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#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}
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/delay.h>
29#include <linux/init.h>
30#include <linux/interrupt.h>
31#include <linux/kernel.h>
32#include <linux/mempolicy.h>
33#include <linux/mm.h>
34#include <linux/memory.h>
35#include <linux/export.h>
36#include <linux/rcupdate.h>
37#include <linux/sched.h>
38#include <linux/sched/deadline.h>
39#include <linux/sched/mm.h>
40#include <linux/sched/task.h>
41#include <linux/security.h>
42#include <linux/spinlock.h>
43#include <linux/oom.h>
44#include <linux/sched/isolation.h>
45#include <linux/cgroup.h>
46#include <linux/wait.h>
47#include <linux/workqueue.h>
48
49DEFINE_STATIC_KEY_FALSE(cpusets_pre_enable_key);
50DEFINE_STATIC_KEY_FALSE(cpusets_enabled_key);
51
52/*
53 * There could be abnormal cpuset configurations for cpu or memory
54 * node binding, add this key to provide a quick low-cost judgment
55 * of the situation.
56 */
57DEFINE_STATIC_KEY_FALSE(cpusets_insane_config_key);
58
59/* See "Frequency meter" comments, below. */
60
61struct fmeter {
62 int cnt; /* unprocessed events count */
63 int val; /* most recent output value */
64 time64_t time; /* clock (secs) when val computed */
65 spinlock_t lock; /* guards read or write of above */
66};
67
68/*
69 * Invalid partition error code
70 */
71enum prs_errcode {
72 PERR_NONE = 0,
73 PERR_INVCPUS,
74 PERR_INVPARENT,
75 PERR_NOTPART,
76 PERR_NOTEXCL,
77 PERR_NOCPUS,
78 PERR_HOTPLUG,
79 PERR_CPUSEMPTY,
80 PERR_HKEEPING,
81};
82
83static const char * const perr_strings[] = {
84 [PERR_INVCPUS] = "Invalid cpu list in cpuset.cpus.exclusive",
85 [PERR_INVPARENT] = "Parent is an invalid partition root",
86 [PERR_NOTPART] = "Parent is not a partition root",
87 [PERR_NOTEXCL] = "Cpu list in cpuset.cpus not exclusive",
88 [PERR_NOCPUS] = "Parent unable to distribute cpu downstream",
89 [PERR_HOTPLUG] = "No cpu available due to hotplug",
90 [PERR_CPUSEMPTY] = "cpuset.cpus is empty",
91 [PERR_HKEEPING] = "partition config conflicts with housekeeping setup",
92};
93
94struct cpuset {
95 struct cgroup_subsys_state css;
96
97 unsigned long flags; /* "unsigned long" so bitops work */
98
99 /*
100 * On default hierarchy:
101 *
102 * The user-configured masks can only be changed by writing to
103 * cpuset.cpus and cpuset.mems, and won't be limited by the
104 * parent masks.
105 *
106 * The effective masks is the real masks that apply to the tasks
107 * in the cpuset. They may be changed if the configured masks are
108 * changed or hotplug happens.
109 *
110 * effective_mask == configured_mask & parent's effective_mask,
111 * and if it ends up empty, it will inherit the parent's mask.
112 *
113 *
114 * On legacy hierarchy:
115 *
116 * The user-configured masks are always the same with effective masks.
117 */
118
119 /* user-configured CPUs and Memory Nodes allow to tasks */
120 cpumask_var_t cpus_allowed;
121 nodemask_t mems_allowed;
122
123 /* effective CPUs and Memory Nodes allow to tasks */
124 cpumask_var_t effective_cpus;
125 nodemask_t effective_mems;
126
127 /*
128 * Exclusive CPUs dedicated to current cgroup (default hierarchy only)
129 *
130 * This exclusive CPUs must be a subset of cpus_allowed. A parent
131 * cgroup can only grant exclusive CPUs to one of its children.
132 *
133 * When the cgroup becomes a valid partition root, effective_xcpus
134 * defaults to cpus_allowed if not set. The effective_cpus of a valid
135 * partition root comes solely from its effective_xcpus and some of the
136 * effective_xcpus may be distributed to sub-partitions below & hence
137 * excluded from its effective_cpus.
138 */
139 cpumask_var_t effective_xcpus;
140
141 /*
142 * Exclusive CPUs as requested by the user (default hierarchy only)
143 */
144 cpumask_var_t exclusive_cpus;
145
146 /*
147 * This is old Memory Nodes tasks took on.
148 *
149 * - top_cpuset.old_mems_allowed is initialized to mems_allowed.
150 * - A new cpuset's old_mems_allowed is initialized when some
151 * task is moved into it.
152 * - old_mems_allowed is used in cpuset_migrate_mm() when we change
153 * cpuset.mems_allowed and have tasks' nodemask updated, and
154 * then old_mems_allowed is updated to mems_allowed.
155 */
156 nodemask_t old_mems_allowed;
157
158 struct fmeter fmeter; /* memory_pressure filter */
159
160 /*
161 * Tasks are being attached to this cpuset. Used to prevent
162 * zeroing cpus/mems_allowed between ->can_attach() and ->attach().
163 */
164 int attach_in_progress;
165
166 /* partition number for rebuild_sched_domains() */
167 int pn;
168
169 /* for custom sched domain */
170 int relax_domain_level;
171
172 /* number of valid sub-partitions */
173 int nr_subparts;
174
175 /* partition root state */
176 int partition_root_state;
177
178 /*
179 * Default hierarchy only:
180 * use_parent_ecpus - set if using parent's effective_cpus
181 * child_ecpus_count - # of children with use_parent_ecpus set
182 */
183 int use_parent_ecpus;
184 int child_ecpus_count;
185
186 /*
187 * number of SCHED_DEADLINE tasks attached to this cpuset, so that we
188 * know when to rebuild associated root domain bandwidth information.
189 */
190 int nr_deadline_tasks;
191 int nr_migrate_dl_tasks;
192 u64 sum_migrate_dl_bw;
193
194 /* Invalid partition error code, not lock protected */
195 enum prs_errcode prs_err;
196
197 /* Handle for cpuset.cpus.partition */
198 struct cgroup_file partition_file;
199
200 /* Remote partition silbling list anchored at remote_children */
201 struct list_head remote_sibling;
202};
203
204/*
205 * Exclusive CPUs distributed out to sub-partitions of top_cpuset
206 */
207static cpumask_var_t subpartitions_cpus;
208
209/*
210 * Exclusive CPUs in isolated partitions
211 */
212static cpumask_var_t isolated_cpus;
213
214/* List of remote partition root children */
215static struct list_head remote_children;
216
217/*
218 * Partition root states:
219 *
220 * 0 - member (not a partition root)
221 * 1 - partition root
222 * 2 - partition root without load balancing (isolated)
223 * -1 - invalid partition root
224 * -2 - invalid isolated partition root
225 */
226#define PRS_MEMBER 0
227#define PRS_ROOT 1
228#define PRS_ISOLATED 2
229#define PRS_INVALID_ROOT -1
230#define PRS_INVALID_ISOLATED -2
231
232static inline bool is_prs_invalid(int prs_state)
233{
234 return prs_state < 0;
235}
236
237/*
238 * Temporary cpumasks for working with partitions that are passed among
239 * functions to avoid memory allocation in inner functions.
240 */
241struct tmpmasks {
242 cpumask_var_t addmask, delmask; /* For partition root */
243 cpumask_var_t new_cpus; /* For update_cpumasks_hier() */
244};
245
246static inline struct cpuset *css_cs(struct cgroup_subsys_state *css)
247{
248 return css ? container_of(css, struct cpuset, css) : NULL;
249}
250
251/* Retrieve the cpuset for a task */
252static inline struct cpuset *task_cs(struct task_struct *task)
253{
254 return css_cs(task_css(task, cpuset_cgrp_id));
255}
256
257static inline struct cpuset *parent_cs(struct cpuset *cs)
258{
259 return css_cs(cs->css.parent);
260}
261
262void inc_dl_tasks_cs(struct task_struct *p)
263{
264 struct cpuset *cs = task_cs(p);
265
266 cs->nr_deadline_tasks++;
267}
268
269void dec_dl_tasks_cs(struct task_struct *p)
270{
271 struct cpuset *cs = task_cs(p);
272
273 cs->nr_deadline_tasks--;
274}
275
276/* bits in struct cpuset flags field */
277typedef enum {
278 CS_ONLINE,
279 CS_CPU_EXCLUSIVE,
280 CS_MEM_EXCLUSIVE,
281 CS_MEM_HARDWALL,
282 CS_MEMORY_MIGRATE,
283 CS_SCHED_LOAD_BALANCE,
284 CS_SPREAD_PAGE,
285 CS_SPREAD_SLAB,
286} cpuset_flagbits_t;
287
288/* convenient tests for these bits */
289static inline bool is_cpuset_online(struct cpuset *cs)
290{
291 return test_bit(CS_ONLINE, &cs->flags) && !css_is_dying(&cs->css);
292}
293
294static inline int is_cpu_exclusive(const struct cpuset *cs)
295{
296 return test_bit(CS_CPU_EXCLUSIVE, &cs->flags);
297}
298
299static inline int is_mem_exclusive(const struct cpuset *cs)
300{
301 return test_bit(CS_MEM_EXCLUSIVE, &cs->flags);
302}
303
304static inline int is_mem_hardwall(const struct cpuset *cs)
305{
306 return test_bit(CS_MEM_HARDWALL, &cs->flags);
307}
308
309static inline int is_sched_load_balance(const struct cpuset *cs)
310{
311 return test_bit(CS_SCHED_LOAD_BALANCE, &cs->flags);
312}
313
314static inline int is_memory_migrate(const struct cpuset *cs)
315{
316 return test_bit(CS_MEMORY_MIGRATE, &cs->flags);
317}
318
319static inline int is_spread_page(const struct cpuset *cs)
320{
321 return test_bit(CS_SPREAD_PAGE, &cs->flags);
322}
323
324static inline int is_spread_slab(const struct cpuset *cs)
325{
326 return test_bit(CS_SPREAD_SLAB, &cs->flags);
327}
328
329static inline int is_partition_valid(const struct cpuset *cs)
330{
331 return cs->partition_root_state > 0;
332}
333
334static inline int is_partition_invalid(const struct cpuset *cs)
335{
336 return cs->partition_root_state < 0;
337}
338
339/*
340 * Callers should hold callback_lock to modify partition_root_state.
341 */
342static inline void make_partition_invalid(struct cpuset *cs)
343{
344 if (cs->partition_root_state > 0)
345 cs->partition_root_state = -cs->partition_root_state;
346}
347
348/*
349 * Send notification event of whenever partition_root_state changes.
350 */
351static inline void notify_partition_change(struct cpuset *cs, int old_prs)
352{
353 if (old_prs == cs->partition_root_state)
354 return;
355 cgroup_file_notify(&cs->partition_file);
356
357 /* Reset prs_err if not invalid */
358 if (is_partition_valid(cs))
359 WRITE_ONCE(cs->prs_err, PERR_NONE);
360}
361
362static struct cpuset top_cpuset = {
363 .flags = ((1 << CS_ONLINE) | (1 << CS_CPU_EXCLUSIVE) |
364 (1 << CS_MEM_EXCLUSIVE)),
365 .partition_root_state = PRS_ROOT,
366 .remote_sibling = LIST_HEAD_INIT(top_cpuset.remote_sibling),
367};
368
369/**
370 * cpuset_for_each_child - traverse online children of a cpuset
371 * @child_cs: loop cursor pointing to the current child
372 * @pos_css: used for iteration
373 * @parent_cs: target cpuset to walk children of
374 *
375 * Walk @child_cs through the online children of @parent_cs. Must be used
376 * with RCU read locked.
377 */
378#define cpuset_for_each_child(child_cs, pos_css, parent_cs) \
379 css_for_each_child((pos_css), &(parent_cs)->css) \
380 if (is_cpuset_online(((child_cs) = css_cs((pos_css)))))
381
382/**
383 * cpuset_for_each_descendant_pre - pre-order walk of a cpuset's descendants
384 * @des_cs: loop cursor pointing to the current descendant
385 * @pos_css: used for iteration
386 * @root_cs: target cpuset to walk ancestor of
387 *
388 * Walk @des_cs through the online descendants of @root_cs. Must be used
389 * with RCU read locked. The caller may modify @pos_css by calling
390 * css_rightmost_descendant() to skip subtree. @root_cs is included in the
391 * iteration and the first node to be visited.
392 */
393#define cpuset_for_each_descendant_pre(des_cs, pos_css, root_cs) \
394 css_for_each_descendant_pre((pos_css), &(root_cs)->css) \
395 if (is_cpuset_online(((des_cs) = css_cs((pos_css)))))
396
397/*
398 * There are two global locks guarding cpuset structures - cpuset_mutex and
399 * callback_lock. We also require taking task_lock() when dereferencing a
400 * task's cpuset pointer. See "The task_lock() exception", at the end of this
401 * comment. The cpuset code uses only cpuset_mutex. Other kernel subsystems
402 * can use cpuset_lock()/cpuset_unlock() to prevent change to cpuset
403 * structures. Note that cpuset_mutex needs to be a mutex as it is used in
404 * paths that rely on priority inheritance (e.g. scheduler - on RT) for
405 * correctness.
406 *
407 * A task must hold both locks to modify cpusets. If a task holds
408 * cpuset_mutex, it blocks others, ensuring that it is the only task able to
409 * also acquire callback_lock and be able to modify cpusets. It can perform
410 * various checks on the cpuset structure first, knowing nothing will change.
411 * It can also allocate memory while just holding cpuset_mutex. While it is
412 * performing these checks, various callback routines can briefly acquire
413 * callback_lock to query cpusets. Once it is ready to make the changes, it
414 * takes callback_lock, blocking everyone else.
415 *
416 * Calls to the kernel memory allocator can not be made while holding
417 * callback_lock, as that would risk double tripping on callback_lock
418 * from one of the callbacks into the cpuset code from within
419 * __alloc_pages().
420 *
421 * If a task is only holding callback_lock, then it has read-only
422 * access to cpusets.
423 *
424 * Now, the task_struct fields mems_allowed and mempolicy may be changed
425 * by other task, we use alloc_lock in the task_struct fields to protect
426 * them.
427 *
428 * The cpuset_common_file_read() handlers only hold callback_lock across
429 * small pieces of code, such as when reading out possibly multi-word
430 * cpumasks and nodemasks.
431 *
432 * Accessing a task's cpuset should be done in accordance with the
433 * guidelines for accessing subsystem state in kernel/cgroup.c
434 */
435
436static DEFINE_MUTEX(cpuset_mutex);
437
438void cpuset_lock(void)
439{
440 mutex_lock(&cpuset_mutex);
441}
442
443void cpuset_unlock(void)
444{
445 mutex_unlock(&cpuset_mutex);
446}
447
448static DEFINE_SPINLOCK(callback_lock);
449
450static struct workqueue_struct *cpuset_migrate_mm_wq;
451
452/*
453 * CPU / memory hotplug is handled asynchronously.
454 */
455static void cpuset_hotplug_workfn(struct work_struct *work);
456static DECLARE_WORK(cpuset_hotplug_work, cpuset_hotplug_workfn);
457
458static DECLARE_WAIT_QUEUE_HEAD(cpuset_attach_wq);
459
460static inline void check_insane_mems_config(nodemask_t *nodes)
461{
462 if (!cpusets_insane_config() &&
463 movable_only_nodes(nodes)) {
464 static_branch_enable(&cpusets_insane_config_key);
465 pr_info("Unsupported (movable nodes only) cpuset configuration detected (nmask=%*pbl)!\n"
466 "Cpuset allocations might fail even with a lot of memory available.\n",
467 nodemask_pr_args(nodes));
468 }
469}
470
471/*
472 * Cgroup v2 behavior is used on the "cpus" and "mems" control files when
473 * on default hierarchy or when the cpuset_v2_mode flag is set by mounting
474 * the v1 cpuset cgroup filesystem with the "cpuset_v2_mode" mount option.
475 * With v2 behavior, "cpus" and "mems" are always what the users have
476 * requested and won't be changed by hotplug events. Only the effective
477 * cpus or mems will be affected.
478 */
479static inline bool is_in_v2_mode(void)
480{
481 return cgroup_subsys_on_dfl(cpuset_cgrp_subsys) ||
482 (cpuset_cgrp_subsys.root->flags & CGRP_ROOT_CPUSET_V2_MODE);
483}
484
485/**
486 * partition_is_populated - check if partition has tasks
487 * @cs: partition root to be checked
488 * @excluded_child: a child cpuset to be excluded in task checking
489 * Return: true if there are tasks, false otherwise
490 *
491 * It is assumed that @cs is a valid partition root. @excluded_child should
492 * be non-NULL when this cpuset is going to become a partition itself.
493 */
494static inline bool partition_is_populated(struct cpuset *cs,
495 struct cpuset *excluded_child)
496{
497 struct cgroup_subsys_state *css;
498 struct cpuset *child;
499
500 if (cs->css.cgroup->nr_populated_csets)
501 return true;
502 if (!excluded_child && !cs->nr_subparts)
503 return cgroup_is_populated(cs->css.cgroup);
504
505 rcu_read_lock();
506 cpuset_for_each_child(child, css, cs) {
507 if (child == excluded_child)
508 continue;
509 if (is_partition_valid(child))
510 continue;
511 if (cgroup_is_populated(child->css.cgroup)) {
512 rcu_read_unlock();
513 return true;
514 }
515 }
516 rcu_read_unlock();
517 return false;
518}
519
520/*
521 * Return in pmask the portion of a task's cpusets's cpus_allowed that
522 * are online and are capable of running the task. If none are found,
523 * walk up the cpuset hierarchy until we find one that does have some
524 * appropriate cpus.
525 *
526 * One way or another, we guarantee to return some non-empty subset
527 * of cpu_online_mask.
528 *
529 * Call with callback_lock or cpuset_mutex held.
530 */
531static void guarantee_online_cpus(struct task_struct *tsk,
532 struct cpumask *pmask)
533{
534 const struct cpumask *possible_mask = task_cpu_possible_mask(tsk);
535 struct cpuset *cs;
536
537 if (WARN_ON(!cpumask_and(pmask, possible_mask, cpu_online_mask)))
538 cpumask_copy(pmask, cpu_online_mask);
539
540 rcu_read_lock();
541 cs = task_cs(tsk);
542
543 while (!cpumask_intersects(cs->effective_cpus, pmask)) {
544 cs = parent_cs(cs);
545 if (unlikely(!cs)) {
546 /*
547 * The top cpuset doesn't have any online cpu as a
548 * consequence of a race between cpuset_hotplug_work
549 * and cpu hotplug notifier. But we know the top
550 * cpuset's effective_cpus is on its way to be
551 * identical to cpu_online_mask.
552 */
553 goto out_unlock;
554 }
555 }
556 cpumask_and(pmask, pmask, cs->effective_cpus);
557
558out_unlock:
559 rcu_read_unlock();
560}
561
562/*
563 * Return in *pmask the portion of a cpusets's mems_allowed that
564 * are online, with memory. If none are online with memory, walk
565 * up the cpuset hierarchy until we find one that does have some
566 * online mems. The top cpuset always has some mems online.
567 *
568 * One way or another, we guarantee to return some non-empty subset
569 * of node_states[N_MEMORY].
570 *
571 * Call with callback_lock or cpuset_mutex held.
572 */
573static void guarantee_online_mems(struct cpuset *cs, nodemask_t *pmask)
574{
575 while (!nodes_intersects(cs->effective_mems, node_states[N_MEMORY]))
576 cs = parent_cs(cs);
577 nodes_and(*pmask, cs->effective_mems, node_states[N_MEMORY]);
578}
579
580/*
581 * update task's spread flag if cpuset's page/slab spread flag is set
582 *
583 * Call with callback_lock or cpuset_mutex held. The check can be skipped
584 * if on default hierarchy.
585 */
586static void cpuset_update_task_spread_flags(struct cpuset *cs,
587 struct task_struct *tsk)
588{
589 if (cgroup_subsys_on_dfl(cpuset_cgrp_subsys))
590 return;
591
592 if (is_spread_page(cs))
593 task_set_spread_page(tsk);
594 else
595 task_clear_spread_page(tsk);
596
597 if (is_spread_slab(cs))
598 task_set_spread_slab(tsk);
599 else
600 task_clear_spread_slab(tsk);
601}
602
603/*
604 * is_cpuset_subset(p, q) - Is cpuset p a subset of cpuset q?
605 *
606 * One cpuset is a subset of another if all its allowed CPUs and
607 * Memory Nodes are a subset of the other, and its exclusive flags
608 * are only set if the other's are set. Call holding cpuset_mutex.
609 */
610
611static int is_cpuset_subset(const struct cpuset *p, const struct cpuset *q)
612{
613 return cpumask_subset(p->cpus_allowed, q->cpus_allowed) &&
614 nodes_subset(p->mems_allowed, q->mems_allowed) &&
615 is_cpu_exclusive(p) <= is_cpu_exclusive(q) &&
616 is_mem_exclusive(p) <= is_mem_exclusive(q);
617}
618
619/**
620 * alloc_cpumasks - allocate three cpumasks for cpuset
621 * @cs: the cpuset that have cpumasks to be allocated.
622 * @tmp: the tmpmasks structure pointer
623 * Return: 0 if successful, -ENOMEM otherwise.
624 *
625 * Only one of the two input arguments should be non-NULL.
626 */
627static inline int alloc_cpumasks(struct cpuset *cs, struct tmpmasks *tmp)
628{
629 cpumask_var_t *pmask1, *pmask2, *pmask3, *pmask4;
630
631 if (cs) {
632 pmask1 = &cs->cpus_allowed;
633 pmask2 = &cs->effective_cpus;
634 pmask3 = &cs->effective_xcpus;
635 pmask4 = &cs->exclusive_cpus;
636 } else {
637 pmask1 = &tmp->new_cpus;
638 pmask2 = &tmp->addmask;
639 pmask3 = &tmp->delmask;
640 pmask4 = NULL;
641 }
642
643 if (!zalloc_cpumask_var(pmask1, GFP_KERNEL))
644 return -ENOMEM;
645
646 if (!zalloc_cpumask_var(pmask2, GFP_KERNEL))
647 goto free_one;
648
649 if (!zalloc_cpumask_var(pmask3, GFP_KERNEL))
650 goto free_two;
651
652 if (pmask4 && !zalloc_cpumask_var(pmask4, GFP_KERNEL))
653 goto free_three;
654
655
656 return 0;
657
658free_three:
659 free_cpumask_var(*pmask3);
660free_two:
661 free_cpumask_var(*pmask2);
662free_one:
663 free_cpumask_var(*pmask1);
664 return -ENOMEM;
665}
666
667/**
668 * free_cpumasks - free cpumasks in a tmpmasks structure
669 * @cs: the cpuset that have cpumasks to be free.
670 * @tmp: the tmpmasks structure pointer
671 */
672static inline void free_cpumasks(struct cpuset *cs, struct tmpmasks *tmp)
673{
674 if (cs) {
675 free_cpumask_var(cs->cpus_allowed);
676 free_cpumask_var(cs->effective_cpus);
677 free_cpumask_var(cs->effective_xcpus);
678 free_cpumask_var(cs->exclusive_cpus);
679 }
680 if (tmp) {
681 free_cpumask_var(tmp->new_cpus);
682 free_cpumask_var(tmp->addmask);
683 free_cpumask_var(tmp->delmask);
684 }
685}
686
687/**
688 * alloc_trial_cpuset - allocate a trial cpuset
689 * @cs: the cpuset that the trial cpuset duplicates
690 */
691static struct cpuset *alloc_trial_cpuset(struct cpuset *cs)
692{
693 struct cpuset *trial;
694
695 trial = kmemdup(cs, sizeof(*cs), GFP_KERNEL);
696 if (!trial)
697 return NULL;
698
699 if (alloc_cpumasks(trial, NULL)) {
700 kfree(trial);
701 return NULL;
702 }
703
704 cpumask_copy(trial->cpus_allowed, cs->cpus_allowed);
705 cpumask_copy(trial->effective_cpus, cs->effective_cpus);
706 cpumask_copy(trial->effective_xcpus, cs->effective_xcpus);
707 cpumask_copy(trial->exclusive_cpus, cs->exclusive_cpus);
708 return trial;
709}
710
711/**
712 * free_cpuset - free the cpuset
713 * @cs: the cpuset to be freed
714 */
715static inline void free_cpuset(struct cpuset *cs)
716{
717 free_cpumasks(cs, NULL);
718 kfree(cs);
719}
720
721static inline struct cpumask *fetch_xcpus(struct cpuset *cs)
722{
723 return !cpumask_empty(cs->exclusive_cpus) ? cs->exclusive_cpus :
724 cpumask_empty(cs->effective_xcpus) ? cs->cpus_allowed
725 : cs->effective_xcpus;
726}
727
728/*
729 * cpusets_are_exclusive() - check if two cpusets are exclusive
730 *
731 * Return true if exclusive, false if not
732 */
733static inline bool cpusets_are_exclusive(struct cpuset *cs1, struct cpuset *cs2)
734{
735 struct cpumask *xcpus1 = fetch_xcpus(cs1);
736 struct cpumask *xcpus2 = fetch_xcpus(cs2);
737
738 if (cpumask_intersects(xcpus1, xcpus2))
739 return false;
740 return true;
741}
742
743/*
744 * validate_change_legacy() - Validate conditions specific to legacy (v1)
745 * behavior.
746 */
747static int validate_change_legacy(struct cpuset *cur, struct cpuset *trial)
748{
749 struct cgroup_subsys_state *css;
750 struct cpuset *c, *par;
751 int ret;
752
753 WARN_ON_ONCE(!rcu_read_lock_held());
754
755 /* Each of our child cpusets must be a subset of us */
756 ret = -EBUSY;
757 cpuset_for_each_child(c, css, cur)
758 if (!is_cpuset_subset(c, trial))
759 goto out;
760
761 /* On legacy hierarchy, we must be a subset of our parent cpuset. */
762 ret = -EACCES;
763 par = parent_cs(cur);
764 if (par && !is_cpuset_subset(trial, par))
765 goto out;
766
767 ret = 0;
768out:
769 return ret;
770}
771
772/*
773 * validate_change() - Used to validate that any proposed cpuset change
774 * follows the structural rules for cpusets.
775 *
776 * If we replaced the flag and mask values of the current cpuset
777 * (cur) with those values in the trial cpuset (trial), would
778 * our various subset and exclusive rules still be valid? Presumes
779 * cpuset_mutex held.
780 *
781 * 'cur' is the address of an actual, in-use cpuset. Operations
782 * such as list traversal that depend on the actual address of the
783 * cpuset in the list must use cur below, not trial.
784 *
785 * 'trial' is the address of bulk structure copy of cur, with
786 * perhaps one or more of the fields cpus_allowed, mems_allowed,
787 * or flags changed to new, trial values.
788 *
789 * Return 0 if valid, -errno if not.
790 */
791
792static int validate_change(struct cpuset *cur, struct cpuset *trial)
793{
794 struct cgroup_subsys_state *css;
795 struct cpuset *c, *par;
796 int ret = 0;
797
798 rcu_read_lock();
799
800 if (!is_in_v2_mode())
801 ret = validate_change_legacy(cur, trial);
802 if (ret)
803 goto out;
804
805 /* Remaining checks don't apply to root cpuset */
806 if (cur == &top_cpuset)
807 goto out;
808
809 par = parent_cs(cur);
810
811 /*
812 * Cpusets with tasks - existing or newly being attached - can't
813 * be changed to have empty cpus_allowed or mems_allowed.
814 */
815 ret = -ENOSPC;
816 if ((cgroup_is_populated(cur->css.cgroup) || cur->attach_in_progress)) {
817 if (!cpumask_empty(cur->cpus_allowed) &&
818 cpumask_empty(trial->cpus_allowed))
819 goto out;
820 if (!nodes_empty(cur->mems_allowed) &&
821 nodes_empty(trial->mems_allowed))
822 goto out;
823 }
824
825 /*
826 * We can't shrink if we won't have enough room for SCHED_DEADLINE
827 * tasks.
828 */
829 ret = -EBUSY;
830 if (is_cpu_exclusive(cur) &&
831 !cpuset_cpumask_can_shrink(cur->cpus_allowed,
832 trial->cpus_allowed))
833 goto out;
834
835 /*
836 * If either I or some sibling (!= me) is exclusive, we can't
837 * overlap
838 */
839 ret = -EINVAL;
840 cpuset_for_each_child(c, css, par) {
841 if ((is_cpu_exclusive(trial) || is_cpu_exclusive(c)) &&
842 c != cur) {
843 if (!cpusets_are_exclusive(trial, c))
844 goto out;
845 }
846 if ((is_mem_exclusive(trial) || is_mem_exclusive(c)) &&
847 c != cur &&
848 nodes_intersects(trial->mems_allowed, c->mems_allowed))
849 goto out;
850 }
851
852 ret = 0;
853out:
854 rcu_read_unlock();
855 return ret;
856}
857
858#ifdef CONFIG_SMP
859/*
860 * Helper routine for generate_sched_domains().
861 * Do cpusets a, b have overlapping effective cpus_allowed masks?
862 */
863static int cpusets_overlap(struct cpuset *a, struct cpuset *b)
864{
865 return cpumask_intersects(a->effective_cpus, b->effective_cpus);
866}
867
868static void
869update_domain_attr(struct sched_domain_attr *dattr, struct cpuset *c)
870{
871 if (dattr->relax_domain_level < c->relax_domain_level)
872 dattr->relax_domain_level = c->relax_domain_level;
873 return;
874}
875
876static void update_domain_attr_tree(struct sched_domain_attr *dattr,
877 struct cpuset *root_cs)
878{
879 struct cpuset *cp;
880 struct cgroup_subsys_state *pos_css;
881
882 rcu_read_lock();
883 cpuset_for_each_descendant_pre(cp, pos_css, root_cs) {
884 /* skip the whole subtree if @cp doesn't have any CPU */
885 if (cpumask_empty(cp->cpus_allowed)) {
886 pos_css = css_rightmost_descendant(pos_css);
887 continue;
888 }
889
890 if (is_sched_load_balance(cp))
891 update_domain_attr(dattr, cp);
892 }
893 rcu_read_unlock();
894}
895
896/* Must be called with cpuset_mutex held. */
897static inline int nr_cpusets(void)
898{
899 /* jump label reference count + the top-level cpuset */
900 return static_key_count(&cpusets_enabled_key.key) + 1;
901}
902
903/*
904 * generate_sched_domains()
905 *
906 * This function builds a partial partition of the systems CPUs
907 * A 'partial partition' is a set of non-overlapping subsets whose
908 * union is a subset of that set.
909 * The output of this function needs to be passed to kernel/sched/core.c
910 * partition_sched_domains() routine, which will rebuild the scheduler's
911 * load balancing domains (sched domains) as specified by that partial
912 * partition.
913 *
914 * See "What is sched_load_balance" in Documentation/admin-guide/cgroup-v1/cpusets.rst
915 * for a background explanation of this.
916 *
917 * Does not return errors, on the theory that the callers of this
918 * routine would rather not worry about failures to rebuild sched
919 * domains when operating in the severe memory shortage situations
920 * that could cause allocation failures below.
921 *
922 * Must be called with cpuset_mutex held.
923 *
924 * The three key local variables below are:
925 * cp - cpuset pointer, used (together with pos_css) to perform a
926 * top-down scan of all cpusets. For our purposes, rebuilding
927 * the schedulers sched domains, we can ignore !is_sched_load_
928 * balance cpusets.
929 * csa - (for CpuSet Array) Array of pointers to all the cpusets
930 * that need to be load balanced, for convenient iterative
931 * access by the subsequent code that finds the best partition,
932 * i.e the set of domains (subsets) of CPUs such that the
933 * cpus_allowed of every cpuset marked is_sched_load_balance
934 * is a subset of one of these domains, while there are as
935 * many such domains as possible, each as small as possible.
936 * doms - Conversion of 'csa' to an array of cpumasks, for passing to
937 * the kernel/sched/core.c routine partition_sched_domains() in a
938 * convenient format, that can be easily compared to the prior
939 * value to determine what partition elements (sched domains)
940 * were changed (added or removed.)
941 *
942 * Finding the best partition (set of domains):
943 * The triple nested loops below over i, j, k scan over the
944 * load balanced cpusets (using the array of cpuset pointers in
945 * csa[]) looking for pairs of cpusets that have overlapping
946 * cpus_allowed, but which don't have the same 'pn' partition
947 * number and gives them in the same partition number. It keeps
948 * looping on the 'restart' label until it can no longer find
949 * any such pairs.
950 *
951 * The union of the cpus_allowed masks from the set of
952 * all cpusets having the same 'pn' value then form the one
953 * element of the partition (one sched domain) to be passed to
954 * partition_sched_domains().
955 */
956static int generate_sched_domains(cpumask_var_t **domains,
957 struct sched_domain_attr **attributes)
958{
959 struct cpuset *cp; /* top-down scan of cpusets */
960 struct cpuset **csa; /* array of all cpuset ptrs */
961 int csn; /* how many cpuset ptrs in csa so far */
962 int i, j, k; /* indices for partition finding loops */
963 cpumask_var_t *doms; /* resulting partition; i.e. sched domains */
964 struct sched_domain_attr *dattr; /* attributes for custom domains */
965 int ndoms = 0; /* number of sched domains in result */
966 int nslot; /* next empty doms[] struct cpumask slot */
967 struct cgroup_subsys_state *pos_css;
968 bool root_load_balance = is_sched_load_balance(&top_cpuset);
969
970 doms = NULL;
971 dattr = NULL;
972 csa = NULL;
973
974 /* Special case for the 99% of systems with one, full, sched domain */
975 if (root_load_balance && !top_cpuset.nr_subparts) {
976 ndoms = 1;
977 doms = alloc_sched_domains(ndoms);
978 if (!doms)
979 goto done;
980
981 dattr = kmalloc(sizeof(struct sched_domain_attr), GFP_KERNEL);
982 if (dattr) {
983 *dattr = SD_ATTR_INIT;
984 update_domain_attr_tree(dattr, &top_cpuset);
985 }
986 cpumask_and(doms[0], top_cpuset.effective_cpus,
987 housekeeping_cpumask(HK_TYPE_DOMAIN));
988
989 goto done;
990 }
991
992 csa = kmalloc_array(nr_cpusets(), sizeof(cp), GFP_KERNEL);
993 if (!csa)
994 goto done;
995 csn = 0;
996
997 rcu_read_lock();
998 if (root_load_balance)
999 csa[csn++] = &top_cpuset;
1000 cpuset_for_each_descendant_pre(cp, pos_css, &top_cpuset) {
1001 if (cp == &top_cpuset)
1002 continue;
1003 /*
1004 * Continue traversing beyond @cp iff @cp has some CPUs and
1005 * isn't load balancing. The former is obvious. The
1006 * latter: All child cpusets contain a subset of the
1007 * parent's cpus, so just skip them, and then we call
1008 * update_domain_attr_tree() to calc relax_domain_level of
1009 * the corresponding sched domain.
1010 *
1011 * If root is load-balancing, we can skip @cp if it
1012 * is a subset of the root's effective_cpus.
1013 */
1014 if (!cpumask_empty(cp->cpus_allowed) &&
1015 !(is_sched_load_balance(cp) &&
1016 cpumask_intersects(cp->cpus_allowed,
1017 housekeeping_cpumask(HK_TYPE_DOMAIN))))
1018 continue;
1019
1020 if (root_load_balance &&
1021 cpumask_subset(cp->cpus_allowed, top_cpuset.effective_cpus))
1022 continue;
1023
1024 if (is_sched_load_balance(cp) &&
1025 !cpumask_empty(cp->effective_cpus))
1026 csa[csn++] = cp;
1027
1028 /* skip @cp's subtree if not a partition root */
1029 if (!is_partition_valid(cp))
1030 pos_css = css_rightmost_descendant(pos_css);
1031 }
1032 rcu_read_unlock();
1033
1034 for (i = 0; i < csn; i++)
1035 csa[i]->pn = i;
1036 ndoms = csn;
1037
1038restart:
1039 /* Find the best partition (set of sched domains) */
1040 for (i = 0; i < csn; i++) {
1041 struct cpuset *a = csa[i];
1042 int apn = a->pn;
1043
1044 for (j = 0; j < csn; j++) {
1045 struct cpuset *b = csa[j];
1046 int bpn = b->pn;
1047
1048 if (apn != bpn && cpusets_overlap(a, b)) {
1049 for (k = 0; k < csn; k++) {
1050 struct cpuset *c = csa[k];
1051
1052 if (c->pn == bpn)
1053 c->pn = apn;
1054 }
1055 ndoms--; /* one less element */
1056 goto restart;
1057 }
1058 }
1059 }
1060
1061 /*
1062 * Now we know how many domains to create.
1063 * Convert <csn, csa> to <ndoms, doms> and populate cpu masks.
1064 */
1065 doms = alloc_sched_domains(ndoms);
1066 if (!doms)
1067 goto done;
1068
1069 /*
1070 * The rest of the code, including the scheduler, can deal with
1071 * dattr==NULL case. No need to abort if alloc fails.
1072 */
1073 dattr = kmalloc_array(ndoms, sizeof(struct sched_domain_attr),
1074 GFP_KERNEL);
1075
1076 for (nslot = 0, i = 0; i < csn; i++) {
1077 struct cpuset *a = csa[i];
1078 struct cpumask *dp;
1079 int apn = a->pn;
1080
1081 if (apn < 0) {
1082 /* Skip completed partitions */
1083 continue;
1084 }
1085
1086 dp = doms[nslot];
1087
1088 if (nslot == ndoms) {
1089 static int warnings = 10;
1090 if (warnings) {
1091 pr_warn("rebuild_sched_domains confused: nslot %d, ndoms %d, csn %d, i %d, apn %d\n",
1092 nslot, ndoms, csn, i, apn);
1093 warnings--;
1094 }
1095 continue;
1096 }
1097
1098 cpumask_clear(dp);
1099 if (dattr)
1100 *(dattr + nslot) = SD_ATTR_INIT;
1101 for (j = i; j < csn; j++) {
1102 struct cpuset *b = csa[j];
1103
1104 if (apn == b->pn) {
1105 cpumask_or(dp, dp, b->effective_cpus);
1106 cpumask_and(dp, dp, housekeeping_cpumask(HK_TYPE_DOMAIN));
1107 if (dattr)
1108 update_domain_attr_tree(dattr + nslot, b);
1109
1110 /* Done with this partition */
1111 b->pn = -1;
1112 }
1113 }
1114 nslot++;
1115 }
1116 BUG_ON(nslot != ndoms);
1117
1118done:
1119 kfree(csa);
1120
1121 /*
1122 * Fallback to the default domain if kmalloc() failed.
1123 * See comments in partition_sched_domains().
1124 */
1125 if (doms == NULL)
1126 ndoms = 1;
1127
1128 *domains = doms;
1129 *attributes = dattr;
1130 return ndoms;
1131}
1132
1133static void dl_update_tasks_root_domain(struct cpuset *cs)
1134{
1135 struct css_task_iter it;
1136 struct task_struct *task;
1137
1138 if (cs->nr_deadline_tasks == 0)
1139 return;
1140
1141 css_task_iter_start(&cs->css, 0, &it);
1142
1143 while ((task = css_task_iter_next(&it)))
1144 dl_add_task_root_domain(task);
1145
1146 css_task_iter_end(&it);
1147}
1148
1149static void dl_rebuild_rd_accounting(void)
1150{
1151 struct cpuset *cs = NULL;
1152 struct cgroup_subsys_state *pos_css;
1153
1154 lockdep_assert_held(&cpuset_mutex);
1155 lockdep_assert_cpus_held();
1156 lockdep_assert_held(&sched_domains_mutex);
1157
1158 rcu_read_lock();
1159
1160 /*
1161 * Clear default root domain DL accounting, it will be computed again
1162 * if a task belongs to it.
1163 */
1164 dl_clear_root_domain(&def_root_domain);
1165
1166 cpuset_for_each_descendant_pre(cs, pos_css, &top_cpuset) {
1167
1168 if (cpumask_empty(cs->effective_cpus)) {
1169 pos_css = css_rightmost_descendant(pos_css);
1170 continue;
1171 }
1172
1173 css_get(&cs->css);
1174
1175 rcu_read_unlock();
1176
1177 dl_update_tasks_root_domain(cs);
1178
1179 rcu_read_lock();
1180 css_put(&cs->css);
1181 }
1182 rcu_read_unlock();
1183}
1184
1185static void
1186partition_and_rebuild_sched_domains(int ndoms_new, cpumask_var_t doms_new[],
1187 struct sched_domain_attr *dattr_new)
1188{
1189 mutex_lock(&sched_domains_mutex);
1190 partition_sched_domains_locked(ndoms_new, doms_new, dattr_new);
1191 dl_rebuild_rd_accounting();
1192 mutex_unlock(&sched_domains_mutex);
1193}
1194
1195/*
1196 * Rebuild scheduler domains.
1197 *
1198 * If the flag 'sched_load_balance' of any cpuset with non-empty
1199 * 'cpus' changes, or if the 'cpus' allowed changes in any cpuset
1200 * which has that flag enabled, or if any cpuset with a non-empty
1201 * 'cpus' is removed, then call this routine to rebuild the
1202 * scheduler's dynamic sched domains.
1203 *
1204 * Call with cpuset_mutex held. Takes cpus_read_lock().
1205 */
1206static void rebuild_sched_domains_locked(void)
1207{
1208 struct cgroup_subsys_state *pos_css;
1209 struct sched_domain_attr *attr;
1210 cpumask_var_t *doms;
1211 struct cpuset *cs;
1212 int ndoms;
1213
1214 lockdep_assert_cpus_held();
1215 lockdep_assert_held(&cpuset_mutex);
1216
1217 /*
1218 * If we have raced with CPU hotplug, return early to avoid
1219 * passing doms with offlined cpu to partition_sched_domains().
1220 * Anyways, cpuset_hotplug_workfn() will rebuild sched domains.
1221 *
1222 * With no CPUs in any subpartitions, top_cpuset's effective CPUs
1223 * should be the same as the active CPUs, so checking only top_cpuset
1224 * is enough to detect racing CPU offlines.
1225 */
1226 if (cpumask_empty(subpartitions_cpus) &&
1227 !cpumask_equal(top_cpuset.effective_cpus, cpu_active_mask))
1228 return;
1229
1230 /*
1231 * With subpartition CPUs, however, the effective CPUs of a partition
1232 * root should be only a subset of the active CPUs. Since a CPU in any
1233 * partition root could be offlined, all must be checked.
1234 */
1235 if (top_cpuset.nr_subparts) {
1236 rcu_read_lock();
1237 cpuset_for_each_descendant_pre(cs, pos_css, &top_cpuset) {
1238 if (!is_partition_valid(cs)) {
1239 pos_css = css_rightmost_descendant(pos_css);
1240 continue;
1241 }
1242 if (!cpumask_subset(cs->effective_cpus,
1243 cpu_active_mask)) {
1244 rcu_read_unlock();
1245 return;
1246 }
1247 }
1248 rcu_read_unlock();
1249 }
1250
1251 /* Generate domain masks and attrs */
1252 ndoms = generate_sched_domains(&doms, &attr);
1253
1254 /* Have scheduler rebuild the domains */
1255 partition_and_rebuild_sched_domains(ndoms, doms, attr);
1256}
1257#else /* !CONFIG_SMP */
1258static void rebuild_sched_domains_locked(void)
1259{
1260}
1261#endif /* CONFIG_SMP */
1262
1263void rebuild_sched_domains(void)
1264{
1265 cpus_read_lock();
1266 mutex_lock(&cpuset_mutex);
1267 rebuild_sched_domains_locked();
1268 mutex_unlock(&cpuset_mutex);
1269 cpus_read_unlock();
1270}
1271
1272/**
1273 * update_tasks_cpumask - Update the cpumasks of tasks in the cpuset.
1274 * @cs: the cpuset in which each task's cpus_allowed mask needs to be changed
1275 * @new_cpus: the temp variable for the new effective_cpus mask
1276 *
1277 * Iterate through each task of @cs updating its cpus_allowed to the
1278 * effective cpuset's. As this function is called with cpuset_mutex held,
1279 * cpuset membership stays stable. For top_cpuset, task_cpu_possible_mask()
1280 * is used instead of effective_cpus to make sure all offline CPUs are also
1281 * included as hotplug code won't update cpumasks for tasks in top_cpuset.
1282 */
1283static void update_tasks_cpumask(struct cpuset *cs, struct cpumask *new_cpus)
1284{
1285 struct css_task_iter it;
1286 struct task_struct *task;
1287 bool top_cs = cs == &top_cpuset;
1288
1289 css_task_iter_start(&cs->css, 0, &it);
1290 while ((task = css_task_iter_next(&it))) {
1291 const struct cpumask *possible_mask = task_cpu_possible_mask(task);
1292
1293 if (top_cs) {
1294 /*
1295 * Percpu kthreads in top_cpuset are ignored
1296 */
1297 if (kthread_is_per_cpu(task))
1298 continue;
1299 cpumask_andnot(new_cpus, possible_mask, subpartitions_cpus);
1300 } else {
1301 cpumask_and(new_cpus, possible_mask, cs->effective_cpus);
1302 }
1303 set_cpus_allowed_ptr(task, new_cpus);
1304 }
1305 css_task_iter_end(&it);
1306}
1307
1308/**
1309 * compute_effective_cpumask - Compute the effective cpumask of the cpuset
1310 * @new_cpus: the temp variable for the new effective_cpus mask
1311 * @cs: the cpuset the need to recompute the new effective_cpus mask
1312 * @parent: the parent cpuset
1313 *
1314 * The result is valid only if the given cpuset isn't a partition root.
1315 */
1316static void compute_effective_cpumask(struct cpumask *new_cpus,
1317 struct cpuset *cs, struct cpuset *parent)
1318{
1319 cpumask_and(new_cpus, cs->cpus_allowed, parent->effective_cpus);
1320}
1321
1322/*
1323 * Commands for update_parent_effective_cpumask
1324 */
1325enum partition_cmd {
1326 partcmd_enable, /* Enable partition root */
1327 partcmd_enablei, /* Enable isolated partition root */
1328 partcmd_disable, /* Disable partition root */
1329 partcmd_update, /* Update parent's effective_cpus */
1330 partcmd_invalidate, /* Make partition invalid */
1331};
1332
1333static int update_flag(cpuset_flagbits_t bit, struct cpuset *cs,
1334 int turning_on);
1335static void update_sibling_cpumasks(struct cpuset *parent, struct cpuset *cs,
1336 struct tmpmasks *tmp);
1337
1338/*
1339 * Update partition exclusive flag
1340 *
1341 * Return: 0 if successful, an error code otherwise
1342 */
1343static int update_partition_exclusive(struct cpuset *cs, int new_prs)
1344{
1345 bool exclusive = (new_prs > 0);
1346
1347 if (exclusive && !is_cpu_exclusive(cs)) {
1348 if (update_flag(CS_CPU_EXCLUSIVE, cs, 1))
1349 return PERR_NOTEXCL;
1350 } else if (!exclusive && is_cpu_exclusive(cs)) {
1351 /* Turning off CS_CPU_EXCLUSIVE will not return error */
1352 update_flag(CS_CPU_EXCLUSIVE, cs, 0);
1353 }
1354 return 0;
1355}
1356
1357/*
1358 * Update partition load balance flag and/or rebuild sched domain
1359 *
1360 * Changing load balance flag will automatically call
1361 * rebuild_sched_domains_locked().
1362 * This function is for cgroup v2 only.
1363 */
1364static void update_partition_sd_lb(struct cpuset *cs, int old_prs)
1365{
1366 int new_prs = cs->partition_root_state;
1367 bool rebuild_domains = (new_prs > 0) || (old_prs > 0);
1368 bool new_lb;
1369
1370 /*
1371 * If cs is not a valid partition root, the load balance state
1372 * will follow its parent.
1373 */
1374 if (new_prs > 0) {
1375 new_lb = (new_prs != PRS_ISOLATED);
1376 } else {
1377 new_lb = is_sched_load_balance(parent_cs(cs));
1378 }
1379 if (new_lb != !!is_sched_load_balance(cs)) {
1380 rebuild_domains = true;
1381 if (new_lb)
1382 set_bit(CS_SCHED_LOAD_BALANCE, &cs->flags);
1383 else
1384 clear_bit(CS_SCHED_LOAD_BALANCE, &cs->flags);
1385 }
1386
1387 if (rebuild_domains)
1388 rebuild_sched_domains_locked();
1389}
1390
1391/*
1392 * tasks_nocpu_error - Return true if tasks will have no effective_cpus
1393 */
1394static bool tasks_nocpu_error(struct cpuset *parent, struct cpuset *cs,
1395 struct cpumask *xcpus)
1396{
1397 /*
1398 * A populated partition (cs or parent) can't have empty effective_cpus
1399 */
1400 return (cpumask_subset(parent->effective_cpus, xcpus) &&
1401 partition_is_populated(parent, cs)) ||
1402 (!cpumask_intersects(xcpus, cpu_active_mask) &&
1403 partition_is_populated(cs, NULL));
1404}
1405
1406static void reset_partition_data(struct cpuset *cs)
1407{
1408 struct cpuset *parent = parent_cs(cs);
1409
1410 if (!cgroup_subsys_on_dfl(cpuset_cgrp_subsys))
1411 return;
1412
1413 lockdep_assert_held(&callback_lock);
1414
1415 cs->nr_subparts = 0;
1416 if (cpumask_empty(cs->exclusive_cpus)) {
1417 cpumask_clear(cs->effective_xcpus);
1418 if (is_cpu_exclusive(cs))
1419 clear_bit(CS_CPU_EXCLUSIVE, &cs->flags);
1420 }
1421 if (!cpumask_and(cs->effective_cpus,
1422 parent->effective_cpus, cs->cpus_allowed)) {
1423 cs->use_parent_ecpus = true;
1424 parent->child_ecpus_count++;
1425 cpumask_copy(cs->effective_cpus, parent->effective_cpus);
1426 }
1427}
1428
1429/*
1430 * partition_xcpus_newstate - Exclusive CPUs state change
1431 * @old_prs: old partition_root_state
1432 * @new_prs: new partition_root_state
1433 * @xcpus: exclusive CPUs with state change
1434 */
1435static void partition_xcpus_newstate(int old_prs, int new_prs, struct cpumask *xcpus)
1436{
1437 WARN_ON_ONCE(old_prs == new_prs);
1438 if (new_prs == PRS_ISOLATED)
1439 cpumask_or(isolated_cpus, isolated_cpus, xcpus);
1440 else
1441 cpumask_andnot(isolated_cpus, isolated_cpus, xcpus);
1442}
1443
1444/*
1445 * partition_xcpus_add - Add new exclusive CPUs to partition
1446 * @new_prs: new partition_root_state
1447 * @parent: parent cpuset
1448 * @xcpus: exclusive CPUs to be added
1449 * Return: true if isolated_cpus modified, false otherwise
1450 *
1451 * Remote partition if parent == NULL
1452 */
1453static bool partition_xcpus_add(int new_prs, struct cpuset *parent,
1454 struct cpumask *xcpus)
1455{
1456 bool isolcpus_updated;
1457
1458 WARN_ON_ONCE(new_prs < 0);
1459 lockdep_assert_held(&callback_lock);
1460 if (!parent)
1461 parent = &top_cpuset;
1462
1463
1464 if (parent == &top_cpuset)
1465 cpumask_or(subpartitions_cpus, subpartitions_cpus, xcpus);
1466
1467 isolcpus_updated = (new_prs != parent->partition_root_state);
1468 if (isolcpus_updated)
1469 partition_xcpus_newstate(parent->partition_root_state, new_prs,
1470 xcpus);
1471
1472 cpumask_andnot(parent->effective_cpus, parent->effective_cpus, xcpus);
1473 return isolcpus_updated;
1474}
1475
1476/*
1477 * partition_xcpus_del - Remove exclusive CPUs from partition
1478 * @old_prs: old partition_root_state
1479 * @parent: parent cpuset
1480 * @xcpus: exclusive CPUs to be removed
1481 * Return: true if isolated_cpus modified, false otherwise
1482 *
1483 * Remote partition if parent == NULL
1484 */
1485static bool partition_xcpus_del(int old_prs, struct cpuset *parent,
1486 struct cpumask *xcpus)
1487{
1488 bool isolcpus_updated;
1489
1490 WARN_ON_ONCE(old_prs < 0);
1491 lockdep_assert_held(&callback_lock);
1492 if (!parent)
1493 parent = &top_cpuset;
1494
1495 if (parent == &top_cpuset)
1496 cpumask_andnot(subpartitions_cpus, subpartitions_cpus, xcpus);
1497
1498 isolcpus_updated = (old_prs != parent->partition_root_state);
1499 if (isolcpus_updated)
1500 partition_xcpus_newstate(old_prs, parent->partition_root_state,
1501 xcpus);
1502
1503 cpumask_and(xcpus, xcpus, cpu_active_mask);
1504 cpumask_or(parent->effective_cpus, parent->effective_cpus, xcpus);
1505 return isolcpus_updated;
1506}
1507
1508static void update_unbound_workqueue_cpumask(bool isolcpus_updated)
1509{
1510 int ret;
1511
1512 lockdep_assert_cpus_held();
1513
1514 if (!isolcpus_updated)
1515 return;
1516
1517 ret = workqueue_unbound_exclude_cpumask(isolated_cpus);
1518 WARN_ON_ONCE(ret < 0);
1519}
1520
1521/**
1522 * cpuset_cpu_is_isolated - Check if the given CPU is isolated
1523 * @cpu: the CPU number to be checked
1524 * Return: true if CPU is used in an isolated partition, false otherwise
1525 */
1526bool cpuset_cpu_is_isolated(int cpu)
1527{
1528 return cpumask_test_cpu(cpu, isolated_cpus);
1529}
1530EXPORT_SYMBOL_GPL(cpuset_cpu_is_isolated);
1531
1532/*
1533 * compute_effective_exclusive_cpumask - compute effective exclusive CPUs
1534 * @cs: cpuset
1535 * @xcpus: effective exclusive CPUs value to be set
1536 * Return: true if xcpus is not empty, false otherwise.
1537 *
1538 * Starting with exclusive_cpus (cpus_allowed if exclusive_cpus is not set),
1539 * it must be a subset of cpus_allowed and parent's effective_xcpus.
1540 */
1541static bool compute_effective_exclusive_cpumask(struct cpuset *cs,
1542 struct cpumask *xcpus)
1543{
1544 struct cpuset *parent = parent_cs(cs);
1545
1546 if (!xcpus)
1547 xcpus = cs->effective_xcpus;
1548
1549 if (!cpumask_empty(cs->exclusive_cpus))
1550 cpumask_and(xcpus, cs->exclusive_cpus, cs->cpus_allowed);
1551 else
1552 cpumask_copy(xcpus, cs->cpus_allowed);
1553
1554 return cpumask_and(xcpus, xcpus, parent->effective_xcpus);
1555}
1556
1557static inline bool is_remote_partition(struct cpuset *cs)
1558{
1559 return !list_empty(&cs->remote_sibling);
1560}
1561
1562static inline bool is_local_partition(struct cpuset *cs)
1563{
1564 return is_partition_valid(cs) && !is_remote_partition(cs);
1565}
1566
1567/*
1568 * remote_partition_enable - Enable current cpuset as a remote partition root
1569 * @cs: the cpuset to update
1570 * @new_prs: new partition_root_state
1571 * @tmp: temparary masks
1572 * Return: 1 if successful, 0 if error
1573 *
1574 * Enable the current cpuset to become a remote partition root taking CPUs
1575 * directly from the top cpuset. cpuset_mutex must be held by the caller.
1576 */
1577static int remote_partition_enable(struct cpuset *cs, int new_prs,
1578 struct tmpmasks *tmp)
1579{
1580 bool isolcpus_updated;
1581
1582 /*
1583 * The user must have sysadmin privilege.
1584 */
1585 if (!capable(CAP_SYS_ADMIN))
1586 return 0;
1587
1588 /*
1589 * The requested exclusive_cpus must not be allocated to other
1590 * partitions and it can't use up all the root's effective_cpus.
1591 *
1592 * Note that if there is any local partition root above it or
1593 * remote partition root underneath it, its exclusive_cpus must
1594 * have overlapped with subpartitions_cpus.
1595 */
1596 compute_effective_exclusive_cpumask(cs, tmp->new_cpus);
1597 if (cpumask_empty(tmp->new_cpus) ||
1598 cpumask_intersects(tmp->new_cpus, subpartitions_cpus) ||
1599 cpumask_subset(top_cpuset.effective_cpus, tmp->new_cpus))
1600 return 0;
1601
1602 spin_lock_irq(&callback_lock);
1603 isolcpus_updated = partition_xcpus_add(new_prs, NULL, tmp->new_cpus);
1604 list_add(&cs->remote_sibling, &remote_children);
1605 if (cs->use_parent_ecpus) {
1606 struct cpuset *parent = parent_cs(cs);
1607
1608 cs->use_parent_ecpus = false;
1609 parent->child_ecpus_count--;
1610 }
1611 spin_unlock_irq(&callback_lock);
1612 update_unbound_workqueue_cpumask(isolcpus_updated);
1613
1614 /*
1615 * Proprogate changes in top_cpuset's effective_cpus down the hierarchy.
1616 */
1617 update_tasks_cpumask(&top_cpuset, tmp->new_cpus);
1618 update_sibling_cpumasks(&top_cpuset, NULL, tmp);
1619 return 1;
1620}
1621
1622/*
1623 * remote_partition_disable - Remove current cpuset from remote partition list
1624 * @cs: the cpuset to update
1625 * @tmp: temparary masks
1626 *
1627 * The effective_cpus is also updated.
1628 *
1629 * cpuset_mutex must be held by the caller.
1630 */
1631static void remote_partition_disable(struct cpuset *cs, struct tmpmasks *tmp)
1632{
1633 bool isolcpus_updated;
1634
1635 compute_effective_exclusive_cpumask(cs, tmp->new_cpus);
1636 WARN_ON_ONCE(!is_remote_partition(cs));
1637 WARN_ON_ONCE(!cpumask_subset(tmp->new_cpus, subpartitions_cpus));
1638
1639 spin_lock_irq(&callback_lock);
1640 list_del_init(&cs->remote_sibling);
1641 isolcpus_updated = partition_xcpus_del(cs->partition_root_state,
1642 NULL, tmp->new_cpus);
1643 cs->partition_root_state = -cs->partition_root_state;
1644 if (!cs->prs_err)
1645 cs->prs_err = PERR_INVCPUS;
1646 reset_partition_data(cs);
1647 spin_unlock_irq(&callback_lock);
1648 update_unbound_workqueue_cpumask(isolcpus_updated);
1649
1650 /*
1651 * Proprogate changes in top_cpuset's effective_cpus down the hierarchy.
1652 */
1653 update_tasks_cpumask(&top_cpuset, tmp->new_cpus);
1654 update_sibling_cpumasks(&top_cpuset, NULL, tmp);
1655}
1656
1657/*
1658 * remote_cpus_update - cpus_exclusive change of remote partition
1659 * @cs: the cpuset to be updated
1660 * @newmask: the new effective_xcpus mask
1661 * @tmp: temparary masks
1662 *
1663 * top_cpuset and subpartitions_cpus will be updated or partition can be
1664 * invalidated.
1665 */
1666static void remote_cpus_update(struct cpuset *cs, struct cpumask *newmask,
1667 struct tmpmasks *tmp)
1668{
1669 bool adding, deleting;
1670 int prs = cs->partition_root_state;
1671 int isolcpus_updated = 0;
1672
1673 if (WARN_ON_ONCE(!is_remote_partition(cs)))
1674 return;
1675
1676 WARN_ON_ONCE(!cpumask_subset(cs->effective_xcpus, subpartitions_cpus));
1677
1678 if (cpumask_empty(newmask))
1679 goto invalidate;
1680
1681 adding = cpumask_andnot(tmp->addmask, newmask, cs->effective_xcpus);
1682 deleting = cpumask_andnot(tmp->delmask, cs->effective_xcpus, newmask);
1683
1684 /*
1685 * Additions of remote CPUs is only allowed if those CPUs are
1686 * not allocated to other partitions and there are effective_cpus
1687 * left in the top cpuset.
1688 */
1689 if (adding && (!capable(CAP_SYS_ADMIN) ||
1690 cpumask_intersects(tmp->addmask, subpartitions_cpus) ||
1691 cpumask_subset(top_cpuset.effective_cpus, tmp->addmask)))
1692 goto invalidate;
1693
1694 spin_lock_irq(&callback_lock);
1695 if (adding)
1696 isolcpus_updated += partition_xcpus_add(prs, NULL, tmp->addmask);
1697 if (deleting)
1698 isolcpus_updated += partition_xcpus_del(prs, NULL, tmp->delmask);
1699 spin_unlock_irq(&callback_lock);
1700 update_unbound_workqueue_cpumask(isolcpus_updated);
1701
1702 /*
1703 * Proprogate changes in top_cpuset's effective_cpus down the hierarchy.
1704 */
1705 update_tasks_cpumask(&top_cpuset, tmp->new_cpus);
1706 update_sibling_cpumasks(&top_cpuset, NULL, tmp);
1707 return;
1708
1709invalidate:
1710 remote_partition_disable(cs, tmp);
1711}
1712
1713/*
1714 * remote_partition_check - check if a child remote partition needs update
1715 * @cs: the cpuset to be updated
1716 * @newmask: the new effective_xcpus mask
1717 * @delmask: temporary mask for deletion (not in tmp)
1718 * @tmp: temparary masks
1719 *
1720 * This should be called before the given cs has updated its cpus_allowed
1721 * and/or effective_xcpus.
1722 */
1723static void remote_partition_check(struct cpuset *cs, struct cpumask *newmask,
1724 struct cpumask *delmask, struct tmpmasks *tmp)
1725{
1726 struct cpuset *child, *next;
1727 int disable_cnt = 0;
1728
1729 /*
1730 * Compute the effective exclusive CPUs that will be deleted.
1731 */
1732 if (!cpumask_andnot(delmask, cs->effective_xcpus, newmask) ||
1733 !cpumask_intersects(delmask, subpartitions_cpus))
1734 return; /* No deletion of exclusive CPUs in partitions */
1735
1736 /*
1737 * Searching the remote children list to look for those that will
1738 * be impacted by the deletion of exclusive CPUs.
1739 *
1740 * Since a cpuset must be removed from the remote children list
1741 * before it can go offline and holding cpuset_mutex will prevent
1742 * any change in cpuset status. RCU read lock isn't needed.
1743 */
1744 lockdep_assert_held(&cpuset_mutex);
1745 list_for_each_entry_safe(child, next, &remote_children, remote_sibling)
1746 if (cpumask_intersects(child->effective_cpus, delmask)) {
1747 remote_partition_disable(child, tmp);
1748 disable_cnt++;
1749 }
1750 if (disable_cnt)
1751 rebuild_sched_domains_locked();
1752}
1753
1754/*
1755 * prstate_housekeeping_conflict - check for partition & housekeeping conflicts
1756 * @prstate: partition root state to be checked
1757 * @new_cpus: cpu mask
1758 * Return: true if there is conflict, false otherwise
1759 *
1760 * CPUs outside of housekeeping_cpumask(HK_TYPE_DOMAIN) can only be used in
1761 * an isolated partition.
1762 */
1763static bool prstate_housekeeping_conflict(int prstate, struct cpumask *new_cpus)
1764{
1765 const struct cpumask *hk_domain = housekeeping_cpumask(HK_TYPE_DOMAIN);
1766 bool all_in_hk = cpumask_subset(new_cpus, hk_domain);
1767
1768 if (!all_in_hk && (prstate != PRS_ISOLATED))
1769 return true;
1770
1771 return false;
1772}
1773
1774/**
1775 * update_parent_effective_cpumask - update effective_cpus mask of parent cpuset
1776 * @cs: The cpuset that requests change in partition root state
1777 * @cmd: Partition root state change command
1778 * @newmask: Optional new cpumask for partcmd_update
1779 * @tmp: Temporary addmask and delmask
1780 * Return: 0 or a partition root state error code
1781 *
1782 * For partcmd_enable*, the cpuset is being transformed from a non-partition
1783 * root to a partition root. The effective_xcpus (cpus_allowed if
1784 * effective_xcpus not set) mask of the given cpuset will be taken away from
1785 * parent's effective_cpus. The function will return 0 if all the CPUs listed
1786 * in effective_xcpus can be granted or an error code will be returned.
1787 *
1788 * For partcmd_disable, the cpuset is being transformed from a partition
1789 * root back to a non-partition root. Any CPUs in effective_xcpus will be
1790 * given back to parent's effective_cpus. 0 will always be returned.
1791 *
1792 * For partcmd_update, if the optional newmask is specified, the cpu list is
1793 * to be changed from effective_xcpus to newmask. Otherwise, effective_xcpus is
1794 * assumed to remain the same. The cpuset should either be a valid or invalid
1795 * partition root. The partition root state may change from valid to invalid
1796 * or vice versa. An error code will be returned if transitioning from
1797 * invalid to valid violates the exclusivity rule.
1798 *
1799 * For partcmd_invalidate, the current partition will be made invalid.
1800 *
1801 * The partcmd_enable* and partcmd_disable commands are used by
1802 * update_prstate(). An error code may be returned and the caller will check
1803 * for error.
1804 *
1805 * The partcmd_update command is used by update_cpumasks_hier() with newmask
1806 * NULL and update_cpumask() with newmask set. The partcmd_invalidate is used
1807 * by update_cpumask() with NULL newmask. In both cases, the callers won't
1808 * check for error and so partition_root_state and prs_error will be updated
1809 * directly.
1810 */
1811static int update_parent_effective_cpumask(struct cpuset *cs, int cmd,
1812 struct cpumask *newmask,
1813 struct tmpmasks *tmp)
1814{
1815 struct cpuset *parent = parent_cs(cs);
1816 int adding; /* Adding cpus to parent's effective_cpus */
1817 int deleting; /* Deleting cpus from parent's effective_cpus */
1818 int old_prs, new_prs;
1819 int part_error = PERR_NONE; /* Partition error? */
1820 int subparts_delta = 0;
1821 struct cpumask *xcpus; /* cs effective_xcpus */
1822 int isolcpus_updated = 0;
1823 bool nocpu;
1824
1825 lockdep_assert_held(&cpuset_mutex);
1826
1827 /*
1828 * new_prs will only be changed for the partcmd_update and
1829 * partcmd_invalidate commands.
1830 */
1831 adding = deleting = false;
1832 old_prs = new_prs = cs->partition_root_state;
1833 xcpus = !cpumask_empty(cs->exclusive_cpus)
1834 ? cs->effective_xcpus : cs->cpus_allowed;
1835
1836 if (cmd == partcmd_invalidate) {
1837 if (is_prs_invalid(old_prs))
1838 return 0;
1839
1840 /*
1841 * Make the current partition invalid.
1842 */
1843 if (is_partition_valid(parent))
1844 adding = cpumask_and(tmp->addmask,
1845 xcpus, parent->effective_xcpus);
1846 if (old_prs > 0) {
1847 new_prs = -old_prs;
1848 subparts_delta--;
1849 }
1850 goto write_error;
1851 }
1852
1853 /*
1854 * The parent must be a partition root.
1855 * The new cpumask, if present, or the current cpus_allowed must
1856 * not be empty.
1857 */
1858 if (!is_partition_valid(parent)) {
1859 return is_partition_invalid(parent)
1860 ? PERR_INVPARENT : PERR_NOTPART;
1861 }
1862 if (!newmask && cpumask_empty(cs->cpus_allowed))
1863 return PERR_CPUSEMPTY;
1864
1865 nocpu = tasks_nocpu_error(parent, cs, xcpus);
1866
1867 if ((cmd == partcmd_enable) || (cmd == partcmd_enablei)) {
1868 /*
1869 * Enabling partition root is not allowed if its
1870 * effective_xcpus is empty or doesn't overlap with
1871 * parent's effective_xcpus.
1872 */
1873 if (cpumask_empty(xcpus) ||
1874 !cpumask_intersects(xcpus, parent->effective_xcpus))
1875 return PERR_INVCPUS;
1876
1877 if (prstate_housekeeping_conflict(new_prs, xcpus))
1878 return PERR_HKEEPING;
1879
1880 /*
1881 * A parent can be left with no CPU as long as there is no
1882 * task directly associated with the parent partition.
1883 */
1884 if (nocpu)
1885 return PERR_NOCPUS;
1886
1887 cpumask_copy(tmp->delmask, xcpus);
1888 deleting = true;
1889 subparts_delta++;
1890 new_prs = (cmd == partcmd_enable) ? PRS_ROOT : PRS_ISOLATED;
1891 } else if (cmd == partcmd_disable) {
1892 /*
1893 * May need to add cpus to parent's effective_cpus for
1894 * valid partition root.
1895 */
1896 adding = !is_prs_invalid(old_prs) &&
1897 cpumask_and(tmp->addmask, xcpus, parent->effective_xcpus);
1898 if (adding)
1899 subparts_delta--;
1900 new_prs = PRS_MEMBER;
1901 } else if (newmask) {
1902 /*
1903 * Empty cpumask is not allowed
1904 */
1905 if (cpumask_empty(newmask)) {
1906 part_error = PERR_CPUSEMPTY;
1907 goto write_error;
1908 }
1909
1910 /*
1911 * partcmd_update with newmask:
1912 *
1913 * Compute add/delete mask to/from effective_cpus
1914 *
1915 * For valid partition:
1916 * addmask = exclusive_cpus & ~newmask
1917 * & parent->effective_xcpus
1918 * delmask = newmask & ~exclusive_cpus
1919 * & parent->effective_xcpus
1920 *
1921 * For invalid partition:
1922 * delmask = newmask & parent->effective_xcpus
1923 */
1924 if (is_prs_invalid(old_prs)) {
1925 adding = false;
1926 deleting = cpumask_and(tmp->delmask,
1927 newmask, parent->effective_xcpus);
1928 } else {
1929 cpumask_andnot(tmp->addmask, xcpus, newmask);
1930 adding = cpumask_and(tmp->addmask, tmp->addmask,
1931 parent->effective_xcpus);
1932
1933 cpumask_andnot(tmp->delmask, newmask, xcpus);
1934 deleting = cpumask_and(tmp->delmask, tmp->delmask,
1935 parent->effective_xcpus);
1936 }
1937 /*
1938 * Make partition invalid if parent's effective_cpus could
1939 * become empty and there are tasks in the parent.
1940 */
1941 if (nocpu && (!adding ||
1942 !cpumask_intersects(tmp->addmask, cpu_active_mask))) {
1943 part_error = PERR_NOCPUS;
1944 deleting = false;
1945 adding = cpumask_and(tmp->addmask,
1946 xcpus, parent->effective_xcpus);
1947 }
1948 } else {
1949 /*
1950 * partcmd_update w/o newmask
1951 *
1952 * delmask = effective_xcpus & parent->effective_cpus
1953 *
1954 * This can be called from:
1955 * 1) update_cpumasks_hier()
1956 * 2) cpuset_hotplug_update_tasks()
1957 *
1958 * Check to see if it can be transitioned from valid to
1959 * invalid partition or vice versa.
1960 *
1961 * A partition error happens when parent has tasks and all
1962 * its effective CPUs will have to be distributed out.
1963 */
1964 WARN_ON_ONCE(!is_partition_valid(parent));
1965 if (nocpu) {
1966 part_error = PERR_NOCPUS;
1967 if (is_partition_valid(cs))
1968 adding = cpumask_and(tmp->addmask,
1969 xcpus, parent->effective_xcpus);
1970 } else if (is_partition_invalid(cs) &&
1971 cpumask_subset(xcpus, parent->effective_xcpus)) {
1972 struct cgroup_subsys_state *css;
1973 struct cpuset *child;
1974 bool exclusive = true;
1975
1976 /*
1977 * Convert invalid partition to valid has to
1978 * pass the cpu exclusivity test.
1979 */
1980 rcu_read_lock();
1981 cpuset_for_each_child(child, css, parent) {
1982 if (child == cs)
1983 continue;
1984 if (!cpusets_are_exclusive(cs, child)) {
1985 exclusive = false;
1986 break;
1987 }
1988 }
1989 rcu_read_unlock();
1990 if (exclusive)
1991 deleting = cpumask_and(tmp->delmask,
1992 xcpus, parent->effective_cpus);
1993 else
1994 part_error = PERR_NOTEXCL;
1995 }
1996 }
1997
1998write_error:
1999 if (part_error)
2000 WRITE_ONCE(cs->prs_err, part_error);
2001
2002 if (cmd == partcmd_update) {
2003 /*
2004 * Check for possible transition between valid and invalid
2005 * partition root.
2006 */
2007 switch (cs->partition_root_state) {
2008 case PRS_ROOT:
2009 case PRS_ISOLATED:
2010 if (part_error) {
2011 new_prs = -old_prs;
2012 subparts_delta--;
2013 }
2014 break;
2015 case PRS_INVALID_ROOT:
2016 case PRS_INVALID_ISOLATED:
2017 if (!part_error) {
2018 new_prs = -old_prs;
2019 subparts_delta++;
2020 }
2021 break;
2022 }
2023 }
2024
2025 if (!adding && !deleting && (new_prs == old_prs))
2026 return 0;
2027
2028 /*
2029 * Transitioning between invalid to valid or vice versa may require
2030 * changing CS_CPU_EXCLUSIVE. In the case of partcmd_update,
2031 * validate_change() has already been successfully called and
2032 * CPU lists in cs haven't been updated yet. So defer it to later.
2033 */
2034 if ((old_prs != new_prs) && (cmd != partcmd_update)) {
2035 int err = update_partition_exclusive(cs, new_prs);
2036
2037 if (err)
2038 return err;
2039 }
2040
2041 /*
2042 * Change the parent's effective_cpus & effective_xcpus (top cpuset
2043 * only).
2044 *
2045 * Newly added CPUs will be removed from effective_cpus and
2046 * newly deleted ones will be added back to effective_cpus.
2047 */
2048 spin_lock_irq(&callback_lock);
2049 if (old_prs != new_prs) {
2050 cs->partition_root_state = new_prs;
2051 if (new_prs <= 0)
2052 cs->nr_subparts = 0;
2053 }
2054 /*
2055 * Adding to parent's effective_cpus means deletion CPUs from cs
2056 * and vice versa.
2057 */
2058 if (adding)
2059 isolcpus_updated += partition_xcpus_del(old_prs, parent,
2060 tmp->addmask);
2061 if (deleting)
2062 isolcpus_updated += partition_xcpus_add(new_prs, parent,
2063 tmp->delmask);
2064
2065 if (is_partition_valid(parent)) {
2066 parent->nr_subparts += subparts_delta;
2067 WARN_ON_ONCE(parent->nr_subparts < 0);
2068 }
2069 spin_unlock_irq(&callback_lock);
2070 update_unbound_workqueue_cpumask(isolcpus_updated);
2071
2072 if ((old_prs != new_prs) && (cmd == partcmd_update))
2073 update_partition_exclusive(cs, new_prs);
2074
2075 if (adding || deleting) {
2076 update_tasks_cpumask(parent, tmp->addmask);
2077 update_sibling_cpumasks(parent, cs, tmp);
2078 }
2079
2080 /*
2081 * For partcmd_update without newmask, it is being called from
2082 * cpuset_hotplug_workfn() where cpus_read_lock() wasn't taken.
2083 * Update the load balance flag and scheduling domain if
2084 * cpus_read_trylock() is successful.
2085 */
2086 if ((cmd == partcmd_update) && !newmask && cpus_read_trylock()) {
2087 update_partition_sd_lb(cs, old_prs);
2088 cpus_read_unlock();
2089 }
2090
2091 notify_partition_change(cs, old_prs);
2092 return 0;
2093}
2094
2095/**
2096 * compute_partition_effective_cpumask - compute effective_cpus for partition
2097 * @cs: partition root cpuset
2098 * @new_ecpus: previously computed effective_cpus to be updated
2099 *
2100 * Compute the effective_cpus of a partition root by scanning effective_xcpus
2101 * of child partition roots and excluding their effective_xcpus.
2102 *
2103 * This has the side effect of invalidating valid child partition roots,
2104 * if necessary. Since it is called from either cpuset_hotplug_update_tasks()
2105 * or update_cpumasks_hier() where parent and children are modified
2106 * successively, we don't need to call update_parent_effective_cpumask()
2107 * and the child's effective_cpus will be updated in later iterations.
2108 *
2109 * Note that rcu_read_lock() is assumed to be held.
2110 */
2111static void compute_partition_effective_cpumask(struct cpuset *cs,
2112 struct cpumask *new_ecpus)
2113{
2114 struct cgroup_subsys_state *css;
2115 struct cpuset *child;
2116 bool populated = partition_is_populated(cs, NULL);
2117
2118 /*
2119 * Check child partition roots to see if they should be
2120 * invalidated when
2121 * 1) child effective_xcpus not a subset of new
2122 * excluisve_cpus
2123 * 2) All the effective_cpus will be used up and cp
2124 * has tasks
2125 */
2126 compute_effective_exclusive_cpumask(cs, new_ecpus);
2127 cpumask_and(new_ecpus, new_ecpus, cpu_active_mask);
2128
2129 rcu_read_lock();
2130 cpuset_for_each_child(child, css, cs) {
2131 if (!is_partition_valid(child))
2132 continue;
2133
2134 child->prs_err = 0;
2135 if (!cpumask_subset(child->effective_xcpus,
2136 cs->effective_xcpus))
2137 child->prs_err = PERR_INVCPUS;
2138 else if (populated &&
2139 cpumask_subset(new_ecpus, child->effective_xcpus))
2140 child->prs_err = PERR_NOCPUS;
2141
2142 if (child->prs_err) {
2143 int old_prs = child->partition_root_state;
2144
2145 /*
2146 * Invalidate child partition
2147 */
2148 spin_lock_irq(&callback_lock);
2149 make_partition_invalid(child);
2150 cs->nr_subparts--;
2151 child->nr_subparts = 0;
2152 spin_unlock_irq(&callback_lock);
2153 notify_partition_change(child, old_prs);
2154 continue;
2155 }
2156 cpumask_andnot(new_ecpus, new_ecpus,
2157 child->effective_xcpus);
2158 }
2159 rcu_read_unlock();
2160}
2161
2162/*
2163 * update_cpumasks_hier() flags
2164 */
2165#define HIER_CHECKALL 0x01 /* Check all cpusets with no skipping */
2166#define HIER_NO_SD_REBUILD 0x02 /* Don't rebuild sched domains */
2167
2168/*
2169 * update_cpumasks_hier - Update effective cpumasks and tasks in the subtree
2170 * @cs: the cpuset to consider
2171 * @tmp: temp variables for calculating effective_cpus & partition setup
2172 * @force: don't skip any descendant cpusets if set
2173 *
2174 * When configured cpumask is changed, the effective cpumasks of this cpuset
2175 * and all its descendants need to be updated.
2176 *
2177 * On legacy hierarchy, effective_cpus will be the same with cpu_allowed.
2178 *
2179 * Called with cpuset_mutex held
2180 */
2181static void update_cpumasks_hier(struct cpuset *cs, struct tmpmasks *tmp,
2182 int flags)
2183{
2184 struct cpuset *cp;
2185 struct cgroup_subsys_state *pos_css;
2186 bool need_rebuild_sched_domains = false;
2187 int old_prs, new_prs;
2188
2189 rcu_read_lock();
2190 cpuset_for_each_descendant_pre(cp, pos_css, cs) {
2191 struct cpuset *parent = parent_cs(cp);
2192 bool remote = is_remote_partition(cp);
2193 bool update_parent = false;
2194
2195 /*
2196 * Skip descendent remote partition that acquires CPUs
2197 * directly from top cpuset unless it is cs.
2198 */
2199 if (remote && (cp != cs)) {
2200 pos_css = css_rightmost_descendant(pos_css);
2201 continue;
2202 }
2203
2204 /*
2205 * Update effective_xcpus if exclusive_cpus set.
2206 * The case when exclusive_cpus isn't set is handled later.
2207 */
2208 if (!cpumask_empty(cp->exclusive_cpus) && (cp != cs)) {
2209 spin_lock_irq(&callback_lock);
2210 compute_effective_exclusive_cpumask(cp, NULL);
2211 spin_unlock_irq(&callback_lock);
2212 }
2213
2214 old_prs = new_prs = cp->partition_root_state;
2215 if (remote || (is_partition_valid(parent) &&
2216 is_partition_valid(cp)))
2217 compute_partition_effective_cpumask(cp, tmp->new_cpus);
2218 else
2219 compute_effective_cpumask(tmp->new_cpus, cp, parent);
2220
2221 /*
2222 * A partition with no effective_cpus is allowed as long as
2223 * there is no task associated with it. Call
2224 * update_parent_effective_cpumask() to check it.
2225 */
2226 if (is_partition_valid(cp) && cpumask_empty(tmp->new_cpus)) {
2227 update_parent = true;
2228 goto update_parent_effective;
2229 }
2230
2231 /*
2232 * If it becomes empty, inherit the effective mask of the
2233 * parent, which is guaranteed to have some CPUs unless
2234 * it is a partition root that has explicitly distributed
2235 * out all its CPUs.
2236 */
2237 if (is_in_v2_mode() && !remote && cpumask_empty(tmp->new_cpus)) {
2238 cpumask_copy(tmp->new_cpus, parent->effective_cpus);
2239 if (!cp->use_parent_ecpus) {
2240 cp->use_parent_ecpus = true;
2241 parent->child_ecpus_count++;
2242 }
2243 } else if (cp->use_parent_ecpus) {
2244 cp->use_parent_ecpus = false;
2245 WARN_ON_ONCE(!parent->child_ecpus_count);
2246 parent->child_ecpus_count--;
2247 }
2248
2249 if (remote)
2250 goto get_css;
2251
2252 /*
2253 * Skip the whole subtree if
2254 * 1) the cpumask remains the same,
2255 * 2) has no partition root state,
2256 * 3) HIER_CHECKALL flag not set, and
2257 * 4) for v2 load balance state same as its parent.
2258 */
2259 if (!cp->partition_root_state && !(flags & HIER_CHECKALL) &&
2260 cpumask_equal(tmp->new_cpus, cp->effective_cpus) &&
2261 (!cgroup_subsys_on_dfl(cpuset_cgrp_subsys) ||
2262 (is_sched_load_balance(parent) == is_sched_load_balance(cp)))) {
2263 pos_css = css_rightmost_descendant(pos_css);
2264 continue;
2265 }
2266
2267update_parent_effective:
2268 /*
2269 * update_parent_effective_cpumask() should have been called
2270 * for cs already in update_cpumask(). We should also call
2271 * update_tasks_cpumask() again for tasks in the parent
2272 * cpuset if the parent's effective_cpus changes.
2273 */
2274 if ((cp != cs) && old_prs) {
2275 switch (parent->partition_root_state) {
2276 case PRS_ROOT:
2277 case PRS_ISOLATED:
2278 update_parent = true;
2279 break;
2280
2281 default:
2282 /*
2283 * When parent is not a partition root or is
2284 * invalid, child partition roots become
2285 * invalid too.
2286 */
2287 if (is_partition_valid(cp))
2288 new_prs = -cp->partition_root_state;
2289 WRITE_ONCE(cp->prs_err,
2290 is_partition_invalid(parent)
2291 ? PERR_INVPARENT : PERR_NOTPART);
2292 break;
2293 }
2294 }
2295get_css:
2296 if (!css_tryget_online(&cp->css))
2297 continue;
2298 rcu_read_unlock();
2299
2300 if (update_parent) {
2301 update_parent_effective_cpumask(cp, partcmd_update, NULL, tmp);
2302 /*
2303 * The cpuset partition_root_state may become
2304 * invalid. Capture it.
2305 */
2306 new_prs = cp->partition_root_state;
2307 }
2308
2309 spin_lock_irq(&callback_lock);
2310 cpumask_copy(cp->effective_cpus, tmp->new_cpus);
2311 cp->partition_root_state = new_prs;
2312 /*
2313 * Make sure effective_xcpus is properly set for a valid
2314 * partition root.
2315 */
2316 if ((new_prs > 0) && cpumask_empty(cp->exclusive_cpus))
2317 cpumask_and(cp->effective_xcpus,
2318 cp->cpus_allowed, parent->effective_xcpus);
2319 else if (new_prs < 0)
2320 reset_partition_data(cp);
2321 spin_unlock_irq(&callback_lock);
2322
2323 notify_partition_change(cp, old_prs);
2324
2325 WARN_ON(!is_in_v2_mode() &&
2326 !cpumask_equal(cp->cpus_allowed, cp->effective_cpus));
2327
2328 update_tasks_cpumask(cp, cp->effective_cpus);
2329
2330 /*
2331 * On default hierarchy, inherit the CS_SCHED_LOAD_BALANCE
2332 * from parent if current cpuset isn't a valid partition root
2333 * and their load balance states differ.
2334 */
2335 if (cgroup_subsys_on_dfl(cpuset_cgrp_subsys) &&
2336 !is_partition_valid(cp) &&
2337 (is_sched_load_balance(parent) != is_sched_load_balance(cp))) {
2338 if (is_sched_load_balance(parent))
2339 set_bit(CS_SCHED_LOAD_BALANCE, &cp->flags);
2340 else
2341 clear_bit(CS_SCHED_LOAD_BALANCE, &cp->flags);
2342 }
2343
2344 /*
2345 * On legacy hierarchy, if the effective cpumask of any non-
2346 * empty cpuset is changed, we need to rebuild sched domains.
2347 * On default hierarchy, the cpuset needs to be a partition
2348 * root as well.
2349 */
2350 if (!cpumask_empty(cp->cpus_allowed) &&
2351 is_sched_load_balance(cp) &&
2352 (!cgroup_subsys_on_dfl(cpuset_cgrp_subsys) ||
2353 is_partition_valid(cp)))
2354 need_rebuild_sched_domains = true;
2355
2356 rcu_read_lock();
2357 css_put(&cp->css);
2358 }
2359 rcu_read_unlock();
2360
2361 if (need_rebuild_sched_domains && !(flags & HIER_NO_SD_REBUILD))
2362 rebuild_sched_domains_locked();
2363}
2364
2365/**
2366 * update_sibling_cpumasks - Update siblings cpumasks
2367 * @parent: Parent cpuset
2368 * @cs: Current cpuset
2369 * @tmp: Temp variables
2370 */
2371static void update_sibling_cpumasks(struct cpuset *parent, struct cpuset *cs,
2372 struct tmpmasks *tmp)
2373{
2374 struct cpuset *sibling;
2375 struct cgroup_subsys_state *pos_css;
2376
2377 lockdep_assert_held(&cpuset_mutex);
2378
2379 /*
2380 * Check all its siblings and call update_cpumasks_hier()
2381 * if their effective_cpus will need to be changed.
2382 *
2383 * With the addition of effective_xcpus which is a subset of
2384 * cpus_allowed. It is possible a change in parent's effective_cpus
2385 * due to a change in a child partition's effective_xcpus will impact
2386 * its siblings even if they do not inherit parent's effective_cpus
2387 * directly.
2388 *
2389 * The update_cpumasks_hier() function may sleep. So we have to
2390 * release the RCU read lock before calling it. HIER_NO_SD_REBUILD
2391 * flag is used to suppress rebuild of sched domains as the callers
2392 * will take care of that.
2393 */
2394 rcu_read_lock();
2395 cpuset_for_each_child(sibling, pos_css, parent) {
2396 if (sibling == cs)
2397 continue;
2398 if (!sibling->use_parent_ecpus &&
2399 !is_partition_valid(sibling)) {
2400 compute_effective_cpumask(tmp->new_cpus, sibling,
2401 parent);
2402 if (cpumask_equal(tmp->new_cpus, sibling->effective_cpus))
2403 continue;
2404 }
2405 if (!css_tryget_online(&sibling->css))
2406 continue;
2407
2408 rcu_read_unlock();
2409 update_cpumasks_hier(sibling, tmp, HIER_NO_SD_REBUILD);
2410 rcu_read_lock();
2411 css_put(&sibling->css);
2412 }
2413 rcu_read_unlock();
2414}
2415
2416/**
2417 * update_cpumask - update the cpus_allowed mask of a cpuset and all tasks in it
2418 * @cs: the cpuset to consider
2419 * @trialcs: trial cpuset
2420 * @buf: buffer of cpu numbers written to this cpuset
2421 */
2422static int update_cpumask(struct cpuset *cs, struct cpuset *trialcs,
2423 const char *buf)
2424{
2425 int retval;
2426 struct tmpmasks tmp;
2427 struct cpuset *parent = parent_cs(cs);
2428 bool invalidate = false;
2429 int hier_flags = 0;
2430 int old_prs = cs->partition_root_state;
2431
2432 /* top_cpuset.cpus_allowed tracks cpu_online_mask; it's read-only */
2433 if (cs == &top_cpuset)
2434 return -EACCES;
2435
2436 /*
2437 * An empty cpus_allowed is ok only if the cpuset has no tasks.
2438 * Since cpulist_parse() fails on an empty mask, we special case
2439 * that parsing. The validate_change() call ensures that cpusets
2440 * with tasks have cpus.
2441 */
2442 if (!*buf) {
2443 cpumask_clear(trialcs->cpus_allowed);
2444 cpumask_clear(trialcs->effective_xcpus);
2445 } else {
2446 retval = cpulist_parse(buf, trialcs->cpus_allowed);
2447 if (retval < 0)
2448 return retval;
2449
2450 if (!cpumask_subset(trialcs->cpus_allowed,
2451 top_cpuset.cpus_allowed))
2452 return -EINVAL;
2453
2454 /*
2455 * When exclusive_cpus isn't explicitly set, it is constrainted
2456 * by cpus_allowed and parent's effective_xcpus. Otherwise,
2457 * trialcs->effective_xcpus is used as a temporary cpumask
2458 * for checking validity of the partition root.
2459 */
2460 if (!cpumask_empty(trialcs->exclusive_cpus) || is_partition_valid(cs))
2461 compute_effective_exclusive_cpumask(trialcs, NULL);
2462 }
2463
2464 /* Nothing to do if the cpus didn't change */
2465 if (cpumask_equal(cs->cpus_allowed, trialcs->cpus_allowed))
2466 return 0;
2467
2468 if (alloc_cpumasks(NULL, &tmp))
2469 return -ENOMEM;
2470
2471 if (old_prs) {
2472 if (is_partition_valid(cs) &&
2473 cpumask_empty(trialcs->effective_xcpus)) {
2474 invalidate = true;
2475 cs->prs_err = PERR_INVCPUS;
2476 } else if (prstate_housekeeping_conflict(old_prs, trialcs->effective_xcpus)) {
2477 invalidate = true;
2478 cs->prs_err = PERR_HKEEPING;
2479 } else if (tasks_nocpu_error(parent, cs, trialcs->effective_xcpus)) {
2480 invalidate = true;
2481 cs->prs_err = PERR_NOCPUS;
2482 }
2483 }
2484
2485 /*
2486 * Check all the descendants in update_cpumasks_hier() if
2487 * effective_xcpus is to be changed.
2488 */
2489 if (!cpumask_equal(cs->effective_xcpus, trialcs->effective_xcpus))
2490 hier_flags = HIER_CHECKALL;
2491
2492 retval = validate_change(cs, trialcs);
2493
2494 if ((retval == -EINVAL) && cgroup_subsys_on_dfl(cpuset_cgrp_subsys)) {
2495 struct cgroup_subsys_state *css;
2496 struct cpuset *cp;
2497
2498 /*
2499 * The -EINVAL error code indicates that partition sibling
2500 * CPU exclusivity rule has been violated. We still allow
2501 * the cpumask change to proceed while invalidating the
2502 * partition. However, any conflicting sibling partitions
2503 * have to be marked as invalid too.
2504 */
2505 invalidate = true;
2506 rcu_read_lock();
2507 cpuset_for_each_child(cp, css, parent) {
2508 struct cpumask *xcpus = fetch_xcpus(trialcs);
2509
2510 if (is_partition_valid(cp) &&
2511 cpumask_intersects(xcpus, cp->effective_xcpus)) {
2512 rcu_read_unlock();
2513 update_parent_effective_cpumask(cp, partcmd_invalidate, NULL, &tmp);
2514 rcu_read_lock();
2515 }
2516 }
2517 rcu_read_unlock();
2518 retval = 0;
2519 }
2520
2521 if (retval < 0)
2522 goto out_free;
2523
2524 if (is_partition_valid(cs) ||
2525 (is_partition_invalid(cs) && !invalidate)) {
2526 struct cpumask *xcpus = trialcs->effective_xcpus;
2527
2528 if (cpumask_empty(xcpus) && is_partition_invalid(cs))
2529 xcpus = trialcs->cpus_allowed;
2530
2531 /*
2532 * Call remote_cpus_update() to handle valid remote partition
2533 */
2534 if (is_remote_partition(cs))
2535 remote_cpus_update(cs, xcpus, &tmp);
2536 else if (invalidate)
2537 update_parent_effective_cpumask(cs, partcmd_invalidate,
2538 NULL, &tmp);
2539 else
2540 update_parent_effective_cpumask(cs, partcmd_update,
2541 xcpus, &tmp);
2542 } else if (!cpumask_empty(cs->exclusive_cpus)) {
2543 /*
2544 * Use trialcs->effective_cpus as a temp cpumask
2545 */
2546 remote_partition_check(cs, trialcs->effective_xcpus,
2547 trialcs->effective_cpus, &tmp);
2548 }
2549
2550 spin_lock_irq(&callback_lock);
2551 cpumask_copy(cs->cpus_allowed, trialcs->cpus_allowed);
2552 cpumask_copy(cs->effective_xcpus, trialcs->effective_xcpus);
2553 if ((old_prs > 0) && !is_partition_valid(cs))
2554 reset_partition_data(cs);
2555 spin_unlock_irq(&callback_lock);
2556
2557 /* effective_cpus/effective_xcpus will be updated here */
2558 update_cpumasks_hier(cs, &tmp, hier_flags);
2559
2560 /* Update CS_SCHED_LOAD_BALANCE and/or sched_domains, if necessary */
2561 if (cs->partition_root_state)
2562 update_partition_sd_lb(cs, old_prs);
2563out_free:
2564 free_cpumasks(NULL, &tmp);
2565 return retval;
2566}
2567
2568/**
2569 * update_exclusive_cpumask - update the exclusive_cpus mask of a cpuset
2570 * @cs: the cpuset to consider
2571 * @trialcs: trial cpuset
2572 * @buf: buffer of cpu numbers written to this cpuset
2573 *
2574 * The tasks' cpumask will be updated if cs is a valid partition root.
2575 */
2576static int update_exclusive_cpumask(struct cpuset *cs, struct cpuset *trialcs,
2577 const char *buf)
2578{
2579 int retval;
2580 struct tmpmasks tmp;
2581 struct cpuset *parent = parent_cs(cs);
2582 bool invalidate = false;
2583 int hier_flags = 0;
2584 int old_prs = cs->partition_root_state;
2585
2586 if (!*buf) {
2587 cpumask_clear(trialcs->exclusive_cpus);
2588 cpumask_clear(trialcs->effective_xcpus);
2589 } else {
2590 retval = cpulist_parse(buf, trialcs->exclusive_cpus);
2591 if (retval < 0)
2592 return retval;
2593 if (!is_cpu_exclusive(cs))
2594 set_bit(CS_CPU_EXCLUSIVE, &trialcs->flags);
2595 }
2596
2597 /* Nothing to do if the CPUs didn't change */
2598 if (cpumask_equal(cs->exclusive_cpus, trialcs->exclusive_cpus))
2599 return 0;
2600
2601 if (*buf)
2602 compute_effective_exclusive_cpumask(trialcs, NULL);
2603
2604 /*
2605 * Check all the descendants in update_cpumasks_hier() if
2606 * effective_xcpus is to be changed.
2607 */
2608 if (!cpumask_equal(cs->effective_xcpus, trialcs->effective_xcpus))
2609 hier_flags = HIER_CHECKALL;
2610
2611 retval = validate_change(cs, trialcs);
2612 if (retval)
2613 return retval;
2614
2615 if (alloc_cpumasks(NULL, &tmp))
2616 return -ENOMEM;
2617
2618 if (old_prs) {
2619 if (cpumask_empty(trialcs->effective_xcpus)) {
2620 invalidate = true;
2621 cs->prs_err = PERR_INVCPUS;
2622 } else if (prstate_housekeeping_conflict(old_prs, trialcs->effective_xcpus)) {
2623 invalidate = true;
2624 cs->prs_err = PERR_HKEEPING;
2625 } else if (tasks_nocpu_error(parent, cs, trialcs->effective_xcpus)) {
2626 invalidate = true;
2627 cs->prs_err = PERR_NOCPUS;
2628 }
2629
2630 if (is_remote_partition(cs)) {
2631 if (invalidate)
2632 remote_partition_disable(cs, &tmp);
2633 else
2634 remote_cpus_update(cs, trialcs->effective_xcpus,
2635 &tmp);
2636 } else if (invalidate) {
2637 update_parent_effective_cpumask(cs, partcmd_invalidate,
2638 NULL, &tmp);
2639 } else {
2640 update_parent_effective_cpumask(cs, partcmd_update,
2641 trialcs->effective_xcpus, &tmp);
2642 }
2643 } else if (!cpumask_empty(trialcs->exclusive_cpus)) {
2644 /*
2645 * Use trialcs->effective_cpus as a temp cpumask
2646 */
2647 remote_partition_check(cs, trialcs->effective_xcpus,
2648 trialcs->effective_cpus, &tmp);
2649 }
2650 spin_lock_irq(&callback_lock);
2651 cpumask_copy(cs->exclusive_cpus, trialcs->exclusive_cpus);
2652 cpumask_copy(cs->effective_xcpus, trialcs->effective_xcpus);
2653 if ((old_prs > 0) && !is_partition_valid(cs))
2654 reset_partition_data(cs);
2655 spin_unlock_irq(&callback_lock);
2656
2657 /*
2658 * Call update_cpumasks_hier() to update effective_cpus/effective_xcpus
2659 * of the subtree when it is a valid partition root or effective_xcpus
2660 * is updated.
2661 */
2662 if (is_partition_valid(cs) || hier_flags)
2663 update_cpumasks_hier(cs, &tmp, hier_flags);
2664
2665 /* Update CS_SCHED_LOAD_BALANCE and/or sched_domains, if necessary */
2666 if (cs->partition_root_state)
2667 update_partition_sd_lb(cs, old_prs);
2668
2669 free_cpumasks(NULL, &tmp);
2670 return 0;
2671}
2672
2673/*
2674 * Migrate memory region from one set of nodes to another. This is
2675 * performed asynchronously as it can be called from process migration path
2676 * holding locks involved in process management. All mm migrations are
2677 * performed in the queued order and can be waited for by flushing
2678 * cpuset_migrate_mm_wq.
2679 */
2680
2681struct cpuset_migrate_mm_work {
2682 struct work_struct work;
2683 struct mm_struct *mm;
2684 nodemask_t from;
2685 nodemask_t to;
2686};
2687
2688static void cpuset_migrate_mm_workfn(struct work_struct *work)
2689{
2690 struct cpuset_migrate_mm_work *mwork =
2691 container_of(work, struct cpuset_migrate_mm_work, work);
2692
2693 /* on a wq worker, no need to worry about %current's mems_allowed */
2694 do_migrate_pages(mwork->mm, &mwork->from, &mwork->to, MPOL_MF_MOVE_ALL);
2695 mmput(mwork->mm);
2696 kfree(mwork);
2697}
2698
2699static void cpuset_migrate_mm(struct mm_struct *mm, const nodemask_t *from,
2700 const nodemask_t *to)
2701{
2702 struct cpuset_migrate_mm_work *mwork;
2703
2704 if (nodes_equal(*from, *to)) {
2705 mmput(mm);
2706 return;
2707 }
2708
2709 mwork = kzalloc(sizeof(*mwork), GFP_KERNEL);
2710 if (mwork) {
2711 mwork->mm = mm;
2712 mwork->from = *from;
2713 mwork->to = *to;
2714 INIT_WORK(&mwork->work, cpuset_migrate_mm_workfn);
2715 queue_work(cpuset_migrate_mm_wq, &mwork->work);
2716 } else {
2717 mmput(mm);
2718 }
2719}
2720
2721static void cpuset_post_attach(void)
2722{
2723 flush_workqueue(cpuset_migrate_mm_wq);
2724}
2725
2726/*
2727 * cpuset_change_task_nodemask - change task's mems_allowed and mempolicy
2728 * @tsk: the task to change
2729 * @newmems: new nodes that the task will be set
2730 *
2731 * We use the mems_allowed_seq seqlock to safely update both tsk->mems_allowed
2732 * and rebind an eventual tasks' mempolicy. If the task is allocating in
2733 * parallel, it might temporarily see an empty intersection, which results in
2734 * a seqlock check and retry before OOM or allocation failure.
2735 */
2736static void cpuset_change_task_nodemask(struct task_struct *tsk,
2737 nodemask_t *newmems)
2738{
2739 task_lock(tsk);
2740
2741 local_irq_disable();
2742 write_seqcount_begin(&tsk->mems_allowed_seq);
2743
2744 nodes_or(tsk->mems_allowed, tsk->mems_allowed, *newmems);
2745 mpol_rebind_task(tsk, newmems);
2746 tsk->mems_allowed = *newmems;
2747
2748 write_seqcount_end(&tsk->mems_allowed_seq);
2749 local_irq_enable();
2750
2751 task_unlock(tsk);
2752}
2753
2754static void *cpuset_being_rebound;
2755
2756/**
2757 * update_tasks_nodemask - Update the nodemasks of tasks in the cpuset.
2758 * @cs: the cpuset in which each task's mems_allowed mask needs to be changed
2759 *
2760 * Iterate through each task of @cs updating its mems_allowed to the
2761 * effective cpuset's. As this function is called with cpuset_mutex held,
2762 * cpuset membership stays stable.
2763 */
2764static void update_tasks_nodemask(struct cpuset *cs)
2765{
2766 static nodemask_t newmems; /* protected by cpuset_mutex */
2767 struct css_task_iter it;
2768 struct task_struct *task;
2769
2770 cpuset_being_rebound = cs; /* causes mpol_dup() rebind */
2771
2772 guarantee_online_mems(cs, &newmems);
2773
2774 /*
2775 * The mpol_rebind_mm() call takes mmap_lock, which we couldn't
2776 * take while holding tasklist_lock. Forks can happen - the
2777 * mpol_dup() cpuset_being_rebound check will catch such forks,
2778 * and rebind their vma mempolicies too. Because we still hold
2779 * the global cpuset_mutex, we know that no other rebind effort
2780 * will be contending for the global variable cpuset_being_rebound.
2781 * It's ok if we rebind the same mm twice; mpol_rebind_mm()
2782 * is idempotent. Also migrate pages in each mm to new nodes.
2783 */
2784 css_task_iter_start(&cs->css, 0, &it);
2785 while ((task = css_task_iter_next(&it))) {
2786 struct mm_struct *mm;
2787 bool migrate;
2788
2789 cpuset_change_task_nodemask(task, &newmems);
2790
2791 mm = get_task_mm(task);
2792 if (!mm)
2793 continue;
2794
2795 migrate = is_memory_migrate(cs);
2796
2797 mpol_rebind_mm(mm, &cs->mems_allowed);
2798 if (migrate)
2799 cpuset_migrate_mm(mm, &cs->old_mems_allowed, &newmems);
2800 else
2801 mmput(mm);
2802 }
2803 css_task_iter_end(&it);
2804
2805 /*
2806 * All the tasks' nodemasks have been updated, update
2807 * cs->old_mems_allowed.
2808 */
2809 cs->old_mems_allowed = newmems;
2810
2811 /* We're done rebinding vmas to this cpuset's new mems_allowed. */
2812 cpuset_being_rebound = NULL;
2813}
2814
2815/*
2816 * update_nodemasks_hier - Update effective nodemasks and tasks in the subtree
2817 * @cs: the cpuset to consider
2818 * @new_mems: a temp variable for calculating new effective_mems
2819 *
2820 * When configured nodemask is changed, the effective nodemasks of this cpuset
2821 * and all its descendants need to be updated.
2822 *
2823 * On legacy hierarchy, effective_mems will be the same with mems_allowed.
2824 *
2825 * Called with cpuset_mutex held
2826 */
2827static void update_nodemasks_hier(struct cpuset *cs, nodemask_t *new_mems)
2828{
2829 struct cpuset *cp;
2830 struct cgroup_subsys_state *pos_css;
2831
2832 rcu_read_lock();
2833 cpuset_for_each_descendant_pre(cp, pos_css, cs) {
2834 struct cpuset *parent = parent_cs(cp);
2835
2836 nodes_and(*new_mems, cp->mems_allowed, parent->effective_mems);
2837
2838 /*
2839 * If it becomes empty, inherit the effective mask of the
2840 * parent, which is guaranteed to have some MEMs.
2841 */
2842 if (is_in_v2_mode() && nodes_empty(*new_mems))
2843 *new_mems = parent->effective_mems;
2844
2845 /* Skip the whole subtree if the nodemask remains the same. */
2846 if (nodes_equal(*new_mems, cp->effective_mems)) {
2847 pos_css = css_rightmost_descendant(pos_css);
2848 continue;
2849 }
2850
2851 if (!css_tryget_online(&cp->css))
2852 continue;
2853 rcu_read_unlock();
2854
2855 spin_lock_irq(&callback_lock);
2856 cp->effective_mems = *new_mems;
2857 spin_unlock_irq(&callback_lock);
2858
2859 WARN_ON(!is_in_v2_mode() &&
2860 !nodes_equal(cp->mems_allowed, cp->effective_mems));
2861
2862 update_tasks_nodemask(cp);
2863
2864 rcu_read_lock();
2865 css_put(&cp->css);
2866 }
2867 rcu_read_unlock();
2868}
2869
2870/*
2871 * Handle user request to change the 'mems' memory placement
2872 * of a cpuset. Needs to validate the request, update the
2873 * cpusets mems_allowed, and for each task in the cpuset,
2874 * update mems_allowed and rebind task's mempolicy and any vma
2875 * mempolicies and if the cpuset is marked 'memory_migrate',
2876 * migrate the tasks pages to the new memory.
2877 *
2878 * Call with cpuset_mutex held. May take callback_lock during call.
2879 * Will take tasklist_lock, scan tasklist for tasks in cpuset cs,
2880 * lock each such tasks mm->mmap_lock, scan its vma's and rebind
2881 * their mempolicies to the cpusets new mems_allowed.
2882 */
2883static int update_nodemask(struct cpuset *cs, struct cpuset *trialcs,
2884 const char *buf)
2885{
2886 int retval;
2887
2888 /*
2889 * top_cpuset.mems_allowed tracks node_stats[N_MEMORY];
2890 * it's read-only
2891 */
2892 if (cs == &top_cpuset) {
2893 retval = -EACCES;
2894 goto done;
2895 }
2896
2897 /*
2898 * An empty mems_allowed is ok iff there are no tasks in the cpuset.
2899 * Since nodelist_parse() fails on an empty mask, we special case
2900 * that parsing. The validate_change() call ensures that cpusets
2901 * with tasks have memory.
2902 */
2903 if (!*buf) {
2904 nodes_clear(trialcs->mems_allowed);
2905 } else {
2906 retval = nodelist_parse(buf, trialcs->mems_allowed);
2907 if (retval < 0)
2908 goto done;
2909
2910 if (!nodes_subset(trialcs->mems_allowed,
2911 top_cpuset.mems_allowed)) {
2912 retval = -EINVAL;
2913 goto done;
2914 }
2915 }
2916
2917 if (nodes_equal(cs->mems_allowed, trialcs->mems_allowed)) {
2918 retval = 0; /* Too easy - nothing to do */
2919 goto done;
2920 }
2921 retval = validate_change(cs, trialcs);
2922 if (retval < 0)
2923 goto done;
2924
2925 check_insane_mems_config(&trialcs->mems_allowed);
2926
2927 spin_lock_irq(&callback_lock);
2928 cs->mems_allowed = trialcs->mems_allowed;
2929 spin_unlock_irq(&callback_lock);
2930
2931 /* use trialcs->mems_allowed as a temp variable */
2932 update_nodemasks_hier(cs, &trialcs->mems_allowed);
2933done:
2934 return retval;
2935}
2936
2937bool current_cpuset_is_being_rebound(void)
2938{
2939 bool ret;
2940
2941 rcu_read_lock();
2942 ret = task_cs(current) == cpuset_being_rebound;
2943 rcu_read_unlock();
2944
2945 return ret;
2946}
2947
2948static int update_relax_domain_level(struct cpuset *cs, s64 val)
2949{
2950#ifdef CONFIG_SMP
2951 if (val < -1 || val >= sched_domain_level_max)
2952 return -EINVAL;
2953#endif
2954
2955 if (val != cs->relax_domain_level) {
2956 cs->relax_domain_level = val;
2957 if (!cpumask_empty(cs->cpus_allowed) &&
2958 is_sched_load_balance(cs))
2959 rebuild_sched_domains_locked();
2960 }
2961
2962 return 0;
2963}
2964
2965/**
2966 * update_tasks_flags - update the spread flags of tasks in the cpuset.
2967 * @cs: the cpuset in which each task's spread flags needs to be changed
2968 *
2969 * Iterate through each task of @cs updating its spread flags. As this
2970 * function is called with cpuset_mutex held, cpuset membership stays
2971 * stable.
2972 */
2973static void update_tasks_flags(struct cpuset *cs)
2974{
2975 struct css_task_iter it;
2976 struct task_struct *task;
2977
2978 css_task_iter_start(&cs->css, 0, &it);
2979 while ((task = css_task_iter_next(&it)))
2980 cpuset_update_task_spread_flags(cs, task);
2981 css_task_iter_end(&it);
2982}
2983
2984/*
2985 * update_flag - read a 0 or a 1 in a file and update associated flag
2986 * bit: the bit to update (see cpuset_flagbits_t)
2987 * cs: the cpuset to update
2988 * turning_on: whether the flag is being set or cleared
2989 *
2990 * Call with cpuset_mutex held.
2991 */
2992
2993static int update_flag(cpuset_flagbits_t bit, struct cpuset *cs,
2994 int turning_on)
2995{
2996 struct cpuset *trialcs;
2997 int balance_flag_changed;
2998 int spread_flag_changed;
2999 int err;
3000
3001 trialcs = alloc_trial_cpuset(cs);
3002 if (!trialcs)
3003 return -ENOMEM;
3004
3005 if (turning_on)
3006 set_bit(bit, &trialcs->flags);
3007 else
3008 clear_bit(bit, &trialcs->flags);
3009
3010 err = validate_change(cs, trialcs);
3011 if (err < 0)
3012 goto out;
3013
3014 balance_flag_changed = (is_sched_load_balance(cs) !=
3015 is_sched_load_balance(trialcs));
3016
3017 spread_flag_changed = ((is_spread_slab(cs) != is_spread_slab(trialcs))
3018 || (is_spread_page(cs) != is_spread_page(trialcs)));
3019
3020 spin_lock_irq(&callback_lock);
3021 cs->flags = trialcs->flags;
3022 spin_unlock_irq(&callback_lock);
3023
3024 if (!cpumask_empty(trialcs->cpus_allowed) && balance_flag_changed)
3025 rebuild_sched_domains_locked();
3026
3027 if (spread_flag_changed)
3028 update_tasks_flags(cs);
3029out:
3030 free_cpuset(trialcs);
3031 return err;
3032}
3033
3034/**
3035 * update_prstate - update partition_root_state
3036 * @cs: the cpuset to update
3037 * @new_prs: new partition root state
3038 * Return: 0 if successful, != 0 if error
3039 *
3040 * Call with cpuset_mutex held.
3041 */
3042static int update_prstate(struct cpuset *cs, int new_prs)
3043{
3044 int err = PERR_NONE, old_prs = cs->partition_root_state;
3045 struct cpuset *parent = parent_cs(cs);
3046 struct tmpmasks tmpmask;
3047 bool new_xcpus_state = false;
3048
3049 if (old_prs == new_prs)
3050 return 0;
3051
3052 /*
3053 * Treat a previously invalid partition root as if it is a "member".
3054 */
3055 if (new_prs && is_prs_invalid(old_prs))
3056 old_prs = PRS_MEMBER;
3057
3058 if (alloc_cpumasks(NULL, &tmpmask))
3059 return -ENOMEM;
3060
3061 /*
3062 * Setup effective_xcpus if not properly set yet, it will be cleared
3063 * later if partition becomes invalid.
3064 */
3065 if ((new_prs > 0) && cpumask_empty(cs->exclusive_cpus)) {
3066 spin_lock_irq(&callback_lock);
3067 cpumask_and(cs->effective_xcpus,
3068 cs->cpus_allowed, parent->effective_xcpus);
3069 spin_unlock_irq(&callback_lock);
3070 }
3071
3072 err = update_partition_exclusive(cs, new_prs);
3073 if (err)
3074 goto out;
3075
3076 if (!old_prs) {
3077 enum partition_cmd cmd = (new_prs == PRS_ROOT)
3078 ? partcmd_enable : partcmd_enablei;
3079
3080 /*
3081 * cpus_allowed cannot be empty.
3082 */
3083 if (cpumask_empty(cs->cpus_allowed)) {
3084 err = PERR_CPUSEMPTY;
3085 goto out;
3086 }
3087
3088 err = update_parent_effective_cpumask(cs, cmd, NULL, &tmpmask);
3089 /*
3090 * If an attempt to become local partition root fails,
3091 * try to become a remote partition root instead.
3092 */
3093 if (err && remote_partition_enable(cs, new_prs, &tmpmask))
3094 err = 0;
3095 } else if (old_prs && new_prs) {
3096 /*
3097 * A change in load balance state only, no change in cpumasks.
3098 */
3099 new_xcpus_state = true;
3100 } else {
3101 /*
3102 * Switching back to member is always allowed even if it
3103 * disables child partitions.
3104 */
3105 if (is_remote_partition(cs))
3106 remote_partition_disable(cs, &tmpmask);
3107 else
3108 update_parent_effective_cpumask(cs, partcmd_disable,
3109 NULL, &tmpmask);
3110
3111 /*
3112 * Invalidation of child partitions will be done in
3113 * update_cpumasks_hier().
3114 */
3115 }
3116out:
3117 /*
3118 * Make partition invalid & disable CS_CPU_EXCLUSIVE if an error
3119 * happens.
3120 */
3121 if (err) {
3122 new_prs = -new_prs;
3123 update_partition_exclusive(cs, new_prs);
3124 }
3125
3126 spin_lock_irq(&callback_lock);
3127 cs->partition_root_state = new_prs;
3128 WRITE_ONCE(cs->prs_err, err);
3129 if (!is_partition_valid(cs))
3130 reset_partition_data(cs);
3131 else if (new_xcpus_state)
3132 partition_xcpus_newstate(old_prs, new_prs, cs->effective_xcpus);
3133 spin_unlock_irq(&callback_lock);
3134 update_unbound_workqueue_cpumask(new_xcpus_state);
3135
3136 /* Force update if switching back to member */
3137 update_cpumasks_hier(cs, &tmpmask, !new_prs ? HIER_CHECKALL : 0);
3138
3139 /* Update sched domains and load balance flag */
3140 update_partition_sd_lb(cs, old_prs);
3141
3142 notify_partition_change(cs, old_prs);
3143 free_cpumasks(NULL, &tmpmask);
3144 return 0;
3145}
3146
3147/*
3148 * Frequency meter - How fast is some event occurring?
3149 *
3150 * These routines manage a digitally filtered, constant time based,
3151 * event frequency meter. There are four routines:
3152 * fmeter_init() - initialize a frequency meter.
3153 * fmeter_markevent() - called each time the event happens.
3154 * fmeter_getrate() - returns the recent rate of such events.
3155 * fmeter_update() - internal routine used to update fmeter.
3156 *
3157 * A common data structure is passed to each of these routines,
3158 * which is used to keep track of the state required to manage the
3159 * frequency meter and its digital filter.
3160 *
3161 * The filter works on the number of events marked per unit time.
3162 * The filter is single-pole low-pass recursive (IIR). The time unit
3163 * is 1 second. Arithmetic is done using 32-bit integers scaled to
3164 * simulate 3 decimal digits of precision (multiplied by 1000).
3165 *
3166 * With an FM_COEF of 933, and a time base of 1 second, the filter
3167 * has a half-life of 10 seconds, meaning that if the events quit
3168 * happening, then the rate returned from the fmeter_getrate()
3169 * will be cut in half each 10 seconds, until it converges to zero.
3170 *
3171 * It is not worth doing a real infinitely recursive filter. If more
3172 * than FM_MAXTICKS ticks have elapsed since the last filter event,
3173 * just compute FM_MAXTICKS ticks worth, by which point the level
3174 * will be stable.
3175 *
3176 * Limit the count of unprocessed events to FM_MAXCNT, so as to avoid
3177 * arithmetic overflow in the fmeter_update() routine.
3178 *
3179 * Given the simple 32 bit integer arithmetic used, this meter works
3180 * best for reporting rates between one per millisecond (msec) and
3181 * one per 32 (approx) seconds. At constant rates faster than one
3182 * per msec it maxes out at values just under 1,000,000. At constant
3183 * rates between one per msec, and one per second it will stabilize
3184 * to a value N*1000, where N is the rate of events per second.
3185 * At constant rates between one per second and one per 32 seconds,
3186 * it will be choppy, moving up on the seconds that have an event,
3187 * and then decaying until the next event. At rates slower than
3188 * about one in 32 seconds, it decays all the way back to zero between
3189 * each event.
3190 */
3191
3192#define FM_COEF 933 /* coefficient for half-life of 10 secs */
3193#define FM_MAXTICKS ((u32)99) /* useless computing more ticks than this */
3194#define FM_MAXCNT 1000000 /* limit cnt to avoid overflow */
3195#define FM_SCALE 1000 /* faux fixed point scale */
3196
3197/* Initialize a frequency meter */
3198static void fmeter_init(struct fmeter *fmp)
3199{
3200 fmp->cnt = 0;
3201 fmp->val = 0;
3202 fmp->time = 0;
3203 spin_lock_init(&fmp->lock);
3204}
3205
3206/* Internal meter update - process cnt events and update value */
3207static void fmeter_update(struct fmeter *fmp)
3208{
3209 time64_t now;
3210 u32 ticks;
3211
3212 now = ktime_get_seconds();
3213 ticks = now - fmp->time;
3214
3215 if (ticks == 0)
3216 return;
3217
3218 ticks = min(FM_MAXTICKS, ticks);
3219 while (ticks-- > 0)
3220 fmp->val = (FM_COEF * fmp->val) / FM_SCALE;
3221 fmp->time = now;
3222
3223 fmp->val += ((FM_SCALE - FM_COEF) * fmp->cnt) / FM_SCALE;
3224 fmp->cnt = 0;
3225}
3226
3227/* Process any previous ticks, then bump cnt by one (times scale). */
3228static void fmeter_markevent(struct fmeter *fmp)
3229{
3230 spin_lock(&fmp->lock);
3231 fmeter_update(fmp);
3232 fmp->cnt = min(FM_MAXCNT, fmp->cnt + FM_SCALE);
3233 spin_unlock(&fmp->lock);
3234}
3235
3236/* Process any previous ticks, then return current value. */
3237static int fmeter_getrate(struct fmeter *fmp)
3238{
3239 int val;
3240
3241 spin_lock(&fmp->lock);
3242 fmeter_update(fmp);
3243 val = fmp->val;
3244 spin_unlock(&fmp->lock);
3245 return val;
3246}
3247
3248static struct cpuset *cpuset_attach_old_cs;
3249
3250/*
3251 * Check to see if a cpuset can accept a new task
3252 * For v1, cpus_allowed and mems_allowed can't be empty.
3253 * For v2, effective_cpus can't be empty.
3254 * Note that in v1, effective_cpus = cpus_allowed.
3255 */
3256static int cpuset_can_attach_check(struct cpuset *cs)
3257{
3258 if (cpumask_empty(cs->effective_cpus) ||
3259 (!is_in_v2_mode() && nodes_empty(cs->mems_allowed)))
3260 return -ENOSPC;
3261 return 0;
3262}
3263
3264static void reset_migrate_dl_data(struct cpuset *cs)
3265{
3266 cs->nr_migrate_dl_tasks = 0;
3267 cs->sum_migrate_dl_bw = 0;
3268}
3269
3270/* Called by cgroups to determine if a cpuset is usable; cpuset_mutex held */
3271static int cpuset_can_attach(struct cgroup_taskset *tset)
3272{
3273 struct cgroup_subsys_state *css;
3274 struct cpuset *cs, *oldcs;
3275 struct task_struct *task;
3276 bool cpus_updated, mems_updated;
3277 int ret;
3278
3279 /* used later by cpuset_attach() */
3280 cpuset_attach_old_cs = task_cs(cgroup_taskset_first(tset, &css));
3281 oldcs = cpuset_attach_old_cs;
3282 cs = css_cs(css);
3283
3284 mutex_lock(&cpuset_mutex);
3285
3286 /* Check to see if task is allowed in the cpuset */
3287 ret = cpuset_can_attach_check(cs);
3288 if (ret)
3289 goto out_unlock;
3290
3291 cpus_updated = !cpumask_equal(cs->effective_cpus, oldcs->effective_cpus);
3292 mems_updated = !nodes_equal(cs->effective_mems, oldcs->effective_mems);
3293
3294 cgroup_taskset_for_each(task, css, tset) {
3295 ret = task_can_attach(task);
3296 if (ret)
3297 goto out_unlock;
3298
3299 /*
3300 * Skip rights over task check in v2 when nothing changes,
3301 * migration permission derives from hierarchy ownership in
3302 * cgroup_procs_write_permission()).
3303 */
3304 if (!cgroup_subsys_on_dfl(cpuset_cgrp_subsys) ||
3305 (cpus_updated || mems_updated)) {
3306 ret = security_task_setscheduler(task);
3307 if (ret)
3308 goto out_unlock;
3309 }
3310
3311 if (dl_task(task)) {
3312 cs->nr_migrate_dl_tasks++;
3313 cs->sum_migrate_dl_bw += task->dl.dl_bw;
3314 }
3315 }
3316
3317 if (!cs->nr_migrate_dl_tasks)
3318 goto out_success;
3319
3320 if (!cpumask_intersects(oldcs->effective_cpus, cs->effective_cpus)) {
3321 int cpu = cpumask_any_and(cpu_active_mask, cs->effective_cpus);
3322
3323 if (unlikely(cpu >= nr_cpu_ids)) {
3324 reset_migrate_dl_data(cs);
3325 ret = -EINVAL;
3326 goto out_unlock;
3327 }
3328
3329 ret = dl_bw_alloc(cpu, cs->sum_migrate_dl_bw);
3330 if (ret) {
3331 reset_migrate_dl_data(cs);
3332 goto out_unlock;
3333 }
3334 }
3335
3336out_success:
3337 /*
3338 * Mark attach is in progress. This makes validate_change() fail
3339 * changes which zero cpus/mems_allowed.
3340 */
3341 cs->attach_in_progress++;
3342out_unlock:
3343 mutex_unlock(&cpuset_mutex);
3344 return ret;
3345}
3346
3347static void cpuset_cancel_attach(struct cgroup_taskset *tset)
3348{
3349 struct cgroup_subsys_state *css;
3350 struct cpuset *cs;
3351
3352 cgroup_taskset_first(tset, &css);
3353 cs = css_cs(css);
3354
3355 mutex_lock(&cpuset_mutex);
3356 cs->attach_in_progress--;
3357 if (!cs->attach_in_progress)
3358 wake_up(&cpuset_attach_wq);
3359
3360 if (cs->nr_migrate_dl_tasks) {
3361 int cpu = cpumask_any(cs->effective_cpus);
3362
3363 dl_bw_free(cpu, cs->sum_migrate_dl_bw);
3364 reset_migrate_dl_data(cs);
3365 }
3366
3367 mutex_unlock(&cpuset_mutex);
3368}
3369
3370/*
3371 * Protected by cpuset_mutex. cpus_attach is used only by cpuset_attach_task()
3372 * but we can't allocate it dynamically there. Define it global and
3373 * allocate from cpuset_init().
3374 */
3375static cpumask_var_t cpus_attach;
3376static nodemask_t cpuset_attach_nodemask_to;
3377
3378static void cpuset_attach_task(struct cpuset *cs, struct task_struct *task)
3379{
3380 lockdep_assert_held(&cpuset_mutex);
3381
3382 if (cs != &top_cpuset)
3383 guarantee_online_cpus(task, cpus_attach);
3384 else
3385 cpumask_andnot(cpus_attach, task_cpu_possible_mask(task),
3386 subpartitions_cpus);
3387 /*
3388 * can_attach beforehand should guarantee that this doesn't
3389 * fail. TODO: have a better way to handle failure here
3390 */
3391 WARN_ON_ONCE(set_cpus_allowed_ptr(task, cpus_attach));
3392
3393 cpuset_change_task_nodemask(task, &cpuset_attach_nodemask_to);
3394 cpuset_update_task_spread_flags(cs, task);
3395}
3396
3397static void cpuset_attach(struct cgroup_taskset *tset)
3398{
3399 struct task_struct *task;
3400 struct task_struct *leader;
3401 struct cgroup_subsys_state *css;
3402 struct cpuset *cs;
3403 struct cpuset *oldcs = cpuset_attach_old_cs;
3404 bool cpus_updated, mems_updated;
3405
3406 cgroup_taskset_first(tset, &css);
3407 cs = css_cs(css);
3408
3409 lockdep_assert_cpus_held(); /* see cgroup_attach_lock() */
3410 mutex_lock(&cpuset_mutex);
3411 cpus_updated = !cpumask_equal(cs->effective_cpus,
3412 oldcs->effective_cpus);
3413 mems_updated = !nodes_equal(cs->effective_mems, oldcs->effective_mems);
3414
3415 /*
3416 * In the default hierarchy, enabling cpuset in the child cgroups
3417 * will trigger a number of cpuset_attach() calls with no change
3418 * in effective cpus and mems. In that case, we can optimize out
3419 * by skipping the task iteration and update.
3420 */
3421 if (cgroup_subsys_on_dfl(cpuset_cgrp_subsys) &&
3422 !cpus_updated && !mems_updated) {
3423 cpuset_attach_nodemask_to = cs->effective_mems;
3424 goto out;
3425 }
3426
3427 guarantee_online_mems(cs, &cpuset_attach_nodemask_to);
3428
3429 cgroup_taskset_for_each(task, css, tset)
3430 cpuset_attach_task(cs, task);
3431
3432 /*
3433 * Change mm for all threadgroup leaders. This is expensive and may
3434 * sleep and should be moved outside migration path proper. Skip it
3435 * if there is no change in effective_mems and CS_MEMORY_MIGRATE is
3436 * not set.
3437 */
3438 cpuset_attach_nodemask_to = cs->effective_mems;
3439 if (!is_memory_migrate(cs) && !mems_updated)
3440 goto out;
3441
3442 cgroup_taskset_for_each_leader(leader, css, tset) {
3443 struct mm_struct *mm = get_task_mm(leader);
3444
3445 if (mm) {
3446 mpol_rebind_mm(mm, &cpuset_attach_nodemask_to);
3447
3448 /*
3449 * old_mems_allowed is the same with mems_allowed
3450 * here, except if this task is being moved
3451 * automatically due to hotplug. In that case
3452 * @mems_allowed has been updated and is empty, so
3453 * @old_mems_allowed is the right nodesets that we
3454 * migrate mm from.
3455 */
3456 if (is_memory_migrate(cs))
3457 cpuset_migrate_mm(mm, &oldcs->old_mems_allowed,
3458 &cpuset_attach_nodemask_to);
3459 else
3460 mmput(mm);
3461 }
3462 }
3463
3464out:
3465 cs->old_mems_allowed = cpuset_attach_nodemask_to;
3466
3467 if (cs->nr_migrate_dl_tasks) {
3468 cs->nr_deadline_tasks += cs->nr_migrate_dl_tasks;
3469 oldcs->nr_deadline_tasks -= cs->nr_migrate_dl_tasks;
3470 reset_migrate_dl_data(cs);
3471 }
3472
3473 cs->attach_in_progress--;
3474 if (!cs->attach_in_progress)
3475 wake_up(&cpuset_attach_wq);
3476
3477 mutex_unlock(&cpuset_mutex);
3478}
3479
3480/* The various types of files and directories in a cpuset file system */
3481
3482typedef enum {
3483 FILE_MEMORY_MIGRATE,
3484 FILE_CPULIST,
3485 FILE_MEMLIST,
3486 FILE_EFFECTIVE_CPULIST,
3487 FILE_EFFECTIVE_MEMLIST,
3488 FILE_SUBPARTS_CPULIST,
3489 FILE_EXCLUSIVE_CPULIST,
3490 FILE_EFFECTIVE_XCPULIST,
3491 FILE_ISOLATED_CPULIST,
3492 FILE_CPU_EXCLUSIVE,
3493 FILE_MEM_EXCLUSIVE,
3494 FILE_MEM_HARDWALL,
3495 FILE_SCHED_LOAD_BALANCE,
3496 FILE_PARTITION_ROOT,
3497 FILE_SCHED_RELAX_DOMAIN_LEVEL,
3498 FILE_MEMORY_PRESSURE_ENABLED,
3499 FILE_MEMORY_PRESSURE,
3500 FILE_SPREAD_PAGE,
3501 FILE_SPREAD_SLAB,
3502} cpuset_filetype_t;
3503
3504static int cpuset_write_u64(struct cgroup_subsys_state *css, struct cftype *cft,
3505 u64 val)
3506{
3507 struct cpuset *cs = css_cs(css);
3508 cpuset_filetype_t type = cft->private;
3509 int retval = 0;
3510
3511 cpus_read_lock();
3512 mutex_lock(&cpuset_mutex);
3513 if (!is_cpuset_online(cs)) {
3514 retval = -ENODEV;
3515 goto out_unlock;
3516 }
3517
3518 switch (type) {
3519 case FILE_CPU_EXCLUSIVE:
3520 retval = update_flag(CS_CPU_EXCLUSIVE, cs, val);
3521 break;
3522 case FILE_MEM_EXCLUSIVE:
3523 retval = update_flag(CS_MEM_EXCLUSIVE, cs, val);
3524 break;
3525 case FILE_MEM_HARDWALL:
3526 retval = update_flag(CS_MEM_HARDWALL, cs, val);
3527 break;
3528 case FILE_SCHED_LOAD_BALANCE:
3529 retval = update_flag(CS_SCHED_LOAD_BALANCE, cs, val);
3530 break;
3531 case FILE_MEMORY_MIGRATE:
3532 retval = update_flag(CS_MEMORY_MIGRATE, cs, val);
3533 break;
3534 case FILE_MEMORY_PRESSURE_ENABLED:
3535 cpuset_memory_pressure_enabled = !!val;
3536 break;
3537 case FILE_SPREAD_PAGE:
3538 retval = update_flag(CS_SPREAD_PAGE, cs, val);
3539 break;
3540 case FILE_SPREAD_SLAB:
3541 retval = update_flag(CS_SPREAD_SLAB, cs, val);
3542 break;
3543 default:
3544 retval = -EINVAL;
3545 break;
3546 }
3547out_unlock:
3548 mutex_unlock(&cpuset_mutex);
3549 cpus_read_unlock();
3550 return retval;
3551}
3552
3553static int cpuset_write_s64(struct cgroup_subsys_state *css, struct cftype *cft,
3554 s64 val)
3555{
3556 struct cpuset *cs = css_cs(css);
3557 cpuset_filetype_t type = cft->private;
3558 int retval = -ENODEV;
3559
3560 cpus_read_lock();
3561 mutex_lock(&cpuset_mutex);
3562 if (!is_cpuset_online(cs))
3563 goto out_unlock;
3564
3565 switch (type) {
3566 case FILE_SCHED_RELAX_DOMAIN_LEVEL:
3567 retval = update_relax_domain_level(cs, val);
3568 break;
3569 default:
3570 retval = -EINVAL;
3571 break;
3572 }
3573out_unlock:
3574 mutex_unlock(&cpuset_mutex);
3575 cpus_read_unlock();
3576 return retval;
3577}
3578
3579/*
3580 * Common handling for a write to a "cpus" or "mems" file.
3581 */
3582static ssize_t cpuset_write_resmask(struct kernfs_open_file *of,
3583 char *buf, size_t nbytes, loff_t off)
3584{
3585 struct cpuset *cs = css_cs(of_css(of));
3586 struct cpuset *trialcs;
3587 int retval = -ENODEV;
3588
3589 buf = strstrip(buf);
3590
3591 /*
3592 * CPU or memory hotunplug may leave @cs w/o any execution
3593 * resources, in which case the hotplug code asynchronously updates
3594 * configuration and transfers all tasks to the nearest ancestor
3595 * which can execute.
3596 *
3597 * As writes to "cpus" or "mems" may restore @cs's execution
3598 * resources, wait for the previously scheduled operations before
3599 * proceeding, so that we don't end up keep removing tasks added
3600 * after execution capability is restored.
3601 *
3602 * cpuset_hotplug_work calls back into cgroup core via
3603 * cgroup_transfer_tasks() and waiting for it from a cgroupfs
3604 * operation like this one can lead to a deadlock through kernfs
3605 * active_ref protection. Let's break the protection. Losing the
3606 * protection is okay as we check whether @cs is online after
3607 * grabbing cpuset_mutex anyway. This only happens on the legacy
3608 * hierarchies.
3609 */
3610 css_get(&cs->css);
3611 kernfs_break_active_protection(of->kn);
3612 flush_work(&cpuset_hotplug_work);
3613
3614 cpus_read_lock();
3615 mutex_lock(&cpuset_mutex);
3616 if (!is_cpuset_online(cs))
3617 goto out_unlock;
3618
3619 trialcs = alloc_trial_cpuset(cs);
3620 if (!trialcs) {
3621 retval = -ENOMEM;
3622 goto out_unlock;
3623 }
3624
3625 switch (of_cft(of)->private) {
3626 case FILE_CPULIST:
3627 retval = update_cpumask(cs, trialcs, buf);
3628 break;
3629 case FILE_EXCLUSIVE_CPULIST:
3630 retval = update_exclusive_cpumask(cs, trialcs, buf);
3631 break;
3632 case FILE_MEMLIST:
3633 retval = update_nodemask(cs, trialcs, buf);
3634 break;
3635 default:
3636 retval = -EINVAL;
3637 break;
3638 }
3639
3640 free_cpuset(trialcs);
3641out_unlock:
3642 mutex_unlock(&cpuset_mutex);
3643 cpus_read_unlock();
3644 kernfs_unbreak_active_protection(of->kn);
3645 css_put(&cs->css);
3646 flush_workqueue(cpuset_migrate_mm_wq);
3647 return retval ?: nbytes;
3648}
3649
3650/*
3651 * These ascii lists should be read in a single call, by using a user
3652 * buffer large enough to hold the entire map. If read in smaller
3653 * chunks, there is no guarantee of atomicity. Since the display format
3654 * used, list of ranges of sequential numbers, is variable length,
3655 * and since these maps can change value dynamically, one could read
3656 * gibberish by doing partial reads while a list was changing.
3657 */
3658static int cpuset_common_seq_show(struct seq_file *sf, void *v)
3659{
3660 struct cpuset *cs = css_cs(seq_css(sf));
3661 cpuset_filetype_t type = seq_cft(sf)->private;
3662 int ret = 0;
3663
3664 spin_lock_irq(&callback_lock);
3665
3666 switch (type) {
3667 case FILE_CPULIST:
3668 seq_printf(sf, "%*pbl\n", cpumask_pr_args(cs->cpus_allowed));
3669 break;
3670 case FILE_MEMLIST:
3671 seq_printf(sf, "%*pbl\n", nodemask_pr_args(&cs->mems_allowed));
3672 break;
3673 case FILE_EFFECTIVE_CPULIST:
3674 seq_printf(sf, "%*pbl\n", cpumask_pr_args(cs->effective_cpus));
3675 break;
3676 case FILE_EFFECTIVE_MEMLIST:
3677 seq_printf(sf, "%*pbl\n", nodemask_pr_args(&cs->effective_mems));
3678 break;
3679 case FILE_EXCLUSIVE_CPULIST:
3680 seq_printf(sf, "%*pbl\n", cpumask_pr_args(cs->exclusive_cpus));
3681 break;
3682 case FILE_EFFECTIVE_XCPULIST:
3683 seq_printf(sf, "%*pbl\n", cpumask_pr_args(cs->effective_xcpus));
3684 break;
3685 case FILE_SUBPARTS_CPULIST:
3686 seq_printf(sf, "%*pbl\n", cpumask_pr_args(subpartitions_cpus));
3687 break;
3688 case FILE_ISOLATED_CPULIST:
3689 seq_printf(sf, "%*pbl\n", cpumask_pr_args(isolated_cpus));
3690 break;
3691 default:
3692 ret = -EINVAL;
3693 }
3694
3695 spin_unlock_irq(&callback_lock);
3696 return ret;
3697}
3698
3699static u64 cpuset_read_u64(struct cgroup_subsys_state *css, struct cftype *cft)
3700{
3701 struct cpuset *cs = css_cs(css);
3702 cpuset_filetype_t type = cft->private;
3703 switch (type) {
3704 case FILE_CPU_EXCLUSIVE:
3705 return is_cpu_exclusive(cs);
3706 case FILE_MEM_EXCLUSIVE:
3707 return is_mem_exclusive(cs);
3708 case FILE_MEM_HARDWALL:
3709 return is_mem_hardwall(cs);
3710 case FILE_SCHED_LOAD_BALANCE:
3711 return is_sched_load_balance(cs);
3712 case FILE_MEMORY_MIGRATE:
3713 return is_memory_migrate(cs);
3714 case FILE_MEMORY_PRESSURE_ENABLED:
3715 return cpuset_memory_pressure_enabled;
3716 case FILE_MEMORY_PRESSURE:
3717 return fmeter_getrate(&cs->fmeter);
3718 case FILE_SPREAD_PAGE:
3719 return is_spread_page(cs);
3720 case FILE_SPREAD_SLAB:
3721 return is_spread_slab(cs);
3722 default:
3723 BUG();
3724 }
3725
3726 /* Unreachable but makes gcc happy */
3727 return 0;
3728}
3729
3730static s64 cpuset_read_s64(struct cgroup_subsys_state *css, struct cftype *cft)
3731{
3732 struct cpuset *cs = css_cs(css);
3733 cpuset_filetype_t type = cft->private;
3734 switch (type) {
3735 case FILE_SCHED_RELAX_DOMAIN_LEVEL:
3736 return cs->relax_domain_level;
3737 default:
3738 BUG();
3739 }
3740
3741 /* Unreachable but makes gcc happy */
3742 return 0;
3743}
3744
3745static int sched_partition_show(struct seq_file *seq, void *v)
3746{
3747 struct cpuset *cs = css_cs(seq_css(seq));
3748 const char *err, *type = NULL;
3749
3750 switch (cs->partition_root_state) {
3751 case PRS_ROOT:
3752 seq_puts(seq, "root\n");
3753 break;
3754 case PRS_ISOLATED:
3755 seq_puts(seq, "isolated\n");
3756 break;
3757 case PRS_MEMBER:
3758 seq_puts(seq, "member\n");
3759 break;
3760 case PRS_INVALID_ROOT:
3761 type = "root";
3762 fallthrough;
3763 case PRS_INVALID_ISOLATED:
3764 if (!type)
3765 type = "isolated";
3766 err = perr_strings[READ_ONCE(cs->prs_err)];
3767 if (err)
3768 seq_printf(seq, "%s invalid (%s)\n", type, err);
3769 else
3770 seq_printf(seq, "%s invalid\n", type);
3771 break;
3772 }
3773 return 0;
3774}
3775
3776static ssize_t sched_partition_write(struct kernfs_open_file *of, char *buf,
3777 size_t nbytes, loff_t off)
3778{
3779 struct cpuset *cs = css_cs(of_css(of));
3780 int val;
3781 int retval = -ENODEV;
3782
3783 buf = strstrip(buf);
3784
3785 /*
3786 * Convert "root" to ENABLED, and convert "member" to DISABLED.
3787 */
3788 if (!strcmp(buf, "root"))
3789 val = PRS_ROOT;
3790 else if (!strcmp(buf, "member"))
3791 val = PRS_MEMBER;
3792 else if (!strcmp(buf, "isolated"))
3793 val = PRS_ISOLATED;
3794 else
3795 return -EINVAL;
3796
3797 css_get(&cs->css);
3798 cpus_read_lock();
3799 mutex_lock(&cpuset_mutex);
3800 if (!is_cpuset_online(cs))
3801 goto out_unlock;
3802
3803 retval = update_prstate(cs, val);
3804out_unlock:
3805 mutex_unlock(&cpuset_mutex);
3806 cpus_read_unlock();
3807 css_put(&cs->css);
3808 return retval ?: nbytes;
3809}
3810
3811/*
3812 * for the common functions, 'private' gives the type of file
3813 */
3814
3815static struct cftype legacy_files[] = {
3816 {
3817 .name = "cpus",
3818 .seq_show = cpuset_common_seq_show,
3819 .write = cpuset_write_resmask,
3820 .max_write_len = (100U + 6 * NR_CPUS),
3821 .private = FILE_CPULIST,
3822 },
3823
3824 {
3825 .name = "mems",
3826 .seq_show = cpuset_common_seq_show,
3827 .write = cpuset_write_resmask,
3828 .max_write_len = (100U + 6 * MAX_NUMNODES),
3829 .private = FILE_MEMLIST,
3830 },
3831
3832 {
3833 .name = "effective_cpus",
3834 .seq_show = cpuset_common_seq_show,
3835 .private = FILE_EFFECTIVE_CPULIST,
3836 },
3837
3838 {
3839 .name = "effective_mems",
3840 .seq_show = cpuset_common_seq_show,
3841 .private = FILE_EFFECTIVE_MEMLIST,
3842 },
3843
3844 {
3845 .name = "cpu_exclusive",
3846 .read_u64 = cpuset_read_u64,
3847 .write_u64 = cpuset_write_u64,
3848 .private = FILE_CPU_EXCLUSIVE,
3849 },
3850
3851 {
3852 .name = "mem_exclusive",
3853 .read_u64 = cpuset_read_u64,
3854 .write_u64 = cpuset_write_u64,
3855 .private = FILE_MEM_EXCLUSIVE,
3856 },
3857
3858 {
3859 .name = "mem_hardwall",
3860 .read_u64 = cpuset_read_u64,
3861 .write_u64 = cpuset_write_u64,
3862 .private = FILE_MEM_HARDWALL,
3863 },
3864
3865 {
3866 .name = "sched_load_balance",
3867 .read_u64 = cpuset_read_u64,
3868 .write_u64 = cpuset_write_u64,
3869 .private = FILE_SCHED_LOAD_BALANCE,
3870 },
3871
3872 {
3873 .name = "sched_relax_domain_level",
3874 .read_s64 = cpuset_read_s64,
3875 .write_s64 = cpuset_write_s64,
3876 .private = FILE_SCHED_RELAX_DOMAIN_LEVEL,
3877 },
3878
3879 {
3880 .name = "memory_migrate",
3881 .read_u64 = cpuset_read_u64,
3882 .write_u64 = cpuset_write_u64,
3883 .private = FILE_MEMORY_MIGRATE,
3884 },
3885
3886 {
3887 .name = "memory_pressure",
3888 .read_u64 = cpuset_read_u64,
3889 .private = FILE_MEMORY_PRESSURE,
3890 },
3891
3892 {
3893 .name = "memory_spread_page",
3894 .read_u64 = cpuset_read_u64,
3895 .write_u64 = cpuset_write_u64,
3896 .private = FILE_SPREAD_PAGE,
3897 },
3898
3899 {
3900 .name = "memory_spread_slab",
3901 .read_u64 = cpuset_read_u64,
3902 .write_u64 = cpuset_write_u64,
3903 .private = FILE_SPREAD_SLAB,
3904 },
3905
3906 {
3907 .name = "memory_pressure_enabled",
3908 .flags = CFTYPE_ONLY_ON_ROOT,
3909 .read_u64 = cpuset_read_u64,
3910 .write_u64 = cpuset_write_u64,
3911 .private = FILE_MEMORY_PRESSURE_ENABLED,
3912 },
3913
3914 { } /* terminate */
3915};
3916
3917/*
3918 * This is currently a minimal set for the default hierarchy. It can be
3919 * expanded later on by migrating more features and control files from v1.
3920 */
3921static struct cftype dfl_files[] = {
3922 {
3923 .name = "cpus",
3924 .seq_show = cpuset_common_seq_show,
3925 .write = cpuset_write_resmask,
3926 .max_write_len = (100U + 6 * NR_CPUS),
3927 .private = FILE_CPULIST,
3928 .flags = CFTYPE_NOT_ON_ROOT,
3929 },
3930
3931 {
3932 .name = "mems",
3933 .seq_show = cpuset_common_seq_show,
3934 .write = cpuset_write_resmask,
3935 .max_write_len = (100U + 6 * MAX_NUMNODES),
3936 .private = FILE_MEMLIST,
3937 .flags = CFTYPE_NOT_ON_ROOT,
3938 },
3939
3940 {
3941 .name = "cpus.effective",
3942 .seq_show = cpuset_common_seq_show,
3943 .private = FILE_EFFECTIVE_CPULIST,
3944 },
3945
3946 {
3947 .name = "mems.effective",
3948 .seq_show = cpuset_common_seq_show,
3949 .private = FILE_EFFECTIVE_MEMLIST,
3950 },
3951
3952 {
3953 .name = "cpus.partition",
3954 .seq_show = sched_partition_show,
3955 .write = sched_partition_write,
3956 .private = FILE_PARTITION_ROOT,
3957 .flags = CFTYPE_NOT_ON_ROOT,
3958 .file_offset = offsetof(struct cpuset, partition_file),
3959 },
3960
3961 {
3962 .name = "cpus.exclusive",
3963 .seq_show = cpuset_common_seq_show,
3964 .write = cpuset_write_resmask,
3965 .max_write_len = (100U + 6 * NR_CPUS),
3966 .private = FILE_EXCLUSIVE_CPULIST,
3967 .flags = CFTYPE_NOT_ON_ROOT,
3968 },
3969
3970 {
3971 .name = "cpus.exclusive.effective",
3972 .seq_show = cpuset_common_seq_show,
3973 .private = FILE_EFFECTIVE_XCPULIST,
3974 .flags = CFTYPE_NOT_ON_ROOT,
3975 },
3976
3977 {
3978 .name = "cpus.subpartitions",
3979 .seq_show = cpuset_common_seq_show,
3980 .private = FILE_SUBPARTS_CPULIST,
3981 .flags = CFTYPE_ONLY_ON_ROOT | CFTYPE_DEBUG,
3982 },
3983
3984 {
3985 .name = "cpus.isolated",
3986 .seq_show = cpuset_common_seq_show,
3987 .private = FILE_ISOLATED_CPULIST,
3988 .flags = CFTYPE_ONLY_ON_ROOT,
3989 },
3990
3991 { } /* terminate */
3992};
3993
3994
3995/**
3996 * cpuset_css_alloc - Allocate a cpuset css
3997 * @parent_css: Parent css of the control group that the new cpuset will be
3998 * part of
3999 * Return: cpuset css on success, -ENOMEM on failure.
4000 *
4001 * Allocate and initialize a new cpuset css, for non-NULL @parent_css, return
4002 * top cpuset css otherwise.
4003 */
4004static struct cgroup_subsys_state *
4005cpuset_css_alloc(struct cgroup_subsys_state *parent_css)
4006{
4007 struct cpuset *cs;
4008
4009 if (!parent_css)
4010 return &top_cpuset.css;
4011
4012 cs = kzalloc(sizeof(*cs), GFP_KERNEL);
4013 if (!cs)
4014 return ERR_PTR(-ENOMEM);
4015
4016 if (alloc_cpumasks(cs, NULL)) {
4017 kfree(cs);
4018 return ERR_PTR(-ENOMEM);
4019 }
4020
4021 __set_bit(CS_SCHED_LOAD_BALANCE, &cs->flags);
4022 nodes_clear(cs->mems_allowed);
4023 nodes_clear(cs->effective_mems);
4024 fmeter_init(&cs->fmeter);
4025 cs->relax_domain_level = -1;
4026 INIT_LIST_HEAD(&cs->remote_sibling);
4027
4028 /* Set CS_MEMORY_MIGRATE for default hierarchy */
4029 if (cgroup_subsys_on_dfl(cpuset_cgrp_subsys))
4030 __set_bit(CS_MEMORY_MIGRATE, &cs->flags);
4031
4032 return &cs->css;
4033}
4034
4035static int cpuset_css_online(struct cgroup_subsys_state *css)
4036{
4037 struct cpuset *cs = css_cs(css);
4038 struct cpuset *parent = parent_cs(cs);
4039 struct cpuset *tmp_cs;
4040 struct cgroup_subsys_state *pos_css;
4041
4042 if (!parent)
4043 return 0;
4044
4045 cpus_read_lock();
4046 mutex_lock(&cpuset_mutex);
4047
4048 set_bit(CS_ONLINE, &cs->flags);
4049 if (is_spread_page(parent))
4050 set_bit(CS_SPREAD_PAGE, &cs->flags);
4051 if (is_spread_slab(parent))
4052 set_bit(CS_SPREAD_SLAB, &cs->flags);
4053
4054 cpuset_inc();
4055
4056 spin_lock_irq(&callback_lock);
4057 if (is_in_v2_mode()) {
4058 cpumask_copy(cs->effective_cpus, parent->effective_cpus);
4059 cs->effective_mems = parent->effective_mems;
4060 cs->use_parent_ecpus = true;
4061 parent->child_ecpus_count++;
4062 /*
4063 * Clear CS_SCHED_LOAD_BALANCE if parent is isolated
4064 */
4065 if (!is_sched_load_balance(parent))
4066 clear_bit(CS_SCHED_LOAD_BALANCE, &cs->flags);
4067 }
4068
4069 /*
4070 * For v2, clear CS_SCHED_LOAD_BALANCE if parent is isolated
4071 */
4072 if (cgroup_subsys_on_dfl(cpuset_cgrp_subsys) &&
4073 !is_sched_load_balance(parent))
4074 clear_bit(CS_SCHED_LOAD_BALANCE, &cs->flags);
4075
4076 spin_unlock_irq(&callback_lock);
4077
4078 if (!test_bit(CGRP_CPUSET_CLONE_CHILDREN, &css->cgroup->flags))
4079 goto out_unlock;
4080
4081 /*
4082 * Clone @parent's configuration if CGRP_CPUSET_CLONE_CHILDREN is
4083 * set. This flag handling is implemented in cgroup core for
4084 * historical reasons - the flag may be specified during mount.
4085 *
4086 * Currently, if any sibling cpusets have exclusive cpus or mem, we
4087 * refuse to clone the configuration - thereby refusing the task to
4088 * be entered, and as a result refusing the sys_unshare() or
4089 * clone() which initiated it. If this becomes a problem for some
4090 * users who wish to allow that scenario, then this could be
4091 * changed to grant parent->cpus_allowed-sibling_cpus_exclusive
4092 * (and likewise for mems) to the new cgroup.
4093 */
4094 rcu_read_lock();
4095 cpuset_for_each_child(tmp_cs, pos_css, parent) {
4096 if (is_mem_exclusive(tmp_cs) || is_cpu_exclusive(tmp_cs)) {
4097 rcu_read_unlock();
4098 goto out_unlock;
4099 }
4100 }
4101 rcu_read_unlock();
4102
4103 spin_lock_irq(&callback_lock);
4104 cs->mems_allowed = parent->mems_allowed;
4105 cs->effective_mems = parent->mems_allowed;
4106 cpumask_copy(cs->cpus_allowed, parent->cpus_allowed);
4107 cpumask_copy(cs->effective_cpus, parent->cpus_allowed);
4108 spin_unlock_irq(&callback_lock);
4109out_unlock:
4110 mutex_unlock(&cpuset_mutex);
4111 cpus_read_unlock();
4112 return 0;
4113}
4114
4115/*
4116 * If the cpuset being removed has its flag 'sched_load_balance'
4117 * enabled, then simulate turning sched_load_balance off, which
4118 * will call rebuild_sched_domains_locked(). That is not needed
4119 * in the default hierarchy where only changes in partition
4120 * will cause repartitioning.
4121 *
4122 * If the cpuset has the 'sched.partition' flag enabled, simulate
4123 * turning 'sched.partition" off.
4124 */
4125
4126static void cpuset_css_offline(struct cgroup_subsys_state *css)
4127{
4128 struct cpuset *cs = css_cs(css);
4129
4130 cpus_read_lock();
4131 mutex_lock(&cpuset_mutex);
4132
4133 if (is_partition_valid(cs))
4134 update_prstate(cs, 0);
4135
4136 if (!cgroup_subsys_on_dfl(cpuset_cgrp_subsys) &&
4137 is_sched_load_balance(cs))
4138 update_flag(CS_SCHED_LOAD_BALANCE, cs, 0);
4139
4140 if (cs->use_parent_ecpus) {
4141 struct cpuset *parent = parent_cs(cs);
4142
4143 cs->use_parent_ecpus = false;
4144 parent->child_ecpus_count--;
4145 }
4146
4147 cpuset_dec();
4148 clear_bit(CS_ONLINE, &cs->flags);
4149
4150 mutex_unlock(&cpuset_mutex);
4151 cpus_read_unlock();
4152}
4153
4154static void cpuset_css_free(struct cgroup_subsys_state *css)
4155{
4156 struct cpuset *cs = css_cs(css);
4157
4158 free_cpuset(cs);
4159}
4160
4161static void cpuset_bind(struct cgroup_subsys_state *root_css)
4162{
4163 mutex_lock(&cpuset_mutex);
4164 spin_lock_irq(&callback_lock);
4165
4166 if (is_in_v2_mode()) {
4167 cpumask_copy(top_cpuset.cpus_allowed, cpu_possible_mask);
4168 cpumask_copy(top_cpuset.effective_xcpus, cpu_possible_mask);
4169 top_cpuset.mems_allowed = node_possible_map;
4170 } else {
4171 cpumask_copy(top_cpuset.cpus_allowed,
4172 top_cpuset.effective_cpus);
4173 top_cpuset.mems_allowed = top_cpuset.effective_mems;
4174 }
4175
4176 spin_unlock_irq(&callback_lock);
4177 mutex_unlock(&cpuset_mutex);
4178}
4179
4180/*
4181 * In case the child is cloned into a cpuset different from its parent,
4182 * additional checks are done to see if the move is allowed.
4183 */
4184static int cpuset_can_fork(struct task_struct *task, struct css_set *cset)
4185{
4186 struct cpuset *cs = css_cs(cset->subsys[cpuset_cgrp_id]);
4187 bool same_cs;
4188 int ret;
4189
4190 rcu_read_lock();
4191 same_cs = (cs == task_cs(current));
4192 rcu_read_unlock();
4193
4194 if (same_cs)
4195 return 0;
4196
4197 lockdep_assert_held(&cgroup_mutex);
4198 mutex_lock(&cpuset_mutex);
4199
4200 /* Check to see if task is allowed in the cpuset */
4201 ret = cpuset_can_attach_check(cs);
4202 if (ret)
4203 goto out_unlock;
4204
4205 ret = task_can_attach(task);
4206 if (ret)
4207 goto out_unlock;
4208
4209 ret = security_task_setscheduler(task);
4210 if (ret)
4211 goto out_unlock;
4212
4213 /*
4214 * Mark attach is in progress. This makes validate_change() fail
4215 * changes which zero cpus/mems_allowed.
4216 */
4217 cs->attach_in_progress++;
4218out_unlock:
4219 mutex_unlock(&cpuset_mutex);
4220 return ret;
4221}
4222
4223static void cpuset_cancel_fork(struct task_struct *task, struct css_set *cset)
4224{
4225 struct cpuset *cs = css_cs(cset->subsys[cpuset_cgrp_id]);
4226 bool same_cs;
4227
4228 rcu_read_lock();
4229 same_cs = (cs == task_cs(current));
4230 rcu_read_unlock();
4231
4232 if (same_cs)
4233 return;
4234
4235 mutex_lock(&cpuset_mutex);
4236 cs->attach_in_progress--;
4237 if (!cs->attach_in_progress)
4238 wake_up(&cpuset_attach_wq);
4239 mutex_unlock(&cpuset_mutex);
4240}
4241
4242/*
4243 * Make sure the new task conform to the current state of its parent,
4244 * which could have been changed by cpuset just after it inherits the
4245 * state from the parent and before it sits on the cgroup's task list.
4246 */
4247static void cpuset_fork(struct task_struct *task)
4248{
4249 struct cpuset *cs;
4250 bool same_cs;
4251
4252 rcu_read_lock();
4253 cs = task_cs(task);
4254 same_cs = (cs == task_cs(current));
4255 rcu_read_unlock();
4256
4257 if (same_cs) {
4258 if (cs == &top_cpuset)
4259 return;
4260
4261 set_cpus_allowed_ptr(task, current->cpus_ptr);
4262 task->mems_allowed = current->mems_allowed;
4263 return;
4264 }
4265
4266 /* CLONE_INTO_CGROUP */
4267 mutex_lock(&cpuset_mutex);
4268 guarantee_online_mems(cs, &cpuset_attach_nodemask_to);
4269 cpuset_attach_task(cs, task);
4270
4271 cs->attach_in_progress--;
4272 if (!cs->attach_in_progress)
4273 wake_up(&cpuset_attach_wq);
4274
4275 mutex_unlock(&cpuset_mutex);
4276}
4277
4278struct cgroup_subsys cpuset_cgrp_subsys = {
4279 .css_alloc = cpuset_css_alloc,
4280 .css_online = cpuset_css_online,
4281 .css_offline = cpuset_css_offline,
4282 .css_free = cpuset_css_free,
4283 .can_attach = cpuset_can_attach,
4284 .cancel_attach = cpuset_cancel_attach,
4285 .attach = cpuset_attach,
4286 .post_attach = cpuset_post_attach,
4287 .bind = cpuset_bind,
4288 .can_fork = cpuset_can_fork,
4289 .cancel_fork = cpuset_cancel_fork,
4290 .fork = cpuset_fork,
4291 .legacy_cftypes = legacy_files,
4292 .dfl_cftypes = dfl_files,
4293 .early_init = true,
4294 .threaded = true,
4295};
4296
4297/**
4298 * cpuset_init - initialize cpusets at system boot
4299 *
4300 * Description: Initialize top_cpuset
4301 **/
4302
4303int __init cpuset_init(void)
4304{
4305 BUG_ON(!alloc_cpumask_var(&top_cpuset.cpus_allowed, GFP_KERNEL));
4306 BUG_ON(!alloc_cpumask_var(&top_cpuset.effective_cpus, GFP_KERNEL));
4307 BUG_ON(!alloc_cpumask_var(&top_cpuset.effective_xcpus, GFP_KERNEL));
4308 BUG_ON(!alloc_cpumask_var(&top_cpuset.exclusive_cpus, GFP_KERNEL));
4309 BUG_ON(!zalloc_cpumask_var(&subpartitions_cpus, GFP_KERNEL));
4310 BUG_ON(!zalloc_cpumask_var(&isolated_cpus, GFP_KERNEL));
4311
4312 cpumask_setall(top_cpuset.cpus_allowed);
4313 nodes_setall(top_cpuset.mems_allowed);
4314 cpumask_setall(top_cpuset.effective_cpus);
4315 cpumask_setall(top_cpuset.effective_xcpus);
4316 cpumask_setall(top_cpuset.exclusive_cpus);
4317 nodes_setall(top_cpuset.effective_mems);
4318
4319 fmeter_init(&top_cpuset.fmeter);
4320 set_bit(CS_SCHED_LOAD_BALANCE, &top_cpuset.flags);
4321 top_cpuset.relax_domain_level = -1;
4322 INIT_LIST_HEAD(&remote_children);
4323
4324 BUG_ON(!alloc_cpumask_var(&cpus_attach, GFP_KERNEL));
4325
4326 return 0;
4327}
4328
4329/*
4330 * If CPU and/or memory hotplug handlers, below, unplug any CPUs
4331 * or memory nodes, we need to walk over the cpuset hierarchy,
4332 * removing that CPU or node from all cpusets. If this removes the
4333 * last CPU or node from a cpuset, then move the tasks in the empty
4334 * cpuset to its next-highest non-empty parent.
4335 */
4336static void remove_tasks_in_empty_cpuset(struct cpuset *cs)
4337{
4338 struct cpuset *parent;
4339
4340 /*
4341 * Find its next-highest non-empty parent, (top cpuset
4342 * has online cpus, so can't be empty).
4343 */
4344 parent = parent_cs(cs);
4345 while (cpumask_empty(parent->cpus_allowed) ||
4346 nodes_empty(parent->mems_allowed))
4347 parent = parent_cs(parent);
4348
4349 if (cgroup_transfer_tasks(parent->css.cgroup, cs->css.cgroup)) {
4350 pr_err("cpuset: failed to transfer tasks out of empty cpuset ");
4351 pr_cont_cgroup_name(cs->css.cgroup);
4352 pr_cont("\n");
4353 }
4354}
4355
4356static void
4357hotplug_update_tasks_legacy(struct cpuset *cs,
4358 struct cpumask *new_cpus, nodemask_t *new_mems,
4359 bool cpus_updated, bool mems_updated)
4360{
4361 bool is_empty;
4362
4363 spin_lock_irq(&callback_lock);
4364 cpumask_copy(cs->cpus_allowed, new_cpus);
4365 cpumask_copy(cs->effective_cpus, new_cpus);
4366 cs->mems_allowed = *new_mems;
4367 cs->effective_mems = *new_mems;
4368 spin_unlock_irq(&callback_lock);
4369
4370 /*
4371 * Don't call update_tasks_cpumask() if the cpuset becomes empty,
4372 * as the tasks will be migrated to an ancestor.
4373 */
4374 if (cpus_updated && !cpumask_empty(cs->cpus_allowed))
4375 update_tasks_cpumask(cs, new_cpus);
4376 if (mems_updated && !nodes_empty(cs->mems_allowed))
4377 update_tasks_nodemask(cs);
4378
4379 is_empty = cpumask_empty(cs->cpus_allowed) ||
4380 nodes_empty(cs->mems_allowed);
4381
4382 /*
4383 * Move tasks to the nearest ancestor with execution resources,
4384 * This is full cgroup operation which will also call back into
4385 * cpuset. Should be done outside any lock.
4386 */
4387 if (is_empty) {
4388 mutex_unlock(&cpuset_mutex);
4389 remove_tasks_in_empty_cpuset(cs);
4390 mutex_lock(&cpuset_mutex);
4391 }
4392}
4393
4394static void
4395hotplug_update_tasks(struct cpuset *cs,
4396 struct cpumask *new_cpus, nodemask_t *new_mems,
4397 bool cpus_updated, bool mems_updated)
4398{
4399 /* A partition root is allowed to have empty effective cpus */
4400 if (cpumask_empty(new_cpus) && !is_partition_valid(cs))
4401 cpumask_copy(new_cpus, parent_cs(cs)->effective_cpus);
4402 if (nodes_empty(*new_mems))
4403 *new_mems = parent_cs(cs)->effective_mems;
4404
4405 spin_lock_irq(&callback_lock);
4406 cpumask_copy(cs->effective_cpus, new_cpus);
4407 cs->effective_mems = *new_mems;
4408 spin_unlock_irq(&callback_lock);
4409
4410 if (cpus_updated)
4411 update_tasks_cpumask(cs, new_cpus);
4412 if (mems_updated)
4413 update_tasks_nodemask(cs);
4414}
4415
4416static bool force_rebuild;
4417
4418void cpuset_force_rebuild(void)
4419{
4420 force_rebuild = true;
4421}
4422
4423/*
4424 * Attempt to acquire a cpus_read_lock while a hotplug operation may be in
4425 * progress.
4426 * Return: true if successful, false otherwise
4427 *
4428 * To avoid circular lock dependency between cpuset_mutex and cpus_read_lock,
4429 * cpus_read_trylock() is used here to acquire the lock.
4430 */
4431static bool cpuset_hotplug_cpus_read_trylock(void)
4432{
4433 int retries = 0;
4434
4435 while (!cpus_read_trylock()) {
4436 /*
4437 * CPU hotplug still in progress. Retry 5 times
4438 * with a 10ms wait before bailing out.
4439 */
4440 if (++retries > 5)
4441 return false;
4442 msleep(10);
4443 }
4444 return true;
4445}
4446
4447/**
4448 * cpuset_hotplug_update_tasks - update tasks in a cpuset for hotunplug
4449 * @cs: cpuset in interest
4450 * @tmp: the tmpmasks structure pointer
4451 *
4452 * Compare @cs's cpu and mem masks against top_cpuset and if some have gone
4453 * offline, update @cs accordingly. If @cs ends up with no CPU or memory,
4454 * all its tasks are moved to the nearest ancestor with both resources.
4455 */
4456static void cpuset_hotplug_update_tasks(struct cpuset *cs, struct tmpmasks *tmp)
4457{
4458 static cpumask_t new_cpus;
4459 static nodemask_t new_mems;
4460 bool cpus_updated;
4461 bool mems_updated;
4462 bool remote;
4463 int partcmd = -1;
4464 struct cpuset *parent;
4465retry:
4466 wait_event(cpuset_attach_wq, cs->attach_in_progress == 0);
4467
4468 mutex_lock(&cpuset_mutex);
4469
4470 /*
4471 * We have raced with task attaching. We wait until attaching
4472 * is finished, so we won't attach a task to an empty cpuset.
4473 */
4474 if (cs->attach_in_progress) {
4475 mutex_unlock(&cpuset_mutex);
4476 goto retry;
4477 }
4478
4479 parent = parent_cs(cs);
4480 compute_effective_cpumask(&new_cpus, cs, parent);
4481 nodes_and(new_mems, cs->mems_allowed, parent->effective_mems);
4482
4483 if (!tmp || !cs->partition_root_state)
4484 goto update_tasks;
4485
4486 /*
4487 * Compute effective_cpus for valid partition root, may invalidate
4488 * child partition roots if necessary.
4489 */
4490 remote = is_remote_partition(cs);
4491 if (remote || (is_partition_valid(cs) && is_partition_valid(parent)))
4492 compute_partition_effective_cpumask(cs, &new_cpus);
4493
4494 if (remote && cpumask_empty(&new_cpus) &&
4495 partition_is_populated(cs, NULL) &&
4496 cpuset_hotplug_cpus_read_trylock()) {
4497 remote_partition_disable(cs, tmp);
4498 compute_effective_cpumask(&new_cpus, cs, parent);
4499 remote = false;
4500 cpuset_force_rebuild();
4501 cpus_read_unlock();
4502 }
4503
4504 /*
4505 * Force the partition to become invalid if either one of
4506 * the following conditions hold:
4507 * 1) empty effective cpus but not valid empty partition.
4508 * 2) parent is invalid or doesn't grant any cpus to child
4509 * partitions.
4510 */
4511 if (is_local_partition(cs) && (!is_partition_valid(parent) ||
4512 tasks_nocpu_error(parent, cs, &new_cpus)))
4513 partcmd = partcmd_invalidate;
4514 /*
4515 * On the other hand, an invalid partition root may be transitioned
4516 * back to a regular one.
4517 */
4518 else if (is_partition_valid(parent) && is_partition_invalid(cs))
4519 partcmd = partcmd_update;
4520
4521 /*
4522 * cpus_read_lock needs to be held before calling
4523 * update_parent_effective_cpumask(). To avoid circular lock
4524 * dependency between cpuset_mutex and cpus_read_lock,
4525 * cpus_read_trylock() is used here to acquire the lock.
4526 */
4527 if (partcmd >= 0) {
4528 if (!cpuset_hotplug_cpus_read_trylock())
4529 goto update_tasks;
4530
4531 update_parent_effective_cpumask(cs, partcmd, NULL, tmp);
4532 cpus_read_unlock();
4533 if ((partcmd == partcmd_invalidate) || is_partition_valid(cs)) {
4534 compute_partition_effective_cpumask(cs, &new_cpus);
4535 cpuset_force_rebuild();
4536 }
4537 }
4538
4539update_tasks:
4540 cpus_updated = !cpumask_equal(&new_cpus, cs->effective_cpus);
4541 mems_updated = !nodes_equal(new_mems, cs->effective_mems);
4542 if (!cpus_updated && !mems_updated)
4543 goto unlock; /* Hotplug doesn't affect this cpuset */
4544
4545 if (mems_updated)
4546 check_insane_mems_config(&new_mems);
4547
4548 if (is_in_v2_mode())
4549 hotplug_update_tasks(cs, &new_cpus, &new_mems,
4550 cpus_updated, mems_updated);
4551 else
4552 hotplug_update_tasks_legacy(cs, &new_cpus, &new_mems,
4553 cpus_updated, mems_updated);
4554
4555unlock:
4556 mutex_unlock(&cpuset_mutex);
4557}
4558
4559/**
4560 * cpuset_hotplug_workfn - handle CPU/memory hotunplug for a cpuset
4561 * @work: unused
4562 *
4563 * This function is called after either CPU or memory configuration has
4564 * changed and updates cpuset accordingly. The top_cpuset is always
4565 * synchronized to cpu_active_mask and N_MEMORY, which is necessary in
4566 * order to make cpusets transparent (of no affect) on systems that are
4567 * actively using CPU hotplug but making no active use of cpusets.
4568 *
4569 * Non-root cpusets are only affected by offlining. If any CPUs or memory
4570 * nodes have been taken down, cpuset_hotplug_update_tasks() is invoked on
4571 * all descendants.
4572 *
4573 * Note that CPU offlining during suspend is ignored. We don't modify
4574 * cpusets across suspend/resume cycles at all.
4575 */
4576static void cpuset_hotplug_workfn(struct work_struct *work)
4577{
4578 static cpumask_t new_cpus;
4579 static nodemask_t new_mems;
4580 bool cpus_updated, mems_updated;
4581 bool on_dfl = is_in_v2_mode();
4582 struct tmpmasks tmp, *ptmp = NULL;
4583
4584 if (on_dfl && !alloc_cpumasks(NULL, &tmp))
4585 ptmp = &tmp;
4586
4587 mutex_lock(&cpuset_mutex);
4588
4589 /* fetch the available cpus/mems and find out which changed how */
4590 cpumask_copy(&new_cpus, cpu_active_mask);
4591 new_mems = node_states[N_MEMORY];
4592
4593 /*
4594 * If subpartitions_cpus is populated, it is likely that the check
4595 * below will produce a false positive on cpus_updated when the cpu
4596 * list isn't changed. It is extra work, but it is better to be safe.
4597 */
4598 cpus_updated = !cpumask_equal(top_cpuset.effective_cpus, &new_cpus) ||
4599 !cpumask_empty(subpartitions_cpus);
4600 mems_updated = !nodes_equal(top_cpuset.effective_mems, new_mems);
4601
4602 /*
4603 * In the rare case that hotplug removes all the cpus in
4604 * subpartitions_cpus, we assumed that cpus are updated.
4605 */
4606 if (!cpus_updated && top_cpuset.nr_subparts)
4607 cpus_updated = true;
4608
4609 /* For v1, synchronize cpus_allowed to cpu_active_mask */
4610 if (cpus_updated) {
4611 spin_lock_irq(&callback_lock);
4612 if (!on_dfl)
4613 cpumask_copy(top_cpuset.cpus_allowed, &new_cpus);
4614 /*
4615 * Make sure that CPUs allocated to child partitions
4616 * do not show up in effective_cpus. If no CPU is left,
4617 * we clear the subpartitions_cpus & let the child partitions
4618 * fight for the CPUs again.
4619 */
4620 if (!cpumask_empty(subpartitions_cpus)) {
4621 if (cpumask_subset(&new_cpus, subpartitions_cpus)) {
4622 top_cpuset.nr_subparts = 0;
4623 cpumask_clear(subpartitions_cpus);
4624 } else {
4625 cpumask_andnot(&new_cpus, &new_cpus,
4626 subpartitions_cpus);
4627 }
4628 }
4629 cpumask_copy(top_cpuset.effective_cpus, &new_cpus);
4630 spin_unlock_irq(&callback_lock);
4631 /* we don't mess with cpumasks of tasks in top_cpuset */
4632 }
4633
4634 /* synchronize mems_allowed to N_MEMORY */
4635 if (mems_updated) {
4636 spin_lock_irq(&callback_lock);
4637 if (!on_dfl)
4638 top_cpuset.mems_allowed = new_mems;
4639 top_cpuset.effective_mems = new_mems;
4640 spin_unlock_irq(&callback_lock);
4641 update_tasks_nodemask(&top_cpuset);
4642 }
4643
4644 mutex_unlock(&cpuset_mutex);
4645
4646 /* if cpus or mems changed, we need to propagate to descendants */
4647 if (cpus_updated || mems_updated) {
4648 struct cpuset *cs;
4649 struct cgroup_subsys_state *pos_css;
4650
4651 rcu_read_lock();
4652 cpuset_for_each_descendant_pre(cs, pos_css, &top_cpuset) {
4653 if (cs == &top_cpuset || !css_tryget_online(&cs->css))
4654 continue;
4655 rcu_read_unlock();
4656
4657 cpuset_hotplug_update_tasks(cs, ptmp);
4658
4659 rcu_read_lock();
4660 css_put(&cs->css);
4661 }
4662 rcu_read_unlock();
4663 }
4664
4665 /* rebuild sched domains if cpus_allowed has changed */
4666 if (cpus_updated || force_rebuild) {
4667 force_rebuild = false;
4668 rebuild_sched_domains();
4669 }
4670
4671 free_cpumasks(NULL, ptmp);
4672}
4673
4674void cpuset_update_active_cpus(void)
4675{
4676 /*
4677 * We're inside cpu hotplug critical region which usually nests
4678 * inside cgroup synchronization. Bounce actual hotplug processing
4679 * to a work item to avoid reverse locking order.
4680 */
4681 schedule_work(&cpuset_hotplug_work);
4682}
4683
4684void cpuset_wait_for_hotplug(void)
4685{
4686 flush_work(&cpuset_hotplug_work);
4687}
4688
4689/*
4690 * Keep top_cpuset.mems_allowed tracking node_states[N_MEMORY].
4691 * Call this routine anytime after node_states[N_MEMORY] changes.
4692 * See cpuset_update_active_cpus() for CPU hotplug handling.
4693 */
4694static int cpuset_track_online_nodes(struct notifier_block *self,
4695 unsigned long action, void *arg)
4696{
4697 schedule_work(&cpuset_hotplug_work);
4698 return NOTIFY_OK;
4699}
4700
4701/**
4702 * cpuset_init_smp - initialize cpus_allowed
4703 *
4704 * Description: Finish top cpuset after cpu, node maps are initialized
4705 */
4706void __init cpuset_init_smp(void)
4707{
4708 /*
4709 * cpus_allowd/mems_allowed set to v2 values in the initial
4710 * cpuset_bind() call will be reset to v1 values in another
4711 * cpuset_bind() call when v1 cpuset is mounted.
4712 */
4713 top_cpuset.old_mems_allowed = top_cpuset.mems_allowed;
4714
4715 cpumask_copy(top_cpuset.effective_cpus, cpu_active_mask);
4716 top_cpuset.effective_mems = node_states[N_MEMORY];
4717
4718 hotplug_memory_notifier(cpuset_track_online_nodes, CPUSET_CALLBACK_PRI);
4719
4720 cpuset_migrate_mm_wq = alloc_ordered_workqueue("cpuset_migrate_mm", 0);
4721 BUG_ON(!cpuset_migrate_mm_wq);
4722}
4723
4724/**
4725 * cpuset_cpus_allowed - return cpus_allowed mask from a tasks cpuset.
4726 * @tsk: pointer to task_struct from which to obtain cpuset->cpus_allowed.
4727 * @pmask: pointer to struct cpumask variable to receive cpus_allowed set.
4728 *
4729 * Description: Returns the cpumask_var_t cpus_allowed of the cpuset
4730 * attached to the specified @tsk. Guaranteed to return some non-empty
4731 * subset of cpu_online_mask, even if this means going outside the
4732 * tasks cpuset, except when the task is in the top cpuset.
4733 **/
4734
4735void cpuset_cpus_allowed(struct task_struct *tsk, struct cpumask *pmask)
4736{
4737 unsigned long flags;
4738 struct cpuset *cs;
4739
4740 spin_lock_irqsave(&callback_lock, flags);
4741 rcu_read_lock();
4742
4743 cs = task_cs(tsk);
4744 if (cs != &top_cpuset)
4745 guarantee_online_cpus(tsk, pmask);
4746 /*
4747 * Tasks in the top cpuset won't get update to their cpumasks
4748 * when a hotplug online/offline event happens. So we include all
4749 * offline cpus in the allowed cpu list.
4750 */
4751 if ((cs == &top_cpuset) || cpumask_empty(pmask)) {
4752 const struct cpumask *possible_mask = task_cpu_possible_mask(tsk);
4753
4754 /*
4755 * We first exclude cpus allocated to partitions. If there is no
4756 * allowable online cpu left, we fall back to all possible cpus.
4757 */
4758 cpumask_andnot(pmask, possible_mask, subpartitions_cpus);
4759 if (!cpumask_intersects(pmask, cpu_online_mask))
4760 cpumask_copy(pmask, possible_mask);
4761 }
4762
4763 rcu_read_unlock();
4764 spin_unlock_irqrestore(&callback_lock, flags);
4765}
4766
4767/**
4768 * cpuset_cpus_allowed_fallback - final fallback before complete catastrophe.
4769 * @tsk: pointer to task_struct with which the scheduler is struggling
4770 *
4771 * Description: In the case that the scheduler cannot find an allowed cpu in
4772 * tsk->cpus_allowed, we fall back to task_cs(tsk)->cpus_allowed. In legacy
4773 * mode however, this value is the same as task_cs(tsk)->effective_cpus,
4774 * which will not contain a sane cpumask during cases such as cpu hotplugging.
4775 * This is the absolute last resort for the scheduler and it is only used if
4776 * _every_ other avenue has been traveled.
4777 *
4778 * Returns true if the affinity of @tsk was changed, false otherwise.
4779 **/
4780
4781bool cpuset_cpus_allowed_fallback(struct task_struct *tsk)
4782{
4783 const struct cpumask *possible_mask = task_cpu_possible_mask(tsk);
4784 const struct cpumask *cs_mask;
4785 bool changed = false;
4786
4787 rcu_read_lock();
4788 cs_mask = task_cs(tsk)->cpus_allowed;
4789 if (is_in_v2_mode() && cpumask_subset(cs_mask, possible_mask)) {
4790 do_set_cpus_allowed(tsk, cs_mask);
4791 changed = true;
4792 }
4793 rcu_read_unlock();
4794
4795 /*
4796 * We own tsk->cpus_allowed, nobody can change it under us.
4797 *
4798 * But we used cs && cs->cpus_allowed lockless and thus can
4799 * race with cgroup_attach_task() or update_cpumask() and get
4800 * the wrong tsk->cpus_allowed. However, both cases imply the
4801 * subsequent cpuset_change_cpumask()->set_cpus_allowed_ptr()
4802 * which takes task_rq_lock().
4803 *
4804 * If we are called after it dropped the lock we must see all
4805 * changes in tsk_cs()->cpus_allowed. Otherwise we can temporary
4806 * set any mask even if it is not right from task_cs() pov,
4807 * the pending set_cpus_allowed_ptr() will fix things.
4808 *
4809 * select_fallback_rq() will fix things ups and set cpu_possible_mask
4810 * if required.
4811 */
4812 return changed;
4813}
4814
4815void __init cpuset_init_current_mems_allowed(void)
4816{
4817 nodes_setall(current->mems_allowed);
4818}
4819
4820/**
4821 * cpuset_mems_allowed - return mems_allowed mask from a tasks cpuset.
4822 * @tsk: pointer to task_struct from which to obtain cpuset->mems_allowed.
4823 *
4824 * Description: Returns the nodemask_t mems_allowed of the cpuset
4825 * attached to the specified @tsk. Guaranteed to return some non-empty
4826 * subset of node_states[N_MEMORY], even if this means going outside the
4827 * tasks cpuset.
4828 **/
4829
4830nodemask_t cpuset_mems_allowed(struct task_struct *tsk)
4831{
4832 nodemask_t mask;
4833 unsigned long flags;
4834
4835 spin_lock_irqsave(&callback_lock, flags);
4836 rcu_read_lock();
4837 guarantee_online_mems(task_cs(tsk), &mask);
4838 rcu_read_unlock();
4839 spin_unlock_irqrestore(&callback_lock, flags);
4840
4841 return mask;
4842}
4843
4844/**
4845 * cpuset_nodemask_valid_mems_allowed - check nodemask vs. current mems_allowed
4846 * @nodemask: the nodemask to be checked
4847 *
4848 * Are any of the nodes in the nodemask allowed in current->mems_allowed?
4849 */
4850int cpuset_nodemask_valid_mems_allowed(nodemask_t *nodemask)
4851{
4852 return nodes_intersects(*nodemask, current->mems_allowed);
4853}
4854
4855/*
4856 * nearest_hardwall_ancestor() - Returns the nearest mem_exclusive or
4857 * mem_hardwall ancestor to the specified cpuset. Call holding
4858 * callback_lock. If no ancestor is mem_exclusive or mem_hardwall
4859 * (an unusual configuration), then returns the root cpuset.
4860 */
4861static struct cpuset *nearest_hardwall_ancestor(struct cpuset *cs)
4862{
4863 while (!(is_mem_exclusive(cs) || is_mem_hardwall(cs)) && parent_cs(cs))
4864 cs = parent_cs(cs);
4865 return cs;
4866}
4867
4868/*
4869 * cpuset_node_allowed - Can we allocate on a memory node?
4870 * @node: is this an allowed node?
4871 * @gfp_mask: memory allocation flags
4872 *
4873 * If we're in interrupt, yes, we can always allocate. If @node is set in
4874 * current's mems_allowed, yes. If it's not a __GFP_HARDWALL request and this
4875 * node is set in the nearest hardwalled cpuset ancestor to current's cpuset,
4876 * yes. If current has access to memory reserves as an oom victim, yes.
4877 * Otherwise, no.
4878 *
4879 * GFP_USER allocations are marked with the __GFP_HARDWALL bit,
4880 * and do not allow allocations outside the current tasks cpuset
4881 * unless the task has been OOM killed.
4882 * GFP_KERNEL allocations are not so marked, so can escape to the
4883 * nearest enclosing hardwalled ancestor cpuset.
4884 *
4885 * Scanning up parent cpusets requires callback_lock. The
4886 * __alloc_pages() routine only calls here with __GFP_HARDWALL bit
4887 * _not_ set if it's a GFP_KERNEL allocation, and all nodes in the
4888 * current tasks mems_allowed came up empty on the first pass over
4889 * the zonelist. So only GFP_KERNEL allocations, if all nodes in the
4890 * cpuset are short of memory, might require taking the callback_lock.
4891 *
4892 * The first call here from mm/page_alloc:get_page_from_freelist()
4893 * has __GFP_HARDWALL set in gfp_mask, enforcing hardwall cpusets,
4894 * so no allocation on a node outside the cpuset is allowed (unless
4895 * in interrupt, of course).
4896 *
4897 * The second pass through get_page_from_freelist() doesn't even call
4898 * here for GFP_ATOMIC calls. For those calls, the __alloc_pages()
4899 * variable 'wait' is not set, and the bit ALLOC_CPUSET is not set
4900 * in alloc_flags. That logic and the checks below have the combined
4901 * affect that:
4902 * in_interrupt - any node ok (current task context irrelevant)
4903 * GFP_ATOMIC - any node ok
4904 * tsk_is_oom_victim - any node ok
4905 * GFP_KERNEL - any node in enclosing hardwalled cpuset ok
4906 * GFP_USER - only nodes in current tasks mems allowed ok.
4907 */
4908bool cpuset_node_allowed(int node, gfp_t gfp_mask)
4909{
4910 struct cpuset *cs; /* current cpuset ancestors */
4911 bool allowed; /* is allocation in zone z allowed? */
4912 unsigned long flags;
4913
4914 if (in_interrupt())
4915 return true;
4916 if (node_isset(node, current->mems_allowed))
4917 return true;
4918 /*
4919 * Allow tasks that have access to memory reserves because they have
4920 * been OOM killed to get memory anywhere.
4921 */
4922 if (unlikely(tsk_is_oom_victim(current)))
4923 return true;
4924 if (gfp_mask & __GFP_HARDWALL) /* If hardwall request, stop here */
4925 return false;
4926
4927 if (current->flags & PF_EXITING) /* Let dying task have memory */
4928 return true;
4929
4930 /* Not hardwall and node outside mems_allowed: scan up cpusets */
4931 spin_lock_irqsave(&callback_lock, flags);
4932
4933 rcu_read_lock();
4934 cs = nearest_hardwall_ancestor(task_cs(current));
4935 allowed = node_isset(node, cs->mems_allowed);
4936 rcu_read_unlock();
4937
4938 spin_unlock_irqrestore(&callback_lock, flags);
4939 return allowed;
4940}
4941
4942/**
4943 * cpuset_spread_node() - On which node to begin search for a page
4944 * @rotor: round robin rotor
4945 *
4946 * If a task is marked PF_SPREAD_PAGE or PF_SPREAD_SLAB (as for
4947 * tasks in a cpuset with is_spread_page or is_spread_slab set),
4948 * and if the memory allocation used cpuset_mem_spread_node()
4949 * to determine on which node to start looking, as it will for
4950 * certain page cache or slab cache pages such as used for file
4951 * system buffers and inode caches, then instead of starting on the
4952 * local node to look for a free page, rather spread the starting
4953 * node around the tasks mems_allowed nodes.
4954 *
4955 * We don't have to worry about the returned node being offline
4956 * because "it can't happen", and even if it did, it would be ok.
4957 *
4958 * The routines calling guarantee_online_mems() are careful to
4959 * only set nodes in task->mems_allowed that are online. So it
4960 * should not be possible for the following code to return an
4961 * offline node. But if it did, that would be ok, as this routine
4962 * is not returning the node where the allocation must be, only
4963 * the node where the search should start. The zonelist passed to
4964 * __alloc_pages() will include all nodes. If the slab allocator
4965 * is passed an offline node, it will fall back to the local node.
4966 * See kmem_cache_alloc_node().
4967 */
4968static int cpuset_spread_node(int *rotor)
4969{
4970 return *rotor = next_node_in(*rotor, current->mems_allowed);
4971}
4972
4973/**
4974 * cpuset_mem_spread_node() - On which node to begin search for a file page
4975 */
4976int cpuset_mem_spread_node(void)
4977{
4978 if (current->cpuset_mem_spread_rotor == NUMA_NO_NODE)
4979 current->cpuset_mem_spread_rotor =
4980 node_random(¤t->mems_allowed);
4981
4982 return cpuset_spread_node(¤t->cpuset_mem_spread_rotor);
4983}
4984
4985/**
4986 * cpuset_slab_spread_node() - On which node to begin search for a slab page
4987 */
4988int cpuset_slab_spread_node(void)
4989{
4990 if (current->cpuset_slab_spread_rotor == NUMA_NO_NODE)
4991 current->cpuset_slab_spread_rotor =
4992 node_random(¤t->mems_allowed);
4993
4994 return cpuset_spread_node(¤t->cpuset_slab_spread_rotor);
4995}
4996EXPORT_SYMBOL_GPL(cpuset_mem_spread_node);
4997
4998/**
4999 * cpuset_mems_allowed_intersects - Does @tsk1's mems_allowed intersect @tsk2's?
5000 * @tsk1: pointer to task_struct of some task.
5001 * @tsk2: pointer to task_struct of some other task.
5002 *
5003 * Description: Return true if @tsk1's mems_allowed intersects the
5004 * mems_allowed of @tsk2. Used by the OOM killer to determine if
5005 * one of the task's memory usage might impact the memory available
5006 * to the other.
5007 **/
5008
5009int cpuset_mems_allowed_intersects(const struct task_struct *tsk1,
5010 const struct task_struct *tsk2)
5011{
5012 return nodes_intersects(tsk1->mems_allowed, tsk2->mems_allowed);
5013}
5014
5015/**
5016 * cpuset_print_current_mems_allowed - prints current's cpuset and mems_allowed
5017 *
5018 * Description: Prints current's name, cpuset name, and cached copy of its
5019 * mems_allowed to the kernel log.
5020 */
5021void cpuset_print_current_mems_allowed(void)
5022{
5023 struct cgroup *cgrp;
5024
5025 rcu_read_lock();
5026
5027 cgrp = task_cs(current)->css.cgroup;
5028 pr_cont(",cpuset=");
5029 pr_cont_cgroup_name(cgrp);
5030 pr_cont(",mems_allowed=%*pbl",
5031 nodemask_pr_args(¤t->mems_allowed));
5032
5033 rcu_read_unlock();
5034}
5035
5036/*
5037 * Collection of memory_pressure is suppressed unless
5038 * this flag is enabled by writing "1" to the special
5039 * cpuset file 'memory_pressure_enabled' in the root cpuset.
5040 */
5041
5042int cpuset_memory_pressure_enabled __read_mostly;
5043
5044/*
5045 * __cpuset_memory_pressure_bump - keep stats of per-cpuset reclaims.
5046 *
5047 * Keep a running average of the rate of synchronous (direct)
5048 * page reclaim efforts initiated by tasks in each cpuset.
5049 *
5050 * This represents the rate at which some task in the cpuset
5051 * ran low on memory on all nodes it was allowed to use, and
5052 * had to enter the kernels page reclaim code in an effort to
5053 * create more free memory by tossing clean pages or swapping
5054 * or writing dirty pages.
5055 *
5056 * Display to user space in the per-cpuset read-only file
5057 * "memory_pressure". Value displayed is an integer
5058 * representing the recent rate of entry into the synchronous
5059 * (direct) page reclaim by any task attached to the cpuset.
5060 */
5061
5062void __cpuset_memory_pressure_bump(void)
5063{
5064 rcu_read_lock();
5065 fmeter_markevent(&task_cs(current)->fmeter);
5066 rcu_read_unlock();
5067}
5068
5069#ifdef CONFIG_PROC_PID_CPUSET
5070/*
5071 * proc_cpuset_show()
5072 * - Print tasks cpuset path into seq_file.
5073 * - Used for /proc/<pid>/cpuset.
5074 * - No need to task_lock(tsk) on this tsk->cpuset reference, as it
5075 * doesn't really matter if tsk->cpuset changes after we read it,
5076 * and we take cpuset_mutex, keeping cpuset_attach() from changing it
5077 * anyway.
5078 */
5079int proc_cpuset_show(struct seq_file *m, struct pid_namespace *ns,
5080 struct pid *pid, struct task_struct *tsk)
5081{
5082 char *buf;
5083 struct cgroup_subsys_state *css;
5084 int retval;
5085
5086 retval = -ENOMEM;
5087 buf = kmalloc(PATH_MAX, GFP_KERNEL);
5088 if (!buf)
5089 goto out;
5090
5091 css = task_get_css(tsk, cpuset_cgrp_id);
5092 retval = cgroup_path_ns(css->cgroup, buf, PATH_MAX,
5093 current->nsproxy->cgroup_ns);
5094 css_put(css);
5095 if (retval == -E2BIG)
5096 retval = -ENAMETOOLONG;
5097 if (retval < 0)
5098 goto out_free;
5099 seq_puts(m, buf);
5100 seq_putc(m, '\n');
5101 retval = 0;
5102out_free:
5103 kfree(buf);
5104out:
5105 return retval;
5106}
5107#endif /* CONFIG_PROC_PID_CPUSET */
5108
5109/* Display task mems_allowed in /proc/<pid>/status file. */
5110void cpuset_task_status_allowed(struct seq_file *m, struct task_struct *task)
5111{
5112 seq_printf(m, "Mems_allowed:\t%*pb\n",
5113 nodemask_pr_args(&task->mems_allowed));
5114 seq_printf(m, "Mems_allowed_list:\t%*pbl\n",
5115 nodemask_pr_args(&task->mems_allowed));
5116}