOSPC;
447 }
448 }
449
450 return 0;
451 }
452
453 #ifdef CONFIG_SMP
454 /*
455 * Helper routine for generate_sched_domains().
456 * Do cpusets a, b have overlapping cpus_allowed masks?
457 */
cpusets_overlap(struct cpuset * a,struct cpuset * b)458 static int cpusets_overlap(struct cpuset *a, struct cpuset *b)
459 {
460 return cpumask_intersects(a->cpus_allowed, b->cpus_allowed);
461 }
462
463 static void
update_domain_attr(struct sched_domain_attr * dattr,struct cpuset * c)464 update_domain_attr(struct sched_domain_attr *dattr, struct cpuset *c)
465 {
466 if (dattr->relax_domain_level < c->relax_domain_level)
467 dattr->relax_domain_level = c->relax_domain_level;
468 return;
469 }
470
471 static void
update_domain_attr_tree(struct sched_domain_attr * dattr,struct cpuset * c)472 update_domain_attr_tree(struct sched_domain_attr *dattr, struct cpuset *c)
473 {
474 LIST_HEAD(q);
475
476 list_add(&c->stack_list, &q);
477 while (!list_empty(&q)) {
478 struct cpuset *cp;
479 struct cgroup *cont;
480 struct cpuset *child;
481
482 cp = list_first_entry(&q, struct cpuset, stack_list);
483 list_del(q.next);
484
485 if (cpumask_empty(cp->cpus_allowed))
486 continue;
487
488 if (is_sched_load_balance(cp))
489 update_domain_attr(dattr, cp);
490
491 list_for_each_entry(cont, &cp->css.cgroup->children, sibling) {
492 child = cgroup_cs(cont);
493 list_add_tail(&child->stack_list, &q);
494 }
495 }
496 }
497
498 /*
499 * generate_sched_domains()
500 *
501 * This function builds a partial partition of the systems CPUs
502 * A 'partial partition' is a set of non-overlapping subsets whose
503 * union is a subset of that set.
504 * The output of this function needs to be passed to kernel/sched.c
505 * partition_sched_domains() routine, which will rebuild the scheduler's
506 * load balancing domains (sched domains) as specified by that partial
507 * partition.
508 *
509 * See "What is sched_load_balance" in Documentation/cgroups/cpusets.txt
510 * for a background explanation of this.
511 *
512 * Does not return errors, on the theory that the callers of this
513 * routine would rather not worry about failures to rebuild sched
514 * domains when operating in the severe memory shortage situations
515 * that could cause allocation failures below.
516 *
517 * Must be called with cgroup_lock held.
518 *
519 * The three key local variables below are:
520 * q - a linked-list queue of cpuset pointers, used to implement a
521 * top-down scan of all cpusets. This scan loads a pointer
522 * to each cpuset marked is_sched_load_balance into the
523 * array 'csa'. For our purposes, rebuilding the schedulers
524 * sched domains, we can ignore !is_sched_load_balance cpusets.
525 * csa - (for CpuSet Array) Array of pointers to all the cpusets
526 * that need to be load balanced, for convenient iterative
527 * access by the subsequent code that finds the best partition,
528 * i.e the set of domains (subsets) of CPUs such that the
529 * cpus_allowed of every cpuset marked is_sched_load_balance
530 * is a subset of one of these domains, while there are as
531 * many such domains as possible, each as small as possible.
532 * doms - Conversion of 'csa' to an array of cpumasks, for passing to
533 * the kernel/sched.c routine partition_sched_domains() in a
534 * convenient format, that can be easily compared to the prior
535 * value to determine what partition elements (sched domains)
536 * were changed (added or removed.)
537 *
538 * Finding the best partition (set of domains):
539 * The triple nested loops below over i, j, k scan over the
540 * load balanced cpusets (using the array of cpuset pointers in
541 * csa[]) looking for pairs of cpusets that have overlapping
542 * cpus_allowed, but which don't have the same 'pn' partition
543 * number and gives them in the same partition number. It keeps
544 * looping on the 'restart' label until it can no longer find
545 * any such pairs.
546 *
547 * The union of the cpus_allowed masks from the set of
548 * all cpusets having the same 'pn' value then form the one
549 * element of the partition (one sched domain) to be passed to
550 * partition_sched_domains().
551 */
generate_sched_domains(cpumask_var_t ** domains,struct sched_domain_attr ** attributes)552 static int generate_sched_domains(cpumask_var_t **domains,
553 struct sched_domain_attr **attributes)
554 {
555 LIST_HEAD(q); /* queue of cpusets to be scanned */
556 struct cpuset *cp; /* scans q */
557 struct cpuset **csa; /* array of all cpuset ptrs */
558 int csn; /* how many cpuset ptrs in csa so far */
559 int i, j, k; /* indices for partition finding loops */
560 cpumask_var_t *doms; /* resulting partition; i.e. sched domains */
561 struct sched_domain_attr *dattr; /* attributes for custom domains */
562 int ndoms = 0; /* number of sched domains in result */
563 int nslot; /* next empty doms[] struct cpumask slot */
564
565 doms = NULL;
566 dattr = NULL;
567 csa = NULL;
568
569 /* Special case for the 99% of systems with one, full, sched domain */
570 if (is_sched_load_balance(&top_cpuset)) {
571 ndoms = 1;
572 doms = alloc_sched_domains(ndoms);
573 if (!doms)
574 goto done;
575
576 dattr = kmalloc(sizeof(struct sched_domain_attr), GFP_KERNEL);
577 if (dattr) {
578 *dattr = SD_ATTR_INIT;
579 update_domain_attr_tree(dattr, &top_cpuset);
580 }
581 cpumask_copy(doms[0], top_cpuset.cpus_allowed);
582
583 goto done;
584 }
585
586 csa = kmalloc(number_of_cpusets * sizeof(cp), GFP_KERNEL);
587 if (!csa)
588 goto done;
589 csn = 0;
590
591 list_add(&top_cpuset.stack_list, &q);
592 while (!list_empty(&q)) {
593 struct cgroup *cont;
594 struct cpuset *child; /* scans child cpusets of cp */
595
596 cp = list_first_entry(&q, struct cpuset, stack_list);
597 list_del(q.next);
598
599 if (cpumask_empty(cp->cpus_allowed))
600 continue;
601
602 /*
603 * All child cpusets contain a subset of the parent's cpus, so
604 * just skip them, and then we call update_domain_attr_tree()
605 * to calc relax_domain_level of the corresponding sched
606 * domain.
607 */
608 if (is_sched_load_balance(cp)) {
609 csa[csn++] = cp;
610 continue;
611 }
612
613 list_for_each_entry(cont, &cp->css.cgroup->children, sibling) {
614 child = cgroup_cs(cont);
615 list_add_tail(&child->stack_list, &q);
616 }
617 }
618
619 for (i = 0; i < csn; i++)
620 csa[i]->pn = i;
621 ndoms = csn;
622
623 restart:
624 /* Find the best partition (set of sched domains) */
625 for (i = 0; i < csn; i++) {
626 struct cpuset *a = csa[i];
627 int apn = a->pn;
628
629 for (j = 0; j < csn; j++) {
630 struct cpuset *b = csa[j];
631 int bpn = b->pn;
632
633 if (apn != bpn && cpusets_overlap(a, b)) {
634 for (k = 0; k < csn; k++) {
635 struct cpuset *c = csa[k];
636
637 if (c->pn == bpn)
638 c->pn = apn;
639 }
640 ndoms--; /* one less element */
641 goto restart;
642 }
643 }
644 }
645
646 /*
647 * Now we know how many domains to create.
648 * Convert <csn, csa> to <ndoms, doms> and populate cpu masks.
649 */
650 doms = alloc_sched_domains(ndoms);
651 if (!doms)
652 goto done;
653
654 /*
655 * The rest of the code, including the scheduler, can deal with
656 * dattr==NULL case. No need to abort if alloc fails.
657 */
658 dattr = kmalloc(ndoms * sizeof(struct sched_domain_attr), GFP_KERNEL);
659
660 for (nslot = 0, i = 0; i < csn; i++) {
661 struct cpuset *a = csa[i];
662 struct cpumask *dp;
663 int apn = a->pn;
664
665 if (apn < 0) {
666 /* Skip completed partitions */
667 continue;
668 }
669
670 dp = doms[nslot];
671
672 if (nslot == ndoms) {
673 static int warnings = 10;
674 if (warnings) {
675 printk(KERN_WARNING
676 "rebuild_sched_domains confused:"
677 " nslot %d, ndoms %d, csn %d, i %d,"
678 " apn %d\n",
679 nslot, ndoms, csn, i, apn);
680 warnings--;
681 }
682 continue;
683 }
684
685 cpumask_clear(dp);
686 if (dattr)
687 *(dattr + nslot) = SD_ATTR_INIT;
688 for (j = i; j < csn; j++) {
689 struct cpuset *b = csa[j];
690
691 if (apn == b->pn) {
692 cpumask_or(dp, dp, b->cpus_allowed);
693 if (dattr)
694 update_domain_attr_tree(dattr + nslot, b);
695
696 /* Done with this partition */
697 b->pn = -1;
698 }
699 }
700 nslot++;
701 }
702 BUG_ON(nslot != ndoms);
703
704 done:
705 kfree(csa);
706
707 /*
708 * Fallback to the default domain if kmalloc() failed.
709 * See comments in partition_sched_domains().
710 */
711 if (doms == NULL)
712 ndoms = 1;
713
714 *domains = doms;
715 *attributes = dattr;
716 return ndoms;
717 }
718
719 /*
720 * Rebuild scheduler domains.
721 *
722 * Call with neither cgroup_mutex held nor within get_online_cpus().
723 * Takes both cgroup_mutex and get_online_cpus().
724 *
725 * Cannot be directly called from cpuset code handling changes
726 * to the cpuset pseudo-filesystem, because it cannot be called
727 * from code that already holds cgroup_mutex.
728 */
do_rebuild_sched_domains(struct work_struct * unused)729 static void do_rebuild_sched_domains(struct work_struct *unused)
730 {
731 struct sched_domain_attr *attr;
732 cpumask_var_t *doms;
733 int ndoms;
734
735 get_online_cpus();
736
737 /* Generate domain masks and attrs */
738 cgroup_lock();
739 ndoms = generate_sched_domains(&doms, &attr);
740 cgroup_unlock();
741
742 /* Have scheduler rebuild the domains */
743 partition_sched_domains(ndoms, doms, attr);
744
745 put_online_cpus();
746 }
747 #else /* !CONFIG_SMP */
do_rebuild_sched_domains(struct work_struct * unused)748 static void do_rebuild_sched_domains(struct work_struct *unused)
749 {
750 }
751
generate_sched_domains(cpumask_var_t ** domains,struct sched_domain_attr ** attributes)752 static int generate_sched_domains(cpumask_var_t **domains,
753 struct sched_domain_attr **attributes)
754 {
755 *domains = NULL;
756 return 1;
757 }
758 #endif /* CONFIG_SMP */
759
760 static DECLARE_WORK(rebuild_sched_domains_work, do_rebuild_sched_domains);
761
762 /*
763 * Rebuild scheduler domains, asynchronously via workqueue.
764 *
765 * If the flag 'sched_load_balance' of any cpuset with non-empty
766 * 'cpus' changes, or if the 'cpus' allowed changes in any cpuset
767 * which has that flag enabled, or if any cpuset with a non-empty
768 * 'cpus' is removed, then call this routine to rebuild the
769 * scheduler's dynamic sched domains.
770 *
771 * The rebuild_sched_domains() and partition_sched_domains()
772 * routines must nest cgroup_lock() inside get_online_cpus(),
773 * but such cpuset changes as these must nest that locking the
774 * other way, holding cgroup_lock() for much of the code.
775 *
776 * So in order to avoid an ABBA deadlock, the cpuset code handling
777 * these user changes delegates the actual sched domain rebuilding
778 * to a separate workqueue thread, which ends up processing the
779 * above do_rebuild_sched_domains() function.
780 */
async_rebuild_sched_domains(void)781 static void async_rebuild_sched_domains(void)
782 {
783 queue_work(cpuset_wq, &rebuild_sched_domains_work);
784 }
785
786 /*
787 * Accomplishes the same scheduler domain rebuild as the above
788 * async_rebuild_sched_domains(), however it directly calls the
789 * rebuild routine synchronously rather than calling it via an
790 * asynchronous work thread.
791 *
792 * This can only be called from code that is not holding
793 * cgroup_mutex (not nested in a cgroup_lock() call.)
794 */
rebuild_sched_domains(void)795 void rebuild_sched_domains(void)
796 {
797 do_rebuild_sched_domains(NULL);
798 }
799
800 /**
801 * cpuset_test_cpumask - test a task's cpus_allowed versus its cpuset's
802 * @tsk: task to test
803 * @scan: struct cgroup_scanner contained in its struct cpuset_hotplug_scanner
804 *
805 * Call with cgroup_mutex held. May take callback_mutex during call.
806 * Called for each task in a cgroup by cgroup_scan_tasks().
807 * Return nonzero if this tasks's cpus_allowed mask should be changed (in other
808 * words, if its mask is not equal to its cpuset's mask).
809 */
cpuset_test_cpumask(struct task_struct * tsk,struct cgroup_scanner * scan)810 static int cpuset_test_cpumask(struct task_struct *tsk,
811 struct cgroup_scanner *scan)
812 {
813 return !cpumask_equal(&tsk->cpus_allowed,
814 (cgroup_cs(scan->cg))->cpus_allowed);
815 }
816
817 /**
818 * cpuset_change_cpumask - make a task's cpus_allowed the same as its cpuset's
819 * @tsk: task to test
820 * @scan: struct cgroup_scanner containing the cgroup of the task
821 *
822 * Called by cgroup_scan_tasks() for each task in a cgroup whose
823 * cpus_allowed mask needs to be changed.
824 *
825 * We don't need to re-check for the cgroup/cpuset membership, since we're
826 * holding cgroup_lock() at this point.
827 */
cpuset_change_cpumask(struct task_struct * tsk,struct cgroup_scanner * scan)828 static void cpuset_change_cpumask(struct task_struct *tsk,
829 struct cgroup_scanner *scan)
830 {
831 set_cpus_allowed_ptr(tsk, ((cgroup_cs(scan->cg))->cpus_allowed));
832 }
833
834 /**
835 * update_tasks_cpumask - Update the cpumasks of tasks in the cpuset.
836 * @cs: the cpuset in which each task's cpus_allowed mask needs to be changed
837 * @heap: if NULL, defer allocating heap memory to cgroup_scan_tasks()
838 *
839 * Called with cgroup_mutex held
840 *
841 * The cgroup_scan_tasks() function will scan all the tasks in a cgroup,
842 * calling callback functions for each.
843 *
844 * No return value. It's guaranteed that cgroup_scan_tasks() always returns 0
845 * if @heap != NULL.
846 */
update_tasks_cpumask(struct cpuset * cs,struct ptr_heap * heap)847 static void update_tasks_cpumask(struct cpuset *cs, struct ptr_heap *heap)
848 {
849 struct cgroup_scanner scan;
850
851 scan.cg = cs->css.cgroup;
852 scan.test_task = cpuset_test_cpumask;
853 scan.process_task = cpuset_change_cpumask;
854 scan.heap = heap;
855 cgroup_scan_tasks(&scan);
856 }
857
858 /**
859 * update_cpumask - update the cpus_allowed mask of a cpuset and all tasks in it
860 * @cs: the cpuset to consider
861 * @buf: buffer of cpu numbers written to this cpuset
862 */
update_cpumask(struct cpuset * cs,struct cpuset * trialcs,const char * buf)863 static int update_cpumask(struct cpuset *cs, struct cpuset *trialcs,
864 const char *buf)
865 {
866 struct ptr_heap heap;
867 int retval;
868 int is_load_balanced;
869
870 /* top_cpuset.cpus_allowed tracks cpu_online_mask; it's read-only */
871 if (cs == &top_cpuset)
872 return -EACCES;
873
874 /*
875 * An empty cpus_allowed is ok only if the cpuset has no tasks.
876 * Since cpulist_parse() fails on an empty mask, we special case
877 * that parsing. The validate_change() call ensures that cpusets
878 * with tasks have cpus.
879 */
880 if (!*buf) {
881 cpumask_clear(trialcs->cpus_allowed);
882 } else {
883 retval = cpulist_parse(buf, trialcs->cpus_allowed);
884 if (retval < 0)
885 return retval;
886
887 if (!cpumask_subset(trialcs->cpus_allowed, cpu_active_mask))
888 return -EINVAL;
889 }
890 retval = validate_change(cs, trialcs);
891 if (retval < 0)
892 return retval;
893
894 /* Nothing to do if the cpus didn't change */
895 if (cpumask_equal(cs->cpus_allowed, trialcs->cpus_allowed))
896 return 0;
897
898 retval = heap_init(&heap, PAGE_SIZE, GFP_KERNEL, NULL);
899 if (retval)
900 return retval;
901
902 is_load_balanced = is_sched_load_balance(trialcs);
903
904 mutex_lock(&callback_mutex);
905 cpumask_copy(cs->cpus_allowed, trialcs->cpus_allowed);
906 mutex_unlock(&callback_mutex);
907
908 /*
909 * Scan tasks in the cpuset, and update the cpumasks of any
910 * that need an update.
911 */
912 update_tasks_cpumask(cs, &heap);
913
914 heap_free(&heap);
915
916 if (is_load_balanced)
917 async_rebuild_sched_domains();
918 return 0;
919 }
920
921 /*
922 * cpuset_migrate_mm
923 *
924 * Migrate memory region from one set of nodes to another.
925 *
926 * Temporarilly set tasks mems_allowed to target nodes of migration,
927 * so that the migration code can allocate pages on these nodes.
928 *
929 * Call holding cgroup_mutex, so current's cpuset won't change
930 * during this call, as manage_mutex holds off any cpuset_attach()
931 * calls. Therefore we don't need to take task_lock around the
932 * call to guarantee_online_mems(), as we know no one is changing
933 * our task's cpuset.
934 *
935 * While the mm_struct we are migrating is typically from some
936 * other task, the task_struct mems_allowed that we are hacking
937 * is for our current task, which must allocate new pages for that
938 * migrating memory region.
939 */
940
cpuset_migrate_mm(struct mm_struct * mm,const nodemask_t * from,const nodemask_t * to)941 static void cpuset_migrate_mm(struct mm_struct *mm, const nodemask_t *from,
942 const nodemask_t *to)
943 {
944 struct task_struct *tsk = current;
945
946 tsk->mems_allowed = *to;
947
948 do_migrate_pages(mm, from, to, MPOL_MF_MOVE_ALL);
949
950 guarantee_online_mems(task_cs(tsk),&tsk->mems_allowed);
951 }
952
953 /*
954 * cpuset_change_task_nodemask - change task's mems_allowed and mempolicy
955 * @tsk: the task to change
956 * @newmems: new nodes that the task will be set
957 *
958 * In order to avoid seeing no nodes if the old and new nodes are disjoint,
959 * we structure updates as setting all new allowed nodes, then clearing newly
960 * disallowed ones.
961 */
cpuset_change_task_nodemask(struct task_struct * tsk,nodemask_t * newmems)962 static void cpuset_change_task_nodemask(struct task_struct *tsk,
963 nodemask_t *newmems)
964 {
965 bool need_loop;
966
967 /*
968 * Allow tasks that have access to memory reserves because they have
969 * been OOM killed to get memory anywhere.
970 */
971 if (unlikely(test_thread_flag(TIF_MEMDIE)))
972 return;
973 if (current->flags & PF_EXITING) /* Let dying task have memory */
974 return;
975
976 task_lock(tsk);
977 /*
978 * Determine if a loop is necessary if another thread is doing
979 * get_mems_allowed(). If at least one node remains unchanged and
980 * tsk does not have a mempolicy, then an empty nodemask will not be
981 * possible when mems_allowed is larger than a word.
982 */
983 need_loop = task_has_mempolicy(tsk) ||
984 !nodes_intersects(*newmems, tsk->mems_allowed);
985
986 if (need_loop) {
987 local_irq_disable();
988 write_seqcount_begin(&tsk->mems_allowed_seq);
989 }
990
991 nodes_or(tsk->mems_allowed, tsk->mems_allowed, *newmems);
992 mpol_rebind_task(tsk, newmems, MPOL_REBIND_STEP1);
993
994 mpol_rebind_task(tsk, newmems, MPOL_REBIND_STEP2);
995 tsk->mems_allowed = *newmems;
996
997 if (need_loop) {
998 write_seqcount_end(&tsk->mems_allowed_seq);
999 local_irq_enable();
1000 }
1001
1002 task_unlock(tsk);
1003 }
1004
1005 /*
1006 * Update task's mems_allowed and rebind its mempolicy and vmas' mempolicy
1007 * of it to cpuset's new mems_allowed, and migrate pages to new nodes if
1008 * memory_migrate flag is set. Called with cgroup_mutex held.
1009 */
cpuset_change_nodemask(struct task_struct * p,struct cgroup_scanner * scan)1010 static void cpuset_change_nodemask(struct task_struct *p,
1011 struct cgroup_scanner *scan)
1012 {
1013 struct mm_struct *mm;
1014 struct cpuset *cs;
1015 int migrate;
1016 const nodemask_t *oldmem = scan->data;
1017 static nodemask_t newmems; /* protected by cgroup_mutex */
1018
1019 cs = cgroup_cs(scan->cg);
1020 guarantee_online_mems(cs, &newmems);
1021
1022 cpuset_change_task_nodemask(p, &newmems);
1023
1024 mm = get_task_mm(p);
1025 if (!mm)
1026 return;
1027
1028 migrate = is_memory_migrate(cs);
1029
1030 mpol_rebind_mm(mm, &cs->mems_allowed);
1031 if (migrate)
1032 cpuset_migrate_mm(mm, oldmem, &cs->mems_allowed);
1033 mmput(mm);
1034 }
1035
1036 static void *cpuset_being_rebound;
1037
1038 /**
1039 * update_tasks_nodemask - Update the nodemasks of tasks in the cpuset.
1040 * @cs: the cpuset in which each task's mems_allowed mask needs to be changed
1041 * @oldmem: old mems_allowed of cpuset cs
1042 * @heap: if NULL, defer allocating heap memory to cgroup_scan_tasks()
1043 *
1044 * Called with cgroup_mutex held
1045 * No return value. It's guaranteed that cgroup_scan_tasks() always returns 0
1046 * if @heap != NULL.
1047 */
update_tasks_nodemask(struct cpuset * cs,const nodemask_t * oldmem,struct ptr_heap * heap)1048 static void update_tasks_nodemask(struct cpuset *cs, const nodemask_t *oldmem,
1049 struct ptr_heap *heap)
1050 {
1051 struct cgroup_scanner scan;
1052
1053 cpuset_being_rebound = cs; /* causes mpol_dup() rebind */
1054
1055 scan.cg = cs->css.cgroup;
1056 scan.test_task = NULL;
1057 scan.process_task = cpuset_change_nodemask;
1058 scan.heap = heap;
1059 scan.data = (nodemask_t *)oldmem;
1060
1061 /*
1062 * The mpol_rebind_mm() call takes mmap_sem, which we couldn't
1063 * take while holding tasklist_lock. Forks can happen - the
1064 * mpol_dup() cpuset_being_rebound check will catch such forks,
1065 * and rebind their vma mempolicies too. Because we still hold
1066 * the global cgroup_mutex, we know that no other rebind effort
1067 * will be contending for the global variable cpuset_being_rebound.
1068 * It's ok if we rebind the same mm twice; mpol_rebind_mm()
1069 * is idempotent. Also migrate pages in each mm to new nodes.
1070 */
1071 cgroup_scan_tasks(&scan);
1072
1073 /* We're done rebinding vmas to this cpuset's new mems_allowed. */
1074 cpuset_being_rebound = NULL;
1075 }
1076
1077 /*
1078 * Handle user request to change the 'mems' memory placement
1079 * of a cpuset. Needs to validate the request, update the
1080 * cpusets mems_allowed, and for each task in the cpuset,
1081 * update mems_allowed and rebind task's mempolicy and any vma
1082 * mempolicies and if the cpuset is marked 'memory_migrate',
1083 * migrate the tasks pages to the new memory.
1084 *
1085 * Call with cgroup_mutex held. May take callback_mutex during call.
1086 * Will take tasklist_lock, scan tasklist for tasks in cpuset cs,
1087 * lock each such tasks mm->mmap_sem, scan its vma's and rebind
1088 * their mempolicies to the cpusets new mems_allowed.
1089 */
update_nodemask(struct cpuset * cs,struct cpuset * trialcs,const char * buf)1090 static int update_nodemask(struct cpuset *cs, struct cpuset *trialcs,
1091 const char *buf)
1092 {
1093 NODEMASK_ALLOC(nodemask_t, oldmem, GFP_KERNEL);
1094 int retval;
1095 struct ptr_heap heap;
1096
1097 if (!oldmem)
1098 return -ENOMEM;
1099
1100 /*
1101 * top_cpuset.mems_allowed tracks node_stats[N_HIGH_MEMORY];
1102 * it's read-only
1103 */
1104 if (cs == &top_cpuset) {
1105 retval = -EACCES;
1106 goto done;
1107 }
1108
1109 /*
1110 * An empty mems_allowed is ok iff there are no tasks in the cpuset.
1111 * Since nodelist_parse() fails on an empty mask, we special case
1112 * that parsing. The validate_change() call ensures that cpusets
1113 * with tasks have memory.
1114 */
1115 if (!*buf) {
1116 nodes_clear(trialcs->mems_allowed);
1117 } else {
1118 retval = nodelist_parse(buf, trialcs->mems_allowed);
1119 if (retval < 0)
1120 goto done;
1121
1122 if (!nodes_subset(trialcs->mems_allowed,
1123 node_states[N_HIGH_MEMORY])) {
1124 retval = -EINVAL;
1125 goto done;
1126 }
1127 }
1128 *oldmem = cs->mems_allowed;
1129 if (nodes_equal(*oldmem, trialcs->mems_allowed)) {
1130 retval = 0; /* Too easy - nothing to do */
1131 goto done;
1132 }
1133 retval = validate_change(cs, trialcs);
1134 if (retval < 0)
1135 goto done;
1136
1137 retval = heap_init(&heap, PAGE_SIZE, GFP_KERNEL, NULL);
1138 if (retval < 0)
1139 goto done;
1140
1141 mutex_lock(&callback_mutex);
1142 cs->mems_allowed = trialcs->mems_allowed;
1143 mutex_unlock(&callback_mutex);
1144
1145 update_tasks_nodemask(cs, oldmem, &heap);
1146
1147 heap_free(&heap);
1148 done:
1149 NODEMASK_FREE(oldmem);
1150 return retval;
1151 }
1152
current_cpuset_is_being_rebound(void)1153 int current_cpuset_is_being_rebound(void)
1154 {
1155 int ret;
1156
1157 rcu_read_lock();
1158 ret = task_cs(current) == cpuset_being_rebound;
1159 rcu_read_unlock();
1160
1161 return ret;
1162 }
1163
update_relax_domain_level(struct cpuset * cs,s64 val)1164 static int update_relax_domain_level(struct cpuset *cs, s64 val)
1165 {
1166 #ifdef CONFIG_SMP
1167 if (val < -1 || val >= sched_domain_level_max)
1168 return -EINVAL;
1169 #endif
1170
1171 if (val != cs->relax_domain_level) {
1172 cs->relax_domain_level = val;
1173 if (!cpumask_empty(cs->cpus_allowed) &&
1174 is_sched_load_balance(cs))
1175 async_rebuild_sched_domains();
1176 }
1177
1178 return 0;
1179 }
1180
1181 /*
1182 * cpuset_change_flag - make a task's spread flags the same as its cpuset's
1183 * @tsk: task to be updated
1184 * @scan: struct cgroup_scanner containing the cgroup of the task
1185 *
1186 * Called by cgroup_scan_tasks() for each task in a cgroup.
1187 *
1188 * We don't need to re-check for the cgroup/cpuset membership, since we're
1189 * holding cgroup_lock() at this point.
1190 */
cpuset_change_flag(struct task_struct * tsk,struct cgroup_scanner * scan)1191 static void cpuset_change_flag(struct task_struct *tsk,
1192 struct cgroup_scanner *scan)
1193 {
1194 cpuset_update_task_spread_flag(cgroup_cs(scan->cg), tsk);
1195 }
1196
1197 /*
1198 * update_tasks_flags - update the spread flags of tasks in the cpuset.
1199 * @cs: the cpuset in which each task's spread flags needs to be changed
1200 * @heap: if NULL, defer allocating heap memory to cgroup_scan_tasks()
1201 *
1202 * Called with cgroup_mutex held
1203 *
1204 * The cgroup_scan_tasks() function will scan all the tasks in a cgroup,
1205 * calling callback functions for each.
1206 *
1207 * No return value. It's guaranteed that cgroup_scan_tasks() always returns 0
1208 * if @heap != NULL.
1209 */
update_tasks_flags(struct cpuset * cs,struct ptr_heap * heap)1210 static void update_tasks_flags(struct cpuset *cs, struct ptr_heap *heap)
1211 {
1212 struct cgroup_scanner scan;
1213
1214 scan.cg = cs->css.cgroup;
1215 scan.test_task = NULL;
1216 scan.process_task = cpuset_change_flag;
1217 scan.heap = heap;
1218 cgroup_scan_tasks(&scan);
1219 }
1220
1221 /*
1222 * update_flag - read a 0 or a 1 in a file and update associated flag
1223 * bit: the bit to update (see cpuset_flagbits_t)
1224 * cs: the cpuset to update
1225 * turning_on: whether the flag is being set or cleared
1226 *
1227 * Call with cgroup_mutex held.
1228 */
1229
update_flag(cpuset_flagbits_t bit,struct cpuset * cs,int turning_on)1230 static int update_flag(cpuset_flagbits_t bit, struct cpuset *cs,
1231 int turning_on)
1232 {
1233 struct cpuset *trialcs;
1234 int balance_flag_changed;
1235 int spread_flag_changed;
1236 struct ptr_heap heap;
1237 int err;
1238
1239 trialcs = alloc_trial_cpuset(cs);
1240 if (!trialcs)
1241 return -ENOMEM;
1242
1243 if (turning_on)
1244 set_bit(bit, &trialcs->flags);
1245 else
1246 clear_bit(bit, &trialcs->flags);
1247
1248 err = validate_change(cs, trialcs);
1249 if (err < 0)
1250 goto out;
1251
1252 err = heap_init(&heap, PAGE_SIZE, GFP_KERNEL, NULL);
1253 if (err < 0)
1254 goto out;
1255
1256 balance_flag_changed = (is_sched_load_balance(cs) !=
1257 is_sched_load_balance(trialcs));
1258
1259 spread_flag_changed = ((is_spread_slab(cs) != is_spread_slab(trialcs))
1260 || (is_spread_page(cs) != is_spread_page(trialcs)));
1261
1262 mutex_lock(&callback_mutex);
1263 cs->flags = trialcs->flags;
1264 mutex_unlock(&callback_mutex);
1265
1266 if (!cpumask_empty(trialcs->cpus_allowed) && balance_flag_changed)
1267 async_rebuild_sched_domains();
1268
1269 if (spread_flag_changed)
1270 update_tasks_flags(cs, &heap);
1271 heap_free(&heap);
1272 out:
1273 free_trial_cpuset(trialcs);
1274 return err;
1275 }
1276
1277 /*
1278 * Frequency meter - How fast is some event occurring?
1279 *
1280 * These routines manage a digitally filtered, constant time based,
1281 * event frequency meter. There are four routines:
1282 * fmeter_init() - initialize a frequency meter.
1283 * fmeter_markevent() - called each time the event happens.
1284 * fmeter_getrate() - returns the recent rate of such events.
1285 * fmeter_update() - internal routine used to update fmeter.
1286 *
1287 * A common data structure is passed to each of these routines,
1288 * which is used to keep track of the state required to manage the
1289 * frequency meter and its digital filter.
1290 *
1291 * The filter works on the number of events marked per unit time.
1292 * The filter is single-pole low-pass recursive (IIR). The time unit
1293 * is 1 second. Arithmetic is done using 32-bit integers scaled to
1294 * simulate 3 decimal digits of precision (multiplied by 1000).
1295 *
1296 * With an FM_COEF of 933, and a time base of 1 second, the filter
1297 * has a half-life of 10 seconds, meaning that if the events quit
1298 * happening, then the rate returned from the fmeter_getrate()
1299 * will be cut in half each 10 seconds, until it converges to zero.
1300 *
1301 * It is not worth doing a real infinitely recursive filter. If more
1302 * than FM_MAXTICKS ticks have elapsed since the last filter event,
1303 * just compute FM_MAXTICKS ticks worth, by which point the level
1304 * will be stable.
1305 *
1306 * Limit the count of unprocessed events to FM_MAXCNT, so as to avoid
1307 * arithmetic overflow in the fmeter_update() routine.
1308 *
1309 * Given the simple 32 bit integer arithmetic used, this meter works
1310 * best for reporting rates between one per millisecond (msec) and
1311 * one per 32 (approx) seconds. At constant rates faster than one
1312 * per msec it maxes out at values just under 1,000,000. At constant
1313 * rates between one per msec, and one per second it will stabilize
1314 * to a value N*1000, where N is the rate of events per second.
1315 * At constant rates between one per second and one per 32 seconds,
1316 * it will be choppy, moving up on the seconds that have an event,
1317 * and then decaying until the next event. At rates slower than
1318 * about one in 32 seconds, it decays all the way back to zero between
1319 * each event.
1320 */
1321
1322 #define FM_COEF 933 /* coefficient for half-life of 10 secs */
1323 #define FM_MAXTICKS ((time_t)99) /* useless computing more ticks than this */
1324 #define FM_MAXCNT 1000000 /* limit cnt to avoid overflow */
1325 #define FM_SCALE 1000 /* faux fixed point scale */
1326
1327 /* Initialize a frequency meter */
fmeter_init(struct fmeter * fmp)1328 static void fmeter_init(struct fmeter *fmp)
1329 {
1330 fmp->cnt = 0;
1331 fmp->val = 0;
1332 fmp->time = 0;
1333 spin_lock_init(&fmp->lock);
1334 }
1335
1336 /* Internal meter update - process cnt events and update value */
fmeter_update(struct fmeter * fmp)1337 static void fmeter_update(struct fmeter *fmp)
1338 {
1339 time_t now = get_seconds();
1340 time_t ticks = now - fmp->time;
1341
1342 if (ticks == 0)
1343 return;
1344
1345 ticks = min(FM_MAXTICKS, ticks);
1346 while (ticks-- > 0)
1347 fmp->val = (FM_COEF * fmp->val) / FM_SCALE;
1348 fmp->time = now;
1349
1350 fmp->val += ((FM_SCALE - FM_COEF) * fmp->cnt) / FM_SCALE;
1351 fmp->cnt = 0;
1352 }
1353
1354 /* Process any previous ticks, then bump cnt by one (times scale). */
fmeter_markevent(struct fmeter * fmp)1355 static void fmeter_markevent(struct fmeter *fmp)
1356 {
1357 spin_lock(&fmp->lock);
1358 fmeter_update(fmp);
1359 fmp->cnt = min(FM_MAXCNT, fmp->cnt + FM_SCALE);
1360 spin_unlock(&fmp->lock);
1361 }
1362
1363 /* Process any previous ticks, then return current value. */
fmeter_getrate(struct fmeter * fmp)1364 static int fmeter_getrate(struct fmeter *fmp)
1365 {
1366 int val;
1367
1368 spin_lock(&fmp->lock);
1369 fmeter_update(fmp);
1370 val = fmp->val;
1371 spin_unlock(&fmp->lock);
1372 return val;
1373 }
1374
1375 /*
1376 * Protected by cgroup_lock. The nodemasks must be stored globally because
1377 * dynamically allocating them is not allowed in can_attach, and they must
1378 * persist until attach.
1379 */
1380 static cpumask_var_t cpus_attach;
1381 static nodemask_t cpuset_attach_nodemask_from;
1382 static nodemask_t cpuset_attach_nodemask_to;
1383
1384 /* Called by cgroups to determine if a cpuset is usable; cgroup_mutex held */
cpuset_can_attach(struct cgroup * cgrp,struct cgroup_taskset * tset)1385 static int cpuset_can_attach(struct cgroup *cgrp, struct cgroup_taskset *tset)
1386 {
1387 struct cpuset *cs = cgroup_cs(cgrp);
1388 struct task_struct *task;
1389 int ret;
1390
1391 if (cpumask_empty(cs->cpus_allowed) || nodes_empty(cs->mems_allowed))
1392 return -ENOSPC;
1393
1394 cgroup_taskset_for_each(task, cgrp, tset) {
1395 /*
1396 * Kthreads bound to specific cpus cannot be moved to a new
1397 * cpuset; we cannot change their cpu affinity and
1398 * isolating such threads by their set of allowed nodes is
1399 * unnecessary. Thus, cpusets are not applicable for such
1400 * threads. This prevents checking for success of
1401 * set_cpus_allowed_ptr() on all attached tasks before
1402 * cpus_allowed may be changed.
1403 */
1404 if (task->flags & PF_THREAD_BOUND)
1405 return -EINVAL;
1406 if ((ret = security_task_setscheduler(task)))
1407 return ret;
1408 }
1409
1410 /* prepare for attach */
1411 if (cs == &top_cpuset)
1412 cpumask_copy(cpus_attach, cpu_possible_mask);
1413 else
1414 guarantee_online_cpus(cs, cpus_attach);
1415
1416 guarantee_online_mems(cs, &cpuset_attach_nodemask_to);
1417
1418 return 0;
1419 }
1420
cpuset_attach(struct cgroup * cgrp,struct cgroup_taskset * tset)1421 static void cpuset_attach(struct cgroup *cgrp, struct cgroup_taskset *tset)
1422 {
1423 struct mm_struct *mm;
1424 struct task_struct *task;
1425 struct task_struct *leader = cgroup_taskset_first(tset);
1426 struct cgroup *oldcgrp = cgroup_taskset_cur_cgroup(tset);
1427 struct cpuset *cs = cgroup_cs(cgrp);
1428 struct cpuset *oldcs = cgroup_cs(oldcgrp);
1429
1430 cgroup_taskset_for_each(task, cgrp, tset) {
1431 /*
1432 * can_attach beforehand should guarantee that this doesn't
1433 * fail. TODO: have a better way to handle failure here
1434 */
1435 WARN_ON_ONCE(set_cpus_allowed_ptr(task, cpus_attach));
1436
1437 cpuset_change_task_nodemask(task, &cpuset_attach_nodemask_to);
1438 cpuset_update_task_spread_flag(cs, task);
1439 }
1440
1441 /*
1442 * Change mm, possibly for multiple threads in a threadgroup. This is
1443 * expensive and may sleep.
1444 */
1445 cpuset_attach_nodemask_from = oldcs->mems_allowed;
1446 cpuset_attach_nodemask_to = cs->mems_allowed;
1447 mm = get_task_mm(leader);
1448 if (mm) {
1449 mpol_rebind_mm(mm, &cpuset_attach_nodemask_to);
1450 if (is_memory_migrate(cs))
1451 cpuset_migrate_mm(mm, &cpuset_attach_nodemask_from,
1452 &cpuset_attach_nodemask_to);
1453 mmput(mm);
1454 }
1455 }
1456
1457 /* The various types of files and directories in a cpuset file system */
1458
1459 typedef enum {
1460 FILE_MEMORY_MIGRATE,
1461 FILE_CPULIST,
1462 FILE_MEMLIST,
1463 FILE_CPU_EXCLUSIVE,
1464 FILE_MEM_EXCLUSIVE,
1465 FILE_MEM_HARDWALL,
1466 FILE_SCHED_LOAD_BALANCE,
1467 FILE_SCHED_RELAX_DOMAIN_LEVEL,
1468 FILE_MEMORY_PRESSURE_ENABLED,
1469 FILE_MEMORY_PRESSURE,
1470 FILE_SPREAD_PAGE,
1471 FILE_SPREAD_SLAB,
1472 } cpuset_filetype_t;
1473
cpuset_write_u64(struct cgroup * cgrp,struct cftype * cft,u64 val)1474 static int cpuset_write_u64(struct cgroup *cgrp, struct cftype *cft, u64 val)
1475 {
1476 int retval = 0;
1477 struct cpuset *cs = cgroup_cs(cgrp);
1478 cpuset_filetype_t type = cft->private;
1479
1480 if (!cgroup_lock_live_group(cgrp))
1481 return -ENODEV;
1482
1483 switch (type) {
1484 case FILE_CPU_EXCLUSIVE:
1485 retval = update_flag(CS_CPU_EXCLUSIVE, cs, val);
1486 break;
1487 case FILE_MEM_EXCLUSIVE:
1488 retval = update_flag(CS_MEM_EXCLUSIVE, cs, val);
1489 break;
1490 case FILE_MEM_HARDWALL:
1491 retval = update_flag(CS_MEM_HARDWALL, cs, val);
1492 break;
1493 case FILE_SCHED_LOAD_BALANCE:
1494 retval = update_flag(CS_SCHED_LOAD_BALANCE, cs, val);
1495 break;
1496 case FILE_MEMORY_MIGRATE:
1497 retval = update_flag(CS_MEMORY_MIGRATE, cs, val);
1498 break;
1499 case FILE_MEMORY_PRESSURE_ENABLED:
1500 cpuset_memory_pressure_enabled = !!val;
1501 break;
1502 case FILE_MEMORY_PRESSURE:
1503 retval = -EACCES;
1504 break;
1505 case FILE_SPREAD_PAGE:
1506 retval = update_flag(CS_SPREAD_PAGE, cs, val);
1507 break;
1508 case FILE_SPREAD_SLAB:
1509 retval = update_flag(CS_SPREAD_SLAB, cs, val);
1510 break;
1511 default:
1512 retval = -EINVAL;
1513 break;
1514 }
1515 cgroup_unlock();
1516 return retval;
1517 }
1518
cpuset_write_s64(struct cgroup * cgrp,struct cftype * cft,s64 val)1519 static int cpuset_write_s64(struct cgroup *cgrp, struct cftype *cft, s64 val)
1520 {
1521 int retval = 0;
1522 struct cpuset *cs = cgroup_cs(cgrp);
1523 cpuset_filetype_t type = cft->private;
1524
1525 if (!cgroup_lock_live_group(cgrp))
1526 return -ENODEV;
1527
1528 switch (type) {
1529 case FILE_SCHED_RELAX_DOMAIN_LEVEL:
1530 retval = update_relax_domain_level(cs, val);
1531 break;
1532 default:
1533 retval = -EINVAL;
1534 break;
1535 }
1536 cgroup_unlock();
1537 return retval;
1538 }
1539
1540 /*
1541 * Common handling for a write to a "cpus" or "mems" file.
1542 */
cpuset_write_resmask(struct cgroup * cgrp,struct cftype * cft,const char * buf)1543 static int cpuset_write_resmask(struct cgroup *cgrp, struct cftype *cft,
1544 const char *buf)
1545 {
1546 int retval = 0;
1547 struct cpuset *cs = cgroup_cs(cgrp);
1548 struct cpuset *trialcs;
1549
1550 if (!cgroup_lock_live_group(cgrp))
1551 return -ENODEV;
1552
1553 trialcs = alloc_trial_cpuset(cs);
1554 if (!trialcs) {
1555 retval = -ENOMEM;
1556 goto out;
1557 }
1558
1559 switch (cft->private) {
1560 case FILE_CPULIST:
1561 retval = update_cpumask(cs, trialcs, buf);
1562 break;
1563 case FILE_MEMLIST:
1564 retval = update_nodemask(cs, trialcs, buf);
1565 break;
1566 default:
1567 retval = -EINVAL;
1568 break;
1569 }
1570
1571 free_trial_cpuset(trialcs);
1572 out:
1573 cgroup_unlock();
1574 return retval;
1575 }
1576
1577 /*
1578 * These ascii lists should be read in a single call, by using a user
1579 * buffer large enough to hold the entire map. If read in smaller
1580 * chunks, there is no guarantee of atomicity. Since the display format
1581 * used, list of ranges of sequential numbers, is variable length,
1582 * and since these maps can change value dynamically, one could read
1583 * gibberish by doing partial reads while a list was changing.
1584 * A single large read to a buffer that crosses a page boundary is
1585 * ok, because the result being copied to user land is not recomputed
1586 * across a page fault.
1587 */
1588
cpuset_sprintf_cpulist(char * page,struct cpuset * cs)1589 static size_t cpuset_sprintf_cpulist(char *page, struct cpuset *cs)
1590 {
1591 size_t count;
1592
1593 mutex_lock(&callback_mutex);
1594 count = cpulist_scnprintf(page, PAGE_SIZE, cs->cpus_allowed);
1595 mutex_unlock(&callback_mutex);
1596
1597 return count;
1598 }
1599
cpuset_sprintf_memlist(char * page,struct cpuset * cs)1600 static size_t cpuset_sprintf_memlist(char *page, struct cpuset *cs)
1601 {
1602 size_t count;
1603
1604 mutex_lock(&callback_mutex);
1605 count = nodelist_scnprintf(page, PAGE_SIZE, cs->mems_allowed);
1606 mutex_unlock(&callback_mutex);
1607
1608 return count;
1609 }
1610
cpuset_common_file_read(struct cgroup * cont,struct cftype * cft,struct file * file,char __user * buf,size_t nbytes,loff_t * ppos)1611 static ssize_t cpuset_common_file_read(struct cgroup *cont,
1612 struct cftype *cft,
1613 struct file *file,
1614 char __user *buf,
1615 size_t nbytes, loff_t *ppos)
1616 {
1617 struct cpuset *cs = cgroup_cs(cont);
1618 cpuset_filetype_t type = cft->private;
1619 char *page;
1620 ssize_t retval = 0;
1621 char *s;
1622
1623 if (!(page = (char *)__get_free_page(GFP_TEMPORARY)))
1624 return -ENOMEM;
1625
1626 s = page;
1627
1628 switch (type) {
1629 case FILE_CPULIST:
1630 s += cpuset_sprintf_cpulist(s, cs);
1631 break;
1632 case FILE_MEMLIST:
1633 s += cpuset_sprintf_memlist(s, cs);
1634 break;
1635 default:
1636 retval = -EINVAL;
1637 goto out;
1638 }
1639 *s++ = '\n';
1640
1641 retval = simple_read_from_buffer(buf, nbytes, ppos, page, s - page);
1642 out:
1643 free_page((unsigned long)page);
1644 return retval;
1645 }
1646
cpuset_read_u64(struct cgroup * cont,struct cftype * cft)1647 static u64 cpuset_read_u64(struct cgroup *cont, struct cftype *cft)
1648 {
1649 struct cpuset *cs = cgroup_cs(cont);
1650 cpuset_filetype_t type = cft->private;
1651 switch (type) {
1652 case FILE_CPU_EXCLUSIVE:
1653 return is_cpu_exclusive(cs);
1654 case FILE_MEM_EXCLUSIVE:
1655 return is_mem_exclusive(cs);
1656 case FILE_MEM_HARDWALL:
1657 return is_mem_hardwall(cs);
1658 case FILE_SCHED_LOAD_BALANCE:
1659 return is_sched_load_balance(cs);
1660 case FILE_MEMORY_MIGRATE:
1661 return is_memory_migrate(cs);
1662 case FILE_MEMORY_PRESSURE_ENABLED:
1663 return cpuset_memory_pressure_enabled;
1664 case FILE_MEMORY_PRESSURE:
1665 return fmeter_getrate(&cs->fmeter);
1666 case FILE_SPREAD_PAGE:
1667 return is_spread_page(cs);
1668 case FILE_SPREAD_SLAB:
1669 return is_spread_slab(cs);
1670 default:
1671 BUG();
1672 }
1673
1674 /* Unreachable but makes gcc happy */
1675 return 0;
1676 }
1677
cpuset_read_s64(struct cgroup * cont,struct cftype * cft)1678 static s64 cpuset_read_s64(struct cgroup *cont, struct cftype *cft)
1679 {
1680 struct cpuset *cs = cgroup_cs(cont);
1681 cpuset_filetype_t type = cft->private;
1682 switch (type) {
1683 case FILE_SCHED_RELAX_DOMAIN_LEVEL:
1684 return cs->relax_domain_level;
1685 default:
1686 BUG();
1687 }
1688
1689 /* Unrechable but makes gcc happy */
1690 return 0;
1691 }
1692
1693
1694 /*
1695 * for the common functions, 'private' gives the type of file
1696 */
1697
1698 static struct cftype files[] = {
1699 {
1700 .name = "cpus",
1701 .read = cpuset_common_file_read,
1702 .write_string = cpuset_write_resmask,
1703 .max_write_len = (100U + 6 * NR_CPUS),
1704 .private = FILE_CPULIST,
1705 },
1706
1707 {
1708 .name = "mems",
1709 .read = cpuset_common_file_read,
1710 .write_string = cpuset_write_resmask,
1711 .max_write_len = (100U + 6 * MAX_NUMNODES),
1712 .private = FILE_MEMLIST,
1713 },
1714
1715 {
1716 .name = "cpu_exclusive",
1717 .read_u64 = cpuset_read_u64,
1718 .write_u64 = cpuset_write_u64,
1719 .private = FILE_CPU_EXCLUSIVE,
1720 },
1721
1722 {
1723 .name = "mem_exclusive",
1724 .read_u64 = cpuset_read_u64,
1725 .write_u64 = cpuset_write_u64,
1726 .private = FILE_MEM_EXCLUSIVE,
1727 },
1728
1729 {
1730 .name = "mem_hardwall",
1731 .read_u64 = cpuset_read_u64,
1732 .write_u64 = cpuset_write_u64,
1733 .private = FILE_MEM_HARDWALL,
1734 },
1735
1736 {
1737 .name = "sched_load_balance",
1738 .read_u64 = cpuset_read_u64,
1739 .write_u64 = cpuset_write_u64,
1740 .private = FILE_SCHED_LOAD_BALANCE,
1741 },
1742
1743 {
1744 .name = "sched_relax_domain_level",
1745 .read_s64 = cpuset_read_s64,
1746 .write_s64 = cpuset_write_s64,
1747 .private = FILE_SCHED_RELAX_DOMAIN_LEVEL,
1748 },
1749
1750 {
1751 .name = "memory_migrate",
1752 .read_u64 = cpuset_read_u64,
1753 .write_u64 = cpuset_write_u64,
1754 .private = FILE_MEMORY_MIGRATE,
1755 },
1756
1757 {
1758 .name = "memory_pressure",
1759 .read_u64 = cpuset_read_u64,
1760 .write_u64 = cpuset_write_u64,
1761 .private = FILE_MEMORY_PRESSURE,
1762 .mode = S_IRUGO,
1763 },
1764
1765 {
1766 .name = "memory_spread_page",
1767 .read_u64 = cpuset_read_u64,
1768 .write_u64 = cpuset_write_u64,
1769 .private = FILE_SPREAD_PAGE,
1770 },
1771
1772 {
1773 .name = "memory_spread_slab",
1774 .read_u64 = cpuset_read_u64,
1775 .write_u64 = cpuset_write_u64,
1776 .private = FILE_SPREAD_SLAB,
1777 },
1778 };
1779
1780 static struct cftype cft_memory_pressure_enabled = {
1781 .name = "memory_pressure_enabled",
1782 .read_u64 = cpuset_read_u64,
1783 .write_u64 = cpuset_write_u64,
1784 .private = FILE_MEMORY_PRESSURE_ENABLED,
1785 };
1786
cpuset_populate(struct cgroup_subsys * ss,struct cgroup * cont)1787 static int cpuset_populate(struct cgroup_subsys *ss, struct cgroup *cont)
1788 {
1789 int err;
1790
1791 err = cgroup_add_files(cont, ss, files, ARRAY_SIZE(files));
1792 if (err)
1793 return err;
1794 /* memory_pressure_enabled is in root cpuset only */
1795 if (!cont->parent)
1796 err = cgroup_add_file(cont, ss,
1797 &cft_memory_pressure_enabled);
1798 return err;
1799 }
1800
1801 /*
1802 * post_clone() is called during cgroup_create() when the
1803 * clone_children mount argument was specified. The cgroup
1804 * can not yet have any tasks.
1805 *
1806 * Currently we refuse to set up the cgroup - thereby
1807 * refusing the task to be entered, and as a result refusing
1808 * the sys_unshare() or clone() which initiated it - if any
1809 * sibling cpusets have exclusive cpus or mem.
1810 *
1811 * If this becomes a problem for some users who wish to
1812 * allow that scenario, then cpuset_post_clone() could be
1813 * changed to grant parent->cpus_allowed-sibling_cpus_exclusive
1814 * (and likewise for mems) to the new cgroup. Called with cgroup_mutex
1815 * held.
1816 */
cpuset_post_clone(struct cgroup * cgroup)1817 static void cpuset_post_clone(struct cgroup *cgroup)
1818 {
1819 struct cgroup *parent, *child;
1820 struct cpuset *cs, *parent_cs;
1821
1822 parent = cgroup->parent;
1823 list_for_each_entry(child, &parent->children, sibling) {
1824 cs = cgroup_cs(child);
1825 if (is_mem_exclusive(cs) || is_cpu_exclusive(cs))
1826 return;
1827 }
1828 cs = cgroup_cs(cgroup);
1829 parent_cs = cgroup_cs(parent);
1830
1831 mutex_lock(&callback_mutex);
1832 cs->mems_allowed = parent_cs->mems_allowed;
1833 cpumask_copy(cs->cpus_allowed, parent_cs->cpus_allowed);
1834 mutex_unlock(&callback_mutex);
1835 return;
1836 }
1837
1838 /*
1839 * cpuset_create - create a cpuset
1840 * cont: control group that the new cpuset will be part of
1841 */
1842
cpuset_create(struct cgroup * cont)1843 static struct cgroup_subsys_state *cpuset_create(struct cgroup *cont)
1844 {
1845 struct cpuset *cs;
1846 struct cpuset *parent;
1847
1848 if (!cont->parent) {
1849 return &top_cpuset.css;
1850 }
1851 parent = cgroup_cs(cont->parent);
1852 cs = kmalloc(sizeof(*cs), GFP_KERNEL);
1853 if (!cs)
1854 return ERR_PTR(-ENOMEM);
1855 if (!alloc_cpumask_var(&cs->cpus_allowed, GFP_KERNEL)) {
1856 kfree(cs);
1857 return ERR_PTR(-ENOMEM);
1858 }
1859
1860 cs->flags = 0;
1861 if (is_spread_page(parent))
1862 set_bit(CS_SPREAD_PAGE, &cs->flags);
1863 if (is_spread_slab(parent))
1864 set_bit(CS_SPREAD_SLAB, &cs->flags);
1865 set_bit(CS_SCHED_LOAD_BALANCE, &cs->flags);
1866 cpumask_clear(cs->cpus_allowed);
1867 nodes_clear(cs->mems_allowed);
1868 fmeter_init(&cs->fmeter);
1869 cs->relax_domain_level = -1;
1870
1871 cs->parent = parent;
1872 number_of_cpusets++;
1873 return &cs->css ;
1874 }
1875
1876 /*
1877 * If the cpuset being removed has its flag 'sched_load_balance'
1878 * enabled, then simulate turning sched_load_balance off, which
1879 * will call async_rebuild_sched_domains().
1880 */
1881
cpuset_destroy(struct cgroup * cont)1882 static void cpuset_destroy(struct cgroup *cont)
1883 {
1884 struct cpuset *cs = cgroup_cs(cont);
1885
1886 if (is_sched_load_balance(cs))
1887 update_flag(CS_SCHED_LOAD_BALANCE, cs, 0);
1888
1889 number_of_cpusets--;
1890 free_cpumask_var(cs->cpus_allowed);
1891 kfree(cs);
1892 }
1893
1894 struct cgroup_subsys cpuset_subsys = {
1895 .name = "cpuset",
1896 .create = cpuset_create,
1897 .destroy = cpuset_destroy,
1898 .can_attach = cpuset_can_attach,
1899 .attach = cpuset_attach,
1900 .populate = cpuset_populate,
1901 .post_clone = cpuset_post_clone,
1902 .subsys_id = cpuset_subsys_id,
1903 .early_init = 1,
1904 };
1905
1906 /**
1907 * cpuset_init - initialize cpusets at system boot
1908 *
1909 * Description: Initialize top_cpuset and the cpuset internal file system,
1910 **/
1911
cpuset_init(void)1912 int __init cpuset_init(void)
1913 {
1914 int err = 0;
1915
1916 if (!alloc_cpumask_var(&top_cpuset.cpus_allowed, GFP_KERNEL))
1917 BUG();
1918
1919 cpumask_setall(top_cpuset.cpus_allowed);
1920 nodes_setall(top_cpuset.mems_allowed);
1921
1922 fmeter_init(&top_cpuset.fmeter);
1923 set_bit(CS_SCHED_LOAD_BALANCE, &top_cpuset.flags);
1924 top_cpuset.relax_domain_level = -1;
1925
1926 err = register_filesystem(&cpuset_fs_type);
1927 if (err < 0)
1928 return err;
1929
1930 if (!alloc_cpumask_var(&cpus_attach, GFP_KERNEL))
1931 BUG();
1932
1933 number_of_cpusets = 1;
1934 return 0;
1935 }
1936
1937 /**
1938 * cpuset_do_move_task - move a given task to another cpuset
1939 * @tsk: pointer to task_struct the task to move
1940 * @scan: struct cgroup_scanner contained in its struct cpuset_hotplug_scanner
1941 *
1942 * Called by cgroup_scan_tasks() for each task in a cgroup.
1943 * Return nonzero to stop the walk through the tasks.
1944 */
cpuset_do_move_task(struct task_struct * tsk,struct cgroup_scanner * scan)1945 static void cpuset_do_move_task(struct task_struct *tsk,
1946 struct cgroup_scanner *scan)
1947 {
1948 struct cgroup *new_cgroup = scan->data;
1949
1950 cgroup_attach_task(new_cgroup, tsk);
1951 }
1952
1953 /**
1954 * move_member_tasks_to_cpuset - move tasks from one cpuset to another
1955 * @from: cpuset in which the tasks currently reside
1956 * @to: cpuset to which the tasks will be moved
1957 *
1958 * Called with cgroup_mutex held
1959 * callback_mutex must not be held, as cpuset_attach() will take it.
1960 *
1961 * The cgroup_scan_tasks() function will scan all the tasks in a cgroup,
1962 * calling callback functions for each.
1963 */
move_member_tasks_to_cpuset(struct cpuset * from,struct cpuset * to)1964 static void move_member_tasks_to_cpuset(struct cpuset *from, struct cpuset *to)
1965 {
1966 struct cgroup_scanner scan;
1967
1968 scan.cg = from->css.cgroup;
1969 scan.test_task = NULL; /* select all tasks in cgroup */
1970 scan.process_task = cpuset_do_move_task;
1971 scan.heap = NULL;
1972 scan.data = to->css.cgroup;
1973
1974 if (cgroup_scan_tasks(&scan))
1975 printk(KERN_ERR "move_member_tasks_to_cpuset: "
1976 "cgroup_scan_tasks failed\n");
1977 }
1978
1979 /*
1980 * If CPU and/or memory hotplug handlers, below, unplug any CPUs
1981 * or memory nodes, we need to walk over the cpuset hierarchy,
1982 * removing that CPU or node from all cpusets. If this removes the
1983 * last CPU or node from a cpuset, then move the tasks in the empty
1984 * cpuset to its next-highest non-empty parent.
1985 *
1986 * Called with cgroup_mutex held
1987 * callback_mutex must not be held, as cpuset_attach() will take it.
1988 */
remove_tasks_in_empty_cpuset(struct cpuset * cs)1989 static void remove_tasks_in_empty_cpuset(struct cpuset *cs)
1990 {
1991 struct cpuset *parent;
1992
1993 /*
1994 * The cgroup's css_sets list is in use if there are tasks
1995 * in the cpuset; the list is empty if there are none;
1996 * the cs->css.refcnt seems always 0.
1997 */
1998 if (list_empty(&cs->css.cgroup->css_sets))
1999 return;
2000
2001 /*
2002 * Find its next-highest non-empty parent, (top cpuset
2003 * has online cpus, so can't be empty).
2004 */
2005 parent = cs->parent;
2006 while (cpumask_empty(parent->cpus_allowed) ||
2007 nodes_empty(parent->mems_allowed))
2008 parent = parent->parent;
2009
2010 move_member_tasks_to_cpuset(cs, parent);
2011 }
2012
2013 /*
2014 * Walk the specified cpuset subtree and look for empty cpusets.
2015 * The tasks of such cpuset must be moved to a parent cpuset.
2016 *
2017 * Called with cgroup_mutex held. We take callback_mutex to modify
2018 * cpus_allowed and mems_allowed.
2019 *
2020 * This walk processes the tree from top to bottom, completing one layer
2021 * before dropping down to the next. It always processes a node before
2022 * any of its children.
2023 *
2024 * For now, since we lack memory hot unplug, we'll never see a cpuset
2025 * that has tasks along with an empty 'mems'. But if we did see such
2026 * a cpuset, we'd handle it just like we do if its 'cpus' was empty.
2027 */
scan_for_empty_cpusets(struct cpuset * root)2028 static void scan_for_empty_cpusets(struct cpuset *root)
2029 {
2030 LIST_HEAD(queue);
2031 struct cpuset *cp; /* scans cpusets being updated */
2032 struct cpuset *child; /* scans child cpusets of cp */
2033 struct cgroup *cont;
2034 static nodemask_t oldmems; /* protected by cgroup_mutex */
2035
2036 list_add_tail((struct list_head *)&root->stack_list, &queue);
2037
2038 while (!list_empty(&queue)) {
2039 cp = list_first_entry(&queue, struct cpuset, stack_list);
2040 list_del(queue.next);
2041 list_for_each_entry(cont, &cp->css.cgroup->children, sibling) {
2042 child = cgroup_cs(cont);
2043 list_add_tail(&child->stack_list, &queue);
2044 }
2045
2046 /* Continue past cpusets with all cpus, mems online */
2047 if (cpumask_subset(cp->cpus_allowed, cpu_active_mask) &&
2048 nodes_subset(cp->mems_allowed, node_states[N_HIGH_MEMORY]))
2049 continue;
2050
2051 oldmems = cp->mems_allowed;
2052
2053 /* Remove offline cpus and mems from this cpuset. */
2054 mutex_lock(&callback_mutex);
2055 cpumask_and(cp->cpus_allowed, cp->cpus_allowed,
2056 cpu_active_mask);
2057 nodes_and(cp->mems_allowed, cp->mems_allowed,
2058 node_states[N_HIGH_MEMORY]);
2059 mutex_unlock(&callback_mutex);
2060
2061 /* Move tasks from the empty cpuset to a parent */
2062 if (cpumask_empty(cp->cpus_allowed) ||
2063 nodes_empty(cp->mems_allowed))
2064 remove_tasks_in_empty_cpuset(cp);
2065 else {
2066 update_tasks_cpumask(cp, NULL);
2067 update_tasks_nodemask(cp, &oldmems, NULL);
2068 }
2069 }
2070 }
2071
2072 /*
2073 * The top_cpuset tracks what CPUs and Memory Nodes are online,
2074 * period. This is necessary in order to make cpusets transparent
2075 * (of no affect) on systems that are actively using CPU hotplug
2076 * but making no active use of cpusets.
2077 *
2078 * The only exception to this is suspend/resume, where we don't
2079 * modify cpusets at all.
2080 *
2081 * This routine ensures that top_cpuset.cpus_allowed tracks
2082 * cpu_active_mask on each CPU hotplug (cpuhp) event.
2083 *
2084 * Called within get_online_cpus(). Needs to call cgroup_lock()
2085 * before calling generate_sched_domains().
2086 */
cpuset_update_active_cpus(void)2087 void cpuset_update_active_cpus(void)
2088 {
2089 struct sched_domain_attr *attr;
2090 cpumask_var_t *doms;
2091 int ndoms;
2092
2093 cgroup_lock();
2094 mutex_lock(&callback_mutex);
2095 cpumask_copy(top_cpuset.cpus_allowed, cpu_active_mask);
2096 mutex_unlock(&callback_mutex);
2097 scan_for_empty_cpusets(&top_cpuset);
2098 ndoms = generate_sched_domains(&doms, &attr);
2099 cgroup_unlock();
2100
2101 /* Have scheduler rebuild the domains */
2102 partition_sched_domains(ndoms, doms, attr);
2103 }
2104
2105 #ifdef CONFIG_MEMORY_HOTPLUG
2106 /*
2107 * Keep top_cpuset.mems_allowed tracking node_states[N_HIGH_MEMORY].
2108 * Call this routine anytime after node_states[N_HIGH_MEMORY] changes.
2109 * See also the previous routine cpuset_track_online_cpus().
2110 */
cpuset_track_online_nodes(struct notifier_block * self,unsigned long action,void * arg)2111 static int cpuset_track_online_nodes(struct notifier_block *self,
2112 unsigned long action, void *arg)
2113 {
2114 static nodemask_t oldmems; /* protected by cgroup_mutex */
2115
2116 cgroup_lock();
2117 switch (action) {
2118 case MEM_ONLINE:
2119 oldmems = top_cpuset.mems_allowed;
2120 mutex_lock(&callback_mutex);
2121 top_cpuset.mems_allowed = node_states[N_HIGH_MEMORY];
2122 mutex_unlock(&callback_mutex);
2123 update_tasks_nodemask(&top_cpuset, &oldmems, NULL);
2124 break;
2125 case MEM_OFFLINE:
2126 /*
2127 * needn't update top_cpuset.mems_allowed explicitly because
2128 * scan_for_empty_cpusets() will update it.
2129 */
2130 scan_for_empty_cpusets(&top_cpuset);
2131 break;
2132 default:
2133 break;
2134 }
2135 cgroup_unlock();
2136
2137 return NOTIFY_OK;
2138 }
2139 #endif
2140
2141 /**
2142 * cpuset_init_smp - initialize cpus_allowed
2143 *
2144 * Description: Finish top cpuset after cpu, node maps are initialized
2145 **/
2146
cpuset_init_smp(void)2147 void __init cpuset_init_smp(void)
2148 {
2149 cpumask_copy(top_cpuset.cpus_allowed, cpu_active_mask);
2150 top_cpuset.mems_allowed = node_states[N_HIGH_MEMORY];
2151
2152 hotplug_memory_notifier(cpuset_track_online_nodes, 10);
2153
2154 cpuset_wq = create_singlethread_workqueue("cpuset");
2155 BUG_ON(!cpuset_wq);
2156 }
2157
2158 /**
2159 * cpuset_cpus_allowed - return cpus_allowed mask from a tasks cpuset.
2160 * @tsk: pointer to task_struct from which to obtain cpuset->cpus_allowed.
2161 * @pmask: pointer to struct cpumask variable to receive cpus_allowed set.
2162 *
2163 * Description: Returns the cpumask_var_t cpus_allowed of the cpuset
2164 * attached to the specified @tsk. Guaranteed to return some non-empty
2165 * subset of cpu_online_mask, even if this means going outside the
2166 * tasks cpuset.
2167 **/
2168
cpuset_cpus_allowed(struct task_struct * tsk,struct cpumask * pmask)2169 void cpuset_cpus_allowed(struct task_struct *tsk, struct cpumask *pmask)
2170 {
2171 mutex_lock(&callback_mutex);
2172 task_lock(tsk);
2173 guarantee_online_cpus(task_cs(tsk), pmask);
2174 task_unlock(tsk);
2175 mutex_unlock(&callback_mutex);
2176 }
2177
cpuset_cpus_allowed_fallback(struct task_struct * tsk)2178 void cpuset_cpus_allowed_fallback(struct task_struct *tsk)
2179 {
2180 const struct cpuset *cs;
2181
2182 rcu_read_lock();
2183 cs = task_cs(tsk);
2184 if (cs)
2185 do_set_cpus_allowed(tsk, cs->cpus_allowed);
2186 rcu_read_unlock();
2187
2188 /*
2189 * We own tsk->cpus_allowed, nobody can change it under us.
2190 *
2191 * But we used cs && cs->cpus_allowed lockless and thus can
2192 * race with cgroup_attach_task() or update_cpumask() and get
2193 * the wrong tsk->cpus_allowed. However, both cases imply the
2194 * subsequent cpuset_change_cpumask()->set_cpus_allowed_ptr()
2195 * which takes task_rq_lock().
2196 *
2197 * If we are called after it dropped the lock we must see all
2198 * changes in tsk_cs()->cpus_allowed. Otherwise we can temporary
2199 * set any mask even if it is not right from task_cs() pov,
2200 * the pending set_cpus_allowed_ptr() will fix things.
2201 *
2202 * select_fallback_rq() will fix things ups and set cpu_possible_mask
2203 * if required.
2204 */
2205 }
2206
cpuset_init_current_mems_allowed(void)2207 void cpuset_init_current_mems_allowed(void)
2208 {
2209 nodes_setall(current->mems_allowed);
2210 }
2211
2212 /**
2213 * cpuset_mems_allowed - return mems_allowed mask from a tasks cpuset.
2214 * @tsk: pointer to task_struct from which to obtain cpuset->mems_allowed.
2215 *
2216 * Description: Returns the nodemask_t mems_allowed of the cpuset
2217 * attached to the specified @tsk. Guaranteed to return some non-empty
2218 * subset of node_states[N_HIGH_MEMORY], even if this means going outside the
2219 * tasks cpuset.
2220 **/
2221
cpuset_mems_allowed(struct task_struct * tsk)2222 nodemask_t cpuset_mems_allowed(struct task_struct *tsk)
2223 {
2224 nodemask_t mask;
2225
2226 mutex_lock(&callback_mutex);
2227 task_lock(tsk);
2228 guarantee_online_mems(task_cs(tsk), &mask);
2229 task_unlock(tsk);
2230 mutex_unlock(&callback_mutex);
2231
2232 return mask;
2233 }
2234
2235 /**
2236 * cpuset_nodemask_valid_mems_allowed - check nodemask vs. curremt mems_allowed
2237 * @nodemask: the nodemask to be checked
2238 *
2239 * Are any of the nodes in the nodemask allowed in current->mems_allowed?
2240 */
cpuset_nodemask_valid_mems_allowed(nodemask_t * nodemask)2241 int cpuset_nodemask_valid_mems_allowed(nodemask_t *nodemask)
2242 {
2243 return nodes_intersects(*nodemask, current->mems_allowed);
2244 }
2245
2246 /*
2247 * nearest_hardwall_ancestor() - Returns the nearest mem_exclusive or
2248 * mem_hardwall ancestor to the specified cpuset. Call holding
2249 * callback_mutex. If no ancestor is mem_exclusive or mem_hardwall
2250 * (an unusual configuration), then returns the root cpuset.
2251 */
nearest_hardwall_ancestor(const struct cpuset * cs)2252 static const struct cpuset *nearest_hardwall_ancestor(const struct cpuset *cs)
2253 {
2254 while (!(is_mem_exclusive(cs) || is_mem_hardwall(cs)) && cs->parent)
2255 cs = cs->parent;
2256 return cs;
2257 }
2258
2259 /**
2260 * cpuset_node_allowed_softwall - Can we allocate on a memory node?
2261 * @node: is this an allowed node?
2262 * @gfp_mask: memory allocation flags
2263 *
2264 * If we're in interrupt, yes, we can always allocate. If __GFP_THISNODE is
2265 * set, yes, we can always allocate. If node is in our task's mems_allowed,
2266 * yes. If it's not a __GFP_HARDWALL request and this node is in the nearest
2267 * hardwalled cpuset ancestor to this task's cpuset, yes. If the task has been
2268 * OOM killed and has access to memory reserves as specified by the TIF_MEMDIE
2269 * flag, yes.
2270 * Otherwise, no.
2271 *
2272 * If __GFP_HARDWALL is set, cpuset_node_allowed_softwall() reduces to
2273 * cpuset_node_allowed_hardwall(). Otherwise, cpuset_node_allowed_softwall()
2274 * might sleep, and might allow a node from an enclosing cpuset.
2275 *
2276 * cpuset_node_allowed_hardwall() only handles the simpler case of hardwall
2277 * cpusets, and never sleeps.
2278 *
2279 * The __GFP_THISNODE placement logic is really handled elsewhere,
2280 * by forcibly using a zonelist starting at a specified node, and by
2281 * (in get_page_from_freelist()) refusing to consider the zones for
2282 * any node on the zonelist except the first. By the time any such
2283 * calls get to this routine, we should just shut up and say 'yes'.
2284 *
2285 * GFP_USER allocations are marked with the __GFP_HARDWALL bit,
2286 * and do not allow allocations outside the current tasks cpuset
2287 * unless the task has been OOM killed as is marked TIF_MEMDIE.
2288 * GFP_KERNEL allocations are not so marked, so can escape to the
2289 * nearest enclosing hardwalled ancestor cpuset.
2290 *
2291 * Scanning up parent cpusets requires callback_mutex. The
2292 * __alloc_pages() routine only calls here with __GFP_HARDWALL bit
2293 * _not_ set if it's a GFP_KERNEL allocation, and all nodes in the
2294 * current tasks mems_allowed came up empty on the first pass over
2295 * the zonelist. So only GFP_KERNEL allocations, if all nodes in the
2296 * cpuset are short of memory, might require taking the callback_mutex
2297 * mutex.
2298 *
2299 * The first call here from mm/page_alloc:get_page_from_freelist()
2300 * has __GFP_HARDWALL set in gfp_mask, enforcing hardwall cpusets,
2301 * so no allocation on a node outside the cpuset is allowed (unless
2302 * in interrupt, of course).
2303 *
2304 * The second pass through get_page_from_freelist() doesn't even call
2305 * here for GFP_ATOMIC calls. For those calls, the __alloc_pages()
2306 * variable 'wait' is not set, and the bit ALLOC_CPUSET is not set
2307 * in alloc_flags. That logic and the checks below have the combined
2308 * affect that:
2309 * in_interrupt - any node ok (current task context irrelevant)
2310 * GFP_ATOMIC - any node ok
2311 * TIF_MEMDIE - any node ok
2312 * GFP_KERNEL - any node in enclosing hardwalled cpuset ok
2313 * GFP_USER - only nodes in current tasks mems allowed ok.
2314 *
2315 * Rule:
2316 * Don't call cpuset_node_allowed_softwall if you can't sleep, unless you
2317 * pass in the __GFP_HARDWALL flag set in gfp_flag, which disables
2318 * the code that might scan up ancestor cpusets and sleep.
2319 */
__cpuset_node_allowed_softwall(int node,gfp_t gfp_mask)2320 int __cpuset_node_allowed_softwall(int node, gfp_t gfp_mask)
2321 {
2322 const struct cpuset *cs; /* current cpuset ancestors */
2323 int allowed; /* is allocation in zone z allowed? */
2324
2325 if (in_interrupt() || (gfp_mask & __GFP_THISNODE))
2326 return 1;
2327 might_sleep_if(!(gfp_mask & __GFP_HARDWALL));
2328 if (node_isset(node, current->mems_allowed))
2329 return 1;
2330 /*
2331 * Allow tasks that have access to memory reserves because they have
2332 * been OOM killed to get memory anywhere.
2333 */
2334 if (unlikely(test_thread_flag(TIF_MEMDIE)))
2335 return 1;
2336 if (gfp_mask & __GFP_HARDWALL) /* If hardwall request, stop here */
2337 return 0;
2338
2339 if (current->flags & PF_EXITING) /* Let dying task have memory */
2340 return 1;
2341
2342 /* Not hardwall and node outside mems_allowed: scan up cpusets */
2343 mutex_lock(&callback_mutex);
2344
2345 task_lock(current);
2346 cs = nearest_hardwall_ancestor(task_cs(current));
2347 allowed = node_isset(node, cs->mems_allowed);
2348 task_unlock(current);
2349
2350 mutex_unlock(&callback_mutex);
2351 return allowed;
2352 }
2353
2354 /*
2355 * cpuset_node_allowed_hardwall - Can we allocate on a memory node?
2356 * @node: is this an allowed node?
2357 * @gfp_mask: memory allocation flags
2358 *
2359 * If we're in interrupt, yes, we can always allocate. If __GFP_THISNODE is
2360 * set, yes, we can always allocate. If node is in our task's mems_allowed,
2361 * yes. If the task has been OOM killed and has access to memory reserves as
2362 * specified by the TIF_MEMDIE flag, yes.
2363 * Otherwise, no.
2364 *
2365 * The __GFP_THISNODE placement logic is really handled elsewhere,
2366 * by forcibly using a zonelist starting at a specified node, and by
2367 * (in get_page_from_freelist()) refusing to consider the zones for
2368 * any node on the zonelist except the first. By the time any such
2369 * calls get to this routine, we should just shut up and say 'yes'.
2370 *
2371 * Unlike the cpuset_node_allowed_softwall() variant, above,
2372 * this variant requires that the node be in the current task's
2373 * mems_allowed or that we're in interrupt. It does not scan up the
2374 * cpuset hierarchy for the nearest enclosing mem_exclusive cpuset.
2375 * It never sleeps.
2376 */
__cpuset_node_allowed_hardwall(int node,gfp_t gfp_mask)2377 int __cpuset_node_allowed_hardwall(int node, gfp_t gfp_mask)
2378 {
2379 if (in_interrupt() || (gfp_mask & __GFP_THISNODE))
2380 return 1;
2381 if (node_isset(node, current->mems_allowed))
2382 return 1;
2383 /*
2384 * Allow tasks that have access to memory reserves because they have
2385 * been OOM killed to get memory anywhere.
2386 */
2387 if (unlikely(test_thread_flag(TIF_MEMDIE)))
2388 return 1;
2389 return 0;
2390 }
2391
2392 /**
2393 * cpuset_unlock - release lock on cpuset changes
2394 *
2395 * Undo the lock taken in a previous cpuset_lock() call.
2396 */
2397
cpuset_unlock(void)2398 void cpuset_unlock(void)
2399 {
2400 mutex_unlock(&callback_mutex);
2401 }
2402
2403 /**
2404 * cpuset_mem_spread_node() - On which node to begin search for a file page
2405 * cpuset_slab_spread_node() - On which node to begin search for a slab page
2406 *
2407 * If a task is marked PF_SPREAD_PAGE or PF_SPREAD_SLAB (as for
2408 * tasks in a cpuset with is_spread_page or is_spread_slab set),
2409 * and if the memory allocation used cpuset_mem_spread_node()
2410 * to determine on which node to start looking, as it will for
2411 * certain page cache or slab cache pages such as used for file
2412 * system buffers and inode caches, then instead of starting on the
2413 * local node to look for a free page, rather spread the starting
2414 * node around the tasks mems_allowed nodes.
2415 *
2416 * We don't have to worry about the returned node being offline
2417 * because "it can't happen", and even if it did, it would be ok.
2418 *
2419 * The routines calling guarantee_online_mems() are careful to
2420 * only set nodes in task->mems_allowed that are online. So it
2421 * should not be possible for the following code to return an
2422 * offline node. But if it did, that would be ok, as this routine
2423 * is not returning the node where the allocation must be, only
2424 * the node where the search should start. The zonelist passed to
2425 * __alloc_pages() will include all nodes. If the slab allocator
2426 * is passed an offline node, it will fall back to the local node.
2427 * See kmem_cache_alloc_node().
2428 */
2429
cpuset_spread_node(int * rotor)2430 static int cpuset_spread_node(int *rotor)
2431 {
2432 int node;
2433
2434 node = next_node(*rotor, current->mems_allowed);
2435 if (node == MAX_NUMNODES)
2436 node = first_node(current->mems_allowed);
2437 *rotor = node;
2438 return node;
2439 }
2440
cpuset_mem_spread_node(void)2441 int cpuset_mem_spread_node(void)
2442 {
2443 if (current->cpuset_mem_spread_rotor == NUMA_NO_NODE)
2444 current->cpuset_mem_spread_rotor =
2445 node_random(¤t->mems_allowed);
2446
2447 return cpuset_spread_node(¤t->cpuset_mem_spread_rotor);
2448 }
2449
cpuset_slab_spread_node(void)2450 int cpuset_slab_spread_node(void)
2451 {
2452 if (current->cpuset_slab_spread_rotor == NUMA_NO_NODE)
2453 current->cpuset_slab_spread_rotor =
2454 node_random(¤t->mems_allowed);
2455
2456 return cpuset_spread_node(¤t->cpuset_slab_spread_rotor);
2457 }
2458
2459 EXPORT_SYMBOL_GPL(cpuset_mem_spread_node);
2460
2461 /**
2462 * cpuset_mems_allowed_intersects - Does @tsk1's mems_allowed intersect @tsk2's?
2463 * @tsk1: pointer to task_struct of some task.
2464 * @tsk2: pointer to task_struct of some other task.
2465 *
2466 * Description: Return true if @tsk1's mems_allowed intersects the
2467 * mems_allowed of @tsk2. Used by the OOM killer to determine if
2468 * one of the task's memory usage might impact the memory available
2469 * to the other.
2470 **/
2471
cpuset_mems_allowed_intersects(const struct task_struct * tsk1,const struct task_struct * tsk2)2472 int cpuset_mems_allowed_intersects(const struct task_struct *tsk1,
2473 const struct task_struct *tsk2)
2474 {
2475 return nodes_intersects(tsk1->mems_allowed, tsk2->mems_allowed);
2476 }
2477
2478 /**
2479 * cpuset_print_task_mems_allowed - prints task's cpuset and mems_allowed
2480 * @task: pointer to task_struct of some task.
2481 *
2482 * Description: Prints @task's name, cpuset name, and cached copy of its
2483 * mems_allowed to the kernel log. Must hold task_lock(task) to allow
2484 * dereferencing task_cs(task).
2485 */
cpuset_print_task_mems_allowed(struct task_struct * tsk)2486 void cpuset_print_task_mems_allowed(struct task_struct *tsk)
2487 {
2488 struct dentry *dentry;
2489
2490 dentry = task_cs(tsk)->css.cgroup->dentry;
2491 spin_lock(&cpuset_buffer_lock);
2492
2493 if (!dentry) {
2494 strcpy(cpuset_name, "/");
2495 } else {
2496 spin_lock(&dentry->d_lock);
2497 strlcpy(cpuset_name, (const char *)dentry->d_name.name,
2498 CPUSET_NAME_LEN);
2499 spin_unlock(&dentry->d_lock);
2500 }
2501
2502 nodelist_scnprintf(cpuset_nodelist, CPUSET_NODELIST_LEN,
2503 tsk->mems_allowed);
2504 printk(KERN_INFO "%s cpuset=%s mems_allowed=%s\n",
2505 tsk->comm, cpuset_name, cpuset_nodelist);
2506 spin_unlock(&cpuset_buffer_lock);
2507 }
2508
2509 /*
2510 * Collection of memory_pressure is suppressed unless
2511 * this flag is enabled by writing "1" to the special
2512 * cpuset file 'memory_pressure_enabled' in the root cpuset.
2513 */
2514
2515 int cpuset_memory_pressure_enabled __read_mostly;
2516
2517 /**
2518 * cpuset_memory_pressure_bump - keep stats of per-cpuset reclaims.
2519 *
2520 * Keep a running average of the rate of synchronous (direct)
2521 * page reclaim efforts initiated by tasks in each cpuset.
2522 *
2523 * This represents the rate at which some task in the cpuset
2524 * ran low on memory on all nodes it was allowed to use, and
2525 * had to enter the kernels page reclaim code in an effort to
2526 * create more free memory by tossing clean pages or swapping
2527 * or writing dirty pages.
2528 *
2529 * Display to user space in the per-cpuset read-only file
2530 * "memory_pressure". Value displayed is an integer
2531 * representing the recent rate of entry into the synchronous
2532 * (direct) page reclaim by any task attached to the cpuset.
2533 **/
2534
__cpuset_memory_pressure_bump(void)2535 void __cpuset_memory_pressure_bump(void)
2536 {
2537 task_lock(current);
2538 fmeter_markevent(&task_cs(current)->fmeter);
2539 task_unlock(current);
2540 }
2541
2542 #ifdef CONFIG_PROC_PID_CPUSET
2543 /*
2544 * proc_cpuset_show()
2545 * - Print tasks cpuset path into seq_file.
2546 * - Used for /proc/<pid>/cpuset.
2547 * - No need to task_lock(tsk) on this tsk->cpuset reference, as it
2548 * doesn't really matter if tsk->cpuset changes after we read it,
2549 * and we take cgroup_mutex, keeping cpuset_attach() from changing it
2550 * anyway.
2551 */
proc_cpuset_show(struct seq_file * m,void * unused_v)2552 static int proc_cpuset_show(struct seq_file *m, void *unused_v)
2553 {
2554 struct pid *pid;
2555 struct task_struct *tsk;
2556 char *buf;
2557 struct cgroup_subsys_state *css;
2558 int retval;
2559
2560 retval = -ENOMEM;
2561 buf = kmalloc(PAGE_SIZE, GFP_KERNEL);
2562 if (!buf)
2563 goto out;
2564
2565 retval = -ESRCH;
2566 pid = m->private;
2567 tsk = get_pid_task(pid, PIDTYPE_PID);
2568 if (!tsk)
2569 goto out_free;
2570
2571 retval = -EINVAL;
2572 cgroup_lock();
2573 css = task_subsys_state(tsk, cpuset_subsys_id);
2574 retval = cgroup_path(css->cgroup, buf, PAGE_SIZE);
2575 if (retval < 0)
2576 goto out_unlock;
2577 seq_puts(m, buf);
2578 seq_putc(m, '\n');
2579 out_unlock:
2580 cgroup_unlock();
2581 put_task_struct(tsk);
2582 out_free:
2583 kfree(buf);
2584 out:
2585 return retval;
2586 }
2587
cpuset_open(struct inode * inode,struct file * file)2588 static int cpuset_open(struct inode *inode, struct file *file)
2589 {
2590 struct pid *pid = PROC_I(inode)->pid;
2591 return single_open(file, proc_cpuset_show, pid);
2592 }
2593
2594 const struct file_operations proc_cpuset_operations = {
2595 .open = cpuset_open,
2596 .read = seq_read,
2597 .llseek = seq_lseek,
2598 .release = single_release,
2599 };
2600 #endif /* CONFIG_PROC_PID_CPUSET */
2601
2602 /* Display task mems_allowed in /proc/<pid>/status file. */
cpuset_task_status_allowed(struct seq_file * m,struct task_struct * task)2603 void cpuset_task_status_allowed(struct seq_file *m, struct task_struct *task)
2604 {
2605 seq_printf(m, "Mems_allowed:\t");
2606 seq_nodemask(m, &task->mems_allowed);
2607 seq_printf(m, "\n");
2608 seq_printf(m, "Mems_allowed_list:\t");
2609 seq_nodemask_list(m, &task->mems_allowed);
2610 seq_printf(m, "\n");
2611 }
2612