Loading...
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#include <linux/wait.h>
63
64/*
65 * Tracks how many cpusets are currently defined in system.
66 * When there is only one cpuset (the root cpuset) we can
67 * short circuit some hooks.
68 */
69int number_of_cpusets __read_mostly;
70
71/* See "Frequency meter" comments, below. */
72
73struct fmeter {
74 int cnt; /* unprocessed events count */
75 int val; /* most recent output value */
76 time_t time; /* clock (secs) when val computed */
77 spinlock_t lock; /* guards read or write of above */
78};
79
80struct cpuset {
81 struct cgroup_subsys_state css;
82
83 unsigned long flags; /* "unsigned long" so bitops work */
84 cpumask_var_t cpus_allowed; /* CPUs allowed to tasks in cpuset */
85 nodemask_t mems_allowed; /* Memory Nodes allowed to tasks */
86
87 /*
88 * This is old Memory Nodes tasks took on.
89 *
90 * - top_cpuset.old_mems_allowed is initialized to mems_allowed.
91 * - A new cpuset's old_mems_allowed is initialized when some
92 * task is moved into it.
93 * - old_mems_allowed is used in cpuset_migrate_mm() when we change
94 * cpuset.mems_allowed and have tasks' nodemask updated, and
95 * then old_mems_allowed is updated to mems_allowed.
96 */
97 nodemask_t old_mems_allowed;
98
99 struct fmeter fmeter; /* memory_pressure filter */
100
101 /*
102 * Tasks are being attached to this cpuset. Used to prevent
103 * zeroing cpus/mems_allowed between ->can_attach() and ->attach().
104 */
105 int attach_in_progress;
106
107 /* partition number for rebuild_sched_domains() */
108 int pn;
109
110 /* for custom sched domain */
111 int relax_domain_level;
112};
113
114static inline struct cpuset *css_cs(struct cgroup_subsys_state *css)
115{
116 return css ? container_of(css, struct cpuset, css) : NULL;
117}
118
119/* Retrieve the cpuset for a task */
120static inline struct cpuset *task_cs(struct task_struct *task)
121{
122 return css_cs(task_css(task, cpuset_cgrp_id));
123}
124
125static inline struct cpuset *parent_cs(struct cpuset *cs)
126{
127 return css_cs(css_parent(&cs->css));
128}
129
130#ifdef CONFIG_NUMA
131static inline bool task_has_mempolicy(struct task_struct *task)
132{
133 return task->mempolicy;
134}
135#else
136static inline bool task_has_mempolicy(struct task_struct *task)
137{
138 return false;
139}
140#endif
141
142
143/* bits in struct cpuset flags field */
144typedef enum {
145 CS_ONLINE,
146 CS_CPU_EXCLUSIVE,
147 CS_MEM_EXCLUSIVE,
148 CS_MEM_HARDWALL,
149 CS_MEMORY_MIGRATE,
150 CS_SCHED_LOAD_BALANCE,
151 CS_SPREAD_PAGE,
152 CS_SPREAD_SLAB,
153} cpuset_flagbits_t;
154
155/* convenient tests for these bits */
156static inline bool is_cpuset_online(const struct cpuset *cs)
157{
158 return test_bit(CS_ONLINE, &cs->flags);
159}
160
161static inline int is_cpu_exclusive(const struct cpuset *cs)
162{
163 return test_bit(CS_CPU_EXCLUSIVE, &cs->flags);
164}
165
166static inline int is_mem_exclusive(const struct cpuset *cs)
167{
168 return test_bit(CS_MEM_EXCLUSIVE, &cs->flags);
169}
170
171static inline int is_mem_hardwall(const struct cpuset *cs)
172{
173 return test_bit(CS_MEM_HARDWALL, &cs->flags);
174}
175
176static inline int is_sched_load_balance(const struct cpuset *cs)
177{
178 return test_bit(CS_SCHED_LOAD_BALANCE, &cs->flags);
179}
180
181static inline int is_memory_migrate(const struct cpuset *cs)
182{
183 return test_bit(CS_MEMORY_MIGRATE, &cs->flags);
184}
185
186static inline int is_spread_page(const struct cpuset *cs)
187{
188 return test_bit(CS_SPREAD_PAGE, &cs->flags);
189}
190
191static inline int is_spread_slab(const struct cpuset *cs)
192{
193 return test_bit(CS_SPREAD_SLAB, &cs->flags);
194}
195
196static struct cpuset top_cpuset = {
197 .flags = ((1 << CS_ONLINE) | (1 << CS_CPU_EXCLUSIVE) |
198 (1 << CS_MEM_EXCLUSIVE)),
199};
200
201/**
202 * cpuset_for_each_child - traverse online children of a cpuset
203 * @child_cs: loop cursor pointing to the current child
204 * @pos_css: used for iteration
205 * @parent_cs: target cpuset to walk children of
206 *
207 * Walk @child_cs through the online children of @parent_cs. Must be used
208 * with RCU read locked.
209 */
210#define cpuset_for_each_child(child_cs, pos_css, parent_cs) \
211 css_for_each_child((pos_css), &(parent_cs)->css) \
212 if (is_cpuset_online(((child_cs) = css_cs((pos_css)))))
213
214/**
215 * cpuset_for_each_descendant_pre - pre-order walk of a cpuset's descendants
216 * @des_cs: loop cursor pointing to the current descendant
217 * @pos_css: used for iteration
218 * @root_cs: target cpuset to walk ancestor of
219 *
220 * Walk @des_cs through the online descendants of @root_cs. Must be used
221 * with RCU read locked. The caller may modify @pos_css by calling
222 * css_rightmost_descendant() to skip subtree. @root_cs is included in the
223 * iteration and the first node to be visited.
224 */
225#define cpuset_for_each_descendant_pre(des_cs, pos_css, root_cs) \
226 css_for_each_descendant_pre((pos_css), &(root_cs)->css) \
227 if (is_cpuset_online(((des_cs) = css_cs((pos_css)))))
228
229/*
230 * There are two global mutexes guarding cpuset structures - cpuset_mutex
231 * and callback_mutex. The latter may nest inside the former. We also
232 * require taking task_lock() when dereferencing a task's cpuset pointer.
233 * See "The task_lock() exception", at the end of this comment.
234 *
235 * A task must hold both mutexes to modify cpusets. If a task holds
236 * cpuset_mutex, then it blocks others wanting that mutex, ensuring that it
237 * is the only task able to also acquire callback_mutex and be able to
238 * modify cpusets. It can perform various checks on the cpuset structure
239 * first, knowing nothing will change. It can also allocate memory while
240 * just holding cpuset_mutex. While it is performing these checks, various
241 * callback routines can briefly acquire callback_mutex to query cpusets.
242 * Once it is ready to make the changes, it takes callback_mutex, blocking
243 * everyone else.
244 *
245 * Calls to the kernel memory allocator can not be made while holding
246 * callback_mutex, as that would risk double tripping on callback_mutex
247 * from one of the callbacks into the cpuset code from within
248 * __alloc_pages().
249 *
250 * If a task is only holding callback_mutex, then it has read-only
251 * access to cpusets.
252 *
253 * Now, the task_struct fields mems_allowed and mempolicy may be changed
254 * by other task, we use alloc_lock in the task_struct fields to protect
255 * them.
256 *
257 * The cpuset_common_file_read() handlers only hold callback_mutex across
258 * small pieces of code, such as when reading out possibly multi-word
259 * cpumasks and nodemasks.
260 *
261 * Accessing a task's cpuset should be done in accordance with the
262 * guidelines for accessing subsystem state in kernel/cgroup.c
263 */
264
265static DEFINE_MUTEX(cpuset_mutex);
266static DEFINE_MUTEX(callback_mutex);
267
268/*
269 * CPU / memory hotplug is handled asynchronously.
270 */
271static void cpuset_hotplug_workfn(struct work_struct *work);
272static DECLARE_WORK(cpuset_hotplug_work, cpuset_hotplug_workfn);
273
274static DECLARE_WAIT_QUEUE_HEAD(cpuset_attach_wq);
275
276/*
277 * This is ugly, but preserves the userspace API for existing cpuset
278 * users. If someone tries to mount the "cpuset" filesystem, we
279 * silently switch it to mount "cgroup" instead
280 */
281static struct dentry *cpuset_mount(struct file_system_type *fs_type,
282 int flags, const char *unused_dev_name, void *data)
283{
284 struct file_system_type *cgroup_fs = get_fs_type("cgroup");
285 struct dentry *ret = ERR_PTR(-ENODEV);
286 if (cgroup_fs) {
287 char mountopts[] =
288 "cpuset,noprefix,"
289 "release_agent=/sbin/cpuset_release_agent";
290 ret = cgroup_fs->mount(cgroup_fs, flags,
291 unused_dev_name, mountopts);
292 put_filesystem(cgroup_fs);
293 }
294 return ret;
295}
296
297static struct file_system_type cpuset_fs_type = {
298 .name = "cpuset",
299 .mount = cpuset_mount,
300};
301
302/*
303 * Return in pmask the portion of a cpusets's cpus_allowed that
304 * are online. If none are online, walk up the cpuset hierarchy
305 * until we find one that does have some online cpus. The top
306 * cpuset always has some cpus online.
307 *
308 * One way or another, we guarantee to return some non-empty subset
309 * of cpu_online_mask.
310 *
311 * Call with callback_mutex held.
312 */
313static void guarantee_online_cpus(struct cpuset *cs, struct cpumask *pmask)
314{
315 while (!cpumask_intersects(cs->cpus_allowed, cpu_online_mask))
316 cs = parent_cs(cs);
317 cpumask_and(pmask, cs->cpus_allowed, cpu_online_mask);
318}
319
320/*
321 * Return in *pmask the portion of a cpusets's mems_allowed that
322 * are online, with memory. If none are online with memory, walk
323 * up the cpuset hierarchy until we find one that does have some
324 * online mems. The top cpuset always has some mems online.
325 *
326 * One way or another, we guarantee to return some non-empty subset
327 * of node_states[N_MEMORY].
328 *
329 * Call with callback_mutex held.
330 */
331static void guarantee_online_mems(struct cpuset *cs, nodemask_t *pmask)
332{
333 while (!nodes_intersects(cs->mems_allowed, node_states[N_MEMORY]))
334 cs = parent_cs(cs);
335 nodes_and(*pmask, cs->mems_allowed, node_states[N_MEMORY]);
336}
337
338/*
339 * update task's spread flag if cpuset's page/slab spread flag is set
340 *
341 * Called with callback_mutex/cpuset_mutex held
342 */
343static void cpuset_update_task_spread_flag(struct cpuset *cs,
344 struct task_struct *tsk)
345{
346 if (is_spread_page(cs))
347 tsk->flags |= PF_SPREAD_PAGE;
348 else
349 tsk->flags &= ~PF_SPREAD_PAGE;
350 if (is_spread_slab(cs))
351 tsk->flags |= PF_SPREAD_SLAB;
352 else
353 tsk->flags &= ~PF_SPREAD_SLAB;
354}
355
356/*
357 * is_cpuset_subset(p, q) - Is cpuset p a subset of cpuset q?
358 *
359 * One cpuset is a subset of another if all its allowed CPUs and
360 * Memory Nodes are a subset of the other, and its exclusive flags
361 * are only set if the other's are set. Call holding cpuset_mutex.
362 */
363
364static int is_cpuset_subset(const struct cpuset *p, const struct cpuset *q)
365{
366 return cpumask_subset(p->cpus_allowed, q->cpus_allowed) &&
367 nodes_subset(p->mems_allowed, q->mems_allowed) &&
368 is_cpu_exclusive(p) <= is_cpu_exclusive(q) &&
369 is_mem_exclusive(p) <= is_mem_exclusive(q);
370}
371
372/**
373 * alloc_trial_cpuset - allocate a trial cpuset
374 * @cs: the cpuset that the trial cpuset duplicates
375 */
376static struct cpuset *alloc_trial_cpuset(struct cpuset *cs)
377{
378 struct cpuset *trial;
379
380 trial = kmemdup(cs, sizeof(*cs), GFP_KERNEL);
381 if (!trial)
382 return NULL;
383
384 if (!alloc_cpumask_var(&trial->cpus_allowed, GFP_KERNEL)) {
385 kfree(trial);
386 return NULL;
387 }
388 cpumask_copy(trial->cpus_allowed, cs->cpus_allowed);
389
390 return trial;
391}
392
393/**
394 * free_trial_cpuset - free the trial cpuset
395 * @trial: the trial cpuset to be freed
396 */
397static void free_trial_cpuset(struct cpuset *trial)
398{
399 free_cpumask_var(trial->cpus_allowed);
400 kfree(trial);
401}
402
403/*
404 * validate_change() - Used to validate that any proposed cpuset change
405 * follows the structural rules for cpusets.
406 *
407 * If we replaced the flag and mask values of the current cpuset
408 * (cur) with those values in the trial cpuset (trial), would
409 * our various subset and exclusive rules still be valid? Presumes
410 * cpuset_mutex held.
411 *
412 * 'cur' is the address of an actual, in-use cpuset. Operations
413 * such as list traversal that depend on the actual address of the
414 * cpuset in the list must use cur below, not trial.
415 *
416 * 'trial' is the address of bulk structure copy of cur, with
417 * perhaps one or more of the fields cpus_allowed, mems_allowed,
418 * or flags changed to new, trial values.
419 *
420 * Return 0 if valid, -errno if not.
421 */
422
423static int validate_change(struct cpuset *cur, struct cpuset *trial)
424{
425 struct cgroup_subsys_state *css;
426 struct cpuset *c, *par;
427 int ret;
428
429 rcu_read_lock();
430
431 /* Each of our child cpusets must be a subset of us */
432 ret = -EBUSY;
433 cpuset_for_each_child(c, css, cur)
434 if (!is_cpuset_subset(c, trial))
435 goto out;
436
437 /* Remaining checks don't apply to root cpuset */
438 ret = 0;
439 if (cur == &top_cpuset)
440 goto out;
441
442 par = parent_cs(cur);
443
444 /* We must be a subset of our parent cpuset */
445 ret = -EACCES;
446 if (!is_cpuset_subset(trial, par))
447 goto out;
448
449 /*
450 * If either I or some sibling (!= me) is exclusive, we can't
451 * overlap
452 */
453 ret = -EINVAL;
454 cpuset_for_each_child(c, css, par) {
455 if ((is_cpu_exclusive(trial) || is_cpu_exclusive(c)) &&
456 c != cur &&
457 cpumask_intersects(trial->cpus_allowed, c->cpus_allowed))
458 goto out;
459 if ((is_mem_exclusive(trial) || is_mem_exclusive(c)) &&
460 c != cur &&
461 nodes_intersects(trial->mems_allowed, c->mems_allowed))
462 goto out;
463 }
464
465 /*
466 * Cpusets with tasks - existing or newly being attached - can't
467 * be changed to have empty cpus_allowed or mems_allowed.
468 */
469 ret = -ENOSPC;
470 if ((cgroup_has_tasks(cur->css.cgroup) || cur->attach_in_progress)) {
471 if (!cpumask_empty(cur->cpus_allowed) &&
472 cpumask_empty(trial->cpus_allowed))
473 goto out;
474 if (!nodes_empty(cur->mems_allowed) &&
475 nodes_empty(trial->mems_allowed))
476 goto out;
477 }
478
479 ret = 0;
480out:
481 rcu_read_unlock();
482 return ret;
483}
484
485#ifdef CONFIG_SMP
486/*
487 * Helper routine for generate_sched_domains().
488 * Do cpusets a, b have overlapping cpus_allowed masks?
489 */
490static int cpusets_overlap(struct cpuset *a, struct cpuset *b)
491{
492 return cpumask_intersects(a->cpus_allowed, b->cpus_allowed);
493}
494
495static void
496update_domain_attr(struct sched_domain_attr *dattr, struct cpuset *c)
497{
498 if (dattr->relax_domain_level < c->relax_domain_level)
499 dattr->relax_domain_level = c->relax_domain_level;
500 return;
501}
502
503static void update_domain_attr_tree(struct sched_domain_attr *dattr,
504 struct cpuset *root_cs)
505{
506 struct cpuset *cp;
507 struct cgroup_subsys_state *pos_css;
508
509 rcu_read_lock();
510 cpuset_for_each_descendant_pre(cp, pos_css, root_cs) {
511 if (cp == root_cs)
512 continue;
513
514 /* skip the whole subtree if @cp doesn't have any CPU */
515 if (cpumask_empty(cp->cpus_allowed)) {
516 pos_css = css_rightmost_descendant(pos_css);
517 continue;
518 }
519
520 if (is_sched_load_balance(cp))
521 update_domain_attr(dattr, cp);
522 }
523 rcu_read_unlock();
524}
525
526/*
527 * generate_sched_domains()
528 *
529 * This function builds a partial partition of the systems CPUs
530 * A 'partial partition' is a set of non-overlapping subsets whose
531 * union is a subset of that set.
532 * The output of this function needs to be passed to kernel/sched/core.c
533 * partition_sched_domains() routine, which will rebuild the scheduler's
534 * load balancing domains (sched domains) as specified by that partial
535 * partition.
536 *
537 * See "What is sched_load_balance" in Documentation/cgroups/cpusets.txt
538 * for a background explanation of this.
539 *
540 * Does not return errors, on the theory that the callers of this
541 * routine would rather not worry about failures to rebuild sched
542 * domains when operating in the severe memory shortage situations
543 * that could cause allocation failures below.
544 *
545 * Must be called with cpuset_mutex held.
546 *
547 * The three key local variables below are:
548 * q - a linked-list queue of cpuset pointers, used to implement a
549 * top-down scan of all cpusets. This scan loads a pointer
550 * to each cpuset marked is_sched_load_balance into the
551 * array 'csa'. For our purposes, rebuilding the schedulers
552 * sched domains, we can ignore !is_sched_load_balance cpusets.
553 * csa - (for CpuSet Array) Array of pointers to all the cpusets
554 * that need to be load balanced, for convenient iterative
555 * access by the subsequent code that finds the best partition,
556 * i.e the set of domains (subsets) of CPUs such that the
557 * cpus_allowed of every cpuset marked is_sched_load_balance
558 * is a subset of one of these domains, while there are as
559 * many such domains as possible, each as small as possible.
560 * doms - Conversion of 'csa' to an array of cpumasks, for passing to
561 * the kernel/sched/core.c routine partition_sched_domains() in a
562 * convenient format, that can be easily compared to the prior
563 * value to determine what partition elements (sched domains)
564 * were changed (added or removed.)
565 *
566 * Finding the best partition (set of domains):
567 * The triple nested loops below over i, j, k scan over the
568 * load balanced cpusets (using the array of cpuset pointers in
569 * csa[]) looking for pairs of cpusets that have overlapping
570 * cpus_allowed, but which don't have the same 'pn' partition
571 * number and gives them in the same partition number. It keeps
572 * looping on the 'restart' label until it can no longer find
573 * any such pairs.
574 *
575 * The union of the cpus_allowed masks from the set of
576 * all cpusets having the same 'pn' value then form the one
577 * element of the partition (one sched domain) to be passed to
578 * partition_sched_domains().
579 */
580static int generate_sched_domains(cpumask_var_t **domains,
581 struct sched_domain_attr **attributes)
582{
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 */
587 cpumask_var_t *doms; /* resulting partition; i.e. sched domains */
588 struct sched_domain_attr *dattr; /* attributes for custom domains */
589 int ndoms = 0; /* number of sched domains in result */
590 int nslot; /* next empty doms[] struct cpumask slot */
591 struct cgroup_subsys_state *pos_css;
592
593 doms = NULL;
594 dattr = NULL;
595 csa = NULL;
596
597 /* Special case for the 99% of systems with one, full, sched domain */
598 if (is_sched_load_balance(&top_cpuset)) {
599 ndoms = 1;
600 doms = alloc_sched_domains(ndoms);
601 if (!doms)
602 goto done;
603
604 dattr = kmalloc(sizeof(struct sched_domain_attr), GFP_KERNEL);
605 if (dattr) {
606 *dattr = SD_ATTR_INIT;
607 update_domain_attr_tree(dattr, &top_cpuset);
608 }
609 cpumask_copy(doms[0], top_cpuset.cpus_allowed);
610
611 goto done;
612 }
613
614 csa = kmalloc(number_of_cpusets * sizeof(cp), GFP_KERNEL);
615 if (!csa)
616 goto done;
617 csn = 0;
618
619 rcu_read_lock();
620 cpuset_for_each_descendant_pre(cp, pos_css, &top_cpuset) {
621 if (cp == &top_cpuset)
622 continue;
623 /*
624 * Continue traversing beyond @cp iff @cp has some CPUs and
625 * isn't load balancing. The former is obvious. The
626 * latter: All child cpusets contain a subset of the
627 * parent's cpus, so just skip them, and then we call
628 * update_domain_attr_tree() to calc relax_domain_level of
629 * the corresponding sched domain.
630 */
631 if (!cpumask_empty(cp->cpus_allowed) &&
632 !is_sched_load_balance(cp))
633 continue;
634
635 if (is_sched_load_balance(cp))
636 csa[csn++] = cp;
637
638 /* skip @cp's subtree */
639 pos_css = css_rightmost_descendant(pos_css);
640 }
641 rcu_read_unlock();
642
643 for (i = 0; i < csn; i++)
644 csa[i]->pn = i;
645 ndoms = csn;
646
647restart:
648 /* Find the best partition (set of sched domains) */
649 for (i = 0; i < csn; i++) {
650 struct cpuset *a = csa[i];
651 int apn = a->pn;
652
653 for (j = 0; j < csn; j++) {
654 struct cpuset *b = csa[j];
655 int bpn = b->pn;
656
657 if (apn != bpn && cpusets_overlap(a, b)) {
658 for (k = 0; k < csn; k++) {
659 struct cpuset *c = csa[k];
660
661 if (c->pn == bpn)
662 c->pn = apn;
663 }
664 ndoms--; /* one less element */
665 goto restart;
666 }
667 }
668 }
669
670 /*
671 * Now we know how many domains to create.
672 * Convert <csn, csa> to <ndoms, doms> and populate cpu masks.
673 */
674 doms = alloc_sched_domains(ndoms);
675 if (!doms)
676 goto done;
677
678 /*
679 * The rest of the code, including the scheduler, can deal with
680 * dattr==NULL case. No need to abort if alloc fails.
681 */
682 dattr = kmalloc(ndoms * sizeof(struct sched_domain_attr), GFP_KERNEL);
683
684 for (nslot = 0, i = 0; i < csn; i++) {
685 struct cpuset *a = csa[i];
686 struct cpumask *dp;
687 int apn = a->pn;
688
689 if (apn < 0) {
690 /* Skip completed partitions */
691 continue;
692 }
693
694 dp = doms[nslot];
695
696 if (nslot == ndoms) {
697 static int warnings = 10;
698 if (warnings) {
699 printk(KERN_WARNING
700 "rebuild_sched_domains confused:"
701 " nslot %d, ndoms %d, csn %d, i %d,"
702 " apn %d\n",
703 nslot, ndoms, csn, i, apn);
704 warnings--;
705 }
706 continue;
707 }
708
709 cpumask_clear(dp);
710 if (dattr)
711 *(dattr + nslot) = SD_ATTR_INIT;
712 for (j = i; j < csn; j++) {
713 struct cpuset *b = csa[j];
714
715 if (apn == b->pn) {
716 cpumask_or(dp, dp, b->cpus_allowed);
717 if (dattr)
718 update_domain_attr_tree(dattr + nslot, b);
719
720 /* Done with this partition */
721 b->pn = -1;
722 }
723 }
724 nslot++;
725 }
726 BUG_ON(nslot != ndoms);
727
728done:
729 kfree(csa);
730
731 /*
732 * Fallback to the default domain if kmalloc() failed.
733 * See comments in partition_sched_domains().
734 */
735 if (doms == NULL)
736 ndoms = 1;
737
738 *domains = doms;
739 *attributes = dattr;
740 return ndoms;
741}
742
743/*
744 * Rebuild scheduler domains.
745 *
746 * If the flag 'sched_load_balance' of any cpuset with non-empty
747 * 'cpus' changes, or if the 'cpus' allowed changes in any cpuset
748 * which has that flag enabled, or if any cpuset with a non-empty
749 * 'cpus' is removed, then call this routine to rebuild the
750 * scheduler's dynamic sched domains.
751 *
752 * Call with cpuset_mutex held. Takes get_online_cpus().
753 */
754static void rebuild_sched_domains_locked(void)
755{
756 struct sched_domain_attr *attr;
757 cpumask_var_t *doms;
758 int ndoms;
759
760 lockdep_assert_held(&cpuset_mutex);
761 get_online_cpus();
762
763 /*
764 * We have raced with CPU hotplug. Don't do anything to avoid
765 * passing doms with offlined cpu to partition_sched_domains().
766 * Anyways, hotplug work item will rebuild sched domains.
767 */
768 if (!cpumask_equal(top_cpuset.cpus_allowed, cpu_active_mask))
769 goto out;
770
771 /* Generate domain masks and attrs */
772 ndoms = generate_sched_domains(&doms, &attr);
773
774 /* Have scheduler rebuild the domains */
775 partition_sched_domains(ndoms, doms, attr);
776out:
777 put_online_cpus();
778}
779#else /* !CONFIG_SMP */
780static void rebuild_sched_domains_locked(void)
781{
782}
783#endif /* CONFIG_SMP */
784
785void rebuild_sched_domains(void)
786{
787 mutex_lock(&cpuset_mutex);
788 rebuild_sched_domains_locked();
789 mutex_unlock(&cpuset_mutex);
790}
791
792/*
793 * effective_cpumask_cpuset - return nearest ancestor with non-empty cpus
794 * @cs: the cpuset in interest
795 *
796 * A cpuset's effective cpumask is the cpumask of the nearest ancestor
797 * with non-empty cpus. We use effective cpumask whenever:
798 * - we update tasks' cpus_allowed. (they take on the ancestor's cpumask
799 * if the cpuset they reside in has no cpus)
800 * - we want to retrieve task_cs(tsk)'s cpus_allowed.
801 *
802 * Called with cpuset_mutex held. cpuset_cpus_allowed_fallback() is an
803 * exception. See comments there.
804 */
805static struct cpuset *effective_cpumask_cpuset(struct cpuset *cs)
806{
807 while (cpumask_empty(cs->cpus_allowed))
808 cs = parent_cs(cs);
809 return cs;
810}
811
812/*
813 * effective_nodemask_cpuset - return nearest ancestor with non-empty mems
814 * @cs: the cpuset in interest
815 *
816 * A cpuset's effective nodemask is the nodemask of the nearest ancestor
817 * with non-empty memss. We use effective nodemask whenever:
818 * - we update tasks' mems_allowed. (they take on the ancestor's nodemask
819 * if the cpuset they reside in has no mems)
820 * - we want to retrieve task_cs(tsk)'s mems_allowed.
821 *
822 * Called with cpuset_mutex held.
823 */
824static struct cpuset *effective_nodemask_cpuset(struct cpuset *cs)
825{
826 while (nodes_empty(cs->mems_allowed))
827 cs = parent_cs(cs);
828 return cs;
829}
830
831/**
832 * update_tasks_cpumask - Update the cpumasks of tasks in the cpuset.
833 * @cs: the cpuset in which each task's cpus_allowed mask needs to be changed
834 *
835 * Iterate through each task of @cs updating its cpus_allowed to the
836 * effective cpuset's. As this function is called with cpuset_mutex held,
837 * cpuset membership stays stable.
838 */
839static void update_tasks_cpumask(struct cpuset *cs)
840{
841 struct cpuset *cpus_cs = effective_cpumask_cpuset(cs);
842 struct css_task_iter it;
843 struct task_struct *task;
844
845 css_task_iter_start(&cs->css, &it);
846 while ((task = css_task_iter_next(&it)))
847 set_cpus_allowed_ptr(task, cpus_cs->cpus_allowed);
848 css_task_iter_end(&it);
849}
850
851/*
852 * update_tasks_cpumask_hier - Update the cpumasks of tasks in the hierarchy.
853 * @root_cs: the root cpuset of the hierarchy
854 * @update_root: update root cpuset or not?
855 *
856 * This will update cpumasks of tasks in @root_cs and all other empty cpusets
857 * which take on cpumask of @root_cs.
858 *
859 * Called with cpuset_mutex held
860 */
861static void update_tasks_cpumask_hier(struct cpuset *root_cs, bool update_root)
862{
863 struct cpuset *cp;
864 struct cgroup_subsys_state *pos_css;
865
866 rcu_read_lock();
867 cpuset_for_each_descendant_pre(cp, pos_css, root_cs) {
868 if (cp == root_cs) {
869 if (!update_root)
870 continue;
871 } else {
872 /* skip the whole subtree if @cp have some CPU */
873 if (!cpumask_empty(cp->cpus_allowed)) {
874 pos_css = css_rightmost_descendant(pos_css);
875 continue;
876 }
877 }
878 if (!css_tryget(&cp->css))
879 continue;
880 rcu_read_unlock();
881
882 update_tasks_cpumask(cp);
883
884 rcu_read_lock();
885 css_put(&cp->css);
886 }
887 rcu_read_unlock();
888}
889
890/**
891 * update_cpumask - update the cpus_allowed mask of a cpuset and all tasks in it
892 * @cs: the cpuset to consider
893 * @buf: buffer of cpu numbers written to this cpuset
894 */
895static int update_cpumask(struct cpuset *cs, struct cpuset *trialcs,
896 const char *buf)
897{
898 int retval;
899 int is_load_balanced;
900
901 /* top_cpuset.cpus_allowed tracks cpu_online_mask; it's read-only */
902 if (cs == &top_cpuset)
903 return -EACCES;
904
905 /*
906 * An empty cpus_allowed is ok only if the cpuset has no tasks.
907 * Since cpulist_parse() fails on an empty mask, we special case
908 * that parsing. The validate_change() call ensures that cpusets
909 * with tasks have cpus.
910 */
911 if (!*buf) {
912 cpumask_clear(trialcs->cpus_allowed);
913 } else {
914 retval = cpulist_parse(buf, trialcs->cpus_allowed);
915 if (retval < 0)
916 return retval;
917
918 if (!cpumask_subset(trialcs->cpus_allowed, cpu_active_mask))
919 return -EINVAL;
920 }
921
922 /* Nothing to do if the cpus didn't change */
923 if (cpumask_equal(cs->cpus_allowed, trialcs->cpus_allowed))
924 return 0;
925
926 retval = validate_change(cs, trialcs);
927 if (retval < 0)
928 return retval;
929
930 is_load_balanced = is_sched_load_balance(trialcs);
931
932 mutex_lock(&callback_mutex);
933 cpumask_copy(cs->cpus_allowed, trialcs->cpus_allowed);
934 mutex_unlock(&callback_mutex);
935
936 update_tasks_cpumask_hier(cs, true);
937
938 if (is_load_balanced)
939 rebuild_sched_domains_locked();
940 return 0;
941}
942
943/*
944 * cpuset_migrate_mm
945 *
946 * Migrate memory region from one set of nodes to another.
947 *
948 * Temporarilly set tasks mems_allowed to target nodes of migration,
949 * so that the migration code can allocate pages on these nodes.
950 *
951 * While the mm_struct we are migrating is typically from some
952 * other task, the task_struct mems_allowed that we are hacking
953 * is for our current task, which must allocate new pages for that
954 * migrating memory region.
955 */
956
957static void cpuset_migrate_mm(struct mm_struct *mm, const nodemask_t *from,
958 const nodemask_t *to)
959{
960 struct task_struct *tsk = current;
961 struct cpuset *mems_cs;
962
963 tsk->mems_allowed = *to;
964
965 do_migrate_pages(mm, from, to, MPOL_MF_MOVE_ALL);
966
967 rcu_read_lock();
968 mems_cs = effective_nodemask_cpuset(task_cs(tsk));
969 guarantee_online_mems(mems_cs, &tsk->mems_allowed);
970 rcu_read_unlock();
971}
972
973/*
974 * cpuset_change_task_nodemask - change task's mems_allowed and mempolicy
975 * @tsk: the task to change
976 * @newmems: new nodes that the task will be set
977 *
978 * In order to avoid seeing no nodes if the old and new nodes are disjoint,
979 * we structure updates as setting all new allowed nodes, then clearing newly
980 * disallowed ones.
981 */
982static void cpuset_change_task_nodemask(struct task_struct *tsk,
983 nodemask_t *newmems)
984{
985 bool need_loop;
986
987 /*
988 * Allow tasks that have access to memory reserves because they have
989 * been OOM killed to get memory anywhere.
990 */
991 if (unlikely(test_thread_flag(TIF_MEMDIE)))
992 return;
993 if (current->flags & PF_EXITING) /* Let dying task have memory */
994 return;
995
996 task_lock(tsk);
997 /*
998 * Determine if a loop is necessary if another thread is doing
999 * read_mems_allowed_begin(). If at least one node remains unchanged and
1000 * tsk does not have a mempolicy, then an empty nodemask will not be
1001 * possible when mems_allowed is larger than a word.
1002 */
1003 need_loop = task_has_mempolicy(tsk) ||
1004 !nodes_intersects(*newmems, tsk->mems_allowed);
1005
1006 if (need_loop) {
1007 local_irq_disable();
1008 write_seqcount_begin(&tsk->mems_allowed_seq);
1009 }
1010
1011 nodes_or(tsk->mems_allowed, tsk->mems_allowed, *newmems);
1012 mpol_rebind_task(tsk, newmems, MPOL_REBIND_STEP1);
1013
1014 mpol_rebind_task(tsk, newmems, MPOL_REBIND_STEP2);
1015 tsk->mems_allowed = *newmems;
1016
1017 if (need_loop) {
1018 write_seqcount_end(&tsk->mems_allowed_seq);
1019 local_irq_enable();
1020 }
1021
1022 task_unlock(tsk);
1023}
1024
1025static void *cpuset_being_rebound;
1026
1027/**
1028 * update_tasks_nodemask - Update the nodemasks of tasks in the cpuset.
1029 * @cs: the cpuset in which each task's mems_allowed mask needs to be changed
1030 *
1031 * Iterate through each task of @cs updating its mems_allowed to the
1032 * effective cpuset's. As this function is called with cpuset_mutex held,
1033 * cpuset membership stays stable.
1034 */
1035static void update_tasks_nodemask(struct cpuset *cs)
1036{
1037 static nodemask_t newmems; /* protected by cpuset_mutex */
1038 struct cpuset *mems_cs = effective_nodemask_cpuset(cs);
1039 struct css_task_iter it;
1040 struct task_struct *task;
1041
1042 cpuset_being_rebound = cs; /* causes mpol_dup() rebind */
1043
1044 guarantee_online_mems(mems_cs, &newmems);
1045
1046 /*
1047 * The mpol_rebind_mm() call takes mmap_sem, which we couldn't
1048 * take while holding tasklist_lock. Forks can happen - the
1049 * mpol_dup() cpuset_being_rebound check will catch such forks,
1050 * and rebind their vma mempolicies too. Because we still hold
1051 * the global cpuset_mutex, we know that no other rebind effort
1052 * will be contending for the global variable cpuset_being_rebound.
1053 * It's ok if we rebind the same mm twice; mpol_rebind_mm()
1054 * is idempotent. Also migrate pages in each mm to new nodes.
1055 */
1056 css_task_iter_start(&cs->css, &it);
1057 while ((task = css_task_iter_next(&it))) {
1058 struct mm_struct *mm;
1059 bool migrate;
1060
1061 cpuset_change_task_nodemask(task, &newmems);
1062
1063 mm = get_task_mm(task);
1064 if (!mm)
1065 continue;
1066
1067 migrate = is_memory_migrate(cs);
1068
1069 mpol_rebind_mm(mm, &cs->mems_allowed);
1070 if (migrate)
1071 cpuset_migrate_mm(mm, &cs->old_mems_allowed, &newmems);
1072 mmput(mm);
1073 }
1074 css_task_iter_end(&it);
1075
1076 /*
1077 * All the tasks' nodemasks have been updated, update
1078 * cs->old_mems_allowed.
1079 */
1080 cs->old_mems_allowed = newmems;
1081
1082 /* We're done rebinding vmas to this cpuset's new mems_allowed. */
1083 cpuset_being_rebound = NULL;
1084}
1085
1086/*
1087 * update_tasks_nodemask_hier - Update the nodemasks of tasks in the hierarchy.
1088 * @cs: the root cpuset of the hierarchy
1089 * @update_root: update the root cpuset or not?
1090 *
1091 * This will update nodemasks of tasks in @root_cs and all other empty cpusets
1092 * which take on nodemask of @root_cs.
1093 *
1094 * Called with cpuset_mutex held
1095 */
1096static void update_tasks_nodemask_hier(struct cpuset *root_cs, bool update_root)
1097{
1098 struct cpuset *cp;
1099 struct cgroup_subsys_state *pos_css;
1100
1101 rcu_read_lock();
1102 cpuset_for_each_descendant_pre(cp, pos_css, root_cs) {
1103 if (cp == root_cs) {
1104 if (!update_root)
1105 continue;
1106 } else {
1107 /* skip the whole subtree if @cp have some CPU */
1108 if (!nodes_empty(cp->mems_allowed)) {
1109 pos_css = css_rightmost_descendant(pos_css);
1110 continue;
1111 }
1112 }
1113 if (!css_tryget(&cp->css))
1114 continue;
1115 rcu_read_unlock();
1116
1117 update_tasks_nodemask(cp);
1118
1119 rcu_read_lock();
1120 css_put(&cp->css);
1121 }
1122 rcu_read_unlock();
1123}
1124
1125/*
1126 * Handle user request to change the 'mems' memory placement
1127 * of a cpuset. Needs to validate the request, update the
1128 * cpusets mems_allowed, and for each task in the cpuset,
1129 * update mems_allowed and rebind task's mempolicy and any vma
1130 * mempolicies and if the cpuset is marked 'memory_migrate',
1131 * migrate the tasks pages to the new memory.
1132 *
1133 * Call with cpuset_mutex held. May take callback_mutex during call.
1134 * Will take tasklist_lock, scan tasklist for tasks in cpuset cs,
1135 * lock each such tasks mm->mmap_sem, scan its vma's and rebind
1136 * their mempolicies to the cpusets new mems_allowed.
1137 */
1138static int update_nodemask(struct cpuset *cs, struct cpuset *trialcs,
1139 const char *buf)
1140{
1141 int retval;
1142
1143 /*
1144 * top_cpuset.mems_allowed tracks node_stats[N_MEMORY];
1145 * it's read-only
1146 */
1147 if (cs == &top_cpuset) {
1148 retval = -EACCES;
1149 goto done;
1150 }
1151
1152 /*
1153 * An empty mems_allowed is ok iff there are no tasks in the cpuset.
1154 * Since nodelist_parse() fails on an empty mask, we special case
1155 * that parsing. The validate_change() call ensures that cpusets
1156 * with tasks have memory.
1157 */
1158 if (!*buf) {
1159 nodes_clear(trialcs->mems_allowed);
1160 } else {
1161 retval = nodelist_parse(buf, trialcs->mems_allowed);
1162 if (retval < 0)
1163 goto done;
1164
1165 if (!nodes_subset(trialcs->mems_allowed,
1166 node_states[N_MEMORY])) {
1167 retval = -EINVAL;
1168 goto done;
1169 }
1170 }
1171
1172 if (nodes_equal(cs->mems_allowed, trialcs->mems_allowed)) {
1173 retval = 0; /* Too easy - nothing to do */
1174 goto done;
1175 }
1176 retval = validate_change(cs, trialcs);
1177 if (retval < 0)
1178 goto done;
1179
1180 mutex_lock(&callback_mutex);
1181 cs->mems_allowed = trialcs->mems_allowed;
1182 mutex_unlock(&callback_mutex);
1183
1184 update_tasks_nodemask_hier(cs, true);
1185done:
1186 return retval;
1187}
1188
1189int current_cpuset_is_being_rebound(void)
1190{
1191 return task_cs(current) == cpuset_being_rebound;
1192}
1193
1194static int update_relax_domain_level(struct cpuset *cs, s64 val)
1195{
1196#ifdef CONFIG_SMP
1197 if (val < -1 || val >= sched_domain_level_max)
1198 return -EINVAL;
1199#endif
1200
1201 if (val != cs->relax_domain_level) {
1202 cs->relax_domain_level = val;
1203 if (!cpumask_empty(cs->cpus_allowed) &&
1204 is_sched_load_balance(cs))
1205 rebuild_sched_domains_locked();
1206 }
1207
1208 return 0;
1209}
1210
1211/**
1212 * update_tasks_flags - update the spread flags of tasks in the cpuset.
1213 * @cs: the cpuset in which each task's spread flags needs to be changed
1214 *
1215 * Iterate through each task of @cs updating its spread flags. As this
1216 * function is called with cpuset_mutex held, cpuset membership stays
1217 * stable.
1218 */
1219static void update_tasks_flags(struct cpuset *cs)
1220{
1221 struct css_task_iter it;
1222 struct task_struct *task;
1223
1224 css_task_iter_start(&cs->css, &it);
1225 while ((task = css_task_iter_next(&it)))
1226 cpuset_update_task_spread_flag(cs, task);
1227 css_task_iter_end(&it);
1228}
1229
1230/*
1231 * update_flag - read a 0 or a 1 in a file and update associated flag
1232 * bit: the bit to update (see cpuset_flagbits_t)
1233 * cs: the cpuset to update
1234 * turning_on: whether the flag is being set or cleared
1235 *
1236 * Call with cpuset_mutex held.
1237 */
1238
1239static int update_flag(cpuset_flagbits_t bit, struct cpuset *cs,
1240 int turning_on)
1241{
1242 struct cpuset *trialcs;
1243 int balance_flag_changed;
1244 int spread_flag_changed;
1245 int err;
1246
1247 trialcs = alloc_trial_cpuset(cs);
1248 if (!trialcs)
1249 return -ENOMEM;
1250
1251 if (turning_on)
1252 set_bit(bit, &trialcs->flags);
1253 else
1254 clear_bit(bit, &trialcs->flags);
1255
1256 err = validate_change(cs, trialcs);
1257 if (err < 0)
1258 goto out;
1259
1260 balance_flag_changed = (is_sched_load_balance(cs) !=
1261 is_sched_load_balance(trialcs));
1262
1263 spread_flag_changed = ((is_spread_slab(cs) != is_spread_slab(trialcs))
1264 || (is_spread_page(cs) != is_spread_page(trialcs)));
1265
1266 mutex_lock(&callback_mutex);
1267 cs->flags = trialcs->flags;
1268 mutex_unlock(&callback_mutex);
1269
1270 if (!cpumask_empty(trialcs->cpus_allowed) && balance_flag_changed)
1271 rebuild_sched_domains_locked();
1272
1273 if (spread_flag_changed)
1274 update_tasks_flags(cs);
1275out:
1276 free_trial_cpuset(trialcs);
1277 return err;
1278}
1279
1280/*
1281 * Frequency meter - How fast is some event occurring?
1282 *
1283 * These routines manage a digitally filtered, constant time based,
1284 * event frequency meter. There are four routines:
1285 * fmeter_init() - initialize a frequency meter.
1286 * fmeter_markevent() - called each time the event happens.
1287 * fmeter_getrate() - returns the recent rate of such events.
1288 * fmeter_update() - internal routine used to update fmeter.
1289 *
1290 * A common data structure is passed to each of these routines,
1291 * which is used to keep track of the state required to manage the
1292 * frequency meter and its digital filter.
1293 *
1294 * The filter works on the number of events marked per unit time.
1295 * The filter is single-pole low-pass recursive (IIR). The time unit
1296 * is 1 second. Arithmetic is done using 32-bit integers scaled to
1297 * simulate 3 decimal digits of precision (multiplied by 1000).
1298 *
1299 * With an FM_COEF of 933, and a time base of 1 second, the filter
1300 * has a half-life of 10 seconds, meaning that if the events quit
1301 * happening, then the rate returned from the fmeter_getrate()
1302 * will be cut in half each 10 seconds, until it converges to zero.
1303 *
1304 * It is not worth doing a real infinitely recursive filter. If more
1305 * than FM_MAXTICKS ticks have elapsed since the last filter event,
1306 * just compute FM_MAXTICKS ticks worth, by which point the level
1307 * will be stable.
1308 *
1309 * Limit the count of unprocessed events to FM_MAXCNT, so as to avoid
1310 * arithmetic overflow in the fmeter_update() routine.
1311 *
1312 * Given the simple 32 bit integer arithmetic used, this meter works
1313 * best for reporting rates between one per millisecond (msec) and
1314 * one per 32 (approx) seconds. At constant rates faster than one
1315 * per msec it maxes out at values just under 1,000,000. At constant
1316 * rates between one per msec, and one per second it will stabilize
1317 * to a value N*1000, where N is the rate of events per second.
1318 * At constant rates between one per second and one per 32 seconds,
1319 * it will be choppy, moving up on the seconds that have an event,
1320 * and then decaying until the next event. At rates slower than
1321 * about one in 32 seconds, it decays all the way back to zero between
1322 * each event.
1323 */
1324
1325#define FM_COEF 933 /* coefficient for half-life of 10 secs */
1326#define FM_MAXTICKS ((time_t)99) /* useless computing more ticks than this */
1327#define FM_MAXCNT 1000000 /* limit cnt to avoid overflow */
1328#define FM_SCALE 1000 /* faux fixed point scale */
1329
1330/* Initialize a frequency meter */
1331static void fmeter_init(struct fmeter *fmp)
1332{
1333 fmp->cnt = 0;
1334 fmp->val = 0;
1335 fmp->time = 0;
1336 spin_lock_init(&fmp->lock);
1337}
1338
1339/* Internal meter update - process cnt events and update value */
1340static void fmeter_update(struct fmeter *fmp)
1341{
1342 time_t now = get_seconds();
1343 time_t ticks = now - fmp->time;
1344
1345 if (ticks == 0)
1346 return;
1347
1348 ticks = min(FM_MAXTICKS, ticks);
1349 while (ticks-- > 0)
1350 fmp->val = (FM_COEF * fmp->val) / FM_SCALE;
1351 fmp->time = now;
1352
1353 fmp->val += ((FM_SCALE - FM_COEF) * fmp->cnt) / FM_SCALE;
1354 fmp->cnt = 0;
1355}
1356
1357/* Process any previous ticks, then bump cnt by one (times scale). */
1358static void fmeter_markevent(struct fmeter *fmp)
1359{
1360 spin_lock(&fmp->lock);
1361 fmeter_update(fmp);
1362 fmp->cnt = min(FM_MAXCNT, fmp->cnt + FM_SCALE);
1363 spin_unlock(&fmp->lock);
1364}
1365
1366/* Process any previous ticks, then return current value. */
1367static int fmeter_getrate(struct fmeter *fmp)
1368{
1369 int val;
1370
1371 spin_lock(&fmp->lock);
1372 fmeter_update(fmp);
1373 val = fmp->val;
1374 spin_unlock(&fmp->lock);
1375 return val;
1376}
1377
1378static struct cpuset *cpuset_attach_old_cs;
1379
1380/* Called by cgroups to determine if a cpuset is usable; cpuset_mutex held */
1381static int cpuset_can_attach(struct cgroup_subsys_state *css,
1382 struct cgroup_taskset *tset)
1383{
1384 struct cpuset *cs = css_cs(css);
1385 struct task_struct *task;
1386 int ret;
1387
1388 /* used later by cpuset_attach() */
1389 cpuset_attach_old_cs = task_cs(cgroup_taskset_first(tset));
1390
1391 mutex_lock(&cpuset_mutex);
1392
1393 /*
1394 * We allow to move tasks into an empty cpuset if sane_behavior
1395 * flag is set.
1396 */
1397 ret = -ENOSPC;
1398 if (!cgroup_sane_behavior(css->cgroup) &&
1399 (cpumask_empty(cs->cpus_allowed) || nodes_empty(cs->mems_allowed)))
1400 goto out_unlock;
1401
1402 cgroup_taskset_for_each(task, tset) {
1403 /*
1404 * Kthreads which disallow setaffinity shouldn't be moved
1405 * to a new cpuset; we don't want to change their cpu
1406 * affinity and isolating such threads by their set of
1407 * allowed nodes is unnecessary. Thus, cpusets are not
1408 * applicable for such threads. This prevents checking for
1409 * success of set_cpus_allowed_ptr() on all attached tasks
1410 * before cpus_allowed may be changed.
1411 */
1412 ret = -EINVAL;
1413 if (task->flags & PF_NO_SETAFFINITY)
1414 goto out_unlock;
1415 ret = security_task_setscheduler(task);
1416 if (ret)
1417 goto out_unlock;
1418 }
1419
1420 /*
1421 * Mark attach is in progress. This makes validate_change() fail
1422 * changes which zero cpus/mems_allowed.
1423 */
1424 cs->attach_in_progress++;
1425 ret = 0;
1426out_unlock:
1427 mutex_unlock(&cpuset_mutex);
1428 return ret;
1429}
1430
1431static void cpuset_cancel_attach(struct cgroup_subsys_state *css,
1432 struct cgroup_taskset *tset)
1433{
1434 mutex_lock(&cpuset_mutex);
1435 css_cs(css)->attach_in_progress--;
1436 mutex_unlock(&cpuset_mutex);
1437}
1438
1439/*
1440 * Protected by cpuset_mutex. cpus_attach is used only by cpuset_attach()
1441 * but we can't allocate it dynamically there. Define it global and
1442 * allocate from cpuset_init().
1443 */
1444static cpumask_var_t cpus_attach;
1445
1446static void cpuset_attach(struct cgroup_subsys_state *css,
1447 struct cgroup_taskset *tset)
1448{
1449 /* static buf protected by cpuset_mutex */
1450 static nodemask_t cpuset_attach_nodemask_to;
1451 struct mm_struct *mm;
1452 struct task_struct *task;
1453 struct task_struct *leader = cgroup_taskset_first(tset);
1454 struct cpuset *cs = css_cs(css);
1455 struct cpuset *oldcs = cpuset_attach_old_cs;
1456 struct cpuset *cpus_cs = effective_cpumask_cpuset(cs);
1457 struct cpuset *mems_cs = effective_nodemask_cpuset(cs);
1458
1459 mutex_lock(&cpuset_mutex);
1460
1461 /* prepare for attach */
1462 if (cs == &top_cpuset)
1463 cpumask_copy(cpus_attach, cpu_possible_mask);
1464 else
1465 guarantee_online_cpus(cpus_cs, cpus_attach);
1466
1467 guarantee_online_mems(mems_cs, &cpuset_attach_nodemask_to);
1468
1469 cgroup_taskset_for_each(task, tset) {
1470 /*
1471 * can_attach beforehand should guarantee that this doesn't
1472 * fail. TODO: have a better way to handle failure here
1473 */
1474 WARN_ON_ONCE(set_cpus_allowed_ptr(task, cpus_attach));
1475
1476 cpuset_change_task_nodemask(task, &cpuset_attach_nodemask_to);
1477 cpuset_update_task_spread_flag(cs, task);
1478 }
1479
1480 /*
1481 * Change mm, possibly for multiple threads in a threadgroup. This is
1482 * expensive and may sleep.
1483 */
1484 cpuset_attach_nodemask_to = cs->mems_allowed;
1485 mm = get_task_mm(leader);
1486 if (mm) {
1487 struct cpuset *mems_oldcs = effective_nodemask_cpuset(oldcs);
1488
1489 mpol_rebind_mm(mm, &cpuset_attach_nodemask_to);
1490
1491 /*
1492 * old_mems_allowed is the same with mems_allowed here, except
1493 * if this task is being moved automatically due to hotplug.
1494 * In that case @mems_allowed has been updated and is empty,
1495 * so @old_mems_allowed is the right nodesets that we migrate
1496 * mm from.
1497 */
1498 if (is_memory_migrate(cs)) {
1499 cpuset_migrate_mm(mm, &mems_oldcs->old_mems_allowed,
1500 &cpuset_attach_nodemask_to);
1501 }
1502 mmput(mm);
1503 }
1504
1505 cs->old_mems_allowed = cpuset_attach_nodemask_to;
1506
1507 cs->attach_in_progress--;
1508 if (!cs->attach_in_progress)
1509 wake_up(&cpuset_attach_wq);
1510
1511 mutex_unlock(&cpuset_mutex);
1512}
1513
1514/* The various types of files and directories in a cpuset file system */
1515
1516typedef enum {
1517 FILE_MEMORY_MIGRATE,
1518 FILE_CPULIST,
1519 FILE_MEMLIST,
1520 FILE_CPU_EXCLUSIVE,
1521 FILE_MEM_EXCLUSIVE,
1522 FILE_MEM_HARDWALL,
1523 FILE_SCHED_LOAD_BALANCE,
1524 FILE_SCHED_RELAX_DOMAIN_LEVEL,
1525 FILE_MEMORY_PRESSURE_ENABLED,
1526 FILE_MEMORY_PRESSURE,
1527 FILE_SPREAD_PAGE,
1528 FILE_SPREAD_SLAB,
1529} cpuset_filetype_t;
1530
1531static int cpuset_write_u64(struct cgroup_subsys_state *css, struct cftype *cft,
1532 u64 val)
1533{
1534 struct cpuset *cs = css_cs(css);
1535 cpuset_filetype_t type = cft->private;
1536 int retval = 0;
1537
1538 mutex_lock(&cpuset_mutex);
1539 if (!is_cpuset_online(cs)) {
1540 retval = -ENODEV;
1541 goto out_unlock;
1542 }
1543
1544 switch (type) {
1545 case FILE_CPU_EXCLUSIVE:
1546 retval = update_flag(CS_CPU_EXCLUSIVE, cs, val);
1547 break;
1548 case FILE_MEM_EXCLUSIVE:
1549 retval = update_flag(CS_MEM_EXCLUSIVE, cs, val);
1550 break;
1551 case FILE_MEM_HARDWALL:
1552 retval = update_flag(CS_MEM_HARDWALL, cs, val);
1553 break;
1554 case FILE_SCHED_LOAD_BALANCE:
1555 retval = update_flag(CS_SCHED_LOAD_BALANCE, cs, val);
1556 break;
1557 case FILE_MEMORY_MIGRATE:
1558 retval = update_flag(CS_MEMORY_MIGRATE, cs, val);
1559 break;
1560 case FILE_MEMORY_PRESSURE_ENABLED:
1561 cpuset_memory_pressure_enabled = !!val;
1562 break;
1563 case FILE_MEMORY_PRESSURE:
1564 retval = -EACCES;
1565 break;
1566 case FILE_SPREAD_PAGE:
1567 retval = update_flag(CS_SPREAD_PAGE, cs, val);
1568 break;
1569 case FILE_SPREAD_SLAB:
1570 retval = update_flag(CS_SPREAD_SLAB, cs, val);
1571 break;
1572 default:
1573 retval = -EINVAL;
1574 break;
1575 }
1576out_unlock:
1577 mutex_unlock(&cpuset_mutex);
1578 return retval;
1579}
1580
1581static int cpuset_write_s64(struct cgroup_subsys_state *css, struct cftype *cft,
1582 s64 val)
1583{
1584 struct cpuset *cs = css_cs(css);
1585 cpuset_filetype_t type = cft->private;
1586 int retval = -ENODEV;
1587
1588 mutex_lock(&cpuset_mutex);
1589 if (!is_cpuset_online(cs))
1590 goto out_unlock;
1591
1592 switch (type) {
1593 case FILE_SCHED_RELAX_DOMAIN_LEVEL:
1594 retval = update_relax_domain_level(cs, val);
1595 break;
1596 default:
1597 retval = -EINVAL;
1598 break;
1599 }
1600out_unlock:
1601 mutex_unlock(&cpuset_mutex);
1602 return retval;
1603}
1604
1605/*
1606 * Common handling for a write to a "cpus" or "mems" file.
1607 */
1608static int cpuset_write_resmask(struct cgroup_subsys_state *css,
1609 struct cftype *cft, char *buf)
1610{
1611 struct cpuset *cs = css_cs(css);
1612 struct cpuset *trialcs;
1613 int retval = -ENODEV;
1614
1615 /*
1616 * CPU or memory hotunplug may leave @cs w/o any execution
1617 * resources, in which case the hotplug code asynchronously updates
1618 * configuration and transfers all tasks to the nearest ancestor
1619 * which can execute.
1620 *
1621 * As writes to "cpus" or "mems" may restore @cs's execution
1622 * resources, wait for the previously scheduled operations before
1623 * proceeding, so that we don't end up keep removing tasks added
1624 * after execution capability is restored.
1625 */
1626 flush_work(&cpuset_hotplug_work);
1627
1628 mutex_lock(&cpuset_mutex);
1629 if (!is_cpuset_online(cs))
1630 goto out_unlock;
1631
1632 trialcs = alloc_trial_cpuset(cs);
1633 if (!trialcs) {
1634 retval = -ENOMEM;
1635 goto out_unlock;
1636 }
1637
1638 switch (cft->private) {
1639 case FILE_CPULIST:
1640 retval = update_cpumask(cs, trialcs, buf);
1641 break;
1642 case FILE_MEMLIST:
1643 retval = update_nodemask(cs, trialcs, buf);
1644 break;
1645 default:
1646 retval = -EINVAL;
1647 break;
1648 }
1649
1650 free_trial_cpuset(trialcs);
1651out_unlock:
1652 mutex_unlock(&cpuset_mutex);
1653 return retval;
1654}
1655
1656/*
1657 * These ascii lists should be read in a single call, by using a user
1658 * buffer large enough to hold the entire map. If read in smaller
1659 * chunks, there is no guarantee of atomicity. Since the display format
1660 * used, list of ranges of sequential numbers, is variable length,
1661 * and since these maps can change value dynamically, one could read
1662 * gibberish by doing partial reads while a list was changing.
1663 */
1664static int cpuset_common_seq_show(struct seq_file *sf, void *v)
1665{
1666 struct cpuset *cs = css_cs(seq_css(sf));
1667 cpuset_filetype_t type = seq_cft(sf)->private;
1668 ssize_t count;
1669 char *buf, *s;
1670 int ret = 0;
1671
1672 count = seq_get_buf(sf, &buf);
1673 s = buf;
1674
1675 mutex_lock(&callback_mutex);
1676
1677 switch (type) {
1678 case FILE_CPULIST:
1679 s += cpulist_scnprintf(s, count, cs->cpus_allowed);
1680 break;
1681 case FILE_MEMLIST:
1682 s += nodelist_scnprintf(s, count, cs->mems_allowed);
1683 break;
1684 default:
1685 ret = -EINVAL;
1686 goto out_unlock;
1687 }
1688
1689 if (s < buf + count - 1) {
1690 *s++ = '\n';
1691 seq_commit(sf, s - buf);
1692 } else {
1693 seq_commit(sf, -1);
1694 }
1695out_unlock:
1696 mutex_unlock(&callback_mutex);
1697 return ret;
1698}
1699
1700static u64 cpuset_read_u64(struct cgroup_subsys_state *css, struct cftype *cft)
1701{
1702 struct cpuset *cs = css_cs(css);
1703 cpuset_filetype_t type = cft->private;
1704 switch (type) {
1705 case FILE_CPU_EXCLUSIVE:
1706 return is_cpu_exclusive(cs);
1707 case FILE_MEM_EXCLUSIVE:
1708 return is_mem_exclusive(cs);
1709 case FILE_MEM_HARDWALL:
1710 return is_mem_hardwall(cs);
1711 case FILE_SCHED_LOAD_BALANCE:
1712 return is_sched_load_balance(cs);
1713 case FILE_MEMORY_MIGRATE:
1714 return is_memory_migrate(cs);
1715 case FILE_MEMORY_PRESSURE_ENABLED:
1716 return cpuset_memory_pressure_enabled;
1717 case FILE_MEMORY_PRESSURE:
1718 return fmeter_getrate(&cs->fmeter);
1719 case FILE_SPREAD_PAGE:
1720 return is_spread_page(cs);
1721 case FILE_SPREAD_SLAB:
1722 return is_spread_slab(cs);
1723 default:
1724 BUG();
1725 }
1726
1727 /* Unreachable but makes gcc happy */
1728 return 0;
1729}
1730
1731static s64 cpuset_read_s64(struct cgroup_subsys_state *css, struct cftype *cft)
1732{
1733 struct cpuset *cs = css_cs(css);
1734 cpuset_filetype_t type = cft->private;
1735 switch (type) {
1736 case FILE_SCHED_RELAX_DOMAIN_LEVEL:
1737 return cs->relax_domain_level;
1738 default:
1739 BUG();
1740 }
1741
1742 /* Unrechable but makes gcc happy */
1743 return 0;
1744}
1745
1746
1747/*
1748 * for the common functions, 'private' gives the type of file
1749 */
1750
1751static struct cftype files[] = {
1752 {
1753 .name = "cpus",
1754 .seq_show = cpuset_common_seq_show,
1755 .write_string = cpuset_write_resmask,
1756 .max_write_len = (100U + 6 * NR_CPUS),
1757 .private = FILE_CPULIST,
1758 },
1759
1760 {
1761 .name = "mems",
1762 .seq_show = cpuset_common_seq_show,
1763 .write_string = cpuset_write_resmask,
1764 .max_write_len = (100U + 6 * MAX_NUMNODES),
1765 .private = FILE_MEMLIST,
1766 },
1767
1768 {
1769 .name = "cpu_exclusive",
1770 .read_u64 = cpuset_read_u64,
1771 .write_u64 = cpuset_write_u64,
1772 .private = FILE_CPU_EXCLUSIVE,
1773 },
1774
1775 {
1776 .name = "mem_exclusive",
1777 .read_u64 = cpuset_read_u64,
1778 .write_u64 = cpuset_write_u64,
1779 .private = FILE_MEM_EXCLUSIVE,
1780 },
1781
1782 {
1783 .name = "mem_hardwall",
1784 .read_u64 = cpuset_read_u64,
1785 .write_u64 = cpuset_write_u64,
1786 .private = FILE_MEM_HARDWALL,
1787 },
1788
1789 {
1790 .name = "sched_load_balance",
1791 .read_u64 = cpuset_read_u64,
1792 .write_u64 = cpuset_write_u64,
1793 .private = FILE_SCHED_LOAD_BALANCE,
1794 },
1795
1796 {
1797 .name = "sched_relax_domain_level",
1798 .read_s64 = cpuset_read_s64,
1799 .write_s64 = cpuset_write_s64,
1800 .private = FILE_SCHED_RELAX_DOMAIN_LEVEL,
1801 },
1802
1803 {
1804 .name = "memory_migrate",
1805 .read_u64 = cpuset_read_u64,
1806 .write_u64 = cpuset_write_u64,
1807 .private = FILE_MEMORY_MIGRATE,
1808 },
1809
1810 {
1811 .name = "memory_pressure",
1812 .read_u64 = cpuset_read_u64,
1813 .write_u64 = cpuset_write_u64,
1814 .private = FILE_MEMORY_PRESSURE,
1815 .mode = S_IRUGO,
1816 },
1817
1818 {
1819 .name = "memory_spread_page",
1820 .read_u64 = cpuset_read_u64,
1821 .write_u64 = cpuset_write_u64,
1822 .private = FILE_SPREAD_PAGE,
1823 },
1824
1825 {
1826 .name = "memory_spread_slab",
1827 .read_u64 = cpuset_read_u64,
1828 .write_u64 = cpuset_write_u64,
1829 .private = FILE_SPREAD_SLAB,
1830 },
1831
1832 {
1833 .name = "memory_pressure_enabled",
1834 .flags = CFTYPE_ONLY_ON_ROOT,
1835 .read_u64 = cpuset_read_u64,
1836 .write_u64 = cpuset_write_u64,
1837 .private = FILE_MEMORY_PRESSURE_ENABLED,
1838 },
1839
1840 { } /* terminate */
1841};
1842
1843/*
1844 * cpuset_css_alloc - allocate a cpuset css
1845 * cgrp: control group that the new cpuset will be part of
1846 */
1847
1848static struct cgroup_subsys_state *
1849cpuset_css_alloc(struct cgroup_subsys_state *parent_css)
1850{
1851 struct cpuset *cs;
1852
1853 if (!parent_css)
1854 return &top_cpuset.css;
1855
1856 cs = kzalloc(sizeof(*cs), GFP_KERNEL);
1857 if (!cs)
1858 return ERR_PTR(-ENOMEM);
1859 if (!alloc_cpumask_var(&cs->cpus_allowed, GFP_KERNEL)) {
1860 kfree(cs);
1861 return ERR_PTR(-ENOMEM);
1862 }
1863
1864 set_bit(CS_SCHED_LOAD_BALANCE, &cs->flags);
1865 cpumask_clear(cs->cpus_allowed);
1866 nodes_clear(cs->mems_allowed);
1867 fmeter_init(&cs->fmeter);
1868 cs->relax_domain_level = -1;
1869
1870 return &cs->css;
1871}
1872
1873static int cpuset_css_online(struct cgroup_subsys_state *css)
1874{
1875 struct cpuset *cs = css_cs(css);
1876 struct cpuset *parent = parent_cs(cs);
1877 struct cpuset *tmp_cs;
1878 struct cgroup_subsys_state *pos_css;
1879
1880 if (!parent)
1881 return 0;
1882
1883 mutex_lock(&cpuset_mutex);
1884
1885 set_bit(CS_ONLINE, &cs->flags);
1886 if (is_spread_page(parent))
1887 set_bit(CS_SPREAD_PAGE, &cs->flags);
1888 if (is_spread_slab(parent))
1889 set_bit(CS_SPREAD_SLAB, &cs->flags);
1890
1891 number_of_cpusets++;
1892
1893 if (!test_bit(CGRP_CPUSET_CLONE_CHILDREN, &css->cgroup->flags))
1894 goto out_unlock;
1895
1896 /*
1897 * Clone @parent's configuration if CGRP_CPUSET_CLONE_CHILDREN is
1898 * set. This flag handling is implemented in cgroup core for
1899 * histrical reasons - the flag may be specified during mount.
1900 *
1901 * Currently, if any sibling cpusets have exclusive cpus or mem, we
1902 * refuse to clone the configuration - thereby refusing the task to
1903 * be entered, and as a result refusing the sys_unshare() or
1904 * clone() which initiated it. If this becomes a problem for some
1905 * users who wish to allow that scenario, then this could be
1906 * changed to grant parent->cpus_allowed-sibling_cpus_exclusive
1907 * (and likewise for mems) to the new cgroup.
1908 */
1909 rcu_read_lock();
1910 cpuset_for_each_child(tmp_cs, pos_css, parent) {
1911 if (is_mem_exclusive(tmp_cs) || is_cpu_exclusive(tmp_cs)) {
1912 rcu_read_unlock();
1913 goto out_unlock;
1914 }
1915 }
1916 rcu_read_unlock();
1917
1918 mutex_lock(&callback_mutex);
1919 cs->mems_allowed = parent->mems_allowed;
1920 cpumask_copy(cs->cpus_allowed, parent->cpus_allowed);
1921 mutex_unlock(&callback_mutex);
1922out_unlock:
1923 mutex_unlock(&cpuset_mutex);
1924 return 0;
1925}
1926
1927/*
1928 * If the cpuset being removed has its flag 'sched_load_balance'
1929 * enabled, then simulate turning sched_load_balance off, which
1930 * will call rebuild_sched_domains_locked().
1931 */
1932
1933static void cpuset_css_offline(struct cgroup_subsys_state *css)
1934{
1935 struct cpuset *cs = css_cs(css);
1936
1937 mutex_lock(&cpuset_mutex);
1938
1939 if (is_sched_load_balance(cs))
1940 update_flag(CS_SCHED_LOAD_BALANCE, cs, 0);
1941
1942 number_of_cpusets--;
1943 clear_bit(CS_ONLINE, &cs->flags);
1944
1945 mutex_unlock(&cpuset_mutex);
1946}
1947
1948static void cpuset_css_free(struct cgroup_subsys_state *css)
1949{
1950 struct cpuset *cs = css_cs(css);
1951
1952 free_cpumask_var(cs->cpus_allowed);
1953 kfree(cs);
1954}
1955
1956struct cgroup_subsys cpuset_cgrp_subsys = {
1957 .css_alloc = cpuset_css_alloc,
1958 .css_online = cpuset_css_online,
1959 .css_offline = cpuset_css_offline,
1960 .css_free = cpuset_css_free,
1961 .can_attach = cpuset_can_attach,
1962 .cancel_attach = cpuset_cancel_attach,
1963 .attach = cpuset_attach,
1964 .base_cftypes = files,
1965 .early_init = 1,
1966};
1967
1968/**
1969 * cpuset_init - initialize cpusets at system boot
1970 *
1971 * Description: Initialize top_cpuset and the cpuset internal file system,
1972 **/
1973
1974int __init cpuset_init(void)
1975{
1976 int err = 0;
1977
1978 if (!alloc_cpumask_var(&top_cpuset.cpus_allowed, GFP_KERNEL))
1979 BUG();
1980
1981 cpumask_setall(top_cpuset.cpus_allowed);
1982 nodes_setall(top_cpuset.mems_allowed);
1983
1984 fmeter_init(&top_cpuset.fmeter);
1985 set_bit(CS_SCHED_LOAD_BALANCE, &top_cpuset.flags);
1986 top_cpuset.relax_domain_level = -1;
1987
1988 err = register_filesystem(&cpuset_fs_type);
1989 if (err < 0)
1990 return err;
1991
1992 if (!alloc_cpumask_var(&cpus_attach, GFP_KERNEL))
1993 BUG();
1994
1995 number_of_cpusets = 1;
1996 return 0;
1997}
1998
1999/*
2000 * If CPU and/or memory hotplug handlers, below, unplug any CPUs
2001 * or memory nodes, we need to walk over the cpuset hierarchy,
2002 * removing that CPU or node from all cpusets. If this removes the
2003 * last CPU or node from a cpuset, then move the tasks in the empty
2004 * cpuset to its next-highest non-empty parent.
2005 */
2006static void remove_tasks_in_empty_cpuset(struct cpuset *cs)
2007{
2008 struct cpuset *parent;
2009
2010 /*
2011 * Find its next-highest non-empty parent, (top cpuset
2012 * has online cpus, so can't be empty).
2013 */
2014 parent = parent_cs(cs);
2015 while (cpumask_empty(parent->cpus_allowed) ||
2016 nodes_empty(parent->mems_allowed))
2017 parent = parent_cs(parent);
2018
2019 if (cgroup_transfer_tasks(parent->css.cgroup, cs->css.cgroup)) {
2020 printk(KERN_ERR "cpuset: failed to transfer tasks out of empty cpuset ");
2021 pr_cont_cgroup_name(cs->css.cgroup);
2022 pr_cont("\n");
2023 }
2024}
2025
2026/**
2027 * cpuset_hotplug_update_tasks - update tasks in a cpuset for hotunplug
2028 * @cs: cpuset in interest
2029 *
2030 * Compare @cs's cpu and mem masks against top_cpuset and if some have gone
2031 * offline, update @cs accordingly. If @cs ends up with no CPU or memory,
2032 * all its tasks are moved to the nearest ancestor with both resources.
2033 */
2034static void cpuset_hotplug_update_tasks(struct cpuset *cs)
2035{
2036 static cpumask_t off_cpus;
2037 static nodemask_t off_mems;
2038 bool is_empty;
2039 bool sane = cgroup_sane_behavior(cs->css.cgroup);
2040
2041retry:
2042 wait_event(cpuset_attach_wq, cs->attach_in_progress == 0);
2043
2044 mutex_lock(&cpuset_mutex);
2045
2046 /*
2047 * We have raced with task attaching. We wait until attaching
2048 * is finished, so we won't attach a task to an empty cpuset.
2049 */
2050 if (cs->attach_in_progress) {
2051 mutex_unlock(&cpuset_mutex);
2052 goto retry;
2053 }
2054
2055 cpumask_andnot(&off_cpus, cs->cpus_allowed, top_cpuset.cpus_allowed);
2056 nodes_andnot(off_mems, cs->mems_allowed, top_cpuset.mems_allowed);
2057
2058 mutex_lock(&callback_mutex);
2059 cpumask_andnot(cs->cpus_allowed, cs->cpus_allowed, &off_cpus);
2060 mutex_unlock(&callback_mutex);
2061
2062 /*
2063 * If sane_behavior flag is set, we need to update tasks' cpumask
2064 * for empty cpuset to take on ancestor's cpumask. Otherwise, don't
2065 * call update_tasks_cpumask() if the cpuset becomes empty, as
2066 * the tasks in it will be migrated to an ancestor.
2067 */
2068 if ((sane && cpumask_empty(cs->cpus_allowed)) ||
2069 (!cpumask_empty(&off_cpus) && !cpumask_empty(cs->cpus_allowed)))
2070 update_tasks_cpumask(cs);
2071
2072 mutex_lock(&callback_mutex);
2073 nodes_andnot(cs->mems_allowed, cs->mems_allowed, off_mems);
2074 mutex_unlock(&callback_mutex);
2075
2076 /*
2077 * If sane_behavior flag is set, we need to update tasks' nodemask
2078 * for empty cpuset to take on ancestor's nodemask. Otherwise, don't
2079 * call update_tasks_nodemask() if the cpuset becomes empty, as
2080 * the tasks in it will be migratd to an ancestor.
2081 */
2082 if ((sane && nodes_empty(cs->mems_allowed)) ||
2083 (!nodes_empty(off_mems) && !nodes_empty(cs->mems_allowed)))
2084 update_tasks_nodemask(cs);
2085
2086 is_empty = cpumask_empty(cs->cpus_allowed) ||
2087 nodes_empty(cs->mems_allowed);
2088
2089 mutex_unlock(&cpuset_mutex);
2090
2091 /*
2092 * If sane_behavior flag is set, we'll keep tasks in empty cpusets.
2093 *
2094 * Otherwise move tasks to the nearest ancestor with execution
2095 * resources. This is full cgroup operation which will
2096 * also call back into cpuset. Should be done outside any lock.
2097 */
2098 if (!sane && is_empty)
2099 remove_tasks_in_empty_cpuset(cs);
2100}
2101
2102/**
2103 * cpuset_hotplug_workfn - handle CPU/memory hotunplug for a cpuset
2104 *
2105 * This function is called after either CPU or memory configuration has
2106 * changed and updates cpuset accordingly. The top_cpuset is always
2107 * synchronized to cpu_active_mask and N_MEMORY, which is necessary in
2108 * order to make cpusets transparent (of no affect) on systems that are
2109 * actively using CPU hotplug but making no active use of cpusets.
2110 *
2111 * Non-root cpusets are only affected by offlining. If any CPUs or memory
2112 * nodes have been taken down, cpuset_hotplug_update_tasks() is invoked on
2113 * all descendants.
2114 *
2115 * Note that CPU offlining during suspend is ignored. We don't modify
2116 * cpusets across suspend/resume cycles at all.
2117 */
2118static void cpuset_hotplug_workfn(struct work_struct *work)
2119{
2120 static cpumask_t new_cpus;
2121 static nodemask_t new_mems;
2122 bool cpus_updated, mems_updated;
2123
2124 mutex_lock(&cpuset_mutex);
2125
2126 /* fetch the available cpus/mems and find out which changed how */
2127 cpumask_copy(&new_cpus, cpu_active_mask);
2128 new_mems = node_states[N_MEMORY];
2129
2130 cpus_updated = !cpumask_equal(top_cpuset.cpus_allowed, &new_cpus);
2131 mems_updated = !nodes_equal(top_cpuset.mems_allowed, new_mems);
2132
2133 /* synchronize cpus_allowed to cpu_active_mask */
2134 if (cpus_updated) {
2135 mutex_lock(&callback_mutex);
2136 cpumask_copy(top_cpuset.cpus_allowed, &new_cpus);
2137 mutex_unlock(&callback_mutex);
2138 /* we don't mess with cpumasks of tasks in top_cpuset */
2139 }
2140
2141 /* synchronize mems_allowed to N_MEMORY */
2142 if (mems_updated) {
2143 mutex_lock(&callback_mutex);
2144 top_cpuset.mems_allowed = new_mems;
2145 mutex_unlock(&callback_mutex);
2146 update_tasks_nodemask(&top_cpuset);
2147 }
2148
2149 mutex_unlock(&cpuset_mutex);
2150
2151 /* if cpus or mems changed, we need to propagate to descendants */
2152 if (cpus_updated || mems_updated) {
2153 struct cpuset *cs;
2154 struct cgroup_subsys_state *pos_css;
2155
2156 rcu_read_lock();
2157 cpuset_for_each_descendant_pre(cs, pos_css, &top_cpuset) {
2158 if (cs == &top_cpuset || !css_tryget(&cs->css))
2159 continue;
2160 rcu_read_unlock();
2161
2162 cpuset_hotplug_update_tasks(cs);
2163
2164 rcu_read_lock();
2165 css_put(&cs->css);
2166 }
2167 rcu_read_unlock();
2168 }
2169
2170 /* rebuild sched domains if cpus_allowed has changed */
2171 if (cpus_updated)
2172 rebuild_sched_domains();
2173}
2174
2175void cpuset_update_active_cpus(bool cpu_online)
2176{
2177 /*
2178 * We're inside cpu hotplug critical region which usually nests
2179 * inside cgroup synchronization. Bounce actual hotplug processing
2180 * to a work item to avoid reverse locking order.
2181 *
2182 * We still need to do partition_sched_domains() synchronously;
2183 * otherwise, the scheduler will get confused and put tasks to the
2184 * dead CPU. Fall back to the default single domain.
2185 * cpuset_hotplug_workfn() will rebuild it as necessary.
2186 */
2187 partition_sched_domains(1, NULL, NULL);
2188 schedule_work(&cpuset_hotplug_work);
2189}
2190
2191/*
2192 * Keep top_cpuset.mems_allowed tracking node_states[N_MEMORY].
2193 * Call this routine anytime after node_states[N_MEMORY] changes.
2194 * See cpuset_update_active_cpus() for CPU hotplug handling.
2195 */
2196static int cpuset_track_online_nodes(struct notifier_block *self,
2197 unsigned long action, void *arg)
2198{
2199 schedule_work(&cpuset_hotplug_work);
2200 return NOTIFY_OK;
2201}
2202
2203static struct notifier_block cpuset_track_online_nodes_nb = {
2204 .notifier_call = cpuset_track_online_nodes,
2205 .priority = 10, /* ??! */
2206};
2207
2208/**
2209 * cpuset_init_smp - initialize cpus_allowed
2210 *
2211 * Description: Finish top cpuset after cpu, node maps are initialized
2212 */
2213void __init cpuset_init_smp(void)
2214{
2215 cpumask_copy(top_cpuset.cpus_allowed, cpu_active_mask);
2216 top_cpuset.mems_allowed = node_states[N_MEMORY];
2217 top_cpuset.old_mems_allowed = top_cpuset.mems_allowed;
2218
2219 register_hotmemory_notifier(&cpuset_track_online_nodes_nb);
2220}
2221
2222/**
2223 * cpuset_cpus_allowed - return cpus_allowed mask from a tasks cpuset.
2224 * @tsk: pointer to task_struct from which to obtain cpuset->cpus_allowed.
2225 * @pmask: pointer to struct cpumask variable to receive cpus_allowed set.
2226 *
2227 * Description: Returns the cpumask_var_t cpus_allowed of the cpuset
2228 * attached to the specified @tsk. Guaranteed to return some non-empty
2229 * subset of cpu_online_mask, even if this means going outside the
2230 * tasks cpuset.
2231 **/
2232
2233void cpuset_cpus_allowed(struct task_struct *tsk, struct cpumask *pmask)
2234{
2235 struct cpuset *cpus_cs;
2236
2237 mutex_lock(&callback_mutex);
2238 rcu_read_lock();
2239 cpus_cs = effective_cpumask_cpuset(task_cs(tsk));
2240 guarantee_online_cpus(cpus_cs, pmask);
2241 rcu_read_unlock();
2242 mutex_unlock(&callback_mutex);
2243}
2244
2245void cpuset_cpus_allowed_fallback(struct task_struct *tsk)
2246{
2247 struct cpuset *cpus_cs;
2248
2249 rcu_read_lock();
2250 cpus_cs = effective_cpumask_cpuset(task_cs(tsk));
2251 do_set_cpus_allowed(tsk, cpus_cs->cpus_allowed);
2252 rcu_read_unlock();
2253
2254 /*
2255 * We own tsk->cpus_allowed, nobody can change it under us.
2256 *
2257 * But we used cs && cs->cpus_allowed lockless and thus can
2258 * race with cgroup_attach_task() or update_cpumask() and get
2259 * the wrong tsk->cpus_allowed. However, both cases imply the
2260 * subsequent cpuset_change_cpumask()->set_cpus_allowed_ptr()
2261 * which takes task_rq_lock().
2262 *
2263 * If we are called after it dropped the lock we must see all
2264 * changes in tsk_cs()->cpus_allowed. Otherwise we can temporary
2265 * set any mask even if it is not right from task_cs() pov,
2266 * the pending set_cpus_allowed_ptr() will fix things.
2267 *
2268 * select_fallback_rq() will fix things ups and set cpu_possible_mask
2269 * if required.
2270 */
2271}
2272
2273void cpuset_init_current_mems_allowed(void)
2274{
2275 nodes_setall(current->mems_allowed);
2276}
2277
2278/**
2279 * cpuset_mems_allowed - return mems_allowed mask from a tasks cpuset.
2280 * @tsk: pointer to task_struct from which to obtain cpuset->mems_allowed.
2281 *
2282 * Description: Returns the nodemask_t mems_allowed of the cpuset
2283 * attached to the specified @tsk. Guaranteed to return some non-empty
2284 * subset of node_states[N_MEMORY], even if this means going outside the
2285 * tasks cpuset.
2286 **/
2287
2288nodemask_t cpuset_mems_allowed(struct task_struct *tsk)
2289{
2290 struct cpuset *mems_cs;
2291 nodemask_t mask;
2292
2293 mutex_lock(&callback_mutex);
2294 rcu_read_lock();
2295 mems_cs = effective_nodemask_cpuset(task_cs(tsk));
2296 guarantee_online_mems(mems_cs, &mask);
2297 rcu_read_unlock();
2298 mutex_unlock(&callback_mutex);
2299
2300 return mask;
2301}
2302
2303/**
2304 * cpuset_nodemask_valid_mems_allowed - check nodemask vs. curremt mems_allowed
2305 * @nodemask: the nodemask to be checked
2306 *
2307 * Are any of the nodes in the nodemask allowed in current->mems_allowed?
2308 */
2309int cpuset_nodemask_valid_mems_allowed(nodemask_t *nodemask)
2310{
2311 return nodes_intersects(*nodemask, current->mems_allowed);
2312}
2313
2314/*
2315 * nearest_hardwall_ancestor() - Returns the nearest mem_exclusive or
2316 * mem_hardwall ancestor to the specified cpuset. Call holding
2317 * callback_mutex. If no ancestor is mem_exclusive or mem_hardwall
2318 * (an unusual configuration), then returns the root cpuset.
2319 */
2320static struct cpuset *nearest_hardwall_ancestor(struct cpuset *cs)
2321{
2322 while (!(is_mem_exclusive(cs) || is_mem_hardwall(cs)) && parent_cs(cs))
2323 cs = parent_cs(cs);
2324 return cs;
2325}
2326
2327/**
2328 * cpuset_node_allowed_softwall - Can we allocate on a memory node?
2329 * @node: is this an allowed node?
2330 * @gfp_mask: memory allocation flags
2331 *
2332 * If we're in interrupt, yes, we can always allocate. If __GFP_THISNODE is
2333 * set, yes, we can always allocate. If node is in our task's mems_allowed,
2334 * yes. If it's not a __GFP_HARDWALL request and this node is in the nearest
2335 * hardwalled cpuset ancestor to this task's cpuset, yes. If the task has been
2336 * OOM killed and has access to memory reserves as specified by the TIF_MEMDIE
2337 * flag, yes.
2338 * Otherwise, no.
2339 *
2340 * If __GFP_HARDWALL is set, cpuset_node_allowed_softwall() reduces to
2341 * cpuset_node_allowed_hardwall(). Otherwise, cpuset_node_allowed_softwall()
2342 * might sleep, and might allow a node from an enclosing cpuset.
2343 *
2344 * cpuset_node_allowed_hardwall() only handles the simpler case of hardwall
2345 * cpusets, and never sleeps.
2346 *
2347 * The __GFP_THISNODE placement logic is really handled elsewhere,
2348 * by forcibly using a zonelist starting at a specified node, and by
2349 * (in get_page_from_freelist()) refusing to consider the zones for
2350 * any node on the zonelist except the first. By the time any such
2351 * calls get to this routine, we should just shut up and say 'yes'.
2352 *
2353 * GFP_USER allocations are marked with the __GFP_HARDWALL bit,
2354 * and do not allow allocations outside the current tasks cpuset
2355 * unless the task has been OOM killed as is marked TIF_MEMDIE.
2356 * GFP_KERNEL allocations are not so marked, so can escape to the
2357 * nearest enclosing hardwalled ancestor cpuset.
2358 *
2359 * Scanning up parent cpusets requires callback_mutex. The
2360 * __alloc_pages() routine only calls here with __GFP_HARDWALL bit
2361 * _not_ set if it's a GFP_KERNEL allocation, and all nodes in the
2362 * current tasks mems_allowed came up empty on the first pass over
2363 * the zonelist. So only GFP_KERNEL allocations, if all nodes in the
2364 * cpuset are short of memory, might require taking the callback_mutex
2365 * mutex.
2366 *
2367 * The first call here from mm/page_alloc:get_page_from_freelist()
2368 * has __GFP_HARDWALL set in gfp_mask, enforcing hardwall cpusets,
2369 * so no allocation on a node outside the cpuset is allowed (unless
2370 * in interrupt, of course).
2371 *
2372 * The second pass through get_page_from_freelist() doesn't even call
2373 * here for GFP_ATOMIC calls. For those calls, the __alloc_pages()
2374 * variable 'wait' is not set, and the bit ALLOC_CPUSET is not set
2375 * in alloc_flags. That logic and the checks below have the combined
2376 * affect that:
2377 * in_interrupt - any node ok (current task context irrelevant)
2378 * GFP_ATOMIC - any node ok
2379 * TIF_MEMDIE - any node ok
2380 * GFP_KERNEL - any node in enclosing hardwalled cpuset ok
2381 * GFP_USER - only nodes in current tasks mems allowed ok.
2382 *
2383 * Rule:
2384 * Don't call cpuset_node_allowed_softwall if you can't sleep, unless you
2385 * pass in the __GFP_HARDWALL flag set in gfp_flag, which disables
2386 * the code that might scan up ancestor cpusets and sleep.
2387 */
2388int __cpuset_node_allowed_softwall(int node, gfp_t gfp_mask)
2389{
2390 struct cpuset *cs; /* current cpuset ancestors */
2391 int allowed; /* is allocation in zone z allowed? */
2392
2393 if (in_interrupt() || (gfp_mask & __GFP_THISNODE))
2394 return 1;
2395 might_sleep_if(!(gfp_mask & __GFP_HARDWALL));
2396 if (node_isset(node, current->mems_allowed))
2397 return 1;
2398 /*
2399 * Allow tasks that have access to memory reserves because they have
2400 * been OOM killed to get memory anywhere.
2401 */
2402 if (unlikely(test_thread_flag(TIF_MEMDIE)))
2403 return 1;
2404 if (gfp_mask & __GFP_HARDWALL) /* If hardwall request, stop here */
2405 return 0;
2406
2407 if (current->flags & PF_EXITING) /* Let dying task have memory */
2408 return 1;
2409
2410 /* Not hardwall and node outside mems_allowed: scan up cpusets */
2411 mutex_lock(&callback_mutex);
2412
2413 rcu_read_lock();
2414 cs = nearest_hardwall_ancestor(task_cs(current));
2415 allowed = node_isset(node, cs->mems_allowed);
2416 rcu_read_unlock();
2417
2418 mutex_unlock(&callback_mutex);
2419 return allowed;
2420}
2421
2422/*
2423 * cpuset_node_allowed_hardwall - Can we allocate on a memory node?
2424 * @node: is this an allowed node?
2425 * @gfp_mask: memory allocation flags
2426 *
2427 * If we're in interrupt, yes, we can always allocate. If __GFP_THISNODE is
2428 * set, yes, we can always allocate. If node is in our task's mems_allowed,
2429 * yes. If the task has been OOM killed and has access to memory reserves as
2430 * specified by the TIF_MEMDIE flag, yes.
2431 * Otherwise, no.
2432 *
2433 * The __GFP_THISNODE placement logic is really handled elsewhere,
2434 * by forcibly using a zonelist starting at a specified node, and by
2435 * (in get_page_from_freelist()) refusing to consider the zones for
2436 * any node on the zonelist except the first. By the time any such
2437 * calls get to this routine, we should just shut up and say 'yes'.
2438 *
2439 * Unlike the cpuset_node_allowed_softwall() variant, above,
2440 * this variant requires that the node be in the current task's
2441 * mems_allowed or that we're in interrupt. It does not scan up the
2442 * cpuset hierarchy for the nearest enclosing mem_exclusive cpuset.
2443 * It never sleeps.
2444 */
2445int __cpuset_node_allowed_hardwall(int node, gfp_t gfp_mask)
2446{
2447 if (in_interrupt() || (gfp_mask & __GFP_THISNODE))
2448 return 1;
2449 if (node_isset(node, current->mems_allowed))
2450 return 1;
2451 /*
2452 * Allow tasks that have access to memory reserves because they have
2453 * been OOM killed to get memory anywhere.
2454 */
2455 if (unlikely(test_thread_flag(TIF_MEMDIE)))
2456 return 1;
2457 return 0;
2458}
2459
2460/**
2461 * cpuset_mem_spread_node() - On which node to begin search for a file page
2462 * cpuset_slab_spread_node() - On which node to begin search for a slab page
2463 *
2464 * If a task is marked PF_SPREAD_PAGE or PF_SPREAD_SLAB (as for
2465 * tasks in a cpuset with is_spread_page or is_spread_slab set),
2466 * and if the memory allocation used cpuset_mem_spread_node()
2467 * to determine on which node to start looking, as it will for
2468 * certain page cache or slab cache pages such as used for file
2469 * system buffers and inode caches, then instead of starting on the
2470 * local node to look for a free page, rather spread the starting
2471 * node around the tasks mems_allowed nodes.
2472 *
2473 * We don't have to worry about the returned node being offline
2474 * because "it can't happen", and even if it did, it would be ok.
2475 *
2476 * The routines calling guarantee_online_mems() are careful to
2477 * only set nodes in task->mems_allowed that are online. So it
2478 * should not be possible for the following code to return an
2479 * offline node. But if it did, that would be ok, as this routine
2480 * is not returning the node where the allocation must be, only
2481 * the node where the search should start. The zonelist passed to
2482 * __alloc_pages() will include all nodes. If the slab allocator
2483 * is passed an offline node, it will fall back to the local node.
2484 * See kmem_cache_alloc_node().
2485 */
2486
2487static int cpuset_spread_node(int *rotor)
2488{
2489 int node;
2490
2491 node = next_node(*rotor, current->mems_allowed);
2492 if (node == MAX_NUMNODES)
2493 node = first_node(current->mems_allowed);
2494 *rotor = node;
2495 return node;
2496}
2497
2498int cpuset_mem_spread_node(void)
2499{
2500 if (current->cpuset_mem_spread_rotor == NUMA_NO_NODE)
2501 current->cpuset_mem_spread_rotor =
2502 node_random(¤t->mems_allowed);
2503
2504 return cpuset_spread_node(¤t->cpuset_mem_spread_rotor);
2505}
2506
2507int cpuset_slab_spread_node(void)
2508{
2509 if (current->cpuset_slab_spread_rotor == NUMA_NO_NODE)
2510 current->cpuset_slab_spread_rotor =
2511 node_random(¤t->mems_allowed);
2512
2513 return cpuset_spread_node(¤t->cpuset_slab_spread_rotor);
2514}
2515
2516EXPORT_SYMBOL_GPL(cpuset_mem_spread_node);
2517
2518/**
2519 * cpuset_mems_allowed_intersects - Does @tsk1's mems_allowed intersect @tsk2's?
2520 * @tsk1: pointer to task_struct of some task.
2521 * @tsk2: pointer to task_struct of some other task.
2522 *
2523 * Description: Return true if @tsk1's mems_allowed intersects the
2524 * mems_allowed of @tsk2. Used by the OOM killer to determine if
2525 * one of the task's memory usage might impact the memory available
2526 * to the other.
2527 **/
2528
2529int cpuset_mems_allowed_intersects(const struct task_struct *tsk1,
2530 const struct task_struct *tsk2)
2531{
2532 return nodes_intersects(tsk1->mems_allowed, tsk2->mems_allowed);
2533}
2534
2535#define CPUSET_NODELIST_LEN (256)
2536
2537/**
2538 * cpuset_print_task_mems_allowed - prints task's cpuset and mems_allowed
2539 * @task: pointer to task_struct of some task.
2540 *
2541 * Description: Prints @task's name, cpuset name, and cached copy of its
2542 * mems_allowed to the kernel log.
2543 */
2544void cpuset_print_task_mems_allowed(struct task_struct *tsk)
2545{
2546 /* Statically allocated to prevent using excess stack. */
2547 static char cpuset_nodelist[CPUSET_NODELIST_LEN];
2548 static DEFINE_SPINLOCK(cpuset_buffer_lock);
2549 struct cgroup *cgrp;
2550
2551 spin_lock(&cpuset_buffer_lock);
2552 rcu_read_lock();
2553
2554 cgrp = task_cs(tsk)->css.cgroup;
2555 nodelist_scnprintf(cpuset_nodelist, CPUSET_NODELIST_LEN,
2556 tsk->mems_allowed);
2557 printk(KERN_INFO "%s cpuset=", tsk->comm);
2558 pr_cont_cgroup_name(cgrp);
2559 pr_cont(" mems_allowed=%s\n", cpuset_nodelist);
2560
2561 rcu_read_unlock();
2562 spin_unlock(&cpuset_buffer_lock);
2563}
2564
2565/*
2566 * Collection of memory_pressure is suppressed unless
2567 * this flag is enabled by writing "1" to the special
2568 * cpuset file 'memory_pressure_enabled' in the root cpuset.
2569 */
2570
2571int cpuset_memory_pressure_enabled __read_mostly;
2572
2573/**
2574 * cpuset_memory_pressure_bump - keep stats of per-cpuset reclaims.
2575 *
2576 * Keep a running average of the rate of synchronous (direct)
2577 * page reclaim efforts initiated by tasks in each cpuset.
2578 *
2579 * This represents the rate at which some task in the cpuset
2580 * ran low on memory on all nodes it was allowed to use, and
2581 * had to enter the kernels page reclaim code in an effort to
2582 * create more free memory by tossing clean pages or swapping
2583 * or writing dirty pages.
2584 *
2585 * Display to user space in the per-cpuset read-only file
2586 * "memory_pressure". Value displayed is an integer
2587 * representing the recent rate of entry into the synchronous
2588 * (direct) page reclaim by any task attached to the cpuset.
2589 **/
2590
2591void __cpuset_memory_pressure_bump(void)
2592{
2593 rcu_read_lock();
2594 fmeter_markevent(&task_cs(current)->fmeter);
2595 rcu_read_unlock();
2596}
2597
2598#ifdef CONFIG_PROC_PID_CPUSET
2599/*
2600 * proc_cpuset_show()
2601 * - Print tasks cpuset path into seq_file.
2602 * - Used for /proc/<pid>/cpuset.
2603 * - No need to task_lock(tsk) on this tsk->cpuset reference, as it
2604 * doesn't really matter if tsk->cpuset changes after we read it,
2605 * and we take cpuset_mutex, keeping cpuset_attach() from changing it
2606 * anyway.
2607 */
2608int proc_cpuset_show(struct seq_file *m, void *unused_v)
2609{
2610 struct pid *pid;
2611 struct task_struct *tsk;
2612 char *buf, *p;
2613 struct cgroup_subsys_state *css;
2614 int retval;
2615
2616 retval = -ENOMEM;
2617 buf = kmalloc(PATH_MAX, GFP_KERNEL);
2618 if (!buf)
2619 goto out;
2620
2621 retval = -ESRCH;
2622 pid = m->private;
2623 tsk = get_pid_task(pid, PIDTYPE_PID);
2624 if (!tsk)
2625 goto out_free;
2626
2627 retval = -ENAMETOOLONG;
2628 rcu_read_lock();
2629 css = task_css(tsk, cpuset_cgrp_id);
2630 p = cgroup_path(css->cgroup, buf, PATH_MAX);
2631 rcu_read_unlock();
2632 if (!p)
2633 goto out_put_task;
2634 seq_puts(m, p);
2635 seq_putc(m, '\n');
2636 retval = 0;
2637out_put_task:
2638 put_task_struct(tsk);
2639out_free:
2640 kfree(buf);
2641out:
2642 return retval;
2643}
2644#endif /* CONFIG_PROC_PID_CPUSET */
2645
2646/* Display task mems_allowed in /proc/<pid>/status file. */
2647void cpuset_task_status_allowed(struct seq_file *m, struct task_struct *task)
2648{
2649 seq_printf(m, "Mems_allowed:\t");
2650 seq_nodemask(m, &task->mems_allowed);
2651 seq_printf(m, "\n");
2652 seq_printf(m, "Mems_allowed_list:\t");
2653 seq_nodemask_list(m, &task->mems_allowed);
2654 seq_printf(m, "\n");
2655}
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/time64.h>
55#include <linux/backing-dev.h>
56#include <linux/sort.h>
57
58#include <linux/uaccess.h>
59#include <linux/atomic.h>
60#include <linux/mutex.h>
61#include <linux/cgroup.h>
62#include <linux/wait.h>
63
64DEFINE_STATIC_KEY_FALSE(cpusets_enabled_key);
65
66/* See "Frequency meter" comments, below. */
67
68struct fmeter {
69 int cnt; /* unprocessed events count */
70 int val; /* most recent output value */
71 time64_t time; /* clock (secs) when val computed */
72 spinlock_t lock; /* guards read or write of above */
73};
74
75struct cpuset {
76 struct cgroup_subsys_state css;
77
78 unsigned long flags; /* "unsigned long" so bitops work */
79
80 /*
81 * On default hierarchy:
82 *
83 * The user-configured masks can only be changed by writing to
84 * cpuset.cpus and cpuset.mems, and won't be limited by the
85 * parent masks.
86 *
87 * The effective masks is the real masks that apply to the tasks
88 * in the cpuset. They may be changed if the configured masks are
89 * changed or hotplug happens.
90 *
91 * effective_mask == configured_mask & parent's effective_mask,
92 * and if it ends up empty, it will inherit the parent's mask.
93 *
94 *
95 * On legacy hierachy:
96 *
97 * The user-configured masks are always the same with effective masks.
98 */
99
100 /* user-configured CPUs and Memory Nodes allow to tasks */
101 cpumask_var_t cpus_allowed;
102 nodemask_t mems_allowed;
103
104 /* effective CPUs and Memory Nodes allow to tasks */
105 cpumask_var_t effective_cpus;
106 nodemask_t effective_mems;
107
108 /*
109 * This is old Memory Nodes tasks took on.
110 *
111 * - top_cpuset.old_mems_allowed is initialized to mems_allowed.
112 * - A new cpuset's old_mems_allowed is initialized when some
113 * task is moved into it.
114 * - old_mems_allowed is used in cpuset_migrate_mm() when we change
115 * cpuset.mems_allowed and have tasks' nodemask updated, and
116 * then old_mems_allowed is updated to mems_allowed.
117 */
118 nodemask_t old_mems_allowed;
119
120 struct fmeter fmeter; /* memory_pressure filter */
121
122 /*
123 * Tasks are being attached to this cpuset. Used to prevent
124 * zeroing cpus/mems_allowed between ->can_attach() and ->attach().
125 */
126 int attach_in_progress;
127
128 /* partition number for rebuild_sched_domains() */
129 int pn;
130
131 /* for custom sched domain */
132 int relax_domain_level;
133};
134
135static inline struct cpuset *css_cs(struct cgroup_subsys_state *css)
136{
137 return css ? container_of(css, struct cpuset, css) : NULL;
138}
139
140/* Retrieve the cpuset for a task */
141static inline struct cpuset *task_cs(struct task_struct *task)
142{
143 return css_cs(task_css(task, cpuset_cgrp_id));
144}
145
146static inline struct cpuset *parent_cs(struct cpuset *cs)
147{
148 return css_cs(cs->css.parent);
149}
150
151#ifdef CONFIG_NUMA
152static inline bool task_has_mempolicy(struct task_struct *task)
153{
154 return task->mempolicy;
155}
156#else
157static inline bool task_has_mempolicy(struct task_struct *task)
158{
159 return false;
160}
161#endif
162
163
164/* bits in struct cpuset flags field */
165typedef enum {
166 CS_ONLINE,
167 CS_CPU_EXCLUSIVE,
168 CS_MEM_EXCLUSIVE,
169 CS_MEM_HARDWALL,
170 CS_MEMORY_MIGRATE,
171 CS_SCHED_LOAD_BALANCE,
172 CS_SPREAD_PAGE,
173 CS_SPREAD_SLAB,
174} cpuset_flagbits_t;
175
176/* convenient tests for these bits */
177static inline bool is_cpuset_online(const struct cpuset *cs)
178{
179 return test_bit(CS_ONLINE, &cs->flags);
180}
181
182static inline int is_cpu_exclusive(const struct cpuset *cs)
183{
184 return test_bit(CS_CPU_EXCLUSIVE, &cs->flags);
185}
186
187static inline int is_mem_exclusive(const struct cpuset *cs)
188{
189 return test_bit(CS_MEM_EXCLUSIVE, &cs->flags);
190}
191
192static inline int is_mem_hardwall(const struct cpuset *cs)
193{
194 return test_bit(CS_MEM_HARDWALL, &cs->flags);
195}
196
197static inline int is_sched_load_balance(const struct cpuset *cs)
198{
199 return test_bit(CS_SCHED_LOAD_BALANCE, &cs->flags);
200}
201
202static inline int is_memory_migrate(const struct cpuset *cs)
203{
204 return test_bit(CS_MEMORY_MIGRATE, &cs->flags);
205}
206
207static inline int is_spread_page(const struct cpuset *cs)
208{
209 return test_bit(CS_SPREAD_PAGE, &cs->flags);
210}
211
212static inline int is_spread_slab(const struct cpuset *cs)
213{
214 return test_bit(CS_SPREAD_SLAB, &cs->flags);
215}
216
217static struct cpuset top_cpuset = {
218 .flags = ((1 << CS_ONLINE) | (1 << CS_CPU_EXCLUSIVE) |
219 (1 << CS_MEM_EXCLUSIVE)),
220};
221
222/**
223 * cpuset_for_each_child - traverse online children of a cpuset
224 * @child_cs: loop cursor pointing to the current child
225 * @pos_css: used for iteration
226 * @parent_cs: target cpuset to walk children of
227 *
228 * Walk @child_cs through the online children of @parent_cs. Must be used
229 * with RCU read locked.
230 */
231#define cpuset_for_each_child(child_cs, pos_css, parent_cs) \
232 css_for_each_child((pos_css), &(parent_cs)->css) \
233 if (is_cpuset_online(((child_cs) = css_cs((pos_css)))))
234
235/**
236 * cpuset_for_each_descendant_pre - pre-order walk of a cpuset's descendants
237 * @des_cs: loop cursor pointing to the current descendant
238 * @pos_css: used for iteration
239 * @root_cs: target cpuset to walk ancestor of
240 *
241 * Walk @des_cs through the online descendants of @root_cs. Must be used
242 * with RCU read locked. The caller may modify @pos_css by calling
243 * css_rightmost_descendant() to skip subtree. @root_cs is included in the
244 * iteration and the first node to be visited.
245 */
246#define cpuset_for_each_descendant_pre(des_cs, pos_css, root_cs) \
247 css_for_each_descendant_pre((pos_css), &(root_cs)->css) \
248 if (is_cpuset_online(((des_cs) = css_cs((pos_css)))))
249
250/*
251 * There are two global locks guarding cpuset structures - cpuset_mutex and
252 * callback_lock. We also require taking task_lock() when dereferencing a
253 * task's cpuset pointer. See "The task_lock() exception", at the end of this
254 * comment.
255 *
256 * A task must hold both locks to modify cpusets. If a task holds
257 * cpuset_mutex, then it blocks others wanting that mutex, ensuring that it
258 * is the only task able to also acquire callback_lock and be able to
259 * modify cpusets. It can perform various checks on the cpuset structure
260 * first, knowing nothing will change. It can also allocate memory while
261 * just holding cpuset_mutex. While it is performing these checks, various
262 * callback routines can briefly acquire callback_lock to query cpusets.
263 * Once it is ready to make the changes, it takes callback_lock, blocking
264 * everyone else.
265 *
266 * Calls to the kernel memory allocator can not be made while holding
267 * callback_lock, as that would risk double tripping on callback_lock
268 * from one of the callbacks into the cpuset code from within
269 * __alloc_pages().
270 *
271 * If a task is only holding callback_lock, then it has read-only
272 * access to cpusets.
273 *
274 * Now, the task_struct fields mems_allowed and mempolicy may be changed
275 * by other task, we use alloc_lock in the task_struct fields to protect
276 * them.
277 *
278 * The cpuset_common_file_read() handlers only hold callback_lock across
279 * small pieces of code, such as when reading out possibly multi-word
280 * cpumasks and nodemasks.
281 *
282 * Accessing a task's cpuset should be done in accordance with the
283 * guidelines for accessing subsystem state in kernel/cgroup.c
284 */
285
286static DEFINE_MUTEX(cpuset_mutex);
287static DEFINE_SPINLOCK(callback_lock);
288
289static struct workqueue_struct *cpuset_migrate_mm_wq;
290
291/*
292 * CPU / memory hotplug is handled asynchronously.
293 */
294static void cpuset_hotplug_workfn(struct work_struct *work);
295static DECLARE_WORK(cpuset_hotplug_work, cpuset_hotplug_workfn);
296
297static DECLARE_WAIT_QUEUE_HEAD(cpuset_attach_wq);
298
299/*
300 * This is ugly, but preserves the userspace API for existing cpuset
301 * users. If someone tries to mount the "cpuset" filesystem, we
302 * silently switch it to mount "cgroup" instead
303 */
304static struct dentry *cpuset_mount(struct file_system_type *fs_type,
305 int flags, const char *unused_dev_name, void *data)
306{
307 struct file_system_type *cgroup_fs = get_fs_type("cgroup");
308 struct dentry *ret = ERR_PTR(-ENODEV);
309 if (cgroup_fs) {
310 char mountopts[] =
311 "cpuset,noprefix,"
312 "release_agent=/sbin/cpuset_release_agent";
313 ret = cgroup_fs->mount(cgroup_fs, flags,
314 unused_dev_name, mountopts);
315 put_filesystem(cgroup_fs);
316 }
317 return ret;
318}
319
320static struct file_system_type cpuset_fs_type = {
321 .name = "cpuset",
322 .mount = cpuset_mount,
323};
324
325/*
326 * Return in pmask the portion of a cpusets's cpus_allowed that
327 * are online. If none are online, walk up the cpuset hierarchy
328 * until we find one that does have some online cpus.
329 *
330 * One way or another, we guarantee to return some non-empty subset
331 * of cpu_online_mask.
332 *
333 * Call with callback_lock or cpuset_mutex held.
334 */
335static void guarantee_online_cpus(struct cpuset *cs, struct cpumask *pmask)
336{
337 while (!cpumask_intersects(cs->effective_cpus, cpu_online_mask)) {
338 cs = parent_cs(cs);
339 if (unlikely(!cs)) {
340 /*
341 * The top cpuset doesn't have any online cpu as a
342 * consequence of a race between cpuset_hotplug_work
343 * and cpu hotplug notifier. But we know the top
344 * cpuset's effective_cpus is on its way to to be
345 * identical to cpu_online_mask.
346 */
347 cpumask_copy(pmask, cpu_online_mask);
348 return;
349 }
350 }
351 cpumask_and(pmask, cs->effective_cpus, cpu_online_mask);
352}
353
354/*
355 * Return in *pmask the portion of a cpusets's mems_allowed that
356 * are online, with memory. If none are online with memory, walk
357 * up the cpuset hierarchy until we find one that does have some
358 * online mems. The top cpuset always has some mems online.
359 *
360 * One way or another, we guarantee to return some non-empty subset
361 * of node_states[N_MEMORY].
362 *
363 * Call with callback_lock or cpuset_mutex held.
364 */
365static void guarantee_online_mems(struct cpuset *cs, nodemask_t *pmask)
366{
367 while (!nodes_intersects(cs->effective_mems, node_states[N_MEMORY]))
368 cs = parent_cs(cs);
369 nodes_and(*pmask, cs->effective_mems, node_states[N_MEMORY]);
370}
371
372/*
373 * update task's spread flag if cpuset's page/slab spread flag is set
374 *
375 * Call with callback_lock or cpuset_mutex held.
376 */
377static void cpuset_update_task_spread_flag(struct cpuset *cs,
378 struct task_struct *tsk)
379{
380 if (is_spread_page(cs))
381 task_set_spread_page(tsk);
382 else
383 task_clear_spread_page(tsk);
384
385 if (is_spread_slab(cs))
386 task_set_spread_slab(tsk);
387 else
388 task_clear_spread_slab(tsk);
389}
390
391/*
392 * is_cpuset_subset(p, q) - Is cpuset p a subset of cpuset q?
393 *
394 * One cpuset is a subset of another if all its allowed CPUs and
395 * Memory Nodes are a subset of the other, and its exclusive flags
396 * are only set if the other's are set. Call holding cpuset_mutex.
397 */
398
399static int is_cpuset_subset(const struct cpuset *p, const struct cpuset *q)
400{
401 return cpumask_subset(p->cpus_allowed, q->cpus_allowed) &&
402 nodes_subset(p->mems_allowed, q->mems_allowed) &&
403 is_cpu_exclusive(p) <= is_cpu_exclusive(q) &&
404 is_mem_exclusive(p) <= is_mem_exclusive(q);
405}
406
407/**
408 * alloc_trial_cpuset - allocate a trial cpuset
409 * @cs: the cpuset that the trial cpuset duplicates
410 */
411static struct cpuset *alloc_trial_cpuset(struct cpuset *cs)
412{
413 struct cpuset *trial;
414
415 trial = kmemdup(cs, sizeof(*cs), GFP_KERNEL);
416 if (!trial)
417 return NULL;
418
419 if (!alloc_cpumask_var(&trial->cpus_allowed, GFP_KERNEL))
420 goto free_cs;
421 if (!alloc_cpumask_var(&trial->effective_cpus, GFP_KERNEL))
422 goto free_cpus;
423
424 cpumask_copy(trial->cpus_allowed, cs->cpus_allowed);
425 cpumask_copy(trial->effective_cpus, cs->effective_cpus);
426 return trial;
427
428free_cpus:
429 free_cpumask_var(trial->cpus_allowed);
430free_cs:
431 kfree(trial);
432 return NULL;
433}
434
435/**
436 * free_trial_cpuset - free the trial cpuset
437 * @trial: the trial cpuset to be freed
438 */
439static void free_trial_cpuset(struct cpuset *trial)
440{
441 free_cpumask_var(trial->effective_cpus);
442 free_cpumask_var(trial->cpus_allowed);
443 kfree(trial);
444}
445
446/*
447 * validate_change() - Used to validate that any proposed cpuset change
448 * follows the structural rules for cpusets.
449 *
450 * If we replaced the flag and mask values of the current cpuset
451 * (cur) with those values in the trial cpuset (trial), would
452 * our various subset and exclusive rules still be valid? Presumes
453 * cpuset_mutex held.
454 *
455 * 'cur' is the address of an actual, in-use cpuset. Operations
456 * such as list traversal that depend on the actual address of the
457 * cpuset in the list must use cur below, not trial.
458 *
459 * 'trial' is the address of bulk structure copy of cur, with
460 * perhaps one or more of the fields cpus_allowed, mems_allowed,
461 * or flags changed to new, trial values.
462 *
463 * Return 0 if valid, -errno if not.
464 */
465
466static int validate_change(struct cpuset *cur, struct cpuset *trial)
467{
468 struct cgroup_subsys_state *css;
469 struct cpuset *c, *par;
470 int ret;
471
472 rcu_read_lock();
473
474 /* Each of our child cpusets must be a subset of us */
475 ret = -EBUSY;
476 cpuset_for_each_child(c, css, cur)
477 if (!is_cpuset_subset(c, trial))
478 goto out;
479
480 /* Remaining checks don't apply to root cpuset */
481 ret = 0;
482 if (cur == &top_cpuset)
483 goto out;
484
485 par = parent_cs(cur);
486
487 /* On legacy hiearchy, we must be a subset of our parent cpuset. */
488 ret = -EACCES;
489 if (!cgroup_subsys_on_dfl(cpuset_cgrp_subsys) &&
490 !is_cpuset_subset(trial, par))
491 goto out;
492
493 /*
494 * If either I or some sibling (!= me) is exclusive, we can't
495 * overlap
496 */
497 ret = -EINVAL;
498 cpuset_for_each_child(c, css, par) {
499 if ((is_cpu_exclusive(trial) || is_cpu_exclusive(c)) &&
500 c != cur &&
501 cpumask_intersects(trial->cpus_allowed, c->cpus_allowed))
502 goto out;
503 if ((is_mem_exclusive(trial) || is_mem_exclusive(c)) &&
504 c != cur &&
505 nodes_intersects(trial->mems_allowed, c->mems_allowed))
506 goto out;
507 }
508
509 /*
510 * Cpusets with tasks - existing or newly being attached - can't
511 * be changed to have empty cpus_allowed or mems_allowed.
512 */
513 ret = -ENOSPC;
514 if ((cgroup_is_populated(cur->css.cgroup) || cur->attach_in_progress)) {
515 if (!cpumask_empty(cur->cpus_allowed) &&
516 cpumask_empty(trial->cpus_allowed))
517 goto out;
518 if (!nodes_empty(cur->mems_allowed) &&
519 nodes_empty(trial->mems_allowed))
520 goto out;
521 }
522
523 /*
524 * We can't shrink if we won't have enough room for SCHED_DEADLINE
525 * tasks.
526 */
527 ret = -EBUSY;
528 if (is_cpu_exclusive(cur) &&
529 !cpuset_cpumask_can_shrink(cur->cpus_allowed,
530 trial->cpus_allowed))
531 goto out;
532
533 ret = 0;
534out:
535 rcu_read_unlock();
536 return ret;
537}
538
539#ifdef CONFIG_SMP
540/*
541 * Helper routine for generate_sched_domains().
542 * Do cpusets a, b have overlapping effective cpus_allowed masks?
543 */
544static int cpusets_overlap(struct cpuset *a, struct cpuset *b)
545{
546 return cpumask_intersects(a->effective_cpus, b->effective_cpus);
547}
548
549static void
550update_domain_attr(struct sched_domain_attr *dattr, struct cpuset *c)
551{
552 if (dattr->relax_domain_level < c->relax_domain_level)
553 dattr->relax_domain_level = c->relax_domain_level;
554 return;
555}
556
557static void update_domain_attr_tree(struct sched_domain_attr *dattr,
558 struct cpuset *root_cs)
559{
560 struct cpuset *cp;
561 struct cgroup_subsys_state *pos_css;
562
563 rcu_read_lock();
564 cpuset_for_each_descendant_pre(cp, pos_css, root_cs) {
565 /* skip the whole subtree if @cp doesn't have any CPU */
566 if (cpumask_empty(cp->cpus_allowed)) {
567 pos_css = css_rightmost_descendant(pos_css);
568 continue;
569 }
570
571 if (is_sched_load_balance(cp))
572 update_domain_attr(dattr, cp);
573 }
574 rcu_read_unlock();
575}
576
577/*
578 * generate_sched_domains()
579 *
580 * This function builds a partial partition of the systems CPUs
581 * A 'partial partition' is a set of non-overlapping subsets whose
582 * union is a subset of that set.
583 * The output of this function needs to be passed to kernel/sched/core.c
584 * partition_sched_domains() routine, which will rebuild the scheduler's
585 * load balancing domains (sched domains) as specified by that partial
586 * partition.
587 *
588 * See "What is sched_load_balance" in Documentation/cgroups/cpusets.txt
589 * for a background explanation of this.
590 *
591 * Does not return errors, on the theory that the callers of this
592 * routine would rather not worry about failures to rebuild sched
593 * domains when operating in the severe memory shortage situations
594 * that could cause allocation failures below.
595 *
596 * Must be called with cpuset_mutex held.
597 *
598 * The three key local variables below are:
599 * q - a linked-list queue of cpuset pointers, used to implement a
600 * top-down scan of all cpusets. This scan loads a pointer
601 * to each cpuset marked is_sched_load_balance into the
602 * array 'csa'. For our purposes, rebuilding the schedulers
603 * sched domains, we can ignore !is_sched_load_balance cpusets.
604 * csa - (for CpuSet Array) Array of pointers to all the cpusets
605 * that need to be load balanced, for convenient iterative
606 * access by the subsequent code that finds the best partition,
607 * i.e the set of domains (subsets) of CPUs such that the
608 * cpus_allowed of every cpuset marked is_sched_load_balance
609 * is a subset of one of these domains, while there are as
610 * many such domains as possible, each as small as possible.
611 * doms - Conversion of 'csa' to an array of cpumasks, for passing to
612 * the kernel/sched/core.c routine partition_sched_domains() in a
613 * convenient format, that can be easily compared to the prior
614 * value to determine what partition elements (sched domains)
615 * were changed (added or removed.)
616 *
617 * Finding the best partition (set of domains):
618 * The triple nested loops below over i, j, k scan over the
619 * load balanced cpusets (using the array of cpuset pointers in
620 * csa[]) looking for pairs of cpusets that have overlapping
621 * cpus_allowed, but which don't have the same 'pn' partition
622 * number and gives them in the same partition number. It keeps
623 * looping on the 'restart' label until it can no longer find
624 * any such pairs.
625 *
626 * The union of the cpus_allowed masks from the set of
627 * all cpusets having the same 'pn' value then form the one
628 * element of the partition (one sched domain) to be passed to
629 * partition_sched_domains().
630 */
631static int generate_sched_domains(cpumask_var_t **domains,
632 struct sched_domain_attr **attributes)
633{
634 struct cpuset *cp; /* scans q */
635 struct cpuset **csa; /* array of all cpuset ptrs */
636 int csn; /* how many cpuset ptrs in csa so far */
637 int i, j, k; /* indices for partition finding loops */
638 cpumask_var_t *doms; /* resulting partition; i.e. sched domains */
639 cpumask_var_t non_isolated_cpus; /* load balanced CPUs */
640 struct sched_domain_attr *dattr; /* attributes for custom domains */
641 int ndoms = 0; /* number of sched domains in result */
642 int nslot; /* next empty doms[] struct cpumask slot */
643 struct cgroup_subsys_state *pos_css;
644
645 doms = NULL;
646 dattr = NULL;
647 csa = NULL;
648
649 if (!alloc_cpumask_var(&non_isolated_cpus, GFP_KERNEL))
650 goto done;
651 cpumask_andnot(non_isolated_cpus, cpu_possible_mask, cpu_isolated_map);
652
653 /* Special case for the 99% of systems with one, full, sched domain */
654 if (is_sched_load_balance(&top_cpuset)) {
655 ndoms = 1;
656 doms = alloc_sched_domains(ndoms);
657 if (!doms)
658 goto done;
659
660 dattr = kmalloc(sizeof(struct sched_domain_attr), GFP_KERNEL);
661 if (dattr) {
662 *dattr = SD_ATTR_INIT;
663 update_domain_attr_tree(dattr, &top_cpuset);
664 }
665 cpumask_and(doms[0], top_cpuset.effective_cpus,
666 non_isolated_cpus);
667
668 goto done;
669 }
670
671 csa = kmalloc(nr_cpusets() * sizeof(cp), GFP_KERNEL);
672 if (!csa)
673 goto done;
674 csn = 0;
675
676 rcu_read_lock();
677 cpuset_for_each_descendant_pre(cp, pos_css, &top_cpuset) {
678 if (cp == &top_cpuset)
679 continue;
680 /*
681 * Continue traversing beyond @cp iff @cp has some CPUs and
682 * isn't load balancing. The former is obvious. The
683 * latter: All child cpusets contain a subset of the
684 * parent's cpus, so just skip them, and then we call
685 * update_domain_attr_tree() to calc relax_domain_level of
686 * the corresponding sched domain.
687 */
688 if (!cpumask_empty(cp->cpus_allowed) &&
689 !(is_sched_load_balance(cp) &&
690 cpumask_intersects(cp->cpus_allowed, non_isolated_cpus)))
691 continue;
692
693 if (is_sched_load_balance(cp))
694 csa[csn++] = cp;
695
696 /* skip @cp's subtree */
697 pos_css = css_rightmost_descendant(pos_css);
698 }
699 rcu_read_unlock();
700
701 for (i = 0; i < csn; i++)
702 csa[i]->pn = i;
703 ndoms = csn;
704
705restart:
706 /* Find the best partition (set of sched domains) */
707 for (i = 0; i < csn; i++) {
708 struct cpuset *a = csa[i];
709 int apn = a->pn;
710
711 for (j = 0; j < csn; j++) {
712 struct cpuset *b = csa[j];
713 int bpn = b->pn;
714
715 if (apn != bpn && cpusets_overlap(a, b)) {
716 for (k = 0; k < csn; k++) {
717 struct cpuset *c = csa[k];
718
719 if (c->pn == bpn)
720 c->pn = apn;
721 }
722 ndoms--; /* one less element */
723 goto restart;
724 }
725 }
726 }
727
728 /*
729 * Now we know how many domains to create.
730 * Convert <csn, csa> to <ndoms, doms> and populate cpu masks.
731 */
732 doms = alloc_sched_domains(ndoms);
733 if (!doms)
734 goto done;
735
736 /*
737 * The rest of the code, including the scheduler, can deal with
738 * dattr==NULL case. No need to abort if alloc fails.
739 */
740 dattr = kmalloc(ndoms * sizeof(struct sched_domain_attr), GFP_KERNEL);
741
742 for (nslot = 0, i = 0; i < csn; i++) {
743 struct cpuset *a = csa[i];
744 struct cpumask *dp;
745 int apn = a->pn;
746
747 if (apn < 0) {
748 /* Skip completed partitions */
749 continue;
750 }
751
752 dp = doms[nslot];
753
754 if (nslot == ndoms) {
755 static int warnings = 10;
756 if (warnings) {
757 pr_warn("rebuild_sched_domains confused: nslot %d, ndoms %d, csn %d, i %d, apn %d\n",
758 nslot, ndoms, csn, i, apn);
759 warnings--;
760 }
761 continue;
762 }
763
764 cpumask_clear(dp);
765 if (dattr)
766 *(dattr + nslot) = SD_ATTR_INIT;
767 for (j = i; j < csn; j++) {
768 struct cpuset *b = csa[j];
769
770 if (apn == b->pn) {
771 cpumask_or(dp, dp, b->effective_cpus);
772 cpumask_and(dp, dp, non_isolated_cpus);
773 if (dattr)
774 update_domain_attr_tree(dattr + nslot, b);
775
776 /* Done with this partition */
777 b->pn = -1;
778 }
779 }
780 nslot++;
781 }
782 BUG_ON(nslot != ndoms);
783
784done:
785 free_cpumask_var(non_isolated_cpus);
786 kfree(csa);
787
788 /*
789 * Fallback to the default domain if kmalloc() failed.
790 * See comments in partition_sched_domains().
791 */
792 if (doms == NULL)
793 ndoms = 1;
794
795 *domains = doms;
796 *attributes = dattr;
797 return ndoms;
798}
799
800/*
801 * Rebuild scheduler domains.
802 *
803 * If the flag 'sched_load_balance' of any cpuset with non-empty
804 * 'cpus' changes, or if the 'cpus' allowed changes in any cpuset
805 * which has that flag enabled, or if any cpuset with a non-empty
806 * 'cpus' is removed, then call this routine to rebuild the
807 * scheduler's dynamic sched domains.
808 *
809 * Call with cpuset_mutex held. Takes get_online_cpus().
810 */
811static void rebuild_sched_domains_locked(void)
812{
813 struct sched_domain_attr *attr;
814 cpumask_var_t *doms;
815 int ndoms;
816
817 lockdep_assert_held(&cpuset_mutex);
818 get_online_cpus();
819
820 /*
821 * We have raced with CPU hotplug. Don't do anything to avoid
822 * passing doms with offlined cpu to partition_sched_domains().
823 * Anyways, hotplug work item will rebuild sched domains.
824 */
825 if (!cpumask_equal(top_cpuset.effective_cpus, cpu_active_mask))
826 goto out;
827
828 /* Generate domain masks and attrs */
829 ndoms = generate_sched_domains(&doms, &attr);
830
831 /* Have scheduler rebuild the domains */
832 partition_sched_domains(ndoms, doms, attr);
833out:
834 put_online_cpus();
835}
836#else /* !CONFIG_SMP */
837static void rebuild_sched_domains_locked(void)
838{
839}
840#endif /* CONFIG_SMP */
841
842void rebuild_sched_domains(void)
843{
844 mutex_lock(&cpuset_mutex);
845 rebuild_sched_domains_locked();
846 mutex_unlock(&cpuset_mutex);
847}
848
849/**
850 * update_tasks_cpumask - Update the cpumasks of tasks in the cpuset.
851 * @cs: the cpuset in which each task's cpus_allowed mask needs to be changed
852 *
853 * Iterate through each task of @cs updating its cpus_allowed to the
854 * effective cpuset's. As this function is called with cpuset_mutex held,
855 * cpuset membership stays stable.
856 */
857static void update_tasks_cpumask(struct cpuset *cs)
858{
859 struct css_task_iter it;
860 struct task_struct *task;
861
862 css_task_iter_start(&cs->css, &it);
863 while ((task = css_task_iter_next(&it)))
864 set_cpus_allowed_ptr(task, cs->effective_cpus);
865 css_task_iter_end(&it);
866}
867
868/*
869 * update_cpumasks_hier - Update effective cpumasks and tasks in the subtree
870 * @cs: the cpuset to consider
871 * @new_cpus: temp variable for calculating new effective_cpus
872 *
873 * When congifured cpumask is changed, the effective cpumasks of this cpuset
874 * and all its descendants need to be updated.
875 *
876 * On legacy hierachy, effective_cpus will be the same with cpu_allowed.
877 *
878 * Called with cpuset_mutex held
879 */
880static void update_cpumasks_hier(struct cpuset *cs, struct cpumask *new_cpus)
881{
882 struct cpuset *cp;
883 struct cgroup_subsys_state *pos_css;
884 bool need_rebuild_sched_domains = false;
885
886 rcu_read_lock();
887 cpuset_for_each_descendant_pre(cp, pos_css, cs) {
888 struct cpuset *parent = parent_cs(cp);
889
890 cpumask_and(new_cpus, cp->cpus_allowed, parent->effective_cpus);
891
892 /*
893 * If it becomes empty, inherit the effective mask of the
894 * parent, which is guaranteed to have some CPUs.
895 */
896 if (cgroup_subsys_on_dfl(cpuset_cgrp_subsys) &&
897 cpumask_empty(new_cpus))
898 cpumask_copy(new_cpus, parent->effective_cpus);
899
900 /* Skip the whole subtree if the cpumask remains the same. */
901 if (cpumask_equal(new_cpus, cp->effective_cpus)) {
902 pos_css = css_rightmost_descendant(pos_css);
903 continue;
904 }
905
906 if (!css_tryget_online(&cp->css))
907 continue;
908 rcu_read_unlock();
909
910 spin_lock_irq(&callback_lock);
911 cpumask_copy(cp->effective_cpus, new_cpus);
912 spin_unlock_irq(&callback_lock);
913
914 WARN_ON(!cgroup_subsys_on_dfl(cpuset_cgrp_subsys) &&
915 !cpumask_equal(cp->cpus_allowed, cp->effective_cpus));
916
917 update_tasks_cpumask(cp);
918
919 /*
920 * If the effective cpumask of any non-empty cpuset is changed,
921 * we need to rebuild sched domains.
922 */
923 if (!cpumask_empty(cp->cpus_allowed) &&
924 is_sched_load_balance(cp))
925 need_rebuild_sched_domains = true;
926
927 rcu_read_lock();
928 css_put(&cp->css);
929 }
930 rcu_read_unlock();
931
932 if (need_rebuild_sched_domains)
933 rebuild_sched_domains_locked();
934}
935
936/**
937 * update_cpumask - update the cpus_allowed mask of a cpuset and all tasks in it
938 * @cs: the cpuset to consider
939 * @trialcs: trial cpuset
940 * @buf: buffer of cpu numbers written to this cpuset
941 */
942static int update_cpumask(struct cpuset *cs, struct cpuset *trialcs,
943 const char *buf)
944{
945 int retval;
946
947 /* top_cpuset.cpus_allowed tracks cpu_online_mask; it's read-only */
948 if (cs == &top_cpuset)
949 return -EACCES;
950
951 /*
952 * An empty cpus_allowed is ok only if the cpuset has no tasks.
953 * Since cpulist_parse() fails on an empty mask, we special case
954 * that parsing. The validate_change() call ensures that cpusets
955 * with tasks have cpus.
956 */
957 if (!*buf) {
958 cpumask_clear(trialcs->cpus_allowed);
959 } else {
960 retval = cpulist_parse(buf, trialcs->cpus_allowed);
961 if (retval < 0)
962 return retval;
963
964 if (!cpumask_subset(trialcs->cpus_allowed,
965 top_cpuset.cpus_allowed))
966 return -EINVAL;
967 }
968
969 /* Nothing to do if the cpus didn't change */
970 if (cpumask_equal(cs->cpus_allowed, trialcs->cpus_allowed))
971 return 0;
972
973 retval = validate_change(cs, trialcs);
974 if (retval < 0)
975 return retval;
976
977 spin_lock_irq(&callback_lock);
978 cpumask_copy(cs->cpus_allowed, trialcs->cpus_allowed);
979 spin_unlock_irq(&callback_lock);
980
981 /* use trialcs->cpus_allowed as a temp variable */
982 update_cpumasks_hier(cs, trialcs->cpus_allowed);
983 return 0;
984}
985
986/*
987 * Migrate memory region from one set of nodes to another. This is
988 * performed asynchronously as it can be called from process migration path
989 * holding locks involved in process management. All mm migrations are
990 * performed in the queued order and can be waited for by flushing
991 * cpuset_migrate_mm_wq.
992 */
993
994struct cpuset_migrate_mm_work {
995 struct work_struct work;
996 struct mm_struct *mm;
997 nodemask_t from;
998 nodemask_t to;
999};
1000
1001static void cpuset_migrate_mm_workfn(struct work_struct *work)
1002{
1003 struct cpuset_migrate_mm_work *mwork =
1004 container_of(work, struct cpuset_migrate_mm_work, work);
1005
1006 /* on a wq worker, no need to worry about %current's mems_allowed */
1007 do_migrate_pages(mwork->mm, &mwork->from, &mwork->to, MPOL_MF_MOVE_ALL);
1008 mmput(mwork->mm);
1009 kfree(mwork);
1010}
1011
1012static void cpuset_migrate_mm(struct mm_struct *mm, const nodemask_t *from,
1013 const nodemask_t *to)
1014{
1015 struct cpuset_migrate_mm_work *mwork;
1016
1017 mwork = kzalloc(sizeof(*mwork), GFP_KERNEL);
1018 if (mwork) {
1019 mwork->mm = mm;
1020 mwork->from = *from;
1021 mwork->to = *to;
1022 INIT_WORK(&mwork->work, cpuset_migrate_mm_workfn);
1023 queue_work(cpuset_migrate_mm_wq, &mwork->work);
1024 } else {
1025 mmput(mm);
1026 }
1027}
1028
1029static void cpuset_post_attach(void)
1030{
1031 flush_workqueue(cpuset_migrate_mm_wq);
1032}
1033
1034/*
1035 * cpuset_change_task_nodemask - change task's mems_allowed and mempolicy
1036 * @tsk: the task to change
1037 * @newmems: new nodes that the task will be set
1038 *
1039 * In order to avoid seeing no nodes if the old and new nodes are disjoint,
1040 * we structure updates as setting all new allowed nodes, then clearing newly
1041 * disallowed ones.
1042 */
1043static void cpuset_change_task_nodemask(struct task_struct *tsk,
1044 nodemask_t *newmems)
1045{
1046 bool need_loop;
1047
1048 task_lock(tsk);
1049 /*
1050 * Determine if a loop is necessary if another thread is doing
1051 * read_mems_allowed_begin(). If at least one node remains unchanged and
1052 * tsk does not have a mempolicy, then an empty nodemask will not be
1053 * possible when mems_allowed is larger than a word.
1054 */
1055 need_loop = task_has_mempolicy(tsk) ||
1056 !nodes_intersects(*newmems, tsk->mems_allowed);
1057
1058 if (need_loop) {
1059 local_irq_disable();
1060 write_seqcount_begin(&tsk->mems_allowed_seq);
1061 }
1062
1063 nodes_or(tsk->mems_allowed, tsk->mems_allowed, *newmems);
1064 mpol_rebind_task(tsk, newmems, MPOL_REBIND_STEP1);
1065
1066 mpol_rebind_task(tsk, newmems, MPOL_REBIND_STEP2);
1067 tsk->mems_allowed = *newmems;
1068
1069 if (need_loop) {
1070 write_seqcount_end(&tsk->mems_allowed_seq);
1071 local_irq_enable();
1072 }
1073
1074 task_unlock(tsk);
1075}
1076
1077static void *cpuset_being_rebound;
1078
1079/**
1080 * update_tasks_nodemask - Update the nodemasks of tasks in the cpuset.
1081 * @cs: the cpuset in which each task's mems_allowed mask needs to be changed
1082 *
1083 * Iterate through each task of @cs updating its mems_allowed to the
1084 * effective cpuset's. As this function is called with cpuset_mutex held,
1085 * cpuset membership stays stable.
1086 */
1087static void update_tasks_nodemask(struct cpuset *cs)
1088{
1089 static nodemask_t newmems; /* protected by cpuset_mutex */
1090 struct css_task_iter it;
1091 struct task_struct *task;
1092
1093 cpuset_being_rebound = cs; /* causes mpol_dup() rebind */
1094
1095 guarantee_online_mems(cs, &newmems);
1096
1097 /*
1098 * The mpol_rebind_mm() call takes mmap_sem, which we couldn't
1099 * take while holding tasklist_lock. Forks can happen - the
1100 * mpol_dup() cpuset_being_rebound check will catch such forks,
1101 * and rebind their vma mempolicies too. Because we still hold
1102 * the global cpuset_mutex, we know that no other rebind effort
1103 * will be contending for the global variable cpuset_being_rebound.
1104 * It's ok if we rebind the same mm twice; mpol_rebind_mm()
1105 * is idempotent. Also migrate pages in each mm to new nodes.
1106 */
1107 css_task_iter_start(&cs->css, &it);
1108 while ((task = css_task_iter_next(&it))) {
1109 struct mm_struct *mm;
1110 bool migrate;
1111
1112 cpuset_change_task_nodemask(task, &newmems);
1113
1114 mm = get_task_mm(task);
1115 if (!mm)
1116 continue;
1117
1118 migrate = is_memory_migrate(cs);
1119
1120 mpol_rebind_mm(mm, &cs->mems_allowed);
1121 if (migrate)
1122 cpuset_migrate_mm(mm, &cs->old_mems_allowed, &newmems);
1123 else
1124 mmput(mm);
1125 }
1126 css_task_iter_end(&it);
1127
1128 /*
1129 * All the tasks' nodemasks have been updated, update
1130 * cs->old_mems_allowed.
1131 */
1132 cs->old_mems_allowed = newmems;
1133
1134 /* We're done rebinding vmas to this cpuset's new mems_allowed. */
1135 cpuset_being_rebound = NULL;
1136}
1137
1138/*
1139 * update_nodemasks_hier - Update effective nodemasks and tasks in the subtree
1140 * @cs: the cpuset to consider
1141 * @new_mems: a temp variable for calculating new effective_mems
1142 *
1143 * When configured nodemask is changed, the effective nodemasks of this cpuset
1144 * and all its descendants need to be updated.
1145 *
1146 * On legacy hiearchy, effective_mems will be the same with mems_allowed.
1147 *
1148 * Called with cpuset_mutex held
1149 */
1150static void update_nodemasks_hier(struct cpuset *cs, nodemask_t *new_mems)
1151{
1152 struct cpuset *cp;
1153 struct cgroup_subsys_state *pos_css;
1154
1155 rcu_read_lock();
1156 cpuset_for_each_descendant_pre(cp, pos_css, cs) {
1157 struct cpuset *parent = parent_cs(cp);
1158
1159 nodes_and(*new_mems, cp->mems_allowed, parent->effective_mems);
1160
1161 /*
1162 * If it becomes empty, inherit the effective mask of the
1163 * parent, which is guaranteed to have some MEMs.
1164 */
1165 if (cgroup_subsys_on_dfl(cpuset_cgrp_subsys) &&
1166 nodes_empty(*new_mems))
1167 *new_mems = parent->effective_mems;
1168
1169 /* Skip the whole subtree if the nodemask remains the same. */
1170 if (nodes_equal(*new_mems, cp->effective_mems)) {
1171 pos_css = css_rightmost_descendant(pos_css);
1172 continue;
1173 }
1174
1175 if (!css_tryget_online(&cp->css))
1176 continue;
1177 rcu_read_unlock();
1178
1179 spin_lock_irq(&callback_lock);
1180 cp->effective_mems = *new_mems;
1181 spin_unlock_irq(&callback_lock);
1182
1183 WARN_ON(!cgroup_subsys_on_dfl(cpuset_cgrp_subsys) &&
1184 !nodes_equal(cp->mems_allowed, cp->effective_mems));
1185
1186 update_tasks_nodemask(cp);
1187
1188 rcu_read_lock();
1189 css_put(&cp->css);
1190 }
1191 rcu_read_unlock();
1192}
1193
1194/*
1195 * Handle user request to change the 'mems' memory placement
1196 * of a cpuset. Needs to validate the request, update the
1197 * cpusets mems_allowed, and for each task in the cpuset,
1198 * update mems_allowed and rebind task's mempolicy and any vma
1199 * mempolicies and if the cpuset is marked 'memory_migrate',
1200 * migrate the tasks pages to the new memory.
1201 *
1202 * Call with cpuset_mutex held. May take callback_lock during call.
1203 * Will take tasklist_lock, scan tasklist for tasks in cpuset cs,
1204 * lock each such tasks mm->mmap_sem, scan its vma's and rebind
1205 * their mempolicies to the cpusets new mems_allowed.
1206 */
1207static int update_nodemask(struct cpuset *cs, struct cpuset *trialcs,
1208 const char *buf)
1209{
1210 int retval;
1211
1212 /*
1213 * top_cpuset.mems_allowed tracks node_stats[N_MEMORY];
1214 * it's read-only
1215 */
1216 if (cs == &top_cpuset) {
1217 retval = -EACCES;
1218 goto done;
1219 }
1220
1221 /*
1222 * An empty mems_allowed is ok iff there are no tasks in the cpuset.
1223 * Since nodelist_parse() fails on an empty mask, we special case
1224 * that parsing. The validate_change() call ensures that cpusets
1225 * with tasks have memory.
1226 */
1227 if (!*buf) {
1228 nodes_clear(trialcs->mems_allowed);
1229 } else {
1230 retval = nodelist_parse(buf, trialcs->mems_allowed);
1231 if (retval < 0)
1232 goto done;
1233
1234 if (!nodes_subset(trialcs->mems_allowed,
1235 top_cpuset.mems_allowed)) {
1236 retval = -EINVAL;
1237 goto done;
1238 }
1239 }
1240
1241 if (nodes_equal(cs->mems_allowed, trialcs->mems_allowed)) {
1242 retval = 0; /* Too easy - nothing to do */
1243 goto done;
1244 }
1245 retval = validate_change(cs, trialcs);
1246 if (retval < 0)
1247 goto done;
1248
1249 spin_lock_irq(&callback_lock);
1250 cs->mems_allowed = trialcs->mems_allowed;
1251 spin_unlock_irq(&callback_lock);
1252
1253 /* use trialcs->mems_allowed as a temp variable */
1254 update_nodemasks_hier(cs, &trialcs->mems_allowed);
1255done:
1256 return retval;
1257}
1258
1259int current_cpuset_is_being_rebound(void)
1260{
1261 int ret;
1262
1263 rcu_read_lock();
1264 ret = task_cs(current) == cpuset_being_rebound;
1265 rcu_read_unlock();
1266
1267 return ret;
1268}
1269
1270static int update_relax_domain_level(struct cpuset *cs, s64 val)
1271{
1272#ifdef CONFIG_SMP
1273 if (val < -1 || val >= sched_domain_level_max)
1274 return -EINVAL;
1275#endif
1276
1277 if (val != cs->relax_domain_level) {
1278 cs->relax_domain_level = val;
1279 if (!cpumask_empty(cs->cpus_allowed) &&
1280 is_sched_load_balance(cs))
1281 rebuild_sched_domains_locked();
1282 }
1283
1284 return 0;
1285}
1286
1287/**
1288 * update_tasks_flags - update the spread flags of tasks in the cpuset.
1289 * @cs: the cpuset in which each task's spread flags needs to be changed
1290 *
1291 * Iterate through each task of @cs updating its spread flags. As this
1292 * function is called with cpuset_mutex held, cpuset membership stays
1293 * stable.
1294 */
1295static void update_tasks_flags(struct cpuset *cs)
1296{
1297 struct css_task_iter it;
1298 struct task_struct *task;
1299
1300 css_task_iter_start(&cs->css, &it);
1301 while ((task = css_task_iter_next(&it)))
1302 cpuset_update_task_spread_flag(cs, task);
1303 css_task_iter_end(&it);
1304}
1305
1306/*
1307 * update_flag - read a 0 or a 1 in a file and update associated flag
1308 * bit: the bit to update (see cpuset_flagbits_t)
1309 * cs: the cpuset to update
1310 * turning_on: whether the flag is being set or cleared
1311 *
1312 * Call with cpuset_mutex held.
1313 */
1314
1315static int update_flag(cpuset_flagbits_t bit, struct cpuset *cs,
1316 int turning_on)
1317{
1318 struct cpuset *trialcs;
1319 int balance_flag_changed;
1320 int spread_flag_changed;
1321 int err;
1322
1323 trialcs = alloc_trial_cpuset(cs);
1324 if (!trialcs)
1325 return -ENOMEM;
1326
1327 if (turning_on)
1328 set_bit(bit, &trialcs->flags);
1329 else
1330 clear_bit(bit, &trialcs->flags);
1331
1332 err = validate_change(cs, trialcs);
1333 if (err < 0)
1334 goto out;
1335
1336 balance_flag_changed = (is_sched_load_balance(cs) !=
1337 is_sched_load_balance(trialcs));
1338
1339 spread_flag_changed = ((is_spread_slab(cs) != is_spread_slab(trialcs))
1340 || (is_spread_page(cs) != is_spread_page(trialcs)));
1341
1342 spin_lock_irq(&callback_lock);
1343 cs->flags = trialcs->flags;
1344 spin_unlock_irq(&callback_lock);
1345
1346 if (!cpumask_empty(trialcs->cpus_allowed) && balance_flag_changed)
1347 rebuild_sched_domains_locked();
1348
1349 if (spread_flag_changed)
1350 update_tasks_flags(cs);
1351out:
1352 free_trial_cpuset(trialcs);
1353 return err;
1354}
1355
1356/*
1357 * Frequency meter - How fast is some event occurring?
1358 *
1359 * These routines manage a digitally filtered, constant time based,
1360 * event frequency meter. There are four routines:
1361 * fmeter_init() - initialize a frequency meter.
1362 * fmeter_markevent() - called each time the event happens.
1363 * fmeter_getrate() - returns the recent rate of such events.
1364 * fmeter_update() - internal routine used to update fmeter.
1365 *
1366 * A common data structure is passed to each of these routines,
1367 * which is used to keep track of the state required to manage the
1368 * frequency meter and its digital filter.
1369 *
1370 * The filter works on the number of events marked per unit time.
1371 * The filter is single-pole low-pass recursive (IIR). The time unit
1372 * is 1 second. Arithmetic is done using 32-bit integers scaled to
1373 * simulate 3 decimal digits of precision (multiplied by 1000).
1374 *
1375 * With an FM_COEF of 933, and a time base of 1 second, the filter
1376 * has a half-life of 10 seconds, meaning that if the events quit
1377 * happening, then the rate returned from the fmeter_getrate()
1378 * will be cut in half each 10 seconds, until it converges to zero.
1379 *
1380 * It is not worth doing a real infinitely recursive filter. If more
1381 * than FM_MAXTICKS ticks have elapsed since the last filter event,
1382 * just compute FM_MAXTICKS ticks worth, by which point the level
1383 * will be stable.
1384 *
1385 * Limit the count of unprocessed events to FM_MAXCNT, so as to avoid
1386 * arithmetic overflow in the fmeter_update() routine.
1387 *
1388 * Given the simple 32 bit integer arithmetic used, this meter works
1389 * best for reporting rates between one per millisecond (msec) and
1390 * one per 32 (approx) seconds. At constant rates faster than one
1391 * per msec it maxes out at values just under 1,000,000. At constant
1392 * rates between one per msec, and one per second it will stabilize
1393 * to a value N*1000, where N is the rate of events per second.
1394 * At constant rates between one per second and one per 32 seconds,
1395 * it will be choppy, moving up on the seconds that have an event,
1396 * and then decaying until the next event. At rates slower than
1397 * about one in 32 seconds, it decays all the way back to zero between
1398 * each event.
1399 */
1400
1401#define FM_COEF 933 /* coefficient for half-life of 10 secs */
1402#define FM_MAXTICKS ((u32)99) /* useless computing more ticks than this */
1403#define FM_MAXCNT 1000000 /* limit cnt to avoid overflow */
1404#define FM_SCALE 1000 /* faux fixed point scale */
1405
1406/* Initialize a frequency meter */
1407static void fmeter_init(struct fmeter *fmp)
1408{
1409 fmp->cnt = 0;
1410 fmp->val = 0;
1411 fmp->time = 0;
1412 spin_lock_init(&fmp->lock);
1413}
1414
1415/* Internal meter update - process cnt events and update value */
1416static void fmeter_update(struct fmeter *fmp)
1417{
1418 time64_t now;
1419 u32 ticks;
1420
1421 now = ktime_get_seconds();
1422 ticks = now - fmp->time;
1423
1424 if (ticks == 0)
1425 return;
1426
1427 ticks = min(FM_MAXTICKS, ticks);
1428 while (ticks-- > 0)
1429 fmp->val = (FM_COEF * fmp->val) / FM_SCALE;
1430 fmp->time = now;
1431
1432 fmp->val += ((FM_SCALE - FM_COEF) * fmp->cnt) / FM_SCALE;
1433 fmp->cnt = 0;
1434}
1435
1436/* Process any previous ticks, then bump cnt by one (times scale). */
1437static void fmeter_markevent(struct fmeter *fmp)
1438{
1439 spin_lock(&fmp->lock);
1440 fmeter_update(fmp);
1441 fmp->cnt = min(FM_MAXCNT, fmp->cnt + FM_SCALE);
1442 spin_unlock(&fmp->lock);
1443}
1444
1445/* Process any previous ticks, then return current value. */
1446static int fmeter_getrate(struct fmeter *fmp)
1447{
1448 int val;
1449
1450 spin_lock(&fmp->lock);
1451 fmeter_update(fmp);
1452 val = fmp->val;
1453 spin_unlock(&fmp->lock);
1454 return val;
1455}
1456
1457static struct cpuset *cpuset_attach_old_cs;
1458
1459/* Called by cgroups to determine if a cpuset is usable; cpuset_mutex held */
1460static int cpuset_can_attach(struct cgroup_taskset *tset)
1461{
1462 struct cgroup_subsys_state *css;
1463 struct cpuset *cs;
1464 struct task_struct *task;
1465 int ret;
1466
1467 /* used later by cpuset_attach() */
1468 cpuset_attach_old_cs = task_cs(cgroup_taskset_first(tset, &css));
1469 cs = css_cs(css);
1470
1471 mutex_lock(&cpuset_mutex);
1472
1473 /* allow moving tasks into an empty cpuset if on default hierarchy */
1474 ret = -ENOSPC;
1475 if (!cgroup_subsys_on_dfl(cpuset_cgrp_subsys) &&
1476 (cpumask_empty(cs->cpus_allowed) || nodes_empty(cs->mems_allowed)))
1477 goto out_unlock;
1478
1479 cgroup_taskset_for_each(task, css, tset) {
1480 ret = task_can_attach(task, cs->cpus_allowed);
1481 if (ret)
1482 goto out_unlock;
1483 ret = security_task_setscheduler(task);
1484 if (ret)
1485 goto out_unlock;
1486 }
1487
1488 /*
1489 * Mark attach is in progress. This makes validate_change() fail
1490 * changes which zero cpus/mems_allowed.
1491 */
1492 cs->attach_in_progress++;
1493 ret = 0;
1494out_unlock:
1495 mutex_unlock(&cpuset_mutex);
1496 return ret;
1497}
1498
1499static void cpuset_cancel_attach(struct cgroup_taskset *tset)
1500{
1501 struct cgroup_subsys_state *css;
1502 struct cpuset *cs;
1503
1504 cgroup_taskset_first(tset, &css);
1505 cs = css_cs(css);
1506
1507 mutex_lock(&cpuset_mutex);
1508 css_cs(css)->attach_in_progress--;
1509 mutex_unlock(&cpuset_mutex);
1510}
1511
1512/*
1513 * Protected by cpuset_mutex. cpus_attach is used only by cpuset_attach()
1514 * but we can't allocate it dynamically there. Define it global and
1515 * allocate from cpuset_init().
1516 */
1517static cpumask_var_t cpus_attach;
1518
1519static void cpuset_attach(struct cgroup_taskset *tset)
1520{
1521 /* static buf protected by cpuset_mutex */
1522 static nodemask_t cpuset_attach_nodemask_to;
1523 struct task_struct *task;
1524 struct task_struct *leader;
1525 struct cgroup_subsys_state *css;
1526 struct cpuset *cs;
1527 struct cpuset *oldcs = cpuset_attach_old_cs;
1528
1529 cgroup_taskset_first(tset, &css);
1530 cs = css_cs(css);
1531
1532 mutex_lock(&cpuset_mutex);
1533
1534 /* prepare for attach */
1535 if (cs == &top_cpuset)
1536 cpumask_copy(cpus_attach, cpu_possible_mask);
1537 else
1538 guarantee_online_cpus(cs, cpus_attach);
1539
1540 guarantee_online_mems(cs, &cpuset_attach_nodemask_to);
1541
1542 cgroup_taskset_for_each(task, css, tset) {
1543 /*
1544 * can_attach beforehand should guarantee that this doesn't
1545 * fail. TODO: have a better way to handle failure here
1546 */
1547 WARN_ON_ONCE(set_cpus_allowed_ptr(task, cpus_attach));
1548
1549 cpuset_change_task_nodemask(task, &cpuset_attach_nodemask_to);
1550 cpuset_update_task_spread_flag(cs, task);
1551 }
1552
1553 /*
1554 * Change mm for all threadgroup leaders. This is expensive and may
1555 * sleep and should be moved outside migration path proper.
1556 */
1557 cpuset_attach_nodemask_to = cs->effective_mems;
1558 cgroup_taskset_for_each_leader(leader, css, tset) {
1559 struct mm_struct *mm = get_task_mm(leader);
1560
1561 if (mm) {
1562 mpol_rebind_mm(mm, &cpuset_attach_nodemask_to);
1563
1564 /*
1565 * old_mems_allowed is the same with mems_allowed
1566 * here, except if this task is being moved
1567 * automatically due to hotplug. In that case
1568 * @mems_allowed has been updated and is empty, so
1569 * @old_mems_allowed is the right nodesets that we
1570 * migrate mm from.
1571 */
1572 if (is_memory_migrate(cs))
1573 cpuset_migrate_mm(mm, &oldcs->old_mems_allowed,
1574 &cpuset_attach_nodemask_to);
1575 else
1576 mmput(mm);
1577 }
1578 }
1579
1580 cs->old_mems_allowed = cpuset_attach_nodemask_to;
1581
1582 cs->attach_in_progress--;
1583 if (!cs->attach_in_progress)
1584 wake_up(&cpuset_attach_wq);
1585
1586 mutex_unlock(&cpuset_mutex);
1587}
1588
1589/* The various types of files and directories in a cpuset file system */
1590
1591typedef enum {
1592 FILE_MEMORY_MIGRATE,
1593 FILE_CPULIST,
1594 FILE_MEMLIST,
1595 FILE_EFFECTIVE_CPULIST,
1596 FILE_EFFECTIVE_MEMLIST,
1597 FILE_CPU_EXCLUSIVE,
1598 FILE_MEM_EXCLUSIVE,
1599 FILE_MEM_HARDWALL,
1600 FILE_SCHED_LOAD_BALANCE,
1601 FILE_SCHED_RELAX_DOMAIN_LEVEL,
1602 FILE_MEMORY_PRESSURE_ENABLED,
1603 FILE_MEMORY_PRESSURE,
1604 FILE_SPREAD_PAGE,
1605 FILE_SPREAD_SLAB,
1606} cpuset_filetype_t;
1607
1608static int cpuset_write_u64(struct cgroup_subsys_state *css, struct cftype *cft,
1609 u64 val)
1610{
1611 struct cpuset *cs = css_cs(css);
1612 cpuset_filetype_t type = cft->private;
1613 int retval = 0;
1614
1615 mutex_lock(&cpuset_mutex);
1616 if (!is_cpuset_online(cs)) {
1617 retval = -ENODEV;
1618 goto out_unlock;
1619 }
1620
1621 switch (type) {
1622 case FILE_CPU_EXCLUSIVE:
1623 retval = update_flag(CS_CPU_EXCLUSIVE, cs, val);
1624 break;
1625 case FILE_MEM_EXCLUSIVE:
1626 retval = update_flag(CS_MEM_EXCLUSIVE, cs, val);
1627 break;
1628 case FILE_MEM_HARDWALL:
1629 retval = update_flag(CS_MEM_HARDWALL, cs, val);
1630 break;
1631 case FILE_SCHED_LOAD_BALANCE:
1632 retval = update_flag(CS_SCHED_LOAD_BALANCE, cs, val);
1633 break;
1634 case FILE_MEMORY_MIGRATE:
1635 retval = update_flag(CS_MEMORY_MIGRATE, cs, val);
1636 break;
1637 case FILE_MEMORY_PRESSURE_ENABLED:
1638 cpuset_memory_pressure_enabled = !!val;
1639 break;
1640 case FILE_SPREAD_PAGE:
1641 retval = update_flag(CS_SPREAD_PAGE, cs, val);
1642 break;
1643 case FILE_SPREAD_SLAB:
1644 retval = update_flag(CS_SPREAD_SLAB, cs, val);
1645 break;
1646 default:
1647 retval = -EINVAL;
1648 break;
1649 }
1650out_unlock:
1651 mutex_unlock(&cpuset_mutex);
1652 return retval;
1653}
1654
1655static int cpuset_write_s64(struct cgroup_subsys_state *css, struct cftype *cft,
1656 s64 val)
1657{
1658 struct cpuset *cs = css_cs(css);
1659 cpuset_filetype_t type = cft->private;
1660 int retval = -ENODEV;
1661
1662 mutex_lock(&cpuset_mutex);
1663 if (!is_cpuset_online(cs))
1664 goto out_unlock;
1665
1666 switch (type) {
1667 case FILE_SCHED_RELAX_DOMAIN_LEVEL:
1668 retval = update_relax_domain_level(cs, val);
1669 break;
1670 default:
1671 retval = -EINVAL;
1672 break;
1673 }
1674out_unlock:
1675 mutex_unlock(&cpuset_mutex);
1676 return retval;
1677}
1678
1679/*
1680 * Common handling for a write to a "cpus" or "mems" file.
1681 */
1682static ssize_t cpuset_write_resmask(struct kernfs_open_file *of,
1683 char *buf, size_t nbytes, loff_t off)
1684{
1685 struct cpuset *cs = css_cs(of_css(of));
1686 struct cpuset *trialcs;
1687 int retval = -ENODEV;
1688
1689 buf = strstrip(buf);
1690
1691 /*
1692 * CPU or memory hotunplug may leave @cs w/o any execution
1693 * resources, in which case the hotplug code asynchronously updates
1694 * configuration and transfers all tasks to the nearest ancestor
1695 * which can execute.
1696 *
1697 * As writes to "cpus" or "mems" may restore @cs's execution
1698 * resources, wait for the previously scheduled operations before
1699 * proceeding, so that we don't end up keep removing tasks added
1700 * after execution capability is restored.
1701 *
1702 * cpuset_hotplug_work calls back into cgroup core via
1703 * cgroup_transfer_tasks() and waiting for it from a cgroupfs
1704 * operation like this one can lead to a deadlock through kernfs
1705 * active_ref protection. Let's break the protection. Losing the
1706 * protection is okay as we check whether @cs is online after
1707 * grabbing cpuset_mutex anyway. This only happens on the legacy
1708 * hierarchies.
1709 */
1710 css_get(&cs->css);
1711 kernfs_break_active_protection(of->kn);
1712 flush_work(&cpuset_hotplug_work);
1713
1714 mutex_lock(&cpuset_mutex);
1715 if (!is_cpuset_online(cs))
1716 goto out_unlock;
1717
1718 trialcs = alloc_trial_cpuset(cs);
1719 if (!trialcs) {
1720 retval = -ENOMEM;
1721 goto out_unlock;
1722 }
1723
1724 switch (of_cft(of)->private) {
1725 case FILE_CPULIST:
1726 retval = update_cpumask(cs, trialcs, buf);
1727 break;
1728 case FILE_MEMLIST:
1729 retval = update_nodemask(cs, trialcs, buf);
1730 break;
1731 default:
1732 retval = -EINVAL;
1733 break;
1734 }
1735
1736 free_trial_cpuset(trialcs);
1737out_unlock:
1738 mutex_unlock(&cpuset_mutex);
1739 kernfs_unbreak_active_protection(of->kn);
1740 css_put(&cs->css);
1741 flush_workqueue(cpuset_migrate_mm_wq);
1742 return retval ?: nbytes;
1743}
1744
1745/*
1746 * These ascii lists should be read in a single call, by using a user
1747 * buffer large enough to hold the entire map. If read in smaller
1748 * chunks, there is no guarantee of atomicity. Since the display format
1749 * used, list of ranges of sequential numbers, is variable length,
1750 * and since these maps can change value dynamically, one could read
1751 * gibberish by doing partial reads while a list was changing.
1752 */
1753static int cpuset_common_seq_show(struct seq_file *sf, void *v)
1754{
1755 struct cpuset *cs = css_cs(seq_css(sf));
1756 cpuset_filetype_t type = seq_cft(sf)->private;
1757 int ret = 0;
1758
1759 spin_lock_irq(&callback_lock);
1760
1761 switch (type) {
1762 case FILE_CPULIST:
1763 seq_printf(sf, "%*pbl\n", cpumask_pr_args(cs->cpus_allowed));
1764 break;
1765 case FILE_MEMLIST:
1766 seq_printf(sf, "%*pbl\n", nodemask_pr_args(&cs->mems_allowed));
1767 break;
1768 case FILE_EFFECTIVE_CPULIST:
1769 seq_printf(sf, "%*pbl\n", cpumask_pr_args(cs->effective_cpus));
1770 break;
1771 case FILE_EFFECTIVE_MEMLIST:
1772 seq_printf(sf, "%*pbl\n", nodemask_pr_args(&cs->effective_mems));
1773 break;
1774 default:
1775 ret = -EINVAL;
1776 }
1777
1778 spin_unlock_irq(&callback_lock);
1779 return ret;
1780}
1781
1782static u64 cpuset_read_u64(struct cgroup_subsys_state *css, struct cftype *cft)
1783{
1784 struct cpuset *cs = css_cs(css);
1785 cpuset_filetype_t type = cft->private;
1786 switch (type) {
1787 case FILE_CPU_EXCLUSIVE:
1788 return is_cpu_exclusive(cs);
1789 case FILE_MEM_EXCLUSIVE:
1790 return is_mem_exclusive(cs);
1791 case FILE_MEM_HARDWALL:
1792 return is_mem_hardwall(cs);
1793 case FILE_SCHED_LOAD_BALANCE:
1794 return is_sched_load_balance(cs);
1795 case FILE_MEMORY_MIGRATE:
1796 return is_memory_migrate(cs);
1797 case FILE_MEMORY_PRESSURE_ENABLED:
1798 return cpuset_memory_pressure_enabled;
1799 case FILE_MEMORY_PRESSURE:
1800 return fmeter_getrate(&cs->fmeter);
1801 case FILE_SPREAD_PAGE:
1802 return is_spread_page(cs);
1803 case FILE_SPREAD_SLAB:
1804 return is_spread_slab(cs);
1805 default:
1806 BUG();
1807 }
1808
1809 /* Unreachable but makes gcc happy */
1810 return 0;
1811}
1812
1813static s64 cpuset_read_s64(struct cgroup_subsys_state *css, struct cftype *cft)
1814{
1815 struct cpuset *cs = css_cs(css);
1816 cpuset_filetype_t type = cft->private;
1817 switch (type) {
1818 case FILE_SCHED_RELAX_DOMAIN_LEVEL:
1819 return cs->relax_domain_level;
1820 default:
1821 BUG();
1822 }
1823
1824 /* Unrechable but makes gcc happy */
1825 return 0;
1826}
1827
1828
1829/*
1830 * for the common functions, 'private' gives the type of file
1831 */
1832
1833static struct cftype files[] = {
1834 {
1835 .name = "cpus",
1836 .seq_show = cpuset_common_seq_show,
1837 .write = cpuset_write_resmask,
1838 .max_write_len = (100U + 6 * NR_CPUS),
1839 .private = FILE_CPULIST,
1840 },
1841
1842 {
1843 .name = "mems",
1844 .seq_show = cpuset_common_seq_show,
1845 .write = cpuset_write_resmask,
1846 .max_write_len = (100U + 6 * MAX_NUMNODES),
1847 .private = FILE_MEMLIST,
1848 },
1849
1850 {
1851 .name = "effective_cpus",
1852 .seq_show = cpuset_common_seq_show,
1853 .private = FILE_EFFECTIVE_CPULIST,
1854 },
1855
1856 {
1857 .name = "effective_mems",
1858 .seq_show = cpuset_common_seq_show,
1859 .private = FILE_EFFECTIVE_MEMLIST,
1860 },
1861
1862 {
1863 .name = "cpu_exclusive",
1864 .read_u64 = cpuset_read_u64,
1865 .write_u64 = cpuset_write_u64,
1866 .private = FILE_CPU_EXCLUSIVE,
1867 },
1868
1869 {
1870 .name = "mem_exclusive",
1871 .read_u64 = cpuset_read_u64,
1872 .write_u64 = cpuset_write_u64,
1873 .private = FILE_MEM_EXCLUSIVE,
1874 },
1875
1876 {
1877 .name = "mem_hardwall",
1878 .read_u64 = cpuset_read_u64,
1879 .write_u64 = cpuset_write_u64,
1880 .private = FILE_MEM_HARDWALL,
1881 },
1882
1883 {
1884 .name = "sched_load_balance",
1885 .read_u64 = cpuset_read_u64,
1886 .write_u64 = cpuset_write_u64,
1887 .private = FILE_SCHED_LOAD_BALANCE,
1888 },
1889
1890 {
1891 .name = "sched_relax_domain_level",
1892 .read_s64 = cpuset_read_s64,
1893 .write_s64 = cpuset_write_s64,
1894 .private = FILE_SCHED_RELAX_DOMAIN_LEVEL,
1895 },
1896
1897 {
1898 .name = "memory_migrate",
1899 .read_u64 = cpuset_read_u64,
1900 .write_u64 = cpuset_write_u64,
1901 .private = FILE_MEMORY_MIGRATE,
1902 },
1903
1904 {
1905 .name = "memory_pressure",
1906 .read_u64 = cpuset_read_u64,
1907 },
1908
1909 {
1910 .name = "memory_spread_page",
1911 .read_u64 = cpuset_read_u64,
1912 .write_u64 = cpuset_write_u64,
1913 .private = FILE_SPREAD_PAGE,
1914 },
1915
1916 {
1917 .name = "memory_spread_slab",
1918 .read_u64 = cpuset_read_u64,
1919 .write_u64 = cpuset_write_u64,
1920 .private = FILE_SPREAD_SLAB,
1921 },
1922
1923 {
1924 .name = "memory_pressure_enabled",
1925 .flags = CFTYPE_ONLY_ON_ROOT,
1926 .read_u64 = cpuset_read_u64,
1927 .write_u64 = cpuset_write_u64,
1928 .private = FILE_MEMORY_PRESSURE_ENABLED,
1929 },
1930
1931 { } /* terminate */
1932};
1933
1934/*
1935 * cpuset_css_alloc - allocate a cpuset css
1936 * cgrp: control group that the new cpuset will be part of
1937 */
1938
1939static struct cgroup_subsys_state *
1940cpuset_css_alloc(struct cgroup_subsys_state *parent_css)
1941{
1942 struct cpuset *cs;
1943
1944 if (!parent_css)
1945 return &top_cpuset.css;
1946
1947 cs = kzalloc(sizeof(*cs), GFP_KERNEL);
1948 if (!cs)
1949 return ERR_PTR(-ENOMEM);
1950 if (!alloc_cpumask_var(&cs->cpus_allowed, GFP_KERNEL))
1951 goto free_cs;
1952 if (!alloc_cpumask_var(&cs->effective_cpus, GFP_KERNEL))
1953 goto free_cpus;
1954
1955 set_bit(CS_SCHED_LOAD_BALANCE, &cs->flags);
1956 cpumask_clear(cs->cpus_allowed);
1957 nodes_clear(cs->mems_allowed);
1958 cpumask_clear(cs->effective_cpus);
1959 nodes_clear(cs->effective_mems);
1960 fmeter_init(&cs->fmeter);
1961 cs->relax_domain_level = -1;
1962
1963 return &cs->css;
1964
1965free_cpus:
1966 free_cpumask_var(cs->cpus_allowed);
1967free_cs:
1968 kfree(cs);
1969 return ERR_PTR(-ENOMEM);
1970}
1971
1972static int cpuset_css_online(struct cgroup_subsys_state *css)
1973{
1974 struct cpuset *cs = css_cs(css);
1975 struct cpuset *parent = parent_cs(cs);
1976 struct cpuset *tmp_cs;
1977 struct cgroup_subsys_state *pos_css;
1978
1979 if (!parent)
1980 return 0;
1981
1982 mutex_lock(&cpuset_mutex);
1983
1984 set_bit(CS_ONLINE, &cs->flags);
1985 if (is_spread_page(parent))
1986 set_bit(CS_SPREAD_PAGE, &cs->flags);
1987 if (is_spread_slab(parent))
1988 set_bit(CS_SPREAD_SLAB, &cs->flags);
1989
1990 cpuset_inc();
1991
1992 spin_lock_irq(&callback_lock);
1993 if (cgroup_subsys_on_dfl(cpuset_cgrp_subsys)) {
1994 cpumask_copy(cs->effective_cpus, parent->effective_cpus);
1995 cs->effective_mems = parent->effective_mems;
1996 }
1997 spin_unlock_irq(&callback_lock);
1998
1999 if (!test_bit(CGRP_CPUSET_CLONE_CHILDREN, &css->cgroup->flags))
2000 goto out_unlock;
2001
2002 /*
2003 * Clone @parent's configuration if CGRP_CPUSET_CLONE_CHILDREN is
2004 * set. This flag handling is implemented in cgroup core for
2005 * histrical reasons - the flag may be specified during mount.
2006 *
2007 * Currently, if any sibling cpusets have exclusive cpus or mem, we
2008 * refuse to clone the configuration - thereby refusing the task to
2009 * be entered, and as a result refusing the sys_unshare() or
2010 * clone() which initiated it. If this becomes a problem for some
2011 * users who wish to allow that scenario, then this could be
2012 * changed to grant parent->cpus_allowed-sibling_cpus_exclusive
2013 * (and likewise for mems) to the new cgroup.
2014 */
2015 rcu_read_lock();
2016 cpuset_for_each_child(tmp_cs, pos_css, parent) {
2017 if (is_mem_exclusive(tmp_cs) || is_cpu_exclusive(tmp_cs)) {
2018 rcu_read_unlock();
2019 goto out_unlock;
2020 }
2021 }
2022 rcu_read_unlock();
2023
2024 spin_lock_irq(&callback_lock);
2025 cs->mems_allowed = parent->mems_allowed;
2026 cs->effective_mems = parent->mems_allowed;
2027 cpumask_copy(cs->cpus_allowed, parent->cpus_allowed);
2028 cpumask_copy(cs->effective_cpus, parent->cpus_allowed);
2029 spin_unlock_irq(&callback_lock);
2030out_unlock:
2031 mutex_unlock(&cpuset_mutex);
2032 return 0;
2033}
2034
2035/*
2036 * If the cpuset being removed has its flag 'sched_load_balance'
2037 * enabled, then simulate turning sched_load_balance off, which
2038 * will call rebuild_sched_domains_locked().
2039 */
2040
2041static void cpuset_css_offline(struct cgroup_subsys_state *css)
2042{
2043 struct cpuset *cs = css_cs(css);
2044
2045 mutex_lock(&cpuset_mutex);
2046
2047 if (is_sched_load_balance(cs))
2048 update_flag(CS_SCHED_LOAD_BALANCE, cs, 0);
2049
2050 cpuset_dec();
2051 clear_bit(CS_ONLINE, &cs->flags);
2052
2053 mutex_unlock(&cpuset_mutex);
2054}
2055
2056static void cpuset_css_free(struct cgroup_subsys_state *css)
2057{
2058 struct cpuset *cs = css_cs(css);
2059
2060 free_cpumask_var(cs->effective_cpus);
2061 free_cpumask_var(cs->cpus_allowed);
2062 kfree(cs);
2063}
2064
2065static void cpuset_bind(struct cgroup_subsys_state *root_css)
2066{
2067 mutex_lock(&cpuset_mutex);
2068 spin_lock_irq(&callback_lock);
2069
2070 if (cgroup_subsys_on_dfl(cpuset_cgrp_subsys)) {
2071 cpumask_copy(top_cpuset.cpus_allowed, cpu_possible_mask);
2072 top_cpuset.mems_allowed = node_possible_map;
2073 } else {
2074 cpumask_copy(top_cpuset.cpus_allowed,
2075 top_cpuset.effective_cpus);
2076 top_cpuset.mems_allowed = top_cpuset.effective_mems;
2077 }
2078
2079 spin_unlock_irq(&callback_lock);
2080 mutex_unlock(&cpuset_mutex);
2081}
2082
2083/*
2084 * Make sure the new task conform to the current state of its parent,
2085 * which could have been changed by cpuset just after it inherits the
2086 * state from the parent and before it sits on the cgroup's task list.
2087 */
2088static void cpuset_fork(struct task_struct *task)
2089{
2090 if (task_css_is_root(task, cpuset_cgrp_id))
2091 return;
2092
2093 set_cpus_allowed_ptr(task, ¤t->cpus_allowed);
2094 task->mems_allowed = current->mems_allowed;
2095}
2096
2097struct cgroup_subsys cpuset_cgrp_subsys = {
2098 .css_alloc = cpuset_css_alloc,
2099 .css_online = cpuset_css_online,
2100 .css_offline = cpuset_css_offline,
2101 .css_free = cpuset_css_free,
2102 .can_attach = cpuset_can_attach,
2103 .cancel_attach = cpuset_cancel_attach,
2104 .attach = cpuset_attach,
2105 .post_attach = cpuset_post_attach,
2106 .bind = cpuset_bind,
2107 .fork = cpuset_fork,
2108 .legacy_cftypes = files,
2109 .early_init = true,
2110};
2111
2112/**
2113 * cpuset_init - initialize cpusets at system boot
2114 *
2115 * Description: Initialize top_cpuset and the cpuset internal file system,
2116 **/
2117
2118int __init cpuset_init(void)
2119{
2120 int err = 0;
2121
2122 if (!alloc_cpumask_var(&top_cpuset.cpus_allowed, GFP_KERNEL))
2123 BUG();
2124 if (!alloc_cpumask_var(&top_cpuset.effective_cpus, GFP_KERNEL))
2125 BUG();
2126
2127 cpumask_setall(top_cpuset.cpus_allowed);
2128 nodes_setall(top_cpuset.mems_allowed);
2129 cpumask_setall(top_cpuset.effective_cpus);
2130 nodes_setall(top_cpuset.effective_mems);
2131
2132 fmeter_init(&top_cpuset.fmeter);
2133 set_bit(CS_SCHED_LOAD_BALANCE, &top_cpuset.flags);
2134 top_cpuset.relax_domain_level = -1;
2135
2136 err = register_filesystem(&cpuset_fs_type);
2137 if (err < 0)
2138 return err;
2139
2140 if (!alloc_cpumask_var(&cpus_attach, GFP_KERNEL))
2141 BUG();
2142
2143 return 0;
2144}
2145
2146/*
2147 * If CPU and/or memory hotplug handlers, below, unplug any CPUs
2148 * or memory nodes, we need to walk over the cpuset hierarchy,
2149 * removing that CPU or node from all cpusets. If this removes the
2150 * last CPU or node from a cpuset, then move the tasks in the empty
2151 * cpuset to its next-highest non-empty parent.
2152 */
2153static void remove_tasks_in_empty_cpuset(struct cpuset *cs)
2154{
2155 struct cpuset *parent;
2156
2157 /*
2158 * Find its next-highest non-empty parent, (top cpuset
2159 * has online cpus, so can't be empty).
2160 */
2161 parent = parent_cs(cs);
2162 while (cpumask_empty(parent->cpus_allowed) ||
2163 nodes_empty(parent->mems_allowed))
2164 parent = parent_cs(parent);
2165
2166 if (cgroup_transfer_tasks(parent->css.cgroup, cs->css.cgroup)) {
2167 pr_err("cpuset: failed to transfer tasks out of empty cpuset ");
2168 pr_cont_cgroup_name(cs->css.cgroup);
2169 pr_cont("\n");
2170 }
2171}
2172
2173static void
2174hotplug_update_tasks_legacy(struct cpuset *cs,
2175 struct cpumask *new_cpus, nodemask_t *new_mems,
2176 bool cpus_updated, bool mems_updated)
2177{
2178 bool is_empty;
2179
2180 spin_lock_irq(&callback_lock);
2181 cpumask_copy(cs->cpus_allowed, new_cpus);
2182 cpumask_copy(cs->effective_cpus, new_cpus);
2183 cs->mems_allowed = *new_mems;
2184 cs->effective_mems = *new_mems;
2185 spin_unlock_irq(&callback_lock);
2186
2187 /*
2188 * Don't call update_tasks_cpumask() if the cpuset becomes empty,
2189 * as the tasks will be migratecd to an ancestor.
2190 */
2191 if (cpus_updated && !cpumask_empty(cs->cpus_allowed))
2192 update_tasks_cpumask(cs);
2193 if (mems_updated && !nodes_empty(cs->mems_allowed))
2194 update_tasks_nodemask(cs);
2195
2196 is_empty = cpumask_empty(cs->cpus_allowed) ||
2197 nodes_empty(cs->mems_allowed);
2198
2199 mutex_unlock(&cpuset_mutex);
2200
2201 /*
2202 * Move tasks to the nearest ancestor with execution resources,
2203 * This is full cgroup operation which will also call back into
2204 * cpuset. Should be done outside any lock.
2205 */
2206 if (is_empty)
2207 remove_tasks_in_empty_cpuset(cs);
2208
2209 mutex_lock(&cpuset_mutex);
2210}
2211
2212static void
2213hotplug_update_tasks(struct cpuset *cs,
2214 struct cpumask *new_cpus, nodemask_t *new_mems,
2215 bool cpus_updated, bool mems_updated)
2216{
2217 if (cpumask_empty(new_cpus))
2218 cpumask_copy(new_cpus, parent_cs(cs)->effective_cpus);
2219 if (nodes_empty(*new_mems))
2220 *new_mems = parent_cs(cs)->effective_mems;
2221
2222 spin_lock_irq(&callback_lock);
2223 cpumask_copy(cs->effective_cpus, new_cpus);
2224 cs->effective_mems = *new_mems;
2225 spin_unlock_irq(&callback_lock);
2226
2227 if (cpus_updated)
2228 update_tasks_cpumask(cs);
2229 if (mems_updated)
2230 update_tasks_nodemask(cs);
2231}
2232
2233/**
2234 * cpuset_hotplug_update_tasks - update tasks in a cpuset for hotunplug
2235 * @cs: cpuset in interest
2236 *
2237 * Compare @cs's cpu and mem masks against top_cpuset and if some have gone
2238 * offline, update @cs accordingly. If @cs ends up with no CPU or memory,
2239 * all its tasks are moved to the nearest ancestor with both resources.
2240 */
2241static void cpuset_hotplug_update_tasks(struct cpuset *cs)
2242{
2243 static cpumask_t new_cpus;
2244 static nodemask_t new_mems;
2245 bool cpus_updated;
2246 bool mems_updated;
2247retry:
2248 wait_event(cpuset_attach_wq, cs->attach_in_progress == 0);
2249
2250 mutex_lock(&cpuset_mutex);
2251
2252 /*
2253 * We have raced with task attaching. We wait until attaching
2254 * is finished, so we won't attach a task to an empty cpuset.
2255 */
2256 if (cs->attach_in_progress) {
2257 mutex_unlock(&cpuset_mutex);
2258 goto retry;
2259 }
2260
2261 cpumask_and(&new_cpus, cs->cpus_allowed, parent_cs(cs)->effective_cpus);
2262 nodes_and(new_mems, cs->mems_allowed, parent_cs(cs)->effective_mems);
2263
2264 cpus_updated = !cpumask_equal(&new_cpus, cs->effective_cpus);
2265 mems_updated = !nodes_equal(new_mems, cs->effective_mems);
2266
2267 if (cgroup_subsys_on_dfl(cpuset_cgrp_subsys))
2268 hotplug_update_tasks(cs, &new_cpus, &new_mems,
2269 cpus_updated, mems_updated);
2270 else
2271 hotplug_update_tasks_legacy(cs, &new_cpus, &new_mems,
2272 cpus_updated, mems_updated);
2273
2274 mutex_unlock(&cpuset_mutex);
2275}
2276
2277/**
2278 * cpuset_hotplug_workfn - handle CPU/memory hotunplug for a cpuset
2279 *
2280 * This function is called after either CPU or memory configuration has
2281 * changed and updates cpuset accordingly. The top_cpuset is always
2282 * synchronized to cpu_active_mask and N_MEMORY, which is necessary in
2283 * order to make cpusets transparent (of no affect) on systems that are
2284 * actively using CPU hotplug but making no active use of cpusets.
2285 *
2286 * Non-root cpusets are only affected by offlining. If any CPUs or memory
2287 * nodes have been taken down, cpuset_hotplug_update_tasks() is invoked on
2288 * all descendants.
2289 *
2290 * Note that CPU offlining during suspend is ignored. We don't modify
2291 * cpusets across suspend/resume cycles at all.
2292 */
2293static void cpuset_hotplug_workfn(struct work_struct *work)
2294{
2295 static cpumask_t new_cpus;
2296 static nodemask_t new_mems;
2297 bool cpus_updated, mems_updated;
2298 bool on_dfl = cgroup_subsys_on_dfl(cpuset_cgrp_subsys);
2299
2300 mutex_lock(&cpuset_mutex);
2301
2302 /* fetch the available cpus/mems and find out which changed how */
2303 cpumask_copy(&new_cpus, cpu_active_mask);
2304 new_mems = node_states[N_MEMORY];
2305
2306 cpus_updated = !cpumask_equal(top_cpuset.effective_cpus, &new_cpus);
2307 mems_updated = !nodes_equal(top_cpuset.effective_mems, new_mems);
2308
2309 /* synchronize cpus_allowed to cpu_active_mask */
2310 if (cpus_updated) {
2311 spin_lock_irq(&callback_lock);
2312 if (!on_dfl)
2313 cpumask_copy(top_cpuset.cpus_allowed, &new_cpus);
2314 cpumask_copy(top_cpuset.effective_cpus, &new_cpus);
2315 spin_unlock_irq(&callback_lock);
2316 /* we don't mess with cpumasks of tasks in top_cpuset */
2317 }
2318
2319 /* synchronize mems_allowed to N_MEMORY */
2320 if (mems_updated) {
2321 spin_lock_irq(&callback_lock);
2322 if (!on_dfl)
2323 top_cpuset.mems_allowed = new_mems;
2324 top_cpuset.effective_mems = new_mems;
2325 spin_unlock_irq(&callback_lock);
2326 update_tasks_nodemask(&top_cpuset);
2327 }
2328
2329 mutex_unlock(&cpuset_mutex);
2330
2331 /* if cpus or mems changed, we need to propagate to descendants */
2332 if (cpus_updated || mems_updated) {
2333 struct cpuset *cs;
2334 struct cgroup_subsys_state *pos_css;
2335
2336 rcu_read_lock();
2337 cpuset_for_each_descendant_pre(cs, pos_css, &top_cpuset) {
2338 if (cs == &top_cpuset || !css_tryget_online(&cs->css))
2339 continue;
2340 rcu_read_unlock();
2341
2342 cpuset_hotplug_update_tasks(cs);
2343
2344 rcu_read_lock();
2345 css_put(&cs->css);
2346 }
2347 rcu_read_unlock();
2348 }
2349
2350 /* rebuild sched domains if cpus_allowed has changed */
2351 if (cpus_updated)
2352 rebuild_sched_domains();
2353}
2354
2355void cpuset_update_active_cpus(bool cpu_online)
2356{
2357 /*
2358 * We're inside cpu hotplug critical region which usually nests
2359 * inside cgroup synchronization. Bounce actual hotplug processing
2360 * to a work item to avoid reverse locking order.
2361 *
2362 * We still need to do partition_sched_domains() synchronously;
2363 * otherwise, the scheduler will get confused and put tasks to the
2364 * dead CPU. Fall back to the default single domain.
2365 * cpuset_hotplug_workfn() will rebuild it as necessary.
2366 */
2367 partition_sched_domains(1, NULL, NULL);
2368 schedule_work(&cpuset_hotplug_work);
2369}
2370
2371/*
2372 * Keep top_cpuset.mems_allowed tracking node_states[N_MEMORY].
2373 * Call this routine anytime after node_states[N_MEMORY] changes.
2374 * See cpuset_update_active_cpus() for CPU hotplug handling.
2375 */
2376static int cpuset_track_online_nodes(struct notifier_block *self,
2377 unsigned long action, void *arg)
2378{
2379 schedule_work(&cpuset_hotplug_work);
2380 return NOTIFY_OK;
2381}
2382
2383static struct notifier_block cpuset_track_online_nodes_nb = {
2384 .notifier_call = cpuset_track_online_nodes,
2385 .priority = 10, /* ??! */
2386};
2387
2388/**
2389 * cpuset_init_smp - initialize cpus_allowed
2390 *
2391 * Description: Finish top cpuset after cpu, node maps are initialized
2392 */
2393void __init cpuset_init_smp(void)
2394{
2395 cpumask_copy(top_cpuset.cpus_allowed, cpu_active_mask);
2396 top_cpuset.mems_allowed = node_states[N_MEMORY];
2397 top_cpuset.old_mems_allowed = top_cpuset.mems_allowed;
2398
2399 cpumask_copy(top_cpuset.effective_cpus, cpu_active_mask);
2400 top_cpuset.effective_mems = node_states[N_MEMORY];
2401
2402 register_hotmemory_notifier(&cpuset_track_online_nodes_nb);
2403
2404 cpuset_migrate_mm_wq = alloc_ordered_workqueue("cpuset_migrate_mm", 0);
2405 BUG_ON(!cpuset_migrate_mm_wq);
2406}
2407
2408/**
2409 * cpuset_cpus_allowed - return cpus_allowed mask from a tasks cpuset.
2410 * @tsk: pointer to task_struct from which to obtain cpuset->cpus_allowed.
2411 * @pmask: pointer to struct cpumask variable to receive cpus_allowed set.
2412 *
2413 * Description: Returns the cpumask_var_t cpus_allowed of the cpuset
2414 * attached to the specified @tsk. Guaranteed to return some non-empty
2415 * subset of cpu_online_mask, even if this means going outside the
2416 * tasks cpuset.
2417 **/
2418
2419void cpuset_cpus_allowed(struct task_struct *tsk, struct cpumask *pmask)
2420{
2421 unsigned long flags;
2422
2423 spin_lock_irqsave(&callback_lock, flags);
2424 rcu_read_lock();
2425 guarantee_online_cpus(task_cs(tsk), pmask);
2426 rcu_read_unlock();
2427 spin_unlock_irqrestore(&callback_lock, flags);
2428}
2429
2430void cpuset_cpus_allowed_fallback(struct task_struct *tsk)
2431{
2432 rcu_read_lock();
2433 do_set_cpus_allowed(tsk, task_cs(tsk)->effective_cpus);
2434 rcu_read_unlock();
2435
2436 /*
2437 * We own tsk->cpus_allowed, nobody can change it under us.
2438 *
2439 * But we used cs && cs->cpus_allowed lockless and thus can
2440 * race with cgroup_attach_task() or update_cpumask() and get
2441 * the wrong tsk->cpus_allowed. However, both cases imply the
2442 * subsequent cpuset_change_cpumask()->set_cpus_allowed_ptr()
2443 * which takes task_rq_lock().
2444 *
2445 * If we are called after it dropped the lock we must see all
2446 * changes in tsk_cs()->cpus_allowed. Otherwise we can temporary
2447 * set any mask even if it is not right from task_cs() pov,
2448 * the pending set_cpus_allowed_ptr() will fix things.
2449 *
2450 * select_fallback_rq() will fix things ups and set cpu_possible_mask
2451 * if required.
2452 */
2453}
2454
2455void __init cpuset_init_current_mems_allowed(void)
2456{
2457 nodes_setall(current->mems_allowed);
2458}
2459
2460/**
2461 * cpuset_mems_allowed - return mems_allowed mask from a tasks cpuset.
2462 * @tsk: pointer to task_struct from which to obtain cpuset->mems_allowed.
2463 *
2464 * Description: Returns the nodemask_t mems_allowed of the cpuset
2465 * attached to the specified @tsk. Guaranteed to return some non-empty
2466 * subset of node_states[N_MEMORY], even if this means going outside the
2467 * tasks cpuset.
2468 **/
2469
2470nodemask_t cpuset_mems_allowed(struct task_struct *tsk)
2471{
2472 nodemask_t mask;
2473 unsigned long flags;
2474
2475 spin_lock_irqsave(&callback_lock, flags);
2476 rcu_read_lock();
2477 guarantee_online_mems(task_cs(tsk), &mask);
2478 rcu_read_unlock();
2479 spin_unlock_irqrestore(&callback_lock, flags);
2480
2481 return mask;
2482}
2483
2484/**
2485 * cpuset_nodemask_valid_mems_allowed - check nodemask vs. curremt mems_allowed
2486 * @nodemask: the nodemask to be checked
2487 *
2488 * Are any of the nodes in the nodemask allowed in current->mems_allowed?
2489 */
2490int cpuset_nodemask_valid_mems_allowed(nodemask_t *nodemask)
2491{
2492 return nodes_intersects(*nodemask, current->mems_allowed);
2493}
2494
2495/*
2496 * nearest_hardwall_ancestor() - Returns the nearest mem_exclusive or
2497 * mem_hardwall ancestor to the specified cpuset. Call holding
2498 * callback_lock. If no ancestor is mem_exclusive or mem_hardwall
2499 * (an unusual configuration), then returns the root cpuset.
2500 */
2501static struct cpuset *nearest_hardwall_ancestor(struct cpuset *cs)
2502{
2503 while (!(is_mem_exclusive(cs) || is_mem_hardwall(cs)) && parent_cs(cs))
2504 cs = parent_cs(cs);
2505 return cs;
2506}
2507
2508/**
2509 * cpuset_node_allowed - Can we allocate on a memory node?
2510 * @node: is this an allowed node?
2511 * @gfp_mask: memory allocation flags
2512 *
2513 * If we're in interrupt, yes, we can always allocate. If @node is set in
2514 * current's mems_allowed, yes. If it's not a __GFP_HARDWALL request and this
2515 * node is set in the nearest hardwalled cpuset ancestor to current's cpuset,
2516 * yes. If current has access to memory reserves due to TIF_MEMDIE, yes.
2517 * Otherwise, no.
2518 *
2519 * GFP_USER allocations are marked with the __GFP_HARDWALL bit,
2520 * and do not allow allocations outside the current tasks cpuset
2521 * unless the task has been OOM killed as is marked TIF_MEMDIE.
2522 * GFP_KERNEL allocations are not so marked, so can escape to the
2523 * nearest enclosing hardwalled ancestor cpuset.
2524 *
2525 * Scanning up parent cpusets requires callback_lock. The
2526 * __alloc_pages() routine only calls here with __GFP_HARDWALL bit
2527 * _not_ set if it's a GFP_KERNEL allocation, and all nodes in the
2528 * current tasks mems_allowed came up empty on the first pass over
2529 * the zonelist. So only GFP_KERNEL allocations, if all nodes in the
2530 * cpuset are short of memory, might require taking the callback_lock.
2531 *
2532 * The first call here from mm/page_alloc:get_page_from_freelist()
2533 * has __GFP_HARDWALL set in gfp_mask, enforcing hardwall cpusets,
2534 * so no allocation on a node outside the cpuset is allowed (unless
2535 * in interrupt, of course).
2536 *
2537 * The second pass through get_page_from_freelist() doesn't even call
2538 * here for GFP_ATOMIC calls. For those calls, the __alloc_pages()
2539 * variable 'wait' is not set, and the bit ALLOC_CPUSET is not set
2540 * in alloc_flags. That logic and the checks below have the combined
2541 * affect that:
2542 * in_interrupt - any node ok (current task context irrelevant)
2543 * GFP_ATOMIC - any node ok
2544 * TIF_MEMDIE - any node ok
2545 * GFP_KERNEL - any node in enclosing hardwalled cpuset ok
2546 * GFP_USER - only nodes in current tasks mems allowed ok.
2547 */
2548bool __cpuset_node_allowed(int node, gfp_t gfp_mask)
2549{
2550 struct cpuset *cs; /* current cpuset ancestors */
2551 int allowed; /* is allocation in zone z allowed? */
2552 unsigned long flags;
2553
2554 if (in_interrupt())
2555 return true;
2556 if (node_isset(node, current->mems_allowed))
2557 return true;
2558 /*
2559 * Allow tasks that have access to memory reserves because they have
2560 * been OOM killed to get memory anywhere.
2561 */
2562 if (unlikely(test_thread_flag(TIF_MEMDIE)))
2563 return true;
2564 if (gfp_mask & __GFP_HARDWALL) /* If hardwall request, stop here */
2565 return false;
2566
2567 if (current->flags & PF_EXITING) /* Let dying task have memory */
2568 return true;
2569
2570 /* Not hardwall and node outside mems_allowed: scan up cpusets */
2571 spin_lock_irqsave(&callback_lock, flags);
2572
2573 rcu_read_lock();
2574 cs = nearest_hardwall_ancestor(task_cs(current));
2575 allowed = node_isset(node, cs->mems_allowed);
2576 rcu_read_unlock();
2577
2578 spin_unlock_irqrestore(&callback_lock, flags);
2579 return allowed;
2580}
2581
2582/**
2583 * cpuset_mem_spread_node() - On which node to begin search for a file page
2584 * cpuset_slab_spread_node() - On which node to begin search for a slab page
2585 *
2586 * If a task is marked PF_SPREAD_PAGE or PF_SPREAD_SLAB (as for
2587 * tasks in a cpuset with is_spread_page or is_spread_slab set),
2588 * and if the memory allocation used cpuset_mem_spread_node()
2589 * to determine on which node to start looking, as it will for
2590 * certain page cache or slab cache pages such as used for file
2591 * system buffers and inode caches, then instead of starting on the
2592 * local node to look for a free page, rather spread the starting
2593 * node around the tasks mems_allowed nodes.
2594 *
2595 * We don't have to worry about the returned node being offline
2596 * because "it can't happen", and even if it did, it would be ok.
2597 *
2598 * The routines calling guarantee_online_mems() are careful to
2599 * only set nodes in task->mems_allowed that are online. So it
2600 * should not be possible for the following code to return an
2601 * offline node. But if it did, that would be ok, as this routine
2602 * is not returning the node where the allocation must be, only
2603 * the node where the search should start. The zonelist passed to
2604 * __alloc_pages() will include all nodes. If the slab allocator
2605 * is passed an offline node, it will fall back to the local node.
2606 * See kmem_cache_alloc_node().
2607 */
2608
2609static int cpuset_spread_node(int *rotor)
2610{
2611 return *rotor = next_node_in(*rotor, current->mems_allowed);
2612}
2613
2614int cpuset_mem_spread_node(void)
2615{
2616 if (current->cpuset_mem_spread_rotor == NUMA_NO_NODE)
2617 current->cpuset_mem_spread_rotor =
2618 node_random(¤t->mems_allowed);
2619
2620 return cpuset_spread_node(¤t->cpuset_mem_spread_rotor);
2621}
2622
2623int cpuset_slab_spread_node(void)
2624{
2625 if (current->cpuset_slab_spread_rotor == NUMA_NO_NODE)
2626 current->cpuset_slab_spread_rotor =
2627 node_random(¤t->mems_allowed);
2628
2629 return cpuset_spread_node(¤t->cpuset_slab_spread_rotor);
2630}
2631
2632EXPORT_SYMBOL_GPL(cpuset_mem_spread_node);
2633
2634/**
2635 * cpuset_mems_allowed_intersects - Does @tsk1's mems_allowed intersect @tsk2's?
2636 * @tsk1: pointer to task_struct of some task.
2637 * @tsk2: pointer to task_struct of some other task.
2638 *
2639 * Description: Return true if @tsk1's mems_allowed intersects the
2640 * mems_allowed of @tsk2. Used by the OOM killer to determine if
2641 * one of the task's memory usage might impact the memory available
2642 * to the other.
2643 **/
2644
2645int cpuset_mems_allowed_intersects(const struct task_struct *tsk1,
2646 const struct task_struct *tsk2)
2647{
2648 return nodes_intersects(tsk1->mems_allowed, tsk2->mems_allowed);
2649}
2650
2651/**
2652 * cpuset_print_current_mems_allowed - prints current's cpuset and mems_allowed
2653 *
2654 * Description: Prints current's name, cpuset name, and cached copy of its
2655 * mems_allowed to the kernel log.
2656 */
2657void cpuset_print_current_mems_allowed(void)
2658{
2659 struct cgroup *cgrp;
2660
2661 rcu_read_lock();
2662
2663 cgrp = task_cs(current)->css.cgroup;
2664 pr_info("%s cpuset=", current->comm);
2665 pr_cont_cgroup_name(cgrp);
2666 pr_cont(" mems_allowed=%*pbl\n",
2667 nodemask_pr_args(¤t->mems_allowed));
2668
2669 rcu_read_unlock();
2670}
2671
2672/*
2673 * Collection of memory_pressure is suppressed unless
2674 * this flag is enabled by writing "1" to the special
2675 * cpuset file 'memory_pressure_enabled' in the root cpuset.
2676 */
2677
2678int cpuset_memory_pressure_enabled __read_mostly;
2679
2680/**
2681 * cpuset_memory_pressure_bump - keep stats of per-cpuset reclaims.
2682 *
2683 * Keep a running average of the rate of synchronous (direct)
2684 * page reclaim efforts initiated by tasks in each cpuset.
2685 *
2686 * This represents the rate at which some task in the cpuset
2687 * ran low on memory on all nodes it was allowed to use, and
2688 * had to enter the kernels page reclaim code in an effort to
2689 * create more free memory by tossing clean pages or swapping
2690 * or writing dirty pages.
2691 *
2692 * Display to user space in the per-cpuset read-only file
2693 * "memory_pressure". Value displayed is an integer
2694 * representing the recent rate of entry into the synchronous
2695 * (direct) page reclaim by any task attached to the cpuset.
2696 **/
2697
2698void __cpuset_memory_pressure_bump(void)
2699{
2700 rcu_read_lock();
2701 fmeter_markevent(&task_cs(current)->fmeter);
2702 rcu_read_unlock();
2703}
2704
2705#ifdef CONFIG_PROC_PID_CPUSET
2706/*
2707 * proc_cpuset_show()
2708 * - Print tasks cpuset path into seq_file.
2709 * - Used for /proc/<pid>/cpuset.
2710 * - No need to task_lock(tsk) on this tsk->cpuset reference, as it
2711 * doesn't really matter if tsk->cpuset changes after we read it,
2712 * and we take cpuset_mutex, keeping cpuset_attach() from changing it
2713 * anyway.
2714 */
2715int proc_cpuset_show(struct seq_file *m, struct pid_namespace *ns,
2716 struct pid *pid, struct task_struct *tsk)
2717{
2718 char *buf;
2719 struct cgroup_subsys_state *css;
2720 int retval;
2721
2722 retval = -ENOMEM;
2723 buf = kmalloc(PATH_MAX, GFP_KERNEL);
2724 if (!buf)
2725 goto out;
2726
2727 css = task_get_css(tsk, cpuset_cgrp_id);
2728 retval = cgroup_path_ns(css->cgroup, buf, PATH_MAX,
2729 current->nsproxy->cgroup_ns);
2730 css_put(css);
2731 if (retval >= PATH_MAX)
2732 retval = -ENAMETOOLONG;
2733 if (retval < 0)
2734 goto out_free;
2735 seq_puts(m, buf);
2736 seq_putc(m, '\n');
2737 retval = 0;
2738out_free:
2739 kfree(buf);
2740out:
2741 return retval;
2742}
2743#endif /* CONFIG_PROC_PID_CPUSET */
2744
2745/* Display task mems_allowed in /proc/<pid>/status file. */
2746void cpuset_task_status_allowed(struct seq_file *m, struct task_struct *task)
2747{
2748 seq_printf(m, "Mems_allowed:\t%*pb\n",
2749 nodemask_pr_args(&task->mems_allowed));
2750 seq_printf(m, "Mems_allowed_list:\t%*pbl\n",
2751 nodemask_pr_args(&task->mems_allowed));
2752}