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/sched/mm.h>
48#include <linux/sched/task.h>
49#include <linux/seq_file.h>
50#include <linux/security.h>
51#include <linux/slab.h>
52#include <linux/spinlock.h>
53#include <linux/stat.h>
54#include <linux/string.h>
55#include <linux/time.h>
56#include <linux/time64.h>
57#include <linux/backing-dev.h>
58#include <linux/sort.h>
59#include <linux/oom.h>
60#include <linux/sched/isolation.h>
61#include <linux/uaccess.h>
62#include <linux/atomic.h>
63#include <linux/mutex.h>
64#include <linux/cgroup.h>
65#include <linux/wait.h>
66
67DEFINE_STATIC_KEY_FALSE(cpusets_pre_enable_key);
68DEFINE_STATIC_KEY_FALSE(cpusets_enabled_key);
69
70/* See "Frequency meter" comments, below. */
71
72struct fmeter {
73 int cnt; /* unprocessed events count */
74 int val; /* most recent output value */
75 time64_t time; /* clock (secs) when val computed */
76 spinlock_t lock; /* guards read or write of above */
77};
78
79struct cpuset {
80 struct cgroup_subsys_state css;
81
82 unsigned long flags; /* "unsigned long" so bitops work */
83
84 /*
85 * On default hierarchy:
86 *
87 * The user-configured masks can only be changed by writing to
88 * cpuset.cpus and cpuset.mems, and won't be limited by the
89 * parent masks.
90 *
91 * The effective masks is the real masks that apply to the tasks
92 * in the cpuset. They may be changed if the configured masks are
93 * changed or hotplug happens.
94 *
95 * effective_mask == configured_mask & parent's effective_mask,
96 * and if it ends up empty, it will inherit the parent's mask.
97 *
98 *
99 * On legacy hierachy:
100 *
101 * The user-configured masks are always the same with effective masks.
102 */
103
104 /* user-configured CPUs and Memory Nodes allow to tasks */
105 cpumask_var_t cpus_allowed;
106 nodemask_t mems_allowed;
107
108 /* effective CPUs and Memory Nodes allow to tasks */
109 cpumask_var_t effective_cpus;
110 nodemask_t effective_mems;
111
112 /*
113 * This is old Memory Nodes tasks took on.
114 *
115 * - top_cpuset.old_mems_allowed is initialized to mems_allowed.
116 * - A new cpuset's old_mems_allowed is initialized when some
117 * task is moved into it.
118 * - old_mems_allowed is used in cpuset_migrate_mm() when we change
119 * cpuset.mems_allowed and have tasks' nodemask updated, and
120 * then old_mems_allowed is updated to mems_allowed.
121 */
122 nodemask_t old_mems_allowed;
123
124 struct fmeter fmeter; /* memory_pressure filter */
125
126 /*
127 * Tasks are being attached to this cpuset. Used to prevent
128 * zeroing cpus/mems_allowed between ->can_attach() and ->attach().
129 */
130 int attach_in_progress;
131
132 /* partition number for rebuild_sched_domains() */
133 int pn;
134
135 /* for custom sched domain */
136 int relax_domain_level;
137};
138
139static inline struct cpuset *css_cs(struct cgroup_subsys_state *css)
140{
141 return css ? container_of(css, struct cpuset, css) : NULL;
142}
143
144/* Retrieve the cpuset for a task */
145static inline struct cpuset *task_cs(struct task_struct *task)
146{
147 return css_cs(task_css(task, cpuset_cgrp_id));
148}
149
150static inline struct cpuset *parent_cs(struct cpuset *cs)
151{
152 return css_cs(cs->css.parent);
153}
154
155#ifdef CONFIG_NUMA
156static inline bool task_has_mempolicy(struct task_struct *task)
157{
158 return task->mempolicy;
159}
160#else
161static inline bool task_has_mempolicy(struct task_struct *task)
162{
163 return false;
164}
165#endif
166
167
168/* bits in struct cpuset flags field */
169typedef enum {
170 CS_ONLINE,
171 CS_CPU_EXCLUSIVE,
172 CS_MEM_EXCLUSIVE,
173 CS_MEM_HARDWALL,
174 CS_MEMORY_MIGRATE,
175 CS_SCHED_LOAD_BALANCE,
176 CS_SPREAD_PAGE,
177 CS_SPREAD_SLAB,
178} cpuset_flagbits_t;
179
180/* convenient tests for these bits */
181static inline bool is_cpuset_online(struct cpuset *cs)
182{
183 return test_bit(CS_ONLINE, &cs->flags) && !css_is_dying(&cs->css);
184}
185
186static inline int is_cpu_exclusive(const struct cpuset *cs)
187{
188 return test_bit(CS_CPU_EXCLUSIVE, &cs->flags);
189}
190
191static inline int is_mem_exclusive(const struct cpuset *cs)
192{
193 return test_bit(CS_MEM_EXCLUSIVE, &cs->flags);
194}
195
196static inline int is_mem_hardwall(const struct cpuset *cs)
197{
198 return test_bit(CS_MEM_HARDWALL, &cs->flags);
199}
200
201static inline int is_sched_load_balance(const struct cpuset *cs)
202{
203 return test_bit(CS_SCHED_LOAD_BALANCE, &cs->flags);
204}
205
206static inline int is_memory_migrate(const struct cpuset *cs)
207{
208 return test_bit(CS_MEMORY_MIGRATE, &cs->flags);
209}
210
211static inline int is_spread_page(const struct cpuset *cs)
212{
213 return test_bit(CS_SPREAD_PAGE, &cs->flags);
214}
215
216static inline int is_spread_slab(const struct cpuset *cs)
217{
218 return test_bit(CS_SPREAD_SLAB, &cs->flags);
219}
220
221static struct cpuset top_cpuset = {
222 .flags = ((1 << CS_ONLINE) | (1 << CS_CPU_EXCLUSIVE) |
223 (1 << CS_MEM_EXCLUSIVE)),
224};
225
226/**
227 * cpuset_for_each_child - traverse online children of a cpuset
228 * @child_cs: loop cursor pointing to the current child
229 * @pos_css: used for iteration
230 * @parent_cs: target cpuset to walk children of
231 *
232 * Walk @child_cs through the online children of @parent_cs. Must be used
233 * with RCU read locked.
234 */
235#define cpuset_for_each_child(child_cs, pos_css, parent_cs) \
236 css_for_each_child((pos_css), &(parent_cs)->css) \
237 if (is_cpuset_online(((child_cs) = css_cs((pos_css)))))
238
239/**
240 * cpuset_for_each_descendant_pre - pre-order walk of a cpuset's descendants
241 * @des_cs: loop cursor pointing to the current descendant
242 * @pos_css: used for iteration
243 * @root_cs: target cpuset to walk ancestor of
244 *
245 * Walk @des_cs through the online descendants of @root_cs. Must be used
246 * with RCU read locked. The caller may modify @pos_css by calling
247 * css_rightmost_descendant() to skip subtree. @root_cs is included in the
248 * iteration and the first node to be visited.
249 */
250#define cpuset_for_each_descendant_pre(des_cs, pos_css, root_cs) \
251 css_for_each_descendant_pre((pos_css), &(root_cs)->css) \
252 if (is_cpuset_online(((des_cs) = css_cs((pos_css)))))
253
254/*
255 * There are two global locks guarding cpuset structures - cpuset_mutex and
256 * callback_lock. We also require taking task_lock() when dereferencing a
257 * task's cpuset pointer. See "The task_lock() exception", at the end of this
258 * comment.
259 *
260 * A task must hold both locks to modify cpusets. If a task holds
261 * cpuset_mutex, then it blocks others wanting that mutex, ensuring that it
262 * is the only task able to also acquire callback_lock and be able to
263 * modify cpusets. It can perform various checks on the cpuset structure
264 * first, knowing nothing will change. It can also allocate memory while
265 * just holding cpuset_mutex. While it is performing these checks, various
266 * callback routines can briefly acquire callback_lock to query cpusets.
267 * Once it is ready to make the changes, it takes callback_lock, blocking
268 * everyone else.
269 *
270 * Calls to the kernel memory allocator can not be made while holding
271 * callback_lock, as that would risk double tripping on callback_lock
272 * from one of the callbacks into the cpuset code from within
273 * __alloc_pages().
274 *
275 * If a task is only holding callback_lock, then it has read-only
276 * access to cpusets.
277 *
278 * Now, the task_struct fields mems_allowed and mempolicy may be changed
279 * by other task, we use alloc_lock in the task_struct fields to protect
280 * them.
281 *
282 * The cpuset_common_file_read() handlers only hold callback_lock across
283 * small pieces of code, such as when reading out possibly multi-word
284 * cpumasks and nodemasks.
285 *
286 * Accessing a task's cpuset should be done in accordance with the
287 * guidelines for accessing subsystem state in kernel/cgroup.c
288 */
289
290static DEFINE_MUTEX(cpuset_mutex);
291static DEFINE_SPINLOCK(callback_lock);
292
293static struct workqueue_struct *cpuset_migrate_mm_wq;
294
295/*
296 * CPU / memory hotplug is handled asynchronously.
297 */
298static void cpuset_hotplug_workfn(struct work_struct *work);
299static DECLARE_WORK(cpuset_hotplug_work, cpuset_hotplug_workfn);
300
301static DECLARE_WAIT_QUEUE_HEAD(cpuset_attach_wq);
302
303/*
304 * Cgroup v2 behavior is used when on default hierarchy or the
305 * cgroup_v2_mode flag is set.
306 */
307static inline bool is_in_v2_mode(void)
308{
309 return cgroup_subsys_on_dfl(cpuset_cgrp_subsys) ||
310 (cpuset_cgrp_subsys.root->flags & CGRP_ROOT_CPUSET_V2_MODE);
311}
312
313/*
314 * This is ugly, but preserves the userspace API for existing cpuset
315 * users. If someone tries to mount the "cpuset" filesystem, we
316 * silently switch it to mount "cgroup" instead
317 */
318static struct dentry *cpuset_mount(struct file_system_type *fs_type,
319 int flags, const char *unused_dev_name, void *data)
320{
321 struct file_system_type *cgroup_fs = get_fs_type("cgroup");
322 struct dentry *ret = ERR_PTR(-ENODEV);
323 if (cgroup_fs) {
324 char mountopts[] =
325 "cpuset,noprefix,"
326 "release_agent=/sbin/cpuset_release_agent";
327 ret = cgroup_fs->mount(cgroup_fs, flags,
328 unused_dev_name, mountopts);
329 put_filesystem(cgroup_fs);
330 }
331 return ret;
332}
333
334static struct file_system_type cpuset_fs_type = {
335 .name = "cpuset",
336 .mount = cpuset_mount,
337};
338
339/*
340 * Return in pmask the portion of a cpusets's cpus_allowed that
341 * are online. If none are online, walk up the cpuset hierarchy
342 * until we find one that does have some online cpus.
343 *
344 * One way or another, we guarantee to return some non-empty subset
345 * of cpu_online_mask.
346 *
347 * Call with callback_lock or cpuset_mutex held.
348 */
349static void guarantee_online_cpus(struct cpuset *cs, struct cpumask *pmask)
350{
351 while (!cpumask_intersects(cs->effective_cpus, cpu_online_mask)) {
352 cs = parent_cs(cs);
353 if (unlikely(!cs)) {
354 /*
355 * The top cpuset doesn't have any online cpu as a
356 * consequence of a race between cpuset_hotplug_work
357 * and cpu hotplug notifier. But we know the top
358 * cpuset's effective_cpus is on its way to to be
359 * identical to cpu_online_mask.
360 */
361 cpumask_copy(pmask, cpu_online_mask);
362 return;
363 }
364 }
365 cpumask_and(pmask, cs->effective_cpus, cpu_online_mask);
366}
367
368/*
369 * Return in *pmask the portion of a cpusets's mems_allowed that
370 * are online, with memory. If none are online with memory, walk
371 * up the cpuset hierarchy until we find one that does have some
372 * online mems. The top cpuset always has some mems online.
373 *
374 * One way or another, we guarantee to return some non-empty subset
375 * of node_states[N_MEMORY].
376 *
377 * Call with callback_lock or cpuset_mutex held.
378 */
379static void guarantee_online_mems(struct cpuset *cs, nodemask_t *pmask)
380{
381 while (!nodes_intersects(cs->effective_mems, node_states[N_MEMORY]))
382 cs = parent_cs(cs);
383 nodes_and(*pmask, cs->effective_mems, node_states[N_MEMORY]);
384}
385
386/*
387 * update task's spread flag if cpuset's page/slab spread flag is set
388 *
389 * Call with callback_lock or cpuset_mutex held.
390 */
391static void cpuset_update_task_spread_flag(struct cpuset *cs,
392 struct task_struct *tsk)
393{
394 if (is_spread_page(cs))
395 task_set_spread_page(tsk);
396 else
397 task_clear_spread_page(tsk);
398
399 if (is_spread_slab(cs))
400 task_set_spread_slab(tsk);
401 else
402 task_clear_spread_slab(tsk);
403}
404
405/*
406 * is_cpuset_subset(p, q) - Is cpuset p a subset of cpuset q?
407 *
408 * One cpuset is a subset of another if all its allowed CPUs and
409 * Memory Nodes are a subset of the other, and its exclusive flags
410 * are only set if the other's are set. Call holding cpuset_mutex.
411 */
412
413static int is_cpuset_subset(const struct cpuset *p, const struct cpuset *q)
414{
415 return cpumask_subset(p->cpus_allowed, q->cpus_allowed) &&
416 nodes_subset(p->mems_allowed, q->mems_allowed) &&
417 is_cpu_exclusive(p) <= is_cpu_exclusive(q) &&
418 is_mem_exclusive(p) <= is_mem_exclusive(q);
419}
420
421/**
422 * alloc_trial_cpuset - allocate a trial cpuset
423 * @cs: the cpuset that the trial cpuset duplicates
424 */
425static struct cpuset *alloc_trial_cpuset(struct cpuset *cs)
426{
427 struct cpuset *trial;
428
429 trial = kmemdup(cs, sizeof(*cs), GFP_KERNEL);
430 if (!trial)
431 return NULL;
432
433 if (!alloc_cpumask_var(&trial->cpus_allowed, GFP_KERNEL))
434 goto free_cs;
435 if (!alloc_cpumask_var(&trial->effective_cpus, GFP_KERNEL))
436 goto free_cpus;
437
438 cpumask_copy(trial->cpus_allowed, cs->cpus_allowed);
439 cpumask_copy(trial->effective_cpus, cs->effective_cpus);
440 return trial;
441
442free_cpus:
443 free_cpumask_var(trial->cpus_allowed);
444free_cs:
445 kfree(trial);
446 return NULL;
447}
448
449/**
450 * free_trial_cpuset - free the trial cpuset
451 * @trial: the trial cpuset to be freed
452 */
453static void free_trial_cpuset(struct cpuset *trial)
454{
455 free_cpumask_var(trial->effective_cpus);
456 free_cpumask_var(trial->cpus_allowed);
457 kfree(trial);
458}
459
460/*
461 * validate_change() - Used to validate that any proposed cpuset change
462 * follows the structural rules for cpusets.
463 *
464 * If we replaced the flag and mask values of the current cpuset
465 * (cur) with those values in the trial cpuset (trial), would
466 * our various subset and exclusive rules still be valid? Presumes
467 * cpuset_mutex held.
468 *
469 * 'cur' is the address of an actual, in-use cpuset. Operations
470 * such as list traversal that depend on the actual address of the
471 * cpuset in the list must use cur below, not trial.
472 *
473 * 'trial' is the address of bulk structure copy of cur, with
474 * perhaps one or more of the fields cpus_allowed, mems_allowed,
475 * or flags changed to new, trial values.
476 *
477 * Return 0 if valid, -errno if not.
478 */
479
480static int validate_change(struct cpuset *cur, struct cpuset *trial)
481{
482 struct cgroup_subsys_state *css;
483 struct cpuset *c, *par;
484 int ret;
485
486 rcu_read_lock();
487
488 /* Each of our child cpusets must be a subset of us */
489 ret = -EBUSY;
490 cpuset_for_each_child(c, css, cur)
491 if (!is_cpuset_subset(c, trial))
492 goto out;
493
494 /* Remaining checks don't apply to root cpuset */
495 ret = 0;
496 if (cur == &top_cpuset)
497 goto out;
498
499 par = parent_cs(cur);
500
501 /* On legacy hiearchy, we must be a subset of our parent cpuset. */
502 ret = -EACCES;
503 if (!is_in_v2_mode() && !is_cpuset_subset(trial, par))
504 goto out;
505
506 /*
507 * If either I or some sibling (!= me) is exclusive, we can't
508 * overlap
509 */
510 ret = -EINVAL;
511 cpuset_for_each_child(c, css, par) {
512 if ((is_cpu_exclusive(trial) || is_cpu_exclusive(c)) &&
513 c != cur &&
514 cpumask_intersects(trial->cpus_allowed, c->cpus_allowed))
515 goto out;
516 if ((is_mem_exclusive(trial) || is_mem_exclusive(c)) &&
517 c != cur &&
518 nodes_intersects(trial->mems_allowed, c->mems_allowed))
519 goto out;
520 }
521
522 /*
523 * Cpusets with tasks - existing or newly being attached - can't
524 * be changed to have empty cpus_allowed or mems_allowed.
525 */
526 ret = -ENOSPC;
527 if ((cgroup_is_populated(cur->css.cgroup) || cur->attach_in_progress)) {
528 if (!cpumask_empty(cur->cpus_allowed) &&
529 cpumask_empty(trial->cpus_allowed))
530 goto out;
531 if (!nodes_empty(cur->mems_allowed) &&
532 nodes_empty(trial->mems_allowed))
533 goto out;
534 }
535
536 /*
537 * We can't shrink if we won't have enough room for SCHED_DEADLINE
538 * tasks.
539 */
540 ret = -EBUSY;
541 if (is_cpu_exclusive(cur) &&
542 !cpuset_cpumask_can_shrink(cur->cpus_allowed,
543 trial->cpus_allowed))
544 goto out;
545
546 ret = 0;
547out:
548 rcu_read_unlock();
549 return ret;
550}
551
552#ifdef CONFIG_SMP
553/*
554 * Helper routine for generate_sched_domains().
555 * Do cpusets a, b have overlapping effective cpus_allowed masks?
556 */
557static int cpusets_overlap(struct cpuset *a, struct cpuset *b)
558{
559 return cpumask_intersects(a->effective_cpus, b->effective_cpus);
560}
561
562static void
563update_domain_attr(struct sched_domain_attr *dattr, struct cpuset *c)
564{
565 if (dattr->relax_domain_level < c->relax_domain_level)
566 dattr->relax_domain_level = c->relax_domain_level;
567 return;
568}
569
570static void update_domain_attr_tree(struct sched_domain_attr *dattr,
571 struct cpuset *root_cs)
572{
573 struct cpuset *cp;
574 struct cgroup_subsys_state *pos_css;
575
576 rcu_read_lock();
577 cpuset_for_each_descendant_pre(cp, pos_css, root_cs) {
578 /* skip the whole subtree if @cp doesn't have any CPU */
579 if (cpumask_empty(cp->cpus_allowed)) {
580 pos_css = css_rightmost_descendant(pos_css);
581 continue;
582 }
583
584 if (is_sched_load_balance(cp))
585 update_domain_attr(dattr, cp);
586 }
587 rcu_read_unlock();
588}
589
590/* Must be called with cpuset_mutex held. */
591static inline int nr_cpusets(void)
592{
593 /* jump label reference count + the top-level cpuset */
594 return static_key_count(&cpusets_enabled_key.key) + 1;
595}
596
597/*
598 * generate_sched_domains()
599 *
600 * This function builds a partial partition of the systems CPUs
601 * A 'partial partition' is a set of non-overlapping subsets whose
602 * union is a subset of that set.
603 * The output of this function needs to be passed to kernel/sched/core.c
604 * partition_sched_domains() routine, which will rebuild the scheduler's
605 * load balancing domains (sched domains) as specified by that partial
606 * partition.
607 *
608 * See "What is sched_load_balance" in Documentation/cgroups/cpusets.txt
609 * for a background explanation of this.
610 *
611 * Does not return errors, on the theory that the callers of this
612 * routine would rather not worry about failures to rebuild sched
613 * domains when operating in the severe memory shortage situations
614 * that could cause allocation failures below.
615 *
616 * Must be called with cpuset_mutex held.
617 *
618 * The three key local variables below are:
619 * q - a linked-list queue of cpuset pointers, used to implement a
620 * top-down scan of all cpusets. This scan loads a pointer
621 * to each cpuset marked is_sched_load_balance into the
622 * array 'csa'. For our purposes, rebuilding the schedulers
623 * sched domains, we can ignore !is_sched_load_balance cpusets.
624 * csa - (for CpuSet Array) Array of pointers to all the cpusets
625 * that need to be load balanced, for convenient iterative
626 * access by the subsequent code that finds the best partition,
627 * i.e the set of domains (subsets) of CPUs such that the
628 * cpus_allowed of every cpuset marked is_sched_load_balance
629 * is a subset of one of these domains, while there are as
630 * many such domains as possible, each as small as possible.
631 * doms - Conversion of 'csa' to an array of cpumasks, for passing to
632 * the kernel/sched/core.c routine partition_sched_domains() in a
633 * convenient format, that can be easily compared to the prior
634 * value to determine what partition elements (sched domains)
635 * were changed (added or removed.)
636 *
637 * Finding the best partition (set of domains):
638 * The triple nested loops below over i, j, k scan over the
639 * load balanced cpusets (using the array of cpuset pointers in
640 * csa[]) looking for pairs of cpusets that have overlapping
641 * cpus_allowed, but which don't have the same 'pn' partition
642 * number and gives them in the same partition number. It keeps
643 * looping on the 'restart' label until it can no longer find
644 * any such pairs.
645 *
646 * The union of the cpus_allowed masks from the set of
647 * all cpusets having the same 'pn' value then form the one
648 * element of the partition (one sched domain) to be passed to
649 * partition_sched_domains().
650 */
651static int generate_sched_domains(cpumask_var_t **domains,
652 struct sched_domain_attr **attributes)
653{
654 struct cpuset *cp; /* scans q */
655 struct cpuset **csa; /* array of all cpuset ptrs */
656 int csn; /* how many cpuset ptrs in csa so far */
657 int i, j, k; /* indices for partition finding loops */
658 cpumask_var_t *doms; /* resulting partition; i.e. sched domains */
659 struct sched_domain_attr *dattr; /* attributes for custom domains */
660 int ndoms = 0; /* number of sched domains in result */
661 int nslot; /* next empty doms[] struct cpumask slot */
662 struct cgroup_subsys_state *pos_css;
663
664 doms = NULL;
665 dattr = NULL;
666 csa = NULL;
667
668 /* Special case for the 99% of systems with one, full, sched domain */
669 if (is_sched_load_balance(&top_cpuset)) {
670 ndoms = 1;
671 doms = alloc_sched_domains(ndoms);
672 if (!doms)
673 goto done;
674
675 dattr = kmalloc(sizeof(struct sched_domain_attr), GFP_KERNEL);
676 if (dattr) {
677 *dattr = SD_ATTR_INIT;
678 update_domain_attr_tree(dattr, &top_cpuset);
679 }
680 cpumask_and(doms[0], top_cpuset.effective_cpus,
681 housekeeping_cpumask(HK_FLAG_DOMAIN));
682
683 goto done;
684 }
685
686 csa = kmalloc(nr_cpusets() * sizeof(cp), GFP_KERNEL);
687 if (!csa)
688 goto done;
689 csn = 0;
690
691 rcu_read_lock();
692 cpuset_for_each_descendant_pre(cp, pos_css, &top_cpuset) {
693 if (cp == &top_cpuset)
694 continue;
695 /*
696 * Continue traversing beyond @cp iff @cp has some CPUs and
697 * isn't load balancing. The former is obvious. The
698 * latter: All child cpusets contain a subset of the
699 * parent's cpus, so just skip them, and then we call
700 * update_domain_attr_tree() to calc relax_domain_level of
701 * the corresponding sched domain.
702 */
703 if (!cpumask_empty(cp->cpus_allowed) &&
704 !(is_sched_load_balance(cp) &&
705 cpumask_intersects(cp->cpus_allowed,
706 housekeeping_cpumask(HK_FLAG_DOMAIN))))
707 continue;
708
709 if (is_sched_load_balance(cp))
710 csa[csn++] = cp;
711
712 /* skip @cp's subtree */
713 pos_css = css_rightmost_descendant(pos_css);
714 }
715 rcu_read_unlock();
716
717 for (i = 0; i < csn; i++)
718 csa[i]->pn = i;
719 ndoms = csn;
720
721restart:
722 /* Find the best partition (set of sched domains) */
723 for (i = 0; i < csn; i++) {
724 struct cpuset *a = csa[i];
725 int apn = a->pn;
726
727 for (j = 0; j < csn; j++) {
728 struct cpuset *b = csa[j];
729 int bpn = b->pn;
730
731 if (apn != bpn && cpusets_overlap(a, b)) {
732 for (k = 0; k < csn; k++) {
733 struct cpuset *c = csa[k];
734
735 if (c->pn == bpn)
736 c->pn = apn;
737 }
738 ndoms--; /* one less element */
739 goto restart;
740 }
741 }
742 }
743
744 /*
745 * Now we know how many domains to create.
746 * Convert <csn, csa> to <ndoms, doms> and populate cpu masks.
747 */
748 doms = alloc_sched_domains(ndoms);
749 if (!doms)
750 goto done;
751
752 /*
753 * The rest of the code, including the scheduler, can deal with
754 * dattr==NULL case. No need to abort if alloc fails.
755 */
756 dattr = kmalloc(ndoms * sizeof(struct sched_domain_attr), GFP_KERNEL);
757
758 for (nslot = 0, i = 0; i < csn; i++) {
759 struct cpuset *a = csa[i];
760 struct cpumask *dp;
761 int apn = a->pn;
762
763 if (apn < 0) {
764 /* Skip completed partitions */
765 continue;
766 }
767
768 dp = doms[nslot];
769
770 if (nslot == ndoms) {
771 static int warnings = 10;
772 if (warnings) {
773 pr_warn("rebuild_sched_domains confused: nslot %d, ndoms %d, csn %d, i %d, apn %d\n",
774 nslot, ndoms, csn, i, apn);
775 warnings--;
776 }
777 continue;
778 }
779
780 cpumask_clear(dp);
781 if (dattr)
782 *(dattr + nslot) = SD_ATTR_INIT;
783 for (j = i; j < csn; j++) {
784 struct cpuset *b = csa[j];
785
786 if (apn == b->pn) {
787 cpumask_or(dp, dp, b->effective_cpus);
788 cpumask_and(dp, dp, housekeeping_cpumask(HK_FLAG_DOMAIN));
789 if (dattr)
790 update_domain_attr_tree(dattr + nslot, b);
791
792 /* Done with this partition */
793 b->pn = -1;
794 }
795 }
796 nslot++;
797 }
798 BUG_ON(nslot != ndoms);
799
800done:
801 kfree(csa);
802
803 /*
804 * Fallback to the default domain if kmalloc() failed.
805 * See comments in partition_sched_domains().
806 */
807 if (doms == NULL)
808 ndoms = 1;
809
810 *domains = doms;
811 *attributes = dattr;
812 return ndoms;
813}
814
815/*
816 * Rebuild scheduler domains.
817 *
818 * If the flag 'sched_load_balance' of any cpuset with non-empty
819 * 'cpus' changes, or if the 'cpus' allowed changes in any cpuset
820 * which has that flag enabled, or if any cpuset with a non-empty
821 * 'cpus' is removed, then call this routine to rebuild the
822 * scheduler's dynamic sched domains.
823 *
824 * Call with cpuset_mutex held. Takes get_online_cpus().
825 */
826static void rebuild_sched_domains_locked(void)
827{
828 struct sched_domain_attr *attr;
829 cpumask_var_t *doms;
830 int ndoms;
831
832 lockdep_assert_held(&cpuset_mutex);
833 get_online_cpus();
834
835 /*
836 * We have raced with CPU hotplug. Don't do anything to avoid
837 * passing doms with offlined cpu to partition_sched_domains().
838 * Anyways, hotplug work item will rebuild sched domains.
839 */
840 if (!cpumask_equal(top_cpuset.effective_cpus, cpu_active_mask))
841 goto out;
842
843 /* Generate domain masks and attrs */
844 ndoms = generate_sched_domains(&doms, &attr);
845
846 /* Have scheduler rebuild the domains */
847 partition_sched_domains(ndoms, doms, attr);
848out:
849 put_online_cpus();
850}
851#else /* !CONFIG_SMP */
852static void rebuild_sched_domains_locked(void)
853{
854}
855#endif /* CONFIG_SMP */
856
857void rebuild_sched_domains(void)
858{
859 mutex_lock(&cpuset_mutex);
860 rebuild_sched_domains_locked();
861 mutex_unlock(&cpuset_mutex);
862}
863
864/**
865 * update_tasks_cpumask - Update the cpumasks of tasks in the cpuset.
866 * @cs: the cpuset in which each task's cpus_allowed mask needs to be changed
867 *
868 * Iterate through each task of @cs updating its cpus_allowed to the
869 * effective cpuset's. As this function is called with cpuset_mutex held,
870 * cpuset membership stays stable.
871 */
872static void update_tasks_cpumask(struct cpuset *cs)
873{
874 struct css_task_iter it;
875 struct task_struct *task;
876
877 css_task_iter_start(&cs->css, 0, &it);
878 while ((task = css_task_iter_next(&it)))
879 set_cpus_allowed_ptr(task, cs->effective_cpus);
880 css_task_iter_end(&it);
881}
882
883/*
884 * update_cpumasks_hier - Update effective cpumasks and tasks in the subtree
885 * @cs: the cpuset to consider
886 * @new_cpus: temp variable for calculating new effective_cpus
887 *
888 * When congifured cpumask is changed, the effective cpumasks of this cpuset
889 * and all its descendants need to be updated.
890 *
891 * On legacy hierachy, effective_cpus will be the same with cpu_allowed.
892 *
893 * Called with cpuset_mutex held
894 */
895static void update_cpumasks_hier(struct cpuset *cs, struct cpumask *new_cpus)
896{
897 struct cpuset *cp;
898 struct cgroup_subsys_state *pos_css;
899 bool need_rebuild_sched_domains = false;
900
901 rcu_read_lock();
902 cpuset_for_each_descendant_pre(cp, pos_css, cs) {
903 struct cpuset *parent = parent_cs(cp);
904
905 cpumask_and(new_cpus, cp->cpus_allowed, parent->effective_cpus);
906
907 /*
908 * If it becomes empty, inherit the effective mask of the
909 * parent, which is guaranteed to have some CPUs.
910 */
911 if (is_in_v2_mode() && cpumask_empty(new_cpus))
912 cpumask_copy(new_cpus, parent->effective_cpus);
913
914 /* Skip the whole subtree if the cpumask remains the same. */
915 if (cpumask_equal(new_cpus, cp->effective_cpus)) {
916 pos_css = css_rightmost_descendant(pos_css);
917 continue;
918 }
919
920 if (!css_tryget_online(&cp->css))
921 continue;
922 rcu_read_unlock();
923
924 spin_lock_irq(&callback_lock);
925 cpumask_copy(cp->effective_cpus, new_cpus);
926 spin_unlock_irq(&callback_lock);
927
928 WARN_ON(!is_in_v2_mode() &&
929 !cpumask_equal(cp->cpus_allowed, cp->effective_cpus));
930
931 update_tasks_cpumask(cp);
932
933 /*
934 * If the effective cpumask of any non-empty cpuset is changed,
935 * we need to rebuild sched domains.
936 */
937 if (!cpumask_empty(cp->cpus_allowed) &&
938 is_sched_load_balance(cp))
939 need_rebuild_sched_domains = true;
940
941 rcu_read_lock();
942 css_put(&cp->css);
943 }
944 rcu_read_unlock();
945
946 if (need_rebuild_sched_domains)
947 rebuild_sched_domains_locked();
948}
949
950/**
951 * update_cpumask - update the cpus_allowed mask of a cpuset and all tasks in it
952 * @cs: the cpuset to consider
953 * @trialcs: trial cpuset
954 * @buf: buffer of cpu numbers written to this cpuset
955 */
956static int update_cpumask(struct cpuset *cs, struct cpuset *trialcs,
957 const char *buf)
958{
959 int retval;
960
961 /* top_cpuset.cpus_allowed tracks cpu_online_mask; it's read-only */
962 if (cs == &top_cpuset)
963 return -EACCES;
964
965 /*
966 * An empty cpus_allowed is ok only if the cpuset has no tasks.
967 * Since cpulist_parse() fails on an empty mask, we special case
968 * that parsing. The validate_change() call ensures that cpusets
969 * with tasks have cpus.
970 */
971 if (!*buf) {
972 cpumask_clear(trialcs->cpus_allowed);
973 } else {
974 retval = cpulist_parse(buf, trialcs->cpus_allowed);
975 if (retval < 0)
976 return retval;
977
978 if (!cpumask_subset(trialcs->cpus_allowed,
979 top_cpuset.cpus_allowed))
980 return -EINVAL;
981 }
982
983 /* Nothing to do if the cpus didn't change */
984 if (cpumask_equal(cs->cpus_allowed, trialcs->cpus_allowed))
985 return 0;
986
987 retval = validate_change(cs, trialcs);
988 if (retval < 0)
989 return retval;
990
991 spin_lock_irq(&callback_lock);
992 cpumask_copy(cs->cpus_allowed, trialcs->cpus_allowed);
993 spin_unlock_irq(&callback_lock);
994
995 /* use trialcs->cpus_allowed as a temp variable */
996 update_cpumasks_hier(cs, trialcs->cpus_allowed);
997 return 0;
998}
999
1000/*
1001 * Migrate memory region from one set of nodes to another. This is
1002 * performed asynchronously as it can be called from process migration path
1003 * holding locks involved in process management. All mm migrations are
1004 * performed in the queued order and can be waited for by flushing
1005 * cpuset_migrate_mm_wq.
1006 */
1007
1008struct cpuset_migrate_mm_work {
1009 struct work_struct work;
1010 struct mm_struct *mm;
1011 nodemask_t from;
1012 nodemask_t to;
1013};
1014
1015static void cpuset_migrate_mm_workfn(struct work_struct *work)
1016{
1017 struct cpuset_migrate_mm_work *mwork =
1018 container_of(work, struct cpuset_migrate_mm_work, work);
1019
1020 /* on a wq worker, no need to worry about %current's mems_allowed */
1021 do_migrate_pages(mwork->mm, &mwork->from, &mwork->to, MPOL_MF_MOVE_ALL);
1022 mmput(mwork->mm);
1023 kfree(mwork);
1024}
1025
1026static void cpuset_migrate_mm(struct mm_struct *mm, const nodemask_t *from,
1027 const nodemask_t *to)
1028{
1029 struct cpuset_migrate_mm_work *mwork;
1030
1031 mwork = kzalloc(sizeof(*mwork), GFP_KERNEL);
1032 if (mwork) {
1033 mwork->mm = mm;
1034 mwork->from = *from;
1035 mwork->to = *to;
1036 INIT_WORK(&mwork->work, cpuset_migrate_mm_workfn);
1037 queue_work(cpuset_migrate_mm_wq, &mwork->work);
1038 } else {
1039 mmput(mm);
1040 }
1041}
1042
1043static void cpuset_post_attach(void)
1044{
1045 flush_workqueue(cpuset_migrate_mm_wq);
1046}
1047
1048/*
1049 * cpuset_change_task_nodemask - change task's mems_allowed and mempolicy
1050 * @tsk: the task to change
1051 * @newmems: new nodes that the task will be set
1052 *
1053 * We use the mems_allowed_seq seqlock to safely update both tsk->mems_allowed
1054 * and rebind an eventual tasks' mempolicy. If the task is allocating in
1055 * parallel, it might temporarily see an empty intersection, which results in
1056 * a seqlock check and retry before OOM or allocation failure.
1057 */
1058static void cpuset_change_task_nodemask(struct task_struct *tsk,
1059 nodemask_t *newmems)
1060{
1061 task_lock(tsk);
1062
1063 local_irq_disable();
1064 write_seqcount_begin(&tsk->mems_allowed_seq);
1065
1066 nodes_or(tsk->mems_allowed, tsk->mems_allowed, *newmems);
1067 mpol_rebind_task(tsk, newmems);
1068 tsk->mems_allowed = *newmems;
1069
1070 write_seqcount_end(&tsk->mems_allowed_seq);
1071 local_irq_enable();
1072
1073 task_unlock(tsk);
1074}
1075
1076static void *cpuset_being_rebound;
1077
1078/**
1079 * update_tasks_nodemask - Update the nodemasks of tasks in the cpuset.
1080 * @cs: the cpuset in which each task's mems_allowed mask needs to be changed
1081 *
1082 * Iterate through each task of @cs updating its mems_allowed to the
1083 * effective cpuset's. As this function is called with cpuset_mutex held,
1084 * cpuset membership stays stable.
1085 */
1086static void update_tasks_nodemask(struct cpuset *cs)
1087{
1088 static nodemask_t newmems; /* protected by cpuset_mutex */
1089 struct css_task_iter it;
1090 struct task_struct *task;
1091
1092 cpuset_being_rebound = cs; /* causes mpol_dup() rebind */
1093
1094 guarantee_online_mems(cs, &newmems);
1095
1096 /*
1097 * The mpol_rebind_mm() call takes mmap_sem, which we couldn't
1098 * take while holding tasklist_lock. Forks can happen - the
1099 * mpol_dup() cpuset_being_rebound check will catch such forks,
1100 * and rebind their vma mempolicies too. Because we still hold
1101 * the global cpuset_mutex, we know that no other rebind effort
1102 * will be contending for the global variable cpuset_being_rebound.
1103 * It's ok if we rebind the same mm twice; mpol_rebind_mm()
1104 * is idempotent. Also migrate pages in each mm to new nodes.
1105 */
1106 css_task_iter_start(&cs->css, 0, &it);
1107 while ((task = css_task_iter_next(&it))) {
1108 struct mm_struct *mm;
1109 bool migrate;
1110
1111 cpuset_change_task_nodemask(task, &newmems);
1112
1113 mm = get_task_mm(task);
1114 if (!mm)
1115 continue;
1116
1117 migrate = is_memory_migrate(cs);
1118
1119 mpol_rebind_mm(mm, &cs->mems_allowed);
1120 if (migrate)
1121 cpuset_migrate_mm(mm, &cs->old_mems_allowed, &newmems);
1122 else
1123 mmput(mm);
1124 }
1125 css_task_iter_end(&it);
1126
1127 /*
1128 * All the tasks' nodemasks have been updated, update
1129 * cs->old_mems_allowed.
1130 */
1131 cs->old_mems_allowed = newmems;
1132
1133 /* We're done rebinding vmas to this cpuset's new mems_allowed. */
1134 cpuset_being_rebound = NULL;
1135}
1136
1137/*
1138 * update_nodemasks_hier - Update effective nodemasks and tasks in the subtree
1139 * @cs: the cpuset to consider
1140 * @new_mems: a temp variable for calculating new effective_mems
1141 *
1142 * When configured nodemask is changed, the effective nodemasks of this cpuset
1143 * and all its descendants need to be updated.
1144 *
1145 * On legacy hiearchy, effective_mems will be the same with mems_allowed.
1146 *
1147 * Called with cpuset_mutex held
1148 */
1149static void update_nodemasks_hier(struct cpuset *cs, nodemask_t *new_mems)
1150{
1151 struct cpuset *cp;
1152 struct cgroup_subsys_state *pos_css;
1153
1154 rcu_read_lock();
1155 cpuset_for_each_descendant_pre(cp, pos_css, cs) {
1156 struct cpuset *parent = parent_cs(cp);
1157
1158 nodes_and(*new_mems, cp->mems_allowed, parent->effective_mems);
1159
1160 /*
1161 * If it becomes empty, inherit the effective mask of the
1162 * parent, which is guaranteed to have some MEMs.
1163 */
1164 if (is_in_v2_mode() && nodes_empty(*new_mems))
1165 *new_mems = parent->effective_mems;
1166
1167 /* Skip the whole subtree if the nodemask remains the same. */
1168 if (nodes_equal(*new_mems, cp->effective_mems)) {
1169 pos_css = css_rightmost_descendant(pos_css);
1170 continue;
1171 }
1172
1173 if (!css_tryget_online(&cp->css))
1174 continue;
1175 rcu_read_unlock();
1176
1177 spin_lock_irq(&callback_lock);
1178 cp->effective_mems = *new_mems;
1179 spin_unlock_irq(&callback_lock);
1180
1181 WARN_ON(!is_in_v2_mode() &&
1182 !nodes_equal(cp->mems_allowed, cp->effective_mems));
1183
1184 update_tasks_nodemask(cp);
1185
1186 rcu_read_lock();
1187 css_put(&cp->css);
1188 }
1189 rcu_read_unlock();
1190}
1191
1192/*
1193 * Handle user request to change the 'mems' memory placement
1194 * of a cpuset. Needs to validate the request, update the
1195 * cpusets mems_allowed, and for each task in the cpuset,
1196 * update mems_allowed and rebind task's mempolicy and any vma
1197 * mempolicies and if the cpuset is marked 'memory_migrate',
1198 * migrate the tasks pages to the new memory.
1199 *
1200 * Call with cpuset_mutex held. May take callback_lock during call.
1201 * Will take tasklist_lock, scan tasklist for tasks in cpuset cs,
1202 * lock each such tasks mm->mmap_sem, scan its vma's and rebind
1203 * their mempolicies to the cpusets new mems_allowed.
1204 */
1205static int update_nodemask(struct cpuset *cs, struct cpuset *trialcs,
1206 const char *buf)
1207{
1208 int retval;
1209
1210 /*
1211 * top_cpuset.mems_allowed tracks node_stats[N_MEMORY];
1212 * it's read-only
1213 */
1214 if (cs == &top_cpuset) {
1215 retval = -EACCES;
1216 goto done;
1217 }
1218
1219 /*
1220 * An empty mems_allowed is ok iff there are no tasks in the cpuset.
1221 * Since nodelist_parse() fails on an empty mask, we special case
1222 * that parsing. The validate_change() call ensures that cpusets
1223 * with tasks have memory.
1224 */
1225 if (!*buf) {
1226 nodes_clear(trialcs->mems_allowed);
1227 } else {
1228 retval = nodelist_parse(buf, trialcs->mems_allowed);
1229 if (retval < 0)
1230 goto done;
1231
1232 if (!nodes_subset(trialcs->mems_allowed,
1233 top_cpuset.mems_allowed)) {
1234 retval = -EINVAL;
1235 goto done;
1236 }
1237 }
1238
1239 if (nodes_equal(cs->mems_allowed, trialcs->mems_allowed)) {
1240 retval = 0; /* Too easy - nothing to do */
1241 goto done;
1242 }
1243 retval = validate_change(cs, trialcs);
1244 if (retval < 0)
1245 goto done;
1246
1247 spin_lock_irq(&callback_lock);
1248 cs->mems_allowed = trialcs->mems_allowed;
1249 spin_unlock_irq(&callback_lock);
1250
1251 /* use trialcs->mems_allowed as a temp variable */
1252 update_nodemasks_hier(cs, &trialcs->mems_allowed);
1253done:
1254 return retval;
1255}
1256
1257bool current_cpuset_is_being_rebound(void)
1258{
1259 bool ret;
1260
1261 rcu_read_lock();
1262 ret = task_cs(current) == cpuset_being_rebound;
1263 rcu_read_unlock();
1264
1265 return ret;
1266}
1267
1268static int update_relax_domain_level(struct cpuset *cs, s64 val)
1269{
1270#ifdef CONFIG_SMP
1271 if (val < -1 || val >= sched_domain_level_max)
1272 return -EINVAL;
1273#endif
1274
1275 if (val != cs->relax_domain_level) {
1276 cs->relax_domain_level = val;
1277 if (!cpumask_empty(cs->cpus_allowed) &&
1278 is_sched_load_balance(cs))
1279 rebuild_sched_domains_locked();
1280 }
1281
1282 return 0;
1283}
1284
1285/**
1286 * update_tasks_flags - update the spread flags of tasks in the cpuset.
1287 * @cs: the cpuset in which each task's spread flags needs to be changed
1288 *
1289 * Iterate through each task of @cs updating its spread flags. As this
1290 * function is called with cpuset_mutex held, cpuset membership stays
1291 * stable.
1292 */
1293static void update_tasks_flags(struct cpuset *cs)
1294{
1295 struct css_task_iter it;
1296 struct task_struct *task;
1297
1298 css_task_iter_start(&cs->css, 0, &it);
1299 while ((task = css_task_iter_next(&it)))
1300 cpuset_update_task_spread_flag(cs, task);
1301 css_task_iter_end(&it);
1302}
1303
1304/*
1305 * update_flag - read a 0 or a 1 in a file and update associated flag
1306 * bit: the bit to update (see cpuset_flagbits_t)
1307 * cs: the cpuset to update
1308 * turning_on: whether the flag is being set or cleared
1309 *
1310 * Call with cpuset_mutex held.
1311 */
1312
1313static int update_flag(cpuset_flagbits_t bit, struct cpuset *cs,
1314 int turning_on)
1315{
1316 struct cpuset *trialcs;
1317 int balance_flag_changed;
1318 int spread_flag_changed;
1319 int err;
1320
1321 trialcs = alloc_trial_cpuset(cs);
1322 if (!trialcs)
1323 return -ENOMEM;
1324
1325 if (turning_on)
1326 set_bit(bit, &trialcs->flags);
1327 else
1328 clear_bit(bit, &trialcs->flags);
1329
1330 err = validate_change(cs, trialcs);
1331 if (err < 0)
1332 goto out;
1333
1334 balance_flag_changed = (is_sched_load_balance(cs) !=
1335 is_sched_load_balance(trialcs));
1336
1337 spread_flag_changed = ((is_spread_slab(cs) != is_spread_slab(trialcs))
1338 || (is_spread_page(cs) != is_spread_page(trialcs)));
1339
1340 spin_lock_irq(&callback_lock);
1341 cs->flags = trialcs->flags;
1342 spin_unlock_irq(&callback_lock);
1343
1344 if (!cpumask_empty(trialcs->cpus_allowed) && balance_flag_changed)
1345 rebuild_sched_domains_locked();
1346
1347 if (spread_flag_changed)
1348 update_tasks_flags(cs);
1349out:
1350 free_trial_cpuset(trialcs);
1351 return err;
1352}
1353
1354/*
1355 * Frequency meter - How fast is some event occurring?
1356 *
1357 * These routines manage a digitally filtered, constant time based,
1358 * event frequency meter. There are four routines:
1359 * fmeter_init() - initialize a frequency meter.
1360 * fmeter_markevent() - called each time the event happens.
1361 * fmeter_getrate() - returns the recent rate of such events.
1362 * fmeter_update() - internal routine used to update fmeter.
1363 *
1364 * A common data structure is passed to each of these routines,
1365 * which is used to keep track of the state required to manage the
1366 * frequency meter and its digital filter.
1367 *
1368 * The filter works on the number of events marked per unit time.
1369 * The filter is single-pole low-pass recursive (IIR). The time unit
1370 * is 1 second. Arithmetic is done using 32-bit integers scaled to
1371 * simulate 3 decimal digits of precision (multiplied by 1000).
1372 *
1373 * With an FM_COEF of 933, and a time base of 1 second, the filter
1374 * has a half-life of 10 seconds, meaning that if the events quit
1375 * happening, then the rate returned from the fmeter_getrate()
1376 * will be cut in half each 10 seconds, until it converges to zero.
1377 *
1378 * It is not worth doing a real infinitely recursive filter. If more
1379 * than FM_MAXTICKS ticks have elapsed since the last filter event,
1380 * just compute FM_MAXTICKS ticks worth, by which point the level
1381 * will be stable.
1382 *
1383 * Limit the count of unprocessed events to FM_MAXCNT, so as to avoid
1384 * arithmetic overflow in the fmeter_update() routine.
1385 *
1386 * Given the simple 32 bit integer arithmetic used, this meter works
1387 * best for reporting rates between one per millisecond (msec) and
1388 * one per 32 (approx) seconds. At constant rates faster than one
1389 * per msec it maxes out at values just under 1,000,000. At constant
1390 * rates between one per msec, and one per second it will stabilize
1391 * to a value N*1000, where N is the rate of events per second.
1392 * At constant rates between one per second and one per 32 seconds,
1393 * it will be choppy, moving up on the seconds that have an event,
1394 * and then decaying until the next event. At rates slower than
1395 * about one in 32 seconds, it decays all the way back to zero between
1396 * each event.
1397 */
1398
1399#define FM_COEF 933 /* coefficient for half-life of 10 secs */
1400#define FM_MAXTICKS ((u32)99) /* useless computing more ticks than this */
1401#define FM_MAXCNT 1000000 /* limit cnt to avoid overflow */
1402#define FM_SCALE 1000 /* faux fixed point scale */
1403
1404/* Initialize a frequency meter */
1405static void fmeter_init(struct fmeter *fmp)
1406{
1407 fmp->cnt = 0;
1408 fmp->val = 0;
1409 fmp->time = 0;
1410 spin_lock_init(&fmp->lock);
1411}
1412
1413/* Internal meter update - process cnt events and update value */
1414static void fmeter_update(struct fmeter *fmp)
1415{
1416 time64_t now;
1417 u32 ticks;
1418
1419 now = ktime_get_seconds();
1420 ticks = now - fmp->time;
1421
1422 if (ticks == 0)
1423 return;
1424
1425 ticks = min(FM_MAXTICKS, ticks);
1426 while (ticks-- > 0)
1427 fmp->val = (FM_COEF * fmp->val) / FM_SCALE;
1428 fmp->time = now;
1429
1430 fmp->val += ((FM_SCALE - FM_COEF) * fmp->cnt) / FM_SCALE;
1431 fmp->cnt = 0;
1432}
1433
1434/* Process any previous ticks, then bump cnt by one (times scale). */
1435static void fmeter_markevent(struct fmeter *fmp)
1436{
1437 spin_lock(&fmp->lock);
1438 fmeter_update(fmp);
1439 fmp->cnt = min(FM_MAXCNT, fmp->cnt + FM_SCALE);
1440 spin_unlock(&fmp->lock);
1441}
1442
1443/* Process any previous ticks, then return current value. */
1444static int fmeter_getrate(struct fmeter *fmp)
1445{
1446 int val;
1447
1448 spin_lock(&fmp->lock);
1449 fmeter_update(fmp);
1450 val = fmp->val;
1451 spin_unlock(&fmp->lock);
1452 return val;
1453}
1454
1455static struct cpuset *cpuset_attach_old_cs;
1456
1457/* Called by cgroups to determine if a cpuset is usable; cpuset_mutex held */
1458static int cpuset_can_attach(struct cgroup_taskset *tset)
1459{
1460 struct cgroup_subsys_state *css;
1461 struct cpuset *cs;
1462 struct task_struct *task;
1463 int ret;
1464
1465 /* used later by cpuset_attach() */
1466 cpuset_attach_old_cs = task_cs(cgroup_taskset_first(tset, &css));
1467 cs = css_cs(css);
1468
1469 mutex_lock(&cpuset_mutex);
1470
1471 /* allow moving tasks into an empty cpuset if on default hierarchy */
1472 ret = -ENOSPC;
1473 if (!is_in_v2_mode() &&
1474 (cpumask_empty(cs->cpus_allowed) || nodes_empty(cs->mems_allowed)))
1475 goto out_unlock;
1476
1477 cgroup_taskset_for_each(task, css, tset) {
1478 ret = task_can_attach(task, cs->cpus_allowed);
1479 if (ret)
1480 goto out_unlock;
1481 ret = security_task_setscheduler(task);
1482 if (ret)
1483 goto out_unlock;
1484 }
1485
1486 /*
1487 * Mark attach is in progress. This makes validate_change() fail
1488 * changes which zero cpus/mems_allowed.
1489 */
1490 cs->attach_in_progress++;
1491 ret = 0;
1492out_unlock:
1493 mutex_unlock(&cpuset_mutex);
1494 return ret;
1495}
1496
1497static void cpuset_cancel_attach(struct cgroup_taskset *tset)
1498{
1499 struct cgroup_subsys_state *css;
1500 struct cpuset *cs;
1501
1502 cgroup_taskset_first(tset, &css);
1503 cs = css_cs(css);
1504
1505 mutex_lock(&cpuset_mutex);
1506 css_cs(css)->attach_in_progress--;
1507 mutex_unlock(&cpuset_mutex);
1508}
1509
1510/*
1511 * Protected by cpuset_mutex. cpus_attach is used only by cpuset_attach()
1512 * but we can't allocate it dynamically there. Define it global and
1513 * allocate from cpuset_init().
1514 */
1515static cpumask_var_t cpus_attach;
1516
1517static void cpuset_attach(struct cgroup_taskset *tset)
1518{
1519 /* static buf protected by cpuset_mutex */
1520 static nodemask_t cpuset_attach_nodemask_to;
1521 struct task_struct *task;
1522 struct task_struct *leader;
1523 struct cgroup_subsys_state *css;
1524 struct cpuset *cs;
1525 struct cpuset *oldcs = cpuset_attach_old_cs;
1526
1527 cgroup_taskset_first(tset, &css);
1528 cs = css_cs(css);
1529
1530 mutex_lock(&cpuset_mutex);
1531
1532 /* prepare for attach */
1533 if (cs == &top_cpuset)
1534 cpumask_copy(cpus_attach, cpu_possible_mask);
1535 else
1536 guarantee_online_cpus(cs, cpus_attach);
1537
1538 guarantee_online_mems(cs, &cpuset_attach_nodemask_to);
1539
1540 cgroup_taskset_for_each(task, css, tset) {
1541 /*
1542 * can_attach beforehand should guarantee that this doesn't
1543 * fail. TODO: have a better way to handle failure here
1544 */
1545 WARN_ON_ONCE(set_cpus_allowed_ptr(task, cpus_attach));
1546
1547 cpuset_change_task_nodemask(task, &cpuset_attach_nodemask_to);
1548 cpuset_update_task_spread_flag(cs, task);
1549 }
1550
1551 /*
1552 * Change mm for all threadgroup leaders. This is expensive and may
1553 * sleep and should be moved outside migration path proper.
1554 */
1555 cpuset_attach_nodemask_to = cs->effective_mems;
1556 cgroup_taskset_for_each_leader(leader, css, tset) {
1557 struct mm_struct *mm = get_task_mm(leader);
1558
1559 if (mm) {
1560 mpol_rebind_mm(mm, &cpuset_attach_nodemask_to);
1561
1562 /*
1563 * old_mems_allowed is the same with mems_allowed
1564 * here, except if this task is being moved
1565 * automatically due to hotplug. In that case
1566 * @mems_allowed has been updated and is empty, so
1567 * @old_mems_allowed is the right nodesets that we
1568 * migrate mm from.
1569 */
1570 if (is_memory_migrate(cs))
1571 cpuset_migrate_mm(mm, &oldcs->old_mems_allowed,
1572 &cpuset_attach_nodemask_to);
1573 else
1574 mmput(mm);
1575 }
1576 }
1577
1578 cs->old_mems_allowed = cpuset_attach_nodemask_to;
1579
1580 cs->attach_in_progress--;
1581 if (!cs->attach_in_progress)
1582 wake_up(&cpuset_attach_wq);
1583
1584 mutex_unlock(&cpuset_mutex);
1585}
1586
1587/* The various types of files and directories in a cpuset file system */
1588
1589typedef enum {
1590 FILE_MEMORY_MIGRATE,
1591 FILE_CPULIST,
1592 FILE_MEMLIST,
1593 FILE_EFFECTIVE_CPULIST,
1594 FILE_EFFECTIVE_MEMLIST,
1595 FILE_CPU_EXCLUSIVE,
1596 FILE_MEM_EXCLUSIVE,
1597 FILE_MEM_HARDWALL,
1598 FILE_SCHED_LOAD_BALANCE,
1599 FILE_SCHED_RELAX_DOMAIN_LEVEL,
1600 FILE_MEMORY_PRESSURE_ENABLED,
1601 FILE_MEMORY_PRESSURE,
1602 FILE_SPREAD_PAGE,
1603 FILE_SPREAD_SLAB,
1604} cpuset_filetype_t;
1605
1606static int cpuset_write_u64(struct cgroup_subsys_state *css, struct cftype *cft,
1607 u64 val)
1608{
1609 struct cpuset *cs = css_cs(css);
1610 cpuset_filetype_t type = cft->private;
1611 int retval = 0;
1612
1613 mutex_lock(&cpuset_mutex);
1614 if (!is_cpuset_online(cs)) {
1615 retval = -ENODEV;
1616 goto out_unlock;
1617 }
1618
1619 switch (type) {
1620 case FILE_CPU_EXCLUSIVE:
1621 retval = update_flag(CS_CPU_EXCLUSIVE, cs, val);
1622 break;
1623 case FILE_MEM_EXCLUSIVE:
1624 retval = update_flag(CS_MEM_EXCLUSIVE, cs, val);
1625 break;
1626 case FILE_MEM_HARDWALL:
1627 retval = update_flag(CS_MEM_HARDWALL, cs, val);
1628 break;
1629 case FILE_SCHED_LOAD_BALANCE:
1630 retval = update_flag(CS_SCHED_LOAD_BALANCE, cs, val);
1631 break;
1632 case FILE_MEMORY_MIGRATE:
1633 retval = update_flag(CS_MEMORY_MIGRATE, cs, val);
1634 break;
1635 case FILE_MEMORY_PRESSURE_ENABLED:
1636 cpuset_memory_pressure_enabled = !!val;
1637 break;
1638 case FILE_SPREAD_PAGE:
1639 retval = update_flag(CS_SPREAD_PAGE, cs, val);
1640 break;
1641 case FILE_SPREAD_SLAB:
1642 retval = update_flag(CS_SPREAD_SLAB, cs, val);
1643 break;
1644 default:
1645 retval = -EINVAL;
1646 break;
1647 }
1648out_unlock:
1649 mutex_unlock(&cpuset_mutex);
1650 return retval;
1651}
1652
1653static int cpuset_write_s64(struct cgroup_subsys_state *css, struct cftype *cft,
1654 s64 val)
1655{
1656 struct cpuset *cs = css_cs(css);
1657 cpuset_filetype_t type = cft->private;
1658 int retval = -ENODEV;
1659
1660 mutex_lock(&cpuset_mutex);
1661 if (!is_cpuset_online(cs))
1662 goto out_unlock;
1663
1664 switch (type) {
1665 case FILE_SCHED_RELAX_DOMAIN_LEVEL:
1666 retval = update_relax_domain_level(cs, val);
1667 break;
1668 default:
1669 retval = -EINVAL;
1670 break;
1671 }
1672out_unlock:
1673 mutex_unlock(&cpuset_mutex);
1674 return retval;
1675}
1676
1677/*
1678 * Common handling for a write to a "cpus" or "mems" file.
1679 */
1680static ssize_t cpuset_write_resmask(struct kernfs_open_file *of,
1681 char *buf, size_t nbytes, loff_t off)
1682{
1683 struct cpuset *cs = css_cs(of_css(of));
1684 struct cpuset *trialcs;
1685 int retval = -ENODEV;
1686
1687 buf = strstrip(buf);
1688
1689 /*
1690 * CPU or memory hotunplug may leave @cs w/o any execution
1691 * resources, in which case the hotplug code asynchronously updates
1692 * configuration and transfers all tasks to the nearest ancestor
1693 * which can execute.
1694 *
1695 * As writes to "cpus" or "mems" may restore @cs's execution
1696 * resources, wait for the previously scheduled operations before
1697 * proceeding, so that we don't end up keep removing tasks added
1698 * after execution capability is restored.
1699 *
1700 * cpuset_hotplug_work calls back into cgroup core via
1701 * cgroup_transfer_tasks() and waiting for it from a cgroupfs
1702 * operation like this one can lead to a deadlock through kernfs
1703 * active_ref protection. Let's break the protection. Losing the
1704 * protection is okay as we check whether @cs is online after
1705 * grabbing cpuset_mutex anyway. This only happens on the legacy
1706 * hierarchies.
1707 */
1708 css_get(&cs->css);
1709 kernfs_break_active_protection(of->kn);
1710 flush_work(&cpuset_hotplug_work);
1711
1712 mutex_lock(&cpuset_mutex);
1713 if (!is_cpuset_online(cs))
1714 goto out_unlock;
1715
1716 trialcs = alloc_trial_cpuset(cs);
1717 if (!trialcs) {
1718 retval = -ENOMEM;
1719 goto out_unlock;
1720 }
1721
1722 switch (of_cft(of)->private) {
1723 case FILE_CPULIST:
1724 retval = update_cpumask(cs, trialcs, buf);
1725 break;
1726 case FILE_MEMLIST:
1727 retval = update_nodemask(cs, trialcs, buf);
1728 break;
1729 default:
1730 retval = -EINVAL;
1731 break;
1732 }
1733
1734 free_trial_cpuset(trialcs);
1735out_unlock:
1736 mutex_unlock(&cpuset_mutex);
1737 kernfs_unbreak_active_protection(of->kn);
1738 css_put(&cs->css);
1739 flush_workqueue(cpuset_migrate_mm_wq);
1740 return retval ?: nbytes;
1741}
1742
1743/*
1744 * These ascii lists should be read in a single call, by using a user
1745 * buffer large enough to hold the entire map. If read in smaller
1746 * chunks, there is no guarantee of atomicity. Since the display format
1747 * used, list of ranges of sequential numbers, is variable length,
1748 * and since these maps can change value dynamically, one could read
1749 * gibberish by doing partial reads while a list was changing.
1750 */
1751static int cpuset_common_seq_show(struct seq_file *sf, void *v)
1752{
1753 struct cpuset *cs = css_cs(seq_css(sf));
1754 cpuset_filetype_t type = seq_cft(sf)->private;
1755 int ret = 0;
1756
1757 spin_lock_irq(&callback_lock);
1758
1759 switch (type) {
1760 case FILE_CPULIST:
1761 seq_printf(sf, "%*pbl\n", cpumask_pr_args(cs->cpus_allowed));
1762 break;
1763 case FILE_MEMLIST:
1764 seq_printf(sf, "%*pbl\n", nodemask_pr_args(&cs->mems_allowed));
1765 break;
1766 case FILE_EFFECTIVE_CPULIST:
1767 seq_printf(sf, "%*pbl\n", cpumask_pr_args(cs->effective_cpus));
1768 break;
1769 case FILE_EFFECTIVE_MEMLIST:
1770 seq_printf(sf, "%*pbl\n", nodemask_pr_args(&cs->effective_mems));
1771 break;
1772 default:
1773 ret = -EINVAL;
1774 }
1775
1776 spin_unlock_irq(&callback_lock);
1777 return ret;
1778}
1779
1780static u64 cpuset_read_u64(struct cgroup_subsys_state *css, struct cftype *cft)
1781{
1782 struct cpuset *cs = css_cs(css);
1783 cpuset_filetype_t type = cft->private;
1784 switch (type) {
1785 case FILE_CPU_EXCLUSIVE:
1786 return is_cpu_exclusive(cs);
1787 case FILE_MEM_EXCLUSIVE:
1788 return is_mem_exclusive(cs);
1789 case FILE_MEM_HARDWALL:
1790 return is_mem_hardwall(cs);
1791 case FILE_SCHED_LOAD_BALANCE:
1792 return is_sched_load_balance(cs);
1793 case FILE_MEMORY_MIGRATE:
1794 return is_memory_migrate(cs);
1795 case FILE_MEMORY_PRESSURE_ENABLED:
1796 return cpuset_memory_pressure_enabled;
1797 case FILE_MEMORY_PRESSURE:
1798 return fmeter_getrate(&cs->fmeter);
1799 case FILE_SPREAD_PAGE:
1800 return is_spread_page(cs);
1801 case FILE_SPREAD_SLAB:
1802 return is_spread_slab(cs);
1803 default:
1804 BUG();
1805 }
1806
1807 /* Unreachable but makes gcc happy */
1808 return 0;
1809}
1810
1811static s64 cpuset_read_s64(struct cgroup_subsys_state *css, struct cftype *cft)
1812{
1813 struct cpuset *cs = css_cs(css);
1814 cpuset_filetype_t type = cft->private;
1815 switch (type) {
1816 case FILE_SCHED_RELAX_DOMAIN_LEVEL:
1817 return cs->relax_domain_level;
1818 default:
1819 BUG();
1820 }
1821
1822 /* Unrechable but makes gcc happy */
1823 return 0;
1824}
1825
1826
1827/*
1828 * for the common functions, 'private' gives the type of file
1829 */
1830
1831static struct cftype files[] = {
1832 {
1833 .name = "cpus",
1834 .seq_show = cpuset_common_seq_show,
1835 .write = cpuset_write_resmask,
1836 .max_write_len = (100U + 6 * NR_CPUS),
1837 .private = FILE_CPULIST,
1838 },
1839
1840 {
1841 .name = "mems",
1842 .seq_show = cpuset_common_seq_show,
1843 .write = cpuset_write_resmask,
1844 .max_write_len = (100U + 6 * MAX_NUMNODES),
1845 .private = FILE_MEMLIST,
1846 },
1847
1848 {
1849 .name = "effective_cpus",
1850 .seq_show = cpuset_common_seq_show,
1851 .private = FILE_EFFECTIVE_CPULIST,
1852 },
1853
1854 {
1855 .name = "effective_mems",
1856 .seq_show = cpuset_common_seq_show,
1857 .private = FILE_EFFECTIVE_MEMLIST,
1858 },
1859
1860 {
1861 .name = "cpu_exclusive",
1862 .read_u64 = cpuset_read_u64,
1863 .write_u64 = cpuset_write_u64,
1864 .private = FILE_CPU_EXCLUSIVE,
1865 },
1866
1867 {
1868 .name = "mem_exclusive",
1869 .read_u64 = cpuset_read_u64,
1870 .write_u64 = cpuset_write_u64,
1871 .private = FILE_MEM_EXCLUSIVE,
1872 },
1873
1874 {
1875 .name = "mem_hardwall",
1876 .read_u64 = cpuset_read_u64,
1877 .write_u64 = cpuset_write_u64,
1878 .private = FILE_MEM_HARDWALL,
1879 },
1880
1881 {
1882 .name = "sched_load_balance",
1883 .read_u64 = cpuset_read_u64,
1884 .write_u64 = cpuset_write_u64,
1885 .private = FILE_SCHED_LOAD_BALANCE,
1886 },
1887
1888 {
1889 .name = "sched_relax_domain_level",
1890 .read_s64 = cpuset_read_s64,
1891 .write_s64 = cpuset_write_s64,
1892 .private = FILE_SCHED_RELAX_DOMAIN_LEVEL,
1893 },
1894
1895 {
1896 .name = "memory_migrate",
1897 .read_u64 = cpuset_read_u64,
1898 .write_u64 = cpuset_write_u64,
1899 .private = FILE_MEMORY_MIGRATE,
1900 },
1901
1902 {
1903 .name = "memory_pressure",
1904 .read_u64 = cpuset_read_u64,
1905 .private = FILE_MEMORY_PRESSURE,
1906 },
1907
1908 {
1909 .name = "memory_spread_page",
1910 .read_u64 = cpuset_read_u64,
1911 .write_u64 = cpuset_write_u64,
1912 .private = FILE_SPREAD_PAGE,
1913 },
1914
1915 {
1916 .name = "memory_spread_slab",
1917 .read_u64 = cpuset_read_u64,
1918 .write_u64 = cpuset_write_u64,
1919 .private = FILE_SPREAD_SLAB,
1920 },
1921
1922 {
1923 .name = "memory_pressure_enabled",
1924 .flags = CFTYPE_ONLY_ON_ROOT,
1925 .read_u64 = cpuset_read_u64,
1926 .write_u64 = cpuset_write_u64,
1927 .private = FILE_MEMORY_PRESSURE_ENABLED,
1928 },
1929
1930 { } /* terminate */
1931};
1932
1933/*
1934 * cpuset_css_alloc - allocate a cpuset css
1935 * cgrp: control group that the new cpuset will be part of
1936 */
1937
1938static struct cgroup_subsys_state *
1939cpuset_css_alloc(struct cgroup_subsys_state *parent_css)
1940{
1941 struct cpuset *cs;
1942
1943 if (!parent_css)
1944 return &top_cpuset.css;
1945
1946 cs = kzalloc(sizeof(*cs), GFP_KERNEL);
1947 if (!cs)
1948 return ERR_PTR(-ENOMEM);
1949 if (!alloc_cpumask_var(&cs->cpus_allowed, GFP_KERNEL))
1950 goto free_cs;
1951 if (!alloc_cpumask_var(&cs->effective_cpus, GFP_KERNEL))
1952 goto free_cpus;
1953
1954 set_bit(CS_SCHED_LOAD_BALANCE, &cs->flags);
1955 cpumask_clear(cs->cpus_allowed);
1956 nodes_clear(cs->mems_allowed);
1957 cpumask_clear(cs->effective_cpus);
1958 nodes_clear(cs->effective_mems);
1959 fmeter_init(&cs->fmeter);
1960 cs->relax_domain_level = -1;
1961
1962 return &cs->css;
1963
1964free_cpus:
1965 free_cpumask_var(cs->cpus_allowed);
1966free_cs:
1967 kfree(cs);
1968 return ERR_PTR(-ENOMEM);
1969}
1970
1971static int cpuset_css_online(struct cgroup_subsys_state *css)
1972{
1973 struct cpuset *cs = css_cs(css);
1974 struct cpuset *parent = parent_cs(cs);
1975 struct cpuset *tmp_cs;
1976 struct cgroup_subsys_state *pos_css;
1977
1978 if (!parent)
1979 return 0;
1980
1981 mutex_lock(&cpuset_mutex);
1982
1983 set_bit(CS_ONLINE, &cs->flags);
1984 if (is_spread_page(parent))
1985 set_bit(CS_SPREAD_PAGE, &cs->flags);
1986 if (is_spread_slab(parent))
1987 set_bit(CS_SPREAD_SLAB, &cs->flags);
1988
1989 cpuset_inc();
1990
1991 spin_lock_irq(&callback_lock);
1992 if (is_in_v2_mode()) {
1993 cpumask_copy(cs->effective_cpus, parent->effective_cpus);
1994 cs->effective_mems = parent->effective_mems;
1995 }
1996 spin_unlock_irq(&callback_lock);
1997
1998 if (!test_bit(CGRP_CPUSET_CLONE_CHILDREN, &css->cgroup->flags))
1999 goto out_unlock;
2000
2001 /*
2002 * Clone @parent's configuration if CGRP_CPUSET_CLONE_CHILDREN is
2003 * set. This flag handling is implemented in cgroup core for
2004 * histrical reasons - the flag may be specified during mount.
2005 *
2006 * Currently, if any sibling cpusets have exclusive cpus or mem, we
2007 * refuse to clone the configuration - thereby refusing the task to
2008 * be entered, and as a result refusing the sys_unshare() or
2009 * clone() which initiated it. If this becomes a problem for some
2010 * users who wish to allow that scenario, then this could be
2011 * changed to grant parent->cpus_allowed-sibling_cpus_exclusive
2012 * (and likewise for mems) to the new cgroup.
2013 */
2014 rcu_read_lock();
2015 cpuset_for_each_child(tmp_cs, pos_css, parent) {
2016 if (is_mem_exclusive(tmp_cs) || is_cpu_exclusive(tmp_cs)) {
2017 rcu_read_unlock();
2018 goto out_unlock;
2019 }
2020 }
2021 rcu_read_unlock();
2022
2023 spin_lock_irq(&callback_lock);
2024 cs->mems_allowed = parent->mems_allowed;
2025 cs->effective_mems = parent->mems_allowed;
2026 cpumask_copy(cs->cpus_allowed, parent->cpus_allowed);
2027 cpumask_copy(cs->effective_cpus, parent->cpus_allowed);
2028 spin_unlock_irq(&callback_lock);
2029out_unlock:
2030 mutex_unlock(&cpuset_mutex);
2031 return 0;
2032}
2033
2034/*
2035 * If the cpuset being removed has its flag 'sched_load_balance'
2036 * enabled, then simulate turning sched_load_balance off, which
2037 * will call rebuild_sched_domains_locked().
2038 */
2039
2040static void cpuset_css_offline(struct cgroup_subsys_state *css)
2041{
2042 struct cpuset *cs = css_cs(css);
2043
2044 mutex_lock(&cpuset_mutex);
2045
2046 if (is_sched_load_balance(cs))
2047 update_flag(CS_SCHED_LOAD_BALANCE, cs, 0);
2048
2049 cpuset_dec();
2050 clear_bit(CS_ONLINE, &cs->flags);
2051
2052 mutex_unlock(&cpuset_mutex);
2053}
2054
2055static void cpuset_css_free(struct cgroup_subsys_state *css)
2056{
2057 struct cpuset *cs = css_cs(css);
2058
2059 free_cpumask_var(cs->effective_cpus);
2060 free_cpumask_var(cs->cpus_allowed);
2061 kfree(cs);
2062}
2063
2064static void cpuset_bind(struct cgroup_subsys_state *root_css)
2065{
2066 mutex_lock(&cpuset_mutex);
2067 spin_lock_irq(&callback_lock);
2068
2069 if (is_in_v2_mode()) {
2070 cpumask_copy(top_cpuset.cpus_allowed, cpu_possible_mask);
2071 top_cpuset.mems_allowed = node_possible_map;
2072 } else {
2073 cpumask_copy(top_cpuset.cpus_allowed,
2074 top_cpuset.effective_cpus);
2075 top_cpuset.mems_allowed = top_cpuset.effective_mems;
2076 }
2077
2078 spin_unlock_irq(&callback_lock);
2079 mutex_unlock(&cpuset_mutex);
2080}
2081
2082/*
2083 * Make sure the new task conform to the current state of its parent,
2084 * which could have been changed by cpuset just after it inherits the
2085 * state from the parent and before it sits on the cgroup's task list.
2086 */
2087static void cpuset_fork(struct task_struct *task)
2088{
2089 if (task_css_is_root(task, cpuset_cgrp_id))
2090 return;
2091
2092 set_cpus_allowed_ptr(task, ¤t->cpus_allowed);
2093 task->mems_allowed = current->mems_allowed;
2094}
2095
2096struct cgroup_subsys cpuset_cgrp_subsys = {
2097 .css_alloc = cpuset_css_alloc,
2098 .css_online = cpuset_css_online,
2099 .css_offline = cpuset_css_offline,
2100 .css_free = cpuset_css_free,
2101 .can_attach = cpuset_can_attach,
2102 .cancel_attach = cpuset_cancel_attach,
2103 .attach = cpuset_attach,
2104 .post_attach = cpuset_post_attach,
2105 .bind = cpuset_bind,
2106 .fork = cpuset_fork,
2107 .legacy_cftypes = files,
2108 .early_init = true,
2109};
2110
2111/**
2112 * cpuset_init - initialize cpusets at system boot
2113 *
2114 * Description: Initialize top_cpuset and the cpuset internal file system,
2115 **/
2116
2117int __init cpuset_init(void)
2118{
2119 int err = 0;
2120
2121 BUG_ON(!alloc_cpumask_var(&top_cpuset.cpus_allowed, GFP_KERNEL));
2122 BUG_ON(!alloc_cpumask_var(&top_cpuset.effective_cpus, GFP_KERNEL));
2123
2124 cpumask_setall(top_cpuset.cpus_allowed);
2125 nodes_setall(top_cpuset.mems_allowed);
2126 cpumask_setall(top_cpuset.effective_cpus);
2127 nodes_setall(top_cpuset.effective_mems);
2128
2129 fmeter_init(&top_cpuset.fmeter);
2130 set_bit(CS_SCHED_LOAD_BALANCE, &top_cpuset.flags);
2131 top_cpuset.relax_domain_level = -1;
2132
2133 err = register_filesystem(&cpuset_fs_type);
2134 if (err < 0)
2135 return err;
2136
2137 BUG_ON(!alloc_cpumask_var(&cpus_attach, GFP_KERNEL));
2138
2139 return 0;
2140}
2141
2142/*
2143 * If CPU and/or memory hotplug handlers, below, unplug any CPUs
2144 * or memory nodes, we need to walk over the cpuset hierarchy,
2145 * removing that CPU or node from all cpusets. If this removes the
2146 * last CPU or node from a cpuset, then move the tasks in the empty
2147 * cpuset to its next-highest non-empty parent.
2148 */
2149static void remove_tasks_in_empty_cpuset(struct cpuset *cs)
2150{
2151 struct cpuset *parent;
2152
2153 /*
2154 * Find its next-highest non-empty parent, (top cpuset
2155 * has online cpus, so can't be empty).
2156 */
2157 parent = parent_cs(cs);
2158 while (cpumask_empty(parent->cpus_allowed) ||
2159 nodes_empty(parent->mems_allowed))
2160 parent = parent_cs(parent);
2161
2162 if (cgroup_transfer_tasks(parent->css.cgroup, cs->css.cgroup)) {
2163 pr_err("cpuset: failed to transfer tasks out of empty cpuset ");
2164 pr_cont_cgroup_name(cs->css.cgroup);
2165 pr_cont("\n");
2166 }
2167}
2168
2169static void
2170hotplug_update_tasks_legacy(struct cpuset *cs,
2171 struct cpumask *new_cpus, nodemask_t *new_mems,
2172 bool cpus_updated, bool mems_updated)
2173{
2174 bool is_empty;
2175
2176 spin_lock_irq(&callback_lock);
2177 cpumask_copy(cs->cpus_allowed, new_cpus);
2178 cpumask_copy(cs->effective_cpus, new_cpus);
2179 cs->mems_allowed = *new_mems;
2180 cs->effective_mems = *new_mems;
2181 spin_unlock_irq(&callback_lock);
2182
2183 /*
2184 * Don't call update_tasks_cpumask() if the cpuset becomes empty,
2185 * as the tasks will be migratecd to an ancestor.
2186 */
2187 if (cpus_updated && !cpumask_empty(cs->cpus_allowed))
2188 update_tasks_cpumask(cs);
2189 if (mems_updated && !nodes_empty(cs->mems_allowed))
2190 update_tasks_nodemask(cs);
2191
2192 is_empty = cpumask_empty(cs->cpus_allowed) ||
2193 nodes_empty(cs->mems_allowed);
2194
2195 mutex_unlock(&cpuset_mutex);
2196
2197 /*
2198 * Move tasks to the nearest ancestor with execution resources,
2199 * This is full cgroup operation which will also call back into
2200 * cpuset. Should be done outside any lock.
2201 */
2202 if (is_empty)
2203 remove_tasks_in_empty_cpuset(cs);
2204
2205 mutex_lock(&cpuset_mutex);
2206}
2207
2208static void
2209hotplug_update_tasks(struct cpuset *cs,
2210 struct cpumask *new_cpus, nodemask_t *new_mems,
2211 bool cpus_updated, bool mems_updated)
2212{
2213 if (cpumask_empty(new_cpus))
2214 cpumask_copy(new_cpus, parent_cs(cs)->effective_cpus);
2215 if (nodes_empty(*new_mems))
2216 *new_mems = parent_cs(cs)->effective_mems;
2217
2218 spin_lock_irq(&callback_lock);
2219 cpumask_copy(cs->effective_cpus, new_cpus);
2220 cs->effective_mems = *new_mems;
2221 spin_unlock_irq(&callback_lock);
2222
2223 if (cpus_updated)
2224 update_tasks_cpumask(cs);
2225 if (mems_updated)
2226 update_tasks_nodemask(cs);
2227}
2228
2229/**
2230 * cpuset_hotplug_update_tasks - update tasks in a cpuset for hotunplug
2231 * @cs: cpuset in interest
2232 *
2233 * Compare @cs's cpu and mem masks against top_cpuset and if some have gone
2234 * offline, update @cs accordingly. If @cs ends up with no CPU or memory,
2235 * all its tasks are moved to the nearest ancestor with both resources.
2236 */
2237static void cpuset_hotplug_update_tasks(struct cpuset *cs)
2238{
2239 static cpumask_t new_cpus;
2240 static nodemask_t new_mems;
2241 bool cpus_updated;
2242 bool mems_updated;
2243retry:
2244 wait_event(cpuset_attach_wq, cs->attach_in_progress == 0);
2245
2246 mutex_lock(&cpuset_mutex);
2247
2248 /*
2249 * We have raced with task attaching. We wait until attaching
2250 * is finished, so we won't attach a task to an empty cpuset.
2251 */
2252 if (cs->attach_in_progress) {
2253 mutex_unlock(&cpuset_mutex);
2254 goto retry;
2255 }
2256
2257 cpumask_and(&new_cpus, cs->cpus_allowed, parent_cs(cs)->effective_cpus);
2258 nodes_and(new_mems, cs->mems_allowed, parent_cs(cs)->effective_mems);
2259
2260 cpus_updated = !cpumask_equal(&new_cpus, cs->effective_cpus);
2261 mems_updated = !nodes_equal(new_mems, cs->effective_mems);
2262
2263 if (is_in_v2_mode())
2264 hotplug_update_tasks(cs, &new_cpus, &new_mems,
2265 cpus_updated, mems_updated);
2266 else
2267 hotplug_update_tasks_legacy(cs, &new_cpus, &new_mems,
2268 cpus_updated, mems_updated);
2269
2270 mutex_unlock(&cpuset_mutex);
2271}
2272
2273static bool force_rebuild;
2274
2275void cpuset_force_rebuild(void)
2276{
2277 force_rebuild = true;
2278}
2279
2280/**
2281 * cpuset_hotplug_workfn - handle CPU/memory hotunplug for a cpuset
2282 *
2283 * This function is called after either CPU or memory configuration has
2284 * changed and updates cpuset accordingly. The top_cpuset is always
2285 * synchronized to cpu_active_mask and N_MEMORY, which is necessary in
2286 * order to make cpusets transparent (of no affect) on systems that are
2287 * actively using CPU hotplug but making no active use of cpusets.
2288 *
2289 * Non-root cpusets are only affected by offlining. If any CPUs or memory
2290 * nodes have been taken down, cpuset_hotplug_update_tasks() is invoked on
2291 * all descendants.
2292 *
2293 * Note that CPU offlining during suspend is ignored. We don't modify
2294 * cpusets across suspend/resume cycles at all.
2295 */
2296static void cpuset_hotplug_workfn(struct work_struct *work)
2297{
2298 static cpumask_t new_cpus;
2299 static nodemask_t new_mems;
2300 bool cpus_updated, mems_updated;
2301 bool on_dfl = is_in_v2_mode();
2302
2303 mutex_lock(&cpuset_mutex);
2304
2305 /* fetch the available cpus/mems and find out which changed how */
2306 cpumask_copy(&new_cpus, cpu_active_mask);
2307 new_mems = node_states[N_MEMORY];
2308
2309 cpus_updated = !cpumask_equal(top_cpuset.effective_cpus, &new_cpus);
2310 mems_updated = !nodes_equal(top_cpuset.effective_mems, new_mems);
2311
2312 /* synchronize cpus_allowed to cpu_active_mask */
2313 if (cpus_updated) {
2314 spin_lock_irq(&callback_lock);
2315 if (!on_dfl)
2316 cpumask_copy(top_cpuset.cpus_allowed, &new_cpus);
2317 cpumask_copy(top_cpuset.effective_cpus, &new_cpus);
2318 spin_unlock_irq(&callback_lock);
2319 /* we don't mess with cpumasks of tasks in top_cpuset */
2320 }
2321
2322 /* synchronize mems_allowed to N_MEMORY */
2323 if (mems_updated) {
2324 spin_lock_irq(&callback_lock);
2325 if (!on_dfl)
2326 top_cpuset.mems_allowed = new_mems;
2327 top_cpuset.effective_mems = new_mems;
2328 spin_unlock_irq(&callback_lock);
2329 update_tasks_nodemask(&top_cpuset);
2330 }
2331
2332 mutex_unlock(&cpuset_mutex);
2333
2334 /* if cpus or mems changed, we need to propagate to descendants */
2335 if (cpus_updated || mems_updated) {
2336 struct cpuset *cs;
2337 struct cgroup_subsys_state *pos_css;
2338
2339 rcu_read_lock();
2340 cpuset_for_each_descendant_pre(cs, pos_css, &top_cpuset) {
2341 if (cs == &top_cpuset || !css_tryget_online(&cs->css))
2342 continue;
2343 rcu_read_unlock();
2344
2345 cpuset_hotplug_update_tasks(cs);
2346
2347 rcu_read_lock();
2348 css_put(&cs->css);
2349 }
2350 rcu_read_unlock();
2351 }
2352
2353 /* rebuild sched domains if cpus_allowed has changed */
2354 if (cpus_updated || force_rebuild) {
2355 force_rebuild = false;
2356 rebuild_sched_domains();
2357 }
2358}
2359
2360void cpuset_update_active_cpus(void)
2361{
2362 /*
2363 * We're inside cpu hotplug critical region which usually nests
2364 * inside cgroup synchronization. Bounce actual hotplug processing
2365 * to a work item to avoid reverse locking order.
2366 */
2367 schedule_work(&cpuset_hotplug_work);
2368}
2369
2370void cpuset_wait_for_hotplug(void)
2371{
2372 flush_work(&cpuset_hotplug_work);
2373}
2374
2375/*
2376 * Keep top_cpuset.mems_allowed tracking node_states[N_MEMORY].
2377 * Call this routine anytime after node_states[N_MEMORY] changes.
2378 * See cpuset_update_active_cpus() for CPU hotplug handling.
2379 */
2380static int cpuset_track_online_nodes(struct notifier_block *self,
2381 unsigned long action, void *arg)
2382{
2383 schedule_work(&cpuset_hotplug_work);
2384 return NOTIFY_OK;
2385}
2386
2387static struct notifier_block cpuset_track_online_nodes_nb = {
2388 .notifier_call = cpuset_track_online_nodes,
2389 .priority = 10, /* ??! */
2390};
2391
2392/**
2393 * cpuset_init_smp - initialize cpus_allowed
2394 *
2395 * Description: Finish top cpuset after cpu, node maps are initialized
2396 */
2397void __init cpuset_init_smp(void)
2398{
2399 cpumask_copy(top_cpuset.cpus_allowed, cpu_active_mask);
2400 top_cpuset.mems_allowed = node_states[N_MEMORY];
2401 top_cpuset.old_mems_allowed = top_cpuset.mems_allowed;
2402
2403 cpumask_copy(top_cpuset.effective_cpus, cpu_active_mask);
2404 top_cpuset.effective_mems = node_states[N_MEMORY];
2405
2406 register_hotmemory_notifier(&cpuset_track_online_nodes_nb);
2407
2408 cpuset_migrate_mm_wq = alloc_ordered_workqueue("cpuset_migrate_mm", 0);
2409 BUG_ON(!cpuset_migrate_mm_wq);
2410}
2411
2412/**
2413 * cpuset_cpus_allowed - return cpus_allowed mask from a tasks cpuset.
2414 * @tsk: pointer to task_struct from which to obtain cpuset->cpus_allowed.
2415 * @pmask: pointer to struct cpumask variable to receive cpus_allowed set.
2416 *
2417 * Description: Returns the cpumask_var_t cpus_allowed of the cpuset
2418 * attached to the specified @tsk. Guaranteed to return some non-empty
2419 * subset of cpu_online_mask, even if this means going outside the
2420 * tasks cpuset.
2421 **/
2422
2423void cpuset_cpus_allowed(struct task_struct *tsk, struct cpumask *pmask)
2424{
2425 unsigned long flags;
2426
2427 spin_lock_irqsave(&callback_lock, flags);
2428 rcu_read_lock();
2429 guarantee_online_cpus(task_cs(tsk), pmask);
2430 rcu_read_unlock();
2431 spin_unlock_irqrestore(&callback_lock, flags);
2432}
2433
2434void cpuset_cpus_allowed_fallback(struct task_struct *tsk)
2435{
2436 rcu_read_lock();
2437 do_set_cpus_allowed(tsk, task_cs(tsk)->effective_cpus);
2438 rcu_read_unlock();
2439
2440 /*
2441 * We own tsk->cpus_allowed, nobody can change it under us.
2442 *
2443 * But we used cs && cs->cpus_allowed lockless and thus can
2444 * race with cgroup_attach_task() or update_cpumask() and get
2445 * the wrong tsk->cpus_allowed. However, both cases imply the
2446 * subsequent cpuset_change_cpumask()->set_cpus_allowed_ptr()
2447 * which takes task_rq_lock().
2448 *
2449 * If we are called after it dropped the lock we must see all
2450 * changes in tsk_cs()->cpus_allowed. Otherwise we can temporary
2451 * set any mask even if it is not right from task_cs() pov,
2452 * the pending set_cpus_allowed_ptr() will fix things.
2453 *
2454 * select_fallback_rq() will fix things ups and set cpu_possible_mask
2455 * if required.
2456 */
2457}
2458
2459void __init cpuset_init_current_mems_allowed(void)
2460{
2461 nodes_setall(current->mems_allowed);
2462}
2463
2464/**
2465 * cpuset_mems_allowed - return mems_allowed mask from a tasks cpuset.
2466 * @tsk: pointer to task_struct from which to obtain cpuset->mems_allowed.
2467 *
2468 * Description: Returns the nodemask_t mems_allowed of the cpuset
2469 * attached to the specified @tsk. Guaranteed to return some non-empty
2470 * subset of node_states[N_MEMORY], even if this means going outside the
2471 * tasks cpuset.
2472 **/
2473
2474nodemask_t cpuset_mems_allowed(struct task_struct *tsk)
2475{
2476 nodemask_t mask;
2477 unsigned long flags;
2478
2479 spin_lock_irqsave(&callback_lock, flags);
2480 rcu_read_lock();
2481 guarantee_online_mems(task_cs(tsk), &mask);
2482 rcu_read_unlock();
2483 spin_unlock_irqrestore(&callback_lock, flags);
2484
2485 return mask;
2486}
2487
2488/**
2489 * cpuset_nodemask_valid_mems_allowed - check nodemask vs. curremt mems_allowed
2490 * @nodemask: the nodemask to be checked
2491 *
2492 * Are any of the nodes in the nodemask allowed in current->mems_allowed?
2493 */
2494int cpuset_nodemask_valid_mems_allowed(nodemask_t *nodemask)
2495{
2496 return nodes_intersects(*nodemask, current->mems_allowed);
2497}
2498
2499/*
2500 * nearest_hardwall_ancestor() - Returns the nearest mem_exclusive or
2501 * mem_hardwall ancestor to the specified cpuset. Call holding
2502 * callback_lock. If no ancestor is mem_exclusive or mem_hardwall
2503 * (an unusual configuration), then returns the root cpuset.
2504 */
2505static struct cpuset *nearest_hardwall_ancestor(struct cpuset *cs)
2506{
2507 while (!(is_mem_exclusive(cs) || is_mem_hardwall(cs)) && parent_cs(cs))
2508 cs = parent_cs(cs);
2509 return cs;
2510}
2511
2512/**
2513 * cpuset_node_allowed - Can we allocate on a memory node?
2514 * @node: is this an allowed node?
2515 * @gfp_mask: memory allocation flags
2516 *
2517 * If we're in interrupt, yes, we can always allocate. If @node is set in
2518 * current's mems_allowed, yes. If it's not a __GFP_HARDWALL request and this
2519 * node is set in the nearest hardwalled cpuset ancestor to current's cpuset,
2520 * yes. If current has access to memory reserves as an oom victim, yes.
2521 * Otherwise, no.
2522 *
2523 * GFP_USER allocations are marked with the __GFP_HARDWALL bit,
2524 * and do not allow allocations outside the current tasks cpuset
2525 * unless the task has been OOM killed.
2526 * GFP_KERNEL allocations are not so marked, so can escape to the
2527 * nearest enclosing hardwalled ancestor cpuset.
2528 *
2529 * Scanning up parent cpusets requires callback_lock. The
2530 * __alloc_pages() routine only calls here with __GFP_HARDWALL bit
2531 * _not_ set if it's a GFP_KERNEL allocation, and all nodes in the
2532 * current tasks mems_allowed came up empty on the first pass over
2533 * the zonelist. So only GFP_KERNEL allocations, if all nodes in the
2534 * cpuset are short of memory, might require taking the callback_lock.
2535 *
2536 * The first call here from mm/page_alloc:get_page_from_freelist()
2537 * has __GFP_HARDWALL set in gfp_mask, enforcing hardwall cpusets,
2538 * so no allocation on a node outside the cpuset is allowed (unless
2539 * in interrupt, of course).
2540 *
2541 * The second pass through get_page_from_freelist() doesn't even call
2542 * here for GFP_ATOMIC calls. For those calls, the __alloc_pages()
2543 * variable 'wait' is not set, and the bit ALLOC_CPUSET is not set
2544 * in alloc_flags. That logic and the checks below have the combined
2545 * affect that:
2546 * in_interrupt - any node ok (current task context irrelevant)
2547 * GFP_ATOMIC - any node ok
2548 * tsk_is_oom_victim - any node ok
2549 * GFP_KERNEL - any node in enclosing hardwalled cpuset ok
2550 * GFP_USER - only nodes in current tasks mems allowed ok.
2551 */
2552bool __cpuset_node_allowed(int node, gfp_t gfp_mask)
2553{
2554 struct cpuset *cs; /* current cpuset ancestors */
2555 int allowed; /* is allocation in zone z allowed? */
2556 unsigned long flags;
2557
2558 if (in_interrupt())
2559 return true;
2560 if (node_isset(node, current->mems_allowed))
2561 return true;
2562 /*
2563 * Allow tasks that have access to memory reserves because they have
2564 * been OOM killed to get memory anywhere.
2565 */
2566 if (unlikely(tsk_is_oom_victim(current)))
2567 return true;
2568 if (gfp_mask & __GFP_HARDWALL) /* If hardwall request, stop here */
2569 return false;
2570
2571 if (current->flags & PF_EXITING) /* Let dying task have memory */
2572 return true;
2573
2574 /* Not hardwall and node outside mems_allowed: scan up cpusets */
2575 spin_lock_irqsave(&callback_lock, flags);
2576
2577 rcu_read_lock();
2578 cs = nearest_hardwall_ancestor(task_cs(current));
2579 allowed = node_isset(node, cs->mems_allowed);
2580 rcu_read_unlock();
2581
2582 spin_unlock_irqrestore(&callback_lock, flags);
2583 return allowed;
2584}
2585
2586/**
2587 * cpuset_mem_spread_node() - On which node to begin search for a file page
2588 * cpuset_slab_spread_node() - On which node to begin search for a slab page
2589 *
2590 * If a task is marked PF_SPREAD_PAGE or PF_SPREAD_SLAB (as for
2591 * tasks in a cpuset with is_spread_page or is_spread_slab set),
2592 * and if the memory allocation used cpuset_mem_spread_node()
2593 * to determine on which node to start looking, as it will for
2594 * certain page cache or slab cache pages such as used for file
2595 * system buffers and inode caches, then instead of starting on the
2596 * local node to look for a free page, rather spread the starting
2597 * node around the tasks mems_allowed nodes.
2598 *
2599 * We don't have to worry about the returned node being offline
2600 * because "it can't happen", and even if it did, it would be ok.
2601 *
2602 * The routines calling guarantee_online_mems() are careful to
2603 * only set nodes in task->mems_allowed that are online. So it
2604 * should not be possible for the following code to return an
2605 * offline node. But if it did, that would be ok, as this routine
2606 * is not returning the node where the allocation must be, only
2607 * the node where the search should start. The zonelist passed to
2608 * __alloc_pages() will include all nodes. If the slab allocator
2609 * is passed an offline node, it will fall back to the local node.
2610 * See kmem_cache_alloc_node().
2611 */
2612
2613static int cpuset_spread_node(int *rotor)
2614{
2615 return *rotor = next_node_in(*rotor, current->mems_allowed);
2616}
2617
2618int cpuset_mem_spread_node(void)
2619{
2620 if (current->cpuset_mem_spread_rotor == NUMA_NO_NODE)
2621 current->cpuset_mem_spread_rotor =
2622 node_random(¤t->mems_allowed);
2623
2624 return cpuset_spread_node(¤t->cpuset_mem_spread_rotor);
2625}
2626
2627int cpuset_slab_spread_node(void)
2628{
2629 if (current->cpuset_slab_spread_rotor == NUMA_NO_NODE)
2630 current->cpuset_slab_spread_rotor =
2631 node_random(¤t->mems_allowed);
2632
2633 return cpuset_spread_node(¤t->cpuset_slab_spread_rotor);
2634}
2635
2636EXPORT_SYMBOL_GPL(cpuset_mem_spread_node);
2637
2638/**
2639 * cpuset_mems_allowed_intersects - Does @tsk1's mems_allowed intersect @tsk2's?
2640 * @tsk1: pointer to task_struct of some task.
2641 * @tsk2: pointer to task_struct of some other task.
2642 *
2643 * Description: Return true if @tsk1's mems_allowed intersects the
2644 * mems_allowed of @tsk2. Used by the OOM killer to determine if
2645 * one of the task's memory usage might impact the memory available
2646 * to the other.
2647 **/
2648
2649int cpuset_mems_allowed_intersects(const struct task_struct *tsk1,
2650 const struct task_struct *tsk2)
2651{
2652 return nodes_intersects(tsk1->mems_allowed, tsk2->mems_allowed);
2653}
2654
2655/**
2656 * cpuset_print_current_mems_allowed - prints current's cpuset and mems_allowed
2657 *
2658 * Description: Prints current's name, cpuset name, and cached copy of its
2659 * mems_allowed to the kernel log.
2660 */
2661void cpuset_print_current_mems_allowed(void)
2662{
2663 struct cgroup *cgrp;
2664
2665 rcu_read_lock();
2666
2667 cgrp = task_cs(current)->css.cgroup;
2668 pr_info("%s cpuset=", current->comm);
2669 pr_cont_cgroup_name(cgrp);
2670 pr_cont(" mems_allowed=%*pbl\n",
2671 nodemask_pr_args(¤t->mems_allowed));
2672
2673 rcu_read_unlock();
2674}
2675
2676/*
2677 * Collection of memory_pressure is suppressed unless
2678 * this flag is enabled by writing "1" to the special
2679 * cpuset file 'memory_pressure_enabled' in the root cpuset.
2680 */
2681
2682int cpuset_memory_pressure_enabled __read_mostly;
2683
2684/**
2685 * cpuset_memory_pressure_bump - keep stats of per-cpuset reclaims.
2686 *
2687 * Keep a running average of the rate of synchronous (direct)
2688 * page reclaim efforts initiated by tasks in each cpuset.
2689 *
2690 * This represents the rate at which some task in the cpuset
2691 * ran low on memory on all nodes it was allowed to use, and
2692 * had to enter the kernels page reclaim code in an effort to
2693 * create more free memory by tossing clean pages or swapping
2694 * or writing dirty pages.
2695 *
2696 * Display to user space in the per-cpuset read-only file
2697 * "memory_pressure". Value displayed is an integer
2698 * representing the recent rate of entry into the synchronous
2699 * (direct) page reclaim by any task attached to the cpuset.
2700 **/
2701
2702void __cpuset_memory_pressure_bump(void)
2703{
2704 rcu_read_lock();
2705 fmeter_markevent(&task_cs(current)->fmeter);
2706 rcu_read_unlock();
2707}
2708
2709#ifdef CONFIG_PROC_PID_CPUSET
2710/*
2711 * proc_cpuset_show()
2712 * - Print tasks cpuset path into seq_file.
2713 * - Used for /proc/<pid>/cpuset.
2714 * - No need to task_lock(tsk) on this tsk->cpuset reference, as it
2715 * doesn't really matter if tsk->cpuset changes after we read it,
2716 * and we take cpuset_mutex, keeping cpuset_attach() from changing it
2717 * anyway.
2718 */
2719int proc_cpuset_show(struct seq_file *m, struct pid_namespace *ns,
2720 struct pid *pid, struct task_struct *tsk)
2721{
2722 char *buf;
2723 struct cgroup_subsys_state *css;
2724 int retval;
2725
2726 retval = -ENOMEM;
2727 buf = kmalloc(PATH_MAX, GFP_KERNEL);
2728 if (!buf)
2729 goto out;
2730
2731 css = task_get_css(tsk, cpuset_cgrp_id);
2732 retval = cgroup_path_ns(css->cgroup, buf, PATH_MAX,
2733 current->nsproxy->cgroup_ns);
2734 css_put(css);
2735 if (retval >= PATH_MAX)
2736 retval = -ENAMETOOLONG;
2737 if (retval < 0)
2738 goto out_free;
2739 seq_puts(m, buf);
2740 seq_putc(m, '\n');
2741 retval = 0;
2742out_free:
2743 kfree(buf);
2744out:
2745 return retval;
2746}
2747#endif /* CONFIG_PROC_PID_CPUSET */
2748
2749/* Display task mems_allowed in /proc/<pid>/status file. */
2750void cpuset_task_status_allowed(struct seq_file *m, struct task_struct *task)
2751{
2752 seq_printf(m, "Mems_allowed:\t%*pb\n",
2753 nodemask_pr_args(&task->mems_allowed));
2754 seq_printf(m, "Mems_allowed_list:\t%*pbl\n",
2755 nodemask_pr_args(&task->mems_allowed));
2756}
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#include "cgroup-internal.h"
25#include "cpuset-internal.h"
26
27#include <linux/init.h>
28#include <linux/interrupt.h>
29#include <linux/kernel.h>
30#include <linux/mempolicy.h>
31#include <linux/mm.h>
32#include <linux/memory.h>
33#include <linux/export.h>
34#include <linux/rcupdate.h>
35#include <linux/sched.h>
36#include <linux/sched/deadline.h>
37#include <linux/sched/mm.h>
38#include <linux/sched/task.h>
39#include <linux/security.h>
40#include <linux/oom.h>
41#include <linux/sched/isolation.h>
42#include <linux/wait.h>
43#include <linux/workqueue.h>
44
45DEFINE_STATIC_KEY_FALSE(cpusets_pre_enable_key);
46DEFINE_STATIC_KEY_FALSE(cpusets_enabled_key);
47
48/*
49 * There could be abnormal cpuset configurations for cpu or memory
50 * node binding, add this key to provide a quick low-cost judgment
51 * of the situation.
52 */
53DEFINE_STATIC_KEY_FALSE(cpusets_insane_config_key);
54
55static const char * const perr_strings[] = {
56 [PERR_INVCPUS] = "Invalid cpu list in cpuset.cpus.exclusive",
57 [PERR_INVPARENT] = "Parent is an invalid partition root",
58 [PERR_NOTPART] = "Parent is not a partition root",
59 [PERR_NOTEXCL] = "Cpu list in cpuset.cpus not exclusive",
60 [PERR_NOCPUS] = "Parent unable to distribute cpu downstream",
61 [PERR_HOTPLUG] = "No cpu available due to hotplug",
62 [PERR_CPUSEMPTY] = "cpuset.cpus and cpuset.cpus.exclusive are empty",
63 [PERR_HKEEPING] = "partition config conflicts with housekeeping setup",
64 [PERR_ACCESS] = "Enable partition not permitted",
65};
66
67/*
68 * Exclusive CPUs distributed out to sub-partitions of top_cpuset
69 */
70static cpumask_var_t subpartitions_cpus;
71
72/*
73 * Exclusive CPUs in isolated partitions
74 */
75static cpumask_var_t isolated_cpus;
76
77/*
78 * Housekeeping (HK_TYPE_DOMAIN) CPUs at boot
79 */
80static cpumask_var_t boot_hk_cpus;
81static bool have_boot_isolcpus;
82
83/* List of remote partition root children */
84static struct list_head remote_children;
85
86/*
87 * A flag to force sched domain rebuild at the end of an operation.
88 * It can be set in
89 * - update_partition_sd_lb()
90 * - remote_partition_check()
91 * - update_cpumasks_hier()
92 * - cpuset_update_flag()
93 * - cpuset_hotplug_update_tasks()
94 * - cpuset_handle_hotplug()
95 *
96 * Protected by cpuset_mutex (with cpus_read_lock held) or cpus_write_lock.
97 *
98 * Note that update_relax_domain_level() in cpuset-v1.c can still call
99 * rebuild_sched_domains_locked() directly without using this flag.
100 */
101static bool force_sd_rebuild;
102
103/*
104 * Partition root states:
105 *
106 * 0 - member (not a partition root)
107 * 1 - partition root
108 * 2 - partition root without load balancing (isolated)
109 * -1 - invalid partition root
110 * -2 - invalid isolated partition root
111 *
112 * There are 2 types of partitions - local or remote. Local partitions are
113 * those whose parents are partition root themselves. Setting of
114 * cpuset.cpus.exclusive are optional in setting up local partitions.
115 * Remote partitions are those whose parents are not partition roots. Passing
116 * down exclusive CPUs by setting cpuset.cpus.exclusive along its ancestor
117 * nodes are mandatory in creating a remote partition.
118 *
119 * For simplicity, a local partition can be created under a local or remote
120 * partition but a remote partition cannot have any partition root in its
121 * ancestor chain except the cgroup root.
122 */
123#define PRS_MEMBER 0
124#define PRS_ROOT 1
125#define PRS_ISOLATED 2
126#define PRS_INVALID_ROOT -1
127#define PRS_INVALID_ISOLATED -2
128
129static inline bool is_prs_invalid(int prs_state)
130{
131 return prs_state < 0;
132}
133
134/*
135 * Temporary cpumasks for working with partitions that are passed among
136 * functions to avoid memory allocation in inner functions.
137 */
138struct tmpmasks {
139 cpumask_var_t addmask, delmask; /* For partition root */
140 cpumask_var_t new_cpus; /* For update_cpumasks_hier() */
141};
142
143void inc_dl_tasks_cs(struct task_struct *p)
144{
145 struct cpuset *cs = task_cs(p);
146
147 cs->nr_deadline_tasks++;
148}
149
150void dec_dl_tasks_cs(struct task_struct *p)
151{
152 struct cpuset *cs = task_cs(p);
153
154 cs->nr_deadline_tasks--;
155}
156
157static inline int is_partition_valid(const struct cpuset *cs)
158{
159 return cs->partition_root_state > 0;
160}
161
162static inline int is_partition_invalid(const struct cpuset *cs)
163{
164 return cs->partition_root_state < 0;
165}
166
167/*
168 * Callers should hold callback_lock to modify partition_root_state.
169 */
170static inline void make_partition_invalid(struct cpuset *cs)
171{
172 if (cs->partition_root_state > 0)
173 cs->partition_root_state = -cs->partition_root_state;
174}
175
176/*
177 * Send notification event of whenever partition_root_state changes.
178 */
179static inline void notify_partition_change(struct cpuset *cs, int old_prs)
180{
181 if (old_prs == cs->partition_root_state)
182 return;
183 cgroup_file_notify(&cs->partition_file);
184
185 /* Reset prs_err if not invalid */
186 if (is_partition_valid(cs))
187 WRITE_ONCE(cs->prs_err, PERR_NONE);
188}
189
190static struct cpuset top_cpuset = {
191 .flags = BIT(CS_ONLINE) | BIT(CS_CPU_EXCLUSIVE) |
192 BIT(CS_MEM_EXCLUSIVE) | BIT(CS_SCHED_LOAD_BALANCE),
193 .partition_root_state = PRS_ROOT,
194 .relax_domain_level = -1,
195 .remote_sibling = LIST_HEAD_INIT(top_cpuset.remote_sibling),
196};
197
198/*
199 * There are two global locks guarding cpuset structures - cpuset_mutex and
200 * callback_lock. The cpuset code uses only cpuset_mutex. Other kernel
201 * subsystems can use cpuset_lock()/cpuset_unlock() to prevent change to cpuset
202 * structures. Note that cpuset_mutex needs to be a mutex as it is used in
203 * paths that rely on priority inheritance (e.g. scheduler - on RT) for
204 * correctness.
205 *
206 * A task must hold both locks to modify cpusets. If a task holds
207 * cpuset_mutex, it blocks others, ensuring that it is the only task able to
208 * also acquire callback_lock and be able to modify cpusets. It can perform
209 * various checks on the cpuset structure first, knowing nothing will change.
210 * It can also allocate memory while just holding cpuset_mutex. While it is
211 * performing these checks, various callback routines can briefly acquire
212 * callback_lock to query cpusets. Once it is ready to make the changes, it
213 * takes callback_lock, blocking everyone else.
214 *
215 * Calls to the kernel memory allocator can not be made while holding
216 * callback_lock, as that would risk double tripping on callback_lock
217 * from one of the callbacks into the cpuset code from within
218 * __alloc_pages().
219 *
220 * If a task is only holding callback_lock, then it has read-only
221 * access to cpusets.
222 *
223 * Now, the task_struct fields mems_allowed and mempolicy may be changed
224 * by other task, we use alloc_lock in the task_struct fields to protect
225 * them.
226 *
227 * The cpuset_common_seq_show() handlers only hold callback_lock across
228 * small pieces of code, such as when reading out possibly multi-word
229 * cpumasks and nodemasks.
230 */
231
232static DEFINE_MUTEX(cpuset_mutex);
233
234void cpuset_lock(void)
235{
236 mutex_lock(&cpuset_mutex);
237}
238
239void cpuset_unlock(void)
240{
241 mutex_unlock(&cpuset_mutex);
242}
243
244static DEFINE_SPINLOCK(callback_lock);
245
246void cpuset_callback_lock_irq(void)
247{
248 spin_lock_irq(&callback_lock);
249}
250
251void cpuset_callback_unlock_irq(void)
252{
253 spin_unlock_irq(&callback_lock);
254}
255
256static struct workqueue_struct *cpuset_migrate_mm_wq;
257
258static DECLARE_WAIT_QUEUE_HEAD(cpuset_attach_wq);
259
260static inline void check_insane_mems_config(nodemask_t *nodes)
261{
262 if (!cpusets_insane_config() &&
263 movable_only_nodes(nodes)) {
264 static_branch_enable(&cpusets_insane_config_key);
265 pr_info("Unsupported (movable nodes only) cpuset configuration detected (nmask=%*pbl)!\n"
266 "Cpuset allocations might fail even with a lot of memory available.\n",
267 nodemask_pr_args(nodes));
268 }
269}
270
271/*
272 * decrease cs->attach_in_progress.
273 * wake_up cpuset_attach_wq if cs->attach_in_progress==0.
274 */
275static inline void dec_attach_in_progress_locked(struct cpuset *cs)
276{
277 lockdep_assert_held(&cpuset_mutex);
278
279 cs->attach_in_progress--;
280 if (!cs->attach_in_progress)
281 wake_up(&cpuset_attach_wq);
282}
283
284static inline void dec_attach_in_progress(struct cpuset *cs)
285{
286 mutex_lock(&cpuset_mutex);
287 dec_attach_in_progress_locked(cs);
288 mutex_unlock(&cpuset_mutex);
289}
290
291static inline bool cpuset_v2(void)
292{
293 return !IS_ENABLED(CONFIG_CPUSETS_V1) ||
294 cgroup_subsys_on_dfl(cpuset_cgrp_subsys);
295}
296
297/*
298 * Cgroup v2 behavior is used on the "cpus" and "mems" control files when
299 * on default hierarchy or when the cpuset_v2_mode flag is set by mounting
300 * the v1 cpuset cgroup filesystem with the "cpuset_v2_mode" mount option.
301 * With v2 behavior, "cpus" and "mems" are always what the users have
302 * requested and won't be changed by hotplug events. Only the effective
303 * cpus or mems will be affected.
304 */
305static inline bool is_in_v2_mode(void)
306{
307 return cpuset_v2() ||
308 (cpuset_cgrp_subsys.root->flags & CGRP_ROOT_CPUSET_V2_MODE);
309}
310
311/**
312 * partition_is_populated - check if partition has tasks
313 * @cs: partition root to be checked
314 * @excluded_child: a child cpuset to be excluded in task checking
315 * Return: true if there are tasks, false otherwise
316 *
317 * It is assumed that @cs is a valid partition root. @excluded_child should
318 * be non-NULL when this cpuset is going to become a partition itself.
319 */
320static inline bool partition_is_populated(struct cpuset *cs,
321 struct cpuset *excluded_child)
322{
323 struct cgroup_subsys_state *css;
324 struct cpuset *child;
325
326 if (cs->css.cgroup->nr_populated_csets)
327 return true;
328 if (!excluded_child && !cs->nr_subparts)
329 return cgroup_is_populated(cs->css.cgroup);
330
331 rcu_read_lock();
332 cpuset_for_each_child(child, css, cs) {
333 if (child == excluded_child)
334 continue;
335 if (is_partition_valid(child))
336 continue;
337 if (cgroup_is_populated(child->css.cgroup)) {
338 rcu_read_unlock();
339 return true;
340 }
341 }
342 rcu_read_unlock();
343 return false;
344}
345
346/*
347 * Return in pmask the portion of a task's cpusets's cpus_allowed that
348 * are online and are capable of running the task. If none are found,
349 * walk up the cpuset hierarchy until we find one that does have some
350 * appropriate cpus.
351 *
352 * One way or another, we guarantee to return some non-empty subset
353 * of cpu_online_mask.
354 *
355 * Call with callback_lock or cpuset_mutex held.
356 */
357static void guarantee_online_cpus(struct task_struct *tsk,
358 struct cpumask *pmask)
359{
360 const struct cpumask *possible_mask = task_cpu_possible_mask(tsk);
361 struct cpuset *cs;
362
363 if (WARN_ON(!cpumask_and(pmask, possible_mask, cpu_online_mask)))
364 cpumask_copy(pmask, cpu_online_mask);
365
366 rcu_read_lock();
367 cs = task_cs(tsk);
368
369 while (!cpumask_intersects(cs->effective_cpus, pmask))
370 cs = parent_cs(cs);
371
372 cpumask_and(pmask, pmask, cs->effective_cpus);
373 rcu_read_unlock();
374}
375
376/*
377 * Return in *pmask the portion of a cpusets's mems_allowed that
378 * are online, with memory. If none are online with memory, walk
379 * up the cpuset hierarchy until we find one that does have some
380 * online mems. The top cpuset always has some mems online.
381 *
382 * One way or another, we guarantee to return some non-empty subset
383 * of node_states[N_MEMORY].
384 *
385 * Call with callback_lock or cpuset_mutex held.
386 */
387static void guarantee_online_mems(struct cpuset *cs, nodemask_t *pmask)
388{
389 while (!nodes_intersects(cs->effective_mems, node_states[N_MEMORY]))
390 cs = parent_cs(cs);
391 nodes_and(*pmask, cs->effective_mems, node_states[N_MEMORY]);
392}
393
394/**
395 * alloc_cpumasks - allocate three cpumasks for cpuset
396 * @cs: the cpuset that have cpumasks to be allocated.
397 * @tmp: the tmpmasks structure pointer
398 * Return: 0 if successful, -ENOMEM otherwise.
399 *
400 * Only one of the two input arguments should be non-NULL.
401 */
402static inline int alloc_cpumasks(struct cpuset *cs, struct tmpmasks *tmp)
403{
404 cpumask_var_t *pmask1, *pmask2, *pmask3, *pmask4;
405
406 if (cs) {
407 pmask1 = &cs->cpus_allowed;
408 pmask2 = &cs->effective_cpus;
409 pmask3 = &cs->effective_xcpus;
410 pmask4 = &cs->exclusive_cpus;
411 } else {
412 pmask1 = &tmp->new_cpus;
413 pmask2 = &tmp->addmask;
414 pmask3 = &tmp->delmask;
415 pmask4 = NULL;
416 }
417
418 if (!zalloc_cpumask_var(pmask1, GFP_KERNEL))
419 return -ENOMEM;
420
421 if (!zalloc_cpumask_var(pmask2, GFP_KERNEL))
422 goto free_one;
423
424 if (!zalloc_cpumask_var(pmask3, GFP_KERNEL))
425 goto free_two;
426
427 if (pmask4 && !zalloc_cpumask_var(pmask4, GFP_KERNEL))
428 goto free_three;
429
430
431 return 0;
432
433free_three:
434 free_cpumask_var(*pmask3);
435free_two:
436 free_cpumask_var(*pmask2);
437free_one:
438 free_cpumask_var(*pmask1);
439 return -ENOMEM;
440}
441
442/**
443 * free_cpumasks - free cpumasks in a tmpmasks structure
444 * @cs: the cpuset that have cpumasks to be free.
445 * @tmp: the tmpmasks structure pointer
446 */
447static inline void free_cpumasks(struct cpuset *cs, struct tmpmasks *tmp)
448{
449 if (cs) {
450 free_cpumask_var(cs->cpus_allowed);
451 free_cpumask_var(cs->effective_cpus);
452 free_cpumask_var(cs->effective_xcpus);
453 free_cpumask_var(cs->exclusive_cpus);
454 }
455 if (tmp) {
456 free_cpumask_var(tmp->new_cpus);
457 free_cpumask_var(tmp->addmask);
458 free_cpumask_var(tmp->delmask);
459 }
460}
461
462/**
463 * alloc_trial_cpuset - allocate a trial cpuset
464 * @cs: the cpuset that the trial cpuset duplicates
465 */
466static struct cpuset *alloc_trial_cpuset(struct cpuset *cs)
467{
468 struct cpuset *trial;
469
470 trial = kmemdup(cs, sizeof(*cs), GFP_KERNEL);
471 if (!trial)
472 return NULL;
473
474 if (alloc_cpumasks(trial, NULL)) {
475 kfree(trial);
476 return NULL;
477 }
478
479 cpumask_copy(trial->cpus_allowed, cs->cpus_allowed);
480 cpumask_copy(trial->effective_cpus, cs->effective_cpus);
481 cpumask_copy(trial->effective_xcpus, cs->effective_xcpus);
482 cpumask_copy(trial->exclusive_cpus, cs->exclusive_cpus);
483 return trial;
484}
485
486/**
487 * free_cpuset - free the cpuset
488 * @cs: the cpuset to be freed
489 */
490static inline void free_cpuset(struct cpuset *cs)
491{
492 free_cpumasks(cs, NULL);
493 kfree(cs);
494}
495
496/* Return user specified exclusive CPUs */
497static inline struct cpumask *user_xcpus(struct cpuset *cs)
498{
499 return cpumask_empty(cs->exclusive_cpus) ? cs->cpus_allowed
500 : cs->exclusive_cpus;
501}
502
503static inline bool xcpus_empty(struct cpuset *cs)
504{
505 return cpumask_empty(cs->cpus_allowed) &&
506 cpumask_empty(cs->exclusive_cpus);
507}
508
509/*
510 * cpusets_are_exclusive() - check if two cpusets are exclusive
511 *
512 * Return true if exclusive, false if not
513 */
514static inline bool cpusets_are_exclusive(struct cpuset *cs1, struct cpuset *cs2)
515{
516 struct cpumask *xcpus1 = user_xcpus(cs1);
517 struct cpumask *xcpus2 = user_xcpus(cs2);
518
519 if (cpumask_intersects(xcpus1, xcpus2))
520 return false;
521 return true;
522}
523
524/*
525 * validate_change() - Used to validate that any proposed cpuset change
526 * follows the structural rules for cpusets.
527 *
528 * If we replaced the flag and mask values of the current cpuset
529 * (cur) with those values in the trial cpuset (trial), would
530 * our various subset and exclusive rules still be valid? Presumes
531 * cpuset_mutex held.
532 *
533 * 'cur' is the address of an actual, in-use cpuset. Operations
534 * such as list traversal that depend on the actual address of the
535 * cpuset in the list must use cur below, not trial.
536 *
537 * 'trial' is the address of bulk structure copy of cur, with
538 * perhaps one or more of the fields cpus_allowed, mems_allowed,
539 * or flags changed to new, trial values.
540 *
541 * Return 0 if valid, -errno if not.
542 */
543
544static int validate_change(struct cpuset *cur, struct cpuset *trial)
545{
546 struct cgroup_subsys_state *css;
547 struct cpuset *c, *par;
548 int ret = 0;
549
550 rcu_read_lock();
551
552 if (!is_in_v2_mode())
553 ret = cpuset1_validate_change(cur, trial);
554 if (ret)
555 goto out;
556
557 /* Remaining checks don't apply to root cpuset */
558 if (cur == &top_cpuset)
559 goto out;
560
561 par = parent_cs(cur);
562
563 /*
564 * Cpusets with tasks - existing or newly being attached - can't
565 * be changed to have empty cpus_allowed or mems_allowed.
566 */
567 ret = -ENOSPC;
568 if ((cgroup_is_populated(cur->css.cgroup) || cur->attach_in_progress)) {
569 if (!cpumask_empty(cur->cpus_allowed) &&
570 cpumask_empty(trial->cpus_allowed))
571 goto out;
572 if (!nodes_empty(cur->mems_allowed) &&
573 nodes_empty(trial->mems_allowed))
574 goto out;
575 }
576
577 /*
578 * We can't shrink if we won't have enough room for SCHED_DEADLINE
579 * tasks. This check is not done when scheduling is disabled as the
580 * users should know what they are doing.
581 *
582 * For v1, effective_cpus == cpus_allowed & user_xcpus() returns
583 * cpus_allowed.
584 *
585 * For v2, is_cpu_exclusive() & is_sched_load_balance() are true only
586 * for non-isolated partition root. At this point, the target
587 * effective_cpus isn't computed yet. user_xcpus() is the best
588 * approximation.
589 *
590 * TBD: May need to precompute the real effective_cpus here in case
591 * incorrect scheduling of SCHED_DEADLINE tasks in a partition
592 * becomes an issue.
593 */
594 ret = -EBUSY;
595 if (is_cpu_exclusive(cur) && is_sched_load_balance(cur) &&
596 !cpuset_cpumask_can_shrink(cur->effective_cpus, user_xcpus(trial)))
597 goto out;
598
599 /*
600 * If either I or some sibling (!= me) is exclusive, we can't
601 * overlap. exclusive_cpus cannot overlap with each other if set.
602 */
603 ret = -EINVAL;
604 cpuset_for_each_child(c, css, par) {
605 bool txset, cxset; /* Are exclusive_cpus set? */
606
607 if (c == cur)
608 continue;
609
610 txset = !cpumask_empty(trial->exclusive_cpus);
611 cxset = !cpumask_empty(c->exclusive_cpus);
612 if (is_cpu_exclusive(trial) || is_cpu_exclusive(c) ||
613 (txset && cxset)) {
614 if (!cpusets_are_exclusive(trial, c))
615 goto out;
616 } else if (txset || cxset) {
617 struct cpumask *xcpus, *acpus;
618
619 /*
620 * When just one of the exclusive_cpus's is set,
621 * cpus_allowed of the other cpuset, if set, cannot be
622 * a subset of it or none of those CPUs will be
623 * available if these exclusive CPUs are activated.
624 */
625 if (txset) {
626 xcpus = trial->exclusive_cpus;
627 acpus = c->cpus_allowed;
628 } else {
629 xcpus = c->exclusive_cpus;
630 acpus = trial->cpus_allowed;
631 }
632 if (!cpumask_empty(acpus) && cpumask_subset(acpus, xcpus))
633 goto out;
634 }
635 if ((is_mem_exclusive(trial) || is_mem_exclusive(c)) &&
636 nodes_intersects(trial->mems_allowed, c->mems_allowed))
637 goto out;
638 }
639
640 ret = 0;
641out:
642 rcu_read_unlock();
643 return ret;
644}
645
646#ifdef CONFIG_SMP
647/*
648 * Helper routine for generate_sched_domains().
649 * Do cpusets a, b have overlapping effective cpus_allowed masks?
650 */
651static int cpusets_overlap(struct cpuset *a, struct cpuset *b)
652{
653 return cpumask_intersects(a->effective_cpus, b->effective_cpus);
654}
655
656static void
657update_domain_attr(struct sched_domain_attr *dattr, struct cpuset *c)
658{
659 if (dattr->relax_domain_level < c->relax_domain_level)
660 dattr->relax_domain_level = c->relax_domain_level;
661 return;
662}
663
664static void update_domain_attr_tree(struct sched_domain_attr *dattr,
665 struct cpuset *root_cs)
666{
667 struct cpuset *cp;
668 struct cgroup_subsys_state *pos_css;
669
670 rcu_read_lock();
671 cpuset_for_each_descendant_pre(cp, pos_css, root_cs) {
672 /* skip the whole subtree if @cp doesn't have any CPU */
673 if (cpumask_empty(cp->cpus_allowed)) {
674 pos_css = css_rightmost_descendant(pos_css);
675 continue;
676 }
677
678 if (is_sched_load_balance(cp))
679 update_domain_attr(dattr, cp);
680 }
681 rcu_read_unlock();
682}
683
684/* Must be called with cpuset_mutex held. */
685static inline int nr_cpusets(void)
686{
687 /* jump label reference count + the top-level cpuset */
688 return static_key_count(&cpusets_enabled_key.key) + 1;
689}
690
691/*
692 * generate_sched_domains()
693 *
694 * This function builds a partial partition of the systems CPUs
695 * A 'partial partition' is a set of non-overlapping subsets whose
696 * union is a subset of that set.
697 * The output of this function needs to be passed to kernel/sched/core.c
698 * partition_sched_domains() routine, which will rebuild the scheduler's
699 * load balancing domains (sched domains) as specified by that partial
700 * partition.
701 *
702 * See "What is sched_load_balance" in Documentation/admin-guide/cgroup-v1/cpusets.rst
703 * for a background explanation of this.
704 *
705 * Does not return errors, on the theory that the callers of this
706 * routine would rather not worry about failures to rebuild sched
707 * domains when operating in the severe memory shortage situations
708 * that could cause allocation failures below.
709 *
710 * Must be called with cpuset_mutex held.
711 *
712 * The three key local variables below are:
713 * cp - cpuset pointer, used (together with pos_css) to perform a
714 * top-down scan of all cpusets. For our purposes, rebuilding
715 * the schedulers sched domains, we can ignore !is_sched_load_
716 * balance cpusets.
717 * csa - (for CpuSet Array) Array of pointers to all the cpusets
718 * that need to be load balanced, for convenient iterative
719 * access by the subsequent code that finds the best partition,
720 * i.e the set of domains (subsets) of CPUs such that the
721 * cpus_allowed of every cpuset marked is_sched_load_balance
722 * is a subset of one of these domains, while there are as
723 * many such domains as possible, each as small as possible.
724 * doms - Conversion of 'csa' to an array of cpumasks, for passing to
725 * the kernel/sched/core.c routine partition_sched_domains() in a
726 * convenient format, that can be easily compared to the prior
727 * value to determine what partition elements (sched domains)
728 * were changed (added or removed.)
729 *
730 * Finding the best partition (set of domains):
731 * The double nested loops below over i, j scan over the load
732 * balanced cpusets (using the array of cpuset pointers in csa[])
733 * looking for pairs of cpusets that have overlapping cpus_allowed
734 * and merging them using a union-find algorithm.
735 *
736 * The union of the cpus_allowed masks from the set of all cpusets
737 * having the same root then form the one element of the partition
738 * (one sched domain) to be passed to partition_sched_domains().
739 *
740 */
741static int generate_sched_domains(cpumask_var_t **domains,
742 struct sched_domain_attr **attributes)
743{
744 struct cpuset *cp; /* top-down scan of cpusets */
745 struct cpuset **csa; /* array of all cpuset ptrs */
746 int csn; /* how many cpuset ptrs in csa so far */
747 int i, j; /* indices for partition finding loops */
748 cpumask_var_t *doms; /* resulting partition; i.e. sched domains */
749 struct sched_domain_attr *dattr; /* attributes for custom domains */
750 int ndoms = 0; /* number of sched domains in result */
751 int nslot; /* next empty doms[] struct cpumask slot */
752 struct cgroup_subsys_state *pos_css;
753 bool root_load_balance = is_sched_load_balance(&top_cpuset);
754 bool cgrpv2 = cpuset_v2();
755 int nslot_update;
756
757 doms = NULL;
758 dattr = NULL;
759 csa = NULL;
760
761 /* Special case for the 99% of systems with one, full, sched domain */
762 if (root_load_balance && cpumask_empty(subpartitions_cpus)) {
763single_root_domain:
764 ndoms = 1;
765 doms = alloc_sched_domains(ndoms);
766 if (!doms)
767 goto done;
768
769 dattr = kmalloc(sizeof(struct sched_domain_attr), GFP_KERNEL);
770 if (dattr) {
771 *dattr = SD_ATTR_INIT;
772 update_domain_attr_tree(dattr, &top_cpuset);
773 }
774 cpumask_and(doms[0], top_cpuset.effective_cpus,
775 housekeeping_cpumask(HK_TYPE_DOMAIN));
776
777 goto done;
778 }
779
780 csa = kmalloc_array(nr_cpusets(), sizeof(cp), GFP_KERNEL);
781 if (!csa)
782 goto done;
783 csn = 0;
784
785 rcu_read_lock();
786 if (root_load_balance)
787 csa[csn++] = &top_cpuset;
788 cpuset_for_each_descendant_pre(cp, pos_css, &top_cpuset) {
789 if (cp == &top_cpuset)
790 continue;
791
792 if (cgrpv2)
793 goto v2;
794
795 /*
796 * v1:
797 * Continue traversing beyond @cp iff @cp has some CPUs and
798 * isn't load balancing. The former is obvious. The
799 * latter: All child cpusets contain a subset of the
800 * parent's cpus, so just skip them, and then we call
801 * update_domain_attr_tree() to calc relax_domain_level of
802 * the corresponding sched domain.
803 */
804 if (!cpumask_empty(cp->cpus_allowed) &&
805 !(is_sched_load_balance(cp) &&
806 cpumask_intersects(cp->cpus_allowed,
807 housekeeping_cpumask(HK_TYPE_DOMAIN))))
808 continue;
809
810 if (is_sched_load_balance(cp) &&
811 !cpumask_empty(cp->effective_cpus))
812 csa[csn++] = cp;
813
814 /* skip @cp's subtree */
815 pos_css = css_rightmost_descendant(pos_css);
816 continue;
817
818v2:
819 /*
820 * Only valid partition roots that are not isolated and with
821 * non-empty effective_cpus will be saved into csn[].
822 */
823 if ((cp->partition_root_state == PRS_ROOT) &&
824 !cpumask_empty(cp->effective_cpus))
825 csa[csn++] = cp;
826
827 /*
828 * Skip @cp's subtree if not a partition root and has no
829 * exclusive CPUs to be granted to child cpusets.
830 */
831 if (!is_partition_valid(cp) && cpumask_empty(cp->exclusive_cpus))
832 pos_css = css_rightmost_descendant(pos_css);
833 }
834 rcu_read_unlock();
835
836 /*
837 * If there are only isolated partitions underneath the cgroup root,
838 * we can optimize out unneeded sched domains scanning.
839 */
840 if (root_load_balance && (csn == 1))
841 goto single_root_domain;
842
843 for (i = 0; i < csn; i++)
844 uf_node_init(&csa[i]->node);
845
846 /* Merge overlapping cpusets */
847 for (i = 0; i < csn; i++) {
848 for (j = i + 1; j < csn; j++) {
849 if (cpusets_overlap(csa[i], csa[j])) {
850 /*
851 * Cgroup v2 shouldn't pass down overlapping
852 * partition root cpusets.
853 */
854 WARN_ON_ONCE(cgrpv2);
855 uf_union(&csa[i]->node, &csa[j]->node);
856 }
857 }
858 }
859
860 /* Count the total number of domains */
861 for (i = 0; i < csn; i++) {
862 if (uf_find(&csa[i]->node) == &csa[i]->node)
863 ndoms++;
864 }
865
866 /*
867 * Now we know how many domains to create.
868 * Convert <csn, csa> to <ndoms, doms> and populate cpu masks.
869 */
870 doms = alloc_sched_domains(ndoms);
871 if (!doms)
872 goto done;
873
874 /*
875 * The rest of the code, including the scheduler, can deal with
876 * dattr==NULL case. No need to abort if alloc fails.
877 */
878 dattr = kmalloc_array(ndoms, sizeof(struct sched_domain_attr),
879 GFP_KERNEL);
880
881 /*
882 * Cgroup v2 doesn't support domain attributes, just set all of them
883 * to SD_ATTR_INIT. Also non-isolating partition root CPUs are a
884 * subset of HK_TYPE_DOMAIN housekeeping CPUs.
885 */
886 if (cgrpv2) {
887 for (i = 0; i < ndoms; i++) {
888 /*
889 * The top cpuset may contain some boot time isolated
890 * CPUs that need to be excluded from the sched domain.
891 */
892 if (csa[i] == &top_cpuset)
893 cpumask_and(doms[i], csa[i]->effective_cpus,
894 housekeeping_cpumask(HK_TYPE_DOMAIN));
895 else
896 cpumask_copy(doms[i], csa[i]->effective_cpus);
897 if (dattr)
898 dattr[i] = SD_ATTR_INIT;
899 }
900 goto done;
901 }
902
903 for (nslot = 0, i = 0; i < csn; i++) {
904 nslot_update = 0;
905 for (j = i; j < csn; j++) {
906 if (uf_find(&csa[j]->node) == &csa[i]->node) {
907 struct cpumask *dp = doms[nslot];
908
909 if (i == j) {
910 nslot_update = 1;
911 cpumask_clear(dp);
912 if (dattr)
913 *(dattr + nslot) = SD_ATTR_INIT;
914 }
915 cpumask_or(dp, dp, csa[j]->effective_cpus);
916 cpumask_and(dp, dp, housekeeping_cpumask(HK_TYPE_DOMAIN));
917 if (dattr)
918 update_domain_attr_tree(dattr + nslot, csa[j]);
919 }
920 }
921 if (nslot_update)
922 nslot++;
923 }
924 BUG_ON(nslot != ndoms);
925
926done:
927 kfree(csa);
928
929 /*
930 * Fallback to the default domain if kmalloc() failed.
931 * See comments in partition_sched_domains().
932 */
933 if (doms == NULL)
934 ndoms = 1;
935
936 *domains = doms;
937 *attributes = dattr;
938 return ndoms;
939}
940
941static void dl_update_tasks_root_domain(struct cpuset *cs)
942{
943 struct css_task_iter it;
944 struct task_struct *task;
945
946 if (cs->nr_deadline_tasks == 0)
947 return;
948
949 css_task_iter_start(&cs->css, 0, &it);
950
951 while ((task = css_task_iter_next(&it)))
952 dl_add_task_root_domain(task);
953
954 css_task_iter_end(&it);
955}
956
957static void dl_rebuild_rd_accounting(void)
958{
959 struct cpuset *cs = NULL;
960 struct cgroup_subsys_state *pos_css;
961
962 lockdep_assert_held(&cpuset_mutex);
963 lockdep_assert_cpus_held();
964 lockdep_assert_held(&sched_domains_mutex);
965
966 rcu_read_lock();
967
968 /*
969 * Clear default root domain DL accounting, it will be computed again
970 * if a task belongs to it.
971 */
972 dl_clear_root_domain(&def_root_domain);
973
974 cpuset_for_each_descendant_pre(cs, pos_css, &top_cpuset) {
975
976 if (cpumask_empty(cs->effective_cpus)) {
977 pos_css = css_rightmost_descendant(pos_css);
978 continue;
979 }
980
981 css_get(&cs->css);
982
983 rcu_read_unlock();
984
985 dl_update_tasks_root_domain(cs);
986
987 rcu_read_lock();
988 css_put(&cs->css);
989 }
990 rcu_read_unlock();
991}
992
993static void
994partition_and_rebuild_sched_domains(int ndoms_new, cpumask_var_t doms_new[],
995 struct sched_domain_attr *dattr_new)
996{
997 mutex_lock(&sched_domains_mutex);
998 partition_sched_domains_locked(ndoms_new, doms_new, dattr_new);
999 dl_rebuild_rd_accounting();
1000 mutex_unlock(&sched_domains_mutex);
1001}
1002
1003/*
1004 * Rebuild scheduler domains.
1005 *
1006 * If the flag 'sched_load_balance' of any cpuset with non-empty
1007 * 'cpus' changes, or if the 'cpus' allowed changes in any cpuset
1008 * which has that flag enabled, or if any cpuset with a non-empty
1009 * 'cpus' is removed, then call this routine to rebuild the
1010 * scheduler's dynamic sched domains.
1011 *
1012 * Call with cpuset_mutex held. Takes cpus_read_lock().
1013 */
1014void rebuild_sched_domains_locked(void)
1015{
1016 struct cgroup_subsys_state *pos_css;
1017 struct sched_domain_attr *attr;
1018 cpumask_var_t *doms;
1019 struct cpuset *cs;
1020 int ndoms;
1021
1022 lockdep_assert_cpus_held();
1023 lockdep_assert_held(&cpuset_mutex);
1024 force_sd_rebuild = false;
1025
1026 /*
1027 * If we have raced with CPU hotplug, return early to avoid
1028 * passing doms with offlined cpu to partition_sched_domains().
1029 * Anyways, cpuset_handle_hotplug() will rebuild sched domains.
1030 *
1031 * With no CPUs in any subpartitions, top_cpuset's effective CPUs
1032 * should be the same as the active CPUs, so checking only top_cpuset
1033 * is enough to detect racing CPU offlines.
1034 */
1035 if (cpumask_empty(subpartitions_cpus) &&
1036 !cpumask_equal(top_cpuset.effective_cpus, cpu_active_mask))
1037 return;
1038
1039 /*
1040 * With subpartition CPUs, however, the effective CPUs of a partition
1041 * root should be only a subset of the active CPUs. Since a CPU in any
1042 * partition root could be offlined, all must be checked.
1043 */
1044 if (!cpumask_empty(subpartitions_cpus)) {
1045 rcu_read_lock();
1046 cpuset_for_each_descendant_pre(cs, pos_css, &top_cpuset) {
1047 if (!is_partition_valid(cs)) {
1048 pos_css = css_rightmost_descendant(pos_css);
1049 continue;
1050 }
1051 if (!cpumask_subset(cs->effective_cpus,
1052 cpu_active_mask)) {
1053 rcu_read_unlock();
1054 return;
1055 }
1056 }
1057 rcu_read_unlock();
1058 }
1059
1060 /* Generate domain masks and attrs */
1061 ndoms = generate_sched_domains(&doms, &attr);
1062
1063 /* Have scheduler rebuild the domains */
1064 partition_and_rebuild_sched_domains(ndoms, doms, attr);
1065}
1066#else /* !CONFIG_SMP */
1067void rebuild_sched_domains_locked(void)
1068{
1069}
1070#endif /* CONFIG_SMP */
1071
1072static void rebuild_sched_domains_cpuslocked(void)
1073{
1074 mutex_lock(&cpuset_mutex);
1075 rebuild_sched_domains_locked();
1076 mutex_unlock(&cpuset_mutex);
1077}
1078
1079void rebuild_sched_domains(void)
1080{
1081 cpus_read_lock();
1082 rebuild_sched_domains_cpuslocked();
1083 cpus_read_unlock();
1084}
1085
1086/**
1087 * cpuset_update_tasks_cpumask - Update the cpumasks of tasks in the cpuset.
1088 * @cs: the cpuset in which each task's cpus_allowed mask needs to be changed
1089 * @new_cpus: the temp variable for the new effective_cpus mask
1090 *
1091 * Iterate through each task of @cs updating its cpus_allowed to the
1092 * effective cpuset's. As this function is called with cpuset_mutex held,
1093 * cpuset membership stays stable. For top_cpuset, task_cpu_possible_mask()
1094 * is used instead of effective_cpus to make sure all offline CPUs are also
1095 * included as hotplug code won't update cpumasks for tasks in top_cpuset.
1096 */
1097void cpuset_update_tasks_cpumask(struct cpuset *cs, struct cpumask *new_cpus)
1098{
1099 struct css_task_iter it;
1100 struct task_struct *task;
1101 bool top_cs = cs == &top_cpuset;
1102
1103 css_task_iter_start(&cs->css, 0, &it);
1104 while ((task = css_task_iter_next(&it))) {
1105 const struct cpumask *possible_mask = task_cpu_possible_mask(task);
1106
1107 if (top_cs) {
1108 /*
1109 * Percpu kthreads in top_cpuset are ignored
1110 */
1111 if (kthread_is_per_cpu(task))
1112 continue;
1113 cpumask_andnot(new_cpus, possible_mask, subpartitions_cpus);
1114 } else {
1115 cpumask_and(new_cpus, possible_mask, cs->effective_cpus);
1116 }
1117 set_cpus_allowed_ptr(task, new_cpus);
1118 }
1119 css_task_iter_end(&it);
1120}
1121
1122/**
1123 * compute_effective_cpumask - Compute the effective cpumask of the cpuset
1124 * @new_cpus: the temp variable for the new effective_cpus mask
1125 * @cs: the cpuset the need to recompute the new effective_cpus mask
1126 * @parent: the parent cpuset
1127 *
1128 * The result is valid only if the given cpuset isn't a partition root.
1129 */
1130static void compute_effective_cpumask(struct cpumask *new_cpus,
1131 struct cpuset *cs, struct cpuset *parent)
1132{
1133 cpumask_and(new_cpus, cs->cpus_allowed, parent->effective_cpus);
1134}
1135
1136/*
1137 * Commands for update_parent_effective_cpumask
1138 */
1139enum partition_cmd {
1140 partcmd_enable, /* Enable partition root */
1141 partcmd_enablei, /* Enable isolated partition root */
1142 partcmd_disable, /* Disable partition root */
1143 partcmd_update, /* Update parent's effective_cpus */
1144 partcmd_invalidate, /* Make partition invalid */
1145};
1146
1147static void update_sibling_cpumasks(struct cpuset *parent, struct cpuset *cs,
1148 struct tmpmasks *tmp);
1149
1150/*
1151 * Update partition exclusive flag
1152 *
1153 * Return: 0 if successful, an error code otherwise
1154 */
1155static int update_partition_exclusive(struct cpuset *cs, int new_prs)
1156{
1157 bool exclusive = (new_prs > PRS_MEMBER);
1158
1159 if (exclusive && !is_cpu_exclusive(cs)) {
1160 if (cpuset_update_flag(CS_CPU_EXCLUSIVE, cs, 1))
1161 return PERR_NOTEXCL;
1162 } else if (!exclusive && is_cpu_exclusive(cs)) {
1163 /* Turning off CS_CPU_EXCLUSIVE will not return error */
1164 cpuset_update_flag(CS_CPU_EXCLUSIVE, cs, 0);
1165 }
1166 return 0;
1167}
1168
1169/*
1170 * Update partition load balance flag and/or rebuild sched domain
1171 *
1172 * Changing load balance flag will automatically call
1173 * rebuild_sched_domains_locked().
1174 * This function is for cgroup v2 only.
1175 */
1176static void update_partition_sd_lb(struct cpuset *cs, int old_prs)
1177{
1178 int new_prs = cs->partition_root_state;
1179 bool rebuild_domains = (new_prs > 0) || (old_prs > 0);
1180 bool new_lb;
1181
1182 /*
1183 * If cs is not a valid partition root, the load balance state
1184 * will follow its parent.
1185 */
1186 if (new_prs > 0) {
1187 new_lb = (new_prs != PRS_ISOLATED);
1188 } else {
1189 new_lb = is_sched_load_balance(parent_cs(cs));
1190 }
1191 if (new_lb != !!is_sched_load_balance(cs)) {
1192 rebuild_domains = true;
1193 if (new_lb)
1194 set_bit(CS_SCHED_LOAD_BALANCE, &cs->flags);
1195 else
1196 clear_bit(CS_SCHED_LOAD_BALANCE, &cs->flags);
1197 }
1198
1199 if (rebuild_domains)
1200 cpuset_force_rebuild();
1201}
1202
1203/*
1204 * tasks_nocpu_error - Return true if tasks will have no effective_cpus
1205 */
1206static bool tasks_nocpu_error(struct cpuset *parent, struct cpuset *cs,
1207 struct cpumask *xcpus)
1208{
1209 /*
1210 * A populated partition (cs or parent) can't have empty effective_cpus
1211 */
1212 return (cpumask_subset(parent->effective_cpus, xcpus) &&
1213 partition_is_populated(parent, cs)) ||
1214 (!cpumask_intersects(xcpus, cpu_active_mask) &&
1215 partition_is_populated(cs, NULL));
1216}
1217
1218static void reset_partition_data(struct cpuset *cs)
1219{
1220 struct cpuset *parent = parent_cs(cs);
1221
1222 if (!cpuset_v2())
1223 return;
1224
1225 lockdep_assert_held(&callback_lock);
1226
1227 cs->nr_subparts = 0;
1228 if (cpumask_empty(cs->exclusive_cpus)) {
1229 cpumask_clear(cs->effective_xcpus);
1230 if (is_cpu_exclusive(cs))
1231 clear_bit(CS_CPU_EXCLUSIVE, &cs->flags);
1232 }
1233 if (!cpumask_and(cs->effective_cpus, parent->effective_cpus, cs->cpus_allowed))
1234 cpumask_copy(cs->effective_cpus, parent->effective_cpus);
1235}
1236
1237/*
1238 * partition_xcpus_newstate - Exclusive CPUs state change
1239 * @old_prs: old partition_root_state
1240 * @new_prs: new partition_root_state
1241 * @xcpus: exclusive CPUs with state change
1242 */
1243static void partition_xcpus_newstate(int old_prs, int new_prs, struct cpumask *xcpus)
1244{
1245 WARN_ON_ONCE(old_prs == new_prs);
1246 if (new_prs == PRS_ISOLATED)
1247 cpumask_or(isolated_cpus, isolated_cpus, xcpus);
1248 else
1249 cpumask_andnot(isolated_cpus, isolated_cpus, xcpus);
1250}
1251
1252/*
1253 * partition_xcpus_add - Add new exclusive CPUs to partition
1254 * @new_prs: new partition_root_state
1255 * @parent: parent cpuset
1256 * @xcpus: exclusive CPUs to be added
1257 * Return: true if isolated_cpus modified, false otherwise
1258 *
1259 * Remote partition if parent == NULL
1260 */
1261static bool partition_xcpus_add(int new_prs, struct cpuset *parent,
1262 struct cpumask *xcpus)
1263{
1264 bool isolcpus_updated;
1265
1266 WARN_ON_ONCE(new_prs < 0);
1267 lockdep_assert_held(&callback_lock);
1268 if (!parent)
1269 parent = &top_cpuset;
1270
1271
1272 if (parent == &top_cpuset)
1273 cpumask_or(subpartitions_cpus, subpartitions_cpus, xcpus);
1274
1275 isolcpus_updated = (new_prs != parent->partition_root_state);
1276 if (isolcpus_updated)
1277 partition_xcpus_newstate(parent->partition_root_state, new_prs,
1278 xcpus);
1279
1280 cpumask_andnot(parent->effective_cpus, parent->effective_cpus, xcpus);
1281 return isolcpus_updated;
1282}
1283
1284/*
1285 * partition_xcpus_del - Remove exclusive CPUs from partition
1286 * @old_prs: old partition_root_state
1287 * @parent: parent cpuset
1288 * @xcpus: exclusive CPUs to be removed
1289 * Return: true if isolated_cpus modified, false otherwise
1290 *
1291 * Remote partition if parent == NULL
1292 */
1293static bool partition_xcpus_del(int old_prs, struct cpuset *parent,
1294 struct cpumask *xcpus)
1295{
1296 bool isolcpus_updated;
1297
1298 WARN_ON_ONCE(old_prs < 0);
1299 lockdep_assert_held(&callback_lock);
1300 if (!parent)
1301 parent = &top_cpuset;
1302
1303 if (parent == &top_cpuset)
1304 cpumask_andnot(subpartitions_cpus, subpartitions_cpus, xcpus);
1305
1306 isolcpus_updated = (old_prs != parent->partition_root_state);
1307 if (isolcpus_updated)
1308 partition_xcpus_newstate(old_prs, parent->partition_root_state,
1309 xcpus);
1310
1311 cpumask_and(xcpus, xcpus, cpu_active_mask);
1312 cpumask_or(parent->effective_cpus, parent->effective_cpus, xcpus);
1313 return isolcpus_updated;
1314}
1315
1316static void update_unbound_workqueue_cpumask(bool isolcpus_updated)
1317{
1318 int ret;
1319
1320 lockdep_assert_cpus_held();
1321
1322 if (!isolcpus_updated)
1323 return;
1324
1325 ret = workqueue_unbound_exclude_cpumask(isolated_cpus);
1326 WARN_ON_ONCE(ret < 0);
1327}
1328
1329/**
1330 * cpuset_cpu_is_isolated - Check if the given CPU is isolated
1331 * @cpu: the CPU number to be checked
1332 * Return: true if CPU is used in an isolated partition, false otherwise
1333 */
1334bool cpuset_cpu_is_isolated(int cpu)
1335{
1336 return cpumask_test_cpu(cpu, isolated_cpus);
1337}
1338EXPORT_SYMBOL_GPL(cpuset_cpu_is_isolated);
1339
1340/*
1341 * compute_effective_exclusive_cpumask - compute effective exclusive CPUs
1342 * @cs: cpuset
1343 * @xcpus: effective exclusive CPUs value to be set
1344 * Return: true if xcpus is not empty, false otherwise.
1345 *
1346 * Starting with exclusive_cpus (cpus_allowed if exclusive_cpus is not set),
1347 * it must be a subset of parent's effective_xcpus.
1348 */
1349static bool compute_effective_exclusive_cpumask(struct cpuset *cs,
1350 struct cpumask *xcpus)
1351{
1352 struct cpuset *parent = parent_cs(cs);
1353
1354 if (!xcpus)
1355 xcpus = cs->effective_xcpus;
1356
1357 return cpumask_and(xcpus, user_xcpus(cs), parent->effective_xcpus);
1358}
1359
1360static inline bool is_remote_partition(struct cpuset *cs)
1361{
1362 return !list_empty(&cs->remote_sibling);
1363}
1364
1365static inline bool is_local_partition(struct cpuset *cs)
1366{
1367 return is_partition_valid(cs) && !is_remote_partition(cs);
1368}
1369
1370/*
1371 * remote_partition_enable - Enable current cpuset as a remote partition root
1372 * @cs: the cpuset to update
1373 * @new_prs: new partition_root_state
1374 * @tmp: temporary masks
1375 * Return: 0 if successful, errcode if error
1376 *
1377 * Enable the current cpuset to become a remote partition root taking CPUs
1378 * directly from the top cpuset. cpuset_mutex must be held by the caller.
1379 */
1380static int remote_partition_enable(struct cpuset *cs, int new_prs,
1381 struct tmpmasks *tmp)
1382{
1383 bool isolcpus_updated;
1384
1385 /*
1386 * The user must have sysadmin privilege.
1387 */
1388 if (!capable(CAP_SYS_ADMIN))
1389 return PERR_ACCESS;
1390
1391 /*
1392 * The requested exclusive_cpus must not be allocated to other
1393 * partitions and it can't use up all the root's effective_cpus.
1394 *
1395 * Note that if there is any local partition root above it or
1396 * remote partition root underneath it, its exclusive_cpus must
1397 * have overlapped with subpartitions_cpus.
1398 */
1399 compute_effective_exclusive_cpumask(cs, tmp->new_cpus);
1400 if (cpumask_empty(tmp->new_cpus) ||
1401 cpumask_intersects(tmp->new_cpus, subpartitions_cpus) ||
1402 cpumask_subset(top_cpuset.effective_cpus, tmp->new_cpus))
1403 return PERR_INVCPUS;
1404
1405 spin_lock_irq(&callback_lock);
1406 isolcpus_updated = partition_xcpus_add(new_prs, NULL, tmp->new_cpus);
1407 list_add(&cs->remote_sibling, &remote_children);
1408 spin_unlock_irq(&callback_lock);
1409 update_unbound_workqueue_cpumask(isolcpus_updated);
1410
1411 /*
1412 * Propagate changes in top_cpuset's effective_cpus down the hierarchy.
1413 */
1414 cpuset_update_tasks_cpumask(&top_cpuset, tmp->new_cpus);
1415 update_sibling_cpumasks(&top_cpuset, NULL, tmp);
1416 return 0;
1417}
1418
1419/*
1420 * remote_partition_disable - Remove current cpuset from remote partition list
1421 * @cs: the cpuset to update
1422 * @tmp: temporary masks
1423 *
1424 * The effective_cpus is also updated.
1425 *
1426 * cpuset_mutex must be held by the caller.
1427 */
1428static void remote_partition_disable(struct cpuset *cs, struct tmpmasks *tmp)
1429{
1430 bool isolcpus_updated;
1431
1432 compute_effective_exclusive_cpumask(cs, tmp->new_cpus);
1433 WARN_ON_ONCE(!is_remote_partition(cs));
1434 WARN_ON_ONCE(!cpumask_subset(tmp->new_cpus, subpartitions_cpus));
1435
1436 spin_lock_irq(&callback_lock);
1437 list_del_init(&cs->remote_sibling);
1438 isolcpus_updated = partition_xcpus_del(cs->partition_root_state,
1439 NULL, tmp->new_cpus);
1440 cs->partition_root_state = -cs->partition_root_state;
1441 if (!cs->prs_err)
1442 cs->prs_err = PERR_INVCPUS;
1443 reset_partition_data(cs);
1444 spin_unlock_irq(&callback_lock);
1445 update_unbound_workqueue_cpumask(isolcpus_updated);
1446
1447 /*
1448 * Propagate changes in top_cpuset's effective_cpus down the hierarchy.
1449 */
1450 cpuset_update_tasks_cpumask(&top_cpuset, tmp->new_cpus);
1451 update_sibling_cpumasks(&top_cpuset, NULL, tmp);
1452}
1453
1454/*
1455 * remote_cpus_update - cpus_exclusive change of remote partition
1456 * @cs: the cpuset to be updated
1457 * @newmask: the new effective_xcpus mask
1458 * @tmp: temporary masks
1459 *
1460 * top_cpuset and subpartitions_cpus will be updated or partition can be
1461 * invalidated.
1462 */
1463static void remote_cpus_update(struct cpuset *cs, struct cpumask *newmask,
1464 struct tmpmasks *tmp)
1465{
1466 bool adding, deleting;
1467 int prs = cs->partition_root_state;
1468 int isolcpus_updated = 0;
1469
1470 if (WARN_ON_ONCE(!is_remote_partition(cs)))
1471 return;
1472
1473 WARN_ON_ONCE(!cpumask_subset(cs->effective_xcpus, subpartitions_cpus));
1474
1475 if (cpumask_empty(newmask))
1476 goto invalidate;
1477
1478 adding = cpumask_andnot(tmp->addmask, newmask, cs->effective_xcpus);
1479 deleting = cpumask_andnot(tmp->delmask, cs->effective_xcpus, newmask);
1480
1481 /*
1482 * Additions of remote CPUs is only allowed if those CPUs are
1483 * not allocated to other partitions and there are effective_cpus
1484 * left in the top cpuset.
1485 */
1486 if (adding && (!capable(CAP_SYS_ADMIN) ||
1487 cpumask_intersects(tmp->addmask, subpartitions_cpus) ||
1488 cpumask_subset(top_cpuset.effective_cpus, tmp->addmask)))
1489 goto invalidate;
1490
1491 spin_lock_irq(&callback_lock);
1492 if (adding)
1493 isolcpus_updated += partition_xcpus_add(prs, NULL, tmp->addmask);
1494 if (deleting)
1495 isolcpus_updated += partition_xcpus_del(prs, NULL, tmp->delmask);
1496 spin_unlock_irq(&callback_lock);
1497 update_unbound_workqueue_cpumask(isolcpus_updated);
1498
1499 /*
1500 * Propagate changes in top_cpuset's effective_cpus down the hierarchy.
1501 */
1502 cpuset_update_tasks_cpumask(&top_cpuset, tmp->new_cpus);
1503 update_sibling_cpumasks(&top_cpuset, NULL, tmp);
1504 return;
1505
1506invalidate:
1507 remote_partition_disable(cs, tmp);
1508}
1509
1510/*
1511 * remote_partition_check - check if a child remote partition needs update
1512 * @cs: the cpuset to be updated
1513 * @newmask: the new effective_xcpus mask
1514 * @delmask: temporary mask for deletion (not in tmp)
1515 * @tmp: temporary masks
1516 *
1517 * This should be called before the given cs has updated its cpus_allowed
1518 * and/or effective_xcpus.
1519 */
1520static void remote_partition_check(struct cpuset *cs, struct cpumask *newmask,
1521 struct cpumask *delmask, struct tmpmasks *tmp)
1522{
1523 struct cpuset *child, *next;
1524 int disable_cnt = 0;
1525
1526 /*
1527 * Compute the effective exclusive CPUs that will be deleted.
1528 */
1529 if (!cpumask_andnot(delmask, cs->effective_xcpus, newmask) ||
1530 !cpumask_intersects(delmask, subpartitions_cpus))
1531 return; /* No deletion of exclusive CPUs in partitions */
1532
1533 /*
1534 * Searching the remote children list to look for those that will
1535 * be impacted by the deletion of exclusive CPUs.
1536 *
1537 * Since a cpuset must be removed from the remote children list
1538 * before it can go offline and holding cpuset_mutex will prevent
1539 * any change in cpuset status. RCU read lock isn't needed.
1540 */
1541 lockdep_assert_held(&cpuset_mutex);
1542 list_for_each_entry_safe(child, next, &remote_children, remote_sibling)
1543 if (cpumask_intersects(child->effective_cpus, delmask)) {
1544 remote_partition_disable(child, tmp);
1545 disable_cnt++;
1546 }
1547 if (disable_cnt)
1548 cpuset_force_rebuild();
1549}
1550
1551/*
1552 * prstate_housekeeping_conflict - check for partition & housekeeping conflicts
1553 * @prstate: partition root state to be checked
1554 * @new_cpus: cpu mask
1555 * Return: true if there is conflict, false otherwise
1556 *
1557 * CPUs outside of boot_hk_cpus, if defined, can only be used in an
1558 * isolated partition.
1559 */
1560static bool prstate_housekeeping_conflict(int prstate, struct cpumask *new_cpus)
1561{
1562 if (!have_boot_isolcpus)
1563 return false;
1564
1565 if ((prstate != PRS_ISOLATED) && !cpumask_subset(new_cpus, boot_hk_cpus))
1566 return true;
1567
1568 return false;
1569}
1570
1571/**
1572 * update_parent_effective_cpumask - update effective_cpus mask of parent cpuset
1573 * @cs: The cpuset that requests change in partition root state
1574 * @cmd: Partition root state change command
1575 * @newmask: Optional new cpumask for partcmd_update
1576 * @tmp: Temporary addmask and delmask
1577 * Return: 0 or a partition root state error code
1578 *
1579 * For partcmd_enable*, the cpuset is being transformed from a non-partition
1580 * root to a partition root. The effective_xcpus (cpus_allowed if
1581 * effective_xcpus not set) mask of the given cpuset will be taken away from
1582 * parent's effective_cpus. The function will return 0 if all the CPUs listed
1583 * in effective_xcpus can be granted or an error code will be returned.
1584 *
1585 * For partcmd_disable, the cpuset is being transformed from a partition
1586 * root back to a non-partition root. Any CPUs in effective_xcpus will be
1587 * given back to parent's effective_cpus. 0 will always be returned.
1588 *
1589 * For partcmd_update, if the optional newmask is specified, the cpu list is
1590 * to be changed from effective_xcpus to newmask. Otherwise, effective_xcpus is
1591 * assumed to remain the same. The cpuset should either be a valid or invalid
1592 * partition root. The partition root state may change from valid to invalid
1593 * or vice versa. An error code will be returned if transitioning from
1594 * invalid to valid violates the exclusivity rule.
1595 *
1596 * For partcmd_invalidate, the current partition will be made invalid.
1597 *
1598 * The partcmd_enable* and partcmd_disable commands are used by
1599 * update_prstate(). An error code may be returned and the caller will check
1600 * for error.
1601 *
1602 * The partcmd_update command is used by update_cpumasks_hier() with newmask
1603 * NULL and update_cpumask() with newmask set. The partcmd_invalidate is used
1604 * by update_cpumask() with NULL newmask. In both cases, the callers won't
1605 * check for error and so partition_root_state and prs_error will be updated
1606 * directly.
1607 */
1608static int update_parent_effective_cpumask(struct cpuset *cs, int cmd,
1609 struct cpumask *newmask,
1610 struct tmpmasks *tmp)
1611{
1612 struct cpuset *parent = parent_cs(cs);
1613 int adding; /* Adding cpus to parent's effective_cpus */
1614 int deleting; /* Deleting cpus from parent's effective_cpus */
1615 int old_prs, new_prs;
1616 int part_error = PERR_NONE; /* Partition error? */
1617 int subparts_delta = 0;
1618 struct cpumask *xcpus; /* cs effective_xcpus */
1619 int isolcpus_updated = 0;
1620 bool nocpu;
1621
1622 lockdep_assert_held(&cpuset_mutex);
1623
1624 /*
1625 * new_prs will only be changed for the partcmd_update and
1626 * partcmd_invalidate commands.
1627 */
1628 adding = deleting = false;
1629 old_prs = new_prs = cs->partition_root_state;
1630 xcpus = user_xcpus(cs);
1631
1632 if (cmd == partcmd_invalidate) {
1633 if (is_prs_invalid(old_prs))
1634 return 0;
1635
1636 /*
1637 * Make the current partition invalid.
1638 */
1639 if (is_partition_valid(parent))
1640 adding = cpumask_and(tmp->addmask,
1641 xcpus, parent->effective_xcpus);
1642 if (old_prs > 0) {
1643 new_prs = -old_prs;
1644 subparts_delta--;
1645 }
1646 goto write_error;
1647 }
1648
1649 /*
1650 * The parent must be a partition root.
1651 * The new cpumask, if present, or the current cpus_allowed must
1652 * not be empty.
1653 */
1654 if (!is_partition_valid(parent)) {
1655 return is_partition_invalid(parent)
1656 ? PERR_INVPARENT : PERR_NOTPART;
1657 }
1658 if (!newmask && xcpus_empty(cs))
1659 return PERR_CPUSEMPTY;
1660
1661 nocpu = tasks_nocpu_error(parent, cs, xcpus);
1662
1663 if ((cmd == partcmd_enable) || (cmd == partcmd_enablei)) {
1664 /*
1665 * Enabling partition root is not allowed if its
1666 * effective_xcpus is empty or doesn't overlap with
1667 * parent's effective_xcpus.
1668 */
1669 if (cpumask_empty(xcpus) ||
1670 !cpumask_intersects(xcpus, parent->effective_xcpus))
1671 return PERR_INVCPUS;
1672
1673 if (prstate_housekeeping_conflict(new_prs, xcpus))
1674 return PERR_HKEEPING;
1675
1676 /*
1677 * A parent can be left with no CPU as long as there is no
1678 * task directly associated with the parent partition.
1679 */
1680 if (nocpu)
1681 return PERR_NOCPUS;
1682
1683 cpumask_copy(tmp->delmask, xcpus);
1684 deleting = true;
1685 subparts_delta++;
1686 new_prs = (cmd == partcmd_enable) ? PRS_ROOT : PRS_ISOLATED;
1687 } else if (cmd == partcmd_disable) {
1688 /*
1689 * May need to add cpus to parent's effective_cpus for
1690 * valid partition root.
1691 */
1692 adding = !is_prs_invalid(old_prs) &&
1693 cpumask_and(tmp->addmask, xcpus, parent->effective_xcpus);
1694 if (adding)
1695 subparts_delta--;
1696 new_prs = PRS_MEMBER;
1697 } else if (newmask) {
1698 /*
1699 * Empty cpumask is not allowed
1700 */
1701 if (cpumask_empty(newmask)) {
1702 part_error = PERR_CPUSEMPTY;
1703 goto write_error;
1704 }
1705 /* Check newmask again, whether cpus are available for parent/cs */
1706 nocpu |= tasks_nocpu_error(parent, cs, newmask);
1707
1708 /*
1709 * partcmd_update with newmask:
1710 *
1711 * Compute add/delete mask to/from effective_cpus
1712 *
1713 * For valid partition:
1714 * addmask = exclusive_cpus & ~newmask
1715 * & parent->effective_xcpus
1716 * delmask = newmask & ~exclusive_cpus
1717 * & parent->effective_xcpus
1718 *
1719 * For invalid partition:
1720 * delmask = newmask & parent->effective_xcpus
1721 */
1722 if (is_prs_invalid(old_prs)) {
1723 adding = false;
1724 deleting = cpumask_and(tmp->delmask,
1725 newmask, parent->effective_xcpus);
1726 } else {
1727 cpumask_andnot(tmp->addmask, xcpus, newmask);
1728 adding = cpumask_and(tmp->addmask, tmp->addmask,
1729 parent->effective_xcpus);
1730
1731 cpumask_andnot(tmp->delmask, newmask, xcpus);
1732 deleting = cpumask_and(tmp->delmask, tmp->delmask,
1733 parent->effective_xcpus);
1734 }
1735 /*
1736 * Make partition invalid if parent's effective_cpus could
1737 * become empty and there are tasks in the parent.
1738 */
1739 if (nocpu && (!adding ||
1740 !cpumask_intersects(tmp->addmask, cpu_active_mask))) {
1741 part_error = PERR_NOCPUS;
1742 deleting = false;
1743 adding = cpumask_and(tmp->addmask,
1744 xcpus, parent->effective_xcpus);
1745 }
1746 } else {
1747 /*
1748 * partcmd_update w/o newmask
1749 *
1750 * delmask = effective_xcpus & parent->effective_cpus
1751 *
1752 * This can be called from:
1753 * 1) update_cpumasks_hier()
1754 * 2) cpuset_hotplug_update_tasks()
1755 *
1756 * Check to see if it can be transitioned from valid to
1757 * invalid partition or vice versa.
1758 *
1759 * A partition error happens when parent has tasks and all
1760 * its effective CPUs will have to be distributed out.
1761 */
1762 WARN_ON_ONCE(!is_partition_valid(parent));
1763 if (nocpu) {
1764 part_error = PERR_NOCPUS;
1765 if (is_partition_valid(cs))
1766 adding = cpumask_and(tmp->addmask,
1767 xcpus, parent->effective_xcpus);
1768 } else if (is_partition_invalid(cs) &&
1769 cpumask_subset(xcpus, parent->effective_xcpus)) {
1770 struct cgroup_subsys_state *css;
1771 struct cpuset *child;
1772 bool exclusive = true;
1773
1774 /*
1775 * Convert invalid partition to valid has to
1776 * pass the cpu exclusivity test.
1777 */
1778 rcu_read_lock();
1779 cpuset_for_each_child(child, css, parent) {
1780 if (child == cs)
1781 continue;
1782 if (!cpusets_are_exclusive(cs, child)) {
1783 exclusive = false;
1784 break;
1785 }
1786 }
1787 rcu_read_unlock();
1788 if (exclusive)
1789 deleting = cpumask_and(tmp->delmask,
1790 xcpus, parent->effective_cpus);
1791 else
1792 part_error = PERR_NOTEXCL;
1793 }
1794 }
1795
1796write_error:
1797 if (part_error)
1798 WRITE_ONCE(cs->prs_err, part_error);
1799
1800 if (cmd == partcmd_update) {
1801 /*
1802 * Check for possible transition between valid and invalid
1803 * partition root.
1804 */
1805 switch (cs->partition_root_state) {
1806 case PRS_ROOT:
1807 case PRS_ISOLATED:
1808 if (part_error) {
1809 new_prs = -old_prs;
1810 subparts_delta--;
1811 }
1812 break;
1813 case PRS_INVALID_ROOT:
1814 case PRS_INVALID_ISOLATED:
1815 if (!part_error) {
1816 new_prs = -old_prs;
1817 subparts_delta++;
1818 }
1819 break;
1820 }
1821 }
1822
1823 if (!adding && !deleting && (new_prs == old_prs))
1824 return 0;
1825
1826 /*
1827 * Transitioning between invalid to valid or vice versa may require
1828 * changing CS_CPU_EXCLUSIVE. In the case of partcmd_update,
1829 * validate_change() has already been successfully called and
1830 * CPU lists in cs haven't been updated yet. So defer it to later.
1831 */
1832 if ((old_prs != new_prs) && (cmd != partcmd_update)) {
1833 int err = update_partition_exclusive(cs, new_prs);
1834
1835 if (err)
1836 return err;
1837 }
1838
1839 /*
1840 * Change the parent's effective_cpus & effective_xcpus (top cpuset
1841 * only).
1842 *
1843 * Newly added CPUs will be removed from effective_cpus and
1844 * newly deleted ones will be added back to effective_cpus.
1845 */
1846 spin_lock_irq(&callback_lock);
1847 if (old_prs != new_prs) {
1848 cs->partition_root_state = new_prs;
1849 if (new_prs <= 0)
1850 cs->nr_subparts = 0;
1851 }
1852 /*
1853 * Adding to parent's effective_cpus means deletion CPUs from cs
1854 * and vice versa.
1855 */
1856 if (adding)
1857 isolcpus_updated += partition_xcpus_del(old_prs, parent,
1858 tmp->addmask);
1859 if (deleting)
1860 isolcpus_updated += partition_xcpus_add(new_prs, parent,
1861 tmp->delmask);
1862
1863 if (is_partition_valid(parent)) {
1864 parent->nr_subparts += subparts_delta;
1865 WARN_ON_ONCE(parent->nr_subparts < 0);
1866 }
1867 spin_unlock_irq(&callback_lock);
1868 update_unbound_workqueue_cpumask(isolcpus_updated);
1869
1870 if ((old_prs != new_prs) && (cmd == partcmd_update))
1871 update_partition_exclusive(cs, new_prs);
1872
1873 if (adding || deleting) {
1874 cpuset_update_tasks_cpumask(parent, tmp->addmask);
1875 update_sibling_cpumasks(parent, cs, tmp);
1876 }
1877
1878 /*
1879 * For partcmd_update without newmask, it is being called from
1880 * cpuset_handle_hotplug(). Update the load balance flag and
1881 * scheduling domain accordingly.
1882 */
1883 if ((cmd == partcmd_update) && !newmask)
1884 update_partition_sd_lb(cs, old_prs);
1885
1886 notify_partition_change(cs, old_prs);
1887 return 0;
1888}
1889
1890/**
1891 * compute_partition_effective_cpumask - compute effective_cpus for partition
1892 * @cs: partition root cpuset
1893 * @new_ecpus: previously computed effective_cpus to be updated
1894 *
1895 * Compute the effective_cpus of a partition root by scanning effective_xcpus
1896 * of child partition roots and excluding their effective_xcpus.
1897 *
1898 * This has the side effect of invalidating valid child partition roots,
1899 * if necessary. Since it is called from either cpuset_hotplug_update_tasks()
1900 * or update_cpumasks_hier() where parent and children are modified
1901 * successively, we don't need to call update_parent_effective_cpumask()
1902 * and the child's effective_cpus will be updated in later iterations.
1903 *
1904 * Note that rcu_read_lock() is assumed to be held.
1905 */
1906static void compute_partition_effective_cpumask(struct cpuset *cs,
1907 struct cpumask *new_ecpus)
1908{
1909 struct cgroup_subsys_state *css;
1910 struct cpuset *child;
1911 bool populated = partition_is_populated(cs, NULL);
1912
1913 /*
1914 * Check child partition roots to see if they should be
1915 * invalidated when
1916 * 1) child effective_xcpus not a subset of new
1917 * excluisve_cpus
1918 * 2) All the effective_cpus will be used up and cp
1919 * has tasks
1920 */
1921 compute_effective_exclusive_cpumask(cs, new_ecpus);
1922 cpumask_and(new_ecpus, new_ecpus, cpu_active_mask);
1923
1924 rcu_read_lock();
1925 cpuset_for_each_child(child, css, cs) {
1926 if (!is_partition_valid(child))
1927 continue;
1928
1929 child->prs_err = 0;
1930 if (!cpumask_subset(child->effective_xcpus,
1931 cs->effective_xcpus))
1932 child->prs_err = PERR_INVCPUS;
1933 else if (populated &&
1934 cpumask_subset(new_ecpus, child->effective_xcpus))
1935 child->prs_err = PERR_NOCPUS;
1936
1937 if (child->prs_err) {
1938 int old_prs = child->partition_root_state;
1939
1940 /*
1941 * Invalidate child partition
1942 */
1943 spin_lock_irq(&callback_lock);
1944 make_partition_invalid(child);
1945 cs->nr_subparts--;
1946 child->nr_subparts = 0;
1947 spin_unlock_irq(&callback_lock);
1948 notify_partition_change(child, old_prs);
1949 continue;
1950 }
1951 cpumask_andnot(new_ecpus, new_ecpus,
1952 child->effective_xcpus);
1953 }
1954 rcu_read_unlock();
1955}
1956
1957/*
1958 * update_cpumasks_hier - Update effective cpumasks and tasks in the subtree
1959 * @cs: the cpuset to consider
1960 * @tmp: temp variables for calculating effective_cpus & partition setup
1961 * @force: don't skip any descendant cpusets if set
1962 *
1963 * When configured cpumask is changed, the effective cpumasks of this cpuset
1964 * and all its descendants need to be updated.
1965 *
1966 * On legacy hierarchy, effective_cpus will be the same with cpu_allowed.
1967 *
1968 * Called with cpuset_mutex held
1969 */
1970static void update_cpumasks_hier(struct cpuset *cs, struct tmpmasks *tmp,
1971 bool force)
1972{
1973 struct cpuset *cp;
1974 struct cgroup_subsys_state *pos_css;
1975 bool need_rebuild_sched_domains = false;
1976 int old_prs, new_prs;
1977
1978 rcu_read_lock();
1979 cpuset_for_each_descendant_pre(cp, pos_css, cs) {
1980 struct cpuset *parent = parent_cs(cp);
1981 bool remote = is_remote_partition(cp);
1982 bool update_parent = false;
1983
1984 /*
1985 * Skip descendent remote partition that acquires CPUs
1986 * directly from top cpuset unless it is cs.
1987 */
1988 if (remote && (cp != cs)) {
1989 pos_css = css_rightmost_descendant(pos_css);
1990 continue;
1991 }
1992
1993 /*
1994 * Update effective_xcpus if exclusive_cpus set.
1995 * The case when exclusive_cpus isn't set is handled later.
1996 */
1997 if (!cpumask_empty(cp->exclusive_cpus) && (cp != cs)) {
1998 spin_lock_irq(&callback_lock);
1999 compute_effective_exclusive_cpumask(cp, NULL);
2000 spin_unlock_irq(&callback_lock);
2001 }
2002
2003 old_prs = new_prs = cp->partition_root_state;
2004 if (remote || (is_partition_valid(parent) &&
2005 is_partition_valid(cp)))
2006 compute_partition_effective_cpumask(cp, tmp->new_cpus);
2007 else
2008 compute_effective_cpumask(tmp->new_cpus, cp, parent);
2009
2010 /*
2011 * A partition with no effective_cpus is allowed as long as
2012 * there is no task associated with it. Call
2013 * update_parent_effective_cpumask() to check it.
2014 */
2015 if (is_partition_valid(cp) && cpumask_empty(tmp->new_cpus)) {
2016 update_parent = true;
2017 goto update_parent_effective;
2018 }
2019
2020 /*
2021 * If it becomes empty, inherit the effective mask of the
2022 * parent, which is guaranteed to have some CPUs unless
2023 * it is a partition root that has explicitly distributed
2024 * out all its CPUs.
2025 */
2026 if (is_in_v2_mode() && !remote && cpumask_empty(tmp->new_cpus))
2027 cpumask_copy(tmp->new_cpus, parent->effective_cpus);
2028
2029 if (remote)
2030 goto get_css;
2031
2032 /*
2033 * Skip the whole subtree if
2034 * 1) the cpumask remains the same,
2035 * 2) has no partition root state,
2036 * 3) force flag not set, and
2037 * 4) for v2 load balance state same as its parent.
2038 */
2039 if (!cp->partition_root_state && !force &&
2040 cpumask_equal(tmp->new_cpus, cp->effective_cpus) &&
2041 (!cpuset_v2() ||
2042 (is_sched_load_balance(parent) == is_sched_load_balance(cp)))) {
2043 pos_css = css_rightmost_descendant(pos_css);
2044 continue;
2045 }
2046
2047update_parent_effective:
2048 /*
2049 * update_parent_effective_cpumask() should have been called
2050 * for cs already in update_cpumask(). We should also call
2051 * cpuset_update_tasks_cpumask() again for tasks in the parent
2052 * cpuset if the parent's effective_cpus changes.
2053 */
2054 if ((cp != cs) && old_prs) {
2055 switch (parent->partition_root_state) {
2056 case PRS_ROOT:
2057 case PRS_ISOLATED:
2058 update_parent = true;
2059 break;
2060
2061 default:
2062 /*
2063 * When parent is not a partition root or is
2064 * invalid, child partition roots become
2065 * invalid too.
2066 */
2067 if (is_partition_valid(cp))
2068 new_prs = -cp->partition_root_state;
2069 WRITE_ONCE(cp->prs_err,
2070 is_partition_invalid(parent)
2071 ? PERR_INVPARENT : PERR_NOTPART);
2072 break;
2073 }
2074 }
2075get_css:
2076 if (!css_tryget_online(&cp->css))
2077 continue;
2078 rcu_read_unlock();
2079
2080 if (update_parent) {
2081 update_parent_effective_cpumask(cp, partcmd_update, NULL, tmp);
2082 /*
2083 * The cpuset partition_root_state may become
2084 * invalid. Capture it.
2085 */
2086 new_prs = cp->partition_root_state;
2087 }
2088
2089 spin_lock_irq(&callback_lock);
2090 cpumask_copy(cp->effective_cpus, tmp->new_cpus);
2091 cp->partition_root_state = new_prs;
2092 /*
2093 * Make sure effective_xcpus is properly set for a valid
2094 * partition root.
2095 */
2096 if ((new_prs > 0) && cpumask_empty(cp->exclusive_cpus))
2097 cpumask_and(cp->effective_xcpus,
2098 cp->cpus_allowed, parent->effective_xcpus);
2099 else if (new_prs < 0)
2100 reset_partition_data(cp);
2101 spin_unlock_irq(&callback_lock);
2102
2103 notify_partition_change(cp, old_prs);
2104
2105 WARN_ON(!is_in_v2_mode() &&
2106 !cpumask_equal(cp->cpus_allowed, cp->effective_cpus));
2107
2108 cpuset_update_tasks_cpumask(cp, cp->effective_cpus);
2109
2110 /*
2111 * On default hierarchy, inherit the CS_SCHED_LOAD_BALANCE
2112 * from parent if current cpuset isn't a valid partition root
2113 * and their load balance states differ.
2114 */
2115 if (cpuset_v2() && !is_partition_valid(cp) &&
2116 (is_sched_load_balance(parent) != is_sched_load_balance(cp))) {
2117 if (is_sched_load_balance(parent))
2118 set_bit(CS_SCHED_LOAD_BALANCE, &cp->flags);
2119 else
2120 clear_bit(CS_SCHED_LOAD_BALANCE, &cp->flags);
2121 }
2122
2123 /*
2124 * On legacy hierarchy, if the effective cpumask of any non-
2125 * empty cpuset is changed, we need to rebuild sched domains.
2126 * On default hierarchy, the cpuset needs to be a partition
2127 * root as well.
2128 */
2129 if (!cpumask_empty(cp->cpus_allowed) &&
2130 is_sched_load_balance(cp) &&
2131 (!cpuset_v2() || is_partition_valid(cp)))
2132 need_rebuild_sched_domains = true;
2133
2134 rcu_read_lock();
2135 css_put(&cp->css);
2136 }
2137 rcu_read_unlock();
2138
2139 if (need_rebuild_sched_domains)
2140 cpuset_force_rebuild();
2141}
2142
2143/**
2144 * update_sibling_cpumasks - Update siblings cpumasks
2145 * @parent: Parent cpuset
2146 * @cs: Current cpuset
2147 * @tmp: Temp variables
2148 */
2149static void update_sibling_cpumasks(struct cpuset *parent, struct cpuset *cs,
2150 struct tmpmasks *tmp)
2151{
2152 struct cpuset *sibling;
2153 struct cgroup_subsys_state *pos_css;
2154
2155 lockdep_assert_held(&cpuset_mutex);
2156
2157 /*
2158 * Check all its siblings and call update_cpumasks_hier()
2159 * if their effective_cpus will need to be changed.
2160 *
2161 * It is possible a change in parent's effective_cpus
2162 * due to a change in a child partition's effective_xcpus will impact
2163 * its siblings even if they do not inherit parent's effective_cpus
2164 * directly.
2165 *
2166 * The update_cpumasks_hier() function may sleep. So we have to
2167 * release the RCU read lock before calling it.
2168 */
2169 rcu_read_lock();
2170 cpuset_for_each_child(sibling, pos_css, parent) {
2171 if (sibling == cs)
2172 continue;
2173 if (!is_partition_valid(sibling)) {
2174 compute_effective_cpumask(tmp->new_cpus, sibling,
2175 parent);
2176 if (cpumask_equal(tmp->new_cpus, sibling->effective_cpus))
2177 continue;
2178 }
2179 if (!css_tryget_online(&sibling->css))
2180 continue;
2181
2182 rcu_read_unlock();
2183 update_cpumasks_hier(sibling, tmp, false);
2184 rcu_read_lock();
2185 css_put(&sibling->css);
2186 }
2187 rcu_read_unlock();
2188}
2189
2190/**
2191 * update_cpumask - update the cpus_allowed mask of a cpuset and all tasks in it
2192 * @cs: the cpuset to consider
2193 * @trialcs: trial cpuset
2194 * @buf: buffer of cpu numbers written to this cpuset
2195 */
2196static int update_cpumask(struct cpuset *cs, struct cpuset *trialcs,
2197 const char *buf)
2198{
2199 int retval;
2200 struct tmpmasks tmp;
2201 struct cpuset *parent = parent_cs(cs);
2202 bool invalidate = false;
2203 bool force = false;
2204 int old_prs = cs->partition_root_state;
2205
2206 /* top_cpuset.cpus_allowed tracks cpu_online_mask; it's read-only */
2207 if (cs == &top_cpuset)
2208 return -EACCES;
2209
2210 /*
2211 * An empty cpus_allowed is ok only if the cpuset has no tasks.
2212 * Since cpulist_parse() fails on an empty mask, we special case
2213 * that parsing. The validate_change() call ensures that cpusets
2214 * with tasks have cpus.
2215 */
2216 if (!*buf) {
2217 cpumask_clear(trialcs->cpus_allowed);
2218 if (cpumask_empty(trialcs->exclusive_cpus))
2219 cpumask_clear(trialcs->effective_xcpus);
2220 } else {
2221 retval = cpulist_parse(buf, trialcs->cpus_allowed);
2222 if (retval < 0)
2223 return retval;
2224
2225 if (!cpumask_subset(trialcs->cpus_allowed,
2226 top_cpuset.cpus_allowed))
2227 return -EINVAL;
2228
2229 /*
2230 * When exclusive_cpus isn't explicitly set, it is constrained
2231 * by cpus_allowed and parent's effective_xcpus. Otherwise,
2232 * trialcs->effective_xcpus is used as a temporary cpumask
2233 * for checking validity of the partition root.
2234 */
2235 if (!cpumask_empty(trialcs->exclusive_cpus) || is_partition_valid(cs))
2236 compute_effective_exclusive_cpumask(trialcs, NULL);
2237 }
2238
2239 /* Nothing to do if the cpus didn't change */
2240 if (cpumask_equal(cs->cpus_allowed, trialcs->cpus_allowed))
2241 return 0;
2242
2243 if (alloc_cpumasks(NULL, &tmp))
2244 return -ENOMEM;
2245
2246 if (old_prs) {
2247 if (is_partition_valid(cs) &&
2248 cpumask_empty(trialcs->effective_xcpus)) {
2249 invalidate = true;
2250 cs->prs_err = PERR_INVCPUS;
2251 } else if (prstate_housekeeping_conflict(old_prs, trialcs->effective_xcpus)) {
2252 invalidate = true;
2253 cs->prs_err = PERR_HKEEPING;
2254 } else if (tasks_nocpu_error(parent, cs, trialcs->effective_xcpus)) {
2255 invalidate = true;
2256 cs->prs_err = PERR_NOCPUS;
2257 }
2258 }
2259
2260 /*
2261 * Check all the descendants in update_cpumasks_hier() if
2262 * effective_xcpus is to be changed.
2263 */
2264 force = !cpumask_equal(cs->effective_xcpus, trialcs->effective_xcpus);
2265
2266 retval = validate_change(cs, trialcs);
2267
2268 if ((retval == -EINVAL) && cpuset_v2()) {
2269 struct cgroup_subsys_state *css;
2270 struct cpuset *cp;
2271
2272 /*
2273 * The -EINVAL error code indicates that partition sibling
2274 * CPU exclusivity rule has been violated. We still allow
2275 * the cpumask change to proceed while invalidating the
2276 * partition. However, any conflicting sibling partitions
2277 * have to be marked as invalid too.
2278 */
2279 invalidate = true;
2280 rcu_read_lock();
2281 cpuset_for_each_child(cp, css, parent) {
2282 struct cpumask *xcpus = user_xcpus(trialcs);
2283
2284 if (is_partition_valid(cp) &&
2285 cpumask_intersects(xcpus, cp->effective_xcpus)) {
2286 rcu_read_unlock();
2287 update_parent_effective_cpumask(cp, partcmd_invalidate, NULL, &tmp);
2288 rcu_read_lock();
2289 }
2290 }
2291 rcu_read_unlock();
2292 retval = 0;
2293 }
2294
2295 if (retval < 0)
2296 goto out_free;
2297
2298 if (is_partition_valid(cs) ||
2299 (is_partition_invalid(cs) && !invalidate)) {
2300 struct cpumask *xcpus = trialcs->effective_xcpus;
2301
2302 if (cpumask_empty(xcpus) && is_partition_invalid(cs))
2303 xcpus = trialcs->cpus_allowed;
2304
2305 /*
2306 * Call remote_cpus_update() to handle valid remote partition
2307 */
2308 if (is_remote_partition(cs))
2309 remote_cpus_update(cs, xcpus, &tmp);
2310 else if (invalidate)
2311 update_parent_effective_cpumask(cs, partcmd_invalidate,
2312 NULL, &tmp);
2313 else
2314 update_parent_effective_cpumask(cs, partcmd_update,
2315 xcpus, &tmp);
2316 } else if (!cpumask_empty(cs->exclusive_cpus)) {
2317 /*
2318 * Use trialcs->effective_cpus as a temp cpumask
2319 */
2320 remote_partition_check(cs, trialcs->effective_xcpus,
2321 trialcs->effective_cpus, &tmp);
2322 }
2323
2324 spin_lock_irq(&callback_lock);
2325 cpumask_copy(cs->cpus_allowed, trialcs->cpus_allowed);
2326 cpumask_copy(cs->effective_xcpus, trialcs->effective_xcpus);
2327 if ((old_prs > 0) && !is_partition_valid(cs))
2328 reset_partition_data(cs);
2329 spin_unlock_irq(&callback_lock);
2330
2331 /* effective_cpus/effective_xcpus will be updated here */
2332 update_cpumasks_hier(cs, &tmp, force);
2333
2334 /* Update CS_SCHED_LOAD_BALANCE and/or sched_domains, if necessary */
2335 if (cs->partition_root_state)
2336 update_partition_sd_lb(cs, old_prs);
2337out_free:
2338 free_cpumasks(NULL, &tmp);
2339 return retval;
2340}
2341
2342/**
2343 * update_exclusive_cpumask - update the exclusive_cpus mask of a cpuset
2344 * @cs: the cpuset to consider
2345 * @trialcs: trial cpuset
2346 * @buf: buffer of cpu numbers written to this cpuset
2347 *
2348 * The tasks' cpumask will be updated if cs is a valid partition root.
2349 */
2350static int update_exclusive_cpumask(struct cpuset *cs, struct cpuset *trialcs,
2351 const char *buf)
2352{
2353 int retval;
2354 struct tmpmasks tmp;
2355 struct cpuset *parent = parent_cs(cs);
2356 bool invalidate = false;
2357 bool force = false;
2358 int old_prs = cs->partition_root_state;
2359
2360 if (!*buf) {
2361 cpumask_clear(trialcs->exclusive_cpus);
2362 cpumask_clear(trialcs->effective_xcpus);
2363 } else {
2364 retval = cpulist_parse(buf, trialcs->exclusive_cpus);
2365 if (retval < 0)
2366 return retval;
2367 }
2368
2369 /* Nothing to do if the CPUs didn't change */
2370 if (cpumask_equal(cs->exclusive_cpus, trialcs->exclusive_cpus))
2371 return 0;
2372
2373 if (*buf)
2374 compute_effective_exclusive_cpumask(trialcs, NULL);
2375
2376 /*
2377 * Check all the descendants in update_cpumasks_hier() if
2378 * effective_xcpus is to be changed.
2379 */
2380 force = !cpumask_equal(cs->effective_xcpus, trialcs->effective_xcpus);
2381
2382 retval = validate_change(cs, trialcs);
2383 if (retval)
2384 return retval;
2385
2386 if (alloc_cpumasks(NULL, &tmp))
2387 return -ENOMEM;
2388
2389 if (old_prs) {
2390 if (cpumask_empty(trialcs->effective_xcpus)) {
2391 invalidate = true;
2392 cs->prs_err = PERR_INVCPUS;
2393 } else if (prstate_housekeeping_conflict(old_prs, trialcs->effective_xcpus)) {
2394 invalidate = true;
2395 cs->prs_err = PERR_HKEEPING;
2396 } else if (tasks_nocpu_error(parent, cs, trialcs->effective_xcpus)) {
2397 invalidate = true;
2398 cs->prs_err = PERR_NOCPUS;
2399 }
2400
2401 if (is_remote_partition(cs)) {
2402 if (invalidate)
2403 remote_partition_disable(cs, &tmp);
2404 else
2405 remote_cpus_update(cs, trialcs->effective_xcpus,
2406 &tmp);
2407 } else if (invalidate) {
2408 update_parent_effective_cpumask(cs, partcmd_invalidate,
2409 NULL, &tmp);
2410 } else {
2411 update_parent_effective_cpumask(cs, partcmd_update,
2412 trialcs->effective_xcpus, &tmp);
2413 }
2414 } else if (!cpumask_empty(trialcs->exclusive_cpus)) {
2415 /*
2416 * Use trialcs->effective_cpus as a temp cpumask
2417 */
2418 remote_partition_check(cs, trialcs->effective_xcpus,
2419 trialcs->effective_cpus, &tmp);
2420 }
2421 spin_lock_irq(&callback_lock);
2422 cpumask_copy(cs->exclusive_cpus, trialcs->exclusive_cpus);
2423 cpumask_copy(cs->effective_xcpus, trialcs->effective_xcpus);
2424 if ((old_prs > 0) && !is_partition_valid(cs))
2425 reset_partition_data(cs);
2426 spin_unlock_irq(&callback_lock);
2427
2428 /*
2429 * Call update_cpumasks_hier() to update effective_cpus/effective_xcpus
2430 * of the subtree when it is a valid partition root or effective_xcpus
2431 * is updated.
2432 */
2433 if (is_partition_valid(cs) || force)
2434 update_cpumasks_hier(cs, &tmp, force);
2435
2436 /* Update CS_SCHED_LOAD_BALANCE and/or sched_domains, if necessary */
2437 if (cs->partition_root_state)
2438 update_partition_sd_lb(cs, old_prs);
2439
2440 free_cpumasks(NULL, &tmp);
2441 return 0;
2442}
2443
2444/*
2445 * Migrate memory region from one set of nodes to another. This is
2446 * performed asynchronously as it can be called from process migration path
2447 * holding locks involved in process management. All mm migrations are
2448 * performed in the queued order and can be waited for by flushing
2449 * cpuset_migrate_mm_wq.
2450 */
2451
2452struct cpuset_migrate_mm_work {
2453 struct work_struct work;
2454 struct mm_struct *mm;
2455 nodemask_t from;
2456 nodemask_t to;
2457};
2458
2459static void cpuset_migrate_mm_workfn(struct work_struct *work)
2460{
2461 struct cpuset_migrate_mm_work *mwork =
2462 container_of(work, struct cpuset_migrate_mm_work, work);
2463
2464 /* on a wq worker, no need to worry about %current's mems_allowed */
2465 do_migrate_pages(mwork->mm, &mwork->from, &mwork->to, MPOL_MF_MOVE_ALL);
2466 mmput(mwork->mm);
2467 kfree(mwork);
2468}
2469
2470static void cpuset_migrate_mm(struct mm_struct *mm, const nodemask_t *from,
2471 const nodemask_t *to)
2472{
2473 struct cpuset_migrate_mm_work *mwork;
2474
2475 if (nodes_equal(*from, *to)) {
2476 mmput(mm);
2477 return;
2478 }
2479
2480 mwork = kzalloc(sizeof(*mwork), GFP_KERNEL);
2481 if (mwork) {
2482 mwork->mm = mm;
2483 mwork->from = *from;
2484 mwork->to = *to;
2485 INIT_WORK(&mwork->work, cpuset_migrate_mm_workfn);
2486 queue_work(cpuset_migrate_mm_wq, &mwork->work);
2487 } else {
2488 mmput(mm);
2489 }
2490}
2491
2492static void cpuset_post_attach(void)
2493{
2494 flush_workqueue(cpuset_migrate_mm_wq);
2495}
2496
2497/*
2498 * cpuset_change_task_nodemask - change task's mems_allowed and mempolicy
2499 * @tsk: the task to change
2500 * @newmems: new nodes that the task will be set
2501 *
2502 * We use the mems_allowed_seq seqlock to safely update both tsk->mems_allowed
2503 * and rebind an eventual tasks' mempolicy. If the task is allocating in
2504 * parallel, it might temporarily see an empty intersection, which results in
2505 * a seqlock check and retry before OOM or allocation failure.
2506 */
2507static void cpuset_change_task_nodemask(struct task_struct *tsk,
2508 nodemask_t *newmems)
2509{
2510 task_lock(tsk);
2511
2512 local_irq_disable();
2513 write_seqcount_begin(&tsk->mems_allowed_seq);
2514
2515 nodes_or(tsk->mems_allowed, tsk->mems_allowed, *newmems);
2516 mpol_rebind_task(tsk, newmems);
2517 tsk->mems_allowed = *newmems;
2518
2519 write_seqcount_end(&tsk->mems_allowed_seq);
2520 local_irq_enable();
2521
2522 task_unlock(tsk);
2523}
2524
2525static void *cpuset_being_rebound;
2526
2527/**
2528 * cpuset_update_tasks_nodemask - Update the nodemasks of tasks in the cpuset.
2529 * @cs: the cpuset in which each task's mems_allowed mask needs to be changed
2530 *
2531 * Iterate through each task of @cs updating its mems_allowed to the
2532 * effective cpuset's. As this function is called with cpuset_mutex held,
2533 * cpuset membership stays stable.
2534 */
2535void cpuset_update_tasks_nodemask(struct cpuset *cs)
2536{
2537 static nodemask_t newmems; /* protected by cpuset_mutex */
2538 struct css_task_iter it;
2539 struct task_struct *task;
2540
2541 cpuset_being_rebound = cs; /* causes mpol_dup() rebind */
2542
2543 guarantee_online_mems(cs, &newmems);
2544
2545 /*
2546 * The mpol_rebind_mm() call takes mmap_lock, which we couldn't
2547 * take while holding tasklist_lock. Forks can happen - the
2548 * mpol_dup() cpuset_being_rebound check will catch such forks,
2549 * and rebind their vma mempolicies too. Because we still hold
2550 * the global cpuset_mutex, we know that no other rebind effort
2551 * will be contending for the global variable cpuset_being_rebound.
2552 * It's ok if we rebind the same mm twice; mpol_rebind_mm()
2553 * is idempotent. Also migrate pages in each mm to new nodes.
2554 */
2555 css_task_iter_start(&cs->css, 0, &it);
2556 while ((task = css_task_iter_next(&it))) {
2557 struct mm_struct *mm;
2558 bool migrate;
2559
2560 cpuset_change_task_nodemask(task, &newmems);
2561
2562 mm = get_task_mm(task);
2563 if (!mm)
2564 continue;
2565
2566 migrate = is_memory_migrate(cs);
2567
2568 mpol_rebind_mm(mm, &cs->mems_allowed);
2569 if (migrate)
2570 cpuset_migrate_mm(mm, &cs->old_mems_allowed, &newmems);
2571 else
2572 mmput(mm);
2573 }
2574 css_task_iter_end(&it);
2575
2576 /*
2577 * All the tasks' nodemasks have been updated, update
2578 * cs->old_mems_allowed.
2579 */
2580 cs->old_mems_allowed = newmems;
2581
2582 /* We're done rebinding vmas to this cpuset's new mems_allowed. */
2583 cpuset_being_rebound = NULL;
2584}
2585
2586/*
2587 * update_nodemasks_hier - Update effective nodemasks and tasks in the subtree
2588 * @cs: the cpuset to consider
2589 * @new_mems: a temp variable for calculating new effective_mems
2590 *
2591 * When configured nodemask is changed, the effective nodemasks of this cpuset
2592 * and all its descendants need to be updated.
2593 *
2594 * On legacy hierarchy, effective_mems will be the same with mems_allowed.
2595 *
2596 * Called with cpuset_mutex held
2597 */
2598static void update_nodemasks_hier(struct cpuset *cs, nodemask_t *new_mems)
2599{
2600 struct cpuset *cp;
2601 struct cgroup_subsys_state *pos_css;
2602
2603 rcu_read_lock();
2604 cpuset_for_each_descendant_pre(cp, pos_css, cs) {
2605 struct cpuset *parent = parent_cs(cp);
2606
2607 nodes_and(*new_mems, cp->mems_allowed, parent->effective_mems);
2608
2609 /*
2610 * If it becomes empty, inherit the effective mask of the
2611 * parent, which is guaranteed to have some MEMs.
2612 */
2613 if (is_in_v2_mode() && nodes_empty(*new_mems))
2614 *new_mems = parent->effective_mems;
2615
2616 /* Skip the whole subtree if the nodemask remains the same. */
2617 if (nodes_equal(*new_mems, cp->effective_mems)) {
2618 pos_css = css_rightmost_descendant(pos_css);
2619 continue;
2620 }
2621
2622 if (!css_tryget_online(&cp->css))
2623 continue;
2624 rcu_read_unlock();
2625
2626 spin_lock_irq(&callback_lock);
2627 cp->effective_mems = *new_mems;
2628 spin_unlock_irq(&callback_lock);
2629
2630 WARN_ON(!is_in_v2_mode() &&
2631 !nodes_equal(cp->mems_allowed, cp->effective_mems));
2632
2633 cpuset_update_tasks_nodemask(cp);
2634
2635 rcu_read_lock();
2636 css_put(&cp->css);
2637 }
2638 rcu_read_unlock();
2639}
2640
2641/*
2642 * Handle user request to change the 'mems' memory placement
2643 * of a cpuset. Needs to validate the request, update the
2644 * cpusets mems_allowed, and for each task in the cpuset,
2645 * update mems_allowed and rebind task's mempolicy and any vma
2646 * mempolicies and if the cpuset is marked 'memory_migrate',
2647 * migrate the tasks pages to the new memory.
2648 *
2649 * Call with cpuset_mutex held. May take callback_lock during call.
2650 * Will take tasklist_lock, scan tasklist for tasks in cpuset cs,
2651 * lock each such tasks mm->mmap_lock, scan its vma's and rebind
2652 * their mempolicies to the cpusets new mems_allowed.
2653 */
2654static int update_nodemask(struct cpuset *cs, struct cpuset *trialcs,
2655 const char *buf)
2656{
2657 int retval;
2658
2659 /*
2660 * top_cpuset.mems_allowed tracks node_stats[N_MEMORY];
2661 * it's read-only
2662 */
2663 if (cs == &top_cpuset) {
2664 retval = -EACCES;
2665 goto done;
2666 }
2667
2668 /*
2669 * An empty mems_allowed is ok iff there are no tasks in the cpuset.
2670 * Since nodelist_parse() fails on an empty mask, we special case
2671 * that parsing. The validate_change() call ensures that cpusets
2672 * with tasks have memory.
2673 */
2674 if (!*buf) {
2675 nodes_clear(trialcs->mems_allowed);
2676 } else {
2677 retval = nodelist_parse(buf, trialcs->mems_allowed);
2678 if (retval < 0)
2679 goto done;
2680
2681 if (!nodes_subset(trialcs->mems_allowed,
2682 top_cpuset.mems_allowed)) {
2683 retval = -EINVAL;
2684 goto done;
2685 }
2686 }
2687
2688 if (nodes_equal(cs->mems_allowed, trialcs->mems_allowed)) {
2689 retval = 0; /* Too easy - nothing to do */
2690 goto done;
2691 }
2692 retval = validate_change(cs, trialcs);
2693 if (retval < 0)
2694 goto done;
2695
2696 check_insane_mems_config(&trialcs->mems_allowed);
2697
2698 spin_lock_irq(&callback_lock);
2699 cs->mems_allowed = trialcs->mems_allowed;
2700 spin_unlock_irq(&callback_lock);
2701
2702 /* use trialcs->mems_allowed as a temp variable */
2703 update_nodemasks_hier(cs, &trialcs->mems_allowed);
2704done:
2705 return retval;
2706}
2707
2708bool current_cpuset_is_being_rebound(void)
2709{
2710 bool ret;
2711
2712 rcu_read_lock();
2713 ret = task_cs(current) == cpuset_being_rebound;
2714 rcu_read_unlock();
2715
2716 return ret;
2717}
2718
2719/*
2720 * cpuset_update_flag - read a 0 or a 1 in a file and update associated flag
2721 * bit: the bit to update (see cpuset_flagbits_t)
2722 * cs: the cpuset to update
2723 * turning_on: whether the flag is being set or cleared
2724 *
2725 * Call with cpuset_mutex held.
2726 */
2727
2728int cpuset_update_flag(cpuset_flagbits_t bit, struct cpuset *cs,
2729 int turning_on)
2730{
2731 struct cpuset *trialcs;
2732 int balance_flag_changed;
2733 int spread_flag_changed;
2734 int err;
2735
2736 trialcs = alloc_trial_cpuset(cs);
2737 if (!trialcs)
2738 return -ENOMEM;
2739
2740 if (turning_on)
2741 set_bit(bit, &trialcs->flags);
2742 else
2743 clear_bit(bit, &trialcs->flags);
2744
2745 err = validate_change(cs, trialcs);
2746 if (err < 0)
2747 goto out;
2748
2749 balance_flag_changed = (is_sched_load_balance(cs) !=
2750 is_sched_load_balance(trialcs));
2751
2752 spread_flag_changed = ((is_spread_slab(cs) != is_spread_slab(trialcs))
2753 || (is_spread_page(cs) != is_spread_page(trialcs)));
2754
2755 spin_lock_irq(&callback_lock);
2756 cs->flags = trialcs->flags;
2757 spin_unlock_irq(&callback_lock);
2758
2759 if (!cpumask_empty(trialcs->cpus_allowed) && balance_flag_changed) {
2760 if (cpuset_v2())
2761 cpuset_force_rebuild();
2762 else
2763 rebuild_sched_domains_locked();
2764 }
2765
2766 if (spread_flag_changed)
2767 cpuset1_update_tasks_flags(cs);
2768out:
2769 free_cpuset(trialcs);
2770 return err;
2771}
2772
2773/**
2774 * update_prstate - update partition_root_state
2775 * @cs: the cpuset to update
2776 * @new_prs: new partition root state
2777 * Return: 0 if successful, != 0 if error
2778 *
2779 * Call with cpuset_mutex held.
2780 */
2781static int update_prstate(struct cpuset *cs, int new_prs)
2782{
2783 int err = PERR_NONE, old_prs = cs->partition_root_state;
2784 struct cpuset *parent = parent_cs(cs);
2785 struct tmpmasks tmpmask;
2786 bool new_xcpus_state = false;
2787
2788 if (old_prs == new_prs)
2789 return 0;
2790
2791 /*
2792 * Treat a previously invalid partition root as if it is a "member".
2793 */
2794 if (new_prs && is_prs_invalid(old_prs))
2795 old_prs = PRS_MEMBER;
2796
2797 if (alloc_cpumasks(NULL, &tmpmask))
2798 return -ENOMEM;
2799
2800 /*
2801 * Setup effective_xcpus if not properly set yet, it will be cleared
2802 * later if partition becomes invalid.
2803 */
2804 if ((new_prs > 0) && cpumask_empty(cs->exclusive_cpus)) {
2805 spin_lock_irq(&callback_lock);
2806 cpumask_and(cs->effective_xcpus,
2807 cs->cpus_allowed, parent->effective_xcpus);
2808 spin_unlock_irq(&callback_lock);
2809 }
2810
2811 err = update_partition_exclusive(cs, new_prs);
2812 if (err)
2813 goto out;
2814
2815 if (!old_prs) {
2816 /*
2817 * cpus_allowed and exclusive_cpus cannot be both empty.
2818 */
2819 if (xcpus_empty(cs)) {
2820 err = PERR_CPUSEMPTY;
2821 goto out;
2822 }
2823
2824 /*
2825 * If parent is valid partition, enable local partiion.
2826 * Otherwise, enable a remote partition.
2827 */
2828 if (is_partition_valid(parent)) {
2829 enum partition_cmd cmd = (new_prs == PRS_ROOT)
2830 ? partcmd_enable : partcmd_enablei;
2831
2832 err = update_parent_effective_cpumask(cs, cmd, NULL, &tmpmask);
2833 } else {
2834 err = remote_partition_enable(cs, new_prs, &tmpmask);
2835 }
2836 } else if (old_prs && new_prs) {
2837 /*
2838 * A change in load balance state only, no change in cpumasks.
2839 */
2840 new_xcpus_state = true;
2841 } else {
2842 /*
2843 * Switching back to member is always allowed even if it
2844 * disables child partitions.
2845 */
2846 if (is_remote_partition(cs))
2847 remote_partition_disable(cs, &tmpmask);
2848 else
2849 update_parent_effective_cpumask(cs, partcmd_disable,
2850 NULL, &tmpmask);
2851
2852 /*
2853 * Invalidation of child partitions will be done in
2854 * update_cpumasks_hier().
2855 */
2856 }
2857out:
2858 /*
2859 * Make partition invalid & disable CS_CPU_EXCLUSIVE if an error
2860 * happens.
2861 */
2862 if (err) {
2863 new_prs = -new_prs;
2864 update_partition_exclusive(cs, new_prs);
2865 }
2866
2867 spin_lock_irq(&callback_lock);
2868 cs->partition_root_state = new_prs;
2869 WRITE_ONCE(cs->prs_err, err);
2870 if (!is_partition_valid(cs))
2871 reset_partition_data(cs);
2872 else if (new_xcpus_state)
2873 partition_xcpus_newstate(old_prs, new_prs, cs->effective_xcpus);
2874 spin_unlock_irq(&callback_lock);
2875 update_unbound_workqueue_cpumask(new_xcpus_state);
2876
2877 /* Force update if switching back to member */
2878 update_cpumasks_hier(cs, &tmpmask, !new_prs);
2879
2880 /* Update sched domains and load balance flag */
2881 update_partition_sd_lb(cs, old_prs);
2882
2883 notify_partition_change(cs, old_prs);
2884 if (force_sd_rebuild)
2885 rebuild_sched_domains_locked();
2886 free_cpumasks(NULL, &tmpmask);
2887 return 0;
2888}
2889
2890static struct cpuset *cpuset_attach_old_cs;
2891
2892/*
2893 * Check to see if a cpuset can accept a new task
2894 * For v1, cpus_allowed and mems_allowed can't be empty.
2895 * For v2, effective_cpus can't be empty.
2896 * Note that in v1, effective_cpus = cpus_allowed.
2897 */
2898static int cpuset_can_attach_check(struct cpuset *cs)
2899{
2900 if (cpumask_empty(cs->effective_cpus) ||
2901 (!is_in_v2_mode() && nodes_empty(cs->mems_allowed)))
2902 return -ENOSPC;
2903 return 0;
2904}
2905
2906static void reset_migrate_dl_data(struct cpuset *cs)
2907{
2908 cs->nr_migrate_dl_tasks = 0;
2909 cs->sum_migrate_dl_bw = 0;
2910}
2911
2912/* Called by cgroups to determine if a cpuset is usable; cpuset_mutex held */
2913static int cpuset_can_attach(struct cgroup_taskset *tset)
2914{
2915 struct cgroup_subsys_state *css;
2916 struct cpuset *cs, *oldcs;
2917 struct task_struct *task;
2918 bool cpus_updated, mems_updated;
2919 int ret;
2920
2921 /* used later by cpuset_attach() */
2922 cpuset_attach_old_cs = task_cs(cgroup_taskset_first(tset, &css));
2923 oldcs = cpuset_attach_old_cs;
2924 cs = css_cs(css);
2925
2926 mutex_lock(&cpuset_mutex);
2927
2928 /* Check to see if task is allowed in the cpuset */
2929 ret = cpuset_can_attach_check(cs);
2930 if (ret)
2931 goto out_unlock;
2932
2933 cpus_updated = !cpumask_equal(cs->effective_cpus, oldcs->effective_cpus);
2934 mems_updated = !nodes_equal(cs->effective_mems, oldcs->effective_mems);
2935
2936 cgroup_taskset_for_each(task, css, tset) {
2937 ret = task_can_attach(task);
2938 if (ret)
2939 goto out_unlock;
2940
2941 /*
2942 * Skip rights over task check in v2 when nothing changes,
2943 * migration permission derives from hierarchy ownership in
2944 * cgroup_procs_write_permission()).
2945 */
2946 if (!cpuset_v2() || (cpus_updated || mems_updated)) {
2947 ret = security_task_setscheduler(task);
2948 if (ret)
2949 goto out_unlock;
2950 }
2951
2952 if (dl_task(task)) {
2953 cs->nr_migrate_dl_tasks++;
2954 cs->sum_migrate_dl_bw += task->dl.dl_bw;
2955 }
2956 }
2957
2958 if (!cs->nr_migrate_dl_tasks)
2959 goto out_success;
2960
2961 if (!cpumask_intersects(oldcs->effective_cpus, cs->effective_cpus)) {
2962 int cpu = cpumask_any_and(cpu_active_mask, cs->effective_cpus);
2963
2964 if (unlikely(cpu >= nr_cpu_ids)) {
2965 reset_migrate_dl_data(cs);
2966 ret = -EINVAL;
2967 goto out_unlock;
2968 }
2969
2970 ret = dl_bw_alloc(cpu, cs->sum_migrate_dl_bw);
2971 if (ret) {
2972 reset_migrate_dl_data(cs);
2973 goto out_unlock;
2974 }
2975 }
2976
2977out_success:
2978 /*
2979 * Mark attach is in progress. This makes validate_change() fail
2980 * changes which zero cpus/mems_allowed.
2981 */
2982 cs->attach_in_progress++;
2983out_unlock:
2984 mutex_unlock(&cpuset_mutex);
2985 return ret;
2986}
2987
2988static void cpuset_cancel_attach(struct cgroup_taskset *tset)
2989{
2990 struct cgroup_subsys_state *css;
2991 struct cpuset *cs;
2992
2993 cgroup_taskset_first(tset, &css);
2994 cs = css_cs(css);
2995
2996 mutex_lock(&cpuset_mutex);
2997 dec_attach_in_progress_locked(cs);
2998
2999 if (cs->nr_migrate_dl_tasks) {
3000 int cpu = cpumask_any(cs->effective_cpus);
3001
3002 dl_bw_free(cpu, cs->sum_migrate_dl_bw);
3003 reset_migrate_dl_data(cs);
3004 }
3005
3006 mutex_unlock(&cpuset_mutex);
3007}
3008
3009/*
3010 * Protected by cpuset_mutex. cpus_attach is used only by cpuset_attach_task()
3011 * but we can't allocate it dynamically there. Define it global and
3012 * allocate from cpuset_init().
3013 */
3014static cpumask_var_t cpus_attach;
3015static nodemask_t cpuset_attach_nodemask_to;
3016
3017static void cpuset_attach_task(struct cpuset *cs, struct task_struct *task)
3018{
3019 lockdep_assert_held(&cpuset_mutex);
3020
3021 if (cs != &top_cpuset)
3022 guarantee_online_cpus(task, cpus_attach);
3023 else
3024 cpumask_andnot(cpus_attach, task_cpu_possible_mask(task),
3025 subpartitions_cpus);
3026 /*
3027 * can_attach beforehand should guarantee that this doesn't
3028 * fail. TODO: have a better way to handle failure here
3029 */
3030 WARN_ON_ONCE(set_cpus_allowed_ptr(task, cpus_attach));
3031
3032 cpuset_change_task_nodemask(task, &cpuset_attach_nodemask_to);
3033 cpuset1_update_task_spread_flags(cs, task);
3034}
3035
3036static void cpuset_attach(struct cgroup_taskset *tset)
3037{
3038 struct task_struct *task;
3039 struct task_struct *leader;
3040 struct cgroup_subsys_state *css;
3041 struct cpuset *cs;
3042 struct cpuset *oldcs = cpuset_attach_old_cs;
3043 bool cpus_updated, mems_updated;
3044
3045 cgroup_taskset_first(tset, &css);
3046 cs = css_cs(css);
3047
3048 lockdep_assert_cpus_held(); /* see cgroup_attach_lock() */
3049 mutex_lock(&cpuset_mutex);
3050 cpus_updated = !cpumask_equal(cs->effective_cpus,
3051 oldcs->effective_cpus);
3052 mems_updated = !nodes_equal(cs->effective_mems, oldcs->effective_mems);
3053
3054 /*
3055 * In the default hierarchy, enabling cpuset in the child cgroups
3056 * will trigger a number of cpuset_attach() calls with no change
3057 * in effective cpus and mems. In that case, we can optimize out
3058 * by skipping the task iteration and update.
3059 */
3060 if (cpuset_v2() && !cpus_updated && !mems_updated) {
3061 cpuset_attach_nodemask_to = cs->effective_mems;
3062 goto out;
3063 }
3064
3065 guarantee_online_mems(cs, &cpuset_attach_nodemask_to);
3066
3067 cgroup_taskset_for_each(task, css, tset)
3068 cpuset_attach_task(cs, task);
3069
3070 /*
3071 * Change mm for all threadgroup leaders. This is expensive and may
3072 * sleep and should be moved outside migration path proper. Skip it
3073 * if there is no change in effective_mems and CS_MEMORY_MIGRATE is
3074 * not set.
3075 */
3076 cpuset_attach_nodemask_to = cs->effective_mems;
3077 if (!is_memory_migrate(cs) && !mems_updated)
3078 goto out;
3079
3080 cgroup_taskset_for_each_leader(leader, css, tset) {
3081 struct mm_struct *mm = get_task_mm(leader);
3082
3083 if (mm) {
3084 mpol_rebind_mm(mm, &cpuset_attach_nodemask_to);
3085
3086 /*
3087 * old_mems_allowed is the same with mems_allowed
3088 * here, except if this task is being moved
3089 * automatically due to hotplug. In that case
3090 * @mems_allowed has been updated and is empty, so
3091 * @old_mems_allowed is the right nodesets that we
3092 * migrate mm from.
3093 */
3094 if (is_memory_migrate(cs))
3095 cpuset_migrate_mm(mm, &oldcs->old_mems_allowed,
3096 &cpuset_attach_nodemask_to);
3097 else
3098 mmput(mm);
3099 }
3100 }
3101
3102out:
3103 cs->old_mems_allowed = cpuset_attach_nodemask_to;
3104
3105 if (cs->nr_migrate_dl_tasks) {
3106 cs->nr_deadline_tasks += cs->nr_migrate_dl_tasks;
3107 oldcs->nr_deadline_tasks -= cs->nr_migrate_dl_tasks;
3108 reset_migrate_dl_data(cs);
3109 }
3110
3111 dec_attach_in_progress_locked(cs);
3112
3113 mutex_unlock(&cpuset_mutex);
3114}
3115
3116/*
3117 * Common handling for a write to a "cpus" or "mems" file.
3118 */
3119ssize_t cpuset_write_resmask(struct kernfs_open_file *of,
3120 char *buf, size_t nbytes, loff_t off)
3121{
3122 struct cpuset *cs = css_cs(of_css(of));
3123 struct cpuset *trialcs;
3124 int retval = -ENODEV;
3125
3126 buf = strstrip(buf);
3127 cpus_read_lock();
3128 mutex_lock(&cpuset_mutex);
3129 if (!is_cpuset_online(cs))
3130 goto out_unlock;
3131
3132 trialcs = alloc_trial_cpuset(cs);
3133 if (!trialcs) {
3134 retval = -ENOMEM;
3135 goto out_unlock;
3136 }
3137
3138 switch (of_cft(of)->private) {
3139 case FILE_CPULIST:
3140 retval = update_cpumask(cs, trialcs, buf);
3141 break;
3142 case FILE_EXCLUSIVE_CPULIST:
3143 retval = update_exclusive_cpumask(cs, trialcs, buf);
3144 break;
3145 case FILE_MEMLIST:
3146 retval = update_nodemask(cs, trialcs, buf);
3147 break;
3148 default:
3149 retval = -EINVAL;
3150 break;
3151 }
3152
3153 free_cpuset(trialcs);
3154 if (force_sd_rebuild)
3155 rebuild_sched_domains_locked();
3156out_unlock:
3157 mutex_unlock(&cpuset_mutex);
3158 cpus_read_unlock();
3159 flush_workqueue(cpuset_migrate_mm_wq);
3160 return retval ?: nbytes;
3161}
3162
3163/*
3164 * These ascii lists should be read in a single call, by using a user
3165 * buffer large enough to hold the entire map. If read in smaller
3166 * chunks, there is no guarantee of atomicity. Since the display format
3167 * used, list of ranges of sequential numbers, is variable length,
3168 * and since these maps can change value dynamically, one could read
3169 * gibberish by doing partial reads while a list was changing.
3170 */
3171int cpuset_common_seq_show(struct seq_file *sf, void *v)
3172{
3173 struct cpuset *cs = css_cs(seq_css(sf));
3174 cpuset_filetype_t type = seq_cft(sf)->private;
3175 int ret = 0;
3176
3177 spin_lock_irq(&callback_lock);
3178
3179 switch (type) {
3180 case FILE_CPULIST:
3181 seq_printf(sf, "%*pbl\n", cpumask_pr_args(cs->cpus_allowed));
3182 break;
3183 case FILE_MEMLIST:
3184 seq_printf(sf, "%*pbl\n", nodemask_pr_args(&cs->mems_allowed));
3185 break;
3186 case FILE_EFFECTIVE_CPULIST:
3187 seq_printf(sf, "%*pbl\n", cpumask_pr_args(cs->effective_cpus));
3188 break;
3189 case FILE_EFFECTIVE_MEMLIST:
3190 seq_printf(sf, "%*pbl\n", nodemask_pr_args(&cs->effective_mems));
3191 break;
3192 case FILE_EXCLUSIVE_CPULIST:
3193 seq_printf(sf, "%*pbl\n", cpumask_pr_args(cs->exclusive_cpus));
3194 break;
3195 case FILE_EFFECTIVE_XCPULIST:
3196 seq_printf(sf, "%*pbl\n", cpumask_pr_args(cs->effective_xcpus));
3197 break;
3198 case FILE_SUBPARTS_CPULIST:
3199 seq_printf(sf, "%*pbl\n", cpumask_pr_args(subpartitions_cpus));
3200 break;
3201 case FILE_ISOLATED_CPULIST:
3202 seq_printf(sf, "%*pbl\n", cpumask_pr_args(isolated_cpus));
3203 break;
3204 default:
3205 ret = -EINVAL;
3206 }
3207
3208 spin_unlock_irq(&callback_lock);
3209 return ret;
3210}
3211
3212static int sched_partition_show(struct seq_file *seq, void *v)
3213{
3214 struct cpuset *cs = css_cs(seq_css(seq));
3215 const char *err, *type = NULL;
3216
3217 switch (cs->partition_root_state) {
3218 case PRS_ROOT:
3219 seq_puts(seq, "root\n");
3220 break;
3221 case PRS_ISOLATED:
3222 seq_puts(seq, "isolated\n");
3223 break;
3224 case PRS_MEMBER:
3225 seq_puts(seq, "member\n");
3226 break;
3227 case PRS_INVALID_ROOT:
3228 type = "root";
3229 fallthrough;
3230 case PRS_INVALID_ISOLATED:
3231 if (!type)
3232 type = "isolated";
3233 err = perr_strings[READ_ONCE(cs->prs_err)];
3234 if (err)
3235 seq_printf(seq, "%s invalid (%s)\n", type, err);
3236 else
3237 seq_printf(seq, "%s invalid\n", type);
3238 break;
3239 }
3240 return 0;
3241}
3242
3243static ssize_t sched_partition_write(struct kernfs_open_file *of, char *buf,
3244 size_t nbytes, loff_t off)
3245{
3246 struct cpuset *cs = css_cs(of_css(of));
3247 int val;
3248 int retval = -ENODEV;
3249
3250 buf = strstrip(buf);
3251
3252 if (!strcmp(buf, "root"))
3253 val = PRS_ROOT;
3254 else if (!strcmp(buf, "member"))
3255 val = PRS_MEMBER;
3256 else if (!strcmp(buf, "isolated"))
3257 val = PRS_ISOLATED;
3258 else
3259 return -EINVAL;
3260
3261 css_get(&cs->css);
3262 cpus_read_lock();
3263 mutex_lock(&cpuset_mutex);
3264 if (!is_cpuset_online(cs))
3265 goto out_unlock;
3266
3267 retval = update_prstate(cs, val);
3268out_unlock:
3269 mutex_unlock(&cpuset_mutex);
3270 cpus_read_unlock();
3271 css_put(&cs->css);
3272 return retval ?: nbytes;
3273}
3274
3275/*
3276 * This is currently a minimal set for the default hierarchy. It can be
3277 * expanded later on by migrating more features and control files from v1.
3278 */
3279static struct cftype dfl_files[] = {
3280 {
3281 .name = "cpus",
3282 .seq_show = cpuset_common_seq_show,
3283 .write = cpuset_write_resmask,
3284 .max_write_len = (100U + 6 * NR_CPUS),
3285 .private = FILE_CPULIST,
3286 .flags = CFTYPE_NOT_ON_ROOT,
3287 },
3288
3289 {
3290 .name = "mems",
3291 .seq_show = cpuset_common_seq_show,
3292 .write = cpuset_write_resmask,
3293 .max_write_len = (100U + 6 * MAX_NUMNODES),
3294 .private = FILE_MEMLIST,
3295 .flags = CFTYPE_NOT_ON_ROOT,
3296 },
3297
3298 {
3299 .name = "cpus.effective",
3300 .seq_show = cpuset_common_seq_show,
3301 .private = FILE_EFFECTIVE_CPULIST,
3302 },
3303
3304 {
3305 .name = "mems.effective",
3306 .seq_show = cpuset_common_seq_show,
3307 .private = FILE_EFFECTIVE_MEMLIST,
3308 },
3309
3310 {
3311 .name = "cpus.partition",
3312 .seq_show = sched_partition_show,
3313 .write = sched_partition_write,
3314 .private = FILE_PARTITION_ROOT,
3315 .flags = CFTYPE_NOT_ON_ROOT,
3316 .file_offset = offsetof(struct cpuset, partition_file),
3317 },
3318
3319 {
3320 .name = "cpus.exclusive",
3321 .seq_show = cpuset_common_seq_show,
3322 .write = cpuset_write_resmask,
3323 .max_write_len = (100U + 6 * NR_CPUS),
3324 .private = FILE_EXCLUSIVE_CPULIST,
3325 .flags = CFTYPE_NOT_ON_ROOT,
3326 },
3327
3328 {
3329 .name = "cpus.exclusive.effective",
3330 .seq_show = cpuset_common_seq_show,
3331 .private = FILE_EFFECTIVE_XCPULIST,
3332 .flags = CFTYPE_NOT_ON_ROOT,
3333 },
3334
3335 {
3336 .name = "cpus.subpartitions",
3337 .seq_show = cpuset_common_seq_show,
3338 .private = FILE_SUBPARTS_CPULIST,
3339 .flags = CFTYPE_ONLY_ON_ROOT | CFTYPE_DEBUG,
3340 },
3341
3342 {
3343 .name = "cpus.isolated",
3344 .seq_show = cpuset_common_seq_show,
3345 .private = FILE_ISOLATED_CPULIST,
3346 .flags = CFTYPE_ONLY_ON_ROOT,
3347 },
3348
3349 { } /* terminate */
3350};
3351
3352
3353/**
3354 * cpuset_css_alloc - Allocate a cpuset css
3355 * @parent_css: Parent css of the control group that the new cpuset will be
3356 * part of
3357 * Return: cpuset css on success, -ENOMEM on failure.
3358 *
3359 * Allocate and initialize a new cpuset css, for non-NULL @parent_css, return
3360 * top cpuset css otherwise.
3361 */
3362static struct cgroup_subsys_state *
3363cpuset_css_alloc(struct cgroup_subsys_state *parent_css)
3364{
3365 struct cpuset *cs;
3366
3367 if (!parent_css)
3368 return &top_cpuset.css;
3369
3370 cs = kzalloc(sizeof(*cs), GFP_KERNEL);
3371 if (!cs)
3372 return ERR_PTR(-ENOMEM);
3373
3374 if (alloc_cpumasks(cs, NULL)) {
3375 kfree(cs);
3376 return ERR_PTR(-ENOMEM);
3377 }
3378
3379 __set_bit(CS_SCHED_LOAD_BALANCE, &cs->flags);
3380 fmeter_init(&cs->fmeter);
3381 cs->relax_domain_level = -1;
3382 INIT_LIST_HEAD(&cs->remote_sibling);
3383
3384 /* Set CS_MEMORY_MIGRATE for default hierarchy */
3385 if (cpuset_v2())
3386 __set_bit(CS_MEMORY_MIGRATE, &cs->flags);
3387
3388 return &cs->css;
3389}
3390
3391static int cpuset_css_online(struct cgroup_subsys_state *css)
3392{
3393 struct cpuset *cs = css_cs(css);
3394 struct cpuset *parent = parent_cs(cs);
3395 struct cpuset *tmp_cs;
3396 struct cgroup_subsys_state *pos_css;
3397
3398 if (!parent)
3399 return 0;
3400
3401 cpus_read_lock();
3402 mutex_lock(&cpuset_mutex);
3403
3404 set_bit(CS_ONLINE, &cs->flags);
3405 if (is_spread_page(parent))
3406 set_bit(CS_SPREAD_PAGE, &cs->flags);
3407 if (is_spread_slab(parent))
3408 set_bit(CS_SPREAD_SLAB, &cs->flags);
3409 /*
3410 * For v2, clear CS_SCHED_LOAD_BALANCE if parent is isolated
3411 */
3412 if (cpuset_v2() && !is_sched_load_balance(parent))
3413 clear_bit(CS_SCHED_LOAD_BALANCE, &cs->flags);
3414
3415 cpuset_inc();
3416
3417 spin_lock_irq(&callback_lock);
3418 if (is_in_v2_mode()) {
3419 cpumask_copy(cs->effective_cpus, parent->effective_cpus);
3420 cs->effective_mems = parent->effective_mems;
3421 }
3422 spin_unlock_irq(&callback_lock);
3423
3424 if (!test_bit(CGRP_CPUSET_CLONE_CHILDREN, &css->cgroup->flags))
3425 goto out_unlock;
3426
3427 /*
3428 * Clone @parent's configuration if CGRP_CPUSET_CLONE_CHILDREN is
3429 * set. This flag handling is implemented in cgroup core for
3430 * historical reasons - the flag may be specified during mount.
3431 *
3432 * Currently, if any sibling cpusets have exclusive cpus or mem, we
3433 * refuse to clone the configuration - thereby refusing the task to
3434 * be entered, and as a result refusing the sys_unshare() or
3435 * clone() which initiated it. If this becomes a problem for some
3436 * users who wish to allow that scenario, then this could be
3437 * changed to grant parent->cpus_allowed-sibling_cpus_exclusive
3438 * (and likewise for mems) to the new cgroup.
3439 */
3440 rcu_read_lock();
3441 cpuset_for_each_child(tmp_cs, pos_css, parent) {
3442 if (is_mem_exclusive(tmp_cs) || is_cpu_exclusive(tmp_cs)) {
3443 rcu_read_unlock();
3444 goto out_unlock;
3445 }
3446 }
3447 rcu_read_unlock();
3448
3449 spin_lock_irq(&callback_lock);
3450 cs->mems_allowed = parent->mems_allowed;
3451 cs->effective_mems = parent->mems_allowed;
3452 cpumask_copy(cs->cpus_allowed, parent->cpus_allowed);
3453 cpumask_copy(cs->effective_cpus, parent->cpus_allowed);
3454 spin_unlock_irq(&callback_lock);
3455out_unlock:
3456 mutex_unlock(&cpuset_mutex);
3457 cpus_read_unlock();
3458 return 0;
3459}
3460
3461/*
3462 * If the cpuset being removed has its flag 'sched_load_balance'
3463 * enabled, then simulate turning sched_load_balance off, which
3464 * will call rebuild_sched_domains_locked(). That is not needed
3465 * in the default hierarchy where only changes in partition
3466 * will cause repartitioning.
3467 *
3468 * If the cpuset has the 'sched.partition' flag enabled, simulate
3469 * turning 'sched.partition" off.
3470 */
3471
3472static void cpuset_css_offline(struct cgroup_subsys_state *css)
3473{
3474 struct cpuset *cs = css_cs(css);
3475
3476 cpus_read_lock();
3477 mutex_lock(&cpuset_mutex);
3478
3479 if (is_partition_valid(cs))
3480 update_prstate(cs, 0);
3481
3482 if (!cpuset_v2() && is_sched_load_balance(cs))
3483 cpuset_update_flag(CS_SCHED_LOAD_BALANCE, cs, 0);
3484
3485 cpuset_dec();
3486 clear_bit(CS_ONLINE, &cs->flags);
3487
3488 mutex_unlock(&cpuset_mutex);
3489 cpus_read_unlock();
3490}
3491
3492static void cpuset_css_free(struct cgroup_subsys_state *css)
3493{
3494 struct cpuset *cs = css_cs(css);
3495
3496 free_cpuset(cs);
3497}
3498
3499static void cpuset_bind(struct cgroup_subsys_state *root_css)
3500{
3501 mutex_lock(&cpuset_mutex);
3502 spin_lock_irq(&callback_lock);
3503
3504 if (is_in_v2_mode()) {
3505 cpumask_copy(top_cpuset.cpus_allowed, cpu_possible_mask);
3506 cpumask_copy(top_cpuset.effective_xcpus, cpu_possible_mask);
3507 top_cpuset.mems_allowed = node_possible_map;
3508 } else {
3509 cpumask_copy(top_cpuset.cpus_allowed,
3510 top_cpuset.effective_cpus);
3511 top_cpuset.mems_allowed = top_cpuset.effective_mems;
3512 }
3513
3514 spin_unlock_irq(&callback_lock);
3515 mutex_unlock(&cpuset_mutex);
3516}
3517
3518/*
3519 * In case the child is cloned into a cpuset different from its parent,
3520 * additional checks are done to see if the move is allowed.
3521 */
3522static int cpuset_can_fork(struct task_struct *task, struct css_set *cset)
3523{
3524 struct cpuset *cs = css_cs(cset->subsys[cpuset_cgrp_id]);
3525 bool same_cs;
3526 int ret;
3527
3528 rcu_read_lock();
3529 same_cs = (cs == task_cs(current));
3530 rcu_read_unlock();
3531
3532 if (same_cs)
3533 return 0;
3534
3535 lockdep_assert_held(&cgroup_mutex);
3536 mutex_lock(&cpuset_mutex);
3537
3538 /* Check to see if task is allowed in the cpuset */
3539 ret = cpuset_can_attach_check(cs);
3540 if (ret)
3541 goto out_unlock;
3542
3543 ret = task_can_attach(task);
3544 if (ret)
3545 goto out_unlock;
3546
3547 ret = security_task_setscheduler(task);
3548 if (ret)
3549 goto out_unlock;
3550
3551 /*
3552 * Mark attach is in progress. This makes validate_change() fail
3553 * changes which zero cpus/mems_allowed.
3554 */
3555 cs->attach_in_progress++;
3556out_unlock:
3557 mutex_unlock(&cpuset_mutex);
3558 return ret;
3559}
3560
3561static void cpuset_cancel_fork(struct task_struct *task, struct css_set *cset)
3562{
3563 struct cpuset *cs = css_cs(cset->subsys[cpuset_cgrp_id]);
3564 bool same_cs;
3565
3566 rcu_read_lock();
3567 same_cs = (cs == task_cs(current));
3568 rcu_read_unlock();
3569
3570 if (same_cs)
3571 return;
3572
3573 dec_attach_in_progress(cs);
3574}
3575
3576/*
3577 * Make sure the new task conform to the current state of its parent,
3578 * which could have been changed by cpuset just after it inherits the
3579 * state from the parent and before it sits on the cgroup's task list.
3580 */
3581static void cpuset_fork(struct task_struct *task)
3582{
3583 struct cpuset *cs;
3584 bool same_cs;
3585
3586 rcu_read_lock();
3587 cs = task_cs(task);
3588 same_cs = (cs == task_cs(current));
3589 rcu_read_unlock();
3590
3591 if (same_cs) {
3592 if (cs == &top_cpuset)
3593 return;
3594
3595 set_cpus_allowed_ptr(task, current->cpus_ptr);
3596 task->mems_allowed = current->mems_allowed;
3597 return;
3598 }
3599
3600 /* CLONE_INTO_CGROUP */
3601 mutex_lock(&cpuset_mutex);
3602 guarantee_online_mems(cs, &cpuset_attach_nodemask_to);
3603 cpuset_attach_task(cs, task);
3604
3605 dec_attach_in_progress_locked(cs);
3606 mutex_unlock(&cpuset_mutex);
3607}
3608
3609struct cgroup_subsys cpuset_cgrp_subsys = {
3610 .css_alloc = cpuset_css_alloc,
3611 .css_online = cpuset_css_online,
3612 .css_offline = cpuset_css_offline,
3613 .css_free = cpuset_css_free,
3614 .can_attach = cpuset_can_attach,
3615 .cancel_attach = cpuset_cancel_attach,
3616 .attach = cpuset_attach,
3617 .post_attach = cpuset_post_attach,
3618 .bind = cpuset_bind,
3619 .can_fork = cpuset_can_fork,
3620 .cancel_fork = cpuset_cancel_fork,
3621 .fork = cpuset_fork,
3622#ifdef CONFIG_CPUSETS_V1
3623 .legacy_cftypes = cpuset1_files,
3624#endif
3625 .dfl_cftypes = dfl_files,
3626 .early_init = true,
3627 .threaded = true,
3628};
3629
3630/**
3631 * cpuset_init - initialize cpusets at system boot
3632 *
3633 * Description: Initialize top_cpuset
3634 **/
3635
3636int __init cpuset_init(void)
3637{
3638 BUG_ON(!alloc_cpumask_var(&top_cpuset.cpus_allowed, GFP_KERNEL));
3639 BUG_ON(!alloc_cpumask_var(&top_cpuset.effective_cpus, GFP_KERNEL));
3640 BUG_ON(!alloc_cpumask_var(&top_cpuset.effective_xcpus, GFP_KERNEL));
3641 BUG_ON(!alloc_cpumask_var(&top_cpuset.exclusive_cpus, GFP_KERNEL));
3642 BUG_ON(!zalloc_cpumask_var(&subpartitions_cpus, GFP_KERNEL));
3643 BUG_ON(!zalloc_cpumask_var(&isolated_cpus, GFP_KERNEL));
3644
3645 cpumask_setall(top_cpuset.cpus_allowed);
3646 nodes_setall(top_cpuset.mems_allowed);
3647 cpumask_setall(top_cpuset.effective_cpus);
3648 cpumask_setall(top_cpuset.effective_xcpus);
3649 cpumask_setall(top_cpuset.exclusive_cpus);
3650 nodes_setall(top_cpuset.effective_mems);
3651
3652 fmeter_init(&top_cpuset.fmeter);
3653 INIT_LIST_HEAD(&remote_children);
3654
3655 BUG_ON(!alloc_cpumask_var(&cpus_attach, GFP_KERNEL));
3656
3657 have_boot_isolcpus = housekeeping_enabled(HK_TYPE_DOMAIN);
3658 if (have_boot_isolcpus) {
3659 BUG_ON(!alloc_cpumask_var(&boot_hk_cpus, GFP_KERNEL));
3660 cpumask_copy(boot_hk_cpus, housekeeping_cpumask(HK_TYPE_DOMAIN));
3661 cpumask_andnot(isolated_cpus, cpu_possible_mask, boot_hk_cpus);
3662 }
3663
3664 return 0;
3665}
3666
3667static void
3668hotplug_update_tasks(struct cpuset *cs,
3669 struct cpumask *new_cpus, nodemask_t *new_mems,
3670 bool cpus_updated, bool mems_updated)
3671{
3672 /* A partition root is allowed to have empty effective cpus */
3673 if (cpumask_empty(new_cpus) && !is_partition_valid(cs))
3674 cpumask_copy(new_cpus, parent_cs(cs)->effective_cpus);
3675 if (nodes_empty(*new_mems))
3676 *new_mems = parent_cs(cs)->effective_mems;
3677
3678 spin_lock_irq(&callback_lock);
3679 cpumask_copy(cs->effective_cpus, new_cpus);
3680 cs->effective_mems = *new_mems;
3681 spin_unlock_irq(&callback_lock);
3682
3683 if (cpus_updated)
3684 cpuset_update_tasks_cpumask(cs, new_cpus);
3685 if (mems_updated)
3686 cpuset_update_tasks_nodemask(cs);
3687}
3688
3689void cpuset_force_rebuild(void)
3690{
3691 force_sd_rebuild = true;
3692}
3693
3694/**
3695 * cpuset_hotplug_update_tasks - update tasks in a cpuset for hotunplug
3696 * @cs: cpuset in interest
3697 * @tmp: the tmpmasks structure pointer
3698 *
3699 * Compare @cs's cpu and mem masks against top_cpuset and if some have gone
3700 * offline, update @cs accordingly. If @cs ends up with no CPU or memory,
3701 * all its tasks are moved to the nearest ancestor with both resources.
3702 */
3703static void cpuset_hotplug_update_tasks(struct cpuset *cs, struct tmpmasks *tmp)
3704{
3705 static cpumask_t new_cpus;
3706 static nodemask_t new_mems;
3707 bool cpus_updated;
3708 bool mems_updated;
3709 bool remote;
3710 int partcmd = -1;
3711 struct cpuset *parent;
3712retry:
3713 wait_event(cpuset_attach_wq, cs->attach_in_progress == 0);
3714
3715 mutex_lock(&cpuset_mutex);
3716
3717 /*
3718 * We have raced with task attaching. We wait until attaching
3719 * is finished, so we won't attach a task to an empty cpuset.
3720 */
3721 if (cs->attach_in_progress) {
3722 mutex_unlock(&cpuset_mutex);
3723 goto retry;
3724 }
3725
3726 parent = parent_cs(cs);
3727 compute_effective_cpumask(&new_cpus, cs, parent);
3728 nodes_and(new_mems, cs->mems_allowed, parent->effective_mems);
3729
3730 if (!tmp || !cs->partition_root_state)
3731 goto update_tasks;
3732
3733 /*
3734 * Compute effective_cpus for valid partition root, may invalidate
3735 * child partition roots if necessary.
3736 */
3737 remote = is_remote_partition(cs);
3738 if (remote || (is_partition_valid(cs) && is_partition_valid(parent)))
3739 compute_partition_effective_cpumask(cs, &new_cpus);
3740
3741 if (remote && cpumask_empty(&new_cpus) &&
3742 partition_is_populated(cs, NULL)) {
3743 remote_partition_disable(cs, tmp);
3744 compute_effective_cpumask(&new_cpus, cs, parent);
3745 remote = false;
3746 cpuset_force_rebuild();
3747 }
3748
3749 /*
3750 * Force the partition to become invalid if either one of
3751 * the following conditions hold:
3752 * 1) empty effective cpus but not valid empty partition.
3753 * 2) parent is invalid or doesn't grant any cpus to child
3754 * partitions.
3755 */
3756 if (is_local_partition(cs) && (!is_partition_valid(parent) ||
3757 tasks_nocpu_error(parent, cs, &new_cpus)))
3758 partcmd = partcmd_invalidate;
3759 /*
3760 * On the other hand, an invalid partition root may be transitioned
3761 * back to a regular one.
3762 */
3763 else if (is_partition_valid(parent) && is_partition_invalid(cs))
3764 partcmd = partcmd_update;
3765
3766 if (partcmd >= 0) {
3767 update_parent_effective_cpumask(cs, partcmd, NULL, tmp);
3768 if ((partcmd == partcmd_invalidate) || is_partition_valid(cs)) {
3769 compute_partition_effective_cpumask(cs, &new_cpus);
3770 cpuset_force_rebuild();
3771 }
3772 }
3773
3774update_tasks:
3775 cpus_updated = !cpumask_equal(&new_cpus, cs->effective_cpus);
3776 mems_updated = !nodes_equal(new_mems, cs->effective_mems);
3777 if (!cpus_updated && !mems_updated)
3778 goto unlock; /* Hotplug doesn't affect this cpuset */
3779
3780 if (mems_updated)
3781 check_insane_mems_config(&new_mems);
3782
3783 if (is_in_v2_mode())
3784 hotplug_update_tasks(cs, &new_cpus, &new_mems,
3785 cpus_updated, mems_updated);
3786 else
3787 cpuset1_hotplug_update_tasks(cs, &new_cpus, &new_mems,
3788 cpus_updated, mems_updated);
3789
3790unlock:
3791 mutex_unlock(&cpuset_mutex);
3792}
3793
3794/**
3795 * cpuset_handle_hotplug - handle CPU/memory hot{,un}plug for a cpuset
3796 *
3797 * This function is called after either CPU or memory configuration has
3798 * changed and updates cpuset accordingly. The top_cpuset is always
3799 * synchronized to cpu_active_mask and N_MEMORY, which is necessary in
3800 * order to make cpusets transparent (of no affect) on systems that are
3801 * actively using CPU hotplug but making no active use of cpusets.
3802 *
3803 * Non-root cpusets are only affected by offlining. If any CPUs or memory
3804 * nodes have been taken down, cpuset_hotplug_update_tasks() is invoked on
3805 * all descendants.
3806 *
3807 * Note that CPU offlining during suspend is ignored. We don't modify
3808 * cpusets across suspend/resume cycles at all.
3809 *
3810 * CPU / memory hotplug is handled synchronously.
3811 */
3812static void cpuset_handle_hotplug(void)
3813{
3814 static cpumask_t new_cpus;
3815 static nodemask_t new_mems;
3816 bool cpus_updated, mems_updated;
3817 bool on_dfl = is_in_v2_mode();
3818 struct tmpmasks tmp, *ptmp = NULL;
3819
3820 if (on_dfl && !alloc_cpumasks(NULL, &tmp))
3821 ptmp = &tmp;
3822
3823 lockdep_assert_cpus_held();
3824 mutex_lock(&cpuset_mutex);
3825
3826 /* fetch the available cpus/mems and find out which changed how */
3827 cpumask_copy(&new_cpus, cpu_active_mask);
3828 new_mems = node_states[N_MEMORY];
3829
3830 /*
3831 * If subpartitions_cpus is populated, it is likely that the check
3832 * below will produce a false positive on cpus_updated when the cpu
3833 * list isn't changed. It is extra work, but it is better to be safe.
3834 */
3835 cpus_updated = !cpumask_equal(top_cpuset.effective_cpus, &new_cpus) ||
3836 !cpumask_empty(subpartitions_cpus);
3837 mems_updated = !nodes_equal(top_cpuset.effective_mems, new_mems);
3838
3839 /* For v1, synchronize cpus_allowed to cpu_active_mask */
3840 if (cpus_updated) {
3841 cpuset_force_rebuild();
3842 spin_lock_irq(&callback_lock);
3843 if (!on_dfl)
3844 cpumask_copy(top_cpuset.cpus_allowed, &new_cpus);
3845 /*
3846 * Make sure that CPUs allocated to child partitions
3847 * do not show up in effective_cpus. If no CPU is left,
3848 * we clear the subpartitions_cpus & let the child partitions
3849 * fight for the CPUs again.
3850 */
3851 if (!cpumask_empty(subpartitions_cpus)) {
3852 if (cpumask_subset(&new_cpus, subpartitions_cpus)) {
3853 top_cpuset.nr_subparts = 0;
3854 cpumask_clear(subpartitions_cpus);
3855 } else {
3856 cpumask_andnot(&new_cpus, &new_cpus,
3857 subpartitions_cpus);
3858 }
3859 }
3860 cpumask_copy(top_cpuset.effective_cpus, &new_cpus);
3861 spin_unlock_irq(&callback_lock);
3862 /* we don't mess with cpumasks of tasks in top_cpuset */
3863 }
3864
3865 /* synchronize mems_allowed to N_MEMORY */
3866 if (mems_updated) {
3867 spin_lock_irq(&callback_lock);
3868 if (!on_dfl)
3869 top_cpuset.mems_allowed = new_mems;
3870 top_cpuset.effective_mems = new_mems;
3871 spin_unlock_irq(&callback_lock);
3872 cpuset_update_tasks_nodemask(&top_cpuset);
3873 }
3874
3875 mutex_unlock(&cpuset_mutex);
3876
3877 /* if cpus or mems changed, we need to propagate to descendants */
3878 if (cpus_updated || mems_updated) {
3879 struct cpuset *cs;
3880 struct cgroup_subsys_state *pos_css;
3881
3882 rcu_read_lock();
3883 cpuset_for_each_descendant_pre(cs, pos_css, &top_cpuset) {
3884 if (cs == &top_cpuset || !css_tryget_online(&cs->css))
3885 continue;
3886 rcu_read_unlock();
3887
3888 cpuset_hotplug_update_tasks(cs, ptmp);
3889
3890 rcu_read_lock();
3891 css_put(&cs->css);
3892 }
3893 rcu_read_unlock();
3894 }
3895
3896 /* rebuild sched domains if necessary */
3897 if (force_sd_rebuild)
3898 rebuild_sched_domains_cpuslocked();
3899
3900 free_cpumasks(NULL, ptmp);
3901}
3902
3903void cpuset_update_active_cpus(void)
3904{
3905 /*
3906 * We're inside cpu hotplug critical region which usually nests
3907 * inside cgroup synchronization. Bounce actual hotplug processing
3908 * to a work item to avoid reverse locking order.
3909 */
3910 cpuset_handle_hotplug();
3911}
3912
3913/*
3914 * Keep top_cpuset.mems_allowed tracking node_states[N_MEMORY].
3915 * Call this routine anytime after node_states[N_MEMORY] changes.
3916 * See cpuset_update_active_cpus() for CPU hotplug handling.
3917 */
3918static int cpuset_track_online_nodes(struct notifier_block *self,
3919 unsigned long action, void *arg)
3920{
3921 cpuset_handle_hotplug();
3922 return NOTIFY_OK;
3923}
3924
3925/**
3926 * cpuset_init_smp - initialize cpus_allowed
3927 *
3928 * Description: Finish top cpuset after cpu, node maps are initialized
3929 */
3930void __init cpuset_init_smp(void)
3931{
3932 /*
3933 * cpus_allowd/mems_allowed set to v2 values in the initial
3934 * cpuset_bind() call will be reset to v1 values in another
3935 * cpuset_bind() call when v1 cpuset is mounted.
3936 */
3937 top_cpuset.old_mems_allowed = top_cpuset.mems_allowed;
3938
3939 cpumask_copy(top_cpuset.effective_cpus, cpu_active_mask);
3940 top_cpuset.effective_mems = node_states[N_MEMORY];
3941
3942 hotplug_memory_notifier(cpuset_track_online_nodes, CPUSET_CALLBACK_PRI);
3943
3944 cpuset_migrate_mm_wq = alloc_ordered_workqueue("cpuset_migrate_mm", 0);
3945 BUG_ON(!cpuset_migrate_mm_wq);
3946}
3947
3948/**
3949 * cpuset_cpus_allowed - return cpus_allowed mask from a tasks cpuset.
3950 * @tsk: pointer to task_struct from which to obtain cpuset->cpus_allowed.
3951 * @pmask: pointer to struct cpumask variable to receive cpus_allowed set.
3952 *
3953 * Description: Returns the cpumask_var_t cpus_allowed of the cpuset
3954 * attached to the specified @tsk. Guaranteed to return some non-empty
3955 * subset of cpu_online_mask, even if this means going outside the
3956 * tasks cpuset, except when the task is in the top cpuset.
3957 **/
3958
3959void cpuset_cpus_allowed(struct task_struct *tsk, struct cpumask *pmask)
3960{
3961 unsigned long flags;
3962 struct cpuset *cs;
3963
3964 spin_lock_irqsave(&callback_lock, flags);
3965 rcu_read_lock();
3966
3967 cs = task_cs(tsk);
3968 if (cs != &top_cpuset)
3969 guarantee_online_cpus(tsk, pmask);
3970 /*
3971 * Tasks in the top cpuset won't get update to their cpumasks
3972 * when a hotplug online/offline event happens. So we include all
3973 * offline cpus in the allowed cpu list.
3974 */
3975 if ((cs == &top_cpuset) || cpumask_empty(pmask)) {
3976 const struct cpumask *possible_mask = task_cpu_possible_mask(tsk);
3977
3978 /*
3979 * We first exclude cpus allocated to partitions. If there is no
3980 * allowable online cpu left, we fall back to all possible cpus.
3981 */
3982 cpumask_andnot(pmask, possible_mask, subpartitions_cpus);
3983 if (!cpumask_intersects(pmask, cpu_online_mask))
3984 cpumask_copy(pmask, possible_mask);
3985 }
3986
3987 rcu_read_unlock();
3988 spin_unlock_irqrestore(&callback_lock, flags);
3989}
3990
3991/**
3992 * cpuset_cpus_allowed_fallback - final fallback before complete catastrophe.
3993 * @tsk: pointer to task_struct with which the scheduler is struggling
3994 *
3995 * Description: In the case that the scheduler cannot find an allowed cpu in
3996 * tsk->cpus_allowed, we fall back to task_cs(tsk)->cpus_allowed. In legacy
3997 * mode however, this value is the same as task_cs(tsk)->effective_cpus,
3998 * which will not contain a sane cpumask during cases such as cpu hotplugging.
3999 * This is the absolute last resort for the scheduler and it is only used if
4000 * _every_ other avenue has been traveled.
4001 *
4002 * Returns true if the affinity of @tsk was changed, false otherwise.
4003 **/
4004
4005bool cpuset_cpus_allowed_fallback(struct task_struct *tsk)
4006{
4007 const struct cpumask *possible_mask = task_cpu_possible_mask(tsk);
4008 const struct cpumask *cs_mask;
4009 bool changed = false;
4010
4011 rcu_read_lock();
4012 cs_mask = task_cs(tsk)->cpus_allowed;
4013 if (is_in_v2_mode() && cpumask_subset(cs_mask, possible_mask)) {
4014 do_set_cpus_allowed(tsk, cs_mask);
4015 changed = true;
4016 }
4017 rcu_read_unlock();
4018
4019 /*
4020 * We own tsk->cpus_allowed, nobody can change it under us.
4021 *
4022 * But we used cs && cs->cpus_allowed lockless and thus can
4023 * race with cgroup_attach_task() or update_cpumask() and get
4024 * the wrong tsk->cpus_allowed. However, both cases imply the
4025 * subsequent cpuset_change_cpumask()->set_cpus_allowed_ptr()
4026 * which takes task_rq_lock().
4027 *
4028 * If we are called after it dropped the lock we must see all
4029 * changes in tsk_cs()->cpus_allowed. Otherwise we can temporary
4030 * set any mask even if it is not right from task_cs() pov,
4031 * the pending set_cpus_allowed_ptr() will fix things.
4032 *
4033 * select_fallback_rq() will fix things ups and set cpu_possible_mask
4034 * if required.
4035 */
4036 return changed;
4037}
4038
4039void __init cpuset_init_current_mems_allowed(void)
4040{
4041 nodes_setall(current->mems_allowed);
4042}
4043
4044/**
4045 * cpuset_mems_allowed - return mems_allowed mask from a tasks cpuset.
4046 * @tsk: pointer to task_struct from which to obtain cpuset->mems_allowed.
4047 *
4048 * Description: Returns the nodemask_t mems_allowed of the cpuset
4049 * attached to the specified @tsk. Guaranteed to return some non-empty
4050 * subset of node_states[N_MEMORY], even if this means going outside the
4051 * tasks cpuset.
4052 **/
4053
4054nodemask_t cpuset_mems_allowed(struct task_struct *tsk)
4055{
4056 nodemask_t mask;
4057 unsigned long flags;
4058
4059 spin_lock_irqsave(&callback_lock, flags);
4060 rcu_read_lock();
4061 guarantee_online_mems(task_cs(tsk), &mask);
4062 rcu_read_unlock();
4063 spin_unlock_irqrestore(&callback_lock, flags);
4064
4065 return mask;
4066}
4067
4068/**
4069 * cpuset_nodemask_valid_mems_allowed - check nodemask vs. current mems_allowed
4070 * @nodemask: the nodemask to be checked
4071 *
4072 * Are any of the nodes in the nodemask allowed in current->mems_allowed?
4073 */
4074int cpuset_nodemask_valid_mems_allowed(nodemask_t *nodemask)
4075{
4076 return nodes_intersects(*nodemask, current->mems_allowed);
4077}
4078
4079/*
4080 * nearest_hardwall_ancestor() - Returns the nearest mem_exclusive or
4081 * mem_hardwall ancestor to the specified cpuset. Call holding
4082 * callback_lock. If no ancestor is mem_exclusive or mem_hardwall
4083 * (an unusual configuration), then returns the root cpuset.
4084 */
4085static struct cpuset *nearest_hardwall_ancestor(struct cpuset *cs)
4086{
4087 while (!(is_mem_exclusive(cs) || is_mem_hardwall(cs)) && parent_cs(cs))
4088 cs = parent_cs(cs);
4089 return cs;
4090}
4091
4092/*
4093 * cpuset_node_allowed - Can we allocate on a memory node?
4094 * @node: is this an allowed node?
4095 * @gfp_mask: memory allocation flags
4096 *
4097 * If we're in interrupt, yes, we can always allocate. If @node is set in
4098 * current's mems_allowed, yes. If it's not a __GFP_HARDWALL request and this
4099 * node is set in the nearest hardwalled cpuset ancestor to current's cpuset,
4100 * yes. If current has access to memory reserves as an oom victim, yes.
4101 * Otherwise, no.
4102 *
4103 * GFP_USER allocations are marked with the __GFP_HARDWALL bit,
4104 * and do not allow allocations outside the current tasks cpuset
4105 * unless the task has been OOM killed.
4106 * GFP_KERNEL allocations are not so marked, so can escape to the
4107 * nearest enclosing hardwalled ancestor cpuset.
4108 *
4109 * Scanning up parent cpusets requires callback_lock. The
4110 * __alloc_pages() routine only calls here with __GFP_HARDWALL bit
4111 * _not_ set if it's a GFP_KERNEL allocation, and all nodes in the
4112 * current tasks mems_allowed came up empty on the first pass over
4113 * the zonelist. So only GFP_KERNEL allocations, if all nodes in the
4114 * cpuset are short of memory, might require taking the callback_lock.
4115 *
4116 * The first call here from mm/page_alloc:get_page_from_freelist()
4117 * has __GFP_HARDWALL set in gfp_mask, enforcing hardwall cpusets,
4118 * so no allocation on a node outside the cpuset is allowed (unless
4119 * in interrupt, of course).
4120 *
4121 * The second pass through get_page_from_freelist() doesn't even call
4122 * here for GFP_ATOMIC calls. For those calls, the __alloc_pages()
4123 * variable 'wait' is not set, and the bit ALLOC_CPUSET is not set
4124 * in alloc_flags. That logic and the checks below have the combined
4125 * affect that:
4126 * in_interrupt - any node ok (current task context irrelevant)
4127 * GFP_ATOMIC - any node ok
4128 * tsk_is_oom_victim - any node ok
4129 * GFP_KERNEL - any node in enclosing hardwalled cpuset ok
4130 * GFP_USER - only nodes in current tasks mems allowed ok.
4131 */
4132bool cpuset_node_allowed(int node, gfp_t gfp_mask)
4133{
4134 struct cpuset *cs; /* current cpuset ancestors */
4135 bool allowed; /* is allocation in zone z allowed? */
4136 unsigned long flags;
4137
4138 if (in_interrupt())
4139 return true;
4140 if (node_isset(node, current->mems_allowed))
4141 return true;
4142 /*
4143 * Allow tasks that have access to memory reserves because they have
4144 * been OOM killed to get memory anywhere.
4145 */
4146 if (unlikely(tsk_is_oom_victim(current)))
4147 return true;
4148 if (gfp_mask & __GFP_HARDWALL) /* If hardwall request, stop here */
4149 return false;
4150
4151 if (current->flags & PF_EXITING) /* Let dying task have memory */
4152 return true;
4153
4154 /* Not hardwall and node outside mems_allowed: scan up cpusets */
4155 spin_lock_irqsave(&callback_lock, flags);
4156
4157 rcu_read_lock();
4158 cs = nearest_hardwall_ancestor(task_cs(current));
4159 allowed = node_isset(node, cs->mems_allowed);
4160 rcu_read_unlock();
4161
4162 spin_unlock_irqrestore(&callback_lock, flags);
4163 return allowed;
4164}
4165
4166/**
4167 * cpuset_spread_node() - On which node to begin search for a page
4168 * @rotor: round robin rotor
4169 *
4170 * If a task is marked PF_SPREAD_PAGE or PF_SPREAD_SLAB (as for
4171 * tasks in a cpuset with is_spread_page or is_spread_slab set),
4172 * and if the memory allocation used cpuset_mem_spread_node()
4173 * to determine on which node to start looking, as it will for
4174 * certain page cache or slab cache pages such as used for file
4175 * system buffers and inode caches, then instead of starting on the
4176 * local node to look for a free page, rather spread the starting
4177 * node around the tasks mems_allowed nodes.
4178 *
4179 * We don't have to worry about the returned node being offline
4180 * because "it can't happen", and even if it did, it would be ok.
4181 *
4182 * The routines calling guarantee_online_mems() are careful to
4183 * only set nodes in task->mems_allowed that are online. So it
4184 * should not be possible for the following code to return an
4185 * offline node. But if it did, that would be ok, as this routine
4186 * is not returning the node where the allocation must be, only
4187 * the node where the search should start. The zonelist passed to
4188 * __alloc_pages() will include all nodes. If the slab allocator
4189 * is passed an offline node, it will fall back to the local node.
4190 * See kmem_cache_alloc_node().
4191 */
4192static int cpuset_spread_node(int *rotor)
4193{
4194 return *rotor = next_node_in(*rotor, current->mems_allowed);
4195}
4196
4197/**
4198 * cpuset_mem_spread_node() - On which node to begin search for a file page
4199 */
4200int cpuset_mem_spread_node(void)
4201{
4202 if (current->cpuset_mem_spread_rotor == NUMA_NO_NODE)
4203 current->cpuset_mem_spread_rotor =
4204 node_random(¤t->mems_allowed);
4205
4206 return cpuset_spread_node(¤t->cpuset_mem_spread_rotor);
4207}
4208
4209/**
4210 * cpuset_mems_allowed_intersects - Does @tsk1's mems_allowed intersect @tsk2's?
4211 * @tsk1: pointer to task_struct of some task.
4212 * @tsk2: pointer to task_struct of some other task.
4213 *
4214 * Description: Return true if @tsk1's mems_allowed intersects the
4215 * mems_allowed of @tsk2. Used by the OOM killer to determine if
4216 * one of the task's memory usage might impact the memory available
4217 * to the other.
4218 **/
4219
4220int cpuset_mems_allowed_intersects(const struct task_struct *tsk1,
4221 const struct task_struct *tsk2)
4222{
4223 return nodes_intersects(tsk1->mems_allowed, tsk2->mems_allowed);
4224}
4225
4226/**
4227 * cpuset_print_current_mems_allowed - prints current's cpuset and mems_allowed
4228 *
4229 * Description: Prints current's name, cpuset name, and cached copy of its
4230 * mems_allowed to the kernel log.
4231 */
4232void cpuset_print_current_mems_allowed(void)
4233{
4234 struct cgroup *cgrp;
4235
4236 rcu_read_lock();
4237
4238 cgrp = task_cs(current)->css.cgroup;
4239 pr_cont(",cpuset=");
4240 pr_cont_cgroup_name(cgrp);
4241 pr_cont(",mems_allowed=%*pbl",
4242 nodemask_pr_args(¤t->mems_allowed));
4243
4244 rcu_read_unlock();
4245}
4246
4247#ifdef CONFIG_PROC_PID_CPUSET
4248/*
4249 * proc_cpuset_show()
4250 * - Print tasks cpuset path into seq_file.
4251 * - Used for /proc/<pid>/cpuset.
4252 * - No need to task_lock(tsk) on this tsk->cpuset reference, as it
4253 * doesn't really matter if tsk->cpuset changes after we read it,
4254 * and we take cpuset_mutex, keeping cpuset_attach() from changing it
4255 * anyway.
4256 */
4257int proc_cpuset_show(struct seq_file *m, struct pid_namespace *ns,
4258 struct pid *pid, struct task_struct *tsk)
4259{
4260 char *buf;
4261 struct cgroup_subsys_state *css;
4262 int retval;
4263
4264 retval = -ENOMEM;
4265 buf = kmalloc(PATH_MAX, GFP_KERNEL);
4266 if (!buf)
4267 goto out;
4268
4269 rcu_read_lock();
4270 spin_lock_irq(&css_set_lock);
4271 css = task_css(tsk, cpuset_cgrp_id);
4272 retval = cgroup_path_ns_locked(css->cgroup, buf, PATH_MAX,
4273 current->nsproxy->cgroup_ns);
4274 spin_unlock_irq(&css_set_lock);
4275 rcu_read_unlock();
4276
4277 if (retval == -E2BIG)
4278 retval = -ENAMETOOLONG;
4279 if (retval < 0)
4280 goto out_free;
4281 seq_puts(m, buf);
4282 seq_putc(m, '\n');
4283 retval = 0;
4284out_free:
4285 kfree(buf);
4286out:
4287 return retval;
4288}
4289#endif /* CONFIG_PROC_PID_CPUSET */
4290
4291/* Display task mems_allowed in /proc/<pid>/status file. */
4292void cpuset_task_status_allowed(struct seq_file *m, struct task_struct *task)
4293{
4294 seq_printf(m, "Mems_allowed:\t%*pb\n",
4295 nodemask_pr_args(&task->mems_allowed));
4296 seq_printf(m, "Mems_allowed_list:\t%*pbl\n",
4297 nodemask_pr_args(&task->mems_allowed));
4298}