xref: /linux/kernel/cgroup/cpuset.c (revision ef69f8d2ff09518657c3ecaf2db8408c16549829)
1 /*
2  *  kernel/cpuset.c
3  *
4  *  Processor and Memory placement constraints for sets of tasks.
5  *
6  *  Copyright (C) 2003 BULL SA.
7  *  Copyright (C) 2004-2007 Silicon Graphics, Inc.
8  *  Copyright (C) 2006 Google, Inc
9  *
10  *  Portions derived from Patrick Mochel's sysfs code.
11  *  sysfs is Copyright (c) 2001-3 Patrick Mochel
12  *
13  *  2003-10-10 Written by Simon Derr.
14  *  2003-10-22 Updates by Stephen Hemminger.
15  *  2004 May-July Rework by Paul Jackson.
16  *  2006 Rework by Paul Menage to use generic cgroups
17  *  2008 Rework of the scheduler domains and CPU hotplug handling
18  *       by Max Krasnyansky
19  *
20  *  This file is subject to the terms and conditions of the GNU General Public
21  *  License.  See the file COPYING in the main directory of the Linux
22  *  distribution for more details.
23  */
24 
25 #include <linux/cpu.h>
26 #include <linux/cpumask.h>
27 #include <linux/cpuset.h>
28 #include <linux/err.h>
29 #include <linux/errno.h>
30 #include <linux/file.h>
31 #include <linux/fs.h>
32 #include <linux/init.h>
33 #include <linux/interrupt.h>
34 #include <linux/kernel.h>
35 #include <linux/kmod.h>
36 #include <linux/list.h>
37 #include <linux/mempolicy.h>
38 #include <linux/mm.h>
39 #include <linux/memory.h>
40 #include <linux/export.h>
41 #include <linux/mount.h>
42 #include <linux/namei.h>
43 #include <linux/pagemap.h>
44 #include <linux/proc_fs.h>
45 #include <linux/rcupdate.h>
46 #include <linux/sched.h>
47 #include <linux/sched/mm.h>
48 #include <linux/sched/task.h>
49 #include <linux/seq_file.h>
50 #include <linux/security.h>
51 #include <linux/slab.h>
52 #include <linux/spinlock.h>
53 #include <linux/stat.h>
54 #include <linux/string.h>
55 #include <linux/time.h>
56 #include <linux/time64.h>
57 #include <linux/backing-dev.h>
58 #include <linux/sort.h>
59 #include <linux/oom.h>
60 #include <linux/sched/isolation.h>
61 #include <linux/uaccess.h>
62 #include <linux/atomic.h>
63 #include <linux/mutex.h>
64 #include <linux/cgroup.h>
65 #include <linux/wait.h>
66 
67 DEFINE_STATIC_KEY_FALSE(cpusets_pre_enable_key);
68 DEFINE_STATIC_KEY_FALSE(cpusets_enabled_key);
69 
70 /* See "Frequency meter" comments, below. */
71 
72 struct fmeter {
73 	int cnt;		/* unprocessed events count */
74 	int val;		/* most recent output value */
75 	time64_t time;		/* clock (secs) when val computed */
76 	spinlock_t lock;	/* guards read or write of above */
77 };
78 
79 struct cpuset {
80 	struct cgroup_subsys_state css;
81 
82 	unsigned long flags;		/* "unsigned long" so bitops work */
83 
84 	/*
85 	 * On default hierarchy:
86 	 *
87 	 * The user-configured masks can only be changed by writing to
88 	 * cpuset.cpus and cpuset.mems, and won't be limited by the
89 	 * parent masks.
90 	 *
91 	 * The effective masks is the real masks that apply to the tasks
92 	 * in the cpuset. They may be changed if the configured masks are
93 	 * changed or hotplug happens.
94 	 *
95 	 * effective_mask == configured_mask & parent's effective_mask,
96 	 * and if it ends up empty, it will inherit the parent's mask.
97 	 *
98 	 *
99 	 * On legacy hierachy:
100 	 *
101 	 * The user-configured masks are always the same with effective masks.
102 	 */
103 
104 	/* user-configured CPUs and Memory Nodes allow to tasks */
105 	cpumask_var_t cpus_allowed;
106 	nodemask_t mems_allowed;
107 
108 	/* effective CPUs and Memory Nodes allow to tasks */
109 	cpumask_var_t effective_cpus;
110 	nodemask_t effective_mems;
111 
112 	/*
113 	 * This is old Memory Nodes tasks took on.
114 	 *
115 	 * - top_cpuset.old_mems_allowed is initialized to mems_allowed.
116 	 * - A new cpuset's old_mems_allowed is initialized when some
117 	 *   task is moved into it.
118 	 * - old_mems_allowed is used in cpuset_migrate_mm() when we change
119 	 *   cpuset.mems_allowed and have tasks' nodemask updated, and
120 	 *   then old_mems_allowed is updated to mems_allowed.
121 	 */
122 	nodemask_t old_mems_allowed;
123 
124 	struct fmeter fmeter;		/* memory_pressure filter */
125 
126 	/*
127 	 * Tasks are being attached to this cpuset.  Used to prevent
128 	 * zeroing cpus/mems_allowed between ->can_attach() and ->attach().
129 	 */
130 	int attach_in_progress;
131 
132 	/* partition number for rebuild_sched_domains() */
133 	int pn;
134 
135 	/* for custom sched domain */
136 	int relax_domain_level;
137 };
138 
139 static inline struct cpuset *css_cs(struct cgroup_subsys_state *css)
140 {
141 	return css ? container_of(css, struct cpuset, css) : NULL;
142 }
143 
144 /* Retrieve the cpuset for a task */
145 static inline struct cpuset *task_cs(struct task_struct *task)
146 {
147 	return css_cs(task_css(task, cpuset_cgrp_id));
148 }
149 
150 static inline struct cpuset *parent_cs(struct cpuset *cs)
151 {
152 	return css_cs(cs->css.parent);
153 }
154 
155 #ifdef CONFIG_NUMA
156 static inline bool task_has_mempolicy(struct task_struct *task)
157 {
158 	return task->mempolicy;
159 }
160 #else
161 static inline bool task_has_mempolicy(struct task_struct *task)
162 {
163 	return false;
164 }
165 #endif
166 
167 
168 /* bits in struct cpuset flags field */
169 typedef enum {
170 	CS_ONLINE,
171 	CS_CPU_EXCLUSIVE,
172 	CS_MEM_EXCLUSIVE,
173 	CS_MEM_HARDWALL,
174 	CS_MEMORY_MIGRATE,
175 	CS_SCHED_LOAD_BALANCE,
176 	CS_SPREAD_PAGE,
177 	CS_SPREAD_SLAB,
178 } cpuset_flagbits_t;
179 
180 /* convenient tests for these bits */
181 static inline bool is_cpuset_online(struct cpuset *cs)
182 {
183 	return test_bit(CS_ONLINE, &cs->flags) && !css_is_dying(&cs->css);
184 }
185 
186 static inline int is_cpu_exclusive(const struct cpuset *cs)
187 {
188 	return test_bit(CS_CPU_EXCLUSIVE, &cs->flags);
189 }
190 
191 static inline int is_mem_exclusive(const struct cpuset *cs)
192 {
193 	return test_bit(CS_MEM_EXCLUSIVE, &cs->flags);
194 }
195 
196 static inline int is_mem_hardwall(const struct cpuset *cs)
197 {
198 	return test_bit(CS_MEM_HARDWALL, &cs->flags);
199 }
200 
201 static inline int is_sched_load_balance(const struct cpuset *cs)
202 {
203 	return test_bit(CS_SCHED_LOAD_BALANCE, &cs->flags);
204 }
205 
206 static inline int is_memory_migrate(const struct cpuset *cs)
207 {
208 	return test_bit(CS_MEMORY_MIGRATE, &cs->flags);
209 }
210 
211 static inline int is_spread_page(const struct cpuset *cs)
212 {
213 	return test_bit(CS_SPREAD_PAGE, &cs->flags);
214 }
215 
216 static inline int is_spread_slab(const struct cpuset *cs)
217 {
218 	return test_bit(CS_SPREAD_SLAB, &cs->flags);
219 }
220 
221 static struct cpuset top_cpuset = {
222 	.flags = ((1 << CS_ONLINE) | (1 << CS_CPU_EXCLUSIVE) |
223 		  (1 << CS_MEM_EXCLUSIVE)),
224 };
225 
226 /**
227  * cpuset_for_each_child - traverse online children of a cpuset
228  * @child_cs: loop cursor pointing to the current child
229  * @pos_css: used for iteration
230  * @parent_cs: target cpuset to walk children of
231  *
232  * Walk @child_cs through the online children of @parent_cs.  Must be used
233  * with RCU read locked.
234  */
235 #define cpuset_for_each_child(child_cs, pos_css, parent_cs)		\
236 	css_for_each_child((pos_css), &(parent_cs)->css)		\
237 		if (is_cpuset_online(((child_cs) = css_cs((pos_css)))))
238 
239 /**
240  * cpuset_for_each_descendant_pre - pre-order walk of a cpuset's descendants
241  * @des_cs: loop cursor pointing to the current descendant
242  * @pos_css: used for iteration
243  * @root_cs: target cpuset to walk ancestor of
244  *
245  * Walk @des_cs through the online descendants of @root_cs.  Must be used
246  * with RCU read locked.  The caller may modify @pos_css by calling
247  * css_rightmost_descendant() to skip subtree.  @root_cs is included in the
248  * iteration and the first node to be visited.
249  */
250 #define cpuset_for_each_descendant_pre(des_cs, pos_css, root_cs)	\
251 	css_for_each_descendant_pre((pos_css), &(root_cs)->css)		\
252 		if (is_cpuset_online(((des_cs) = css_cs((pos_css)))))
253 
254 /*
255  * There are two global locks guarding cpuset structures - cpuset_mutex and
256  * callback_lock. We also require taking task_lock() when dereferencing a
257  * task's cpuset pointer. See "The task_lock() exception", at the end of this
258  * comment.
259  *
260  * A task must hold both locks to modify cpusets.  If a task holds
261  * cpuset_mutex, then it blocks others wanting that mutex, ensuring that it
262  * is the only task able to also acquire callback_lock and be able to
263  * modify cpusets.  It can perform various checks on the cpuset structure
264  * first, knowing nothing will change.  It can also allocate memory while
265  * just holding cpuset_mutex.  While it is performing these checks, various
266  * callback routines can briefly acquire callback_lock to query cpusets.
267  * Once it is ready to make the changes, it takes callback_lock, blocking
268  * everyone else.
269  *
270  * Calls to the kernel memory allocator can not be made while holding
271  * callback_lock, as that would risk double tripping on callback_lock
272  * from one of the callbacks into the cpuset code from within
273  * __alloc_pages().
274  *
275  * If a task is only holding callback_lock, then it has read-only
276  * access to cpusets.
277  *
278  * Now, the task_struct fields mems_allowed and mempolicy may be changed
279  * by other task, we use alloc_lock in the task_struct fields to protect
280  * them.
281  *
282  * The cpuset_common_file_read() handlers only hold callback_lock across
283  * small pieces of code, such as when reading out possibly multi-word
284  * cpumasks and nodemasks.
285  *
286  * Accessing a task's cpuset should be done in accordance with the
287  * guidelines for accessing subsystem state in kernel/cgroup.c
288  */
289 
290 static DEFINE_MUTEX(cpuset_mutex);
291 static DEFINE_SPINLOCK(callback_lock);
292 
293 static struct workqueue_struct *cpuset_migrate_mm_wq;
294 
295 /*
296  * CPU / memory hotplug is handled asynchronously.
297  */
298 static void cpuset_hotplug_workfn(struct work_struct *work);
299 static DECLARE_WORK(cpuset_hotplug_work, cpuset_hotplug_workfn);
300 
301 static DECLARE_WAIT_QUEUE_HEAD(cpuset_attach_wq);
302 
303 /*
304  * Cgroup v2 behavior is used when on default hierarchy or the
305  * cgroup_v2_mode flag is set.
306  */
307 static inline bool is_in_v2_mode(void)
308 {
309 	return cgroup_subsys_on_dfl(cpuset_cgrp_subsys) ||
310 	      (cpuset_cgrp_subsys.root->flags & CGRP_ROOT_CPUSET_V2_MODE);
311 }
312 
313 /*
314  * This is ugly, but preserves the userspace API for existing cpuset
315  * users. If someone tries to mount the "cpuset" filesystem, we
316  * silently switch it to mount "cgroup" instead
317  */
318 static struct dentry *cpuset_mount(struct file_system_type *fs_type,
319 			 int flags, const char *unused_dev_name, void *data)
320 {
321 	struct file_system_type *cgroup_fs = get_fs_type("cgroup");
322 	struct dentry *ret = ERR_PTR(-ENODEV);
323 	if (cgroup_fs) {
324 		char mountopts[] =
325 			"cpuset,noprefix,"
326 			"release_agent=/sbin/cpuset_release_agent";
327 		ret = cgroup_fs->mount(cgroup_fs, flags,
328 					   unused_dev_name, mountopts);
329 		put_filesystem(cgroup_fs);
330 	}
331 	return ret;
332 }
333 
334 static struct file_system_type cpuset_fs_type = {
335 	.name = "cpuset",
336 	.mount = cpuset_mount,
337 };
338 
339 /*
340  * Return in pmask the portion of a cpusets's cpus_allowed that
341  * are online.  If none are online, walk up the cpuset hierarchy
342  * until we find one that does have some online cpus.
343  *
344  * One way or another, we guarantee to return some non-empty subset
345  * of cpu_online_mask.
346  *
347  * Call with callback_lock or cpuset_mutex held.
348  */
349 static void guarantee_online_cpus(struct cpuset *cs, struct cpumask *pmask)
350 {
351 	while (!cpumask_intersects(cs->effective_cpus, cpu_online_mask)) {
352 		cs = parent_cs(cs);
353 		if (unlikely(!cs)) {
354 			/*
355 			 * The top cpuset doesn't have any online cpu as a
356 			 * consequence of a race between cpuset_hotplug_work
357 			 * and cpu hotplug notifier.  But we know the top
358 			 * cpuset's effective_cpus is on its way to to be
359 			 * identical to cpu_online_mask.
360 			 */
361 			cpumask_copy(pmask, cpu_online_mask);
362 			return;
363 		}
364 	}
365 	cpumask_and(pmask, cs->effective_cpus, cpu_online_mask);
366 }
367 
368 /*
369  * Return in *pmask the portion of a cpusets's mems_allowed that
370  * are online, with memory.  If none are online with memory, walk
371  * up the cpuset hierarchy until we find one that does have some
372  * online mems.  The top cpuset always has some mems online.
373  *
374  * One way or another, we guarantee to return some non-empty subset
375  * of node_states[N_MEMORY].
376  *
377  * Call with callback_lock or cpuset_mutex held.
378  */
379 static void guarantee_online_mems(struct cpuset *cs, nodemask_t *pmask)
380 {
381 	while (!nodes_intersects(cs->effective_mems, node_states[N_MEMORY]))
382 		cs = parent_cs(cs);
383 	nodes_and(*pmask, cs->effective_mems, node_states[N_MEMORY]);
384 }
385 
386 /*
387  * update task's spread flag if cpuset's page/slab spread flag is set
388  *
389  * Call with callback_lock or cpuset_mutex held.
390  */
391 static void cpuset_update_task_spread_flag(struct cpuset *cs,
392 					struct task_struct *tsk)
393 {
394 	if (is_spread_page(cs))
395 		task_set_spread_page(tsk);
396 	else
397 		task_clear_spread_page(tsk);
398 
399 	if (is_spread_slab(cs))
400 		task_set_spread_slab(tsk);
401 	else
402 		task_clear_spread_slab(tsk);
403 }
404 
405 /*
406  * is_cpuset_subset(p, q) - Is cpuset p a subset of cpuset q?
407  *
408  * One cpuset is a subset of another if all its allowed CPUs and
409  * Memory Nodes are a subset of the other, and its exclusive flags
410  * are only set if the other's are set.  Call holding cpuset_mutex.
411  */
412 
413 static int is_cpuset_subset(const struct cpuset *p, const struct cpuset *q)
414 {
415 	return	cpumask_subset(p->cpus_allowed, q->cpus_allowed) &&
416 		nodes_subset(p->mems_allowed, q->mems_allowed) &&
417 		is_cpu_exclusive(p) <= is_cpu_exclusive(q) &&
418 		is_mem_exclusive(p) <= is_mem_exclusive(q);
419 }
420 
421 /**
422  * alloc_trial_cpuset - allocate a trial cpuset
423  * @cs: the cpuset that the trial cpuset duplicates
424  */
425 static struct cpuset *alloc_trial_cpuset(struct cpuset *cs)
426 {
427 	struct cpuset *trial;
428 
429 	trial = kmemdup(cs, sizeof(*cs), GFP_KERNEL);
430 	if (!trial)
431 		return NULL;
432 
433 	if (!alloc_cpumask_var(&trial->cpus_allowed, GFP_KERNEL))
434 		goto free_cs;
435 	if (!alloc_cpumask_var(&trial->effective_cpus, GFP_KERNEL))
436 		goto free_cpus;
437 
438 	cpumask_copy(trial->cpus_allowed, cs->cpus_allowed);
439 	cpumask_copy(trial->effective_cpus, cs->effective_cpus);
440 	return trial;
441 
442 free_cpus:
443 	free_cpumask_var(trial->cpus_allowed);
444 free_cs:
445 	kfree(trial);
446 	return NULL;
447 }
448 
449 /**
450  * free_trial_cpuset - free the trial cpuset
451  * @trial: the trial cpuset to be freed
452  */
453 static void free_trial_cpuset(struct cpuset *trial)
454 {
455 	free_cpumask_var(trial->effective_cpus);
456 	free_cpumask_var(trial->cpus_allowed);
457 	kfree(trial);
458 }
459 
460 /*
461  * validate_change() - Used to validate that any proposed cpuset change
462  *		       follows the structural rules for cpusets.
463  *
464  * If we replaced the flag and mask values of the current cpuset
465  * (cur) with those values in the trial cpuset (trial), would
466  * our various subset and exclusive rules still be valid?  Presumes
467  * cpuset_mutex held.
468  *
469  * 'cur' is the address of an actual, in-use cpuset.  Operations
470  * such as list traversal that depend on the actual address of the
471  * cpuset in the list must use cur below, not trial.
472  *
473  * 'trial' is the address of bulk structure copy of cur, with
474  * perhaps one or more of the fields cpus_allowed, mems_allowed,
475  * or flags changed to new, trial values.
476  *
477  * Return 0 if valid, -errno if not.
478  */
479 
480 static int validate_change(struct cpuset *cur, struct cpuset *trial)
481 {
482 	struct cgroup_subsys_state *css;
483 	struct cpuset *c, *par;
484 	int ret;
485 
486 	rcu_read_lock();
487 
488 	/* Each of our child cpusets must be a subset of us */
489 	ret = -EBUSY;
490 	cpuset_for_each_child(c, css, cur)
491 		if (!is_cpuset_subset(c, trial))
492 			goto out;
493 
494 	/* Remaining checks don't apply to root cpuset */
495 	ret = 0;
496 	if (cur == &top_cpuset)
497 		goto out;
498 
499 	par = parent_cs(cur);
500 
501 	/* On legacy hiearchy, we must be a subset of our parent cpuset. */
502 	ret = -EACCES;
503 	if (!is_in_v2_mode() && !is_cpuset_subset(trial, par))
504 		goto out;
505 
506 	/*
507 	 * If either I or some sibling (!= me) is exclusive, we can't
508 	 * overlap
509 	 */
510 	ret = -EINVAL;
511 	cpuset_for_each_child(c, css, par) {
512 		if ((is_cpu_exclusive(trial) || is_cpu_exclusive(c)) &&
513 		    c != cur &&
514 		    cpumask_intersects(trial->cpus_allowed, c->cpus_allowed))
515 			goto out;
516 		if ((is_mem_exclusive(trial) || is_mem_exclusive(c)) &&
517 		    c != cur &&
518 		    nodes_intersects(trial->mems_allowed, c->mems_allowed))
519 			goto out;
520 	}
521 
522 	/*
523 	 * Cpusets with tasks - existing or newly being attached - can't
524 	 * be changed to have empty cpus_allowed or mems_allowed.
525 	 */
526 	ret = -ENOSPC;
527 	if ((cgroup_is_populated(cur->css.cgroup) || cur->attach_in_progress)) {
528 		if (!cpumask_empty(cur->cpus_allowed) &&
529 		    cpumask_empty(trial->cpus_allowed))
530 			goto out;
531 		if (!nodes_empty(cur->mems_allowed) &&
532 		    nodes_empty(trial->mems_allowed))
533 			goto out;
534 	}
535 
536 	/*
537 	 * We can't shrink if we won't have enough room for SCHED_DEADLINE
538 	 * tasks.
539 	 */
540 	ret = -EBUSY;
541 	if (is_cpu_exclusive(cur) &&
542 	    !cpuset_cpumask_can_shrink(cur->cpus_allowed,
543 				       trial->cpus_allowed))
544 		goto out;
545 
546 	ret = 0;
547 out:
548 	rcu_read_unlock();
549 	return ret;
550 }
551 
552 #ifdef CONFIG_SMP
553 /*
554  * Helper routine for generate_sched_domains().
555  * Do cpusets a, b have overlapping effective cpus_allowed masks?
556  */
557 static int cpusets_overlap(struct cpuset *a, struct cpuset *b)
558 {
559 	return cpumask_intersects(a->effective_cpus, b->effective_cpus);
560 }
561 
562 static void
563 update_domain_attr(struct sched_domain_attr *dattr, struct cpuset *c)
564 {
565 	if (dattr->relax_domain_level < c->relax_domain_level)
566 		dattr->relax_domain_level = c->relax_domain_level;
567 	return;
568 }
569 
570 static void update_domain_attr_tree(struct sched_domain_attr *dattr,
571 				    struct cpuset *root_cs)
572 {
573 	struct cpuset *cp;
574 	struct cgroup_subsys_state *pos_css;
575 
576 	rcu_read_lock();
577 	cpuset_for_each_descendant_pre(cp, pos_css, root_cs) {
578 		/* skip the whole subtree if @cp doesn't have any CPU */
579 		if (cpumask_empty(cp->cpus_allowed)) {
580 			pos_css = css_rightmost_descendant(pos_css);
581 			continue;
582 		}
583 
584 		if (is_sched_load_balance(cp))
585 			update_domain_attr(dattr, cp);
586 	}
587 	rcu_read_unlock();
588 }
589 
590 /* Must be called with cpuset_mutex held.  */
591 static inline int nr_cpusets(void)
592 {
593 	/* jump label reference count + the top-level cpuset */
594 	return static_key_count(&cpusets_enabled_key.key) + 1;
595 }
596 
597 /*
598  * generate_sched_domains()
599  *
600  * This function builds a partial partition of the systems CPUs
601  * A 'partial partition' is a set of non-overlapping subsets whose
602  * union is a subset of that set.
603  * The output of this function needs to be passed to kernel/sched/core.c
604  * partition_sched_domains() routine, which will rebuild the scheduler's
605  * load balancing domains (sched domains) as specified by that partial
606  * partition.
607  *
608  * See "What is sched_load_balance" in Documentation/cgroups/cpusets.txt
609  * for a background explanation of this.
610  *
611  * Does not return errors, on the theory that the callers of this
612  * routine would rather not worry about failures to rebuild sched
613  * domains when operating in the severe memory shortage situations
614  * that could cause allocation failures below.
615  *
616  * Must be called with cpuset_mutex held.
617  *
618  * The three key local variables below are:
619  *    q  - a linked-list queue of cpuset pointers, used to implement a
620  *	   top-down scan of all cpusets.  This scan loads a pointer
621  *	   to each cpuset marked is_sched_load_balance into the
622  *	   array 'csa'.  For our purposes, rebuilding the schedulers
623  *	   sched domains, we can ignore !is_sched_load_balance cpusets.
624  *  csa  - (for CpuSet Array) Array of pointers to all the cpusets
625  *	   that need to be load balanced, for convenient iterative
626  *	   access by the subsequent code that finds the best partition,
627  *	   i.e the set of domains (subsets) of CPUs such that the
628  *	   cpus_allowed of every cpuset marked is_sched_load_balance
629  *	   is a subset of one of these domains, while there are as
630  *	   many such domains as possible, each as small as possible.
631  * doms  - Conversion of 'csa' to an array of cpumasks, for passing to
632  *	   the kernel/sched/core.c routine partition_sched_domains() in a
633  *	   convenient format, that can be easily compared to the prior
634  *	   value to determine what partition elements (sched domains)
635  *	   were changed (added or removed.)
636  *
637  * Finding the best partition (set of domains):
638  *	The triple nested loops below over i, j, k scan over the
639  *	load balanced cpusets (using the array of cpuset pointers in
640  *	csa[]) looking for pairs of cpusets that have overlapping
641  *	cpus_allowed, but which don't have the same 'pn' partition
642  *	number and gives them in the same partition number.  It keeps
643  *	looping on the 'restart' label until it can no longer find
644  *	any such pairs.
645  *
646  *	The union of the cpus_allowed masks from the set of
647  *	all cpusets having the same 'pn' value then form the one
648  *	element of the partition (one sched domain) to be passed to
649  *	partition_sched_domains().
650  */
651 static int generate_sched_domains(cpumask_var_t **domains,
652 			struct sched_domain_attr **attributes)
653 {
654 	struct cpuset *cp;	/* scans q */
655 	struct cpuset **csa;	/* array of all cpuset ptrs */
656 	int csn;		/* how many cpuset ptrs in csa so far */
657 	int i, j, k;		/* indices for partition finding loops */
658 	cpumask_var_t *doms;	/* resulting partition; i.e. sched domains */
659 	struct sched_domain_attr *dattr;  /* attributes for custom domains */
660 	int ndoms = 0;		/* number of sched domains in result */
661 	int nslot;		/* next empty doms[] struct cpumask slot */
662 	struct cgroup_subsys_state *pos_css;
663 
664 	doms = NULL;
665 	dattr = NULL;
666 	csa = NULL;
667 
668 	/* Special case for the 99% of systems with one, full, sched domain */
669 	if (is_sched_load_balance(&top_cpuset)) {
670 		ndoms = 1;
671 		doms = alloc_sched_domains(ndoms);
672 		if (!doms)
673 			goto done;
674 
675 		dattr = kmalloc(sizeof(struct sched_domain_attr), GFP_KERNEL);
676 		if (dattr) {
677 			*dattr = SD_ATTR_INIT;
678 			update_domain_attr_tree(dattr, &top_cpuset);
679 		}
680 		cpumask_and(doms[0], top_cpuset.effective_cpus,
681 			    housekeeping_cpumask(HK_FLAG_DOMAIN));
682 
683 		goto done;
684 	}
685 
686 	csa = kmalloc(nr_cpusets() * sizeof(cp), GFP_KERNEL);
687 	if (!csa)
688 		goto done;
689 	csn = 0;
690 
691 	rcu_read_lock();
692 	cpuset_for_each_descendant_pre(cp, pos_css, &top_cpuset) {
693 		if (cp == &top_cpuset)
694 			continue;
695 		/*
696 		 * Continue traversing beyond @cp iff @cp has some CPUs and
697 		 * isn't load balancing.  The former is obvious.  The
698 		 * latter: All child cpusets contain a subset of the
699 		 * parent's cpus, so just skip them, and then we call
700 		 * update_domain_attr_tree() to calc relax_domain_level of
701 		 * the corresponding sched domain.
702 		 */
703 		if (!cpumask_empty(cp->cpus_allowed) &&
704 		    !(is_sched_load_balance(cp) &&
705 		      cpumask_intersects(cp->cpus_allowed,
706 					 housekeeping_cpumask(HK_FLAG_DOMAIN))))
707 			continue;
708 
709 		if (is_sched_load_balance(cp))
710 			csa[csn++] = cp;
711 
712 		/* skip @cp's subtree */
713 		pos_css = css_rightmost_descendant(pos_css);
714 	}
715 	rcu_read_unlock();
716 
717 	for (i = 0; i < csn; i++)
718 		csa[i]->pn = i;
719 	ndoms = csn;
720 
721 restart:
722 	/* Find the best partition (set of sched domains) */
723 	for (i = 0; i < csn; i++) {
724 		struct cpuset *a = csa[i];
725 		int apn = a->pn;
726 
727 		for (j = 0; j < csn; j++) {
728 			struct cpuset *b = csa[j];
729 			int bpn = b->pn;
730 
731 			if (apn != bpn && cpusets_overlap(a, b)) {
732 				for (k = 0; k < csn; k++) {
733 					struct cpuset *c = csa[k];
734 
735 					if (c->pn == bpn)
736 						c->pn = apn;
737 				}
738 				ndoms--;	/* one less element */
739 				goto restart;
740 			}
741 		}
742 	}
743 
744 	/*
745 	 * Now we know how many domains to create.
746 	 * Convert <csn, csa> to <ndoms, doms> and populate cpu masks.
747 	 */
748 	doms = alloc_sched_domains(ndoms);
749 	if (!doms)
750 		goto done;
751 
752 	/*
753 	 * The rest of the code, including the scheduler, can deal with
754 	 * dattr==NULL case. No need to abort if alloc fails.
755 	 */
756 	dattr = kmalloc(ndoms * sizeof(struct sched_domain_attr), GFP_KERNEL);
757 
758 	for (nslot = 0, i = 0; i < csn; i++) {
759 		struct cpuset *a = csa[i];
760 		struct cpumask *dp;
761 		int apn = a->pn;
762 
763 		if (apn < 0) {
764 			/* Skip completed partitions */
765 			continue;
766 		}
767 
768 		dp = doms[nslot];
769 
770 		if (nslot == ndoms) {
771 			static int warnings = 10;
772 			if (warnings) {
773 				pr_warn("rebuild_sched_domains confused: nslot %d, ndoms %d, csn %d, i %d, apn %d\n",
774 					nslot, ndoms, csn, i, apn);
775 				warnings--;
776 			}
777 			continue;
778 		}
779 
780 		cpumask_clear(dp);
781 		if (dattr)
782 			*(dattr + nslot) = SD_ATTR_INIT;
783 		for (j = i; j < csn; j++) {
784 			struct cpuset *b = csa[j];
785 
786 			if (apn == b->pn) {
787 				cpumask_or(dp, dp, b->effective_cpus);
788 				cpumask_and(dp, dp, housekeeping_cpumask(HK_FLAG_DOMAIN));
789 				if (dattr)
790 					update_domain_attr_tree(dattr + nslot, b);
791 
792 				/* Done with this partition */
793 				b->pn = -1;
794 			}
795 		}
796 		nslot++;
797 	}
798 	BUG_ON(nslot != ndoms);
799 
800 done:
801 	kfree(csa);
802 
803 	/*
804 	 * Fallback to the default domain if kmalloc() failed.
805 	 * See comments in partition_sched_domains().
806 	 */
807 	if (doms == NULL)
808 		ndoms = 1;
809 
810 	*domains    = doms;
811 	*attributes = dattr;
812 	return ndoms;
813 }
814 
815 /*
816  * Rebuild scheduler domains.
817  *
818  * If the flag 'sched_load_balance' of any cpuset with non-empty
819  * 'cpus' changes, or if the 'cpus' allowed changes in any cpuset
820  * which has that flag enabled, or if any cpuset with a non-empty
821  * 'cpus' is removed, then call this routine to rebuild the
822  * scheduler's dynamic sched domains.
823  *
824  * Call with cpuset_mutex held.  Takes get_online_cpus().
825  */
826 static void rebuild_sched_domains_locked(void)
827 {
828 	struct sched_domain_attr *attr;
829 	cpumask_var_t *doms;
830 	int ndoms;
831 
832 	lockdep_assert_held(&cpuset_mutex);
833 	get_online_cpus();
834 
835 	/*
836 	 * We have raced with CPU hotplug. Don't do anything to avoid
837 	 * passing doms with offlined cpu to partition_sched_domains().
838 	 * Anyways, hotplug work item will rebuild sched domains.
839 	 */
840 	if (!cpumask_equal(top_cpuset.effective_cpus, cpu_active_mask))
841 		goto out;
842 
843 	/* Generate domain masks and attrs */
844 	ndoms = generate_sched_domains(&doms, &attr);
845 
846 	/* Have scheduler rebuild the domains */
847 	partition_sched_domains(ndoms, doms, attr);
848 out:
849 	put_online_cpus();
850 }
851 #else /* !CONFIG_SMP */
852 static void rebuild_sched_domains_locked(void)
853 {
854 }
855 #endif /* CONFIG_SMP */
856 
857 void rebuild_sched_domains(void)
858 {
859 	mutex_lock(&cpuset_mutex);
860 	rebuild_sched_domains_locked();
861 	mutex_unlock(&cpuset_mutex);
862 }
863 
864 /**
865  * update_tasks_cpumask - Update the cpumasks of tasks in the cpuset.
866  * @cs: the cpuset in which each task's cpus_allowed mask needs to be changed
867  *
868  * Iterate through each task of @cs updating its cpus_allowed to the
869  * effective cpuset's.  As this function is called with cpuset_mutex held,
870  * cpuset membership stays stable.
871  */
872 static void update_tasks_cpumask(struct cpuset *cs)
873 {
874 	struct css_task_iter it;
875 	struct task_struct *task;
876 
877 	css_task_iter_start(&cs->css, 0, &it);
878 	while ((task = css_task_iter_next(&it)))
879 		set_cpus_allowed_ptr(task, cs->effective_cpus);
880 	css_task_iter_end(&it);
881 }
882 
883 /*
884  * update_cpumasks_hier - Update effective cpumasks and tasks in the subtree
885  * @cs: the cpuset to consider
886  * @new_cpus: temp variable for calculating new effective_cpus
887  *
888  * When congifured cpumask is changed, the effective cpumasks of this cpuset
889  * and all its descendants need to be updated.
890  *
891  * On legacy hierachy, effective_cpus will be the same with cpu_allowed.
892  *
893  * Called with cpuset_mutex held
894  */
895 static void update_cpumasks_hier(struct cpuset *cs, struct cpumask *new_cpus)
896 {
897 	struct cpuset *cp;
898 	struct cgroup_subsys_state *pos_css;
899 	bool need_rebuild_sched_domains = false;
900 
901 	rcu_read_lock();
902 	cpuset_for_each_descendant_pre(cp, pos_css, cs) {
903 		struct cpuset *parent = parent_cs(cp);
904 
905 		cpumask_and(new_cpus, cp->cpus_allowed, parent->effective_cpus);
906 
907 		/*
908 		 * If it becomes empty, inherit the effective mask of the
909 		 * parent, which is guaranteed to have some CPUs.
910 		 */
911 		if (is_in_v2_mode() && cpumask_empty(new_cpus))
912 			cpumask_copy(new_cpus, parent->effective_cpus);
913 
914 		/* Skip the whole subtree if the cpumask remains the same. */
915 		if (cpumask_equal(new_cpus, cp->effective_cpus)) {
916 			pos_css = css_rightmost_descendant(pos_css);
917 			continue;
918 		}
919 
920 		if (!css_tryget_online(&cp->css))
921 			continue;
922 		rcu_read_unlock();
923 
924 		spin_lock_irq(&callback_lock);
925 		cpumask_copy(cp->effective_cpus, new_cpus);
926 		spin_unlock_irq(&callback_lock);
927 
928 		WARN_ON(!is_in_v2_mode() &&
929 			!cpumask_equal(cp->cpus_allowed, cp->effective_cpus));
930 
931 		update_tasks_cpumask(cp);
932 
933 		/*
934 		 * If the effective cpumask of any non-empty cpuset is changed,
935 		 * we need to rebuild sched domains.
936 		 */
937 		if (!cpumask_empty(cp->cpus_allowed) &&
938 		    is_sched_load_balance(cp))
939 			need_rebuild_sched_domains = true;
940 
941 		rcu_read_lock();
942 		css_put(&cp->css);
943 	}
944 	rcu_read_unlock();
945 
946 	if (need_rebuild_sched_domains)
947 		rebuild_sched_domains_locked();
948 }
949 
950 /**
951  * update_cpumask - update the cpus_allowed mask of a cpuset and all tasks in it
952  * @cs: the cpuset to consider
953  * @trialcs: trial cpuset
954  * @buf: buffer of cpu numbers written to this cpuset
955  */
956 static int update_cpumask(struct cpuset *cs, struct cpuset *trialcs,
957 			  const char *buf)
958 {
959 	int retval;
960 
961 	/* top_cpuset.cpus_allowed tracks cpu_online_mask; it's read-only */
962 	if (cs == &top_cpuset)
963 		return -EACCES;
964 
965 	/*
966 	 * An empty cpus_allowed is ok only if the cpuset has no tasks.
967 	 * Since cpulist_parse() fails on an empty mask, we special case
968 	 * that parsing.  The validate_change() call ensures that cpusets
969 	 * with tasks have cpus.
970 	 */
971 	if (!*buf) {
972 		cpumask_clear(trialcs->cpus_allowed);
973 	} else {
974 		retval = cpulist_parse(buf, trialcs->cpus_allowed);
975 		if (retval < 0)
976 			return retval;
977 
978 		if (!cpumask_subset(trialcs->cpus_allowed,
979 				    top_cpuset.cpus_allowed))
980 			return -EINVAL;
981 	}
982 
983 	/* Nothing to do if the cpus didn't change */
984 	if (cpumask_equal(cs->cpus_allowed, trialcs->cpus_allowed))
985 		return 0;
986 
987 	retval = validate_change(cs, trialcs);
988 	if (retval < 0)
989 		return retval;
990 
991 	spin_lock_irq(&callback_lock);
992 	cpumask_copy(cs->cpus_allowed, trialcs->cpus_allowed);
993 	spin_unlock_irq(&callback_lock);
994 
995 	/* use trialcs->cpus_allowed as a temp variable */
996 	update_cpumasks_hier(cs, trialcs->cpus_allowed);
997 	return 0;
998 }
999 
1000 /*
1001  * Migrate memory region from one set of nodes to another.  This is
1002  * performed asynchronously as it can be called from process migration path
1003  * holding locks involved in process management.  All mm migrations are
1004  * performed in the queued order and can be waited for by flushing
1005  * cpuset_migrate_mm_wq.
1006  */
1007 
1008 struct cpuset_migrate_mm_work {
1009 	struct work_struct	work;
1010 	struct mm_struct	*mm;
1011 	nodemask_t		from;
1012 	nodemask_t		to;
1013 };
1014 
1015 static void cpuset_migrate_mm_workfn(struct work_struct *work)
1016 {
1017 	struct cpuset_migrate_mm_work *mwork =
1018 		container_of(work, struct cpuset_migrate_mm_work, work);
1019 
1020 	/* on a wq worker, no need to worry about %current's mems_allowed */
1021 	do_migrate_pages(mwork->mm, &mwork->from, &mwork->to, MPOL_MF_MOVE_ALL);
1022 	mmput(mwork->mm);
1023 	kfree(mwork);
1024 }
1025 
1026 static void cpuset_migrate_mm(struct mm_struct *mm, const nodemask_t *from,
1027 							const nodemask_t *to)
1028 {
1029 	struct cpuset_migrate_mm_work *mwork;
1030 
1031 	mwork = kzalloc(sizeof(*mwork), GFP_KERNEL);
1032 	if (mwork) {
1033 		mwork->mm = mm;
1034 		mwork->from = *from;
1035 		mwork->to = *to;
1036 		INIT_WORK(&mwork->work, cpuset_migrate_mm_workfn);
1037 		queue_work(cpuset_migrate_mm_wq, &mwork->work);
1038 	} else {
1039 		mmput(mm);
1040 	}
1041 }
1042 
1043 static void cpuset_post_attach(void)
1044 {
1045 	flush_workqueue(cpuset_migrate_mm_wq);
1046 }
1047 
1048 /*
1049  * cpuset_change_task_nodemask - change task's mems_allowed and mempolicy
1050  * @tsk: the task to change
1051  * @newmems: new nodes that the task will be set
1052  *
1053  * We use the mems_allowed_seq seqlock to safely update both tsk->mems_allowed
1054  * and rebind an eventual tasks' mempolicy. If the task is allocating in
1055  * parallel, it might temporarily see an empty intersection, which results in
1056  * a seqlock check and retry before OOM or allocation failure.
1057  */
1058 static void cpuset_change_task_nodemask(struct task_struct *tsk,
1059 					nodemask_t *newmems)
1060 {
1061 	task_lock(tsk);
1062 
1063 	local_irq_disable();
1064 	write_seqcount_begin(&tsk->mems_allowed_seq);
1065 
1066 	nodes_or(tsk->mems_allowed, tsk->mems_allowed, *newmems);
1067 	mpol_rebind_task(tsk, newmems);
1068 	tsk->mems_allowed = *newmems;
1069 
1070 	write_seqcount_end(&tsk->mems_allowed_seq);
1071 	local_irq_enable();
1072 
1073 	task_unlock(tsk);
1074 }
1075 
1076 static void *cpuset_being_rebound;
1077 
1078 /**
1079  * update_tasks_nodemask - Update the nodemasks of tasks in the cpuset.
1080  * @cs: the cpuset in which each task's mems_allowed mask needs to be changed
1081  *
1082  * Iterate through each task of @cs updating its mems_allowed to the
1083  * effective cpuset's.  As this function is called with cpuset_mutex held,
1084  * cpuset membership stays stable.
1085  */
1086 static void update_tasks_nodemask(struct cpuset *cs)
1087 {
1088 	static nodemask_t newmems;	/* protected by cpuset_mutex */
1089 	struct css_task_iter it;
1090 	struct task_struct *task;
1091 
1092 	cpuset_being_rebound = cs;		/* causes mpol_dup() rebind */
1093 
1094 	guarantee_online_mems(cs, &newmems);
1095 
1096 	/*
1097 	 * The mpol_rebind_mm() call takes mmap_sem, which we couldn't
1098 	 * take while holding tasklist_lock.  Forks can happen - the
1099 	 * mpol_dup() cpuset_being_rebound check will catch such forks,
1100 	 * and rebind their vma mempolicies too.  Because we still hold
1101 	 * the global cpuset_mutex, we know that no other rebind effort
1102 	 * will be contending for the global variable cpuset_being_rebound.
1103 	 * It's ok if we rebind the same mm twice; mpol_rebind_mm()
1104 	 * is idempotent.  Also migrate pages in each mm to new nodes.
1105 	 */
1106 	css_task_iter_start(&cs->css, 0, &it);
1107 	while ((task = css_task_iter_next(&it))) {
1108 		struct mm_struct *mm;
1109 		bool migrate;
1110 
1111 		cpuset_change_task_nodemask(task, &newmems);
1112 
1113 		mm = get_task_mm(task);
1114 		if (!mm)
1115 			continue;
1116 
1117 		migrate = is_memory_migrate(cs);
1118 
1119 		mpol_rebind_mm(mm, &cs->mems_allowed);
1120 		if (migrate)
1121 			cpuset_migrate_mm(mm, &cs->old_mems_allowed, &newmems);
1122 		else
1123 			mmput(mm);
1124 	}
1125 	css_task_iter_end(&it);
1126 
1127 	/*
1128 	 * All the tasks' nodemasks have been updated, update
1129 	 * cs->old_mems_allowed.
1130 	 */
1131 	cs->old_mems_allowed = newmems;
1132 
1133 	/* We're done rebinding vmas to this cpuset's new mems_allowed. */
1134 	cpuset_being_rebound = NULL;
1135 }
1136 
1137 /*
1138  * update_nodemasks_hier - Update effective nodemasks and tasks in the subtree
1139  * @cs: the cpuset to consider
1140  * @new_mems: a temp variable for calculating new effective_mems
1141  *
1142  * When configured nodemask is changed, the effective nodemasks of this cpuset
1143  * and all its descendants need to be updated.
1144  *
1145  * On legacy hiearchy, effective_mems will be the same with mems_allowed.
1146  *
1147  * Called with cpuset_mutex held
1148  */
1149 static void update_nodemasks_hier(struct cpuset *cs, nodemask_t *new_mems)
1150 {
1151 	struct cpuset *cp;
1152 	struct cgroup_subsys_state *pos_css;
1153 
1154 	rcu_read_lock();
1155 	cpuset_for_each_descendant_pre(cp, pos_css, cs) {
1156 		struct cpuset *parent = parent_cs(cp);
1157 
1158 		nodes_and(*new_mems, cp->mems_allowed, parent->effective_mems);
1159 
1160 		/*
1161 		 * If it becomes empty, inherit the effective mask of the
1162 		 * parent, which is guaranteed to have some MEMs.
1163 		 */
1164 		if (is_in_v2_mode() && nodes_empty(*new_mems))
1165 			*new_mems = parent->effective_mems;
1166 
1167 		/* Skip the whole subtree if the nodemask remains the same. */
1168 		if (nodes_equal(*new_mems, cp->effective_mems)) {
1169 			pos_css = css_rightmost_descendant(pos_css);
1170 			continue;
1171 		}
1172 
1173 		if (!css_tryget_online(&cp->css))
1174 			continue;
1175 		rcu_read_unlock();
1176 
1177 		spin_lock_irq(&callback_lock);
1178 		cp->effective_mems = *new_mems;
1179 		spin_unlock_irq(&callback_lock);
1180 
1181 		WARN_ON(!is_in_v2_mode() &&
1182 			!nodes_equal(cp->mems_allowed, cp->effective_mems));
1183 
1184 		update_tasks_nodemask(cp);
1185 
1186 		rcu_read_lock();
1187 		css_put(&cp->css);
1188 	}
1189 	rcu_read_unlock();
1190 }
1191 
1192 /*
1193  * Handle user request to change the 'mems' memory placement
1194  * of a cpuset.  Needs to validate the request, update the
1195  * cpusets mems_allowed, and for each task in the cpuset,
1196  * update mems_allowed and rebind task's mempolicy and any vma
1197  * mempolicies and if the cpuset is marked 'memory_migrate',
1198  * migrate the tasks pages to the new memory.
1199  *
1200  * Call with cpuset_mutex held. May take callback_lock during call.
1201  * Will take tasklist_lock, scan tasklist for tasks in cpuset cs,
1202  * lock each such tasks mm->mmap_sem, scan its vma's and rebind
1203  * their mempolicies to the cpusets new mems_allowed.
1204  */
1205 static int update_nodemask(struct cpuset *cs, struct cpuset *trialcs,
1206 			   const char *buf)
1207 {
1208 	int retval;
1209 
1210 	/*
1211 	 * top_cpuset.mems_allowed tracks node_stats[N_MEMORY];
1212 	 * it's read-only
1213 	 */
1214 	if (cs == &top_cpuset) {
1215 		retval = -EACCES;
1216 		goto done;
1217 	}
1218 
1219 	/*
1220 	 * An empty mems_allowed is ok iff there are no tasks in the cpuset.
1221 	 * Since nodelist_parse() fails on an empty mask, we special case
1222 	 * that parsing.  The validate_change() call ensures that cpusets
1223 	 * with tasks have memory.
1224 	 */
1225 	if (!*buf) {
1226 		nodes_clear(trialcs->mems_allowed);
1227 	} else {
1228 		retval = nodelist_parse(buf, trialcs->mems_allowed);
1229 		if (retval < 0)
1230 			goto done;
1231 
1232 		if (!nodes_subset(trialcs->mems_allowed,
1233 				  top_cpuset.mems_allowed)) {
1234 			retval = -EINVAL;
1235 			goto done;
1236 		}
1237 	}
1238 
1239 	if (nodes_equal(cs->mems_allowed, trialcs->mems_allowed)) {
1240 		retval = 0;		/* Too easy - nothing to do */
1241 		goto done;
1242 	}
1243 	retval = validate_change(cs, trialcs);
1244 	if (retval < 0)
1245 		goto done;
1246 
1247 	spin_lock_irq(&callback_lock);
1248 	cs->mems_allowed = trialcs->mems_allowed;
1249 	spin_unlock_irq(&callback_lock);
1250 
1251 	/* use trialcs->mems_allowed as a temp variable */
1252 	update_nodemasks_hier(cs, &trialcs->mems_allowed);
1253 done:
1254 	return retval;
1255 }
1256 
1257 int current_cpuset_is_being_rebound(void)
1258 {
1259 	int ret;
1260 
1261 	rcu_read_lock();
1262 	ret = task_cs(current) == cpuset_being_rebound;
1263 	rcu_read_unlock();
1264 
1265 	return ret;
1266 }
1267 
1268 static int update_relax_domain_level(struct cpuset *cs, s64 val)
1269 {
1270 #ifdef CONFIG_SMP
1271 	if (val < -1 || val >= sched_domain_level_max)
1272 		return -EINVAL;
1273 #endif
1274 
1275 	if (val != cs->relax_domain_level) {
1276 		cs->relax_domain_level = val;
1277 		if (!cpumask_empty(cs->cpus_allowed) &&
1278 		    is_sched_load_balance(cs))
1279 			rebuild_sched_domains_locked();
1280 	}
1281 
1282 	return 0;
1283 }
1284 
1285 /**
1286  * update_tasks_flags - update the spread flags of tasks in the cpuset.
1287  * @cs: the cpuset in which each task's spread flags needs to be changed
1288  *
1289  * Iterate through each task of @cs updating its spread flags.  As this
1290  * function is called with cpuset_mutex held, cpuset membership stays
1291  * stable.
1292  */
1293 static void update_tasks_flags(struct cpuset *cs)
1294 {
1295 	struct css_task_iter it;
1296 	struct task_struct *task;
1297 
1298 	css_task_iter_start(&cs->css, 0, &it);
1299 	while ((task = css_task_iter_next(&it)))
1300 		cpuset_update_task_spread_flag(cs, task);
1301 	css_task_iter_end(&it);
1302 }
1303 
1304 /*
1305  * update_flag - read a 0 or a 1 in a file and update associated flag
1306  * bit:		the bit to update (see cpuset_flagbits_t)
1307  * cs:		the cpuset to update
1308  * turning_on: 	whether the flag is being set or cleared
1309  *
1310  * Call with cpuset_mutex held.
1311  */
1312 
1313 static int update_flag(cpuset_flagbits_t bit, struct cpuset *cs,
1314 		       int turning_on)
1315 {
1316 	struct cpuset *trialcs;
1317 	int balance_flag_changed;
1318 	int spread_flag_changed;
1319 	int err;
1320 
1321 	trialcs = alloc_trial_cpuset(cs);
1322 	if (!trialcs)
1323 		return -ENOMEM;
1324 
1325 	if (turning_on)
1326 		set_bit(bit, &trialcs->flags);
1327 	else
1328 		clear_bit(bit, &trialcs->flags);
1329 
1330 	err = validate_change(cs, trialcs);
1331 	if (err < 0)
1332 		goto out;
1333 
1334 	balance_flag_changed = (is_sched_load_balance(cs) !=
1335 				is_sched_load_balance(trialcs));
1336 
1337 	spread_flag_changed = ((is_spread_slab(cs) != is_spread_slab(trialcs))
1338 			|| (is_spread_page(cs) != is_spread_page(trialcs)));
1339 
1340 	spin_lock_irq(&callback_lock);
1341 	cs->flags = trialcs->flags;
1342 	spin_unlock_irq(&callback_lock);
1343 
1344 	if (!cpumask_empty(trialcs->cpus_allowed) && balance_flag_changed)
1345 		rebuild_sched_domains_locked();
1346 
1347 	if (spread_flag_changed)
1348 		update_tasks_flags(cs);
1349 out:
1350 	free_trial_cpuset(trialcs);
1351 	return err;
1352 }
1353 
1354 /*
1355  * Frequency meter - How fast is some event occurring?
1356  *
1357  * These routines manage a digitally filtered, constant time based,
1358  * event frequency meter.  There are four routines:
1359  *   fmeter_init() - initialize a frequency meter.
1360  *   fmeter_markevent() - called each time the event happens.
1361  *   fmeter_getrate() - returns the recent rate of such events.
1362  *   fmeter_update() - internal routine used to update fmeter.
1363  *
1364  * A common data structure is passed to each of these routines,
1365  * which is used to keep track of the state required to manage the
1366  * frequency meter and its digital filter.
1367  *
1368  * The filter works on the number of events marked per unit time.
1369  * The filter is single-pole low-pass recursive (IIR).  The time unit
1370  * is 1 second.  Arithmetic is done using 32-bit integers scaled to
1371  * simulate 3 decimal digits of precision (multiplied by 1000).
1372  *
1373  * With an FM_COEF of 933, and a time base of 1 second, the filter
1374  * has a half-life of 10 seconds, meaning that if the events quit
1375  * happening, then the rate returned from the fmeter_getrate()
1376  * will be cut in half each 10 seconds, until it converges to zero.
1377  *
1378  * It is not worth doing a real infinitely recursive filter.  If more
1379  * than FM_MAXTICKS ticks have elapsed since the last filter event,
1380  * just compute FM_MAXTICKS ticks worth, by which point the level
1381  * will be stable.
1382  *
1383  * Limit the count of unprocessed events to FM_MAXCNT, so as to avoid
1384  * arithmetic overflow in the fmeter_update() routine.
1385  *
1386  * Given the simple 32 bit integer arithmetic used, this meter works
1387  * best for reporting rates between one per millisecond (msec) and
1388  * one per 32 (approx) seconds.  At constant rates faster than one
1389  * per msec it maxes out at values just under 1,000,000.  At constant
1390  * rates between one per msec, and one per second it will stabilize
1391  * to a value N*1000, where N is the rate of events per second.
1392  * At constant rates between one per second and one per 32 seconds,
1393  * it will be choppy, moving up on the seconds that have an event,
1394  * and then decaying until the next event.  At rates slower than
1395  * about one in 32 seconds, it decays all the way back to zero between
1396  * each event.
1397  */
1398 
1399 #define FM_COEF 933		/* coefficient for half-life of 10 secs */
1400 #define FM_MAXTICKS ((u32)99)   /* useless computing more ticks than this */
1401 #define FM_MAXCNT 1000000	/* limit cnt to avoid overflow */
1402 #define FM_SCALE 1000		/* faux fixed point scale */
1403 
1404 /* Initialize a frequency meter */
1405 static void fmeter_init(struct fmeter *fmp)
1406 {
1407 	fmp->cnt = 0;
1408 	fmp->val = 0;
1409 	fmp->time = 0;
1410 	spin_lock_init(&fmp->lock);
1411 }
1412 
1413 /* Internal meter update - process cnt events and update value */
1414 static void fmeter_update(struct fmeter *fmp)
1415 {
1416 	time64_t now;
1417 	u32 ticks;
1418 
1419 	now = ktime_get_seconds();
1420 	ticks = now - fmp->time;
1421 
1422 	if (ticks == 0)
1423 		return;
1424 
1425 	ticks = min(FM_MAXTICKS, ticks);
1426 	while (ticks-- > 0)
1427 		fmp->val = (FM_COEF * fmp->val) / FM_SCALE;
1428 	fmp->time = now;
1429 
1430 	fmp->val += ((FM_SCALE - FM_COEF) * fmp->cnt) / FM_SCALE;
1431 	fmp->cnt = 0;
1432 }
1433 
1434 /* Process any previous ticks, then bump cnt by one (times scale). */
1435 static void fmeter_markevent(struct fmeter *fmp)
1436 {
1437 	spin_lock(&fmp->lock);
1438 	fmeter_update(fmp);
1439 	fmp->cnt = min(FM_MAXCNT, fmp->cnt + FM_SCALE);
1440 	spin_unlock(&fmp->lock);
1441 }
1442 
1443 /* Process any previous ticks, then return current value. */
1444 static int fmeter_getrate(struct fmeter *fmp)
1445 {
1446 	int val;
1447 
1448 	spin_lock(&fmp->lock);
1449 	fmeter_update(fmp);
1450 	val = fmp->val;
1451 	spin_unlock(&fmp->lock);
1452 	return val;
1453 }
1454 
1455 static struct cpuset *cpuset_attach_old_cs;
1456 
1457 /* Called by cgroups to determine if a cpuset is usable; cpuset_mutex held */
1458 static int cpuset_can_attach(struct cgroup_taskset *tset)
1459 {
1460 	struct cgroup_subsys_state *css;
1461 	struct cpuset *cs;
1462 	struct task_struct *task;
1463 	int ret;
1464 
1465 	/* used later by cpuset_attach() */
1466 	cpuset_attach_old_cs = task_cs(cgroup_taskset_first(tset, &css));
1467 	cs = css_cs(css);
1468 
1469 	mutex_lock(&cpuset_mutex);
1470 
1471 	/* allow moving tasks into an empty cpuset if on default hierarchy */
1472 	ret = -ENOSPC;
1473 	if (!is_in_v2_mode() &&
1474 	    (cpumask_empty(cs->cpus_allowed) || nodes_empty(cs->mems_allowed)))
1475 		goto out_unlock;
1476 
1477 	cgroup_taskset_for_each(task, css, tset) {
1478 		ret = task_can_attach(task, cs->cpus_allowed);
1479 		if (ret)
1480 			goto out_unlock;
1481 		ret = security_task_setscheduler(task);
1482 		if (ret)
1483 			goto out_unlock;
1484 	}
1485 
1486 	/*
1487 	 * Mark attach is in progress.  This makes validate_change() fail
1488 	 * changes which zero cpus/mems_allowed.
1489 	 */
1490 	cs->attach_in_progress++;
1491 	ret = 0;
1492 out_unlock:
1493 	mutex_unlock(&cpuset_mutex);
1494 	return ret;
1495 }
1496 
1497 static void cpuset_cancel_attach(struct cgroup_taskset *tset)
1498 {
1499 	struct cgroup_subsys_state *css;
1500 	struct cpuset *cs;
1501 
1502 	cgroup_taskset_first(tset, &css);
1503 	cs = css_cs(css);
1504 
1505 	mutex_lock(&cpuset_mutex);
1506 	css_cs(css)->attach_in_progress--;
1507 	mutex_unlock(&cpuset_mutex);
1508 }
1509 
1510 /*
1511  * Protected by cpuset_mutex.  cpus_attach is used only by cpuset_attach()
1512  * but we can't allocate it dynamically there.  Define it global and
1513  * allocate from cpuset_init().
1514  */
1515 static cpumask_var_t cpus_attach;
1516 
1517 static void cpuset_attach(struct cgroup_taskset *tset)
1518 {
1519 	/* static buf protected by cpuset_mutex */
1520 	static nodemask_t cpuset_attach_nodemask_to;
1521 	struct task_struct *task;
1522 	struct task_struct *leader;
1523 	struct cgroup_subsys_state *css;
1524 	struct cpuset *cs;
1525 	struct cpuset *oldcs = cpuset_attach_old_cs;
1526 
1527 	cgroup_taskset_first(tset, &css);
1528 	cs = css_cs(css);
1529 
1530 	mutex_lock(&cpuset_mutex);
1531 
1532 	/* prepare for attach */
1533 	if (cs == &top_cpuset)
1534 		cpumask_copy(cpus_attach, cpu_possible_mask);
1535 	else
1536 		guarantee_online_cpus(cs, cpus_attach);
1537 
1538 	guarantee_online_mems(cs, &cpuset_attach_nodemask_to);
1539 
1540 	cgroup_taskset_for_each(task, css, tset) {
1541 		/*
1542 		 * can_attach beforehand should guarantee that this doesn't
1543 		 * fail.  TODO: have a better way to handle failure here
1544 		 */
1545 		WARN_ON_ONCE(set_cpus_allowed_ptr(task, cpus_attach));
1546 
1547 		cpuset_change_task_nodemask(task, &cpuset_attach_nodemask_to);
1548 		cpuset_update_task_spread_flag(cs, task);
1549 	}
1550 
1551 	/*
1552 	 * Change mm for all threadgroup leaders. This is expensive and may
1553 	 * sleep and should be moved outside migration path proper.
1554 	 */
1555 	cpuset_attach_nodemask_to = cs->effective_mems;
1556 	cgroup_taskset_for_each_leader(leader, css, tset) {
1557 		struct mm_struct *mm = get_task_mm(leader);
1558 
1559 		if (mm) {
1560 			mpol_rebind_mm(mm, &cpuset_attach_nodemask_to);
1561 
1562 			/*
1563 			 * old_mems_allowed is the same with mems_allowed
1564 			 * here, except if this task is being moved
1565 			 * automatically due to hotplug.  In that case
1566 			 * @mems_allowed has been updated and is empty, so
1567 			 * @old_mems_allowed is the right nodesets that we
1568 			 * migrate mm from.
1569 			 */
1570 			if (is_memory_migrate(cs))
1571 				cpuset_migrate_mm(mm, &oldcs->old_mems_allowed,
1572 						  &cpuset_attach_nodemask_to);
1573 			else
1574 				mmput(mm);
1575 		}
1576 	}
1577 
1578 	cs->old_mems_allowed = cpuset_attach_nodemask_to;
1579 
1580 	cs->attach_in_progress--;
1581 	if (!cs->attach_in_progress)
1582 		wake_up(&cpuset_attach_wq);
1583 
1584 	mutex_unlock(&cpuset_mutex);
1585 }
1586 
1587 /* The various types of files and directories in a cpuset file system */
1588 
1589 typedef enum {
1590 	FILE_MEMORY_MIGRATE,
1591 	FILE_CPULIST,
1592 	FILE_MEMLIST,
1593 	FILE_EFFECTIVE_CPULIST,
1594 	FILE_EFFECTIVE_MEMLIST,
1595 	FILE_CPU_EXCLUSIVE,
1596 	FILE_MEM_EXCLUSIVE,
1597 	FILE_MEM_HARDWALL,
1598 	FILE_SCHED_LOAD_BALANCE,
1599 	FILE_SCHED_RELAX_DOMAIN_LEVEL,
1600 	FILE_MEMORY_PRESSURE_ENABLED,
1601 	FILE_MEMORY_PRESSURE,
1602 	FILE_SPREAD_PAGE,
1603 	FILE_SPREAD_SLAB,
1604 } cpuset_filetype_t;
1605 
1606 static int cpuset_write_u64(struct cgroup_subsys_state *css, struct cftype *cft,
1607 			    u64 val)
1608 {
1609 	struct cpuset *cs = css_cs(css);
1610 	cpuset_filetype_t type = cft->private;
1611 	int retval = 0;
1612 
1613 	mutex_lock(&cpuset_mutex);
1614 	if (!is_cpuset_online(cs)) {
1615 		retval = -ENODEV;
1616 		goto out_unlock;
1617 	}
1618 
1619 	switch (type) {
1620 	case FILE_CPU_EXCLUSIVE:
1621 		retval = update_flag(CS_CPU_EXCLUSIVE, cs, val);
1622 		break;
1623 	case FILE_MEM_EXCLUSIVE:
1624 		retval = update_flag(CS_MEM_EXCLUSIVE, cs, val);
1625 		break;
1626 	case FILE_MEM_HARDWALL:
1627 		retval = update_flag(CS_MEM_HARDWALL, cs, val);
1628 		break;
1629 	case FILE_SCHED_LOAD_BALANCE:
1630 		retval = update_flag(CS_SCHED_LOAD_BALANCE, cs, val);
1631 		break;
1632 	case FILE_MEMORY_MIGRATE:
1633 		retval = update_flag(CS_MEMORY_MIGRATE, cs, val);
1634 		break;
1635 	case FILE_MEMORY_PRESSURE_ENABLED:
1636 		cpuset_memory_pressure_enabled = !!val;
1637 		break;
1638 	case FILE_SPREAD_PAGE:
1639 		retval = update_flag(CS_SPREAD_PAGE, cs, val);
1640 		break;
1641 	case FILE_SPREAD_SLAB:
1642 		retval = update_flag(CS_SPREAD_SLAB, cs, val);
1643 		break;
1644 	default:
1645 		retval = -EINVAL;
1646 		break;
1647 	}
1648 out_unlock:
1649 	mutex_unlock(&cpuset_mutex);
1650 	return retval;
1651 }
1652 
1653 static int cpuset_write_s64(struct cgroup_subsys_state *css, struct cftype *cft,
1654 			    s64 val)
1655 {
1656 	struct cpuset *cs = css_cs(css);
1657 	cpuset_filetype_t type = cft->private;
1658 	int retval = -ENODEV;
1659 
1660 	mutex_lock(&cpuset_mutex);
1661 	if (!is_cpuset_online(cs))
1662 		goto out_unlock;
1663 
1664 	switch (type) {
1665 	case FILE_SCHED_RELAX_DOMAIN_LEVEL:
1666 		retval = update_relax_domain_level(cs, val);
1667 		break;
1668 	default:
1669 		retval = -EINVAL;
1670 		break;
1671 	}
1672 out_unlock:
1673 	mutex_unlock(&cpuset_mutex);
1674 	return retval;
1675 }
1676 
1677 /*
1678  * Common handling for a write to a "cpus" or "mems" file.
1679  */
1680 static ssize_t cpuset_write_resmask(struct kernfs_open_file *of,
1681 				    char *buf, size_t nbytes, loff_t off)
1682 {
1683 	struct cpuset *cs = css_cs(of_css(of));
1684 	struct cpuset *trialcs;
1685 	int retval = -ENODEV;
1686 
1687 	buf = strstrip(buf);
1688 
1689 	/*
1690 	 * CPU or memory hotunplug may leave @cs w/o any execution
1691 	 * resources, in which case the hotplug code asynchronously updates
1692 	 * configuration and transfers all tasks to the nearest ancestor
1693 	 * which can execute.
1694 	 *
1695 	 * As writes to "cpus" or "mems" may restore @cs's execution
1696 	 * resources, wait for the previously scheduled operations before
1697 	 * proceeding, so that we don't end up keep removing tasks added
1698 	 * after execution capability is restored.
1699 	 *
1700 	 * cpuset_hotplug_work calls back into cgroup core via
1701 	 * cgroup_transfer_tasks() and waiting for it from a cgroupfs
1702 	 * operation like this one can lead to a deadlock through kernfs
1703 	 * active_ref protection.  Let's break the protection.  Losing the
1704 	 * protection is okay as we check whether @cs is online after
1705 	 * grabbing cpuset_mutex anyway.  This only happens on the legacy
1706 	 * hierarchies.
1707 	 */
1708 	css_get(&cs->css);
1709 	kernfs_break_active_protection(of->kn);
1710 	flush_work(&cpuset_hotplug_work);
1711 
1712 	mutex_lock(&cpuset_mutex);
1713 	if (!is_cpuset_online(cs))
1714 		goto out_unlock;
1715 
1716 	trialcs = alloc_trial_cpuset(cs);
1717 	if (!trialcs) {
1718 		retval = -ENOMEM;
1719 		goto out_unlock;
1720 	}
1721 
1722 	switch (of_cft(of)->private) {
1723 	case FILE_CPULIST:
1724 		retval = update_cpumask(cs, trialcs, buf);
1725 		break;
1726 	case FILE_MEMLIST:
1727 		retval = update_nodemask(cs, trialcs, buf);
1728 		break;
1729 	default:
1730 		retval = -EINVAL;
1731 		break;
1732 	}
1733 
1734 	free_trial_cpuset(trialcs);
1735 out_unlock:
1736 	mutex_unlock(&cpuset_mutex);
1737 	kernfs_unbreak_active_protection(of->kn);
1738 	css_put(&cs->css);
1739 	flush_workqueue(cpuset_migrate_mm_wq);
1740 	return retval ?: nbytes;
1741 }
1742 
1743 /*
1744  * These ascii lists should be read in a single call, by using a user
1745  * buffer large enough to hold the entire map.  If read in smaller
1746  * chunks, there is no guarantee of atomicity.  Since the display format
1747  * used, list of ranges of sequential numbers, is variable length,
1748  * and since these maps can change value dynamically, one could read
1749  * gibberish by doing partial reads while a list was changing.
1750  */
1751 static int cpuset_common_seq_show(struct seq_file *sf, void *v)
1752 {
1753 	struct cpuset *cs = css_cs(seq_css(sf));
1754 	cpuset_filetype_t type = seq_cft(sf)->private;
1755 	int ret = 0;
1756 
1757 	spin_lock_irq(&callback_lock);
1758 
1759 	switch (type) {
1760 	case FILE_CPULIST:
1761 		seq_printf(sf, "%*pbl\n", cpumask_pr_args(cs->cpus_allowed));
1762 		break;
1763 	case FILE_MEMLIST:
1764 		seq_printf(sf, "%*pbl\n", nodemask_pr_args(&cs->mems_allowed));
1765 		break;
1766 	case FILE_EFFECTIVE_CPULIST:
1767 		seq_printf(sf, "%*pbl\n", cpumask_pr_args(cs->effective_cpus));
1768 		break;
1769 	case FILE_EFFECTIVE_MEMLIST:
1770 		seq_printf(sf, "%*pbl\n", nodemask_pr_args(&cs->effective_mems));
1771 		break;
1772 	default:
1773 		ret = -EINVAL;
1774 	}
1775 
1776 	spin_unlock_irq(&callback_lock);
1777 	return ret;
1778 }
1779 
1780 static u64 cpuset_read_u64(struct cgroup_subsys_state *css, struct cftype *cft)
1781 {
1782 	struct cpuset *cs = css_cs(css);
1783 	cpuset_filetype_t type = cft->private;
1784 	switch (type) {
1785 	case FILE_CPU_EXCLUSIVE:
1786 		return is_cpu_exclusive(cs);
1787 	case FILE_MEM_EXCLUSIVE:
1788 		return is_mem_exclusive(cs);
1789 	case FILE_MEM_HARDWALL:
1790 		return is_mem_hardwall(cs);
1791 	case FILE_SCHED_LOAD_BALANCE:
1792 		return is_sched_load_balance(cs);
1793 	case FILE_MEMORY_MIGRATE:
1794 		return is_memory_migrate(cs);
1795 	case FILE_MEMORY_PRESSURE_ENABLED:
1796 		return cpuset_memory_pressure_enabled;
1797 	case FILE_MEMORY_PRESSURE:
1798 		return fmeter_getrate(&cs->fmeter);
1799 	case FILE_SPREAD_PAGE:
1800 		return is_spread_page(cs);
1801 	case FILE_SPREAD_SLAB:
1802 		return is_spread_slab(cs);
1803 	default:
1804 		BUG();
1805 	}
1806 
1807 	/* Unreachable but makes gcc happy */
1808 	return 0;
1809 }
1810 
1811 static s64 cpuset_read_s64(struct cgroup_subsys_state *css, struct cftype *cft)
1812 {
1813 	struct cpuset *cs = css_cs(css);
1814 	cpuset_filetype_t type = cft->private;
1815 	switch (type) {
1816 	case FILE_SCHED_RELAX_DOMAIN_LEVEL:
1817 		return cs->relax_domain_level;
1818 	default:
1819 		BUG();
1820 	}
1821 
1822 	/* Unrechable but makes gcc happy */
1823 	return 0;
1824 }
1825 
1826 
1827 /*
1828  * for the common functions, 'private' gives the type of file
1829  */
1830 
1831 static struct cftype files[] = {
1832 	{
1833 		.name = "cpus",
1834 		.seq_show = cpuset_common_seq_show,
1835 		.write = cpuset_write_resmask,
1836 		.max_write_len = (100U + 6 * NR_CPUS),
1837 		.private = FILE_CPULIST,
1838 	},
1839 
1840 	{
1841 		.name = "mems",
1842 		.seq_show = cpuset_common_seq_show,
1843 		.write = cpuset_write_resmask,
1844 		.max_write_len = (100U + 6 * MAX_NUMNODES),
1845 		.private = FILE_MEMLIST,
1846 	},
1847 
1848 	{
1849 		.name = "effective_cpus",
1850 		.seq_show = cpuset_common_seq_show,
1851 		.private = FILE_EFFECTIVE_CPULIST,
1852 	},
1853 
1854 	{
1855 		.name = "effective_mems",
1856 		.seq_show = cpuset_common_seq_show,
1857 		.private = FILE_EFFECTIVE_MEMLIST,
1858 	},
1859 
1860 	{
1861 		.name = "cpu_exclusive",
1862 		.read_u64 = cpuset_read_u64,
1863 		.write_u64 = cpuset_write_u64,
1864 		.private = FILE_CPU_EXCLUSIVE,
1865 	},
1866 
1867 	{
1868 		.name = "mem_exclusive",
1869 		.read_u64 = cpuset_read_u64,
1870 		.write_u64 = cpuset_write_u64,
1871 		.private = FILE_MEM_EXCLUSIVE,
1872 	},
1873 
1874 	{
1875 		.name = "mem_hardwall",
1876 		.read_u64 = cpuset_read_u64,
1877 		.write_u64 = cpuset_write_u64,
1878 		.private = FILE_MEM_HARDWALL,
1879 	},
1880 
1881 	{
1882 		.name = "sched_load_balance",
1883 		.read_u64 = cpuset_read_u64,
1884 		.write_u64 = cpuset_write_u64,
1885 		.private = FILE_SCHED_LOAD_BALANCE,
1886 	},
1887 
1888 	{
1889 		.name = "sched_relax_domain_level",
1890 		.read_s64 = cpuset_read_s64,
1891 		.write_s64 = cpuset_write_s64,
1892 		.private = FILE_SCHED_RELAX_DOMAIN_LEVEL,
1893 	},
1894 
1895 	{
1896 		.name = "memory_migrate",
1897 		.read_u64 = cpuset_read_u64,
1898 		.write_u64 = cpuset_write_u64,
1899 		.private = FILE_MEMORY_MIGRATE,
1900 	},
1901 
1902 	{
1903 		.name = "memory_pressure",
1904 		.read_u64 = cpuset_read_u64,
1905 		.private = FILE_MEMORY_PRESSURE,
1906 	},
1907 
1908 	{
1909 		.name = "memory_spread_page",
1910 		.read_u64 = cpuset_read_u64,
1911 		.write_u64 = cpuset_write_u64,
1912 		.private = FILE_SPREAD_PAGE,
1913 	},
1914 
1915 	{
1916 		.name = "memory_spread_slab",
1917 		.read_u64 = cpuset_read_u64,
1918 		.write_u64 = cpuset_write_u64,
1919 		.private = FILE_SPREAD_SLAB,
1920 	},
1921 
1922 	{
1923 		.name = "memory_pressure_enabled",
1924 		.flags = CFTYPE_ONLY_ON_ROOT,
1925 		.read_u64 = cpuset_read_u64,
1926 		.write_u64 = cpuset_write_u64,
1927 		.private = FILE_MEMORY_PRESSURE_ENABLED,
1928 	},
1929 
1930 	{ }	/* terminate */
1931 };
1932 
1933 /*
1934  *	cpuset_css_alloc - allocate a cpuset css
1935  *	cgrp:	control group that the new cpuset will be part of
1936  */
1937 
1938 static struct cgroup_subsys_state *
1939 cpuset_css_alloc(struct cgroup_subsys_state *parent_css)
1940 {
1941 	struct cpuset *cs;
1942 
1943 	if (!parent_css)
1944 		return &top_cpuset.css;
1945 
1946 	cs = kzalloc(sizeof(*cs), GFP_KERNEL);
1947 	if (!cs)
1948 		return ERR_PTR(-ENOMEM);
1949 	if (!alloc_cpumask_var(&cs->cpus_allowed, GFP_KERNEL))
1950 		goto free_cs;
1951 	if (!alloc_cpumask_var(&cs->effective_cpus, GFP_KERNEL))
1952 		goto free_cpus;
1953 
1954 	set_bit(CS_SCHED_LOAD_BALANCE, &cs->flags);
1955 	cpumask_clear(cs->cpus_allowed);
1956 	nodes_clear(cs->mems_allowed);
1957 	cpumask_clear(cs->effective_cpus);
1958 	nodes_clear(cs->effective_mems);
1959 	fmeter_init(&cs->fmeter);
1960 	cs->relax_domain_level = -1;
1961 
1962 	return &cs->css;
1963 
1964 free_cpus:
1965 	free_cpumask_var(cs->cpus_allowed);
1966 free_cs:
1967 	kfree(cs);
1968 	return ERR_PTR(-ENOMEM);
1969 }
1970 
1971 static int cpuset_css_online(struct cgroup_subsys_state *css)
1972 {
1973 	struct cpuset *cs = css_cs(css);
1974 	struct cpuset *parent = parent_cs(cs);
1975 	struct cpuset *tmp_cs;
1976 	struct cgroup_subsys_state *pos_css;
1977 
1978 	if (!parent)
1979 		return 0;
1980 
1981 	mutex_lock(&cpuset_mutex);
1982 
1983 	set_bit(CS_ONLINE, &cs->flags);
1984 	if (is_spread_page(parent))
1985 		set_bit(CS_SPREAD_PAGE, &cs->flags);
1986 	if (is_spread_slab(parent))
1987 		set_bit(CS_SPREAD_SLAB, &cs->flags);
1988 
1989 	cpuset_inc();
1990 
1991 	spin_lock_irq(&callback_lock);
1992 	if (is_in_v2_mode()) {
1993 		cpumask_copy(cs->effective_cpus, parent->effective_cpus);
1994 		cs->effective_mems = parent->effective_mems;
1995 	}
1996 	spin_unlock_irq(&callback_lock);
1997 
1998 	if (!test_bit(CGRP_CPUSET_CLONE_CHILDREN, &css->cgroup->flags))
1999 		goto out_unlock;
2000 
2001 	/*
2002 	 * Clone @parent's configuration if CGRP_CPUSET_CLONE_CHILDREN is
2003 	 * set.  This flag handling is implemented in cgroup core for
2004 	 * histrical reasons - the flag may be specified during mount.
2005 	 *
2006 	 * Currently, if any sibling cpusets have exclusive cpus or mem, we
2007 	 * refuse to clone the configuration - thereby refusing the task to
2008 	 * be entered, and as a result refusing the sys_unshare() or
2009 	 * clone() which initiated it.  If this becomes a problem for some
2010 	 * users who wish to allow that scenario, then this could be
2011 	 * changed to grant parent->cpus_allowed-sibling_cpus_exclusive
2012 	 * (and likewise for mems) to the new cgroup.
2013 	 */
2014 	rcu_read_lock();
2015 	cpuset_for_each_child(tmp_cs, pos_css, parent) {
2016 		if (is_mem_exclusive(tmp_cs) || is_cpu_exclusive(tmp_cs)) {
2017 			rcu_read_unlock();
2018 			goto out_unlock;
2019 		}
2020 	}
2021 	rcu_read_unlock();
2022 
2023 	spin_lock_irq(&callback_lock);
2024 	cs->mems_allowed = parent->mems_allowed;
2025 	cs->effective_mems = parent->mems_allowed;
2026 	cpumask_copy(cs->cpus_allowed, parent->cpus_allowed);
2027 	cpumask_copy(cs->effective_cpus, parent->cpus_allowed);
2028 	spin_unlock_irq(&callback_lock);
2029 out_unlock:
2030 	mutex_unlock(&cpuset_mutex);
2031 	return 0;
2032 }
2033 
2034 /*
2035  * If the cpuset being removed has its flag 'sched_load_balance'
2036  * enabled, then simulate turning sched_load_balance off, which
2037  * will call rebuild_sched_domains_locked().
2038  */
2039 
2040 static void cpuset_css_offline(struct cgroup_subsys_state *css)
2041 {
2042 	struct cpuset *cs = css_cs(css);
2043 
2044 	mutex_lock(&cpuset_mutex);
2045 
2046 	if (is_sched_load_balance(cs))
2047 		update_flag(CS_SCHED_LOAD_BALANCE, cs, 0);
2048 
2049 	cpuset_dec();
2050 	clear_bit(CS_ONLINE, &cs->flags);
2051 
2052 	mutex_unlock(&cpuset_mutex);
2053 }
2054 
2055 static void cpuset_css_free(struct cgroup_subsys_state *css)
2056 {
2057 	struct cpuset *cs = css_cs(css);
2058 
2059 	free_cpumask_var(cs->effective_cpus);
2060 	free_cpumask_var(cs->cpus_allowed);
2061 	kfree(cs);
2062 }
2063 
2064 static void cpuset_bind(struct cgroup_subsys_state *root_css)
2065 {
2066 	mutex_lock(&cpuset_mutex);
2067 	spin_lock_irq(&callback_lock);
2068 
2069 	if (is_in_v2_mode()) {
2070 		cpumask_copy(top_cpuset.cpus_allowed, cpu_possible_mask);
2071 		top_cpuset.mems_allowed = node_possible_map;
2072 	} else {
2073 		cpumask_copy(top_cpuset.cpus_allowed,
2074 			     top_cpuset.effective_cpus);
2075 		top_cpuset.mems_allowed = top_cpuset.effective_mems;
2076 	}
2077 
2078 	spin_unlock_irq(&callback_lock);
2079 	mutex_unlock(&cpuset_mutex);
2080 }
2081 
2082 /*
2083  * Make sure the new task conform to the current state of its parent,
2084  * which could have been changed by cpuset just after it inherits the
2085  * state from the parent and before it sits on the cgroup's task list.
2086  */
2087 static void cpuset_fork(struct task_struct *task)
2088 {
2089 	if (task_css_is_root(task, cpuset_cgrp_id))
2090 		return;
2091 
2092 	set_cpus_allowed_ptr(task, &current->cpus_allowed);
2093 	task->mems_allowed = current->mems_allowed;
2094 }
2095 
2096 struct cgroup_subsys cpuset_cgrp_subsys = {
2097 	.css_alloc	= cpuset_css_alloc,
2098 	.css_online	= cpuset_css_online,
2099 	.css_offline	= cpuset_css_offline,
2100 	.css_free	= cpuset_css_free,
2101 	.can_attach	= cpuset_can_attach,
2102 	.cancel_attach	= cpuset_cancel_attach,
2103 	.attach		= cpuset_attach,
2104 	.post_attach	= cpuset_post_attach,
2105 	.bind		= cpuset_bind,
2106 	.fork		= cpuset_fork,
2107 	.legacy_cftypes	= files,
2108 	.early_init	= true,
2109 };
2110 
2111 /**
2112  * cpuset_init - initialize cpusets at system boot
2113  *
2114  * Description: Initialize top_cpuset and the cpuset internal file system,
2115  **/
2116 
2117 int __init cpuset_init(void)
2118 {
2119 	int err = 0;
2120 
2121 	BUG_ON(!alloc_cpumask_var(&top_cpuset.cpus_allowed, GFP_KERNEL));
2122 	BUG_ON(!alloc_cpumask_var(&top_cpuset.effective_cpus, GFP_KERNEL));
2123 
2124 	cpumask_setall(top_cpuset.cpus_allowed);
2125 	nodes_setall(top_cpuset.mems_allowed);
2126 	cpumask_setall(top_cpuset.effective_cpus);
2127 	nodes_setall(top_cpuset.effective_mems);
2128 
2129 	fmeter_init(&top_cpuset.fmeter);
2130 	set_bit(CS_SCHED_LOAD_BALANCE, &top_cpuset.flags);
2131 	top_cpuset.relax_domain_level = -1;
2132 
2133 	err = register_filesystem(&cpuset_fs_type);
2134 	if (err < 0)
2135 		return err;
2136 
2137 	BUG_ON(!alloc_cpumask_var(&cpus_attach, GFP_KERNEL));
2138 
2139 	return 0;
2140 }
2141 
2142 /*
2143  * If CPU and/or memory hotplug handlers, below, unplug any CPUs
2144  * or memory nodes, we need to walk over the cpuset hierarchy,
2145  * removing that CPU or node from all cpusets.  If this removes the
2146  * last CPU or node from a cpuset, then move the tasks in the empty
2147  * cpuset to its next-highest non-empty parent.
2148  */
2149 static void remove_tasks_in_empty_cpuset(struct cpuset *cs)
2150 {
2151 	struct cpuset *parent;
2152 
2153 	/*
2154 	 * Find its next-highest non-empty parent, (top cpuset
2155 	 * has online cpus, so can't be empty).
2156 	 */
2157 	parent = parent_cs(cs);
2158 	while (cpumask_empty(parent->cpus_allowed) ||
2159 			nodes_empty(parent->mems_allowed))
2160 		parent = parent_cs(parent);
2161 
2162 	if (cgroup_transfer_tasks(parent->css.cgroup, cs->css.cgroup)) {
2163 		pr_err("cpuset: failed to transfer tasks out of empty cpuset ");
2164 		pr_cont_cgroup_name(cs->css.cgroup);
2165 		pr_cont("\n");
2166 	}
2167 }
2168 
2169 static void
2170 hotplug_update_tasks_legacy(struct cpuset *cs,
2171 			    struct cpumask *new_cpus, nodemask_t *new_mems,
2172 			    bool cpus_updated, bool mems_updated)
2173 {
2174 	bool is_empty;
2175 
2176 	spin_lock_irq(&callback_lock);
2177 	cpumask_copy(cs->cpus_allowed, new_cpus);
2178 	cpumask_copy(cs->effective_cpus, new_cpus);
2179 	cs->mems_allowed = *new_mems;
2180 	cs->effective_mems = *new_mems;
2181 	spin_unlock_irq(&callback_lock);
2182 
2183 	/*
2184 	 * Don't call update_tasks_cpumask() if the cpuset becomes empty,
2185 	 * as the tasks will be migratecd to an ancestor.
2186 	 */
2187 	if (cpus_updated && !cpumask_empty(cs->cpus_allowed))
2188 		update_tasks_cpumask(cs);
2189 	if (mems_updated && !nodes_empty(cs->mems_allowed))
2190 		update_tasks_nodemask(cs);
2191 
2192 	is_empty = cpumask_empty(cs->cpus_allowed) ||
2193 		   nodes_empty(cs->mems_allowed);
2194 
2195 	mutex_unlock(&cpuset_mutex);
2196 
2197 	/*
2198 	 * Move tasks to the nearest ancestor with execution resources,
2199 	 * This is full cgroup operation which will also call back into
2200 	 * cpuset. Should be done outside any lock.
2201 	 */
2202 	if (is_empty)
2203 		remove_tasks_in_empty_cpuset(cs);
2204 
2205 	mutex_lock(&cpuset_mutex);
2206 }
2207 
2208 static void
2209 hotplug_update_tasks(struct cpuset *cs,
2210 		     struct cpumask *new_cpus, nodemask_t *new_mems,
2211 		     bool cpus_updated, bool mems_updated)
2212 {
2213 	if (cpumask_empty(new_cpus))
2214 		cpumask_copy(new_cpus, parent_cs(cs)->effective_cpus);
2215 	if (nodes_empty(*new_mems))
2216 		*new_mems = parent_cs(cs)->effective_mems;
2217 
2218 	spin_lock_irq(&callback_lock);
2219 	cpumask_copy(cs->effective_cpus, new_cpus);
2220 	cs->effective_mems = *new_mems;
2221 	spin_unlock_irq(&callback_lock);
2222 
2223 	if (cpus_updated)
2224 		update_tasks_cpumask(cs);
2225 	if (mems_updated)
2226 		update_tasks_nodemask(cs);
2227 }
2228 
2229 /**
2230  * cpuset_hotplug_update_tasks - update tasks in a cpuset for hotunplug
2231  * @cs: cpuset in interest
2232  *
2233  * Compare @cs's cpu and mem masks against top_cpuset and if some have gone
2234  * offline, update @cs accordingly.  If @cs ends up with no CPU or memory,
2235  * all its tasks are moved to the nearest ancestor with both resources.
2236  */
2237 static void cpuset_hotplug_update_tasks(struct cpuset *cs)
2238 {
2239 	static cpumask_t new_cpus;
2240 	static nodemask_t new_mems;
2241 	bool cpus_updated;
2242 	bool mems_updated;
2243 retry:
2244 	wait_event(cpuset_attach_wq, cs->attach_in_progress == 0);
2245 
2246 	mutex_lock(&cpuset_mutex);
2247 
2248 	/*
2249 	 * We have raced with task attaching. We wait until attaching
2250 	 * is finished, so we won't attach a task to an empty cpuset.
2251 	 */
2252 	if (cs->attach_in_progress) {
2253 		mutex_unlock(&cpuset_mutex);
2254 		goto retry;
2255 	}
2256 
2257 	cpumask_and(&new_cpus, cs->cpus_allowed, parent_cs(cs)->effective_cpus);
2258 	nodes_and(new_mems, cs->mems_allowed, parent_cs(cs)->effective_mems);
2259 
2260 	cpus_updated = !cpumask_equal(&new_cpus, cs->effective_cpus);
2261 	mems_updated = !nodes_equal(new_mems, cs->effective_mems);
2262 
2263 	if (is_in_v2_mode())
2264 		hotplug_update_tasks(cs, &new_cpus, &new_mems,
2265 				     cpus_updated, mems_updated);
2266 	else
2267 		hotplug_update_tasks_legacy(cs, &new_cpus, &new_mems,
2268 					    cpus_updated, mems_updated);
2269 
2270 	mutex_unlock(&cpuset_mutex);
2271 }
2272 
2273 static bool force_rebuild;
2274 
2275 void cpuset_force_rebuild(void)
2276 {
2277 	force_rebuild = true;
2278 }
2279 
2280 /**
2281  * cpuset_hotplug_workfn - handle CPU/memory hotunplug for a cpuset
2282  *
2283  * This function is called after either CPU or memory configuration has
2284  * changed and updates cpuset accordingly.  The top_cpuset is always
2285  * synchronized to cpu_active_mask and N_MEMORY, which is necessary in
2286  * order to make cpusets transparent (of no affect) on systems that are
2287  * actively using CPU hotplug but making no active use of cpusets.
2288  *
2289  * Non-root cpusets are only affected by offlining.  If any CPUs or memory
2290  * nodes have been taken down, cpuset_hotplug_update_tasks() is invoked on
2291  * all descendants.
2292  *
2293  * Note that CPU offlining during suspend is ignored.  We don't modify
2294  * cpusets across suspend/resume cycles at all.
2295  */
2296 static void cpuset_hotplug_workfn(struct work_struct *work)
2297 {
2298 	static cpumask_t new_cpus;
2299 	static nodemask_t new_mems;
2300 	bool cpus_updated, mems_updated;
2301 	bool on_dfl = is_in_v2_mode();
2302 
2303 	mutex_lock(&cpuset_mutex);
2304 
2305 	/* fetch the available cpus/mems and find out which changed how */
2306 	cpumask_copy(&new_cpus, cpu_active_mask);
2307 	new_mems = node_states[N_MEMORY];
2308 
2309 	cpus_updated = !cpumask_equal(top_cpuset.effective_cpus, &new_cpus);
2310 	mems_updated = !nodes_equal(top_cpuset.effective_mems, new_mems);
2311 
2312 	/* synchronize cpus_allowed to cpu_active_mask */
2313 	if (cpus_updated) {
2314 		spin_lock_irq(&callback_lock);
2315 		if (!on_dfl)
2316 			cpumask_copy(top_cpuset.cpus_allowed, &new_cpus);
2317 		cpumask_copy(top_cpuset.effective_cpus, &new_cpus);
2318 		spin_unlock_irq(&callback_lock);
2319 		/* we don't mess with cpumasks of tasks in top_cpuset */
2320 	}
2321 
2322 	/* synchronize mems_allowed to N_MEMORY */
2323 	if (mems_updated) {
2324 		spin_lock_irq(&callback_lock);
2325 		if (!on_dfl)
2326 			top_cpuset.mems_allowed = new_mems;
2327 		top_cpuset.effective_mems = new_mems;
2328 		spin_unlock_irq(&callback_lock);
2329 		update_tasks_nodemask(&top_cpuset);
2330 	}
2331 
2332 	mutex_unlock(&cpuset_mutex);
2333 
2334 	/* if cpus or mems changed, we need to propagate to descendants */
2335 	if (cpus_updated || mems_updated) {
2336 		struct cpuset *cs;
2337 		struct cgroup_subsys_state *pos_css;
2338 
2339 		rcu_read_lock();
2340 		cpuset_for_each_descendant_pre(cs, pos_css, &top_cpuset) {
2341 			if (cs == &top_cpuset || !css_tryget_online(&cs->css))
2342 				continue;
2343 			rcu_read_unlock();
2344 
2345 			cpuset_hotplug_update_tasks(cs);
2346 
2347 			rcu_read_lock();
2348 			css_put(&cs->css);
2349 		}
2350 		rcu_read_unlock();
2351 	}
2352 
2353 	/* rebuild sched domains if cpus_allowed has changed */
2354 	if (cpus_updated || force_rebuild) {
2355 		force_rebuild = false;
2356 		rebuild_sched_domains();
2357 	}
2358 }
2359 
2360 void cpuset_update_active_cpus(void)
2361 {
2362 	/*
2363 	 * We're inside cpu hotplug critical region which usually nests
2364 	 * inside cgroup synchronization.  Bounce actual hotplug processing
2365 	 * to a work item to avoid reverse locking order.
2366 	 */
2367 	schedule_work(&cpuset_hotplug_work);
2368 }
2369 
2370 void cpuset_wait_for_hotplug(void)
2371 {
2372 	flush_work(&cpuset_hotplug_work);
2373 }
2374 
2375 /*
2376  * Keep top_cpuset.mems_allowed tracking node_states[N_MEMORY].
2377  * Call this routine anytime after node_states[N_MEMORY] changes.
2378  * See cpuset_update_active_cpus() for CPU hotplug handling.
2379  */
2380 static int cpuset_track_online_nodes(struct notifier_block *self,
2381 				unsigned long action, void *arg)
2382 {
2383 	schedule_work(&cpuset_hotplug_work);
2384 	return NOTIFY_OK;
2385 }
2386 
2387 static struct notifier_block cpuset_track_online_nodes_nb = {
2388 	.notifier_call = cpuset_track_online_nodes,
2389 	.priority = 10,		/* ??! */
2390 };
2391 
2392 /**
2393  * cpuset_init_smp - initialize cpus_allowed
2394  *
2395  * Description: Finish top cpuset after cpu, node maps are initialized
2396  */
2397 void __init cpuset_init_smp(void)
2398 {
2399 	cpumask_copy(top_cpuset.cpus_allowed, cpu_active_mask);
2400 	top_cpuset.mems_allowed = node_states[N_MEMORY];
2401 	top_cpuset.old_mems_allowed = top_cpuset.mems_allowed;
2402 
2403 	cpumask_copy(top_cpuset.effective_cpus, cpu_active_mask);
2404 	top_cpuset.effective_mems = node_states[N_MEMORY];
2405 
2406 	register_hotmemory_notifier(&cpuset_track_online_nodes_nb);
2407 
2408 	cpuset_migrate_mm_wq = alloc_ordered_workqueue("cpuset_migrate_mm", 0);
2409 	BUG_ON(!cpuset_migrate_mm_wq);
2410 }
2411 
2412 /**
2413  * cpuset_cpus_allowed - return cpus_allowed mask from a tasks cpuset.
2414  * @tsk: pointer to task_struct from which to obtain cpuset->cpus_allowed.
2415  * @pmask: pointer to struct cpumask variable to receive cpus_allowed set.
2416  *
2417  * Description: Returns the cpumask_var_t cpus_allowed of the cpuset
2418  * attached to the specified @tsk.  Guaranteed to return some non-empty
2419  * subset of cpu_online_mask, even if this means going outside the
2420  * tasks cpuset.
2421  **/
2422 
2423 void cpuset_cpus_allowed(struct task_struct *tsk, struct cpumask *pmask)
2424 {
2425 	unsigned long flags;
2426 
2427 	spin_lock_irqsave(&callback_lock, flags);
2428 	rcu_read_lock();
2429 	guarantee_online_cpus(task_cs(tsk), pmask);
2430 	rcu_read_unlock();
2431 	spin_unlock_irqrestore(&callback_lock, flags);
2432 }
2433 
2434 void cpuset_cpus_allowed_fallback(struct task_struct *tsk)
2435 {
2436 	rcu_read_lock();
2437 	do_set_cpus_allowed(tsk, task_cs(tsk)->effective_cpus);
2438 	rcu_read_unlock();
2439 
2440 	/*
2441 	 * We own tsk->cpus_allowed, nobody can change it under us.
2442 	 *
2443 	 * But we used cs && cs->cpus_allowed lockless and thus can
2444 	 * race with cgroup_attach_task() or update_cpumask() and get
2445 	 * the wrong tsk->cpus_allowed. However, both cases imply the
2446 	 * subsequent cpuset_change_cpumask()->set_cpus_allowed_ptr()
2447 	 * which takes task_rq_lock().
2448 	 *
2449 	 * If we are called after it dropped the lock we must see all
2450 	 * changes in tsk_cs()->cpus_allowed. Otherwise we can temporary
2451 	 * set any mask even if it is not right from task_cs() pov,
2452 	 * the pending set_cpus_allowed_ptr() will fix things.
2453 	 *
2454 	 * select_fallback_rq() will fix things ups and set cpu_possible_mask
2455 	 * if required.
2456 	 */
2457 }
2458 
2459 void __init cpuset_init_current_mems_allowed(void)
2460 {
2461 	nodes_setall(current->mems_allowed);
2462 }
2463 
2464 /**
2465  * cpuset_mems_allowed - return mems_allowed mask from a tasks cpuset.
2466  * @tsk: pointer to task_struct from which to obtain cpuset->mems_allowed.
2467  *
2468  * Description: Returns the nodemask_t mems_allowed of the cpuset
2469  * attached to the specified @tsk.  Guaranteed to return some non-empty
2470  * subset of node_states[N_MEMORY], even if this means going outside the
2471  * tasks cpuset.
2472  **/
2473 
2474 nodemask_t cpuset_mems_allowed(struct task_struct *tsk)
2475 {
2476 	nodemask_t mask;
2477 	unsigned long flags;
2478 
2479 	spin_lock_irqsave(&callback_lock, flags);
2480 	rcu_read_lock();
2481 	guarantee_online_mems(task_cs(tsk), &mask);
2482 	rcu_read_unlock();
2483 	spin_unlock_irqrestore(&callback_lock, flags);
2484 
2485 	return mask;
2486 }
2487 
2488 /**
2489  * cpuset_nodemask_valid_mems_allowed - check nodemask vs. curremt mems_allowed
2490  * @nodemask: the nodemask to be checked
2491  *
2492  * Are any of the nodes in the nodemask allowed in current->mems_allowed?
2493  */
2494 int cpuset_nodemask_valid_mems_allowed(nodemask_t *nodemask)
2495 {
2496 	return nodes_intersects(*nodemask, current->mems_allowed);
2497 }
2498 
2499 /*
2500  * nearest_hardwall_ancestor() - Returns the nearest mem_exclusive or
2501  * mem_hardwall ancestor to the specified cpuset.  Call holding
2502  * callback_lock.  If no ancestor is mem_exclusive or mem_hardwall
2503  * (an unusual configuration), then returns the root cpuset.
2504  */
2505 static struct cpuset *nearest_hardwall_ancestor(struct cpuset *cs)
2506 {
2507 	while (!(is_mem_exclusive(cs) || is_mem_hardwall(cs)) && parent_cs(cs))
2508 		cs = parent_cs(cs);
2509 	return cs;
2510 }
2511 
2512 /**
2513  * cpuset_node_allowed - Can we allocate on a memory node?
2514  * @node: is this an allowed node?
2515  * @gfp_mask: memory allocation flags
2516  *
2517  * If we're in interrupt, yes, we can always allocate.  If @node is set in
2518  * current's mems_allowed, yes.  If it's not a __GFP_HARDWALL request and this
2519  * node is set in the nearest hardwalled cpuset ancestor to current's cpuset,
2520  * yes.  If current has access to memory reserves as an oom victim, yes.
2521  * Otherwise, no.
2522  *
2523  * GFP_USER allocations are marked with the __GFP_HARDWALL bit,
2524  * and do not allow allocations outside the current tasks cpuset
2525  * unless the task has been OOM killed.
2526  * GFP_KERNEL allocations are not so marked, so can escape to the
2527  * nearest enclosing hardwalled ancestor cpuset.
2528  *
2529  * Scanning up parent cpusets requires callback_lock.  The
2530  * __alloc_pages() routine only calls here with __GFP_HARDWALL bit
2531  * _not_ set if it's a GFP_KERNEL allocation, and all nodes in the
2532  * current tasks mems_allowed came up empty on the first pass over
2533  * the zonelist.  So only GFP_KERNEL allocations, if all nodes in the
2534  * cpuset are short of memory, might require taking the callback_lock.
2535  *
2536  * The first call here from mm/page_alloc:get_page_from_freelist()
2537  * has __GFP_HARDWALL set in gfp_mask, enforcing hardwall cpusets,
2538  * so no allocation on a node outside the cpuset is allowed (unless
2539  * in interrupt, of course).
2540  *
2541  * The second pass through get_page_from_freelist() doesn't even call
2542  * here for GFP_ATOMIC calls.  For those calls, the __alloc_pages()
2543  * variable 'wait' is not set, and the bit ALLOC_CPUSET is not set
2544  * in alloc_flags.  That logic and the checks below have the combined
2545  * affect that:
2546  *	in_interrupt - any node ok (current task context irrelevant)
2547  *	GFP_ATOMIC   - any node ok
2548  *	tsk_is_oom_victim   - any node ok
2549  *	GFP_KERNEL   - any node in enclosing hardwalled cpuset ok
2550  *	GFP_USER     - only nodes in current tasks mems allowed ok.
2551  */
2552 bool __cpuset_node_allowed(int node, gfp_t gfp_mask)
2553 {
2554 	struct cpuset *cs;		/* current cpuset ancestors */
2555 	int allowed;			/* is allocation in zone z allowed? */
2556 	unsigned long flags;
2557 
2558 	if (in_interrupt())
2559 		return true;
2560 	if (node_isset(node, current->mems_allowed))
2561 		return true;
2562 	/*
2563 	 * Allow tasks that have access to memory reserves because they have
2564 	 * been OOM killed to get memory anywhere.
2565 	 */
2566 	if (unlikely(tsk_is_oom_victim(current)))
2567 		return true;
2568 	if (gfp_mask & __GFP_HARDWALL)	/* If hardwall request, stop here */
2569 		return false;
2570 
2571 	if (current->flags & PF_EXITING) /* Let dying task have memory */
2572 		return true;
2573 
2574 	/* Not hardwall and node outside mems_allowed: scan up cpusets */
2575 	spin_lock_irqsave(&callback_lock, flags);
2576 
2577 	rcu_read_lock();
2578 	cs = nearest_hardwall_ancestor(task_cs(current));
2579 	allowed = node_isset(node, cs->mems_allowed);
2580 	rcu_read_unlock();
2581 
2582 	spin_unlock_irqrestore(&callback_lock, flags);
2583 	return allowed;
2584 }
2585 
2586 /**
2587  * cpuset_mem_spread_node() - On which node to begin search for a file page
2588  * cpuset_slab_spread_node() - On which node to begin search for a slab page
2589  *
2590  * If a task is marked PF_SPREAD_PAGE or PF_SPREAD_SLAB (as for
2591  * tasks in a cpuset with is_spread_page or is_spread_slab set),
2592  * and if the memory allocation used cpuset_mem_spread_node()
2593  * to determine on which node to start looking, as it will for
2594  * certain page cache or slab cache pages such as used for file
2595  * system buffers and inode caches, then instead of starting on the
2596  * local node to look for a free page, rather spread the starting
2597  * node around the tasks mems_allowed nodes.
2598  *
2599  * We don't have to worry about the returned node being offline
2600  * because "it can't happen", and even if it did, it would be ok.
2601  *
2602  * The routines calling guarantee_online_mems() are careful to
2603  * only set nodes in task->mems_allowed that are online.  So it
2604  * should not be possible for the following code to return an
2605  * offline node.  But if it did, that would be ok, as this routine
2606  * is not returning the node where the allocation must be, only
2607  * the node where the search should start.  The zonelist passed to
2608  * __alloc_pages() will include all nodes.  If the slab allocator
2609  * is passed an offline node, it will fall back to the local node.
2610  * See kmem_cache_alloc_node().
2611  */
2612 
2613 static int cpuset_spread_node(int *rotor)
2614 {
2615 	return *rotor = next_node_in(*rotor, current->mems_allowed);
2616 }
2617 
2618 int cpuset_mem_spread_node(void)
2619 {
2620 	if (current->cpuset_mem_spread_rotor == NUMA_NO_NODE)
2621 		current->cpuset_mem_spread_rotor =
2622 			node_random(&current->mems_allowed);
2623 
2624 	return cpuset_spread_node(&current->cpuset_mem_spread_rotor);
2625 }
2626 
2627 int cpuset_slab_spread_node(void)
2628 {
2629 	if (current->cpuset_slab_spread_rotor == NUMA_NO_NODE)
2630 		current->cpuset_slab_spread_rotor =
2631 			node_random(&current->mems_allowed);
2632 
2633 	return cpuset_spread_node(&current->cpuset_slab_spread_rotor);
2634 }
2635 
2636 EXPORT_SYMBOL_GPL(cpuset_mem_spread_node);
2637 
2638 /**
2639  * cpuset_mems_allowed_intersects - Does @tsk1's mems_allowed intersect @tsk2's?
2640  * @tsk1: pointer to task_struct of some task.
2641  * @tsk2: pointer to task_struct of some other task.
2642  *
2643  * Description: Return true if @tsk1's mems_allowed intersects the
2644  * mems_allowed of @tsk2.  Used by the OOM killer to determine if
2645  * one of the task's memory usage might impact the memory available
2646  * to the other.
2647  **/
2648 
2649 int cpuset_mems_allowed_intersects(const struct task_struct *tsk1,
2650 				   const struct task_struct *tsk2)
2651 {
2652 	return nodes_intersects(tsk1->mems_allowed, tsk2->mems_allowed);
2653 }
2654 
2655 /**
2656  * cpuset_print_current_mems_allowed - prints current's cpuset and mems_allowed
2657  *
2658  * Description: Prints current's name, cpuset name, and cached copy of its
2659  * mems_allowed to the kernel log.
2660  */
2661 void cpuset_print_current_mems_allowed(void)
2662 {
2663 	struct cgroup *cgrp;
2664 
2665 	rcu_read_lock();
2666 
2667 	cgrp = task_cs(current)->css.cgroup;
2668 	pr_info("%s cpuset=", current->comm);
2669 	pr_cont_cgroup_name(cgrp);
2670 	pr_cont(" mems_allowed=%*pbl\n",
2671 		nodemask_pr_args(&current->mems_allowed));
2672 
2673 	rcu_read_unlock();
2674 }
2675 
2676 /*
2677  * Collection of memory_pressure is suppressed unless
2678  * this flag is enabled by writing "1" to the special
2679  * cpuset file 'memory_pressure_enabled' in the root cpuset.
2680  */
2681 
2682 int cpuset_memory_pressure_enabled __read_mostly;
2683 
2684 /**
2685  * cpuset_memory_pressure_bump - keep stats of per-cpuset reclaims.
2686  *
2687  * Keep a running average of the rate of synchronous (direct)
2688  * page reclaim efforts initiated by tasks in each cpuset.
2689  *
2690  * This represents the rate at which some task in the cpuset
2691  * ran low on memory on all nodes it was allowed to use, and
2692  * had to enter the kernels page reclaim code in an effort to
2693  * create more free memory by tossing clean pages or swapping
2694  * or writing dirty pages.
2695  *
2696  * Display to user space in the per-cpuset read-only file
2697  * "memory_pressure".  Value displayed is an integer
2698  * representing the recent rate of entry into the synchronous
2699  * (direct) page reclaim by any task attached to the cpuset.
2700  **/
2701 
2702 void __cpuset_memory_pressure_bump(void)
2703 {
2704 	rcu_read_lock();
2705 	fmeter_markevent(&task_cs(current)->fmeter);
2706 	rcu_read_unlock();
2707 }
2708 
2709 #ifdef CONFIG_PROC_PID_CPUSET
2710 /*
2711  * proc_cpuset_show()
2712  *  - Print tasks cpuset path into seq_file.
2713  *  - Used for /proc/<pid>/cpuset.
2714  *  - No need to task_lock(tsk) on this tsk->cpuset reference, as it
2715  *    doesn't really matter if tsk->cpuset changes after we read it,
2716  *    and we take cpuset_mutex, keeping cpuset_attach() from changing it
2717  *    anyway.
2718  */
2719 int proc_cpuset_show(struct seq_file *m, struct pid_namespace *ns,
2720 		     struct pid *pid, struct task_struct *tsk)
2721 {
2722 	char *buf;
2723 	struct cgroup_subsys_state *css;
2724 	int retval;
2725 
2726 	retval = -ENOMEM;
2727 	buf = kmalloc(PATH_MAX, GFP_KERNEL);
2728 	if (!buf)
2729 		goto out;
2730 
2731 	css = task_get_css(tsk, cpuset_cgrp_id);
2732 	retval = cgroup_path_ns(css->cgroup, buf, PATH_MAX,
2733 				current->nsproxy->cgroup_ns);
2734 	css_put(css);
2735 	if (retval >= PATH_MAX)
2736 		retval = -ENAMETOOLONG;
2737 	if (retval < 0)
2738 		goto out_free;
2739 	seq_puts(m, buf);
2740 	seq_putc(m, '\n');
2741 	retval = 0;
2742 out_free:
2743 	kfree(buf);
2744 out:
2745 	return retval;
2746 }
2747 #endif /* CONFIG_PROC_PID_CPUSET */
2748 
2749 /* Display task mems_allowed in /proc/<pid>/status file. */
2750 void cpuset_task_status_allowed(struct seq_file *m, struct task_struct *task)
2751 {
2752 	seq_printf(m, "Mems_allowed:\t%*pb\n",
2753 		   nodemask_pr_args(&task->mems_allowed));
2754 	seq_printf(m, "Mems_allowed_list:\t%*pbl\n",
2755 		   nodemask_pr_args(&task->mems_allowed));
2756 }
2757