1 /*
2 * kernel/cpuset.c
3 *
4 * Processor and Memory placement constraints for sets of tasks.
5 *
6 * Copyright (C) 2003 BULL SA.
7 * Copyright (C) 2004-2007 Silicon Graphics, Inc.
8 * Copyright (C) 2006 Google, Inc
9 *
10 * Portions derived from Patrick Mochel's sysfs code.
11 * sysfs is Copyright (c) 2001-3 Patrick Mochel
12 *
13 * 2003-10-10 Written by Simon Derr.
14 * 2003-10-22 Updates by Stephen Hemminger.
15 * 2004 May-July Rework by Paul Jackson.
16 * 2006 Rework by Paul Menage to use generic cgroups
17 * 2008 Rework of the scheduler domains and CPU hotplug handling
18 * by Max Krasnyansky
19 *
20 * This file is subject to the terms and conditions of the GNU General Public
21 * License. See the file COPYING in the main directory of the Linux
22 * distribution for more details.
23 */
24
25 #include <linux/cpu.h>
26 #include <linux/cpumask.h>
27 #include <linux/cpuset.h>
28 #include <linux/err.h>
29 #include <linux/errno.h>
30 #include <linux/file.h>
31 #include <linux/fs.h>
32 #include <linux/init.h>
33 #include <linux/interrupt.h>
34 #include <linux/kernel.h>
35 #include <linux/kmod.h>
36 #include <linux/list.h>
37 #include <linux/mempolicy.h>
38 #include <linux/mm.h>
39 #include <linux/memory.h>
40 #include <linux/export.h>
41 #include <linux/mount.h>
42 #include <linux/namei.h>
43 #include <linux/pagemap.h>
44 #include <linux/proc_fs.h>
45 #include <linux/rcupdate.h>
46 #include <linux/sched.h>
47 #include <linux/seq_file.h>
48 #include <linux/security.h>
49 #include <linux/slab.h>
50 #include <linux/spinlock.h>
51 #include <linux/stat.h>
52 #include <linux/string.h>
53 #include <linux/time.h>
54 #include <linux/backing-dev.h>
55 #include <linux/sort.h>
56
57 #include <asm/uaccess.h>
58 #include <linux/atomic.h>
59 #include <linux/mutex.h>
60 #include <linux/cgroup.h>
61 #include <linux/wait.h>
62
63 struct static_key cpusets_pre_enable_key __read_mostly = STATIC_KEY_INIT_FALSE;
64 struct static_key cpusets_enabled_key __read_mostly = STATIC_KEY_INIT_FALSE;
65
66 /* See "Frequency meter" comments, below. */
67
68 struct fmeter {
69 int cnt; /* unprocessed events count */
70 int val; /* most recent output value */
71 time_t time; /* clock (secs) when val computed */
72 spinlock_t lock; /* guards read or write of above */
73 };
74
75 struct cpuset {
76 struct cgroup_subsys_state css;
77
78 unsigned long flags; /* "unsigned long" so bitops work */
79
80 /*
81 * On default hierarchy:
82 *
83 * The user-configured masks can only be changed by writing to
84 * cpuset.cpus and cpuset.mems, and won't be limited by the
85 * parent masks.
86 *
87 * The effective masks is the real masks that apply to the tasks
88 * in the cpuset. They may be changed if the configured masks are
89 * changed or hotplug happens.
90 *
91 * effective_mask == configured_mask & parent's effective_mask,
92 * and if it ends up empty, it will inherit the parent's mask.
93 *
94 *
95 * On legacy hierachy:
96 *
97 * The user-configured masks are always the same with effective masks.
98 */
99
100 /* user-configured CPUs and Memory Nodes allow to tasks */
101 cpumask_var_t cpus_allowed;
102 cpumask_var_t cpus_requested;
103 nodemask_t mems_allowed;
104
105 /* effective CPUs and Memory Nodes allow to tasks */
106 cpumask_var_t effective_cpus;
107 nodemask_t effective_mems;
108
109 /*
110 * This is old Memory Nodes tasks took on.
111 *
112 * - top_cpuset.old_mems_allowed is initialized to mems_allowed.
113 * - A new cpuset's old_mems_allowed is initialized when some
114 * task is moved into it.
115 * - old_mems_allowed is used in cpuset_migrate_mm() when we change
116 * cpuset.mems_allowed and have tasks' nodemask updated, and
117 * then old_mems_allowed is updated to mems_allowed.
118 */
119 nodemask_t old_mems_allowed;
120
121 struct fmeter fmeter; /* memory_pressure filter */
122
123 /*
124 * Tasks are being attached to this cpuset. Used to prevent
125 * zeroing cpus/mems_allowed between ->can_attach() and ->attach().
126 */
127 int attach_in_progress;
128
129 /* partition number for rebuild_sched_domains() */
130 int pn;
131
132 /* for custom sched domain */
133 int relax_domain_level;
134 };
135
css_cs(struct cgroup_subsys_state * css)136 static inline struct cpuset *css_cs(struct cgroup_subsys_state *css)
137 {
138 return css ? container_of(css, struct cpuset, css) : NULL;
139 }
140
141 /* Retrieve the cpuset for a task */
task_cs(struct task_struct * task)142 static inline struct cpuset *task_cs(struct task_struct *task)
143 {
144 return css_cs(task_css(task, cpuset_cgrp_id));
145 }
146
parent_cs(struct cpuset * cs)147 static inline struct cpuset *parent_cs(struct cpuset *cs)
148 {
149 return css_cs(cs->css.parent);
150 }
151
152 #ifdef CONFIG_NUMA
task_has_mempolicy(struct task_struct * task)153 static inline bool task_has_mempolicy(struct task_struct *task)
154 {
155 return task->mempolicy;
156 }
157 #else
task_has_mempolicy(struct task_struct * task)158 static inline bool task_has_mempolicy(struct task_struct *task)
159 {
160 return false;
161 }
162 #endif
163
164
165 /* bits in struct cpuset flags field */
166 typedef enum {
167 CS_ONLINE,
168 CS_CPU_EXCLUSIVE,
169 CS_MEM_EXCLUSIVE,
170 CS_MEM_HARDWALL,
171 CS_MEMORY_MIGRATE,
172 CS_SCHED_LOAD_BALANCE,
173 CS_SPREAD_PAGE,
174 CS_SPREAD_SLAB,
175 } cpuset_flagbits_t;
176
177 /* convenient tests for these bits */
is_cpuset_online(struct cpuset * cs)178 static inline bool is_cpuset_online(struct cpuset *cs)
179 {
180 return test_bit(CS_ONLINE, &cs->flags) && !css_is_dying(&cs->css);
181 }
182
is_cpu_exclusive(const struct cpuset * cs)183 static inline int is_cpu_exclusive(const struct cpuset *cs)
184 {
185 return test_bit(CS_CPU_EXCLUSIVE, &cs->flags);
186 }
187
is_mem_exclusive(const struct cpuset * cs)188 static inline int is_mem_exclusive(const struct cpuset *cs)
189 {
190 return test_bit(CS_MEM_EXCLUSIVE, &cs->flags);
191 }
192
is_mem_hardwall(const struct cpuset * cs)193 static inline int is_mem_hardwall(const struct cpuset *cs)
194 {
195 return test_bit(CS_MEM_HARDWALL, &cs->flags);
196 }
197
is_sched_load_balance(const struct cpuset * cs)198 static inline int is_sched_load_balance(const struct cpuset *cs)
199 {
200 return test_bit(CS_SCHED_LOAD_BALANCE, &cs->flags);
201 }
202
is_memory_migrate(const struct cpuset * cs)203 static inline int is_memory_migrate(const struct cpuset *cs)
204 {
205 return test_bit(CS_MEMORY_MIGRATE, &cs->flags);
206 }
207
is_spread_page(const struct cpuset * cs)208 static inline int is_spread_page(const struct cpuset *cs)
209 {
210 return test_bit(CS_SPREAD_PAGE, &cs->flags);
211 }
212
is_spread_slab(const struct cpuset * cs)213 static inline int is_spread_slab(const struct cpuset *cs)
214 {
215 return test_bit(CS_SPREAD_SLAB, &cs->flags);
216 }
217
218 static struct cpuset top_cpuset = {
219 .flags = ((1 << CS_ONLINE) | (1 << CS_CPU_EXCLUSIVE) |
220 (1 << CS_MEM_EXCLUSIVE)),
221 };
222
223 /**
224 * cpuset_for_each_child - traverse online children of a cpuset
225 * @child_cs: loop cursor pointing to the current child
226 * @pos_css: used for iteration
227 * @parent_cs: target cpuset to walk children of
228 *
229 * Walk @child_cs through the online children of @parent_cs. Must be used
230 * with RCU read locked.
231 */
232 #define cpuset_for_each_child(child_cs, pos_css, parent_cs) \
233 css_for_each_child((pos_css), &(parent_cs)->css) \
234 if (is_cpuset_online(((child_cs) = css_cs((pos_css)))))
235
236 /**
237 * cpuset_for_each_descendant_pre - pre-order walk of a cpuset's descendants
238 * @des_cs: loop cursor pointing to the current descendant
239 * @pos_css: used for iteration
240 * @root_cs: target cpuset to walk ancestor of
241 *
242 * Walk @des_cs through the online descendants of @root_cs. Must be used
243 * with RCU read locked. The caller may modify @pos_css by calling
244 * css_rightmost_descendant() to skip subtree. @root_cs is included in the
245 * iteration and the first node to be visited.
246 */
247 #define cpuset_for_each_descendant_pre(des_cs, pos_css, root_cs) \
248 css_for_each_descendant_pre((pos_css), &(root_cs)->css) \
249 if (is_cpuset_online(((des_cs) = css_cs((pos_css)))))
250
251 /*
252 * There are two global locks guarding cpuset structures - cpuset_mutex and
253 * callback_lock. We also require taking task_lock() when dereferencing a
254 * task's cpuset pointer. See "The task_lock() exception", at the end of this
255 * comment.
256 *
257 * A task must hold both locks to modify cpusets. If a task holds
258 * cpuset_mutex, then it blocks others wanting that mutex, ensuring that it
259 * is the only task able to also acquire callback_lock and be able to
260 * modify cpusets. It can perform various checks on the cpuset structure
261 * first, knowing nothing will change. It can also allocate memory while
262 * just holding cpuset_mutex. While it is performing these checks, various
263 * callback routines can briefly acquire callback_lock to query cpusets.
264 * Once it is ready to make the changes, it takes callback_lock, blocking
265 * everyone else.
266 *
267 * Calls to the kernel memory allocator can not be made while holding
268 * callback_lock, as that would risk double tripping on callback_lock
269 * from one of the callbacks into the cpuset code from within
270 * __alloc_pages().
271 *
272 * If a task is only holding callback_lock, then it has read-only
273 * access to cpusets.
274 *
275 * Now, the task_struct fields mems_allowed and mempolicy may be changed
276 * by other task, we use alloc_lock in the task_struct fields to protect
277 * them.
278 *
279 * The cpuset_common_file_read() handlers only hold callback_lock across
280 * small pieces of code, such as when reading out possibly multi-word
281 * cpumasks and nodemasks.
282 *
283 * Accessing a task's cpuset should be done in accordance with the
284 * guidelines for accessing subsystem state in kernel/cgroup.c
285 */
286
287 static DEFINE_MUTEX(cpuset_mutex);
288 static DEFINE_SPINLOCK(callback_lock);
289
290 static struct workqueue_struct *cpuset_migrate_mm_wq;
291
292 /*
293 * CPU / memory hotplug is handled asynchronously.
294 */
295 static void cpuset_hotplug_workfn(struct work_struct *work);
296 static DECLARE_WORK(cpuset_hotplug_work, cpuset_hotplug_workfn);
297
298 static DECLARE_WAIT_QUEUE_HEAD(cpuset_attach_wq);
299
300 /*
301 * This is ugly, but preserves the userspace API for existing cpuset
302 * users. If someone tries to mount the "cpuset" filesystem, we
303 * silently switch it to mount "cgroup" instead
304 */
cpuset_mount(struct file_system_type * fs_type,int flags,const char * unused_dev_name,void * data)305 static struct dentry *cpuset_mount(struct file_system_type *fs_type,
306 int flags, const char *unused_dev_name, void *data)
307 {
308 struct file_system_type *cgroup_fs = get_fs_type("cgroup");
309 struct dentry *ret = ERR_PTR(-ENODEV);
310 if (cgroup_fs) {
311 char mountopts[] =
312 "cpuset,noprefix,"
313 "release_agent=/sbin/cpuset_release_agent";
314 ret = cgroup_fs->mount(cgroup_fs, flags,
315 unused_dev_name, mountopts);
316 put_filesystem(cgroup_fs);
317 }
318 return ret;
319 }
320
321 static struct file_system_type cpuset_fs_type = {
322 .name = "cpuset",
323 .mount = cpuset_mount,
324 };
325
326 /*
327 * Return in pmask the portion of a cpusets's cpus_allowed that
328 * are online. If none are online, walk up the cpuset hierarchy
329 * until we find one that does have some online cpus.
330 *
331 * One way or another, we guarantee to return some non-empty subset
332 * of cpu_online_mask.
333 *
334 * Call with callback_lock or cpuset_mutex held.
335 */
guarantee_online_cpus(struct cpuset * cs,struct cpumask * pmask)336 static void guarantee_online_cpus(struct cpuset *cs, struct cpumask *pmask)
337 {
338 while (!cpumask_intersects(cs->effective_cpus, cpu_online_mask)) {
339 cs = parent_cs(cs);
340 if (unlikely(!cs)) {
341 /*
342 * The top cpuset doesn't have any online cpu as a
343 * consequence of a race between cpuset_hotplug_work
344 * and cpu hotplug notifier. But we know the top
345 * cpuset's effective_cpus is on its way to to be
346 * identical to cpu_online_mask.
347 */
348 cpumask_copy(pmask, cpu_online_mask);
349 return;
350 }
351 }
352 cpumask_and(pmask, cs->effective_cpus, cpu_online_mask);
353 }
354
355 /*
356 * Return in *pmask the portion of a cpusets's mems_allowed that
357 * are online, with memory. If none are online with memory, walk
358 * up the cpuset hierarchy until we find one that does have some
359 * online mems. The top cpuset always has some mems online.
360 *
361 * One way or another, we guarantee to return some non-empty subset
362 * of node_states[N_MEMORY].
363 *
364 * Call with callback_lock or cpuset_mutex held.
365 */
guarantee_online_mems(struct cpuset * cs,nodemask_t * pmask)366 static void guarantee_online_mems(struct cpuset *cs, nodemask_t *pmask)
367 {
368 while (!nodes_intersects(cs->effective_mems, node_states[N_MEMORY]))
369 cs = parent_cs(cs);
370 nodes_and(*pmask, cs->effective_mems, node_states[N_MEMORY]);
371 }
372
373 /*
374 * update task's spread flag if cpuset's page/slab spread flag is set
375 *
376 * Call with callback_lock or cpuset_mutex held.
377 */
cpuset_update_task_spread_flag(struct cpuset * cs,struct task_struct * tsk)378 static void cpuset_update_task_spread_flag(struct cpuset *cs,
379 struct task_struct *tsk)
380 {
381 if (is_spread_page(cs))
382 task_set_spread_page(tsk);
383 else
384 task_clear_spread_page(tsk);
385
386 if (is_spread_slab(cs))
387 task_set_spread_slab(tsk);
388 else
389 task_clear_spread_slab(tsk);
390 }
391
392 /*
393 * is_cpuset_subset(p, q) - Is cpuset p a subset of cpuset q?
394 *
395 * One cpuset is a subset of another if all its allowed CPUs and
396 * Memory Nodes are a subset of the other, and its exclusive flags
397 * are only set if the other's are set. Call holding cpuset_mutex.
398 */
399
is_cpuset_subset(const struct cpuset * p,const struct cpuset * q)400 static int is_cpuset_subset(const struct cpuset *p, const struct cpuset *q)
401 {
402 return cpumask_subset(p->cpus_requested, q->cpus_requested) &&
403 nodes_subset(p->mems_allowed, q->mems_allowed) &&
404 is_cpu_exclusive(p) <= is_cpu_exclusive(q) &&
405 is_mem_exclusive(p) <= is_mem_exclusive(q);
406 }
407
408 /**
409 * alloc_trial_cpuset - allocate a trial cpuset
410 * @cs: the cpuset that the trial cpuset duplicates
411 */
alloc_trial_cpuset(struct cpuset * cs)412 static struct cpuset *alloc_trial_cpuset(struct cpuset *cs)
413 {
414 struct cpuset *trial;
415
416 trial = kmemdup(cs, sizeof(*cs), GFP_KERNEL);
417 if (!trial)
418 return NULL;
419
420 if (!alloc_cpumask_var(&trial->cpus_allowed, GFP_KERNEL))
421 goto free_cs;
422 if (!alloc_cpumask_var(&trial->effective_cpus, GFP_KERNEL))
423 goto free_cpus;
424
425 cpumask_copy(trial->cpus_allowed, cs->cpus_allowed);
426 cpumask_copy(trial->effective_cpus, cs->effective_cpus);
427 return trial;
428
429 free_cpus:
430 free_cpumask_var(trial->cpus_allowed);
431 free_cs:
432 kfree(trial);
433 return NULL;
434 }
435
436 /**
437 * free_trial_cpuset - free the trial cpuset
438 * @trial: the trial cpuset to be freed
439 */
free_trial_cpuset(struct cpuset * trial)440 static void free_trial_cpuset(struct cpuset *trial)
441 {
442 free_cpumask_var(trial->effective_cpus);
443 free_cpumask_var(trial->cpus_allowed);
444 kfree(trial);
445 }
446
447 /*
448 * validate_change() - Used to validate that any proposed cpuset change
449 * follows the structural rules for cpusets.
450 *
451 * If we replaced the flag and mask values of the current cpuset
452 * (cur) with those values in the trial cpuset (trial), would
453 * our various subset and exclusive rules still be valid? Presumes
454 * cpuset_mutex held.
455 *
456 * 'cur' is the address of an actual, in-use cpuset. Operations
457 * such as list traversal that depend on the actual address of the
458 * cpuset in the list must use cur below, not trial.
459 *
460 * 'trial' is the address of bulk structure copy of cur, with
461 * perhaps one or more of the fields cpus_allowed, mems_allowed,
462 * or flags changed to new, trial values.
463 *
464 * Return 0 if valid, -errno if not.
465 */
466
validate_change(struct cpuset * cur,struct cpuset * trial)467 static int validate_change(struct cpuset *cur, struct cpuset *trial)
468 {
469 struct cgroup_subsys_state *css;
470 struct cpuset *c, *par;
471 int ret;
472
473 rcu_read_lock();
474
475 /* Each of our child cpusets must be a subset of us */
476 ret = -EBUSY;
477 cpuset_for_each_child(c, css, cur)
478 if (!is_cpuset_subset(c, trial))
479 goto out;
480
481 /* Remaining checks don't apply to root cpuset */
482 ret = 0;
483 if (cur == &top_cpuset)
484 goto out;
485
486 par = parent_cs(cur);
487
488 /* On legacy hiearchy, we must be a subset of our parent cpuset. */
489 ret = -EACCES;
490 if (!cgroup_subsys_on_dfl(cpuset_cgrp_subsys) &&
491 !is_cpuset_subset(trial, par))
492 goto out;
493
494 /*
495 * If either I or some sibling (!= me) is exclusive, we can't
496 * overlap
497 */
498 ret = -EINVAL;
499 cpuset_for_each_child(c, css, par) {
500 if ((is_cpu_exclusive(trial) || is_cpu_exclusive(c)) &&
501 c != cur &&
502 cpumask_intersects(trial->cpus_requested, c->cpus_requested))
503 goto out;
504 if ((is_mem_exclusive(trial) || is_mem_exclusive(c)) &&
505 c != cur &&
506 nodes_intersects(trial->mems_allowed, c->mems_allowed))
507 goto out;
508 }
509
510 /*
511 * Cpusets with tasks - existing or newly being attached - can't
512 * be changed to have empty cpus_allowed or mems_allowed.
513 */
514 ret = -ENOSPC;
515 if ((cgroup_is_populated(cur->css.cgroup) || cur->attach_in_progress)) {
516 if (!cpumask_empty(cur->cpus_allowed) &&
517 cpumask_empty(trial->cpus_allowed))
518 goto out;
519 if (!nodes_empty(cur->mems_allowed) &&
520 nodes_empty(trial->mems_allowed))
521 goto out;
522 }
523
524 /*
525 * We can't shrink if we won't have enough room for SCHED_DEADLINE
526 * tasks.
527 */
528 ret = -EBUSY;
529 if (is_cpu_exclusive(cur) &&
530 !cpuset_cpumask_can_shrink(cur->cpus_allowed,
531 trial->cpus_allowed))
532 goto out;
533
534 ret = 0;
535 out:
536 rcu_read_unlock();
537 return ret;
538 }
539
540 #ifdef CONFIG_SMP
541 /*
542 * Helper routine for generate_sched_domains().
543 * Do cpusets a, b have overlapping effective cpus_allowed masks?
544 */
cpusets_overlap(struct cpuset * a,struct cpuset * b)545 static int cpusets_overlap(struct cpuset *a, struct cpuset *b)
546 {
547 return cpumask_intersects(a->effective_cpus, b->effective_cpus);
548 }
549
550 static void
update_domain_attr(struct sched_domain_attr * dattr,struct cpuset * c)551 update_domain_attr(struct sched_domain_attr *dattr, struct cpuset *c)
552 {
553 if (dattr->relax_domain_level < c->relax_domain_level)
554 dattr->relax_domain_level = c->relax_domain_level;
555 return;
556 }
557
update_domain_attr_tree(struct sched_domain_attr * dattr,struct cpuset * root_cs)558 static void update_domain_attr_tree(struct sched_domain_attr *dattr,
559 struct cpuset *root_cs)
560 {
561 struct cpuset *cp;
562 struct cgroup_subsys_state *pos_css;
563
564 rcu_read_lock();
565 cpuset_for_each_descendant_pre(cp, pos_css, root_cs) {
566 /* skip the whole subtree if @cp doesn't have any CPU */
567 if (cpumask_empty(cp->cpus_allowed)) {
568 pos_css = css_rightmost_descendant(pos_css);
569 continue;
570 }
571
572 if (is_sched_load_balance(cp))
573 update_domain_attr(dattr, cp);
574 }
575 rcu_read_unlock();
576 }
577
578 /*
579 * generate_sched_domains()
580 *
581 * This function builds a partial partition of the systems CPUs
582 * A 'partial partition' is a set of non-overlapping subsets whose
583 * union is a subset of that set.
584 * The output of this function needs to be passed to kernel/sched/core.c
585 * partition_sched_domains() routine, which will rebuild the scheduler's
586 * load balancing domains (sched domains) as specified by that partial
587 * partition.
588 *
589 * See "What is sched_load_balance" in Documentation/cgroups/cpusets.txt
590 * for a background explanation of this.
591 *
592 * Does not return errors, on the theory that the callers of this
593 * routine would rather not worry about failures to rebuild sched
594 * domains when operating in the severe memory shortage situations
595 * that could cause allocation failures below.
596 *
597 * Must be called with cpuset_mutex held.
598 *
599 * The three key local variables below are:
600 * q - a linked-list queue of cpuset pointers, used to implement a
601 * top-down scan of all cpusets. This scan loads a pointer
602 * to each cpuset marked is_sched_load_balance into the
603 * array 'csa'. For our purposes, rebuilding the schedulers
604 * sched domains, we can ignore !is_sched_load_balance cpusets.
605 * csa - (for CpuSet Array) Array of pointers to all the cpusets
606 * that need to be load balanced, for convenient iterative
607 * access by the subsequent code that finds the best partition,
608 * i.e the set of domains (subsets) of CPUs such that the
609 * cpus_allowed of every cpuset marked is_sched_load_balance
610 * is a subset of one of these domains, while there are as
611 * many such domains as possible, each as small as possible.
612 * doms - Conversion of 'csa' to an array of cpumasks, for passing to
613 * the kernel/sched/core.c routine partition_sched_domains() in a
614 * convenient format, that can be easily compared to the prior
615 * value to determine what partition elements (sched domains)
616 * were changed (added or removed.)
617 *
618 * Finding the best partition (set of domains):
619 * The triple nested loops below over i, j, k scan over the
620 * load balanced cpusets (using the array of cpuset pointers in
621 * csa[]) looking for pairs of cpusets that have overlapping
622 * cpus_allowed, but which don't have the same 'pn' partition
623 * number and gives them in the same partition number. It keeps
624 * looping on the 'restart' label until it can no longer find
625 * any such pairs.
626 *
627 * The union of the cpus_allowed masks from the set of
628 * all cpusets having the same 'pn' value then form the one
629 * element of the partition (one sched domain) to be passed to
630 * partition_sched_domains().
631 */
generate_sched_domains(cpumask_var_t ** domains,struct sched_domain_attr ** attributes)632 static int generate_sched_domains(cpumask_var_t **domains,
633 struct sched_domain_attr **attributes)
634 {
635 struct cpuset *cp; /* scans q */
636 struct cpuset **csa; /* array of all cpuset ptrs */
637 int csn; /* how many cpuset ptrs in csa so far */
638 int i, j, k; /* indices for partition finding loops */
639 cpumask_var_t *doms; /* resulting partition; i.e. sched domains */
640 cpumask_var_t non_isolated_cpus; /* load balanced CPUs */
641 struct sched_domain_attr *dattr; /* attributes for custom domains */
642 int ndoms = 0; /* number of sched domains in result */
643 int nslot; /* next empty doms[] struct cpumask slot */
644 struct cgroup_subsys_state *pos_css;
645
646 doms = NULL;
647 dattr = NULL;
648 csa = NULL;
649
650 if (!alloc_cpumask_var(&non_isolated_cpus, GFP_KERNEL))
651 goto done;
652 cpumask_andnot(non_isolated_cpus, cpu_possible_mask, cpu_isolated_map);
653
654 /* Special case for the 99% of systems with one, full, sched domain */
655 if (is_sched_load_balance(&top_cpuset)) {
656 ndoms = 1;
657 doms = alloc_sched_domains(ndoms);
658 if (!doms)
659 goto done;
660
661 dattr = kmalloc(sizeof(struct sched_domain_attr), GFP_KERNEL);
662 if (dattr) {
663 *dattr = SD_ATTR_INIT;
664 update_domain_attr_tree(dattr, &top_cpuset);
665 }
666 cpumask_and(doms[0], top_cpuset.effective_cpus,
667 non_isolated_cpus);
668
669 goto done;
670 }
671
672 csa = kmalloc(nr_cpusets() * sizeof(cp), GFP_KERNEL);
673 if (!csa)
674 goto done;
675 csn = 0;
676
677 rcu_read_lock();
678 cpuset_for_each_descendant_pre(cp, pos_css, &top_cpuset) {
679 if (cp == &top_cpuset)
680 continue;
681 /*
682 * Continue traversing beyond @cp iff @cp has some CPUs and
683 * isn't load balancing. The former is obvious. The
684 * latter: All child cpusets contain a subset of the
685 * parent's cpus, so just skip them, and then we call
686 * update_domain_attr_tree() to calc relax_domain_level of
687 * the corresponding sched domain.
688 */
689 if (!cpumask_empty(cp->cpus_allowed) &&
690 !(is_sched_load_balance(cp) &&
691 cpumask_intersects(cp->cpus_allowed, non_isolated_cpus)))
692 continue;
693
694 if (is_sched_load_balance(cp))
695 csa[csn++] = cp;
696
697 /* skip @cp's subtree */
698 pos_css = css_rightmost_descendant(pos_css);
699 }
700 rcu_read_unlock();
701
702 for (i = 0; i < csn; i++)
703 csa[i]->pn = i;
704 ndoms = csn;
705
706 restart:
707 /* Find the best partition (set of sched domains) */
708 for (i = 0; i < csn; i++) {
709 struct cpuset *a = csa[i];
710 int apn = a->pn;
711
712 for (j = 0; j < csn; j++) {
713 struct cpuset *b = csa[j];
714 int bpn = b->pn;
715
716 if (apn != bpn && cpusets_overlap(a, b)) {
717 for (k = 0; k < csn; k++) {
718 struct cpuset *c = csa[k];
719
720 if (c->pn == bpn)
721 c->pn = apn;
722 }
723 ndoms--; /* one less element */
724 goto restart;
725 }
726 }
727 }
728
729 /*
730 * Now we know how many domains to create.
731 * Convert <csn, csa> to <ndoms, doms> and populate cpu masks.
732 */
733 doms = alloc_sched_domains(ndoms);
734 if (!doms)
735 goto done;
736
737 /*
738 * The rest of the code, including the scheduler, can deal with
739 * dattr==NULL case. No need to abort if alloc fails.
740 */
741 dattr = kmalloc(ndoms * sizeof(struct sched_domain_attr), GFP_KERNEL);
742
743 for (nslot = 0, i = 0; i < csn; i++) {
744 struct cpuset *a = csa[i];
745 struct cpumask *dp;
746 int apn = a->pn;
747
748 if (apn < 0) {
749 /* Skip completed partitions */
750 continue;
751 }
752
753 dp = doms[nslot];
754
755 if (nslot == ndoms) {
756 static int warnings = 10;
757 if (warnings) {
758 pr_warn("rebuild_sched_domains confused: nslot %d, ndoms %d, csn %d, i %d, apn %d\n",
759 nslot, ndoms, csn, i, apn);
760 warnings--;
761 }
762 continue;
763 }
764
765 cpumask_clear(dp);
766 if (dattr)
767 *(dattr + nslot) = SD_ATTR_INIT;
768 for (j = i; j < csn; j++) {
769 struct cpuset *b = csa[j];
770
771 if (apn == b->pn) {
772 cpumask_or(dp, dp, b->effective_cpus);
773 cpumask_and(dp, dp, non_isolated_cpus);
774 if (dattr)
775 update_domain_attr_tree(dattr + nslot, b);
776
777 /* Done with this partition */
778 b->pn = -1;
779 }
780 }
781 nslot++;
782 }
783 BUG_ON(nslot != ndoms);
784
785 done:
786 free_cpumask_var(non_isolated_cpus);
787 kfree(csa);
788
789 /*
790 * Fallback to the default domain if kmalloc() failed.
791 * See comments in partition_sched_domains().
792 */
793 if (doms == NULL)
794 ndoms = 1;
795
796 *domains = doms;
797 *attributes = dattr;
798 return ndoms;
799 }
800
801 /*
802 * Rebuild scheduler domains.
803 *
804 * If the flag 'sched_load_balance' of any cpuset with non-empty
805 * 'cpus' changes, or if the 'cpus' allowed changes in any cpuset
806 * which has that flag enabled, or if any cpuset with a non-empty
807 * 'cpus' is removed, then call this routine to rebuild the
808 * scheduler's dynamic sched domains.
809 *
810 * Call with cpuset_mutex held. Takes get_online_cpus().
811 */
rebuild_sched_domains_locked(void)812 static void rebuild_sched_domains_locked(void)
813 {
814 struct sched_domain_attr *attr;
815 cpumask_var_t *doms;
816 int ndoms;
817
818 lockdep_assert_held(&cpuset_mutex);
819 get_online_cpus();
820
821 /*
822 * We have raced with CPU hotplug. Don't do anything to avoid
823 * passing doms with offlined cpu to partition_sched_domains().
824 * Anyways, hotplug work item will rebuild sched domains.
825 */
826 if (!cpumask_equal(top_cpuset.effective_cpus, cpu_active_mask))
827 goto out;
828
829 /* Generate domain masks and attrs */
830 ndoms = generate_sched_domains(&doms, &attr);
831
832 /* Have scheduler rebuild the domains */
833 partition_sched_domains(ndoms, doms, attr);
834 out:
835 put_online_cpus();
836 }
837 #else /* !CONFIG_SMP */
rebuild_sched_domains_locked(void)838 static void rebuild_sched_domains_locked(void)
839 {
840 }
841 #endif /* CONFIG_SMP */
842
rebuild_sched_domains(void)843 void rebuild_sched_domains(void)
844 {
845 mutex_lock(&cpuset_mutex);
846 rebuild_sched_domains_locked();
847 mutex_unlock(&cpuset_mutex);
848 }
849
850 /**
851 * update_tasks_cpumask - Update the cpumasks of tasks in the cpuset.
852 * @cs: the cpuset in which each task's cpus_allowed mask needs to be changed
853 *
854 * Iterate through each task of @cs updating its cpus_allowed to the
855 * effective cpuset's. As this function is called with cpuset_mutex held,
856 * cpuset membership stays stable.
857 */
update_tasks_cpumask(struct cpuset * cs)858 static void update_tasks_cpumask(struct cpuset *cs)
859 {
860 struct css_task_iter it;
861 struct task_struct *task;
862
863 css_task_iter_start(&cs->css, &it);
864 while ((task = css_task_iter_next(&it)))
865 set_cpus_allowed_ptr(task, cs->effective_cpus);
866 css_task_iter_end(&it);
867 }
868
869 /*
870 * update_cpumasks_hier - Update effective cpumasks and tasks in the subtree
871 * @cs: the cpuset to consider
872 * @new_cpus: temp variable for calculating new effective_cpus
873 *
874 * When congifured cpumask is changed, the effective cpumasks of this cpuset
875 * and all its descendants need to be updated.
876 *
877 * On legacy hierachy, effective_cpus will be the same with cpu_allowed.
878 *
879 * Called with cpuset_mutex held
880 */
update_cpumasks_hier(struct cpuset * cs,struct cpumask * new_cpus)881 static void update_cpumasks_hier(struct cpuset *cs, struct cpumask *new_cpus)
882 {
883 struct cpuset *cp;
884 struct cgroup_subsys_state *pos_css;
885 bool need_rebuild_sched_domains = false;
886
887 rcu_read_lock();
888 cpuset_for_each_descendant_pre(cp, pos_css, cs) {
889 struct cpuset *parent = parent_cs(cp);
890
891 cpumask_and(new_cpus, cp->cpus_allowed, parent->effective_cpus);
892
893 /*
894 * If it becomes empty, inherit the effective mask of the
895 * parent, which is guaranteed to have some CPUs.
896 */
897 if (cgroup_subsys_on_dfl(cpuset_cgrp_subsys) &&
898 cpumask_empty(new_cpus))
899 cpumask_copy(new_cpus, parent->effective_cpus);
900
901 /* Skip the whole subtree if the cpumask remains the same. */
902 if (cpumask_equal(new_cpus, cp->effective_cpus)) {
903 pos_css = css_rightmost_descendant(pos_css);
904 continue;
905 }
906
907 if (!css_tryget_online(&cp->css))
908 continue;
909 rcu_read_unlock();
910
911 spin_lock_irq(&callback_lock);
912 cpumask_copy(cp->effective_cpus, new_cpus);
913 spin_unlock_irq(&callback_lock);
914
915 WARN_ON(!cgroup_subsys_on_dfl(cpuset_cgrp_subsys) &&
916 !cpumask_equal(cp->cpus_allowed, cp->effective_cpus));
917
918 update_tasks_cpumask(cp);
919
920 /*
921 * If the effective cpumask of any non-empty cpuset is changed,
922 * we need to rebuild sched domains.
923 */
924 if (!cpumask_empty(cp->cpus_allowed) &&
925 is_sched_load_balance(cp))
926 need_rebuild_sched_domains = true;
927
928 rcu_read_lock();
929 css_put(&cp->css);
930 }
931 rcu_read_unlock();
932
933 if (need_rebuild_sched_domains)
934 rebuild_sched_domains_locked();
935 }
936
937 /**
938 * update_cpumask - update the cpus_allowed mask of a cpuset and all tasks in it
939 * @cs: the cpuset to consider
940 * @trialcs: trial cpuset
941 * @buf: buffer of cpu numbers written to this cpuset
942 */
update_cpumask(struct cpuset * cs,struct cpuset * trialcs,const char * buf)943 static int update_cpumask(struct cpuset *cs, struct cpuset *trialcs,
944 const char *buf)
945 {
946 int retval;
947
948 /* top_cpuset.cpus_allowed tracks cpu_online_mask; it's read-only */
949 if (cs == &top_cpuset)
950 return -EACCES;
951
952 /*
953 * An empty cpus_allowed is ok only if the cpuset has no tasks.
954 * Since cpulist_parse() fails on an empty mask, we special case
955 * that parsing. The validate_change() call ensures that cpusets
956 * with tasks have cpus.
957 */
958 if (!*buf) {
959 cpumask_clear(trialcs->cpus_allowed);
960 } else {
961 retval = cpulist_parse(buf, trialcs->cpus_requested);
962 if (retval < 0)
963 return retval;
964
965 if (!cpumask_subset(trialcs->cpus_requested, cpu_present_mask))
966 return -EINVAL;
967
968 cpumask_and(trialcs->cpus_allowed, trialcs->cpus_requested, cpu_active_mask);
969 }
970
971 /* Nothing to do if the cpus didn't change */
972 if (cpumask_equal(cs->cpus_requested, trialcs->cpus_requested))
973 return 0;
974
975 retval = validate_change(cs, trialcs);
976 if (retval < 0)
977 return retval;
978
979 spin_lock_irq(&callback_lock);
980 cpumask_copy(cs->cpus_allowed, trialcs->cpus_allowed);
981 cpumask_copy(cs->cpus_requested, trialcs->cpus_requested);
982 spin_unlock_irq(&callback_lock);
983
984 /* use trialcs->cpus_allowed as a temp variable */
985 update_cpumasks_hier(cs, trialcs->cpus_allowed);
986 return 0;
987 }
988
989 /*
990 * Migrate memory region from one set of nodes to another. This is
991 * performed asynchronously as it can be called from process migration path
992 * holding locks involved in process management. All mm migrations are
993 * performed in the queued order and can be waited for by flushing
994 * cpuset_migrate_mm_wq.
995 */
996
997 struct cpuset_migrate_mm_work {
998 struct work_struct work;
999 struct mm_struct *mm;
1000 nodemask_t from;
1001 nodemask_t to;
1002 };
1003
cpuset_migrate_mm_workfn(struct work_struct * work)1004 static void cpuset_migrate_mm_workfn(struct work_struct *work)
1005 {
1006 struct cpuset_migrate_mm_work *mwork =
1007 container_of(work, struct cpuset_migrate_mm_work, work);
1008
1009 /* on a wq worker, no need to worry about %current's mems_allowed */
1010 do_migrate_pages(mwork->mm, &mwork->from, &mwork->to, MPOL_MF_MOVE_ALL);
1011 mmput(mwork->mm);
1012 kfree(mwork);
1013 }
1014
cpuset_migrate_mm(struct mm_struct * mm,const nodemask_t * from,const nodemask_t * to)1015 static void cpuset_migrate_mm(struct mm_struct *mm, const nodemask_t *from,
1016 const nodemask_t *to)
1017 {
1018 struct cpuset_migrate_mm_work *mwork;
1019
1020 mwork = kzalloc(sizeof(*mwork), GFP_KERNEL);
1021 if (mwork) {
1022 mwork->mm = mm;
1023 mwork->from = *from;
1024 mwork->to = *to;
1025 INIT_WORK(&mwork->work, cpuset_migrate_mm_workfn);
1026 queue_work(cpuset_migrate_mm_wq, &mwork->work);
1027 } else {
1028 mmput(mm);
1029 }
1030 }
1031
cpuset_post_attach(void)1032 static void cpuset_post_attach(void)
1033 {
1034 flush_workqueue(cpuset_migrate_mm_wq);
1035 }
1036
1037 /*
1038 * cpuset_change_task_nodemask - change task's mems_allowed and mempolicy
1039 * @tsk: the task to change
1040 * @newmems: new nodes that the task will be set
1041 *
1042 * In order to avoid seeing no nodes if the old and new nodes are disjoint,
1043 * we structure updates as setting all new allowed nodes, then clearing newly
1044 * disallowed ones.
1045 */
cpuset_change_task_nodemask(struct task_struct * tsk,nodemask_t * newmems)1046 static void cpuset_change_task_nodemask(struct task_struct *tsk,
1047 nodemask_t *newmems)
1048 {
1049 bool need_loop;
1050
1051 /*
1052 * Allow tasks that have access to memory reserves because they have
1053 * been OOM killed to get memory anywhere.
1054 */
1055 if (unlikely(test_thread_flag(TIF_MEMDIE)))
1056 return;
1057 if (current->flags & PF_EXITING) /* Let dying task have memory */
1058 return;
1059
1060 task_lock(tsk);
1061 /*
1062 * Determine if a loop is necessary if another thread is doing
1063 * read_mems_allowed_begin(). If at least one node remains unchanged and
1064 * tsk does not have a mempolicy, then an empty nodemask will not be
1065 * possible when mems_allowed is larger than a word.
1066 */
1067 need_loop = task_has_mempolicy(tsk) ||
1068 !nodes_intersects(*newmems, tsk->mems_allowed);
1069
1070 if (need_loop) {
1071 local_irq_disable();
1072 write_seqcount_begin(&tsk->mems_allowed_seq);
1073 }
1074
1075 nodes_or(tsk->mems_allowed, tsk->mems_allowed, *newmems);
1076 mpol_rebind_task(tsk, newmems, MPOL_REBIND_STEP1);
1077
1078 mpol_rebind_task(tsk, newmems, MPOL_REBIND_STEP2);
1079 tsk->mems_allowed = *newmems;
1080
1081 if (need_loop) {
1082 write_seqcount_end(&tsk->mems_allowed_seq);
1083 local_irq_enable();
1084 }
1085
1086 task_unlock(tsk);
1087 }
1088
1089 static void *cpuset_being_rebound;
1090
1091 /**
1092 * update_tasks_nodemask - Update the nodemasks of tasks in the cpuset.
1093 * @cs: the cpuset in which each task's mems_allowed mask needs to be changed
1094 *
1095 * Iterate through each task of @cs updating its mems_allowed to the
1096 * effective cpuset's. As this function is called with cpuset_mutex held,
1097 * cpuset membership stays stable.
1098 */
update_tasks_nodemask(struct cpuset * cs)1099 static void update_tasks_nodemask(struct cpuset *cs)
1100 {
1101 static nodemask_t newmems; /* protected by cpuset_mutex */
1102 struct css_task_iter it;
1103 struct task_struct *task;
1104
1105 cpuset_being_rebound = cs; /* causes mpol_dup() rebind */
1106
1107 guarantee_online_mems(cs, &newmems);
1108
1109 /*
1110 * The mpol_rebind_mm() call takes mmap_sem, which we couldn't
1111 * take while holding tasklist_lock. Forks can happen - the
1112 * mpol_dup() cpuset_being_rebound check will catch such forks,
1113 * and rebind their vma mempolicies too. Because we still hold
1114 * the global cpuset_mutex, we know that no other rebind effort
1115 * will be contending for the global variable cpuset_being_rebound.
1116 * It's ok if we rebind the same mm twice; mpol_rebind_mm()
1117 * is idempotent. Also migrate pages in each mm to new nodes.
1118 */
1119 css_task_iter_start(&cs->css, &it);
1120 while ((task = css_task_iter_next(&it))) {
1121 struct mm_struct *mm;
1122 bool migrate;
1123
1124 cpuset_change_task_nodemask(task, &newmems);
1125
1126 mm = get_task_mm(task);
1127 if (!mm)
1128 continue;
1129
1130 migrate = is_memory_migrate(cs);
1131
1132 mpol_rebind_mm(mm, &cs->mems_allowed);
1133 if (migrate)
1134 cpuset_migrate_mm(mm, &cs->old_mems_allowed, &newmems);
1135 else
1136 mmput(mm);
1137 }
1138 css_task_iter_end(&it);
1139
1140 /*
1141 * All the tasks' nodemasks have been updated, update
1142 * cs->old_mems_allowed.
1143 */
1144 cs->old_mems_allowed = newmems;
1145
1146 /* We're done rebinding vmas to this cpuset's new mems_allowed. */
1147 cpuset_being_rebound = NULL;
1148 }
1149
1150 /*
1151 * update_nodemasks_hier - Update effective nodemasks and tasks in the subtree
1152 * @cs: the cpuset to consider
1153 * @new_mems: a temp variable for calculating new effective_mems
1154 *
1155 * When configured nodemask is changed, the effective nodemasks of this cpuset
1156 * and all its descendants need to be updated.
1157 *
1158 * On legacy hiearchy, effective_mems will be the same with mems_allowed.
1159 *
1160 * Called with cpuset_mutex held
1161 */
update_nodemasks_hier(struct cpuset * cs,nodemask_t * new_mems)1162 static void update_nodemasks_hier(struct cpuset *cs, nodemask_t *new_mems)
1163 {
1164 struct cpuset *cp;
1165 struct cgroup_subsys_state *pos_css;
1166
1167 rcu_read_lock();
1168 cpuset_for_each_descendant_pre(cp, pos_css, cs) {
1169 struct cpuset *parent = parent_cs(cp);
1170
1171 nodes_and(*new_mems, cp->mems_allowed, parent->effective_mems);
1172
1173 /*
1174 * If it becomes empty, inherit the effective mask of the
1175 * parent, which is guaranteed to have some MEMs.
1176 */
1177 if (cgroup_subsys_on_dfl(cpuset_cgrp_subsys) &&
1178 nodes_empty(*new_mems))
1179 *new_mems = parent->effective_mems;
1180
1181 /* Skip the whole subtree if the nodemask remains the same. */
1182 if (nodes_equal(*new_mems, cp->effective_mems)) {
1183 pos_css = css_rightmost_descendant(pos_css);
1184 continue;
1185 }
1186
1187 if (!css_tryget_online(&cp->css))
1188 continue;
1189 rcu_read_unlock();
1190
1191 spin_lock_irq(&callback_lock);
1192 cp->effective_mems = *new_mems;
1193 spin_unlock_irq(&callback_lock);
1194
1195 WARN_ON(!cgroup_subsys_on_dfl(cpuset_cgrp_subsys) &&
1196 !nodes_equal(cp->mems_allowed, cp->effective_mems));
1197
1198 update_tasks_nodemask(cp);
1199
1200 rcu_read_lock();
1201 css_put(&cp->css);
1202 }
1203 rcu_read_unlock();
1204 }
1205
1206 /*
1207 * Handle user request to change the 'mems' memory placement
1208 * of a cpuset. Needs to validate the request, update the
1209 * cpusets mems_allowed, and for each task in the cpuset,
1210 * update mems_allowed and rebind task's mempolicy and any vma
1211 * mempolicies and if the cpuset is marked 'memory_migrate',
1212 * migrate the tasks pages to the new memory.
1213 *
1214 * Call with cpuset_mutex held. May take callback_lock during call.
1215 * Will take tasklist_lock, scan tasklist for tasks in cpuset cs,
1216 * lock each such tasks mm->mmap_sem, scan its vma's and rebind
1217 * their mempolicies to the cpusets new mems_allowed.
1218 */
update_nodemask(struct cpuset * cs,struct cpuset * trialcs,const char * buf)1219 static int update_nodemask(struct cpuset *cs, struct cpuset *trialcs,
1220 const char *buf)
1221 {
1222 int retval;
1223
1224 /*
1225 * top_cpuset.mems_allowed tracks node_stats[N_MEMORY];
1226 * it's read-only
1227 */
1228 if (cs == &top_cpuset) {
1229 retval = -EACCES;
1230 goto done;
1231 }
1232
1233 /*
1234 * An empty mems_allowed is ok iff there are no tasks in the cpuset.
1235 * Since nodelist_parse() fails on an empty mask, we special case
1236 * that parsing. The validate_change() call ensures that cpusets
1237 * with tasks have memory.
1238 */
1239 if (!*buf) {
1240 nodes_clear(trialcs->mems_allowed);
1241 } else {
1242 retval = nodelist_parse(buf, trialcs->mems_allowed);
1243 if (retval < 0)
1244 goto done;
1245
1246 if (!nodes_subset(trialcs->mems_allowed,
1247 top_cpuset.mems_allowed)) {
1248 retval = -EINVAL;
1249 goto done;
1250 }
1251 }
1252
1253 if (nodes_equal(cs->mems_allowed, trialcs->mems_allowed)) {
1254 retval = 0; /* Too easy - nothing to do */
1255 goto done;
1256 }
1257 retval = validate_change(cs, trialcs);
1258 if (retval < 0)
1259 goto done;
1260
1261 spin_lock_irq(&callback_lock);
1262 cs->mems_allowed = trialcs->mems_allowed;
1263 spin_unlock_irq(&callback_lock);
1264
1265 /* use trialcs->mems_allowed as a temp variable */
1266 update_nodemasks_hier(cs, &trialcs->mems_allowed);
1267 done:
1268 return retval;
1269 }
1270
current_cpuset_is_being_rebound(void)1271 int current_cpuset_is_being_rebound(void)
1272 {
1273 int ret;
1274
1275 rcu_read_lock();
1276 ret = task_cs(current) == cpuset_being_rebound;
1277 rcu_read_unlock();
1278
1279 return ret;
1280 }
1281
update_relax_domain_level(struct cpuset * cs,s64 val)1282 static int update_relax_domain_level(struct cpuset *cs, s64 val)
1283 {
1284 #ifdef CONFIG_SMP
1285 if (val < -1 || val >= sched_domain_level_max)
1286 return -EINVAL;
1287 #endif
1288
1289 if (val != cs->relax_domain_level) {
1290 cs->relax_domain_level = val;
1291 if (!cpumask_empty(cs->cpus_allowed) &&
1292 is_sched_load_balance(cs))
1293 rebuild_sched_domains_locked();
1294 }
1295
1296 return 0;
1297 }
1298
1299 /**
1300 * update_tasks_flags - update the spread flags of tasks in the cpuset.
1301 * @cs: the cpuset in which each task's spread flags needs to be changed
1302 *
1303 * Iterate through each task of @cs updating its spread flags. As this
1304 * function is called with cpuset_mutex held, cpuset membership stays
1305 * stable.
1306 */
update_tasks_flags(struct cpuset * cs)1307 static void update_tasks_flags(struct cpuset *cs)
1308 {
1309 struct css_task_iter it;
1310 struct task_struct *task;
1311
1312 css_task_iter_start(&cs->css, &it);
1313 while ((task = css_task_iter_next(&it)))
1314 cpuset_update_task_spread_flag(cs, task);
1315 css_task_iter_end(&it);
1316 }
1317
1318 /*
1319 * update_flag - read a 0 or a 1 in a file and update associated flag
1320 * bit: the bit to update (see cpuset_flagbits_t)
1321 * cs: the cpuset to update
1322 * turning_on: whether the flag is being set or cleared
1323 *
1324 * Call with cpuset_mutex held.
1325 */
1326
update_flag(cpuset_flagbits_t bit,struct cpuset * cs,int turning_on)1327 static int update_flag(cpuset_flagbits_t bit, struct cpuset *cs,
1328 int turning_on)
1329 {
1330 struct cpuset *trialcs;
1331 int balance_flag_changed;
1332 int spread_flag_changed;
1333 int err;
1334
1335 trialcs = alloc_trial_cpuset(cs);
1336 if (!trialcs)
1337 return -ENOMEM;
1338
1339 if (turning_on)
1340 set_bit(bit, &trialcs->flags);
1341 else
1342 clear_bit(bit, &trialcs->flags);
1343
1344 err = validate_change(cs, trialcs);
1345 if (err < 0)
1346 goto out;
1347
1348 balance_flag_changed = (is_sched_load_balance(cs) !=
1349 is_sched_load_balance(trialcs));
1350
1351 spread_flag_changed = ((is_spread_slab(cs) != is_spread_slab(trialcs))
1352 || (is_spread_page(cs) != is_spread_page(trialcs)));
1353
1354 spin_lock_irq(&callback_lock);
1355 cs->flags = trialcs->flags;
1356 spin_unlock_irq(&callback_lock);
1357
1358 if (!cpumask_empty(trialcs->cpus_allowed) && balance_flag_changed)
1359 rebuild_sched_domains_locked();
1360
1361 if (spread_flag_changed)
1362 update_tasks_flags(cs);
1363 out:
1364 free_trial_cpuset(trialcs);
1365 return err;
1366 }
1367
1368 /*
1369 * Frequency meter - How fast is some event occurring?
1370 *
1371 * These routines manage a digitally filtered, constant time based,
1372 * event frequency meter. There are four routines:
1373 * fmeter_init() - initialize a frequency meter.
1374 * fmeter_markevent() - called each time the event happens.
1375 * fmeter_getrate() - returns the recent rate of such events.
1376 * fmeter_update() - internal routine used to update fmeter.
1377 *
1378 * A common data structure is passed to each of these routines,
1379 * which is used to keep track of the state required to manage the
1380 * frequency meter and its digital filter.
1381 *
1382 * The filter works on the number of events marked per unit time.
1383 * The filter is single-pole low-pass recursive (IIR). The time unit
1384 * is 1 second. Arithmetic is done using 32-bit integers scaled to
1385 * simulate 3 decimal digits of precision (multiplied by 1000).
1386 *
1387 * With an FM_COEF of 933, and a time base of 1 second, the filter
1388 * has a half-life of 10 seconds, meaning that if the events quit
1389 * happening, then the rate returned from the fmeter_getrate()
1390 * will be cut in half each 10 seconds, until it converges to zero.
1391 *
1392 * It is not worth doing a real infinitely recursive filter. If more
1393 * than FM_MAXTICKS ticks have elapsed since the last filter event,
1394 * just compute FM_MAXTICKS ticks worth, by which point the level
1395 * will be stable.
1396 *
1397 * Limit the count of unprocessed events to FM_MAXCNT, so as to avoid
1398 * arithmetic overflow in the fmeter_update() routine.
1399 *
1400 * Given the simple 32 bit integer arithmetic used, this meter works
1401 * best for reporting rates between one per millisecond (msec) and
1402 * one per 32 (approx) seconds. At constant rates faster than one
1403 * per msec it maxes out at values just under 1,000,000. At constant
1404 * rates between one per msec, and one per second it will stabilize
1405 * to a value N*1000, where N is the rate of events per second.
1406 * At constant rates between one per second and one per 32 seconds,
1407 * it will be choppy, moving up on the seconds that have an event,
1408 * and then decaying until the next event. At rates slower than
1409 * about one in 32 seconds, it decays all the way back to zero between
1410 * each event.
1411 */
1412
1413 #define FM_COEF 933 /* coefficient for half-life of 10 secs */
1414 #define FM_MAXTICKS ((time_t)99) /* useless computing more ticks than this */
1415 #define FM_MAXCNT 1000000 /* limit cnt to avoid overflow */
1416 #define FM_SCALE 1000 /* faux fixed point scale */
1417
1418 /* Initialize a frequency meter */
fmeter_init(struct fmeter * fmp)1419 static void fmeter_init(struct fmeter *fmp)
1420 {
1421 fmp->cnt = 0;
1422 fmp->val = 0;
1423 fmp->time = 0;
1424 spin_lock_init(&fmp->lock);
1425 }
1426
1427 /* Internal meter update - process cnt events and update value */
fmeter_update(struct fmeter * fmp)1428 static void fmeter_update(struct fmeter *fmp)
1429 {
1430 time_t now = get_seconds();
1431 time_t ticks = now - fmp->time;
1432
1433 if (ticks == 0)
1434 return;
1435
1436 ticks = min(FM_MAXTICKS, ticks);
1437 while (ticks-- > 0)
1438 fmp->val = (FM_COEF * fmp->val) / FM_SCALE;
1439 fmp->time = now;
1440
1441 fmp->val += ((FM_SCALE - FM_COEF) * fmp->cnt) / FM_SCALE;
1442 fmp->cnt = 0;
1443 }
1444
1445 /* Process any previous ticks, then bump cnt by one (times scale). */
fmeter_markevent(struct fmeter * fmp)1446 static void fmeter_markevent(struct fmeter *fmp)
1447 {
1448 spin_lock(&fmp->lock);
1449 fmeter_update(fmp);
1450 fmp->cnt = min(FM_MAXCNT, fmp->cnt + FM_SCALE);
1451 spin_unlock(&fmp->lock);
1452 }
1453
1454 /* Process any previous ticks, then return current value. */
fmeter_getrate(struct fmeter * fmp)1455 static int fmeter_getrate(struct fmeter *fmp)
1456 {
1457 int val;
1458
1459 spin_lock(&fmp->lock);
1460 fmeter_update(fmp);
1461 val = fmp->val;
1462 spin_unlock(&fmp->lock);
1463 return val;
1464 }
1465
1466 static struct cpuset *cpuset_attach_old_cs;
1467
1468 /* Called by cgroups to determine if a cpuset is usable; cpuset_mutex held */
cpuset_can_attach(struct cgroup_taskset * tset)1469 static int cpuset_can_attach(struct cgroup_taskset *tset)
1470 {
1471 struct cgroup_subsys_state *css;
1472 struct cpuset *cs;
1473 struct task_struct *task;
1474 int ret;
1475
1476 /* used later by cpuset_attach() */
1477 cpuset_attach_old_cs = task_cs(cgroup_taskset_first(tset, &css));
1478 cs = css_cs(css);
1479
1480 mutex_lock(&cpuset_mutex);
1481
1482 /* allow moving tasks into an empty cpuset if on default hierarchy */
1483 ret = -ENOSPC;
1484 if (!cgroup_subsys_on_dfl(cpuset_cgrp_subsys) &&
1485 (cpumask_empty(cs->cpus_allowed) || nodes_empty(cs->mems_allowed)))
1486 goto out_unlock;
1487
1488 cgroup_taskset_for_each(task, css, tset) {
1489 ret = task_can_attach(task, cs->cpus_allowed);
1490 if (ret)
1491 goto out_unlock;
1492 ret = security_task_setscheduler(task);
1493 if (ret)
1494 goto out_unlock;
1495 }
1496
1497 /*
1498 * Mark attach is in progress. This makes validate_change() fail
1499 * changes which zero cpus/mems_allowed.
1500 */
1501 cs->attach_in_progress++;
1502 ret = 0;
1503 out_unlock:
1504 mutex_unlock(&cpuset_mutex);
1505 return ret;
1506 }
1507
cpuset_cancel_attach(struct cgroup_taskset * tset)1508 static void cpuset_cancel_attach(struct cgroup_taskset *tset)
1509 {
1510 struct cgroup_subsys_state *css;
1511 struct cpuset *cs;
1512
1513 cgroup_taskset_first(tset, &css);
1514 cs = css_cs(css);
1515
1516 mutex_lock(&cpuset_mutex);
1517 css_cs(css)->attach_in_progress--;
1518 mutex_unlock(&cpuset_mutex);
1519 }
1520
1521 /*
1522 * Protected by cpuset_mutex. cpus_attach is used only by cpuset_attach()
1523 * but we can't allocate it dynamically there. Define it global and
1524 * allocate from cpuset_init().
1525 */
1526 static cpumask_var_t cpus_attach;
1527
cpuset_attach(struct cgroup_taskset * tset)1528 static void cpuset_attach(struct cgroup_taskset *tset)
1529 {
1530 /* static buf protected by cpuset_mutex */
1531 static nodemask_t cpuset_attach_nodemask_to;
1532 struct task_struct *task;
1533 struct task_struct *leader;
1534 struct cgroup_subsys_state *css;
1535 struct cpuset *cs;
1536 struct cpuset *oldcs = cpuset_attach_old_cs;
1537
1538 cgroup_taskset_first(tset, &css);
1539 cs = css_cs(css);
1540
1541 mutex_lock(&cpuset_mutex);
1542
1543 /* prepare for attach */
1544 if (cs == &top_cpuset)
1545 cpumask_copy(cpus_attach, cpu_possible_mask);
1546 else
1547 guarantee_online_cpus(cs, cpus_attach);
1548
1549 guarantee_online_mems(cs, &cpuset_attach_nodemask_to);
1550
1551 cgroup_taskset_for_each(task, css, tset) {
1552 /*
1553 * can_attach beforehand should guarantee that this doesn't
1554 * fail. TODO: have a better way to handle failure here
1555 */
1556 WARN_ON_ONCE(set_cpus_allowed_ptr(task, cpus_attach));
1557
1558 cpuset_change_task_nodemask(task, &cpuset_attach_nodemask_to);
1559 cpuset_update_task_spread_flag(cs, task);
1560 }
1561
1562 /*
1563 * Change mm for all threadgroup leaders. This is expensive and may
1564 * sleep and should be moved outside migration path proper.
1565 */
1566 cpuset_attach_nodemask_to = cs->effective_mems;
1567 cgroup_taskset_for_each_leader(leader, css, tset) {
1568 struct mm_struct *mm = get_task_mm(leader);
1569
1570 if (mm) {
1571 mpol_rebind_mm(mm, &cpuset_attach_nodemask_to);
1572
1573 /*
1574 * old_mems_allowed is the same with mems_allowed
1575 * here, except if this task is being moved
1576 * automatically due to hotplug. In that case
1577 * @mems_allowed has been updated and is empty, so
1578 * @old_mems_allowed is the right nodesets that we
1579 * migrate mm from.
1580 */
1581 if (is_memory_migrate(cs))
1582 cpuset_migrate_mm(mm, &oldcs->old_mems_allowed,
1583 &cpuset_attach_nodemask_to);
1584 else
1585 mmput(mm);
1586 }
1587 }
1588
1589 cs->old_mems_allowed = cpuset_attach_nodemask_to;
1590
1591 cs->attach_in_progress--;
1592 if (!cs->attach_in_progress)
1593 wake_up(&cpuset_attach_wq);
1594
1595 mutex_unlock(&cpuset_mutex);
1596 }
1597
1598 /* The various types of files and directories in a cpuset file system */
1599
1600 typedef enum {
1601 FILE_MEMORY_MIGRATE,
1602 FILE_CPULIST,
1603 FILE_MEMLIST,
1604 FILE_EFFECTIVE_CPULIST,
1605 FILE_EFFECTIVE_MEMLIST,
1606 FILE_CPU_EXCLUSIVE,
1607 FILE_MEM_EXCLUSIVE,
1608 FILE_MEM_HARDWALL,
1609 FILE_SCHED_LOAD_BALANCE,
1610 FILE_SCHED_RELAX_DOMAIN_LEVEL,
1611 FILE_MEMORY_PRESSURE_ENABLED,
1612 FILE_MEMORY_PRESSURE,
1613 FILE_SPREAD_PAGE,
1614 FILE_SPREAD_SLAB,
1615 } cpuset_filetype_t;
1616
cpuset_write_u64(struct cgroup_subsys_state * css,struct cftype * cft,u64 val)1617 static int cpuset_write_u64(struct cgroup_subsys_state *css, struct cftype *cft,
1618 u64 val)
1619 {
1620 struct cpuset *cs = css_cs(css);
1621 cpuset_filetype_t type = cft->private;
1622 int retval = 0;
1623
1624 mutex_lock(&cpuset_mutex);
1625 if (!is_cpuset_online(cs)) {
1626 retval = -ENODEV;
1627 goto out_unlock;
1628 }
1629
1630 switch (type) {
1631 case FILE_CPU_EXCLUSIVE:
1632 retval = update_flag(CS_CPU_EXCLUSIVE, cs, val);
1633 break;
1634 case FILE_MEM_EXCLUSIVE:
1635 retval = update_flag(CS_MEM_EXCLUSIVE, cs, val);
1636 break;
1637 case FILE_MEM_HARDWALL:
1638 retval = update_flag(CS_MEM_HARDWALL, cs, val);
1639 break;
1640 case FILE_SCHED_LOAD_BALANCE:
1641 retval = update_flag(CS_SCHED_LOAD_BALANCE, cs, val);
1642 break;
1643 case FILE_MEMORY_MIGRATE:
1644 retval = update_flag(CS_MEMORY_MIGRATE, cs, val);
1645 break;
1646 case FILE_MEMORY_PRESSURE_ENABLED:
1647 cpuset_memory_pressure_enabled = !!val;
1648 break;
1649 case FILE_SPREAD_PAGE:
1650 retval = update_flag(CS_SPREAD_PAGE, cs, val);
1651 break;
1652 case FILE_SPREAD_SLAB:
1653 retval = update_flag(CS_SPREAD_SLAB, cs, val);
1654 break;
1655 default:
1656 retval = -EINVAL;
1657 break;
1658 }
1659 out_unlock:
1660 mutex_unlock(&cpuset_mutex);
1661 return retval;
1662 }
1663
cpuset_write_s64(struct cgroup_subsys_state * css,struct cftype * cft,s64 val)1664 static int cpuset_write_s64(struct cgroup_subsys_state *css, struct cftype *cft,
1665 s64 val)
1666 {
1667 struct cpuset *cs = css_cs(css);
1668 cpuset_filetype_t type = cft->private;
1669 int retval = -ENODEV;
1670
1671 mutex_lock(&cpuset_mutex);
1672 if (!is_cpuset_online(cs))
1673 goto out_unlock;
1674
1675 switch (type) {
1676 case FILE_SCHED_RELAX_DOMAIN_LEVEL:
1677 retval = update_relax_domain_level(cs, val);
1678 break;
1679 default:
1680 retval = -EINVAL;
1681 break;
1682 }
1683 out_unlock:
1684 mutex_unlock(&cpuset_mutex);
1685 return retval;
1686 }
1687
1688 /*
1689 * Common handling for a write to a "cpus" or "mems" file.
1690 */
cpuset_write_resmask(struct kernfs_open_file * of,char * buf,size_t nbytes,loff_t off)1691 static ssize_t cpuset_write_resmask(struct kernfs_open_file *of,
1692 char *buf, size_t nbytes, loff_t off)
1693 {
1694 struct cpuset *cs = css_cs(of_css(of));
1695 struct cpuset *trialcs;
1696 int retval = -ENODEV;
1697
1698 buf = strstrip(buf);
1699
1700 /*
1701 * CPU or memory hotunplug may leave @cs w/o any execution
1702 * resources, in which case the hotplug code asynchronously updates
1703 * configuration and transfers all tasks to the nearest ancestor
1704 * which can execute.
1705 *
1706 * As writes to "cpus" or "mems" may restore @cs's execution
1707 * resources, wait for the previously scheduled operations before
1708 * proceeding, so that we don't end up keep removing tasks added
1709 * after execution capability is restored.
1710 *
1711 * cpuset_hotplug_work calls back into cgroup core via
1712 * cgroup_transfer_tasks() and waiting for it from a cgroupfs
1713 * operation like this one can lead to a deadlock through kernfs
1714 * active_ref protection. Let's break the protection. Losing the
1715 * protection is okay as we check whether @cs is online after
1716 * grabbing cpuset_mutex anyway. This only happens on the legacy
1717 * hierarchies.
1718 */
1719 css_get(&cs->css);
1720 kernfs_break_active_protection(of->kn);
1721 flush_work(&cpuset_hotplug_work);
1722
1723 mutex_lock(&cpuset_mutex);
1724 if (!is_cpuset_online(cs))
1725 goto out_unlock;
1726
1727 trialcs = alloc_trial_cpuset(cs);
1728 if (!trialcs) {
1729 retval = -ENOMEM;
1730 goto out_unlock;
1731 }
1732
1733 switch (of_cft(of)->private) {
1734 case FILE_CPULIST:
1735 retval = update_cpumask(cs, trialcs, buf);
1736 break;
1737 case FILE_MEMLIST:
1738 retval = update_nodemask(cs, trialcs, buf);
1739 break;
1740 default:
1741 retval = -EINVAL;
1742 break;
1743 }
1744
1745 free_trial_cpuset(trialcs);
1746 out_unlock:
1747 mutex_unlock(&cpuset_mutex);
1748 kernfs_unbreak_active_protection(of->kn);
1749 css_put(&cs->css);
1750 flush_workqueue(cpuset_migrate_mm_wq);
1751 return retval ?: nbytes;
1752 }
1753
1754 /*
1755 * These ascii lists should be read in a single call, by using a user
1756 * buffer large enough to hold the entire map. If read in smaller
1757 * chunks, there is no guarantee of atomicity. Since the display format
1758 * used, list of ranges of sequential numbers, is variable length,
1759 * and since these maps can change value dynamically, one could read
1760 * gibberish by doing partial reads while a list was changing.
1761 */
cpuset_common_seq_show(struct seq_file * sf,void * v)1762 static int cpuset_common_seq_show(struct seq_file *sf, void *v)
1763 {
1764 struct cpuset *cs = css_cs(seq_css(sf));
1765 cpuset_filetype_t type = seq_cft(sf)->private;
1766 int ret = 0;
1767
1768 spin_lock_irq(&callback_lock);
1769
1770 switch (type) {
1771 case FILE_CPULIST:
1772 seq_printf(sf, "%*pbl\n", cpumask_pr_args(cs->cpus_requested));
1773 break;
1774 case FILE_MEMLIST:
1775 seq_printf(sf, "%*pbl\n", nodemask_pr_args(&cs->mems_allowed));
1776 break;
1777 case FILE_EFFECTIVE_CPULIST:
1778 seq_printf(sf, "%*pbl\n", cpumask_pr_args(cs->effective_cpus));
1779 break;
1780 case FILE_EFFECTIVE_MEMLIST:
1781 seq_printf(sf, "%*pbl\n", nodemask_pr_args(&cs->effective_mems));
1782 break;
1783 default:
1784 ret = -EINVAL;
1785 }
1786
1787 spin_unlock_irq(&callback_lock);
1788 return ret;
1789 }
1790
cpuset_read_u64(struct cgroup_subsys_state * css,struct cftype * cft)1791 static u64 cpuset_read_u64(struct cgroup_subsys_state *css, struct cftype *cft)
1792 {
1793 struct cpuset *cs = css_cs(css);
1794 cpuset_filetype_t type = cft->private;
1795 switch (type) {
1796 case FILE_CPU_EXCLUSIVE:
1797 return is_cpu_exclusive(cs);
1798 case FILE_MEM_EXCLUSIVE:
1799 return is_mem_exclusive(cs);
1800 case FILE_MEM_HARDWALL:
1801 return is_mem_hardwall(cs);
1802 case FILE_SCHED_LOAD_BALANCE:
1803 return is_sched_load_balance(cs);
1804 case FILE_MEMORY_MIGRATE:
1805 return is_memory_migrate(cs);
1806 case FILE_MEMORY_PRESSURE_ENABLED:
1807 return cpuset_memory_pressure_enabled;
1808 case FILE_MEMORY_PRESSURE:
1809 return fmeter_getrate(&cs->fmeter);
1810 case FILE_SPREAD_PAGE:
1811 return is_spread_page(cs);
1812 case FILE_SPREAD_SLAB:
1813 return is_spread_slab(cs);
1814 default:
1815 BUG();
1816 }
1817
1818 /* Unreachable but makes gcc happy */
1819 return 0;
1820 }
1821
cpuset_read_s64(struct cgroup_subsys_state * css,struct cftype * cft)1822 static s64 cpuset_read_s64(struct cgroup_subsys_state *css, struct cftype *cft)
1823 {
1824 struct cpuset *cs = css_cs(css);
1825 cpuset_filetype_t type = cft->private;
1826 switch (type) {
1827 case FILE_SCHED_RELAX_DOMAIN_LEVEL:
1828 return cs->relax_domain_level;
1829 default:
1830 BUG();
1831 }
1832
1833 /* Unrechable but makes gcc happy */
1834 return 0;
1835 }
1836
1837
1838 /*
1839 * for the common functions, 'private' gives the type of file
1840 */
1841
1842 static struct cftype files[] = {
1843 {
1844 .name = "cpus",
1845 .seq_show = cpuset_common_seq_show,
1846 .write = cpuset_write_resmask,
1847 .max_write_len = (100U + 6 * NR_CPUS),
1848 .private = FILE_CPULIST,
1849 },
1850
1851 {
1852 .name = "mems",
1853 .seq_show = cpuset_common_seq_show,
1854 .write = cpuset_write_resmask,
1855 .max_write_len = (100U + 6 * MAX_NUMNODES),
1856 .private = FILE_MEMLIST,
1857 },
1858
1859 {
1860 .name = "effective_cpus",
1861 .seq_show = cpuset_common_seq_show,
1862 .private = FILE_EFFECTIVE_CPULIST,
1863 },
1864
1865 {
1866 .name = "effective_mems",
1867 .seq_show = cpuset_common_seq_show,
1868 .private = FILE_EFFECTIVE_MEMLIST,
1869 },
1870
1871 {
1872 .name = "cpu_exclusive",
1873 .read_u64 = cpuset_read_u64,
1874 .write_u64 = cpuset_write_u64,
1875 .private = FILE_CPU_EXCLUSIVE,
1876 },
1877
1878 {
1879 .name = "mem_exclusive",
1880 .read_u64 = cpuset_read_u64,
1881 .write_u64 = cpuset_write_u64,
1882 .private = FILE_MEM_EXCLUSIVE,
1883 },
1884
1885 {
1886 .name = "mem_hardwall",
1887 .read_u64 = cpuset_read_u64,
1888 .write_u64 = cpuset_write_u64,
1889 .private = FILE_MEM_HARDWALL,
1890 },
1891
1892 {
1893 .name = "sched_load_balance",
1894 .read_u64 = cpuset_read_u64,
1895 .write_u64 = cpuset_write_u64,
1896 .private = FILE_SCHED_LOAD_BALANCE,
1897 },
1898
1899 {
1900 .name = "sched_relax_domain_level",
1901 .read_s64 = cpuset_read_s64,
1902 .write_s64 = cpuset_write_s64,
1903 .private = FILE_SCHED_RELAX_DOMAIN_LEVEL,
1904 },
1905
1906 {
1907 .name = "memory_migrate",
1908 .read_u64 = cpuset_read_u64,
1909 .write_u64 = cpuset_write_u64,
1910 .private = FILE_MEMORY_MIGRATE,
1911 },
1912
1913 {
1914 .name = "memory_pressure",
1915 .read_u64 = cpuset_read_u64,
1916 .private = FILE_MEMORY_PRESSURE,
1917 },
1918
1919 {
1920 .name = "memory_spread_page",
1921 .read_u64 = cpuset_read_u64,
1922 .write_u64 = cpuset_write_u64,
1923 .private = FILE_SPREAD_PAGE,
1924 },
1925
1926 {
1927 .name = "memory_spread_slab",
1928 .read_u64 = cpuset_read_u64,
1929 .write_u64 = cpuset_write_u64,
1930 .private = FILE_SPREAD_SLAB,
1931 },
1932
1933 {
1934 .name = "memory_pressure_enabled",
1935 .flags = CFTYPE_ONLY_ON_ROOT,
1936 .read_u64 = cpuset_read_u64,
1937 .write_u64 = cpuset_write_u64,
1938 .private = FILE_MEMORY_PRESSURE_ENABLED,
1939 },
1940
1941 { } /* terminate */
1942 };
1943
1944 /*
1945 * cpuset_css_alloc - allocate a cpuset css
1946 * cgrp: control group that the new cpuset will be part of
1947 */
1948
1949 static struct cgroup_subsys_state *
cpuset_css_alloc(struct cgroup_subsys_state * parent_css)1950 cpuset_css_alloc(struct cgroup_subsys_state *parent_css)
1951 {
1952 struct cpuset *cs;
1953
1954 if (!parent_css)
1955 return &top_cpuset.css;
1956
1957 cs = kzalloc(sizeof(*cs), GFP_KERNEL);
1958 if (!cs)
1959 return ERR_PTR(-ENOMEM);
1960 if (!alloc_cpumask_var(&cs->cpus_allowed, GFP_KERNEL))
1961 goto free_cs;
1962 if (!alloc_cpumask_var(&cs->cpus_requested, GFP_KERNEL))
1963 goto free_allowed;
1964 if (!alloc_cpumask_var(&cs->effective_cpus, GFP_KERNEL))
1965 goto free_requested;
1966
1967 set_bit(CS_SCHED_LOAD_BALANCE, &cs->flags);
1968 cpumask_clear(cs->cpus_allowed);
1969 cpumask_clear(cs->cpus_requested);
1970 nodes_clear(cs->mems_allowed);
1971 cpumask_clear(cs->effective_cpus);
1972 nodes_clear(cs->effective_mems);
1973 fmeter_init(&cs->fmeter);
1974 cs->relax_domain_level = -1;
1975
1976 return &cs->css;
1977
1978 free_requested:
1979 free_cpumask_var(cs->cpus_requested);
1980 free_allowed:
1981 free_cpumask_var(cs->cpus_allowed);
1982 free_cs:
1983 kfree(cs);
1984 return ERR_PTR(-ENOMEM);
1985 }
1986
cpuset_css_online(struct cgroup_subsys_state * css)1987 static int cpuset_css_online(struct cgroup_subsys_state *css)
1988 {
1989 struct cpuset *cs = css_cs(css);
1990 struct cpuset *parent = parent_cs(cs);
1991 struct cpuset *tmp_cs;
1992 struct cgroup_subsys_state *pos_css;
1993
1994 if (!parent)
1995 return 0;
1996
1997 mutex_lock(&cpuset_mutex);
1998
1999 set_bit(CS_ONLINE, &cs->flags);
2000 if (is_spread_page(parent))
2001 set_bit(CS_SPREAD_PAGE, &cs->flags);
2002 if (is_spread_slab(parent))
2003 set_bit(CS_SPREAD_SLAB, &cs->flags);
2004
2005 cpuset_inc();
2006
2007 spin_lock_irq(&callback_lock);
2008 if (cgroup_subsys_on_dfl(cpuset_cgrp_subsys)) {
2009 cpumask_copy(cs->effective_cpus, parent->effective_cpus);
2010 cs->effective_mems = parent->effective_mems;
2011 }
2012 spin_unlock_irq(&callback_lock);
2013
2014 if (!test_bit(CGRP_CPUSET_CLONE_CHILDREN, &css->cgroup->flags))
2015 goto out_unlock;
2016
2017 /*
2018 * Clone @parent's configuration if CGRP_CPUSET_CLONE_CHILDREN is
2019 * set. This flag handling is implemented in cgroup core for
2020 * histrical reasons - the flag may be specified during mount.
2021 *
2022 * Currently, if any sibling cpusets have exclusive cpus or mem, we
2023 * refuse to clone the configuration - thereby refusing the task to
2024 * be entered, and as a result refusing the sys_unshare() or
2025 * clone() which initiated it. If this becomes a problem for some
2026 * users who wish to allow that scenario, then this could be
2027 * changed to grant parent->cpus_allowed-sibling_cpus_exclusive
2028 * (and likewise for mems) to the new cgroup.
2029 */
2030 rcu_read_lock();
2031 cpuset_for_each_child(tmp_cs, pos_css, parent) {
2032 if (is_mem_exclusive(tmp_cs) || is_cpu_exclusive(tmp_cs)) {
2033 rcu_read_unlock();
2034 goto out_unlock;
2035 }
2036 }
2037 rcu_read_unlock();
2038
2039 spin_lock_irq(&callback_lock);
2040 cs->mems_allowed = parent->mems_allowed;
2041 cs->effective_mems = parent->mems_allowed;
2042 cpumask_copy(cs->cpus_allowed, parent->cpus_allowed);
2043 cpumask_copy(cs->cpus_requested, parent->cpus_requested);
2044 cpumask_copy(cs->effective_cpus, parent->cpus_allowed);
2045 spin_unlock_irq(&callback_lock);
2046 out_unlock:
2047 mutex_unlock(&cpuset_mutex);
2048 return 0;
2049 }
2050
2051 /*
2052 * If the cpuset being removed has its flag 'sched_load_balance'
2053 * enabled, then simulate turning sched_load_balance off, which
2054 * will call rebuild_sched_domains_locked().
2055 */
2056
cpuset_css_offline(struct cgroup_subsys_state * css)2057 static void cpuset_css_offline(struct cgroup_subsys_state *css)
2058 {
2059 struct cpuset *cs = css_cs(css);
2060
2061 mutex_lock(&cpuset_mutex);
2062
2063 if (is_sched_load_balance(cs))
2064 update_flag(CS_SCHED_LOAD_BALANCE, cs, 0);
2065
2066 cpuset_dec();
2067 clear_bit(CS_ONLINE, &cs->flags);
2068
2069 mutex_unlock(&cpuset_mutex);
2070 }
2071
cpuset_css_free(struct cgroup_subsys_state * css)2072 static void cpuset_css_free(struct cgroup_subsys_state *css)
2073 {
2074 struct cpuset *cs = css_cs(css);
2075
2076 free_cpumask_var(cs->effective_cpus);
2077 free_cpumask_var(cs->cpus_allowed);
2078 free_cpumask_var(cs->cpus_requested);
2079 kfree(cs);
2080 }
2081
cpuset_bind(struct cgroup_subsys_state * root_css)2082 static void cpuset_bind(struct cgroup_subsys_state *root_css)
2083 {
2084 mutex_lock(&cpuset_mutex);
2085 spin_lock_irq(&callback_lock);
2086
2087 if (cgroup_subsys_on_dfl(cpuset_cgrp_subsys)) {
2088 cpumask_copy(top_cpuset.cpus_allowed, cpu_possible_mask);
2089 top_cpuset.mems_allowed = node_possible_map;
2090 } else {
2091 cpumask_copy(top_cpuset.cpus_allowed,
2092 top_cpuset.effective_cpus);
2093 top_cpuset.mems_allowed = top_cpuset.effective_mems;
2094 }
2095
2096 spin_unlock_irq(&callback_lock);
2097 mutex_unlock(&cpuset_mutex);
2098 }
2099
2100 /*
2101 * Make sure the new task conform to the current state of its parent,
2102 * which could have been changed by cpuset just after it inherits the
2103 * state from the parent and before it sits on the cgroup's task list.
2104 */
cpuset_fork(struct task_struct * task,void * priv)2105 void cpuset_fork(struct task_struct *task, void *priv)
2106 {
2107 if (task_css_is_root(task, cpuset_cgrp_id))
2108 return;
2109
2110 set_cpus_allowed_ptr(task, ¤t->cpus_allowed);
2111 task->mems_allowed = current->mems_allowed;
2112 }
2113
2114 struct cgroup_subsys cpuset_cgrp_subsys = {
2115 .css_alloc = cpuset_css_alloc,
2116 .css_online = cpuset_css_online,
2117 .css_offline = cpuset_css_offline,
2118 .css_free = cpuset_css_free,
2119 .can_attach = cpuset_can_attach,
2120 .cancel_attach = cpuset_cancel_attach,
2121 .attach = cpuset_attach,
2122 .post_attach = cpuset_post_attach,
2123 .bind = cpuset_bind,
2124 .fork = cpuset_fork,
2125 .legacy_cftypes = files,
2126 .early_init = 1,
2127 };
2128
2129 /**
2130 * cpuset_init - initialize cpusets at system boot
2131 *
2132 * Description: Initialize top_cpuset and the cpuset internal file system,
2133 **/
2134
cpuset_init(void)2135 int __init cpuset_init(void)
2136 {
2137 int err = 0;
2138
2139 if (!alloc_cpumask_var(&top_cpuset.cpus_allowed, GFP_KERNEL))
2140 BUG();
2141 if (!alloc_cpumask_var(&top_cpuset.effective_cpus, GFP_KERNEL))
2142 BUG();
2143 if (!alloc_cpumask_var(&top_cpuset.cpus_requested, GFP_KERNEL))
2144 BUG();
2145
2146 cpumask_setall(top_cpuset.cpus_allowed);
2147 cpumask_setall(top_cpuset.cpus_requested);
2148 nodes_setall(top_cpuset.mems_allowed);
2149 cpumask_setall(top_cpuset.effective_cpus);
2150 nodes_setall(top_cpuset.effective_mems);
2151
2152 fmeter_init(&top_cpuset.fmeter);
2153 set_bit(CS_SCHED_LOAD_BALANCE, &top_cpuset.flags);
2154 top_cpuset.relax_domain_level = -1;
2155
2156 err = register_filesystem(&cpuset_fs_type);
2157 if (err < 0)
2158 return err;
2159
2160 if (!alloc_cpumask_var(&cpus_attach, GFP_KERNEL))
2161 BUG();
2162
2163 return 0;
2164 }
2165
2166 /*
2167 * If CPU and/or memory hotplug handlers, below, unplug any CPUs
2168 * or memory nodes, we need to walk over the cpuset hierarchy,
2169 * removing that CPU or node from all cpusets. If this removes the
2170 * last CPU or node from a cpuset, then move the tasks in the empty
2171 * cpuset to its next-highest non-empty parent.
2172 */
remove_tasks_in_empty_cpuset(struct cpuset * cs)2173 static void remove_tasks_in_empty_cpuset(struct cpuset *cs)
2174 {
2175 struct cpuset *parent;
2176
2177 /*
2178 * Find its next-highest non-empty parent, (top cpuset
2179 * has online cpus, so can't be empty).
2180 */
2181 parent = parent_cs(cs);
2182 while (cpumask_empty(parent->cpus_allowed) ||
2183 nodes_empty(parent->mems_allowed))
2184 parent = parent_cs(parent);
2185
2186 if (cgroup_transfer_tasks(parent->css.cgroup, cs->css.cgroup)) {
2187 pr_err("cpuset: failed to transfer tasks out of empty cpuset ");
2188 pr_cont_cgroup_name(cs->css.cgroup);
2189 pr_cont("\n");
2190 }
2191 }
2192
2193 static void
hotplug_update_tasks_legacy(struct cpuset * cs,struct cpumask * new_cpus,nodemask_t * new_mems,bool cpus_updated,bool mems_updated)2194 hotplug_update_tasks_legacy(struct cpuset *cs,
2195 struct cpumask *new_cpus, nodemask_t *new_mems,
2196 bool cpus_updated, bool mems_updated)
2197 {
2198 bool is_empty;
2199
2200 spin_lock_irq(&callback_lock);
2201 cpumask_copy(cs->cpus_allowed, new_cpus);
2202 cpumask_copy(cs->effective_cpus, new_cpus);
2203 cs->mems_allowed = *new_mems;
2204 cs->effective_mems = *new_mems;
2205 spin_unlock_irq(&callback_lock);
2206
2207 /*
2208 * Don't call update_tasks_cpumask() if the cpuset becomes empty,
2209 * as the tasks will be migratecd to an ancestor.
2210 */
2211 if (cpus_updated && !cpumask_empty(cs->cpus_allowed))
2212 update_tasks_cpumask(cs);
2213 if (mems_updated && !nodes_empty(cs->mems_allowed))
2214 update_tasks_nodemask(cs);
2215
2216 is_empty = cpumask_empty(cs->cpus_allowed) ||
2217 nodes_empty(cs->mems_allowed);
2218
2219 mutex_unlock(&cpuset_mutex);
2220
2221 /*
2222 * Move tasks to the nearest ancestor with execution resources,
2223 * This is full cgroup operation which will also call back into
2224 * cpuset. Should be done outside any lock.
2225 */
2226 if (is_empty)
2227 remove_tasks_in_empty_cpuset(cs);
2228
2229 mutex_lock(&cpuset_mutex);
2230 }
2231
2232 static void
hotplug_update_tasks(struct cpuset * cs,struct cpumask * new_cpus,nodemask_t * new_mems,bool cpus_updated,bool mems_updated)2233 hotplug_update_tasks(struct cpuset *cs,
2234 struct cpumask *new_cpus, nodemask_t *new_mems,
2235 bool cpus_updated, bool mems_updated)
2236 {
2237 if (cpumask_empty(new_cpus))
2238 cpumask_copy(new_cpus, parent_cs(cs)->effective_cpus);
2239 if (nodes_empty(*new_mems))
2240 *new_mems = parent_cs(cs)->effective_mems;
2241
2242 spin_lock_irq(&callback_lock);
2243 cpumask_copy(cs->effective_cpus, new_cpus);
2244 cs->effective_mems = *new_mems;
2245 spin_unlock_irq(&callback_lock);
2246
2247 if (cpus_updated)
2248 update_tasks_cpumask(cs);
2249 if (mems_updated)
2250 update_tasks_nodemask(cs);
2251 }
2252
2253 /**
2254 * cpuset_hotplug_update_tasks - update tasks in a cpuset for hotunplug
2255 * @cs: cpuset in interest
2256 *
2257 * Compare @cs's cpu and mem masks against top_cpuset and if some have gone
2258 * offline, update @cs accordingly. If @cs ends up with no CPU or memory,
2259 * all its tasks are moved to the nearest ancestor with both resources.
2260 */
cpuset_hotplug_update_tasks(struct cpuset * cs)2261 static void cpuset_hotplug_update_tasks(struct cpuset *cs)
2262 {
2263 static cpumask_t new_cpus;
2264 static nodemask_t new_mems;
2265 bool cpus_updated;
2266 bool mems_updated;
2267 retry:
2268 wait_event(cpuset_attach_wq, cs->attach_in_progress == 0);
2269
2270 mutex_lock(&cpuset_mutex);
2271
2272 /*
2273 * We have raced with task attaching. We wait until attaching
2274 * is finished, so we won't attach a task to an empty cpuset.
2275 */
2276 if (cs->attach_in_progress) {
2277 mutex_unlock(&cpuset_mutex);
2278 goto retry;
2279 }
2280
2281 cpumask_and(&new_cpus, cs->cpus_requested, parent_cs(cs)->effective_cpus);
2282 nodes_and(new_mems, cs->mems_allowed, parent_cs(cs)->effective_mems);
2283
2284 cpus_updated = !cpumask_equal(&new_cpus, cs->effective_cpus);
2285 mems_updated = !nodes_equal(new_mems, cs->effective_mems);
2286
2287 if (cgroup_subsys_on_dfl(cpuset_cgrp_subsys))
2288 hotplug_update_tasks(cs, &new_cpus, &new_mems,
2289 cpus_updated, mems_updated);
2290 else
2291 hotplug_update_tasks_legacy(cs, &new_cpus, &new_mems,
2292 cpus_updated, mems_updated);
2293
2294 mutex_unlock(&cpuset_mutex);
2295 }
2296
2297 static bool force_rebuild;
2298
cpuset_force_rebuild(void)2299 void cpuset_force_rebuild(void)
2300 {
2301 force_rebuild = true;
2302 }
2303
2304 /**
2305 * cpuset_hotplug_workfn - handle CPU/memory hotunplug for a cpuset
2306 *
2307 * This function is called after either CPU or memory configuration has
2308 * changed and updates cpuset accordingly. The top_cpuset is always
2309 * synchronized to cpu_active_mask and N_MEMORY, which is necessary in
2310 * order to make cpusets transparent (of no affect) on systems that are
2311 * actively using CPU hotplug but making no active use of cpusets.
2312 *
2313 * Non-root cpusets are only affected by offlining. If any CPUs or memory
2314 * nodes have been taken down, cpuset_hotplug_update_tasks() is invoked on
2315 * all descendants.
2316 *
2317 * Note that CPU offlining during suspend is ignored. We don't modify
2318 * cpusets across suspend/resume cycles at all.
2319 */
cpuset_hotplug_workfn(struct work_struct * work)2320 static void cpuset_hotplug_workfn(struct work_struct *work)
2321 {
2322 static cpumask_t new_cpus;
2323 static nodemask_t new_mems;
2324 bool cpus_updated, mems_updated;
2325 bool on_dfl = cgroup_subsys_on_dfl(cpuset_cgrp_subsys);
2326
2327 mutex_lock(&cpuset_mutex);
2328
2329 /* fetch the available cpus/mems and find out which changed how */
2330 cpumask_copy(&new_cpus, cpu_active_mask);
2331 new_mems = node_states[N_MEMORY];
2332
2333 cpus_updated = !cpumask_equal(top_cpuset.effective_cpus, &new_cpus);
2334 mems_updated = !nodes_equal(top_cpuset.effective_mems, new_mems);
2335
2336 /* synchronize cpus_allowed to cpu_active_mask */
2337 if (cpus_updated) {
2338 spin_lock_irq(&callback_lock);
2339 if (!on_dfl)
2340 cpumask_copy(top_cpuset.cpus_allowed, &new_cpus);
2341 cpumask_copy(top_cpuset.effective_cpus, &new_cpus);
2342 spin_unlock_irq(&callback_lock);
2343 /* we don't mess with cpumasks of tasks in top_cpuset */
2344 }
2345
2346 /* synchronize mems_allowed to N_MEMORY */
2347 if (mems_updated) {
2348 spin_lock_irq(&callback_lock);
2349 if (!on_dfl)
2350 top_cpuset.mems_allowed = new_mems;
2351 top_cpuset.effective_mems = new_mems;
2352 spin_unlock_irq(&callback_lock);
2353 update_tasks_nodemask(&top_cpuset);
2354 }
2355
2356 mutex_unlock(&cpuset_mutex);
2357
2358 /* if cpus or mems changed, we need to propagate to descendants */
2359 if (cpus_updated || mems_updated) {
2360 struct cpuset *cs;
2361 struct cgroup_subsys_state *pos_css;
2362
2363 rcu_read_lock();
2364 cpuset_for_each_descendant_pre(cs, pos_css, &top_cpuset) {
2365 if (cs == &top_cpuset || !css_tryget_online(&cs->css))
2366 continue;
2367 rcu_read_unlock();
2368
2369 cpuset_hotplug_update_tasks(cs);
2370
2371 rcu_read_lock();
2372 css_put(&cs->css);
2373 }
2374 rcu_read_unlock();
2375 }
2376
2377 /* rebuild sched domains if cpus_allowed has changed */
2378 if (cpus_updated || force_rebuild) {
2379 force_rebuild = false;
2380 rebuild_sched_domains();
2381 }
2382 }
2383
cpuset_update_active_cpus(bool cpu_online)2384 void cpuset_update_active_cpus(bool cpu_online)
2385 {
2386 /*
2387 * We're inside cpu hotplug critical region which usually nests
2388 * inside cgroup synchronization. Bounce actual hotplug processing
2389 * to a work item to avoid reverse locking order.
2390 *
2391 * We still need to do partition_sched_domains() synchronously;
2392 * otherwise, the scheduler will get confused and put tasks to the
2393 * dead CPU. Fall back to the default single domain.
2394 * cpuset_hotplug_workfn() will rebuild it as necessary.
2395 */
2396 partition_sched_domains(1, NULL, NULL);
2397 schedule_work(&cpuset_hotplug_work);
2398 }
2399
cpuset_wait_for_hotplug(void)2400 void cpuset_wait_for_hotplug(void)
2401 {
2402 flush_work(&cpuset_hotplug_work);
2403 }
2404
2405 /*
2406 * Keep top_cpuset.mems_allowed tracking node_states[N_MEMORY].
2407 * Call this routine anytime after node_states[N_MEMORY] changes.
2408 * See cpuset_update_active_cpus() for CPU hotplug handling.
2409 */
cpuset_track_online_nodes(struct notifier_block * self,unsigned long action,void * arg)2410 static int cpuset_track_online_nodes(struct notifier_block *self,
2411 unsigned long action, void *arg)
2412 {
2413 schedule_work(&cpuset_hotplug_work);
2414 return NOTIFY_OK;
2415 }
2416
2417 static struct notifier_block cpuset_track_online_nodes_nb = {
2418 .notifier_call = cpuset_track_online_nodes,
2419 .priority = 10, /* ??! */
2420 };
2421
2422 /**
2423 * cpuset_init_smp - initialize cpus_allowed
2424 *
2425 * Description: Finish top cpuset after cpu, node maps are initialized
2426 */
cpuset_init_smp(void)2427 void __init cpuset_init_smp(void)
2428 {
2429 cpumask_copy(top_cpuset.cpus_allowed, cpu_active_mask);
2430 top_cpuset.mems_allowed = node_states[N_MEMORY];
2431 top_cpuset.old_mems_allowed = top_cpuset.mems_allowed;
2432
2433 cpumask_copy(top_cpuset.effective_cpus, cpu_active_mask);
2434 top_cpuset.effective_mems = node_states[N_MEMORY];
2435
2436 register_hotmemory_notifier(&cpuset_track_online_nodes_nb);
2437
2438 cpuset_migrate_mm_wq = alloc_ordered_workqueue("cpuset_migrate_mm", 0);
2439 BUG_ON(!cpuset_migrate_mm_wq);
2440 }
2441
2442 /**
2443 * cpuset_cpus_allowed - return cpus_allowed mask from a tasks cpuset.
2444 * @tsk: pointer to task_struct from which to obtain cpuset->cpus_allowed.
2445 * @pmask: pointer to struct cpumask variable to receive cpus_allowed set.
2446 *
2447 * Description: Returns the cpumask_var_t cpus_allowed of the cpuset
2448 * attached to the specified @tsk. Guaranteed to return some non-empty
2449 * subset of cpu_online_mask, even if this means going outside the
2450 * tasks cpuset.
2451 **/
2452
cpuset_cpus_allowed(struct task_struct * tsk,struct cpumask * pmask)2453 void cpuset_cpus_allowed(struct task_struct *tsk, struct cpumask *pmask)
2454 {
2455 unsigned long flags;
2456
2457 spin_lock_irqsave(&callback_lock, flags);
2458 rcu_read_lock();
2459 guarantee_online_cpus(task_cs(tsk), pmask);
2460 rcu_read_unlock();
2461 spin_unlock_irqrestore(&callback_lock, flags);
2462 }
2463
cpuset_cpus_allowed_fallback(struct task_struct * tsk)2464 void cpuset_cpus_allowed_fallback(struct task_struct *tsk)
2465 {
2466 rcu_read_lock();
2467 do_set_cpus_allowed(tsk, task_cs(tsk)->effective_cpus);
2468 rcu_read_unlock();
2469
2470 /*
2471 * We own tsk->cpus_allowed, nobody can change it under us.
2472 *
2473 * But we used cs && cs->cpus_allowed lockless and thus can
2474 * race with cgroup_attach_task() or update_cpumask() and get
2475 * the wrong tsk->cpus_allowed. However, both cases imply the
2476 * subsequent cpuset_change_cpumask()->set_cpus_allowed_ptr()
2477 * which takes task_rq_lock().
2478 *
2479 * If we are called after it dropped the lock we must see all
2480 * changes in tsk_cs()->cpus_allowed. Otherwise we can temporary
2481 * set any mask even if it is not right from task_cs() pov,
2482 * the pending set_cpus_allowed_ptr() will fix things.
2483 *
2484 * select_fallback_rq() will fix things ups and set cpu_possible_mask
2485 * if required.
2486 */
2487 }
2488
cpuset_init_current_mems_allowed(void)2489 void __init cpuset_init_current_mems_allowed(void)
2490 {
2491 nodes_setall(current->mems_allowed);
2492 }
2493
2494 /**
2495 * cpuset_mems_allowed - return mems_allowed mask from a tasks cpuset.
2496 * @tsk: pointer to task_struct from which to obtain cpuset->mems_allowed.
2497 *
2498 * Description: Returns the nodemask_t mems_allowed of the cpuset
2499 * attached to the specified @tsk. Guaranteed to return some non-empty
2500 * subset of node_states[N_MEMORY], even if this means going outside the
2501 * tasks cpuset.
2502 **/
2503
cpuset_mems_allowed(struct task_struct * tsk)2504 nodemask_t cpuset_mems_allowed(struct task_struct *tsk)
2505 {
2506 nodemask_t mask;
2507 unsigned long flags;
2508
2509 spin_lock_irqsave(&callback_lock, flags);
2510 rcu_read_lock();
2511 guarantee_online_mems(task_cs(tsk), &mask);
2512 rcu_read_unlock();
2513 spin_unlock_irqrestore(&callback_lock, flags);
2514
2515 return mask;
2516 }
2517
2518 /**
2519 * cpuset_nodemask_valid_mems_allowed - check nodemask vs. curremt mems_allowed
2520 * @nodemask: the nodemask to be checked
2521 *
2522 * Are any of the nodes in the nodemask allowed in current->mems_allowed?
2523 */
cpuset_nodemask_valid_mems_allowed(nodemask_t * nodemask)2524 int cpuset_nodemask_valid_mems_allowed(nodemask_t *nodemask)
2525 {
2526 return nodes_intersects(*nodemask, current->mems_allowed);
2527 }
2528
2529 /*
2530 * nearest_hardwall_ancestor() - Returns the nearest mem_exclusive or
2531 * mem_hardwall ancestor to the specified cpuset. Call holding
2532 * callback_lock. If no ancestor is mem_exclusive or mem_hardwall
2533 * (an unusual configuration), then returns the root cpuset.
2534 */
nearest_hardwall_ancestor(struct cpuset * cs)2535 static struct cpuset *nearest_hardwall_ancestor(struct cpuset *cs)
2536 {
2537 while (!(is_mem_exclusive(cs) || is_mem_hardwall(cs)) && parent_cs(cs))
2538 cs = parent_cs(cs);
2539 return cs;
2540 }
2541
2542 /**
2543 * cpuset_node_allowed - Can we allocate on a memory node?
2544 * @node: is this an allowed node?
2545 * @gfp_mask: memory allocation flags
2546 *
2547 * If we're in interrupt, yes, we can always allocate. If @node is set in
2548 * current's mems_allowed, yes. If it's not a __GFP_HARDWALL request and this
2549 * node is set in the nearest hardwalled cpuset ancestor to current's cpuset,
2550 * yes. If current has access to memory reserves due to TIF_MEMDIE, yes.
2551 * Otherwise, no.
2552 *
2553 * GFP_USER allocations are marked with the __GFP_HARDWALL bit,
2554 * and do not allow allocations outside the current tasks cpuset
2555 * unless the task has been OOM killed as is marked TIF_MEMDIE.
2556 * GFP_KERNEL allocations are not so marked, so can escape to the
2557 * nearest enclosing hardwalled ancestor cpuset.
2558 *
2559 * Scanning up parent cpusets requires callback_lock. The
2560 * __alloc_pages() routine only calls here with __GFP_HARDWALL bit
2561 * _not_ set if it's a GFP_KERNEL allocation, and all nodes in the
2562 * current tasks mems_allowed came up empty on the first pass over
2563 * the zonelist. So only GFP_KERNEL allocations, if all nodes in the
2564 * cpuset are short of memory, might require taking the callback_lock.
2565 *
2566 * The first call here from mm/page_alloc:get_page_from_freelist()
2567 * has __GFP_HARDWALL set in gfp_mask, enforcing hardwall cpusets,
2568 * so no allocation on a node outside the cpuset is allowed (unless
2569 * in interrupt, of course).
2570 *
2571 * The second pass through get_page_from_freelist() doesn't even call
2572 * here for GFP_ATOMIC calls. For those calls, the __alloc_pages()
2573 * variable 'wait' is not set, and the bit ALLOC_CPUSET is not set
2574 * in alloc_flags. That logic and the checks below have the combined
2575 * affect that:
2576 * in_interrupt - any node ok (current task context irrelevant)
2577 * GFP_ATOMIC - any node ok
2578 * TIF_MEMDIE - any node ok
2579 * GFP_KERNEL - any node in enclosing hardwalled cpuset ok
2580 * GFP_USER - only nodes in current tasks mems allowed ok.
2581 */
__cpuset_node_allowed(int node,gfp_t gfp_mask)2582 int __cpuset_node_allowed(int node, gfp_t gfp_mask)
2583 {
2584 struct cpuset *cs; /* current cpuset ancestors */
2585 int allowed; /* is allocation in zone z allowed? */
2586 unsigned long flags;
2587
2588 if (in_interrupt())
2589 return 1;
2590 if (node_isset(node, current->mems_allowed))
2591 return 1;
2592 /*
2593 * Allow tasks that have access to memory reserves because they have
2594 * been OOM killed to get memory anywhere.
2595 */
2596 if (unlikely(test_thread_flag(TIF_MEMDIE)))
2597 return 1;
2598 if (gfp_mask & __GFP_HARDWALL) /* If hardwall request, stop here */
2599 return 0;
2600
2601 if (current->flags & PF_EXITING) /* Let dying task have memory */
2602 return 1;
2603
2604 /* Not hardwall and node outside mems_allowed: scan up cpusets */
2605 spin_lock_irqsave(&callback_lock, flags);
2606
2607 rcu_read_lock();
2608 cs = nearest_hardwall_ancestor(task_cs(current));
2609 allowed = node_isset(node, cs->mems_allowed);
2610 rcu_read_unlock();
2611
2612 spin_unlock_irqrestore(&callback_lock, flags);
2613 return allowed;
2614 }
2615
2616 /**
2617 * cpuset_mem_spread_node() - On which node to begin search for a file page
2618 * cpuset_slab_spread_node() - On which node to begin search for a slab page
2619 *
2620 * If a task is marked PF_SPREAD_PAGE or PF_SPREAD_SLAB (as for
2621 * tasks in a cpuset with is_spread_page or is_spread_slab set),
2622 * and if the memory allocation used cpuset_mem_spread_node()
2623 * to determine on which node to start looking, as it will for
2624 * certain page cache or slab cache pages such as used for file
2625 * system buffers and inode caches, then instead of starting on the
2626 * local node to look for a free page, rather spread the starting
2627 * node around the tasks mems_allowed nodes.
2628 *
2629 * We don't have to worry about the returned node being offline
2630 * because "it can't happen", and even if it did, it would be ok.
2631 *
2632 * The routines calling guarantee_online_mems() are careful to
2633 * only set nodes in task->mems_allowed that are online. So it
2634 * should not be possible for the following code to return an
2635 * offline node. But if it did, that would be ok, as this routine
2636 * is not returning the node where the allocation must be, only
2637 * the node where the search should start. The zonelist passed to
2638 * __alloc_pages() will include all nodes. If the slab allocator
2639 * is passed an offline node, it will fall back to the local node.
2640 * See kmem_cache_alloc_node().
2641 */
2642
cpuset_spread_node(int * rotor)2643 static int cpuset_spread_node(int *rotor)
2644 {
2645 int node;
2646
2647 node = next_node(*rotor, current->mems_allowed);
2648 if (node == MAX_NUMNODES)
2649 node = first_node(current->mems_allowed);
2650 *rotor = node;
2651 return node;
2652 }
2653
cpuset_mem_spread_node(void)2654 int cpuset_mem_spread_node(void)
2655 {
2656 if (current->cpuset_mem_spread_rotor == NUMA_NO_NODE)
2657 current->cpuset_mem_spread_rotor =
2658 node_random(¤t->mems_allowed);
2659
2660 return cpuset_spread_node(¤t->cpuset_mem_spread_rotor);
2661 }
2662
cpuset_slab_spread_node(void)2663 int cpuset_slab_spread_node(void)
2664 {
2665 if (current->cpuset_slab_spread_rotor == NUMA_NO_NODE)
2666 current->cpuset_slab_spread_rotor =
2667 node_random(¤t->mems_allowed);
2668
2669 return cpuset_spread_node(¤t->cpuset_slab_spread_rotor);
2670 }
2671
2672 EXPORT_SYMBOL_GPL(cpuset_mem_spread_node);
2673
2674 /**
2675 * cpuset_mems_allowed_intersects - Does @tsk1's mems_allowed intersect @tsk2's?
2676 * @tsk1: pointer to task_struct of some task.
2677 * @tsk2: pointer to task_struct of some other task.
2678 *
2679 * Description: Return true if @tsk1's mems_allowed intersects the
2680 * mems_allowed of @tsk2. Used by the OOM killer to determine if
2681 * one of the task's memory usage might impact the memory available
2682 * to the other.
2683 **/
2684
cpuset_mems_allowed_intersects(const struct task_struct * tsk1,const struct task_struct * tsk2)2685 int cpuset_mems_allowed_intersects(const struct task_struct *tsk1,
2686 const struct task_struct *tsk2)
2687 {
2688 return nodes_intersects(tsk1->mems_allowed, tsk2->mems_allowed);
2689 }
2690
2691 /**
2692 * cpuset_print_current_mems_allowed - prints current's cpuset and mems_allowed
2693 *
2694 * Description: Prints current's name, cpuset name, and cached copy of its
2695 * mems_allowed to the kernel log.
2696 */
cpuset_print_current_mems_allowed(void)2697 void cpuset_print_current_mems_allowed(void)
2698 {
2699 struct cgroup *cgrp;
2700
2701 rcu_read_lock();
2702
2703 cgrp = task_cs(current)->css.cgroup;
2704 pr_info("%s cpuset=", current->comm);
2705 pr_cont_cgroup_name(cgrp);
2706 pr_cont(" mems_allowed=%*pbl\n",
2707 nodemask_pr_args(¤t->mems_allowed));
2708
2709 rcu_read_unlock();
2710 }
2711
2712 /*
2713 * Collection of memory_pressure is suppressed unless
2714 * this flag is enabled by writing "1" to the special
2715 * cpuset file 'memory_pressure_enabled' in the root cpuset.
2716 */
2717
2718 int cpuset_memory_pressure_enabled __read_mostly;
2719
2720 /**
2721 * cpuset_memory_pressure_bump - keep stats of per-cpuset reclaims.
2722 *
2723 * Keep a running average of the rate of synchronous (direct)
2724 * page reclaim efforts initiated by tasks in each cpuset.
2725 *
2726 * This represents the rate at which some task in the cpuset
2727 * ran low on memory on all nodes it was allowed to use, and
2728 * had to enter the kernels page reclaim code in an effort to
2729 * create more free memory by tossing clean pages or swapping
2730 * or writing dirty pages.
2731 *
2732 * Display to user space in the per-cpuset read-only file
2733 * "memory_pressure". Value displayed is an integer
2734 * representing the recent rate of entry into the synchronous
2735 * (direct) page reclaim by any task attached to the cpuset.
2736 **/
2737
__cpuset_memory_pressure_bump(void)2738 void __cpuset_memory_pressure_bump(void)
2739 {
2740 rcu_read_lock();
2741 fmeter_markevent(&task_cs(current)->fmeter);
2742 rcu_read_unlock();
2743 }
2744
2745 #ifdef CONFIG_PROC_PID_CPUSET
2746 /*
2747 * proc_cpuset_show()
2748 * - Print tasks cpuset path into seq_file.
2749 * - Used for /proc/<pid>/cpuset.
2750 * - No need to task_lock(tsk) on this tsk->cpuset reference, as it
2751 * doesn't really matter if tsk->cpuset changes after we read it,
2752 * and we take cpuset_mutex, keeping cpuset_attach() from changing it
2753 * anyway.
2754 */
proc_cpuset_show(struct seq_file * m,struct pid_namespace * ns,struct pid * pid,struct task_struct * tsk)2755 int proc_cpuset_show(struct seq_file *m, struct pid_namespace *ns,
2756 struct pid *pid, struct task_struct *tsk)
2757 {
2758 char *buf, *p;
2759 struct cgroup_subsys_state *css;
2760 int retval;
2761
2762 retval = -ENOMEM;
2763 buf = kmalloc(PATH_MAX, GFP_KERNEL);
2764 if (!buf)
2765 goto out;
2766
2767 retval = -ENAMETOOLONG;
2768 rcu_read_lock();
2769 css = task_css(tsk, cpuset_cgrp_id);
2770 p = cgroup_path(css->cgroup, buf, PATH_MAX);
2771 rcu_read_unlock();
2772 if (!p)
2773 goto out_free;
2774 seq_puts(m, p);
2775 seq_putc(m, '\n');
2776 retval = 0;
2777 out_free:
2778 kfree(buf);
2779 out:
2780 return retval;
2781 }
2782 #endif /* CONFIG_PROC_PID_CPUSET */
2783
2784 /* Display task mems_allowed in /proc/<pid>/status file. */
cpuset_task_status_allowed(struct seq_file * m,struct task_struct * task)2785 void cpuset_task_status_allowed(struct seq_file *m, struct task_struct *task)
2786 {
2787 seq_printf(m, "Mems_allowed:\t%*pb\n",
2788 nodemask_pr_args(&task->mems_allowed));
2789 seq_printf(m, "Mems_allowed_list:\t%*pbl\n",
2790 nodemask_pr_args(&task->mems_allowed));
2791 }
2792