xref: /openbmc/linux/kernel/cgroup/cpuset.c (revision c8dbaa22)
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 
60 #include <linux/uaccess.h>
61 #include <linux/atomic.h>
62 #include <linux/mutex.h>
63 #include <linux/cgroup.h>
64 #include <linux/wait.h>
65 
66 DEFINE_STATIC_KEY_FALSE(cpusets_pre_enable_key);
67 DEFINE_STATIC_KEY_FALSE(cpusets_enabled_key);
68 
69 /* See "Frequency meter" comments, below. */
70 
71 struct fmeter {
72 	int cnt;		/* unprocessed events count */
73 	int val;		/* most recent output value */
74 	time64_t time;		/* clock (secs) when val computed */
75 	spinlock_t lock;	/* guards read or write of above */
76 };
77 
78 struct cpuset {
79 	struct cgroup_subsys_state css;
80 
81 	unsigned long flags;		/* "unsigned long" so bitops work */
82 
83 	/*
84 	 * On default hierarchy:
85 	 *
86 	 * The user-configured masks can only be changed by writing to
87 	 * cpuset.cpus and cpuset.mems, and won't be limited by the
88 	 * parent masks.
89 	 *
90 	 * The effective masks is the real masks that apply to the tasks
91 	 * in the cpuset. They may be changed if the configured masks are
92 	 * changed or hotplug happens.
93 	 *
94 	 * effective_mask == configured_mask & parent's effective_mask,
95 	 * and if it ends up empty, it will inherit the parent's mask.
96 	 *
97 	 *
98 	 * On legacy hierachy:
99 	 *
100 	 * The user-configured masks are always the same with effective masks.
101 	 */
102 
103 	/* user-configured CPUs and Memory Nodes allow to tasks */
104 	cpumask_var_t cpus_allowed;
105 	nodemask_t mems_allowed;
106 
107 	/* effective CPUs and Memory Nodes allow to tasks */
108 	cpumask_var_t effective_cpus;
109 	nodemask_t effective_mems;
110 
111 	/*
112 	 * This is old Memory Nodes tasks took on.
113 	 *
114 	 * - top_cpuset.old_mems_allowed is initialized to mems_allowed.
115 	 * - A new cpuset's old_mems_allowed is initialized when some
116 	 *   task is moved into it.
117 	 * - old_mems_allowed is used in cpuset_migrate_mm() when we change
118 	 *   cpuset.mems_allowed and have tasks' nodemask updated, and
119 	 *   then old_mems_allowed is updated to mems_allowed.
120 	 */
121 	nodemask_t old_mems_allowed;
122 
123 	struct fmeter fmeter;		/* memory_pressure filter */
124 
125 	/*
126 	 * Tasks are being attached to this cpuset.  Used to prevent
127 	 * zeroing cpus/mems_allowed between ->can_attach() and ->attach().
128 	 */
129 	int attach_in_progress;
130 
131 	/* partition number for rebuild_sched_domains() */
132 	int pn;
133 
134 	/* for custom sched domain */
135 	int relax_domain_level;
136 };
137 
138 static inline struct cpuset *css_cs(struct cgroup_subsys_state *css)
139 {
140 	return css ? container_of(css, struct cpuset, css) : NULL;
141 }
142 
143 /* Retrieve the cpuset for a task */
144 static inline struct cpuset *task_cs(struct task_struct *task)
145 {
146 	return css_cs(task_css(task, cpuset_cgrp_id));
147 }
148 
149 static inline struct cpuset *parent_cs(struct cpuset *cs)
150 {
151 	return css_cs(cs->css.parent);
152 }
153 
154 #ifdef CONFIG_NUMA
155 static inline bool task_has_mempolicy(struct task_struct *task)
156 {
157 	return task->mempolicy;
158 }
159 #else
160 static inline bool task_has_mempolicy(struct task_struct *task)
161 {
162 	return false;
163 }
164 #endif
165 
166 
167 /* bits in struct cpuset flags field */
168 typedef enum {
169 	CS_ONLINE,
170 	CS_CPU_EXCLUSIVE,
171 	CS_MEM_EXCLUSIVE,
172 	CS_MEM_HARDWALL,
173 	CS_MEMORY_MIGRATE,
174 	CS_SCHED_LOAD_BALANCE,
175 	CS_SPREAD_PAGE,
176 	CS_SPREAD_SLAB,
177 } cpuset_flagbits_t;
178 
179 /* convenient tests for these bits */
180 static inline bool is_cpuset_online(struct cpuset *cs)
181 {
182 	return test_bit(CS_ONLINE, &cs->flags) && !css_is_dying(&cs->css);
183 }
184 
185 static inline int is_cpu_exclusive(const struct cpuset *cs)
186 {
187 	return test_bit(CS_CPU_EXCLUSIVE, &cs->flags);
188 }
189 
190 static inline int is_mem_exclusive(const struct cpuset *cs)
191 {
192 	return test_bit(CS_MEM_EXCLUSIVE, &cs->flags);
193 }
194 
195 static inline int is_mem_hardwall(const struct cpuset *cs)
196 {
197 	return test_bit(CS_MEM_HARDWALL, &cs->flags);
198 }
199 
200 static inline int is_sched_load_balance(const struct cpuset *cs)
201 {
202 	return test_bit(CS_SCHED_LOAD_BALANCE, &cs->flags);
203 }
204 
205 static inline int is_memory_migrate(const struct cpuset *cs)
206 {
207 	return test_bit(CS_MEMORY_MIGRATE, &cs->flags);
208 }
209 
210 static inline int is_spread_page(const struct cpuset *cs)
211 {
212 	return test_bit(CS_SPREAD_PAGE, &cs->flags);
213 }
214 
215 static inline int is_spread_slab(const struct cpuset *cs)
216 {
217 	return test_bit(CS_SPREAD_SLAB, &cs->flags);
218 }
219 
220 static struct cpuset top_cpuset = {
221 	.flags = ((1 << CS_ONLINE) | (1 << CS_CPU_EXCLUSIVE) |
222 		  (1 << CS_MEM_EXCLUSIVE)),
223 };
224 
225 /**
226  * cpuset_for_each_child - traverse online children of a cpuset
227  * @child_cs: loop cursor pointing to the current child
228  * @pos_css: used for iteration
229  * @parent_cs: target cpuset to walk children of
230  *
231  * Walk @child_cs through the online children of @parent_cs.  Must be used
232  * with RCU read locked.
233  */
234 #define cpuset_for_each_child(child_cs, pos_css, parent_cs)		\
235 	css_for_each_child((pos_css), &(parent_cs)->css)		\
236 		if (is_cpuset_online(((child_cs) = css_cs((pos_css)))))
237 
238 /**
239  * cpuset_for_each_descendant_pre - pre-order walk of a cpuset's descendants
240  * @des_cs: loop cursor pointing to the current descendant
241  * @pos_css: used for iteration
242  * @root_cs: target cpuset to walk ancestor of
243  *
244  * Walk @des_cs through the online descendants of @root_cs.  Must be used
245  * with RCU read locked.  The caller may modify @pos_css by calling
246  * css_rightmost_descendant() to skip subtree.  @root_cs is included in the
247  * iteration and the first node to be visited.
248  */
249 #define cpuset_for_each_descendant_pre(des_cs, pos_css, root_cs)	\
250 	css_for_each_descendant_pre((pos_css), &(root_cs)->css)		\
251 		if (is_cpuset_online(((des_cs) = css_cs((pos_css)))))
252 
253 /*
254  * There are two global locks guarding cpuset structures - cpuset_mutex and
255  * callback_lock. We also require taking task_lock() when dereferencing a
256  * task's cpuset pointer. See "The task_lock() exception", at the end of this
257  * comment.
258  *
259  * A task must hold both locks to modify cpusets.  If a task holds
260  * cpuset_mutex, then it blocks others wanting that mutex, ensuring that it
261  * is the only task able to also acquire callback_lock and be able to
262  * modify cpusets.  It can perform various checks on the cpuset structure
263  * first, knowing nothing will change.  It can also allocate memory while
264  * just holding cpuset_mutex.  While it is performing these checks, various
265  * callback routines can briefly acquire callback_lock to query cpusets.
266  * Once it is ready to make the changes, it takes callback_lock, blocking
267  * everyone else.
268  *
269  * Calls to the kernel memory allocator can not be made while holding
270  * callback_lock, as that would risk double tripping on callback_lock
271  * from one of the callbacks into the cpuset code from within
272  * __alloc_pages().
273  *
274  * If a task is only holding callback_lock, then it has read-only
275  * access to cpusets.
276  *
277  * Now, the task_struct fields mems_allowed and mempolicy may be changed
278  * by other task, we use alloc_lock in the task_struct fields to protect
279  * them.
280  *
281  * The cpuset_common_file_read() handlers only hold callback_lock across
282  * small pieces of code, such as when reading out possibly multi-word
283  * cpumasks and nodemasks.
284  *
285  * Accessing a task's cpuset should be done in accordance with the
286  * guidelines for accessing subsystem state in kernel/cgroup.c
287  */
288 
289 static DEFINE_MUTEX(cpuset_mutex);
290 static DEFINE_SPINLOCK(callback_lock);
291 
292 static struct workqueue_struct *cpuset_migrate_mm_wq;
293 
294 /*
295  * CPU / memory hotplug is handled asynchronously.
296  */
297 static void cpuset_hotplug_workfn(struct work_struct *work);
298 static DECLARE_WORK(cpuset_hotplug_work, cpuset_hotplug_workfn);
299 
300 static DECLARE_WAIT_QUEUE_HEAD(cpuset_attach_wq);
301 
302 /*
303  * This is ugly, but preserves the userspace API for existing cpuset
304  * users. If someone tries to mount the "cpuset" filesystem, we
305  * silently switch it to mount "cgroup" instead
306  */
307 static struct dentry *cpuset_mount(struct file_system_type *fs_type,
308 			 int flags, const char *unused_dev_name, void *data)
309 {
310 	struct file_system_type *cgroup_fs = get_fs_type("cgroup");
311 	struct dentry *ret = ERR_PTR(-ENODEV);
312 	if (cgroup_fs) {
313 		char mountopts[] =
314 			"cpuset,noprefix,"
315 			"release_agent=/sbin/cpuset_release_agent";
316 		ret = cgroup_fs->mount(cgroup_fs, flags,
317 					   unused_dev_name, mountopts);
318 		put_filesystem(cgroup_fs);
319 	}
320 	return ret;
321 }
322 
323 static struct file_system_type cpuset_fs_type = {
324 	.name = "cpuset",
325 	.mount = cpuset_mount,
326 };
327 
328 /*
329  * Return in pmask the portion of a cpusets's cpus_allowed that
330  * are online.  If none are online, walk up the cpuset hierarchy
331  * until we find one that does have some online cpus.
332  *
333  * One way or another, we guarantee to return some non-empty subset
334  * of cpu_online_mask.
335  *
336  * Call with callback_lock or cpuset_mutex held.
337  */
338 static void guarantee_online_cpus(struct cpuset *cs, struct cpumask *pmask)
339 {
340 	while (!cpumask_intersects(cs->effective_cpus, cpu_online_mask)) {
341 		cs = parent_cs(cs);
342 		if (unlikely(!cs)) {
343 			/*
344 			 * The top cpuset doesn't have any online cpu as a
345 			 * consequence of a race between cpuset_hotplug_work
346 			 * and cpu hotplug notifier.  But we know the top
347 			 * cpuset's effective_cpus is on its way to to be
348 			 * identical to cpu_online_mask.
349 			 */
350 			cpumask_copy(pmask, cpu_online_mask);
351 			return;
352 		}
353 	}
354 	cpumask_and(pmask, cs->effective_cpus, cpu_online_mask);
355 }
356 
357 /*
358  * Return in *pmask the portion of a cpusets's mems_allowed that
359  * are online, with memory.  If none are online with memory, walk
360  * up the cpuset hierarchy until we find one that does have some
361  * online mems.  The top cpuset always has some mems online.
362  *
363  * One way or another, we guarantee to return some non-empty subset
364  * of node_states[N_MEMORY].
365  *
366  * Call with callback_lock or cpuset_mutex held.
367  */
368 static void guarantee_online_mems(struct cpuset *cs, nodemask_t *pmask)
369 {
370 	while (!nodes_intersects(cs->effective_mems, node_states[N_MEMORY]))
371 		cs = parent_cs(cs);
372 	nodes_and(*pmask, cs->effective_mems, node_states[N_MEMORY]);
373 }
374 
375 /*
376  * update task's spread flag if cpuset's page/slab spread flag is set
377  *
378  * Call with callback_lock or cpuset_mutex held.
379  */
380 static void cpuset_update_task_spread_flag(struct cpuset *cs,
381 					struct task_struct *tsk)
382 {
383 	if (is_spread_page(cs))
384 		task_set_spread_page(tsk);
385 	else
386 		task_clear_spread_page(tsk);
387 
388 	if (is_spread_slab(cs))
389 		task_set_spread_slab(tsk);
390 	else
391 		task_clear_spread_slab(tsk);
392 }
393 
394 /*
395  * is_cpuset_subset(p, q) - Is cpuset p a subset of cpuset q?
396  *
397  * One cpuset is a subset of another if all its allowed CPUs and
398  * Memory Nodes are a subset of the other, and its exclusive flags
399  * are only set if the other's are set.  Call holding cpuset_mutex.
400  */
401 
402 static int is_cpuset_subset(const struct cpuset *p, const struct cpuset *q)
403 {
404 	return	cpumask_subset(p->cpus_allowed, q->cpus_allowed) &&
405 		nodes_subset(p->mems_allowed, q->mems_allowed) &&
406 		is_cpu_exclusive(p) <= is_cpu_exclusive(q) &&
407 		is_mem_exclusive(p) <= is_mem_exclusive(q);
408 }
409 
410 /**
411  * alloc_trial_cpuset - allocate a trial cpuset
412  * @cs: the cpuset that the trial cpuset duplicates
413  */
414 static struct cpuset *alloc_trial_cpuset(struct cpuset *cs)
415 {
416 	struct cpuset *trial;
417 
418 	trial = kmemdup(cs, sizeof(*cs), GFP_KERNEL);
419 	if (!trial)
420 		return NULL;
421 
422 	if (!alloc_cpumask_var(&trial->cpus_allowed, GFP_KERNEL))
423 		goto free_cs;
424 	if (!alloc_cpumask_var(&trial->effective_cpus, GFP_KERNEL))
425 		goto free_cpus;
426 
427 	cpumask_copy(trial->cpus_allowed, cs->cpus_allowed);
428 	cpumask_copy(trial->effective_cpus, cs->effective_cpus);
429 	return trial;
430 
431 free_cpus:
432 	free_cpumask_var(trial->cpus_allowed);
433 free_cs:
434 	kfree(trial);
435 	return NULL;
436 }
437 
438 /**
439  * free_trial_cpuset - free the trial cpuset
440  * @trial: the trial cpuset to be freed
441  */
442 static void free_trial_cpuset(struct cpuset *trial)
443 {
444 	free_cpumask_var(trial->effective_cpus);
445 	free_cpumask_var(trial->cpus_allowed);
446 	kfree(trial);
447 }
448 
449 /*
450  * validate_change() - Used to validate that any proposed cpuset change
451  *		       follows the structural rules for cpusets.
452  *
453  * If we replaced the flag and mask values of the current cpuset
454  * (cur) with those values in the trial cpuset (trial), would
455  * our various subset and exclusive rules still be valid?  Presumes
456  * cpuset_mutex held.
457  *
458  * 'cur' is the address of an actual, in-use cpuset.  Operations
459  * such as list traversal that depend on the actual address of the
460  * cpuset in the list must use cur below, not trial.
461  *
462  * 'trial' is the address of bulk structure copy of cur, with
463  * perhaps one or more of the fields cpus_allowed, mems_allowed,
464  * or flags changed to new, trial values.
465  *
466  * Return 0 if valid, -errno if not.
467  */
468 
469 static int validate_change(struct cpuset *cur, struct cpuset *trial)
470 {
471 	struct cgroup_subsys_state *css;
472 	struct cpuset *c, *par;
473 	int ret;
474 
475 	rcu_read_lock();
476 
477 	/* Each of our child cpusets must be a subset of us */
478 	ret = -EBUSY;
479 	cpuset_for_each_child(c, css, cur)
480 		if (!is_cpuset_subset(c, trial))
481 			goto out;
482 
483 	/* Remaining checks don't apply to root cpuset */
484 	ret = 0;
485 	if (cur == &top_cpuset)
486 		goto out;
487 
488 	par = parent_cs(cur);
489 
490 	/* On legacy hiearchy, we must be a subset of our parent cpuset. */
491 	ret = -EACCES;
492 	if (!cgroup_subsys_on_dfl(cpuset_cgrp_subsys) &&
493 	    !is_cpuset_subset(trial, par))
494 		goto out;
495 
496 	/*
497 	 * If either I or some sibling (!= me) is exclusive, we can't
498 	 * overlap
499 	 */
500 	ret = -EINVAL;
501 	cpuset_for_each_child(c, css, par) {
502 		if ((is_cpu_exclusive(trial) || is_cpu_exclusive(c)) &&
503 		    c != cur &&
504 		    cpumask_intersects(trial->cpus_allowed, c->cpus_allowed))
505 			goto out;
506 		if ((is_mem_exclusive(trial) || is_mem_exclusive(c)) &&
507 		    c != cur &&
508 		    nodes_intersects(trial->mems_allowed, c->mems_allowed))
509 			goto out;
510 	}
511 
512 	/*
513 	 * Cpusets with tasks - existing or newly being attached - can't
514 	 * be changed to have empty cpus_allowed or mems_allowed.
515 	 */
516 	ret = -ENOSPC;
517 	if ((cgroup_is_populated(cur->css.cgroup) || cur->attach_in_progress)) {
518 		if (!cpumask_empty(cur->cpus_allowed) &&
519 		    cpumask_empty(trial->cpus_allowed))
520 			goto out;
521 		if (!nodes_empty(cur->mems_allowed) &&
522 		    nodes_empty(trial->mems_allowed))
523 			goto out;
524 	}
525 
526 	/*
527 	 * We can't shrink if we won't have enough room for SCHED_DEADLINE
528 	 * tasks.
529 	 */
530 	ret = -EBUSY;
531 	if (is_cpu_exclusive(cur) &&
532 	    !cpuset_cpumask_can_shrink(cur->cpus_allowed,
533 				       trial->cpus_allowed))
534 		goto out;
535 
536 	ret = 0;
537 out:
538 	rcu_read_unlock();
539 	return ret;
540 }
541 
542 #ifdef CONFIG_SMP
543 /*
544  * Helper routine for generate_sched_domains().
545  * Do cpusets a, b have overlapping effective cpus_allowed masks?
546  */
547 static int cpusets_overlap(struct cpuset *a, struct cpuset *b)
548 {
549 	return cpumask_intersects(a->effective_cpus, b->effective_cpus);
550 }
551 
552 static void
553 update_domain_attr(struct sched_domain_attr *dattr, struct cpuset *c)
554 {
555 	if (dattr->relax_domain_level < c->relax_domain_level)
556 		dattr->relax_domain_level = c->relax_domain_level;
557 	return;
558 }
559 
560 static void update_domain_attr_tree(struct sched_domain_attr *dattr,
561 				    struct cpuset *root_cs)
562 {
563 	struct cpuset *cp;
564 	struct cgroup_subsys_state *pos_css;
565 
566 	rcu_read_lock();
567 	cpuset_for_each_descendant_pre(cp, pos_css, root_cs) {
568 		/* skip the whole subtree if @cp doesn't have any CPU */
569 		if (cpumask_empty(cp->cpus_allowed)) {
570 			pos_css = css_rightmost_descendant(pos_css);
571 			continue;
572 		}
573 
574 		if (is_sched_load_balance(cp))
575 			update_domain_attr(dattr, cp);
576 	}
577 	rcu_read_unlock();
578 }
579 
580 /*
581  * generate_sched_domains()
582  *
583  * This function builds a partial partition of the systems CPUs
584  * A 'partial partition' is a set of non-overlapping subsets whose
585  * union is a subset of that set.
586  * The output of this function needs to be passed to kernel/sched/core.c
587  * partition_sched_domains() routine, which will rebuild the scheduler's
588  * load balancing domains (sched domains) as specified by that partial
589  * partition.
590  *
591  * See "What is sched_load_balance" in Documentation/cgroups/cpusets.txt
592  * for a background explanation of this.
593  *
594  * Does not return errors, on the theory that the callers of this
595  * routine would rather not worry about failures to rebuild sched
596  * domains when operating in the severe memory shortage situations
597  * that could cause allocation failures below.
598  *
599  * Must be called with cpuset_mutex held.
600  *
601  * The three key local variables below are:
602  *    q  - a linked-list queue of cpuset pointers, used to implement a
603  *	   top-down scan of all cpusets.  This scan loads a pointer
604  *	   to each cpuset marked is_sched_load_balance into the
605  *	   array 'csa'.  For our purposes, rebuilding the schedulers
606  *	   sched domains, we can ignore !is_sched_load_balance cpusets.
607  *  csa  - (for CpuSet Array) Array of pointers to all the cpusets
608  *	   that need to be load balanced, for convenient iterative
609  *	   access by the subsequent code that finds the best partition,
610  *	   i.e the set of domains (subsets) of CPUs such that the
611  *	   cpus_allowed of every cpuset marked is_sched_load_balance
612  *	   is a subset of one of these domains, while there are as
613  *	   many such domains as possible, each as small as possible.
614  * doms  - Conversion of 'csa' to an array of cpumasks, for passing to
615  *	   the kernel/sched/core.c routine partition_sched_domains() in a
616  *	   convenient format, that can be easily compared to the prior
617  *	   value to determine what partition elements (sched domains)
618  *	   were changed (added or removed.)
619  *
620  * Finding the best partition (set of domains):
621  *	The triple nested loops below over i, j, k scan over the
622  *	load balanced cpusets (using the array of cpuset pointers in
623  *	csa[]) looking for pairs of cpusets that have overlapping
624  *	cpus_allowed, but which don't have the same 'pn' partition
625  *	number and gives them in the same partition number.  It keeps
626  *	looping on the 'restart' label until it can no longer find
627  *	any such pairs.
628  *
629  *	The union of the cpus_allowed masks from the set of
630  *	all cpusets having the same 'pn' value then form the one
631  *	element of the partition (one sched domain) to be passed to
632  *	partition_sched_domains().
633  */
634 static int generate_sched_domains(cpumask_var_t **domains,
635 			struct sched_domain_attr **attributes)
636 {
637 	struct cpuset *cp;	/* scans q */
638 	struct cpuset **csa;	/* array of all cpuset ptrs */
639 	int csn;		/* how many cpuset ptrs in csa so far */
640 	int i, j, k;		/* indices for partition finding loops */
641 	cpumask_var_t *doms;	/* resulting partition; i.e. sched domains */
642 	cpumask_var_t non_isolated_cpus;  /* load balanced CPUs */
643 	struct sched_domain_attr *dattr;  /* attributes for custom domains */
644 	int ndoms = 0;		/* number of sched domains in result */
645 	int nslot;		/* next empty doms[] struct cpumask slot */
646 	struct cgroup_subsys_state *pos_css;
647 
648 	doms = NULL;
649 	dattr = NULL;
650 	csa = NULL;
651 
652 	if (!alloc_cpumask_var(&non_isolated_cpus, GFP_KERNEL))
653 		goto done;
654 	cpumask_andnot(non_isolated_cpus, cpu_possible_mask, cpu_isolated_map);
655 
656 	/* Special case for the 99% of systems with one, full, sched domain */
657 	if (is_sched_load_balance(&top_cpuset)) {
658 		ndoms = 1;
659 		doms = alloc_sched_domains(ndoms);
660 		if (!doms)
661 			goto done;
662 
663 		dattr = kmalloc(sizeof(struct sched_domain_attr), GFP_KERNEL);
664 		if (dattr) {
665 			*dattr = SD_ATTR_INIT;
666 			update_domain_attr_tree(dattr, &top_cpuset);
667 		}
668 		cpumask_and(doms[0], top_cpuset.effective_cpus,
669 				     non_isolated_cpus);
670 
671 		goto done;
672 	}
673 
674 	csa = kmalloc(nr_cpusets() * sizeof(cp), GFP_KERNEL);
675 	if (!csa)
676 		goto done;
677 	csn = 0;
678 
679 	rcu_read_lock();
680 	cpuset_for_each_descendant_pre(cp, pos_css, &top_cpuset) {
681 		if (cp == &top_cpuset)
682 			continue;
683 		/*
684 		 * Continue traversing beyond @cp iff @cp has some CPUs and
685 		 * isn't load balancing.  The former is obvious.  The
686 		 * latter: All child cpusets contain a subset of the
687 		 * parent's cpus, so just skip them, and then we call
688 		 * update_domain_attr_tree() to calc relax_domain_level of
689 		 * the corresponding sched domain.
690 		 */
691 		if (!cpumask_empty(cp->cpus_allowed) &&
692 		    !(is_sched_load_balance(cp) &&
693 		      cpumask_intersects(cp->cpus_allowed, non_isolated_cpus)))
694 			continue;
695 
696 		if (is_sched_load_balance(cp))
697 			csa[csn++] = cp;
698 
699 		/* skip @cp's subtree */
700 		pos_css = css_rightmost_descendant(pos_css);
701 	}
702 	rcu_read_unlock();
703 
704 	for (i = 0; i < csn; i++)
705 		csa[i]->pn = i;
706 	ndoms = csn;
707 
708 restart:
709 	/* Find the best partition (set of sched domains) */
710 	for (i = 0; i < csn; i++) {
711 		struct cpuset *a = csa[i];
712 		int apn = a->pn;
713 
714 		for (j = 0; j < csn; j++) {
715 			struct cpuset *b = csa[j];
716 			int bpn = b->pn;
717 
718 			if (apn != bpn && cpusets_overlap(a, b)) {
719 				for (k = 0; k < csn; k++) {
720 					struct cpuset *c = csa[k];
721 
722 					if (c->pn == bpn)
723 						c->pn = apn;
724 				}
725 				ndoms--;	/* one less element */
726 				goto restart;
727 			}
728 		}
729 	}
730 
731 	/*
732 	 * Now we know how many domains to create.
733 	 * Convert <csn, csa> to <ndoms, doms> and populate cpu masks.
734 	 */
735 	doms = alloc_sched_domains(ndoms);
736 	if (!doms)
737 		goto done;
738 
739 	/*
740 	 * The rest of the code, including the scheduler, can deal with
741 	 * dattr==NULL case. No need to abort if alloc fails.
742 	 */
743 	dattr = kmalloc(ndoms * sizeof(struct sched_domain_attr), GFP_KERNEL);
744 
745 	for (nslot = 0, i = 0; i < csn; i++) {
746 		struct cpuset *a = csa[i];
747 		struct cpumask *dp;
748 		int apn = a->pn;
749 
750 		if (apn < 0) {
751 			/* Skip completed partitions */
752 			continue;
753 		}
754 
755 		dp = doms[nslot];
756 
757 		if (nslot == ndoms) {
758 			static int warnings = 10;
759 			if (warnings) {
760 				pr_warn("rebuild_sched_domains confused: nslot %d, ndoms %d, csn %d, i %d, apn %d\n",
761 					nslot, ndoms, csn, i, apn);
762 				warnings--;
763 			}
764 			continue;
765 		}
766 
767 		cpumask_clear(dp);
768 		if (dattr)
769 			*(dattr + nslot) = SD_ATTR_INIT;
770 		for (j = i; j < csn; j++) {
771 			struct cpuset *b = csa[j];
772 
773 			if (apn == b->pn) {
774 				cpumask_or(dp, dp, b->effective_cpus);
775 				cpumask_and(dp, dp, non_isolated_cpus);
776 				if (dattr)
777 					update_domain_attr_tree(dattr + nslot, b);
778 
779 				/* Done with this partition */
780 				b->pn = -1;
781 			}
782 		}
783 		nslot++;
784 	}
785 	BUG_ON(nslot != ndoms);
786 
787 done:
788 	free_cpumask_var(non_isolated_cpus);
789 	kfree(csa);
790 
791 	/*
792 	 * Fallback to the default domain if kmalloc() failed.
793 	 * See comments in partition_sched_domains().
794 	 */
795 	if (doms == NULL)
796 		ndoms = 1;
797 
798 	*domains    = doms;
799 	*attributes = dattr;
800 	return ndoms;
801 }
802 
803 /*
804  * Rebuild scheduler domains.
805  *
806  * If the flag 'sched_load_balance' of any cpuset with non-empty
807  * 'cpus' changes, or if the 'cpus' allowed changes in any cpuset
808  * which has that flag enabled, or if any cpuset with a non-empty
809  * 'cpus' is removed, then call this routine to rebuild the
810  * scheduler's dynamic sched domains.
811  *
812  * Call with cpuset_mutex held.  Takes get_online_cpus().
813  */
814 static void rebuild_sched_domains_locked(void)
815 {
816 	struct sched_domain_attr *attr;
817 	cpumask_var_t *doms;
818 	int ndoms;
819 
820 	lockdep_assert_held(&cpuset_mutex);
821 	get_online_cpus();
822 
823 	/*
824 	 * We have raced with CPU hotplug. Don't do anything to avoid
825 	 * passing doms with offlined cpu to partition_sched_domains().
826 	 * Anyways, hotplug work item will rebuild sched domains.
827 	 */
828 	if (!cpumask_equal(top_cpuset.effective_cpus, cpu_active_mask))
829 		goto out;
830 
831 	/* Generate domain masks and attrs */
832 	ndoms = generate_sched_domains(&doms, &attr);
833 
834 	/* Have scheduler rebuild the domains */
835 	partition_sched_domains(ndoms, doms, attr);
836 out:
837 	put_online_cpus();
838 }
839 #else /* !CONFIG_SMP */
840 static void rebuild_sched_domains_locked(void)
841 {
842 }
843 #endif /* CONFIG_SMP */
844 
845 void rebuild_sched_domains(void)
846 {
847 	mutex_lock(&cpuset_mutex);
848 	rebuild_sched_domains_locked();
849 	mutex_unlock(&cpuset_mutex);
850 }
851 
852 /**
853  * update_tasks_cpumask - Update the cpumasks of tasks in the cpuset.
854  * @cs: the cpuset in which each task's cpus_allowed mask needs to be changed
855  *
856  * Iterate through each task of @cs updating its cpus_allowed to the
857  * effective cpuset's.  As this function is called with cpuset_mutex held,
858  * cpuset membership stays stable.
859  */
860 static void update_tasks_cpumask(struct cpuset *cs)
861 {
862 	struct css_task_iter it;
863 	struct task_struct *task;
864 
865 	css_task_iter_start(&cs->css, &it);
866 	while ((task = css_task_iter_next(&it)))
867 		set_cpus_allowed_ptr(task, cs->effective_cpus);
868 	css_task_iter_end(&it);
869 }
870 
871 /*
872  * update_cpumasks_hier - Update effective cpumasks and tasks in the subtree
873  * @cs: the cpuset to consider
874  * @new_cpus: temp variable for calculating new effective_cpus
875  *
876  * When congifured cpumask is changed, the effective cpumasks of this cpuset
877  * and all its descendants need to be updated.
878  *
879  * On legacy hierachy, effective_cpus will be the same with cpu_allowed.
880  *
881  * Called with cpuset_mutex held
882  */
883 static void update_cpumasks_hier(struct cpuset *cs, struct cpumask *new_cpus)
884 {
885 	struct cpuset *cp;
886 	struct cgroup_subsys_state *pos_css;
887 	bool need_rebuild_sched_domains = false;
888 
889 	rcu_read_lock();
890 	cpuset_for_each_descendant_pre(cp, pos_css, cs) {
891 		struct cpuset *parent = parent_cs(cp);
892 
893 		cpumask_and(new_cpus, cp->cpus_allowed, parent->effective_cpus);
894 
895 		/*
896 		 * If it becomes empty, inherit the effective mask of the
897 		 * parent, which is guaranteed to have some CPUs.
898 		 */
899 		if (cgroup_subsys_on_dfl(cpuset_cgrp_subsys) &&
900 		    cpumask_empty(new_cpus))
901 			cpumask_copy(new_cpus, parent->effective_cpus);
902 
903 		/* Skip the whole subtree if the cpumask remains the same. */
904 		if (cpumask_equal(new_cpus, cp->effective_cpus)) {
905 			pos_css = css_rightmost_descendant(pos_css);
906 			continue;
907 		}
908 
909 		if (!css_tryget_online(&cp->css))
910 			continue;
911 		rcu_read_unlock();
912 
913 		spin_lock_irq(&callback_lock);
914 		cpumask_copy(cp->effective_cpus, new_cpus);
915 		spin_unlock_irq(&callback_lock);
916 
917 		WARN_ON(!cgroup_subsys_on_dfl(cpuset_cgrp_subsys) &&
918 			!cpumask_equal(cp->cpus_allowed, cp->effective_cpus));
919 
920 		update_tasks_cpumask(cp);
921 
922 		/*
923 		 * If the effective cpumask of any non-empty cpuset is changed,
924 		 * we need to rebuild sched domains.
925 		 */
926 		if (!cpumask_empty(cp->cpus_allowed) &&
927 		    is_sched_load_balance(cp))
928 			need_rebuild_sched_domains = true;
929 
930 		rcu_read_lock();
931 		css_put(&cp->css);
932 	}
933 	rcu_read_unlock();
934 
935 	if (need_rebuild_sched_domains)
936 		rebuild_sched_domains_locked();
937 }
938 
939 /**
940  * update_cpumask - update the cpus_allowed mask of a cpuset and all tasks in it
941  * @cs: the cpuset to consider
942  * @trialcs: trial cpuset
943  * @buf: buffer of cpu numbers written to this cpuset
944  */
945 static int update_cpumask(struct cpuset *cs, struct cpuset *trialcs,
946 			  const char *buf)
947 {
948 	int retval;
949 
950 	/* top_cpuset.cpus_allowed tracks cpu_online_mask; it's read-only */
951 	if (cs == &top_cpuset)
952 		return -EACCES;
953 
954 	/*
955 	 * An empty cpus_allowed is ok only if the cpuset has no tasks.
956 	 * Since cpulist_parse() fails on an empty mask, we special case
957 	 * that parsing.  The validate_change() call ensures that cpusets
958 	 * with tasks have cpus.
959 	 */
960 	if (!*buf) {
961 		cpumask_clear(trialcs->cpus_allowed);
962 	} else {
963 		retval = cpulist_parse(buf, trialcs->cpus_allowed);
964 		if (retval < 0)
965 			return retval;
966 
967 		if (!cpumask_subset(trialcs->cpus_allowed,
968 				    top_cpuset.cpus_allowed))
969 			return -EINVAL;
970 	}
971 
972 	/* Nothing to do if the cpus didn't change */
973 	if (cpumask_equal(cs->cpus_allowed, trialcs->cpus_allowed))
974 		return 0;
975 
976 	retval = validate_change(cs, trialcs);
977 	if (retval < 0)
978 		return retval;
979 
980 	spin_lock_irq(&callback_lock);
981 	cpumask_copy(cs->cpus_allowed, trialcs->cpus_allowed);
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 
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 
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 
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  * We use the mems_allowed_seq seqlock to safely update both tsk->mems_allowed
1043  * and rebind an eventual tasks' mempolicy. If the task is allocating in
1044  * parallel, it might temporarily see an empty intersection, which results in
1045  * a seqlock check and retry before OOM or allocation failure.
1046  */
1047 static void cpuset_change_task_nodemask(struct task_struct *tsk,
1048 					nodemask_t *newmems)
1049 {
1050 	task_lock(tsk);
1051 
1052 	local_irq_disable();
1053 	write_seqcount_begin(&tsk->mems_allowed_seq);
1054 
1055 	nodes_or(tsk->mems_allowed, tsk->mems_allowed, *newmems);
1056 	mpol_rebind_task(tsk, newmems);
1057 	tsk->mems_allowed = *newmems;
1058 
1059 	write_seqcount_end(&tsk->mems_allowed_seq);
1060 	local_irq_enable();
1061 
1062 	task_unlock(tsk);
1063 }
1064 
1065 static void *cpuset_being_rebound;
1066 
1067 /**
1068  * update_tasks_nodemask - Update the nodemasks of tasks in the cpuset.
1069  * @cs: the cpuset in which each task's mems_allowed mask needs to be changed
1070  *
1071  * Iterate through each task of @cs updating its mems_allowed to the
1072  * effective cpuset's.  As this function is called with cpuset_mutex held,
1073  * cpuset membership stays stable.
1074  */
1075 static void update_tasks_nodemask(struct cpuset *cs)
1076 {
1077 	static nodemask_t newmems;	/* protected by cpuset_mutex */
1078 	struct css_task_iter it;
1079 	struct task_struct *task;
1080 
1081 	cpuset_being_rebound = cs;		/* causes mpol_dup() rebind */
1082 
1083 	guarantee_online_mems(cs, &newmems);
1084 
1085 	/*
1086 	 * The mpol_rebind_mm() call takes mmap_sem, which we couldn't
1087 	 * take while holding tasklist_lock.  Forks can happen - the
1088 	 * mpol_dup() cpuset_being_rebound check will catch such forks,
1089 	 * and rebind their vma mempolicies too.  Because we still hold
1090 	 * the global cpuset_mutex, we know that no other rebind effort
1091 	 * will be contending for the global variable cpuset_being_rebound.
1092 	 * It's ok if we rebind the same mm twice; mpol_rebind_mm()
1093 	 * is idempotent.  Also migrate pages in each mm to new nodes.
1094 	 */
1095 	css_task_iter_start(&cs->css, &it);
1096 	while ((task = css_task_iter_next(&it))) {
1097 		struct mm_struct *mm;
1098 		bool migrate;
1099 
1100 		cpuset_change_task_nodemask(task, &newmems);
1101 
1102 		mm = get_task_mm(task);
1103 		if (!mm)
1104 			continue;
1105 
1106 		migrate = is_memory_migrate(cs);
1107 
1108 		mpol_rebind_mm(mm, &cs->mems_allowed);
1109 		if (migrate)
1110 			cpuset_migrate_mm(mm, &cs->old_mems_allowed, &newmems);
1111 		else
1112 			mmput(mm);
1113 	}
1114 	css_task_iter_end(&it);
1115 
1116 	/*
1117 	 * All the tasks' nodemasks have been updated, update
1118 	 * cs->old_mems_allowed.
1119 	 */
1120 	cs->old_mems_allowed = newmems;
1121 
1122 	/* We're done rebinding vmas to this cpuset's new mems_allowed. */
1123 	cpuset_being_rebound = NULL;
1124 }
1125 
1126 /*
1127  * update_nodemasks_hier - Update effective nodemasks and tasks in the subtree
1128  * @cs: the cpuset to consider
1129  * @new_mems: a temp variable for calculating new effective_mems
1130  *
1131  * When configured nodemask is changed, the effective nodemasks of this cpuset
1132  * and all its descendants need to be updated.
1133  *
1134  * On legacy hiearchy, effective_mems will be the same with mems_allowed.
1135  *
1136  * Called with cpuset_mutex held
1137  */
1138 static void update_nodemasks_hier(struct cpuset *cs, nodemask_t *new_mems)
1139 {
1140 	struct cpuset *cp;
1141 	struct cgroup_subsys_state *pos_css;
1142 
1143 	rcu_read_lock();
1144 	cpuset_for_each_descendant_pre(cp, pos_css, cs) {
1145 		struct cpuset *parent = parent_cs(cp);
1146 
1147 		nodes_and(*new_mems, cp->mems_allowed, parent->effective_mems);
1148 
1149 		/*
1150 		 * If it becomes empty, inherit the effective mask of the
1151 		 * parent, which is guaranteed to have some MEMs.
1152 		 */
1153 		if (cgroup_subsys_on_dfl(cpuset_cgrp_subsys) &&
1154 		    nodes_empty(*new_mems))
1155 			*new_mems = parent->effective_mems;
1156 
1157 		/* Skip the whole subtree if the nodemask remains the same. */
1158 		if (nodes_equal(*new_mems, cp->effective_mems)) {
1159 			pos_css = css_rightmost_descendant(pos_css);
1160 			continue;
1161 		}
1162 
1163 		if (!css_tryget_online(&cp->css))
1164 			continue;
1165 		rcu_read_unlock();
1166 
1167 		spin_lock_irq(&callback_lock);
1168 		cp->effective_mems = *new_mems;
1169 		spin_unlock_irq(&callback_lock);
1170 
1171 		WARN_ON(!cgroup_subsys_on_dfl(cpuset_cgrp_subsys) &&
1172 			!nodes_equal(cp->mems_allowed, cp->effective_mems));
1173 
1174 		update_tasks_nodemask(cp);
1175 
1176 		rcu_read_lock();
1177 		css_put(&cp->css);
1178 	}
1179 	rcu_read_unlock();
1180 }
1181 
1182 /*
1183  * Handle user request to change the 'mems' memory placement
1184  * of a cpuset.  Needs to validate the request, update the
1185  * cpusets mems_allowed, and for each task in the cpuset,
1186  * update mems_allowed and rebind task's mempolicy and any vma
1187  * mempolicies and if the cpuset is marked 'memory_migrate',
1188  * migrate the tasks pages to the new memory.
1189  *
1190  * Call with cpuset_mutex held. May take callback_lock during call.
1191  * Will take tasklist_lock, scan tasklist for tasks in cpuset cs,
1192  * lock each such tasks mm->mmap_sem, scan its vma's and rebind
1193  * their mempolicies to the cpusets new mems_allowed.
1194  */
1195 static int update_nodemask(struct cpuset *cs, struct cpuset *trialcs,
1196 			   const char *buf)
1197 {
1198 	int retval;
1199 
1200 	/*
1201 	 * top_cpuset.mems_allowed tracks node_stats[N_MEMORY];
1202 	 * it's read-only
1203 	 */
1204 	if (cs == &top_cpuset) {
1205 		retval = -EACCES;
1206 		goto done;
1207 	}
1208 
1209 	/*
1210 	 * An empty mems_allowed is ok iff there are no tasks in the cpuset.
1211 	 * Since nodelist_parse() fails on an empty mask, we special case
1212 	 * that parsing.  The validate_change() call ensures that cpusets
1213 	 * with tasks have memory.
1214 	 */
1215 	if (!*buf) {
1216 		nodes_clear(trialcs->mems_allowed);
1217 	} else {
1218 		retval = nodelist_parse(buf, trialcs->mems_allowed);
1219 		if (retval < 0)
1220 			goto done;
1221 
1222 		if (!nodes_subset(trialcs->mems_allowed,
1223 				  top_cpuset.mems_allowed)) {
1224 			retval = -EINVAL;
1225 			goto done;
1226 		}
1227 	}
1228 
1229 	if (nodes_equal(cs->mems_allowed, trialcs->mems_allowed)) {
1230 		retval = 0;		/* Too easy - nothing to do */
1231 		goto done;
1232 	}
1233 	retval = validate_change(cs, trialcs);
1234 	if (retval < 0)
1235 		goto done;
1236 
1237 	spin_lock_irq(&callback_lock);
1238 	cs->mems_allowed = trialcs->mems_allowed;
1239 	spin_unlock_irq(&callback_lock);
1240 
1241 	/* use trialcs->mems_allowed as a temp variable */
1242 	update_nodemasks_hier(cs, &trialcs->mems_allowed);
1243 done:
1244 	return retval;
1245 }
1246 
1247 int current_cpuset_is_being_rebound(void)
1248 {
1249 	int ret;
1250 
1251 	rcu_read_lock();
1252 	ret = task_cs(current) == cpuset_being_rebound;
1253 	rcu_read_unlock();
1254 
1255 	return ret;
1256 }
1257 
1258 static int update_relax_domain_level(struct cpuset *cs, s64 val)
1259 {
1260 #ifdef CONFIG_SMP
1261 	if (val < -1 || val >= sched_domain_level_max)
1262 		return -EINVAL;
1263 #endif
1264 
1265 	if (val != cs->relax_domain_level) {
1266 		cs->relax_domain_level = val;
1267 		if (!cpumask_empty(cs->cpus_allowed) &&
1268 		    is_sched_load_balance(cs))
1269 			rebuild_sched_domains_locked();
1270 	}
1271 
1272 	return 0;
1273 }
1274 
1275 /**
1276  * update_tasks_flags - update the spread flags of tasks in the cpuset.
1277  * @cs: the cpuset in which each task's spread flags needs to be changed
1278  *
1279  * Iterate through each task of @cs updating its spread flags.  As this
1280  * function is called with cpuset_mutex held, cpuset membership stays
1281  * stable.
1282  */
1283 static void update_tasks_flags(struct cpuset *cs)
1284 {
1285 	struct css_task_iter it;
1286 	struct task_struct *task;
1287 
1288 	css_task_iter_start(&cs->css, &it);
1289 	while ((task = css_task_iter_next(&it)))
1290 		cpuset_update_task_spread_flag(cs, task);
1291 	css_task_iter_end(&it);
1292 }
1293 
1294 /*
1295  * update_flag - read a 0 or a 1 in a file and update associated flag
1296  * bit:		the bit to update (see cpuset_flagbits_t)
1297  * cs:		the cpuset to update
1298  * turning_on: 	whether the flag is being set or cleared
1299  *
1300  * Call with cpuset_mutex held.
1301  */
1302 
1303 static int update_flag(cpuset_flagbits_t bit, struct cpuset *cs,
1304 		       int turning_on)
1305 {
1306 	struct cpuset *trialcs;
1307 	int balance_flag_changed;
1308 	int spread_flag_changed;
1309 	int err;
1310 
1311 	trialcs = alloc_trial_cpuset(cs);
1312 	if (!trialcs)
1313 		return -ENOMEM;
1314 
1315 	if (turning_on)
1316 		set_bit(bit, &trialcs->flags);
1317 	else
1318 		clear_bit(bit, &trialcs->flags);
1319 
1320 	err = validate_change(cs, trialcs);
1321 	if (err < 0)
1322 		goto out;
1323 
1324 	balance_flag_changed = (is_sched_load_balance(cs) !=
1325 				is_sched_load_balance(trialcs));
1326 
1327 	spread_flag_changed = ((is_spread_slab(cs) != is_spread_slab(trialcs))
1328 			|| (is_spread_page(cs) != is_spread_page(trialcs)));
1329 
1330 	spin_lock_irq(&callback_lock);
1331 	cs->flags = trialcs->flags;
1332 	spin_unlock_irq(&callback_lock);
1333 
1334 	if (!cpumask_empty(trialcs->cpus_allowed) && balance_flag_changed)
1335 		rebuild_sched_domains_locked();
1336 
1337 	if (spread_flag_changed)
1338 		update_tasks_flags(cs);
1339 out:
1340 	free_trial_cpuset(trialcs);
1341 	return err;
1342 }
1343 
1344 /*
1345  * Frequency meter - How fast is some event occurring?
1346  *
1347  * These routines manage a digitally filtered, constant time based,
1348  * event frequency meter.  There are four routines:
1349  *   fmeter_init() - initialize a frequency meter.
1350  *   fmeter_markevent() - called each time the event happens.
1351  *   fmeter_getrate() - returns the recent rate of such events.
1352  *   fmeter_update() - internal routine used to update fmeter.
1353  *
1354  * A common data structure is passed to each of these routines,
1355  * which is used to keep track of the state required to manage the
1356  * frequency meter and its digital filter.
1357  *
1358  * The filter works on the number of events marked per unit time.
1359  * The filter is single-pole low-pass recursive (IIR).  The time unit
1360  * is 1 second.  Arithmetic is done using 32-bit integers scaled to
1361  * simulate 3 decimal digits of precision (multiplied by 1000).
1362  *
1363  * With an FM_COEF of 933, and a time base of 1 second, the filter
1364  * has a half-life of 10 seconds, meaning that if the events quit
1365  * happening, then the rate returned from the fmeter_getrate()
1366  * will be cut in half each 10 seconds, until it converges to zero.
1367  *
1368  * It is not worth doing a real infinitely recursive filter.  If more
1369  * than FM_MAXTICKS ticks have elapsed since the last filter event,
1370  * just compute FM_MAXTICKS ticks worth, by which point the level
1371  * will be stable.
1372  *
1373  * Limit the count of unprocessed events to FM_MAXCNT, so as to avoid
1374  * arithmetic overflow in the fmeter_update() routine.
1375  *
1376  * Given the simple 32 bit integer arithmetic used, this meter works
1377  * best for reporting rates between one per millisecond (msec) and
1378  * one per 32 (approx) seconds.  At constant rates faster than one
1379  * per msec it maxes out at values just under 1,000,000.  At constant
1380  * rates between one per msec, and one per second it will stabilize
1381  * to a value N*1000, where N is the rate of events per second.
1382  * At constant rates between one per second and one per 32 seconds,
1383  * it will be choppy, moving up on the seconds that have an event,
1384  * and then decaying until the next event.  At rates slower than
1385  * about one in 32 seconds, it decays all the way back to zero between
1386  * each event.
1387  */
1388 
1389 #define FM_COEF 933		/* coefficient for half-life of 10 secs */
1390 #define FM_MAXTICKS ((u32)99)   /* useless computing more ticks than this */
1391 #define FM_MAXCNT 1000000	/* limit cnt to avoid overflow */
1392 #define FM_SCALE 1000		/* faux fixed point scale */
1393 
1394 /* Initialize a frequency meter */
1395 static void fmeter_init(struct fmeter *fmp)
1396 {
1397 	fmp->cnt = 0;
1398 	fmp->val = 0;
1399 	fmp->time = 0;
1400 	spin_lock_init(&fmp->lock);
1401 }
1402 
1403 /* Internal meter update - process cnt events and update value */
1404 static void fmeter_update(struct fmeter *fmp)
1405 {
1406 	time64_t now;
1407 	u32 ticks;
1408 
1409 	now = ktime_get_seconds();
1410 	ticks = now - fmp->time;
1411 
1412 	if (ticks == 0)
1413 		return;
1414 
1415 	ticks = min(FM_MAXTICKS, ticks);
1416 	while (ticks-- > 0)
1417 		fmp->val = (FM_COEF * fmp->val) / FM_SCALE;
1418 	fmp->time = now;
1419 
1420 	fmp->val += ((FM_SCALE - FM_COEF) * fmp->cnt) / FM_SCALE;
1421 	fmp->cnt = 0;
1422 }
1423 
1424 /* Process any previous ticks, then bump cnt by one (times scale). */
1425 static void fmeter_markevent(struct fmeter *fmp)
1426 {
1427 	spin_lock(&fmp->lock);
1428 	fmeter_update(fmp);
1429 	fmp->cnt = min(FM_MAXCNT, fmp->cnt + FM_SCALE);
1430 	spin_unlock(&fmp->lock);
1431 }
1432 
1433 /* Process any previous ticks, then return current value. */
1434 static int fmeter_getrate(struct fmeter *fmp)
1435 {
1436 	int val;
1437 
1438 	spin_lock(&fmp->lock);
1439 	fmeter_update(fmp);
1440 	val = fmp->val;
1441 	spin_unlock(&fmp->lock);
1442 	return val;
1443 }
1444 
1445 static struct cpuset *cpuset_attach_old_cs;
1446 
1447 /* Called by cgroups to determine if a cpuset is usable; cpuset_mutex held */
1448 static int cpuset_can_attach(struct cgroup_taskset *tset)
1449 {
1450 	struct cgroup_subsys_state *css;
1451 	struct cpuset *cs;
1452 	struct task_struct *task;
1453 	int ret;
1454 
1455 	/* used later by cpuset_attach() */
1456 	cpuset_attach_old_cs = task_cs(cgroup_taskset_first(tset, &css));
1457 	cs = css_cs(css);
1458 
1459 	mutex_lock(&cpuset_mutex);
1460 
1461 	/* allow moving tasks into an empty cpuset if on default hierarchy */
1462 	ret = -ENOSPC;
1463 	if (!cgroup_subsys_on_dfl(cpuset_cgrp_subsys) &&
1464 	    (cpumask_empty(cs->cpus_allowed) || nodes_empty(cs->mems_allowed)))
1465 		goto out_unlock;
1466 
1467 	cgroup_taskset_for_each(task, css, tset) {
1468 		ret = task_can_attach(task, cs->cpus_allowed);
1469 		if (ret)
1470 			goto out_unlock;
1471 		ret = security_task_setscheduler(task);
1472 		if (ret)
1473 			goto out_unlock;
1474 	}
1475 
1476 	/*
1477 	 * Mark attach is in progress.  This makes validate_change() fail
1478 	 * changes which zero cpus/mems_allowed.
1479 	 */
1480 	cs->attach_in_progress++;
1481 	ret = 0;
1482 out_unlock:
1483 	mutex_unlock(&cpuset_mutex);
1484 	return ret;
1485 }
1486 
1487 static void cpuset_cancel_attach(struct cgroup_taskset *tset)
1488 {
1489 	struct cgroup_subsys_state *css;
1490 	struct cpuset *cs;
1491 
1492 	cgroup_taskset_first(tset, &css);
1493 	cs = css_cs(css);
1494 
1495 	mutex_lock(&cpuset_mutex);
1496 	css_cs(css)->attach_in_progress--;
1497 	mutex_unlock(&cpuset_mutex);
1498 }
1499 
1500 /*
1501  * Protected by cpuset_mutex.  cpus_attach is used only by cpuset_attach()
1502  * but we can't allocate it dynamically there.  Define it global and
1503  * allocate from cpuset_init().
1504  */
1505 static cpumask_var_t cpus_attach;
1506 
1507 static void cpuset_attach(struct cgroup_taskset *tset)
1508 {
1509 	/* static buf protected by cpuset_mutex */
1510 	static nodemask_t cpuset_attach_nodemask_to;
1511 	struct task_struct *task;
1512 	struct task_struct *leader;
1513 	struct cgroup_subsys_state *css;
1514 	struct cpuset *cs;
1515 	struct cpuset *oldcs = cpuset_attach_old_cs;
1516 
1517 	cgroup_taskset_first(tset, &css);
1518 	cs = css_cs(css);
1519 
1520 	mutex_lock(&cpuset_mutex);
1521 
1522 	/* prepare for attach */
1523 	if (cs == &top_cpuset)
1524 		cpumask_copy(cpus_attach, cpu_possible_mask);
1525 	else
1526 		guarantee_online_cpus(cs, cpus_attach);
1527 
1528 	guarantee_online_mems(cs, &cpuset_attach_nodemask_to);
1529 
1530 	cgroup_taskset_for_each(task, css, tset) {
1531 		/*
1532 		 * can_attach beforehand should guarantee that this doesn't
1533 		 * fail.  TODO: have a better way to handle failure here
1534 		 */
1535 		WARN_ON_ONCE(set_cpus_allowed_ptr(task, cpus_attach));
1536 
1537 		cpuset_change_task_nodemask(task, &cpuset_attach_nodemask_to);
1538 		cpuset_update_task_spread_flag(cs, task);
1539 	}
1540 
1541 	/*
1542 	 * Change mm for all threadgroup leaders. This is expensive and may
1543 	 * sleep and should be moved outside migration path proper.
1544 	 */
1545 	cpuset_attach_nodemask_to = cs->effective_mems;
1546 	cgroup_taskset_for_each_leader(leader, css, tset) {
1547 		struct mm_struct *mm = get_task_mm(leader);
1548 
1549 		if (mm) {
1550 			mpol_rebind_mm(mm, &cpuset_attach_nodemask_to);
1551 
1552 			/*
1553 			 * old_mems_allowed is the same with mems_allowed
1554 			 * here, except if this task is being moved
1555 			 * automatically due to hotplug.  In that case
1556 			 * @mems_allowed has been updated and is empty, so
1557 			 * @old_mems_allowed is the right nodesets that we
1558 			 * migrate mm from.
1559 			 */
1560 			if (is_memory_migrate(cs))
1561 				cpuset_migrate_mm(mm, &oldcs->old_mems_allowed,
1562 						  &cpuset_attach_nodemask_to);
1563 			else
1564 				mmput(mm);
1565 		}
1566 	}
1567 
1568 	cs->old_mems_allowed = cpuset_attach_nodemask_to;
1569 
1570 	cs->attach_in_progress--;
1571 	if (!cs->attach_in_progress)
1572 		wake_up(&cpuset_attach_wq);
1573 
1574 	mutex_unlock(&cpuset_mutex);
1575 }
1576 
1577 /* The various types of files and directories in a cpuset file system */
1578 
1579 typedef enum {
1580 	FILE_MEMORY_MIGRATE,
1581 	FILE_CPULIST,
1582 	FILE_MEMLIST,
1583 	FILE_EFFECTIVE_CPULIST,
1584 	FILE_EFFECTIVE_MEMLIST,
1585 	FILE_CPU_EXCLUSIVE,
1586 	FILE_MEM_EXCLUSIVE,
1587 	FILE_MEM_HARDWALL,
1588 	FILE_SCHED_LOAD_BALANCE,
1589 	FILE_SCHED_RELAX_DOMAIN_LEVEL,
1590 	FILE_MEMORY_PRESSURE_ENABLED,
1591 	FILE_MEMORY_PRESSURE,
1592 	FILE_SPREAD_PAGE,
1593 	FILE_SPREAD_SLAB,
1594 } cpuset_filetype_t;
1595 
1596 static int cpuset_write_u64(struct cgroup_subsys_state *css, struct cftype *cft,
1597 			    u64 val)
1598 {
1599 	struct cpuset *cs = css_cs(css);
1600 	cpuset_filetype_t type = cft->private;
1601 	int retval = 0;
1602 
1603 	mutex_lock(&cpuset_mutex);
1604 	if (!is_cpuset_online(cs)) {
1605 		retval = -ENODEV;
1606 		goto out_unlock;
1607 	}
1608 
1609 	switch (type) {
1610 	case FILE_CPU_EXCLUSIVE:
1611 		retval = update_flag(CS_CPU_EXCLUSIVE, cs, val);
1612 		break;
1613 	case FILE_MEM_EXCLUSIVE:
1614 		retval = update_flag(CS_MEM_EXCLUSIVE, cs, val);
1615 		break;
1616 	case FILE_MEM_HARDWALL:
1617 		retval = update_flag(CS_MEM_HARDWALL, cs, val);
1618 		break;
1619 	case FILE_SCHED_LOAD_BALANCE:
1620 		retval = update_flag(CS_SCHED_LOAD_BALANCE, cs, val);
1621 		break;
1622 	case FILE_MEMORY_MIGRATE:
1623 		retval = update_flag(CS_MEMORY_MIGRATE, cs, val);
1624 		break;
1625 	case FILE_MEMORY_PRESSURE_ENABLED:
1626 		cpuset_memory_pressure_enabled = !!val;
1627 		break;
1628 	case FILE_SPREAD_PAGE:
1629 		retval = update_flag(CS_SPREAD_PAGE, cs, val);
1630 		break;
1631 	case FILE_SPREAD_SLAB:
1632 		retval = update_flag(CS_SPREAD_SLAB, cs, val);
1633 		break;
1634 	default:
1635 		retval = -EINVAL;
1636 		break;
1637 	}
1638 out_unlock:
1639 	mutex_unlock(&cpuset_mutex);
1640 	return retval;
1641 }
1642 
1643 static int cpuset_write_s64(struct cgroup_subsys_state *css, struct cftype *cft,
1644 			    s64 val)
1645 {
1646 	struct cpuset *cs = css_cs(css);
1647 	cpuset_filetype_t type = cft->private;
1648 	int retval = -ENODEV;
1649 
1650 	mutex_lock(&cpuset_mutex);
1651 	if (!is_cpuset_online(cs))
1652 		goto out_unlock;
1653 
1654 	switch (type) {
1655 	case FILE_SCHED_RELAX_DOMAIN_LEVEL:
1656 		retval = update_relax_domain_level(cs, val);
1657 		break;
1658 	default:
1659 		retval = -EINVAL;
1660 		break;
1661 	}
1662 out_unlock:
1663 	mutex_unlock(&cpuset_mutex);
1664 	return retval;
1665 }
1666 
1667 /*
1668  * Common handling for a write to a "cpus" or "mems" file.
1669  */
1670 static ssize_t cpuset_write_resmask(struct kernfs_open_file *of,
1671 				    char *buf, size_t nbytes, loff_t off)
1672 {
1673 	struct cpuset *cs = css_cs(of_css(of));
1674 	struct cpuset *trialcs;
1675 	int retval = -ENODEV;
1676 
1677 	buf = strstrip(buf);
1678 
1679 	/*
1680 	 * CPU or memory hotunplug may leave @cs w/o any execution
1681 	 * resources, in which case the hotplug code asynchronously updates
1682 	 * configuration and transfers all tasks to the nearest ancestor
1683 	 * which can execute.
1684 	 *
1685 	 * As writes to "cpus" or "mems" may restore @cs's execution
1686 	 * resources, wait for the previously scheduled operations before
1687 	 * proceeding, so that we don't end up keep removing tasks added
1688 	 * after execution capability is restored.
1689 	 *
1690 	 * cpuset_hotplug_work calls back into cgroup core via
1691 	 * cgroup_transfer_tasks() and waiting for it from a cgroupfs
1692 	 * operation like this one can lead to a deadlock through kernfs
1693 	 * active_ref protection.  Let's break the protection.  Losing the
1694 	 * protection is okay as we check whether @cs is online after
1695 	 * grabbing cpuset_mutex anyway.  This only happens on the legacy
1696 	 * hierarchies.
1697 	 */
1698 	css_get(&cs->css);
1699 	kernfs_break_active_protection(of->kn);
1700 	flush_work(&cpuset_hotplug_work);
1701 
1702 	mutex_lock(&cpuset_mutex);
1703 	if (!is_cpuset_online(cs))
1704 		goto out_unlock;
1705 
1706 	trialcs = alloc_trial_cpuset(cs);
1707 	if (!trialcs) {
1708 		retval = -ENOMEM;
1709 		goto out_unlock;
1710 	}
1711 
1712 	switch (of_cft(of)->private) {
1713 	case FILE_CPULIST:
1714 		retval = update_cpumask(cs, trialcs, buf);
1715 		break;
1716 	case FILE_MEMLIST:
1717 		retval = update_nodemask(cs, trialcs, buf);
1718 		break;
1719 	default:
1720 		retval = -EINVAL;
1721 		break;
1722 	}
1723 
1724 	free_trial_cpuset(trialcs);
1725 out_unlock:
1726 	mutex_unlock(&cpuset_mutex);
1727 	kernfs_unbreak_active_protection(of->kn);
1728 	css_put(&cs->css);
1729 	flush_workqueue(cpuset_migrate_mm_wq);
1730 	return retval ?: nbytes;
1731 }
1732 
1733 /*
1734  * These ascii lists should be read in a single call, by using a user
1735  * buffer large enough to hold the entire map.  If read in smaller
1736  * chunks, there is no guarantee of atomicity.  Since the display format
1737  * used, list of ranges of sequential numbers, is variable length,
1738  * and since these maps can change value dynamically, one could read
1739  * gibberish by doing partial reads while a list was changing.
1740  */
1741 static int cpuset_common_seq_show(struct seq_file *sf, void *v)
1742 {
1743 	struct cpuset *cs = css_cs(seq_css(sf));
1744 	cpuset_filetype_t type = seq_cft(sf)->private;
1745 	int ret = 0;
1746 
1747 	spin_lock_irq(&callback_lock);
1748 
1749 	switch (type) {
1750 	case FILE_CPULIST:
1751 		seq_printf(sf, "%*pbl\n", cpumask_pr_args(cs->cpus_allowed));
1752 		break;
1753 	case FILE_MEMLIST:
1754 		seq_printf(sf, "%*pbl\n", nodemask_pr_args(&cs->mems_allowed));
1755 		break;
1756 	case FILE_EFFECTIVE_CPULIST:
1757 		seq_printf(sf, "%*pbl\n", cpumask_pr_args(cs->effective_cpus));
1758 		break;
1759 	case FILE_EFFECTIVE_MEMLIST:
1760 		seq_printf(sf, "%*pbl\n", nodemask_pr_args(&cs->effective_mems));
1761 		break;
1762 	default:
1763 		ret = -EINVAL;
1764 	}
1765 
1766 	spin_unlock_irq(&callback_lock);
1767 	return ret;
1768 }
1769 
1770 static u64 cpuset_read_u64(struct cgroup_subsys_state *css, struct cftype *cft)
1771 {
1772 	struct cpuset *cs = css_cs(css);
1773 	cpuset_filetype_t type = cft->private;
1774 	switch (type) {
1775 	case FILE_CPU_EXCLUSIVE:
1776 		return is_cpu_exclusive(cs);
1777 	case FILE_MEM_EXCLUSIVE:
1778 		return is_mem_exclusive(cs);
1779 	case FILE_MEM_HARDWALL:
1780 		return is_mem_hardwall(cs);
1781 	case FILE_SCHED_LOAD_BALANCE:
1782 		return is_sched_load_balance(cs);
1783 	case FILE_MEMORY_MIGRATE:
1784 		return is_memory_migrate(cs);
1785 	case FILE_MEMORY_PRESSURE_ENABLED:
1786 		return cpuset_memory_pressure_enabled;
1787 	case FILE_MEMORY_PRESSURE:
1788 		return fmeter_getrate(&cs->fmeter);
1789 	case FILE_SPREAD_PAGE:
1790 		return is_spread_page(cs);
1791 	case FILE_SPREAD_SLAB:
1792 		return is_spread_slab(cs);
1793 	default:
1794 		BUG();
1795 	}
1796 
1797 	/* Unreachable but makes gcc happy */
1798 	return 0;
1799 }
1800 
1801 static s64 cpuset_read_s64(struct cgroup_subsys_state *css, struct cftype *cft)
1802 {
1803 	struct cpuset *cs = css_cs(css);
1804 	cpuset_filetype_t type = cft->private;
1805 	switch (type) {
1806 	case FILE_SCHED_RELAX_DOMAIN_LEVEL:
1807 		return cs->relax_domain_level;
1808 	default:
1809 		BUG();
1810 	}
1811 
1812 	/* Unrechable but makes gcc happy */
1813 	return 0;
1814 }
1815 
1816 
1817 /*
1818  * for the common functions, 'private' gives the type of file
1819  */
1820 
1821 static struct cftype files[] = {
1822 	{
1823 		.name = "cpus",
1824 		.seq_show = cpuset_common_seq_show,
1825 		.write = cpuset_write_resmask,
1826 		.max_write_len = (100U + 6 * NR_CPUS),
1827 		.private = FILE_CPULIST,
1828 	},
1829 
1830 	{
1831 		.name = "mems",
1832 		.seq_show = cpuset_common_seq_show,
1833 		.write = cpuset_write_resmask,
1834 		.max_write_len = (100U + 6 * MAX_NUMNODES),
1835 		.private = FILE_MEMLIST,
1836 	},
1837 
1838 	{
1839 		.name = "effective_cpus",
1840 		.seq_show = cpuset_common_seq_show,
1841 		.private = FILE_EFFECTIVE_CPULIST,
1842 	},
1843 
1844 	{
1845 		.name = "effective_mems",
1846 		.seq_show = cpuset_common_seq_show,
1847 		.private = FILE_EFFECTIVE_MEMLIST,
1848 	},
1849 
1850 	{
1851 		.name = "cpu_exclusive",
1852 		.read_u64 = cpuset_read_u64,
1853 		.write_u64 = cpuset_write_u64,
1854 		.private = FILE_CPU_EXCLUSIVE,
1855 	},
1856 
1857 	{
1858 		.name = "mem_exclusive",
1859 		.read_u64 = cpuset_read_u64,
1860 		.write_u64 = cpuset_write_u64,
1861 		.private = FILE_MEM_EXCLUSIVE,
1862 	},
1863 
1864 	{
1865 		.name = "mem_hardwall",
1866 		.read_u64 = cpuset_read_u64,
1867 		.write_u64 = cpuset_write_u64,
1868 		.private = FILE_MEM_HARDWALL,
1869 	},
1870 
1871 	{
1872 		.name = "sched_load_balance",
1873 		.read_u64 = cpuset_read_u64,
1874 		.write_u64 = cpuset_write_u64,
1875 		.private = FILE_SCHED_LOAD_BALANCE,
1876 	},
1877 
1878 	{
1879 		.name = "sched_relax_domain_level",
1880 		.read_s64 = cpuset_read_s64,
1881 		.write_s64 = cpuset_write_s64,
1882 		.private = FILE_SCHED_RELAX_DOMAIN_LEVEL,
1883 	},
1884 
1885 	{
1886 		.name = "memory_migrate",
1887 		.read_u64 = cpuset_read_u64,
1888 		.write_u64 = cpuset_write_u64,
1889 		.private = FILE_MEMORY_MIGRATE,
1890 	},
1891 
1892 	{
1893 		.name = "memory_pressure",
1894 		.read_u64 = cpuset_read_u64,
1895 	},
1896 
1897 	{
1898 		.name = "memory_spread_page",
1899 		.read_u64 = cpuset_read_u64,
1900 		.write_u64 = cpuset_write_u64,
1901 		.private = FILE_SPREAD_PAGE,
1902 	},
1903 
1904 	{
1905 		.name = "memory_spread_slab",
1906 		.read_u64 = cpuset_read_u64,
1907 		.write_u64 = cpuset_write_u64,
1908 		.private = FILE_SPREAD_SLAB,
1909 	},
1910 
1911 	{
1912 		.name = "memory_pressure_enabled",
1913 		.flags = CFTYPE_ONLY_ON_ROOT,
1914 		.read_u64 = cpuset_read_u64,
1915 		.write_u64 = cpuset_write_u64,
1916 		.private = FILE_MEMORY_PRESSURE_ENABLED,
1917 	},
1918 
1919 	{ }	/* terminate */
1920 };
1921 
1922 /*
1923  *	cpuset_css_alloc - allocate a cpuset css
1924  *	cgrp:	control group that the new cpuset will be part of
1925  */
1926 
1927 static struct cgroup_subsys_state *
1928 cpuset_css_alloc(struct cgroup_subsys_state *parent_css)
1929 {
1930 	struct cpuset *cs;
1931 
1932 	if (!parent_css)
1933 		return &top_cpuset.css;
1934 
1935 	cs = kzalloc(sizeof(*cs), GFP_KERNEL);
1936 	if (!cs)
1937 		return ERR_PTR(-ENOMEM);
1938 	if (!alloc_cpumask_var(&cs->cpus_allowed, GFP_KERNEL))
1939 		goto free_cs;
1940 	if (!alloc_cpumask_var(&cs->effective_cpus, GFP_KERNEL))
1941 		goto free_cpus;
1942 
1943 	set_bit(CS_SCHED_LOAD_BALANCE, &cs->flags);
1944 	cpumask_clear(cs->cpus_allowed);
1945 	nodes_clear(cs->mems_allowed);
1946 	cpumask_clear(cs->effective_cpus);
1947 	nodes_clear(cs->effective_mems);
1948 	fmeter_init(&cs->fmeter);
1949 	cs->relax_domain_level = -1;
1950 
1951 	return &cs->css;
1952 
1953 free_cpus:
1954 	free_cpumask_var(cs->cpus_allowed);
1955 free_cs:
1956 	kfree(cs);
1957 	return ERR_PTR(-ENOMEM);
1958 }
1959 
1960 static int cpuset_css_online(struct cgroup_subsys_state *css)
1961 {
1962 	struct cpuset *cs = css_cs(css);
1963 	struct cpuset *parent = parent_cs(cs);
1964 	struct cpuset *tmp_cs;
1965 	struct cgroup_subsys_state *pos_css;
1966 
1967 	if (!parent)
1968 		return 0;
1969 
1970 	mutex_lock(&cpuset_mutex);
1971 
1972 	set_bit(CS_ONLINE, &cs->flags);
1973 	if (is_spread_page(parent))
1974 		set_bit(CS_SPREAD_PAGE, &cs->flags);
1975 	if (is_spread_slab(parent))
1976 		set_bit(CS_SPREAD_SLAB, &cs->flags);
1977 
1978 	cpuset_inc();
1979 
1980 	spin_lock_irq(&callback_lock);
1981 	if (cgroup_subsys_on_dfl(cpuset_cgrp_subsys)) {
1982 		cpumask_copy(cs->effective_cpus, parent->effective_cpus);
1983 		cs->effective_mems = parent->effective_mems;
1984 	}
1985 	spin_unlock_irq(&callback_lock);
1986 
1987 	if (!test_bit(CGRP_CPUSET_CLONE_CHILDREN, &css->cgroup->flags))
1988 		goto out_unlock;
1989 
1990 	/*
1991 	 * Clone @parent's configuration if CGRP_CPUSET_CLONE_CHILDREN is
1992 	 * set.  This flag handling is implemented in cgroup core for
1993 	 * histrical reasons - the flag may be specified during mount.
1994 	 *
1995 	 * Currently, if any sibling cpusets have exclusive cpus or mem, we
1996 	 * refuse to clone the configuration - thereby refusing the task to
1997 	 * be entered, and as a result refusing the sys_unshare() or
1998 	 * clone() which initiated it.  If this becomes a problem for some
1999 	 * users who wish to allow that scenario, then this could be
2000 	 * changed to grant parent->cpus_allowed-sibling_cpus_exclusive
2001 	 * (and likewise for mems) to the new cgroup.
2002 	 */
2003 	rcu_read_lock();
2004 	cpuset_for_each_child(tmp_cs, pos_css, parent) {
2005 		if (is_mem_exclusive(tmp_cs) || is_cpu_exclusive(tmp_cs)) {
2006 			rcu_read_unlock();
2007 			goto out_unlock;
2008 		}
2009 	}
2010 	rcu_read_unlock();
2011 
2012 	spin_lock_irq(&callback_lock);
2013 	cs->mems_allowed = parent->mems_allowed;
2014 	cs->effective_mems = parent->mems_allowed;
2015 	cpumask_copy(cs->cpus_allowed, parent->cpus_allowed);
2016 	cpumask_copy(cs->effective_cpus, parent->cpus_allowed);
2017 	spin_unlock_irq(&callback_lock);
2018 out_unlock:
2019 	mutex_unlock(&cpuset_mutex);
2020 	return 0;
2021 }
2022 
2023 /*
2024  * If the cpuset being removed has its flag 'sched_load_balance'
2025  * enabled, then simulate turning sched_load_balance off, which
2026  * will call rebuild_sched_domains_locked().
2027  */
2028 
2029 static void cpuset_css_offline(struct cgroup_subsys_state *css)
2030 {
2031 	struct cpuset *cs = css_cs(css);
2032 
2033 	mutex_lock(&cpuset_mutex);
2034 
2035 	if (is_sched_load_balance(cs))
2036 		update_flag(CS_SCHED_LOAD_BALANCE, cs, 0);
2037 
2038 	cpuset_dec();
2039 	clear_bit(CS_ONLINE, &cs->flags);
2040 
2041 	mutex_unlock(&cpuset_mutex);
2042 }
2043 
2044 static void cpuset_css_free(struct cgroup_subsys_state *css)
2045 {
2046 	struct cpuset *cs = css_cs(css);
2047 
2048 	free_cpumask_var(cs->effective_cpus);
2049 	free_cpumask_var(cs->cpus_allowed);
2050 	kfree(cs);
2051 }
2052 
2053 static void cpuset_bind(struct cgroup_subsys_state *root_css)
2054 {
2055 	mutex_lock(&cpuset_mutex);
2056 	spin_lock_irq(&callback_lock);
2057 
2058 	if (cgroup_subsys_on_dfl(cpuset_cgrp_subsys)) {
2059 		cpumask_copy(top_cpuset.cpus_allowed, cpu_possible_mask);
2060 		top_cpuset.mems_allowed = node_possible_map;
2061 	} else {
2062 		cpumask_copy(top_cpuset.cpus_allowed,
2063 			     top_cpuset.effective_cpus);
2064 		top_cpuset.mems_allowed = top_cpuset.effective_mems;
2065 	}
2066 
2067 	spin_unlock_irq(&callback_lock);
2068 	mutex_unlock(&cpuset_mutex);
2069 }
2070 
2071 /*
2072  * Make sure the new task conform to the current state of its parent,
2073  * which could have been changed by cpuset just after it inherits the
2074  * state from the parent and before it sits on the cgroup's task list.
2075  */
2076 static void cpuset_fork(struct task_struct *task)
2077 {
2078 	if (task_css_is_root(task, cpuset_cgrp_id))
2079 		return;
2080 
2081 	set_cpus_allowed_ptr(task, &current->cpus_allowed);
2082 	task->mems_allowed = current->mems_allowed;
2083 }
2084 
2085 struct cgroup_subsys cpuset_cgrp_subsys = {
2086 	.css_alloc	= cpuset_css_alloc,
2087 	.css_online	= cpuset_css_online,
2088 	.css_offline	= cpuset_css_offline,
2089 	.css_free	= cpuset_css_free,
2090 	.can_attach	= cpuset_can_attach,
2091 	.cancel_attach	= cpuset_cancel_attach,
2092 	.attach		= cpuset_attach,
2093 	.post_attach	= cpuset_post_attach,
2094 	.bind		= cpuset_bind,
2095 	.fork		= cpuset_fork,
2096 	.legacy_cftypes	= files,
2097 	.early_init	= true,
2098 };
2099 
2100 /**
2101  * cpuset_init - initialize cpusets at system boot
2102  *
2103  * Description: Initialize top_cpuset and the cpuset internal file system,
2104  **/
2105 
2106 int __init cpuset_init(void)
2107 {
2108 	int err = 0;
2109 
2110 	BUG_ON(!alloc_cpumask_var(&top_cpuset.cpus_allowed, GFP_KERNEL));
2111 	BUG_ON(!alloc_cpumask_var(&top_cpuset.effective_cpus, GFP_KERNEL));
2112 
2113 	cpumask_setall(top_cpuset.cpus_allowed);
2114 	nodes_setall(top_cpuset.mems_allowed);
2115 	cpumask_setall(top_cpuset.effective_cpus);
2116 	nodes_setall(top_cpuset.effective_mems);
2117 
2118 	fmeter_init(&top_cpuset.fmeter);
2119 	set_bit(CS_SCHED_LOAD_BALANCE, &top_cpuset.flags);
2120 	top_cpuset.relax_domain_level = -1;
2121 
2122 	err = register_filesystem(&cpuset_fs_type);
2123 	if (err < 0)
2124 		return err;
2125 
2126 	BUG_ON(!alloc_cpumask_var(&cpus_attach, GFP_KERNEL));
2127 
2128 	return 0;
2129 }
2130 
2131 /*
2132  * If CPU and/or memory hotplug handlers, below, unplug any CPUs
2133  * or memory nodes, we need to walk over the cpuset hierarchy,
2134  * removing that CPU or node from all cpusets.  If this removes the
2135  * last CPU or node from a cpuset, then move the tasks in the empty
2136  * cpuset to its next-highest non-empty parent.
2137  */
2138 static void remove_tasks_in_empty_cpuset(struct cpuset *cs)
2139 {
2140 	struct cpuset *parent;
2141 
2142 	/*
2143 	 * Find its next-highest non-empty parent, (top cpuset
2144 	 * has online cpus, so can't be empty).
2145 	 */
2146 	parent = parent_cs(cs);
2147 	while (cpumask_empty(parent->cpus_allowed) ||
2148 			nodes_empty(parent->mems_allowed))
2149 		parent = parent_cs(parent);
2150 
2151 	if (cgroup_transfer_tasks(parent->css.cgroup, cs->css.cgroup)) {
2152 		pr_err("cpuset: failed to transfer tasks out of empty cpuset ");
2153 		pr_cont_cgroup_name(cs->css.cgroup);
2154 		pr_cont("\n");
2155 	}
2156 }
2157 
2158 static void
2159 hotplug_update_tasks_legacy(struct cpuset *cs,
2160 			    struct cpumask *new_cpus, nodemask_t *new_mems,
2161 			    bool cpus_updated, bool mems_updated)
2162 {
2163 	bool is_empty;
2164 
2165 	spin_lock_irq(&callback_lock);
2166 	cpumask_copy(cs->cpus_allowed, new_cpus);
2167 	cpumask_copy(cs->effective_cpus, new_cpus);
2168 	cs->mems_allowed = *new_mems;
2169 	cs->effective_mems = *new_mems;
2170 	spin_unlock_irq(&callback_lock);
2171 
2172 	/*
2173 	 * Don't call update_tasks_cpumask() if the cpuset becomes empty,
2174 	 * as the tasks will be migratecd to an ancestor.
2175 	 */
2176 	if (cpus_updated && !cpumask_empty(cs->cpus_allowed))
2177 		update_tasks_cpumask(cs);
2178 	if (mems_updated && !nodes_empty(cs->mems_allowed))
2179 		update_tasks_nodemask(cs);
2180 
2181 	is_empty = cpumask_empty(cs->cpus_allowed) ||
2182 		   nodes_empty(cs->mems_allowed);
2183 
2184 	mutex_unlock(&cpuset_mutex);
2185 
2186 	/*
2187 	 * Move tasks to the nearest ancestor with execution resources,
2188 	 * This is full cgroup operation which will also call back into
2189 	 * cpuset. Should be done outside any lock.
2190 	 */
2191 	if (is_empty)
2192 		remove_tasks_in_empty_cpuset(cs);
2193 
2194 	mutex_lock(&cpuset_mutex);
2195 }
2196 
2197 static void
2198 hotplug_update_tasks(struct cpuset *cs,
2199 		     struct cpumask *new_cpus, nodemask_t *new_mems,
2200 		     bool cpus_updated, bool mems_updated)
2201 {
2202 	if (cpumask_empty(new_cpus))
2203 		cpumask_copy(new_cpus, parent_cs(cs)->effective_cpus);
2204 	if (nodes_empty(*new_mems))
2205 		*new_mems = parent_cs(cs)->effective_mems;
2206 
2207 	spin_lock_irq(&callback_lock);
2208 	cpumask_copy(cs->effective_cpus, new_cpus);
2209 	cs->effective_mems = *new_mems;
2210 	spin_unlock_irq(&callback_lock);
2211 
2212 	if (cpus_updated)
2213 		update_tasks_cpumask(cs);
2214 	if (mems_updated)
2215 		update_tasks_nodemask(cs);
2216 }
2217 
2218 /**
2219  * cpuset_hotplug_update_tasks - update tasks in a cpuset for hotunplug
2220  * @cs: cpuset in interest
2221  *
2222  * Compare @cs's cpu and mem masks against top_cpuset and if some have gone
2223  * offline, update @cs accordingly.  If @cs ends up with no CPU or memory,
2224  * all its tasks are moved to the nearest ancestor with both resources.
2225  */
2226 static void cpuset_hotplug_update_tasks(struct cpuset *cs)
2227 {
2228 	static cpumask_t new_cpus;
2229 	static nodemask_t new_mems;
2230 	bool cpus_updated;
2231 	bool mems_updated;
2232 retry:
2233 	wait_event(cpuset_attach_wq, cs->attach_in_progress == 0);
2234 
2235 	mutex_lock(&cpuset_mutex);
2236 
2237 	/*
2238 	 * We have raced with task attaching. We wait until attaching
2239 	 * is finished, so we won't attach a task to an empty cpuset.
2240 	 */
2241 	if (cs->attach_in_progress) {
2242 		mutex_unlock(&cpuset_mutex);
2243 		goto retry;
2244 	}
2245 
2246 	cpumask_and(&new_cpus, cs->cpus_allowed, parent_cs(cs)->effective_cpus);
2247 	nodes_and(new_mems, cs->mems_allowed, parent_cs(cs)->effective_mems);
2248 
2249 	cpus_updated = !cpumask_equal(&new_cpus, cs->effective_cpus);
2250 	mems_updated = !nodes_equal(new_mems, cs->effective_mems);
2251 
2252 	if (cgroup_subsys_on_dfl(cpuset_cgrp_subsys))
2253 		hotplug_update_tasks(cs, &new_cpus, &new_mems,
2254 				     cpus_updated, mems_updated);
2255 	else
2256 		hotplug_update_tasks_legacy(cs, &new_cpus, &new_mems,
2257 					    cpus_updated, mems_updated);
2258 
2259 	mutex_unlock(&cpuset_mutex);
2260 }
2261 
2262 /**
2263  * cpuset_hotplug_workfn - handle CPU/memory hotunplug for a cpuset
2264  *
2265  * This function is called after either CPU or memory configuration has
2266  * changed and updates cpuset accordingly.  The top_cpuset is always
2267  * synchronized to cpu_active_mask and N_MEMORY, which is necessary in
2268  * order to make cpusets transparent (of no affect) on systems that are
2269  * actively using CPU hotplug but making no active use of cpusets.
2270  *
2271  * Non-root cpusets are only affected by offlining.  If any CPUs or memory
2272  * nodes have been taken down, cpuset_hotplug_update_tasks() is invoked on
2273  * all descendants.
2274  *
2275  * Note that CPU offlining during suspend is ignored.  We don't modify
2276  * cpusets across suspend/resume cycles at all.
2277  */
2278 static void cpuset_hotplug_workfn(struct work_struct *work)
2279 {
2280 	static cpumask_t new_cpus;
2281 	static nodemask_t new_mems;
2282 	bool cpus_updated, mems_updated;
2283 	bool on_dfl = cgroup_subsys_on_dfl(cpuset_cgrp_subsys);
2284 
2285 	mutex_lock(&cpuset_mutex);
2286 
2287 	/* fetch the available cpus/mems and find out which changed how */
2288 	cpumask_copy(&new_cpus, cpu_active_mask);
2289 	new_mems = node_states[N_MEMORY];
2290 
2291 	cpus_updated = !cpumask_equal(top_cpuset.effective_cpus, &new_cpus);
2292 	mems_updated = !nodes_equal(top_cpuset.effective_mems, new_mems);
2293 
2294 	/* synchronize cpus_allowed to cpu_active_mask */
2295 	if (cpus_updated) {
2296 		spin_lock_irq(&callback_lock);
2297 		if (!on_dfl)
2298 			cpumask_copy(top_cpuset.cpus_allowed, &new_cpus);
2299 		cpumask_copy(top_cpuset.effective_cpus, &new_cpus);
2300 		spin_unlock_irq(&callback_lock);
2301 		/* we don't mess with cpumasks of tasks in top_cpuset */
2302 	}
2303 
2304 	/* synchronize mems_allowed to N_MEMORY */
2305 	if (mems_updated) {
2306 		spin_lock_irq(&callback_lock);
2307 		if (!on_dfl)
2308 			top_cpuset.mems_allowed = new_mems;
2309 		top_cpuset.effective_mems = new_mems;
2310 		spin_unlock_irq(&callback_lock);
2311 		update_tasks_nodemask(&top_cpuset);
2312 	}
2313 
2314 	mutex_unlock(&cpuset_mutex);
2315 
2316 	/* if cpus or mems changed, we need to propagate to descendants */
2317 	if (cpus_updated || mems_updated) {
2318 		struct cpuset *cs;
2319 		struct cgroup_subsys_state *pos_css;
2320 
2321 		rcu_read_lock();
2322 		cpuset_for_each_descendant_pre(cs, pos_css, &top_cpuset) {
2323 			if (cs == &top_cpuset || !css_tryget_online(&cs->css))
2324 				continue;
2325 			rcu_read_unlock();
2326 
2327 			cpuset_hotplug_update_tasks(cs);
2328 
2329 			rcu_read_lock();
2330 			css_put(&cs->css);
2331 		}
2332 		rcu_read_unlock();
2333 	}
2334 
2335 	/* rebuild sched domains if cpus_allowed has changed */
2336 	if (cpus_updated)
2337 		rebuild_sched_domains();
2338 }
2339 
2340 void cpuset_update_active_cpus(void)
2341 {
2342 	/*
2343 	 * We're inside cpu hotplug critical region which usually nests
2344 	 * inside cgroup synchronization.  Bounce actual hotplug processing
2345 	 * to a work item to avoid reverse locking order.
2346 	 *
2347 	 * We still need to do partition_sched_domains() synchronously;
2348 	 * otherwise, the scheduler will get confused and put tasks to the
2349 	 * dead CPU.  Fall back to the default single domain.
2350 	 * cpuset_hotplug_workfn() will rebuild it as necessary.
2351 	 */
2352 	partition_sched_domains(1, NULL, NULL);
2353 	schedule_work(&cpuset_hotplug_work);
2354 }
2355 
2356 /*
2357  * Keep top_cpuset.mems_allowed tracking node_states[N_MEMORY].
2358  * Call this routine anytime after node_states[N_MEMORY] changes.
2359  * See cpuset_update_active_cpus() for CPU hotplug handling.
2360  */
2361 static int cpuset_track_online_nodes(struct notifier_block *self,
2362 				unsigned long action, void *arg)
2363 {
2364 	schedule_work(&cpuset_hotplug_work);
2365 	return NOTIFY_OK;
2366 }
2367 
2368 static struct notifier_block cpuset_track_online_nodes_nb = {
2369 	.notifier_call = cpuset_track_online_nodes,
2370 	.priority = 10,		/* ??! */
2371 };
2372 
2373 /**
2374  * cpuset_init_smp - initialize cpus_allowed
2375  *
2376  * Description: Finish top cpuset after cpu, node maps are initialized
2377  */
2378 void __init cpuset_init_smp(void)
2379 {
2380 	cpumask_copy(top_cpuset.cpus_allowed, cpu_active_mask);
2381 	top_cpuset.mems_allowed = node_states[N_MEMORY];
2382 	top_cpuset.old_mems_allowed = top_cpuset.mems_allowed;
2383 
2384 	cpumask_copy(top_cpuset.effective_cpus, cpu_active_mask);
2385 	top_cpuset.effective_mems = node_states[N_MEMORY];
2386 
2387 	register_hotmemory_notifier(&cpuset_track_online_nodes_nb);
2388 
2389 	cpuset_migrate_mm_wq = alloc_ordered_workqueue("cpuset_migrate_mm", 0);
2390 	BUG_ON(!cpuset_migrate_mm_wq);
2391 }
2392 
2393 /**
2394  * cpuset_cpus_allowed - return cpus_allowed mask from a tasks cpuset.
2395  * @tsk: pointer to task_struct from which to obtain cpuset->cpus_allowed.
2396  * @pmask: pointer to struct cpumask variable to receive cpus_allowed set.
2397  *
2398  * Description: Returns the cpumask_var_t cpus_allowed of the cpuset
2399  * attached to the specified @tsk.  Guaranteed to return some non-empty
2400  * subset of cpu_online_mask, even if this means going outside the
2401  * tasks cpuset.
2402  **/
2403 
2404 void cpuset_cpus_allowed(struct task_struct *tsk, struct cpumask *pmask)
2405 {
2406 	unsigned long flags;
2407 
2408 	spin_lock_irqsave(&callback_lock, flags);
2409 	rcu_read_lock();
2410 	guarantee_online_cpus(task_cs(tsk), pmask);
2411 	rcu_read_unlock();
2412 	spin_unlock_irqrestore(&callback_lock, flags);
2413 }
2414 
2415 void cpuset_cpus_allowed_fallback(struct task_struct *tsk)
2416 {
2417 	rcu_read_lock();
2418 	do_set_cpus_allowed(tsk, task_cs(tsk)->effective_cpus);
2419 	rcu_read_unlock();
2420 
2421 	/*
2422 	 * We own tsk->cpus_allowed, nobody can change it under us.
2423 	 *
2424 	 * But we used cs && cs->cpus_allowed lockless and thus can
2425 	 * race with cgroup_attach_task() or update_cpumask() and get
2426 	 * the wrong tsk->cpus_allowed. However, both cases imply the
2427 	 * subsequent cpuset_change_cpumask()->set_cpus_allowed_ptr()
2428 	 * which takes task_rq_lock().
2429 	 *
2430 	 * If we are called after it dropped the lock we must see all
2431 	 * changes in tsk_cs()->cpus_allowed. Otherwise we can temporary
2432 	 * set any mask even if it is not right from task_cs() pov,
2433 	 * the pending set_cpus_allowed_ptr() will fix things.
2434 	 *
2435 	 * select_fallback_rq() will fix things ups and set cpu_possible_mask
2436 	 * if required.
2437 	 */
2438 }
2439 
2440 void __init cpuset_init_current_mems_allowed(void)
2441 {
2442 	nodes_setall(current->mems_allowed);
2443 }
2444 
2445 /**
2446  * cpuset_mems_allowed - return mems_allowed mask from a tasks cpuset.
2447  * @tsk: pointer to task_struct from which to obtain cpuset->mems_allowed.
2448  *
2449  * Description: Returns the nodemask_t mems_allowed of the cpuset
2450  * attached to the specified @tsk.  Guaranteed to return some non-empty
2451  * subset of node_states[N_MEMORY], even if this means going outside the
2452  * tasks cpuset.
2453  **/
2454 
2455 nodemask_t cpuset_mems_allowed(struct task_struct *tsk)
2456 {
2457 	nodemask_t mask;
2458 	unsigned long flags;
2459 
2460 	spin_lock_irqsave(&callback_lock, flags);
2461 	rcu_read_lock();
2462 	guarantee_online_mems(task_cs(tsk), &mask);
2463 	rcu_read_unlock();
2464 	spin_unlock_irqrestore(&callback_lock, flags);
2465 
2466 	return mask;
2467 }
2468 
2469 /**
2470  * cpuset_nodemask_valid_mems_allowed - check nodemask vs. curremt mems_allowed
2471  * @nodemask: the nodemask to be checked
2472  *
2473  * Are any of the nodes in the nodemask allowed in current->mems_allowed?
2474  */
2475 int cpuset_nodemask_valid_mems_allowed(nodemask_t *nodemask)
2476 {
2477 	return nodes_intersects(*nodemask, current->mems_allowed);
2478 }
2479 
2480 /*
2481  * nearest_hardwall_ancestor() - Returns the nearest mem_exclusive or
2482  * mem_hardwall ancestor to the specified cpuset.  Call holding
2483  * callback_lock.  If no ancestor is mem_exclusive or mem_hardwall
2484  * (an unusual configuration), then returns the root cpuset.
2485  */
2486 static struct cpuset *nearest_hardwall_ancestor(struct cpuset *cs)
2487 {
2488 	while (!(is_mem_exclusive(cs) || is_mem_hardwall(cs)) && parent_cs(cs))
2489 		cs = parent_cs(cs);
2490 	return cs;
2491 }
2492 
2493 /**
2494  * cpuset_node_allowed - Can we allocate on a memory node?
2495  * @node: is this an allowed node?
2496  * @gfp_mask: memory allocation flags
2497  *
2498  * If we're in interrupt, yes, we can always allocate.  If @node is set in
2499  * current's mems_allowed, yes.  If it's not a __GFP_HARDWALL request and this
2500  * node is set in the nearest hardwalled cpuset ancestor to current's cpuset,
2501  * yes.  If current has access to memory reserves due to TIF_MEMDIE, yes.
2502  * Otherwise, no.
2503  *
2504  * GFP_USER allocations are marked with the __GFP_HARDWALL bit,
2505  * and do not allow allocations outside the current tasks cpuset
2506  * unless the task has been OOM killed as is marked TIF_MEMDIE.
2507  * GFP_KERNEL allocations are not so marked, so can escape to the
2508  * nearest enclosing hardwalled ancestor cpuset.
2509  *
2510  * Scanning up parent cpusets requires callback_lock.  The
2511  * __alloc_pages() routine only calls here with __GFP_HARDWALL bit
2512  * _not_ set if it's a GFP_KERNEL allocation, and all nodes in the
2513  * current tasks mems_allowed came up empty on the first pass over
2514  * the zonelist.  So only GFP_KERNEL allocations, if all nodes in the
2515  * cpuset are short of memory, might require taking the callback_lock.
2516  *
2517  * The first call here from mm/page_alloc:get_page_from_freelist()
2518  * has __GFP_HARDWALL set in gfp_mask, enforcing hardwall cpusets,
2519  * so no allocation on a node outside the cpuset is allowed (unless
2520  * in interrupt, of course).
2521  *
2522  * The second pass through get_page_from_freelist() doesn't even call
2523  * here for GFP_ATOMIC calls.  For those calls, the __alloc_pages()
2524  * variable 'wait' is not set, and the bit ALLOC_CPUSET is not set
2525  * in alloc_flags.  That logic and the checks below have the combined
2526  * affect that:
2527  *	in_interrupt - any node ok (current task context irrelevant)
2528  *	GFP_ATOMIC   - any node ok
2529  *	TIF_MEMDIE   - any node ok
2530  *	GFP_KERNEL   - any node in enclosing hardwalled cpuset ok
2531  *	GFP_USER     - only nodes in current tasks mems allowed ok.
2532  */
2533 bool __cpuset_node_allowed(int node, gfp_t gfp_mask)
2534 {
2535 	struct cpuset *cs;		/* current cpuset ancestors */
2536 	int allowed;			/* is allocation in zone z allowed? */
2537 	unsigned long flags;
2538 
2539 	if (in_interrupt())
2540 		return true;
2541 	if (node_isset(node, current->mems_allowed))
2542 		return true;
2543 	/*
2544 	 * Allow tasks that have access to memory reserves because they have
2545 	 * been OOM killed to get memory anywhere.
2546 	 */
2547 	if (unlikely(test_thread_flag(TIF_MEMDIE)))
2548 		return true;
2549 	if (gfp_mask & __GFP_HARDWALL)	/* If hardwall request, stop here */
2550 		return false;
2551 
2552 	if (current->flags & PF_EXITING) /* Let dying task have memory */
2553 		return true;
2554 
2555 	/* Not hardwall and node outside mems_allowed: scan up cpusets */
2556 	spin_lock_irqsave(&callback_lock, flags);
2557 
2558 	rcu_read_lock();
2559 	cs = nearest_hardwall_ancestor(task_cs(current));
2560 	allowed = node_isset(node, cs->mems_allowed);
2561 	rcu_read_unlock();
2562 
2563 	spin_unlock_irqrestore(&callback_lock, flags);
2564 	return allowed;
2565 }
2566 
2567 /**
2568  * cpuset_mem_spread_node() - On which node to begin search for a file page
2569  * cpuset_slab_spread_node() - On which node to begin search for a slab page
2570  *
2571  * If a task is marked PF_SPREAD_PAGE or PF_SPREAD_SLAB (as for
2572  * tasks in a cpuset with is_spread_page or is_spread_slab set),
2573  * and if the memory allocation used cpuset_mem_spread_node()
2574  * to determine on which node to start looking, as it will for
2575  * certain page cache or slab cache pages such as used for file
2576  * system buffers and inode caches, then instead of starting on the
2577  * local node to look for a free page, rather spread the starting
2578  * node around the tasks mems_allowed nodes.
2579  *
2580  * We don't have to worry about the returned node being offline
2581  * because "it can't happen", and even if it did, it would be ok.
2582  *
2583  * The routines calling guarantee_online_mems() are careful to
2584  * only set nodes in task->mems_allowed that are online.  So it
2585  * should not be possible for the following code to return an
2586  * offline node.  But if it did, that would be ok, as this routine
2587  * is not returning the node where the allocation must be, only
2588  * the node where the search should start.  The zonelist passed to
2589  * __alloc_pages() will include all nodes.  If the slab allocator
2590  * is passed an offline node, it will fall back to the local node.
2591  * See kmem_cache_alloc_node().
2592  */
2593 
2594 static int cpuset_spread_node(int *rotor)
2595 {
2596 	return *rotor = next_node_in(*rotor, current->mems_allowed);
2597 }
2598 
2599 int cpuset_mem_spread_node(void)
2600 {
2601 	if (current->cpuset_mem_spread_rotor == NUMA_NO_NODE)
2602 		current->cpuset_mem_spread_rotor =
2603 			node_random(&current->mems_allowed);
2604 
2605 	return cpuset_spread_node(&current->cpuset_mem_spread_rotor);
2606 }
2607 
2608 int cpuset_slab_spread_node(void)
2609 {
2610 	if (current->cpuset_slab_spread_rotor == NUMA_NO_NODE)
2611 		current->cpuset_slab_spread_rotor =
2612 			node_random(&current->mems_allowed);
2613 
2614 	return cpuset_spread_node(&current->cpuset_slab_spread_rotor);
2615 }
2616 
2617 EXPORT_SYMBOL_GPL(cpuset_mem_spread_node);
2618 
2619 /**
2620  * cpuset_mems_allowed_intersects - Does @tsk1's mems_allowed intersect @tsk2's?
2621  * @tsk1: pointer to task_struct of some task.
2622  * @tsk2: pointer to task_struct of some other task.
2623  *
2624  * Description: Return true if @tsk1's mems_allowed intersects the
2625  * mems_allowed of @tsk2.  Used by the OOM killer to determine if
2626  * one of the task's memory usage might impact the memory available
2627  * to the other.
2628  **/
2629 
2630 int cpuset_mems_allowed_intersects(const struct task_struct *tsk1,
2631 				   const struct task_struct *tsk2)
2632 {
2633 	return nodes_intersects(tsk1->mems_allowed, tsk2->mems_allowed);
2634 }
2635 
2636 /**
2637  * cpuset_print_current_mems_allowed - prints current's cpuset and mems_allowed
2638  *
2639  * Description: Prints current's name, cpuset name, and cached copy of its
2640  * mems_allowed to the kernel log.
2641  */
2642 void cpuset_print_current_mems_allowed(void)
2643 {
2644 	struct cgroup *cgrp;
2645 
2646 	rcu_read_lock();
2647 
2648 	cgrp = task_cs(current)->css.cgroup;
2649 	pr_info("%s cpuset=", current->comm);
2650 	pr_cont_cgroup_name(cgrp);
2651 	pr_cont(" mems_allowed=%*pbl\n",
2652 		nodemask_pr_args(&current->mems_allowed));
2653 
2654 	rcu_read_unlock();
2655 }
2656 
2657 /*
2658  * Collection of memory_pressure is suppressed unless
2659  * this flag is enabled by writing "1" to the special
2660  * cpuset file 'memory_pressure_enabled' in the root cpuset.
2661  */
2662 
2663 int cpuset_memory_pressure_enabled __read_mostly;
2664 
2665 /**
2666  * cpuset_memory_pressure_bump - keep stats of per-cpuset reclaims.
2667  *
2668  * Keep a running average of the rate of synchronous (direct)
2669  * page reclaim efforts initiated by tasks in each cpuset.
2670  *
2671  * This represents the rate at which some task in the cpuset
2672  * ran low on memory on all nodes it was allowed to use, and
2673  * had to enter the kernels page reclaim code in an effort to
2674  * create more free memory by tossing clean pages or swapping
2675  * or writing dirty pages.
2676  *
2677  * Display to user space in the per-cpuset read-only file
2678  * "memory_pressure".  Value displayed is an integer
2679  * representing the recent rate of entry into the synchronous
2680  * (direct) page reclaim by any task attached to the cpuset.
2681  **/
2682 
2683 void __cpuset_memory_pressure_bump(void)
2684 {
2685 	rcu_read_lock();
2686 	fmeter_markevent(&task_cs(current)->fmeter);
2687 	rcu_read_unlock();
2688 }
2689 
2690 #ifdef CONFIG_PROC_PID_CPUSET
2691 /*
2692  * proc_cpuset_show()
2693  *  - Print tasks cpuset path into seq_file.
2694  *  - Used for /proc/<pid>/cpuset.
2695  *  - No need to task_lock(tsk) on this tsk->cpuset reference, as it
2696  *    doesn't really matter if tsk->cpuset changes after we read it,
2697  *    and we take cpuset_mutex, keeping cpuset_attach() from changing it
2698  *    anyway.
2699  */
2700 int proc_cpuset_show(struct seq_file *m, struct pid_namespace *ns,
2701 		     struct pid *pid, struct task_struct *tsk)
2702 {
2703 	char *buf;
2704 	struct cgroup_subsys_state *css;
2705 	int retval;
2706 
2707 	retval = -ENOMEM;
2708 	buf = kmalloc(PATH_MAX, GFP_KERNEL);
2709 	if (!buf)
2710 		goto out;
2711 
2712 	css = task_get_css(tsk, cpuset_cgrp_id);
2713 	retval = cgroup_path_ns(css->cgroup, buf, PATH_MAX,
2714 				current->nsproxy->cgroup_ns);
2715 	css_put(css);
2716 	if (retval >= PATH_MAX)
2717 		retval = -ENAMETOOLONG;
2718 	if (retval < 0)
2719 		goto out_free;
2720 	seq_puts(m, buf);
2721 	seq_putc(m, '\n');
2722 	retval = 0;
2723 out_free:
2724 	kfree(buf);
2725 out:
2726 	return retval;
2727 }
2728 #endif /* CONFIG_PROC_PID_CPUSET */
2729 
2730 /* Display task mems_allowed in /proc/<pid>/status file. */
2731 void cpuset_task_status_allowed(struct seq_file *m, struct task_struct *task)
2732 {
2733 	seq_printf(m, "Mems_allowed:\t%*pb\n",
2734 		   nodemask_pr_args(&task->mems_allowed));
2735 	seq_printf(m, "Mems_allowed_list:\t%*pbl\n",
2736 		   nodemask_pr_args(&task->mems_allowed));
2737 }
2738