1 /*
2 * kernel/cpuset.c
3 *
4 * Processor and Memory placement constraints for sets of tasks.
5 *
6 * Copyright (C) 2003 BULL SA.
7 * Copyright (C) 2004-2007 Silicon Graphics, Inc.
8 * Copyright (C) 2006 Google, Inc
9 *
10 * Portions derived from Patrick Mochel's sysfs code.
11 * sysfs is Copyright (c) 2001-3 Patrick Mochel
12 *
13 * 2003-10-10 Written by Simon Derr.
14 * 2003-10-22 Updates by Stephen Hemminger.
15 * 2004 May-July Rework by Paul Jackson.
16 * 2006 Rework by Paul Menage to use generic cgroups
17 * 2008 Rework of the scheduler domains and CPU hotplug handling
18 * by Max Krasnyansky
19 *
20 * This file is subject to the terms and conditions of the GNU General Public
21 * License. See the file COPYING in the main directory of the Linux
22 * distribution for more details.
23 */
24
25 #include <linux/cpu.h>
26 #include <linux/cpumask.h>
27 #include <linux/cpuset.h>
28 #include <linux/err.h>
29 #include <linux/errno.h>
30 #include <linux/file.h>
31 #include <linux/fs.h>
32 #include <linux/init.h>
33 #include <linux/interrupt.h>
34 #include <linux/kernel.h>
35 #include <linux/kmod.h>
36 #include <linux/kthread.h>
37 #include <linux/list.h>
38 #include <linux/mempolicy.h>
39 #include <linux/mm.h>
40 #include <linux/memory.h>
41 #include <linux/export.h>
42 #include <linux/mount.h>
43 #include <linux/fs_context.h>
44 #include <linux/namei.h>
45 #include <linux/pagemap.h>
46 #include <linux/proc_fs.h>
47 #include <linux/rcupdate.h>
48 #include <linux/sched.h>
49 #include <linux/sched/deadline.h>
50 #include <linux/sched/mm.h>
51 #include <linux/sched/task.h>
52 #include <linux/seq_file.h>
53 #include <linux/security.h>
54 #include <linux/slab.h>
55 #include <linux/spinlock.h>
56 #include <linux/stat.h>
57 #include <linux/string.h>
58 #include <linux/time.h>
59 #include <linux/time64.h>
60 #include <linux/backing-dev.h>
61 #include <linux/sort.h>
62 #include <linux/oom.h>
63 #include <linux/sched/isolation.h>
64 #include <linux/uaccess.h>
65 #include <linux/atomic.h>
66 #include <linux/mutex.h>
67 #include <linux/cgroup.h>
68 #include <linux/wait.h>
69
70 #include <trace/hooks/cgroup.h>
71 #include <trace/hooks/sched.h>
72
73 DEFINE_STATIC_KEY_FALSE(cpusets_pre_enable_key);
74 DEFINE_STATIC_KEY_FALSE(cpusets_enabled_key);
75
76 /*
77 * There could be abnormal cpuset configurations for cpu or memory
78 * node binding, add this key to provide a quick low-cost judgment
79 * of the situation.
80 */
81 DEFINE_STATIC_KEY_FALSE(cpusets_insane_config_key);
82
83 /* See "Frequency meter" comments, below. */
84
85 struct fmeter {
86 int cnt; /* unprocessed events count */
87 int val; /* most recent output value */
88 time64_t time; /* clock (secs) when val computed */
89 spinlock_t lock; /* guards read or write of above */
90 };
91
92 /*
93 * Invalid partition error code
94 */
95 enum prs_errcode {
96 PERR_NONE = 0,
97 PERR_INVCPUS,
98 PERR_INVPARENT,
99 PERR_NOTPART,
100 PERR_NOTEXCL,
101 PERR_NOCPUS,
102 PERR_HOTPLUG,
103 PERR_CPUSEMPTY,
104 };
105
106 static const char * const perr_strings[] = {
107 [PERR_INVCPUS] = "Invalid cpu list in cpuset.cpus",
108 [PERR_INVPARENT] = "Parent is an invalid partition root",
109 [PERR_NOTPART] = "Parent is not a partition root",
110 [PERR_NOTEXCL] = "Cpu list in cpuset.cpus not exclusive",
111 [PERR_NOCPUS] = "Parent unable to distribute cpu downstream",
112 [PERR_HOTPLUG] = "No cpu available due to hotplug",
113 [PERR_CPUSEMPTY] = "cpuset.cpus is empty",
114 };
115
116 struct cpuset {
117 struct cgroup_subsys_state css;
118
119 unsigned long flags; /* "unsigned long" so bitops work */
120
121 /*
122 * On default hierarchy:
123 *
124 * The user-configured masks can only be changed by writing to
125 * cpuset.cpus and cpuset.mems, and won't be limited by the
126 * parent masks.
127 *
128 * The effective masks is the real masks that apply to the tasks
129 * in the cpuset. They may be changed if the configured masks are
130 * changed or hotplug happens.
131 *
132 * effective_mask == configured_mask & parent's effective_mask,
133 * and if it ends up empty, it will inherit the parent's mask.
134 *
135 *
136 * On legacy hierarchy:
137 *
138 * The user-configured masks are always the same with effective masks.
139 */
140
141 /* user-configured CPUs and Memory Nodes allow to tasks */
142 cpumask_var_t cpus_allowed;
143 cpumask_var_t cpus_requested;
144 nodemask_t mems_allowed;
145
146 /* effective CPUs and Memory Nodes allow to tasks */
147 cpumask_var_t effective_cpus;
148 nodemask_t effective_mems;
149
150 /*
151 * CPUs allocated to child sub-partitions (default hierarchy only)
152 * - CPUs granted by the parent = effective_cpus U subparts_cpus
153 * - effective_cpus and subparts_cpus are mutually exclusive.
154 *
155 * effective_cpus contains only onlined CPUs, but subparts_cpus
156 * may have offlined ones.
157 */
158 cpumask_var_t subparts_cpus;
159
160 /*
161 * This is old Memory Nodes tasks took on.
162 *
163 * - top_cpuset.old_mems_allowed is initialized to mems_allowed.
164 * - A new cpuset's old_mems_allowed is initialized when some
165 * task is moved into it.
166 * - old_mems_allowed is used in cpuset_migrate_mm() when we change
167 * cpuset.mems_allowed and have tasks' nodemask updated, and
168 * then old_mems_allowed is updated to mems_allowed.
169 */
170 nodemask_t old_mems_allowed;
171
172 struct fmeter fmeter; /* memory_pressure filter */
173
174 /*
175 * Tasks are being attached to this cpuset. Used to prevent
176 * zeroing cpus/mems_allowed between ->can_attach() and ->attach().
177 */
178 int attach_in_progress;
179
180 /* partition number for rebuild_sched_domains() */
181 int pn;
182
183 /* for custom sched domain */
184 int relax_domain_level;
185
186 /* number of CPUs in subparts_cpus */
187 int nr_subparts_cpus;
188
189 /* partition root state */
190 int partition_root_state;
191
192 /*
193 * Default hierarchy only:
194 * use_parent_ecpus - set if using parent's effective_cpus
195 * child_ecpus_count - # of children with use_parent_ecpus set
196 */
197 int use_parent_ecpus;
198 int child_ecpus_count;
199
200 /*
201 * number of SCHED_DEADLINE tasks attached to this cpuset, so that we
202 * know when to rebuild associated root domain bandwidth information.
203 */
204 int nr_deadline_tasks;
205 int nr_migrate_dl_tasks;
206 u64 sum_migrate_dl_bw;
207
208 /* Invalid partition error code, not lock protected */
209 enum prs_errcode prs_err;
210
211 /* Handle for cpuset.cpus.partition */
212 struct cgroup_file partition_file;
213 };
214
215 /*
216 * Partition root states:
217 *
218 * 0 - member (not a partition root)
219 * 1 - partition root
220 * 2 - partition root without load balancing (isolated)
221 * -1 - invalid partition root
222 * -2 - invalid isolated partition root
223 */
224 #define PRS_MEMBER 0
225 #define PRS_ROOT 1
226 #define PRS_ISOLATED 2
227 #define PRS_INVALID_ROOT -1
228 #define PRS_INVALID_ISOLATED -2
229
is_prs_invalid(int prs_state)230 static inline bool is_prs_invalid(int prs_state)
231 {
232 return prs_state < 0;
233 }
234
235 /*
236 * Temporary cpumasks for working with partitions that are passed among
237 * functions to avoid memory allocation in inner functions.
238 */
239 struct tmpmasks {
240 cpumask_var_t addmask, delmask; /* For partition root */
241 cpumask_var_t new_cpus; /* For update_cpumasks_hier() */
242 };
243
css_cs(struct cgroup_subsys_state * css)244 static inline struct cpuset *css_cs(struct cgroup_subsys_state *css)
245 {
246 return css ? container_of(css, struct cpuset, css) : NULL;
247 }
248
249 /* Retrieve the cpuset for a task */
task_cs(struct task_struct * task)250 static inline struct cpuset *task_cs(struct task_struct *task)
251 {
252 return css_cs(task_css(task, cpuset_cgrp_id));
253 }
254
parent_cs(struct cpuset * cs)255 static inline struct cpuset *parent_cs(struct cpuset *cs)
256 {
257 return css_cs(cs->css.parent);
258 }
259
inc_dl_tasks_cs(struct task_struct * p)260 void inc_dl_tasks_cs(struct task_struct *p)
261 {
262 struct cpuset *cs = task_cs(p);
263
264 cs->nr_deadline_tasks++;
265 }
266
dec_dl_tasks_cs(struct task_struct * p)267 void dec_dl_tasks_cs(struct task_struct *p)
268 {
269 struct cpuset *cs = task_cs(p);
270
271 cs->nr_deadline_tasks--;
272 }
273
274 /* bits in struct cpuset flags field */
275 typedef enum {
276 CS_ONLINE,
277 CS_CPU_EXCLUSIVE,
278 CS_MEM_EXCLUSIVE,
279 CS_MEM_HARDWALL,
280 CS_MEMORY_MIGRATE,
281 CS_SCHED_LOAD_BALANCE,
282 CS_SPREAD_PAGE,
283 CS_SPREAD_SLAB,
284 } cpuset_flagbits_t;
285
286 /* convenient tests for these bits */
is_cpuset_online(struct cpuset * cs)287 static inline bool is_cpuset_online(struct cpuset *cs)
288 {
289 return test_bit(CS_ONLINE, &cs->flags) && !css_is_dying(&cs->css);
290 }
291
is_cpu_exclusive(const struct cpuset * cs)292 static inline int is_cpu_exclusive(const struct cpuset *cs)
293 {
294 return test_bit(CS_CPU_EXCLUSIVE, &cs->flags);
295 }
296
is_mem_exclusive(const struct cpuset * cs)297 static inline int is_mem_exclusive(const struct cpuset *cs)
298 {
299 return test_bit(CS_MEM_EXCLUSIVE, &cs->flags);
300 }
301
is_mem_hardwall(const struct cpuset * cs)302 static inline int is_mem_hardwall(const struct cpuset *cs)
303 {
304 return test_bit(CS_MEM_HARDWALL, &cs->flags);
305 }
306
is_sched_load_balance(const struct cpuset * cs)307 static inline int is_sched_load_balance(const struct cpuset *cs)
308 {
309 return test_bit(CS_SCHED_LOAD_BALANCE, &cs->flags);
310 }
311
is_memory_migrate(const struct cpuset * cs)312 static inline int is_memory_migrate(const struct cpuset *cs)
313 {
314 return test_bit(CS_MEMORY_MIGRATE, &cs->flags);
315 }
316
is_spread_page(const struct cpuset * cs)317 static inline int is_spread_page(const struct cpuset *cs)
318 {
319 return test_bit(CS_SPREAD_PAGE, &cs->flags);
320 }
321
is_spread_slab(const struct cpuset * cs)322 static inline int is_spread_slab(const struct cpuset *cs)
323 {
324 return test_bit(CS_SPREAD_SLAB, &cs->flags);
325 }
326
is_partition_valid(const struct cpuset * cs)327 static inline int is_partition_valid(const struct cpuset *cs)
328 {
329 return cs->partition_root_state > 0;
330 }
331
is_partition_invalid(const struct cpuset * cs)332 static inline int is_partition_invalid(const struct cpuset *cs)
333 {
334 return cs->partition_root_state < 0;
335 }
336
337 /*
338 * Callers should hold callback_lock to modify partition_root_state.
339 */
make_partition_invalid(struct cpuset * cs)340 static inline void make_partition_invalid(struct cpuset *cs)
341 {
342 if (is_partition_valid(cs))
343 cs->partition_root_state = -cs->partition_root_state;
344 }
345
346 /*
347 * Send notification event of whenever partition_root_state changes.
348 */
notify_partition_change(struct cpuset * cs,int old_prs)349 static inline void notify_partition_change(struct cpuset *cs, int old_prs)
350 {
351 if (old_prs == cs->partition_root_state)
352 return;
353 cgroup_file_notify(&cs->partition_file);
354
355 /* Reset prs_err if not invalid */
356 if (is_partition_valid(cs))
357 WRITE_ONCE(cs->prs_err, PERR_NONE);
358 }
359
360 static struct cpuset top_cpuset = {
361 .flags = ((1 << CS_ONLINE) | (1 << CS_CPU_EXCLUSIVE) |
362 (1 << CS_MEM_EXCLUSIVE)),
363 .partition_root_state = PRS_ROOT,
364 };
365
366 /**
367 * cpuset_for_each_child - traverse online children of a cpuset
368 * @child_cs: loop cursor pointing to the current child
369 * @pos_css: used for iteration
370 * @parent_cs: target cpuset to walk children of
371 *
372 * Walk @child_cs through the online children of @parent_cs. Must be used
373 * with RCU read locked.
374 */
375 #define cpuset_for_each_child(child_cs, pos_css, parent_cs) \
376 css_for_each_child((pos_css), &(parent_cs)->css) \
377 if (is_cpuset_online(((child_cs) = css_cs((pos_css)))))
378
379 /**
380 * cpuset_for_each_descendant_pre - pre-order walk of a cpuset's descendants
381 * @des_cs: loop cursor pointing to the current descendant
382 * @pos_css: used for iteration
383 * @root_cs: target cpuset to walk ancestor of
384 *
385 * Walk @des_cs through the online descendants of @root_cs. Must be used
386 * with RCU read locked. The caller may modify @pos_css by calling
387 * css_rightmost_descendant() to skip subtree. @root_cs is included in the
388 * iteration and the first node to be visited.
389 */
390 #define cpuset_for_each_descendant_pre(des_cs, pos_css, root_cs) \
391 css_for_each_descendant_pre((pos_css), &(root_cs)->css) \
392 if (is_cpuset_online(((des_cs) = css_cs((pos_css)))))
393
394 /*
395 * There are two global locks guarding cpuset structures - cpuset_mutex and
396 * callback_lock. We also require taking task_lock() when dereferencing a
397 * task's cpuset pointer. See "The task_lock() exception", at the end of this
398 * comment. The cpuset code uses only cpuset_mutex. Other kernel subsystems
399 * can use cpuset_lock()/cpuset_unlock() to prevent change to cpuset
400 * structures. Note that cpuset_mutex needs to be a mutex as it is used in
401 * paths that rely on priority inheritance (e.g. scheduler - on RT) for
402 * correctness.
403 *
404 * A task must hold both locks to modify cpusets. If a task holds
405 * cpuset_mutex, it blocks others, ensuring that it is the only task able to
406 * also acquire callback_lock and be able to modify cpusets. It can perform
407 * various checks on the cpuset structure first, knowing nothing will change.
408 * It can also allocate memory while just holding cpuset_mutex. While it is
409 * performing these checks, various callback routines can briefly acquire
410 * callback_lock to query cpusets. Once it is ready to make the changes, it
411 * takes callback_lock, blocking everyone else.
412 *
413 * Calls to the kernel memory allocator can not be made while holding
414 * callback_lock, as that would risk double tripping on callback_lock
415 * from one of the callbacks into the cpuset code from within
416 * __alloc_pages().
417 *
418 * If a task is only holding callback_lock, then it has read-only
419 * access to cpusets.
420 *
421 * Now, the task_struct fields mems_allowed and mempolicy may be changed
422 * by other task, we use alloc_lock in the task_struct fields to protect
423 * them.
424 *
425 * The cpuset_common_file_read() handlers only hold callback_lock across
426 * small pieces of code, such as when reading out possibly multi-word
427 * cpumasks and nodemasks.
428 *
429 * Accessing a task's cpuset should be done in accordance with the
430 * guidelines for accessing subsystem state in kernel/cgroup.c
431 */
432
433 DEFINE_STATIC_PERCPU_RWSEM(cpuset_rwsem);
434
435 static DEFINE_MUTEX(cpuset_mutex);
436
cpuset_lock(void)437 void cpuset_lock(void)
438 {
439 mutex_lock(&cpuset_mutex);
440 }
441
cpuset_unlock(void)442 void cpuset_unlock(void)
443 {
444 mutex_unlock(&cpuset_mutex);
445 }
446
447 static DEFINE_SPINLOCK(callback_lock);
448
449 static struct workqueue_struct *cpuset_migrate_mm_wq;
450
451 /*
452 * CPU / memory hotplug is handled asynchronously.
453 */
454 static void cpuset_hotplug_workfn(struct work_struct *work);
455 static DECLARE_WORK(cpuset_hotplug_work, cpuset_hotplug_workfn);
456
457 static DECLARE_WAIT_QUEUE_HEAD(cpuset_attach_wq);
458
check_insane_mems_config(nodemask_t * nodes)459 static inline void check_insane_mems_config(nodemask_t *nodes)
460 {
461 if (!cpusets_insane_config() &&
462 movable_only_nodes(nodes)) {
463 static_branch_enable(&cpusets_insane_config_key);
464 pr_info("Unsupported (movable nodes only) cpuset configuration detected (nmask=%*pbl)!\n"
465 "Cpuset allocations might fail even with a lot of memory available.\n",
466 nodemask_pr_args(nodes));
467 }
468 }
469
470 /*
471 * Cgroup v2 behavior is used on the "cpus" and "mems" control files when
472 * on default hierarchy or when the cpuset_v2_mode flag is set by mounting
473 * the v1 cpuset cgroup filesystem with the "cpuset_v2_mode" mount option.
474 * With v2 behavior, "cpus" and "mems" are always what the users have
475 * requested and won't be changed by hotplug events. Only the effective
476 * cpus or mems will be affected.
477 */
is_in_v2_mode(void)478 static inline bool is_in_v2_mode(void)
479 {
480 return cgroup_subsys_on_dfl(cpuset_cgrp_subsys) ||
481 (cpuset_cgrp_subsys.root->flags & CGRP_ROOT_CPUSET_V2_MODE);
482 }
483
484 /**
485 * partition_is_populated - check if partition has tasks
486 * @cs: partition root to be checked
487 * @excluded_child: a child cpuset to be excluded in task checking
488 * Return: true if there are tasks, false otherwise
489 *
490 * It is assumed that @cs is a valid partition root. @excluded_child should
491 * be non-NULL when this cpuset is going to become a partition itself.
492 */
partition_is_populated(struct cpuset * cs,struct cpuset * excluded_child)493 static inline bool partition_is_populated(struct cpuset *cs,
494 struct cpuset *excluded_child)
495 {
496 struct cgroup_subsys_state *css;
497 struct cpuset *child;
498
499 if (cs->css.cgroup->nr_populated_csets)
500 return true;
501 if (!excluded_child && !cs->nr_subparts_cpus)
502 return cgroup_is_populated(cs->css.cgroup);
503
504 rcu_read_lock();
505 cpuset_for_each_child(child, css, cs) {
506 if (child == excluded_child)
507 continue;
508 if (is_partition_valid(child))
509 continue;
510 if (cgroup_is_populated(child->css.cgroup)) {
511 rcu_read_unlock();
512 return true;
513 }
514 }
515 rcu_read_unlock();
516 return false;
517 }
518
519 /*
520 * Return in pmask the portion of a task's cpusets's cpus_allowed that
521 * are online and are capable of running the task. If none are found,
522 * walk up the cpuset hierarchy until we find one that does have some
523 * appropriate cpus.
524 *
525 * One way or another, we guarantee to return some non-empty subset
526 * of cpu_online_mask.
527 *
528 * Call with callback_lock or cpuset_mutex held.
529 */
guarantee_online_cpus(struct task_struct * tsk,struct cpumask * pmask)530 static void guarantee_online_cpus(struct task_struct *tsk,
531 struct cpumask *pmask)
532 {
533 const struct cpumask *possible_mask = task_cpu_possible_mask(tsk);
534 struct cpuset *cs;
535
536 if (WARN_ON(!cpumask_and(pmask, possible_mask, cpu_online_mask)))
537 cpumask_copy(pmask, cpu_online_mask);
538
539 rcu_read_lock();
540 cs = task_cs(tsk);
541
542 while (!cpumask_intersects(cs->effective_cpus, pmask)) {
543 cs = parent_cs(cs);
544 if (unlikely(!cs)) {
545 /*
546 * The top cpuset doesn't have any online cpu as a
547 * consequence of a race between cpuset_hotplug_work
548 * and cpu hotplug notifier. But we know the top
549 * cpuset's effective_cpus is on its way to be
550 * identical to cpu_online_mask.
551 */
552 goto out_unlock;
553 }
554 }
555 cpumask_and(pmask, pmask, cs->effective_cpus);
556
557 out_unlock:
558 rcu_read_unlock();
559 }
560
561 /*
562 * Return in *pmask the portion of a cpusets's mems_allowed that
563 * are online, with memory. If none are online with memory, walk
564 * up the cpuset hierarchy until we find one that does have some
565 * online mems. The top cpuset always has some mems online.
566 *
567 * One way or another, we guarantee to return some non-empty subset
568 * of node_states[N_MEMORY].
569 *
570 * Call with callback_lock or cpuset_mutex held.
571 */
guarantee_online_mems(struct cpuset * cs,nodemask_t * pmask)572 static void guarantee_online_mems(struct cpuset *cs, nodemask_t *pmask)
573 {
574 while (!nodes_intersects(cs->effective_mems, node_states[N_MEMORY]))
575 cs = parent_cs(cs);
576 nodes_and(*pmask, cs->effective_mems, node_states[N_MEMORY]);
577 }
578
579 /*
580 * update task's spread flag if cpuset's page/slab spread flag is set
581 *
582 * Call with callback_lock or cpuset_mutex held. The check can be skipped
583 * if on default hierarchy.
584 */
cpuset_update_task_spread_flags(struct cpuset * cs,struct task_struct * tsk)585 static void cpuset_update_task_spread_flags(struct cpuset *cs,
586 struct task_struct *tsk)
587 {
588 if (cgroup_subsys_on_dfl(cpuset_cgrp_subsys))
589 return;
590
591 if (is_spread_page(cs))
592 task_set_spread_page(tsk);
593 else
594 task_clear_spread_page(tsk);
595
596 if (is_spread_slab(cs))
597 task_set_spread_slab(tsk);
598 else
599 task_clear_spread_slab(tsk);
600 }
601
602 /*
603 * is_cpuset_subset(p, q) - Is cpuset p a subset of cpuset q?
604 *
605 * One cpuset is a subset of another if all its allowed CPUs and
606 * Memory Nodes are a subset of the other, and its exclusive flags
607 * are only set if the other's are set. Call holding cpuset_mutex.
608 */
609
is_cpuset_subset(const struct cpuset * p,const struct cpuset * q)610 static int is_cpuset_subset(const struct cpuset *p, const struct cpuset *q)
611 {
612 return cpumask_subset(p->cpus_requested, q->cpus_requested) &&
613 nodes_subset(p->mems_allowed, q->mems_allowed) &&
614 is_cpu_exclusive(p) <= is_cpu_exclusive(q) &&
615 is_mem_exclusive(p) <= is_mem_exclusive(q);
616 }
617
618 /**
619 * alloc_cpumasks - allocate three cpumasks for cpuset
620 * @cs: the cpuset that have cpumasks to be allocated.
621 * @tmp: the tmpmasks structure pointer
622 * Return: 0 if successful, -ENOMEM otherwise.
623 *
624 * Only one of the two input arguments should be non-NULL.
625 */
alloc_cpumasks(struct cpuset * cs,struct tmpmasks * tmp)626 static inline int alloc_cpumasks(struct cpuset *cs, struct tmpmasks *tmp)
627 {
628 cpumask_var_t *pmask1, *pmask2, *pmask3;
629
630 if (cs) {
631 pmask1 = &cs->cpus_allowed;
632 pmask2 = &cs->effective_cpus;
633 pmask3 = &cs->subparts_cpus;
634 } else {
635 pmask1 = &tmp->new_cpus;
636 pmask2 = &tmp->addmask;
637 pmask3 = &tmp->delmask;
638 }
639
640 if (!zalloc_cpumask_var(pmask1, GFP_KERNEL))
641 return -ENOMEM;
642
643 if (!zalloc_cpumask_var(pmask2, GFP_KERNEL))
644 goto free_one;
645
646 if (!zalloc_cpumask_var(pmask3, GFP_KERNEL))
647 goto free_two;
648
649 if (cs && !zalloc_cpumask_var(&cs->cpus_requested, GFP_KERNEL))
650 goto free_three;
651
652 return 0;
653
654 free_three:
655 free_cpumask_var(*pmask3);
656 free_two:
657 free_cpumask_var(*pmask2);
658 free_one:
659 free_cpumask_var(*pmask1);
660 return -ENOMEM;
661 }
662
663 /**
664 * free_cpumasks - free cpumasks in a tmpmasks structure
665 * @cs: the cpuset that have cpumasks to be free.
666 * @tmp: the tmpmasks structure pointer
667 */
free_cpumasks(struct cpuset * cs,struct tmpmasks * tmp)668 static inline void free_cpumasks(struct cpuset *cs, struct tmpmasks *tmp)
669 {
670 if (cs) {
671 free_cpumask_var(cs->cpus_allowed);
672 free_cpumask_var(cs->cpus_requested);
673 free_cpumask_var(cs->effective_cpus);
674 free_cpumask_var(cs->subparts_cpus);
675 }
676 if (tmp) {
677 free_cpumask_var(tmp->new_cpus);
678 free_cpumask_var(tmp->addmask);
679 free_cpumask_var(tmp->delmask);
680 }
681 }
682
683 /**
684 * alloc_trial_cpuset - allocate a trial cpuset
685 * @cs: the cpuset that the trial cpuset duplicates
686 */
alloc_trial_cpuset(struct cpuset * cs)687 static struct cpuset *alloc_trial_cpuset(struct cpuset *cs)
688 {
689 struct cpuset *trial;
690
691 trial = kmemdup(cs, sizeof(*cs), GFP_KERNEL);
692 if (!trial)
693 return NULL;
694
695 if (alloc_cpumasks(trial, NULL)) {
696 kfree(trial);
697 return NULL;
698 }
699
700 cpumask_copy(trial->cpus_allowed, cs->cpus_allowed);
701 cpumask_copy(trial->cpus_requested, cs->cpus_requested);
702 cpumask_copy(trial->effective_cpus, cs->effective_cpus);
703 return trial;
704 }
705
706 /**
707 * free_cpuset - free the cpuset
708 * @cs: the cpuset to be freed
709 */
free_cpuset(struct cpuset * cs)710 static inline void free_cpuset(struct cpuset *cs)
711 {
712 free_cpumasks(cs, NULL);
713 kfree(cs);
714 }
715
716 /*
717 * validate_change_legacy() - Validate conditions specific to legacy (v1)
718 * behavior.
719 */
validate_change_legacy(struct cpuset * cur,struct cpuset * trial)720 static int validate_change_legacy(struct cpuset *cur, struct cpuset *trial)
721 {
722 struct cgroup_subsys_state *css;
723 struct cpuset *c, *par;
724 int ret;
725
726 WARN_ON_ONCE(!rcu_read_lock_held());
727
728 /* Each of our child cpusets must be a subset of us */
729 ret = -EBUSY;
730 cpuset_for_each_child(c, css, cur)
731 if (!is_cpuset_subset(c, trial))
732 goto out;
733
734 /* On legacy hierarchy, we must be a subset of our parent cpuset. */
735 ret = -EACCES;
736 par = parent_cs(cur);
737 if (par && !is_cpuset_subset(trial, par))
738 goto out;
739
740 ret = 0;
741 out:
742 return ret;
743 }
744
745 /*
746 * validate_change() - Used to validate that any proposed cpuset change
747 * follows the structural rules for cpusets.
748 *
749 * If we replaced the flag and mask values of the current cpuset
750 * (cur) with those values in the trial cpuset (trial), would
751 * our various subset and exclusive rules still be valid? Presumes
752 * cpuset_mutex held.
753 *
754 * 'cur' is the address of an actual, in-use cpuset. Operations
755 * such as list traversal that depend on the actual address of the
756 * cpuset in the list must use cur below, not trial.
757 *
758 * 'trial' is the address of bulk structure copy of cur, with
759 * perhaps one or more of the fields cpus_allowed, mems_allowed,
760 * or flags changed to new, trial values.
761 *
762 * Return 0 if valid, -errno if not.
763 */
764
validate_change(struct cpuset * cur,struct cpuset * trial)765 static int validate_change(struct cpuset *cur, struct cpuset *trial)
766 {
767 struct cgroup_subsys_state *css;
768 struct cpuset *c, *par;
769 int ret = 0;
770
771 rcu_read_lock();
772
773 if (!is_in_v2_mode())
774 ret = validate_change_legacy(cur, trial);
775 if (ret)
776 goto out;
777
778 /* Remaining checks don't apply to root cpuset */
779 if (cur == &top_cpuset)
780 goto out;
781
782 par = parent_cs(cur);
783
784 /*
785 * Cpusets with tasks - existing or newly being attached - can't
786 * be changed to have empty cpus_allowed or mems_allowed.
787 */
788 ret = -ENOSPC;
789 if ((cgroup_is_populated(cur->css.cgroup) || cur->attach_in_progress)) {
790 if (!cpumask_empty(cur->cpus_allowed) &&
791 cpumask_empty(trial->cpus_allowed))
792 goto out;
793 if (!nodes_empty(cur->mems_allowed) &&
794 nodes_empty(trial->mems_allowed))
795 goto out;
796 }
797
798 /*
799 * We can't shrink if we won't have enough room for SCHED_DEADLINE
800 * tasks.
801 */
802 ret = -EBUSY;
803 if (is_cpu_exclusive(cur) &&
804 !cpuset_cpumask_can_shrink(cur->cpus_allowed,
805 trial->cpus_allowed))
806 goto out;
807
808 /*
809 * If either I or some sibling (!= me) is exclusive, we can't
810 * overlap
811 */
812 ret = -EINVAL;
813 cpuset_for_each_child(c, css, par) {
814 if ((is_cpu_exclusive(trial) || is_cpu_exclusive(c)) &&
815 c != cur &&
816 cpumask_intersects(trial->cpus_requested, c->cpus_requested))
817 goto out;
818 if ((is_mem_exclusive(trial) || is_mem_exclusive(c)) &&
819 c != cur &&
820 nodes_intersects(trial->mems_allowed, c->mems_allowed))
821 goto out;
822 }
823
824 ret = 0;
825 out:
826 rcu_read_unlock();
827 return ret;
828 }
829
830 #ifdef CONFIG_SMP
831 /*
832 * Helper routine for generate_sched_domains().
833 * Do cpusets a, b have overlapping effective cpus_allowed masks?
834 */
cpusets_overlap(struct cpuset * a,struct cpuset * b)835 static int cpusets_overlap(struct cpuset *a, struct cpuset *b)
836 {
837 return cpumask_intersects(a->effective_cpus, b->effective_cpus);
838 }
839
840 static void
update_domain_attr(struct sched_domain_attr * dattr,struct cpuset * c)841 update_domain_attr(struct sched_domain_attr *dattr, struct cpuset *c)
842 {
843 if (dattr->relax_domain_level < c->relax_domain_level)
844 dattr->relax_domain_level = c->relax_domain_level;
845 return;
846 }
847
update_domain_attr_tree(struct sched_domain_attr * dattr,struct cpuset * root_cs)848 static void update_domain_attr_tree(struct sched_domain_attr *dattr,
849 struct cpuset *root_cs)
850 {
851 struct cpuset *cp;
852 struct cgroup_subsys_state *pos_css;
853
854 rcu_read_lock();
855 cpuset_for_each_descendant_pre(cp, pos_css, root_cs) {
856 /* skip the whole subtree if @cp doesn't have any CPU */
857 if (cpumask_empty(cp->cpus_allowed)) {
858 pos_css = css_rightmost_descendant(pos_css);
859 continue;
860 }
861
862 if (is_sched_load_balance(cp))
863 update_domain_attr(dattr, cp);
864 }
865 rcu_read_unlock();
866 }
867
868 /* Must be called with cpuset_mutex held. */
nr_cpusets(void)869 static inline int nr_cpusets(void)
870 {
871 /* jump label reference count + the top-level cpuset */
872 return static_key_count(&cpusets_enabled_key.key) + 1;
873 }
874
875 /*
876 * generate_sched_domains()
877 *
878 * This function builds a partial partition of the systems CPUs
879 * A 'partial partition' is a set of non-overlapping subsets whose
880 * union is a subset of that set.
881 * The output of this function needs to be passed to kernel/sched/core.c
882 * partition_sched_domains() routine, which will rebuild the scheduler's
883 * load balancing domains (sched domains) as specified by that partial
884 * partition.
885 *
886 * See "What is sched_load_balance" in Documentation/admin-guide/cgroup-v1/cpusets.rst
887 * for a background explanation of this.
888 *
889 * Does not return errors, on the theory that the callers of this
890 * routine would rather not worry about failures to rebuild sched
891 * domains when operating in the severe memory shortage situations
892 * that could cause allocation failures below.
893 *
894 * Must be called with cpuset_mutex held.
895 *
896 * The three key local variables below are:
897 * cp - cpuset pointer, used (together with pos_css) to perform a
898 * top-down scan of all cpusets. For our purposes, rebuilding
899 * the schedulers sched domains, we can ignore !is_sched_load_
900 * balance cpusets.
901 * csa - (for CpuSet Array) Array of pointers to all the cpusets
902 * that need to be load balanced, for convenient iterative
903 * access by the subsequent code that finds the best partition,
904 * i.e the set of domains (subsets) of CPUs such that the
905 * cpus_allowed of every cpuset marked is_sched_load_balance
906 * is a subset of one of these domains, while there are as
907 * many such domains as possible, each as small as possible.
908 * doms - Conversion of 'csa' to an array of cpumasks, for passing to
909 * the kernel/sched/core.c routine partition_sched_domains() in a
910 * convenient format, that can be easily compared to the prior
911 * value to determine what partition elements (sched domains)
912 * were changed (added or removed.)
913 *
914 * Finding the best partition (set of domains):
915 * The triple nested loops below over i, j, k scan over the
916 * load balanced cpusets (using the array of cpuset pointers in
917 * csa[]) looking for pairs of cpusets that have overlapping
918 * cpus_allowed, but which don't have the same 'pn' partition
919 * number and gives them in the same partition number. It keeps
920 * looping on the 'restart' label until it can no longer find
921 * any such pairs.
922 *
923 * The union of the cpus_allowed masks from the set of
924 * all cpusets having the same 'pn' value then form the one
925 * element of the partition (one sched domain) to be passed to
926 * partition_sched_domains().
927 */
generate_sched_domains(cpumask_var_t ** domains,struct sched_domain_attr ** attributes)928 static int generate_sched_domains(cpumask_var_t **domains,
929 struct sched_domain_attr **attributes)
930 {
931 struct cpuset *cp; /* top-down scan of cpusets */
932 struct cpuset **csa; /* array of all cpuset ptrs */
933 int csn; /* how many cpuset ptrs in csa so far */
934 int i, j, k; /* indices for partition finding loops */
935 cpumask_var_t *doms; /* resulting partition; i.e. sched domains */
936 struct sched_domain_attr *dattr; /* attributes for custom domains */
937 int ndoms = 0; /* number of sched domains in result */
938 int nslot; /* next empty doms[] struct cpumask slot */
939 struct cgroup_subsys_state *pos_css;
940 bool root_load_balance = is_sched_load_balance(&top_cpuset);
941
942 doms = NULL;
943 dattr = NULL;
944 csa = NULL;
945
946 /* Special case for the 99% of systems with one, full, sched domain */
947 if (root_load_balance && !top_cpuset.nr_subparts_cpus) {
948 ndoms = 1;
949 doms = alloc_sched_domains(ndoms);
950 if (!doms)
951 goto done;
952
953 dattr = kmalloc(sizeof(struct sched_domain_attr), GFP_KERNEL);
954 if (dattr) {
955 *dattr = SD_ATTR_INIT;
956 update_domain_attr_tree(dattr, &top_cpuset);
957 }
958 cpumask_and(doms[0], top_cpuset.effective_cpus,
959 housekeeping_cpumask(HK_TYPE_DOMAIN));
960
961 goto done;
962 }
963
964 csa = kmalloc_array(nr_cpusets(), sizeof(cp), GFP_KERNEL);
965 if (!csa)
966 goto done;
967 csn = 0;
968
969 rcu_read_lock();
970 if (root_load_balance)
971 csa[csn++] = &top_cpuset;
972 cpuset_for_each_descendant_pre(cp, pos_css, &top_cpuset) {
973 if (cp == &top_cpuset)
974 continue;
975 /*
976 * Continue traversing beyond @cp iff @cp has some CPUs and
977 * isn't load balancing. The former is obvious. The
978 * latter: All child cpusets contain a subset of the
979 * parent's cpus, so just skip them, and then we call
980 * update_domain_attr_tree() to calc relax_domain_level of
981 * the corresponding sched domain.
982 *
983 * If root is load-balancing, we can skip @cp if it
984 * is a subset of the root's effective_cpus.
985 */
986 if (!cpumask_empty(cp->cpus_allowed) &&
987 !(is_sched_load_balance(cp) &&
988 cpumask_intersects(cp->cpus_allowed,
989 housekeeping_cpumask(HK_TYPE_DOMAIN))))
990 continue;
991
992 if (root_load_balance &&
993 cpumask_subset(cp->cpus_allowed, top_cpuset.effective_cpus))
994 continue;
995
996 if (is_sched_load_balance(cp) &&
997 !cpumask_empty(cp->effective_cpus))
998 csa[csn++] = cp;
999
1000 /* skip @cp's subtree if not a partition root */
1001 if (!is_partition_valid(cp))
1002 pos_css = css_rightmost_descendant(pos_css);
1003 }
1004 rcu_read_unlock();
1005
1006 for (i = 0; i < csn; i++)
1007 csa[i]->pn = i;
1008 ndoms = csn;
1009
1010 restart:
1011 /* Find the best partition (set of sched domains) */
1012 for (i = 0; i < csn; i++) {
1013 struct cpuset *a = csa[i];
1014 int apn = a->pn;
1015
1016 for (j = 0; j < csn; j++) {
1017 struct cpuset *b = csa[j];
1018 int bpn = b->pn;
1019
1020 if (apn != bpn && cpusets_overlap(a, b)) {
1021 for (k = 0; k < csn; k++) {
1022 struct cpuset *c = csa[k];
1023
1024 if (c->pn == bpn)
1025 c->pn = apn;
1026 }
1027 ndoms--; /* one less element */
1028 goto restart;
1029 }
1030 }
1031 }
1032
1033 /*
1034 * Now we know how many domains to create.
1035 * Convert <csn, csa> to <ndoms, doms> and populate cpu masks.
1036 */
1037 doms = alloc_sched_domains(ndoms);
1038 if (!doms)
1039 goto done;
1040
1041 /*
1042 * The rest of the code, including the scheduler, can deal with
1043 * dattr==NULL case. No need to abort if alloc fails.
1044 */
1045 dattr = kmalloc_array(ndoms, sizeof(struct sched_domain_attr),
1046 GFP_KERNEL);
1047
1048 for (nslot = 0, i = 0; i < csn; i++) {
1049 struct cpuset *a = csa[i];
1050 struct cpumask *dp;
1051 int apn = a->pn;
1052
1053 if (apn < 0) {
1054 /* Skip completed partitions */
1055 continue;
1056 }
1057
1058 dp = doms[nslot];
1059
1060 if (nslot == ndoms) {
1061 static int warnings = 10;
1062 if (warnings) {
1063 pr_warn("rebuild_sched_domains confused: nslot %d, ndoms %d, csn %d, i %d, apn %d\n",
1064 nslot, ndoms, csn, i, apn);
1065 warnings--;
1066 }
1067 continue;
1068 }
1069
1070 cpumask_clear(dp);
1071 if (dattr)
1072 *(dattr + nslot) = SD_ATTR_INIT;
1073 for (j = i; j < csn; j++) {
1074 struct cpuset *b = csa[j];
1075
1076 if (apn == b->pn) {
1077 cpumask_or(dp, dp, b->effective_cpus);
1078 cpumask_and(dp, dp, housekeeping_cpumask(HK_TYPE_DOMAIN));
1079 if (dattr)
1080 update_domain_attr_tree(dattr + nslot, b);
1081
1082 /* Done with this partition */
1083 b->pn = -1;
1084 }
1085 }
1086 nslot++;
1087 }
1088 BUG_ON(nslot != ndoms);
1089
1090 done:
1091 kfree(csa);
1092
1093 /*
1094 * Fallback to the default domain if kmalloc() failed.
1095 * See comments in partition_sched_domains().
1096 */
1097 if (doms == NULL)
1098 ndoms = 1;
1099
1100 *domains = doms;
1101 *attributes = dattr;
1102 return ndoms;
1103 }
1104
dl_update_tasks_root_domain(struct cpuset * cs)1105 static void dl_update_tasks_root_domain(struct cpuset *cs)
1106 {
1107 struct css_task_iter it;
1108 struct task_struct *task;
1109
1110 if (cs->nr_deadline_tasks == 0)
1111 return;
1112
1113 css_task_iter_start(&cs->css, 0, &it);
1114
1115 while ((task = css_task_iter_next(&it)))
1116 dl_add_task_root_domain(task);
1117
1118 css_task_iter_end(&it);
1119 }
1120
dl_rebuild_rd_accounting(void)1121 static void dl_rebuild_rd_accounting(void)
1122 {
1123 struct cpuset *cs = NULL;
1124 struct cgroup_subsys_state *pos_css;
1125
1126 lockdep_assert_held(&cpuset_mutex);
1127 lockdep_assert_cpus_held();
1128 lockdep_assert_held(&sched_domains_mutex);
1129
1130 rcu_read_lock();
1131
1132 /*
1133 * Clear default root domain DL accounting, it will be computed again
1134 * if a task belongs to it.
1135 */
1136 dl_clear_root_domain(&def_root_domain);
1137
1138 cpuset_for_each_descendant_pre(cs, pos_css, &top_cpuset) {
1139
1140 if (cpumask_empty(cs->effective_cpus)) {
1141 pos_css = css_rightmost_descendant(pos_css);
1142 continue;
1143 }
1144
1145 css_get(&cs->css);
1146
1147 rcu_read_unlock();
1148
1149 dl_update_tasks_root_domain(cs);
1150
1151 rcu_read_lock();
1152 css_put(&cs->css);
1153 }
1154 rcu_read_unlock();
1155 }
1156
1157 static void
partition_and_rebuild_sched_domains(int ndoms_new,cpumask_var_t doms_new[],struct sched_domain_attr * dattr_new)1158 partition_and_rebuild_sched_domains(int ndoms_new, cpumask_var_t doms_new[],
1159 struct sched_domain_attr *dattr_new)
1160 {
1161 mutex_lock(&sched_domains_mutex);
1162 partition_sched_domains_locked(ndoms_new, doms_new, dattr_new);
1163 dl_rebuild_rd_accounting();
1164 mutex_unlock(&sched_domains_mutex);
1165 }
1166
1167 /*
1168 * Rebuild scheduler domains.
1169 *
1170 * If the flag 'sched_load_balance' of any cpuset with non-empty
1171 * 'cpus' changes, or if the 'cpus' allowed changes in any cpuset
1172 * which has that flag enabled, or if any cpuset with a non-empty
1173 * 'cpus' is removed, then call this routine to rebuild the
1174 * scheduler's dynamic sched domains.
1175 *
1176 * Call with cpuset_mutex held. Takes cpus_read_lock().
1177 */
rebuild_sched_domains_locked(void)1178 static void rebuild_sched_domains_locked(void)
1179 {
1180 struct cgroup_subsys_state *pos_css;
1181 struct sched_domain_attr *attr;
1182 cpumask_var_t *doms;
1183 struct cpuset *cs;
1184 int ndoms;
1185
1186 lockdep_assert_cpus_held();
1187 lockdep_assert_held(&cpuset_mutex);
1188
1189 /*
1190 * If we have raced with CPU hotplug, return early to avoid
1191 * passing doms with offlined cpu to partition_sched_domains().
1192 * Anyways, cpuset_hotplug_workfn() will rebuild sched domains.
1193 *
1194 * With no CPUs in any subpartitions, top_cpuset's effective CPUs
1195 * should be the same as the active CPUs, so checking only top_cpuset
1196 * is enough to detect racing CPU offlines.
1197 */
1198 if (!top_cpuset.nr_subparts_cpus &&
1199 !cpumask_equal(top_cpuset.effective_cpus, cpu_active_mask))
1200 return;
1201
1202 /*
1203 * With subpartition CPUs, however, the effective CPUs of a partition
1204 * root should be only a subset of the active CPUs. Since a CPU in any
1205 * partition root could be offlined, all must be checked.
1206 */
1207 if (top_cpuset.nr_subparts_cpus) {
1208 rcu_read_lock();
1209 cpuset_for_each_descendant_pre(cs, pos_css, &top_cpuset) {
1210 if (!is_partition_valid(cs)) {
1211 pos_css = css_rightmost_descendant(pos_css);
1212 continue;
1213 }
1214 if (!cpumask_subset(cs->effective_cpus,
1215 cpu_active_mask)) {
1216 rcu_read_unlock();
1217 return;
1218 }
1219 }
1220 rcu_read_unlock();
1221 }
1222
1223 /* Generate domain masks and attrs */
1224 ndoms = generate_sched_domains(&doms, &attr);
1225
1226 /* Have scheduler rebuild the domains */
1227 partition_and_rebuild_sched_domains(ndoms, doms, attr);
1228 }
1229 #else /* !CONFIG_SMP */
rebuild_sched_domains_locked(void)1230 static void rebuild_sched_domains_locked(void)
1231 {
1232 }
1233 #endif /* CONFIG_SMP */
1234
rebuild_sched_domains(void)1235 void rebuild_sched_domains(void)
1236 {
1237 cpus_read_lock();
1238 mutex_lock(&cpuset_mutex);
1239 rebuild_sched_domains_locked();
1240 mutex_unlock(&cpuset_mutex);
1241 cpus_read_unlock();
1242 }
1243 EXPORT_SYMBOL_GPL(rebuild_sched_domains);
1244
update_cpus_allowed(struct cpuset * cs,struct task_struct * p,const struct cpumask * new_mask)1245 static int update_cpus_allowed(struct cpuset *cs, struct task_struct *p,
1246 const struct cpumask *new_mask)
1247 {
1248 int ret = -EINVAL;
1249
1250 trace_android_rvh_update_cpus_allowed(p, cs->cpus_requested, new_mask, &ret);
1251 if (!ret)
1252 return ret;
1253
1254 return set_cpus_allowed_ptr(p, new_mask);
1255 }
1256
1257 /**
1258 * update_tasks_cpumask - Update the cpumasks of tasks in the cpuset.
1259 * @cs: the cpuset in which each task's cpus_allowed mask needs to be changed
1260 * @new_cpus: the temp variable for the new effective_cpus mask
1261 *
1262 * Iterate through each task of @cs updating its cpus_allowed to the
1263 * effective cpuset's. As this function is called with cpuset_mutex held,
1264 * cpuset membership stays stable.
1265 */
update_tasks_cpumask(struct cpuset * cs,struct cpumask * new_cpus)1266 static void update_tasks_cpumask(struct cpuset *cs, struct cpumask *new_cpus)
1267 {
1268 struct css_task_iter it;
1269 struct task_struct *task;
1270 bool top_cs = cs == &top_cpuset;
1271
1272 css_task_iter_start(&cs->css, 0, &it);
1273 while ((task = css_task_iter_next(&it))) {
1274 /*
1275 * Percpu kthreads in top_cpuset are ignored
1276 */
1277 if (top_cs && (task->flags & PF_KTHREAD) &&
1278 kthread_is_per_cpu(task))
1279 continue;
1280
1281 cpumask_and(new_cpus, cs->effective_cpus,
1282 task_cpu_possible_mask(task));
1283 update_cpus_allowed(cs, task, new_cpus);
1284 }
1285 css_task_iter_end(&it);
1286 }
1287
1288 /**
1289 * compute_effective_cpumask - Compute the effective cpumask of the cpuset
1290 * @new_cpus: the temp variable for the new effective_cpus mask
1291 * @cs: the cpuset the need to recompute the new effective_cpus mask
1292 * @parent: the parent cpuset
1293 *
1294 * If the parent has subpartition CPUs, include them in the list of
1295 * allowable CPUs in computing the new effective_cpus mask. Since offlined
1296 * CPUs are not removed from subparts_cpus, we have to use cpu_active_mask
1297 * to mask those out.
1298 */
compute_effective_cpumask(struct cpumask * new_cpus,struct cpuset * cs,struct cpuset * parent)1299 static void compute_effective_cpumask(struct cpumask *new_cpus,
1300 struct cpuset *cs, struct cpuset *parent)
1301 {
1302 if (parent->nr_subparts_cpus) {
1303 cpumask_or(new_cpus, parent->effective_cpus,
1304 parent->subparts_cpus);
1305 cpumask_and(new_cpus, new_cpus, cs->cpus_requested);
1306 cpumask_and(new_cpus, new_cpus, cpu_active_mask);
1307 } else {
1308 cpumask_and(new_cpus, cs->cpus_requested, parent_cs(cs)->effective_cpus);
1309 }
1310 }
1311
1312 /*
1313 * Commands for update_parent_subparts_cpumask
1314 */
1315 enum subparts_cmd {
1316 partcmd_enable, /* Enable partition root */
1317 partcmd_disable, /* Disable partition root */
1318 partcmd_update, /* Update parent's subparts_cpus */
1319 partcmd_invalidate, /* Make partition invalid */
1320 };
1321
1322 static int update_flag(cpuset_flagbits_t bit, struct cpuset *cs,
1323 int turning_on);
1324 /**
1325 * update_parent_subparts_cpumask - update subparts_cpus mask of parent cpuset
1326 * @cpuset: The cpuset that requests change in partition root state
1327 * @cmd: Partition root state change command
1328 * @newmask: Optional new cpumask for partcmd_update
1329 * @tmp: Temporary addmask and delmask
1330 * Return: 0 or a partition root state error code
1331 *
1332 * For partcmd_enable, the cpuset is being transformed from a non-partition
1333 * root to a partition root. The cpus_allowed mask of the given cpuset will
1334 * be put into parent's subparts_cpus and taken away from parent's
1335 * effective_cpus. The function will return 0 if all the CPUs listed in
1336 * cpus_allowed can be granted or an error code will be returned.
1337 *
1338 * For partcmd_disable, the cpuset is being transformed from a partition
1339 * root back to a non-partition root. Any CPUs in cpus_allowed that are in
1340 * parent's subparts_cpus will be taken away from that cpumask and put back
1341 * into parent's effective_cpus. 0 will always be returned.
1342 *
1343 * For partcmd_update, if the optional newmask is specified, the cpu list is
1344 * to be changed from cpus_allowed to newmask. Otherwise, cpus_allowed is
1345 * assumed to remain the same. The cpuset should either be a valid or invalid
1346 * partition root. The partition root state may change from valid to invalid
1347 * or vice versa. An error code will only be returned if transitioning from
1348 * invalid to valid violates the exclusivity rule.
1349 *
1350 * For partcmd_invalidate, the current partition will be made invalid.
1351 *
1352 * The partcmd_enable and partcmd_disable commands are used by
1353 * update_prstate(). An error code may be returned and the caller will check
1354 * for error.
1355 *
1356 * The partcmd_update command is used by update_cpumasks_hier() with newmask
1357 * NULL and update_cpumask() with newmask set. The partcmd_invalidate is used
1358 * by update_cpumask() with NULL newmask. In both cases, the callers won't
1359 * check for error and so partition_root_state and prs_error will be updated
1360 * directly.
1361 */
update_parent_subparts_cpumask(struct cpuset * cs,int cmd,struct cpumask * newmask,struct tmpmasks * tmp)1362 static int update_parent_subparts_cpumask(struct cpuset *cs, int cmd,
1363 struct cpumask *newmask,
1364 struct tmpmasks *tmp)
1365 {
1366 struct cpuset *parent = parent_cs(cs);
1367 int adding; /* Moving cpus from effective_cpus to subparts_cpus */
1368 int deleting; /* Moving cpus from subparts_cpus to effective_cpus */
1369 int old_prs, new_prs;
1370 int part_error = PERR_NONE; /* Partition error? */
1371
1372 lockdep_assert_held(&cpuset_mutex);
1373
1374 /*
1375 * The parent must be a partition root.
1376 * The new cpumask, if present, or the current cpus_allowed must
1377 * not be empty.
1378 */
1379 if (!is_partition_valid(parent)) {
1380 return is_partition_invalid(parent)
1381 ? PERR_INVPARENT : PERR_NOTPART;
1382 }
1383 if ((newmask && cpumask_empty(newmask)) ||
1384 (!newmask && cpumask_empty(cs->cpus_allowed)))
1385 return PERR_CPUSEMPTY;
1386
1387 /*
1388 * new_prs will only be changed for the partcmd_update and
1389 * partcmd_invalidate commands.
1390 */
1391 adding = deleting = false;
1392 old_prs = new_prs = cs->partition_root_state;
1393 if (cmd == partcmd_enable) {
1394 /*
1395 * Enabling partition root is not allowed if cpus_allowed
1396 * doesn't overlap parent's cpus_allowed.
1397 */
1398 if (!cpumask_intersects(cs->cpus_allowed, parent->cpus_allowed))
1399 return PERR_INVCPUS;
1400
1401 /*
1402 * A parent can be left with no CPU as long as there is no
1403 * task directly associated with the parent partition.
1404 */
1405 if (cpumask_subset(parent->effective_cpus, cs->cpus_allowed) &&
1406 partition_is_populated(parent, cs))
1407 return PERR_NOCPUS;
1408
1409 cpumask_copy(tmp->addmask, cs->cpus_allowed);
1410 adding = true;
1411 } else if (cmd == partcmd_disable) {
1412 /*
1413 * Need to remove cpus from parent's subparts_cpus for valid
1414 * partition root.
1415 */
1416 deleting = !is_prs_invalid(old_prs) &&
1417 cpumask_and(tmp->delmask, cs->cpus_allowed,
1418 parent->subparts_cpus);
1419 } else if (cmd == partcmd_invalidate) {
1420 if (is_prs_invalid(old_prs))
1421 return 0;
1422
1423 /*
1424 * Make the current partition invalid. It is assumed that
1425 * invalidation is caused by violating cpu exclusivity rule.
1426 */
1427 deleting = cpumask_and(tmp->delmask, cs->cpus_allowed,
1428 parent->subparts_cpus);
1429 if (old_prs > 0) {
1430 new_prs = -old_prs;
1431 part_error = PERR_NOTEXCL;
1432 }
1433 } else if (newmask) {
1434 /*
1435 * partcmd_update with newmask:
1436 *
1437 * Compute add/delete mask to/from subparts_cpus
1438 *
1439 * delmask = cpus_allowed & ~newmask & parent->subparts_cpus
1440 * addmask = newmask & parent->cpus_allowed
1441 * & ~parent->subparts_cpus
1442 */
1443 cpumask_andnot(tmp->delmask, cs->cpus_allowed, newmask);
1444 deleting = cpumask_and(tmp->delmask, tmp->delmask,
1445 parent->subparts_cpus);
1446
1447 cpumask_and(tmp->addmask, newmask, parent->cpus_allowed);
1448 adding = cpumask_andnot(tmp->addmask, tmp->addmask,
1449 parent->subparts_cpus);
1450 /*
1451 * Make partition invalid if parent's effective_cpus could
1452 * become empty and there are tasks in the parent.
1453 */
1454 if (adding &&
1455 cpumask_subset(parent->effective_cpus, tmp->addmask) &&
1456 !cpumask_intersects(tmp->delmask, cpu_active_mask) &&
1457 partition_is_populated(parent, cs)) {
1458 part_error = PERR_NOCPUS;
1459 adding = false;
1460 deleting = cpumask_and(tmp->delmask, cs->cpus_allowed,
1461 parent->subparts_cpus);
1462 }
1463 } else {
1464 /*
1465 * partcmd_update w/o newmask:
1466 *
1467 * delmask = cpus_allowed & parent->subparts_cpus
1468 * addmask = cpus_allowed & parent->cpus_allowed
1469 * & ~parent->subparts_cpus
1470 *
1471 * This gets invoked either due to a hotplug event or from
1472 * update_cpumasks_hier(). This can cause the state of a
1473 * partition root to transition from valid to invalid or vice
1474 * versa. So we still need to compute the addmask and delmask.
1475
1476 * A partition error happens when:
1477 * 1) Cpuset is valid partition, but parent does not distribute
1478 * out any CPUs.
1479 * 2) Parent has tasks and all its effective CPUs will have
1480 * to be distributed out.
1481 */
1482 cpumask_and(tmp->addmask, cs->cpus_allowed,
1483 parent->cpus_allowed);
1484 adding = cpumask_andnot(tmp->addmask, tmp->addmask,
1485 parent->subparts_cpus);
1486
1487 if ((is_partition_valid(cs) && !parent->nr_subparts_cpus) ||
1488 (adding &&
1489 cpumask_subset(parent->effective_cpus, tmp->addmask) &&
1490 partition_is_populated(parent, cs))) {
1491 part_error = PERR_NOCPUS;
1492 adding = false;
1493 }
1494
1495 if (part_error && is_partition_valid(cs) &&
1496 parent->nr_subparts_cpus)
1497 deleting = cpumask_and(tmp->delmask, cs->cpus_allowed,
1498 parent->subparts_cpus);
1499 }
1500 if (part_error)
1501 WRITE_ONCE(cs->prs_err, part_error);
1502
1503 if (cmd == partcmd_update) {
1504 /*
1505 * Check for possible transition between valid and invalid
1506 * partition root.
1507 */
1508 switch (cs->partition_root_state) {
1509 case PRS_ROOT:
1510 case PRS_ISOLATED:
1511 if (part_error)
1512 new_prs = -old_prs;
1513 break;
1514 case PRS_INVALID_ROOT:
1515 case PRS_INVALID_ISOLATED:
1516 if (!part_error)
1517 new_prs = -old_prs;
1518 break;
1519 }
1520 }
1521
1522 if (!adding && !deleting && (new_prs == old_prs))
1523 return 0;
1524
1525 /*
1526 * Transitioning between invalid to valid or vice versa may require
1527 * changing CS_CPU_EXCLUSIVE and CS_SCHED_LOAD_BALANCE.
1528 */
1529 if (old_prs != new_prs) {
1530 if (is_prs_invalid(old_prs) && !is_cpu_exclusive(cs) &&
1531 (update_flag(CS_CPU_EXCLUSIVE, cs, 1) < 0))
1532 return PERR_NOTEXCL;
1533 if (is_prs_invalid(new_prs) && is_cpu_exclusive(cs))
1534 update_flag(CS_CPU_EXCLUSIVE, cs, 0);
1535 }
1536
1537 /*
1538 * Change the parent's subparts_cpus.
1539 * Newly added CPUs will be removed from effective_cpus and
1540 * newly deleted ones will be added back to effective_cpus.
1541 */
1542 spin_lock_irq(&callback_lock);
1543 if (adding) {
1544 cpumask_or(parent->subparts_cpus,
1545 parent->subparts_cpus, tmp->addmask);
1546 cpumask_andnot(parent->effective_cpus,
1547 parent->effective_cpus, tmp->addmask);
1548 }
1549 if (deleting) {
1550 cpumask_andnot(parent->subparts_cpus,
1551 parent->subparts_cpus, tmp->delmask);
1552 /*
1553 * Some of the CPUs in subparts_cpus might have been offlined.
1554 */
1555 cpumask_and(tmp->delmask, tmp->delmask, cpu_active_mask);
1556 cpumask_or(parent->effective_cpus,
1557 parent->effective_cpus, tmp->delmask);
1558 }
1559
1560 parent->nr_subparts_cpus = cpumask_weight(parent->subparts_cpus);
1561
1562 if (old_prs != new_prs)
1563 cs->partition_root_state = new_prs;
1564
1565 spin_unlock_irq(&callback_lock);
1566
1567 if (adding || deleting)
1568 update_tasks_cpumask(parent, tmp->addmask);
1569
1570 /*
1571 * Set or clear CS_SCHED_LOAD_BALANCE when partcmd_update, if necessary.
1572 * rebuild_sched_domains_locked() may be called.
1573 */
1574 if (old_prs != new_prs) {
1575 if (old_prs == PRS_ISOLATED)
1576 update_flag(CS_SCHED_LOAD_BALANCE, cs, 1);
1577 else if (new_prs == PRS_ISOLATED)
1578 update_flag(CS_SCHED_LOAD_BALANCE, cs, 0);
1579 }
1580 notify_partition_change(cs, old_prs);
1581 return 0;
1582 }
1583
1584 /*
1585 * update_cpumasks_hier - Update effective cpumasks and tasks in the subtree
1586 * @cs: the cpuset to consider
1587 * @tmp: temp variables for calculating effective_cpus & partition setup
1588 * @force: don't skip any descendant cpusets if set
1589 *
1590 * When configured cpumask is changed, the effective cpumasks of this cpuset
1591 * and all its descendants need to be updated.
1592 *
1593 * On legacy hierarchy, effective_cpus will be the same with cpu_allowed.
1594 *
1595 * Called with cpuset_mutex held
1596 */
update_cpumasks_hier(struct cpuset * cs,struct tmpmasks * tmp,bool force)1597 static void update_cpumasks_hier(struct cpuset *cs, struct tmpmasks *tmp,
1598 bool force)
1599 {
1600 struct cpuset *cp;
1601 struct cgroup_subsys_state *pos_css;
1602 bool need_rebuild_sched_domains = false;
1603 int old_prs, new_prs;
1604
1605 rcu_read_lock();
1606 cpuset_for_each_descendant_pre(cp, pos_css, cs) {
1607 struct cpuset *parent = parent_cs(cp);
1608 bool update_parent = false;
1609
1610 compute_effective_cpumask(tmp->new_cpus, cp, parent);
1611
1612 /*
1613 * If it becomes empty, inherit the effective mask of the
1614 * parent, which is guaranteed to have some CPUs unless
1615 * it is a partition root that has explicitly distributed
1616 * out all its CPUs.
1617 */
1618 if (is_in_v2_mode() && cpumask_empty(tmp->new_cpus)) {
1619 if (is_partition_valid(cp) &&
1620 cpumask_equal(cp->cpus_allowed, cp->subparts_cpus))
1621 goto update_parent_subparts;
1622
1623 cpumask_copy(tmp->new_cpus, parent->effective_cpus);
1624 if (!cp->use_parent_ecpus) {
1625 cp->use_parent_ecpus = true;
1626 parent->child_ecpus_count++;
1627 }
1628 } else if (cp->use_parent_ecpus) {
1629 cp->use_parent_ecpus = false;
1630 WARN_ON_ONCE(!parent->child_ecpus_count);
1631 parent->child_ecpus_count--;
1632 }
1633
1634 /*
1635 * Skip the whole subtree if
1636 * 1) the cpumask remains the same,
1637 * 2) has no partition root state,
1638 * 3) force flag not set, and
1639 * 4) for v2 load balance state same as its parent.
1640 */
1641 if (!cp->partition_root_state && !force &&
1642 cpumask_equal(tmp->new_cpus, cp->effective_cpus) &&
1643 (!cgroup_subsys_on_dfl(cpuset_cgrp_subsys) ||
1644 (is_sched_load_balance(parent) == is_sched_load_balance(cp)))) {
1645 pos_css = css_rightmost_descendant(pos_css);
1646 continue;
1647 }
1648
1649 update_parent_subparts:
1650 /*
1651 * update_parent_subparts_cpumask() should have been called
1652 * for cs already in update_cpumask(). We should also call
1653 * update_tasks_cpumask() again for tasks in the parent
1654 * cpuset if the parent's subparts_cpus changes.
1655 */
1656 old_prs = new_prs = cp->partition_root_state;
1657 if ((cp != cs) && old_prs) {
1658 switch (parent->partition_root_state) {
1659 case PRS_ROOT:
1660 case PRS_ISOLATED:
1661 update_parent = true;
1662 break;
1663
1664 default:
1665 /*
1666 * When parent is not a partition root or is
1667 * invalid, child partition roots become
1668 * invalid too.
1669 */
1670 if (is_partition_valid(cp))
1671 new_prs = -cp->partition_root_state;
1672 WRITE_ONCE(cp->prs_err,
1673 is_partition_invalid(parent)
1674 ? PERR_INVPARENT : PERR_NOTPART);
1675 break;
1676 }
1677 }
1678
1679 if (!css_tryget_online(&cp->css))
1680 continue;
1681 rcu_read_unlock();
1682
1683 if (update_parent) {
1684 update_parent_subparts_cpumask(cp, partcmd_update, NULL,
1685 tmp);
1686 /*
1687 * The cpuset partition_root_state may become
1688 * invalid. Capture it.
1689 */
1690 new_prs = cp->partition_root_state;
1691 }
1692
1693 spin_lock_irq(&callback_lock);
1694
1695 if (cp->nr_subparts_cpus && !is_partition_valid(cp)) {
1696 /*
1697 * Put all active subparts_cpus back to effective_cpus.
1698 */
1699 cpumask_or(tmp->new_cpus, tmp->new_cpus,
1700 cp->subparts_cpus);
1701 cpumask_and(tmp->new_cpus, tmp->new_cpus,
1702 cpu_active_mask);
1703 cp->nr_subparts_cpus = 0;
1704 cpumask_clear(cp->subparts_cpus);
1705 }
1706
1707 cpumask_copy(cp->effective_cpus, tmp->new_cpus);
1708 if (cp->nr_subparts_cpus) {
1709 /*
1710 * Make sure that effective_cpus & subparts_cpus
1711 * are mutually exclusive.
1712 */
1713 cpumask_andnot(cp->effective_cpus, cp->effective_cpus,
1714 cp->subparts_cpus);
1715 }
1716
1717 cp->partition_root_state = new_prs;
1718 spin_unlock_irq(&callback_lock);
1719
1720 notify_partition_change(cp, old_prs);
1721
1722 WARN_ON(!is_in_v2_mode() &&
1723 !cpumask_equal(cp->cpus_allowed, cp->effective_cpus));
1724
1725 update_tasks_cpumask(cp, tmp->new_cpus);
1726
1727 /*
1728 * On default hierarchy, inherit the CS_SCHED_LOAD_BALANCE
1729 * from parent if current cpuset isn't a valid partition root
1730 * and their load balance states differ.
1731 */
1732 if (cgroup_subsys_on_dfl(cpuset_cgrp_subsys) &&
1733 !is_partition_valid(cp) &&
1734 (is_sched_load_balance(parent) != is_sched_load_balance(cp))) {
1735 if (is_sched_load_balance(parent))
1736 set_bit(CS_SCHED_LOAD_BALANCE, &cp->flags);
1737 else
1738 clear_bit(CS_SCHED_LOAD_BALANCE, &cp->flags);
1739 }
1740
1741 /*
1742 * On legacy hierarchy, if the effective cpumask of any non-
1743 * empty cpuset is changed, we need to rebuild sched domains.
1744 * On default hierarchy, the cpuset needs to be a partition
1745 * root as well.
1746 */
1747 if (!cpumask_empty(cp->cpus_allowed) &&
1748 is_sched_load_balance(cp) &&
1749 (!cgroup_subsys_on_dfl(cpuset_cgrp_subsys) ||
1750 is_partition_valid(cp)))
1751 need_rebuild_sched_domains = true;
1752
1753 rcu_read_lock();
1754 css_put(&cp->css);
1755 }
1756 rcu_read_unlock();
1757
1758 if (need_rebuild_sched_domains)
1759 rebuild_sched_domains_locked();
1760 }
1761
1762 /**
1763 * update_sibling_cpumasks - Update siblings cpumasks
1764 * @parent: Parent cpuset
1765 * @cs: Current cpuset
1766 * @tmp: Temp variables
1767 */
update_sibling_cpumasks(struct cpuset * parent,struct cpuset * cs,struct tmpmasks * tmp)1768 static void update_sibling_cpumasks(struct cpuset *parent, struct cpuset *cs,
1769 struct tmpmasks *tmp)
1770 {
1771 struct cpuset *sibling;
1772 struct cgroup_subsys_state *pos_css;
1773
1774 lockdep_assert_held(&cpuset_mutex);
1775
1776 /*
1777 * Check all its siblings and call update_cpumasks_hier()
1778 * if their use_parent_ecpus flag is set in order for them
1779 * to use the right effective_cpus value.
1780 *
1781 * The update_cpumasks_hier() function may sleep. So we have to
1782 * release the RCU read lock before calling it.
1783 */
1784 rcu_read_lock();
1785 cpuset_for_each_child(sibling, pos_css, parent) {
1786 if (sibling == cs)
1787 continue;
1788 if (!sibling->use_parent_ecpus)
1789 continue;
1790 if (!css_tryget_online(&sibling->css))
1791 continue;
1792
1793 rcu_read_unlock();
1794 update_cpumasks_hier(sibling, tmp, false);
1795 rcu_read_lock();
1796 css_put(&sibling->css);
1797 }
1798 rcu_read_unlock();
1799 }
1800
1801 /**
1802 * update_cpumask - update the cpus_allowed mask of a cpuset and all tasks in it
1803 * @cs: the cpuset to consider
1804 * @trialcs: trial cpuset
1805 * @buf: buffer of cpu numbers written to this cpuset
1806 */
update_cpumask(struct cpuset * cs,struct cpuset * trialcs,const char * buf)1807 static int update_cpumask(struct cpuset *cs, struct cpuset *trialcs,
1808 const char *buf)
1809 {
1810 int retval;
1811 struct tmpmasks tmp;
1812 bool invalidate = false;
1813
1814 /* top_cpuset.cpus_allowed tracks cpu_online_mask; it's read-only */
1815 if (cs == &top_cpuset)
1816 return -EACCES;
1817
1818 /*
1819 * An empty cpus_requested is ok only if the cpuset has no tasks.
1820 * Since cpulist_parse() fails on an empty mask, we special case
1821 * that parsing. The validate_change() call ensures that cpusets
1822 * with tasks have cpus.
1823 */
1824 if (!*buf) {
1825 cpumask_clear(trialcs->cpus_requested);
1826 } else {
1827 retval = cpulist_parse(buf, trialcs->cpus_requested);
1828 if (retval < 0)
1829 return retval;
1830 }
1831
1832 if (!cpumask_subset(trialcs->cpus_requested, cpu_present_mask))
1833 return -EINVAL;
1834
1835 cpumask_and(trialcs->cpus_allowed, trialcs->cpus_requested, cpu_active_mask);
1836
1837 /* Nothing to do if the cpus didn't change */
1838 if (cpumask_equal(cs->cpus_requested, trialcs->cpus_requested))
1839 return 0;
1840
1841 #ifdef CONFIG_CPUMASK_OFFSTACK
1842 /*
1843 * Use the cpumasks in trialcs for tmpmasks when they are pointers
1844 * to allocated cpumasks.
1845 *
1846 * Note that update_parent_subparts_cpumask() uses only addmask &
1847 * delmask, but not new_cpus.
1848 */
1849 tmp.addmask = trialcs->subparts_cpus;
1850 tmp.delmask = trialcs->effective_cpus;
1851 tmp.new_cpus = NULL;
1852 #endif
1853
1854 retval = validate_change(cs, trialcs);
1855
1856 if ((retval == -EINVAL) && cgroup_subsys_on_dfl(cpuset_cgrp_subsys)) {
1857 struct cpuset *cp, *parent;
1858 struct cgroup_subsys_state *css;
1859
1860 /*
1861 * The -EINVAL error code indicates that partition sibling
1862 * CPU exclusivity rule has been violated. We still allow
1863 * the cpumask change to proceed while invalidating the
1864 * partition. However, any conflicting sibling partitions
1865 * have to be marked as invalid too.
1866 */
1867 invalidate = true;
1868 rcu_read_lock();
1869 parent = parent_cs(cs);
1870 cpuset_for_each_child(cp, css, parent)
1871 if (is_partition_valid(cp) &&
1872 cpumask_intersects(trialcs->cpus_allowed, cp->cpus_allowed)) {
1873 rcu_read_unlock();
1874 update_parent_subparts_cpumask(cp, partcmd_invalidate, NULL, &tmp);
1875 rcu_read_lock();
1876 }
1877 rcu_read_unlock();
1878 retval = 0;
1879 }
1880 if (retval < 0)
1881 return retval;
1882
1883 if (cs->partition_root_state) {
1884 if (invalidate)
1885 update_parent_subparts_cpumask(cs, partcmd_invalidate,
1886 NULL, &tmp);
1887 else
1888 update_parent_subparts_cpumask(cs, partcmd_update,
1889 trialcs->cpus_allowed, &tmp);
1890 }
1891
1892 compute_effective_cpumask(trialcs->effective_cpus, trialcs,
1893 parent_cs(cs));
1894 spin_lock_irq(&callback_lock);
1895 cpumask_copy(cs->cpus_allowed, trialcs->cpus_allowed);
1896 cpumask_copy(cs->cpus_requested, trialcs->cpus_requested);
1897
1898 /*
1899 * Make sure that subparts_cpus, if not empty, is a subset of
1900 * cpus_allowed. Clear subparts_cpus if partition not valid or
1901 * empty effective cpus with tasks.
1902 */
1903 if (cs->nr_subparts_cpus) {
1904 if (!is_partition_valid(cs) ||
1905 (cpumask_subset(trialcs->effective_cpus, cs->subparts_cpus) &&
1906 partition_is_populated(cs, NULL))) {
1907 cs->nr_subparts_cpus = 0;
1908 cpumask_clear(cs->subparts_cpus);
1909 } else {
1910 cpumask_and(cs->subparts_cpus, cs->subparts_cpus,
1911 cs->cpus_allowed);
1912 cs->nr_subparts_cpus = cpumask_weight(cs->subparts_cpus);
1913 }
1914 }
1915 spin_unlock_irq(&callback_lock);
1916
1917 #ifdef CONFIG_CPUMASK_OFFSTACK
1918 /* Now trialcs->cpus_allowed is available */
1919 tmp.new_cpus = trialcs->cpus_allowed;
1920 #endif
1921
1922 /* effective_cpus will be updated here */
1923 update_cpumasks_hier(cs, &tmp, false);
1924
1925 if (cs->partition_root_state) {
1926 struct cpuset *parent = parent_cs(cs);
1927
1928 /*
1929 * For partition root, update the cpumasks of sibling
1930 * cpusets if they use parent's effective_cpus.
1931 */
1932 if (parent->child_ecpus_count)
1933 update_sibling_cpumasks(parent, cs, &tmp);
1934 }
1935 return 0;
1936 }
1937
1938 /*
1939 * Migrate memory region from one set of nodes to another. This is
1940 * performed asynchronously as it can be called from process migration path
1941 * holding locks involved in process management. All mm migrations are
1942 * performed in the queued order and can be waited for by flushing
1943 * cpuset_migrate_mm_wq.
1944 */
1945
1946 struct cpuset_migrate_mm_work {
1947 struct work_struct work;
1948 struct mm_struct *mm;
1949 nodemask_t from;
1950 nodemask_t to;
1951 };
1952
cpuset_migrate_mm_workfn(struct work_struct * work)1953 static void cpuset_migrate_mm_workfn(struct work_struct *work)
1954 {
1955 struct cpuset_migrate_mm_work *mwork =
1956 container_of(work, struct cpuset_migrate_mm_work, work);
1957
1958 /* on a wq worker, no need to worry about %current's mems_allowed */
1959 do_migrate_pages(mwork->mm, &mwork->from, &mwork->to, MPOL_MF_MOVE_ALL);
1960 mmput(mwork->mm);
1961 kfree(mwork);
1962 }
1963
cpuset_migrate_mm(struct mm_struct * mm,const nodemask_t * from,const nodemask_t * to)1964 static void cpuset_migrate_mm(struct mm_struct *mm, const nodemask_t *from,
1965 const nodemask_t *to)
1966 {
1967 struct cpuset_migrate_mm_work *mwork;
1968
1969 if (nodes_equal(*from, *to)) {
1970 mmput(mm);
1971 return;
1972 }
1973
1974 mwork = kzalloc(sizeof(*mwork), GFP_KERNEL);
1975 if (mwork) {
1976 mwork->mm = mm;
1977 mwork->from = *from;
1978 mwork->to = *to;
1979 INIT_WORK(&mwork->work, cpuset_migrate_mm_workfn);
1980 queue_work(cpuset_migrate_mm_wq, &mwork->work);
1981 } else {
1982 mmput(mm);
1983 }
1984 }
1985
cpuset_post_attach(void)1986 static void cpuset_post_attach(void)
1987 {
1988 flush_workqueue(cpuset_migrate_mm_wq);
1989 }
1990
1991 /*
1992 * cpuset_change_task_nodemask - change task's mems_allowed and mempolicy
1993 * @tsk: the task to change
1994 * @newmems: new nodes that the task will be set
1995 *
1996 * We use the mems_allowed_seq seqlock to safely update both tsk->mems_allowed
1997 * and rebind an eventual tasks' mempolicy. If the task is allocating in
1998 * parallel, it might temporarily see an empty intersection, which results in
1999 * a seqlock check and retry before OOM or allocation failure.
2000 */
cpuset_change_task_nodemask(struct task_struct * tsk,nodemask_t * newmems)2001 static void cpuset_change_task_nodemask(struct task_struct *tsk,
2002 nodemask_t *newmems)
2003 {
2004 task_lock(tsk);
2005
2006 local_irq_disable();
2007 write_seqcount_begin(&tsk->mems_allowed_seq);
2008
2009 nodes_or(tsk->mems_allowed, tsk->mems_allowed, *newmems);
2010 mpol_rebind_task(tsk, newmems);
2011 tsk->mems_allowed = *newmems;
2012
2013 write_seqcount_end(&tsk->mems_allowed_seq);
2014 local_irq_enable();
2015
2016 task_unlock(tsk);
2017 }
2018
2019 static void *cpuset_being_rebound;
2020
2021 /**
2022 * update_tasks_nodemask - Update the nodemasks of tasks in the cpuset.
2023 * @cs: the cpuset in which each task's mems_allowed mask needs to be changed
2024 *
2025 * Iterate through each task of @cs updating its mems_allowed to the
2026 * effective cpuset's. As this function is called with cpuset_mutex held,
2027 * cpuset membership stays stable.
2028 */
update_tasks_nodemask(struct cpuset * cs)2029 static void update_tasks_nodemask(struct cpuset *cs)
2030 {
2031 static nodemask_t newmems; /* protected by cpuset_mutex */
2032 struct css_task_iter it;
2033 struct task_struct *task;
2034
2035 cpuset_being_rebound = cs; /* causes mpol_dup() rebind */
2036
2037 guarantee_online_mems(cs, &newmems);
2038
2039 /*
2040 * The mpol_rebind_mm() call takes mmap_lock, which we couldn't
2041 * take while holding tasklist_lock. Forks can happen - the
2042 * mpol_dup() cpuset_being_rebound check will catch such forks,
2043 * and rebind their vma mempolicies too. Because we still hold
2044 * the global cpuset_mutex, we know that no other rebind effort
2045 * will be contending for the global variable cpuset_being_rebound.
2046 * It's ok if we rebind the same mm twice; mpol_rebind_mm()
2047 * is idempotent. Also migrate pages in each mm to new nodes.
2048 */
2049 css_task_iter_start(&cs->css, 0, &it);
2050 while ((task = css_task_iter_next(&it))) {
2051 struct mm_struct *mm;
2052 bool migrate;
2053
2054 cpuset_change_task_nodemask(task, &newmems);
2055
2056 mm = get_task_mm(task);
2057 if (!mm)
2058 continue;
2059
2060 migrate = is_memory_migrate(cs);
2061
2062 mpol_rebind_mm(mm, &cs->mems_allowed);
2063 if (migrate)
2064 cpuset_migrate_mm(mm, &cs->old_mems_allowed, &newmems);
2065 else
2066 mmput(mm);
2067 }
2068 css_task_iter_end(&it);
2069
2070 /*
2071 * All the tasks' nodemasks have been updated, update
2072 * cs->old_mems_allowed.
2073 */
2074 cs->old_mems_allowed = newmems;
2075
2076 /* We're done rebinding vmas to this cpuset's new mems_allowed. */
2077 cpuset_being_rebound = NULL;
2078 }
2079
2080 /*
2081 * update_nodemasks_hier - Update effective nodemasks and tasks in the subtree
2082 * @cs: the cpuset to consider
2083 * @new_mems: a temp variable for calculating new effective_mems
2084 *
2085 * When configured nodemask is changed, the effective nodemasks of this cpuset
2086 * and all its descendants need to be updated.
2087 *
2088 * On legacy hierarchy, effective_mems will be the same with mems_allowed.
2089 *
2090 * Called with cpuset_mutex held
2091 */
update_nodemasks_hier(struct cpuset * cs,nodemask_t * new_mems)2092 static void update_nodemasks_hier(struct cpuset *cs, nodemask_t *new_mems)
2093 {
2094 struct cpuset *cp;
2095 struct cgroup_subsys_state *pos_css;
2096
2097 rcu_read_lock();
2098 cpuset_for_each_descendant_pre(cp, pos_css, cs) {
2099 struct cpuset *parent = parent_cs(cp);
2100
2101 nodes_and(*new_mems, cp->mems_allowed, parent->effective_mems);
2102
2103 /*
2104 * If it becomes empty, inherit the effective mask of the
2105 * parent, which is guaranteed to have some MEMs.
2106 */
2107 if (is_in_v2_mode() && nodes_empty(*new_mems))
2108 *new_mems = parent->effective_mems;
2109
2110 /* Skip the whole subtree if the nodemask remains the same. */
2111 if (nodes_equal(*new_mems, cp->effective_mems)) {
2112 pos_css = css_rightmost_descendant(pos_css);
2113 continue;
2114 }
2115
2116 if (!css_tryget_online(&cp->css))
2117 continue;
2118 rcu_read_unlock();
2119
2120 spin_lock_irq(&callback_lock);
2121 cp->effective_mems = *new_mems;
2122 spin_unlock_irq(&callback_lock);
2123
2124 WARN_ON(!is_in_v2_mode() &&
2125 !nodes_equal(cp->mems_allowed, cp->effective_mems));
2126
2127 update_tasks_nodemask(cp);
2128
2129 rcu_read_lock();
2130 css_put(&cp->css);
2131 }
2132 rcu_read_unlock();
2133 }
2134
2135 /*
2136 * Handle user request to change the 'mems' memory placement
2137 * of a cpuset. Needs to validate the request, update the
2138 * cpusets mems_allowed, and for each task in the cpuset,
2139 * update mems_allowed and rebind task's mempolicy and any vma
2140 * mempolicies and if the cpuset is marked 'memory_migrate',
2141 * migrate the tasks pages to the new memory.
2142 *
2143 * Call with cpuset_mutex held. May take callback_lock during call.
2144 * Will take tasklist_lock, scan tasklist for tasks in cpuset cs,
2145 * lock each such tasks mm->mmap_lock, scan its vma's and rebind
2146 * their mempolicies to the cpusets new mems_allowed.
2147 */
update_nodemask(struct cpuset * cs,struct cpuset * trialcs,const char * buf)2148 static int update_nodemask(struct cpuset *cs, struct cpuset *trialcs,
2149 const char *buf)
2150 {
2151 int retval;
2152
2153 /*
2154 * top_cpuset.mems_allowed tracks node_stats[N_MEMORY];
2155 * it's read-only
2156 */
2157 if (cs == &top_cpuset) {
2158 retval = -EACCES;
2159 goto done;
2160 }
2161
2162 /*
2163 * An empty mems_allowed is ok iff there are no tasks in the cpuset.
2164 * Since nodelist_parse() fails on an empty mask, we special case
2165 * that parsing. The validate_change() call ensures that cpusets
2166 * with tasks have memory.
2167 */
2168 if (!*buf) {
2169 nodes_clear(trialcs->mems_allowed);
2170 } else {
2171 retval = nodelist_parse(buf, trialcs->mems_allowed);
2172 if (retval < 0)
2173 goto done;
2174
2175 if (!nodes_subset(trialcs->mems_allowed,
2176 top_cpuset.mems_allowed)) {
2177 retval = -EINVAL;
2178 goto done;
2179 }
2180 }
2181
2182 if (nodes_equal(cs->mems_allowed, trialcs->mems_allowed)) {
2183 retval = 0; /* Too easy - nothing to do */
2184 goto done;
2185 }
2186 retval = validate_change(cs, trialcs);
2187 if (retval < 0)
2188 goto done;
2189
2190 check_insane_mems_config(&trialcs->mems_allowed);
2191
2192 spin_lock_irq(&callback_lock);
2193 cs->mems_allowed = trialcs->mems_allowed;
2194 spin_unlock_irq(&callback_lock);
2195
2196 /* use trialcs->mems_allowed as a temp variable */
2197 update_nodemasks_hier(cs, &trialcs->mems_allowed);
2198 done:
2199 return retval;
2200 }
2201
current_cpuset_is_being_rebound(void)2202 bool current_cpuset_is_being_rebound(void)
2203 {
2204 bool ret;
2205
2206 rcu_read_lock();
2207 ret = task_cs(current) == cpuset_being_rebound;
2208 rcu_read_unlock();
2209
2210 return ret;
2211 }
2212
update_relax_domain_level(struct cpuset * cs,s64 val)2213 static int update_relax_domain_level(struct cpuset *cs, s64 val)
2214 {
2215 #ifdef CONFIG_SMP
2216 if (val < -1 || val >= sched_domain_level_max)
2217 return -EINVAL;
2218 #endif
2219
2220 if (val != cs->relax_domain_level) {
2221 cs->relax_domain_level = val;
2222 if (!cpumask_empty(cs->cpus_allowed) &&
2223 is_sched_load_balance(cs))
2224 rebuild_sched_domains_locked();
2225 }
2226
2227 return 0;
2228 }
2229
2230 /**
2231 * update_tasks_flags - update the spread flags of tasks in the cpuset.
2232 * @cs: the cpuset in which each task's spread flags needs to be changed
2233 *
2234 * Iterate through each task of @cs updating its spread flags. As this
2235 * function is called with cpuset_mutex held, cpuset membership stays
2236 * stable.
2237 */
update_tasks_flags(struct cpuset * cs)2238 static void update_tasks_flags(struct cpuset *cs)
2239 {
2240 struct css_task_iter it;
2241 struct task_struct *task;
2242
2243 css_task_iter_start(&cs->css, 0, &it);
2244 while ((task = css_task_iter_next(&it)))
2245 cpuset_update_task_spread_flags(cs, task);
2246 css_task_iter_end(&it);
2247 }
2248
2249 /*
2250 * update_flag - read a 0 or a 1 in a file and update associated flag
2251 * bit: the bit to update (see cpuset_flagbits_t)
2252 * cs: the cpuset to update
2253 * turning_on: whether the flag is being set or cleared
2254 *
2255 * Call with cpuset_mutex held.
2256 */
2257
update_flag(cpuset_flagbits_t bit,struct cpuset * cs,int turning_on)2258 static int update_flag(cpuset_flagbits_t bit, struct cpuset *cs,
2259 int turning_on)
2260 {
2261 struct cpuset *trialcs;
2262 int balance_flag_changed;
2263 int spread_flag_changed;
2264 int err;
2265
2266 trialcs = alloc_trial_cpuset(cs);
2267 if (!trialcs)
2268 return -ENOMEM;
2269
2270 if (turning_on)
2271 set_bit(bit, &trialcs->flags);
2272 else
2273 clear_bit(bit, &trialcs->flags);
2274
2275 err = validate_change(cs, trialcs);
2276 if (err < 0)
2277 goto out;
2278
2279 balance_flag_changed = (is_sched_load_balance(cs) !=
2280 is_sched_load_balance(trialcs));
2281
2282 spread_flag_changed = ((is_spread_slab(cs) != is_spread_slab(trialcs))
2283 || (is_spread_page(cs) != is_spread_page(trialcs)));
2284
2285 spin_lock_irq(&callback_lock);
2286 cs->flags = trialcs->flags;
2287 spin_unlock_irq(&callback_lock);
2288
2289 if (!cpumask_empty(trialcs->cpus_allowed) && balance_flag_changed)
2290 rebuild_sched_domains_locked();
2291
2292 if (spread_flag_changed)
2293 update_tasks_flags(cs);
2294 out:
2295 free_cpuset(trialcs);
2296 return err;
2297 }
2298
2299 /**
2300 * update_prstate - update partition_root_state
2301 * @cs: the cpuset to update
2302 * @new_prs: new partition root state
2303 * Return: 0 if successful, != 0 if error
2304 *
2305 * Call with cpuset_mutex held.
2306 */
update_prstate(struct cpuset * cs,int new_prs)2307 static int update_prstate(struct cpuset *cs, int new_prs)
2308 {
2309 int err = PERR_NONE, old_prs = cs->partition_root_state;
2310 bool sched_domain_rebuilt = false;
2311 struct cpuset *parent = parent_cs(cs);
2312 struct tmpmasks tmpmask;
2313
2314 if (old_prs == new_prs)
2315 return 0;
2316
2317 /*
2318 * For a previously invalid partition root, leave it at being
2319 * invalid if new_prs is not "member".
2320 */
2321 if (new_prs && is_prs_invalid(old_prs)) {
2322 cs->partition_root_state = -new_prs;
2323 return 0;
2324 }
2325
2326 if (alloc_cpumasks(NULL, &tmpmask))
2327 return -ENOMEM;
2328
2329 if (!old_prs) {
2330 /*
2331 * Turning on partition root requires setting the
2332 * CS_CPU_EXCLUSIVE bit implicitly as well and cpus_allowed
2333 * cannot be empty.
2334 */
2335 if (cpumask_empty(cs->cpus_allowed)) {
2336 err = PERR_CPUSEMPTY;
2337 goto out;
2338 }
2339
2340 err = update_flag(CS_CPU_EXCLUSIVE, cs, 1);
2341 if (err) {
2342 err = PERR_NOTEXCL;
2343 goto out;
2344 }
2345
2346 err = update_parent_subparts_cpumask(cs, partcmd_enable,
2347 NULL, &tmpmask);
2348 if (err) {
2349 update_flag(CS_CPU_EXCLUSIVE, cs, 0);
2350 goto out;
2351 }
2352
2353 if (new_prs == PRS_ISOLATED) {
2354 /*
2355 * Disable the load balance flag should not return an
2356 * error unless the system is running out of memory.
2357 */
2358 update_flag(CS_SCHED_LOAD_BALANCE, cs, 0);
2359 sched_domain_rebuilt = true;
2360 }
2361 } else if (old_prs && new_prs) {
2362 /*
2363 * A change in load balance state only, no change in cpumasks.
2364 */
2365 update_flag(CS_SCHED_LOAD_BALANCE, cs, (new_prs != PRS_ISOLATED));
2366 sched_domain_rebuilt = true;
2367 goto out; /* Sched domain is rebuilt in update_flag() */
2368 } else {
2369 /*
2370 * Switching back to member is always allowed even if it
2371 * disables child partitions.
2372 */
2373 update_parent_subparts_cpumask(cs, partcmd_disable, NULL,
2374 &tmpmask);
2375
2376 /*
2377 * If there are child partitions, they will all become invalid.
2378 */
2379 if (unlikely(cs->nr_subparts_cpus)) {
2380 spin_lock_irq(&callback_lock);
2381 cs->nr_subparts_cpus = 0;
2382 cpumask_clear(cs->subparts_cpus);
2383 compute_effective_cpumask(cs->effective_cpus, cs, parent);
2384 spin_unlock_irq(&callback_lock);
2385 }
2386
2387 /* Turning off CS_CPU_EXCLUSIVE will not return error */
2388 update_flag(CS_CPU_EXCLUSIVE, cs, 0);
2389
2390 if (!is_sched_load_balance(cs)) {
2391 /* Make sure load balance is on */
2392 update_flag(CS_SCHED_LOAD_BALANCE, cs, 1);
2393 sched_domain_rebuilt = true;
2394 }
2395 }
2396
2397 update_tasks_cpumask(parent, tmpmask.new_cpus);
2398
2399 if (parent->child_ecpus_count)
2400 update_sibling_cpumasks(parent, cs, &tmpmask);
2401
2402 if (!sched_domain_rebuilt)
2403 rebuild_sched_domains_locked();
2404 out:
2405 /*
2406 * Make partition invalid if an error happen
2407 */
2408 if (err)
2409 new_prs = -new_prs;
2410 spin_lock_irq(&callback_lock);
2411 cs->partition_root_state = new_prs;
2412 WRITE_ONCE(cs->prs_err, err);
2413 spin_unlock_irq(&callback_lock);
2414 /*
2415 * Update child cpusets, if present.
2416 * Force update if switching back to member.
2417 */
2418 if (!list_empty(&cs->css.children))
2419 update_cpumasks_hier(cs, &tmpmask, !new_prs);
2420
2421 notify_partition_change(cs, old_prs);
2422 free_cpumasks(NULL, &tmpmask);
2423 return 0;
2424 }
2425
2426 /*
2427 * Frequency meter - How fast is some event occurring?
2428 *
2429 * These routines manage a digitally filtered, constant time based,
2430 * event frequency meter. There are four routines:
2431 * fmeter_init() - initialize a frequency meter.
2432 * fmeter_markevent() - called each time the event happens.
2433 * fmeter_getrate() - returns the recent rate of such events.
2434 * fmeter_update() - internal routine used to update fmeter.
2435 *
2436 * A common data structure is passed to each of these routines,
2437 * which is used to keep track of the state required to manage the
2438 * frequency meter and its digital filter.
2439 *
2440 * The filter works on the number of events marked per unit time.
2441 * The filter is single-pole low-pass recursive (IIR). The time unit
2442 * is 1 second. Arithmetic is done using 32-bit integers scaled to
2443 * simulate 3 decimal digits of precision (multiplied by 1000).
2444 *
2445 * With an FM_COEF of 933, and a time base of 1 second, the filter
2446 * has a half-life of 10 seconds, meaning that if the events quit
2447 * happening, then the rate returned from the fmeter_getrate()
2448 * will be cut in half each 10 seconds, until it converges to zero.
2449 *
2450 * It is not worth doing a real infinitely recursive filter. If more
2451 * than FM_MAXTICKS ticks have elapsed since the last filter event,
2452 * just compute FM_MAXTICKS ticks worth, by which point the level
2453 * will be stable.
2454 *
2455 * Limit the count of unprocessed events to FM_MAXCNT, so as to avoid
2456 * arithmetic overflow in the fmeter_update() routine.
2457 *
2458 * Given the simple 32 bit integer arithmetic used, this meter works
2459 * best for reporting rates between one per millisecond (msec) and
2460 * one per 32 (approx) seconds. At constant rates faster than one
2461 * per msec it maxes out at values just under 1,000,000. At constant
2462 * rates between one per msec, and one per second it will stabilize
2463 * to a value N*1000, where N is the rate of events per second.
2464 * At constant rates between one per second and one per 32 seconds,
2465 * it will be choppy, moving up on the seconds that have an event,
2466 * and then decaying until the next event. At rates slower than
2467 * about one in 32 seconds, it decays all the way back to zero between
2468 * each event.
2469 */
2470
2471 #define FM_COEF 933 /* coefficient for half-life of 10 secs */
2472 #define FM_MAXTICKS ((u32)99) /* useless computing more ticks than this */
2473 #define FM_MAXCNT 1000000 /* limit cnt to avoid overflow */
2474 #define FM_SCALE 1000 /* faux fixed point scale */
2475
2476 /* Initialize a frequency meter */
fmeter_init(struct fmeter * fmp)2477 static void fmeter_init(struct fmeter *fmp)
2478 {
2479 fmp->cnt = 0;
2480 fmp->val = 0;
2481 fmp->time = 0;
2482 spin_lock_init(&fmp->lock);
2483 }
2484
2485 /* Internal meter update - process cnt events and update value */
fmeter_update(struct fmeter * fmp)2486 static void fmeter_update(struct fmeter *fmp)
2487 {
2488 time64_t now;
2489 u32 ticks;
2490
2491 now = ktime_get_seconds();
2492 ticks = now - fmp->time;
2493
2494 if (ticks == 0)
2495 return;
2496
2497 ticks = min(FM_MAXTICKS, ticks);
2498 while (ticks-- > 0)
2499 fmp->val = (FM_COEF * fmp->val) / FM_SCALE;
2500 fmp->time = now;
2501
2502 fmp->val += ((FM_SCALE - FM_COEF) * fmp->cnt) / FM_SCALE;
2503 fmp->cnt = 0;
2504 }
2505
2506 /* Process any previous ticks, then bump cnt by one (times scale). */
fmeter_markevent(struct fmeter * fmp)2507 static void fmeter_markevent(struct fmeter *fmp)
2508 {
2509 spin_lock(&fmp->lock);
2510 fmeter_update(fmp);
2511 fmp->cnt = min(FM_MAXCNT, fmp->cnt + FM_SCALE);
2512 spin_unlock(&fmp->lock);
2513 }
2514
2515 /* Process any previous ticks, then return current value. */
fmeter_getrate(struct fmeter * fmp)2516 static int fmeter_getrate(struct fmeter *fmp)
2517 {
2518 int val;
2519
2520 spin_lock(&fmp->lock);
2521 fmeter_update(fmp);
2522 val = fmp->val;
2523 spin_unlock(&fmp->lock);
2524 return val;
2525 }
2526
2527 static struct cpuset *cpuset_attach_old_cs;
2528
2529 /*
2530 * Check to see if a cpuset can accept a new task
2531 * For v1, cpus_allowed and mems_allowed can't be empty.
2532 * For v2, effective_cpus can't be empty.
2533 * Note that in v1, effective_cpus = cpus_allowed.
2534 */
cpuset_can_attach_check(struct cpuset * cs)2535 static int cpuset_can_attach_check(struct cpuset *cs)
2536 {
2537 if (cpumask_empty(cs->effective_cpus) ||
2538 (!is_in_v2_mode() && nodes_empty(cs->mems_allowed)))
2539 return -ENOSPC;
2540 return 0;
2541 }
2542
reset_migrate_dl_data(struct cpuset * cs)2543 static void reset_migrate_dl_data(struct cpuset *cs)
2544 {
2545 cs->nr_migrate_dl_tasks = 0;
2546 cs->sum_migrate_dl_bw = 0;
2547 }
2548
2549 /* Called by cgroups to determine if a cpuset is usable; cpuset_mutex held */
cpuset_can_attach(struct cgroup_taskset * tset)2550 static int cpuset_can_attach(struct cgroup_taskset *tset)
2551 {
2552 struct cgroup_subsys_state *css;
2553 struct cpuset *cs, *oldcs;
2554 struct task_struct *task;
2555 int ret;
2556
2557 /* used later by cpuset_attach() */
2558 cpuset_attach_old_cs = task_cs(cgroup_taskset_first(tset, &css));
2559 oldcs = cpuset_attach_old_cs;
2560 cs = css_cs(css);
2561
2562 mutex_lock(&cpuset_mutex);
2563
2564 /* Check to see if task is allowed in the cpuset */
2565 ret = cpuset_can_attach_check(cs);
2566 if (ret)
2567 goto out_unlock;
2568
2569 cgroup_taskset_for_each(task, css, tset) {
2570 ret = task_can_attach(task);
2571 if (ret)
2572 goto out_unlock;
2573 ret = security_task_setscheduler(task);
2574 if (ret)
2575 goto out_unlock;
2576
2577 if (dl_task(task)) {
2578 cs->nr_migrate_dl_tasks++;
2579 cs->sum_migrate_dl_bw += task->dl.dl_bw;
2580 }
2581 }
2582
2583 if (!cs->nr_migrate_dl_tasks)
2584 goto out_success;
2585
2586 if (!cpumask_intersects(oldcs->effective_cpus, cs->effective_cpus)) {
2587 int cpu = cpumask_any_and(cpu_active_mask, cs->effective_cpus);
2588
2589 if (unlikely(cpu >= nr_cpu_ids)) {
2590 reset_migrate_dl_data(cs);
2591 ret = -EINVAL;
2592 goto out_unlock;
2593 }
2594
2595 ret = dl_bw_alloc(cpu, cs->sum_migrate_dl_bw);
2596 if (ret) {
2597 reset_migrate_dl_data(cs);
2598 goto out_unlock;
2599 }
2600 }
2601
2602 out_success:
2603 /*
2604 * Mark attach is in progress. This makes validate_change() fail
2605 * changes which zero cpus/mems_allowed.
2606 */
2607 cs->attach_in_progress++;
2608 out_unlock:
2609 mutex_unlock(&cpuset_mutex);
2610 return ret;
2611 }
2612
cpuset_cancel_attach(struct cgroup_taskset * tset)2613 static void cpuset_cancel_attach(struct cgroup_taskset *tset)
2614 {
2615 struct cgroup_subsys_state *css;
2616 struct cpuset *cs;
2617
2618 cgroup_taskset_first(tset, &css);
2619 cs = css_cs(css);
2620
2621 mutex_lock(&cpuset_mutex);
2622 cs->attach_in_progress--;
2623 if (!cs->attach_in_progress)
2624 wake_up(&cpuset_attach_wq);
2625
2626 if (cs->nr_migrate_dl_tasks) {
2627 int cpu = cpumask_any(cs->effective_cpus);
2628
2629 dl_bw_free(cpu, cs->sum_migrate_dl_bw);
2630 reset_migrate_dl_data(cs);
2631 }
2632
2633 mutex_unlock(&cpuset_mutex);
2634 }
2635
2636 /*
2637 * Protected by cpuset_mutex. cpus_attach is used only by cpuset_attach_task()
2638 * but we can't allocate it dynamically there. Define it global and
2639 * allocate from cpuset_init().
2640 */
2641 static cpumask_var_t cpus_attach;
2642 static nodemask_t cpuset_attach_nodemask_to;
2643
cpuset_attach_task(struct cpuset * cs,struct task_struct * task)2644 static void cpuset_attach_task(struct cpuset *cs, struct task_struct *task)
2645 {
2646 lockdep_assert_held(&cpuset_mutex);
2647
2648 if (cs != &top_cpuset)
2649 guarantee_online_cpus(task, cpus_attach);
2650 else
2651 cpumask_copy(cpus_attach, task_cpu_possible_mask(task));
2652 /*
2653 * can_attach beforehand should guarantee that this doesn't
2654 * fail. TODO: have a better way to handle failure here
2655 */
2656 WARN_ON_ONCE(update_cpus_allowed(cs, task, cpus_attach));
2657
2658 cpuset_change_task_nodemask(task, &cpuset_attach_nodemask_to);
2659 cpuset_update_task_spread_flags(cs, task);
2660 }
2661
cpuset_attach(struct cgroup_taskset * tset)2662 static void cpuset_attach(struct cgroup_taskset *tset)
2663 {
2664 struct task_struct *task;
2665 struct task_struct *leader;
2666 struct cgroup_subsys_state *css;
2667 struct cpuset *cs;
2668 struct cpuset *oldcs = cpuset_attach_old_cs;
2669
2670 cgroup_taskset_first(tset, &css);
2671 cs = css_cs(css);
2672
2673 lockdep_assert_cpus_held(); /* see cgroup_attach_lock() */
2674 mutex_lock(&cpuset_mutex);
2675
2676 guarantee_online_mems(cs, &cpuset_attach_nodemask_to);
2677
2678 cgroup_taskset_for_each(task, css, tset)
2679 cpuset_attach_task(cs, task);
2680
2681 /*
2682 * Change mm for all threadgroup leaders. This is expensive and may
2683 * sleep and should be moved outside migration path proper.
2684 */
2685 cpuset_attach_nodemask_to = cs->effective_mems;
2686 cgroup_taskset_for_each_leader(leader, css, tset) {
2687 struct mm_struct *mm = get_task_mm(leader);
2688
2689 if (mm) {
2690 mpol_rebind_mm(mm, &cpuset_attach_nodemask_to);
2691
2692 /*
2693 * old_mems_allowed is the same with mems_allowed
2694 * here, except if this task is being moved
2695 * automatically due to hotplug. In that case
2696 * @mems_allowed has been updated and is empty, so
2697 * @old_mems_allowed is the right nodesets that we
2698 * migrate mm from.
2699 */
2700 if (is_memory_migrate(cs))
2701 cpuset_migrate_mm(mm, &oldcs->old_mems_allowed,
2702 &cpuset_attach_nodemask_to);
2703 else
2704 mmput(mm);
2705 }
2706 }
2707
2708 cs->old_mems_allowed = cpuset_attach_nodemask_to;
2709
2710 if (cs->nr_migrate_dl_tasks) {
2711 cs->nr_deadline_tasks += cs->nr_migrate_dl_tasks;
2712 oldcs->nr_deadline_tasks -= cs->nr_migrate_dl_tasks;
2713 reset_migrate_dl_data(cs);
2714 }
2715
2716 cs->attach_in_progress--;
2717 if (!cs->attach_in_progress)
2718 wake_up(&cpuset_attach_wq);
2719
2720 mutex_unlock(&cpuset_mutex);
2721 }
2722
2723 /* The various types of files and directories in a cpuset file system */
2724
2725 typedef enum {
2726 FILE_MEMORY_MIGRATE,
2727 FILE_CPULIST,
2728 FILE_MEMLIST,
2729 FILE_EFFECTIVE_CPULIST,
2730 FILE_EFFECTIVE_MEMLIST,
2731 FILE_SUBPARTS_CPULIST,
2732 FILE_CPU_EXCLUSIVE,
2733 FILE_MEM_EXCLUSIVE,
2734 FILE_MEM_HARDWALL,
2735 FILE_SCHED_LOAD_BALANCE,
2736 FILE_PARTITION_ROOT,
2737 FILE_SCHED_RELAX_DOMAIN_LEVEL,
2738 FILE_MEMORY_PRESSURE_ENABLED,
2739 FILE_MEMORY_PRESSURE,
2740 FILE_SPREAD_PAGE,
2741 FILE_SPREAD_SLAB,
2742 } cpuset_filetype_t;
2743
cpuset_write_u64(struct cgroup_subsys_state * css,struct cftype * cft,u64 val)2744 static int cpuset_write_u64(struct cgroup_subsys_state *css, struct cftype *cft,
2745 u64 val)
2746 {
2747 struct cpuset *cs = css_cs(css);
2748 cpuset_filetype_t type = cft->private;
2749 int retval = 0;
2750
2751 cpus_read_lock();
2752 mutex_lock(&cpuset_mutex);
2753 if (!is_cpuset_online(cs)) {
2754 retval = -ENODEV;
2755 goto out_unlock;
2756 }
2757
2758 switch (type) {
2759 case FILE_CPU_EXCLUSIVE:
2760 retval = update_flag(CS_CPU_EXCLUSIVE, cs, val);
2761 break;
2762 case FILE_MEM_EXCLUSIVE:
2763 retval = update_flag(CS_MEM_EXCLUSIVE, cs, val);
2764 break;
2765 case FILE_MEM_HARDWALL:
2766 retval = update_flag(CS_MEM_HARDWALL, cs, val);
2767 break;
2768 case FILE_SCHED_LOAD_BALANCE:
2769 retval = update_flag(CS_SCHED_LOAD_BALANCE, cs, val);
2770 break;
2771 case FILE_MEMORY_MIGRATE:
2772 retval = update_flag(CS_MEMORY_MIGRATE, cs, val);
2773 break;
2774 case FILE_MEMORY_PRESSURE_ENABLED:
2775 cpuset_memory_pressure_enabled = !!val;
2776 break;
2777 case FILE_SPREAD_PAGE:
2778 retval = update_flag(CS_SPREAD_PAGE, cs, val);
2779 break;
2780 case FILE_SPREAD_SLAB:
2781 retval = update_flag(CS_SPREAD_SLAB, cs, val);
2782 break;
2783 default:
2784 retval = -EINVAL;
2785 break;
2786 }
2787 out_unlock:
2788 mutex_unlock(&cpuset_mutex);
2789 cpus_read_unlock();
2790 return retval;
2791 }
2792
cpuset_write_s64(struct cgroup_subsys_state * css,struct cftype * cft,s64 val)2793 static int cpuset_write_s64(struct cgroup_subsys_state *css, struct cftype *cft,
2794 s64 val)
2795 {
2796 struct cpuset *cs = css_cs(css);
2797 cpuset_filetype_t type = cft->private;
2798 int retval = -ENODEV;
2799
2800 cpus_read_lock();
2801 mutex_lock(&cpuset_mutex);
2802 if (!is_cpuset_online(cs))
2803 goto out_unlock;
2804
2805 switch (type) {
2806 case FILE_SCHED_RELAX_DOMAIN_LEVEL:
2807 retval = update_relax_domain_level(cs, val);
2808 break;
2809 default:
2810 retval = -EINVAL;
2811 break;
2812 }
2813 out_unlock:
2814 mutex_unlock(&cpuset_mutex);
2815 cpus_read_unlock();
2816 return retval;
2817 }
2818
2819 /*
2820 * Common handling for a write to a "cpus" or "mems" file.
2821 */
cpuset_write_resmask(struct kernfs_open_file * of,char * buf,size_t nbytes,loff_t off)2822 static ssize_t cpuset_write_resmask(struct kernfs_open_file *of,
2823 char *buf, size_t nbytes, loff_t off)
2824 {
2825 struct cpuset *cs = css_cs(of_css(of));
2826 struct cpuset *trialcs;
2827 int retval = -ENODEV;
2828
2829 buf = strstrip(buf);
2830
2831 /*
2832 * CPU or memory hotunplug may leave @cs w/o any execution
2833 * resources, in which case the hotplug code asynchronously updates
2834 * configuration and transfers all tasks to the nearest ancestor
2835 * which can execute.
2836 *
2837 * As writes to "cpus" or "mems" may restore @cs's execution
2838 * resources, wait for the previously scheduled operations before
2839 * proceeding, so that we don't end up keep removing tasks added
2840 * after execution capability is restored.
2841 *
2842 * cpuset_hotplug_work calls back into cgroup core via
2843 * cgroup_transfer_tasks() and waiting for it from a cgroupfs
2844 * operation like this one can lead to a deadlock through kernfs
2845 * active_ref protection. Let's break the protection. Losing the
2846 * protection is okay as we check whether @cs is online after
2847 * grabbing cpuset_mutex anyway. This only happens on the legacy
2848 * hierarchies.
2849 */
2850 css_get(&cs->css);
2851 kernfs_break_active_protection(of->kn);
2852 flush_work(&cpuset_hotplug_work);
2853
2854 cpus_read_lock();
2855 mutex_lock(&cpuset_mutex);
2856 if (!is_cpuset_online(cs))
2857 goto out_unlock;
2858
2859 trialcs = alloc_trial_cpuset(cs);
2860 if (!trialcs) {
2861 retval = -ENOMEM;
2862 goto out_unlock;
2863 }
2864
2865 switch (of_cft(of)->private) {
2866 case FILE_CPULIST:
2867 retval = update_cpumask(cs, trialcs, buf);
2868 break;
2869 case FILE_MEMLIST:
2870 retval = update_nodemask(cs, trialcs, buf);
2871 break;
2872 default:
2873 retval = -EINVAL;
2874 break;
2875 }
2876
2877 free_cpuset(trialcs);
2878 out_unlock:
2879 mutex_unlock(&cpuset_mutex);
2880 cpus_read_unlock();
2881 kernfs_unbreak_active_protection(of->kn);
2882 css_put(&cs->css);
2883 flush_workqueue(cpuset_migrate_mm_wq);
2884 return retval ?: nbytes;
2885 }
2886
2887 /*
2888 * These ascii lists should be read in a single call, by using a user
2889 * buffer large enough to hold the entire map. If read in smaller
2890 * chunks, there is no guarantee of atomicity. Since the display format
2891 * used, list of ranges of sequential numbers, is variable length,
2892 * and since these maps can change value dynamically, one could read
2893 * gibberish by doing partial reads while a list was changing.
2894 */
cpuset_common_seq_show(struct seq_file * sf,void * v)2895 static int cpuset_common_seq_show(struct seq_file *sf, void *v)
2896 {
2897 struct cpuset *cs = css_cs(seq_css(sf));
2898 cpuset_filetype_t type = seq_cft(sf)->private;
2899 int ret = 0;
2900
2901 spin_lock_irq(&callback_lock);
2902
2903 switch (type) {
2904 case FILE_CPULIST:
2905 seq_printf(sf, "%*pbl\n", cpumask_pr_args(cs->cpus_requested));
2906 break;
2907 case FILE_MEMLIST:
2908 seq_printf(sf, "%*pbl\n", nodemask_pr_args(&cs->mems_allowed));
2909 break;
2910 case FILE_EFFECTIVE_CPULIST:
2911 seq_printf(sf, "%*pbl\n", cpumask_pr_args(cs->effective_cpus));
2912 break;
2913 case FILE_EFFECTIVE_MEMLIST:
2914 seq_printf(sf, "%*pbl\n", nodemask_pr_args(&cs->effective_mems));
2915 break;
2916 case FILE_SUBPARTS_CPULIST:
2917 seq_printf(sf, "%*pbl\n", cpumask_pr_args(cs->subparts_cpus));
2918 break;
2919 default:
2920 ret = -EINVAL;
2921 }
2922
2923 spin_unlock_irq(&callback_lock);
2924 return ret;
2925 }
2926
cpuset_read_u64(struct cgroup_subsys_state * css,struct cftype * cft)2927 static u64 cpuset_read_u64(struct cgroup_subsys_state *css, struct cftype *cft)
2928 {
2929 struct cpuset *cs = css_cs(css);
2930 cpuset_filetype_t type = cft->private;
2931 switch (type) {
2932 case FILE_CPU_EXCLUSIVE:
2933 return is_cpu_exclusive(cs);
2934 case FILE_MEM_EXCLUSIVE:
2935 return is_mem_exclusive(cs);
2936 case FILE_MEM_HARDWALL:
2937 return is_mem_hardwall(cs);
2938 case FILE_SCHED_LOAD_BALANCE:
2939 return is_sched_load_balance(cs);
2940 case FILE_MEMORY_MIGRATE:
2941 return is_memory_migrate(cs);
2942 case FILE_MEMORY_PRESSURE_ENABLED:
2943 return cpuset_memory_pressure_enabled;
2944 case FILE_MEMORY_PRESSURE:
2945 return fmeter_getrate(&cs->fmeter);
2946 case FILE_SPREAD_PAGE:
2947 return is_spread_page(cs);
2948 case FILE_SPREAD_SLAB:
2949 return is_spread_slab(cs);
2950 default:
2951 BUG();
2952 }
2953
2954 /* Unreachable but makes gcc happy */
2955 return 0;
2956 }
2957
cpuset_read_s64(struct cgroup_subsys_state * css,struct cftype * cft)2958 static s64 cpuset_read_s64(struct cgroup_subsys_state *css, struct cftype *cft)
2959 {
2960 struct cpuset *cs = css_cs(css);
2961 cpuset_filetype_t type = cft->private;
2962 switch (type) {
2963 case FILE_SCHED_RELAX_DOMAIN_LEVEL:
2964 return cs->relax_domain_level;
2965 default:
2966 BUG();
2967 }
2968
2969 /* Unreachable but makes gcc happy */
2970 return 0;
2971 }
2972
sched_partition_show(struct seq_file * seq,void * v)2973 static int sched_partition_show(struct seq_file *seq, void *v)
2974 {
2975 struct cpuset *cs = css_cs(seq_css(seq));
2976 const char *err, *type = NULL;
2977
2978 switch (cs->partition_root_state) {
2979 case PRS_ROOT:
2980 seq_puts(seq, "root\n");
2981 break;
2982 case PRS_ISOLATED:
2983 seq_puts(seq, "isolated\n");
2984 break;
2985 case PRS_MEMBER:
2986 seq_puts(seq, "member\n");
2987 break;
2988 case PRS_INVALID_ROOT:
2989 type = "root";
2990 fallthrough;
2991 case PRS_INVALID_ISOLATED:
2992 if (!type)
2993 type = "isolated";
2994 err = perr_strings[READ_ONCE(cs->prs_err)];
2995 if (err)
2996 seq_printf(seq, "%s invalid (%s)\n", type, err);
2997 else
2998 seq_printf(seq, "%s invalid\n", type);
2999 break;
3000 }
3001 return 0;
3002 }
3003
sched_partition_write(struct kernfs_open_file * of,char * buf,size_t nbytes,loff_t off)3004 static ssize_t sched_partition_write(struct kernfs_open_file *of, char *buf,
3005 size_t nbytes, loff_t off)
3006 {
3007 struct cpuset *cs = css_cs(of_css(of));
3008 int val;
3009 int retval = -ENODEV;
3010
3011 buf = strstrip(buf);
3012
3013 /*
3014 * Convert "root" to ENABLED, and convert "member" to DISABLED.
3015 */
3016 if (!strcmp(buf, "root"))
3017 val = PRS_ROOT;
3018 else if (!strcmp(buf, "member"))
3019 val = PRS_MEMBER;
3020 else if (!strcmp(buf, "isolated"))
3021 val = PRS_ISOLATED;
3022 else
3023 return -EINVAL;
3024
3025 css_get(&cs->css);
3026 cpus_read_lock();
3027 mutex_lock(&cpuset_mutex);
3028 if (!is_cpuset_online(cs))
3029 goto out_unlock;
3030
3031 retval = update_prstate(cs, val);
3032 out_unlock:
3033 mutex_unlock(&cpuset_mutex);
3034 cpus_read_unlock();
3035 css_put(&cs->css);
3036 return retval ?: nbytes;
3037 }
3038
3039 /*
3040 * for the common functions, 'private' gives the type of file
3041 */
3042
3043 static struct cftype legacy_files[] = {
3044 {
3045 .name = "cpus",
3046 .seq_show = cpuset_common_seq_show,
3047 .write = cpuset_write_resmask,
3048 .max_write_len = (100U + 6 * NR_CPUS),
3049 .private = FILE_CPULIST,
3050 },
3051
3052 {
3053 .name = "mems",
3054 .seq_show = cpuset_common_seq_show,
3055 .write = cpuset_write_resmask,
3056 .max_write_len = (100U + 6 * MAX_NUMNODES),
3057 .private = FILE_MEMLIST,
3058 },
3059
3060 {
3061 .name = "effective_cpus",
3062 .seq_show = cpuset_common_seq_show,
3063 .private = FILE_EFFECTIVE_CPULIST,
3064 },
3065
3066 {
3067 .name = "effective_mems",
3068 .seq_show = cpuset_common_seq_show,
3069 .private = FILE_EFFECTIVE_MEMLIST,
3070 },
3071
3072 {
3073 .name = "cpu_exclusive",
3074 .read_u64 = cpuset_read_u64,
3075 .write_u64 = cpuset_write_u64,
3076 .private = FILE_CPU_EXCLUSIVE,
3077 },
3078
3079 {
3080 .name = "mem_exclusive",
3081 .read_u64 = cpuset_read_u64,
3082 .write_u64 = cpuset_write_u64,
3083 .private = FILE_MEM_EXCLUSIVE,
3084 },
3085
3086 {
3087 .name = "mem_hardwall",
3088 .read_u64 = cpuset_read_u64,
3089 .write_u64 = cpuset_write_u64,
3090 .private = FILE_MEM_HARDWALL,
3091 },
3092
3093 {
3094 .name = "sched_load_balance",
3095 .read_u64 = cpuset_read_u64,
3096 .write_u64 = cpuset_write_u64,
3097 .private = FILE_SCHED_LOAD_BALANCE,
3098 },
3099
3100 {
3101 .name = "sched_relax_domain_level",
3102 .read_s64 = cpuset_read_s64,
3103 .write_s64 = cpuset_write_s64,
3104 .private = FILE_SCHED_RELAX_DOMAIN_LEVEL,
3105 },
3106
3107 {
3108 .name = "memory_migrate",
3109 .read_u64 = cpuset_read_u64,
3110 .write_u64 = cpuset_write_u64,
3111 .private = FILE_MEMORY_MIGRATE,
3112 },
3113
3114 {
3115 .name = "memory_pressure",
3116 .read_u64 = cpuset_read_u64,
3117 .private = FILE_MEMORY_PRESSURE,
3118 },
3119
3120 {
3121 .name = "memory_spread_page",
3122 .read_u64 = cpuset_read_u64,
3123 .write_u64 = cpuset_write_u64,
3124 .private = FILE_SPREAD_PAGE,
3125 },
3126
3127 {
3128 .name = "memory_spread_slab",
3129 .read_u64 = cpuset_read_u64,
3130 .write_u64 = cpuset_write_u64,
3131 .private = FILE_SPREAD_SLAB,
3132 },
3133
3134 {
3135 .name = "memory_pressure_enabled",
3136 .flags = CFTYPE_ONLY_ON_ROOT,
3137 .read_u64 = cpuset_read_u64,
3138 .write_u64 = cpuset_write_u64,
3139 .private = FILE_MEMORY_PRESSURE_ENABLED,
3140 },
3141
3142 { } /* terminate */
3143 };
3144
3145 /*
3146 * This is currently a minimal set for the default hierarchy. It can be
3147 * expanded later on by migrating more features and control files from v1.
3148 */
3149 static struct cftype dfl_files[] = {
3150 {
3151 .name = "cpus",
3152 .seq_show = cpuset_common_seq_show,
3153 .write = cpuset_write_resmask,
3154 .max_write_len = (100U + 6 * NR_CPUS),
3155 .private = FILE_CPULIST,
3156 .flags = CFTYPE_NOT_ON_ROOT,
3157 },
3158
3159 {
3160 .name = "mems",
3161 .seq_show = cpuset_common_seq_show,
3162 .write = cpuset_write_resmask,
3163 .max_write_len = (100U + 6 * MAX_NUMNODES),
3164 .private = FILE_MEMLIST,
3165 .flags = CFTYPE_NOT_ON_ROOT,
3166 },
3167
3168 {
3169 .name = "cpus.effective",
3170 .seq_show = cpuset_common_seq_show,
3171 .private = FILE_EFFECTIVE_CPULIST,
3172 },
3173
3174 {
3175 .name = "mems.effective",
3176 .seq_show = cpuset_common_seq_show,
3177 .private = FILE_EFFECTIVE_MEMLIST,
3178 },
3179
3180 {
3181 .name = "cpus.partition",
3182 .seq_show = sched_partition_show,
3183 .write = sched_partition_write,
3184 .private = FILE_PARTITION_ROOT,
3185 .flags = CFTYPE_NOT_ON_ROOT,
3186 .file_offset = offsetof(struct cpuset, partition_file),
3187 },
3188
3189 {
3190 .name = "cpus.subpartitions",
3191 .seq_show = cpuset_common_seq_show,
3192 .private = FILE_SUBPARTS_CPULIST,
3193 .flags = CFTYPE_DEBUG,
3194 },
3195
3196 { } /* terminate */
3197 };
3198
3199
3200 /*
3201 * cpuset_css_alloc - allocate a cpuset css
3202 * cgrp: control group that the new cpuset will be part of
3203 */
3204
3205 static struct cgroup_subsys_state *
cpuset_css_alloc(struct cgroup_subsys_state * parent_css)3206 cpuset_css_alloc(struct cgroup_subsys_state *parent_css)
3207 {
3208 struct cpuset *cs;
3209
3210 if (!parent_css)
3211 return &top_cpuset.css;
3212
3213 cs = kzalloc(sizeof(*cs), GFP_KERNEL);
3214 if (!cs)
3215 return ERR_PTR(-ENOMEM);
3216
3217 if (alloc_cpumasks(cs, NULL)) {
3218 kfree(cs);
3219 return ERR_PTR(-ENOMEM);
3220 }
3221
3222 __set_bit(CS_SCHED_LOAD_BALANCE, &cs->flags);
3223 nodes_clear(cs->mems_allowed);
3224 nodes_clear(cs->effective_mems);
3225 fmeter_init(&cs->fmeter);
3226 cs->relax_domain_level = -1;
3227
3228 /* Set CS_MEMORY_MIGRATE for default hierarchy */
3229 if (cgroup_subsys_on_dfl(cpuset_cgrp_subsys))
3230 __set_bit(CS_MEMORY_MIGRATE, &cs->flags);
3231
3232 return &cs->css;
3233 }
3234
cpuset_css_online(struct cgroup_subsys_state * css)3235 static int cpuset_css_online(struct cgroup_subsys_state *css)
3236 {
3237 struct cpuset *cs = css_cs(css);
3238 struct cpuset *parent = parent_cs(cs);
3239 struct cpuset *tmp_cs;
3240 struct cgroup_subsys_state *pos_css;
3241
3242 if (!parent)
3243 return 0;
3244
3245 cpus_read_lock();
3246 mutex_lock(&cpuset_mutex);
3247
3248 set_bit(CS_ONLINE, &cs->flags);
3249 if (is_spread_page(parent))
3250 set_bit(CS_SPREAD_PAGE, &cs->flags);
3251 if (is_spread_slab(parent))
3252 set_bit(CS_SPREAD_SLAB, &cs->flags);
3253
3254 cpuset_inc();
3255
3256 spin_lock_irq(&callback_lock);
3257 if (is_in_v2_mode()) {
3258 cpumask_copy(cs->effective_cpus, parent->effective_cpus);
3259 cs->effective_mems = parent->effective_mems;
3260 cs->use_parent_ecpus = true;
3261 parent->child_ecpus_count++;
3262 }
3263
3264 /*
3265 * For v2, clear CS_SCHED_LOAD_BALANCE if parent is isolated
3266 */
3267 if (cgroup_subsys_on_dfl(cpuset_cgrp_subsys) &&
3268 !is_sched_load_balance(parent))
3269 clear_bit(CS_SCHED_LOAD_BALANCE, &cs->flags);
3270
3271 spin_unlock_irq(&callback_lock);
3272
3273 if (!test_bit(CGRP_CPUSET_CLONE_CHILDREN, &css->cgroup->flags))
3274 goto out_unlock;
3275
3276 /*
3277 * Clone @parent's configuration if CGRP_CPUSET_CLONE_CHILDREN is
3278 * set. This flag handling is implemented in cgroup core for
3279 * historical reasons - the flag may be specified during mount.
3280 *
3281 * Currently, if any sibling cpusets have exclusive cpus or mem, we
3282 * refuse to clone the configuration - thereby refusing the task to
3283 * be entered, and as a result refusing the sys_unshare() or
3284 * clone() which initiated it. If this becomes a problem for some
3285 * users who wish to allow that scenario, then this could be
3286 * changed to grant parent->cpus_allowed-sibling_cpus_exclusive
3287 * (and likewise for mems) to the new cgroup.
3288 */
3289 rcu_read_lock();
3290 cpuset_for_each_child(tmp_cs, pos_css, parent) {
3291 if (is_mem_exclusive(tmp_cs) || is_cpu_exclusive(tmp_cs)) {
3292 rcu_read_unlock();
3293 goto out_unlock;
3294 }
3295 }
3296 rcu_read_unlock();
3297
3298 spin_lock_irq(&callback_lock);
3299 cs->mems_allowed = parent->mems_allowed;
3300 cs->effective_mems = parent->mems_allowed;
3301 cpumask_copy(cs->cpus_allowed, parent->cpus_allowed);
3302 cpumask_copy(cs->cpus_requested, parent->cpus_requested);
3303 cpumask_copy(cs->effective_cpus, parent->cpus_allowed);
3304 spin_unlock_irq(&callback_lock);
3305 out_unlock:
3306 mutex_unlock(&cpuset_mutex);
3307 cpus_read_unlock();
3308 return 0;
3309 }
3310
3311 /*
3312 * If the cpuset being removed has its flag 'sched_load_balance'
3313 * enabled, then simulate turning sched_load_balance off, which
3314 * will call rebuild_sched_domains_locked(). That is not needed
3315 * in the default hierarchy where only changes in partition
3316 * will cause repartitioning.
3317 *
3318 * If the cpuset has the 'sched.partition' flag enabled, simulate
3319 * turning 'sched.partition" off.
3320 */
3321
cpuset_css_offline(struct cgroup_subsys_state * css)3322 static void cpuset_css_offline(struct cgroup_subsys_state *css)
3323 {
3324 struct cpuset *cs = css_cs(css);
3325
3326 cpus_read_lock();
3327 mutex_lock(&cpuset_mutex);
3328
3329 if (is_partition_valid(cs))
3330 update_prstate(cs, 0);
3331
3332 if (!cgroup_subsys_on_dfl(cpuset_cgrp_subsys) &&
3333 is_sched_load_balance(cs))
3334 update_flag(CS_SCHED_LOAD_BALANCE, cs, 0);
3335
3336 if (cs->use_parent_ecpus) {
3337 struct cpuset *parent = parent_cs(cs);
3338
3339 cs->use_parent_ecpus = false;
3340 parent->child_ecpus_count--;
3341 }
3342
3343 cpuset_dec();
3344 clear_bit(CS_ONLINE, &cs->flags);
3345
3346 mutex_unlock(&cpuset_mutex);
3347 cpus_read_unlock();
3348 }
3349
cpuset_css_free(struct cgroup_subsys_state * css)3350 static void cpuset_css_free(struct cgroup_subsys_state *css)
3351 {
3352 struct cpuset *cs = css_cs(css);
3353
3354 free_cpuset(cs);
3355 }
3356
cpuset_bind(struct cgroup_subsys_state * root_css)3357 static void cpuset_bind(struct cgroup_subsys_state *root_css)
3358 {
3359 mutex_lock(&cpuset_mutex);
3360 spin_lock_irq(&callback_lock);
3361
3362 if (is_in_v2_mode()) {
3363 cpumask_copy(top_cpuset.cpus_allowed, cpu_possible_mask);
3364 top_cpuset.mems_allowed = node_possible_map;
3365 } else {
3366 cpumask_copy(top_cpuset.cpus_allowed,
3367 top_cpuset.effective_cpus);
3368 top_cpuset.mems_allowed = top_cpuset.effective_mems;
3369 }
3370
3371 spin_unlock_irq(&callback_lock);
3372 mutex_unlock(&cpuset_mutex);
3373 }
3374
3375 /*
3376 * In case the child is cloned into a cpuset different from its parent,
3377 * additional checks are done to see if the move is allowed.
3378 */
cpuset_can_fork(struct task_struct * task,struct css_set * cset)3379 static int cpuset_can_fork(struct task_struct *task, struct css_set *cset)
3380 {
3381 struct cpuset *cs = css_cs(cset->subsys[cpuset_cgrp_id]);
3382 bool same_cs;
3383 int ret;
3384
3385 rcu_read_lock();
3386 same_cs = (cs == task_cs(current));
3387 rcu_read_unlock();
3388
3389 if (same_cs)
3390 return 0;
3391
3392 lockdep_assert_held(&cgroup_mutex);
3393 mutex_lock(&cpuset_mutex);
3394
3395 /* Check to see if task is allowed in the cpuset */
3396 ret = cpuset_can_attach_check(cs);
3397 if (ret)
3398 goto out_unlock;
3399
3400 ret = task_can_attach(task);
3401 if (ret)
3402 goto out_unlock;
3403
3404 ret = security_task_setscheduler(task);
3405 if (ret)
3406 goto out_unlock;
3407
3408 /*
3409 * Mark attach is in progress. This makes validate_change() fail
3410 * changes which zero cpus/mems_allowed.
3411 */
3412 cs->attach_in_progress++;
3413 out_unlock:
3414 mutex_unlock(&cpuset_mutex);
3415 return ret;
3416 }
3417
cpuset_cancel_fork(struct task_struct * task,struct css_set * cset)3418 static void cpuset_cancel_fork(struct task_struct *task, struct css_set *cset)
3419 {
3420 struct cpuset *cs = css_cs(cset->subsys[cpuset_cgrp_id]);
3421 bool same_cs;
3422
3423 rcu_read_lock();
3424 same_cs = (cs == task_cs(current));
3425 rcu_read_unlock();
3426
3427 if (same_cs)
3428 return;
3429
3430 mutex_lock(&cpuset_mutex);
3431 cs->attach_in_progress--;
3432 if (!cs->attach_in_progress)
3433 wake_up(&cpuset_attach_wq);
3434 mutex_unlock(&cpuset_mutex);
3435 }
3436
3437 /*
3438 * Make sure the new task conform to the current state of its parent,
3439 * which could have been changed by cpuset just after it inherits the
3440 * state from the parent and before it sits on the cgroup's task list.
3441 */
cpuset_fork(struct task_struct * task)3442 static void cpuset_fork(struct task_struct *task)
3443 {
3444 struct cpuset *cs;
3445 bool same_cs, inherit_cpus = false;
3446
3447 rcu_read_lock();
3448 cs = task_cs(task);
3449 same_cs = (cs == task_cs(current));
3450 rcu_read_unlock();
3451 if (same_cs) {
3452 if (cs == &top_cpuset)
3453 return;
3454 trace_android_rvh_cpuset_fork(task, &inherit_cpus);
3455 if (!inherit_cpus)
3456 set_cpus_allowed_ptr(task, current->cpus_ptr);
3457 task->mems_allowed = current->mems_allowed;
3458 return;
3459 }
3460
3461 /* CLONE_INTO_CGROUP */
3462 mutex_lock(&cpuset_mutex);
3463 guarantee_online_mems(cs, &cpuset_attach_nodemask_to);
3464 cpuset_attach_task(cs, task);
3465
3466 cs->attach_in_progress--;
3467 if (!cs->attach_in_progress)
3468 wake_up(&cpuset_attach_wq);
3469
3470 mutex_unlock(&cpuset_mutex);
3471 }
3472
3473 struct cgroup_subsys cpuset_cgrp_subsys = {
3474 .css_alloc = cpuset_css_alloc,
3475 .css_online = cpuset_css_online,
3476 .css_offline = cpuset_css_offline,
3477 .css_free = cpuset_css_free,
3478 .can_attach = cpuset_can_attach,
3479 .cancel_attach = cpuset_cancel_attach,
3480 .attach = cpuset_attach,
3481 .post_attach = cpuset_post_attach,
3482 .bind = cpuset_bind,
3483 .can_fork = cpuset_can_fork,
3484 .cancel_fork = cpuset_cancel_fork,
3485 .fork = cpuset_fork,
3486 .legacy_cftypes = legacy_files,
3487 .dfl_cftypes = dfl_files,
3488 .early_init = true,
3489 .threaded = true,
3490 };
3491
3492 /**
3493 * cpuset_init - initialize cpusets at system boot
3494 *
3495 * Description: Initialize top_cpuset
3496 **/
3497
cpuset_init(void)3498 int __init cpuset_init(void)
3499 {
3500 BUG_ON(!alloc_cpumask_var(&top_cpuset.cpus_allowed, GFP_KERNEL));
3501 BUG_ON(!alloc_cpumask_var(&top_cpuset.effective_cpus, GFP_KERNEL));
3502 BUG_ON(!zalloc_cpumask_var(&top_cpuset.subparts_cpus, GFP_KERNEL));
3503 BUG_ON(!alloc_cpumask_var(&top_cpuset.cpus_requested, GFP_KERNEL));
3504
3505 cpumask_setall(top_cpuset.cpus_allowed);
3506 cpumask_setall(top_cpuset.cpus_requested);
3507 nodes_setall(top_cpuset.mems_allowed);
3508 cpumask_setall(top_cpuset.effective_cpus);
3509 nodes_setall(top_cpuset.effective_mems);
3510
3511 fmeter_init(&top_cpuset.fmeter);
3512 set_bit(CS_SCHED_LOAD_BALANCE, &top_cpuset.flags);
3513 top_cpuset.relax_domain_level = -1;
3514
3515 BUG_ON(!alloc_cpumask_var(&cpus_attach, GFP_KERNEL));
3516
3517 return 0;
3518 }
3519
3520 /*
3521 * If CPU and/or memory hotplug handlers, below, unplug any CPUs
3522 * or memory nodes, we need to walk over the cpuset hierarchy,
3523 * removing that CPU or node from all cpusets. If this removes the
3524 * last CPU or node from a cpuset, then move the tasks in the empty
3525 * cpuset to its next-highest non-empty parent.
3526 */
remove_tasks_in_empty_cpuset(struct cpuset * cs)3527 static void remove_tasks_in_empty_cpuset(struct cpuset *cs)
3528 {
3529 struct cpuset *parent;
3530
3531 /*
3532 * Find its next-highest non-empty parent, (top cpuset
3533 * has online cpus, so can't be empty).
3534 */
3535 parent = parent_cs(cs);
3536 while (cpumask_empty(parent->cpus_allowed) ||
3537 nodes_empty(parent->mems_allowed))
3538 parent = parent_cs(parent);
3539
3540 if (cgroup_transfer_tasks(parent->css.cgroup, cs->css.cgroup)) {
3541 pr_err("cpuset: failed to transfer tasks out of empty cpuset ");
3542 pr_cont_cgroup_name(cs->css.cgroup);
3543 pr_cont("\n");
3544 }
3545 }
3546
3547 static void
hotplug_update_tasks_legacy(struct cpuset * cs,struct cpumask * new_cpus,nodemask_t * new_mems,bool cpus_updated,bool mems_updated)3548 hotplug_update_tasks_legacy(struct cpuset *cs,
3549 struct cpumask *new_cpus, nodemask_t *new_mems,
3550 bool cpus_updated, bool mems_updated)
3551 {
3552 bool is_empty;
3553
3554 spin_lock_irq(&callback_lock);
3555 cpumask_copy(cs->cpus_allowed, new_cpus);
3556 cpumask_copy(cs->effective_cpus, new_cpus);
3557 cs->mems_allowed = *new_mems;
3558 cs->effective_mems = *new_mems;
3559 spin_unlock_irq(&callback_lock);
3560
3561 /*
3562 * Don't call update_tasks_cpumask() if the cpuset becomes empty,
3563 * as the tasks will be migrated to an ancestor.
3564 */
3565 if (cpus_updated && !cpumask_empty(cs->cpus_allowed))
3566 update_tasks_cpumask(cs, new_cpus);
3567 if (mems_updated && !nodes_empty(cs->mems_allowed))
3568 update_tasks_nodemask(cs);
3569
3570 is_empty = cpumask_empty(cs->cpus_allowed) ||
3571 nodes_empty(cs->mems_allowed);
3572
3573 mutex_unlock(&cpuset_mutex);
3574
3575 /*
3576 * Move tasks to the nearest ancestor with execution resources,
3577 * This is full cgroup operation which will also call back into
3578 * cpuset. Should be done outside any lock.
3579 */
3580 if (is_empty)
3581 remove_tasks_in_empty_cpuset(cs);
3582
3583 mutex_lock(&cpuset_mutex);
3584 }
3585
3586 static void
hotplug_update_tasks(struct cpuset * cs,struct cpumask * new_cpus,nodemask_t * new_mems,bool cpus_updated,bool mems_updated)3587 hotplug_update_tasks(struct cpuset *cs,
3588 struct cpumask *new_cpus, nodemask_t *new_mems,
3589 bool cpus_updated, bool mems_updated)
3590 {
3591 /* A partition root is allowed to have empty effective cpus */
3592 if (cpumask_empty(new_cpus) && !is_partition_valid(cs))
3593 cpumask_copy(new_cpus, parent_cs(cs)->effective_cpus);
3594 if (nodes_empty(*new_mems))
3595 *new_mems = parent_cs(cs)->effective_mems;
3596
3597 spin_lock_irq(&callback_lock);
3598 cpumask_copy(cs->effective_cpus, new_cpus);
3599 cs->effective_mems = *new_mems;
3600 spin_unlock_irq(&callback_lock);
3601
3602 if (cpus_updated)
3603 update_tasks_cpumask(cs, new_cpus);
3604 if (mems_updated)
3605 update_tasks_nodemask(cs);
3606 }
3607
3608 static bool force_rebuild;
3609
cpuset_force_rebuild(void)3610 void cpuset_force_rebuild(void)
3611 {
3612 force_rebuild = true;
3613 }
3614
3615 /**
3616 * cpuset_hotplug_update_tasks - update tasks in a cpuset for hotunplug
3617 * @cs: cpuset in interest
3618 * @tmp: the tmpmasks structure pointer
3619 *
3620 * Compare @cs's cpu and mem masks against top_cpuset and if some have gone
3621 * offline, update @cs accordingly. If @cs ends up with no CPU or memory,
3622 * all its tasks are moved to the nearest ancestor with both resources.
3623 */
cpuset_hotplug_update_tasks(struct cpuset * cs,struct tmpmasks * tmp)3624 static void cpuset_hotplug_update_tasks(struct cpuset *cs, struct tmpmasks *tmp)
3625 {
3626 static cpumask_t new_cpus;
3627 static nodemask_t new_mems;
3628 bool cpus_updated;
3629 bool mems_updated;
3630 struct cpuset *parent;
3631 retry:
3632 wait_event(cpuset_attach_wq, cs->attach_in_progress == 0);
3633
3634 mutex_lock(&cpuset_mutex);
3635
3636 /*
3637 * We have raced with task attaching. We wait until attaching
3638 * is finished, so we won't attach a task to an empty cpuset.
3639 */
3640 if (cs->attach_in_progress) {
3641 mutex_unlock(&cpuset_mutex);
3642 goto retry;
3643 }
3644
3645 parent = parent_cs(cs);
3646 compute_effective_cpumask(&new_cpus, cs, parent);
3647 nodes_and(new_mems, cs->mems_allowed, parent->effective_mems);
3648
3649 if (cs->nr_subparts_cpus)
3650 /*
3651 * Make sure that CPUs allocated to child partitions
3652 * do not show up in effective_cpus.
3653 */
3654 cpumask_andnot(&new_cpus, &new_cpus, cs->subparts_cpus);
3655
3656 if (!tmp || !cs->partition_root_state)
3657 goto update_tasks;
3658
3659 /*
3660 * In the unlikely event that a partition root has empty
3661 * effective_cpus with tasks, we will have to invalidate child
3662 * partitions, if present, by setting nr_subparts_cpus to 0 to
3663 * reclaim their cpus.
3664 */
3665 if (cs->nr_subparts_cpus && is_partition_valid(cs) &&
3666 cpumask_empty(&new_cpus) && partition_is_populated(cs, NULL)) {
3667 spin_lock_irq(&callback_lock);
3668 cs->nr_subparts_cpus = 0;
3669 cpumask_clear(cs->subparts_cpus);
3670 spin_unlock_irq(&callback_lock);
3671 compute_effective_cpumask(&new_cpus, cs, parent);
3672 }
3673
3674 /*
3675 * Force the partition to become invalid if either one of
3676 * the following conditions hold:
3677 * 1) empty effective cpus but not valid empty partition.
3678 * 2) parent is invalid or doesn't grant any cpus to child
3679 * partitions.
3680 */
3681 if (is_partition_valid(cs) && (!parent->nr_subparts_cpus ||
3682 (cpumask_empty(&new_cpus) && partition_is_populated(cs, NULL)))) {
3683 int old_prs, parent_prs;
3684
3685 update_parent_subparts_cpumask(cs, partcmd_disable, NULL, tmp);
3686 if (cs->nr_subparts_cpus) {
3687 spin_lock_irq(&callback_lock);
3688 cs->nr_subparts_cpus = 0;
3689 cpumask_clear(cs->subparts_cpus);
3690 spin_unlock_irq(&callback_lock);
3691 compute_effective_cpumask(&new_cpus, cs, parent);
3692 }
3693
3694 old_prs = cs->partition_root_state;
3695 parent_prs = parent->partition_root_state;
3696 if (is_partition_valid(cs)) {
3697 spin_lock_irq(&callback_lock);
3698 make_partition_invalid(cs);
3699 spin_unlock_irq(&callback_lock);
3700 if (is_prs_invalid(parent_prs))
3701 WRITE_ONCE(cs->prs_err, PERR_INVPARENT);
3702 else if (!parent_prs)
3703 WRITE_ONCE(cs->prs_err, PERR_NOTPART);
3704 else
3705 WRITE_ONCE(cs->prs_err, PERR_HOTPLUG);
3706 notify_partition_change(cs, old_prs);
3707 }
3708 cpuset_force_rebuild();
3709 }
3710
3711 /*
3712 * On the other hand, an invalid partition root may be transitioned
3713 * back to a regular one.
3714 */
3715 else if (is_partition_valid(parent) && is_partition_invalid(cs)) {
3716 update_parent_subparts_cpumask(cs, partcmd_update, NULL, tmp);
3717 if (is_partition_valid(cs))
3718 cpuset_force_rebuild();
3719 }
3720
3721 update_tasks:
3722 cpus_updated = !cpumask_equal(&new_cpus, cs->effective_cpus);
3723 mems_updated = !nodes_equal(new_mems, cs->effective_mems);
3724
3725 if (mems_updated)
3726 check_insane_mems_config(&new_mems);
3727
3728 if (is_in_v2_mode())
3729 hotplug_update_tasks(cs, &new_cpus, &new_mems,
3730 cpus_updated, mems_updated);
3731 else
3732 hotplug_update_tasks_legacy(cs, &new_cpus, &new_mems,
3733 cpus_updated, mems_updated);
3734
3735 mutex_unlock(&cpuset_mutex);
3736 }
3737
3738 /**
3739 * cpuset_hotplug_workfn - handle CPU/memory hotunplug for a cpuset
3740 *
3741 * This function is called after either CPU or memory configuration has
3742 * changed and updates cpuset accordingly. The top_cpuset is always
3743 * synchronized to cpu_active_mask and N_MEMORY, which is necessary in
3744 * order to make cpusets transparent (of no affect) on systems that are
3745 * actively using CPU hotplug but making no active use of cpusets.
3746 *
3747 * Non-root cpusets are only affected by offlining. If any CPUs or memory
3748 * nodes have been taken down, cpuset_hotplug_update_tasks() is invoked on
3749 * all descendants.
3750 *
3751 * Note that CPU offlining during suspend is ignored. We don't modify
3752 * cpusets across suspend/resume cycles at all.
3753 */
cpuset_hotplug_workfn(struct work_struct * work)3754 static void cpuset_hotplug_workfn(struct work_struct *work)
3755 {
3756 static cpumask_t new_cpus;
3757 static nodemask_t new_mems;
3758 bool cpus_updated, mems_updated;
3759 bool on_dfl = is_in_v2_mode();
3760 struct tmpmasks tmp, *ptmp = NULL;
3761
3762 if (on_dfl && !alloc_cpumasks(NULL, &tmp))
3763 ptmp = &tmp;
3764
3765 mutex_lock(&cpuset_mutex);
3766
3767 /* fetch the available cpus/mems and find out which changed how */
3768 cpumask_copy(&new_cpus, cpu_active_mask);
3769 new_mems = node_states[N_MEMORY];
3770
3771 /*
3772 * If subparts_cpus is populated, it is likely that the check below
3773 * will produce a false positive on cpus_updated when the cpu list
3774 * isn't changed. It is extra work, but it is better to be safe.
3775 */
3776 cpus_updated = !cpumask_equal(top_cpuset.effective_cpus, &new_cpus);
3777 mems_updated = !nodes_equal(top_cpuset.effective_mems, new_mems);
3778
3779 /*
3780 * In the rare case that hotplug removes all the cpus in subparts_cpus,
3781 * we assumed that cpus are updated.
3782 */
3783 if (!cpus_updated && top_cpuset.nr_subparts_cpus)
3784 cpus_updated = true;
3785
3786 /* synchronize cpus_allowed to cpu_active_mask */
3787 if (cpus_updated) {
3788 spin_lock_irq(&callback_lock);
3789 if (!on_dfl)
3790 cpumask_copy(top_cpuset.cpus_allowed, &new_cpus);
3791 /*
3792 * Make sure that CPUs allocated to child partitions
3793 * do not show up in effective_cpus. If no CPU is left,
3794 * we clear the subparts_cpus & let the child partitions
3795 * fight for the CPUs again.
3796 */
3797 if (top_cpuset.nr_subparts_cpus) {
3798 if (cpumask_subset(&new_cpus,
3799 top_cpuset.subparts_cpus)) {
3800 top_cpuset.nr_subparts_cpus = 0;
3801 cpumask_clear(top_cpuset.subparts_cpus);
3802 } else {
3803 cpumask_andnot(&new_cpus, &new_cpus,
3804 top_cpuset.subparts_cpus);
3805 }
3806 }
3807 cpumask_copy(top_cpuset.effective_cpus, &new_cpus);
3808 spin_unlock_irq(&callback_lock);
3809 /* we don't mess with cpumasks of tasks in top_cpuset */
3810 }
3811
3812 /* synchronize mems_allowed to N_MEMORY */
3813 if (mems_updated) {
3814 spin_lock_irq(&callback_lock);
3815 if (!on_dfl)
3816 top_cpuset.mems_allowed = new_mems;
3817 top_cpuset.effective_mems = new_mems;
3818 spin_unlock_irq(&callback_lock);
3819 update_tasks_nodemask(&top_cpuset);
3820 }
3821
3822 mutex_unlock(&cpuset_mutex);
3823
3824 /* if cpus or mems changed, we need to propagate to descendants */
3825 if (cpus_updated || mems_updated) {
3826 struct cpuset *cs;
3827 struct cgroup_subsys_state *pos_css;
3828
3829 rcu_read_lock();
3830 cpuset_for_each_descendant_pre(cs, pos_css, &top_cpuset) {
3831 if (cs == &top_cpuset || !css_tryget_online(&cs->css))
3832 continue;
3833 rcu_read_unlock();
3834
3835 cpuset_hotplug_update_tasks(cs, ptmp);
3836
3837 rcu_read_lock();
3838 css_put(&cs->css);
3839 }
3840 rcu_read_unlock();
3841 }
3842
3843 /* rebuild sched domains if cpus_allowed has changed */
3844 if (cpus_updated || force_rebuild) {
3845 force_rebuild = false;
3846 rebuild_sched_domains();
3847 }
3848
3849 free_cpumasks(NULL, ptmp);
3850 }
3851
cpuset_update_active_cpus(void)3852 void cpuset_update_active_cpus(void)
3853 {
3854 /*
3855 * We're inside cpu hotplug critical region which usually nests
3856 * inside cgroup synchronization. Bounce actual hotplug processing
3857 * to a work item to avoid reverse locking order.
3858 */
3859 schedule_work(&cpuset_hotplug_work);
3860 }
3861
cpuset_wait_for_hotplug(void)3862 void cpuset_wait_for_hotplug(void)
3863 {
3864 flush_work(&cpuset_hotplug_work);
3865 }
3866
3867 /*
3868 * Keep top_cpuset.mems_allowed tracking node_states[N_MEMORY].
3869 * Call this routine anytime after node_states[N_MEMORY] changes.
3870 * See cpuset_update_active_cpus() for CPU hotplug handling.
3871 */
cpuset_track_online_nodes(struct notifier_block * self,unsigned long action,void * arg)3872 static int cpuset_track_online_nodes(struct notifier_block *self,
3873 unsigned long action, void *arg)
3874 {
3875 schedule_work(&cpuset_hotplug_work);
3876 return NOTIFY_OK;
3877 }
3878
3879 static struct notifier_block cpuset_track_online_nodes_nb = {
3880 .notifier_call = cpuset_track_online_nodes,
3881 .priority = 10, /* ??! */
3882 };
3883
3884 /**
3885 * cpuset_init_smp - initialize cpus_allowed
3886 *
3887 * Description: Finish top cpuset after cpu, node maps are initialized
3888 */
cpuset_init_smp(void)3889 void __init cpuset_init_smp(void)
3890 {
3891 /*
3892 * cpus_allowd/mems_allowed set to v2 values in the initial
3893 * cpuset_bind() call will be reset to v1 values in another
3894 * cpuset_bind() call when v1 cpuset is mounted.
3895 */
3896 top_cpuset.old_mems_allowed = top_cpuset.mems_allowed;
3897
3898 cpumask_copy(top_cpuset.effective_cpus, cpu_active_mask);
3899 top_cpuset.effective_mems = node_states[N_MEMORY];
3900
3901 register_hotmemory_notifier(&cpuset_track_online_nodes_nb);
3902
3903 cpuset_migrate_mm_wq = alloc_ordered_workqueue("cpuset_migrate_mm", 0);
3904 BUG_ON(!cpuset_migrate_mm_wq);
3905 }
3906
3907 /**
3908 * cpuset_cpus_allowed - return cpus_allowed mask from a tasks cpuset.
3909 * @tsk: pointer to task_struct from which to obtain cpuset->cpus_allowed.
3910 * @pmask: pointer to struct cpumask variable to receive cpus_allowed set.
3911 *
3912 * Description: Returns the cpumask_var_t cpus_allowed of the cpuset
3913 * attached to the specified @tsk. Guaranteed to return some non-empty
3914 * subset of cpu_online_mask, even if this means going outside the
3915 * tasks cpuset.
3916 **/
3917
cpuset_cpus_allowed(struct task_struct * tsk,struct cpumask * pmask)3918 void cpuset_cpus_allowed(struct task_struct *tsk, struct cpumask *pmask)
3919 {
3920 unsigned long flags;
3921
3922 spin_lock_irqsave(&callback_lock, flags);
3923 guarantee_online_cpus(tsk, pmask);
3924 spin_unlock_irqrestore(&callback_lock, flags);
3925 }
3926 EXPORT_SYMBOL_GPL(cpuset_cpus_allowed);
3927 /**
3928 * cpuset_cpus_allowed_fallback - final fallback before complete catastrophe.
3929 * @tsk: pointer to task_struct with which the scheduler is struggling
3930 *
3931 * Description: In the case that the scheduler cannot find an allowed cpu in
3932 * tsk->cpus_allowed, we fall back to task_cs(tsk)->cpus_allowed. In legacy
3933 * mode however, this value is the same as task_cs(tsk)->effective_cpus,
3934 * which will not contain a sane cpumask during cases such as cpu hotplugging.
3935 * This is the absolute last resort for the scheduler and it is only used if
3936 * _every_ other avenue has been traveled.
3937 *
3938 * Returns true if the affinity of @tsk was changed, false otherwise.
3939 **/
3940
cpuset_cpus_allowed_fallback(struct task_struct * tsk)3941 bool cpuset_cpus_allowed_fallback(struct task_struct *tsk)
3942 {
3943 const struct cpumask *possible_mask = task_cpu_possible_mask(tsk);
3944 const struct cpumask *cs_mask;
3945 bool changed = false;
3946
3947 rcu_read_lock();
3948 cs_mask = task_cs(tsk)->cpus_allowed;
3949 if (is_in_v2_mode() && cpumask_subset(cs_mask, possible_mask)) {
3950 do_set_cpus_allowed(tsk, cs_mask);
3951 changed = true;
3952 }
3953 rcu_read_unlock();
3954
3955 /*
3956 * We own tsk->cpus_allowed, nobody can change it under us.
3957 *
3958 * But we used cs && cs->cpus_allowed lockless and thus can
3959 * race with cgroup_attach_task() or update_cpumask() and get
3960 * the wrong tsk->cpus_allowed. However, both cases imply the
3961 * subsequent cpuset_change_cpumask()->set_cpus_allowed_ptr()
3962 * which takes task_rq_lock().
3963 *
3964 * If we are called after it dropped the lock we must see all
3965 * changes in tsk_cs()->cpus_allowed. Otherwise we can temporary
3966 * set any mask even if it is not right from task_cs() pov,
3967 * the pending set_cpus_allowed_ptr() will fix things.
3968 *
3969 * select_fallback_rq() will fix things ups and set cpu_possible_mask
3970 * if required.
3971 */
3972 return changed;
3973 }
3974
cpuset_init_current_mems_allowed(void)3975 void __init cpuset_init_current_mems_allowed(void)
3976 {
3977 nodes_setall(current->mems_allowed);
3978 }
3979
3980 /**
3981 * cpuset_mems_allowed - return mems_allowed mask from a tasks cpuset.
3982 * @tsk: pointer to task_struct from which to obtain cpuset->mems_allowed.
3983 *
3984 * Description: Returns the nodemask_t mems_allowed of the cpuset
3985 * attached to the specified @tsk. Guaranteed to return some non-empty
3986 * subset of node_states[N_MEMORY], even if this means going outside the
3987 * tasks cpuset.
3988 **/
3989
cpuset_mems_allowed(struct task_struct * tsk)3990 nodemask_t cpuset_mems_allowed(struct task_struct *tsk)
3991 {
3992 nodemask_t mask;
3993 unsigned long flags;
3994
3995 spin_lock_irqsave(&callback_lock, flags);
3996 rcu_read_lock();
3997 guarantee_online_mems(task_cs(tsk), &mask);
3998 rcu_read_unlock();
3999 spin_unlock_irqrestore(&callback_lock, flags);
4000
4001 return mask;
4002 }
4003
4004 /**
4005 * cpuset_nodemask_valid_mems_allowed - check nodemask vs. current mems_allowed
4006 * @nodemask: the nodemask to be checked
4007 *
4008 * Are any of the nodes in the nodemask allowed in current->mems_allowed?
4009 */
cpuset_nodemask_valid_mems_allowed(nodemask_t * nodemask)4010 int cpuset_nodemask_valid_mems_allowed(nodemask_t *nodemask)
4011 {
4012 return nodes_intersects(*nodemask, current->mems_allowed);
4013 }
4014
4015 /*
4016 * nearest_hardwall_ancestor() - Returns the nearest mem_exclusive or
4017 * mem_hardwall ancestor to the specified cpuset. Call holding
4018 * callback_lock. If no ancestor is mem_exclusive or mem_hardwall
4019 * (an unusual configuration), then returns the root cpuset.
4020 */
nearest_hardwall_ancestor(struct cpuset * cs)4021 static struct cpuset *nearest_hardwall_ancestor(struct cpuset *cs)
4022 {
4023 while (!(is_mem_exclusive(cs) || is_mem_hardwall(cs)) && parent_cs(cs))
4024 cs = parent_cs(cs);
4025 return cs;
4026 }
4027
4028 /*
4029 * __cpuset_node_allowed - Can we allocate on a memory node?
4030 * @node: is this an allowed node?
4031 * @gfp_mask: memory allocation flags
4032 *
4033 * If we're in interrupt, yes, we can always allocate. If @node is set in
4034 * current's mems_allowed, yes. If it's not a __GFP_HARDWALL request and this
4035 * node is set in the nearest hardwalled cpuset ancestor to current's cpuset,
4036 * yes. If current has access to memory reserves as an oom victim, yes.
4037 * Otherwise, no.
4038 *
4039 * GFP_USER allocations are marked with the __GFP_HARDWALL bit,
4040 * and do not allow allocations outside the current tasks cpuset
4041 * unless the task has been OOM killed.
4042 * GFP_KERNEL allocations are not so marked, so can escape to the
4043 * nearest enclosing hardwalled ancestor cpuset.
4044 *
4045 * Scanning up parent cpusets requires callback_lock. The
4046 * __alloc_pages() routine only calls here with __GFP_HARDWALL bit
4047 * _not_ set if it's a GFP_KERNEL allocation, and all nodes in the
4048 * current tasks mems_allowed came up empty on the first pass over
4049 * the zonelist. So only GFP_KERNEL allocations, if all nodes in the
4050 * cpuset are short of memory, might require taking the callback_lock.
4051 *
4052 * The first call here from mm/page_alloc:get_page_from_freelist()
4053 * has __GFP_HARDWALL set in gfp_mask, enforcing hardwall cpusets,
4054 * so no allocation on a node outside the cpuset is allowed (unless
4055 * in interrupt, of course).
4056 *
4057 * The second pass through get_page_from_freelist() doesn't even call
4058 * here for GFP_ATOMIC calls. For those calls, the __alloc_pages()
4059 * variable 'wait' is not set, and the bit ALLOC_CPUSET is not set
4060 * in alloc_flags. That logic and the checks below have the combined
4061 * affect that:
4062 * in_interrupt - any node ok (current task context irrelevant)
4063 * GFP_ATOMIC - any node ok
4064 * tsk_is_oom_victim - any node ok
4065 * GFP_KERNEL - any node in enclosing hardwalled cpuset ok
4066 * GFP_USER - only nodes in current tasks mems allowed ok.
4067 */
__cpuset_node_allowed(int node,gfp_t gfp_mask)4068 bool __cpuset_node_allowed(int node, gfp_t gfp_mask)
4069 {
4070 struct cpuset *cs; /* current cpuset ancestors */
4071 bool allowed; /* is allocation in zone z allowed? */
4072 unsigned long flags;
4073
4074 if (in_interrupt())
4075 return true;
4076 if (node_isset(node, current->mems_allowed))
4077 return true;
4078 /*
4079 * Allow tasks that have access to memory reserves because they have
4080 * been OOM killed to get memory anywhere.
4081 */
4082 if (unlikely(tsk_is_oom_victim(current)))
4083 return true;
4084 if (gfp_mask & __GFP_HARDWALL) /* If hardwall request, stop here */
4085 return false;
4086
4087 if (current->flags & PF_EXITING) /* Let dying task have memory */
4088 return true;
4089
4090 /* Not hardwall and node outside mems_allowed: scan up cpusets */
4091 spin_lock_irqsave(&callback_lock, flags);
4092
4093 rcu_read_lock();
4094 cs = nearest_hardwall_ancestor(task_cs(current));
4095 allowed = node_isset(node, cs->mems_allowed);
4096 rcu_read_unlock();
4097
4098 spin_unlock_irqrestore(&callback_lock, flags);
4099 return allowed;
4100 }
4101
4102 /**
4103 * cpuset_mem_spread_node() - On which node to begin search for a file page
4104 * cpuset_slab_spread_node() - On which node to begin search for a slab page
4105 *
4106 * If a task is marked PF_SPREAD_PAGE or PF_SPREAD_SLAB (as for
4107 * tasks in a cpuset with is_spread_page or is_spread_slab set),
4108 * and if the memory allocation used cpuset_mem_spread_node()
4109 * to determine on which node to start looking, as it will for
4110 * certain page cache or slab cache pages such as used for file
4111 * system buffers and inode caches, then instead of starting on the
4112 * local node to look for a free page, rather spread the starting
4113 * node around the tasks mems_allowed nodes.
4114 *
4115 * We don't have to worry about the returned node being offline
4116 * because "it can't happen", and even if it did, it would be ok.
4117 *
4118 * The routines calling guarantee_online_mems() are careful to
4119 * only set nodes in task->mems_allowed that are online. So it
4120 * should not be possible for the following code to return an
4121 * offline node. But if it did, that would be ok, as this routine
4122 * is not returning the node where the allocation must be, only
4123 * the node where the search should start. The zonelist passed to
4124 * __alloc_pages() will include all nodes. If the slab allocator
4125 * is passed an offline node, it will fall back to the local node.
4126 * See kmem_cache_alloc_node().
4127 */
4128
cpuset_spread_node(int * rotor)4129 static int cpuset_spread_node(int *rotor)
4130 {
4131 return *rotor = next_node_in(*rotor, current->mems_allowed);
4132 }
4133
cpuset_mem_spread_node(void)4134 int cpuset_mem_spread_node(void)
4135 {
4136 if (current->cpuset_mem_spread_rotor == NUMA_NO_NODE)
4137 current->cpuset_mem_spread_rotor =
4138 node_random(¤t->mems_allowed);
4139
4140 return cpuset_spread_node(¤t->cpuset_mem_spread_rotor);
4141 }
4142
cpuset_slab_spread_node(void)4143 int cpuset_slab_spread_node(void)
4144 {
4145 if (current->cpuset_slab_spread_rotor == NUMA_NO_NODE)
4146 current->cpuset_slab_spread_rotor =
4147 node_random(¤t->mems_allowed);
4148
4149 return cpuset_spread_node(¤t->cpuset_slab_spread_rotor);
4150 }
4151
4152 EXPORT_SYMBOL_GPL(cpuset_mem_spread_node);
4153
4154 /**
4155 * cpuset_mems_allowed_intersects - Does @tsk1's mems_allowed intersect @tsk2's?
4156 * @tsk1: pointer to task_struct of some task.
4157 * @tsk2: pointer to task_struct of some other task.
4158 *
4159 * Description: Return true if @tsk1's mems_allowed intersects the
4160 * mems_allowed of @tsk2. Used by the OOM killer to determine if
4161 * one of the task's memory usage might impact the memory available
4162 * to the other.
4163 **/
4164
cpuset_mems_allowed_intersects(const struct task_struct * tsk1,const struct task_struct * tsk2)4165 int cpuset_mems_allowed_intersects(const struct task_struct *tsk1,
4166 const struct task_struct *tsk2)
4167 {
4168 return nodes_intersects(tsk1->mems_allowed, tsk2->mems_allowed);
4169 }
4170
4171 /**
4172 * cpuset_print_current_mems_allowed - prints current's cpuset and mems_allowed
4173 *
4174 * Description: Prints current's name, cpuset name, and cached copy of its
4175 * mems_allowed to the kernel log.
4176 */
cpuset_print_current_mems_allowed(void)4177 void cpuset_print_current_mems_allowed(void)
4178 {
4179 struct cgroup *cgrp;
4180
4181 rcu_read_lock();
4182
4183 cgrp = task_cs(current)->css.cgroup;
4184 pr_cont(",cpuset=");
4185 pr_cont_cgroup_name(cgrp);
4186 pr_cont(",mems_allowed=%*pbl",
4187 nodemask_pr_args(¤t->mems_allowed));
4188
4189 rcu_read_unlock();
4190 }
4191
4192 /*
4193 * Collection of memory_pressure is suppressed unless
4194 * this flag is enabled by writing "1" to the special
4195 * cpuset file 'memory_pressure_enabled' in the root cpuset.
4196 */
4197
4198 int cpuset_memory_pressure_enabled __read_mostly;
4199
4200 /*
4201 * __cpuset_memory_pressure_bump - keep stats of per-cpuset reclaims.
4202 *
4203 * Keep a running average of the rate of synchronous (direct)
4204 * page reclaim efforts initiated by tasks in each cpuset.
4205 *
4206 * This represents the rate at which some task in the cpuset
4207 * ran low on memory on all nodes it was allowed to use, and
4208 * had to enter the kernels page reclaim code in an effort to
4209 * create more free memory by tossing clean pages or swapping
4210 * or writing dirty pages.
4211 *
4212 * Display to user space in the per-cpuset read-only file
4213 * "memory_pressure". Value displayed is an integer
4214 * representing the recent rate of entry into the synchronous
4215 * (direct) page reclaim by any task attached to the cpuset.
4216 */
4217
__cpuset_memory_pressure_bump(void)4218 void __cpuset_memory_pressure_bump(void)
4219 {
4220 rcu_read_lock();
4221 fmeter_markevent(&task_cs(current)->fmeter);
4222 rcu_read_unlock();
4223 }
4224
4225 #ifdef CONFIG_PROC_PID_CPUSET
4226 /*
4227 * proc_cpuset_show()
4228 * - Print tasks cpuset path into seq_file.
4229 * - Used for /proc/<pid>/cpuset.
4230 * - No need to task_lock(tsk) on this tsk->cpuset reference, as it
4231 * doesn't really matter if tsk->cpuset changes after we read it,
4232 * and we take cpuset_mutex, keeping cpuset_attach() from changing it
4233 * anyway.
4234 */
proc_cpuset_show(struct seq_file * m,struct pid_namespace * ns,struct pid * pid,struct task_struct * tsk)4235 int proc_cpuset_show(struct seq_file *m, struct pid_namespace *ns,
4236 struct pid *pid, struct task_struct *tsk)
4237 {
4238 char *buf;
4239 struct cgroup_subsys_state *css;
4240 int retval;
4241
4242 retval = -ENOMEM;
4243 buf = kmalloc(PATH_MAX, GFP_KERNEL);
4244 if (!buf)
4245 goto out;
4246
4247 css = task_get_css(tsk, cpuset_cgrp_id);
4248 retval = cgroup_path_ns(css->cgroup, buf, PATH_MAX,
4249 current->nsproxy->cgroup_ns);
4250 css_put(css);
4251 if (retval >= PATH_MAX)
4252 retval = -ENAMETOOLONG;
4253 if (retval < 0)
4254 goto out_free;
4255 seq_puts(m, buf);
4256 seq_putc(m, '\n');
4257 retval = 0;
4258 out_free:
4259 kfree(buf);
4260 out:
4261 return retval;
4262 }
4263 #endif /* CONFIG_PROC_PID_CPUSET */
4264
4265 /* Display task mems_allowed in /proc/<pid>/status file. */
cpuset_task_status_allowed(struct seq_file * m,struct task_struct * task)4266 void cpuset_task_status_allowed(struct seq_file *m, struct task_struct *task)
4267 {
4268 seq_printf(m, "Mems_allowed:\t%*pb\n",
4269 nodemask_pr_args(&task->mems_allowed));
4270 seq_printf(m, "Mems_allowed_list:\t%*pbl\n",
4271 nodemask_pr_args(&task->mems_allowed));
4272 }
4273